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 hyp_idmap_start;
29 static unsigned long hyp_idmap_end;
30 static phys_addr_t hyp_idmap_vector;
32 static unsigned long io_map_base;
36 * Release kvm_mmu_lock periodically if the memory region is large. Otherwise,
37 * we may see kernel panics with CONFIG_DETECT_HUNG_TASK,
38 * CONFIG_LOCKUP_DETECTOR, CONFIG_LOCKDEP. Additionally, holding the lock too
39 * long will also starve other vCPUs. We have to also make sure that the page
40 * tables are not freed while we released the lock.
42 static int stage2_apply_range(struct kvm *kvm, phys_addr_t addr,
44 int (*fn)(struct kvm_pgtable *, u64, u64),
51 struct kvm_pgtable *pgt = kvm->arch.mmu.pgt;
55 next = stage2_pgd_addr_end(kvm, addr, end);
56 ret = fn(pgt, addr, next - addr);
60 if (resched && next != end)
61 cond_resched_rwlock_write(&kvm->mmu_lock);
62 } while (addr = next, addr != end);
67 #define stage2_apply_range_resched(kvm, addr, end, fn) \
68 stage2_apply_range(kvm, addr, end, fn, true)
70 static bool memslot_is_logging(struct kvm_memory_slot *memslot)
72 return memslot->dirty_bitmap && !(memslot->flags & KVM_MEM_READONLY);
76 * kvm_flush_remote_tlbs() - flush all VM TLB entries for v7/8
77 * @kvm: pointer to kvm structure.
79 * Interface to HYP function to flush all VM TLB entries
81 void kvm_flush_remote_tlbs(struct kvm *kvm)
83 ++kvm->stat.generic.remote_tlb_flush_requests;
84 kvm_call_hyp(__kvm_tlb_flush_vmid, &kvm->arch.mmu);
87 static bool kvm_is_device_pfn(unsigned long pfn)
89 return !pfn_is_map_memory(pfn);
92 static void *stage2_memcache_zalloc_page(void *arg)
94 struct kvm_mmu_memory_cache *mc = arg;
97 /* Allocated with __GFP_ZERO, so no need to zero */
98 virt = kvm_mmu_memory_cache_alloc(mc);
100 kvm_account_pgtable_pages(virt, 1);
104 static void *kvm_host_zalloc_pages_exact(size_t size)
106 return alloc_pages_exact(size, GFP_KERNEL_ACCOUNT | __GFP_ZERO);
109 static void *kvm_s2_zalloc_pages_exact(size_t size)
111 void *virt = kvm_host_zalloc_pages_exact(size);
114 kvm_account_pgtable_pages(virt, (size >> PAGE_SHIFT));
118 static void kvm_s2_free_pages_exact(void *virt, size_t size)
120 kvm_account_pgtable_pages(virt, -(size >> PAGE_SHIFT));
121 free_pages_exact(virt, size);
124 static void kvm_host_get_page(void *addr)
126 get_page(virt_to_page(addr));
129 static void kvm_host_put_page(void *addr)
131 put_page(virt_to_page(addr));
134 static void kvm_s2_put_page(void *addr)
136 struct page *p = virt_to_page(addr);
137 /* Dropping last refcount, the page will be freed */
138 if (page_count(p) == 1)
139 kvm_account_pgtable_pages(addr, -1);
143 static int kvm_host_page_count(void *addr)
145 return page_count(virt_to_page(addr));
148 static phys_addr_t kvm_host_pa(void *addr)
153 static void *kvm_host_va(phys_addr_t phys)
158 static void clean_dcache_guest_page(void *va, size_t size)
160 __clean_dcache_guest_page(va, size);
163 static void invalidate_icache_guest_page(void *va, size_t size)
165 __invalidate_icache_guest_page(va, size);
169 * Unmapping vs dcache management:
171 * If a guest maps certain memory pages as uncached, all writes will
172 * bypass the data cache and go directly to RAM. However, the CPUs
173 * can still speculate reads (not writes) and fill cache lines with
176 * Those cache lines will be *clean* cache lines though, so a
177 * clean+invalidate operation is equivalent to an invalidate
178 * operation, because no cache lines are marked dirty.
180 * Those clean cache lines could be filled prior to an uncached write
181 * by the guest, and the cache coherent IO subsystem would therefore
182 * end up writing old data to disk.
184 * This is why right after unmapping a page/section and invalidating
185 * the corresponding TLBs, we flush to make sure the IO subsystem will
186 * never hit in the cache.
188 * This is all avoided on systems that have ARM64_HAS_STAGE2_FWB, as
189 * we then fully enforce cacheability of RAM, no matter what the guest
193 * unmap_stage2_range -- Clear stage2 page table entries to unmap a range
194 * @mmu: The KVM stage-2 MMU pointer
195 * @start: The intermediate physical base address of the range to unmap
196 * @size: The size of the area to unmap
197 * @may_block: Whether or not we are permitted to block
199 * Clear a range of stage-2 mappings, lowering the various ref-counts. Must
200 * be called while holding mmu_lock (unless for freeing the stage2 pgd before
201 * destroying the VM), otherwise another faulting VCPU may come in and mess
202 * with things behind our backs.
204 static void __unmap_stage2_range(struct kvm_s2_mmu *mmu, phys_addr_t start, u64 size,
207 struct kvm *kvm = kvm_s2_mmu_to_kvm(mmu);
208 phys_addr_t end = start + size;
210 lockdep_assert_held_write(&kvm->mmu_lock);
211 WARN_ON(size & ~PAGE_MASK);
212 WARN_ON(stage2_apply_range(kvm, start, end, kvm_pgtable_stage2_unmap,
216 static void unmap_stage2_range(struct kvm_s2_mmu *mmu, phys_addr_t start, u64 size)
218 __unmap_stage2_range(mmu, start, size, true);
221 static void stage2_flush_memslot(struct kvm *kvm,
222 struct kvm_memory_slot *memslot)
224 phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
225 phys_addr_t end = addr + PAGE_SIZE * memslot->npages;
227 stage2_apply_range_resched(kvm, addr, end, kvm_pgtable_stage2_flush);
231 * stage2_flush_vm - Invalidate cache for pages mapped in stage 2
232 * @kvm: The struct kvm pointer
234 * Go through the stage 2 page tables and invalidate any cache lines
235 * backing memory already mapped to the VM.
237 static void stage2_flush_vm(struct kvm *kvm)
239 struct kvm_memslots *slots;
240 struct kvm_memory_slot *memslot;
243 idx = srcu_read_lock(&kvm->srcu);
244 write_lock(&kvm->mmu_lock);
246 slots = kvm_memslots(kvm);
247 kvm_for_each_memslot(memslot, bkt, slots)
248 stage2_flush_memslot(kvm, memslot);
250 write_unlock(&kvm->mmu_lock);
251 srcu_read_unlock(&kvm->srcu, idx);
255 * free_hyp_pgds - free Hyp-mode page tables
257 void free_hyp_pgds(void)
259 mutex_lock(&kvm_hyp_pgd_mutex);
261 kvm_pgtable_hyp_destroy(hyp_pgtable);
265 mutex_unlock(&kvm_hyp_pgd_mutex);
268 static bool kvm_host_owns_hyp_mappings(void)
270 if (is_kernel_in_hyp_mode())
273 if (static_branch_likely(&kvm_protected_mode_initialized))
277 * This can happen at boot time when __create_hyp_mappings() is called
278 * after the hyp protection has been enabled, but the static key has
279 * not been flipped yet.
281 if (!hyp_pgtable && is_protected_kvm_enabled())
284 WARN_ON(!hyp_pgtable);
289 int __create_hyp_mappings(unsigned long start, unsigned long size,
290 unsigned long phys, enum kvm_pgtable_prot prot)
294 if (WARN_ON(!kvm_host_owns_hyp_mappings()))
297 mutex_lock(&kvm_hyp_pgd_mutex);
298 err = kvm_pgtable_hyp_map(hyp_pgtable, start, size, phys, prot);
299 mutex_unlock(&kvm_hyp_pgd_mutex);
304 static phys_addr_t kvm_kaddr_to_phys(void *kaddr)
306 if (!is_vmalloc_addr(kaddr)) {
307 BUG_ON(!virt_addr_valid(kaddr));
310 return page_to_phys(vmalloc_to_page(kaddr)) +
311 offset_in_page(kaddr);
315 struct hyp_shared_pfn {
321 static DEFINE_MUTEX(hyp_shared_pfns_lock);
322 static struct rb_root hyp_shared_pfns = RB_ROOT;
324 static struct hyp_shared_pfn *find_shared_pfn(u64 pfn, struct rb_node ***node,
325 struct rb_node **parent)
327 struct hyp_shared_pfn *this;
329 *node = &hyp_shared_pfns.rb_node;
332 this = container_of(**node, struct hyp_shared_pfn, node);
335 *node = &((**node)->rb_left);
336 else if (this->pfn > pfn)
337 *node = &((**node)->rb_right);
345 static int share_pfn_hyp(u64 pfn)
347 struct rb_node **node, *parent;
348 struct hyp_shared_pfn *this;
351 mutex_lock(&hyp_shared_pfns_lock);
352 this = find_shared_pfn(pfn, &node, &parent);
358 this = kzalloc(sizeof(*this), GFP_KERNEL);
366 rb_link_node(&this->node, parent, node);
367 rb_insert_color(&this->node, &hyp_shared_pfns);
368 ret = kvm_call_hyp_nvhe(__pkvm_host_share_hyp, pfn, 1);
370 mutex_unlock(&hyp_shared_pfns_lock);
375 static int unshare_pfn_hyp(u64 pfn)
377 struct rb_node **node, *parent;
378 struct hyp_shared_pfn *this;
381 mutex_lock(&hyp_shared_pfns_lock);
382 this = find_shared_pfn(pfn, &node, &parent);
383 if (WARN_ON(!this)) {
392 rb_erase(&this->node, &hyp_shared_pfns);
394 ret = kvm_call_hyp_nvhe(__pkvm_host_unshare_hyp, pfn, 1);
396 mutex_unlock(&hyp_shared_pfns_lock);
401 int kvm_share_hyp(void *from, void *to)
403 phys_addr_t start, end, cur;
407 if (is_kernel_in_hyp_mode())
411 * The share hcall maps things in the 'fixed-offset' region of the hyp
412 * VA space, so we can only share physically contiguous data-structures
415 if (is_vmalloc_or_module_addr(from) || is_vmalloc_or_module_addr(to))
418 if (kvm_host_owns_hyp_mappings())
419 return create_hyp_mappings(from, to, PAGE_HYP);
421 start = ALIGN_DOWN(__pa(from), PAGE_SIZE);
422 end = PAGE_ALIGN(__pa(to));
423 for (cur = start; cur < end; cur += PAGE_SIZE) {
424 pfn = __phys_to_pfn(cur);
425 ret = share_pfn_hyp(pfn);
433 void kvm_unshare_hyp(void *from, void *to)
435 phys_addr_t start, end, cur;
438 if (is_kernel_in_hyp_mode() || kvm_host_owns_hyp_mappings() || !from)
441 start = ALIGN_DOWN(__pa(from), PAGE_SIZE);
442 end = PAGE_ALIGN(__pa(to));
443 for (cur = start; cur < end; cur += PAGE_SIZE) {
444 pfn = __phys_to_pfn(cur);
445 WARN_ON(unshare_pfn_hyp(pfn));
450 * create_hyp_mappings - duplicate a kernel virtual address range in Hyp mode
451 * @from: The virtual kernel start address of the range
452 * @to: The virtual kernel end address of the range (exclusive)
453 * @prot: The protection to be applied to this range
455 * The same virtual address as the kernel virtual address is also used
456 * in Hyp-mode mapping (modulo HYP_PAGE_OFFSET) to the same underlying
459 int create_hyp_mappings(void *from, void *to, enum kvm_pgtable_prot prot)
461 phys_addr_t phys_addr;
462 unsigned long virt_addr;
463 unsigned long start = kern_hyp_va((unsigned long)from);
464 unsigned long end = kern_hyp_va((unsigned long)to);
466 if (is_kernel_in_hyp_mode())
469 if (!kvm_host_owns_hyp_mappings())
472 start = start & PAGE_MASK;
473 end = PAGE_ALIGN(end);
475 for (virt_addr = start; virt_addr < end; virt_addr += PAGE_SIZE) {
478 phys_addr = kvm_kaddr_to_phys(from + virt_addr - start);
479 err = __create_hyp_mappings(virt_addr, PAGE_SIZE, phys_addr,
490 * hyp_alloc_private_va_range - Allocates a private VA range.
491 * @size: The size of the VA range to reserve.
492 * @haddr: The hypervisor virtual start address of the allocation.
494 * The private virtual address (VA) range is allocated below io_map_base
495 * and aligned based on the order of @size.
497 * Return: 0 on success or negative error code on failure.
499 int hyp_alloc_private_va_range(size_t size, unsigned long *haddr)
504 mutex_lock(&kvm_hyp_pgd_mutex);
507 * This assumes that we have enough space below the idmap
508 * page to allocate our VAs. If not, the check below will
509 * kick. A potential alternative would be to detect that
510 * overflow and switch to an allocation above the idmap.
512 * The allocated size is always a multiple of PAGE_SIZE.
514 base = io_map_base - PAGE_ALIGN(size);
516 /* Align the allocation based on the order of its size */
517 base = ALIGN_DOWN(base, PAGE_SIZE << get_order(size));
520 * Verify that BIT(VA_BITS - 1) hasn't been flipped by
521 * allocating the new area, as it would indicate we've
522 * overflowed the idmap/IO address range.
524 if ((base ^ io_map_base) & BIT(VA_BITS - 1))
527 *haddr = io_map_base = base;
529 mutex_unlock(&kvm_hyp_pgd_mutex);
534 static int __create_hyp_private_mapping(phys_addr_t phys_addr, size_t size,
535 unsigned long *haddr,
536 enum kvm_pgtable_prot prot)
541 if (!kvm_host_owns_hyp_mappings()) {
542 addr = kvm_call_hyp_nvhe(__pkvm_create_private_mapping,
543 phys_addr, size, prot);
544 if (IS_ERR_VALUE(addr))
551 size = PAGE_ALIGN(size + offset_in_page(phys_addr));
552 ret = hyp_alloc_private_va_range(size, &addr);
556 ret = __create_hyp_mappings(addr, size, phys_addr, prot);
560 *haddr = addr + offset_in_page(phys_addr);
565 * create_hyp_io_mappings - Map IO into both kernel and HYP
566 * @phys_addr: The physical start address which gets mapped
567 * @size: Size of the region being mapped
568 * @kaddr: Kernel VA for this mapping
569 * @haddr: HYP VA for this mapping
571 int create_hyp_io_mappings(phys_addr_t phys_addr, size_t size,
572 void __iomem **kaddr,
573 void __iomem **haddr)
578 if (is_protected_kvm_enabled())
581 *kaddr = ioremap(phys_addr, size);
585 if (is_kernel_in_hyp_mode()) {
590 ret = __create_hyp_private_mapping(phys_addr, size,
591 &addr, PAGE_HYP_DEVICE);
599 *haddr = (void __iomem *)addr;
604 * create_hyp_exec_mappings - Map an executable range into HYP
605 * @phys_addr: The physical start address which gets mapped
606 * @size: Size of the region being mapped
607 * @haddr: HYP VA for this mapping
609 int create_hyp_exec_mappings(phys_addr_t phys_addr, size_t size,
615 BUG_ON(is_kernel_in_hyp_mode());
617 ret = __create_hyp_private_mapping(phys_addr, size,
618 &addr, PAGE_HYP_EXEC);
624 *haddr = (void *)addr;
628 static struct kvm_pgtable_mm_ops kvm_user_mm_ops = {
629 /* We shouldn't need any other callback to walk the PT */
630 .phys_to_virt = kvm_host_va,
633 static int get_user_mapping_size(struct kvm *kvm, u64 addr)
635 struct kvm_pgtable pgt = {
636 .pgd = (kvm_pte_t *)kvm->mm->pgd,
638 .start_level = (KVM_PGTABLE_MAX_LEVELS -
639 CONFIG_PGTABLE_LEVELS),
640 .mm_ops = &kvm_user_mm_ops,
642 kvm_pte_t pte = 0; /* Keep GCC quiet... */
646 ret = kvm_pgtable_get_leaf(&pgt, addr, &pte, &level);
648 VM_BUG_ON(level >= KVM_PGTABLE_MAX_LEVELS);
649 VM_BUG_ON(!(pte & PTE_VALID));
651 return BIT(ARM64_HW_PGTABLE_LEVEL_SHIFT(level));
654 static struct kvm_pgtable_mm_ops kvm_s2_mm_ops = {
655 .zalloc_page = stage2_memcache_zalloc_page,
656 .zalloc_pages_exact = kvm_s2_zalloc_pages_exact,
657 .free_pages_exact = kvm_s2_free_pages_exact,
658 .get_page = kvm_host_get_page,
659 .put_page = kvm_s2_put_page,
660 .page_count = kvm_host_page_count,
661 .phys_to_virt = kvm_host_va,
662 .virt_to_phys = kvm_host_pa,
663 .dcache_clean_inval_poc = clean_dcache_guest_page,
664 .icache_inval_pou = invalidate_icache_guest_page,
668 * kvm_init_stage2_mmu - Initialise a S2 MMU structure
669 * @kvm: The pointer to the KVM structure
670 * @mmu: The pointer to the s2 MMU structure
672 * Allocates only the stage-2 HW PGD level table(s).
673 * Note we don't need locking here as this is only called when the VM is
674 * created, which can only be done once.
676 int kvm_init_stage2_mmu(struct kvm *kvm, struct kvm_s2_mmu *mmu)
679 struct kvm_pgtable *pgt;
681 if (mmu->pgt != NULL) {
682 kvm_err("kvm_arch already initialized?\n");
686 pgt = kzalloc(sizeof(*pgt), GFP_KERNEL_ACCOUNT);
690 mmu->arch = &kvm->arch;
691 err = kvm_pgtable_stage2_init(pgt, mmu, &kvm_s2_mm_ops);
693 goto out_free_pgtable;
695 mmu->last_vcpu_ran = alloc_percpu(typeof(*mmu->last_vcpu_ran));
696 if (!mmu->last_vcpu_ran) {
698 goto out_destroy_pgtable;
701 for_each_possible_cpu(cpu)
702 *per_cpu_ptr(mmu->last_vcpu_ran, cpu) = -1;
705 mmu->pgd_phys = __pa(pgt->pgd);
709 kvm_pgtable_stage2_destroy(pgt);
715 static void stage2_unmap_memslot(struct kvm *kvm,
716 struct kvm_memory_slot *memslot)
718 hva_t hva = memslot->userspace_addr;
719 phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
720 phys_addr_t size = PAGE_SIZE * memslot->npages;
721 hva_t reg_end = hva + size;
724 * A memory region could potentially cover multiple VMAs, and any holes
725 * between them, so iterate over all of them to find out if we should
728 * +--------------------------------------------+
729 * +---------------+----------------+ +----------------+
730 * | : VMA 1 | VMA 2 | | VMA 3 : |
731 * +---------------+----------------+ +----------------+
733 * +--------------------------------------------+
736 struct vm_area_struct *vma;
737 hva_t vm_start, vm_end;
739 vma = find_vma_intersection(current->mm, hva, reg_end);
744 * Take the intersection of this VMA with the memory region
746 vm_start = max(hva, vma->vm_start);
747 vm_end = min(reg_end, vma->vm_end);
749 if (!(vma->vm_flags & VM_PFNMAP)) {
750 gpa_t gpa = addr + (vm_start - memslot->userspace_addr);
751 unmap_stage2_range(&kvm->arch.mmu, gpa, vm_end - vm_start);
754 } while (hva < reg_end);
758 * stage2_unmap_vm - Unmap Stage-2 RAM mappings
759 * @kvm: The struct kvm pointer
761 * Go through the memregions and unmap any regular RAM
762 * backing memory already mapped to the VM.
764 void stage2_unmap_vm(struct kvm *kvm)
766 struct kvm_memslots *slots;
767 struct kvm_memory_slot *memslot;
770 idx = srcu_read_lock(&kvm->srcu);
771 mmap_read_lock(current->mm);
772 write_lock(&kvm->mmu_lock);
774 slots = kvm_memslots(kvm);
775 kvm_for_each_memslot(memslot, bkt, slots)
776 stage2_unmap_memslot(kvm, memslot);
778 write_unlock(&kvm->mmu_lock);
779 mmap_read_unlock(current->mm);
780 srcu_read_unlock(&kvm->srcu, idx);
783 void kvm_free_stage2_pgd(struct kvm_s2_mmu *mmu)
785 struct kvm *kvm = kvm_s2_mmu_to_kvm(mmu);
786 struct kvm_pgtable *pgt = NULL;
788 write_lock(&kvm->mmu_lock);
793 free_percpu(mmu->last_vcpu_ran);
795 write_unlock(&kvm->mmu_lock);
798 kvm_pgtable_stage2_destroy(pgt);
804 * kvm_phys_addr_ioremap - map a device range to guest IPA
806 * @kvm: The KVM pointer
807 * @guest_ipa: The IPA at which to insert the mapping
808 * @pa: The physical address of the device
809 * @size: The size of the mapping
810 * @writable: Whether or not to create a writable mapping
812 int kvm_phys_addr_ioremap(struct kvm *kvm, phys_addr_t guest_ipa,
813 phys_addr_t pa, unsigned long size, bool writable)
817 struct kvm_mmu_memory_cache cache = { .gfp_zero = __GFP_ZERO };
818 struct kvm_pgtable *pgt = kvm->arch.mmu.pgt;
819 enum kvm_pgtable_prot prot = KVM_PGTABLE_PROT_DEVICE |
821 (writable ? KVM_PGTABLE_PROT_W : 0);
823 if (is_protected_kvm_enabled())
826 size += offset_in_page(guest_ipa);
827 guest_ipa &= PAGE_MASK;
829 for (addr = guest_ipa; addr < guest_ipa + size; addr += PAGE_SIZE) {
830 ret = kvm_mmu_topup_memory_cache(&cache,
831 kvm_mmu_cache_min_pages(kvm));
835 write_lock(&kvm->mmu_lock);
836 ret = kvm_pgtable_stage2_map(pgt, addr, PAGE_SIZE, pa, prot,
838 write_unlock(&kvm->mmu_lock);
845 kvm_mmu_free_memory_cache(&cache);
850 * stage2_wp_range() - write protect stage2 memory region range
851 * @mmu: The KVM stage-2 MMU pointer
852 * @addr: Start address of range
853 * @end: End address of range
855 static void stage2_wp_range(struct kvm_s2_mmu *mmu, phys_addr_t addr, phys_addr_t end)
857 struct kvm *kvm = kvm_s2_mmu_to_kvm(mmu);
858 stage2_apply_range_resched(kvm, addr, end, kvm_pgtable_stage2_wrprotect);
862 * kvm_mmu_wp_memory_region() - write protect stage 2 entries for memory slot
863 * @kvm: The KVM pointer
864 * @slot: The memory slot to write protect
866 * Called to start logging dirty pages after memory region
867 * KVM_MEM_LOG_DIRTY_PAGES operation is called. After this function returns
868 * all present PUD, PMD and PTEs are write protected in the memory region.
869 * Afterwards read of dirty page log can be called.
871 * Acquires kvm_mmu_lock. Called with kvm->slots_lock mutex acquired,
872 * serializing operations for VM memory regions.
874 static void kvm_mmu_wp_memory_region(struct kvm *kvm, int slot)
876 struct kvm_memslots *slots = kvm_memslots(kvm);
877 struct kvm_memory_slot *memslot = id_to_memslot(slots, slot);
878 phys_addr_t start, end;
880 if (WARN_ON_ONCE(!memslot))
883 start = memslot->base_gfn << PAGE_SHIFT;
884 end = (memslot->base_gfn + memslot->npages) << PAGE_SHIFT;
886 write_lock(&kvm->mmu_lock);
887 stage2_wp_range(&kvm->arch.mmu, start, end);
888 write_unlock(&kvm->mmu_lock);
889 kvm_flush_remote_tlbs(kvm);
893 * kvm_mmu_write_protect_pt_masked() - write protect dirty pages
894 * @kvm: The KVM pointer
895 * @slot: The memory slot associated with mask
896 * @gfn_offset: The gfn offset in memory slot
897 * @mask: The mask of dirty pages at offset 'gfn_offset' in this memory
898 * slot to be write protected
900 * Walks bits set in mask write protects the associated pte's. Caller must
901 * acquire kvm_mmu_lock.
903 static void kvm_mmu_write_protect_pt_masked(struct kvm *kvm,
904 struct kvm_memory_slot *slot,
905 gfn_t gfn_offset, unsigned long mask)
907 phys_addr_t base_gfn = slot->base_gfn + gfn_offset;
908 phys_addr_t start = (base_gfn + __ffs(mask)) << PAGE_SHIFT;
909 phys_addr_t end = (base_gfn + __fls(mask) + 1) << PAGE_SHIFT;
911 stage2_wp_range(&kvm->arch.mmu, start, end);
915 * kvm_arch_mmu_enable_log_dirty_pt_masked - enable dirty logging for selected
918 * It calls kvm_mmu_write_protect_pt_masked to write protect selected pages to
919 * enable dirty logging for them.
921 void kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm *kvm,
922 struct kvm_memory_slot *slot,
923 gfn_t gfn_offset, unsigned long mask)
925 kvm_mmu_write_protect_pt_masked(kvm, slot, gfn_offset, mask);
928 static void kvm_send_hwpoison_signal(unsigned long address, short lsb)
930 send_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, lsb, current);
933 static bool fault_supports_stage2_huge_mapping(struct kvm_memory_slot *memslot,
935 unsigned long map_size)
938 hva_t uaddr_start, uaddr_end;
941 /* The memslot and the VMA are guaranteed to be aligned to PAGE_SIZE */
942 if (map_size == PAGE_SIZE)
945 size = memslot->npages * PAGE_SIZE;
947 gpa_start = memslot->base_gfn << PAGE_SHIFT;
949 uaddr_start = memslot->userspace_addr;
950 uaddr_end = uaddr_start + size;
953 * Pages belonging to memslots that don't have the same alignment
954 * within a PMD/PUD for userspace and IPA cannot be mapped with stage-2
955 * PMD/PUD entries, because we'll end up mapping the wrong pages.
957 * Consider a layout like the following:
959 * memslot->userspace_addr:
960 * +-----+--------------------+--------------------+---+
961 * |abcde|fgh Stage-1 block | Stage-1 block tv|xyz|
962 * +-----+--------------------+--------------------+---+
964 * memslot->base_gfn << PAGE_SHIFT:
965 * +---+--------------------+--------------------+-----+
966 * |abc|def Stage-2 block | Stage-2 block |tvxyz|
967 * +---+--------------------+--------------------+-----+
969 * If we create those stage-2 blocks, we'll end up with this incorrect
975 if ((gpa_start & (map_size - 1)) != (uaddr_start & (map_size - 1)))
979 * Next, let's make sure we're not trying to map anything not covered
980 * by the memslot. This means we have to prohibit block size mappings
981 * for the beginning and end of a non-block aligned and non-block sized
982 * memory slot (illustrated by the head and tail parts of the
983 * userspace view above containing pages 'abcde' and 'xyz',
986 * Note that it doesn't matter if we do the check using the
987 * userspace_addr or the base_gfn, as both are equally aligned (per
988 * the check above) and equally sized.
990 return (hva & ~(map_size - 1)) >= uaddr_start &&
991 (hva & ~(map_size - 1)) + map_size <= uaddr_end;
995 * Check if the given hva is backed by a transparent huge page (THP) and
996 * whether it can be mapped using block mapping in stage2. If so, adjust
997 * the stage2 PFN and IPA accordingly. Only PMD_SIZE THPs are currently
998 * supported. This will need to be updated to support other THP sizes.
1000 * Returns the size of the mapping.
1002 static unsigned long
1003 transparent_hugepage_adjust(struct kvm *kvm, struct kvm_memory_slot *memslot,
1004 unsigned long hva, kvm_pfn_t *pfnp,
1007 kvm_pfn_t pfn = *pfnp;
1010 * Make sure the adjustment is done only for THP pages. Also make
1011 * sure that the HVA and IPA are sufficiently aligned and that the
1012 * block map is contained within the memslot.
1014 if (fault_supports_stage2_huge_mapping(memslot, hva, PMD_SIZE) &&
1015 get_user_mapping_size(kvm, hva) >= PMD_SIZE) {
1017 * The address we faulted on is backed by a transparent huge
1018 * page. However, because we map the compound huge page and
1019 * not the individual tail page, we need to transfer the
1020 * refcount to the head page. We have to be careful that the
1021 * THP doesn't start to split while we are adjusting the
1024 * We are sure this doesn't happen, because mmu_invalidate_retry
1025 * was successful and we are holding the mmu_lock, so if this
1026 * THP is trying to split, it will be blocked in the mmu
1027 * notifier before touching any of the pages, specifically
1028 * before being able to call __split_huge_page_refcount().
1030 * We can therefore safely transfer the refcount from PG_tail
1031 * to PG_head and switch the pfn from a tail page to the head
1035 kvm_release_pfn_clean(pfn);
1036 pfn &= ~(PTRS_PER_PMD - 1);
1037 get_page(pfn_to_page(pfn));
1043 /* Use page mapping if we cannot use block mapping. */
1047 static int get_vma_page_shift(struct vm_area_struct *vma, unsigned long hva)
1051 if (is_vm_hugetlb_page(vma) && !(vma->vm_flags & VM_PFNMAP))
1052 return huge_page_shift(hstate_vma(vma));
1054 if (!(vma->vm_flags & VM_PFNMAP))
1057 VM_BUG_ON(is_vm_hugetlb_page(vma));
1059 pa = (vma->vm_pgoff << PAGE_SHIFT) + (hva - vma->vm_start);
1061 #ifndef __PAGETABLE_PMD_FOLDED
1062 if ((hva & (PUD_SIZE - 1)) == (pa & (PUD_SIZE - 1)) &&
1063 ALIGN_DOWN(hva, PUD_SIZE) >= vma->vm_start &&
1064 ALIGN(hva, PUD_SIZE) <= vma->vm_end)
1068 if ((hva & (PMD_SIZE - 1)) == (pa & (PMD_SIZE - 1)) &&
1069 ALIGN_DOWN(hva, PMD_SIZE) >= vma->vm_start &&
1070 ALIGN(hva, PMD_SIZE) <= vma->vm_end)
1077 * The page will be mapped in stage 2 as Normal Cacheable, so the VM will be
1078 * able to see the page's tags and therefore they must be initialised first. If
1079 * PG_mte_tagged is set, tags have already been initialised.
1081 * The race in the test/set of the PG_mte_tagged flag is handled by:
1082 * - preventing VM_SHARED mappings in a memslot with MTE preventing two VMs
1083 * racing to santise the same page
1084 * - mmap_lock protects between a VM faulting a page in and the VMM performing
1085 * an mprotect() to add VM_MTE
1087 static int sanitise_mte_tags(struct kvm *kvm, kvm_pfn_t pfn,
1090 unsigned long i, nr_pages = size >> PAGE_SHIFT;
1093 if (!kvm_has_mte(kvm))
1097 * pfn_to_online_page() is used to reject ZONE_DEVICE pages
1098 * that may not support tags.
1100 page = pfn_to_online_page(pfn);
1105 for (i = 0; i < nr_pages; i++, page++) {
1106 if (!test_bit(PG_mte_tagged, &page->flags)) {
1107 mte_clear_page_tags(page_address(page));
1108 set_bit(PG_mte_tagged, &page->flags);
1115 static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
1116 struct kvm_memory_slot *memslot, unsigned long hva,
1117 unsigned long fault_status)
1120 bool write_fault, writable, force_pte = false;
1122 bool device = false;
1124 unsigned long mmu_seq;
1125 struct kvm *kvm = vcpu->kvm;
1126 struct kvm_mmu_memory_cache *memcache = &vcpu->arch.mmu_page_cache;
1127 struct vm_area_struct *vma;
1131 bool logging_active = memslot_is_logging(memslot);
1132 bool use_read_lock = false;
1133 unsigned long fault_level = kvm_vcpu_trap_get_fault_level(vcpu);
1134 unsigned long vma_pagesize, fault_granule;
1135 enum kvm_pgtable_prot prot = KVM_PGTABLE_PROT_R;
1136 struct kvm_pgtable *pgt;
1138 fault_granule = 1UL << ARM64_HW_PGTABLE_LEVEL_SHIFT(fault_level);
1139 write_fault = kvm_is_write_fault(vcpu);
1140 exec_fault = kvm_vcpu_trap_is_exec_fault(vcpu);
1141 VM_BUG_ON(write_fault && exec_fault);
1143 if (fault_status == FSC_PERM && !write_fault && !exec_fault) {
1144 kvm_err("Unexpected L2 read permission error\n");
1149 * Let's check if we will get back a huge page backed by hugetlbfs, or
1150 * get block mapping for device MMIO region.
1152 mmap_read_lock(current->mm);
1153 vma = vma_lookup(current->mm, hva);
1154 if (unlikely(!vma)) {
1155 kvm_err("Failed to find VMA for hva 0x%lx\n", hva);
1156 mmap_read_unlock(current->mm);
1161 * logging_active is guaranteed to never be true for VM_PFNMAP
1164 if (logging_active) {
1166 vma_shift = PAGE_SHIFT;
1167 use_read_lock = (fault_status == FSC_PERM && write_fault &&
1168 fault_granule == PAGE_SIZE);
1170 vma_shift = get_vma_page_shift(vma, hva);
1173 shared = (vma->vm_flags & VM_SHARED);
1175 switch (vma_shift) {
1176 #ifndef __PAGETABLE_PMD_FOLDED
1178 if (fault_supports_stage2_huge_mapping(memslot, hva, PUD_SIZE))
1182 case CONT_PMD_SHIFT:
1183 vma_shift = PMD_SHIFT;
1186 if (fault_supports_stage2_huge_mapping(memslot, hva, PMD_SIZE))
1189 case CONT_PTE_SHIFT:
1190 vma_shift = PAGE_SHIFT;
1196 WARN_ONCE(1, "Unknown vma_shift %d", vma_shift);
1199 vma_pagesize = 1UL << vma_shift;
1200 if (vma_pagesize == PMD_SIZE || vma_pagesize == PUD_SIZE)
1201 fault_ipa &= ~(vma_pagesize - 1);
1203 gfn = fault_ipa >> PAGE_SHIFT;
1204 mmap_read_unlock(current->mm);
1207 * Permission faults just need to update the existing leaf entry,
1208 * and so normally don't require allocations from the memcache. The
1209 * only exception to this is when dirty logging is enabled at runtime
1210 * and a write fault needs to collapse a block entry into a table.
1212 if (fault_status != FSC_PERM || (logging_active && write_fault)) {
1213 ret = kvm_mmu_topup_memory_cache(memcache,
1214 kvm_mmu_cache_min_pages(kvm));
1219 mmu_seq = vcpu->kvm->mmu_invalidate_seq;
1221 * Ensure the read of mmu_invalidate_seq happens before we call
1222 * gfn_to_pfn_prot (which calls get_user_pages), so that we don't risk
1223 * the page we just got a reference to gets unmapped before we have a
1224 * chance to grab the mmu_lock, which ensure that if the page gets
1225 * unmapped afterwards, the call to kvm_unmap_gfn will take it away
1226 * from us again properly. This smp_rmb() interacts with the smp_wmb()
1227 * in kvm_mmu_notifier_invalidate_<page|range_end>.
1229 * Besides, __gfn_to_pfn_memslot() instead of gfn_to_pfn_prot() is
1230 * used to avoid unnecessary overhead introduced to locate the memory
1231 * slot because it's always fixed even @gfn is adjusted for huge pages.
1235 pfn = __gfn_to_pfn_memslot(memslot, gfn, false, NULL,
1236 write_fault, &writable, NULL);
1237 if (pfn == KVM_PFN_ERR_HWPOISON) {
1238 kvm_send_hwpoison_signal(hva, vma_shift);
1241 if (is_error_noslot_pfn(pfn))
1244 if (kvm_is_device_pfn(pfn)) {
1246 * If the page was identified as device early by looking at
1247 * the VMA flags, vma_pagesize is already representing the
1248 * largest quantity we can map. If instead it was mapped
1249 * via gfn_to_pfn_prot(), vma_pagesize is set to PAGE_SIZE
1250 * and must not be upgraded.
1252 * In both cases, we don't let transparent_hugepage_adjust()
1253 * change things at the last minute.
1256 } else if (logging_active && !write_fault) {
1258 * Only actually map the page as writable if this was a write
1264 if (exec_fault && device)
1268 * To reduce MMU contentions and enhance concurrency during dirty
1269 * logging dirty logging, only acquire read lock for permission
1273 read_lock(&kvm->mmu_lock);
1275 write_lock(&kvm->mmu_lock);
1276 pgt = vcpu->arch.hw_mmu->pgt;
1277 if (mmu_invalidate_retry(kvm, mmu_seq))
1281 * If we are not forced to use page mapping, check if we are
1282 * backed by a THP and thus use block mapping if possible.
1284 if (vma_pagesize == PAGE_SIZE && !(force_pte || device)) {
1285 if (fault_status == FSC_PERM && fault_granule > PAGE_SIZE)
1286 vma_pagesize = fault_granule;
1288 vma_pagesize = transparent_hugepage_adjust(kvm, memslot,
1293 if (fault_status != FSC_PERM && !device && kvm_has_mte(kvm)) {
1294 /* Check the VMM hasn't introduced a new VM_SHARED VMA */
1296 ret = sanitise_mte_tags(kvm, pfn, vma_pagesize);
1304 prot |= KVM_PGTABLE_PROT_W;
1307 prot |= KVM_PGTABLE_PROT_X;
1310 prot |= KVM_PGTABLE_PROT_DEVICE;
1311 else if (cpus_have_const_cap(ARM64_HAS_CACHE_DIC))
1312 prot |= KVM_PGTABLE_PROT_X;
1315 * Under the premise of getting a FSC_PERM fault, we just need to relax
1316 * permissions only if vma_pagesize equals fault_granule. Otherwise,
1317 * kvm_pgtable_stage2_map() should be called to change block size.
1319 if (fault_status == FSC_PERM && vma_pagesize == fault_granule) {
1320 ret = kvm_pgtable_stage2_relax_perms(pgt, fault_ipa, prot);
1322 WARN_ONCE(use_read_lock, "Attempted stage-2 map outside of write lock\n");
1324 ret = kvm_pgtable_stage2_map(pgt, fault_ipa, vma_pagesize,
1325 __pfn_to_phys(pfn), prot,
1329 /* Mark the page dirty only if the fault is handled successfully */
1330 if (writable && !ret) {
1331 kvm_set_pfn_dirty(pfn);
1332 mark_page_dirty_in_slot(kvm, memslot, gfn);
1337 read_unlock(&kvm->mmu_lock);
1339 write_unlock(&kvm->mmu_lock);
1340 kvm_set_pfn_accessed(pfn);
1341 kvm_release_pfn_clean(pfn);
1342 return ret != -EAGAIN ? ret : 0;
1345 /* Resolve the access fault by making the page young again. */
1346 static void handle_access_fault(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa)
1350 struct kvm_s2_mmu *mmu;
1352 trace_kvm_access_fault(fault_ipa);
1354 write_lock(&vcpu->kvm->mmu_lock);
1355 mmu = vcpu->arch.hw_mmu;
1356 kpte = kvm_pgtable_stage2_mkyoung(mmu->pgt, fault_ipa);
1357 write_unlock(&vcpu->kvm->mmu_lock);
1361 kvm_set_pfn_accessed(pte_pfn(pte));
1365 * kvm_handle_guest_abort - handles all 2nd stage aborts
1366 * @vcpu: the VCPU pointer
1368 * Any abort that gets to the host is almost guaranteed to be caused by a
1369 * missing second stage translation table entry, which can mean that either the
1370 * guest simply needs more memory and we must allocate an appropriate page or it
1371 * can mean that the guest tried to access I/O memory, which is emulated by user
1372 * space. The distinction is based on the IPA causing the fault and whether this
1373 * memory region has been registered as standard RAM by user space.
1375 int kvm_handle_guest_abort(struct kvm_vcpu *vcpu)
1377 unsigned long fault_status;
1378 phys_addr_t fault_ipa;
1379 struct kvm_memory_slot *memslot;
1381 bool is_iabt, write_fault, writable;
1385 fault_status = kvm_vcpu_trap_get_fault_type(vcpu);
1387 fault_ipa = kvm_vcpu_get_fault_ipa(vcpu);
1388 is_iabt = kvm_vcpu_trap_is_iabt(vcpu);
1390 if (fault_status == FSC_FAULT) {
1391 /* Beyond sanitised PARange (which is the IPA limit) */
1392 if (fault_ipa >= BIT_ULL(get_kvm_ipa_limit())) {
1393 kvm_inject_size_fault(vcpu);
1397 /* Falls between the IPA range and the PARange? */
1398 if (fault_ipa >= BIT_ULL(vcpu->arch.hw_mmu->pgt->ia_bits)) {
1399 fault_ipa |= kvm_vcpu_get_hfar(vcpu) & GENMASK(11, 0);
1402 kvm_inject_pabt(vcpu, fault_ipa);
1404 kvm_inject_dabt(vcpu, fault_ipa);
1409 /* Synchronous External Abort? */
1410 if (kvm_vcpu_abt_issea(vcpu)) {
1412 * For RAS the host kernel may handle this abort.
1413 * There is no need to pass the error into the guest.
1415 if (kvm_handle_guest_sea(fault_ipa, kvm_vcpu_get_esr(vcpu)))
1416 kvm_inject_vabt(vcpu);
1421 trace_kvm_guest_fault(*vcpu_pc(vcpu), kvm_vcpu_get_esr(vcpu),
1422 kvm_vcpu_get_hfar(vcpu), fault_ipa);
1424 /* Check the stage-2 fault is trans. fault or write fault */
1425 if (fault_status != FSC_FAULT && fault_status != FSC_PERM &&
1426 fault_status != FSC_ACCESS) {
1427 kvm_err("Unsupported FSC: EC=%#x xFSC=%#lx ESR_EL2=%#lx\n",
1428 kvm_vcpu_trap_get_class(vcpu),
1429 (unsigned long)kvm_vcpu_trap_get_fault(vcpu),
1430 (unsigned long)kvm_vcpu_get_esr(vcpu));
1434 idx = srcu_read_lock(&vcpu->kvm->srcu);
1436 gfn = fault_ipa >> PAGE_SHIFT;
1437 memslot = gfn_to_memslot(vcpu->kvm, gfn);
1438 hva = gfn_to_hva_memslot_prot(memslot, gfn, &writable);
1439 write_fault = kvm_is_write_fault(vcpu);
1440 if (kvm_is_error_hva(hva) || (write_fault && !writable)) {
1442 * The guest has put either its instructions or its page-tables
1443 * somewhere it shouldn't have. Userspace won't be able to do
1444 * anything about this (there's no syndrome for a start), so
1445 * re-inject the abort back into the guest.
1452 if (kvm_vcpu_abt_iss1tw(vcpu)) {
1453 kvm_inject_dabt(vcpu, kvm_vcpu_get_hfar(vcpu));
1459 * Check for a cache maintenance operation. Since we
1460 * ended-up here, we know it is outside of any memory
1461 * slot. But we can't find out if that is for a device,
1462 * or if the guest is just being stupid. The only thing
1463 * we know for sure is that this range cannot be cached.
1465 * So let's assume that the guest is just being
1466 * cautious, and skip the instruction.
1468 if (kvm_is_error_hva(hva) && kvm_vcpu_dabt_is_cm(vcpu)) {
1475 * The IPA is reported as [MAX:12], so we need to
1476 * complement it with the bottom 12 bits from the
1477 * faulting VA. This is always 12 bits, irrespective
1480 fault_ipa |= kvm_vcpu_get_hfar(vcpu) & ((1 << 12) - 1);
1481 ret = io_mem_abort(vcpu, fault_ipa);
1485 /* Userspace should not be able to register out-of-bounds IPAs */
1486 VM_BUG_ON(fault_ipa >= kvm_phys_size(vcpu->kvm));
1488 if (fault_status == FSC_ACCESS) {
1489 handle_access_fault(vcpu, fault_ipa);
1494 ret = user_mem_abort(vcpu, fault_ipa, memslot, hva, fault_status);
1498 if (ret == -ENOEXEC) {
1499 kvm_inject_pabt(vcpu, kvm_vcpu_get_hfar(vcpu));
1503 srcu_read_unlock(&vcpu->kvm->srcu, idx);
1507 bool kvm_unmap_gfn_range(struct kvm *kvm, struct kvm_gfn_range *range)
1509 if (!kvm->arch.mmu.pgt)
1512 __unmap_stage2_range(&kvm->arch.mmu, range->start << PAGE_SHIFT,
1513 (range->end - range->start) << PAGE_SHIFT,
1519 bool kvm_set_spte_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
1521 kvm_pfn_t pfn = pte_pfn(range->pte);
1524 if (!kvm->arch.mmu.pgt)
1527 WARN_ON(range->end - range->start != 1);
1529 ret = sanitise_mte_tags(kvm, pfn, PAGE_SIZE);
1534 * We've moved a page around, probably through CoW, so let's treat
1535 * it just like a translation fault and the map handler will clean
1536 * the cache to the PoC.
1538 * The MMU notifiers will have unmapped a huge PMD before calling
1539 * ->change_pte() (which in turn calls kvm_set_spte_gfn()) and
1540 * therefore we never need to clear out a huge PMD through this
1541 * calling path and a memcache is not required.
1543 kvm_pgtable_stage2_map(kvm->arch.mmu.pgt, range->start << PAGE_SHIFT,
1544 PAGE_SIZE, __pfn_to_phys(pfn),
1545 KVM_PGTABLE_PROT_R, NULL);
1550 bool kvm_age_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
1552 u64 size = (range->end - range->start) << PAGE_SHIFT;
1556 if (!kvm->arch.mmu.pgt)
1559 WARN_ON(size != PAGE_SIZE && size != PMD_SIZE && size != PUD_SIZE);
1561 kpte = kvm_pgtable_stage2_mkold(kvm->arch.mmu.pgt,
1562 range->start << PAGE_SHIFT);
1564 return pte_valid(pte) && pte_young(pte);
1567 bool kvm_test_age_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
1569 if (!kvm->arch.mmu.pgt)
1572 return kvm_pgtable_stage2_is_young(kvm->arch.mmu.pgt,
1573 range->start << PAGE_SHIFT);
1576 phys_addr_t kvm_mmu_get_httbr(void)
1578 return __pa(hyp_pgtable->pgd);
1581 phys_addr_t kvm_get_idmap_vector(void)
1583 return hyp_idmap_vector;
1586 static int kvm_map_idmap_text(void)
1588 unsigned long size = hyp_idmap_end - hyp_idmap_start;
1589 int err = __create_hyp_mappings(hyp_idmap_start, size, hyp_idmap_start,
1592 kvm_err("Failed to idmap %lx-%lx\n",
1593 hyp_idmap_start, hyp_idmap_end);
1598 static void *kvm_hyp_zalloc_page(void *arg)
1600 return (void *)get_zeroed_page(GFP_KERNEL);
1603 static struct kvm_pgtable_mm_ops kvm_hyp_mm_ops = {
1604 .zalloc_page = kvm_hyp_zalloc_page,
1605 .get_page = kvm_host_get_page,
1606 .put_page = kvm_host_put_page,
1607 .phys_to_virt = kvm_host_va,
1608 .virt_to_phys = kvm_host_pa,
1611 int kvm_mmu_init(u32 *hyp_va_bits)
1615 hyp_idmap_start = __pa_symbol(__hyp_idmap_text_start);
1616 hyp_idmap_start = ALIGN_DOWN(hyp_idmap_start, PAGE_SIZE);
1617 hyp_idmap_end = __pa_symbol(__hyp_idmap_text_end);
1618 hyp_idmap_end = ALIGN(hyp_idmap_end, PAGE_SIZE);
1619 hyp_idmap_vector = __pa_symbol(__kvm_hyp_init);
1622 * We rely on the linker script to ensure at build time that the HYP
1623 * init code does not cross a page boundary.
1625 BUG_ON((hyp_idmap_start ^ (hyp_idmap_end - 1)) & PAGE_MASK);
1627 *hyp_va_bits = 64 - ((idmap_t0sz & TCR_T0SZ_MASK) >> TCR_T0SZ_OFFSET);
1628 kvm_debug("Using %u-bit virtual addresses at EL2\n", *hyp_va_bits);
1629 kvm_debug("IDMAP page: %lx\n", hyp_idmap_start);
1630 kvm_debug("HYP VA range: %lx:%lx\n",
1631 kern_hyp_va(PAGE_OFFSET),
1632 kern_hyp_va((unsigned long)high_memory - 1));
1634 if (hyp_idmap_start >= kern_hyp_va(PAGE_OFFSET) &&
1635 hyp_idmap_start < kern_hyp_va((unsigned long)high_memory - 1) &&
1636 hyp_idmap_start != (unsigned long)__hyp_idmap_text_start) {
1638 * The idmap page is intersecting with the VA space,
1639 * it is not safe to continue further.
1641 kvm_err("IDMAP intersecting with HYP VA, unable to continue\n");
1646 hyp_pgtable = kzalloc(sizeof(*hyp_pgtable), GFP_KERNEL);
1648 kvm_err("Hyp mode page-table not allocated\n");
1653 err = kvm_pgtable_hyp_init(hyp_pgtable, *hyp_va_bits, &kvm_hyp_mm_ops);
1655 goto out_free_pgtable;
1657 err = kvm_map_idmap_text();
1659 goto out_destroy_pgtable;
1661 io_map_base = hyp_idmap_start;
1664 out_destroy_pgtable:
1665 kvm_pgtable_hyp_destroy(hyp_pgtable);
1673 void kvm_arch_commit_memory_region(struct kvm *kvm,
1674 struct kvm_memory_slot *old,
1675 const struct kvm_memory_slot *new,
1676 enum kvm_mr_change change)
1679 * At this point memslot has been committed and there is an
1680 * allocated dirty_bitmap[], dirty pages will be tracked while the
1681 * memory slot is write protected.
1683 if (change != KVM_MR_DELETE && new->flags & KVM_MEM_LOG_DIRTY_PAGES) {
1685 * If we're with initial-all-set, we don't need to write
1686 * protect any pages because they're all reported as dirty.
1687 * Huge pages and normal pages will be write protect gradually.
1689 if (!kvm_dirty_log_manual_protect_and_init_set(kvm)) {
1690 kvm_mmu_wp_memory_region(kvm, new->id);
1695 int kvm_arch_prepare_memory_region(struct kvm *kvm,
1696 const struct kvm_memory_slot *old,
1697 struct kvm_memory_slot *new,
1698 enum kvm_mr_change change)
1703 if (change != KVM_MR_CREATE && change != KVM_MR_MOVE &&
1704 change != KVM_MR_FLAGS_ONLY)
1708 * Prevent userspace from creating a memory region outside of the IPA
1709 * space addressable by the KVM guest IPA space.
1711 if ((new->base_gfn + new->npages) > (kvm_phys_size(kvm) >> PAGE_SHIFT))
1714 hva = new->userspace_addr;
1715 reg_end = hva + (new->npages << PAGE_SHIFT);
1717 mmap_read_lock(current->mm);
1719 * A memory region could potentially cover multiple VMAs, and any holes
1720 * between them, so iterate over all of them.
1722 * +--------------------------------------------+
1723 * +---------------+----------------+ +----------------+
1724 * | : VMA 1 | VMA 2 | | VMA 3 : |
1725 * +---------------+----------------+ +----------------+
1727 * +--------------------------------------------+
1730 struct vm_area_struct *vma;
1732 vma = find_vma_intersection(current->mm, hva, reg_end);
1737 * VM_SHARED mappings are not allowed with MTE to avoid races
1738 * when updating the PG_mte_tagged page flag, see
1739 * sanitise_mte_tags for more details.
1741 if (kvm_has_mte(kvm) && vma->vm_flags & VM_SHARED) {
1746 if (vma->vm_flags & VM_PFNMAP) {
1747 /* IO region dirty page logging not allowed */
1748 if (new->flags & KVM_MEM_LOG_DIRTY_PAGES) {
1753 hva = min(reg_end, vma->vm_end);
1754 } while (hva < reg_end);
1756 mmap_read_unlock(current->mm);
1760 void kvm_arch_free_memslot(struct kvm *kvm, struct kvm_memory_slot *slot)
1764 void kvm_arch_memslots_updated(struct kvm *kvm, u64 gen)
1768 void kvm_arch_flush_shadow_all(struct kvm *kvm)
1770 kvm_free_stage2_pgd(&kvm->arch.mmu);
1773 void kvm_arch_flush_shadow_memslot(struct kvm *kvm,
1774 struct kvm_memory_slot *slot)
1776 gpa_t gpa = slot->base_gfn << PAGE_SHIFT;
1777 phys_addr_t size = slot->npages << PAGE_SHIFT;
1779 write_lock(&kvm->mmu_lock);
1780 unmap_stage2_range(&kvm->arch.mmu, gpa, size);
1781 write_unlock(&kvm->mmu_lock);
1785 * See note at ARMv7 ARM B1.14.4 (TL;DR: S/W ops are not easily virtualized).
1788 * - S/W ops are local to a CPU (not broadcast)
1789 * - We have line migration behind our back (speculation)
1790 * - System caches don't support S/W at all (damn!)
1792 * In the face of the above, the best we can do is to try and convert
1793 * S/W ops to VA ops. Because the guest is not allowed to infer the
1794 * S/W to PA mapping, it can only use S/W to nuke the whole cache,
1795 * which is a rather good thing for us.
1797 * Also, it is only used when turning caches on/off ("The expected
1798 * usage of the cache maintenance instructions that operate by set/way
1799 * is associated with the cache maintenance instructions associated
1800 * with the powerdown and powerup of caches, if this is required by
1801 * the implementation.").
1803 * We use the following policy:
1805 * - If we trap a S/W operation, we enable VM trapping to detect
1806 * caches being turned on/off, and do a full clean.
1808 * - We flush the caches on both caches being turned on and off.
1810 * - Once the caches are enabled, we stop trapping VM ops.
1812 void kvm_set_way_flush(struct kvm_vcpu *vcpu)
1814 unsigned long hcr = *vcpu_hcr(vcpu);
1817 * If this is the first time we do a S/W operation
1818 * (i.e. HCR_TVM not set) flush the whole memory, and set the
1821 * Otherwise, rely on the VM trapping to wait for the MMU +
1822 * Caches to be turned off. At that point, we'll be able to
1823 * clean the caches again.
1825 if (!(hcr & HCR_TVM)) {
1826 trace_kvm_set_way_flush(*vcpu_pc(vcpu),
1827 vcpu_has_cache_enabled(vcpu));
1828 stage2_flush_vm(vcpu->kvm);
1829 *vcpu_hcr(vcpu) = hcr | HCR_TVM;
1833 void kvm_toggle_cache(struct kvm_vcpu *vcpu, bool was_enabled)
1835 bool now_enabled = vcpu_has_cache_enabled(vcpu);
1838 * If switching the MMU+caches on, need to invalidate the caches.
1839 * If switching it off, need to clean the caches.
1840 * Clean + invalidate does the trick always.
1842 if (now_enabled != was_enabled)
1843 stage2_flush_vm(vcpu->kvm);
1845 /* Caches are now on, stop trapping VM ops (until a S/W op) */
1847 *vcpu_hcr(vcpu) &= ~HCR_TVM;
1849 trace_kvm_toggle_cache(*vcpu_pc(vcpu), was_enabled, now_enabled);