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;
34 #define KVM_S2PTE_FLAG_IS_IOMAP (1UL << 0)
35 #define KVM_S2_FLAG_LOGGING_ACTIVE (1UL << 1)
37 static bool is_iomap(unsigned long flags)
39 return flags & KVM_S2PTE_FLAG_IS_IOMAP;
42 static bool memslot_is_logging(struct kvm_memory_slot *memslot)
44 return memslot->dirty_bitmap && !(memslot->flags & KVM_MEM_READONLY);
48 * kvm_flush_remote_tlbs() - flush all VM TLB entries for v7/8
49 * @kvm: pointer to kvm structure.
51 * Interface to HYP function to flush all VM TLB entries
53 void kvm_flush_remote_tlbs(struct kvm *kvm)
55 kvm_call_hyp(__kvm_tlb_flush_vmid, &kvm->arch.mmu);
58 static void kvm_tlb_flush_vmid_ipa(struct kvm_s2_mmu *mmu, phys_addr_t ipa,
61 kvm_call_hyp(__kvm_tlb_flush_vmid_ipa, mmu, ipa, level);
65 * D-Cache management functions. They take the page table entries by
66 * value, as they are flushing the cache using the kernel mapping (or
69 static void kvm_flush_dcache_pte(pte_t pte)
71 __kvm_flush_dcache_pte(pte);
74 static void kvm_flush_dcache_pmd(pmd_t pmd)
76 __kvm_flush_dcache_pmd(pmd);
79 static void kvm_flush_dcache_pud(pud_t pud)
81 __kvm_flush_dcache_pud(pud);
84 static bool kvm_is_device_pfn(unsigned long pfn)
86 return !pfn_valid(pfn);
90 * stage2_dissolve_pmd() - clear and flush huge PMD entry
91 * @mmu: pointer to mmu structure to operate on
93 * @pmd: pmd pointer for IPA
95 * Function clears a PMD entry, flushes addr 1st and 2nd stage TLBs.
97 static void stage2_dissolve_pmd(struct kvm_s2_mmu *mmu, phys_addr_t addr, pmd_t *pmd)
99 if (!pmd_thp_or_huge(*pmd))
103 kvm_tlb_flush_vmid_ipa(mmu, addr, S2_PMD_LEVEL);
104 put_page(virt_to_page(pmd));
108 * stage2_dissolve_pud() - clear and flush huge PUD entry
109 * @mmu: pointer to mmu structure to operate on
111 * @pud: pud pointer for IPA
113 * Function clears a PUD entry, flushes addr 1st and 2nd stage TLBs.
115 static void stage2_dissolve_pud(struct kvm_s2_mmu *mmu, phys_addr_t addr, pud_t *pudp)
117 struct kvm *kvm = mmu->kvm;
119 if (!stage2_pud_huge(kvm, *pudp))
122 stage2_pud_clear(kvm, pudp);
123 kvm_tlb_flush_vmid_ipa(mmu, addr, S2_PUD_LEVEL);
124 put_page(virt_to_page(pudp));
127 static void clear_stage2_pgd_entry(struct kvm_s2_mmu *mmu, pgd_t *pgd, phys_addr_t addr)
129 struct kvm *kvm = mmu->kvm;
130 p4d_t *p4d_table __maybe_unused = stage2_p4d_offset(kvm, pgd, 0UL);
131 stage2_pgd_clear(kvm, pgd);
132 kvm_tlb_flush_vmid_ipa(mmu, addr, S2_NO_LEVEL_HINT);
133 stage2_p4d_free(kvm, p4d_table);
134 put_page(virt_to_page(pgd));
137 static void clear_stage2_p4d_entry(struct kvm_s2_mmu *mmu, p4d_t *p4d, phys_addr_t addr)
139 struct kvm *kvm = mmu->kvm;
140 pud_t *pud_table __maybe_unused = stage2_pud_offset(kvm, p4d, 0);
141 stage2_p4d_clear(kvm, p4d);
142 kvm_tlb_flush_vmid_ipa(mmu, addr, S2_NO_LEVEL_HINT);
143 stage2_pud_free(kvm, pud_table);
144 put_page(virt_to_page(p4d));
147 static void clear_stage2_pud_entry(struct kvm_s2_mmu *mmu, pud_t *pud, phys_addr_t addr)
149 struct kvm *kvm = mmu->kvm;
150 pmd_t *pmd_table __maybe_unused = stage2_pmd_offset(kvm, pud, 0);
152 VM_BUG_ON(stage2_pud_huge(kvm, *pud));
153 stage2_pud_clear(kvm, pud);
154 kvm_tlb_flush_vmid_ipa(mmu, addr, S2_NO_LEVEL_HINT);
155 stage2_pmd_free(kvm, pmd_table);
156 put_page(virt_to_page(pud));
159 static void clear_stage2_pmd_entry(struct kvm_s2_mmu *mmu, pmd_t *pmd, phys_addr_t addr)
161 pte_t *pte_table = pte_offset_kernel(pmd, 0);
162 VM_BUG_ON(pmd_thp_or_huge(*pmd));
164 kvm_tlb_flush_vmid_ipa(mmu, addr, S2_NO_LEVEL_HINT);
165 free_page((unsigned long)pte_table);
166 put_page(virt_to_page(pmd));
169 static inline void kvm_set_pte(pte_t *ptep, pte_t new_pte)
171 WRITE_ONCE(*ptep, new_pte);
175 static inline void kvm_set_pmd(pmd_t *pmdp, pmd_t new_pmd)
177 WRITE_ONCE(*pmdp, new_pmd);
181 static inline void kvm_pmd_populate(pmd_t *pmdp, pte_t *ptep)
183 kvm_set_pmd(pmdp, kvm_mk_pmd(ptep));
186 static inline void kvm_pud_populate(pud_t *pudp, pmd_t *pmdp)
188 WRITE_ONCE(*pudp, kvm_mk_pud(pmdp));
192 static inline void kvm_p4d_populate(p4d_t *p4dp, pud_t *pudp)
194 WRITE_ONCE(*p4dp, kvm_mk_p4d(pudp));
198 static inline void kvm_pgd_populate(pgd_t *pgdp, p4d_t *p4dp)
200 #ifndef __PAGETABLE_P4D_FOLDED
201 WRITE_ONCE(*pgdp, kvm_mk_pgd(p4dp));
207 * Unmapping vs dcache management:
209 * If a guest maps certain memory pages as uncached, all writes will
210 * bypass the data cache and go directly to RAM. However, the CPUs
211 * can still speculate reads (not writes) and fill cache lines with
214 * Those cache lines will be *clean* cache lines though, so a
215 * clean+invalidate operation is equivalent to an invalidate
216 * operation, because no cache lines are marked dirty.
218 * Those clean cache lines could be filled prior to an uncached write
219 * by the guest, and the cache coherent IO subsystem would therefore
220 * end up writing old data to disk.
222 * This is why right after unmapping a page/section and invalidating
223 * the corresponding TLBs, we call kvm_flush_dcache_p*() to make sure
224 * the IO subsystem will never hit in the cache.
226 * This is all avoided on systems that have ARM64_HAS_STAGE2_FWB, as
227 * we then fully enforce cacheability of RAM, no matter what the guest
230 static void unmap_stage2_ptes(struct kvm_s2_mmu *mmu, pmd_t *pmd,
231 phys_addr_t addr, phys_addr_t end)
233 phys_addr_t start_addr = addr;
234 pte_t *pte, *start_pte;
236 start_pte = pte = pte_offset_kernel(pmd, addr);
238 if (!pte_none(*pte)) {
239 pte_t old_pte = *pte;
241 kvm_set_pte(pte, __pte(0));
242 kvm_tlb_flush_vmid_ipa(mmu, addr, S2_PTE_LEVEL);
244 /* No need to invalidate the cache for device mappings */
245 if (!kvm_is_device_pfn(pte_pfn(old_pte)))
246 kvm_flush_dcache_pte(old_pte);
248 put_page(virt_to_page(pte));
250 } while (pte++, addr += PAGE_SIZE, addr != end);
252 if (stage2_pte_table_empty(mmu->kvm, start_pte))
253 clear_stage2_pmd_entry(mmu, pmd, start_addr);
256 static void unmap_stage2_pmds(struct kvm_s2_mmu *mmu, pud_t *pud,
257 phys_addr_t addr, phys_addr_t end)
259 struct kvm *kvm = mmu->kvm;
260 phys_addr_t next, start_addr = addr;
261 pmd_t *pmd, *start_pmd;
263 start_pmd = pmd = stage2_pmd_offset(kvm, pud, addr);
265 next = stage2_pmd_addr_end(kvm, addr, end);
266 if (!pmd_none(*pmd)) {
267 if (pmd_thp_or_huge(*pmd)) {
268 pmd_t old_pmd = *pmd;
271 kvm_tlb_flush_vmid_ipa(mmu, addr, S2_PMD_LEVEL);
273 kvm_flush_dcache_pmd(old_pmd);
275 put_page(virt_to_page(pmd));
277 unmap_stage2_ptes(mmu, pmd, addr, next);
280 } while (pmd++, addr = next, addr != end);
282 if (stage2_pmd_table_empty(kvm, start_pmd))
283 clear_stage2_pud_entry(mmu, pud, start_addr);
286 static void unmap_stage2_puds(struct kvm_s2_mmu *mmu, p4d_t *p4d,
287 phys_addr_t addr, phys_addr_t end)
289 struct kvm *kvm = mmu->kvm;
290 phys_addr_t next, start_addr = addr;
291 pud_t *pud, *start_pud;
293 start_pud = pud = stage2_pud_offset(kvm, p4d, addr);
295 next = stage2_pud_addr_end(kvm, addr, end);
296 if (!stage2_pud_none(kvm, *pud)) {
297 if (stage2_pud_huge(kvm, *pud)) {
298 pud_t old_pud = *pud;
300 stage2_pud_clear(kvm, pud);
301 kvm_tlb_flush_vmid_ipa(mmu, addr, S2_PUD_LEVEL);
302 kvm_flush_dcache_pud(old_pud);
303 put_page(virt_to_page(pud));
305 unmap_stage2_pmds(mmu, pud, addr, next);
308 } while (pud++, addr = next, addr != end);
310 if (stage2_pud_table_empty(kvm, start_pud))
311 clear_stage2_p4d_entry(mmu, p4d, start_addr);
314 static void unmap_stage2_p4ds(struct kvm_s2_mmu *mmu, pgd_t *pgd,
315 phys_addr_t addr, phys_addr_t end)
317 struct kvm *kvm = mmu->kvm;
318 phys_addr_t next, start_addr = addr;
319 p4d_t *p4d, *start_p4d;
321 start_p4d = p4d = stage2_p4d_offset(kvm, pgd, addr);
323 next = stage2_p4d_addr_end(kvm, addr, end);
324 if (!stage2_p4d_none(kvm, *p4d))
325 unmap_stage2_puds(mmu, p4d, addr, next);
326 } while (p4d++, addr = next, addr != end);
328 if (stage2_p4d_table_empty(kvm, start_p4d))
329 clear_stage2_pgd_entry(mmu, pgd, start_addr);
333 * unmap_stage2_range -- Clear stage2 page table entries to unmap a range
334 * @kvm: The VM pointer
335 * @start: The intermediate physical base address of the range to unmap
336 * @size: The size of the area to unmap
338 * Clear a range of stage-2 mappings, lowering the various ref-counts. Must
339 * be called while holding mmu_lock (unless for freeing the stage2 pgd before
340 * destroying the VM), otherwise another faulting VCPU may come in and mess
341 * with things behind our backs.
343 static void __unmap_stage2_range(struct kvm_s2_mmu *mmu, phys_addr_t start, u64 size,
346 struct kvm *kvm = mmu->kvm;
348 phys_addr_t addr = start, end = start + size;
351 assert_spin_locked(&kvm->mmu_lock);
352 WARN_ON(size & ~PAGE_MASK);
354 pgd = mmu->pgd + stage2_pgd_index(kvm, addr);
357 * Make sure the page table is still active, as another thread
358 * could have possibly freed the page table, while we released
361 if (!READ_ONCE(mmu->pgd))
363 next = stage2_pgd_addr_end(kvm, addr, end);
364 if (!stage2_pgd_none(kvm, *pgd))
365 unmap_stage2_p4ds(mmu, pgd, addr, next);
367 * If the range is too large, release the kvm->mmu_lock
368 * to prevent starvation and lockup detector warnings.
370 if (may_block && next != end)
371 cond_resched_lock(&kvm->mmu_lock);
372 } while (pgd++, addr = next, addr != end);
375 static void unmap_stage2_range(struct kvm_s2_mmu *mmu, phys_addr_t start, u64 size)
377 __unmap_stage2_range(mmu, start, size, true);
380 static void stage2_flush_ptes(struct kvm_s2_mmu *mmu, pmd_t *pmd,
381 phys_addr_t addr, phys_addr_t end)
385 pte = pte_offset_kernel(pmd, addr);
387 if (!pte_none(*pte) && !kvm_is_device_pfn(pte_pfn(*pte)))
388 kvm_flush_dcache_pte(*pte);
389 } while (pte++, addr += PAGE_SIZE, addr != end);
392 static void stage2_flush_pmds(struct kvm_s2_mmu *mmu, pud_t *pud,
393 phys_addr_t addr, phys_addr_t end)
395 struct kvm *kvm = mmu->kvm;
399 pmd = stage2_pmd_offset(kvm, pud, addr);
401 next = stage2_pmd_addr_end(kvm, addr, end);
402 if (!pmd_none(*pmd)) {
403 if (pmd_thp_or_huge(*pmd))
404 kvm_flush_dcache_pmd(*pmd);
406 stage2_flush_ptes(mmu, pmd, addr, next);
408 } while (pmd++, addr = next, addr != end);
411 static void stage2_flush_puds(struct kvm_s2_mmu *mmu, p4d_t *p4d,
412 phys_addr_t addr, phys_addr_t end)
414 struct kvm *kvm = mmu->kvm;
418 pud = stage2_pud_offset(kvm, p4d, addr);
420 next = stage2_pud_addr_end(kvm, addr, end);
421 if (!stage2_pud_none(kvm, *pud)) {
422 if (stage2_pud_huge(kvm, *pud))
423 kvm_flush_dcache_pud(*pud);
425 stage2_flush_pmds(mmu, pud, addr, next);
427 } while (pud++, addr = next, addr != end);
430 static void stage2_flush_p4ds(struct kvm_s2_mmu *mmu, pgd_t *pgd,
431 phys_addr_t addr, phys_addr_t end)
433 struct kvm *kvm = mmu->kvm;
437 p4d = stage2_p4d_offset(kvm, pgd, addr);
439 next = stage2_p4d_addr_end(kvm, addr, end);
440 if (!stage2_p4d_none(kvm, *p4d))
441 stage2_flush_puds(mmu, p4d, addr, next);
442 } while (p4d++, addr = next, addr != end);
445 static void stage2_flush_memslot(struct kvm *kvm,
446 struct kvm_memory_slot *memslot)
448 struct kvm_s2_mmu *mmu = &kvm->arch.mmu;
449 phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
450 phys_addr_t end = addr + PAGE_SIZE * memslot->npages;
454 pgd = mmu->pgd + stage2_pgd_index(kvm, addr);
456 next = stage2_pgd_addr_end(kvm, addr, end);
457 if (!stage2_pgd_none(kvm, *pgd))
458 stage2_flush_p4ds(mmu, pgd, addr, next);
461 cond_resched_lock(&kvm->mmu_lock);
462 } while (pgd++, addr = next, addr != end);
466 * stage2_flush_vm - Invalidate cache for pages mapped in stage 2
467 * @kvm: The struct kvm pointer
469 * Go through the stage 2 page tables and invalidate any cache lines
470 * backing memory already mapped to the VM.
472 static void stage2_flush_vm(struct kvm *kvm)
474 struct kvm_memslots *slots;
475 struct kvm_memory_slot *memslot;
478 idx = srcu_read_lock(&kvm->srcu);
479 spin_lock(&kvm->mmu_lock);
481 slots = kvm_memslots(kvm);
482 kvm_for_each_memslot(memslot, slots)
483 stage2_flush_memslot(kvm, memslot);
485 spin_unlock(&kvm->mmu_lock);
486 srcu_read_unlock(&kvm->srcu, idx);
490 * free_hyp_pgds - free Hyp-mode page tables
492 void free_hyp_pgds(void)
494 mutex_lock(&kvm_hyp_pgd_mutex);
496 kvm_pgtable_hyp_destroy(hyp_pgtable);
499 mutex_unlock(&kvm_hyp_pgd_mutex);
502 static int __create_hyp_mappings(unsigned long start, unsigned long size,
503 unsigned long phys, enum kvm_pgtable_prot prot)
507 mutex_lock(&kvm_hyp_pgd_mutex);
508 err = kvm_pgtable_hyp_map(hyp_pgtable, start, size, phys, prot);
509 mutex_unlock(&kvm_hyp_pgd_mutex);
514 static phys_addr_t kvm_kaddr_to_phys(void *kaddr)
516 if (!is_vmalloc_addr(kaddr)) {
517 BUG_ON(!virt_addr_valid(kaddr));
520 return page_to_phys(vmalloc_to_page(kaddr)) +
521 offset_in_page(kaddr);
526 * create_hyp_mappings - duplicate a kernel virtual address range in Hyp mode
527 * @from: The virtual kernel start address of the range
528 * @to: The virtual kernel end address of the range (exclusive)
529 * @prot: The protection to be applied to this range
531 * The same virtual address as the kernel virtual address is also used
532 * in Hyp-mode mapping (modulo HYP_PAGE_OFFSET) to the same underlying
535 int create_hyp_mappings(void *from, void *to, enum kvm_pgtable_prot prot)
537 phys_addr_t phys_addr;
538 unsigned long virt_addr;
539 unsigned long start = kern_hyp_va((unsigned long)from);
540 unsigned long end = kern_hyp_va((unsigned long)to);
542 if (is_kernel_in_hyp_mode())
545 start = start & PAGE_MASK;
546 end = PAGE_ALIGN(end);
548 for (virt_addr = start; virt_addr < end; virt_addr += PAGE_SIZE) {
551 phys_addr = kvm_kaddr_to_phys(from + virt_addr - start);
552 err = __create_hyp_mappings(virt_addr, PAGE_SIZE, phys_addr,
561 static int __create_hyp_private_mapping(phys_addr_t phys_addr, size_t size,
562 unsigned long *haddr,
563 enum kvm_pgtable_prot prot)
568 mutex_lock(&kvm_hyp_pgd_mutex);
571 * This assumes that we have enough space below the idmap
572 * page to allocate our VAs. If not, the check below will
573 * kick. A potential alternative would be to detect that
574 * overflow and switch to an allocation above the idmap.
576 * The allocated size is always a multiple of PAGE_SIZE.
578 size = PAGE_ALIGN(size + offset_in_page(phys_addr));
579 base = io_map_base - size;
582 * Verify that BIT(VA_BITS - 1) hasn't been flipped by
583 * allocating the new area, as it would indicate we've
584 * overflowed the idmap/IO address range.
586 if ((base ^ io_map_base) & BIT(VA_BITS - 1))
591 mutex_unlock(&kvm_hyp_pgd_mutex);
596 ret = __create_hyp_mappings(base, size, phys_addr, prot);
600 *haddr = base + offset_in_page(phys_addr);
606 * create_hyp_io_mappings - Map IO into both kernel and HYP
607 * @phys_addr: The physical start address which gets mapped
608 * @size: Size of the region being mapped
609 * @kaddr: Kernel VA for this mapping
610 * @haddr: HYP VA for this mapping
612 int create_hyp_io_mappings(phys_addr_t phys_addr, size_t size,
613 void __iomem **kaddr,
614 void __iomem **haddr)
619 *kaddr = ioremap(phys_addr, size);
623 if (is_kernel_in_hyp_mode()) {
628 ret = __create_hyp_private_mapping(phys_addr, size,
629 &addr, PAGE_HYP_DEVICE);
637 *haddr = (void __iomem *)addr;
642 * create_hyp_exec_mappings - Map an executable range into HYP
643 * @phys_addr: The physical start address which gets mapped
644 * @size: Size of the region being mapped
645 * @haddr: HYP VA for this mapping
647 int create_hyp_exec_mappings(phys_addr_t phys_addr, size_t size,
653 BUG_ON(is_kernel_in_hyp_mode());
655 ret = __create_hyp_private_mapping(phys_addr, size,
656 &addr, PAGE_HYP_EXEC);
662 *haddr = (void *)addr;
667 * kvm_init_stage2_mmu - Initialise a S2 MMU strucrure
668 * @kvm: The pointer to the KVM structure
669 * @mmu: The pointer to the s2 MMU structure
671 * Allocates only the stage-2 HW PGD level table(s).
672 * Note we don't need locking here as this is only called when the VM is
673 * created, which can only be done once.
675 int kvm_init_stage2_mmu(struct kvm *kvm, struct kvm_s2_mmu *mmu)
678 struct kvm_pgtable *pgt;
680 if (mmu->pgt != NULL) {
681 kvm_err("kvm_arch already initialized?\n");
685 pgt = kzalloc(sizeof(*pgt), GFP_KERNEL);
689 err = kvm_pgtable_stage2_init(pgt, kvm);
691 goto out_free_pgtable;
693 mmu->last_vcpu_ran = alloc_percpu(typeof(*mmu->last_vcpu_ran));
694 if (!mmu->last_vcpu_ran) {
696 goto out_destroy_pgtable;
699 for_each_possible_cpu(cpu)
700 *per_cpu_ptr(mmu->last_vcpu_ran, cpu) = -1;
704 mmu->pgd_phys = __pa(pgt->pgd);
705 mmu->pgd = (void *)pgt->pgd;
706 mmu->vmid.vmid_gen = 0;
710 kvm_pgtable_stage2_destroy(pgt);
716 static void stage2_unmap_memslot(struct kvm *kvm,
717 struct kvm_memory_slot *memslot)
719 hva_t hva = memslot->userspace_addr;
720 phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
721 phys_addr_t size = PAGE_SIZE * memslot->npages;
722 hva_t reg_end = hva + size;
725 * A memory region could potentially cover multiple VMAs, and any holes
726 * between them, so iterate over all of them to find out if we should
729 * +--------------------------------------------+
730 * +---------------+----------------+ +----------------+
731 * | : VMA 1 | VMA 2 | | VMA 3 : |
732 * +---------------+----------------+ +----------------+
734 * +--------------------------------------------+
737 struct vm_area_struct *vma = find_vma(current->mm, hva);
738 hva_t vm_start, vm_end;
740 if (!vma || vma->vm_start >= 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 spin_lock(&kvm->mmu_lock);
774 slots = kvm_memslots(kvm);
775 kvm_for_each_memslot(memslot, slots)
776 stage2_unmap_memslot(kvm, memslot);
778 spin_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 = mmu->kvm;
786 struct kvm_pgtable *pgt = NULL;
788 spin_lock(&kvm->mmu_lock);
794 free_percpu(mmu->last_vcpu_ran);
796 spin_unlock(&kvm->mmu_lock);
799 kvm_pgtable_stage2_destroy(pgt);
804 static p4d_t *stage2_get_p4d(struct kvm_s2_mmu *mmu, struct kvm_mmu_memory_cache *cache,
807 struct kvm *kvm = mmu->kvm;
811 pgd = mmu->pgd + stage2_pgd_index(kvm, addr);
812 if (stage2_pgd_none(kvm, *pgd)) {
815 p4d = kvm_mmu_memory_cache_alloc(cache);
816 stage2_pgd_populate(kvm, pgd, p4d);
817 get_page(virt_to_page(pgd));
820 return stage2_p4d_offset(kvm, pgd, addr);
823 static pud_t *stage2_get_pud(struct kvm_s2_mmu *mmu, struct kvm_mmu_memory_cache *cache,
826 struct kvm *kvm = mmu->kvm;
830 p4d = stage2_get_p4d(mmu, cache, addr);
831 if (stage2_p4d_none(kvm, *p4d)) {
834 pud = kvm_mmu_memory_cache_alloc(cache);
835 stage2_p4d_populate(kvm, p4d, pud);
836 get_page(virt_to_page(p4d));
839 return stage2_pud_offset(kvm, p4d, addr);
842 static pmd_t *stage2_get_pmd(struct kvm_s2_mmu *mmu, struct kvm_mmu_memory_cache *cache,
845 struct kvm *kvm = mmu->kvm;
849 pud = stage2_get_pud(mmu, cache, addr);
850 if (!pud || stage2_pud_huge(kvm, *pud))
853 if (stage2_pud_none(kvm, *pud)) {
856 pmd = kvm_mmu_memory_cache_alloc(cache);
857 stage2_pud_populate(kvm, pud, pmd);
858 get_page(virt_to_page(pud));
861 return stage2_pmd_offset(kvm, pud, addr);
864 static int stage2_set_pmd_huge(struct kvm_s2_mmu *mmu,
865 struct kvm_mmu_memory_cache *cache,
866 phys_addr_t addr, const pmd_t *new_pmd)
871 pmd = stage2_get_pmd(mmu, cache, addr);
876 * Multiple vcpus faulting on the same PMD entry, can
877 * lead to them sequentially updating the PMD with the
878 * same value. Following the break-before-make
879 * (pmd_clear() followed by tlb_flush()) process can
880 * hinder forward progress due to refaults generated
881 * on missing translations.
883 * Skip updating the page table if the entry is
886 if (pmd_val(old_pmd) == pmd_val(*new_pmd))
889 if (pmd_present(old_pmd)) {
891 * If we already have PTE level mapping for this block,
892 * we must unmap it to avoid inconsistent TLB state and
893 * leaking the table page. We could end up in this situation
894 * if the memory slot was marked for dirty logging and was
895 * reverted, leaving PTE level mappings for the pages accessed
896 * during the period. So, unmap the PTE level mapping for this
897 * block and retry, as we could have released the upper level
898 * table in the process.
900 * Normal THP split/merge follows mmu_notifier callbacks and do
901 * get handled accordingly.
903 if (!pmd_thp_or_huge(old_pmd)) {
904 unmap_stage2_range(mmu, addr & S2_PMD_MASK, S2_PMD_SIZE);
908 * Mapping in huge pages should only happen through a
909 * fault. If a page is merged into a transparent huge
910 * page, the individual subpages of that huge page
911 * should be unmapped through MMU notifiers before we
914 * Merging of CompoundPages is not supported; they
915 * should become splitting first, unmapped, merged,
916 * and mapped back in on-demand.
918 WARN_ON_ONCE(pmd_pfn(old_pmd) != pmd_pfn(*new_pmd));
920 kvm_tlb_flush_vmid_ipa(mmu, addr, S2_PMD_LEVEL);
922 get_page(virt_to_page(pmd));
925 kvm_set_pmd(pmd, *new_pmd);
929 static int stage2_set_pud_huge(struct kvm_s2_mmu *mmu,
930 struct kvm_mmu_memory_cache *cache,
931 phys_addr_t addr, const pud_t *new_pudp)
933 struct kvm *kvm = mmu->kvm;
934 pud_t *pudp, old_pud;
937 pudp = stage2_get_pud(mmu, cache, addr);
943 * A large number of vcpus faulting on the same stage 2 entry,
944 * can lead to a refault due to the stage2_pud_clear()/tlb_flush().
945 * Skip updating the page tables if there is no change.
947 if (pud_val(old_pud) == pud_val(*new_pudp))
950 if (stage2_pud_present(kvm, old_pud)) {
952 * If we already have table level mapping for this block, unmap
953 * the range for this block and retry.
955 if (!stage2_pud_huge(kvm, old_pud)) {
956 unmap_stage2_range(mmu, addr & S2_PUD_MASK, S2_PUD_SIZE);
960 WARN_ON_ONCE(kvm_pud_pfn(old_pud) != kvm_pud_pfn(*new_pudp));
961 stage2_pud_clear(kvm, pudp);
962 kvm_tlb_flush_vmid_ipa(mmu, addr, S2_PUD_LEVEL);
964 get_page(virt_to_page(pudp));
967 kvm_set_pud(pudp, *new_pudp);
972 * stage2_get_leaf_entry - walk the stage2 VM page tables and return
973 * true if a valid and present leaf-entry is found. A pointer to the
974 * leaf-entry is returned in the appropriate level variable - pudpp,
977 static bool stage2_get_leaf_entry(struct kvm_s2_mmu *mmu, phys_addr_t addr,
978 pud_t **pudpp, pmd_t **pmdpp, pte_t **ptepp)
980 struct kvm *kvm = mmu->kvm;
989 pudp = stage2_get_pud(mmu, NULL, addr);
990 if (!pudp || stage2_pud_none(kvm, *pudp) || !stage2_pud_present(kvm, *pudp))
993 if (stage2_pud_huge(kvm, *pudp)) {
998 pmdp = stage2_pmd_offset(kvm, pudp, addr);
999 if (!pmdp || pmd_none(*pmdp) || !pmd_present(*pmdp))
1002 if (pmd_thp_or_huge(*pmdp)) {
1007 ptep = pte_offset_kernel(pmdp, addr);
1008 if (!ptep || pte_none(*ptep) || !pte_present(*ptep))
1015 static bool stage2_is_exec(struct kvm_s2_mmu *mmu, phys_addr_t addr, unsigned long sz)
1022 found = stage2_get_leaf_entry(mmu, addr, &pudp, &pmdp, &ptep);
1027 return sz <= PUD_SIZE && kvm_s2pud_exec(pudp);
1029 return sz <= PMD_SIZE && kvm_s2pmd_exec(pmdp);
1031 return sz == PAGE_SIZE && kvm_s2pte_exec(ptep);
1034 static int stage2_set_pte(struct kvm_s2_mmu *mmu,
1035 struct kvm_mmu_memory_cache *cache,
1036 phys_addr_t addr, const pte_t *new_pte,
1037 unsigned long flags)
1039 struct kvm *kvm = mmu->kvm;
1042 pte_t *pte, old_pte;
1043 bool iomap = flags & KVM_S2PTE_FLAG_IS_IOMAP;
1044 bool logging_active = flags & KVM_S2_FLAG_LOGGING_ACTIVE;
1046 VM_BUG_ON(logging_active && !cache);
1048 /* Create stage-2 page table mapping - Levels 0 and 1 */
1049 pud = stage2_get_pud(mmu, cache, addr);
1052 * Ignore calls from kvm_set_spte_hva for unallocated
1059 * While dirty page logging - dissolve huge PUD, then continue
1060 * on to allocate page.
1063 stage2_dissolve_pud(mmu, addr, pud);
1065 if (stage2_pud_none(kvm, *pud)) {
1067 return 0; /* ignore calls from kvm_set_spte_hva */
1068 pmd = kvm_mmu_memory_cache_alloc(cache);
1069 stage2_pud_populate(kvm, pud, pmd);
1070 get_page(virt_to_page(pud));
1073 pmd = stage2_pmd_offset(kvm, pud, addr);
1076 * Ignore calls from kvm_set_spte_hva for unallocated
1083 * While dirty page logging - dissolve huge PMD, then continue on to
1087 stage2_dissolve_pmd(mmu, addr, pmd);
1089 /* Create stage-2 page mappings - Level 2 */
1090 if (pmd_none(*pmd)) {
1092 return 0; /* ignore calls from kvm_set_spte_hva */
1093 pte = kvm_mmu_memory_cache_alloc(cache);
1094 kvm_pmd_populate(pmd, pte);
1095 get_page(virt_to_page(pmd));
1098 pte = pte_offset_kernel(pmd, addr);
1100 if (iomap && pte_present(*pte))
1103 /* Create 2nd stage page table mapping - Level 3 */
1105 if (pte_present(old_pte)) {
1106 /* Skip page table update if there is no change */
1107 if (pte_val(old_pte) == pte_val(*new_pte))
1110 kvm_set_pte(pte, __pte(0));
1111 kvm_tlb_flush_vmid_ipa(mmu, addr, S2_PTE_LEVEL);
1113 get_page(virt_to_page(pte));
1116 kvm_set_pte(pte, *new_pte);
1120 #ifndef __HAVE_ARCH_PTEP_TEST_AND_CLEAR_YOUNG
1121 static int stage2_ptep_test_and_clear_young(pte_t *pte)
1123 if (pte_young(*pte)) {
1124 *pte = pte_mkold(*pte);
1130 static int stage2_ptep_test_and_clear_young(pte_t *pte)
1132 return __ptep_test_and_clear_young(pte);
1136 static int stage2_pmdp_test_and_clear_young(pmd_t *pmd)
1138 return stage2_ptep_test_and_clear_young((pte_t *)pmd);
1141 static int stage2_pudp_test_and_clear_young(pud_t *pud)
1143 return stage2_ptep_test_and_clear_young((pte_t *)pud);
1147 * kvm_phys_addr_ioremap - map a device range to guest IPA
1149 * @kvm: The KVM pointer
1150 * @guest_ipa: The IPA at which to insert the mapping
1151 * @pa: The physical address of the device
1152 * @size: The size of the mapping
1154 int kvm_phys_addr_ioremap(struct kvm *kvm, phys_addr_t guest_ipa,
1155 phys_addr_t pa, unsigned long size, bool writable)
1159 struct kvm_mmu_memory_cache cache = { 0, __GFP_ZERO, NULL, };
1160 struct kvm_pgtable *pgt = kvm->arch.mmu.pgt;
1161 enum kvm_pgtable_prot prot = KVM_PGTABLE_PROT_DEVICE |
1162 KVM_PGTABLE_PROT_R |
1163 (writable ? KVM_PGTABLE_PROT_W : 0);
1165 size += offset_in_page(guest_ipa);
1166 guest_ipa &= PAGE_MASK;
1168 for (addr = guest_ipa; addr < guest_ipa + size; addr += PAGE_SIZE) {
1169 ret = kvm_mmu_topup_memory_cache(&cache,
1170 kvm_mmu_cache_min_pages(kvm));
1174 spin_lock(&kvm->mmu_lock);
1175 ret = kvm_pgtable_stage2_map(pgt, addr, PAGE_SIZE, pa, prot,
1177 spin_unlock(&kvm->mmu_lock);
1184 kvm_mmu_free_memory_cache(&cache);
1189 * stage2_wp_ptes - write protect PMD range
1190 * @pmd: pointer to pmd entry
1191 * @addr: range start address
1192 * @end: range end address
1194 static void stage2_wp_ptes(pmd_t *pmd, phys_addr_t addr, phys_addr_t end)
1198 pte = pte_offset_kernel(pmd, addr);
1200 if (!pte_none(*pte)) {
1201 if (!kvm_s2pte_readonly(pte))
1202 kvm_set_s2pte_readonly(pte);
1204 } while (pte++, addr += PAGE_SIZE, addr != end);
1208 * stage2_wp_pmds - write protect PUD range
1209 * kvm: kvm instance for the VM
1210 * @pud: pointer to pud entry
1211 * @addr: range start address
1212 * @end: range end address
1214 static void stage2_wp_pmds(struct kvm_s2_mmu *mmu, pud_t *pud,
1215 phys_addr_t addr, phys_addr_t end)
1217 struct kvm *kvm = mmu->kvm;
1221 pmd = stage2_pmd_offset(kvm, pud, addr);
1224 next = stage2_pmd_addr_end(kvm, addr, end);
1225 if (!pmd_none(*pmd)) {
1226 if (pmd_thp_or_huge(*pmd)) {
1227 if (!kvm_s2pmd_readonly(pmd))
1228 kvm_set_s2pmd_readonly(pmd);
1230 stage2_wp_ptes(pmd, addr, next);
1233 } while (pmd++, addr = next, addr != end);
1237 * stage2_wp_puds - write protect P4D range
1238 * @p4d: pointer to p4d entry
1239 * @addr: range start address
1240 * @end: range end address
1242 static void stage2_wp_puds(struct kvm_s2_mmu *mmu, p4d_t *p4d,
1243 phys_addr_t addr, phys_addr_t end)
1245 struct kvm *kvm = mmu->kvm;
1249 pud = stage2_pud_offset(kvm, p4d, addr);
1251 next = stage2_pud_addr_end(kvm, addr, end);
1252 if (!stage2_pud_none(kvm, *pud)) {
1253 if (stage2_pud_huge(kvm, *pud)) {
1254 if (!kvm_s2pud_readonly(pud))
1255 kvm_set_s2pud_readonly(pud);
1257 stage2_wp_pmds(mmu, pud, addr, next);
1260 } while (pud++, addr = next, addr != end);
1264 * stage2_wp_p4ds - write protect PGD range
1265 * @pgd: pointer to pgd entry
1266 * @addr: range start address
1267 * @end: range end address
1269 static void stage2_wp_p4ds(struct kvm_s2_mmu *mmu, pgd_t *pgd,
1270 phys_addr_t addr, phys_addr_t end)
1272 struct kvm *kvm = mmu->kvm;
1276 p4d = stage2_p4d_offset(kvm, pgd, addr);
1278 next = stage2_p4d_addr_end(kvm, addr, end);
1279 if (!stage2_p4d_none(kvm, *p4d))
1280 stage2_wp_puds(mmu, p4d, addr, next);
1281 } while (p4d++, addr = next, addr != end);
1285 * stage2_wp_range() - write protect stage2 memory region range
1286 * @kvm: The KVM pointer
1287 * @addr: Start address of range
1288 * @end: End address of range
1290 static void stage2_wp_range(struct kvm_s2_mmu *mmu, phys_addr_t addr, phys_addr_t end)
1292 struct kvm *kvm = mmu->kvm;
1296 pgd = mmu->pgd + stage2_pgd_index(kvm, addr);
1299 * Release kvm_mmu_lock periodically if the memory region is
1300 * large. Otherwise, we may see kernel panics with
1301 * CONFIG_DETECT_HUNG_TASK, CONFIG_LOCKUP_DETECTOR,
1302 * CONFIG_LOCKDEP. Additionally, holding the lock too long
1303 * will also starve other vCPUs. We have to also make sure
1304 * that the page tables are not freed while we released
1307 cond_resched_lock(&kvm->mmu_lock);
1308 if (!READ_ONCE(mmu->pgd))
1310 next = stage2_pgd_addr_end(kvm, addr, end);
1311 if (stage2_pgd_present(kvm, *pgd))
1312 stage2_wp_p4ds(mmu, pgd, addr, next);
1313 } while (pgd++, addr = next, addr != end);
1317 * kvm_mmu_wp_memory_region() - write protect stage 2 entries for memory slot
1318 * @kvm: The KVM pointer
1319 * @slot: The memory slot to write protect
1321 * Called to start logging dirty pages after memory region
1322 * KVM_MEM_LOG_DIRTY_PAGES operation is called. After this function returns
1323 * all present PUD, PMD and PTEs are write protected in the memory region.
1324 * Afterwards read of dirty page log can be called.
1326 * Acquires kvm_mmu_lock. Called with kvm->slots_lock mutex acquired,
1327 * serializing operations for VM memory regions.
1329 void kvm_mmu_wp_memory_region(struct kvm *kvm, int slot)
1331 struct kvm_memslots *slots = kvm_memslots(kvm);
1332 struct kvm_memory_slot *memslot = id_to_memslot(slots, slot);
1333 phys_addr_t start, end;
1335 if (WARN_ON_ONCE(!memslot))
1338 start = memslot->base_gfn << PAGE_SHIFT;
1339 end = (memslot->base_gfn + memslot->npages) << PAGE_SHIFT;
1341 spin_lock(&kvm->mmu_lock);
1342 stage2_wp_range(&kvm->arch.mmu, start, end);
1343 spin_unlock(&kvm->mmu_lock);
1344 kvm_flush_remote_tlbs(kvm);
1348 * kvm_mmu_write_protect_pt_masked() - write protect dirty pages
1349 * @kvm: The KVM pointer
1350 * @slot: The memory slot associated with mask
1351 * @gfn_offset: The gfn offset in memory slot
1352 * @mask: The mask of dirty pages at offset 'gfn_offset' in this memory
1353 * slot to be write protected
1355 * Walks bits set in mask write protects the associated pte's. Caller must
1356 * acquire kvm_mmu_lock.
1358 static void kvm_mmu_write_protect_pt_masked(struct kvm *kvm,
1359 struct kvm_memory_slot *slot,
1360 gfn_t gfn_offset, unsigned long mask)
1362 phys_addr_t base_gfn = slot->base_gfn + gfn_offset;
1363 phys_addr_t start = (base_gfn + __ffs(mask)) << PAGE_SHIFT;
1364 phys_addr_t end = (base_gfn + __fls(mask) + 1) << PAGE_SHIFT;
1366 stage2_wp_range(&kvm->arch.mmu, start, end);
1370 * kvm_arch_mmu_enable_log_dirty_pt_masked - enable dirty logging for selected
1373 * It calls kvm_mmu_write_protect_pt_masked to write protect selected pages to
1374 * enable dirty logging for them.
1376 void kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm *kvm,
1377 struct kvm_memory_slot *slot,
1378 gfn_t gfn_offset, unsigned long mask)
1380 kvm_mmu_write_protect_pt_masked(kvm, slot, gfn_offset, mask);
1383 static void clean_dcache_guest_page(kvm_pfn_t pfn, unsigned long size)
1385 __clean_dcache_guest_page(pfn, size);
1388 static void invalidate_icache_guest_page(kvm_pfn_t pfn, unsigned long size)
1390 __invalidate_icache_guest_page(pfn, size);
1393 static void kvm_send_hwpoison_signal(unsigned long address, short lsb)
1395 send_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, lsb, current);
1398 static bool fault_supports_stage2_huge_mapping(struct kvm_memory_slot *memslot,
1400 unsigned long map_size)
1403 hva_t uaddr_start, uaddr_end;
1406 /* The memslot and the VMA are guaranteed to be aligned to PAGE_SIZE */
1407 if (map_size == PAGE_SIZE)
1410 size = memslot->npages * PAGE_SIZE;
1412 gpa_start = memslot->base_gfn << PAGE_SHIFT;
1414 uaddr_start = memslot->userspace_addr;
1415 uaddr_end = uaddr_start + size;
1418 * Pages belonging to memslots that don't have the same alignment
1419 * within a PMD/PUD for userspace and IPA cannot be mapped with stage-2
1420 * PMD/PUD entries, because we'll end up mapping the wrong pages.
1422 * Consider a layout like the following:
1424 * memslot->userspace_addr:
1425 * +-----+--------------------+--------------------+---+
1426 * |abcde|fgh Stage-1 block | Stage-1 block tv|xyz|
1427 * +-----+--------------------+--------------------+---+
1429 * memslot->base_gfn << PAGE_SHIFT:
1430 * +---+--------------------+--------------------+-----+
1431 * |abc|def Stage-2 block | Stage-2 block |tvxyz|
1432 * +---+--------------------+--------------------+-----+
1434 * If we create those stage-2 blocks, we'll end up with this incorrect
1440 if ((gpa_start & (map_size - 1)) != (uaddr_start & (map_size - 1)))
1444 * Next, let's make sure we're not trying to map anything not covered
1445 * by the memslot. This means we have to prohibit block size mappings
1446 * for the beginning and end of a non-block aligned and non-block sized
1447 * memory slot (illustrated by the head and tail parts of the
1448 * userspace view above containing pages 'abcde' and 'xyz',
1451 * Note that it doesn't matter if we do the check using the
1452 * userspace_addr or the base_gfn, as both are equally aligned (per
1453 * the check above) and equally sized.
1455 return (hva & ~(map_size - 1)) >= uaddr_start &&
1456 (hva & ~(map_size - 1)) + map_size <= uaddr_end;
1460 * Check if the given hva is backed by a transparent huge page (THP) and
1461 * whether it can be mapped using block mapping in stage2. If so, adjust
1462 * the stage2 PFN and IPA accordingly. Only PMD_SIZE THPs are currently
1463 * supported. This will need to be updated to support other THP sizes.
1465 * Returns the size of the mapping.
1467 static unsigned long
1468 transparent_hugepage_adjust(struct kvm_memory_slot *memslot,
1469 unsigned long hva, kvm_pfn_t *pfnp,
1472 kvm_pfn_t pfn = *pfnp;
1475 * Make sure the adjustment is done only for THP pages. Also make
1476 * sure that the HVA and IPA are sufficiently aligned and that the
1477 * block map is contained within the memslot.
1479 if (kvm_is_transparent_hugepage(pfn) &&
1480 fault_supports_stage2_huge_mapping(memslot, hva, PMD_SIZE)) {
1482 * The address we faulted on is backed by a transparent huge
1483 * page. However, because we map the compound huge page and
1484 * not the individual tail page, we need to transfer the
1485 * refcount to the head page. We have to be careful that the
1486 * THP doesn't start to split while we are adjusting the
1489 * We are sure this doesn't happen, because mmu_notifier_retry
1490 * was successful and we are holding the mmu_lock, so if this
1491 * THP is trying to split, it will be blocked in the mmu
1492 * notifier before touching any of the pages, specifically
1493 * before being able to call __split_huge_page_refcount().
1495 * We can therefore safely transfer the refcount from PG_tail
1496 * to PG_head and switch the pfn from a tail page to the head
1500 kvm_release_pfn_clean(pfn);
1501 pfn &= ~(PTRS_PER_PMD - 1);
1508 /* Use page mapping if we cannot use block mapping. */
1512 static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
1513 struct kvm_memory_slot *memslot, unsigned long hva,
1514 unsigned long fault_status)
1517 bool write_fault, writable, force_pte = false;
1518 bool exec_fault, needs_exec;
1519 unsigned long mmu_seq;
1520 gfn_t gfn = fault_ipa >> PAGE_SHIFT;
1521 struct kvm *kvm = vcpu->kvm;
1522 struct kvm_mmu_memory_cache *memcache = &vcpu->arch.mmu_page_cache;
1523 struct vm_area_struct *vma;
1526 pgprot_t mem_type = PAGE_S2;
1527 bool logging_active = memslot_is_logging(memslot);
1528 unsigned long vma_pagesize, flags = 0;
1529 struct kvm_s2_mmu *mmu = vcpu->arch.hw_mmu;
1531 write_fault = kvm_is_write_fault(vcpu);
1532 exec_fault = kvm_vcpu_trap_is_iabt(vcpu);
1533 VM_BUG_ON(write_fault && exec_fault);
1535 if (fault_status == FSC_PERM && !write_fault && !exec_fault) {
1536 kvm_err("Unexpected L2 read permission error\n");
1540 /* Let's check if we will get back a huge page backed by hugetlbfs */
1541 mmap_read_lock(current->mm);
1542 vma = find_vma_intersection(current->mm, hva, hva + 1);
1543 if (unlikely(!vma)) {
1544 kvm_err("Failed to find VMA for hva 0x%lx\n", hva);
1545 mmap_read_unlock(current->mm);
1549 if (is_vm_hugetlb_page(vma))
1550 vma_shift = huge_page_shift(hstate_vma(vma));
1552 vma_shift = PAGE_SHIFT;
1554 vma_pagesize = 1ULL << vma_shift;
1555 if (logging_active ||
1556 (vma->vm_flags & VM_PFNMAP) ||
1557 !fault_supports_stage2_huge_mapping(memslot, hva, vma_pagesize)) {
1559 vma_pagesize = PAGE_SIZE;
1563 * The stage2 has a minimum of 2 level table (For arm64 see
1564 * kvm_arm_setup_stage2()). Hence, we are guaranteed that we can
1565 * use PMD_SIZE huge mappings (even when the PMD is folded into PGD).
1566 * As for PUD huge maps, we must make sure that we have at least
1567 * 3 levels, i.e, PMD is not folded.
1569 if (vma_pagesize == PMD_SIZE ||
1570 (vma_pagesize == PUD_SIZE && kvm_stage2_has_pmd(kvm)))
1571 gfn = (fault_ipa & huge_page_mask(hstate_vma(vma))) >> PAGE_SHIFT;
1572 mmap_read_unlock(current->mm);
1574 /* We need minimum second+third level pages */
1575 ret = kvm_mmu_topup_memory_cache(memcache, kvm_mmu_cache_min_pages(kvm));
1579 mmu_seq = vcpu->kvm->mmu_notifier_seq;
1581 * Ensure the read of mmu_notifier_seq happens before we call
1582 * gfn_to_pfn_prot (which calls get_user_pages), so that we don't risk
1583 * the page we just got a reference to gets unmapped before we have a
1584 * chance to grab the mmu_lock, which ensure that if the page gets
1585 * unmapped afterwards, the call to kvm_unmap_hva will take it away
1586 * from us again properly. This smp_rmb() interacts with the smp_wmb()
1587 * in kvm_mmu_notifier_invalidate_<page|range_end>.
1591 pfn = gfn_to_pfn_prot(kvm, gfn, write_fault, &writable);
1592 if (pfn == KVM_PFN_ERR_HWPOISON) {
1593 kvm_send_hwpoison_signal(hva, vma_shift);
1596 if (is_error_noslot_pfn(pfn))
1599 if (kvm_is_device_pfn(pfn)) {
1600 mem_type = PAGE_S2_DEVICE;
1601 flags |= KVM_S2PTE_FLAG_IS_IOMAP;
1602 } else if (logging_active) {
1604 * Faults on pages in a memslot with logging enabled
1605 * should not be mapped with huge pages (it introduces churn
1606 * and performance degradation), so force a pte mapping.
1608 flags |= KVM_S2_FLAG_LOGGING_ACTIVE;
1611 * Only actually map the page as writable if this was a write
1618 if (exec_fault && is_iomap(flags))
1621 spin_lock(&kvm->mmu_lock);
1622 if (mmu_notifier_retry(kvm, mmu_seq))
1626 * If we are not forced to use page mapping, check if we are
1627 * backed by a THP and thus use block mapping if possible.
1629 if (vma_pagesize == PAGE_SIZE && !force_pte)
1630 vma_pagesize = transparent_hugepage_adjust(memslot, hva,
1633 kvm_set_pfn_dirty(pfn);
1635 if (fault_status != FSC_PERM && !is_iomap(flags))
1636 clean_dcache_guest_page(pfn, vma_pagesize);
1639 invalidate_icache_guest_page(pfn, vma_pagesize);
1642 * If we took an execution fault we have made the
1643 * icache/dcache coherent above and should now let the s2
1644 * mapping be executable.
1646 * Write faults (!exec_fault && FSC_PERM) are orthogonal to
1647 * execute permissions, and we preserve whatever we have.
1649 needs_exec = exec_fault ||
1650 (fault_status == FSC_PERM &&
1651 stage2_is_exec(mmu, fault_ipa, vma_pagesize));
1653 if (vma_pagesize == PUD_SIZE) {
1654 pud_t new_pud = kvm_pfn_pud(pfn, mem_type);
1656 new_pud = kvm_pud_mkhuge(new_pud);
1658 new_pud = kvm_s2pud_mkwrite(new_pud);
1661 new_pud = kvm_s2pud_mkexec(new_pud);
1663 ret = stage2_set_pud_huge(mmu, memcache, fault_ipa, &new_pud);
1664 } else if (vma_pagesize == PMD_SIZE) {
1665 pmd_t new_pmd = kvm_pfn_pmd(pfn, mem_type);
1667 new_pmd = kvm_pmd_mkhuge(new_pmd);
1670 new_pmd = kvm_s2pmd_mkwrite(new_pmd);
1673 new_pmd = kvm_s2pmd_mkexec(new_pmd);
1675 ret = stage2_set_pmd_huge(mmu, memcache, fault_ipa, &new_pmd);
1677 pte_t new_pte = kvm_pfn_pte(pfn, mem_type);
1680 new_pte = kvm_s2pte_mkwrite(new_pte);
1681 mark_page_dirty(kvm, gfn);
1685 new_pte = kvm_s2pte_mkexec(new_pte);
1687 ret = stage2_set_pte(mmu, memcache, fault_ipa, &new_pte, flags);
1691 spin_unlock(&kvm->mmu_lock);
1692 kvm_set_pfn_accessed(pfn);
1693 kvm_release_pfn_clean(pfn);
1698 * Resolve the access fault by making the page young again.
1699 * Note that because the faulting entry is guaranteed not to be
1700 * cached in the TLB, we don't need to invalidate anything.
1701 * Only the HW Access Flag updates are supported for Stage 2 (no DBM),
1702 * so there is no need for atomic (pte|pmd)_mkyoung operations.
1704 static void handle_access_fault(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa)
1710 bool pfn_valid = false;
1712 trace_kvm_access_fault(fault_ipa);
1714 spin_lock(&vcpu->kvm->mmu_lock);
1716 if (!stage2_get_leaf_entry(vcpu->arch.hw_mmu, fault_ipa, &pud, &pmd, &pte))
1719 if (pud) { /* HugeTLB */
1720 *pud = kvm_s2pud_mkyoung(*pud);
1721 pfn = kvm_pud_pfn(*pud);
1723 } else if (pmd) { /* THP, HugeTLB */
1724 *pmd = pmd_mkyoung(*pmd);
1725 pfn = pmd_pfn(*pmd);
1728 *pte = pte_mkyoung(*pte); /* Just a page... */
1729 pfn = pte_pfn(*pte);
1734 spin_unlock(&vcpu->kvm->mmu_lock);
1736 kvm_set_pfn_accessed(pfn);
1740 * kvm_handle_guest_abort - handles all 2nd stage aborts
1741 * @vcpu: the VCPU pointer
1743 * Any abort that gets to the host is almost guaranteed to be caused by a
1744 * missing second stage translation table entry, which can mean that either the
1745 * guest simply needs more memory and we must allocate an appropriate page or it
1746 * can mean that the guest tried to access I/O memory, which is emulated by user
1747 * space. The distinction is based on the IPA causing the fault and whether this
1748 * memory region has been registered as standard RAM by user space.
1750 int kvm_handle_guest_abort(struct kvm_vcpu *vcpu)
1752 unsigned long fault_status;
1753 phys_addr_t fault_ipa;
1754 struct kvm_memory_slot *memslot;
1756 bool is_iabt, write_fault, writable;
1760 fault_status = kvm_vcpu_trap_get_fault_type(vcpu);
1762 fault_ipa = kvm_vcpu_get_fault_ipa(vcpu);
1763 is_iabt = kvm_vcpu_trap_is_iabt(vcpu);
1765 /* Synchronous External Abort? */
1766 if (kvm_vcpu_abt_issea(vcpu)) {
1768 * For RAS the host kernel may handle this abort.
1769 * There is no need to pass the error into the guest.
1771 if (kvm_handle_guest_sea(fault_ipa, kvm_vcpu_get_esr(vcpu)))
1772 kvm_inject_vabt(vcpu);
1777 trace_kvm_guest_fault(*vcpu_pc(vcpu), kvm_vcpu_get_esr(vcpu),
1778 kvm_vcpu_get_hfar(vcpu), fault_ipa);
1780 /* Check the stage-2 fault is trans. fault or write fault */
1781 if (fault_status != FSC_FAULT && fault_status != FSC_PERM &&
1782 fault_status != FSC_ACCESS) {
1783 kvm_err("Unsupported FSC: EC=%#x xFSC=%#lx ESR_EL2=%#lx\n",
1784 kvm_vcpu_trap_get_class(vcpu),
1785 (unsigned long)kvm_vcpu_trap_get_fault(vcpu),
1786 (unsigned long)kvm_vcpu_get_esr(vcpu));
1790 idx = srcu_read_lock(&vcpu->kvm->srcu);
1792 gfn = fault_ipa >> PAGE_SHIFT;
1793 memslot = gfn_to_memslot(vcpu->kvm, gfn);
1794 hva = gfn_to_hva_memslot_prot(memslot, gfn, &writable);
1795 write_fault = kvm_is_write_fault(vcpu);
1796 if (kvm_is_error_hva(hva) || (write_fault && !writable)) {
1798 * The guest has put either its instructions or its page-tables
1799 * somewhere it shouldn't have. Userspace won't be able to do
1800 * anything about this (there's no syndrome for a start), so
1801 * re-inject the abort back into the guest.
1808 if (kvm_vcpu_dabt_iss1tw(vcpu)) {
1809 kvm_inject_dabt(vcpu, kvm_vcpu_get_hfar(vcpu));
1815 * Check for a cache maintenance operation. Since we
1816 * ended-up here, we know it is outside of any memory
1817 * slot. But we can't find out if that is for a device,
1818 * or if the guest is just being stupid. The only thing
1819 * we know for sure is that this range cannot be cached.
1821 * So let's assume that the guest is just being
1822 * cautious, and skip the instruction.
1824 if (kvm_is_error_hva(hva) && kvm_vcpu_dabt_is_cm(vcpu)) {
1825 kvm_skip_instr(vcpu, kvm_vcpu_trap_il_is32bit(vcpu));
1831 * The IPA is reported as [MAX:12], so we need to
1832 * complement it with the bottom 12 bits from the
1833 * faulting VA. This is always 12 bits, irrespective
1836 fault_ipa |= kvm_vcpu_get_hfar(vcpu) & ((1 << 12) - 1);
1837 ret = io_mem_abort(vcpu, fault_ipa);
1841 /* Userspace should not be able to register out-of-bounds IPAs */
1842 VM_BUG_ON(fault_ipa >= kvm_phys_size(vcpu->kvm));
1844 if (fault_status == FSC_ACCESS) {
1845 handle_access_fault(vcpu, fault_ipa);
1850 ret = user_mem_abort(vcpu, fault_ipa, memslot, hva, fault_status);
1854 if (ret == -ENOEXEC) {
1855 kvm_inject_pabt(vcpu, kvm_vcpu_get_hfar(vcpu));
1859 srcu_read_unlock(&vcpu->kvm->srcu, idx);
1863 static int handle_hva_to_gpa(struct kvm *kvm,
1864 unsigned long start,
1866 int (*handler)(struct kvm *kvm,
1867 gpa_t gpa, u64 size,
1871 struct kvm_memslots *slots;
1872 struct kvm_memory_slot *memslot;
1875 slots = kvm_memslots(kvm);
1877 /* we only care about the pages that the guest sees */
1878 kvm_for_each_memslot(memslot, slots) {
1879 unsigned long hva_start, hva_end;
1882 hva_start = max(start, memslot->userspace_addr);
1883 hva_end = min(end, memslot->userspace_addr +
1884 (memslot->npages << PAGE_SHIFT));
1885 if (hva_start >= hva_end)
1888 gpa = hva_to_gfn_memslot(hva_start, memslot) << PAGE_SHIFT;
1889 ret |= handler(kvm, gpa, (u64)(hva_end - hva_start), data);
1895 static int kvm_unmap_hva_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
1897 unsigned flags = *(unsigned *)data;
1898 bool may_block = flags & MMU_NOTIFIER_RANGE_BLOCKABLE;
1900 __unmap_stage2_range(&kvm->arch.mmu, gpa, size, may_block);
1904 int kvm_unmap_hva_range(struct kvm *kvm,
1905 unsigned long start, unsigned long end, unsigned flags)
1907 if (!kvm->arch.mmu.pgd)
1910 trace_kvm_unmap_hva_range(start, end);
1911 handle_hva_to_gpa(kvm, start, end, &kvm_unmap_hva_handler, &flags);
1915 static int kvm_set_spte_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
1917 kvm_pfn_t *pfn = (kvm_pfn_t *)data;
1919 WARN_ON(size != PAGE_SIZE);
1922 * The MMU notifiers will have unmapped a huge PMD before calling
1923 * ->change_pte() (which in turn calls kvm_set_spte_hva()) and
1924 * therefore we never need to clear out a huge PMD through this
1925 * calling path and a memcache is not required.
1927 kvm_pgtable_stage2_map(kvm->arch.mmu.pgt, gpa, PAGE_SIZE,
1928 __pfn_to_phys(*pfn), KVM_PGTABLE_PROT_R, NULL);
1932 int kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
1934 unsigned long end = hva + PAGE_SIZE;
1935 kvm_pfn_t pfn = pte_pfn(pte);
1937 if (!kvm->arch.mmu.pgt)
1940 trace_kvm_set_spte_hva(hva);
1943 * We've moved a page around, probably through CoW, so let's treat it
1944 * just like a translation fault and clean the cache to the PoC.
1946 clean_dcache_guest_page(pfn, PAGE_SIZE);
1947 handle_hva_to_gpa(kvm, hva, end, &kvm_set_spte_handler, &pfn);
1951 static int kvm_age_hva_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
1957 WARN_ON(size != PAGE_SIZE && size != PMD_SIZE && size != PUD_SIZE);
1958 if (!stage2_get_leaf_entry(&kvm->arch.mmu, gpa, &pud, &pmd, &pte))
1962 return stage2_pudp_test_and_clear_young(pud);
1964 return stage2_pmdp_test_and_clear_young(pmd);
1966 return stage2_ptep_test_and_clear_young(pte);
1969 static int kvm_test_age_hva_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
1975 WARN_ON(size != PAGE_SIZE && size != PMD_SIZE && size != PUD_SIZE);
1976 if (!stage2_get_leaf_entry(&kvm->arch.mmu, gpa, &pud, &pmd, &pte))
1980 return kvm_s2pud_young(*pud);
1982 return pmd_young(*pmd);
1984 return pte_young(*pte);
1987 int kvm_age_hva(struct kvm *kvm, unsigned long start, unsigned long end)
1989 if (!kvm->arch.mmu.pgd)
1991 trace_kvm_age_hva(start, end);
1992 return handle_hva_to_gpa(kvm, start, end, kvm_age_hva_handler, NULL);
1995 int kvm_test_age_hva(struct kvm *kvm, unsigned long hva)
1997 if (!kvm->arch.mmu.pgd)
1999 trace_kvm_test_age_hva(hva);
2000 return handle_hva_to_gpa(kvm, hva, hva + PAGE_SIZE,
2001 kvm_test_age_hva_handler, NULL);
2004 phys_addr_t kvm_mmu_get_httbr(void)
2006 return __pa(hyp_pgtable->pgd);
2009 phys_addr_t kvm_get_idmap_vector(void)
2011 return hyp_idmap_vector;
2014 static int kvm_map_idmap_text(void)
2016 unsigned long size = hyp_idmap_end - hyp_idmap_start;
2017 int err = __create_hyp_mappings(hyp_idmap_start, size, hyp_idmap_start,
2020 kvm_err("Failed to idmap %lx-%lx\n",
2021 hyp_idmap_start, hyp_idmap_end);
2026 int kvm_mmu_init(void)
2031 hyp_idmap_start = __pa_symbol(__hyp_idmap_text_start);
2032 hyp_idmap_start = ALIGN_DOWN(hyp_idmap_start, PAGE_SIZE);
2033 hyp_idmap_end = __pa_symbol(__hyp_idmap_text_end);
2034 hyp_idmap_end = ALIGN(hyp_idmap_end, PAGE_SIZE);
2035 hyp_idmap_vector = __pa_symbol(__kvm_hyp_init);
2038 * We rely on the linker script to ensure at build time that the HYP
2039 * init code does not cross a page boundary.
2041 BUG_ON((hyp_idmap_start ^ (hyp_idmap_end - 1)) & PAGE_MASK);
2043 hyp_va_bits = 64 - ((idmap_t0sz & TCR_T0SZ_MASK) >> TCR_T0SZ_OFFSET);
2044 kvm_debug("Using %u-bit virtual addresses at EL2\n", hyp_va_bits);
2045 kvm_debug("IDMAP page: %lx\n", hyp_idmap_start);
2046 kvm_debug("HYP VA range: %lx:%lx\n",
2047 kern_hyp_va(PAGE_OFFSET),
2048 kern_hyp_va((unsigned long)high_memory - 1));
2050 if (hyp_idmap_start >= kern_hyp_va(PAGE_OFFSET) &&
2051 hyp_idmap_start < kern_hyp_va((unsigned long)high_memory - 1) &&
2052 hyp_idmap_start != (unsigned long)__hyp_idmap_text_start) {
2054 * The idmap page is intersecting with the VA space,
2055 * it is not safe to continue further.
2057 kvm_err("IDMAP intersecting with HYP VA, unable to continue\n");
2062 hyp_pgtable = kzalloc(sizeof(*hyp_pgtable), GFP_KERNEL);
2064 kvm_err("Hyp mode page-table not allocated\n");
2069 err = kvm_pgtable_hyp_init(hyp_pgtable, hyp_va_bits);
2071 goto out_free_pgtable;
2073 err = kvm_map_idmap_text();
2075 goto out_destroy_pgtable;
2077 io_map_base = hyp_idmap_start;
2080 out_destroy_pgtable:
2081 kvm_pgtable_hyp_destroy(hyp_pgtable);
2089 void kvm_arch_commit_memory_region(struct kvm *kvm,
2090 const struct kvm_userspace_memory_region *mem,
2091 struct kvm_memory_slot *old,
2092 const struct kvm_memory_slot *new,
2093 enum kvm_mr_change change)
2096 * At this point memslot has been committed and there is an
2097 * allocated dirty_bitmap[], dirty pages will be tracked while the
2098 * memory slot is write protected.
2100 if (change != KVM_MR_DELETE && mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
2102 * If we're with initial-all-set, we don't need to write
2103 * protect any pages because they're all reported as dirty.
2104 * Huge pages and normal pages will be write protect gradually.
2106 if (!kvm_dirty_log_manual_protect_and_init_set(kvm)) {
2107 kvm_mmu_wp_memory_region(kvm, mem->slot);
2112 int kvm_arch_prepare_memory_region(struct kvm *kvm,
2113 struct kvm_memory_slot *memslot,
2114 const struct kvm_userspace_memory_region *mem,
2115 enum kvm_mr_change change)
2117 hva_t hva = mem->userspace_addr;
2118 hva_t reg_end = hva + mem->memory_size;
2119 bool writable = !(mem->flags & KVM_MEM_READONLY);
2122 if (change != KVM_MR_CREATE && change != KVM_MR_MOVE &&
2123 change != KVM_MR_FLAGS_ONLY)
2127 * Prevent userspace from creating a memory region outside of the IPA
2128 * space addressable by the KVM guest IPA space.
2130 if (memslot->base_gfn + memslot->npages >=
2131 (kvm_phys_size(kvm) >> PAGE_SHIFT))
2134 mmap_read_lock(current->mm);
2136 * A memory region could potentially cover multiple VMAs, and any holes
2137 * between them, so iterate over all of them to find out if we can map
2138 * any of them right now.
2140 * +--------------------------------------------+
2141 * +---------------+----------------+ +----------------+
2142 * | : VMA 1 | VMA 2 | | VMA 3 : |
2143 * +---------------+----------------+ +----------------+
2145 * +--------------------------------------------+
2148 struct vm_area_struct *vma = find_vma(current->mm, hva);
2149 hva_t vm_start, vm_end;
2151 if (!vma || vma->vm_start >= reg_end)
2155 * Take the intersection of this VMA with the memory region
2157 vm_start = max(hva, vma->vm_start);
2158 vm_end = min(reg_end, vma->vm_end);
2160 if (vma->vm_flags & VM_PFNMAP) {
2161 gpa_t gpa = mem->guest_phys_addr +
2162 (vm_start - mem->userspace_addr);
2165 pa = (phys_addr_t)vma->vm_pgoff << PAGE_SHIFT;
2166 pa += vm_start - vma->vm_start;
2168 /* IO region dirty page logging not allowed */
2169 if (memslot->flags & KVM_MEM_LOG_DIRTY_PAGES) {
2174 ret = kvm_phys_addr_ioremap(kvm, gpa, pa,
2181 } while (hva < reg_end);
2183 if (change == KVM_MR_FLAGS_ONLY)
2186 spin_lock(&kvm->mmu_lock);
2188 unmap_stage2_range(&kvm->arch.mmu, mem->guest_phys_addr, mem->memory_size);
2190 stage2_flush_memslot(kvm, memslot);
2191 spin_unlock(&kvm->mmu_lock);
2193 mmap_read_unlock(current->mm);
2197 void kvm_arch_free_memslot(struct kvm *kvm, struct kvm_memory_slot *slot)
2201 void kvm_arch_memslots_updated(struct kvm *kvm, u64 gen)
2205 void kvm_arch_flush_shadow_all(struct kvm *kvm)
2207 kvm_free_stage2_pgd(&kvm->arch.mmu);
2210 void kvm_arch_flush_shadow_memslot(struct kvm *kvm,
2211 struct kvm_memory_slot *slot)
2213 gpa_t gpa = slot->base_gfn << PAGE_SHIFT;
2214 phys_addr_t size = slot->npages << PAGE_SHIFT;
2216 spin_lock(&kvm->mmu_lock);
2217 unmap_stage2_range(&kvm->arch.mmu, gpa, size);
2218 spin_unlock(&kvm->mmu_lock);
2222 * See note at ARMv7 ARM B1.14.4 (TL;DR: S/W ops are not easily virtualized).
2225 * - S/W ops are local to a CPU (not broadcast)
2226 * - We have line migration behind our back (speculation)
2227 * - System caches don't support S/W at all (damn!)
2229 * In the face of the above, the best we can do is to try and convert
2230 * S/W ops to VA ops. Because the guest is not allowed to infer the
2231 * S/W to PA mapping, it can only use S/W to nuke the whole cache,
2232 * which is a rather good thing for us.
2234 * Also, it is only used when turning caches on/off ("The expected
2235 * usage of the cache maintenance instructions that operate by set/way
2236 * is associated with the cache maintenance instructions associated
2237 * with the powerdown and powerup of caches, if this is required by
2238 * the implementation.").
2240 * We use the following policy:
2242 * - If we trap a S/W operation, we enable VM trapping to detect
2243 * caches being turned on/off, and do a full clean.
2245 * - We flush the caches on both caches being turned on and off.
2247 * - Once the caches are enabled, we stop trapping VM ops.
2249 void kvm_set_way_flush(struct kvm_vcpu *vcpu)
2251 unsigned long hcr = *vcpu_hcr(vcpu);
2254 * If this is the first time we do a S/W operation
2255 * (i.e. HCR_TVM not set) flush the whole memory, and set the
2258 * Otherwise, rely on the VM trapping to wait for the MMU +
2259 * Caches to be turned off. At that point, we'll be able to
2260 * clean the caches again.
2262 if (!(hcr & HCR_TVM)) {
2263 trace_kvm_set_way_flush(*vcpu_pc(vcpu),
2264 vcpu_has_cache_enabled(vcpu));
2265 stage2_flush_vm(vcpu->kvm);
2266 *vcpu_hcr(vcpu) = hcr | HCR_TVM;
2270 void kvm_toggle_cache(struct kvm_vcpu *vcpu, bool was_enabled)
2272 bool now_enabled = vcpu_has_cache_enabled(vcpu);
2275 * If switching the MMU+caches on, need to invalidate the caches.
2276 * If switching it off, need to clean the caches.
2277 * Clean + invalidate does the trick always.
2279 if (now_enabled != was_enabled)
2280 stage2_flush_vm(vcpu->kvm);
2282 /* Caches are now on, stop trapping VM ops (until a S/W op) */
2284 *vcpu_hcr(vcpu) &= ~HCR_TVM;
2286 trace_kvm_toggle_cache(*vcpu_pc(vcpu), was_enabled, now_enabled);