1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 2012 - Virtual Open Systems and Columbia University
4 * Author: Christoffer Dall <c.dall@virtualopensystems.com>
7 #include <linux/mman.h>
8 #include <linux/kvm_host.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>
25 static struct kvm_pgtable *hyp_pgtable;
26 static DEFINE_MUTEX(kvm_hyp_pgd_mutex);
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;
32 static unsigned long __ro_after_init io_map_base;
34 static phys_addr_t __stage2_range_addr_end(phys_addr_t addr, phys_addr_t end,
37 phys_addr_t boundary = ALIGN_DOWN(addr + size, size);
39 return (boundary - 1 < end - 1) ? boundary : end;
42 static phys_addr_t stage2_range_addr_end(phys_addr_t addr, phys_addr_t end)
44 phys_addr_t size = kvm_granule_size(KVM_PGTABLE_MIN_BLOCK_LEVEL);
46 return __stage2_range_addr_end(addr, end, size);
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.
56 static int stage2_apply_range(struct kvm_s2_mmu *mmu, phys_addr_t addr,
58 int (*fn)(struct kvm_pgtable *, u64, u64),
61 struct kvm *kvm = kvm_s2_mmu_to_kvm(mmu);
66 struct kvm_pgtable *pgt = mmu->pgt;
70 next = stage2_range_addr_end(addr, end);
71 ret = fn(pgt, addr, next - addr);
75 if (resched && next != end)
76 cond_resched_rwlock_write(&kvm->mmu_lock);
77 } while (addr = next, addr != end);
82 #define stage2_apply_range_resched(mmu, addr, end, fn) \
83 stage2_apply_range(mmu, addr, end, fn, true)
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.
90 static int kvm_mmu_split_nr_page_tables(u64 range)
94 if (KVM_PGTABLE_MIN_BLOCK_LEVEL < 2)
95 n += DIV_ROUND_UP(range, PUD_SIZE);
96 n += DIV_ROUND_UP(range, PMD_SIZE);
100 static bool need_split_memcache_topup_or_resched(struct kvm *kvm)
102 struct kvm_mmu_memory_cache *cache;
105 if (need_resched() || rwlock_needbreak(&kvm->mmu_lock))
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;
114 static int kvm_mmu_split_huge_pages(struct kvm *kvm, phys_addr_t addr,
117 struct kvm_mmu_memory_cache *cache;
118 struct kvm_pgtable *pgt;
119 int ret, cache_capacity;
120 u64 next, chunk_size;
122 lockdep_assert_held_write(&kvm->mmu_lock);
124 chunk_size = kvm->arch.mmu.split_page_chunk_size;
125 cache_capacity = kvm_mmu_split_nr_page_tables(chunk_size);
130 cache = &kvm->arch.mmu.split_page_cache;
133 if (need_split_memcache_topup_or_resched(kvm)) {
134 write_unlock(&kvm->mmu_lock);
136 /* Eager page splitting is best-effort. */
137 ret = __kvm_mmu_topup_memory_cache(cache,
140 write_lock(&kvm->mmu_lock);
145 pgt = kvm->arch.mmu.pgt;
149 next = __stage2_range_addr_end(addr, end, chunk_size);
150 ret = kvm_pgtable_stage2_split(pgt, addr, next - addr, cache);
153 } while (addr = next, addr != end);
158 static bool memslot_is_logging(struct kvm_memory_slot *memslot)
160 return memslot->dirty_bitmap && !(memslot->flags & KVM_MEM_READONLY);
164 * kvm_flush_remote_tlbs() - flush all VM TLB entries for v7/8
165 * @kvm: pointer to kvm structure.
167 * Interface to HYP function to flush all VM TLB entries
169 void kvm_flush_remote_tlbs(struct kvm *kvm)
171 ++kvm->stat.generic.remote_tlb_flush_requests;
172 kvm_call_hyp(__kvm_tlb_flush_vmid, &kvm->arch.mmu);
175 static bool kvm_is_device_pfn(unsigned long pfn)
177 return !pfn_is_map_memory(pfn);
180 static void *stage2_memcache_zalloc_page(void *arg)
182 struct kvm_mmu_memory_cache *mc = arg;
185 /* Allocated with __GFP_ZERO, so no need to zero */
186 virt = kvm_mmu_memory_cache_alloc(mc);
188 kvm_account_pgtable_pages(virt, 1);
192 static void *kvm_host_zalloc_pages_exact(size_t size)
194 return alloc_pages_exact(size, GFP_KERNEL_ACCOUNT | __GFP_ZERO);
197 static void *kvm_s2_zalloc_pages_exact(size_t size)
199 void *virt = kvm_host_zalloc_pages_exact(size);
202 kvm_account_pgtable_pages(virt, (size >> PAGE_SHIFT));
206 static void kvm_s2_free_pages_exact(void *virt, size_t size)
208 kvm_account_pgtable_pages(virt, -(size >> PAGE_SHIFT));
209 free_pages_exact(virt, size);
212 static struct kvm_pgtable_mm_ops kvm_s2_mm_ops;
214 static void stage2_free_unlinked_table_rcu_cb(struct rcu_head *head)
216 struct page *page = container_of(head, struct page, rcu_head);
217 void *pgtable = page_to_virt(page);
218 u32 level = page_private(page);
220 kvm_pgtable_stage2_free_unlinked(&kvm_s2_mm_ops, pgtable, level);
223 static void stage2_free_unlinked_table(void *addr, u32 level)
225 struct page *page = virt_to_page(addr);
227 set_page_private(page, (unsigned long)level);
228 call_rcu(&page->rcu_head, stage2_free_unlinked_table_rcu_cb);
231 static void kvm_host_get_page(void *addr)
233 get_page(virt_to_page(addr));
236 static void kvm_host_put_page(void *addr)
238 put_page(virt_to_page(addr));
241 static void kvm_s2_put_page(void *addr)
243 struct page *p = virt_to_page(addr);
244 /* Dropping last refcount, the page will be freed */
245 if (page_count(p) == 1)
246 kvm_account_pgtable_pages(addr, -1);
250 static int kvm_host_page_count(void *addr)
252 return page_count(virt_to_page(addr));
255 static phys_addr_t kvm_host_pa(void *addr)
260 static void *kvm_host_va(phys_addr_t phys)
265 static void clean_dcache_guest_page(void *va, size_t size)
267 __clean_dcache_guest_page(va, size);
270 static void invalidate_icache_guest_page(void *va, size_t size)
272 __invalidate_icache_guest_page(va, size);
276 * Unmapping vs dcache management:
278 * If a guest maps certain memory pages as uncached, all writes will
279 * bypass the data cache and go directly to RAM. However, the CPUs
280 * can still speculate reads (not writes) and fill cache lines with
283 * Those cache lines will be *clean* cache lines though, so a
284 * clean+invalidate operation is equivalent to an invalidate
285 * operation, because no cache lines are marked dirty.
287 * Those clean cache lines could be filled prior to an uncached write
288 * by the guest, and the cache coherent IO subsystem would therefore
289 * end up writing old data to disk.
291 * This is why right after unmapping a page/section and invalidating
292 * the corresponding TLBs, we flush to make sure the IO subsystem will
293 * never hit in the cache.
295 * This is all avoided on systems that have ARM64_HAS_STAGE2_FWB, as
296 * we then fully enforce cacheability of RAM, no matter what the guest
300 * unmap_stage2_range -- Clear stage2 page table entries to unmap a range
301 * @mmu: The KVM stage-2 MMU pointer
302 * @start: The intermediate physical base address of the range to unmap
303 * @size: The size of the area to unmap
304 * @may_block: Whether or not we are permitted to block
306 * Clear a range of stage-2 mappings, lowering the various ref-counts. Must
307 * be called while holding mmu_lock (unless for freeing the stage2 pgd before
308 * destroying the VM), otherwise another faulting VCPU may come in and mess
309 * with things behind our backs.
311 static void __unmap_stage2_range(struct kvm_s2_mmu *mmu, phys_addr_t start, u64 size,
314 struct kvm *kvm = kvm_s2_mmu_to_kvm(mmu);
315 phys_addr_t end = start + size;
317 lockdep_assert_held_write(&kvm->mmu_lock);
318 WARN_ON(size & ~PAGE_MASK);
319 WARN_ON(stage2_apply_range(mmu, start, end, kvm_pgtable_stage2_unmap,
323 static void unmap_stage2_range(struct kvm_s2_mmu *mmu, phys_addr_t start, u64 size)
325 __unmap_stage2_range(mmu, start, size, true);
328 static void stage2_flush_memslot(struct kvm *kvm,
329 struct kvm_memory_slot *memslot)
331 phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
332 phys_addr_t end = addr + PAGE_SIZE * memslot->npages;
334 stage2_apply_range_resched(&kvm->arch.mmu, addr, end, kvm_pgtable_stage2_flush);
338 * stage2_flush_vm - Invalidate cache for pages mapped in stage 2
339 * @kvm: The struct kvm pointer
341 * Go through the stage 2 page tables and invalidate any cache lines
342 * backing memory already mapped to the VM.
344 static void stage2_flush_vm(struct kvm *kvm)
346 struct kvm_memslots *slots;
347 struct kvm_memory_slot *memslot;
350 idx = srcu_read_lock(&kvm->srcu);
351 write_lock(&kvm->mmu_lock);
353 slots = kvm_memslots(kvm);
354 kvm_for_each_memslot(memslot, bkt, slots)
355 stage2_flush_memslot(kvm, memslot);
357 write_unlock(&kvm->mmu_lock);
358 srcu_read_unlock(&kvm->srcu, idx);
362 * free_hyp_pgds - free Hyp-mode page tables
364 void __init free_hyp_pgds(void)
366 mutex_lock(&kvm_hyp_pgd_mutex);
368 kvm_pgtable_hyp_destroy(hyp_pgtable);
372 mutex_unlock(&kvm_hyp_pgd_mutex);
375 static bool kvm_host_owns_hyp_mappings(void)
377 if (is_kernel_in_hyp_mode())
380 if (static_branch_likely(&kvm_protected_mode_initialized))
384 * This can happen at boot time when __create_hyp_mappings() is called
385 * after the hyp protection has been enabled, but the static key has
386 * not been flipped yet.
388 if (!hyp_pgtable && is_protected_kvm_enabled())
391 WARN_ON(!hyp_pgtable);
396 int __create_hyp_mappings(unsigned long start, unsigned long size,
397 unsigned long phys, enum kvm_pgtable_prot prot)
401 if (WARN_ON(!kvm_host_owns_hyp_mappings()))
404 mutex_lock(&kvm_hyp_pgd_mutex);
405 err = kvm_pgtable_hyp_map(hyp_pgtable, start, size, phys, prot);
406 mutex_unlock(&kvm_hyp_pgd_mutex);
411 static phys_addr_t kvm_kaddr_to_phys(void *kaddr)
413 if (!is_vmalloc_addr(kaddr)) {
414 BUG_ON(!virt_addr_valid(kaddr));
417 return page_to_phys(vmalloc_to_page(kaddr)) +
418 offset_in_page(kaddr);
422 struct hyp_shared_pfn {
428 static DEFINE_MUTEX(hyp_shared_pfns_lock);
429 static struct rb_root hyp_shared_pfns = RB_ROOT;
431 static struct hyp_shared_pfn *find_shared_pfn(u64 pfn, struct rb_node ***node,
432 struct rb_node **parent)
434 struct hyp_shared_pfn *this;
436 *node = &hyp_shared_pfns.rb_node;
439 this = container_of(**node, struct hyp_shared_pfn, node);
442 *node = &((**node)->rb_left);
443 else if (this->pfn > pfn)
444 *node = &((**node)->rb_right);
452 static int share_pfn_hyp(u64 pfn)
454 struct rb_node **node, *parent;
455 struct hyp_shared_pfn *this;
458 mutex_lock(&hyp_shared_pfns_lock);
459 this = find_shared_pfn(pfn, &node, &parent);
465 this = kzalloc(sizeof(*this), GFP_KERNEL);
473 rb_link_node(&this->node, parent, node);
474 rb_insert_color(&this->node, &hyp_shared_pfns);
475 ret = kvm_call_hyp_nvhe(__pkvm_host_share_hyp, pfn, 1);
477 mutex_unlock(&hyp_shared_pfns_lock);
482 static int unshare_pfn_hyp(u64 pfn)
484 struct rb_node **node, *parent;
485 struct hyp_shared_pfn *this;
488 mutex_lock(&hyp_shared_pfns_lock);
489 this = find_shared_pfn(pfn, &node, &parent);
490 if (WARN_ON(!this)) {
499 rb_erase(&this->node, &hyp_shared_pfns);
501 ret = kvm_call_hyp_nvhe(__pkvm_host_unshare_hyp, pfn, 1);
503 mutex_unlock(&hyp_shared_pfns_lock);
508 int kvm_share_hyp(void *from, void *to)
510 phys_addr_t start, end, cur;
514 if (is_kernel_in_hyp_mode())
518 * The share hcall maps things in the 'fixed-offset' region of the hyp
519 * VA space, so we can only share physically contiguous data-structures
522 if (is_vmalloc_or_module_addr(from) || is_vmalloc_or_module_addr(to))
525 if (kvm_host_owns_hyp_mappings())
526 return create_hyp_mappings(from, to, PAGE_HYP);
528 start = ALIGN_DOWN(__pa(from), PAGE_SIZE);
529 end = PAGE_ALIGN(__pa(to));
530 for (cur = start; cur < end; cur += PAGE_SIZE) {
531 pfn = __phys_to_pfn(cur);
532 ret = share_pfn_hyp(pfn);
540 void kvm_unshare_hyp(void *from, void *to)
542 phys_addr_t start, end, cur;
545 if (is_kernel_in_hyp_mode() || kvm_host_owns_hyp_mappings() || !from)
548 start = ALIGN_DOWN(__pa(from), PAGE_SIZE);
549 end = PAGE_ALIGN(__pa(to));
550 for (cur = start; cur < end; cur += PAGE_SIZE) {
551 pfn = __phys_to_pfn(cur);
552 WARN_ON(unshare_pfn_hyp(pfn));
557 * create_hyp_mappings - duplicate a kernel virtual address range in Hyp mode
558 * @from: The virtual kernel start address of the range
559 * @to: The virtual kernel end address of the range (exclusive)
560 * @prot: The protection to be applied to this range
562 * The same virtual address as the kernel virtual address is also used
563 * in Hyp-mode mapping (modulo HYP_PAGE_OFFSET) to the same underlying
566 int create_hyp_mappings(void *from, void *to, enum kvm_pgtable_prot prot)
568 phys_addr_t phys_addr;
569 unsigned long virt_addr;
570 unsigned long start = kern_hyp_va((unsigned long)from);
571 unsigned long end = kern_hyp_va((unsigned long)to);
573 if (is_kernel_in_hyp_mode())
576 if (!kvm_host_owns_hyp_mappings())
579 start = start & PAGE_MASK;
580 end = PAGE_ALIGN(end);
582 for (virt_addr = start; virt_addr < end; virt_addr += PAGE_SIZE) {
585 phys_addr = kvm_kaddr_to_phys(from + virt_addr - start);
586 err = __create_hyp_mappings(virt_addr, PAGE_SIZE, phys_addr,
597 * hyp_alloc_private_va_range - Allocates a private VA range.
598 * @size: The size of the VA range to reserve.
599 * @haddr: The hypervisor virtual start address of the allocation.
601 * The private virtual address (VA) range is allocated below io_map_base
602 * and aligned based on the order of @size.
604 * Return: 0 on success or negative error code on failure.
606 int hyp_alloc_private_va_range(size_t size, unsigned long *haddr)
611 mutex_lock(&kvm_hyp_pgd_mutex);
614 * This assumes that we have enough space below the idmap
615 * page to allocate our VAs. If not, the check below will
616 * kick. A potential alternative would be to detect that
617 * overflow and switch to an allocation above the idmap.
619 * The allocated size is always a multiple of PAGE_SIZE.
621 base = io_map_base - PAGE_ALIGN(size);
623 /* Align the allocation based on the order of its size */
624 base = ALIGN_DOWN(base, PAGE_SIZE << get_order(size));
627 * Verify that BIT(VA_BITS - 1) hasn't been flipped by
628 * allocating the new area, as it would indicate we've
629 * overflowed the idmap/IO address range.
631 if ((base ^ io_map_base) & BIT(VA_BITS - 1))
634 *haddr = io_map_base = base;
636 mutex_unlock(&kvm_hyp_pgd_mutex);
641 static int __create_hyp_private_mapping(phys_addr_t phys_addr, size_t size,
642 unsigned long *haddr,
643 enum kvm_pgtable_prot prot)
648 if (!kvm_host_owns_hyp_mappings()) {
649 addr = kvm_call_hyp_nvhe(__pkvm_create_private_mapping,
650 phys_addr, size, prot);
651 if (IS_ERR_VALUE(addr))
658 size = PAGE_ALIGN(size + offset_in_page(phys_addr));
659 ret = hyp_alloc_private_va_range(size, &addr);
663 ret = __create_hyp_mappings(addr, size, phys_addr, prot);
667 *haddr = addr + offset_in_page(phys_addr);
672 * create_hyp_io_mappings - Map IO into both kernel and HYP
673 * @phys_addr: The physical start address which gets mapped
674 * @size: Size of the region being mapped
675 * @kaddr: Kernel VA for this mapping
676 * @haddr: HYP VA for this mapping
678 int create_hyp_io_mappings(phys_addr_t phys_addr, size_t size,
679 void __iomem **kaddr,
680 void __iomem **haddr)
685 if (is_protected_kvm_enabled())
688 *kaddr = ioremap(phys_addr, size);
692 if (is_kernel_in_hyp_mode()) {
697 ret = __create_hyp_private_mapping(phys_addr, size,
698 &addr, PAGE_HYP_DEVICE);
706 *haddr = (void __iomem *)addr;
711 * create_hyp_exec_mappings - Map an executable range into HYP
712 * @phys_addr: The physical start address which gets mapped
713 * @size: Size of the region being mapped
714 * @haddr: HYP VA for this mapping
716 int create_hyp_exec_mappings(phys_addr_t phys_addr, size_t size,
722 BUG_ON(is_kernel_in_hyp_mode());
724 ret = __create_hyp_private_mapping(phys_addr, size,
725 &addr, PAGE_HYP_EXEC);
731 *haddr = (void *)addr;
735 static struct kvm_pgtable_mm_ops kvm_user_mm_ops = {
736 /* We shouldn't need any other callback to walk the PT */
737 .phys_to_virt = kvm_host_va,
740 static int get_user_mapping_size(struct kvm *kvm, u64 addr)
742 struct kvm_pgtable pgt = {
743 .pgd = (kvm_pteref_t)kvm->mm->pgd,
744 .ia_bits = vabits_actual,
745 .start_level = (KVM_PGTABLE_MAX_LEVELS -
746 CONFIG_PGTABLE_LEVELS),
747 .mm_ops = &kvm_user_mm_ops,
750 kvm_pte_t pte = 0; /* Keep GCC quiet... */
755 * Disable IRQs so that we hazard against a concurrent
756 * teardown of the userspace page tables (which relies on
759 local_irq_save(flags);
760 ret = kvm_pgtable_get_leaf(&pgt, addr, &pte, &level);
761 local_irq_restore(flags);
767 * Not seeing an error, but not updating level? Something went
770 if (WARN_ON(level >= KVM_PGTABLE_MAX_LEVELS))
773 /* Oops, the userspace PTs are gone... Replay the fault */
774 if (!kvm_pte_valid(pte))
777 return BIT(ARM64_HW_PGTABLE_LEVEL_SHIFT(level));
780 static struct kvm_pgtable_mm_ops kvm_s2_mm_ops = {
781 .zalloc_page = stage2_memcache_zalloc_page,
782 .zalloc_pages_exact = kvm_s2_zalloc_pages_exact,
783 .free_pages_exact = kvm_s2_free_pages_exact,
784 .free_unlinked_table = stage2_free_unlinked_table,
785 .get_page = kvm_host_get_page,
786 .put_page = kvm_s2_put_page,
787 .page_count = kvm_host_page_count,
788 .phys_to_virt = kvm_host_va,
789 .virt_to_phys = kvm_host_pa,
790 .dcache_clean_inval_poc = clean_dcache_guest_page,
791 .icache_inval_pou = invalidate_icache_guest_page,
795 * kvm_init_stage2_mmu - Initialise a S2 MMU structure
796 * @kvm: The pointer to the KVM structure
797 * @mmu: The pointer to the s2 MMU structure
798 * @type: The machine type of the virtual machine
800 * Allocates only the stage-2 HW PGD level table(s).
801 * Note we don't need locking here as this is only called when the VM is
802 * created, which can only be done once.
804 int kvm_init_stage2_mmu(struct kvm *kvm, struct kvm_s2_mmu *mmu, unsigned long type)
806 u32 kvm_ipa_limit = get_kvm_ipa_limit();
808 struct kvm_pgtable *pgt;
812 if (type & ~KVM_VM_TYPE_ARM_IPA_SIZE_MASK)
815 phys_shift = KVM_VM_TYPE_ARM_IPA_SIZE(type);
816 if (is_protected_kvm_enabled()) {
817 phys_shift = kvm_ipa_limit;
818 } else if (phys_shift) {
819 if (phys_shift > kvm_ipa_limit ||
820 phys_shift < ARM64_MIN_PARANGE_BITS)
823 phys_shift = KVM_PHYS_SHIFT;
824 if (phys_shift > kvm_ipa_limit) {
825 pr_warn_once("%s using unsupported default IPA limit, upgrade your VMM\n",
831 mmfr0 = read_sanitised_ftr_reg(SYS_ID_AA64MMFR0_EL1);
832 mmfr1 = read_sanitised_ftr_reg(SYS_ID_AA64MMFR1_EL1);
833 kvm->arch.vtcr = kvm_get_vtcr(mmfr0, mmfr1, phys_shift);
835 if (mmu->pgt != NULL) {
836 kvm_err("kvm_arch already initialized?\n");
840 pgt = kzalloc(sizeof(*pgt), GFP_KERNEL_ACCOUNT);
844 mmu->arch = &kvm->arch;
845 err = kvm_pgtable_stage2_init(pgt, mmu, &kvm_s2_mm_ops);
847 goto out_free_pgtable;
849 mmu->last_vcpu_ran = alloc_percpu(typeof(*mmu->last_vcpu_ran));
850 if (!mmu->last_vcpu_ran) {
852 goto out_destroy_pgtable;
855 for_each_possible_cpu(cpu)
856 *per_cpu_ptr(mmu->last_vcpu_ran, cpu) = -1;
858 /* The eager page splitting is disabled by default */
859 mmu->split_page_chunk_size = KVM_ARM_EAGER_SPLIT_CHUNK_SIZE_DEFAULT;
860 mmu->split_page_cache.gfp_zero = __GFP_ZERO;
863 mmu->pgd_phys = __pa(pgt->pgd);
867 kvm_pgtable_stage2_destroy(pgt);
873 void kvm_uninit_stage2_mmu(struct kvm *kvm)
875 kvm_free_stage2_pgd(&kvm->arch.mmu);
876 kvm_mmu_free_memory_cache(&kvm->arch.mmu.split_page_cache);
879 static void stage2_unmap_memslot(struct kvm *kvm,
880 struct kvm_memory_slot *memslot)
882 hva_t hva = memslot->userspace_addr;
883 phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
884 phys_addr_t size = PAGE_SIZE * memslot->npages;
885 hva_t reg_end = hva + size;
888 * A memory region could potentially cover multiple VMAs, and any holes
889 * between them, so iterate over all of them to find out if we should
892 * +--------------------------------------------+
893 * +---------------+----------------+ +----------------+
894 * | : VMA 1 | VMA 2 | | VMA 3 : |
895 * +---------------+----------------+ +----------------+
897 * +--------------------------------------------+
900 struct vm_area_struct *vma;
901 hva_t vm_start, vm_end;
903 vma = find_vma_intersection(current->mm, hva, reg_end);
908 * Take the intersection of this VMA with the memory region
910 vm_start = max(hva, vma->vm_start);
911 vm_end = min(reg_end, vma->vm_end);
913 if (!(vma->vm_flags & VM_PFNMAP)) {
914 gpa_t gpa = addr + (vm_start - memslot->userspace_addr);
915 unmap_stage2_range(&kvm->arch.mmu, gpa, vm_end - vm_start);
918 } while (hva < reg_end);
922 * stage2_unmap_vm - Unmap Stage-2 RAM mappings
923 * @kvm: The struct kvm pointer
925 * Go through the memregions and unmap any regular RAM
926 * backing memory already mapped to the VM.
928 void stage2_unmap_vm(struct kvm *kvm)
930 struct kvm_memslots *slots;
931 struct kvm_memory_slot *memslot;
934 idx = srcu_read_lock(&kvm->srcu);
935 mmap_read_lock(current->mm);
936 write_lock(&kvm->mmu_lock);
938 slots = kvm_memslots(kvm);
939 kvm_for_each_memslot(memslot, bkt, slots)
940 stage2_unmap_memslot(kvm, memslot);
942 write_unlock(&kvm->mmu_lock);
943 mmap_read_unlock(current->mm);
944 srcu_read_unlock(&kvm->srcu, idx);
947 void kvm_free_stage2_pgd(struct kvm_s2_mmu *mmu)
949 struct kvm *kvm = kvm_s2_mmu_to_kvm(mmu);
950 struct kvm_pgtable *pgt = NULL;
952 write_lock(&kvm->mmu_lock);
957 free_percpu(mmu->last_vcpu_ran);
959 write_unlock(&kvm->mmu_lock);
962 kvm_pgtable_stage2_destroy(pgt);
967 static void hyp_mc_free_fn(void *addr, void *unused)
969 free_page((unsigned long)addr);
972 static void *hyp_mc_alloc_fn(void *unused)
974 return (void *)__get_free_page(GFP_KERNEL_ACCOUNT);
977 void free_hyp_memcache(struct kvm_hyp_memcache *mc)
979 if (is_protected_kvm_enabled())
980 __free_hyp_memcache(mc, hyp_mc_free_fn,
984 int topup_hyp_memcache(struct kvm_hyp_memcache *mc, unsigned long min_pages)
986 if (!is_protected_kvm_enabled())
989 return __topup_hyp_memcache(mc, min_pages, hyp_mc_alloc_fn,
994 * kvm_phys_addr_ioremap - map a device range to guest IPA
996 * @kvm: The KVM pointer
997 * @guest_ipa: The IPA at which to insert the mapping
998 * @pa: The physical address of the device
999 * @size: The size of the mapping
1000 * @writable: Whether or not to create a writable mapping
1002 int kvm_phys_addr_ioremap(struct kvm *kvm, phys_addr_t guest_ipa,
1003 phys_addr_t pa, unsigned long size, bool writable)
1007 struct kvm_mmu_memory_cache cache = { .gfp_zero = __GFP_ZERO };
1008 struct kvm_pgtable *pgt = kvm->arch.mmu.pgt;
1009 enum kvm_pgtable_prot prot = KVM_PGTABLE_PROT_DEVICE |
1010 KVM_PGTABLE_PROT_R |
1011 (writable ? KVM_PGTABLE_PROT_W : 0);
1013 if (is_protected_kvm_enabled())
1016 size += offset_in_page(guest_ipa);
1017 guest_ipa &= PAGE_MASK;
1019 for (addr = guest_ipa; addr < guest_ipa + size; addr += PAGE_SIZE) {
1020 ret = kvm_mmu_topup_memory_cache(&cache,
1021 kvm_mmu_cache_min_pages(kvm));
1025 write_lock(&kvm->mmu_lock);
1026 ret = kvm_pgtable_stage2_map(pgt, addr, PAGE_SIZE, pa, prot,
1028 write_unlock(&kvm->mmu_lock);
1035 kvm_mmu_free_memory_cache(&cache);
1040 * stage2_wp_range() - write protect stage2 memory region range
1041 * @mmu: The KVM stage-2 MMU pointer
1042 * @addr: Start address of range
1043 * @end: End address of range
1045 static void stage2_wp_range(struct kvm_s2_mmu *mmu, phys_addr_t addr, phys_addr_t end)
1047 stage2_apply_range_resched(mmu, addr, end, kvm_pgtable_stage2_wrprotect);
1051 * kvm_mmu_wp_memory_region() - write protect stage 2 entries for memory slot
1052 * @kvm: The KVM pointer
1053 * @slot: The memory slot to write protect
1055 * Called to start logging dirty pages after memory region
1056 * KVM_MEM_LOG_DIRTY_PAGES operation is called. After this function returns
1057 * all present PUD, PMD and PTEs are write protected in the memory region.
1058 * Afterwards read of dirty page log can be called.
1060 * Acquires kvm_mmu_lock. Called with kvm->slots_lock mutex acquired,
1061 * serializing operations for VM memory regions.
1063 static void kvm_mmu_wp_memory_region(struct kvm *kvm, int slot)
1065 struct kvm_memslots *slots = kvm_memslots(kvm);
1066 struct kvm_memory_slot *memslot = id_to_memslot(slots, slot);
1067 phys_addr_t start, end;
1069 if (WARN_ON_ONCE(!memslot))
1072 start = memslot->base_gfn << PAGE_SHIFT;
1073 end = (memslot->base_gfn + memslot->npages) << PAGE_SHIFT;
1075 write_lock(&kvm->mmu_lock);
1076 stage2_wp_range(&kvm->arch.mmu, start, end);
1077 write_unlock(&kvm->mmu_lock);
1078 kvm_flush_remote_tlbs(kvm);
1082 * kvm_mmu_split_memory_region() - split the stage 2 blocks into PAGE_SIZE
1083 * pages for memory slot
1084 * @kvm: The KVM pointer
1085 * @slot: The memory slot to split
1087 * Acquires kvm->mmu_lock. Called with kvm->slots_lock mutex acquired,
1088 * serializing operations for VM memory regions.
1090 static void kvm_mmu_split_memory_region(struct kvm *kvm, int slot)
1092 struct kvm_memslots *slots;
1093 struct kvm_memory_slot *memslot;
1094 phys_addr_t start, end;
1096 lockdep_assert_held(&kvm->slots_lock);
1098 slots = kvm_memslots(kvm);
1099 memslot = id_to_memslot(slots, slot);
1101 start = memslot->base_gfn << PAGE_SHIFT;
1102 end = (memslot->base_gfn + memslot->npages) << PAGE_SHIFT;
1104 write_lock(&kvm->mmu_lock);
1105 kvm_mmu_split_huge_pages(kvm, start, end);
1106 write_unlock(&kvm->mmu_lock);
1110 * kvm_arch_mmu_enable_log_dirty_pt_masked() - enable dirty logging for selected pages.
1111 * @kvm: The KVM pointer
1112 * @slot: The memory slot associated with mask
1113 * @gfn_offset: The gfn offset in memory slot
1114 * @mask: The mask of pages at offset 'gfn_offset' in this memory
1115 * slot to enable dirty logging on
1117 * Writes protect selected pages to enable dirty logging, and then
1118 * splits them to PAGE_SIZE. Caller must acquire kvm->mmu_lock.
1120 void kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm *kvm,
1121 struct kvm_memory_slot *slot,
1122 gfn_t gfn_offset, unsigned long mask)
1124 phys_addr_t base_gfn = slot->base_gfn + gfn_offset;
1125 phys_addr_t start = (base_gfn + __ffs(mask)) << PAGE_SHIFT;
1126 phys_addr_t end = (base_gfn + __fls(mask) + 1) << PAGE_SHIFT;
1128 lockdep_assert_held_write(&kvm->mmu_lock);
1130 stage2_wp_range(&kvm->arch.mmu, start, end);
1133 * Eager-splitting is done when manual-protect is set. We
1134 * also check for initially-all-set because we can avoid
1135 * eager-splitting if initially-all-set is false.
1136 * Initially-all-set equal false implies that huge-pages were
1137 * already split when enabling dirty logging: no need to do it
1140 if (kvm_dirty_log_manual_protect_and_init_set(kvm))
1141 kvm_mmu_split_huge_pages(kvm, start, end);
1144 static void kvm_send_hwpoison_signal(unsigned long address, short lsb)
1146 send_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, lsb, current);
1149 static bool fault_supports_stage2_huge_mapping(struct kvm_memory_slot *memslot,
1151 unsigned long map_size)
1154 hva_t uaddr_start, uaddr_end;
1157 /* The memslot and the VMA are guaranteed to be aligned to PAGE_SIZE */
1158 if (map_size == PAGE_SIZE)
1161 size = memslot->npages * PAGE_SIZE;
1163 gpa_start = memslot->base_gfn << PAGE_SHIFT;
1165 uaddr_start = memslot->userspace_addr;
1166 uaddr_end = uaddr_start + size;
1169 * Pages belonging to memslots that don't have the same alignment
1170 * within a PMD/PUD for userspace and IPA cannot be mapped with stage-2
1171 * PMD/PUD entries, because we'll end up mapping the wrong pages.
1173 * Consider a layout like the following:
1175 * memslot->userspace_addr:
1176 * +-----+--------------------+--------------------+---+
1177 * |abcde|fgh Stage-1 block | Stage-1 block tv|xyz|
1178 * +-----+--------------------+--------------------+---+
1180 * memslot->base_gfn << PAGE_SHIFT:
1181 * +---+--------------------+--------------------+-----+
1182 * |abc|def Stage-2 block | Stage-2 block |tvxyz|
1183 * +---+--------------------+--------------------+-----+
1185 * If we create those stage-2 blocks, we'll end up with this incorrect
1191 if ((gpa_start & (map_size - 1)) != (uaddr_start & (map_size - 1)))
1195 * Next, let's make sure we're not trying to map anything not covered
1196 * by the memslot. This means we have to prohibit block size mappings
1197 * for the beginning and end of a non-block aligned and non-block sized
1198 * memory slot (illustrated by the head and tail parts of the
1199 * userspace view above containing pages 'abcde' and 'xyz',
1202 * Note that it doesn't matter if we do the check using the
1203 * userspace_addr or the base_gfn, as both are equally aligned (per
1204 * the check above) and equally sized.
1206 return (hva & ~(map_size - 1)) >= uaddr_start &&
1207 (hva & ~(map_size - 1)) + map_size <= uaddr_end;
1211 * Check if the given hva is backed by a transparent huge page (THP) and
1212 * whether it can be mapped using block mapping in stage2. If so, adjust
1213 * the stage2 PFN and IPA accordingly. Only PMD_SIZE THPs are currently
1214 * supported. This will need to be updated to support other THP sizes.
1216 * Returns the size of the mapping.
1219 transparent_hugepage_adjust(struct kvm *kvm, struct kvm_memory_slot *memslot,
1220 unsigned long hva, kvm_pfn_t *pfnp,
1223 kvm_pfn_t pfn = *pfnp;
1226 * Make sure the adjustment is done only for THP pages. Also make
1227 * sure that the HVA and IPA are sufficiently aligned and that the
1228 * block map is contained within the memslot.
1230 if (fault_supports_stage2_huge_mapping(memslot, hva, PMD_SIZE)) {
1231 int sz = get_user_mapping_size(kvm, hva);
1240 * The address we faulted on is backed by a transparent huge
1241 * page. However, because we map the compound huge page and
1242 * not the individual tail page, we need to transfer the
1243 * refcount to the head page. We have to be careful that the
1244 * THP doesn't start to split while we are adjusting the
1247 * We are sure this doesn't happen, because mmu_invalidate_retry
1248 * was successful and we are holding the mmu_lock, so if this
1249 * THP is trying to split, it will be blocked in the mmu
1250 * notifier before touching any of the pages, specifically
1251 * before being able to call __split_huge_page_refcount().
1253 * We can therefore safely transfer the refcount from PG_tail
1254 * to PG_head and switch the pfn from a tail page to the head
1258 kvm_release_pfn_clean(pfn);
1259 pfn &= ~(PTRS_PER_PMD - 1);
1260 get_page(pfn_to_page(pfn));
1266 /* Use page mapping if we cannot use block mapping. */
1270 static int get_vma_page_shift(struct vm_area_struct *vma, unsigned long hva)
1274 if (is_vm_hugetlb_page(vma) && !(vma->vm_flags & VM_PFNMAP))
1275 return huge_page_shift(hstate_vma(vma));
1277 if (!(vma->vm_flags & VM_PFNMAP))
1280 VM_BUG_ON(is_vm_hugetlb_page(vma));
1282 pa = (vma->vm_pgoff << PAGE_SHIFT) + (hva - vma->vm_start);
1284 #ifndef __PAGETABLE_PMD_FOLDED
1285 if ((hva & (PUD_SIZE - 1)) == (pa & (PUD_SIZE - 1)) &&
1286 ALIGN_DOWN(hva, PUD_SIZE) >= vma->vm_start &&
1287 ALIGN(hva, PUD_SIZE) <= vma->vm_end)
1291 if ((hva & (PMD_SIZE - 1)) == (pa & (PMD_SIZE - 1)) &&
1292 ALIGN_DOWN(hva, PMD_SIZE) >= vma->vm_start &&
1293 ALIGN(hva, PMD_SIZE) <= vma->vm_end)
1300 * The page will be mapped in stage 2 as Normal Cacheable, so the VM will be
1301 * able to see the page's tags and therefore they must be initialised first. If
1302 * PG_mte_tagged is set, tags have already been initialised.
1304 * The race in the test/set of the PG_mte_tagged flag is handled by:
1305 * - preventing VM_SHARED mappings in a memslot with MTE preventing two VMs
1306 * racing to santise the same page
1307 * - mmap_lock protects between a VM faulting a page in and the VMM performing
1308 * an mprotect() to add VM_MTE
1310 static void sanitise_mte_tags(struct kvm *kvm, kvm_pfn_t pfn,
1313 unsigned long i, nr_pages = size >> PAGE_SHIFT;
1314 struct page *page = pfn_to_page(pfn);
1316 if (!kvm_has_mte(kvm))
1319 for (i = 0; i < nr_pages; i++, page++) {
1320 if (try_page_mte_tagging(page)) {
1321 mte_clear_page_tags(page_address(page));
1322 set_page_mte_tagged(page);
1327 static bool kvm_vma_mte_allowed(struct vm_area_struct *vma)
1329 return vma->vm_flags & VM_MTE_ALLOWED;
1332 static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
1333 struct kvm_memory_slot *memslot, unsigned long hva,
1334 unsigned long fault_status)
1337 bool write_fault, writable, force_pte = false;
1338 bool exec_fault, mte_allowed;
1339 bool device = false;
1340 unsigned long mmu_seq;
1341 struct kvm *kvm = vcpu->kvm;
1342 struct kvm_mmu_memory_cache *memcache = &vcpu->arch.mmu_page_cache;
1343 struct vm_area_struct *vma;
1347 bool logging_active = memslot_is_logging(memslot);
1348 unsigned long fault_level = kvm_vcpu_trap_get_fault_level(vcpu);
1349 long vma_pagesize, fault_granule;
1350 enum kvm_pgtable_prot prot = KVM_PGTABLE_PROT_R;
1351 struct kvm_pgtable *pgt;
1353 fault_granule = 1UL << ARM64_HW_PGTABLE_LEVEL_SHIFT(fault_level);
1354 write_fault = kvm_is_write_fault(vcpu);
1355 exec_fault = kvm_vcpu_trap_is_exec_fault(vcpu);
1356 VM_BUG_ON(write_fault && exec_fault);
1358 if (fault_status == ESR_ELx_FSC_PERM && !write_fault && !exec_fault) {
1359 kvm_err("Unexpected L2 read permission error\n");
1364 * Permission faults just need to update the existing leaf entry,
1365 * and so normally don't require allocations from the memcache. The
1366 * only exception to this is when dirty logging is enabled at runtime
1367 * and a write fault needs to collapse a block entry into a table.
1369 if (fault_status != ESR_ELx_FSC_PERM ||
1370 (logging_active && write_fault)) {
1371 ret = kvm_mmu_topup_memory_cache(memcache,
1372 kvm_mmu_cache_min_pages(kvm));
1378 * Let's check if we will get back a huge page backed by hugetlbfs, or
1379 * get block mapping for device MMIO region.
1381 mmap_read_lock(current->mm);
1382 vma = vma_lookup(current->mm, hva);
1383 if (unlikely(!vma)) {
1384 kvm_err("Failed to find VMA for hva 0x%lx\n", hva);
1385 mmap_read_unlock(current->mm);
1390 * logging_active is guaranteed to never be true for VM_PFNMAP
1393 if (logging_active) {
1395 vma_shift = PAGE_SHIFT;
1397 vma_shift = get_vma_page_shift(vma, hva);
1400 switch (vma_shift) {
1401 #ifndef __PAGETABLE_PMD_FOLDED
1403 if (fault_supports_stage2_huge_mapping(memslot, hva, PUD_SIZE))
1407 case CONT_PMD_SHIFT:
1408 vma_shift = PMD_SHIFT;
1411 if (fault_supports_stage2_huge_mapping(memslot, hva, PMD_SIZE))
1414 case CONT_PTE_SHIFT:
1415 vma_shift = PAGE_SHIFT;
1421 WARN_ONCE(1, "Unknown vma_shift %d", vma_shift);
1424 vma_pagesize = 1UL << vma_shift;
1425 if (vma_pagesize == PMD_SIZE || vma_pagesize == PUD_SIZE)
1426 fault_ipa &= ~(vma_pagesize - 1);
1428 gfn = fault_ipa >> PAGE_SHIFT;
1429 mte_allowed = kvm_vma_mte_allowed(vma);
1431 /* Don't use the VMA after the unlock -- it may have vanished */
1435 * Read mmu_invalidate_seq so that KVM can detect if the results of
1436 * vma_lookup() or __gfn_to_pfn_memslot() become stale prior to
1437 * acquiring kvm->mmu_lock.
1439 * Rely on mmap_read_unlock() for an implicit smp_rmb(), which pairs
1440 * with the smp_wmb() in kvm_mmu_invalidate_end().
1442 mmu_seq = vcpu->kvm->mmu_invalidate_seq;
1443 mmap_read_unlock(current->mm);
1445 pfn = __gfn_to_pfn_memslot(memslot, gfn, false, false, NULL,
1446 write_fault, &writable, NULL);
1447 if (pfn == KVM_PFN_ERR_HWPOISON) {
1448 kvm_send_hwpoison_signal(hva, vma_shift);
1451 if (is_error_noslot_pfn(pfn))
1454 if (kvm_is_device_pfn(pfn)) {
1456 * If the page was identified as device early by looking at
1457 * the VMA flags, vma_pagesize is already representing the
1458 * largest quantity we can map. If instead it was mapped
1459 * via gfn_to_pfn_prot(), vma_pagesize is set to PAGE_SIZE
1460 * and must not be upgraded.
1462 * In both cases, we don't let transparent_hugepage_adjust()
1463 * change things at the last minute.
1466 } else if (logging_active && !write_fault) {
1468 * Only actually map the page as writable if this was a write
1474 if (exec_fault && device)
1477 read_lock(&kvm->mmu_lock);
1478 pgt = vcpu->arch.hw_mmu->pgt;
1479 if (mmu_invalidate_retry(kvm, mmu_seq))
1483 * If we are not forced to use page mapping, check if we are
1484 * backed by a THP and thus use block mapping if possible.
1486 if (vma_pagesize == PAGE_SIZE && !(force_pte || device)) {
1487 if (fault_status == ESR_ELx_FSC_PERM &&
1488 fault_granule > PAGE_SIZE)
1489 vma_pagesize = fault_granule;
1491 vma_pagesize = transparent_hugepage_adjust(kvm, memslot,
1495 if (vma_pagesize < 0) {
1501 if (fault_status != ESR_ELx_FSC_PERM && !device && kvm_has_mte(kvm)) {
1502 /* Check the VMM hasn't introduced a new disallowed VMA */
1504 sanitise_mte_tags(kvm, pfn, vma_pagesize);
1512 prot |= KVM_PGTABLE_PROT_W;
1515 prot |= KVM_PGTABLE_PROT_X;
1518 prot |= KVM_PGTABLE_PROT_DEVICE;
1519 else if (cpus_have_const_cap(ARM64_HAS_CACHE_DIC))
1520 prot |= KVM_PGTABLE_PROT_X;
1523 * Under the premise of getting a FSC_PERM fault, we just need to relax
1524 * permissions only if vma_pagesize equals fault_granule. Otherwise,
1525 * kvm_pgtable_stage2_map() should be called to change block size.
1527 if (fault_status == ESR_ELx_FSC_PERM && vma_pagesize == fault_granule)
1528 ret = kvm_pgtable_stage2_relax_perms(pgt, fault_ipa, prot);
1530 ret = kvm_pgtable_stage2_map(pgt, fault_ipa, vma_pagesize,
1531 __pfn_to_phys(pfn), prot,
1533 KVM_PGTABLE_WALK_HANDLE_FAULT |
1534 KVM_PGTABLE_WALK_SHARED);
1536 /* Mark the page dirty only if the fault is handled successfully */
1537 if (writable && !ret) {
1538 kvm_set_pfn_dirty(pfn);
1539 mark_page_dirty_in_slot(kvm, memslot, gfn);
1543 read_unlock(&kvm->mmu_lock);
1544 kvm_set_pfn_accessed(pfn);
1545 kvm_release_pfn_clean(pfn);
1546 return ret != -EAGAIN ? ret : 0;
1549 /* Resolve the access fault by making the page young again. */
1550 static void handle_access_fault(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa)
1553 struct kvm_s2_mmu *mmu;
1555 trace_kvm_access_fault(fault_ipa);
1557 read_lock(&vcpu->kvm->mmu_lock);
1558 mmu = vcpu->arch.hw_mmu;
1559 pte = kvm_pgtable_stage2_mkyoung(mmu->pgt, fault_ipa);
1560 read_unlock(&vcpu->kvm->mmu_lock);
1562 if (kvm_pte_valid(pte))
1563 kvm_set_pfn_accessed(kvm_pte_to_pfn(pte));
1567 * kvm_handle_guest_abort - handles all 2nd stage aborts
1568 * @vcpu: the VCPU pointer
1570 * Any abort that gets to the host is almost guaranteed to be caused by a
1571 * missing second stage translation table entry, which can mean that either the
1572 * guest simply needs more memory and we must allocate an appropriate page or it
1573 * can mean that the guest tried to access I/O memory, which is emulated by user
1574 * space. The distinction is based on the IPA causing the fault and whether this
1575 * memory region has been registered as standard RAM by user space.
1577 int kvm_handle_guest_abort(struct kvm_vcpu *vcpu)
1579 unsigned long fault_status;
1580 phys_addr_t fault_ipa;
1581 struct kvm_memory_slot *memslot;
1583 bool is_iabt, write_fault, writable;
1587 fault_status = kvm_vcpu_trap_get_fault_type(vcpu);
1589 fault_ipa = kvm_vcpu_get_fault_ipa(vcpu);
1590 is_iabt = kvm_vcpu_trap_is_iabt(vcpu);
1592 if (fault_status == ESR_ELx_FSC_FAULT) {
1593 /* Beyond sanitised PARange (which is the IPA limit) */
1594 if (fault_ipa >= BIT_ULL(get_kvm_ipa_limit())) {
1595 kvm_inject_size_fault(vcpu);
1599 /* Falls between the IPA range and the PARange? */
1600 if (fault_ipa >= BIT_ULL(vcpu->arch.hw_mmu->pgt->ia_bits)) {
1601 fault_ipa |= kvm_vcpu_get_hfar(vcpu) & GENMASK(11, 0);
1604 kvm_inject_pabt(vcpu, fault_ipa);
1606 kvm_inject_dabt(vcpu, fault_ipa);
1611 /* Synchronous External Abort? */
1612 if (kvm_vcpu_abt_issea(vcpu)) {
1614 * For RAS the host kernel may handle this abort.
1615 * There is no need to pass the error into the guest.
1617 if (kvm_handle_guest_sea(fault_ipa, kvm_vcpu_get_esr(vcpu)))
1618 kvm_inject_vabt(vcpu);
1623 trace_kvm_guest_fault(*vcpu_pc(vcpu), kvm_vcpu_get_esr(vcpu),
1624 kvm_vcpu_get_hfar(vcpu), fault_ipa);
1626 /* Check the stage-2 fault is trans. fault or write fault */
1627 if (fault_status != ESR_ELx_FSC_FAULT &&
1628 fault_status != ESR_ELx_FSC_PERM &&
1629 fault_status != ESR_ELx_FSC_ACCESS) {
1630 kvm_err("Unsupported FSC: EC=%#x xFSC=%#lx ESR_EL2=%#lx\n",
1631 kvm_vcpu_trap_get_class(vcpu),
1632 (unsigned long)kvm_vcpu_trap_get_fault(vcpu),
1633 (unsigned long)kvm_vcpu_get_esr(vcpu));
1637 idx = srcu_read_lock(&vcpu->kvm->srcu);
1639 gfn = fault_ipa >> PAGE_SHIFT;
1640 memslot = gfn_to_memslot(vcpu->kvm, gfn);
1641 hva = gfn_to_hva_memslot_prot(memslot, gfn, &writable);
1642 write_fault = kvm_is_write_fault(vcpu);
1643 if (kvm_is_error_hva(hva) || (write_fault && !writable)) {
1645 * The guest has put either its instructions or its page-tables
1646 * somewhere it shouldn't have. Userspace won't be able to do
1647 * anything about this (there's no syndrome for a start), so
1648 * re-inject the abort back into the guest.
1655 if (kvm_vcpu_abt_iss1tw(vcpu)) {
1656 kvm_inject_dabt(vcpu, kvm_vcpu_get_hfar(vcpu));
1662 * Check for a cache maintenance operation. Since we
1663 * ended-up here, we know it is outside of any memory
1664 * slot. But we can't find out if that is for a device,
1665 * or if the guest is just being stupid. The only thing
1666 * we know for sure is that this range cannot be cached.
1668 * So let's assume that the guest is just being
1669 * cautious, and skip the instruction.
1671 if (kvm_is_error_hva(hva) && kvm_vcpu_dabt_is_cm(vcpu)) {
1678 * The IPA is reported as [MAX:12], so we need to
1679 * complement it with the bottom 12 bits from the
1680 * faulting VA. This is always 12 bits, irrespective
1683 fault_ipa |= kvm_vcpu_get_hfar(vcpu) & ((1 << 12) - 1);
1684 ret = io_mem_abort(vcpu, fault_ipa);
1688 /* Userspace should not be able to register out-of-bounds IPAs */
1689 VM_BUG_ON(fault_ipa >= kvm_phys_size(vcpu->kvm));
1691 if (fault_status == ESR_ELx_FSC_ACCESS) {
1692 handle_access_fault(vcpu, fault_ipa);
1697 ret = user_mem_abort(vcpu, fault_ipa, memslot, hva, fault_status);
1701 if (ret == -ENOEXEC) {
1702 kvm_inject_pabt(vcpu, kvm_vcpu_get_hfar(vcpu));
1706 srcu_read_unlock(&vcpu->kvm->srcu, idx);
1710 bool kvm_unmap_gfn_range(struct kvm *kvm, struct kvm_gfn_range *range)
1712 if (!kvm->arch.mmu.pgt)
1715 __unmap_stage2_range(&kvm->arch.mmu, range->start << PAGE_SHIFT,
1716 (range->end - range->start) << PAGE_SHIFT,
1722 bool kvm_set_spte_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
1724 kvm_pfn_t pfn = pte_pfn(range->pte);
1726 if (!kvm->arch.mmu.pgt)
1729 WARN_ON(range->end - range->start != 1);
1732 * If the page isn't tagged, defer to user_mem_abort() for sanitising
1733 * the MTE tags. The S2 pte should have been unmapped by
1734 * mmu_notifier_invalidate_range_end().
1736 if (kvm_has_mte(kvm) && !page_mte_tagged(pfn_to_page(pfn)))
1740 * We've moved a page around, probably through CoW, so let's treat
1741 * it just like a translation fault and the map handler will clean
1742 * the cache to the PoC.
1744 * The MMU notifiers will have unmapped a huge PMD before calling
1745 * ->change_pte() (which in turn calls kvm_set_spte_gfn()) and
1746 * therefore we never need to clear out a huge PMD through this
1747 * calling path and a memcache is not required.
1749 kvm_pgtable_stage2_map(kvm->arch.mmu.pgt, range->start << PAGE_SHIFT,
1750 PAGE_SIZE, __pfn_to_phys(pfn),
1751 KVM_PGTABLE_PROT_R, NULL, 0);
1756 bool kvm_age_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
1758 u64 size = (range->end - range->start) << PAGE_SHIFT;
1762 if (!kvm->arch.mmu.pgt)
1765 WARN_ON(size != PAGE_SIZE && size != PMD_SIZE && size != PUD_SIZE);
1767 kpte = kvm_pgtable_stage2_mkold(kvm->arch.mmu.pgt,
1768 range->start << PAGE_SHIFT);
1770 return pte_valid(pte) && pte_young(pte);
1773 bool kvm_test_age_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
1775 if (!kvm->arch.mmu.pgt)
1778 return kvm_pgtable_stage2_is_young(kvm->arch.mmu.pgt,
1779 range->start << PAGE_SHIFT);
1782 phys_addr_t kvm_mmu_get_httbr(void)
1784 return __pa(hyp_pgtable->pgd);
1787 phys_addr_t kvm_get_idmap_vector(void)
1789 return hyp_idmap_vector;
1792 static int kvm_map_idmap_text(void)
1794 unsigned long size = hyp_idmap_end - hyp_idmap_start;
1795 int err = __create_hyp_mappings(hyp_idmap_start, size, hyp_idmap_start,
1798 kvm_err("Failed to idmap %lx-%lx\n",
1799 hyp_idmap_start, hyp_idmap_end);
1804 static void *kvm_hyp_zalloc_page(void *arg)
1806 return (void *)get_zeroed_page(GFP_KERNEL);
1809 static struct kvm_pgtable_mm_ops kvm_hyp_mm_ops = {
1810 .zalloc_page = kvm_hyp_zalloc_page,
1811 .get_page = kvm_host_get_page,
1812 .put_page = kvm_host_put_page,
1813 .phys_to_virt = kvm_host_va,
1814 .virt_to_phys = kvm_host_pa,
1817 int __init kvm_mmu_init(u32 *hyp_va_bits)
1823 hyp_idmap_start = __pa_symbol(__hyp_idmap_text_start);
1824 hyp_idmap_start = ALIGN_DOWN(hyp_idmap_start, PAGE_SIZE);
1825 hyp_idmap_end = __pa_symbol(__hyp_idmap_text_end);
1826 hyp_idmap_end = ALIGN(hyp_idmap_end, PAGE_SIZE);
1827 hyp_idmap_vector = __pa_symbol(__kvm_hyp_init);
1830 * We rely on the linker script to ensure at build time that the HYP
1831 * init code does not cross a page boundary.
1833 BUG_ON((hyp_idmap_start ^ (hyp_idmap_end - 1)) & PAGE_MASK);
1836 * The ID map may be configured to use an extended virtual address
1837 * range. This is only the case if system RAM is out of range for the
1838 * currently configured page size and VA_BITS_MIN, in which case we will
1839 * also need the extended virtual range for the HYP ID map, or we won't
1840 * be able to enable the EL2 MMU.
1842 * However, in some cases the ID map may be configured for fewer than
1843 * the number of VA bits used by the regular kernel stage 1. This
1844 * happens when VA_BITS=52 and the kernel image is placed in PA space
1847 * At EL2, there is only one TTBR register, and we can't switch between
1848 * translation tables *and* update TCR_EL2.T0SZ at the same time. Bottom
1849 * line: we need to use the extended range with *both* our translation
1852 * So use the maximum of the idmap VA bits and the regular kernel stage
1853 * 1 VA bits to assure that the hypervisor can both ID map its code page
1854 * and map any kernel memory.
1856 idmap_bits = 64 - ((idmap_t0sz & TCR_T0SZ_MASK) >> TCR_T0SZ_OFFSET);
1857 kernel_bits = vabits_actual;
1858 *hyp_va_bits = max(idmap_bits, kernel_bits);
1860 kvm_debug("Using %u-bit virtual addresses at EL2\n", *hyp_va_bits);
1861 kvm_debug("IDMAP page: %lx\n", hyp_idmap_start);
1862 kvm_debug("HYP VA range: %lx:%lx\n",
1863 kern_hyp_va(PAGE_OFFSET),
1864 kern_hyp_va((unsigned long)high_memory - 1));
1866 if (hyp_idmap_start >= kern_hyp_va(PAGE_OFFSET) &&
1867 hyp_idmap_start < kern_hyp_va((unsigned long)high_memory - 1) &&
1868 hyp_idmap_start != (unsigned long)__hyp_idmap_text_start) {
1870 * The idmap page is intersecting with the VA space,
1871 * it is not safe to continue further.
1873 kvm_err("IDMAP intersecting with HYP VA, unable to continue\n");
1878 hyp_pgtable = kzalloc(sizeof(*hyp_pgtable), GFP_KERNEL);
1880 kvm_err("Hyp mode page-table not allocated\n");
1885 err = kvm_pgtable_hyp_init(hyp_pgtable, *hyp_va_bits, &kvm_hyp_mm_ops);
1887 goto out_free_pgtable;
1889 err = kvm_map_idmap_text();
1891 goto out_destroy_pgtable;
1893 io_map_base = hyp_idmap_start;
1896 out_destroy_pgtable:
1897 kvm_pgtable_hyp_destroy(hyp_pgtable);
1905 void kvm_arch_commit_memory_region(struct kvm *kvm,
1906 struct kvm_memory_slot *old,
1907 const struct kvm_memory_slot *new,
1908 enum kvm_mr_change change)
1910 bool log_dirty_pages = new && new->flags & KVM_MEM_LOG_DIRTY_PAGES;
1913 * At this point memslot has been committed and there is an
1914 * allocated dirty_bitmap[], dirty pages will be tracked while the
1915 * memory slot is write protected.
1917 if (log_dirty_pages) {
1919 if (change == KVM_MR_DELETE)
1923 * Huge and normal pages are write-protected and split
1924 * on either of these two cases:
1926 * 1. with initial-all-set: gradually with CLEAR ioctls,
1928 if (kvm_dirty_log_manual_protect_and_init_set(kvm))
1932 * 2. without initial-all-set: all in one shot when
1933 * enabling dirty logging.
1935 kvm_mmu_wp_memory_region(kvm, new->id);
1936 kvm_mmu_split_memory_region(kvm, new->id);
1939 * Free any leftovers from the eager page splitting cache. Do
1940 * this when deleting, moving, disabling dirty logging, or
1941 * creating the memslot (a nop). Doing it for deletes makes
1942 * sure we don't leak memory, and there's no need to keep the
1943 * cache around for any of the other cases.
1945 kvm_mmu_free_memory_cache(&kvm->arch.mmu.split_page_cache);
1949 int kvm_arch_prepare_memory_region(struct kvm *kvm,
1950 const struct kvm_memory_slot *old,
1951 struct kvm_memory_slot *new,
1952 enum kvm_mr_change change)
1957 if (change != KVM_MR_CREATE && change != KVM_MR_MOVE &&
1958 change != KVM_MR_FLAGS_ONLY)
1962 * Prevent userspace from creating a memory region outside of the IPA
1963 * space addressable by the KVM guest IPA space.
1965 if ((new->base_gfn + new->npages) > (kvm_phys_size(kvm) >> PAGE_SHIFT))
1968 hva = new->userspace_addr;
1969 reg_end = hva + (new->npages << PAGE_SHIFT);
1971 mmap_read_lock(current->mm);
1973 * A memory region could potentially cover multiple VMAs, and any holes
1974 * between them, so iterate over all of them.
1976 * +--------------------------------------------+
1977 * +---------------+----------------+ +----------------+
1978 * | : VMA 1 | VMA 2 | | VMA 3 : |
1979 * +---------------+----------------+ +----------------+
1981 * +--------------------------------------------+
1984 struct vm_area_struct *vma;
1986 vma = find_vma_intersection(current->mm, hva, reg_end);
1990 if (kvm_has_mte(kvm) && !kvm_vma_mte_allowed(vma)) {
1995 if (vma->vm_flags & VM_PFNMAP) {
1996 /* IO region dirty page logging not allowed */
1997 if (new->flags & KVM_MEM_LOG_DIRTY_PAGES) {
2002 hva = min(reg_end, vma->vm_end);
2003 } while (hva < reg_end);
2005 mmap_read_unlock(current->mm);
2009 void kvm_arch_free_memslot(struct kvm *kvm, struct kvm_memory_slot *slot)
2013 void kvm_arch_memslots_updated(struct kvm *kvm, u64 gen)
2017 void kvm_arch_flush_shadow_all(struct kvm *kvm)
2019 kvm_uninit_stage2_mmu(kvm);
2022 void kvm_arch_flush_shadow_memslot(struct kvm *kvm,
2023 struct kvm_memory_slot *slot)
2025 gpa_t gpa = slot->base_gfn << PAGE_SHIFT;
2026 phys_addr_t size = slot->npages << PAGE_SHIFT;
2028 write_lock(&kvm->mmu_lock);
2029 unmap_stage2_range(&kvm->arch.mmu, gpa, size);
2030 write_unlock(&kvm->mmu_lock);
2034 * See note at ARMv7 ARM B1.14.4 (TL;DR: S/W ops are not easily virtualized).
2037 * - S/W ops are local to a CPU (not broadcast)
2038 * - We have line migration behind our back (speculation)
2039 * - System caches don't support S/W at all (damn!)
2041 * In the face of the above, the best we can do is to try and convert
2042 * S/W ops to VA ops. Because the guest is not allowed to infer the
2043 * S/W to PA mapping, it can only use S/W to nuke the whole cache,
2044 * which is a rather good thing for us.
2046 * Also, it is only used when turning caches on/off ("The expected
2047 * usage of the cache maintenance instructions that operate by set/way
2048 * is associated with the cache maintenance instructions associated
2049 * with the powerdown and powerup of caches, if this is required by
2050 * the implementation.").
2052 * We use the following policy:
2054 * - If we trap a S/W operation, we enable VM trapping to detect
2055 * caches being turned on/off, and do a full clean.
2057 * - We flush the caches on both caches being turned on and off.
2059 * - Once the caches are enabled, we stop trapping VM ops.
2061 void kvm_set_way_flush(struct kvm_vcpu *vcpu)
2063 unsigned long hcr = *vcpu_hcr(vcpu);
2066 * If this is the first time we do a S/W operation
2067 * (i.e. HCR_TVM not set) flush the whole memory, and set the
2070 * Otherwise, rely on the VM trapping to wait for the MMU +
2071 * Caches to be turned off. At that point, we'll be able to
2072 * clean the caches again.
2074 if (!(hcr & HCR_TVM)) {
2075 trace_kvm_set_way_flush(*vcpu_pc(vcpu),
2076 vcpu_has_cache_enabled(vcpu));
2077 stage2_flush_vm(vcpu->kvm);
2078 *vcpu_hcr(vcpu) = hcr | HCR_TVM;
2082 void kvm_toggle_cache(struct kvm_vcpu *vcpu, bool was_enabled)
2084 bool now_enabled = vcpu_has_cache_enabled(vcpu);
2087 * If switching the MMU+caches on, need to invalidate the caches.
2088 * If switching it off, need to clean the caches.
2089 * Clean + invalidate does the trick always.
2091 if (now_enabled != was_enabled)
2092 stage2_flush_vm(vcpu->kvm);
2094 /* Caches are now on, stop trapping VM ops (until a S/W op) */
2096 *vcpu_hcr(vcpu) &= ~HCR_TVM;
2098 trace_kvm_toggle_cache(*vcpu_pc(vcpu), was_enabled, now_enabled);