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);
310 } while (pgd++, addr = next, addr != end);
313 static void stage2_flush_ptes(struct kvm *kvm, pmd_t *pmd,
314 phys_addr_t addr, phys_addr_t end)
318 pte = pte_offset_kernel(pmd, addr);
320 if (!pte_none(*pte) && !kvm_is_device_pfn(pte_pfn(*pte)))
321 kvm_flush_dcache_pte(*pte);
322 } while (pte++, addr += PAGE_SIZE, addr != end);
325 static void stage2_flush_pmds(struct kvm *kvm, pud_t *pud,
326 phys_addr_t addr, phys_addr_t end)
331 pmd = stage2_pmd_offset(pud, addr);
333 next = stage2_pmd_addr_end(addr, end);
334 if (!pmd_none(*pmd)) {
335 if (pmd_thp_or_huge(*pmd))
336 kvm_flush_dcache_pmd(*pmd);
338 stage2_flush_ptes(kvm, pmd, addr, next);
340 } while (pmd++, addr = next, addr != end);
343 static void stage2_flush_puds(struct kvm *kvm, pgd_t *pgd,
344 phys_addr_t addr, phys_addr_t end)
349 pud = stage2_pud_offset(pgd, addr);
351 next = stage2_pud_addr_end(addr, end);
352 if (!stage2_pud_none(*pud)) {
353 if (stage2_pud_huge(*pud))
354 kvm_flush_dcache_pud(*pud);
356 stage2_flush_pmds(kvm, pud, addr, next);
358 } while (pud++, addr = next, addr != end);
361 static void stage2_flush_memslot(struct kvm *kvm,
362 struct kvm_memory_slot *memslot)
364 phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
365 phys_addr_t end = addr + PAGE_SIZE * memslot->npages;
369 pgd = kvm->arch.pgd + stage2_pgd_index(addr);
371 next = stage2_pgd_addr_end(addr, end);
372 if (!stage2_pgd_none(*pgd))
373 stage2_flush_puds(kvm, pgd, addr, next);
374 } while (pgd++, addr = next, addr != end);
378 * stage2_flush_vm - Invalidate cache for pages mapped in stage 2
379 * @kvm: The struct kvm pointer
381 * Go through the stage 2 page tables and invalidate any cache lines
382 * backing memory already mapped to the VM.
384 static void stage2_flush_vm(struct kvm *kvm)
386 struct kvm_memslots *slots;
387 struct kvm_memory_slot *memslot;
390 idx = srcu_read_lock(&kvm->srcu);
391 spin_lock(&kvm->mmu_lock);
393 slots = kvm_memslots(kvm);
394 kvm_for_each_memslot(memslot, slots)
395 stage2_flush_memslot(kvm, memslot);
397 spin_unlock(&kvm->mmu_lock);
398 srcu_read_unlock(&kvm->srcu, idx);
401 static void clear_hyp_pgd_entry(pgd_t *pgd)
403 pud_t *pud_table __maybe_unused = pud_offset(pgd, 0UL);
405 pud_free(NULL, pud_table);
406 put_page(virt_to_page(pgd));
409 static void clear_hyp_pud_entry(pud_t *pud)
411 pmd_t *pmd_table __maybe_unused = pmd_offset(pud, 0);
412 VM_BUG_ON(pud_huge(*pud));
414 pmd_free(NULL, pmd_table);
415 put_page(virt_to_page(pud));
418 static void clear_hyp_pmd_entry(pmd_t *pmd)
420 pte_t *pte_table = pte_offset_kernel(pmd, 0);
421 VM_BUG_ON(pmd_thp_or_huge(*pmd));
423 pte_free_kernel(NULL, pte_table);
424 put_page(virt_to_page(pmd));
427 static void unmap_hyp_ptes(pmd_t *pmd, phys_addr_t addr, phys_addr_t end)
429 pte_t *pte, *start_pte;
431 start_pte = pte = pte_offset_kernel(pmd, addr);
433 if (!pte_none(*pte)) {
434 kvm_set_pte(pte, __pte(0));
435 put_page(virt_to_page(pte));
437 } while (pte++, addr += PAGE_SIZE, addr != end);
439 if (hyp_pte_table_empty(start_pte))
440 clear_hyp_pmd_entry(pmd);
443 static void unmap_hyp_pmds(pud_t *pud, phys_addr_t addr, phys_addr_t end)
446 pmd_t *pmd, *start_pmd;
448 start_pmd = pmd = pmd_offset(pud, addr);
450 next = pmd_addr_end(addr, end);
451 /* Hyp doesn't use huge pmds */
453 unmap_hyp_ptes(pmd, addr, next);
454 } while (pmd++, addr = next, addr != end);
456 if (hyp_pmd_table_empty(start_pmd))
457 clear_hyp_pud_entry(pud);
460 static void unmap_hyp_puds(pgd_t *pgd, phys_addr_t addr, phys_addr_t end)
463 pud_t *pud, *start_pud;
465 start_pud = pud = pud_offset(pgd, addr);
467 next = pud_addr_end(addr, end);
468 /* Hyp doesn't use huge puds */
470 unmap_hyp_pmds(pud, addr, next);
471 } while (pud++, addr = next, addr != end);
473 if (hyp_pud_table_empty(start_pud))
474 clear_hyp_pgd_entry(pgd);
477 static void unmap_hyp_range(pgd_t *pgdp, phys_addr_t start, u64 size)
480 phys_addr_t addr = start, end = start + size;
484 * We don't unmap anything from HYP, except at the hyp tear down.
485 * Hence, we don't have to invalidate the TLBs here.
487 pgd = pgdp + pgd_index(addr);
489 next = pgd_addr_end(addr, end);
491 unmap_hyp_puds(pgd, addr, next);
492 } while (pgd++, addr = next, addr != end);
496 * free_hyp_pgds - free Hyp-mode page tables
498 * Assumes hyp_pgd is a page table used strictly in Hyp-mode and
499 * therefore contains either mappings in the kernel memory area (above
500 * PAGE_OFFSET), or device mappings in the vmalloc range (from
501 * VMALLOC_START to VMALLOC_END).
503 * boot_hyp_pgd should only map two pages for the init code.
505 void free_hyp_pgds(void)
507 mutex_lock(&kvm_hyp_pgd_mutex);
510 unmap_hyp_range(boot_hyp_pgd, hyp_idmap_start, PAGE_SIZE);
511 free_pages((unsigned long)boot_hyp_pgd, hyp_pgd_order);
516 unmap_hyp_range(hyp_pgd, hyp_idmap_start, PAGE_SIZE);
517 unmap_hyp_range(hyp_pgd, kern_hyp_va(PAGE_OFFSET),
518 (uintptr_t)high_memory - PAGE_OFFSET);
519 unmap_hyp_range(hyp_pgd, kern_hyp_va(VMALLOC_START),
520 VMALLOC_END - VMALLOC_START);
522 free_pages((unsigned long)hyp_pgd, hyp_pgd_order);
525 if (merged_hyp_pgd) {
526 clear_page(merged_hyp_pgd);
527 free_page((unsigned long)merged_hyp_pgd);
528 merged_hyp_pgd = NULL;
531 mutex_unlock(&kvm_hyp_pgd_mutex);
534 static void create_hyp_pte_mappings(pmd_t *pmd, unsigned long start,
535 unsigned long end, unsigned long pfn,
543 pte = pte_offset_kernel(pmd, addr);
544 kvm_set_pte(pte, pfn_pte(pfn, prot));
545 get_page(virt_to_page(pte));
546 kvm_flush_dcache_to_poc(pte, sizeof(*pte));
548 } while (addr += PAGE_SIZE, addr != end);
551 static int create_hyp_pmd_mappings(pud_t *pud, unsigned long start,
552 unsigned long end, unsigned long pfn,
557 unsigned long addr, next;
561 pmd = pmd_offset(pud, addr);
563 BUG_ON(pmd_sect(*pmd));
565 if (pmd_none(*pmd)) {
566 pte = pte_alloc_one_kernel(NULL, addr);
568 kvm_err("Cannot allocate Hyp pte\n");
571 pmd_populate_kernel(NULL, pmd, pte);
572 get_page(virt_to_page(pmd));
573 kvm_flush_dcache_to_poc(pmd, sizeof(*pmd));
576 next = pmd_addr_end(addr, end);
578 create_hyp_pte_mappings(pmd, addr, next, pfn, prot);
579 pfn += (next - addr) >> PAGE_SHIFT;
580 } while (addr = next, addr != end);
585 static int create_hyp_pud_mappings(pgd_t *pgd, unsigned long start,
586 unsigned long end, unsigned long pfn,
591 unsigned long addr, next;
596 pud = pud_offset(pgd, addr);
598 if (pud_none_or_clear_bad(pud)) {
599 pmd = pmd_alloc_one(NULL, addr);
601 kvm_err("Cannot allocate Hyp pmd\n");
604 pud_populate(NULL, pud, pmd);
605 get_page(virt_to_page(pud));
606 kvm_flush_dcache_to_poc(pud, sizeof(*pud));
609 next = pud_addr_end(addr, end);
610 ret = create_hyp_pmd_mappings(pud, addr, next, pfn, prot);
613 pfn += (next - addr) >> PAGE_SHIFT;
614 } while (addr = next, addr != end);
619 static int __create_hyp_mappings(pgd_t *pgdp,
620 unsigned long start, unsigned long end,
621 unsigned long pfn, pgprot_t prot)
625 unsigned long addr, next;
628 mutex_lock(&kvm_hyp_pgd_mutex);
629 addr = start & PAGE_MASK;
630 end = PAGE_ALIGN(end);
632 pgd = pgdp + pgd_index(addr);
634 if (pgd_none(*pgd)) {
635 pud = pud_alloc_one(NULL, addr);
637 kvm_err("Cannot allocate Hyp pud\n");
641 pgd_populate(NULL, pgd, pud);
642 get_page(virt_to_page(pgd));
643 kvm_flush_dcache_to_poc(pgd, sizeof(*pgd));
646 next = pgd_addr_end(addr, end);
647 err = create_hyp_pud_mappings(pgd, addr, next, pfn, prot);
650 pfn += (next - addr) >> PAGE_SHIFT;
651 } while (addr = next, addr != end);
653 mutex_unlock(&kvm_hyp_pgd_mutex);
657 static phys_addr_t kvm_kaddr_to_phys(void *kaddr)
659 if (!is_vmalloc_addr(kaddr)) {
660 BUG_ON(!virt_addr_valid(kaddr));
663 return page_to_phys(vmalloc_to_page(kaddr)) +
664 offset_in_page(kaddr);
669 * create_hyp_mappings - duplicate a kernel virtual address range in Hyp mode
670 * @from: The virtual kernel start address of the range
671 * @to: The virtual kernel end address of the range (exclusive)
672 * @prot: The protection to be applied to this range
674 * The same virtual address as the kernel virtual address is also used
675 * in Hyp-mode mapping (modulo HYP_PAGE_OFFSET) to the same underlying
678 int create_hyp_mappings(void *from, void *to, pgprot_t prot)
680 phys_addr_t phys_addr;
681 unsigned long virt_addr;
682 unsigned long start = kern_hyp_va((unsigned long)from);
683 unsigned long end = kern_hyp_va((unsigned long)to);
685 if (is_kernel_in_hyp_mode())
688 start = start & PAGE_MASK;
689 end = PAGE_ALIGN(end);
691 for (virt_addr = start; virt_addr < end; virt_addr += PAGE_SIZE) {
694 phys_addr = kvm_kaddr_to_phys(from + virt_addr - start);
695 err = __create_hyp_mappings(hyp_pgd, virt_addr,
696 virt_addr + PAGE_SIZE,
697 __phys_to_pfn(phys_addr),
707 * create_hyp_io_mappings - duplicate a kernel IO mapping into Hyp mode
708 * @from: The kernel start VA of the range
709 * @to: The kernel end VA of the range (exclusive)
710 * @phys_addr: The physical start address which gets mapped
712 * The resulting HYP VA is the same as the kernel VA, modulo
715 int create_hyp_io_mappings(void *from, void *to, phys_addr_t phys_addr)
717 unsigned long start = kern_hyp_va((unsigned long)from);
718 unsigned long end = kern_hyp_va((unsigned long)to);
720 if (is_kernel_in_hyp_mode())
723 /* Check for a valid kernel IO mapping */
724 if (!is_vmalloc_addr(from) || !is_vmalloc_addr(to - 1))
727 return __create_hyp_mappings(hyp_pgd, start, end,
728 __phys_to_pfn(phys_addr), PAGE_HYP_DEVICE);
732 * kvm_alloc_stage2_pgd - allocate level-1 table for stage-2 translation.
733 * @kvm: The KVM struct pointer for the VM.
735 * Allocates only the stage-2 HW PGD level table(s) (can support either full
736 * 40-bit input addresses or limited to 32-bit input addresses). Clears the
739 * Note we don't need locking here as this is only called when the VM is
740 * created, which can only be done once.
742 int kvm_alloc_stage2_pgd(struct kvm *kvm)
746 if (kvm->arch.pgd != NULL) {
747 kvm_err("kvm_arch already initialized?\n");
751 /* Allocate the HW PGD, making sure that each page gets its own refcount */
752 pgd = alloc_pages_exact(S2_PGD_SIZE, GFP_KERNEL | __GFP_ZERO);
760 static void stage2_unmap_memslot(struct kvm *kvm,
761 struct kvm_memory_slot *memslot)
763 hva_t hva = memslot->userspace_addr;
764 phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
765 phys_addr_t size = PAGE_SIZE * memslot->npages;
766 hva_t reg_end = hva + size;
769 * A memory region could potentially cover multiple VMAs, and any holes
770 * between them, so iterate over all of them to find out if we should
773 * +--------------------------------------------+
774 * +---------------+----------------+ +----------------+
775 * | : VMA 1 | VMA 2 | | VMA 3 : |
776 * +---------------+----------------+ +----------------+
778 * +--------------------------------------------+
781 struct vm_area_struct *vma = find_vma(current->mm, hva);
782 hva_t vm_start, vm_end;
784 if (!vma || vma->vm_start >= reg_end)
788 * Take the intersection of this VMA with the memory region
790 vm_start = max(hva, vma->vm_start);
791 vm_end = min(reg_end, vma->vm_end);
793 if (!(vma->vm_flags & VM_PFNMAP)) {
794 gpa_t gpa = addr + (vm_start - memslot->userspace_addr);
795 unmap_stage2_range(kvm, gpa, vm_end - vm_start);
798 } while (hva < reg_end);
802 * stage2_unmap_vm - Unmap Stage-2 RAM mappings
803 * @kvm: The struct kvm pointer
805 * Go through the memregions and unmap any reguler RAM
806 * backing memory already mapped to the VM.
808 void stage2_unmap_vm(struct kvm *kvm)
810 struct kvm_memslots *slots;
811 struct kvm_memory_slot *memslot;
814 idx = srcu_read_lock(&kvm->srcu);
815 down_read(¤t->mm->mmap_sem);
816 spin_lock(&kvm->mmu_lock);
818 slots = kvm_memslots(kvm);
819 kvm_for_each_memslot(memslot, slots)
820 stage2_unmap_memslot(kvm, memslot);
822 spin_unlock(&kvm->mmu_lock);
823 up_read(¤t->mm->mmap_sem);
824 srcu_read_unlock(&kvm->srcu, idx);
828 * kvm_free_stage2_pgd - free all stage-2 tables
829 * @kvm: The KVM struct pointer for the VM.
831 * Walks the level-1 page table pointed to by kvm->arch.pgd and frees all
832 * underlying level-2 and level-3 tables before freeing the actual level-1 table
833 * and setting the struct pointer to NULL.
835 void kvm_free_stage2_pgd(struct kvm *kvm)
839 spin_lock(&kvm->mmu_lock);
841 unmap_stage2_range(kvm, 0, KVM_PHYS_SIZE);
842 pgd = READ_ONCE(kvm->arch.pgd);
843 kvm->arch.pgd = NULL;
845 spin_unlock(&kvm->mmu_lock);
847 /* Free the HW pgd, one page at a time */
849 free_pages_exact(pgd, S2_PGD_SIZE);
852 static pud_t *stage2_get_pud(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
858 pgd = kvm->arch.pgd + stage2_pgd_index(addr);
859 if (WARN_ON(stage2_pgd_none(*pgd))) {
862 pud = mmu_memory_cache_alloc(cache);
863 stage2_pgd_populate(pgd, pud);
864 get_page(virt_to_page(pgd));
867 return stage2_pud_offset(pgd, addr);
870 static pmd_t *stage2_get_pmd(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
876 pud = stage2_get_pud(kvm, cache, addr);
880 if (stage2_pud_none(*pud)) {
883 pmd = mmu_memory_cache_alloc(cache);
884 stage2_pud_populate(pud, pmd);
885 get_page(virt_to_page(pud));
888 return stage2_pmd_offset(pud, addr);
891 static int stage2_set_pmd_huge(struct kvm *kvm, struct kvm_mmu_memory_cache
892 *cache, phys_addr_t addr, const pmd_t *new_pmd)
896 pmd = stage2_get_pmd(kvm, cache, addr);
900 if (pmd_present(old_pmd)) {
902 * Multiple vcpus faulting on the same PMD entry, can
903 * lead to them sequentially updating the PMD with the
904 * same value. Following the break-before-make
905 * (pmd_clear() followed by tlb_flush()) process can
906 * hinder forward progress due to refaults generated
907 * on missing translations.
909 * Skip updating the page table if the entry is
912 if (pmd_val(old_pmd) == pmd_val(*new_pmd))
916 * Mapping in huge pages should only happen through a
917 * fault. If a page is merged into a transparent huge
918 * page, the individual subpages of that huge page
919 * should be unmapped through MMU notifiers before we
922 * Merging of CompoundPages is not supported; they
923 * should become splitting first, unmapped, merged,
924 * and mapped back in on-demand.
926 VM_BUG_ON(pmd_pfn(old_pmd) != pmd_pfn(*new_pmd));
929 kvm_tlb_flush_vmid_ipa(kvm, addr);
931 get_page(virt_to_page(pmd));
934 kvm_set_pmd(pmd, *new_pmd);
938 static int stage2_set_pte(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
939 phys_addr_t addr, const pte_t *new_pte,
944 bool iomap = flags & KVM_S2PTE_FLAG_IS_IOMAP;
945 bool logging_active = flags & KVM_S2_FLAG_LOGGING_ACTIVE;
947 VM_BUG_ON(logging_active && !cache);
949 /* Create stage-2 page table mapping - Levels 0 and 1 */
950 pmd = stage2_get_pmd(kvm, cache, addr);
953 * Ignore calls from kvm_set_spte_hva for unallocated
960 * While dirty page logging - dissolve huge PMD, then continue on to
964 stage2_dissolve_pmd(kvm, addr, pmd);
966 /* Create stage-2 page mappings - Level 2 */
967 if (pmd_none(*pmd)) {
969 return 0; /* ignore calls from kvm_set_spte_hva */
970 pte = mmu_memory_cache_alloc(cache);
971 pmd_populate_kernel(NULL, pmd, pte);
972 get_page(virt_to_page(pmd));
975 pte = pte_offset_kernel(pmd, addr);
977 if (iomap && pte_present(*pte))
980 /* Create 2nd stage page table mapping - Level 3 */
982 if (pte_present(old_pte)) {
983 /* Skip page table update if there is no change */
984 if (pte_val(old_pte) == pte_val(*new_pte))
987 kvm_set_pte(pte, __pte(0));
988 kvm_tlb_flush_vmid_ipa(kvm, addr);
990 get_page(virt_to_page(pte));
993 kvm_set_pte(pte, *new_pte);
997 #ifndef __HAVE_ARCH_PTEP_TEST_AND_CLEAR_YOUNG
998 static int stage2_ptep_test_and_clear_young(pte_t *pte)
1000 if (pte_young(*pte)) {
1001 *pte = pte_mkold(*pte);
1007 static int stage2_ptep_test_and_clear_young(pte_t *pte)
1009 return __ptep_test_and_clear_young(pte);
1013 static int stage2_pmdp_test_and_clear_young(pmd_t *pmd)
1015 return stage2_ptep_test_and_clear_young((pte_t *)pmd);
1019 * kvm_phys_addr_ioremap - map a device range to guest IPA
1021 * @kvm: The KVM pointer
1022 * @guest_ipa: The IPA at which to insert the mapping
1023 * @pa: The physical address of the device
1024 * @size: The size of the mapping
1026 int kvm_phys_addr_ioremap(struct kvm *kvm, phys_addr_t guest_ipa,
1027 phys_addr_t pa, unsigned long size, bool writable)
1029 phys_addr_t addr, end;
1032 struct kvm_mmu_memory_cache cache = { 0, };
1034 end = (guest_ipa + size + PAGE_SIZE - 1) & PAGE_MASK;
1035 pfn = __phys_to_pfn(pa);
1037 for (addr = guest_ipa; addr < end; addr += PAGE_SIZE) {
1038 pte_t pte = pfn_pte(pfn, PAGE_S2_DEVICE);
1041 pte = kvm_s2pte_mkwrite(pte);
1043 ret = mmu_topup_memory_cache(&cache, KVM_MMU_CACHE_MIN_PAGES,
1047 spin_lock(&kvm->mmu_lock);
1048 ret = stage2_set_pte(kvm, &cache, addr, &pte,
1049 KVM_S2PTE_FLAG_IS_IOMAP);
1050 spin_unlock(&kvm->mmu_lock);
1058 mmu_free_memory_cache(&cache);
1062 static bool transparent_hugepage_adjust(kvm_pfn_t *pfnp, phys_addr_t *ipap)
1064 kvm_pfn_t pfn = *pfnp;
1065 gfn_t gfn = *ipap >> PAGE_SHIFT;
1066 struct page *page = pfn_to_page(pfn);
1069 * PageTransCompoungMap() returns true for THP and
1070 * hugetlbfs. Make sure the adjustment is done only for THP
1073 if (!PageHuge(page) && PageTransCompoundMap(page)) {
1076 * The address we faulted on is backed by a transparent huge
1077 * page. However, because we map the compound huge page and
1078 * not the individual tail page, we need to transfer the
1079 * refcount to the head page. We have to be careful that the
1080 * THP doesn't start to split while we are adjusting the
1083 * We are sure this doesn't happen, because mmu_notifier_retry
1084 * was successful and we are holding the mmu_lock, so if this
1085 * THP is trying to split, it will be blocked in the mmu
1086 * notifier before touching any of the pages, specifically
1087 * before being able to call __split_huge_page_refcount().
1089 * We can therefore safely transfer the refcount from PG_tail
1090 * to PG_head and switch the pfn from a tail page to the head
1093 mask = PTRS_PER_PMD - 1;
1094 VM_BUG_ON((gfn & mask) != (pfn & mask));
1097 kvm_release_pfn_clean(pfn);
1109 static bool kvm_is_write_fault(struct kvm_vcpu *vcpu)
1111 if (kvm_vcpu_trap_is_iabt(vcpu))
1114 return kvm_vcpu_dabt_iswrite(vcpu);
1118 * stage2_wp_ptes - write protect PMD range
1119 * @pmd: pointer to pmd entry
1120 * @addr: range start address
1121 * @end: range end address
1123 static void stage2_wp_ptes(pmd_t *pmd, phys_addr_t addr, phys_addr_t end)
1127 pte = pte_offset_kernel(pmd, addr);
1129 if (!pte_none(*pte)) {
1130 if (!kvm_s2pte_readonly(pte))
1131 kvm_set_s2pte_readonly(pte);
1133 } while (pte++, addr += PAGE_SIZE, addr != end);
1137 * stage2_wp_pmds - write protect PUD range
1138 * @pud: pointer to pud entry
1139 * @addr: range start address
1140 * @end: range end address
1142 static void stage2_wp_pmds(pud_t *pud, phys_addr_t addr, phys_addr_t end)
1147 pmd = stage2_pmd_offset(pud, addr);
1150 next = stage2_pmd_addr_end(addr, end);
1151 if (!pmd_none(*pmd)) {
1152 if (pmd_thp_or_huge(*pmd)) {
1153 if (!kvm_s2pmd_readonly(pmd))
1154 kvm_set_s2pmd_readonly(pmd);
1156 stage2_wp_ptes(pmd, addr, next);
1159 } while (pmd++, addr = next, addr != end);
1163 * stage2_wp_puds - write protect PGD range
1164 * @pgd: pointer to pgd entry
1165 * @addr: range start address
1166 * @end: range end address
1168 * Process PUD entries, for a huge PUD we cause a panic.
1170 static void stage2_wp_puds(pgd_t *pgd, phys_addr_t addr, phys_addr_t end)
1175 pud = stage2_pud_offset(pgd, addr);
1177 next = stage2_pud_addr_end(addr, end);
1178 if (!stage2_pud_none(*pud)) {
1179 /* TODO:PUD not supported, revisit later if supported */
1180 BUG_ON(stage2_pud_huge(*pud));
1181 stage2_wp_pmds(pud, addr, next);
1183 } while (pud++, addr = next, addr != end);
1187 * stage2_wp_range() - write protect stage2 memory region range
1188 * @kvm: The KVM pointer
1189 * @addr: Start address of range
1190 * @end: End address of range
1192 static void stage2_wp_range(struct kvm *kvm, phys_addr_t addr, phys_addr_t end)
1197 pgd = kvm->arch.pgd + stage2_pgd_index(addr);
1200 * Release kvm_mmu_lock periodically if the memory region is
1201 * large. Otherwise, we may see kernel panics with
1202 * CONFIG_DETECT_HUNG_TASK, CONFIG_LOCKUP_DETECTOR,
1203 * CONFIG_LOCKDEP. Additionally, holding the lock too long
1204 * will also starve other vCPUs. We have to also make sure
1205 * that the page tables are not freed while we released
1208 cond_resched_lock(&kvm->mmu_lock);
1209 if (!READ_ONCE(kvm->arch.pgd))
1211 next = stage2_pgd_addr_end(addr, end);
1212 if (stage2_pgd_present(*pgd))
1213 stage2_wp_puds(pgd, addr, next);
1214 } while (pgd++, addr = next, addr != end);
1218 * kvm_mmu_wp_memory_region() - write protect stage 2 entries for memory slot
1219 * @kvm: The KVM pointer
1220 * @slot: The memory slot to write protect
1222 * Called to start logging dirty pages after memory region
1223 * KVM_MEM_LOG_DIRTY_PAGES operation is called. After this function returns
1224 * all present PMD and PTEs are write protected in the memory region.
1225 * Afterwards read of dirty page log can be called.
1227 * Acquires kvm_mmu_lock. Called with kvm->slots_lock mutex acquired,
1228 * serializing operations for VM memory regions.
1230 void kvm_mmu_wp_memory_region(struct kvm *kvm, int slot)
1232 struct kvm_memslots *slots = kvm_memslots(kvm);
1233 struct kvm_memory_slot *memslot = id_to_memslot(slots, slot);
1234 phys_addr_t start = memslot->base_gfn << PAGE_SHIFT;
1235 phys_addr_t end = (memslot->base_gfn + memslot->npages) << PAGE_SHIFT;
1237 spin_lock(&kvm->mmu_lock);
1238 stage2_wp_range(kvm, start, end);
1239 spin_unlock(&kvm->mmu_lock);
1240 kvm_flush_remote_tlbs(kvm);
1244 * kvm_mmu_write_protect_pt_masked() - write protect dirty pages
1245 * @kvm: The KVM pointer
1246 * @slot: The memory slot associated with mask
1247 * @gfn_offset: The gfn offset in memory slot
1248 * @mask: The mask of dirty pages at offset 'gfn_offset' in this memory
1249 * slot to be write protected
1251 * Walks bits set in mask write protects the associated pte's. Caller must
1252 * acquire kvm_mmu_lock.
1254 static void kvm_mmu_write_protect_pt_masked(struct kvm *kvm,
1255 struct kvm_memory_slot *slot,
1256 gfn_t gfn_offset, unsigned long mask)
1258 phys_addr_t base_gfn = slot->base_gfn + gfn_offset;
1259 phys_addr_t start = (base_gfn + __ffs(mask)) << PAGE_SHIFT;
1260 phys_addr_t end = (base_gfn + __fls(mask) + 1) << PAGE_SHIFT;
1262 stage2_wp_range(kvm, start, end);
1266 * kvm_arch_mmu_enable_log_dirty_pt_masked - enable dirty logging for selected
1269 * It calls kvm_mmu_write_protect_pt_masked to write protect selected pages to
1270 * enable dirty logging for them.
1272 void kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm *kvm,
1273 struct kvm_memory_slot *slot,
1274 gfn_t gfn_offset, unsigned long mask)
1276 kvm_mmu_write_protect_pt_masked(kvm, slot, gfn_offset, mask);
1279 static void coherent_cache_guest_page(struct kvm_vcpu *vcpu, kvm_pfn_t pfn,
1282 __coherent_cache_guest_page(vcpu, pfn, size);
1285 static void kvm_send_hwpoison_signal(unsigned long address,
1286 struct vm_area_struct *vma)
1290 info.si_signo = SIGBUS;
1292 info.si_code = BUS_MCEERR_AR;
1293 info.si_addr = (void __user *)address;
1295 if (is_vm_hugetlb_page(vma))
1296 info.si_addr_lsb = huge_page_shift(hstate_vma(vma));
1298 info.si_addr_lsb = PAGE_SHIFT;
1300 send_sig_info(SIGBUS, &info, current);
1303 static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
1304 struct kvm_memory_slot *memslot, unsigned long hva,
1305 unsigned long fault_status)
1308 bool write_fault, writable, hugetlb = false, force_pte = false;
1309 unsigned long mmu_seq;
1310 gfn_t gfn = fault_ipa >> PAGE_SHIFT;
1311 struct kvm *kvm = vcpu->kvm;
1312 struct kvm_mmu_memory_cache *memcache = &vcpu->arch.mmu_page_cache;
1313 struct vm_area_struct *vma;
1315 pgprot_t mem_type = PAGE_S2;
1316 bool logging_active = memslot_is_logging(memslot);
1317 unsigned long flags = 0;
1319 write_fault = kvm_is_write_fault(vcpu);
1320 if (fault_status == FSC_PERM && !write_fault) {
1321 kvm_err("Unexpected L2 read permission error\n");
1325 /* Let's check if we will get back a huge page backed by hugetlbfs */
1326 down_read(¤t->mm->mmap_sem);
1327 vma = find_vma_intersection(current->mm, hva, hva + 1);
1328 if (unlikely(!vma)) {
1329 kvm_err("Failed to find VMA for hva 0x%lx\n", hva);
1330 up_read(¤t->mm->mmap_sem);
1334 if (vma_kernel_pagesize(vma) == PMD_SIZE && !logging_active) {
1336 gfn = (fault_ipa & PMD_MASK) >> PAGE_SHIFT;
1339 * Pages belonging to memslots that don't have the same
1340 * alignment for userspace and IPA cannot be mapped using
1341 * block descriptors even if the pages belong to a THP for
1342 * the process, because the stage-2 block descriptor will
1343 * cover more than a single THP and we loose atomicity for
1344 * unmapping, updates, and splits of the THP or other pages
1345 * in the stage-2 block range.
1347 if ((memslot->userspace_addr & ~PMD_MASK) !=
1348 ((memslot->base_gfn << PAGE_SHIFT) & ~PMD_MASK))
1351 up_read(¤t->mm->mmap_sem);
1353 /* We need minimum second+third level pages */
1354 ret = mmu_topup_memory_cache(memcache, KVM_MMU_CACHE_MIN_PAGES,
1359 mmu_seq = vcpu->kvm->mmu_notifier_seq;
1361 * Ensure the read of mmu_notifier_seq happens before we call
1362 * gfn_to_pfn_prot (which calls get_user_pages), so that we don't risk
1363 * the page we just got a reference to gets unmapped before we have a
1364 * chance to grab the mmu_lock, which ensure that if the page gets
1365 * unmapped afterwards, the call to kvm_unmap_hva will take it away
1366 * from us again properly. This smp_rmb() interacts with the smp_wmb()
1367 * in kvm_mmu_notifier_invalidate_<page|range_end>.
1371 pfn = gfn_to_pfn_prot(kvm, gfn, write_fault, &writable);
1372 if (pfn == KVM_PFN_ERR_HWPOISON) {
1373 kvm_send_hwpoison_signal(hva, vma);
1376 if (is_error_noslot_pfn(pfn))
1379 if (kvm_is_device_pfn(pfn)) {
1380 mem_type = PAGE_S2_DEVICE;
1381 flags |= KVM_S2PTE_FLAG_IS_IOMAP;
1382 } else if (logging_active) {
1384 * Faults on pages in a memslot with logging enabled
1385 * should not be mapped with huge pages (it introduces churn
1386 * and performance degradation), so force a pte mapping.
1389 flags |= KVM_S2_FLAG_LOGGING_ACTIVE;
1392 * Only actually map the page as writable if this was a write
1399 spin_lock(&kvm->mmu_lock);
1400 if (mmu_notifier_retry(kvm, mmu_seq))
1403 if (!hugetlb && !force_pte)
1404 hugetlb = transparent_hugepage_adjust(&pfn, &fault_ipa);
1407 pmd_t new_pmd = pfn_pmd(pfn, mem_type);
1408 new_pmd = pmd_mkhuge(new_pmd);
1410 new_pmd = kvm_s2pmd_mkwrite(new_pmd);
1411 kvm_set_pfn_dirty(pfn);
1413 coherent_cache_guest_page(vcpu, pfn, PMD_SIZE);
1414 ret = stage2_set_pmd_huge(kvm, memcache, fault_ipa, &new_pmd);
1416 pte_t new_pte = pfn_pte(pfn, mem_type);
1419 new_pte = kvm_s2pte_mkwrite(new_pte);
1420 kvm_set_pfn_dirty(pfn);
1421 mark_page_dirty(kvm, gfn);
1423 coherent_cache_guest_page(vcpu, pfn, PAGE_SIZE);
1424 ret = stage2_set_pte(kvm, memcache, fault_ipa, &new_pte, flags);
1428 spin_unlock(&kvm->mmu_lock);
1429 kvm_set_pfn_accessed(pfn);
1430 kvm_release_pfn_clean(pfn);
1435 * Resolve the access fault by making the page young again.
1436 * Note that because the faulting entry is guaranteed not to be
1437 * cached in the TLB, we don't need to invalidate anything.
1438 * Only the HW Access Flag updates are supported for Stage 2 (no DBM),
1439 * so there is no need for atomic (pte|pmd)_mkyoung operations.
1441 static void handle_access_fault(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa)
1446 bool pfn_valid = false;
1448 trace_kvm_access_fault(fault_ipa);
1450 spin_lock(&vcpu->kvm->mmu_lock);
1452 pmd = stage2_get_pmd(vcpu->kvm, NULL, fault_ipa);
1453 if (!pmd || pmd_none(*pmd)) /* Nothing there */
1456 if (pmd_thp_or_huge(*pmd)) { /* THP, HugeTLB */
1457 *pmd = pmd_mkyoung(*pmd);
1458 pfn = pmd_pfn(*pmd);
1463 pte = pte_offset_kernel(pmd, fault_ipa);
1464 if (pte_none(*pte)) /* Nothing there either */
1467 *pte = pte_mkyoung(*pte); /* Just a page... */
1468 pfn = pte_pfn(*pte);
1471 spin_unlock(&vcpu->kvm->mmu_lock);
1473 kvm_set_pfn_accessed(pfn);
1477 * kvm_handle_guest_abort - handles all 2nd stage aborts
1478 * @vcpu: the VCPU pointer
1479 * @run: the kvm_run structure
1481 * Any abort that gets to the host is almost guaranteed to be caused by a
1482 * missing second stage translation table entry, which can mean that either the
1483 * guest simply needs more memory and we must allocate an appropriate page or it
1484 * can mean that the guest tried to access I/O memory, which is emulated by user
1485 * space. The distinction is based on the IPA causing the fault and whether this
1486 * memory region has been registered as standard RAM by user space.
1488 int kvm_handle_guest_abort(struct kvm_vcpu *vcpu, struct kvm_run *run)
1490 unsigned long fault_status;
1491 phys_addr_t fault_ipa;
1492 struct kvm_memory_slot *memslot;
1494 bool is_iabt, write_fault, writable;
1498 fault_status = kvm_vcpu_trap_get_fault_type(vcpu);
1500 fault_ipa = kvm_vcpu_get_fault_ipa(vcpu);
1501 is_iabt = kvm_vcpu_trap_is_iabt(vcpu);
1503 /* Synchronous External Abort? */
1504 if (kvm_vcpu_dabt_isextabt(vcpu)) {
1506 * For RAS the host kernel may handle this abort.
1507 * There is no need to pass the error into the guest.
1509 if (!handle_guest_sea(fault_ipa, kvm_vcpu_get_hsr(vcpu)))
1512 if (unlikely(!is_iabt)) {
1513 kvm_inject_vabt(vcpu);
1518 trace_kvm_guest_fault(*vcpu_pc(vcpu), kvm_vcpu_get_hsr(vcpu),
1519 kvm_vcpu_get_hfar(vcpu), fault_ipa);
1521 /* Check the stage-2 fault is trans. fault or write fault */
1522 if (fault_status != FSC_FAULT && fault_status != FSC_PERM &&
1523 fault_status != FSC_ACCESS) {
1524 kvm_err("Unsupported FSC: EC=%#x xFSC=%#lx ESR_EL2=%#lx\n",
1525 kvm_vcpu_trap_get_class(vcpu),
1526 (unsigned long)kvm_vcpu_trap_get_fault(vcpu),
1527 (unsigned long)kvm_vcpu_get_hsr(vcpu));
1531 idx = srcu_read_lock(&vcpu->kvm->srcu);
1533 gfn = fault_ipa >> PAGE_SHIFT;
1534 memslot = gfn_to_memslot(vcpu->kvm, gfn);
1535 hva = gfn_to_hva_memslot_prot(memslot, gfn, &writable);
1536 write_fault = kvm_is_write_fault(vcpu);
1537 if (kvm_is_error_hva(hva) || (write_fault && !writable)) {
1539 /* Prefetch Abort on I/O address */
1540 kvm_inject_pabt(vcpu, kvm_vcpu_get_hfar(vcpu));
1546 * Check for a cache maintenance operation. Since we
1547 * ended-up here, we know it is outside of any memory
1548 * slot. But we can't find out if that is for a device,
1549 * or if the guest is just being stupid. The only thing
1550 * we know for sure is that this range cannot be cached.
1552 * So let's assume that the guest is just being
1553 * cautious, and skip the instruction.
1555 if (kvm_vcpu_dabt_is_cm(vcpu)) {
1556 kvm_skip_instr(vcpu, kvm_vcpu_trap_il_is32bit(vcpu));
1562 * The IPA is reported as [MAX:12], so we need to
1563 * complement it with the bottom 12 bits from the
1564 * faulting VA. This is always 12 bits, irrespective
1567 fault_ipa |= kvm_vcpu_get_hfar(vcpu) & ((1 << 12) - 1);
1568 ret = io_mem_abort(vcpu, run, fault_ipa);
1572 /* Userspace should not be able to register out-of-bounds IPAs */
1573 VM_BUG_ON(fault_ipa >= KVM_PHYS_SIZE);
1575 if (fault_status == FSC_ACCESS) {
1576 handle_access_fault(vcpu, fault_ipa);
1581 ret = user_mem_abort(vcpu, fault_ipa, memslot, hva, fault_status);
1585 srcu_read_unlock(&vcpu->kvm->srcu, idx);
1589 static int handle_hva_to_gpa(struct kvm *kvm,
1590 unsigned long start,
1592 int (*handler)(struct kvm *kvm,
1593 gpa_t gpa, u64 size,
1597 struct kvm_memslots *slots;
1598 struct kvm_memory_slot *memslot;
1601 slots = kvm_memslots(kvm);
1603 /* we only care about the pages that the guest sees */
1604 kvm_for_each_memslot(memslot, slots) {
1605 unsigned long hva_start, hva_end;
1608 hva_start = max(start, memslot->userspace_addr);
1609 hva_end = min(end, memslot->userspace_addr +
1610 (memslot->npages << PAGE_SHIFT));
1611 if (hva_start >= hva_end)
1614 gpa = hva_to_gfn_memslot(hva_start, memslot) << PAGE_SHIFT;
1615 ret |= handler(kvm, gpa, (u64)(hva_end - hva_start), data);
1621 static int kvm_unmap_hva_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
1623 unmap_stage2_range(kvm, gpa, size);
1627 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
1629 unsigned long end = hva + PAGE_SIZE;
1634 trace_kvm_unmap_hva(hva);
1635 handle_hva_to_gpa(kvm, hva, end, &kvm_unmap_hva_handler, NULL);
1639 int kvm_unmap_hva_range(struct kvm *kvm,
1640 unsigned long start, unsigned long end)
1645 trace_kvm_unmap_hva_range(start, end);
1646 handle_hva_to_gpa(kvm, start, end, &kvm_unmap_hva_handler, NULL);
1650 static int kvm_set_spte_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
1652 pte_t *pte = (pte_t *)data;
1654 WARN_ON(size != PAGE_SIZE);
1656 * We can always call stage2_set_pte with KVM_S2PTE_FLAG_LOGGING_ACTIVE
1657 * flag clear because MMU notifiers will have unmapped a huge PMD before
1658 * calling ->change_pte() (which in turn calls kvm_set_spte_hva()) and
1659 * therefore stage2_set_pte() never needs to clear out a huge PMD
1660 * through this calling path.
1662 stage2_set_pte(kvm, NULL, gpa, pte, 0);
1667 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
1669 unsigned long end = hva + PAGE_SIZE;
1675 trace_kvm_set_spte_hva(hva);
1676 stage2_pte = pfn_pte(pte_pfn(pte), PAGE_S2);
1677 handle_hva_to_gpa(kvm, hva, end, &kvm_set_spte_handler, &stage2_pte);
1680 static int kvm_age_hva_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
1685 WARN_ON(size != PAGE_SIZE && size != PMD_SIZE);
1686 pmd = stage2_get_pmd(kvm, NULL, gpa);
1687 if (!pmd || pmd_none(*pmd)) /* Nothing there */
1690 if (pmd_thp_or_huge(*pmd)) /* THP, HugeTLB */
1691 return stage2_pmdp_test_and_clear_young(pmd);
1693 pte = pte_offset_kernel(pmd, gpa);
1697 return stage2_ptep_test_and_clear_young(pte);
1700 static int kvm_test_age_hva_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
1705 WARN_ON(size != PAGE_SIZE && size != PMD_SIZE);
1706 pmd = stage2_get_pmd(kvm, NULL, gpa);
1707 if (!pmd || pmd_none(*pmd)) /* Nothing there */
1710 if (pmd_thp_or_huge(*pmd)) /* THP, HugeTLB */
1711 return pmd_young(*pmd);
1713 pte = pte_offset_kernel(pmd, gpa);
1714 if (!pte_none(*pte)) /* Just a page... */
1715 return pte_young(*pte);
1720 int kvm_age_hva(struct kvm *kvm, unsigned long start, unsigned long end)
1724 trace_kvm_age_hva(start, end);
1725 return handle_hva_to_gpa(kvm, start, end, kvm_age_hva_handler, NULL);
1728 int kvm_test_age_hva(struct kvm *kvm, unsigned long hva)
1732 trace_kvm_test_age_hva(hva);
1733 return handle_hva_to_gpa(kvm, hva, hva + PAGE_SIZE,
1734 kvm_test_age_hva_handler, NULL);
1737 void kvm_mmu_free_memory_caches(struct kvm_vcpu *vcpu)
1739 mmu_free_memory_cache(&vcpu->arch.mmu_page_cache);
1742 phys_addr_t kvm_mmu_get_httbr(void)
1744 if (__kvm_cpu_uses_extended_idmap())
1745 return virt_to_phys(merged_hyp_pgd);
1747 return virt_to_phys(hyp_pgd);
1750 phys_addr_t kvm_get_idmap_vector(void)
1752 return hyp_idmap_vector;
1755 static int kvm_map_idmap_text(pgd_t *pgd)
1759 /* Create the idmap in the boot page tables */
1760 err = __create_hyp_mappings(pgd,
1761 hyp_idmap_start, hyp_idmap_end,
1762 __phys_to_pfn(hyp_idmap_start),
1765 kvm_err("Failed to idmap %lx-%lx\n",
1766 hyp_idmap_start, hyp_idmap_end);
1771 int kvm_mmu_init(void)
1775 hyp_idmap_start = kvm_virt_to_phys(__hyp_idmap_text_start);
1776 hyp_idmap_end = kvm_virt_to_phys(__hyp_idmap_text_end);
1777 hyp_idmap_vector = kvm_virt_to_phys(__kvm_hyp_init);
1780 * We rely on the linker script to ensure at build time that the HYP
1781 * init code does not cross a page boundary.
1783 BUG_ON((hyp_idmap_start ^ (hyp_idmap_end - 1)) & PAGE_MASK);
1785 kvm_debug("IDMAP page: %lx\n", hyp_idmap_start);
1786 kvm_debug("HYP VA range: %lx:%lx\n",
1787 kern_hyp_va(PAGE_OFFSET), kern_hyp_va(~0UL));
1789 if (hyp_idmap_start >= kern_hyp_va(PAGE_OFFSET) &&
1790 hyp_idmap_start < kern_hyp_va(~0UL) &&
1791 hyp_idmap_start != (unsigned long)__hyp_idmap_text_start) {
1793 * The idmap page is intersecting with the VA space,
1794 * it is not safe to continue further.
1796 kvm_err("IDMAP intersecting with HYP VA, unable to continue\n");
1801 hyp_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, hyp_pgd_order);
1803 kvm_err("Hyp mode PGD not allocated\n");
1808 if (__kvm_cpu_uses_extended_idmap()) {
1809 boot_hyp_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
1811 if (!boot_hyp_pgd) {
1812 kvm_err("Hyp boot PGD not allocated\n");
1817 err = kvm_map_idmap_text(boot_hyp_pgd);
1821 merged_hyp_pgd = (pgd_t *)__get_free_page(GFP_KERNEL | __GFP_ZERO);
1822 if (!merged_hyp_pgd) {
1823 kvm_err("Failed to allocate extra HYP pgd\n");
1826 __kvm_extend_hypmap(boot_hyp_pgd, hyp_pgd, merged_hyp_pgd,
1829 err = kvm_map_idmap_text(hyp_pgd);
1840 void kvm_arch_commit_memory_region(struct kvm *kvm,
1841 const struct kvm_userspace_memory_region *mem,
1842 const struct kvm_memory_slot *old,
1843 const struct kvm_memory_slot *new,
1844 enum kvm_mr_change change)
1847 * At this point memslot has been committed and there is an
1848 * allocated dirty_bitmap[], dirty pages will be be tracked while the
1849 * memory slot is write protected.
1851 if (change != KVM_MR_DELETE && mem->flags & KVM_MEM_LOG_DIRTY_PAGES)
1852 kvm_mmu_wp_memory_region(kvm, mem->slot);
1855 int kvm_arch_prepare_memory_region(struct kvm *kvm,
1856 struct kvm_memory_slot *memslot,
1857 const struct kvm_userspace_memory_region *mem,
1858 enum kvm_mr_change change)
1860 hva_t hva = mem->userspace_addr;
1861 hva_t reg_end = hva + mem->memory_size;
1862 bool writable = !(mem->flags & KVM_MEM_READONLY);
1865 if (change != KVM_MR_CREATE && change != KVM_MR_MOVE &&
1866 change != KVM_MR_FLAGS_ONLY)
1870 * Prevent userspace from creating a memory region outside of the IPA
1871 * space addressable by the KVM guest IPA space.
1873 if (memslot->base_gfn + memslot->npages >
1874 (KVM_PHYS_SIZE >> PAGE_SHIFT))
1877 down_read(¤t->mm->mmap_sem);
1879 * A memory region could potentially cover multiple VMAs, and any holes
1880 * between them, so iterate over all of them to find out if we can map
1881 * any of them right now.
1883 * +--------------------------------------------+
1884 * +---------------+----------------+ +----------------+
1885 * | : VMA 1 | VMA 2 | | VMA 3 : |
1886 * +---------------+----------------+ +----------------+
1888 * +--------------------------------------------+
1891 struct vm_area_struct *vma = find_vma(current->mm, hva);
1892 hva_t vm_start, vm_end;
1894 if (!vma || vma->vm_start >= reg_end)
1898 * Mapping a read-only VMA is only allowed if the
1899 * memory region is configured as read-only.
1901 if (writable && !(vma->vm_flags & VM_WRITE)) {
1907 * Take the intersection of this VMA with the memory region
1909 vm_start = max(hva, vma->vm_start);
1910 vm_end = min(reg_end, vma->vm_end);
1912 if (vma->vm_flags & VM_PFNMAP) {
1913 gpa_t gpa = mem->guest_phys_addr +
1914 (vm_start - mem->userspace_addr);
1917 pa = (phys_addr_t)vma->vm_pgoff << PAGE_SHIFT;
1918 pa += vm_start - vma->vm_start;
1920 /* IO region dirty page logging not allowed */
1921 if (memslot->flags & KVM_MEM_LOG_DIRTY_PAGES) {
1926 ret = kvm_phys_addr_ioremap(kvm, gpa, pa,
1933 } while (hva < reg_end);
1935 if (change == KVM_MR_FLAGS_ONLY)
1938 spin_lock(&kvm->mmu_lock);
1940 unmap_stage2_range(kvm, mem->guest_phys_addr, mem->memory_size);
1942 stage2_flush_memslot(kvm, memslot);
1943 spin_unlock(&kvm->mmu_lock);
1945 up_read(¤t->mm->mmap_sem);
1949 void kvm_arch_free_memslot(struct kvm *kvm, struct kvm_memory_slot *free,
1950 struct kvm_memory_slot *dont)
1954 int kvm_arch_create_memslot(struct kvm *kvm, struct kvm_memory_slot *slot,
1955 unsigned long npages)
1960 void kvm_arch_memslots_updated(struct kvm *kvm, u64 gen)
1964 void kvm_arch_flush_shadow_all(struct kvm *kvm)
1966 kvm_free_stage2_pgd(kvm);
1969 void kvm_arch_flush_shadow_memslot(struct kvm *kvm,
1970 struct kvm_memory_slot *slot)
1972 gpa_t gpa = slot->base_gfn << PAGE_SHIFT;
1973 phys_addr_t size = slot->npages << PAGE_SHIFT;
1975 spin_lock(&kvm->mmu_lock);
1976 unmap_stage2_range(kvm, gpa, size);
1977 spin_unlock(&kvm->mmu_lock);
1981 * See note at ARMv7 ARM B1.14.4 (TL;DR: S/W ops are not easily virtualized).
1984 * - S/W ops are local to a CPU (not broadcast)
1985 * - We have line migration behind our back (speculation)
1986 * - System caches don't support S/W at all (damn!)
1988 * In the face of the above, the best we can do is to try and convert
1989 * S/W ops to VA ops. Because the guest is not allowed to infer the
1990 * S/W to PA mapping, it can only use S/W to nuke the whole cache,
1991 * which is a rather good thing for us.
1993 * Also, it is only used when turning caches on/off ("The expected
1994 * usage of the cache maintenance instructions that operate by set/way
1995 * is associated with the cache maintenance instructions associated
1996 * with the powerdown and powerup of caches, if this is required by
1997 * the implementation.").
1999 * We use the following policy:
2001 * - If we trap a S/W operation, we enable VM trapping to detect
2002 * caches being turned on/off, and do a full clean.
2004 * - We flush the caches on both caches being turned on and off.
2006 * - Once the caches are enabled, we stop trapping VM ops.
2008 void kvm_set_way_flush(struct kvm_vcpu *vcpu)
2010 unsigned long hcr = vcpu_get_hcr(vcpu);
2013 * If this is the first time we do a S/W operation
2014 * (i.e. HCR_TVM not set) flush the whole memory, and set the
2017 * Otherwise, rely on the VM trapping to wait for the MMU +
2018 * Caches to be turned off. At that point, we'll be able to
2019 * clean the caches again.
2021 if (!(hcr & HCR_TVM)) {
2022 trace_kvm_set_way_flush(*vcpu_pc(vcpu),
2023 vcpu_has_cache_enabled(vcpu));
2024 stage2_flush_vm(vcpu->kvm);
2025 vcpu_set_hcr(vcpu, hcr | HCR_TVM);
2029 void kvm_toggle_cache(struct kvm_vcpu *vcpu, bool was_enabled)
2031 bool now_enabled = vcpu_has_cache_enabled(vcpu);
2034 * If switching the MMU+caches on, need to invalidate the caches.
2035 * If switching it off, need to clean the caches.
2036 * Clean + invalidate does the trick always.
2038 if (now_enabled != was_enabled)
2039 stage2_flush_vm(vcpu->kvm);
2041 /* Caches are now on, stop trapping VM ops (until a S/W op) */
2043 vcpu_set_hcr(vcpu, vcpu_get_hcr(vcpu) & ~HCR_TVM);
2045 trace_kvm_toggle_cache(*vcpu_pc(vcpu), was_enabled, now_enabled);