1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 2012 - Virtual Open Systems and Columbia University
4 * Author: Christoffer Dall <c.dall@virtualopensystems.com>
7 #include <linux/mman.h>
8 #include <linux/kvm_host.h>
10 #include <linux/hugetlb.h>
11 #include <linux/sched/signal.h>
12 #include <trace/events/kvm.h>
13 #include <asm/pgalloc.h>
14 #include <asm/cacheflush.h>
15 #include <asm/kvm_arm.h>
16 #include <asm/kvm_mmu.h>
17 #include <asm/kvm_ras.h>
18 #include <asm/kvm_asm.h>
19 #include <asm/kvm_emulate.h>
24 static pgd_t *boot_hyp_pgd;
25 static pgd_t *hyp_pgd;
26 static pgd_t *merged_hyp_pgd;
27 static DEFINE_MUTEX(kvm_hyp_pgd_mutex);
29 static unsigned long hyp_idmap_start;
30 static unsigned long hyp_idmap_end;
31 static phys_addr_t hyp_idmap_vector;
33 static unsigned long io_map_base;
35 #define hyp_pgd_order get_order(PTRS_PER_PGD * sizeof(pgd_t))
37 #define KVM_S2PTE_FLAG_IS_IOMAP (1UL << 0)
38 #define KVM_S2_FLAG_LOGGING_ACTIVE (1UL << 1)
40 static bool is_iomap(unsigned long flags)
42 return flags & KVM_S2PTE_FLAG_IS_IOMAP;
45 static bool memslot_is_logging(struct kvm_memory_slot *memslot)
47 return memslot->dirty_bitmap && !(memslot->flags & KVM_MEM_READONLY);
51 * kvm_flush_remote_tlbs() - flush all VM TLB entries for v7/8
52 * @kvm: pointer to kvm structure.
54 * Interface to HYP function to flush all VM TLB entries
56 void kvm_flush_remote_tlbs(struct kvm *kvm)
58 kvm_call_hyp(__kvm_tlb_flush_vmid, kvm);
61 static void kvm_tlb_flush_vmid_ipa(struct kvm *kvm, phys_addr_t ipa)
63 kvm_call_hyp(__kvm_tlb_flush_vmid_ipa, kvm, ipa);
67 * D-Cache management functions. They take the page table entries by
68 * value, as they are flushing the cache using the kernel mapping (or
71 static void kvm_flush_dcache_pte(pte_t pte)
73 __kvm_flush_dcache_pte(pte);
76 static void kvm_flush_dcache_pmd(pmd_t pmd)
78 __kvm_flush_dcache_pmd(pmd);
81 static void kvm_flush_dcache_pud(pud_t pud)
83 __kvm_flush_dcache_pud(pud);
86 static bool kvm_is_device_pfn(unsigned long pfn)
88 return !pfn_valid(pfn);
92 * stage2_dissolve_pmd() - clear and flush huge PMD entry
93 * @kvm: pointer to kvm structure.
95 * @pmd: pmd pointer for IPA
97 * Function clears a PMD entry, flushes addr 1st and 2nd stage TLBs.
99 static void stage2_dissolve_pmd(struct kvm *kvm, phys_addr_t addr, pmd_t *pmd)
101 if (!pmd_thp_or_huge(*pmd))
105 kvm_tlb_flush_vmid_ipa(kvm, addr);
106 put_page(virt_to_page(pmd));
110 * stage2_dissolve_pud() - clear and flush huge PUD entry
111 * @kvm: pointer to kvm structure.
113 * @pud: pud pointer for IPA
115 * Function clears a PUD entry, flushes addr 1st and 2nd stage TLBs.
117 static void stage2_dissolve_pud(struct kvm *kvm, phys_addr_t addr, pud_t *pudp)
119 if (!stage2_pud_huge(kvm, *pudp))
122 stage2_pud_clear(kvm, pudp);
123 kvm_tlb_flush_vmid_ipa(kvm, addr);
124 put_page(virt_to_page(pudp));
127 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
132 BUG_ON(max > KVM_NR_MEM_OBJS);
133 if (cache->nobjs >= min)
135 while (cache->nobjs < max) {
136 page = (void *)__get_free_page(GFP_PGTABLE_USER);
139 cache->objects[cache->nobjs++] = page;
144 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
147 free_page((unsigned long)mc->objects[--mc->nobjs]);
150 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc)
154 BUG_ON(!mc || !mc->nobjs);
155 p = mc->objects[--mc->nobjs];
159 static void clear_stage2_pgd_entry(struct kvm *kvm, pgd_t *pgd, phys_addr_t addr)
161 pud_t *pud_table __maybe_unused = stage2_pud_offset(kvm, pgd, 0UL);
162 stage2_pgd_clear(kvm, pgd);
163 kvm_tlb_flush_vmid_ipa(kvm, addr);
164 stage2_pud_free(kvm, pud_table);
165 put_page(virt_to_page(pgd));
168 static void clear_stage2_pud_entry(struct kvm *kvm, pud_t *pud, phys_addr_t addr)
170 pmd_t *pmd_table __maybe_unused = stage2_pmd_offset(kvm, pud, 0);
171 VM_BUG_ON(stage2_pud_huge(kvm, *pud));
172 stage2_pud_clear(kvm, pud);
173 kvm_tlb_flush_vmid_ipa(kvm, addr);
174 stage2_pmd_free(kvm, pmd_table);
175 put_page(virt_to_page(pud));
178 static void clear_stage2_pmd_entry(struct kvm *kvm, pmd_t *pmd, phys_addr_t addr)
180 pte_t *pte_table = pte_offset_kernel(pmd, 0);
181 VM_BUG_ON(pmd_thp_or_huge(*pmd));
183 kvm_tlb_flush_vmid_ipa(kvm, addr);
184 free_page((unsigned long)pte_table);
185 put_page(virt_to_page(pmd));
188 static inline void kvm_set_pte(pte_t *ptep, pte_t new_pte)
190 WRITE_ONCE(*ptep, new_pte);
194 static inline void kvm_set_pmd(pmd_t *pmdp, pmd_t new_pmd)
196 WRITE_ONCE(*pmdp, new_pmd);
200 static inline void kvm_pmd_populate(pmd_t *pmdp, pte_t *ptep)
202 kvm_set_pmd(pmdp, kvm_mk_pmd(ptep));
205 static inline void kvm_pud_populate(pud_t *pudp, pmd_t *pmdp)
207 WRITE_ONCE(*pudp, kvm_mk_pud(pmdp));
211 static inline void kvm_pgd_populate(pgd_t *pgdp, pud_t *pudp)
213 WRITE_ONCE(*pgdp, kvm_mk_pgd(pudp));
218 * Unmapping vs dcache management:
220 * If a guest maps certain memory pages as uncached, all writes will
221 * bypass the data cache and go directly to RAM. However, the CPUs
222 * can still speculate reads (not writes) and fill cache lines with
225 * Those cache lines will be *clean* cache lines though, so a
226 * clean+invalidate operation is equivalent to an invalidate
227 * operation, because no cache lines are marked dirty.
229 * Those clean cache lines could be filled prior to an uncached write
230 * by the guest, and the cache coherent IO subsystem would therefore
231 * end up writing old data to disk.
233 * This is why right after unmapping a page/section and invalidating
234 * the corresponding TLBs, we call kvm_flush_dcache_p*() to make sure
235 * the IO subsystem will never hit in the cache.
237 * This is all avoided on systems that have ARM64_HAS_STAGE2_FWB, as
238 * we then fully enforce cacheability of RAM, no matter what the guest
241 static void unmap_stage2_ptes(struct kvm *kvm, pmd_t *pmd,
242 phys_addr_t addr, phys_addr_t end)
244 phys_addr_t start_addr = addr;
245 pte_t *pte, *start_pte;
247 start_pte = pte = pte_offset_kernel(pmd, addr);
249 if (!pte_none(*pte)) {
250 pte_t old_pte = *pte;
252 kvm_set_pte(pte, __pte(0));
253 kvm_tlb_flush_vmid_ipa(kvm, addr);
255 /* No need to invalidate the cache for device mappings */
256 if (!kvm_is_device_pfn(pte_pfn(old_pte)))
257 kvm_flush_dcache_pte(old_pte);
259 put_page(virt_to_page(pte));
261 } while (pte++, addr += PAGE_SIZE, addr != end);
263 if (stage2_pte_table_empty(kvm, start_pte))
264 clear_stage2_pmd_entry(kvm, pmd, start_addr);
267 static void unmap_stage2_pmds(struct kvm *kvm, pud_t *pud,
268 phys_addr_t addr, phys_addr_t end)
270 phys_addr_t next, start_addr = addr;
271 pmd_t *pmd, *start_pmd;
273 start_pmd = pmd = stage2_pmd_offset(kvm, pud, addr);
275 next = stage2_pmd_addr_end(kvm, addr, end);
276 if (!pmd_none(*pmd)) {
277 if (pmd_thp_or_huge(*pmd)) {
278 pmd_t old_pmd = *pmd;
281 kvm_tlb_flush_vmid_ipa(kvm, addr);
283 kvm_flush_dcache_pmd(old_pmd);
285 put_page(virt_to_page(pmd));
287 unmap_stage2_ptes(kvm, pmd, addr, next);
290 } while (pmd++, addr = next, addr != end);
292 if (stage2_pmd_table_empty(kvm, start_pmd))
293 clear_stage2_pud_entry(kvm, pud, start_addr);
296 static void unmap_stage2_puds(struct kvm *kvm, pgd_t *pgd,
297 phys_addr_t addr, phys_addr_t end)
299 phys_addr_t next, start_addr = addr;
300 pud_t *pud, *start_pud;
302 start_pud = pud = stage2_pud_offset(kvm, pgd, addr);
304 next = stage2_pud_addr_end(kvm, addr, end);
305 if (!stage2_pud_none(kvm, *pud)) {
306 if (stage2_pud_huge(kvm, *pud)) {
307 pud_t old_pud = *pud;
309 stage2_pud_clear(kvm, pud);
310 kvm_tlb_flush_vmid_ipa(kvm, addr);
311 kvm_flush_dcache_pud(old_pud);
312 put_page(virt_to_page(pud));
314 unmap_stage2_pmds(kvm, pud, addr, next);
317 } while (pud++, addr = next, addr != end);
319 if (stage2_pud_table_empty(kvm, start_pud))
320 clear_stage2_pgd_entry(kvm, pgd, start_addr);
324 * unmap_stage2_range -- Clear stage2 page table entries to unmap a range
325 * @kvm: The VM pointer
326 * @start: The intermediate physical base address of the range to unmap
327 * @size: The size of the area to unmap
329 * Clear a range of stage-2 mappings, lowering the various ref-counts. Must
330 * be called while holding mmu_lock (unless for freeing the stage2 pgd before
331 * destroying the VM), otherwise another faulting VCPU may come in and mess
332 * with things behind our backs.
334 static void unmap_stage2_range(struct kvm *kvm, phys_addr_t start, u64 size)
337 phys_addr_t addr = start, end = start + size;
340 assert_spin_locked(&kvm->mmu_lock);
341 WARN_ON(size & ~PAGE_MASK);
343 pgd = kvm->arch.pgd + stage2_pgd_index(kvm, addr);
346 * Make sure the page table is still active, as another thread
347 * could have possibly freed the page table, while we released
350 if (!READ_ONCE(kvm->arch.pgd))
352 next = stage2_pgd_addr_end(kvm, addr, end);
353 if (!stage2_pgd_none(kvm, *pgd))
354 unmap_stage2_puds(kvm, pgd, addr, next);
356 * If the range is too large, release the kvm->mmu_lock
357 * to prevent starvation and lockup detector warnings.
360 cond_resched_lock(&kvm->mmu_lock);
361 } while (pgd++, addr = next, addr != end);
364 static void stage2_flush_ptes(struct kvm *kvm, pmd_t *pmd,
365 phys_addr_t addr, phys_addr_t end)
369 pte = pte_offset_kernel(pmd, addr);
371 if (!pte_none(*pte) && !kvm_is_device_pfn(pte_pfn(*pte)))
372 kvm_flush_dcache_pte(*pte);
373 } while (pte++, addr += PAGE_SIZE, addr != end);
376 static void stage2_flush_pmds(struct kvm *kvm, pud_t *pud,
377 phys_addr_t addr, phys_addr_t end)
382 pmd = stage2_pmd_offset(kvm, pud, addr);
384 next = stage2_pmd_addr_end(kvm, addr, end);
385 if (!pmd_none(*pmd)) {
386 if (pmd_thp_or_huge(*pmd))
387 kvm_flush_dcache_pmd(*pmd);
389 stage2_flush_ptes(kvm, pmd, addr, next);
391 } while (pmd++, addr = next, addr != end);
394 static void stage2_flush_puds(struct kvm *kvm, pgd_t *pgd,
395 phys_addr_t addr, phys_addr_t end)
400 pud = stage2_pud_offset(kvm, pgd, addr);
402 next = stage2_pud_addr_end(kvm, addr, end);
403 if (!stage2_pud_none(kvm, *pud)) {
404 if (stage2_pud_huge(kvm, *pud))
405 kvm_flush_dcache_pud(*pud);
407 stage2_flush_pmds(kvm, pud, addr, next);
409 } while (pud++, addr = next, addr != end);
412 static void stage2_flush_memslot(struct kvm *kvm,
413 struct kvm_memory_slot *memslot)
415 phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
416 phys_addr_t end = addr + PAGE_SIZE * memslot->npages;
420 pgd = kvm->arch.pgd + stage2_pgd_index(kvm, addr);
422 next = stage2_pgd_addr_end(kvm, addr, end);
423 if (!stage2_pgd_none(kvm, *pgd))
424 stage2_flush_puds(kvm, pgd, addr, next);
425 } while (pgd++, addr = next, addr != end);
429 * stage2_flush_vm - Invalidate cache for pages mapped in stage 2
430 * @kvm: The struct kvm pointer
432 * Go through the stage 2 page tables and invalidate any cache lines
433 * backing memory already mapped to the VM.
435 static void stage2_flush_vm(struct kvm *kvm)
437 struct kvm_memslots *slots;
438 struct kvm_memory_slot *memslot;
441 idx = srcu_read_lock(&kvm->srcu);
442 spin_lock(&kvm->mmu_lock);
444 slots = kvm_memslots(kvm);
445 kvm_for_each_memslot(memslot, slots)
446 stage2_flush_memslot(kvm, memslot);
448 spin_unlock(&kvm->mmu_lock);
449 srcu_read_unlock(&kvm->srcu, idx);
452 static void clear_hyp_pgd_entry(pgd_t *pgd)
454 pud_t *pud_table __maybe_unused = pud_offset(pgd, 0UL);
456 pud_free(NULL, pud_table);
457 put_page(virt_to_page(pgd));
460 static void clear_hyp_pud_entry(pud_t *pud)
462 pmd_t *pmd_table __maybe_unused = pmd_offset(pud, 0);
463 VM_BUG_ON(pud_huge(*pud));
465 pmd_free(NULL, pmd_table);
466 put_page(virt_to_page(pud));
469 static void clear_hyp_pmd_entry(pmd_t *pmd)
471 pte_t *pte_table = pte_offset_kernel(pmd, 0);
472 VM_BUG_ON(pmd_thp_or_huge(*pmd));
474 pte_free_kernel(NULL, pte_table);
475 put_page(virt_to_page(pmd));
478 static void unmap_hyp_ptes(pmd_t *pmd, phys_addr_t addr, phys_addr_t end)
480 pte_t *pte, *start_pte;
482 start_pte = pte = pte_offset_kernel(pmd, addr);
484 if (!pte_none(*pte)) {
485 kvm_set_pte(pte, __pte(0));
486 put_page(virt_to_page(pte));
488 } while (pte++, addr += PAGE_SIZE, addr != end);
490 if (hyp_pte_table_empty(start_pte))
491 clear_hyp_pmd_entry(pmd);
494 static void unmap_hyp_pmds(pud_t *pud, phys_addr_t addr, phys_addr_t end)
497 pmd_t *pmd, *start_pmd;
499 start_pmd = pmd = pmd_offset(pud, addr);
501 next = pmd_addr_end(addr, end);
502 /* Hyp doesn't use huge pmds */
504 unmap_hyp_ptes(pmd, addr, next);
505 } while (pmd++, addr = next, addr != end);
507 if (hyp_pmd_table_empty(start_pmd))
508 clear_hyp_pud_entry(pud);
511 static void unmap_hyp_puds(pgd_t *pgd, phys_addr_t addr, phys_addr_t end)
514 pud_t *pud, *start_pud;
516 start_pud = pud = pud_offset(pgd, addr);
518 next = pud_addr_end(addr, end);
519 /* Hyp doesn't use huge puds */
521 unmap_hyp_pmds(pud, addr, next);
522 } while (pud++, addr = next, addr != end);
524 if (hyp_pud_table_empty(start_pud))
525 clear_hyp_pgd_entry(pgd);
528 static unsigned int kvm_pgd_index(unsigned long addr, unsigned int ptrs_per_pgd)
530 return (addr >> PGDIR_SHIFT) & (ptrs_per_pgd - 1);
533 static void __unmap_hyp_range(pgd_t *pgdp, unsigned long ptrs_per_pgd,
534 phys_addr_t start, u64 size)
537 phys_addr_t addr = start, end = start + size;
541 * We don't unmap anything from HYP, except at the hyp tear down.
542 * Hence, we don't have to invalidate the TLBs here.
544 pgd = pgdp + kvm_pgd_index(addr, ptrs_per_pgd);
546 next = pgd_addr_end(addr, end);
548 unmap_hyp_puds(pgd, addr, next);
549 } while (pgd++, addr = next, addr != end);
552 static void unmap_hyp_range(pgd_t *pgdp, phys_addr_t start, u64 size)
554 __unmap_hyp_range(pgdp, PTRS_PER_PGD, start, size);
557 static void unmap_hyp_idmap_range(pgd_t *pgdp, phys_addr_t start, u64 size)
559 __unmap_hyp_range(pgdp, __kvm_idmap_ptrs_per_pgd(), start, size);
563 * free_hyp_pgds - free Hyp-mode page tables
565 * Assumes hyp_pgd is a page table used strictly in Hyp-mode and
566 * therefore contains either mappings in the kernel memory area (above
567 * PAGE_OFFSET), or device mappings in the idmap range.
569 * boot_hyp_pgd should only map the idmap range, and is only used in
570 * the extended idmap case.
572 void free_hyp_pgds(void)
576 mutex_lock(&kvm_hyp_pgd_mutex);
578 id_pgd = boot_hyp_pgd ? boot_hyp_pgd : hyp_pgd;
581 /* In case we never called hyp_mmu_init() */
583 io_map_base = hyp_idmap_start;
584 unmap_hyp_idmap_range(id_pgd, io_map_base,
585 hyp_idmap_start + PAGE_SIZE - io_map_base);
589 free_pages((unsigned long)boot_hyp_pgd, hyp_pgd_order);
594 unmap_hyp_range(hyp_pgd, kern_hyp_va(PAGE_OFFSET),
595 (uintptr_t)high_memory - PAGE_OFFSET);
597 free_pages((unsigned long)hyp_pgd, hyp_pgd_order);
600 if (merged_hyp_pgd) {
601 clear_page(merged_hyp_pgd);
602 free_page((unsigned long)merged_hyp_pgd);
603 merged_hyp_pgd = NULL;
606 mutex_unlock(&kvm_hyp_pgd_mutex);
609 static void create_hyp_pte_mappings(pmd_t *pmd, unsigned long start,
610 unsigned long end, unsigned long pfn,
618 pte = pte_offset_kernel(pmd, addr);
619 kvm_set_pte(pte, kvm_pfn_pte(pfn, prot));
620 get_page(virt_to_page(pte));
622 } while (addr += PAGE_SIZE, addr != end);
625 static int create_hyp_pmd_mappings(pud_t *pud, unsigned long start,
626 unsigned long end, unsigned long pfn,
631 unsigned long addr, next;
635 pmd = pmd_offset(pud, addr);
637 BUG_ON(pmd_sect(*pmd));
639 if (pmd_none(*pmd)) {
640 pte = pte_alloc_one_kernel(NULL);
642 kvm_err("Cannot allocate Hyp pte\n");
645 kvm_pmd_populate(pmd, pte);
646 get_page(virt_to_page(pmd));
649 next = pmd_addr_end(addr, end);
651 create_hyp_pte_mappings(pmd, addr, next, pfn, prot);
652 pfn += (next - addr) >> PAGE_SHIFT;
653 } while (addr = next, addr != end);
658 static int create_hyp_pud_mappings(pgd_t *pgd, unsigned long start,
659 unsigned long end, unsigned long pfn,
664 unsigned long addr, next;
669 pud = pud_offset(pgd, addr);
671 if (pud_none_or_clear_bad(pud)) {
672 pmd = pmd_alloc_one(NULL, addr);
674 kvm_err("Cannot allocate Hyp pmd\n");
677 kvm_pud_populate(pud, pmd);
678 get_page(virt_to_page(pud));
681 next = pud_addr_end(addr, end);
682 ret = create_hyp_pmd_mappings(pud, addr, next, pfn, prot);
685 pfn += (next - addr) >> PAGE_SHIFT;
686 } while (addr = next, addr != end);
691 static int __create_hyp_mappings(pgd_t *pgdp, unsigned long ptrs_per_pgd,
692 unsigned long start, unsigned long end,
693 unsigned long pfn, pgprot_t prot)
697 unsigned long addr, next;
700 mutex_lock(&kvm_hyp_pgd_mutex);
701 addr = start & PAGE_MASK;
702 end = PAGE_ALIGN(end);
704 pgd = pgdp + kvm_pgd_index(addr, ptrs_per_pgd);
706 if (pgd_none(*pgd)) {
707 pud = pud_alloc_one(NULL, addr);
709 kvm_err("Cannot allocate Hyp pud\n");
713 kvm_pgd_populate(pgd, pud);
714 get_page(virt_to_page(pgd));
717 next = pgd_addr_end(addr, end);
718 err = create_hyp_pud_mappings(pgd, addr, next, pfn, prot);
721 pfn += (next - addr) >> PAGE_SHIFT;
722 } while (addr = next, addr != end);
724 mutex_unlock(&kvm_hyp_pgd_mutex);
728 static phys_addr_t kvm_kaddr_to_phys(void *kaddr)
730 if (!is_vmalloc_addr(kaddr)) {
731 BUG_ON(!virt_addr_valid(kaddr));
734 return page_to_phys(vmalloc_to_page(kaddr)) +
735 offset_in_page(kaddr);
740 * create_hyp_mappings - duplicate a kernel virtual address range in Hyp mode
741 * @from: The virtual kernel start address of the range
742 * @to: The virtual kernel end address of the range (exclusive)
743 * @prot: The protection to be applied to this range
745 * The same virtual address as the kernel virtual address is also used
746 * in Hyp-mode mapping (modulo HYP_PAGE_OFFSET) to the same underlying
749 int create_hyp_mappings(void *from, void *to, pgprot_t prot)
751 phys_addr_t phys_addr;
752 unsigned long virt_addr;
753 unsigned long start = kern_hyp_va((unsigned long)from);
754 unsigned long end = kern_hyp_va((unsigned long)to);
756 if (is_kernel_in_hyp_mode())
759 start = start & PAGE_MASK;
760 end = PAGE_ALIGN(end);
762 for (virt_addr = start; virt_addr < end; virt_addr += PAGE_SIZE) {
765 phys_addr = kvm_kaddr_to_phys(from + virt_addr - start);
766 err = __create_hyp_mappings(hyp_pgd, PTRS_PER_PGD,
767 virt_addr, virt_addr + PAGE_SIZE,
768 __phys_to_pfn(phys_addr),
777 static int __create_hyp_private_mapping(phys_addr_t phys_addr, size_t size,
778 unsigned long *haddr, pgprot_t prot)
780 pgd_t *pgd = hyp_pgd;
784 mutex_lock(&kvm_hyp_pgd_mutex);
787 * This assumes that we we have enough space below the idmap
788 * page to allocate our VAs. If not, the check below will
789 * kick. A potential alternative would be to detect that
790 * overflow and switch to an allocation above the idmap.
792 * The allocated size is always a multiple of PAGE_SIZE.
794 size = PAGE_ALIGN(size + offset_in_page(phys_addr));
795 base = io_map_base - size;
798 * Verify that BIT(VA_BITS - 1) hasn't been flipped by
799 * allocating the new area, as it would indicate we've
800 * overflowed the idmap/IO address range.
802 if ((base ^ io_map_base) & BIT(VA_BITS - 1))
807 mutex_unlock(&kvm_hyp_pgd_mutex);
812 if (__kvm_cpu_uses_extended_idmap())
815 ret = __create_hyp_mappings(pgd, __kvm_idmap_ptrs_per_pgd(),
817 __phys_to_pfn(phys_addr), prot);
821 *haddr = base + offset_in_page(phys_addr);
828 * create_hyp_io_mappings - Map IO into both kernel and HYP
829 * @phys_addr: The physical start address which gets mapped
830 * @size: Size of the region being mapped
831 * @kaddr: Kernel VA for this mapping
832 * @haddr: HYP VA for this mapping
834 int create_hyp_io_mappings(phys_addr_t phys_addr, size_t size,
835 void __iomem **kaddr,
836 void __iomem **haddr)
841 *kaddr = ioremap(phys_addr, size);
845 if (is_kernel_in_hyp_mode()) {
850 ret = __create_hyp_private_mapping(phys_addr, size,
851 &addr, PAGE_HYP_DEVICE);
859 *haddr = (void __iomem *)addr;
864 * create_hyp_exec_mappings - Map an executable range into HYP
865 * @phys_addr: The physical start address which gets mapped
866 * @size: Size of the region being mapped
867 * @haddr: HYP VA for this mapping
869 int create_hyp_exec_mappings(phys_addr_t phys_addr, size_t size,
875 BUG_ON(is_kernel_in_hyp_mode());
877 ret = __create_hyp_private_mapping(phys_addr, size,
878 &addr, PAGE_HYP_EXEC);
884 *haddr = (void *)addr;
889 * kvm_alloc_stage2_pgd - allocate level-1 table for stage-2 translation.
890 * @kvm: The KVM struct pointer for the VM.
892 * Allocates only the stage-2 HW PGD level table(s) of size defined by
893 * stage2_pgd_size(kvm).
895 * Note we don't need locking here as this is only called when the VM is
896 * created, which can only be done once.
898 int kvm_alloc_stage2_pgd(struct kvm *kvm)
900 phys_addr_t pgd_phys;
903 if (kvm->arch.pgd != NULL) {
904 kvm_err("kvm_arch already initialized?\n");
908 /* Allocate the HW PGD, making sure that each page gets its own refcount */
909 pgd = alloc_pages_exact(stage2_pgd_size(kvm), GFP_KERNEL | __GFP_ZERO);
913 pgd_phys = virt_to_phys(pgd);
914 if (WARN_ON(pgd_phys & ~kvm_vttbr_baddr_mask(kvm)))
918 kvm->arch.pgd_phys = pgd_phys;
922 static void stage2_unmap_memslot(struct kvm *kvm,
923 struct kvm_memory_slot *memslot)
925 hva_t hva = memslot->userspace_addr;
926 phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
927 phys_addr_t size = PAGE_SIZE * memslot->npages;
928 hva_t reg_end = hva + size;
931 * A memory region could potentially cover multiple VMAs, and any holes
932 * between them, so iterate over all of them to find out if we should
935 * +--------------------------------------------+
936 * +---------------+----------------+ +----------------+
937 * | : VMA 1 | VMA 2 | | VMA 3 : |
938 * +---------------+----------------+ +----------------+
940 * +--------------------------------------------+
943 struct vm_area_struct *vma = find_vma(current->mm, hva);
944 hva_t vm_start, vm_end;
946 if (!vma || vma->vm_start >= reg_end)
950 * Take the intersection of this VMA with the memory region
952 vm_start = max(hva, vma->vm_start);
953 vm_end = min(reg_end, vma->vm_end);
955 if (!(vma->vm_flags & VM_PFNMAP)) {
956 gpa_t gpa = addr + (vm_start - memslot->userspace_addr);
957 unmap_stage2_range(kvm, gpa, vm_end - vm_start);
960 } while (hva < reg_end);
964 * stage2_unmap_vm - Unmap Stage-2 RAM mappings
965 * @kvm: The struct kvm pointer
967 * Go through the memregions and unmap any reguler RAM
968 * backing memory already mapped to the VM.
970 void stage2_unmap_vm(struct kvm *kvm)
972 struct kvm_memslots *slots;
973 struct kvm_memory_slot *memslot;
976 idx = srcu_read_lock(&kvm->srcu);
977 down_read(¤t->mm->mmap_sem);
978 spin_lock(&kvm->mmu_lock);
980 slots = kvm_memslots(kvm);
981 kvm_for_each_memslot(memslot, slots)
982 stage2_unmap_memslot(kvm, memslot);
984 spin_unlock(&kvm->mmu_lock);
985 up_read(¤t->mm->mmap_sem);
986 srcu_read_unlock(&kvm->srcu, idx);
990 * kvm_free_stage2_pgd - free all stage-2 tables
991 * @kvm: The KVM struct pointer for the VM.
993 * Walks the level-1 page table pointed to by kvm->arch.pgd and frees all
994 * underlying level-2 and level-3 tables before freeing the actual level-1 table
995 * and setting the struct pointer to NULL.
997 void kvm_free_stage2_pgd(struct kvm *kvm)
1001 spin_lock(&kvm->mmu_lock);
1002 if (kvm->arch.pgd) {
1003 unmap_stage2_range(kvm, 0, kvm_phys_size(kvm));
1004 pgd = READ_ONCE(kvm->arch.pgd);
1005 kvm->arch.pgd = NULL;
1006 kvm->arch.pgd_phys = 0;
1008 spin_unlock(&kvm->mmu_lock);
1010 /* Free the HW pgd, one page at a time */
1012 free_pages_exact(pgd, stage2_pgd_size(kvm));
1015 static pud_t *stage2_get_pud(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
1021 pgd = kvm->arch.pgd + stage2_pgd_index(kvm, addr);
1022 if (stage2_pgd_none(kvm, *pgd)) {
1025 pud = mmu_memory_cache_alloc(cache);
1026 stage2_pgd_populate(kvm, pgd, pud);
1027 get_page(virt_to_page(pgd));
1030 return stage2_pud_offset(kvm, pgd, addr);
1033 static pmd_t *stage2_get_pmd(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
1039 pud = stage2_get_pud(kvm, cache, addr);
1040 if (!pud || stage2_pud_huge(kvm, *pud))
1043 if (stage2_pud_none(kvm, *pud)) {
1046 pmd = mmu_memory_cache_alloc(cache);
1047 stage2_pud_populate(kvm, pud, pmd);
1048 get_page(virt_to_page(pud));
1051 return stage2_pmd_offset(kvm, pud, addr);
1054 static int stage2_set_pmd_huge(struct kvm *kvm, struct kvm_mmu_memory_cache
1055 *cache, phys_addr_t addr, const pmd_t *new_pmd)
1057 pmd_t *pmd, old_pmd;
1060 pmd = stage2_get_pmd(kvm, cache, addr);
1065 * Multiple vcpus faulting on the same PMD entry, can
1066 * lead to them sequentially updating the PMD with the
1067 * same value. Following the break-before-make
1068 * (pmd_clear() followed by tlb_flush()) process can
1069 * hinder forward progress due to refaults generated
1070 * on missing translations.
1072 * Skip updating the page table if the entry is
1075 if (pmd_val(old_pmd) == pmd_val(*new_pmd))
1078 if (pmd_present(old_pmd)) {
1080 * If we already have PTE level mapping for this block,
1081 * we must unmap it to avoid inconsistent TLB state and
1082 * leaking the table page. We could end up in this situation
1083 * if the memory slot was marked for dirty logging and was
1084 * reverted, leaving PTE level mappings for the pages accessed
1085 * during the period. So, unmap the PTE level mapping for this
1086 * block and retry, as we could have released the upper level
1087 * table in the process.
1089 * Normal THP split/merge follows mmu_notifier callbacks and do
1090 * get handled accordingly.
1092 if (!pmd_thp_or_huge(old_pmd)) {
1093 unmap_stage2_range(kvm, addr & S2_PMD_MASK, S2_PMD_SIZE);
1097 * Mapping in huge pages should only happen through a
1098 * fault. If a page is merged into a transparent huge
1099 * page, the individual subpages of that huge page
1100 * should be unmapped through MMU notifiers before we
1103 * Merging of CompoundPages is not supported; they
1104 * should become splitting first, unmapped, merged,
1105 * and mapped back in on-demand.
1107 WARN_ON_ONCE(pmd_pfn(old_pmd) != pmd_pfn(*new_pmd));
1109 kvm_tlb_flush_vmid_ipa(kvm, addr);
1111 get_page(virt_to_page(pmd));
1114 kvm_set_pmd(pmd, *new_pmd);
1118 static int stage2_set_pud_huge(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
1119 phys_addr_t addr, const pud_t *new_pudp)
1121 pud_t *pudp, old_pud;
1124 pudp = stage2_get_pud(kvm, cache, addr);
1130 * A large number of vcpus faulting on the same stage 2 entry,
1131 * can lead to a refault due to the stage2_pud_clear()/tlb_flush().
1132 * Skip updating the page tables if there is no change.
1134 if (pud_val(old_pud) == pud_val(*new_pudp))
1137 if (stage2_pud_present(kvm, old_pud)) {
1139 * If we already have table level mapping for this block, unmap
1140 * the range for this block and retry.
1142 if (!stage2_pud_huge(kvm, old_pud)) {
1143 unmap_stage2_range(kvm, addr & S2_PUD_MASK, S2_PUD_SIZE);
1147 WARN_ON_ONCE(kvm_pud_pfn(old_pud) != kvm_pud_pfn(*new_pudp));
1148 stage2_pud_clear(kvm, pudp);
1149 kvm_tlb_flush_vmid_ipa(kvm, addr);
1151 get_page(virt_to_page(pudp));
1154 kvm_set_pud(pudp, *new_pudp);
1159 * stage2_get_leaf_entry - walk the stage2 VM page tables and return
1160 * true if a valid and present leaf-entry is found. A pointer to the
1161 * leaf-entry is returned in the appropriate level variable - pudpp,
1164 static bool stage2_get_leaf_entry(struct kvm *kvm, phys_addr_t addr,
1165 pud_t **pudpp, pmd_t **pmdpp, pte_t **ptepp)
1175 pudp = stage2_get_pud(kvm, NULL, addr);
1176 if (!pudp || stage2_pud_none(kvm, *pudp) || !stage2_pud_present(kvm, *pudp))
1179 if (stage2_pud_huge(kvm, *pudp)) {
1184 pmdp = stage2_pmd_offset(kvm, pudp, addr);
1185 if (!pmdp || pmd_none(*pmdp) || !pmd_present(*pmdp))
1188 if (pmd_thp_or_huge(*pmdp)) {
1193 ptep = pte_offset_kernel(pmdp, addr);
1194 if (!ptep || pte_none(*ptep) || !pte_present(*ptep))
1201 static bool stage2_is_exec(struct kvm *kvm, phys_addr_t addr)
1208 found = stage2_get_leaf_entry(kvm, addr, &pudp, &pmdp, &ptep);
1213 return kvm_s2pud_exec(pudp);
1215 return kvm_s2pmd_exec(pmdp);
1217 return kvm_s2pte_exec(ptep);
1220 static int stage2_set_pte(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
1221 phys_addr_t addr, const pte_t *new_pte,
1222 unsigned long flags)
1226 pte_t *pte, old_pte;
1227 bool iomap = flags & KVM_S2PTE_FLAG_IS_IOMAP;
1228 bool logging_active = flags & KVM_S2_FLAG_LOGGING_ACTIVE;
1230 VM_BUG_ON(logging_active && !cache);
1232 /* Create stage-2 page table mapping - Levels 0 and 1 */
1233 pud = stage2_get_pud(kvm, cache, addr);
1236 * Ignore calls from kvm_set_spte_hva for unallocated
1243 * While dirty page logging - dissolve huge PUD, then continue
1244 * on to allocate page.
1247 stage2_dissolve_pud(kvm, addr, pud);
1249 if (stage2_pud_none(kvm, *pud)) {
1251 return 0; /* ignore calls from kvm_set_spte_hva */
1252 pmd = mmu_memory_cache_alloc(cache);
1253 stage2_pud_populate(kvm, pud, pmd);
1254 get_page(virt_to_page(pud));
1257 pmd = stage2_pmd_offset(kvm, pud, addr);
1260 * Ignore calls from kvm_set_spte_hva for unallocated
1267 * While dirty page logging - dissolve huge PMD, then continue on to
1271 stage2_dissolve_pmd(kvm, addr, pmd);
1273 /* Create stage-2 page mappings - Level 2 */
1274 if (pmd_none(*pmd)) {
1276 return 0; /* ignore calls from kvm_set_spte_hva */
1277 pte = mmu_memory_cache_alloc(cache);
1278 kvm_pmd_populate(pmd, pte);
1279 get_page(virt_to_page(pmd));
1282 pte = pte_offset_kernel(pmd, addr);
1284 if (iomap && pte_present(*pte))
1287 /* Create 2nd stage page table mapping - Level 3 */
1289 if (pte_present(old_pte)) {
1290 /* Skip page table update if there is no change */
1291 if (pte_val(old_pte) == pte_val(*new_pte))
1294 kvm_set_pte(pte, __pte(0));
1295 kvm_tlb_flush_vmid_ipa(kvm, addr);
1297 get_page(virt_to_page(pte));
1300 kvm_set_pte(pte, *new_pte);
1304 #ifndef __HAVE_ARCH_PTEP_TEST_AND_CLEAR_YOUNG
1305 static int stage2_ptep_test_and_clear_young(pte_t *pte)
1307 if (pte_young(*pte)) {
1308 *pte = pte_mkold(*pte);
1314 static int stage2_ptep_test_and_clear_young(pte_t *pte)
1316 return __ptep_test_and_clear_young(pte);
1320 static int stage2_pmdp_test_and_clear_young(pmd_t *pmd)
1322 return stage2_ptep_test_and_clear_young((pte_t *)pmd);
1325 static int stage2_pudp_test_and_clear_young(pud_t *pud)
1327 return stage2_ptep_test_and_clear_young((pte_t *)pud);
1331 * kvm_phys_addr_ioremap - map a device range to guest IPA
1333 * @kvm: The KVM pointer
1334 * @guest_ipa: The IPA at which to insert the mapping
1335 * @pa: The physical address of the device
1336 * @size: The size of the mapping
1338 int kvm_phys_addr_ioremap(struct kvm *kvm, phys_addr_t guest_ipa,
1339 phys_addr_t pa, unsigned long size, bool writable)
1341 phys_addr_t addr, end;
1344 struct kvm_mmu_memory_cache cache = { 0, };
1346 end = (guest_ipa + size + PAGE_SIZE - 1) & PAGE_MASK;
1347 pfn = __phys_to_pfn(pa);
1349 for (addr = guest_ipa; addr < end; addr += PAGE_SIZE) {
1350 pte_t pte = kvm_pfn_pte(pfn, PAGE_S2_DEVICE);
1353 pte = kvm_s2pte_mkwrite(pte);
1355 ret = mmu_topup_memory_cache(&cache,
1356 kvm_mmu_cache_min_pages(kvm),
1360 spin_lock(&kvm->mmu_lock);
1361 ret = stage2_set_pte(kvm, &cache, addr, &pte,
1362 KVM_S2PTE_FLAG_IS_IOMAP);
1363 spin_unlock(&kvm->mmu_lock);
1371 mmu_free_memory_cache(&cache);
1375 static bool transparent_hugepage_adjust(kvm_pfn_t *pfnp, phys_addr_t *ipap)
1377 kvm_pfn_t pfn = *pfnp;
1378 gfn_t gfn = *ipap >> PAGE_SHIFT;
1380 if (kvm_is_transparent_hugepage(pfn)) {
1383 * The address we faulted on is backed by a transparent huge
1384 * page. However, because we map the compound huge page and
1385 * not the individual tail page, we need to transfer the
1386 * refcount to the head page. We have to be careful that the
1387 * THP doesn't start to split while we are adjusting the
1390 * We are sure this doesn't happen, because mmu_notifier_retry
1391 * was successful and we are holding the mmu_lock, so if this
1392 * THP is trying to split, it will be blocked in the mmu
1393 * notifier before touching any of the pages, specifically
1394 * before being able to call __split_huge_page_refcount().
1396 * We can therefore safely transfer the refcount from PG_tail
1397 * to PG_head and switch the pfn from a tail page to the head
1400 mask = PTRS_PER_PMD - 1;
1401 VM_BUG_ON((gfn & mask) != (pfn & mask));
1404 kvm_release_pfn_clean(pfn);
1417 * stage2_wp_ptes - write protect PMD range
1418 * @pmd: pointer to pmd entry
1419 * @addr: range start address
1420 * @end: range end address
1422 static void stage2_wp_ptes(pmd_t *pmd, phys_addr_t addr, phys_addr_t end)
1426 pte = pte_offset_kernel(pmd, addr);
1428 if (!pte_none(*pte)) {
1429 if (!kvm_s2pte_readonly(pte))
1430 kvm_set_s2pte_readonly(pte);
1432 } while (pte++, addr += PAGE_SIZE, addr != end);
1436 * stage2_wp_pmds - write protect PUD range
1437 * kvm: kvm instance for the VM
1438 * @pud: pointer to pud entry
1439 * @addr: range start address
1440 * @end: range end address
1442 static void stage2_wp_pmds(struct kvm *kvm, pud_t *pud,
1443 phys_addr_t addr, phys_addr_t end)
1448 pmd = stage2_pmd_offset(kvm, pud, addr);
1451 next = stage2_pmd_addr_end(kvm, addr, end);
1452 if (!pmd_none(*pmd)) {
1453 if (pmd_thp_or_huge(*pmd)) {
1454 if (!kvm_s2pmd_readonly(pmd))
1455 kvm_set_s2pmd_readonly(pmd);
1457 stage2_wp_ptes(pmd, addr, next);
1460 } while (pmd++, addr = next, addr != end);
1464 * stage2_wp_puds - write protect PGD range
1465 * @pgd: pointer to pgd entry
1466 * @addr: range start address
1467 * @end: range end address
1469 static void stage2_wp_puds(struct kvm *kvm, pgd_t *pgd,
1470 phys_addr_t addr, phys_addr_t end)
1475 pud = stage2_pud_offset(kvm, pgd, addr);
1477 next = stage2_pud_addr_end(kvm, addr, end);
1478 if (!stage2_pud_none(kvm, *pud)) {
1479 if (stage2_pud_huge(kvm, *pud)) {
1480 if (!kvm_s2pud_readonly(pud))
1481 kvm_set_s2pud_readonly(pud);
1483 stage2_wp_pmds(kvm, pud, addr, next);
1486 } while (pud++, addr = next, addr != end);
1490 * stage2_wp_range() - write protect stage2 memory region range
1491 * @kvm: The KVM pointer
1492 * @addr: Start address of range
1493 * @end: End address of range
1495 static void stage2_wp_range(struct kvm *kvm, phys_addr_t addr, phys_addr_t end)
1500 pgd = kvm->arch.pgd + stage2_pgd_index(kvm, addr);
1503 * Release kvm_mmu_lock periodically if the memory region is
1504 * large. Otherwise, we may see kernel panics with
1505 * CONFIG_DETECT_HUNG_TASK, CONFIG_LOCKUP_DETECTOR,
1506 * CONFIG_LOCKDEP. Additionally, holding the lock too long
1507 * will also starve other vCPUs. We have to also make sure
1508 * that the page tables are not freed while we released
1511 cond_resched_lock(&kvm->mmu_lock);
1512 if (!READ_ONCE(kvm->arch.pgd))
1514 next = stage2_pgd_addr_end(kvm, addr, end);
1515 if (stage2_pgd_present(kvm, *pgd))
1516 stage2_wp_puds(kvm, pgd, addr, next);
1517 } while (pgd++, addr = next, addr != end);
1521 * kvm_mmu_wp_memory_region() - write protect stage 2 entries for memory slot
1522 * @kvm: The KVM pointer
1523 * @slot: The memory slot to write protect
1525 * Called to start logging dirty pages after memory region
1526 * KVM_MEM_LOG_DIRTY_PAGES operation is called. After this function returns
1527 * all present PUD, PMD and PTEs are write protected in the memory region.
1528 * Afterwards read of dirty page log can be called.
1530 * Acquires kvm_mmu_lock. Called with kvm->slots_lock mutex acquired,
1531 * serializing operations for VM memory regions.
1533 void kvm_mmu_wp_memory_region(struct kvm *kvm, int slot)
1535 struct kvm_memslots *slots = kvm_memslots(kvm);
1536 struct kvm_memory_slot *memslot = id_to_memslot(slots, slot);
1537 phys_addr_t start = memslot->base_gfn << PAGE_SHIFT;
1538 phys_addr_t end = (memslot->base_gfn + memslot->npages) << PAGE_SHIFT;
1540 spin_lock(&kvm->mmu_lock);
1541 stage2_wp_range(kvm, start, end);
1542 spin_unlock(&kvm->mmu_lock);
1543 kvm_flush_remote_tlbs(kvm);
1547 * kvm_mmu_write_protect_pt_masked() - write protect dirty pages
1548 * @kvm: The KVM pointer
1549 * @slot: The memory slot associated with mask
1550 * @gfn_offset: The gfn offset in memory slot
1551 * @mask: The mask of dirty pages at offset 'gfn_offset' in this memory
1552 * slot to be write protected
1554 * Walks bits set in mask write protects the associated pte's. Caller must
1555 * acquire kvm_mmu_lock.
1557 static void kvm_mmu_write_protect_pt_masked(struct kvm *kvm,
1558 struct kvm_memory_slot *slot,
1559 gfn_t gfn_offset, unsigned long mask)
1561 phys_addr_t base_gfn = slot->base_gfn + gfn_offset;
1562 phys_addr_t start = (base_gfn + __ffs(mask)) << PAGE_SHIFT;
1563 phys_addr_t end = (base_gfn + __fls(mask) + 1) << PAGE_SHIFT;
1565 stage2_wp_range(kvm, start, end);
1569 * kvm_arch_mmu_enable_log_dirty_pt_masked - enable dirty logging for selected
1572 * It calls kvm_mmu_write_protect_pt_masked to write protect selected pages to
1573 * enable dirty logging for them.
1575 void kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm *kvm,
1576 struct kvm_memory_slot *slot,
1577 gfn_t gfn_offset, unsigned long mask)
1579 kvm_mmu_write_protect_pt_masked(kvm, slot, gfn_offset, mask);
1582 static void clean_dcache_guest_page(kvm_pfn_t pfn, unsigned long size)
1584 __clean_dcache_guest_page(pfn, size);
1587 static void invalidate_icache_guest_page(kvm_pfn_t pfn, unsigned long size)
1589 __invalidate_icache_guest_page(pfn, size);
1592 static void kvm_send_hwpoison_signal(unsigned long address, short lsb)
1594 send_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, lsb, current);
1597 static bool fault_supports_stage2_huge_mapping(struct kvm_memory_slot *memslot,
1599 unsigned long map_size)
1602 hva_t uaddr_start, uaddr_end;
1605 size = memslot->npages * PAGE_SIZE;
1607 gpa_start = memslot->base_gfn << PAGE_SHIFT;
1609 uaddr_start = memslot->userspace_addr;
1610 uaddr_end = uaddr_start + size;
1613 * Pages belonging to memslots that don't have the same alignment
1614 * within a PMD/PUD for userspace and IPA cannot be mapped with stage-2
1615 * PMD/PUD entries, because we'll end up mapping the wrong pages.
1617 * Consider a layout like the following:
1619 * memslot->userspace_addr:
1620 * +-----+--------------------+--------------------+---+
1621 * |abcde|fgh Stage-1 block | Stage-1 block tv|xyz|
1622 * +-----+--------------------+--------------------+---+
1624 * memslot->base_gfn << PAGE_SIZE:
1625 * +---+--------------------+--------------------+-----+
1626 * |abc|def Stage-2 block | Stage-2 block |tvxyz|
1627 * +---+--------------------+--------------------+-----+
1629 * If we create those stage-2 blocks, we'll end up with this incorrect
1635 if ((gpa_start & (map_size - 1)) != (uaddr_start & (map_size - 1)))
1639 * Next, let's make sure we're not trying to map anything not covered
1640 * by the memslot. This means we have to prohibit block size mappings
1641 * for the beginning and end of a non-block aligned and non-block sized
1642 * memory slot (illustrated by the head and tail parts of the
1643 * userspace view above containing pages 'abcde' and 'xyz',
1646 * Note that it doesn't matter if we do the check using the
1647 * userspace_addr or the base_gfn, as both are equally aligned (per
1648 * the check above) and equally sized.
1650 return (hva & ~(map_size - 1)) >= uaddr_start &&
1651 (hva & ~(map_size - 1)) + map_size <= uaddr_end;
1654 static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
1655 struct kvm_memory_slot *memslot, unsigned long hva,
1656 unsigned long fault_status)
1659 bool write_fault, writable, force_pte = false;
1660 bool exec_fault, needs_exec;
1661 unsigned long mmu_seq;
1662 gfn_t gfn = fault_ipa >> PAGE_SHIFT;
1663 struct kvm *kvm = vcpu->kvm;
1664 struct kvm_mmu_memory_cache *memcache = &vcpu->arch.mmu_page_cache;
1665 struct vm_area_struct *vma;
1668 pgprot_t mem_type = PAGE_S2;
1669 bool logging_active = memslot_is_logging(memslot);
1670 unsigned long vma_pagesize, flags = 0;
1672 write_fault = kvm_is_write_fault(vcpu);
1673 exec_fault = kvm_vcpu_trap_is_iabt(vcpu);
1674 VM_BUG_ON(write_fault && exec_fault);
1676 if (fault_status == FSC_PERM && !write_fault && !exec_fault) {
1677 kvm_err("Unexpected L2 read permission error\n");
1681 /* Let's check if we will get back a huge page backed by hugetlbfs */
1682 down_read(¤t->mm->mmap_sem);
1683 vma = find_vma_intersection(current->mm, hva, hva + 1);
1684 if (unlikely(!vma)) {
1685 kvm_err("Failed to find VMA for hva 0x%lx\n", hva);
1686 up_read(¤t->mm->mmap_sem);
1690 if (is_vm_hugetlb_page(vma))
1691 vma_shift = huge_page_shift(hstate_vma(vma));
1693 vma_shift = PAGE_SHIFT;
1695 vma_pagesize = 1ULL << vma_shift;
1696 if (logging_active ||
1697 (vma->vm_flags & VM_PFNMAP) ||
1698 !fault_supports_stage2_huge_mapping(memslot, hva, vma_pagesize)) {
1700 vma_pagesize = PAGE_SIZE;
1704 * The stage2 has a minimum of 2 level table (For arm64 see
1705 * kvm_arm_setup_stage2()). Hence, we are guaranteed that we can
1706 * use PMD_SIZE huge mappings (even when the PMD is folded into PGD).
1707 * As for PUD huge maps, we must make sure that we have at least
1708 * 3 levels, i.e, PMD is not folded.
1710 if (vma_pagesize == PMD_SIZE ||
1711 (vma_pagesize == PUD_SIZE && kvm_stage2_has_pmd(kvm)))
1712 gfn = (fault_ipa & huge_page_mask(hstate_vma(vma))) >> PAGE_SHIFT;
1713 up_read(¤t->mm->mmap_sem);
1715 /* We need minimum second+third level pages */
1716 ret = mmu_topup_memory_cache(memcache, kvm_mmu_cache_min_pages(kvm),
1721 mmu_seq = vcpu->kvm->mmu_notifier_seq;
1723 * Ensure the read of mmu_notifier_seq happens before we call
1724 * gfn_to_pfn_prot (which calls get_user_pages), so that we don't risk
1725 * the page we just got a reference to gets unmapped before we have a
1726 * chance to grab the mmu_lock, which ensure that if the page gets
1727 * unmapped afterwards, the call to kvm_unmap_hva will take it away
1728 * from us again properly. This smp_rmb() interacts with the smp_wmb()
1729 * in kvm_mmu_notifier_invalidate_<page|range_end>.
1733 pfn = gfn_to_pfn_prot(kvm, gfn, write_fault, &writable);
1734 if (pfn == KVM_PFN_ERR_HWPOISON) {
1735 kvm_send_hwpoison_signal(hva, vma_shift);
1738 if (is_error_noslot_pfn(pfn))
1741 if (kvm_is_device_pfn(pfn)) {
1742 mem_type = PAGE_S2_DEVICE;
1743 flags |= KVM_S2PTE_FLAG_IS_IOMAP;
1744 } else if (logging_active) {
1746 * Faults on pages in a memslot with logging enabled
1747 * should not be mapped with huge pages (it introduces churn
1748 * and performance degradation), so force a pte mapping.
1750 flags |= KVM_S2_FLAG_LOGGING_ACTIVE;
1753 * Only actually map the page as writable if this was a write
1760 if (exec_fault && is_iomap(flags))
1763 spin_lock(&kvm->mmu_lock);
1764 if (mmu_notifier_retry(kvm, mmu_seq))
1767 if (vma_pagesize == PAGE_SIZE && !force_pte) {
1769 * Only PMD_SIZE transparent hugepages(THP) are
1770 * currently supported. This code will need to be
1771 * updated to support other THP sizes.
1773 * Make sure the host VA and the guest IPA are sufficiently
1774 * aligned and that the block is contained within the memslot.
1776 if (fault_supports_stage2_huge_mapping(memslot, hva, PMD_SIZE) &&
1777 transparent_hugepage_adjust(&pfn, &fault_ipa))
1778 vma_pagesize = PMD_SIZE;
1782 kvm_set_pfn_dirty(pfn);
1784 if (fault_status != FSC_PERM && !is_iomap(flags))
1785 clean_dcache_guest_page(pfn, vma_pagesize);
1788 invalidate_icache_guest_page(pfn, vma_pagesize);
1791 * If we took an execution fault we have made the
1792 * icache/dcache coherent above and should now let the s2
1793 * mapping be executable.
1795 * Write faults (!exec_fault && FSC_PERM) are orthogonal to
1796 * execute permissions, and we preserve whatever we have.
1798 needs_exec = exec_fault ||
1799 (fault_status == FSC_PERM && stage2_is_exec(kvm, fault_ipa));
1801 if (vma_pagesize == PUD_SIZE) {
1802 pud_t new_pud = kvm_pfn_pud(pfn, mem_type);
1804 new_pud = kvm_pud_mkhuge(new_pud);
1806 new_pud = kvm_s2pud_mkwrite(new_pud);
1809 new_pud = kvm_s2pud_mkexec(new_pud);
1811 ret = stage2_set_pud_huge(kvm, memcache, fault_ipa, &new_pud);
1812 } else if (vma_pagesize == PMD_SIZE) {
1813 pmd_t new_pmd = kvm_pfn_pmd(pfn, mem_type);
1815 new_pmd = kvm_pmd_mkhuge(new_pmd);
1818 new_pmd = kvm_s2pmd_mkwrite(new_pmd);
1821 new_pmd = kvm_s2pmd_mkexec(new_pmd);
1823 ret = stage2_set_pmd_huge(kvm, memcache, fault_ipa, &new_pmd);
1825 pte_t new_pte = kvm_pfn_pte(pfn, mem_type);
1828 new_pte = kvm_s2pte_mkwrite(new_pte);
1829 mark_page_dirty(kvm, gfn);
1833 new_pte = kvm_s2pte_mkexec(new_pte);
1835 ret = stage2_set_pte(kvm, memcache, fault_ipa, &new_pte, flags);
1839 spin_unlock(&kvm->mmu_lock);
1840 kvm_set_pfn_accessed(pfn);
1841 kvm_release_pfn_clean(pfn);
1846 * Resolve the access fault by making the page young again.
1847 * Note that because the faulting entry is guaranteed not to be
1848 * cached in the TLB, we don't need to invalidate anything.
1849 * Only the HW Access Flag updates are supported for Stage 2 (no DBM),
1850 * so there is no need for atomic (pte|pmd)_mkyoung operations.
1852 static void handle_access_fault(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa)
1858 bool pfn_valid = false;
1860 trace_kvm_access_fault(fault_ipa);
1862 spin_lock(&vcpu->kvm->mmu_lock);
1864 if (!stage2_get_leaf_entry(vcpu->kvm, fault_ipa, &pud, &pmd, &pte))
1867 if (pud) { /* HugeTLB */
1868 *pud = kvm_s2pud_mkyoung(*pud);
1869 pfn = kvm_pud_pfn(*pud);
1871 } else if (pmd) { /* THP, HugeTLB */
1872 *pmd = pmd_mkyoung(*pmd);
1873 pfn = pmd_pfn(*pmd);
1876 *pte = pte_mkyoung(*pte); /* Just a page... */
1877 pfn = pte_pfn(*pte);
1882 spin_unlock(&vcpu->kvm->mmu_lock);
1884 kvm_set_pfn_accessed(pfn);
1888 * kvm_handle_guest_abort - handles all 2nd stage aborts
1889 * @vcpu: the VCPU pointer
1890 * @run: the kvm_run structure
1892 * Any abort that gets to the host is almost guaranteed to be caused by a
1893 * missing second stage translation table entry, which can mean that either the
1894 * guest simply needs more memory and we must allocate an appropriate page or it
1895 * can mean that the guest tried to access I/O memory, which is emulated by user
1896 * space. The distinction is based on the IPA causing the fault and whether this
1897 * memory region has been registered as standard RAM by user space.
1899 int kvm_handle_guest_abort(struct kvm_vcpu *vcpu, struct kvm_run *run)
1901 unsigned long fault_status;
1902 phys_addr_t fault_ipa;
1903 struct kvm_memory_slot *memslot;
1905 bool is_iabt, write_fault, writable;
1909 fault_status = kvm_vcpu_trap_get_fault_type(vcpu);
1911 fault_ipa = kvm_vcpu_get_fault_ipa(vcpu);
1912 is_iabt = kvm_vcpu_trap_is_iabt(vcpu);
1914 /* Synchronous External Abort? */
1915 if (kvm_vcpu_dabt_isextabt(vcpu)) {
1917 * For RAS the host kernel may handle this abort.
1918 * There is no need to pass the error into the guest.
1920 if (!kvm_handle_guest_sea(fault_ipa, kvm_vcpu_get_hsr(vcpu)))
1923 if (unlikely(!is_iabt)) {
1924 kvm_inject_vabt(vcpu);
1929 trace_kvm_guest_fault(*vcpu_pc(vcpu), kvm_vcpu_get_hsr(vcpu),
1930 kvm_vcpu_get_hfar(vcpu), fault_ipa);
1932 /* Check the stage-2 fault is trans. fault or write fault */
1933 if (fault_status != FSC_FAULT && fault_status != FSC_PERM &&
1934 fault_status != FSC_ACCESS) {
1935 kvm_err("Unsupported FSC: EC=%#x xFSC=%#lx ESR_EL2=%#lx\n",
1936 kvm_vcpu_trap_get_class(vcpu),
1937 (unsigned long)kvm_vcpu_trap_get_fault(vcpu),
1938 (unsigned long)kvm_vcpu_get_hsr(vcpu));
1942 idx = srcu_read_lock(&vcpu->kvm->srcu);
1944 gfn = fault_ipa >> PAGE_SHIFT;
1945 memslot = gfn_to_memslot(vcpu->kvm, gfn);
1946 hva = gfn_to_hva_memslot_prot(memslot, gfn, &writable);
1947 write_fault = kvm_is_write_fault(vcpu);
1948 if (kvm_is_error_hva(hva) || (write_fault && !writable)) {
1950 /* Prefetch Abort on I/O address */
1956 * Check for a cache maintenance operation. Since we
1957 * ended-up here, we know it is outside of any memory
1958 * slot. But we can't find out if that is for a device,
1959 * or if the guest is just being stupid. The only thing
1960 * we know for sure is that this range cannot be cached.
1962 * So let's assume that the guest is just being
1963 * cautious, and skip the instruction.
1965 if (kvm_vcpu_dabt_is_cm(vcpu)) {
1966 kvm_skip_instr(vcpu, kvm_vcpu_trap_il_is32bit(vcpu));
1972 * The IPA is reported as [MAX:12], so we need to
1973 * complement it with the bottom 12 bits from the
1974 * faulting VA. This is always 12 bits, irrespective
1977 fault_ipa |= kvm_vcpu_get_hfar(vcpu) & ((1 << 12) - 1);
1978 ret = io_mem_abort(vcpu, run, fault_ipa);
1982 /* Userspace should not be able to register out-of-bounds IPAs */
1983 VM_BUG_ON(fault_ipa >= kvm_phys_size(vcpu->kvm));
1985 if (fault_status == FSC_ACCESS) {
1986 handle_access_fault(vcpu, fault_ipa);
1991 ret = user_mem_abort(vcpu, fault_ipa, memslot, hva, fault_status);
1995 if (ret == -ENOEXEC) {
1996 kvm_inject_pabt(vcpu, kvm_vcpu_get_hfar(vcpu));
2000 srcu_read_unlock(&vcpu->kvm->srcu, idx);
2004 static int handle_hva_to_gpa(struct kvm *kvm,
2005 unsigned long start,
2007 int (*handler)(struct kvm *kvm,
2008 gpa_t gpa, u64 size,
2012 struct kvm_memslots *slots;
2013 struct kvm_memory_slot *memslot;
2016 slots = kvm_memslots(kvm);
2018 /* we only care about the pages that the guest sees */
2019 kvm_for_each_memslot(memslot, slots) {
2020 unsigned long hva_start, hva_end;
2023 hva_start = max(start, memslot->userspace_addr);
2024 hva_end = min(end, memslot->userspace_addr +
2025 (memslot->npages << PAGE_SHIFT));
2026 if (hva_start >= hva_end)
2029 gpa = hva_to_gfn_memslot(hva_start, memslot) << PAGE_SHIFT;
2030 ret |= handler(kvm, gpa, (u64)(hva_end - hva_start), data);
2036 static int kvm_unmap_hva_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
2038 unmap_stage2_range(kvm, gpa, size);
2042 int kvm_unmap_hva_range(struct kvm *kvm,
2043 unsigned long start, unsigned long end)
2048 trace_kvm_unmap_hva_range(start, end);
2049 handle_hva_to_gpa(kvm, start, end, &kvm_unmap_hva_handler, NULL);
2053 static int kvm_set_spte_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
2055 pte_t *pte = (pte_t *)data;
2057 WARN_ON(size != PAGE_SIZE);
2059 * We can always call stage2_set_pte with KVM_S2PTE_FLAG_LOGGING_ACTIVE
2060 * flag clear because MMU notifiers will have unmapped a huge PMD before
2061 * calling ->change_pte() (which in turn calls kvm_set_spte_hva()) and
2062 * therefore stage2_set_pte() never needs to clear out a huge PMD
2063 * through this calling path.
2065 stage2_set_pte(kvm, NULL, gpa, pte, 0);
2070 int kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
2072 unsigned long end = hva + PAGE_SIZE;
2073 kvm_pfn_t pfn = pte_pfn(pte);
2079 trace_kvm_set_spte_hva(hva);
2082 * We've moved a page around, probably through CoW, so let's treat it
2083 * just like a translation fault and clean the cache to the PoC.
2085 clean_dcache_guest_page(pfn, PAGE_SIZE);
2086 stage2_pte = kvm_pfn_pte(pfn, PAGE_S2);
2087 handle_hva_to_gpa(kvm, hva, end, &kvm_set_spte_handler, &stage2_pte);
2092 static int kvm_age_hva_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
2098 WARN_ON(size != PAGE_SIZE && size != PMD_SIZE && size != PUD_SIZE);
2099 if (!stage2_get_leaf_entry(kvm, gpa, &pud, &pmd, &pte))
2103 return stage2_pudp_test_and_clear_young(pud);
2105 return stage2_pmdp_test_and_clear_young(pmd);
2107 return stage2_ptep_test_and_clear_young(pte);
2110 static int kvm_test_age_hva_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
2116 WARN_ON(size != PAGE_SIZE && size != PMD_SIZE && size != PUD_SIZE);
2117 if (!stage2_get_leaf_entry(kvm, gpa, &pud, &pmd, &pte))
2121 return kvm_s2pud_young(*pud);
2123 return pmd_young(*pmd);
2125 return pte_young(*pte);
2128 int kvm_age_hva(struct kvm *kvm, unsigned long start, unsigned long end)
2132 trace_kvm_age_hva(start, end);
2133 return handle_hva_to_gpa(kvm, start, end, kvm_age_hva_handler, NULL);
2136 int kvm_test_age_hva(struct kvm *kvm, unsigned long hva)
2140 trace_kvm_test_age_hva(hva);
2141 return handle_hva_to_gpa(kvm, hva, hva + PAGE_SIZE,
2142 kvm_test_age_hva_handler, NULL);
2145 void kvm_mmu_free_memory_caches(struct kvm_vcpu *vcpu)
2147 mmu_free_memory_cache(&vcpu->arch.mmu_page_cache);
2150 phys_addr_t kvm_mmu_get_httbr(void)
2152 if (__kvm_cpu_uses_extended_idmap())
2153 return virt_to_phys(merged_hyp_pgd);
2155 return virt_to_phys(hyp_pgd);
2158 phys_addr_t kvm_get_idmap_vector(void)
2160 return hyp_idmap_vector;
2163 static int kvm_map_idmap_text(pgd_t *pgd)
2167 /* Create the idmap in the boot page tables */
2168 err = __create_hyp_mappings(pgd, __kvm_idmap_ptrs_per_pgd(),
2169 hyp_idmap_start, hyp_idmap_end,
2170 __phys_to_pfn(hyp_idmap_start),
2173 kvm_err("Failed to idmap %lx-%lx\n",
2174 hyp_idmap_start, hyp_idmap_end);
2179 int kvm_mmu_init(void)
2183 hyp_idmap_start = kvm_virt_to_phys(__hyp_idmap_text_start);
2184 hyp_idmap_start = ALIGN_DOWN(hyp_idmap_start, PAGE_SIZE);
2185 hyp_idmap_end = kvm_virt_to_phys(__hyp_idmap_text_end);
2186 hyp_idmap_end = ALIGN(hyp_idmap_end, PAGE_SIZE);
2187 hyp_idmap_vector = kvm_virt_to_phys(__kvm_hyp_init);
2190 * We rely on the linker script to ensure at build time that the HYP
2191 * init code does not cross a page boundary.
2193 BUG_ON((hyp_idmap_start ^ (hyp_idmap_end - 1)) & PAGE_MASK);
2195 kvm_debug("IDMAP page: %lx\n", hyp_idmap_start);
2196 kvm_debug("HYP VA range: %lx:%lx\n",
2197 kern_hyp_va(PAGE_OFFSET),
2198 kern_hyp_va((unsigned long)high_memory - 1));
2200 if (hyp_idmap_start >= kern_hyp_va(PAGE_OFFSET) &&
2201 hyp_idmap_start < kern_hyp_va((unsigned long)high_memory - 1) &&
2202 hyp_idmap_start != (unsigned long)__hyp_idmap_text_start) {
2204 * The idmap page is intersecting with the VA space,
2205 * it is not safe to continue further.
2207 kvm_err("IDMAP intersecting with HYP VA, unable to continue\n");
2212 hyp_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, hyp_pgd_order);
2214 kvm_err("Hyp mode PGD not allocated\n");
2219 if (__kvm_cpu_uses_extended_idmap()) {
2220 boot_hyp_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
2222 if (!boot_hyp_pgd) {
2223 kvm_err("Hyp boot PGD not allocated\n");
2228 err = kvm_map_idmap_text(boot_hyp_pgd);
2232 merged_hyp_pgd = (pgd_t *)__get_free_page(GFP_KERNEL | __GFP_ZERO);
2233 if (!merged_hyp_pgd) {
2234 kvm_err("Failed to allocate extra HYP pgd\n");
2237 __kvm_extend_hypmap(boot_hyp_pgd, hyp_pgd, merged_hyp_pgd,
2240 err = kvm_map_idmap_text(hyp_pgd);
2245 io_map_base = hyp_idmap_start;
2252 void kvm_arch_commit_memory_region(struct kvm *kvm,
2253 const struct kvm_userspace_memory_region *mem,
2254 const struct kvm_memory_slot *old,
2255 const struct kvm_memory_slot *new,
2256 enum kvm_mr_change change)
2259 * At this point memslot has been committed and there is an
2260 * allocated dirty_bitmap[], dirty pages will be be tracked while the
2261 * memory slot is write protected.
2263 if (change != KVM_MR_DELETE && mem->flags & KVM_MEM_LOG_DIRTY_PAGES)
2264 kvm_mmu_wp_memory_region(kvm, mem->slot);
2267 int kvm_arch_prepare_memory_region(struct kvm *kvm,
2268 struct kvm_memory_slot *memslot,
2269 const struct kvm_userspace_memory_region *mem,
2270 enum kvm_mr_change change)
2272 hva_t hva = mem->userspace_addr;
2273 hva_t reg_end = hva + mem->memory_size;
2274 bool writable = !(mem->flags & KVM_MEM_READONLY);
2277 if (change != KVM_MR_CREATE && change != KVM_MR_MOVE &&
2278 change != KVM_MR_FLAGS_ONLY)
2282 * Prevent userspace from creating a memory region outside of the IPA
2283 * space addressable by the KVM guest IPA space.
2285 if (memslot->base_gfn + memslot->npages >=
2286 (kvm_phys_size(kvm) >> PAGE_SHIFT))
2289 down_read(¤t->mm->mmap_sem);
2291 * A memory region could potentially cover multiple VMAs, and any holes
2292 * between them, so iterate over all of them to find out if we can map
2293 * any of them right now.
2295 * +--------------------------------------------+
2296 * +---------------+----------------+ +----------------+
2297 * | : VMA 1 | VMA 2 | | VMA 3 : |
2298 * +---------------+----------------+ +----------------+
2300 * +--------------------------------------------+
2303 struct vm_area_struct *vma = find_vma(current->mm, hva);
2304 hva_t vm_start, vm_end;
2306 if (!vma || vma->vm_start >= reg_end)
2310 * Take the intersection of this VMA with the memory region
2312 vm_start = max(hva, vma->vm_start);
2313 vm_end = min(reg_end, vma->vm_end);
2315 if (vma->vm_flags & VM_PFNMAP) {
2316 gpa_t gpa = mem->guest_phys_addr +
2317 (vm_start - mem->userspace_addr);
2320 pa = (phys_addr_t)vma->vm_pgoff << PAGE_SHIFT;
2321 pa += vm_start - vma->vm_start;
2323 /* IO region dirty page logging not allowed */
2324 if (memslot->flags & KVM_MEM_LOG_DIRTY_PAGES) {
2329 ret = kvm_phys_addr_ioremap(kvm, gpa, pa,
2336 } while (hva < reg_end);
2338 if (change == KVM_MR_FLAGS_ONLY)
2341 spin_lock(&kvm->mmu_lock);
2343 unmap_stage2_range(kvm, mem->guest_phys_addr, mem->memory_size);
2345 stage2_flush_memslot(kvm, memslot);
2346 spin_unlock(&kvm->mmu_lock);
2348 up_read(¤t->mm->mmap_sem);
2352 void kvm_arch_free_memslot(struct kvm *kvm, struct kvm_memory_slot *free,
2353 struct kvm_memory_slot *dont)
2357 int kvm_arch_create_memslot(struct kvm *kvm, struct kvm_memory_slot *slot,
2358 unsigned long npages)
2363 void kvm_arch_memslots_updated(struct kvm *kvm, u64 gen)
2367 void kvm_arch_flush_shadow_all(struct kvm *kvm)
2369 kvm_free_stage2_pgd(kvm);
2372 void kvm_arch_flush_shadow_memslot(struct kvm *kvm,
2373 struct kvm_memory_slot *slot)
2375 gpa_t gpa = slot->base_gfn << PAGE_SHIFT;
2376 phys_addr_t size = slot->npages << PAGE_SHIFT;
2378 spin_lock(&kvm->mmu_lock);
2379 unmap_stage2_range(kvm, gpa, size);
2380 spin_unlock(&kvm->mmu_lock);
2384 * See note at ARMv7 ARM B1.14.4 (TL;DR: S/W ops are not easily virtualized).
2387 * - S/W ops are local to a CPU (not broadcast)
2388 * - We have line migration behind our back (speculation)
2389 * - System caches don't support S/W at all (damn!)
2391 * In the face of the above, the best we can do is to try and convert
2392 * S/W ops to VA ops. Because the guest is not allowed to infer the
2393 * S/W to PA mapping, it can only use S/W to nuke the whole cache,
2394 * which is a rather good thing for us.
2396 * Also, it is only used when turning caches on/off ("The expected
2397 * usage of the cache maintenance instructions that operate by set/way
2398 * is associated with the cache maintenance instructions associated
2399 * with the powerdown and powerup of caches, if this is required by
2400 * the implementation.").
2402 * We use the following policy:
2404 * - If we trap a S/W operation, we enable VM trapping to detect
2405 * caches being turned on/off, and do a full clean.
2407 * - We flush the caches on both caches being turned on and off.
2409 * - Once the caches are enabled, we stop trapping VM ops.
2411 void kvm_set_way_flush(struct kvm_vcpu *vcpu)
2413 unsigned long hcr = *vcpu_hcr(vcpu);
2416 * If this is the first time we do a S/W operation
2417 * (i.e. HCR_TVM not set) flush the whole memory, and set the
2420 * Otherwise, rely on the VM trapping to wait for the MMU +
2421 * Caches to be turned off. At that point, we'll be able to
2422 * clean the caches again.
2424 if (!(hcr & HCR_TVM)) {
2425 trace_kvm_set_way_flush(*vcpu_pc(vcpu),
2426 vcpu_has_cache_enabled(vcpu));
2427 stage2_flush_vm(vcpu->kvm);
2428 *vcpu_hcr(vcpu) = hcr | HCR_TVM;
2432 void kvm_toggle_cache(struct kvm_vcpu *vcpu, bool was_enabled)
2434 bool now_enabled = vcpu_has_cache_enabled(vcpu);
2437 * If switching the MMU+caches on, need to invalidate the caches.
2438 * If switching it off, need to clean the caches.
2439 * Clean + invalidate does the trick always.
2441 if (now_enabled != was_enabled)
2442 stage2_flush_vm(vcpu->kvm);
2444 /* Caches are now on, stop trapping VM ops (until a S/W op) */
2446 *vcpu_hcr(vcpu) &= ~HCR_TVM;
2448 trace_kvm_toggle_cache(*vcpu_pc(vcpu), was_enabled, now_enabled);