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 static unsigned long io_map_base;
48 #define S2_PGD_SIZE (PTRS_PER_S2_PGD * sizeof(pgd_t))
49 #define hyp_pgd_order get_order(PTRS_PER_PGD * sizeof(pgd_t))
51 #define KVM_S2PTE_FLAG_IS_IOMAP (1UL << 0)
52 #define KVM_S2_FLAG_LOGGING_ACTIVE (1UL << 1)
54 static bool memslot_is_logging(struct kvm_memory_slot *memslot)
56 return memslot->dirty_bitmap && !(memslot->flags & KVM_MEM_READONLY);
60 * kvm_flush_remote_tlbs() - flush all VM TLB entries for v7/8
61 * @kvm: pointer to kvm structure.
63 * Interface to HYP function to flush all VM TLB entries
65 void kvm_flush_remote_tlbs(struct kvm *kvm)
67 kvm_call_hyp(__kvm_tlb_flush_vmid, kvm);
70 static void kvm_tlb_flush_vmid_ipa(struct kvm *kvm, phys_addr_t ipa)
72 kvm_call_hyp(__kvm_tlb_flush_vmid_ipa, kvm, ipa);
76 * D-Cache management functions. They take the page table entries by
77 * value, as they are flushing the cache using the kernel mapping (or
80 static void kvm_flush_dcache_pte(pte_t pte)
82 __kvm_flush_dcache_pte(pte);
85 static void kvm_flush_dcache_pmd(pmd_t pmd)
87 __kvm_flush_dcache_pmd(pmd);
90 static void kvm_flush_dcache_pud(pud_t pud)
92 __kvm_flush_dcache_pud(pud);
95 static bool kvm_is_device_pfn(unsigned long pfn)
97 return !pfn_valid(pfn);
101 * stage2_dissolve_pmd() - clear and flush huge PMD entry
102 * @kvm: pointer to kvm structure.
104 * @pmd: pmd pointer for IPA
106 * Function clears a PMD entry, flushes addr 1st and 2nd stage TLBs. Marks all
107 * pages in the range dirty.
109 static void stage2_dissolve_pmd(struct kvm *kvm, phys_addr_t addr, pmd_t *pmd)
111 if (!pmd_thp_or_huge(*pmd))
115 kvm_tlb_flush_vmid_ipa(kvm, addr);
116 put_page(virt_to_page(pmd));
119 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
124 BUG_ON(max > KVM_NR_MEM_OBJS);
125 if (cache->nobjs >= min)
127 while (cache->nobjs < max) {
128 page = (void *)__get_free_page(PGALLOC_GFP);
131 cache->objects[cache->nobjs++] = page;
136 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
139 free_page((unsigned long)mc->objects[--mc->nobjs]);
142 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc)
146 BUG_ON(!mc || !mc->nobjs);
147 p = mc->objects[--mc->nobjs];
151 static void clear_stage2_pgd_entry(struct kvm *kvm, pgd_t *pgd, phys_addr_t addr)
153 pud_t *pud_table __maybe_unused = stage2_pud_offset(pgd, 0UL);
154 stage2_pgd_clear(pgd);
155 kvm_tlb_flush_vmid_ipa(kvm, addr);
156 stage2_pud_free(pud_table);
157 put_page(virt_to_page(pgd));
160 static void clear_stage2_pud_entry(struct kvm *kvm, pud_t *pud, phys_addr_t addr)
162 pmd_t *pmd_table __maybe_unused = stage2_pmd_offset(pud, 0);
163 VM_BUG_ON(stage2_pud_huge(*pud));
164 stage2_pud_clear(pud);
165 kvm_tlb_flush_vmid_ipa(kvm, addr);
166 stage2_pmd_free(pmd_table);
167 put_page(virt_to_page(pud));
170 static void clear_stage2_pmd_entry(struct kvm *kvm, pmd_t *pmd, phys_addr_t addr)
172 pte_t *pte_table = pte_offset_kernel(pmd, 0);
173 VM_BUG_ON(pmd_thp_or_huge(*pmd));
175 kvm_tlb_flush_vmid_ipa(kvm, addr);
176 pte_free_kernel(NULL, pte_table);
177 put_page(virt_to_page(pmd));
180 static inline void kvm_set_pte(pte_t *ptep, pte_t new_pte)
182 WRITE_ONCE(*ptep, new_pte);
186 static inline void kvm_set_pmd(pmd_t *pmdp, pmd_t new_pmd)
188 WRITE_ONCE(*pmdp, new_pmd);
192 static inline void kvm_pmd_populate(pmd_t *pmdp, pte_t *ptep)
194 kvm_set_pmd(pmdp, kvm_mk_pmd(ptep));
197 static inline void kvm_pud_populate(pud_t *pudp, pmd_t *pmdp)
199 WRITE_ONCE(*pudp, kvm_mk_pud(pmdp));
203 static inline void kvm_pgd_populate(pgd_t *pgdp, pud_t *pudp)
205 WRITE_ONCE(*pgdp, kvm_mk_pgd(pudp));
210 * Unmapping vs dcache management:
212 * If a guest maps certain memory pages as uncached, all writes will
213 * bypass the data cache and go directly to RAM. However, the CPUs
214 * can still speculate reads (not writes) and fill cache lines with
217 * Those cache lines will be *clean* cache lines though, so a
218 * clean+invalidate operation is equivalent to an invalidate
219 * operation, because no cache lines are marked dirty.
221 * Those clean cache lines could be filled prior to an uncached write
222 * by the guest, and the cache coherent IO subsystem would therefore
223 * end up writing old data to disk.
225 * This is why right after unmapping a page/section and invalidating
226 * the corresponding TLBs, we call kvm_flush_dcache_p*() to make sure
227 * the IO subsystem will never hit in the cache.
229 * This is all avoided on systems that have ARM64_HAS_STAGE2_FWB, as
230 * we then fully enforce cacheability of RAM, no matter what the guest
233 static void unmap_stage2_ptes(struct kvm *kvm, pmd_t *pmd,
234 phys_addr_t addr, phys_addr_t end)
236 phys_addr_t start_addr = addr;
237 pte_t *pte, *start_pte;
239 start_pte = pte = pte_offset_kernel(pmd, addr);
241 if (!pte_none(*pte)) {
242 pte_t old_pte = *pte;
244 kvm_set_pte(pte, __pte(0));
245 kvm_tlb_flush_vmid_ipa(kvm, addr);
247 /* No need to invalidate the cache for device mappings */
248 if (!kvm_is_device_pfn(pte_pfn(old_pte)))
249 kvm_flush_dcache_pte(old_pte);
251 put_page(virt_to_page(pte));
253 } while (pte++, addr += PAGE_SIZE, addr != end);
255 if (stage2_pte_table_empty(start_pte))
256 clear_stage2_pmd_entry(kvm, pmd, start_addr);
259 static void unmap_stage2_pmds(struct kvm *kvm, pud_t *pud,
260 phys_addr_t addr, phys_addr_t end)
262 phys_addr_t next, start_addr = addr;
263 pmd_t *pmd, *start_pmd;
265 start_pmd = pmd = stage2_pmd_offset(pud, addr);
267 next = stage2_pmd_addr_end(addr, end);
268 if (!pmd_none(*pmd)) {
269 if (pmd_thp_or_huge(*pmd)) {
270 pmd_t old_pmd = *pmd;
273 kvm_tlb_flush_vmid_ipa(kvm, addr);
275 kvm_flush_dcache_pmd(old_pmd);
277 put_page(virt_to_page(pmd));
279 unmap_stage2_ptes(kvm, pmd, addr, next);
282 } while (pmd++, addr = next, addr != end);
284 if (stage2_pmd_table_empty(start_pmd))
285 clear_stage2_pud_entry(kvm, pud, start_addr);
288 static void unmap_stage2_puds(struct kvm *kvm, pgd_t *pgd,
289 phys_addr_t addr, phys_addr_t end)
291 phys_addr_t next, start_addr = addr;
292 pud_t *pud, *start_pud;
294 start_pud = pud = stage2_pud_offset(pgd, addr);
296 next = stage2_pud_addr_end(addr, end);
297 if (!stage2_pud_none(*pud)) {
298 if (stage2_pud_huge(*pud)) {
299 pud_t old_pud = *pud;
301 stage2_pud_clear(pud);
302 kvm_tlb_flush_vmid_ipa(kvm, addr);
303 kvm_flush_dcache_pud(old_pud);
304 put_page(virt_to_page(pud));
306 unmap_stage2_pmds(kvm, pud, addr, next);
309 } while (pud++, addr = next, addr != end);
311 if (stage2_pud_table_empty(start_pud))
312 clear_stage2_pgd_entry(kvm, pgd, start_addr);
316 * unmap_stage2_range -- Clear stage2 page table entries to unmap a range
317 * @kvm: The VM pointer
318 * @start: The intermediate physical base address of the range to unmap
319 * @size: The size of the area to unmap
321 * Clear a range of stage-2 mappings, lowering the various ref-counts. Must
322 * be called while holding mmu_lock (unless for freeing the stage2 pgd before
323 * destroying the VM), otherwise another faulting VCPU may come in and mess
324 * with things behind our backs.
326 static void unmap_stage2_range(struct kvm *kvm, phys_addr_t start, u64 size)
329 phys_addr_t addr = start, end = start + size;
332 assert_spin_locked(&kvm->mmu_lock);
333 WARN_ON(size & ~PAGE_MASK);
335 pgd = kvm->arch.pgd + stage2_pgd_index(addr);
338 * Make sure the page table is still active, as another thread
339 * could have possibly freed the page table, while we released
342 if (!READ_ONCE(kvm->arch.pgd))
344 next = stage2_pgd_addr_end(addr, end);
345 if (!stage2_pgd_none(*pgd))
346 unmap_stage2_puds(kvm, pgd, addr, next);
348 * If the range is too large, release the kvm->mmu_lock
349 * to prevent starvation and lockup detector warnings.
352 cond_resched_lock(&kvm->mmu_lock);
353 } while (pgd++, addr = next, addr != end);
356 static void stage2_flush_ptes(struct kvm *kvm, pmd_t *pmd,
357 phys_addr_t addr, phys_addr_t end)
361 pte = pte_offset_kernel(pmd, addr);
363 if (!pte_none(*pte) && !kvm_is_device_pfn(pte_pfn(*pte)))
364 kvm_flush_dcache_pte(*pte);
365 } while (pte++, addr += PAGE_SIZE, addr != end);
368 static void stage2_flush_pmds(struct kvm *kvm, pud_t *pud,
369 phys_addr_t addr, phys_addr_t end)
374 pmd = stage2_pmd_offset(pud, addr);
376 next = stage2_pmd_addr_end(addr, end);
377 if (!pmd_none(*pmd)) {
378 if (pmd_thp_or_huge(*pmd))
379 kvm_flush_dcache_pmd(*pmd);
381 stage2_flush_ptes(kvm, pmd, addr, next);
383 } while (pmd++, addr = next, addr != end);
386 static void stage2_flush_puds(struct kvm *kvm, pgd_t *pgd,
387 phys_addr_t addr, phys_addr_t end)
392 pud = stage2_pud_offset(pgd, addr);
394 next = stage2_pud_addr_end(addr, end);
395 if (!stage2_pud_none(*pud)) {
396 if (stage2_pud_huge(*pud))
397 kvm_flush_dcache_pud(*pud);
399 stage2_flush_pmds(kvm, pud, addr, next);
401 } while (pud++, addr = next, addr != end);
404 static void stage2_flush_memslot(struct kvm *kvm,
405 struct kvm_memory_slot *memslot)
407 phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
408 phys_addr_t end = addr + PAGE_SIZE * memslot->npages;
412 pgd = kvm->arch.pgd + stage2_pgd_index(addr);
414 next = stage2_pgd_addr_end(addr, end);
415 stage2_flush_puds(kvm, pgd, addr, next);
416 } while (pgd++, addr = next, addr != end);
420 * stage2_flush_vm - Invalidate cache for pages mapped in stage 2
421 * @kvm: The struct kvm pointer
423 * Go through the stage 2 page tables and invalidate any cache lines
424 * backing memory already mapped to the VM.
426 static void stage2_flush_vm(struct kvm *kvm)
428 struct kvm_memslots *slots;
429 struct kvm_memory_slot *memslot;
432 idx = srcu_read_lock(&kvm->srcu);
433 spin_lock(&kvm->mmu_lock);
435 slots = kvm_memslots(kvm);
436 kvm_for_each_memslot(memslot, slots)
437 stage2_flush_memslot(kvm, memslot);
439 spin_unlock(&kvm->mmu_lock);
440 srcu_read_unlock(&kvm->srcu, idx);
443 static void clear_hyp_pgd_entry(pgd_t *pgd)
445 pud_t *pud_table __maybe_unused = pud_offset(pgd, 0UL);
447 pud_free(NULL, pud_table);
448 put_page(virt_to_page(pgd));
451 static void clear_hyp_pud_entry(pud_t *pud)
453 pmd_t *pmd_table __maybe_unused = pmd_offset(pud, 0);
454 VM_BUG_ON(pud_huge(*pud));
456 pmd_free(NULL, pmd_table);
457 put_page(virt_to_page(pud));
460 static void clear_hyp_pmd_entry(pmd_t *pmd)
462 pte_t *pte_table = pte_offset_kernel(pmd, 0);
463 VM_BUG_ON(pmd_thp_or_huge(*pmd));
465 pte_free_kernel(NULL, pte_table);
466 put_page(virt_to_page(pmd));
469 static void unmap_hyp_ptes(pmd_t *pmd, phys_addr_t addr, phys_addr_t end)
471 pte_t *pte, *start_pte;
473 start_pte = pte = pte_offset_kernel(pmd, addr);
475 if (!pte_none(*pte)) {
476 kvm_set_pte(pte, __pte(0));
477 put_page(virt_to_page(pte));
479 } while (pte++, addr += PAGE_SIZE, addr != end);
481 if (hyp_pte_table_empty(start_pte))
482 clear_hyp_pmd_entry(pmd);
485 static void unmap_hyp_pmds(pud_t *pud, phys_addr_t addr, phys_addr_t end)
488 pmd_t *pmd, *start_pmd;
490 start_pmd = pmd = pmd_offset(pud, addr);
492 next = pmd_addr_end(addr, end);
493 /* Hyp doesn't use huge pmds */
495 unmap_hyp_ptes(pmd, addr, next);
496 } while (pmd++, addr = next, addr != end);
498 if (hyp_pmd_table_empty(start_pmd))
499 clear_hyp_pud_entry(pud);
502 static void unmap_hyp_puds(pgd_t *pgd, phys_addr_t addr, phys_addr_t end)
505 pud_t *pud, *start_pud;
507 start_pud = pud = pud_offset(pgd, addr);
509 next = pud_addr_end(addr, end);
510 /* Hyp doesn't use huge puds */
512 unmap_hyp_pmds(pud, addr, next);
513 } while (pud++, addr = next, addr != end);
515 if (hyp_pud_table_empty(start_pud))
516 clear_hyp_pgd_entry(pgd);
519 static unsigned int kvm_pgd_index(unsigned long addr, unsigned int ptrs_per_pgd)
521 return (addr >> PGDIR_SHIFT) & (ptrs_per_pgd - 1);
524 static void __unmap_hyp_range(pgd_t *pgdp, unsigned long ptrs_per_pgd,
525 phys_addr_t start, u64 size)
528 phys_addr_t addr = start, end = start + size;
532 * We don't unmap anything from HYP, except at the hyp tear down.
533 * Hence, we don't have to invalidate the TLBs here.
535 pgd = pgdp + kvm_pgd_index(addr, ptrs_per_pgd);
537 next = pgd_addr_end(addr, end);
539 unmap_hyp_puds(pgd, addr, next);
540 } while (pgd++, addr = next, addr != end);
543 static void unmap_hyp_range(pgd_t *pgdp, phys_addr_t start, u64 size)
545 __unmap_hyp_range(pgdp, PTRS_PER_PGD, start, size);
548 static void unmap_hyp_idmap_range(pgd_t *pgdp, phys_addr_t start, u64 size)
550 __unmap_hyp_range(pgdp, __kvm_idmap_ptrs_per_pgd(), start, size);
554 * free_hyp_pgds - free Hyp-mode page tables
556 * Assumes hyp_pgd is a page table used strictly in Hyp-mode and
557 * therefore contains either mappings in the kernel memory area (above
558 * PAGE_OFFSET), or device mappings in the idmap range.
560 * boot_hyp_pgd should only map the idmap range, and is only used in
561 * the extended idmap case.
563 void free_hyp_pgds(void)
567 mutex_lock(&kvm_hyp_pgd_mutex);
569 id_pgd = boot_hyp_pgd ? boot_hyp_pgd : hyp_pgd;
572 /* In case we never called hyp_mmu_init() */
574 io_map_base = hyp_idmap_start;
575 unmap_hyp_idmap_range(id_pgd, io_map_base,
576 hyp_idmap_start + PAGE_SIZE - io_map_base);
580 free_pages((unsigned long)boot_hyp_pgd, hyp_pgd_order);
585 unmap_hyp_range(hyp_pgd, kern_hyp_va(PAGE_OFFSET),
586 (uintptr_t)high_memory - PAGE_OFFSET);
588 free_pages((unsigned long)hyp_pgd, hyp_pgd_order);
591 if (merged_hyp_pgd) {
592 clear_page(merged_hyp_pgd);
593 free_page((unsigned long)merged_hyp_pgd);
594 merged_hyp_pgd = NULL;
597 mutex_unlock(&kvm_hyp_pgd_mutex);
600 static void create_hyp_pte_mappings(pmd_t *pmd, unsigned long start,
601 unsigned long end, unsigned long pfn,
609 pte = pte_offset_kernel(pmd, addr);
610 kvm_set_pte(pte, pfn_pte(pfn, prot));
611 get_page(virt_to_page(pte));
613 } while (addr += PAGE_SIZE, addr != end);
616 static int create_hyp_pmd_mappings(pud_t *pud, unsigned long start,
617 unsigned long end, unsigned long pfn,
622 unsigned long addr, next;
626 pmd = pmd_offset(pud, addr);
628 BUG_ON(pmd_sect(*pmd));
630 if (pmd_none(*pmd)) {
631 pte = pte_alloc_one_kernel(NULL, addr);
633 kvm_err("Cannot allocate Hyp pte\n");
636 kvm_pmd_populate(pmd, pte);
637 get_page(virt_to_page(pmd));
640 next = pmd_addr_end(addr, end);
642 create_hyp_pte_mappings(pmd, addr, next, pfn, prot);
643 pfn += (next - addr) >> PAGE_SHIFT;
644 } while (addr = next, addr != end);
649 static int create_hyp_pud_mappings(pgd_t *pgd, unsigned long start,
650 unsigned long end, unsigned long pfn,
655 unsigned long addr, next;
660 pud = pud_offset(pgd, addr);
662 if (pud_none_or_clear_bad(pud)) {
663 pmd = pmd_alloc_one(NULL, addr);
665 kvm_err("Cannot allocate Hyp pmd\n");
668 kvm_pud_populate(pud, pmd);
669 get_page(virt_to_page(pud));
672 next = pud_addr_end(addr, end);
673 ret = create_hyp_pmd_mappings(pud, addr, next, pfn, prot);
676 pfn += (next - addr) >> PAGE_SHIFT;
677 } while (addr = next, addr != end);
682 static int __create_hyp_mappings(pgd_t *pgdp, unsigned long ptrs_per_pgd,
683 unsigned long start, unsigned long end,
684 unsigned long pfn, pgprot_t prot)
688 unsigned long addr, next;
691 mutex_lock(&kvm_hyp_pgd_mutex);
692 addr = start & PAGE_MASK;
693 end = PAGE_ALIGN(end);
695 pgd = pgdp + kvm_pgd_index(addr, ptrs_per_pgd);
697 if (pgd_none(*pgd)) {
698 pud = pud_alloc_one(NULL, addr);
700 kvm_err("Cannot allocate Hyp pud\n");
704 kvm_pgd_populate(pgd, pud);
705 get_page(virt_to_page(pgd));
708 next = pgd_addr_end(addr, end);
709 err = create_hyp_pud_mappings(pgd, addr, next, pfn, prot);
712 pfn += (next - addr) >> PAGE_SHIFT;
713 } while (addr = next, addr != end);
715 mutex_unlock(&kvm_hyp_pgd_mutex);
719 static phys_addr_t kvm_kaddr_to_phys(void *kaddr)
721 if (!is_vmalloc_addr(kaddr)) {
722 BUG_ON(!virt_addr_valid(kaddr));
725 return page_to_phys(vmalloc_to_page(kaddr)) +
726 offset_in_page(kaddr);
731 * create_hyp_mappings - duplicate a kernel virtual address range in Hyp mode
732 * @from: The virtual kernel start address of the range
733 * @to: The virtual kernel end address of the range (exclusive)
734 * @prot: The protection to be applied to this range
736 * The same virtual address as the kernel virtual address is also used
737 * in Hyp-mode mapping (modulo HYP_PAGE_OFFSET) to the same underlying
740 int create_hyp_mappings(void *from, void *to, pgprot_t prot)
742 phys_addr_t phys_addr;
743 unsigned long virt_addr;
744 unsigned long start = kern_hyp_va((unsigned long)from);
745 unsigned long end = kern_hyp_va((unsigned long)to);
747 if (is_kernel_in_hyp_mode())
750 start = start & PAGE_MASK;
751 end = PAGE_ALIGN(end);
753 for (virt_addr = start; virt_addr < end; virt_addr += PAGE_SIZE) {
756 phys_addr = kvm_kaddr_to_phys(from + virt_addr - start);
757 err = __create_hyp_mappings(hyp_pgd, PTRS_PER_PGD,
758 virt_addr, virt_addr + PAGE_SIZE,
759 __phys_to_pfn(phys_addr),
768 static int __create_hyp_private_mapping(phys_addr_t phys_addr, size_t size,
769 unsigned long *haddr, pgprot_t prot)
771 pgd_t *pgd = hyp_pgd;
775 mutex_lock(&kvm_hyp_pgd_mutex);
778 * This assumes that we we have enough space below the idmap
779 * page to allocate our VAs. If not, the check below will
780 * kick. A potential alternative would be to detect that
781 * overflow and switch to an allocation above the idmap.
783 * The allocated size is always a multiple of PAGE_SIZE.
785 size = PAGE_ALIGN(size + offset_in_page(phys_addr));
786 base = io_map_base - size;
789 * Verify that BIT(VA_BITS - 1) hasn't been flipped by
790 * allocating the new area, as it would indicate we've
791 * overflowed the idmap/IO address range.
793 if ((base ^ io_map_base) & BIT(VA_BITS - 1))
798 mutex_unlock(&kvm_hyp_pgd_mutex);
803 if (__kvm_cpu_uses_extended_idmap())
806 ret = __create_hyp_mappings(pgd, __kvm_idmap_ptrs_per_pgd(),
808 __phys_to_pfn(phys_addr), prot);
812 *haddr = base + offset_in_page(phys_addr);
819 * create_hyp_io_mappings - Map IO into both kernel and HYP
820 * @phys_addr: The physical start address which gets mapped
821 * @size: Size of the region being mapped
822 * @kaddr: Kernel VA for this mapping
823 * @haddr: HYP VA for this mapping
825 int create_hyp_io_mappings(phys_addr_t phys_addr, size_t size,
826 void __iomem **kaddr,
827 void __iomem **haddr)
832 *kaddr = ioremap(phys_addr, size);
836 if (is_kernel_in_hyp_mode()) {
841 ret = __create_hyp_private_mapping(phys_addr, size,
842 &addr, PAGE_HYP_DEVICE);
850 *haddr = (void __iomem *)addr;
855 * create_hyp_exec_mappings - Map an executable range into HYP
856 * @phys_addr: The physical start address which gets mapped
857 * @size: Size of the region being mapped
858 * @haddr: HYP VA for this mapping
860 int create_hyp_exec_mappings(phys_addr_t phys_addr, size_t size,
866 BUG_ON(is_kernel_in_hyp_mode());
868 ret = __create_hyp_private_mapping(phys_addr, size,
869 &addr, PAGE_HYP_EXEC);
875 *haddr = (void *)addr;
880 * kvm_alloc_stage2_pgd - allocate level-1 table for stage-2 translation.
881 * @kvm: The KVM struct pointer for the VM.
883 * Allocates only the stage-2 HW PGD level table(s) (can support either full
884 * 40-bit input addresses or limited to 32-bit input addresses). Clears the
887 * Note we don't need locking here as this is only called when the VM is
888 * created, which can only be done once.
890 int kvm_alloc_stage2_pgd(struct kvm *kvm)
894 if (kvm->arch.pgd != NULL) {
895 kvm_err("kvm_arch already initialized?\n");
899 /* Allocate the HW PGD, making sure that each page gets its own refcount */
900 pgd = alloc_pages_exact(S2_PGD_SIZE, GFP_KERNEL | __GFP_ZERO);
908 static void stage2_unmap_memslot(struct kvm *kvm,
909 struct kvm_memory_slot *memslot)
911 hva_t hva = memslot->userspace_addr;
912 phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
913 phys_addr_t size = PAGE_SIZE * memslot->npages;
914 hva_t reg_end = hva + size;
917 * A memory region could potentially cover multiple VMAs, and any holes
918 * between them, so iterate over all of them to find out if we should
921 * +--------------------------------------------+
922 * +---------------+----------------+ +----------------+
923 * | : VMA 1 | VMA 2 | | VMA 3 : |
924 * +---------------+----------------+ +----------------+
926 * +--------------------------------------------+
929 struct vm_area_struct *vma = find_vma(current->mm, hva);
930 hva_t vm_start, vm_end;
932 if (!vma || vma->vm_start >= reg_end)
936 * Take the intersection of this VMA with the memory region
938 vm_start = max(hva, vma->vm_start);
939 vm_end = min(reg_end, vma->vm_end);
941 if (!(vma->vm_flags & VM_PFNMAP)) {
942 gpa_t gpa = addr + (vm_start - memslot->userspace_addr);
943 unmap_stage2_range(kvm, gpa, vm_end - vm_start);
946 } while (hva < reg_end);
950 * stage2_unmap_vm - Unmap Stage-2 RAM mappings
951 * @kvm: The struct kvm pointer
953 * Go through the memregions and unmap any reguler RAM
954 * backing memory already mapped to the VM.
956 void stage2_unmap_vm(struct kvm *kvm)
958 struct kvm_memslots *slots;
959 struct kvm_memory_slot *memslot;
962 idx = srcu_read_lock(&kvm->srcu);
963 down_read(¤t->mm->mmap_sem);
964 spin_lock(&kvm->mmu_lock);
966 slots = kvm_memslots(kvm);
967 kvm_for_each_memslot(memslot, slots)
968 stage2_unmap_memslot(kvm, memslot);
970 spin_unlock(&kvm->mmu_lock);
971 up_read(¤t->mm->mmap_sem);
972 srcu_read_unlock(&kvm->srcu, idx);
976 * kvm_free_stage2_pgd - free all stage-2 tables
977 * @kvm: The KVM struct pointer for the VM.
979 * Walks the level-1 page table pointed to by kvm->arch.pgd and frees all
980 * underlying level-2 and level-3 tables before freeing the actual level-1 table
981 * and setting the struct pointer to NULL.
983 void kvm_free_stage2_pgd(struct kvm *kvm)
987 spin_lock(&kvm->mmu_lock);
989 unmap_stage2_range(kvm, 0, KVM_PHYS_SIZE);
990 pgd = READ_ONCE(kvm->arch.pgd);
991 kvm->arch.pgd = NULL;
993 spin_unlock(&kvm->mmu_lock);
995 /* Free the HW pgd, one page at a time */
997 free_pages_exact(pgd, S2_PGD_SIZE);
1000 static pud_t *stage2_get_pud(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
1006 pgd = kvm->arch.pgd + stage2_pgd_index(addr);
1007 if (WARN_ON(stage2_pgd_none(*pgd))) {
1010 pud = mmu_memory_cache_alloc(cache);
1011 stage2_pgd_populate(pgd, pud);
1012 get_page(virt_to_page(pgd));
1015 return stage2_pud_offset(pgd, addr);
1018 static pmd_t *stage2_get_pmd(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
1024 pud = stage2_get_pud(kvm, cache, addr);
1028 if (stage2_pud_none(*pud)) {
1031 pmd = mmu_memory_cache_alloc(cache);
1032 stage2_pud_populate(pud, pmd);
1033 get_page(virt_to_page(pud));
1036 return stage2_pmd_offset(pud, addr);
1039 static int stage2_set_pmd_huge(struct kvm *kvm, struct kvm_mmu_memory_cache
1040 *cache, phys_addr_t addr, const pmd_t *new_pmd)
1042 pmd_t *pmd, old_pmd;
1044 pmd = stage2_get_pmd(kvm, cache, addr);
1048 if (pmd_present(old_pmd)) {
1050 * Multiple vcpus faulting on the same PMD entry, can
1051 * lead to them sequentially updating the PMD with the
1052 * same value. Following the break-before-make
1053 * (pmd_clear() followed by tlb_flush()) process can
1054 * hinder forward progress due to refaults generated
1055 * on missing translations.
1057 * Skip updating the page table if the entry is
1060 if (pmd_val(old_pmd) == pmd_val(*new_pmd))
1064 * Mapping in huge pages should only happen through a
1065 * fault. If a page is merged into a transparent huge
1066 * page, the individual subpages of that huge page
1067 * should be unmapped through MMU notifiers before we
1070 * Merging of CompoundPages is not supported; they
1071 * should become splitting first, unmapped, merged,
1072 * and mapped back in on-demand.
1074 VM_BUG_ON(pmd_pfn(old_pmd) != pmd_pfn(*new_pmd));
1077 kvm_tlb_flush_vmid_ipa(kvm, addr);
1079 get_page(virt_to_page(pmd));
1082 kvm_set_pmd(pmd, *new_pmd);
1086 static bool stage2_is_exec(struct kvm *kvm, phys_addr_t addr)
1091 pmdp = stage2_get_pmd(kvm, NULL, addr);
1092 if (!pmdp || pmd_none(*pmdp) || !pmd_present(*pmdp))
1095 if (pmd_thp_or_huge(*pmdp))
1096 return kvm_s2pmd_exec(pmdp);
1098 ptep = pte_offset_kernel(pmdp, addr);
1099 if (!ptep || pte_none(*ptep) || !pte_present(*ptep))
1102 return kvm_s2pte_exec(ptep);
1105 static int stage2_set_pte(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
1106 phys_addr_t addr, const pte_t *new_pte,
1107 unsigned long flags)
1110 pte_t *pte, old_pte;
1111 bool iomap = flags & KVM_S2PTE_FLAG_IS_IOMAP;
1112 bool logging_active = flags & KVM_S2_FLAG_LOGGING_ACTIVE;
1114 VM_BUG_ON(logging_active && !cache);
1116 /* Create stage-2 page table mapping - Levels 0 and 1 */
1117 pmd = stage2_get_pmd(kvm, cache, addr);
1120 * Ignore calls from kvm_set_spte_hva for unallocated
1127 * While dirty page logging - dissolve huge PMD, then continue on to
1131 stage2_dissolve_pmd(kvm, addr, pmd);
1133 /* Create stage-2 page mappings - Level 2 */
1134 if (pmd_none(*pmd)) {
1136 return 0; /* ignore calls from kvm_set_spte_hva */
1137 pte = mmu_memory_cache_alloc(cache);
1138 kvm_pmd_populate(pmd, pte);
1139 get_page(virt_to_page(pmd));
1142 pte = pte_offset_kernel(pmd, addr);
1144 if (iomap && pte_present(*pte))
1147 /* Create 2nd stage page table mapping - Level 3 */
1149 if (pte_present(old_pte)) {
1150 /* Skip page table update if there is no change */
1151 if (pte_val(old_pte) == pte_val(*new_pte))
1154 kvm_set_pte(pte, __pte(0));
1155 kvm_tlb_flush_vmid_ipa(kvm, addr);
1157 get_page(virt_to_page(pte));
1160 kvm_set_pte(pte, *new_pte);
1164 #ifndef __HAVE_ARCH_PTEP_TEST_AND_CLEAR_YOUNG
1165 static int stage2_ptep_test_and_clear_young(pte_t *pte)
1167 if (pte_young(*pte)) {
1168 *pte = pte_mkold(*pte);
1174 static int stage2_ptep_test_and_clear_young(pte_t *pte)
1176 return __ptep_test_and_clear_young(pte);
1180 static int stage2_pmdp_test_and_clear_young(pmd_t *pmd)
1182 return stage2_ptep_test_and_clear_young((pte_t *)pmd);
1186 * kvm_phys_addr_ioremap - map a device range to guest IPA
1188 * @kvm: The KVM pointer
1189 * @guest_ipa: The IPA at which to insert the mapping
1190 * @pa: The physical address of the device
1191 * @size: The size of the mapping
1193 int kvm_phys_addr_ioremap(struct kvm *kvm, phys_addr_t guest_ipa,
1194 phys_addr_t pa, unsigned long size, bool writable)
1196 phys_addr_t addr, end;
1199 struct kvm_mmu_memory_cache cache = { 0, };
1201 end = (guest_ipa + size + PAGE_SIZE - 1) & PAGE_MASK;
1202 pfn = __phys_to_pfn(pa);
1204 for (addr = guest_ipa; addr < end; addr += PAGE_SIZE) {
1205 pte_t pte = pfn_pte(pfn, PAGE_S2_DEVICE);
1208 pte = kvm_s2pte_mkwrite(pte);
1210 ret = mmu_topup_memory_cache(&cache, KVM_MMU_CACHE_MIN_PAGES,
1214 spin_lock(&kvm->mmu_lock);
1215 ret = stage2_set_pte(kvm, &cache, addr, &pte,
1216 KVM_S2PTE_FLAG_IS_IOMAP);
1217 spin_unlock(&kvm->mmu_lock);
1225 mmu_free_memory_cache(&cache);
1229 static bool transparent_hugepage_adjust(kvm_pfn_t *pfnp, phys_addr_t *ipap)
1231 kvm_pfn_t pfn = *pfnp;
1232 gfn_t gfn = *ipap >> PAGE_SHIFT;
1234 if (PageTransCompoundMap(pfn_to_page(pfn))) {
1237 * The address we faulted on is backed by a transparent huge
1238 * page. However, because we map the compound huge page and
1239 * not the individual tail page, we need to transfer the
1240 * refcount to the head page. We have to be careful that the
1241 * THP doesn't start to split while we are adjusting the
1244 * We are sure this doesn't happen, because mmu_notifier_retry
1245 * was successful and we are holding the mmu_lock, so if this
1246 * THP is trying to split, it will be blocked in the mmu
1247 * notifier before touching any of the pages, specifically
1248 * before being able to call __split_huge_page_refcount().
1250 * We can therefore safely transfer the refcount from PG_tail
1251 * to PG_head and switch the pfn from a tail page to the head
1254 mask = PTRS_PER_PMD - 1;
1255 VM_BUG_ON((gfn & mask) != (pfn & mask));
1258 kvm_release_pfn_clean(pfn);
1270 static bool kvm_is_write_fault(struct kvm_vcpu *vcpu)
1272 if (kvm_vcpu_trap_is_iabt(vcpu))
1275 return kvm_vcpu_dabt_iswrite(vcpu);
1279 * stage2_wp_ptes - write protect PMD range
1280 * @pmd: pointer to pmd entry
1281 * @addr: range start address
1282 * @end: range end address
1284 static void stage2_wp_ptes(pmd_t *pmd, phys_addr_t addr, phys_addr_t end)
1288 pte = pte_offset_kernel(pmd, addr);
1290 if (!pte_none(*pte)) {
1291 if (!kvm_s2pte_readonly(pte))
1292 kvm_set_s2pte_readonly(pte);
1294 } while (pte++, addr += PAGE_SIZE, addr != end);
1298 * stage2_wp_pmds - write protect PUD range
1299 * @pud: pointer to pud entry
1300 * @addr: range start address
1301 * @end: range end address
1303 static void stage2_wp_pmds(pud_t *pud, phys_addr_t addr, phys_addr_t end)
1308 pmd = stage2_pmd_offset(pud, addr);
1311 next = stage2_pmd_addr_end(addr, end);
1312 if (!pmd_none(*pmd)) {
1313 if (pmd_thp_or_huge(*pmd)) {
1314 if (!kvm_s2pmd_readonly(pmd))
1315 kvm_set_s2pmd_readonly(pmd);
1317 stage2_wp_ptes(pmd, addr, next);
1320 } while (pmd++, addr = next, addr != end);
1324 * stage2_wp_puds - write protect PGD range
1325 * @pgd: pointer to pgd entry
1326 * @addr: range start address
1327 * @end: range end address
1329 * Process PUD entries, for a huge PUD we cause a panic.
1331 static void stage2_wp_puds(pgd_t *pgd, phys_addr_t addr, phys_addr_t end)
1336 pud = stage2_pud_offset(pgd, addr);
1338 next = stage2_pud_addr_end(addr, end);
1339 if (!stage2_pud_none(*pud)) {
1340 /* TODO:PUD not supported, revisit later if supported */
1341 BUG_ON(stage2_pud_huge(*pud));
1342 stage2_wp_pmds(pud, addr, next);
1344 } while (pud++, addr = next, addr != end);
1348 * stage2_wp_range() - write protect stage2 memory region range
1349 * @kvm: The KVM pointer
1350 * @addr: Start address of range
1351 * @end: End address of range
1353 static void stage2_wp_range(struct kvm *kvm, phys_addr_t addr, phys_addr_t end)
1358 pgd = kvm->arch.pgd + stage2_pgd_index(addr);
1361 * Release kvm_mmu_lock periodically if the memory region is
1362 * large. Otherwise, we may see kernel panics with
1363 * CONFIG_DETECT_HUNG_TASK, CONFIG_LOCKUP_DETECTOR,
1364 * CONFIG_LOCKDEP. Additionally, holding the lock too long
1365 * will also starve other vCPUs. We have to also make sure
1366 * that the page tables are not freed while we released
1369 cond_resched_lock(&kvm->mmu_lock);
1370 if (!READ_ONCE(kvm->arch.pgd))
1372 next = stage2_pgd_addr_end(addr, end);
1373 if (stage2_pgd_present(*pgd))
1374 stage2_wp_puds(pgd, addr, next);
1375 } while (pgd++, addr = next, addr != end);
1379 * kvm_mmu_wp_memory_region() - write protect stage 2 entries for memory slot
1380 * @kvm: The KVM pointer
1381 * @slot: The memory slot to write protect
1383 * Called to start logging dirty pages after memory region
1384 * KVM_MEM_LOG_DIRTY_PAGES operation is called. After this function returns
1385 * all present PMD and PTEs are write protected in the memory region.
1386 * Afterwards read of dirty page log can be called.
1388 * Acquires kvm_mmu_lock. Called with kvm->slots_lock mutex acquired,
1389 * serializing operations for VM memory regions.
1391 void kvm_mmu_wp_memory_region(struct kvm *kvm, int slot)
1393 struct kvm_memslots *slots = kvm_memslots(kvm);
1394 struct kvm_memory_slot *memslot = id_to_memslot(slots, slot);
1395 phys_addr_t start = memslot->base_gfn << PAGE_SHIFT;
1396 phys_addr_t end = (memslot->base_gfn + memslot->npages) << PAGE_SHIFT;
1398 spin_lock(&kvm->mmu_lock);
1399 stage2_wp_range(kvm, start, end);
1400 spin_unlock(&kvm->mmu_lock);
1401 kvm_flush_remote_tlbs(kvm);
1405 * kvm_mmu_write_protect_pt_masked() - write protect dirty pages
1406 * @kvm: The KVM pointer
1407 * @slot: The memory slot associated with mask
1408 * @gfn_offset: The gfn offset in memory slot
1409 * @mask: The mask of dirty pages at offset 'gfn_offset' in this memory
1410 * slot to be write protected
1412 * Walks bits set in mask write protects the associated pte's. Caller must
1413 * acquire kvm_mmu_lock.
1415 static void kvm_mmu_write_protect_pt_masked(struct kvm *kvm,
1416 struct kvm_memory_slot *slot,
1417 gfn_t gfn_offset, unsigned long mask)
1419 phys_addr_t base_gfn = slot->base_gfn + gfn_offset;
1420 phys_addr_t start = (base_gfn + __ffs(mask)) << PAGE_SHIFT;
1421 phys_addr_t end = (base_gfn + __fls(mask) + 1) << PAGE_SHIFT;
1423 stage2_wp_range(kvm, start, end);
1427 * kvm_arch_mmu_enable_log_dirty_pt_masked - enable dirty logging for selected
1430 * It calls kvm_mmu_write_protect_pt_masked to write protect selected pages to
1431 * enable dirty logging for them.
1433 void kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm *kvm,
1434 struct kvm_memory_slot *slot,
1435 gfn_t gfn_offset, unsigned long mask)
1437 kvm_mmu_write_protect_pt_masked(kvm, slot, gfn_offset, mask);
1440 static void clean_dcache_guest_page(kvm_pfn_t pfn, unsigned long size)
1442 __clean_dcache_guest_page(pfn, size);
1445 static void invalidate_icache_guest_page(kvm_pfn_t pfn, unsigned long size)
1447 __invalidate_icache_guest_page(pfn, size);
1450 static void kvm_send_hwpoison_signal(unsigned long address,
1451 struct vm_area_struct *vma)
1455 if (is_vm_hugetlb_page(vma))
1456 lsb = huge_page_shift(hstate_vma(vma));
1460 send_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, lsb, current);
1463 static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
1464 struct kvm_memory_slot *memslot, unsigned long hva,
1465 unsigned long fault_status)
1468 bool write_fault, exec_fault, writable, hugetlb = false, force_pte = false;
1469 unsigned long mmu_seq;
1470 gfn_t gfn = fault_ipa >> PAGE_SHIFT;
1471 struct kvm *kvm = vcpu->kvm;
1472 struct kvm_mmu_memory_cache *memcache = &vcpu->arch.mmu_page_cache;
1473 struct vm_area_struct *vma;
1475 pgprot_t mem_type = PAGE_S2;
1476 bool logging_active = memslot_is_logging(memslot);
1477 unsigned long flags = 0;
1479 write_fault = kvm_is_write_fault(vcpu);
1480 exec_fault = kvm_vcpu_trap_is_iabt(vcpu);
1481 VM_BUG_ON(write_fault && exec_fault);
1483 if (fault_status == FSC_PERM && !write_fault && !exec_fault) {
1484 kvm_err("Unexpected L2 read permission error\n");
1488 /* Let's check if we will get back a huge page backed by hugetlbfs */
1489 down_read(¤t->mm->mmap_sem);
1490 vma = find_vma_intersection(current->mm, hva, hva + 1);
1491 if (unlikely(!vma)) {
1492 kvm_err("Failed to find VMA for hva 0x%lx\n", hva);
1493 up_read(¤t->mm->mmap_sem);
1497 if (vma_kernel_pagesize(vma) == PMD_SIZE && !logging_active) {
1499 gfn = (fault_ipa & PMD_MASK) >> PAGE_SHIFT;
1502 * Pages belonging to memslots that don't have the same
1503 * alignment for userspace and IPA cannot be mapped using
1504 * block descriptors even if the pages belong to a THP for
1505 * the process, because the stage-2 block descriptor will
1506 * cover more than a single THP and we loose atomicity for
1507 * unmapping, updates, and splits of the THP or other pages
1508 * in the stage-2 block range.
1510 if ((memslot->userspace_addr & ~PMD_MASK) !=
1511 ((memslot->base_gfn << PAGE_SHIFT) & ~PMD_MASK))
1514 up_read(¤t->mm->mmap_sem);
1516 /* We need minimum second+third level pages */
1517 ret = mmu_topup_memory_cache(memcache, KVM_MMU_CACHE_MIN_PAGES,
1522 mmu_seq = vcpu->kvm->mmu_notifier_seq;
1524 * Ensure the read of mmu_notifier_seq happens before we call
1525 * gfn_to_pfn_prot (which calls get_user_pages), so that we don't risk
1526 * the page we just got a reference to gets unmapped before we have a
1527 * chance to grab the mmu_lock, which ensure that if the page gets
1528 * unmapped afterwards, the call to kvm_unmap_hva will take it away
1529 * from us again properly. This smp_rmb() interacts with the smp_wmb()
1530 * in kvm_mmu_notifier_invalidate_<page|range_end>.
1534 pfn = gfn_to_pfn_prot(kvm, gfn, write_fault, &writable);
1535 if (pfn == KVM_PFN_ERR_HWPOISON) {
1536 kvm_send_hwpoison_signal(hva, vma);
1539 if (is_error_noslot_pfn(pfn))
1542 if (kvm_is_device_pfn(pfn)) {
1543 mem_type = PAGE_S2_DEVICE;
1544 flags |= KVM_S2PTE_FLAG_IS_IOMAP;
1545 } else if (logging_active) {
1547 * Faults on pages in a memslot with logging enabled
1548 * should not be mapped with huge pages (it introduces churn
1549 * and performance degradation), so force a pte mapping.
1552 flags |= KVM_S2_FLAG_LOGGING_ACTIVE;
1555 * Only actually map the page as writable if this was a write
1562 spin_lock(&kvm->mmu_lock);
1563 if (mmu_notifier_retry(kvm, mmu_seq))
1566 if (!hugetlb && !force_pte)
1567 hugetlb = transparent_hugepage_adjust(&pfn, &fault_ipa);
1570 pmd_t new_pmd = pfn_pmd(pfn, mem_type);
1571 new_pmd = pmd_mkhuge(new_pmd);
1573 new_pmd = kvm_s2pmd_mkwrite(new_pmd);
1574 kvm_set_pfn_dirty(pfn);
1577 if (fault_status != FSC_PERM)
1578 clean_dcache_guest_page(pfn, PMD_SIZE);
1581 new_pmd = kvm_s2pmd_mkexec(new_pmd);
1582 invalidate_icache_guest_page(pfn, PMD_SIZE);
1583 } else if (fault_status == FSC_PERM) {
1584 /* Preserve execute if XN was already cleared */
1585 if (stage2_is_exec(kvm, fault_ipa))
1586 new_pmd = kvm_s2pmd_mkexec(new_pmd);
1589 ret = stage2_set_pmd_huge(kvm, memcache, fault_ipa, &new_pmd);
1591 pte_t new_pte = pfn_pte(pfn, mem_type);
1594 new_pte = kvm_s2pte_mkwrite(new_pte);
1595 kvm_set_pfn_dirty(pfn);
1596 mark_page_dirty(kvm, gfn);
1599 if (fault_status != FSC_PERM)
1600 clean_dcache_guest_page(pfn, PAGE_SIZE);
1603 new_pte = kvm_s2pte_mkexec(new_pte);
1604 invalidate_icache_guest_page(pfn, PAGE_SIZE);
1605 } else if (fault_status == FSC_PERM) {
1606 /* Preserve execute if XN was already cleared */
1607 if (stage2_is_exec(kvm, fault_ipa))
1608 new_pte = kvm_s2pte_mkexec(new_pte);
1611 ret = stage2_set_pte(kvm, memcache, fault_ipa, &new_pte, flags);
1615 spin_unlock(&kvm->mmu_lock);
1616 kvm_set_pfn_accessed(pfn);
1617 kvm_release_pfn_clean(pfn);
1622 * Resolve the access fault by making the page young again.
1623 * Note that because the faulting entry is guaranteed not to be
1624 * cached in the TLB, we don't need to invalidate anything.
1625 * Only the HW Access Flag updates are supported for Stage 2 (no DBM),
1626 * so there is no need for atomic (pte|pmd)_mkyoung operations.
1628 static void handle_access_fault(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa)
1633 bool pfn_valid = false;
1635 trace_kvm_access_fault(fault_ipa);
1637 spin_lock(&vcpu->kvm->mmu_lock);
1639 pmd = stage2_get_pmd(vcpu->kvm, NULL, fault_ipa);
1640 if (!pmd || pmd_none(*pmd)) /* Nothing there */
1643 if (pmd_thp_or_huge(*pmd)) { /* THP, HugeTLB */
1644 *pmd = pmd_mkyoung(*pmd);
1645 pfn = pmd_pfn(*pmd);
1650 pte = pte_offset_kernel(pmd, fault_ipa);
1651 if (pte_none(*pte)) /* Nothing there either */
1654 *pte = pte_mkyoung(*pte); /* Just a page... */
1655 pfn = pte_pfn(*pte);
1658 spin_unlock(&vcpu->kvm->mmu_lock);
1660 kvm_set_pfn_accessed(pfn);
1664 * kvm_handle_guest_abort - handles all 2nd stage aborts
1665 * @vcpu: the VCPU pointer
1666 * @run: the kvm_run structure
1668 * Any abort that gets to the host is almost guaranteed to be caused by a
1669 * missing second stage translation table entry, which can mean that either the
1670 * guest simply needs more memory and we must allocate an appropriate page or it
1671 * can mean that the guest tried to access I/O memory, which is emulated by user
1672 * space. The distinction is based on the IPA causing the fault and whether this
1673 * memory region has been registered as standard RAM by user space.
1675 int kvm_handle_guest_abort(struct kvm_vcpu *vcpu, struct kvm_run *run)
1677 unsigned long fault_status;
1678 phys_addr_t fault_ipa;
1679 struct kvm_memory_slot *memslot;
1681 bool is_iabt, write_fault, writable;
1685 fault_status = kvm_vcpu_trap_get_fault_type(vcpu);
1687 fault_ipa = kvm_vcpu_get_fault_ipa(vcpu);
1688 is_iabt = kvm_vcpu_trap_is_iabt(vcpu);
1690 /* Synchronous External Abort? */
1691 if (kvm_vcpu_dabt_isextabt(vcpu)) {
1693 * For RAS the host kernel may handle this abort.
1694 * There is no need to pass the error into the guest.
1696 if (!handle_guest_sea(fault_ipa, kvm_vcpu_get_hsr(vcpu)))
1699 if (unlikely(!is_iabt)) {
1700 kvm_inject_vabt(vcpu);
1705 trace_kvm_guest_fault(*vcpu_pc(vcpu), kvm_vcpu_get_hsr(vcpu),
1706 kvm_vcpu_get_hfar(vcpu), fault_ipa);
1708 /* Check the stage-2 fault is trans. fault or write fault */
1709 if (fault_status != FSC_FAULT && fault_status != FSC_PERM &&
1710 fault_status != FSC_ACCESS) {
1711 kvm_err("Unsupported FSC: EC=%#x xFSC=%#lx ESR_EL2=%#lx\n",
1712 kvm_vcpu_trap_get_class(vcpu),
1713 (unsigned long)kvm_vcpu_trap_get_fault(vcpu),
1714 (unsigned long)kvm_vcpu_get_hsr(vcpu));
1718 idx = srcu_read_lock(&vcpu->kvm->srcu);
1720 gfn = fault_ipa >> PAGE_SHIFT;
1721 memslot = gfn_to_memslot(vcpu->kvm, gfn);
1722 hva = gfn_to_hva_memslot_prot(memslot, gfn, &writable);
1723 write_fault = kvm_is_write_fault(vcpu);
1724 if (kvm_is_error_hva(hva) || (write_fault && !writable)) {
1726 /* Prefetch Abort on I/O address */
1727 kvm_inject_pabt(vcpu, kvm_vcpu_get_hfar(vcpu));
1733 * Check for a cache maintenance operation. Since we
1734 * ended-up here, we know it is outside of any memory
1735 * slot. But we can't find out if that is for a device,
1736 * or if the guest is just being stupid. The only thing
1737 * we know for sure is that this range cannot be cached.
1739 * So let's assume that the guest is just being
1740 * cautious, and skip the instruction.
1742 if (kvm_vcpu_dabt_is_cm(vcpu)) {
1743 kvm_skip_instr(vcpu, kvm_vcpu_trap_il_is32bit(vcpu));
1749 * The IPA is reported as [MAX:12], so we need to
1750 * complement it with the bottom 12 bits from the
1751 * faulting VA. This is always 12 bits, irrespective
1754 fault_ipa |= kvm_vcpu_get_hfar(vcpu) & ((1 << 12) - 1);
1755 ret = io_mem_abort(vcpu, run, fault_ipa);
1759 /* Userspace should not be able to register out-of-bounds IPAs */
1760 VM_BUG_ON(fault_ipa >= KVM_PHYS_SIZE);
1762 if (fault_status == FSC_ACCESS) {
1763 handle_access_fault(vcpu, fault_ipa);
1768 ret = user_mem_abort(vcpu, fault_ipa, memslot, hva, fault_status);
1772 srcu_read_unlock(&vcpu->kvm->srcu, idx);
1776 static int handle_hva_to_gpa(struct kvm *kvm,
1777 unsigned long start,
1779 int (*handler)(struct kvm *kvm,
1780 gpa_t gpa, u64 size,
1784 struct kvm_memslots *slots;
1785 struct kvm_memory_slot *memslot;
1788 slots = kvm_memslots(kvm);
1790 /* we only care about the pages that the guest sees */
1791 kvm_for_each_memslot(memslot, slots) {
1792 unsigned long hva_start, hva_end;
1795 hva_start = max(start, memslot->userspace_addr);
1796 hva_end = min(end, memslot->userspace_addr +
1797 (memslot->npages << PAGE_SHIFT));
1798 if (hva_start >= hva_end)
1801 gpa = hva_to_gfn_memslot(hva_start, memslot) << PAGE_SHIFT;
1802 ret |= handler(kvm, gpa, (u64)(hva_end - hva_start), data);
1808 static int kvm_unmap_hva_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
1810 unmap_stage2_range(kvm, gpa, size);
1814 int kvm_unmap_hva_range(struct kvm *kvm,
1815 unsigned long start, unsigned long end)
1820 trace_kvm_unmap_hva_range(start, end);
1821 handle_hva_to_gpa(kvm, start, end, &kvm_unmap_hva_handler, NULL);
1825 static int kvm_set_spte_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
1827 pte_t *pte = (pte_t *)data;
1829 WARN_ON(size != PAGE_SIZE);
1831 * We can always call stage2_set_pte with KVM_S2PTE_FLAG_LOGGING_ACTIVE
1832 * flag clear because MMU notifiers will have unmapped a huge PMD before
1833 * calling ->change_pte() (which in turn calls kvm_set_spte_hva()) and
1834 * therefore stage2_set_pte() never needs to clear out a huge PMD
1835 * through this calling path.
1837 stage2_set_pte(kvm, NULL, gpa, pte, 0);
1842 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
1844 unsigned long end = hva + PAGE_SIZE;
1845 kvm_pfn_t pfn = pte_pfn(pte);
1851 trace_kvm_set_spte_hva(hva);
1854 * We've moved a page around, probably through CoW, so let's treat it
1855 * just like a translation fault and clean the cache to the PoC.
1857 clean_dcache_guest_page(pfn, PAGE_SIZE);
1858 stage2_pte = pfn_pte(pfn, PAGE_S2);
1859 handle_hva_to_gpa(kvm, hva, end, &kvm_set_spte_handler, &stage2_pte);
1862 static int kvm_age_hva_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
1867 WARN_ON(size != PAGE_SIZE && size != PMD_SIZE);
1868 pmd = stage2_get_pmd(kvm, NULL, gpa);
1869 if (!pmd || pmd_none(*pmd)) /* Nothing there */
1872 if (pmd_thp_or_huge(*pmd)) /* THP, HugeTLB */
1873 return stage2_pmdp_test_and_clear_young(pmd);
1875 pte = pte_offset_kernel(pmd, gpa);
1879 return stage2_ptep_test_and_clear_young(pte);
1882 static int kvm_test_age_hva_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
1887 WARN_ON(size != PAGE_SIZE && size != PMD_SIZE);
1888 pmd = stage2_get_pmd(kvm, NULL, gpa);
1889 if (!pmd || pmd_none(*pmd)) /* Nothing there */
1892 if (pmd_thp_or_huge(*pmd)) /* THP, HugeTLB */
1893 return pmd_young(*pmd);
1895 pte = pte_offset_kernel(pmd, gpa);
1896 if (!pte_none(*pte)) /* Just a page... */
1897 return pte_young(*pte);
1902 int kvm_age_hva(struct kvm *kvm, unsigned long start, unsigned long end)
1906 trace_kvm_age_hva(start, end);
1907 return handle_hva_to_gpa(kvm, start, end, kvm_age_hva_handler, NULL);
1910 int kvm_test_age_hva(struct kvm *kvm, unsigned long hva)
1914 trace_kvm_test_age_hva(hva);
1915 return handle_hva_to_gpa(kvm, hva, hva, kvm_test_age_hva_handler, NULL);
1918 void kvm_mmu_free_memory_caches(struct kvm_vcpu *vcpu)
1920 mmu_free_memory_cache(&vcpu->arch.mmu_page_cache);
1923 phys_addr_t kvm_mmu_get_httbr(void)
1925 if (__kvm_cpu_uses_extended_idmap())
1926 return virt_to_phys(merged_hyp_pgd);
1928 return virt_to_phys(hyp_pgd);
1931 phys_addr_t kvm_get_idmap_vector(void)
1933 return hyp_idmap_vector;
1936 static int kvm_map_idmap_text(pgd_t *pgd)
1940 /* Create the idmap in the boot page tables */
1941 err = __create_hyp_mappings(pgd, __kvm_idmap_ptrs_per_pgd(),
1942 hyp_idmap_start, hyp_idmap_end,
1943 __phys_to_pfn(hyp_idmap_start),
1946 kvm_err("Failed to idmap %lx-%lx\n",
1947 hyp_idmap_start, hyp_idmap_end);
1952 int kvm_mmu_init(void)
1956 hyp_idmap_start = kvm_virt_to_phys(__hyp_idmap_text_start);
1957 hyp_idmap_start = ALIGN_DOWN(hyp_idmap_start, PAGE_SIZE);
1958 hyp_idmap_end = kvm_virt_to_phys(__hyp_idmap_text_end);
1959 hyp_idmap_end = ALIGN(hyp_idmap_end, PAGE_SIZE);
1960 hyp_idmap_vector = kvm_virt_to_phys(__kvm_hyp_init);
1963 * We rely on the linker script to ensure at build time that the HYP
1964 * init code does not cross a page boundary.
1966 BUG_ON((hyp_idmap_start ^ (hyp_idmap_end - 1)) & PAGE_MASK);
1968 kvm_debug("IDMAP page: %lx\n", hyp_idmap_start);
1969 kvm_debug("HYP VA range: %lx:%lx\n",
1970 kern_hyp_va(PAGE_OFFSET),
1971 kern_hyp_va((unsigned long)high_memory - 1));
1973 if (hyp_idmap_start >= kern_hyp_va(PAGE_OFFSET) &&
1974 hyp_idmap_start < kern_hyp_va((unsigned long)high_memory - 1) &&
1975 hyp_idmap_start != (unsigned long)__hyp_idmap_text_start) {
1977 * The idmap page is intersecting with the VA space,
1978 * it is not safe to continue further.
1980 kvm_err("IDMAP intersecting with HYP VA, unable to continue\n");
1985 hyp_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, hyp_pgd_order);
1987 kvm_err("Hyp mode PGD not allocated\n");
1992 if (__kvm_cpu_uses_extended_idmap()) {
1993 boot_hyp_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
1995 if (!boot_hyp_pgd) {
1996 kvm_err("Hyp boot PGD not allocated\n");
2001 err = kvm_map_idmap_text(boot_hyp_pgd);
2005 merged_hyp_pgd = (pgd_t *)__get_free_page(GFP_KERNEL | __GFP_ZERO);
2006 if (!merged_hyp_pgd) {
2007 kvm_err("Failed to allocate extra HYP pgd\n");
2010 __kvm_extend_hypmap(boot_hyp_pgd, hyp_pgd, merged_hyp_pgd,
2013 err = kvm_map_idmap_text(hyp_pgd);
2018 io_map_base = hyp_idmap_start;
2025 void kvm_arch_commit_memory_region(struct kvm *kvm,
2026 const struct kvm_userspace_memory_region *mem,
2027 const struct kvm_memory_slot *old,
2028 const struct kvm_memory_slot *new,
2029 enum kvm_mr_change change)
2032 * At this point memslot has been committed and there is an
2033 * allocated dirty_bitmap[], dirty pages will be be tracked while the
2034 * memory slot is write protected.
2036 if (change != KVM_MR_DELETE && mem->flags & KVM_MEM_LOG_DIRTY_PAGES)
2037 kvm_mmu_wp_memory_region(kvm, mem->slot);
2040 int kvm_arch_prepare_memory_region(struct kvm *kvm,
2041 struct kvm_memory_slot *memslot,
2042 const struct kvm_userspace_memory_region *mem,
2043 enum kvm_mr_change change)
2045 hva_t hva = mem->userspace_addr;
2046 hva_t reg_end = hva + mem->memory_size;
2047 bool writable = !(mem->flags & KVM_MEM_READONLY);
2050 if (change != KVM_MR_CREATE && change != KVM_MR_MOVE &&
2051 change != KVM_MR_FLAGS_ONLY)
2055 * Prevent userspace from creating a memory region outside of the IPA
2056 * space addressable by the KVM guest IPA space.
2058 if (memslot->base_gfn + memslot->npages >=
2059 (KVM_PHYS_SIZE >> PAGE_SHIFT))
2062 down_read(¤t->mm->mmap_sem);
2064 * A memory region could potentially cover multiple VMAs, and any holes
2065 * between them, so iterate over all of them to find out if we can map
2066 * any of them right now.
2068 * +--------------------------------------------+
2069 * +---------------+----------------+ +----------------+
2070 * | : VMA 1 | VMA 2 | | VMA 3 : |
2071 * +---------------+----------------+ +----------------+
2073 * +--------------------------------------------+
2076 struct vm_area_struct *vma = find_vma(current->mm, hva);
2077 hva_t vm_start, vm_end;
2079 if (!vma || vma->vm_start >= reg_end)
2083 * Mapping a read-only VMA is only allowed if the
2084 * memory region is configured as read-only.
2086 if (writable && !(vma->vm_flags & VM_WRITE)) {
2092 * Take the intersection of this VMA with the memory region
2094 vm_start = max(hva, vma->vm_start);
2095 vm_end = min(reg_end, vma->vm_end);
2097 if (vma->vm_flags & VM_PFNMAP) {
2098 gpa_t gpa = mem->guest_phys_addr +
2099 (vm_start - mem->userspace_addr);
2102 pa = (phys_addr_t)vma->vm_pgoff << PAGE_SHIFT;
2103 pa += vm_start - vma->vm_start;
2105 /* IO region dirty page logging not allowed */
2106 if (memslot->flags & KVM_MEM_LOG_DIRTY_PAGES) {
2111 ret = kvm_phys_addr_ioremap(kvm, gpa, pa,
2118 } while (hva < reg_end);
2120 if (change == KVM_MR_FLAGS_ONLY)
2123 spin_lock(&kvm->mmu_lock);
2125 unmap_stage2_range(kvm, mem->guest_phys_addr, mem->memory_size);
2127 stage2_flush_memslot(kvm, memslot);
2128 spin_unlock(&kvm->mmu_lock);
2130 up_read(¤t->mm->mmap_sem);
2134 void kvm_arch_free_memslot(struct kvm *kvm, struct kvm_memory_slot *free,
2135 struct kvm_memory_slot *dont)
2139 int kvm_arch_create_memslot(struct kvm *kvm, struct kvm_memory_slot *slot,
2140 unsigned long npages)
2145 void kvm_arch_memslots_updated(struct kvm *kvm, struct kvm_memslots *slots)
2149 void kvm_arch_flush_shadow_all(struct kvm *kvm)
2151 kvm_free_stage2_pgd(kvm);
2154 void kvm_arch_flush_shadow_memslot(struct kvm *kvm,
2155 struct kvm_memory_slot *slot)
2157 gpa_t gpa = slot->base_gfn << PAGE_SHIFT;
2158 phys_addr_t size = slot->npages << PAGE_SHIFT;
2160 spin_lock(&kvm->mmu_lock);
2161 unmap_stage2_range(kvm, gpa, size);
2162 spin_unlock(&kvm->mmu_lock);
2166 * See note at ARMv7 ARM B1.14.4 (TL;DR: S/W ops are not easily virtualized).
2169 * - S/W ops are local to a CPU (not broadcast)
2170 * - We have line migration behind our back (speculation)
2171 * - System caches don't support S/W at all (damn!)
2173 * In the face of the above, the best we can do is to try and convert
2174 * S/W ops to VA ops. Because the guest is not allowed to infer the
2175 * S/W to PA mapping, it can only use S/W to nuke the whole cache,
2176 * which is a rather good thing for us.
2178 * Also, it is only used when turning caches on/off ("The expected
2179 * usage of the cache maintenance instructions that operate by set/way
2180 * is associated with the cache maintenance instructions associated
2181 * with the powerdown and powerup of caches, if this is required by
2182 * the implementation.").
2184 * We use the following policy:
2186 * - If we trap a S/W operation, we enable VM trapping to detect
2187 * caches being turned on/off, and do a full clean.
2189 * - We flush the caches on both caches being turned on and off.
2191 * - Once the caches are enabled, we stop trapping VM ops.
2193 void kvm_set_way_flush(struct kvm_vcpu *vcpu)
2195 unsigned long hcr = *vcpu_hcr(vcpu);
2198 * If this is the first time we do a S/W operation
2199 * (i.e. HCR_TVM not set) flush the whole memory, and set the
2202 * Otherwise, rely on the VM trapping to wait for the MMU +
2203 * Caches to be turned off. At that point, we'll be able to
2204 * clean the caches again.
2206 if (!(hcr & HCR_TVM)) {
2207 trace_kvm_set_way_flush(*vcpu_pc(vcpu),
2208 vcpu_has_cache_enabled(vcpu));
2209 stage2_flush_vm(vcpu->kvm);
2210 *vcpu_hcr(vcpu) = hcr | HCR_TVM;
2214 void kvm_toggle_cache(struct kvm_vcpu *vcpu, bool was_enabled)
2216 bool now_enabled = vcpu_has_cache_enabled(vcpu);
2219 * If switching the MMU+caches on, need to invalidate the caches.
2220 * If switching it off, need to clean the caches.
2221 * Clean + invalidate does the trick always.
2223 if (now_enabled != was_enabled)
2224 stage2_flush_vm(vcpu->kvm);
2226 /* Caches are now on, stop trapping VM ops (until a S/W op) */
2228 *vcpu_hcr(vcpu) &= ~HCR_TVM;
2230 trace_kvm_toggle_cache(*vcpu_pc(vcpu), was_enabled, now_enabled);