2 * Copyright (C) 2012 - Virtual Open Systems and Columbia University
3 * Author: Christoffer Dall <c.dall@virtualopensystems.com>
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License, version 2, as
7 * published by the Free Software Foundation.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
19 #include <linux/mman.h>
20 #include <linux/kvm_host.h>
22 #include <linux/hugetlb.h>
23 #include <linux/sched/signal.h>
24 #include <trace/events/kvm.h>
25 #include <asm/pgalloc.h>
26 #include <asm/cacheflush.h>
27 #include <asm/kvm_arm.h>
28 #include <asm/kvm_mmu.h>
29 #include <asm/kvm_mmio.h>
30 #include <asm/kvm_asm.h>
31 #include <asm/kvm_emulate.h>
33 #include <asm/system_misc.h>
37 static pgd_t *boot_hyp_pgd;
38 static pgd_t *hyp_pgd;
39 static pgd_t *merged_hyp_pgd;
40 static DEFINE_MUTEX(kvm_hyp_pgd_mutex);
42 static unsigned long hyp_idmap_start;
43 static unsigned long hyp_idmap_end;
44 static phys_addr_t hyp_idmap_vector;
46 #define S2_PGD_SIZE (PTRS_PER_S2_PGD * sizeof(pgd_t))
47 #define hyp_pgd_order get_order(PTRS_PER_PGD * sizeof(pgd_t))
49 #define KVM_S2PTE_FLAG_IS_IOMAP (1UL << 0)
50 #define KVM_S2_FLAG_LOGGING_ACTIVE (1UL << 1)
52 static bool memslot_is_logging(struct kvm_memory_slot *memslot)
54 return memslot->dirty_bitmap && !(memslot->flags & KVM_MEM_READONLY);
58 * kvm_flush_remote_tlbs() - flush all VM TLB entries for v7/8
59 * @kvm: pointer to kvm structure.
61 * Interface to HYP function to flush all VM TLB entries
63 void kvm_flush_remote_tlbs(struct kvm *kvm)
65 kvm_call_hyp(__kvm_tlb_flush_vmid, kvm);
68 static void kvm_tlb_flush_vmid_ipa(struct kvm *kvm, phys_addr_t ipa)
70 kvm_call_hyp(__kvm_tlb_flush_vmid_ipa, kvm, ipa);
74 * D-Cache management functions. They take the page table entries by
75 * value, as they are flushing the cache using the kernel mapping (or
78 static void kvm_flush_dcache_pte(pte_t pte)
80 __kvm_flush_dcache_pte(pte);
83 static void kvm_flush_dcache_pmd(pmd_t pmd)
85 __kvm_flush_dcache_pmd(pmd);
88 static void kvm_flush_dcache_pud(pud_t pud)
90 __kvm_flush_dcache_pud(pud);
93 static bool kvm_is_device_pfn(unsigned long pfn)
95 return !pfn_valid(pfn);
99 * stage2_dissolve_pmd() - clear and flush huge PMD entry
100 * @kvm: pointer to kvm structure.
102 * @pmd: pmd pointer for IPA
104 * Function clears a PMD entry, flushes addr 1st and 2nd stage TLBs. Marks all
105 * pages in the range dirty.
107 static void stage2_dissolve_pmd(struct kvm *kvm, phys_addr_t addr, pmd_t *pmd)
109 if (!pmd_thp_or_huge(*pmd))
113 kvm_tlb_flush_vmid_ipa(kvm, addr);
114 put_page(virt_to_page(pmd));
117 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
122 BUG_ON(max > KVM_NR_MEM_OBJS);
123 if (cache->nobjs >= min)
125 while (cache->nobjs < max) {
126 page = (void *)__get_free_page(PGALLOC_GFP);
129 cache->objects[cache->nobjs++] = page;
134 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
137 free_page((unsigned long)mc->objects[--mc->nobjs]);
140 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc)
144 BUG_ON(!mc || !mc->nobjs);
145 p = mc->objects[--mc->nobjs];
149 static void clear_stage2_pgd_entry(struct kvm *kvm, pgd_t *pgd, phys_addr_t addr)
151 pud_t *pud_table __maybe_unused = stage2_pud_offset(pgd, 0UL);
152 stage2_pgd_clear(pgd);
153 kvm_tlb_flush_vmid_ipa(kvm, addr);
154 stage2_pud_free(pud_table);
155 put_page(virt_to_page(pgd));
158 static void clear_stage2_pud_entry(struct kvm *kvm, pud_t *pud, phys_addr_t addr)
160 pmd_t *pmd_table __maybe_unused = stage2_pmd_offset(pud, 0);
161 VM_BUG_ON(stage2_pud_huge(*pud));
162 stage2_pud_clear(pud);
163 kvm_tlb_flush_vmid_ipa(kvm, addr);
164 stage2_pmd_free(pmd_table);
165 put_page(virt_to_page(pud));
168 static void clear_stage2_pmd_entry(struct kvm *kvm, pmd_t *pmd, phys_addr_t addr)
170 pte_t *pte_table = pte_offset_kernel(pmd, 0);
171 VM_BUG_ON(pmd_thp_or_huge(*pmd));
173 kvm_tlb_flush_vmid_ipa(kvm, addr);
174 pte_free_kernel(NULL, pte_table);
175 put_page(virt_to_page(pmd));
179 * Unmapping vs dcache management:
181 * If a guest maps certain memory pages as uncached, all writes will
182 * bypass the data cache and go directly to RAM. However, the CPUs
183 * can still speculate reads (not writes) and fill cache lines with
186 * Those cache lines will be *clean* cache lines though, so a
187 * clean+invalidate operation is equivalent to an invalidate
188 * operation, because no cache lines are marked dirty.
190 * Those clean cache lines could be filled prior to an uncached write
191 * by the guest, and the cache coherent IO subsystem would therefore
192 * end up writing old data to disk.
194 * This is why right after unmapping a page/section and invalidating
195 * the corresponding TLBs, we call kvm_flush_dcache_p*() to make sure
196 * the IO subsystem will never hit in the cache.
198 static void unmap_stage2_ptes(struct kvm *kvm, pmd_t *pmd,
199 phys_addr_t addr, phys_addr_t end)
201 phys_addr_t start_addr = addr;
202 pte_t *pte, *start_pte;
204 start_pte = pte = pte_offset_kernel(pmd, addr);
206 if (!pte_none(*pte)) {
207 pte_t old_pte = *pte;
209 kvm_set_pte(pte, __pte(0));
210 kvm_tlb_flush_vmid_ipa(kvm, addr);
212 /* No need to invalidate the cache for device mappings */
213 if (!kvm_is_device_pfn(pte_pfn(old_pte)))
214 kvm_flush_dcache_pte(old_pte);
216 put_page(virt_to_page(pte));
218 } while (pte++, addr += PAGE_SIZE, addr != end);
220 if (stage2_pte_table_empty(start_pte))
221 clear_stage2_pmd_entry(kvm, pmd, start_addr);
224 static void unmap_stage2_pmds(struct kvm *kvm, pud_t *pud,
225 phys_addr_t addr, phys_addr_t end)
227 phys_addr_t next, start_addr = addr;
228 pmd_t *pmd, *start_pmd;
230 start_pmd = pmd = stage2_pmd_offset(pud, addr);
232 next = stage2_pmd_addr_end(addr, end);
233 if (!pmd_none(*pmd)) {
234 if (pmd_thp_or_huge(*pmd)) {
235 pmd_t old_pmd = *pmd;
238 kvm_tlb_flush_vmid_ipa(kvm, addr);
240 kvm_flush_dcache_pmd(old_pmd);
242 put_page(virt_to_page(pmd));
244 unmap_stage2_ptes(kvm, pmd, addr, next);
247 } while (pmd++, addr = next, addr != end);
249 if (stage2_pmd_table_empty(start_pmd))
250 clear_stage2_pud_entry(kvm, pud, start_addr);
253 static void unmap_stage2_puds(struct kvm *kvm, pgd_t *pgd,
254 phys_addr_t addr, phys_addr_t end)
256 phys_addr_t next, start_addr = addr;
257 pud_t *pud, *start_pud;
259 start_pud = pud = stage2_pud_offset(pgd, addr);
261 next = stage2_pud_addr_end(addr, end);
262 if (!stage2_pud_none(*pud)) {
263 if (stage2_pud_huge(*pud)) {
264 pud_t old_pud = *pud;
266 stage2_pud_clear(pud);
267 kvm_tlb_flush_vmid_ipa(kvm, addr);
268 kvm_flush_dcache_pud(old_pud);
269 put_page(virt_to_page(pud));
271 unmap_stage2_pmds(kvm, pud, addr, next);
274 } while (pud++, addr = next, addr != end);
276 if (stage2_pud_table_empty(start_pud))
277 clear_stage2_pgd_entry(kvm, pgd, start_addr);
281 * unmap_stage2_range -- Clear stage2 page table entries to unmap a range
282 * @kvm: The VM pointer
283 * @start: The intermediate physical base address of the range to unmap
284 * @size: The size of the area to unmap
286 * Clear a range of stage-2 mappings, lowering the various ref-counts. Must
287 * be called while holding mmu_lock (unless for freeing the stage2 pgd before
288 * destroying the VM), otherwise another faulting VCPU may come in and mess
289 * with things behind our backs.
291 static void unmap_stage2_range(struct kvm *kvm, phys_addr_t start, u64 size)
294 phys_addr_t addr = start, end = start + size;
297 assert_spin_locked(&kvm->mmu_lock);
298 pgd = kvm->arch.pgd + stage2_pgd_index(addr);
301 * Make sure the page table is still active, as another thread
302 * could have possibly freed the page table, while we released
305 if (!READ_ONCE(kvm->arch.pgd))
307 next = stage2_pgd_addr_end(addr, end);
308 if (!stage2_pgd_none(*pgd))
309 unmap_stage2_puds(kvm, pgd, addr, next);
311 * If the range is too large, release the kvm->mmu_lock
312 * to prevent starvation and lockup detector warnings.
315 cond_resched_lock(&kvm->mmu_lock);
316 } while (pgd++, addr = next, addr != end);
319 static void stage2_flush_ptes(struct kvm *kvm, pmd_t *pmd,
320 phys_addr_t addr, phys_addr_t end)
324 pte = pte_offset_kernel(pmd, addr);
326 if (!pte_none(*pte) && !kvm_is_device_pfn(pte_pfn(*pte)))
327 kvm_flush_dcache_pte(*pte);
328 } while (pte++, addr += PAGE_SIZE, addr != end);
331 static void stage2_flush_pmds(struct kvm *kvm, pud_t *pud,
332 phys_addr_t addr, phys_addr_t end)
337 pmd = stage2_pmd_offset(pud, addr);
339 next = stage2_pmd_addr_end(addr, end);
340 if (!pmd_none(*pmd)) {
341 if (pmd_thp_or_huge(*pmd))
342 kvm_flush_dcache_pmd(*pmd);
344 stage2_flush_ptes(kvm, pmd, addr, next);
346 } while (pmd++, addr = next, addr != end);
349 static void stage2_flush_puds(struct kvm *kvm, pgd_t *pgd,
350 phys_addr_t addr, phys_addr_t end)
355 pud = stage2_pud_offset(pgd, addr);
357 next = stage2_pud_addr_end(addr, end);
358 if (!stage2_pud_none(*pud)) {
359 if (stage2_pud_huge(*pud))
360 kvm_flush_dcache_pud(*pud);
362 stage2_flush_pmds(kvm, pud, addr, next);
364 } while (pud++, addr = next, addr != end);
367 static void stage2_flush_memslot(struct kvm *kvm,
368 struct kvm_memory_slot *memslot)
370 phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
371 phys_addr_t end = addr + PAGE_SIZE * memslot->npages;
375 pgd = kvm->arch.pgd + stage2_pgd_index(addr);
377 next = stage2_pgd_addr_end(addr, end);
378 stage2_flush_puds(kvm, pgd, addr, next);
379 } while (pgd++, addr = next, addr != end);
383 * stage2_flush_vm - Invalidate cache for pages mapped in stage 2
384 * @kvm: The struct kvm pointer
386 * Go through the stage 2 page tables and invalidate any cache lines
387 * backing memory already mapped to the VM.
389 static void stage2_flush_vm(struct kvm *kvm)
391 struct kvm_memslots *slots;
392 struct kvm_memory_slot *memslot;
395 idx = srcu_read_lock(&kvm->srcu);
396 spin_lock(&kvm->mmu_lock);
398 slots = kvm_memslots(kvm);
399 kvm_for_each_memslot(memslot, slots)
400 stage2_flush_memslot(kvm, memslot);
402 spin_unlock(&kvm->mmu_lock);
403 srcu_read_unlock(&kvm->srcu, idx);
406 static void clear_hyp_pgd_entry(pgd_t *pgd)
408 pud_t *pud_table __maybe_unused = pud_offset(pgd, 0UL);
410 pud_free(NULL, pud_table);
411 put_page(virt_to_page(pgd));
414 static void clear_hyp_pud_entry(pud_t *pud)
416 pmd_t *pmd_table __maybe_unused = pmd_offset(pud, 0);
417 VM_BUG_ON(pud_huge(*pud));
419 pmd_free(NULL, pmd_table);
420 put_page(virt_to_page(pud));
423 static void clear_hyp_pmd_entry(pmd_t *pmd)
425 pte_t *pte_table = pte_offset_kernel(pmd, 0);
426 VM_BUG_ON(pmd_thp_or_huge(*pmd));
428 pte_free_kernel(NULL, pte_table);
429 put_page(virt_to_page(pmd));
432 static void unmap_hyp_ptes(pmd_t *pmd, phys_addr_t addr, phys_addr_t end)
434 pte_t *pte, *start_pte;
436 start_pte = pte = pte_offset_kernel(pmd, addr);
438 if (!pte_none(*pte)) {
439 kvm_set_pte(pte, __pte(0));
440 put_page(virt_to_page(pte));
442 } while (pte++, addr += PAGE_SIZE, addr != end);
444 if (hyp_pte_table_empty(start_pte))
445 clear_hyp_pmd_entry(pmd);
448 static void unmap_hyp_pmds(pud_t *pud, phys_addr_t addr, phys_addr_t end)
451 pmd_t *pmd, *start_pmd;
453 start_pmd = pmd = pmd_offset(pud, addr);
455 next = pmd_addr_end(addr, end);
456 /* Hyp doesn't use huge pmds */
458 unmap_hyp_ptes(pmd, addr, next);
459 } while (pmd++, addr = next, addr != end);
461 if (hyp_pmd_table_empty(start_pmd))
462 clear_hyp_pud_entry(pud);
465 static void unmap_hyp_puds(pgd_t *pgd, phys_addr_t addr, phys_addr_t end)
468 pud_t *pud, *start_pud;
470 start_pud = pud = pud_offset(pgd, addr);
472 next = pud_addr_end(addr, end);
473 /* Hyp doesn't use huge puds */
475 unmap_hyp_pmds(pud, addr, next);
476 } while (pud++, addr = next, addr != end);
478 if (hyp_pud_table_empty(start_pud))
479 clear_hyp_pgd_entry(pgd);
482 static void unmap_hyp_range(pgd_t *pgdp, phys_addr_t start, u64 size)
485 phys_addr_t addr = start, end = start + size;
489 * We don't unmap anything from HYP, except at the hyp tear down.
490 * Hence, we don't have to invalidate the TLBs here.
492 pgd = pgdp + pgd_index(addr);
494 next = pgd_addr_end(addr, end);
496 unmap_hyp_puds(pgd, addr, next);
497 } while (pgd++, addr = next, addr != end);
501 * free_hyp_pgds - free Hyp-mode page tables
503 * Assumes hyp_pgd is a page table used strictly in Hyp-mode and
504 * therefore contains either mappings in the kernel memory area (above
505 * PAGE_OFFSET), or device mappings in the vmalloc range (from
506 * VMALLOC_START to VMALLOC_END).
508 * boot_hyp_pgd should only map two pages for the init code.
510 void free_hyp_pgds(void)
512 mutex_lock(&kvm_hyp_pgd_mutex);
515 unmap_hyp_range(boot_hyp_pgd, hyp_idmap_start, PAGE_SIZE);
516 free_pages((unsigned long)boot_hyp_pgd, hyp_pgd_order);
521 unmap_hyp_range(hyp_pgd, hyp_idmap_start, PAGE_SIZE);
522 unmap_hyp_range(hyp_pgd, kern_hyp_va(PAGE_OFFSET),
523 (uintptr_t)high_memory - PAGE_OFFSET);
524 unmap_hyp_range(hyp_pgd, kern_hyp_va(VMALLOC_START),
525 VMALLOC_END - VMALLOC_START);
527 free_pages((unsigned long)hyp_pgd, hyp_pgd_order);
530 if (merged_hyp_pgd) {
531 clear_page(merged_hyp_pgd);
532 free_page((unsigned long)merged_hyp_pgd);
533 merged_hyp_pgd = NULL;
536 mutex_unlock(&kvm_hyp_pgd_mutex);
539 static void create_hyp_pte_mappings(pmd_t *pmd, unsigned long start,
540 unsigned long end, unsigned long pfn,
548 pte = pte_offset_kernel(pmd, addr);
549 kvm_set_pte(pte, pfn_pte(pfn, prot));
550 get_page(virt_to_page(pte));
551 kvm_flush_dcache_to_poc(pte, sizeof(*pte));
553 } while (addr += PAGE_SIZE, addr != end);
556 static int create_hyp_pmd_mappings(pud_t *pud, unsigned long start,
557 unsigned long end, unsigned long pfn,
562 unsigned long addr, next;
566 pmd = pmd_offset(pud, addr);
568 BUG_ON(pmd_sect(*pmd));
570 if (pmd_none(*pmd)) {
571 pte = pte_alloc_one_kernel(NULL, addr);
573 kvm_err("Cannot allocate Hyp pte\n");
576 pmd_populate_kernel(NULL, pmd, pte);
577 get_page(virt_to_page(pmd));
578 kvm_flush_dcache_to_poc(pmd, sizeof(*pmd));
581 next = pmd_addr_end(addr, end);
583 create_hyp_pte_mappings(pmd, addr, next, pfn, prot);
584 pfn += (next - addr) >> PAGE_SHIFT;
585 } while (addr = next, addr != end);
590 static int create_hyp_pud_mappings(pgd_t *pgd, unsigned long start,
591 unsigned long end, unsigned long pfn,
596 unsigned long addr, next;
601 pud = pud_offset(pgd, addr);
603 if (pud_none_or_clear_bad(pud)) {
604 pmd = pmd_alloc_one(NULL, addr);
606 kvm_err("Cannot allocate Hyp pmd\n");
609 pud_populate(NULL, pud, pmd);
610 get_page(virt_to_page(pud));
611 kvm_flush_dcache_to_poc(pud, sizeof(*pud));
614 next = pud_addr_end(addr, end);
615 ret = create_hyp_pmd_mappings(pud, addr, next, pfn, prot);
618 pfn += (next - addr) >> PAGE_SHIFT;
619 } while (addr = next, addr != end);
624 static int __create_hyp_mappings(pgd_t *pgdp, unsigned long ptrs_per_pgd,
625 unsigned long start, unsigned long end,
626 unsigned long pfn, pgprot_t prot)
630 unsigned long addr, next;
633 mutex_lock(&kvm_hyp_pgd_mutex);
634 addr = start & PAGE_MASK;
635 end = PAGE_ALIGN(end);
637 pgd = pgdp + ((addr >> PGDIR_SHIFT) & (ptrs_per_pgd - 1));
639 if (pgd_none(*pgd)) {
640 pud = pud_alloc_one(NULL, addr);
642 kvm_err("Cannot allocate Hyp pud\n");
646 pgd_populate(NULL, pgd, pud);
647 get_page(virt_to_page(pgd));
648 kvm_flush_dcache_to_poc(pgd, sizeof(*pgd));
651 next = pgd_addr_end(addr, end);
652 err = create_hyp_pud_mappings(pgd, addr, next, pfn, prot);
655 pfn += (next - addr) >> PAGE_SHIFT;
656 } while (addr = next, addr != end);
658 mutex_unlock(&kvm_hyp_pgd_mutex);
662 static phys_addr_t kvm_kaddr_to_phys(void *kaddr)
664 if (!is_vmalloc_addr(kaddr)) {
665 BUG_ON(!virt_addr_valid(kaddr));
668 return page_to_phys(vmalloc_to_page(kaddr)) +
669 offset_in_page(kaddr);
674 * create_hyp_mappings - duplicate a kernel virtual address range in Hyp mode
675 * @from: The virtual kernel start address of the range
676 * @to: The virtual kernel end address of the range (exclusive)
677 * @prot: The protection to be applied to this range
679 * The same virtual address as the kernel virtual address is also used
680 * in Hyp-mode mapping (modulo HYP_PAGE_OFFSET) to the same underlying
683 int create_hyp_mappings(void *from, void *to, pgprot_t prot)
685 phys_addr_t phys_addr;
686 unsigned long virt_addr;
687 unsigned long start = kern_hyp_va((unsigned long)from);
688 unsigned long end = kern_hyp_va((unsigned long)to);
690 if (is_kernel_in_hyp_mode())
693 start = start & PAGE_MASK;
694 end = PAGE_ALIGN(end);
696 for (virt_addr = start; virt_addr < end; virt_addr += PAGE_SIZE) {
699 phys_addr = kvm_kaddr_to_phys(from + virt_addr - start);
700 err = __create_hyp_mappings(hyp_pgd, PTRS_PER_PGD,
701 virt_addr, virt_addr + PAGE_SIZE,
702 __phys_to_pfn(phys_addr),
712 * create_hyp_io_mappings - Map IO into both kernel and HYP
713 * @phys_addr: The physical start address which gets mapped
714 * @size: Size of the region being mapped
715 * @kaddr: Kernel VA for this mapping
716 * @haddr: HYP VA for this mapping
718 int create_hyp_io_mappings(phys_addr_t phys_addr, size_t size,
719 void __iomem **kaddr,
720 void __iomem **haddr)
722 unsigned long start, end;
725 *kaddr = ioremap(phys_addr, size);
729 if (is_kernel_in_hyp_mode()) {
735 start = kern_hyp_va((unsigned long)*kaddr);
736 end = kern_hyp_va((unsigned long)*kaddr + size);
737 ret = __create_hyp_mappings(hyp_pgd, PTRS_PER_PGD, start, end,
738 __phys_to_pfn(phys_addr), PAGE_HYP_DEVICE);
746 *haddr = (void __iomem *)start;
751 * kvm_alloc_stage2_pgd - allocate level-1 table for stage-2 translation.
752 * @kvm: The KVM struct pointer for the VM.
754 * Allocates only the stage-2 HW PGD level table(s) (can support either full
755 * 40-bit input addresses or limited to 32-bit input addresses). Clears the
758 * Note we don't need locking here as this is only called when the VM is
759 * created, which can only be done once.
761 int kvm_alloc_stage2_pgd(struct kvm *kvm)
765 if (kvm->arch.pgd != NULL) {
766 kvm_err("kvm_arch already initialized?\n");
770 /* Allocate the HW PGD, making sure that each page gets its own refcount */
771 pgd = alloc_pages_exact(S2_PGD_SIZE, GFP_KERNEL | __GFP_ZERO);
779 static void stage2_unmap_memslot(struct kvm *kvm,
780 struct kvm_memory_slot *memslot)
782 hva_t hva = memslot->userspace_addr;
783 phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
784 phys_addr_t size = PAGE_SIZE * memslot->npages;
785 hva_t reg_end = hva + size;
788 * A memory region could potentially cover multiple VMAs, and any holes
789 * between them, so iterate over all of them to find out if we should
792 * +--------------------------------------------+
793 * +---------------+----------------+ +----------------+
794 * | : VMA 1 | VMA 2 | | VMA 3 : |
795 * +---------------+----------------+ +----------------+
797 * +--------------------------------------------+
800 struct vm_area_struct *vma = find_vma(current->mm, hva);
801 hva_t vm_start, vm_end;
803 if (!vma || vma->vm_start >= reg_end)
807 * Take the intersection of this VMA with the memory region
809 vm_start = max(hva, vma->vm_start);
810 vm_end = min(reg_end, vma->vm_end);
812 if (!(vma->vm_flags & VM_PFNMAP)) {
813 gpa_t gpa = addr + (vm_start - memslot->userspace_addr);
814 unmap_stage2_range(kvm, gpa, vm_end - vm_start);
817 } while (hva < reg_end);
821 * stage2_unmap_vm - Unmap Stage-2 RAM mappings
822 * @kvm: The struct kvm pointer
824 * Go through the memregions and unmap any reguler RAM
825 * backing memory already mapped to the VM.
827 void stage2_unmap_vm(struct kvm *kvm)
829 struct kvm_memslots *slots;
830 struct kvm_memory_slot *memslot;
833 idx = srcu_read_lock(&kvm->srcu);
834 down_read(¤t->mm->mmap_sem);
835 spin_lock(&kvm->mmu_lock);
837 slots = kvm_memslots(kvm);
838 kvm_for_each_memslot(memslot, slots)
839 stage2_unmap_memslot(kvm, memslot);
841 spin_unlock(&kvm->mmu_lock);
842 up_read(¤t->mm->mmap_sem);
843 srcu_read_unlock(&kvm->srcu, idx);
847 * kvm_free_stage2_pgd - free all stage-2 tables
848 * @kvm: The KVM struct pointer for the VM.
850 * Walks the level-1 page table pointed to by kvm->arch.pgd and frees all
851 * underlying level-2 and level-3 tables before freeing the actual level-1 table
852 * and setting the struct pointer to NULL.
854 void kvm_free_stage2_pgd(struct kvm *kvm)
858 spin_lock(&kvm->mmu_lock);
860 unmap_stage2_range(kvm, 0, KVM_PHYS_SIZE);
861 pgd = READ_ONCE(kvm->arch.pgd);
862 kvm->arch.pgd = NULL;
864 spin_unlock(&kvm->mmu_lock);
866 /* Free the HW pgd, one page at a time */
868 free_pages_exact(pgd, S2_PGD_SIZE);
871 static pud_t *stage2_get_pud(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
877 pgd = kvm->arch.pgd + stage2_pgd_index(addr);
878 if (WARN_ON(stage2_pgd_none(*pgd))) {
881 pud = mmu_memory_cache_alloc(cache);
882 stage2_pgd_populate(pgd, pud);
883 get_page(virt_to_page(pgd));
886 return stage2_pud_offset(pgd, addr);
889 static pmd_t *stage2_get_pmd(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
895 pud = stage2_get_pud(kvm, cache, addr);
899 if (stage2_pud_none(*pud)) {
902 pmd = mmu_memory_cache_alloc(cache);
903 stage2_pud_populate(pud, pmd);
904 get_page(virt_to_page(pud));
907 return stage2_pmd_offset(pud, addr);
910 static int stage2_set_pmd_huge(struct kvm *kvm, struct kvm_mmu_memory_cache
911 *cache, phys_addr_t addr, const pmd_t *new_pmd)
915 pmd = stage2_get_pmd(kvm, cache, addr);
919 * Mapping in huge pages should only happen through a fault. If a
920 * page is merged into a transparent huge page, the individual
921 * subpages of that huge page should be unmapped through MMU
922 * notifiers before we get here.
924 * Merging of CompoundPages is not supported; they should become
925 * splitting first, unmapped, merged, and mapped back in on-demand.
927 VM_BUG_ON(pmd_present(*pmd) && pmd_pfn(*pmd) != pmd_pfn(*new_pmd));
930 if (pmd_present(old_pmd)) {
932 kvm_tlb_flush_vmid_ipa(kvm, addr);
934 get_page(virt_to_page(pmd));
937 kvm_set_pmd(pmd, *new_pmd);
941 static bool stage2_is_exec(struct kvm *kvm, phys_addr_t addr)
946 pmdp = stage2_get_pmd(kvm, NULL, addr);
947 if (!pmdp || pmd_none(*pmdp) || !pmd_present(*pmdp))
950 if (pmd_thp_or_huge(*pmdp))
951 return kvm_s2pmd_exec(pmdp);
953 ptep = pte_offset_kernel(pmdp, addr);
954 if (!ptep || pte_none(*ptep) || !pte_present(*ptep))
957 return kvm_s2pte_exec(ptep);
960 static int stage2_set_pte(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
961 phys_addr_t addr, const pte_t *new_pte,
966 bool iomap = flags & KVM_S2PTE_FLAG_IS_IOMAP;
967 bool logging_active = flags & KVM_S2_FLAG_LOGGING_ACTIVE;
969 VM_BUG_ON(logging_active && !cache);
971 /* Create stage-2 page table mapping - Levels 0 and 1 */
972 pmd = stage2_get_pmd(kvm, cache, addr);
975 * Ignore calls from kvm_set_spte_hva for unallocated
982 * While dirty page logging - dissolve huge PMD, then continue on to
986 stage2_dissolve_pmd(kvm, addr, pmd);
988 /* Create stage-2 page mappings - Level 2 */
989 if (pmd_none(*pmd)) {
991 return 0; /* ignore calls from kvm_set_spte_hva */
992 pte = mmu_memory_cache_alloc(cache);
993 pmd_populate_kernel(NULL, pmd, pte);
994 get_page(virt_to_page(pmd));
997 pte = pte_offset_kernel(pmd, addr);
999 if (iomap && pte_present(*pte))
1002 /* Create 2nd stage page table mapping - Level 3 */
1004 if (pte_present(old_pte)) {
1005 kvm_set_pte(pte, __pte(0));
1006 kvm_tlb_flush_vmid_ipa(kvm, addr);
1008 get_page(virt_to_page(pte));
1011 kvm_set_pte(pte, *new_pte);
1015 #ifndef __HAVE_ARCH_PTEP_TEST_AND_CLEAR_YOUNG
1016 static int stage2_ptep_test_and_clear_young(pte_t *pte)
1018 if (pte_young(*pte)) {
1019 *pte = pte_mkold(*pte);
1025 static int stage2_ptep_test_and_clear_young(pte_t *pte)
1027 return __ptep_test_and_clear_young(pte);
1031 static int stage2_pmdp_test_and_clear_young(pmd_t *pmd)
1033 return stage2_ptep_test_and_clear_young((pte_t *)pmd);
1037 * kvm_phys_addr_ioremap - map a device range to guest IPA
1039 * @kvm: The KVM pointer
1040 * @guest_ipa: The IPA at which to insert the mapping
1041 * @pa: The physical address of the device
1042 * @size: The size of the mapping
1044 int kvm_phys_addr_ioremap(struct kvm *kvm, phys_addr_t guest_ipa,
1045 phys_addr_t pa, unsigned long size, bool writable)
1047 phys_addr_t addr, end;
1050 struct kvm_mmu_memory_cache cache = { 0, };
1052 end = (guest_ipa + size + PAGE_SIZE - 1) & PAGE_MASK;
1053 pfn = __phys_to_pfn(pa);
1055 for (addr = guest_ipa; addr < end; addr += PAGE_SIZE) {
1056 pte_t pte = pfn_pte(pfn, PAGE_S2_DEVICE);
1059 pte = kvm_s2pte_mkwrite(pte);
1061 ret = mmu_topup_memory_cache(&cache, KVM_MMU_CACHE_MIN_PAGES,
1065 spin_lock(&kvm->mmu_lock);
1066 ret = stage2_set_pte(kvm, &cache, addr, &pte,
1067 KVM_S2PTE_FLAG_IS_IOMAP);
1068 spin_unlock(&kvm->mmu_lock);
1076 mmu_free_memory_cache(&cache);
1080 static bool transparent_hugepage_adjust(kvm_pfn_t *pfnp, phys_addr_t *ipap)
1082 kvm_pfn_t pfn = *pfnp;
1083 gfn_t gfn = *ipap >> PAGE_SHIFT;
1085 if (PageTransCompoundMap(pfn_to_page(pfn))) {
1088 * The address we faulted on is backed by a transparent huge
1089 * page. However, because we map the compound huge page and
1090 * not the individual tail page, we need to transfer the
1091 * refcount to the head page. We have to be careful that the
1092 * THP doesn't start to split while we are adjusting the
1095 * We are sure this doesn't happen, because mmu_notifier_retry
1096 * was successful and we are holding the mmu_lock, so if this
1097 * THP is trying to split, it will be blocked in the mmu
1098 * notifier before touching any of the pages, specifically
1099 * before being able to call __split_huge_page_refcount().
1101 * We can therefore safely transfer the refcount from PG_tail
1102 * to PG_head and switch the pfn from a tail page to the head
1105 mask = PTRS_PER_PMD - 1;
1106 VM_BUG_ON((gfn & mask) != (pfn & mask));
1109 kvm_release_pfn_clean(pfn);
1121 static bool kvm_is_write_fault(struct kvm_vcpu *vcpu)
1123 if (kvm_vcpu_trap_is_iabt(vcpu))
1126 return kvm_vcpu_dabt_iswrite(vcpu);
1130 * stage2_wp_ptes - write protect PMD range
1131 * @pmd: pointer to pmd entry
1132 * @addr: range start address
1133 * @end: range end address
1135 static void stage2_wp_ptes(pmd_t *pmd, phys_addr_t addr, phys_addr_t end)
1139 pte = pte_offset_kernel(pmd, addr);
1141 if (!pte_none(*pte)) {
1142 if (!kvm_s2pte_readonly(pte))
1143 kvm_set_s2pte_readonly(pte);
1145 } while (pte++, addr += PAGE_SIZE, addr != end);
1149 * stage2_wp_pmds - write protect PUD range
1150 * @pud: pointer to pud entry
1151 * @addr: range start address
1152 * @end: range end address
1154 static void stage2_wp_pmds(pud_t *pud, phys_addr_t addr, phys_addr_t end)
1159 pmd = stage2_pmd_offset(pud, addr);
1162 next = stage2_pmd_addr_end(addr, end);
1163 if (!pmd_none(*pmd)) {
1164 if (pmd_thp_or_huge(*pmd)) {
1165 if (!kvm_s2pmd_readonly(pmd))
1166 kvm_set_s2pmd_readonly(pmd);
1168 stage2_wp_ptes(pmd, addr, next);
1171 } while (pmd++, addr = next, addr != end);
1175 * stage2_wp_puds - write protect PGD range
1176 * @pgd: pointer to pgd entry
1177 * @addr: range start address
1178 * @end: range end address
1180 * Process PUD entries, for a huge PUD we cause a panic.
1182 static void stage2_wp_puds(pgd_t *pgd, phys_addr_t addr, phys_addr_t end)
1187 pud = stage2_pud_offset(pgd, addr);
1189 next = stage2_pud_addr_end(addr, end);
1190 if (!stage2_pud_none(*pud)) {
1191 /* TODO:PUD not supported, revisit later if supported */
1192 BUG_ON(stage2_pud_huge(*pud));
1193 stage2_wp_pmds(pud, addr, next);
1195 } while (pud++, addr = next, addr != end);
1199 * stage2_wp_range() - write protect stage2 memory region range
1200 * @kvm: The KVM pointer
1201 * @addr: Start address of range
1202 * @end: End address of range
1204 static void stage2_wp_range(struct kvm *kvm, phys_addr_t addr, phys_addr_t end)
1209 pgd = kvm->arch.pgd + stage2_pgd_index(addr);
1212 * Release kvm_mmu_lock periodically if the memory region is
1213 * large. Otherwise, we may see kernel panics with
1214 * CONFIG_DETECT_HUNG_TASK, CONFIG_LOCKUP_DETECTOR,
1215 * CONFIG_LOCKDEP. Additionally, holding the lock too long
1216 * will also starve other vCPUs. We have to also make sure
1217 * that the page tables are not freed while we released
1220 cond_resched_lock(&kvm->mmu_lock);
1221 if (!READ_ONCE(kvm->arch.pgd))
1223 next = stage2_pgd_addr_end(addr, end);
1224 if (stage2_pgd_present(*pgd))
1225 stage2_wp_puds(pgd, addr, next);
1226 } while (pgd++, addr = next, addr != end);
1230 * kvm_mmu_wp_memory_region() - write protect stage 2 entries for memory slot
1231 * @kvm: The KVM pointer
1232 * @slot: The memory slot to write protect
1234 * Called to start logging dirty pages after memory region
1235 * KVM_MEM_LOG_DIRTY_PAGES operation is called. After this function returns
1236 * all present PMD and PTEs are write protected in the memory region.
1237 * Afterwards read of dirty page log can be called.
1239 * Acquires kvm_mmu_lock. Called with kvm->slots_lock mutex acquired,
1240 * serializing operations for VM memory regions.
1242 void kvm_mmu_wp_memory_region(struct kvm *kvm, int slot)
1244 struct kvm_memslots *slots = kvm_memslots(kvm);
1245 struct kvm_memory_slot *memslot = id_to_memslot(slots, slot);
1246 phys_addr_t start = memslot->base_gfn << PAGE_SHIFT;
1247 phys_addr_t end = (memslot->base_gfn + memslot->npages) << PAGE_SHIFT;
1249 spin_lock(&kvm->mmu_lock);
1250 stage2_wp_range(kvm, start, end);
1251 spin_unlock(&kvm->mmu_lock);
1252 kvm_flush_remote_tlbs(kvm);
1256 * kvm_mmu_write_protect_pt_masked() - write protect dirty pages
1257 * @kvm: The KVM pointer
1258 * @slot: The memory slot associated with mask
1259 * @gfn_offset: The gfn offset in memory slot
1260 * @mask: The mask of dirty pages at offset 'gfn_offset' in this memory
1261 * slot to be write protected
1263 * Walks bits set in mask write protects the associated pte's. Caller must
1264 * acquire kvm_mmu_lock.
1266 static void kvm_mmu_write_protect_pt_masked(struct kvm *kvm,
1267 struct kvm_memory_slot *slot,
1268 gfn_t gfn_offset, unsigned long mask)
1270 phys_addr_t base_gfn = slot->base_gfn + gfn_offset;
1271 phys_addr_t start = (base_gfn + __ffs(mask)) << PAGE_SHIFT;
1272 phys_addr_t end = (base_gfn + __fls(mask) + 1) << PAGE_SHIFT;
1274 stage2_wp_range(kvm, start, end);
1278 * kvm_arch_mmu_enable_log_dirty_pt_masked - enable dirty logging for selected
1281 * It calls kvm_mmu_write_protect_pt_masked to write protect selected pages to
1282 * enable dirty logging for them.
1284 void kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm *kvm,
1285 struct kvm_memory_slot *slot,
1286 gfn_t gfn_offset, unsigned long mask)
1288 kvm_mmu_write_protect_pt_masked(kvm, slot, gfn_offset, mask);
1291 static void clean_dcache_guest_page(kvm_pfn_t pfn, unsigned long size)
1293 __clean_dcache_guest_page(pfn, size);
1296 static void invalidate_icache_guest_page(kvm_pfn_t pfn, unsigned long size)
1298 __invalidate_icache_guest_page(pfn, size);
1301 static void kvm_send_hwpoison_signal(unsigned long address,
1302 struct vm_area_struct *vma)
1306 info.si_signo = SIGBUS;
1308 info.si_code = BUS_MCEERR_AR;
1309 info.si_addr = (void __user *)address;
1311 if (is_vm_hugetlb_page(vma))
1312 info.si_addr_lsb = huge_page_shift(hstate_vma(vma));
1314 info.si_addr_lsb = PAGE_SHIFT;
1316 send_sig_info(SIGBUS, &info, current);
1319 static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
1320 struct kvm_memory_slot *memslot, unsigned long hva,
1321 unsigned long fault_status)
1324 bool write_fault, exec_fault, writable, hugetlb = false, force_pte = false;
1325 unsigned long mmu_seq;
1326 gfn_t gfn = fault_ipa >> PAGE_SHIFT;
1327 struct kvm *kvm = vcpu->kvm;
1328 struct kvm_mmu_memory_cache *memcache = &vcpu->arch.mmu_page_cache;
1329 struct vm_area_struct *vma;
1331 pgprot_t mem_type = PAGE_S2;
1332 bool logging_active = memslot_is_logging(memslot);
1333 unsigned long flags = 0;
1335 write_fault = kvm_is_write_fault(vcpu);
1336 exec_fault = kvm_vcpu_trap_is_iabt(vcpu);
1337 VM_BUG_ON(write_fault && exec_fault);
1339 if (fault_status == FSC_PERM && !write_fault && !exec_fault) {
1340 kvm_err("Unexpected L2 read permission error\n");
1344 /* Let's check if we will get back a huge page backed by hugetlbfs */
1345 down_read(¤t->mm->mmap_sem);
1346 vma = find_vma_intersection(current->mm, hva, hva + 1);
1347 if (unlikely(!vma)) {
1348 kvm_err("Failed to find VMA for hva 0x%lx\n", hva);
1349 up_read(¤t->mm->mmap_sem);
1353 if (vma_kernel_pagesize(vma) == PMD_SIZE && !logging_active) {
1355 gfn = (fault_ipa & PMD_MASK) >> PAGE_SHIFT;
1358 * Pages belonging to memslots that don't have the same
1359 * alignment for userspace and IPA cannot be mapped using
1360 * block descriptors even if the pages belong to a THP for
1361 * the process, because the stage-2 block descriptor will
1362 * cover more than a single THP and we loose atomicity for
1363 * unmapping, updates, and splits of the THP or other pages
1364 * in the stage-2 block range.
1366 if ((memslot->userspace_addr & ~PMD_MASK) !=
1367 ((memslot->base_gfn << PAGE_SHIFT) & ~PMD_MASK))
1370 up_read(¤t->mm->mmap_sem);
1372 /* We need minimum second+third level pages */
1373 ret = mmu_topup_memory_cache(memcache, KVM_MMU_CACHE_MIN_PAGES,
1378 mmu_seq = vcpu->kvm->mmu_notifier_seq;
1380 * Ensure the read of mmu_notifier_seq happens before we call
1381 * gfn_to_pfn_prot (which calls get_user_pages), so that we don't risk
1382 * the page we just got a reference to gets unmapped before we have a
1383 * chance to grab the mmu_lock, which ensure that if the page gets
1384 * unmapped afterwards, the call to kvm_unmap_hva will take it away
1385 * from us again properly. This smp_rmb() interacts with the smp_wmb()
1386 * in kvm_mmu_notifier_invalidate_<page|range_end>.
1390 pfn = gfn_to_pfn_prot(kvm, gfn, write_fault, &writable);
1391 if (pfn == KVM_PFN_ERR_HWPOISON) {
1392 kvm_send_hwpoison_signal(hva, vma);
1395 if (is_error_noslot_pfn(pfn))
1398 if (kvm_is_device_pfn(pfn)) {
1399 mem_type = PAGE_S2_DEVICE;
1400 flags |= KVM_S2PTE_FLAG_IS_IOMAP;
1401 } else if (logging_active) {
1403 * Faults on pages in a memslot with logging enabled
1404 * should not be mapped with huge pages (it introduces churn
1405 * and performance degradation), so force a pte mapping.
1408 flags |= KVM_S2_FLAG_LOGGING_ACTIVE;
1411 * Only actually map the page as writable if this was a write
1418 spin_lock(&kvm->mmu_lock);
1419 if (mmu_notifier_retry(kvm, mmu_seq))
1422 if (!hugetlb && !force_pte)
1423 hugetlb = transparent_hugepage_adjust(&pfn, &fault_ipa);
1426 pmd_t new_pmd = pfn_pmd(pfn, mem_type);
1427 new_pmd = pmd_mkhuge(new_pmd);
1429 new_pmd = kvm_s2pmd_mkwrite(new_pmd);
1430 kvm_set_pfn_dirty(pfn);
1433 if (fault_status != FSC_PERM)
1434 clean_dcache_guest_page(pfn, PMD_SIZE);
1437 new_pmd = kvm_s2pmd_mkexec(new_pmd);
1438 invalidate_icache_guest_page(pfn, PMD_SIZE);
1439 } else if (fault_status == FSC_PERM) {
1440 /* Preserve execute if XN was already cleared */
1441 if (stage2_is_exec(kvm, fault_ipa))
1442 new_pmd = kvm_s2pmd_mkexec(new_pmd);
1445 ret = stage2_set_pmd_huge(kvm, memcache, fault_ipa, &new_pmd);
1447 pte_t new_pte = pfn_pte(pfn, mem_type);
1450 new_pte = kvm_s2pte_mkwrite(new_pte);
1451 kvm_set_pfn_dirty(pfn);
1452 mark_page_dirty(kvm, gfn);
1455 if (fault_status != FSC_PERM)
1456 clean_dcache_guest_page(pfn, PAGE_SIZE);
1459 new_pte = kvm_s2pte_mkexec(new_pte);
1460 invalidate_icache_guest_page(pfn, PAGE_SIZE);
1461 } else if (fault_status == FSC_PERM) {
1462 /* Preserve execute if XN was already cleared */
1463 if (stage2_is_exec(kvm, fault_ipa))
1464 new_pte = kvm_s2pte_mkexec(new_pte);
1467 ret = stage2_set_pte(kvm, memcache, fault_ipa, &new_pte, flags);
1471 spin_unlock(&kvm->mmu_lock);
1472 kvm_set_pfn_accessed(pfn);
1473 kvm_release_pfn_clean(pfn);
1478 * Resolve the access fault by making the page young again.
1479 * Note that because the faulting entry is guaranteed not to be
1480 * cached in the TLB, we don't need to invalidate anything.
1481 * Only the HW Access Flag updates are supported for Stage 2 (no DBM),
1482 * so there is no need for atomic (pte|pmd)_mkyoung operations.
1484 static void handle_access_fault(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa)
1489 bool pfn_valid = false;
1491 trace_kvm_access_fault(fault_ipa);
1493 spin_lock(&vcpu->kvm->mmu_lock);
1495 pmd = stage2_get_pmd(vcpu->kvm, NULL, fault_ipa);
1496 if (!pmd || pmd_none(*pmd)) /* Nothing there */
1499 if (pmd_thp_or_huge(*pmd)) { /* THP, HugeTLB */
1500 *pmd = pmd_mkyoung(*pmd);
1501 pfn = pmd_pfn(*pmd);
1506 pte = pte_offset_kernel(pmd, fault_ipa);
1507 if (pte_none(*pte)) /* Nothing there either */
1510 *pte = pte_mkyoung(*pte); /* Just a page... */
1511 pfn = pte_pfn(*pte);
1514 spin_unlock(&vcpu->kvm->mmu_lock);
1516 kvm_set_pfn_accessed(pfn);
1520 * kvm_handle_guest_abort - handles all 2nd stage aborts
1521 * @vcpu: the VCPU pointer
1522 * @run: the kvm_run structure
1524 * Any abort that gets to the host is almost guaranteed to be caused by a
1525 * missing second stage translation table entry, which can mean that either the
1526 * guest simply needs more memory and we must allocate an appropriate page or it
1527 * can mean that the guest tried to access I/O memory, which is emulated by user
1528 * space. The distinction is based on the IPA causing the fault and whether this
1529 * memory region has been registered as standard RAM by user space.
1531 int kvm_handle_guest_abort(struct kvm_vcpu *vcpu, struct kvm_run *run)
1533 unsigned long fault_status;
1534 phys_addr_t fault_ipa;
1535 struct kvm_memory_slot *memslot;
1537 bool is_iabt, write_fault, writable;
1541 fault_status = kvm_vcpu_trap_get_fault_type(vcpu);
1543 fault_ipa = kvm_vcpu_get_fault_ipa(vcpu);
1544 is_iabt = kvm_vcpu_trap_is_iabt(vcpu);
1546 /* Synchronous External Abort? */
1547 if (kvm_vcpu_dabt_isextabt(vcpu)) {
1549 * For RAS the host kernel may handle this abort.
1550 * There is no need to pass the error into the guest.
1552 if (!handle_guest_sea(fault_ipa, kvm_vcpu_get_hsr(vcpu)))
1555 if (unlikely(!is_iabt)) {
1556 kvm_inject_vabt(vcpu);
1561 trace_kvm_guest_fault(*vcpu_pc(vcpu), kvm_vcpu_get_hsr(vcpu),
1562 kvm_vcpu_get_hfar(vcpu), fault_ipa);
1564 /* Check the stage-2 fault is trans. fault or write fault */
1565 if (fault_status != FSC_FAULT && fault_status != FSC_PERM &&
1566 fault_status != FSC_ACCESS) {
1567 kvm_err("Unsupported FSC: EC=%#x xFSC=%#lx ESR_EL2=%#lx\n",
1568 kvm_vcpu_trap_get_class(vcpu),
1569 (unsigned long)kvm_vcpu_trap_get_fault(vcpu),
1570 (unsigned long)kvm_vcpu_get_hsr(vcpu));
1574 idx = srcu_read_lock(&vcpu->kvm->srcu);
1576 gfn = fault_ipa >> PAGE_SHIFT;
1577 memslot = gfn_to_memslot(vcpu->kvm, gfn);
1578 hva = gfn_to_hva_memslot_prot(memslot, gfn, &writable);
1579 write_fault = kvm_is_write_fault(vcpu);
1580 if (kvm_is_error_hva(hva) || (write_fault && !writable)) {
1582 /* Prefetch Abort on I/O address */
1583 kvm_inject_pabt(vcpu, kvm_vcpu_get_hfar(vcpu));
1589 * Check for a cache maintenance operation. Since we
1590 * ended-up here, we know it is outside of any memory
1591 * slot. But we can't find out if that is for a device,
1592 * or if the guest is just being stupid. The only thing
1593 * we know for sure is that this range cannot be cached.
1595 * So let's assume that the guest is just being
1596 * cautious, and skip the instruction.
1598 if (kvm_vcpu_dabt_is_cm(vcpu)) {
1599 kvm_skip_instr(vcpu, kvm_vcpu_trap_il_is32bit(vcpu));
1605 * The IPA is reported as [MAX:12], so we need to
1606 * complement it with the bottom 12 bits from the
1607 * faulting VA. This is always 12 bits, irrespective
1610 fault_ipa |= kvm_vcpu_get_hfar(vcpu) & ((1 << 12) - 1);
1611 ret = io_mem_abort(vcpu, run, fault_ipa);
1615 /* Userspace should not be able to register out-of-bounds IPAs */
1616 VM_BUG_ON(fault_ipa >= KVM_PHYS_SIZE);
1618 if (fault_status == FSC_ACCESS) {
1619 handle_access_fault(vcpu, fault_ipa);
1624 ret = user_mem_abort(vcpu, fault_ipa, memslot, hva, fault_status);
1628 srcu_read_unlock(&vcpu->kvm->srcu, idx);
1632 static int handle_hva_to_gpa(struct kvm *kvm,
1633 unsigned long start,
1635 int (*handler)(struct kvm *kvm,
1636 gpa_t gpa, u64 size,
1640 struct kvm_memslots *slots;
1641 struct kvm_memory_slot *memslot;
1644 slots = kvm_memslots(kvm);
1646 /* we only care about the pages that the guest sees */
1647 kvm_for_each_memslot(memslot, slots) {
1648 unsigned long hva_start, hva_end;
1651 hva_start = max(start, memslot->userspace_addr);
1652 hva_end = min(end, memslot->userspace_addr +
1653 (memslot->npages << PAGE_SHIFT));
1654 if (hva_start >= hva_end)
1657 gpa = hva_to_gfn_memslot(hva_start, memslot) << PAGE_SHIFT;
1658 ret |= handler(kvm, gpa, (u64)(hva_end - hva_start), data);
1664 static int kvm_unmap_hva_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
1666 unmap_stage2_range(kvm, gpa, size);
1670 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
1672 unsigned long end = hva + PAGE_SIZE;
1677 trace_kvm_unmap_hva(hva);
1678 handle_hva_to_gpa(kvm, hva, end, &kvm_unmap_hva_handler, NULL);
1682 int kvm_unmap_hva_range(struct kvm *kvm,
1683 unsigned long start, unsigned long end)
1688 trace_kvm_unmap_hva_range(start, end);
1689 handle_hva_to_gpa(kvm, start, end, &kvm_unmap_hva_handler, NULL);
1693 static int kvm_set_spte_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
1695 pte_t *pte = (pte_t *)data;
1697 WARN_ON(size != PAGE_SIZE);
1699 * We can always call stage2_set_pte with KVM_S2PTE_FLAG_LOGGING_ACTIVE
1700 * flag clear because MMU notifiers will have unmapped a huge PMD before
1701 * calling ->change_pte() (which in turn calls kvm_set_spte_hva()) and
1702 * therefore stage2_set_pte() never needs to clear out a huge PMD
1703 * through this calling path.
1705 stage2_set_pte(kvm, NULL, gpa, pte, 0);
1710 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
1712 unsigned long end = hva + PAGE_SIZE;
1718 trace_kvm_set_spte_hva(hva);
1719 stage2_pte = pfn_pte(pte_pfn(pte), PAGE_S2);
1720 handle_hva_to_gpa(kvm, hva, end, &kvm_set_spte_handler, &stage2_pte);
1723 static int kvm_age_hva_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
1728 WARN_ON(size != PAGE_SIZE && size != PMD_SIZE);
1729 pmd = stage2_get_pmd(kvm, NULL, gpa);
1730 if (!pmd || pmd_none(*pmd)) /* Nothing there */
1733 if (pmd_thp_or_huge(*pmd)) /* THP, HugeTLB */
1734 return stage2_pmdp_test_and_clear_young(pmd);
1736 pte = pte_offset_kernel(pmd, gpa);
1740 return stage2_ptep_test_and_clear_young(pte);
1743 static int kvm_test_age_hva_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
1748 WARN_ON(size != PAGE_SIZE && size != PMD_SIZE);
1749 pmd = stage2_get_pmd(kvm, NULL, gpa);
1750 if (!pmd || pmd_none(*pmd)) /* Nothing there */
1753 if (pmd_thp_or_huge(*pmd)) /* THP, HugeTLB */
1754 return pmd_young(*pmd);
1756 pte = pte_offset_kernel(pmd, gpa);
1757 if (!pte_none(*pte)) /* Just a page... */
1758 return pte_young(*pte);
1763 int kvm_age_hva(struct kvm *kvm, unsigned long start, unsigned long end)
1767 trace_kvm_age_hva(start, end);
1768 return handle_hva_to_gpa(kvm, start, end, kvm_age_hva_handler, NULL);
1771 int kvm_test_age_hva(struct kvm *kvm, unsigned long hva)
1775 trace_kvm_test_age_hva(hva);
1776 return handle_hva_to_gpa(kvm, hva, hva, kvm_test_age_hva_handler, NULL);
1779 void kvm_mmu_free_memory_caches(struct kvm_vcpu *vcpu)
1781 mmu_free_memory_cache(&vcpu->arch.mmu_page_cache);
1784 phys_addr_t kvm_mmu_get_httbr(void)
1786 if (__kvm_cpu_uses_extended_idmap())
1787 return virt_to_phys(merged_hyp_pgd);
1789 return virt_to_phys(hyp_pgd);
1792 phys_addr_t kvm_get_idmap_vector(void)
1794 return hyp_idmap_vector;
1797 static int kvm_map_idmap_text(pgd_t *pgd)
1801 /* Create the idmap in the boot page tables */
1802 err = __create_hyp_mappings(pgd, __kvm_idmap_ptrs_per_pgd(),
1803 hyp_idmap_start, hyp_idmap_end,
1804 __phys_to_pfn(hyp_idmap_start),
1807 kvm_err("Failed to idmap %lx-%lx\n",
1808 hyp_idmap_start, hyp_idmap_end);
1813 int kvm_mmu_init(void)
1817 hyp_idmap_start = kvm_virt_to_phys(__hyp_idmap_text_start);
1818 hyp_idmap_end = kvm_virt_to_phys(__hyp_idmap_text_end);
1819 hyp_idmap_vector = kvm_virt_to_phys(__kvm_hyp_init);
1822 * We rely on the linker script to ensure at build time that the HYP
1823 * init code does not cross a page boundary.
1825 BUG_ON((hyp_idmap_start ^ (hyp_idmap_end - 1)) & PAGE_MASK);
1827 kvm_debug("IDMAP page: %lx\n", hyp_idmap_start);
1828 kvm_debug("HYP VA range: %lx:%lx\n",
1829 kern_hyp_va(PAGE_OFFSET),
1830 kern_hyp_va((unsigned long)high_memory - 1));
1832 if (hyp_idmap_start >= kern_hyp_va(PAGE_OFFSET) &&
1833 hyp_idmap_start < kern_hyp_va(~0UL) &&
1834 hyp_idmap_start != (unsigned long)__hyp_idmap_text_start) {
1836 * The idmap page is intersecting with the VA space,
1837 * it is not safe to continue further.
1839 kvm_err("IDMAP intersecting with HYP VA, unable to continue\n");
1844 hyp_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, hyp_pgd_order);
1846 kvm_err("Hyp mode PGD not allocated\n");
1851 if (__kvm_cpu_uses_extended_idmap()) {
1852 boot_hyp_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
1854 if (!boot_hyp_pgd) {
1855 kvm_err("Hyp boot PGD not allocated\n");
1860 err = kvm_map_idmap_text(boot_hyp_pgd);
1864 merged_hyp_pgd = (pgd_t *)__get_free_page(GFP_KERNEL | __GFP_ZERO);
1865 if (!merged_hyp_pgd) {
1866 kvm_err("Failed to allocate extra HYP pgd\n");
1869 __kvm_extend_hypmap(boot_hyp_pgd, hyp_pgd, merged_hyp_pgd,
1872 err = kvm_map_idmap_text(hyp_pgd);
1883 void kvm_arch_commit_memory_region(struct kvm *kvm,
1884 const struct kvm_userspace_memory_region *mem,
1885 const struct kvm_memory_slot *old,
1886 const struct kvm_memory_slot *new,
1887 enum kvm_mr_change change)
1890 * At this point memslot has been committed and there is an
1891 * allocated dirty_bitmap[], dirty pages will be be tracked while the
1892 * memory slot is write protected.
1894 if (change != KVM_MR_DELETE && mem->flags & KVM_MEM_LOG_DIRTY_PAGES)
1895 kvm_mmu_wp_memory_region(kvm, mem->slot);
1898 int kvm_arch_prepare_memory_region(struct kvm *kvm,
1899 struct kvm_memory_slot *memslot,
1900 const struct kvm_userspace_memory_region *mem,
1901 enum kvm_mr_change change)
1903 hva_t hva = mem->userspace_addr;
1904 hva_t reg_end = hva + mem->memory_size;
1905 bool writable = !(mem->flags & KVM_MEM_READONLY);
1908 if (change != KVM_MR_CREATE && change != KVM_MR_MOVE &&
1909 change != KVM_MR_FLAGS_ONLY)
1913 * Prevent userspace from creating a memory region outside of the IPA
1914 * space addressable by the KVM guest IPA space.
1916 if (memslot->base_gfn + memslot->npages >=
1917 (KVM_PHYS_SIZE >> PAGE_SHIFT))
1920 down_read(¤t->mm->mmap_sem);
1922 * A memory region could potentially cover multiple VMAs, and any holes
1923 * between them, so iterate over all of them to find out if we can map
1924 * any of them right now.
1926 * +--------------------------------------------+
1927 * +---------------+----------------+ +----------------+
1928 * | : VMA 1 | VMA 2 | | VMA 3 : |
1929 * +---------------+----------------+ +----------------+
1931 * +--------------------------------------------+
1934 struct vm_area_struct *vma = find_vma(current->mm, hva);
1935 hva_t vm_start, vm_end;
1937 if (!vma || vma->vm_start >= reg_end)
1941 * Mapping a read-only VMA is only allowed if the
1942 * memory region is configured as read-only.
1944 if (writable && !(vma->vm_flags & VM_WRITE)) {
1950 * Take the intersection of this VMA with the memory region
1952 vm_start = max(hva, vma->vm_start);
1953 vm_end = min(reg_end, vma->vm_end);
1955 if (vma->vm_flags & VM_PFNMAP) {
1956 gpa_t gpa = mem->guest_phys_addr +
1957 (vm_start - mem->userspace_addr);
1960 pa = (phys_addr_t)vma->vm_pgoff << PAGE_SHIFT;
1961 pa += vm_start - vma->vm_start;
1963 /* IO region dirty page logging not allowed */
1964 if (memslot->flags & KVM_MEM_LOG_DIRTY_PAGES) {
1969 ret = kvm_phys_addr_ioremap(kvm, gpa, pa,
1976 } while (hva < reg_end);
1978 if (change == KVM_MR_FLAGS_ONLY)
1981 spin_lock(&kvm->mmu_lock);
1983 unmap_stage2_range(kvm, mem->guest_phys_addr, mem->memory_size);
1985 stage2_flush_memslot(kvm, memslot);
1986 spin_unlock(&kvm->mmu_lock);
1988 up_read(¤t->mm->mmap_sem);
1992 void kvm_arch_free_memslot(struct kvm *kvm, struct kvm_memory_slot *free,
1993 struct kvm_memory_slot *dont)
1997 int kvm_arch_create_memslot(struct kvm *kvm, struct kvm_memory_slot *slot,
1998 unsigned long npages)
2003 void kvm_arch_memslots_updated(struct kvm *kvm, struct kvm_memslots *slots)
2007 void kvm_arch_flush_shadow_all(struct kvm *kvm)
2009 kvm_free_stage2_pgd(kvm);
2012 void kvm_arch_flush_shadow_memslot(struct kvm *kvm,
2013 struct kvm_memory_slot *slot)
2015 gpa_t gpa = slot->base_gfn << PAGE_SHIFT;
2016 phys_addr_t size = slot->npages << PAGE_SHIFT;
2018 spin_lock(&kvm->mmu_lock);
2019 unmap_stage2_range(kvm, gpa, size);
2020 spin_unlock(&kvm->mmu_lock);
2024 * See note at ARMv7 ARM B1.14.4 (TL;DR: S/W ops are not easily virtualized).
2027 * - S/W ops are local to a CPU (not broadcast)
2028 * - We have line migration behind our back (speculation)
2029 * - System caches don't support S/W at all (damn!)
2031 * In the face of the above, the best we can do is to try and convert
2032 * S/W ops to VA ops. Because the guest is not allowed to infer the
2033 * S/W to PA mapping, it can only use S/W to nuke the whole cache,
2034 * which is a rather good thing for us.
2036 * Also, it is only used when turning caches on/off ("The expected
2037 * usage of the cache maintenance instructions that operate by set/way
2038 * is associated with the cache maintenance instructions associated
2039 * with the powerdown and powerup of caches, if this is required by
2040 * the implementation.").
2042 * We use the following policy:
2044 * - If we trap a S/W operation, we enable VM trapping to detect
2045 * caches being turned on/off, and do a full clean.
2047 * - We flush the caches on both caches being turned on and off.
2049 * - Once the caches are enabled, we stop trapping VM ops.
2051 void kvm_set_way_flush(struct kvm_vcpu *vcpu)
2053 unsigned long hcr = *vcpu_hcr(vcpu);
2056 * If this is the first time we do a S/W operation
2057 * (i.e. HCR_TVM not set) flush the whole memory, and set the
2060 * Otherwise, rely on the VM trapping to wait for the MMU +
2061 * Caches to be turned off. At that point, we'll be able to
2062 * clean the caches again.
2064 if (!(hcr & HCR_TVM)) {
2065 trace_kvm_set_way_flush(*vcpu_pc(vcpu),
2066 vcpu_has_cache_enabled(vcpu));
2067 stage2_flush_vm(vcpu->kvm);
2068 *vcpu_hcr(vcpu) = hcr | HCR_TVM;
2072 void kvm_toggle_cache(struct kvm_vcpu *vcpu, bool was_enabled)
2074 bool now_enabled = vcpu_has_cache_enabled(vcpu);
2077 * If switching the MMU+caches on, need to invalidate the caches.
2078 * If switching it off, need to clean the caches.
2079 * Clean + invalidate does the trick always.
2081 if (now_enabled != was_enabled)
2082 stage2_flush_vm(vcpu->kvm);
2084 /* Caches are now on, stop trapping VM ops (until a S/W op) */
2086 *vcpu_hcr(vcpu) &= ~HCR_TVM;
2088 trace_kvm_toggle_cache(*vcpu_pc(vcpu), was_enabled, now_enabled);