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_lock(&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_call_hyp(__kvm_tlb_flush_vmid, &kvm->arch.mmu);
86 static bool kvm_is_device_pfn(unsigned long pfn)
88 return !pfn_valid(pfn);
91 static void *stage2_memcache_zalloc_page(void *arg)
93 struct kvm_mmu_memory_cache *mc = arg;
95 /* Allocated with __GFP_ZERO, so no need to zero */
96 return kvm_mmu_memory_cache_alloc(mc);
99 static void *kvm_host_zalloc_pages_exact(size_t size)
101 return alloc_pages_exact(size, GFP_KERNEL_ACCOUNT | __GFP_ZERO);
104 static void kvm_host_get_page(void *addr)
106 get_page(virt_to_page(addr));
109 static void kvm_host_put_page(void *addr)
111 put_page(virt_to_page(addr));
114 static int kvm_host_page_count(void *addr)
116 return page_count(virt_to_page(addr));
119 static phys_addr_t kvm_host_pa(void *addr)
124 static void *kvm_host_va(phys_addr_t phys)
130 * Unmapping vs dcache management:
132 * If a guest maps certain memory pages as uncached, all writes will
133 * bypass the data cache and go directly to RAM. However, the CPUs
134 * can still speculate reads (not writes) and fill cache lines with
137 * Those cache lines will be *clean* cache lines though, so a
138 * clean+invalidate operation is equivalent to an invalidate
139 * operation, because no cache lines are marked dirty.
141 * Those clean cache lines could be filled prior to an uncached write
142 * by the guest, and the cache coherent IO subsystem would therefore
143 * end up writing old data to disk.
145 * This is why right after unmapping a page/section and invalidating
146 * the corresponding TLBs, we flush to make sure the IO subsystem will
147 * never hit in the cache.
149 * This is all avoided on systems that have ARM64_HAS_STAGE2_FWB, as
150 * we then fully enforce cacheability of RAM, no matter what the guest
154 * unmap_stage2_range -- Clear stage2 page table entries to unmap a range
155 * @mmu: The KVM stage-2 MMU pointer
156 * @start: The intermediate physical base address of the range to unmap
157 * @size: The size of the area to unmap
158 * @may_block: Whether or not we are permitted to block
160 * Clear a range of stage-2 mappings, lowering the various ref-counts. Must
161 * be called while holding mmu_lock (unless for freeing the stage2 pgd before
162 * destroying the VM), otherwise another faulting VCPU may come in and mess
163 * with things behind our backs.
165 static void __unmap_stage2_range(struct kvm_s2_mmu *mmu, phys_addr_t start, u64 size,
168 struct kvm *kvm = kvm_s2_mmu_to_kvm(mmu);
169 phys_addr_t end = start + size;
171 assert_spin_locked(&kvm->mmu_lock);
172 WARN_ON(size & ~PAGE_MASK);
173 WARN_ON(stage2_apply_range(kvm, start, end, kvm_pgtable_stage2_unmap,
177 static void unmap_stage2_range(struct kvm_s2_mmu *mmu, phys_addr_t start, u64 size)
179 __unmap_stage2_range(mmu, start, size, true);
182 static void stage2_flush_memslot(struct kvm *kvm,
183 struct kvm_memory_slot *memslot)
185 phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
186 phys_addr_t end = addr + PAGE_SIZE * memslot->npages;
188 stage2_apply_range_resched(kvm, addr, end, kvm_pgtable_stage2_flush);
192 * stage2_flush_vm - Invalidate cache for pages mapped in stage 2
193 * @kvm: The struct kvm pointer
195 * Go through the stage 2 page tables and invalidate any cache lines
196 * backing memory already mapped to the VM.
198 static void stage2_flush_vm(struct kvm *kvm)
200 struct kvm_memslots *slots;
201 struct kvm_memory_slot *memslot;
204 idx = srcu_read_lock(&kvm->srcu);
205 spin_lock(&kvm->mmu_lock);
207 slots = kvm_memslots(kvm);
208 kvm_for_each_memslot(memslot, slots)
209 stage2_flush_memslot(kvm, memslot);
211 spin_unlock(&kvm->mmu_lock);
212 srcu_read_unlock(&kvm->srcu, idx);
216 * free_hyp_pgds - free Hyp-mode page tables
218 void free_hyp_pgds(void)
220 mutex_lock(&kvm_hyp_pgd_mutex);
222 kvm_pgtable_hyp_destroy(hyp_pgtable);
226 mutex_unlock(&kvm_hyp_pgd_mutex);
229 static bool kvm_host_owns_hyp_mappings(void)
231 if (static_branch_likely(&kvm_protected_mode_initialized))
235 * This can happen at boot time when __create_hyp_mappings() is called
236 * after the hyp protection has been enabled, but the static key has
237 * not been flipped yet.
239 if (!hyp_pgtable && is_protected_kvm_enabled())
242 WARN_ON(!hyp_pgtable);
247 static int __create_hyp_mappings(unsigned long start, unsigned long size,
248 unsigned long phys, enum kvm_pgtable_prot prot)
252 if (!kvm_host_owns_hyp_mappings()) {
253 return kvm_call_hyp_nvhe(__pkvm_create_mappings,
254 start, size, phys, prot);
257 mutex_lock(&kvm_hyp_pgd_mutex);
258 err = kvm_pgtable_hyp_map(hyp_pgtable, start, size, phys, prot);
259 mutex_unlock(&kvm_hyp_pgd_mutex);
264 static phys_addr_t kvm_kaddr_to_phys(void *kaddr)
266 if (!is_vmalloc_addr(kaddr)) {
267 BUG_ON(!virt_addr_valid(kaddr));
270 return page_to_phys(vmalloc_to_page(kaddr)) +
271 offset_in_page(kaddr);
276 * create_hyp_mappings - duplicate a kernel virtual address range in Hyp mode
277 * @from: The virtual kernel start address of the range
278 * @to: The virtual kernel end address of the range (exclusive)
279 * @prot: The protection to be applied to this range
281 * The same virtual address as the kernel virtual address is also used
282 * in Hyp-mode mapping (modulo HYP_PAGE_OFFSET) to the same underlying
285 int create_hyp_mappings(void *from, void *to, enum kvm_pgtable_prot prot)
287 phys_addr_t phys_addr;
288 unsigned long virt_addr;
289 unsigned long start = kern_hyp_va((unsigned long)from);
290 unsigned long end = kern_hyp_va((unsigned long)to);
292 if (is_kernel_in_hyp_mode())
295 start = start & PAGE_MASK;
296 end = PAGE_ALIGN(end);
298 for (virt_addr = start; virt_addr < end; virt_addr += PAGE_SIZE) {
301 phys_addr = kvm_kaddr_to_phys(from + virt_addr - start);
302 err = __create_hyp_mappings(virt_addr, PAGE_SIZE, phys_addr,
311 static int __create_hyp_private_mapping(phys_addr_t phys_addr, size_t size,
312 unsigned long *haddr,
313 enum kvm_pgtable_prot prot)
318 if (!kvm_host_owns_hyp_mappings()) {
319 base = kvm_call_hyp_nvhe(__pkvm_create_private_mapping,
320 phys_addr, size, prot);
321 if (IS_ERR_OR_NULL((void *)base))
322 return PTR_ERR((void *)base);
328 mutex_lock(&kvm_hyp_pgd_mutex);
331 * This assumes that we have enough space below the idmap
332 * page to allocate our VAs. If not, the check below will
333 * kick. A potential alternative would be to detect that
334 * overflow and switch to an allocation above the idmap.
336 * The allocated size is always a multiple of PAGE_SIZE.
338 size = PAGE_ALIGN(size + offset_in_page(phys_addr));
339 base = io_map_base - size;
342 * Verify that BIT(VA_BITS - 1) hasn't been flipped by
343 * allocating the new area, as it would indicate we've
344 * overflowed the idmap/IO address range.
346 if ((base ^ io_map_base) & BIT(VA_BITS - 1))
351 mutex_unlock(&kvm_hyp_pgd_mutex);
356 ret = __create_hyp_mappings(base, size, phys_addr, prot);
360 *haddr = base + offset_in_page(phys_addr);
366 * create_hyp_io_mappings - Map IO into both kernel and HYP
367 * @phys_addr: The physical start address which gets mapped
368 * @size: Size of the region being mapped
369 * @kaddr: Kernel VA for this mapping
370 * @haddr: HYP VA for this mapping
372 int create_hyp_io_mappings(phys_addr_t phys_addr, size_t size,
373 void __iomem **kaddr,
374 void __iomem **haddr)
379 *kaddr = ioremap(phys_addr, size);
383 if (is_kernel_in_hyp_mode()) {
388 ret = __create_hyp_private_mapping(phys_addr, size,
389 &addr, PAGE_HYP_DEVICE);
397 *haddr = (void __iomem *)addr;
402 * create_hyp_exec_mappings - Map an executable range into HYP
403 * @phys_addr: The physical start address which gets mapped
404 * @size: Size of the region being mapped
405 * @haddr: HYP VA for this mapping
407 int create_hyp_exec_mappings(phys_addr_t phys_addr, size_t size,
413 BUG_ON(is_kernel_in_hyp_mode());
415 ret = __create_hyp_private_mapping(phys_addr, size,
416 &addr, PAGE_HYP_EXEC);
422 *haddr = (void *)addr;
426 static struct kvm_pgtable_mm_ops kvm_s2_mm_ops = {
427 .zalloc_page = stage2_memcache_zalloc_page,
428 .zalloc_pages_exact = kvm_host_zalloc_pages_exact,
429 .free_pages_exact = free_pages_exact,
430 .get_page = kvm_host_get_page,
431 .put_page = kvm_host_put_page,
432 .page_count = kvm_host_page_count,
433 .phys_to_virt = kvm_host_va,
434 .virt_to_phys = kvm_host_pa,
438 * kvm_init_stage2_mmu - Initialise a S2 MMU strucrure
439 * @kvm: The pointer to the KVM structure
440 * @mmu: The pointer to the s2 MMU structure
442 * Allocates only the stage-2 HW PGD level table(s).
443 * Note we don't need locking here as this is only called when the VM is
444 * created, which can only be done once.
446 int kvm_init_stage2_mmu(struct kvm *kvm, struct kvm_s2_mmu *mmu)
449 struct kvm_pgtable *pgt;
451 if (mmu->pgt != NULL) {
452 kvm_err("kvm_arch already initialized?\n");
456 pgt = kzalloc(sizeof(*pgt), GFP_KERNEL);
460 err = kvm_pgtable_stage2_init(pgt, &kvm->arch, &kvm_s2_mm_ops);
462 goto out_free_pgtable;
464 mmu->last_vcpu_ran = alloc_percpu(typeof(*mmu->last_vcpu_ran));
465 if (!mmu->last_vcpu_ran) {
467 goto out_destroy_pgtable;
470 for_each_possible_cpu(cpu)
471 *per_cpu_ptr(mmu->last_vcpu_ran, cpu) = -1;
473 mmu->arch = &kvm->arch;
475 mmu->pgd_phys = __pa(pgt->pgd);
476 mmu->vmid.vmid_gen = 0;
480 kvm_pgtable_stage2_destroy(pgt);
486 static void stage2_unmap_memslot(struct kvm *kvm,
487 struct kvm_memory_slot *memslot)
489 hva_t hva = memslot->userspace_addr;
490 phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
491 phys_addr_t size = PAGE_SIZE * memslot->npages;
492 hva_t reg_end = hva + size;
495 * A memory region could potentially cover multiple VMAs, and any holes
496 * between them, so iterate over all of them to find out if we should
499 * +--------------------------------------------+
500 * +---------------+----------------+ +----------------+
501 * | : VMA 1 | VMA 2 | | VMA 3 : |
502 * +---------------+----------------+ +----------------+
504 * +--------------------------------------------+
507 struct vm_area_struct *vma;
508 hva_t vm_start, vm_end;
510 vma = find_vma_intersection(current->mm, hva, reg_end);
515 * Take the intersection of this VMA with the memory region
517 vm_start = max(hva, vma->vm_start);
518 vm_end = min(reg_end, vma->vm_end);
520 if (!(vma->vm_flags & VM_PFNMAP)) {
521 gpa_t gpa = addr + (vm_start - memslot->userspace_addr);
522 unmap_stage2_range(&kvm->arch.mmu, gpa, vm_end - vm_start);
525 } while (hva < reg_end);
529 * stage2_unmap_vm - Unmap Stage-2 RAM mappings
530 * @kvm: The struct kvm pointer
532 * Go through the memregions and unmap any regular RAM
533 * backing memory already mapped to the VM.
535 void stage2_unmap_vm(struct kvm *kvm)
537 struct kvm_memslots *slots;
538 struct kvm_memory_slot *memslot;
541 idx = srcu_read_lock(&kvm->srcu);
542 mmap_read_lock(current->mm);
543 spin_lock(&kvm->mmu_lock);
545 slots = kvm_memslots(kvm);
546 kvm_for_each_memslot(memslot, slots)
547 stage2_unmap_memslot(kvm, memslot);
549 spin_unlock(&kvm->mmu_lock);
550 mmap_read_unlock(current->mm);
551 srcu_read_unlock(&kvm->srcu, idx);
554 void kvm_free_stage2_pgd(struct kvm_s2_mmu *mmu)
556 struct kvm *kvm = kvm_s2_mmu_to_kvm(mmu);
557 struct kvm_pgtable *pgt = NULL;
559 spin_lock(&kvm->mmu_lock);
564 free_percpu(mmu->last_vcpu_ran);
566 spin_unlock(&kvm->mmu_lock);
569 kvm_pgtable_stage2_destroy(pgt);
575 * kvm_phys_addr_ioremap - map a device range to guest IPA
577 * @kvm: The KVM pointer
578 * @guest_ipa: The IPA at which to insert the mapping
579 * @pa: The physical address of the device
580 * @size: The size of the mapping
581 * @writable: Whether or not to create a writable mapping
583 int kvm_phys_addr_ioremap(struct kvm *kvm, phys_addr_t guest_ipa,
584 phys_addr_t pa, unsigned long size, bool writable)
588 struct kvm_mmu_memory_cache cache = { 0, __GFP_ZERO, NULL, };
589 struct kvm_pgtable *pgt = kvm->arch.mmu.pgt;
590 enum kvm_pgtable_prot prot = KVM_PGTABLE_PROT_DEVICE |
592 (writable ? KVM_PGTABLE_PROT_W : 0);
594 size += offset_in_page(guest_ipa);
595 guest_ipa &= PAGE_MASK;
597 for (addr = guest_ipa; addr < guest_ipa + size; addr += PAGE_SIZE) {
598 ret = kvm_mmu_topup_memory_cache(&cache,
599 kvm_mmu_cache_min_pages(kvm));
603 spin_lock(&kvm->mmu_lock);
604 ret = kvm_pgtable_stage2_map(pgt, addr, PAGE_SIZE, pa, prot,
606 spin_unlock(&kvm->mmu_lock);
613 kvm_mmu_free_memory_cache(&cache);
618 * stage2_wp_range() - write protect stage2 memory region range
619 * @mmu: The KVM stage-2 MMU pointer
620 * @addr: Start address of range
621 * @end: End address of range
623 static void stage2_wp_range(struct kvm_s2_mmu *mmu, phys_addr_t addr, phys_addr_t end)
625 struct kvm *kvm = kvm_s2_mmu_to_kvm(mmu);
626 stage2_apply_range_resched(kvm, addr, end, kvm_pgtable_stage2_wrprotect);
630 * kvm_mmu_wp_memory_region() - write protect stage 2 entries for memory slot
631 * @kvm: The KVM pointer
632 * @slot: The memory slot to write protect
634 * Called to start logging dirty pages after memory region
635 * KVM_MEM_LOG_DIRTY_PAGES operation is called. After this function returns
636 * all present PUD, PMD and PTEs are write protected in the memory region.
637 * Afterwards read of dirty page log can be called.
639 * Acquires kvm_mmu_lock. Called with kvm->slots_lock mutex acquired,
640 * serializing operations for VM memory regions.
642 static void kvm_mmu_wp_memory_region(struct kvm *kvm, int slot)
644 struct kvm_memslots *slots = kvm_memslots(kvm);
645 struct kvm_memory_slot *memslot = id_to_memslot(slots, slot);
646 phys_addr_t start, end;
648 if (WARN_ON_ONCE(!memslot))
651 start = memslot->base_gfn << PAGE_SHIFT;
652 end = (memslot->base_gfn + memslot->npages) << PAGE_SHIFT;
654 spin_lock(&kvm->mmu_lock);
655 stage2_wp_range(&kvm->arch.mmu, start, end);
656 spin_unlock(&kvm->mmu_lock);
657 kvm_flush_remote_tlbs(kvm);
661 * kvm_mmu_write_protect_pt_masked() - write protect dirty pages
662 * @kvm: The KVM pointer
663 * @slot: The memory slot associated with mask
664 * @gfn_offset: The gfn offset in memory slot
665 * @mask: The mask of dirty pages at offset 'gfn_offset' in this memory
666 * slot to be write protected
668 * Walks bits set in mask write protects the associated pte's. Caller must
669 * acquire kvm_mmu_lock.
671 static void kvm_mmu_write_protect_pt_masked(struct kvm *kvm,
672 struct kvm_memory_slot *slot,
673 gfn_t gfn_offset, unsigned long mask)
675 phys_addr_t base_gfn = slot->base_gfn + gfn_offset;
676 phys_addr_t start = (base_gfn + __ffs(mask)) << PAGE_SHIFT;
677 phys_addr_t end = (base_gfn + __fls(mask) + 1) << PAGE_SHIFT;
679 stage2_wp_range(&kvm->arch.mmu, start, end);
683 * kvm_arch_mmu_enable_log_dirty_pt_masked - enable dirty logging for selected
686 * It calls kvm_mmu_write_protect_pt_masked to write protect selected pages to
687 * enable dirty logging for them.
689 void kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm *kvm,
690 struct kvm_memory_slot *slot,
691 gfn_t gfn_offset, unsigned long mask)
693 kvm_mmu_write_protect_pt_masked(kvm, slot, gfn_offset, mask);
696 static void clean_dcache_guest_page(kvm_pfn_t pfn, unsigned long size)
698 __clean_dcache_guest_page(pfn, size);
701 static void invalidate_icache_guest_page(kvm_pfn_t pfn, unsigned long size)
703 __invalidate_icache_guest_page(pfn, size);
706 static void kvm_send_hwpoison_signal(unsigned long address, short lsb)
708 send_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, lsb, current);
711 static bool fault_supports_stage2_huge_mapping(struct kvm_memory_slot *memslot,
713 unsigned long map_size)
716 hva_t uaddr_start, uaddr_end;
719 /* The memslot and the VMA are guaranteed to be aligned to PAGE_SIZE */
720 if (map_size == PAGE_SIZE)
723 size = memslot->npages * PAGE_SIZE;
725 gpa_start = memslot->base_gfn << PAGE_SHIFT;
727 uaddr_start = memslot->userspace_addr;
728 uaddr_end = uaddr_start + size;
731 * Pages belonging to memslots that don't have the same alignment
732 * within a PMD/PUD for userspace and IPA cannot be mapped with stage-2
733 * PMD/PUD entries, because we'll end up mapping the wrong pages.
735 * Consider a layout like the following:
737 * memslot->userspace_addr:
738 * +-----+--------------------+--------------------+---+
739 * |abcde|fgh Stage-1 block | Stage-1 block tv|xyz|
740 * +-----+--------------------+--------------------+---+
742 * memslot->base_gfn << PAGE_SHIFT:
743 * +---+--------------------+--------------------+-----+
744 * |abc|def Stage-2 block | Stage-2 block |tvxyz|
745 * +---+--------------------+--------------------+-----+
747 * If we create those stage-2 blocks, we'll end up with this incorrect
753 if ((gpa_start & (map_size - 1)) != (uaddr_start & (map_size - 1)))
757 * Next, let's make sure we're not trying to map anything not covered
758 * by the memslot. This means we have to prohibit block size mappings
759 * for the beginning and end of a non-block aligned and non-block sized
760 * memory slot (illustrated by the head and tail parts of the
761 * userspace view above containing pages 'abcde' and 'xyz',
764 * Note that it doesn't matter if we do the check using the
765 * userspace_addr or the base_gfn, as both are equally aligned (per
766 * the check above) and equally sized.
768 return (hva & ~(map_size - 1)) >= uaddr_start &&
769 (hva & ~(map_size - 1)) + map_size <= uaddr_end;
773 * Check if the given hva is backed by a transparent huge page (THP) and
774 * whether it can be mapped using block mapping in stage2. If so, adjust
775 * the stage2 PFN and IPA accordingly. Only PMD_SIZE THPs are currently
776 * supported. This will need to be updated to support other THP sizes.
778 * Returns the size of the mapping.
781 transparent_hugepage_adjust(struct kvm_memory_slot *memslot,
782 unsigned long hva, kvm_pfn_t *pfnp,
785 kvm_pfn_t pfn = *pfnp;
788 * Make sure the adjustment is done only for THP pages. Also make
789 * sure that the HVA and IPA are sufficiently aligned and that the
790 * block map is contained within the memslot.
792 if (kvm_is_transparent_hugepage(pfn) &&
793 fault_supports_stage2_huge_mapping(memslot, hva, PMD_SIZE)) {
795 * The address we faulted on is backed by a transparent huge
796 * page. However, because we map the compound huge page and
797 * not the individual tail page, we need to transfer the
798 * refcount to the head page. We have to be careful that the
799 * THP doesn't start to split while we are adjusting the
802 * We are sure this doesn't happen, because mmu_notifier_retry
803 * was successful and we are holding the mmu_lock, so if this
804 * THP is trying to split, it will be blocked in the mmu
805 * notifier before touching any of the pages, specifically
806 * before being able to call __split_huge_page_refcount().
808 * We can therefore safely transfer the refcount from PG_tail
809 * to PG_head and switch the pfn from a tail page to the head
813 kvm_release_pfn_clean(pfn);
814 pfn &= ~(PTRS_PER_PMD - 1);
821 /* Use page mapping if we cannot use block mapping. */
825 static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
826 struct kvm_memory_slot *memslot, unsigned long hva,
827 unsigned long fault_status)
830 bool write_fault, writable, force_pte = false;
833 unsigned long mmu_seq;
834 struct kvm *kvm = vcpu->kvm;
835 struct kvm_mmu_memory_cache *memcache = &vcpu->arch.mmu_page_cache;
836 struct vm_area_struct *vma;
840 bool logging_active = memslot_is_logging(memslot);
841 unsigned long fault_level = kvm_vcpu_trap_get_fault_level(vcpu);
842 unsigned long vma_pagesize, fault_granule;
843 enum kvm_pgtable_prot prot = KVM_PGTABLE_PROT_R;
844 struct kvm_pgtable *pgt;
846 fault_granule = 1UL << ARM64_HW_PGTABLE_LEVEL_SHIFT(fault_level);
847 write_fault = kvm_is_write_fault(vcpu);
848 exec_fault = kvm_vcpu_trap_is_exec_fault(vcpu);
849 VM_BUG_ON(write_fault && exec_fault);
851 if (fault_status == FSC_PERM && !write_fault && !exec_fault) {
852 kvm_err("Unexpected L2 read permission error\n");
856 /* Let's check if we will get back a huge page backed by hugetlbfs */
857 mmap_read_lock(current->mm);
858 vma = find_vma_intersection(current->mm, hva, hva + 1);
859 if (unlikely(!vma)) {
860 kvm_err("Failed to find VMA for hva 0x%lx\n", hva);
861 mmap_read_unlock(current->mm);
865 if (is_vm_hugetlb_page(vma))
866 vma_shift = huge_page_shift(hstate_vma(vma));
868 vma_shift = PAGE_SHIFT;
870 if (logging_active ||
871 (vma->vm_flags & VM_PFNMAP)) {
873 vma_shift = PAGE_SHIFT;
877 #ifndef __PAGETABLE_PMD_FOLDED
879 if (fault_supports_stage2_huge_mapping(memslot, hva, PUD_SIZE))
884 vma_shift = PMD_SHIFT;
887 if (fault_supports_stage2_huge_mapping(memslot, hva, PMD_SIZE))
891 vma_shift = PAGE_SHIFT;
897 WARN_ONCE(1, "Unknown vma_shift %d", vma_shift);
900 vma_pagesize = 1UL << vma_shift;
901 if (vma_pagesize == PMD_SIZE || vma_pagesize == PUD_SIZE)
902 fault_ipa &= ~(vma_pagesize - 1);
904 gfn = fault_ipa >> PAGE_SHIFT;
905 mmap_read_unlock(current->mm);
908 * Permission faults just need to update the existing leaf entry,
909 * and so normally don't require allocations from the memcache. The
910 * only exception to this is when dirty logging is enabled at runtime
911 * and a write fault needs to collapse a block entry into a table.
913 if (fault_status != FSC_PERM || (logging_active && write_fault)) {
914 ret = kvm_mmu_topup_memory_cache(memcache,
915 kvm_mmu_cache_min_pages(kvm));
920 mmu_seq = vcpu->kvm->mmu_notifier_seq;
922 * Ensure the read of mmu_notifier_seq happens before we call
923 * gfn_to_pfn_prot (which calls get_user_pages), so that we don't risk
924 * the page we just got a reference to gets unmapped before we have a
925 * chance to grab the mmu_lock, which ensure that if the page gets
926 * unmapped afterwards, the call to kvm_unmap_gfn will take it away
927 * from us again properly. This smp_rmb() interacts with the smp_wmb()
928 * in kvm_mmu_notifier_invalidate_<page|range_end>.
930 * Besides, __gfn_to_pfn_memslot() instead of gfn_to_pfn_prot() is
931 * used to avoid unnecessary overhead introduced to locate the memory
932 * slot because it's always fixed even @gfn is adjusted for huge pages.
936 pfn = __gfn_to_pfn_memslot(memslot, gfn, false, NULL,
937 write_fault, &writable, NULL);
938 if (pfn == KVM_PFN_ERR_HWPOISON) {
939 kvm_send_hwpoison_signal(hva, vma_shift);
942 if (is_error_noslot_pfn(pfn))
945 if (kvm_is_device_pfn(pfn)) {
948 } else if (logging_active && !write_fault) {
950 * Only actually map the page as writable if this was a write
956 if (exec_fault && device)
959 spin_lock(&kvm->mmu_lock);
960 pgt = vcpu->arch.hw_mmu->pgt;
961 if (mmu_notifier_retry(kvm, mmu_seq))
965 * If we are not forced to use page mapping, check if we are
966 * backed by a THP and thus use block mapping if possible.
968 if (vma_pagesize == PAGE_SIZE && !force_pte)
969 vma_pagesize = transparent_hugepage_adjust(memslot, hva,
972 prot |= KVM_PGTABLE_PROT_W;
974 if (fault_status != FSC_PERM && !device)
975 clean_dcache_guest_page(pfn, vma_pagesize);
978 prot |= KVM_PGTABLE_PROT_X;
979 invalidate_icache_guest_page(pfn, vma_pagesize);
983 prot |= KVM_PGTABLE_PROT_DEVICE;
984 else if (cpus_have_const_cap(ARM64_HAS_CACHE_DIC))
985 prot |= KVM_PGTABLE_PROT_X;
988 * Under the premise of getting a FSC_PERM fault, we just need to relax
989 * permissions only if vma_pagesize equals fault_granule. Otherwise,
990 * kvm_pgtable_stage2_map() should be called to change block size.
992 if (fault_status == FSC_PERM && vma_pagesize == fault_granule) {
993 ret = kvm_pgtable_stage2_relax_perms(pgt, fault_ipa, prot);
995 ret = kvm_pgtable_stage2_map(pgt, fault_ipa, vma_pagesize,
996 __pfn_to_phys(pfn), prot,
1000 /* Mark the page dirty only if the fault is handled successfully */
1001 if (writable && !ret) {
1002 kvm_set_pfn_dirty(pfn);
1003 mark_page_dirty_in_slot(kvm, memslot, gfn);
1007 spin_unlock(&kvm->mmu_lock);
1008 kvm_set_pfn_accessed(pfn);
1009 kvm_release_pfn_clean(pfn);
1010 return ret != -EAGAIN ? ret : 0;
1013 /* Resolve the access fault by making the page young again. */
1014 static void handle_access_fault(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa)
1018 struct kvm_s2_mmu *mmu;
1020 trace_kvm_access_fault(fault_ipa);
1022 spin_lock(&vcpu->kvm->mmu_lock);
1023 mmu = vcpu->arch.hw_mmu;
1024 kpte = kvm_pgtable_stage2_mkyoung(mmu->pgt, fault_ipa);
1025 spin_unlock(&vcpu->kvm->mmu_lock);
1029 kvm_set_pfn_accessed(pte_pfn(pte));
1033 * kvm_handle_guest_abort - handles all 2nd stage aborts
1034 * @vcpu: the VCPU pointer
1036 * Any abort that gets to the host is almost guaranteed to be caused by a
1037 * missing second stage translation table entry, which can mean that either the
1038 * guest simply needs more memory and we must allocate an appropriate page or it
1039 * can mean that the guest tried to access I/O memory, which is emulated by user
1040 * space. The distinction is based on the IPA causing the fault and whether this
1041 * memory region has been registered as standard RAM by user space.
1043 int kvm_handle_guest_abort(struct kvm_vcpu *vcpu)
1045 unsigned long fault_status;
1046 phys_addr_t fault_ipa;
1047 struct kvm_memory_slot *memslot;
1049 bool is_iabt, write_fault, writable;
1053 fault_status = kvm_vcpu_trap_get_fault_type(vcpu);
1055 fault_ipa = kvm_vcpu_get_fault_ipa(vcpu);
1056 is_iabt = kvm_vcpu_trap_is_iabt(vcpu);
1058 /* Synchronous External Abort? */
1059 if (kvm_vcpu_abt_issea(vcpu)) {
1061 * For RAS the host kernel may handle this abort.
1062 * There is no need to pass the error into the guest.
1064 if (kvm_handle_guest_sea(fault_ipa, kvm_vcpu_get_esr(vcpu)))
1065 kvm_inject_vabt(vcpu);
1070 trace_kvm_guest_fault(*vcpu_pc(vcpu), kvm_vcpu_get_esr(vcpu),
1071 kvm_vcpu_get_hfar(vcpu), fault_ipa);
1073 /* Check the stage-2 fault is trans. fault or write fault */
1074 if (fault_status != FSC_FAULT && fault_status != FSC_PERM &&
1075 fault_status != FSC_ACCESS) {
1076 kvm_err("Unsupported FSC: EC=%#x xFSC=%#lx ESR_EL2=%#lx\n",
1077 kvm_vcpu_trap_get_class(vcpu),
1078 (unsigned long)kvm_vcpu_trap_get_fault(vcpu),
1079 (unsigned long)kvm_vcpu_get_esr(vcpu));
1083 idx = srcu_read_lock(&vcpu->kvm->srcu);
1085 gfn = fault_ipa >> PAGE_SHIFT;
1086 memslot = gfn_to_memslot(vcpu->kvm, gfn);
1087 hva = gfn_to_hva_memslot_prot(memslot, gfn, &writable);
1088 write_fault = kvm_is_write_fault(vcpu);
1089 if (kvm_is_error_hva(hva) || (write_fault && !writable)) {
1091 * The guest has put either its instructions or its page-tables
1092 * somewhere it shouldn't have. Userspace won't be able to do
1093 * anything about this (there's no syndrome for a start), so
1094 * re-inject the abort back into the guest.
1101 if (kvm_vcpu_abt_iss1tw(vcpu)) {
1102 kvm_inject_dabt(vcpu, kvm_vcpu_get_hfar(vcpu));
1108 * Check for a cache maintenance operation. Since we
1109 * ended-up here, we know it is outside of any memory
1110 * slot. But we can't find out if that is for a device,
1111 * or if the guest is just being stupid. The only thing
1112 * we know for sure is that this range cannot be cached.
1114 * So let's assume that the guest is just being
1115 * cautious, and skip the instruction.
1117 if (kvm_is_error_hva(hva) && kvm_vcpu_dabt_is_cm(vcpu)) {
1124 * The IPA is reported as [MAX:12], so we need to
1125 * complement it with the bottom 12 bits from the
1126 * faulting VA. This is always 12 bits, irrespective
1129 fault_ipa |= kvm_vcpu_get_hfar(vcpu) & ((1 << 12) - 1);
1130 ret = io_mem_abort(vcpu, fault_ipa);
1134 /* Userspace should not be able to register out-of-bounds IPAs */
1135 VM_BUG_ON(fault_ipa >= kvm_phys_size(vcpu->kvm));
1137 if (fault_status == FSC_ACCESS) {
1138 handle_access_fault(vcpu, fault_ipa);
1143 ret = user_mem_abort(vcpu, fault_ipa, memslot, hva, fault_status);
1147 if (ret == -ENOEXEC) {
1148 kvm_inject_pabt(vcpu, kvm_vcpu_get_hfar(vcpu));
1152 srcu_read_unlock(&vcpu->kvm->srcu, idx);
1156 bool kvm_unmap_gfn_range(struct kvm *kvm, struct kvm_gfn_range *range)
1158 if (!kvm->arch.mmu.pgt)
1161 __unmap_stage2_range(&kvm->arch.mmu, range->start << PAGE_SHIFT,
1162 (range->end - range->start) << PAGE_SHIFT,
1168 bool kvm_set_spte_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
1170 kvm_pfn_t pfn = pte_pfn(range->pte);
1172 if (!kvm->arch.mmu.pgt)
1175 WARN_ON(range->end - range->start != 1);
1178 * We've moved a page around, probably through CoW, so let's treat it
1179 * just like a translation fault and clean the cache to the PoC.
1181 clean_dcache_guest_page(pfn, PAGE_SIZE);
1184 * The MMU notifiers will have unmapped a huge PMD before calling
1185 * ->change_pte() (which in turn calls kvm_set_spte_gfn()) and
1186 * therefore we never need to clear out a huge PMD through this
1187 * calling path and a memcache is not required.
1189 kvm_pgtable_stage2_map(kvm->arch.mmu.pgt, range->start << PAGE_SHIFT,
1190 PAGE_SIZE, __pfn_to_phys(pfn),
1191 KVM_PGTABLE_PROT_R, NULL);
1196 bool kvm_age_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
1198 u64 size = (range->end - range->start) << PAGE_SHIFT;
1202 if (!kvm->arch.mmu.pgt)
1205 WARN_ON(size != PAGE_SIZE && size != PMD_SIZE && size != PUD_SIZE);
1207 kpte = kvm_pgtable_stage2_mkold(kvm->arch.mmu.pgt,
1208 range->start << PAGE_SHIFT);
1210 return pte_valid(pte) && pte_young(pte);
1213 bool kvm_test_age_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
1215 if (!kvm->arch.mmu.pgt)
1218 return kvm_pgtable_stage2_is_young(kvm->arch.mmu.pgt,
1219 range->start << PAGE_SHIFT);
1222 phys_addr_t kvm_mmu_get_httbr(void)
1224 return __pa(hyp_pgtable->pgd);
1227 phys_addr_t kvm_get_idmap_vector(void)
1229 return hyp_idmap_vector;
1232 static int kvm_map_idmap_text(void)
1234 unsigned long size = hyp_idmap_end - hyp_idmap_start;
1235 int err = __create_hyp_mappings(hyp_idmap_start, size, hyp_idmap_start,
1238 kvm_err("Failed to idmap %lx-%lx\n",
1239 hyp_idmap_start, hyp_idmap_end);
1244 static void *kvm_hyp_zalloc_page(void *arg)
1246 return (void *)get_zeroed_page(GFP_KERNEL);
1249 static struct kvm_pgtable_mm_ops kvm_hyp_mm_ops = {
1250 .zalloc_page = kvm_hyp_zalloc_page,
1251 .get_page = kvm_host_get_page,
1252 .put_page = kvm_host_put_page,
1253 .phys_to_virt = kvm_host_va,
1254 .virt_to_phys = kvm_host_pa,
1257 int kvm_mmu_init(u32 *hyp_va_bits)
1261 hyp_idmap_start = __pa_symbol(__hyp_idmap_text_start);
1262 hyp_idmap_start = ALIGN_DOWN(hyp_idmap_start, PAGE_SIZE);
1263 hyp_idmap_end = __pa_symbol(__hyp_idmap_text_end);
1264 hyp_idmap_end = ALIGN(hyp_idmap_end, PAGE_SIZE);
1265 hyp_idmap_vector = __pa_symbol(__kvm_hyp_init);
1268 * We rely on the linker script to ensure at build time that the HYP
1269 * init code does not cross a page boundary.
1271 BUG_ON((hyp_idmap_start ^ (hyp_idmap_end - 1)) & PAGE_MASK);
1273 *hyp_va_bits = 64 - ((idmap_t0sz & TCR_T0SZ_MASK) >> TCR_T0SZ_OFFSET);
1274 kvm_debug("Using %u-bit virtual addresses at EL2\n", *hyp_va_bits);
1275 kvm_debug("IDMAP page: %lx\n", hyp_idmap_start);
1276 kvm_debug("HYP VA range: %lx:%lx\n",
1277 kern_hyp_va(PAGE_OFFSET),
1278 kern_hyp_va((unsigned long)high_memory - 1));
1280 if (hyp_idmap_start >= kern_hyp_va(PAGE_OFFSET) &&
1281 hyp_idmap_start < kern_hyp_va((unsigned long)high_memory - 1) &&
1282 hyp_idmap_start != (unsigned long)__hyp_idmap_text_start) {
1284 * The idmap page is intersecting with the VA space,
1285 * it is not safe to continue further.
1287 kvm_err("IDMAP intersecting with HYP VA, unable to continue\n");
1292 hyp_pgtable = kzalloc(sizeof(*hyp_pgtable), GFP_KERNEL);
1294 kvm_err("Hyp mode page-table not allocated\n");
1299 err = kvm_pgtable_hyp_init(hyp_pgtable, *hyp_va_bits, &kvm_hyp_mm_ops);
1301 goto out_free_pgtable;
1303 err = kvm_map_idmap_text();
1305 goto out_destroy_pgtable;
1307 io_map_base = hyp_idmap_start;
1310 out_destroy_pgtable:
1311 kvm_pgtable_hyp_destroy(hyp_pgtable);
1319 void kvm_arch_commit_memory_region(struct kvm *kvm,
1320 const struct kvm_userspace_memory_region *mem,
1321 struct kvm_memory_slot *old,
1322 const struct kvm_memory_slot *new,
1323 enum kvm_mr_change change)
1326 * At this point memslot has been committed and there is an
1327 * allocated dirty_bitmap[], dirty pages will be tracked while the
1328 * memory slot is write protected.
1330 if (change != KVM_MR_DELETE && mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
1332 * If we're with initial-all-set, we don't need to write
1333 * protect any pages because they're all reported as dirty.
1334 * Huge pages and normal pages will be write protect gradually.
1336 if (!kvm_dirty_log_manual_protect_and_init_set(kvm)) {
1337 kvm_mmu_wp_memory_region(kvm, mem->slot);
1342 int kvm_arch_prepare_memory_region(struct kvm *kvm,
1343 struct kvm_memory_slot *memslot,
1344 const struct kvm_userspace_memory_region *mem,
1345 enum kvm_mr_change change)
1347 hva_t hva = mem->userspace_addr;
1348 hva_t reg_end = hva + mem->memory_size;
1349 bool writable = !(mem->flags & KVM_MEM_READONLY);
1352 if (change != KVM_MR_CREATE && change != KVM_MR_MOVE &&
1353 change != KVM_MR_FLAGS_ONLY)
1357 * Prevent userspace from creating a memory region outside of the IPA
1358 * space addressable by the KVM guest IPA space.
1360 if ((memslot->base_gfn + memslot->npages) > (kvm_phys_size(kvm) >> PAGE_SHIFT))
1363 mmap_read_lock(current->mm);
1365 * A memory region could potentially cover multiple VMAs, and any holes
1366 * between them, so iterate over all of them to find out if we can map
1367 * any of them right now.
1369 * +--------------------------------------------+
1370 * +---------------+----------------+ +----------------+
1371 * | : VMA 1 | VMA 2 | | VMA 3 : |
1372 * +---------------+----------------+ +----------------+
1374 * +--------------------------------------------+
1377 struct vm_area_struct *vma;
1378 hva_t vm_start, vm_end;
1380 vma = find_vma_intersection(current->mm, hva, reg_end);
1385 * Take the intersection of this VMA with the memory region
1387 vm_start = max(hva, vma->vm_start);
1388 vm_end = min(reg_end, vma->vm_end);
1390 if (vma->vm_flags & VM_PFNMAP) {
1391 gpa_t gpa = mem->guest_phys_addr +
1392 (vm_start - mem->userspace_addr);
1395 pa = (phys_addr_t)vma->vm_pgoff << PAGE_SHIFT;
1396 pa += vm_start - vma->vm_start;
1398 /* IO region dirty page logging not allowed */
1399 if (memslot->flags & KVM_MEM_LOG_DIRTY_PAGES) {
1404 ret = kvm_phys_addr_ioremap(kvm, gpa, pa,
1411 } while (hva < reg_end);
1413 if (change == KVM_MR_FLAGS_ONLY)
1416 spin_lock(&kvm->mmu_lock);
1418 unmap_stage2_range(&kvm->arch.mmu, mem->guest_phys_addr, mem->memory_size);
1419 else if (!cpus_have_final_cap(ARM64_HAS_STAGE2_FWB))
1420 stage2_flush_memslot(kvm, memslot);
1421 spin_unlock(&kvm->mmu_lock);
1423 mmap_read_unlock(current->mm);
1427 void kvm_arch_free_memslot(struct kvm *kvm, struct kvm_memory_slot *slot)
1431 void kvm_arch_memslots_updated(struct kvm *kvm, u64 gen)
1435 void kvm_arch_flush_shadow_all(struct kvm *kvm)
1437 kvm_free_stage2_pgd(&kvm->arch.mmu);
1440 void kvm_arch_flush_shadow_memslot(struct kvm *kvm,
1441 struct kvm_memory_slot *slot)
1443 gpa_t gpa = slot->base_gfn << PAGE_SHIFT;
1444 phys_addr_t size = slot->npages << PAGE_SHIFT;
1446 spin_lock(&kvm->mmu_lock);
1447 unmap_stage2_range(&kvm->arch.mmu, gpa, size);
1448 spin_unlock(&kvm->mmu_lock);
1452 * See note at ARMv7 ARM B1.14.4 (TL;DR: S/W ops are not easily virtualized).
1455 * - S/W ops are local to a CPU (not broadcast)
1456 * - We have line migration behind our back (speculation)
1457 * - System caches don't support S/W at all (damn!)
1459 * In the face of the above, the best we can do is to try and convert
1460 * S/W ops to VA ops. Because the guest is not allowed to infer the
1461 * S/W to PA mapping, it can only use S/W to nuke the whole cache,
1462 * which is a rather good thing for us.
1464 * Also, it is only used when turning caches on/off ("The expected
1465 * usage of the cache maintenance instructions that operate by set/way
1466 * is associated with the cache maintenance instructions associated
1467 * with the powerdown and powerup of caches, if this is required by
1468 * the implementation.").
1470 * We use the following policy:
1472 * - If we trap a S/W operation, we enable VM trapping to detect
1473 * caches being turned on/off, and do a full clean.
1475 * - We flush the caches on both caches being turned on and off.
1477 * - Once the caches are enabled, we stop trapping VM ops.
1479 void kvm_set_way_flush(struct kvm_vcpu *vcpu)
1481 unsigned long hcr = *vcpu_hcr(vcpu);
1484 * If this is the first time we do a S/W operation
1485 * (i.e. HCR_TVM not set) flush the whole memory, and set the
1488 * Otherwise, rely on the VM trapping to wait for the MMU +
1489 * Caches to be turned off. At that point, we'll be able to
1490 * clean the caches again.
1492 if (!(hcr & HCR_TVM)) {
1493 trace_kvm_set_way_flush(*vcpu_pc(vcpu),
1494 vcpu_has_cache_enabled(vcpu));
1495 stage2_flush_vm(vcpu->kvm);
1496 *vcpu_hcr(vcpu) = hcr | HCR_TVM;
1500 void kvm_toggle_cache(struct kvm_vcpu *vcpu, bool was_enabled)
1502 bool now_enabled = vcpu_has_cache_enabled(vcpu);
1505 * If switching the MMU+caches on, need to invalidate the caches.
1506 * If switching it off, need to clean the caches.
1507 * Clean + invalidate does the trick always.
1509 if (now_enabled != was_enabled)
1510 stage2_flush_vm(vcpu->kvm);
1512 /* Caches are now on, stop trapping VM ops (until a S/W op) */
1514 *vcpu_hcr(vcpu) &= ~HCR_TVM;
1516 trace_kvm_toggle_cache(*vcpu_pc(vcpu), was_enabled, now_enabled);