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,
330 phys_addr_t addr = start, end = start + size;
333 assert_spin_locked(&kvm->mmu_lock);
334 WARN_ON(size & ~PAGE_MASK);
336 pgd = kvm->arch.pgd + stage2_pgd_index(addr);
339 * Make sure the page table is still active, as another thread
340 * could have possibly freed the page table, while we released
343 if (!READ_ONCE(kvm->arch.pgd))
345 next = stage2_pgd_addr_end(addr, end);
346 if (!stage2_pgd_none(*pgd))
347 unmap_stage2_puds(kvm, pgd, addr, next);
349 * If the range is too large, release the kvm->mmu_lock
350 * to prevent starvation and lockup detector warnings.
352 if (may_block && next != end)
353 cond_resched_lock(&kvm->mmu_lock);
354 } while (pgd++, addr = next, addr != end);
357 static void unmap_stage2_range(struct kvm *kvm, phys_addr_t start, u64 size)
359 __unmap_stage2_range(kvm, start, size, true);
362 static void stage2_flush_ptes(struct kvm *kvm, pmd_t *pmd,
363 phys_addr_t addr, phys_addr_t end)
367 pte = pte_offset_kernel(pmd, addr);
369 if (!pte_none(*pte) && !kvm_is_device_pfn(pte_pfn(*pte)))
370 kvm_flush_dcache_pte(*pte);
371 } while (pte++, addr += PAGE_SIZE, addr != end);
374 static void stage2_flush_pmds(struct kvm *kvm, pud_t *pud,
375 phys_addr_t addr, phys_addr_t end)
380 pmd = stage2_pmd_offset(pud, addr);
382 next = stage2_pmd_addr_end(addr, end);
383 if (!pmd_none(*pmd)) {
384 if (pmd_thp_or_huge(*pmd))
385 kvm_flush_dcache_pmd(*pmd);
387 stage2_flush_ptes(kvm, pmd, addr, next);
389 } while (pmd++, addr = next, addr != end);
392 static void stage2_flush_puds(struct kvm *kvm, pgd_t *pgd,
393 phys_addr_t addr, phys_addr_t end)
398 pud = stage2_pud_offset(pgd, addr);
400 next = stage2_pud_addr_end(addr, end);
401 if (!stage2_pud_none(*pud)) {
402 if (stage2_pud_huge(*pud))
403 kvm_flush_dcache_pud(*pud);
405 stage2_flush_pmds(kvm, pud, addr, next);
407 } while (pud++, addr = next, addr != end);
410 static void stage2_flush_memslot(struct kvm *kvm,
411 struct kvm_memory_slot *memslot)
413 phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
414 phys_addr_t end = addr + PAGE_SIZE * memslot->npages;
418 pgd = kvm->arch.pgd + stage2_pgd_index(addr);
420 next = stage2_pgd_addr_end(addr, end);
421 if (!stage2_pgd_none(*pgd))
422 stage2_flush_puds(kvm, pgd, addr, next);
423 } while (pgd++, addr = next, addr != end);
427 * stage2_flush_vm - Invalidate cache for pages mapped in stage 2
428 * @kvm: The struct kvm pointer
430 * Go through the stage 2 page tables and invalidate any cache lines
431 * backing memory already mapped to the VM.
433 static void stage2_flush_vm(struct kvm *kvm)
435 struct kvm_memslots *slots;
436 struct kvm_memory_slot *memslot;
439 idx = srcu_read_lock(&kvm->srcu);
440 spin_lock(&kvm->mmu_lock);
442 slots = kvm_memslots(kvm);
443 kvm_for_each_memslot(memslot, slots)
444 stage2_flush_memslot(kvm, memslot);
446 spin_unlock(&kvm->mmu_lock);
447 srcu_read_unlock(&kvm->srcu, idx);
450 static void clear_hyp_pgd_entry(pgd_t *pgd)
452 pud_t *pud_table __maybe_unused = pud_offset(pgd, 0UL);
454 pud_free(NULL, pud_table);
455 put_page(virt_to_page(pgd));
458 static void clear_hyp_pud_entry(pud_t *pud)
460 pmd_t *pmd_table __maybe_unused = pmd_offset(pud, 0);
461 VM_BUG_ON(pud_huge(*pud));
463 pmd_free(NULL, pmd_table);
464 put_page(virt_to_page(pud));
467 static void clear_hyp_pmd_entry(pmd_t *pmd)
469 pte_t *pte_table = pte_offset_kernel(pmd, 0);
470 VM_BUG_ON(pmd_thp_or_huge(*pmd));
472 pte_free_kernel(NULL, pte_table);
473 put_page(virt_to_page(pmd));
476 static void unmap_hyp_ptes(pmd_t *pmd, phys_addr_t addr, phys_addr_t end)
478 pte_t *pte, *start_pte;
480 start_pte = pte = pte_offset_kernel(pmd, addr);
482 if (!pte_none(*pte)) {
483 kvm_set_pte(pte, __pte(0));
484 put_page(virt_to_page(pte));
486 } while (pte++, addr += PAGE_SIZE, addr != end);
488 if (hyp_pte_table_empty(start_pte))
489 clear_hyp_pmd_entry(pmd);
492 static void unmap_hyp_pmds(pud_t *pud, phys_addr_t addr, phys_addr_t end)
495 pmd_t *pmd, *start_pmd;
497 start_pmd = pmd = pmd_offset(pud, addr);
499 next = pmd_addr_end(addr, end);
500 /* Hyp doesn't use huge pmds */
502 unmap_hyp_ptes(pmd, addr, next);
503 } while (pmd++, addr = next, addr != end);
505 if (hyp_pmd_table_empty(start_pmd))
506 clear_hyp_pud_entry(pud);
509 static void unmap_hyp_puds(pgd_t *pgd, phys_addr_t addr, phys_addr_t end)
512 pud_t *pud, *start_pud;
514 start_pud = pud = pud_offset(pgd, addr);
516 next = pud_addr_end(addr, end);
517 /* Hyp doesn't use huge puds */
519 unmap_hyp_pmds(pud, addr, next);
520 } while (pud++, addr = next, addr != end);
522 if (hyp_pud_table_empty(start_pud))
523 clear_hyp_pgd_entry(pgd);
526 static unsigned int kvm_pgd_index(unsigned long addr, unsigned int ptrs_per_pgd)
528 return (addr >> PGDIR_SHIFT) & (ptrs_per_pgd - 1);
531 static void __unmap_hyp_range(pgd_t *pgdp, unsigned long ptrs_per_pgd,
532 phys_addr_t start, u64 size)
535 phys_addr_t addr = start, end = start + size;
539 * We don't unmap anything from HYP, except at the hyp tear down.
540 * Hence, we don't have to invalidate the TLBs here.
542 pgd = pgdp + kvm_pgd_index(addr, ptrs_per_pgd);
544 next = pgd_addr_end(addr, end);
546 unmap_hyp_puds(pgd, addr, next);
547 } while (pgd++, addr = next, addr != end);
550 static void unmap_hyp_range(pgd_t *pgdp, phys_addr_t start, u64 size)
552 __unmap_hyp_range(pgdp, PTRS_PER_PGD, start, size);
555 static void unmap_hyp_idmap_range(pgd_t *pgdp, phys_addr_t start, u64 size)
557 __unmap_hyp_range(pgdp, __kvm_idmap_ptrs_per_pgd(), start, size);
561 * free_hyp_pgds - free Hyp-mode page tables
563 * Assumes hyp_pgd is a page table used strictly in Hyp-mode and
564 * therefore contains either mappings in the kernel memory area (above
565 * PAGE_OFFSET), or device mappings in the idmap range.
567 * boot_hyp_pgd should only map the idmap range, and is only used in
568 * the extended idmap case.
570 void free_hyp_pgds(void)
574 mutex_lock(&kvm_hyp_pgd_mutex);
576 id_pgd = boot_hyp_pgd ? boot_hyp_pgd : hyp_pgd;
579 /* In case we never called hyp_mmu_init() */
581 io_map_base = hyp_idmap_start;
582 unmap_hyp_idmap_range(id_pgd, io_map_base,
583 hyp_idmap_start + PAGE_SIZE - io_map_base);
587 free_pages((unsigned long)boot_hyp_pgd, hyp_pgd_order);
592 unmap_hyp_range(hyp_pgd, kern_hyp_va(PAGE_OFFSET),
593 (uintptr_t)high_memory - PAGE_OFFSET);
595 free_pages((unsigned long)hyp_pgd, hyp_pgd_order);
598 if (merged_hyp_pgd) {
599 clear_page(merged_hyp_pgd);
600 free_page((unsigned long)merged_hyp_pgd);
601 merged_hyp_pgd = NULL;
604 mutex_unlock(&kvm_hyp_pgd_mutex);
607 static void create_hyp_pte_mappings(pmd_t *pmd, unsigned long start,
608 unsigned long end, unsigned long pfn,
616 pte = pte_offset_kernel(pmd, addr);
617 kvm_set_pte(pte, pfn_pte(pfn, prot));
618 get_page(virt_to_page(pte));
620 } while (addr += PAGE_SIZE, addr != end);
623 static int create_hyp_pmd_mappings(pud_t *pud, unsigned long start,
624 unsigned long end, unsigned long pfn,
629 unsigned long addr, next;
633 pmd = pmd_offset(pud, addr);
635 BUG_ON(pmd_sect(*pmd));
637 if (pmd_none(*pmd)) {
638 pte = pte_alloc_one_kernel(NULL, addr);
640 kvm_err("Cannot allocate Hyp pte\n");
643 kvm_pmd_populate(pmd, pte);
644 get_page(virt_to_page(pmd));
647 next = pmd_addr_end(addr, end);
649 create_hyp_pte_mappings(pmd, addr, next, pfn, prot);
650 pfn += (next - addr) >> PAGE_SHIFT;
651 } while (addr = next, addr != end);
656 static int create_hyp_pud_mappings(pgd_t *pgd, unsigned long start,
657 unsigned long end, unsigned long pfn,
662 unsigned long addr, next;
667 pud = pud_offset(pgd, addr);
669 if (pud_none_or_clear_bad(pud)) {
670 pmd = pmd_alloc_one(NULL, addr);
672 kvm_err("Cannot allocate Hyp pmd\n");
675 kvm_pud_populate(pud, pmd);
676 get_page(virt_to_page(pud));
679 next = pud_addr_end(addr, end);
680 ret = create_hyp_pmd_mappings(pud, addr, next, pfn, prot);
683 pfn += (next - addr) >> PAGE_SHIFT;
684 } while (addr = next, addr != end);
689 static int __create_hyp_mappings(pgd_t *pgdp, unsigned long ptrs_per_pgd,
690 unsigned long start, unsigned long end,
691 unsigned long pfn, pgprot_t prot)
695 unsigned long addr, next;
698 mutex_lock(&kvm_hyp_pgd_mutex);
699 addr = start & PAGE_MASK;
700 end = PAGE_ALIGN(end);
702 pgd = pgdp + kvm_pgd_index(addr, ptrs_per_pgd);
704 if (pgd_none(*pgd)) {
705 pud = pud_alloc_one(NULL, addr);
707 kvm_err("Cannot allocate Hyp pud\n");
711 kvm_pgd_populate(pgd, pud);
712 get_page(virt_to_page(pgd));
715 next = pgd_addr_end(addr, end);
716 err = create_hyp_pud_mappings(pgd, addr, next, pfn, prot);
719 pfn += (next - addr) >> PAGE_SHIFT;
720 } while (addr = next, addr != end);
722 mutex_unlock(&kvm_hyp_pgd_mutex);
726 static phys_addr_t kvm_kaddr_to_phys(void *kaddr)
728 if (!is_vmalloc_addr(kaddr)) {
729 BUG_ON(!virt_addr_valid(kaddr));
732 return page_to_phys(vmalloc_to_page(kaddr)) +
733 offset_in_page(kaddr);
738 * create_hyp_mappings - duplicate a kernel virtual address range in Hyp mode
739 * @from: The virtual kernel start address of the range
740 * @to: The virtual kernel end address of the range (exclusive)
741 * @prot: The protection to be applied to this range
743 * The same virtual address as the kernel virtual address is also used
744 * in Hyp-mode mapping (modulo HYP_PAGE_OFFSET) to the same underlying
747 int create_hyp_mappings(void *from, void *to, pgprot_t prot)
749 phys_addr_t phys_addr;
750 unsigned long virt_addr;
751 unsigned long start = kern_hyp_va((unsigned long)from);
752 unsigned long end = kern_hyp_va((unsigned long)to);
754 if (is_kernel_in_hyp_mode())
757 start = start & PAGE_MASK;
758 end = PAGE_ALIGN(end);
760 for (virt_addr = start; virt_addr < end; virt_addr += PAGE_SIZE) {
763 phys_addr = kvm_kaddr_to_phys(from + virt_addr - start);
764 err = __create_hyp_mappings(hyp_pgd, PTRS_PER_PGD,
765 virt_addr, virt_addr + PAGE_SIZE,
766 __phys_to_pfn(phys_addr),
775 static int __create_hyp_private_mapping(phys_addr_t phys_addr, size_t size,
776 unsigned long *haddr, pgprot_t prot)
778 pgd_t *pgd = hyp_pgd;
782 mutex_lock(&kvm_hyp_pgd_mutex);
785 * This assumes that we we have enough space below the idmap
786 * page to allocate our VAs. If not, the check below will
787 * kick. A potential alternative would be to detect that
788 * overflow and switch to an allocation above the idmap.
790 * The allocated size is always a multiple of PAGE_SIZE.
792 size = PAGE_ALIGN(size + offset_in_page(phys_addr));
793 base = io_map_base - size;
796 * Verify that BIT(VA_BITS - 1) hasn't been flipped by
797 * allocating the new area, as it would indicate we've
798 * overflowed the idmap/IO address range.
800 if ((base ^ io_map_base) & BIT(VA_BITS - 1))
805 mutex_unlock(&kvm_hyp_pgd_mutex);
810 if (__kvm_cpu_uses_extended_idmap())
813 ret = __create_hyp_mappings(pgd, __kvm_idmap_ptrs_per_pgd(),
815 __phys_to_pfn(phys_addr), prot);
819 *haddr = base + offset_in_page(phys_addr);
826 * create_hyp_io_mappings - Map IO into both kernel and HYP
827 * @phys_addr: The physical start address which gets mapped
828 * @size: Size of the region being mapped
829 * @kaddr: Kernel VA for this mapping
830 * @haddr: HYP VA for this mapping
832 int create_hyp_io_mappings(phys_addr_t phys_addr, size_t size,
833 void __iomem **kaddr,
834 void __iomem **haddr)
839 *kaddr = ioremap(phys_addr, size);
843 if (is_kernel_in_hyp_mode()) {
848 ret = __create_hyp_private_mapping(phys_addr, size,
849 &addr, PAGE_HYP_DEVICE);
857 *haddr = (void __iomem *)addr;
862 * create_hyp_exec_mappings - Map an executable range into HYP
863 * @phys_addr: The physical start address which gets mapped
864 * @size: Size of the region being mapped
865 * @haddr: HYP VA for this mapping
867 int create_hyp_exec_mappings(phys_addr_t phys_addr, size_t size,
873 BUG_ON(is_kernel_in_hyp_mode());
875 ret = __create_hyp_private_mapping(phys_addr, size,
876 &addr, PAGE_HYP_EXEC);
882 *haddr = (void *)addr;
887 * kvm_alloc_stage2_pgd - allocate level-1 table for stage-2 translation.
888 * @kvm: The KVM struct pointer for the VM.
890 * Allocates only the stage-2 HW PGD level table(s) (can support either full
891 * 40-bit input addresses or limited to 32-bit input addresses). Clears the
894 * Note we don't need locking here as this is only called when the VM is
895 * created, which can only be done once.
897 int kvm_alloc_stage2_pgd(struct kvm *kvm)
901 if (kvm->arch.pgd != NULL) {
902 kvm_err("kvm_arch already initialized?\n");
906 /* Allocate the HW PGD, making sure that each page gets its own refcount */
907 pgd = alloc_pages_exact(S2_PGD_SIZE, GFP_KERNEL | __GFP_ZERO);
915 static void stage2_unmap_memslot(struct kvm *kvm,
916 struct kvm_memory_slot *memslot)
918 hva_t hva = memslot->userspace_addr;
919 phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
920 phys_addr_t size = PAGE_SIZE * memslot->npages;
921 hva_t reg_end = hva + size;
924 * A memory region could potentially cover multiple VMAs, and any holes
925 * between them, so iterate over all of them to find out if we should
928 * +--------------------------------------------+
929 * +---------------+----------------+ +----------------+
930 * | : VMA 1 | VMA 2 | | VMA 3 : |
931 * +---------------+----------------+ +----------------+
933 * +--------------------------------------------+
936 struct vm_area_struct *vma = find_vma(current->mm, hva);
937 hva_t vm_start, vm_end;
939 if (!vma || vma->vm_start >= reg_end)
943 * Take the intersection of this VMA with the memory region
945 vm_start = max(hva, vma->vm_start);
946 vm_end = min(reg_end, vma->vm_end);
948 if (!(vma->vm_flags & VM_PFNMAP)) {
949 gpa_t gpa = addr + (vm_start - memslot->userspace_addr);
950 unmap_stage2_range(kvm, gpa, vm_end - vm_start);
953 } while (hva < reg_end);
957 * stage2_unmap_vm - Unmap Stage-2 RAM mappings
958 * @kvm: The struct kvm pointer
960 * Go through the memregions and unmap any reguler RAM
961 * backing memory already mapped to the VM.
963 void stage2_unmap_vm(struct kvm *kvm)
965 struct kvm_memslots *slots;
966 struct kvm_memory_slot *memslot;
969 idx = srcu_read_lock(&kvm->srcu);
970 down_read(¤t->mm->mmap_sem);
971 spin_lock(&kvm->mmu_lock);
973 slots = kvm_memslots(kvm);
974 kvm_for_each_memslot(memslot, slots)
975 stage2_unmap_memslot(kvm, memslot);
977 spin_unlock(&kvm->mmu_lock);
978 up_read(¤t->mm->mmap_sem);
979 srcu_read_unlock(&kvm->srcu, idx);
983 * kvm_free_stage2_pgd - free all stage-2 tables
984 * @kvm: The KVM struct pointer for the VM.
986 * Walks the level-1 page table pointed to by kvm->arch.pgd and frees all
987 * underlying level-2 and level-3 tables before freeing the actual level-1 table
988 * and setting the struct pointer to NULL.
990 void kvm_free_stage2_pgd(struct kvm *kvm)
994 spin_lock(&kvm->mmu_lock);
996 unmap_stage2_range(kvm, 0, KVM_PHYS_SIZE);
997 pgd = READ_ONCE(kvm->arch.pgd);
998 kvm->arch.pgd = NULL;
1000 spin_unlock(&kvm->mmu_lock);
1002 /* Free the HW pgd, one page at a time */
1004 free_pages_exact(pgd, S2_PGD_SIZE);
1007 static pud_t *stage2_get_pud(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
1013 pgd = kvm->arch.pgd + stage2_pgd_index(addr);
1014 if (WARN_ON(stage2_pgd_none(*pgd))) {
1017 pud = mmu_memory_cache_alloc(cache);
1018 stage2_pgd_populate(pgd, pud);
1019 get_page(virt_to_page(pgd));
1022 return stage2_pud_offset(pgd, addr);
1025 static pmd_t *stage2_get_pmd(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
1031 pud = stage2_get_pud(kvm, cache, addr);
1035 if (stage2_pud_none(*pud)) {
1038 pmd = mmu_memory_cache_alloc(cache);
1039 stage2_pud_populate(pud, pmd);
1040 get_page(virt_to_page(pud));
1043 return stage2_pmd_offset(pud, addr);
1046 static int stage2_set_pmd_huge(struct kvm *kvm, struct kvm_mmu_memory_cache
1047 *cache, phys_addr_t addr, const pmd_t *new_pmd)
1049 pmd_t *pmd, old_pmd;
1051 pmd = stage2_get_pmd(kvm, cache, addr);
1055 if (pmd_present(old_pmd)) {
1057 * Multiple vcpus faulting on the same PMD entry, can
1058 * lead to them sequentially updating the PMD with the
1059 * same value. Following the break-before-make
1060 * (pmd_clear() followed by tlb_flush()) process can
1061 * hinder forward progress due to refaults generated
1062 * on missing translations.
1064 * Skip updating the page table if the entry is
1067 if (pmd_val(old_pmd) == pmd_val(*new_pmd))
1071 * Mapping in huge pages should only happen through a
1072 * fault. If a page is merged into a transparent huge
1073 * page, the individual subpages of that huge page
1074 * should be unmapped through MMU notifiers before we
1077 * Merging of CompoundPages is not supported; they
1078 * should become splitting first, unmapped, merged,
1079 * and mapped back in on-demand.
1081 VM_BUG_ON(pmd_pfn(old_pmd) != pmd_pfn(*new_pmd));
1084 kvm_tlb_flush_vmid_ipa(kvm, addr);
1086 get_page(virt_to_page(pmd));
1089 kvm_set_pmd(pmd, *new_pmd);
1093 static bool stage2_is_exec(struct kvm *kvm, phys_addr_t addr)
1098 pmdp = stage2_get_pmd(kvm, NULL, addr);
1099 if (!pmdp || pmd_none(*pmdp) || !pmd_present(*pmdp))
1102 if (pmd_thp_or_huge(*pmdp))
1103 return kvm_s2pmd_exec(pmdp);
1105 ptep = pte_offset_kernel(pmdp, addr);
1106 if (!ptep || pte_none(*ptep) || !pte_present(*ptep))
1109 return kvm_s2pte_exec(ptep);
1112 static int stage2_set_pte(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
1113 phys_addr_t addr, const pte_t *new_pte,
1114 unsigned long flags)
1117 pte_t *pte, old_pte;
1118 bool iomap = flags & KVM_S2PTE_FLAG_IS_IOMAP;
1119 bool logging_active = flags & KVM_S2_FLAG_LOGGING_ACTIVE;
1121 VM_BUG_ON(logging_active && !cache);
1123 /* Create stage-2 page table mapping - Levels 0 and 1 */
1124 pmd = stage2_get_pmd(kvm, cache, addr);
1127 * Ignore calls from kvm_set_spte_hva for unallocated
1134 * While dirty page logging - dissolve huge PMD, then continue on to
1138 stage2_dissolve_pmd(kvm, addr, pmd);
1140 /* Create stage-2 page mappings - Level 2 */
1141 if (pmd_none(*pmd)) {
1143 return 0; /* ignore calls from kvm_set_spte_hva */
1144 pte = mmu_memory_cache_alloc(cache);
1145 kvm_pmd_populate(pmd, pte);
1146 get_page(virt_to_page(pmd));
1149 pte = pte_offset_kernel(pmd, addr);
1151 if (iomap && pte_present(*pte))
1154 /* Create 2nd stage page table mapping - Level 3 */
1156 if (pte_present(old_pte)) {
1157 /* Skip page table update if there is no change */
1158 if (pte_val(old_pte) == pte_val(*new_pte))
1161 kvm_set_pte(pte, __pte(0));
1162 kvm_tlb_flush_vmid_ipa(kvm, addr);
1164 get_page(virt_to_page(pte));
1167 kvm_set_pte(pte, *new_pte);
1171 #ifndef __HAVE_ARCH_PTEP_TEST_AND_CLEAR_YOUNG
1172 static int stage2_ptep_test_and_clear_young(pte_t *pte)
1174 if (pte_young(*pte)) {
1175 *pte = pte_mkold(*pte);
1181 static int stage2_ptep_test_and_clear_young(pte_t *pte)
1183 return __ptep_test_and_clear_young(pte);
1187 static int stage2_pmdp_test_and_clear_young(pmd_t *pmd)
1189 return stage2_ptep_test_and_clear_young((pte_t *)pmd);
1193 * kvm_phys_addr_ioremap - map a device range to guest IPA
1195 * @kvm: The KVM pointer
1196 * @guest_ipa: The IPA at which to insert the mapping
1197 * @pa: The physical address of the device
1198 * @size: The size of the mapping
1200 int kvm_phys_addr_ioremap(struct kvm *kvm, phys_addr_t guest_ipa,
1201 phys_addr_t pa, unsigned long size, bool writable)
1203 phys_addr_t addr, end;
1206 struct kvm_mmu_memory_cache cache = { 0, };
1208 end = (guest_ipa + size + PAGE_SIZE - 1) & PAGE_MASK;
1209 pfn = __phys_to_pfn(pa);
1211 for (addr = guest_ipa; addr < end; addr += PAGE_SIZE) {
1212 pte_t pte = pfn_pte(pfn, PAGE_S2_DEVICE);
1215 pte = kvm_s2pte_mkwrite(pte);
1217 ret = mmu_topup_memory_cache(&cache, KVM_MMU_CACHE_MIN_PAGES,
1221 spin_lock(&kvm->mmu_lock);
1222 ret = stage2_set_pte(kvm, &cache, addr, &pte,
1223 KVM_S2PTE_FLAG_IS_IOMAP);
1224 spin_unlock(&kvm->mmu_lock);
1232 mmu_free_memory_cache(&cache);
1236 static bool transparent_hugepage_adjust(kvm_pfn_t *pfnp, phys_addr_t *ipap)
1238 kvm_pfn_t pfn = *pfnp;
1239 gfn_t gfn = *ipap >> PAGE_SHIFT;
1240 struct page *page = pfn_to_page(pfn);
1243 * PageTransCompoungMap() returns true for THP and
1244 * hugetlbfs. Make sure the adjustment is done only for THP
1247 if (!PageHuge(page) && PageTransCompoundMap(page)) {
1250 * The address we faulted on is backed by a transparent huge
1251 * page. However, because we map the compound huge page and
1252 * not the individual tail page, we need to transfer the
1253 * refcount to the head page. We have to be careful that the
1254 * THP doesn't start to split while we are adjusting the
1257 * We are sure this doesn't happen, because mmu_notifier_retry
1258 * was successful and we are holding the mmu_lock, so if this
1259 * THP is trying to split, it will be blocked in the mmu
1260 * notifier before touching any of the pages, specifically
1261 * before being able to call __split_huge_page_refcount().
1263 * We can therefore safely transfer the refcount from PG_tail
1264 * to PG_head and switch the pfn from a tail page to the head
1267 mask = PTRS_PER_PMD - 1;
1268 VM_BUG_ON((gfn & mask) != (pfn & mask));
1271 kvm_release_pfn_clean(pfn);
1283 static bool kvm_is_write_fault(struct kvm_vcpu *vcpu)
1285 if (kvm_vcpu_abt_iss1tw(vcpu))
1288 if (kvm_vcpu_trap_is_iabt(vcpu))
1291 return kvm_vcpu_dabt_iswrite(vcpu);
1295 * stage2_wp_ptes - write protect PMD range
1296 * @pmd: pointer to pmd entry
1297 * @addr: range start address
1298 * @end: range end address
1300 static void stage2_wp_ptes(pmd_t *pmd, phys_addr_t addr, phys_addr_t end)
1304 pte = pte_offset_kernel(pmd, addr);
1306 if (!pte_none(*pte)) {
1307 if (!kvm_s2pte_readonly(pte))
1308 kvm_set_s2pte_readonly(pte);
1310 } while (pte++, addr += PAGE_SIZE, addr != end);
1314 * stage2_wp_pmds - write protect PUD range
1315 * @pud: pointer to pud entry
1316 * @addr: range start address
1317 * @end: range end address
1319 static void stage2_wp_pmds(pud_t *pud, phys_addr_t addr, phys_addr_t end)
1324 pmd = stage2_pmd_offset(pud, addr);
1327 next = stage2_pmd_addr_end(addr, end);
1328 if (!pmd_none(*pmd)) {
1329 if (pmd_thp_or_huge(*pmd)) {
1330 if (!kvm_s2pmd_readonly(pmd))
1331 kvm_set_s2pmd_readonly(pmd);
1333 stage2_wp_ptes(pmd, addr, next);
1336 } while (pmd++, addr = next, addr != end);
1340 * stage2_wp_puds - write protect PGD range
1341 * @pgd: pointer to pgd entry
1342 * @addr: range start address
1343 * @end: range end address
1345 * Process PUD entries, for a huge PUD we cause a panic.
1347 static void stage2_wp_puds(pgd_t *pgd, phys_addr_t addr, phys_addr_t end)
1352 pud = stage2_pud_offset(pgd, addr);
1354 next = stage2_pud_addr_end(addr, end);
1355 if (!stage2_pud_none(*pud)) {
1356 /* TODO:PUD not supported, revisit later if supported */
1357 BUG_ON(stage2_pud_huge(*pud));
1358 stage2_wp_pmds(pud, addr, next);
1360 } while (pud++, addr = next, addr != end);
1364 * stage2_wp_range() - write protect stage2 memory region range
1365 * @kvm: The KVM pointer
1366 * @addr: Start address of range
1367 * @end: End address of range
1369 static void stage2_wp_range(struct kvm *kvm, phys_addr_t addr, phys_addr_t end)
1374 pgd = kvm->arch.pgd + stage2_pgd_index(addr);
1377 * Release kvm_mmu_lock periodically if the memory region is
1378 * large. Otherwise, we may see kernel panics with
1379 * CONFIG_DETECT_HUNG_TASK, CONFIG_LOCKUP_DETECTOR,
1380 * CONFIG_LOCKDEP. Additionally, holding the lock too long
1381 * will also starve other vCPUs. We have to also make sure
1382 * that the page tables are not freed while we released
1385 cond_resched_lock(&kvm->mmu_lock);
1386 if (!READ_ONCE(kvm->arch.pgd))
1388 next = stage2_pgd_addr_end(addr, end);
1389 if (stage2_pgd_present(*pgd))
1390 stage2_wp_puds(pgd, addr, next);
1391 } while (pgd++, addr = next, addr != end);
1395 * kvm_mmu_wp_memory_region() - write protect stage 2 entries for memory slot
1396 * @kvm: The KVM pointer
1397 * @slot: The memory slot to write protect
1399 * Called to start logging dirty pages after memory region
1400 * KVM_MEM_LOG_DIRTY_PAGES operation is called. After this function returns
1401 * all present PMD and PTEs are write protected in the memory region.
1402 * Afterwards read of dirty page log can be called.
1404 * Acquires kvm_mmu_lock. Called with kvm->slots_lock mutex acquired,
1405 * serializing operations for VM memory regions.
1407 void kvm_mmu_wp_memory_region(struct kvm *kvm, int slot)
1409 struct kvm_memslots *slots = kvm_memslots(kvm);
1410 struct kvm_memory_slot *memslot = id_to_memslot(slots, slot);
1411 phys_addr_t start = memslot->base_gfn << PAGE_SHIFT;
1412 phys_addr_t end = (memslot->base_gfn + memslot->npages) << PAGE_SHIFT;
1414 spin_lock(&kvm->mmu_lock);
1415 stage2_wp_range(kvm, start, end);
1416 spin_unlock(&kvm->mmu_lock);
1417 kvm_flush_remote_tlbs(kvm);
1421 * kvm_mmu_write_protect_pt_masked() - write protect dirty pages
1422 * @kvm: The KVM pointer
1423 * @slot: The memory slot associated with mask
1424 * @gfn_offset: The gfn offset in memory slot
1425 * @mask: The mask of dirty pages at offset 'gfn_offset' in this memory
1426 * slot to be write protected
1428 * Walks bits set in mask write protects the associated pte's. Caller must
1429 * acquire kvm_mmu_lock.
1431 static void kvm_mmu_write_protect_pt_masked(struct kvm *kvm,
1432 struct kvm_memory_slot *slot,
1433 gfn_t gfn_offset, unsigned long mask)
1435 phys_addr_t base_gfn = slot->base_gfn + gfn_offset;
1436 phys_addr_t start = (base_gfn + __ffs(mask)) << PAGE_SHIFT;
1437 phys_addr_t end = (base_gfn + __fls(mask) + 1) << PAGE_SHIFT;
1439 stage2_wp_range(kvm, start, end);
1443 * kvm_arch_mmu_enable_log_dirty_pt_masked - enable dirty logging for selected
1446 * It calls kvm_mmu_write_protect_pt_masked to write protect selected pages to
1447 * enable dirty logging for them.
1449 void kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm *kvm,
1450 struct kvm_memory_slot *slot,
1451 gfn_t gfn_offset, unsigned long mask)
1453 kvm_mmu_write_protect_pt_masked(kvm, slot, gfn_offset, mask);
1456 static void clean_dcache_guest_page(kvm_pfn_t pfn, unsigned long size)
1458 __clean_dcache_guest_page(pfn, size);
1461 static void invalidate_icache_guest_page(kvm_pfn_t pfn, unsigned long size)
1463 __invalidate_icache_guest_page(pfn, size);
1466 static void kvm_send_hwpoison_signal(unsigned long address,
1467 struct vm_area_struct *vma)
1471 clear_siginfo(&info);
1472 info.si_signo = SIGBUS;
1474 info.si_code = BUS_MCEERR_AR;
1475 info.si_addr = (void __user *)address;
1477 if (is_vm_hugetlb_page(vma))
1478 info.si_addr_lsb = huge_page_shift(hstate_vma(vma));
1480 info.si_addr_lsb = PAGE_SHIFT;
1482 send_sig_info(SIGBUS, &info, current);
1485 static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
1486 struct kvm_memory_slot *memslot, unsigned long hva,
1487 unsigned long fault_status)
1490 bool write_fault, exec_fault, writable, hugetlb = false, force_pte = false;
1491 unsigned long mmu_seq;
1492 gfn_t gfn = fault_ipa >> PAGE_SHIFT;
1493 struct kvm *kvm = vcpu->kvm;
1494 struct kvm_mmu_memory_cache *memcache = &vcpu->arch.mmu_page_cache;
1495 struct vm_area_struct *vma;
1497 pgprot_t mem_type = PAGE_S2;
1498 bool logging_active = memslot_is_logging(memslot);
1499 unsigned long flags = 0;
1501 write_fault = kvm_is_write_fault(vcpu);
1502 exec_fault = kvm_vcpu_trap_is_exec_fault(vcpu);
1503 VM_BUG_ON(write_fault && exec_fault);
1505 if (fault_status == FSC_PERM && !write_fault && !exec_fault) {
1506 kvm_err("Unexpected L2 read permission error\n");
1510 /* Let's check if we will get back a huge page backed by hugetlbfs */
1511 down_read(¤t->mm->mmap_sem);
1512 vma = find_vma_intersection(current->mm, hva, hva + 1);
1513 if (unlikely(!vma)) {
1514 kvm_err("Failed to find VMA for hva 0x%lx\n", hva);
1515 up_read(¤t->mm->mmap_sem);
1519 if (vma_kernel_pagesize(vma) == PMD_SIZE && !logging_active) {
1521 gfn = (fault_ipa & PMD_MASK) >> PAGE_SHIFT;
1524 * Pages belonging to memslots that don't have the same
1525 * alignment for userspace and IPA cannot be mapped using
1526 * block descriptors even if the pages belong to a THP for
1527 * the process, because the stage-2 block descriptor will
1528 * cover more than a single THP and we loose atomicity for
1529 * unmapping, updates, and splits of the THP or other pages
1530 * in the stage-2 block range.
1532 if ((memslot->userspace_addr & ~PMD_MASK) !=
1533 ((memslot->base_gfn << PAGE_SHIFT) & ~PMD_MASK))
1536 up_read(¤t->mm->mmap_sem);
1538 /* We need minimum second+third level pages */
1539 ret = mmu_topup_memory_cache(memcache, KVM_MMU_CACHE_MIN_PAGES,
1544 mmu_seq = vcpu->kvm->mmu_notifier_seq;
1546 * Ensure the read of mmu_notifier_seq happens before we call
1547 * gfn_to_pfn_prot (which calls get_user_pages), so that we don't risk
1548 * the page we just got a reference to gets unmapped before we have a
1549 * chance to grab the mmu_lock, which ensure that if the page gets
1550 * unmapped afterwards, the call to kvm_unmap_hva will take it away
1551 * from us again properly. This smp_rmb() interacts with the smp_wmb()
1552 * in kvm_mmu_notifier_invalidate_<page|range_end>.
1556 pfn = gfn_to_pfn_prot(kvm, gfn, write_fault, &writable);
1557 if (pfn == KVM_PFN_ERR_HWPOISON) {
1558 kvm_send_hwpoison_signal(hva, vma);
1561 if (is_error_noslot_pfn(pfn))
1564 if (kvm_is_device_pfn(pfn)) {
1565 mem_type = PAGE_S2_DEVICE;
1566 flags |= KVM_S2PTE_FLAG_IS_IOMAP;
1567 } else if (logging_active) {
1569 * Faults on pages in a memslot with logging enabled
1570 * should not be mapped with huge pages (it introduces churn
1571 * and performance degradation), so force a pte mapping.
1574 flags |= KVM_S2_FLAG_LOGGING_ACTIVE;
1577 * Only actually map the page as writable if this was a write
1584 spin_lock(&kvm->mmu_lock);
1585 if (mmu_notifier_retry(kvm, mmu_seq))
1588 if (!hugetlb && !force_pte)
1589 hugetlb = transparent_hugepage_adjust(&pfn, &fault_ipa);
1592 pmd_t new_pmd = pfn_pmd(pfn, mem_type);
1593 new_pmd = pmd_mkhuge(new_pmd);
1595 new_pmd = kvm_s2pmd_mkwrite(new_pmd);
1596 kvm_set_pfn_dirty(pfn);
1599 if (fault_status != FSC_PERM)
1600 clean_dcache_guest_page(pfn, PMD_SIZE);
1603 new_pmd = kvm_s2pmd_mkexec(new_pmd);
1604 invalidate_icache_guest_page(pfn, PMD_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_pmd = kvm_s2pmd_mkexec(new_pmd);
1611 ret = stage2_set_pmd_huge(kvm, memcache, fault_ipa, &new_pmd);
1613 pte_t new_pte = pfn_pte(pfn, mem_type);
1616 new_pte = kvm_s2pte_mkwrite(new_pte);
1617 kvm_set_pfn_dirty(pfn);
1618 mark_page_dirty(kvm, gfn);
1621 if (fault_status != FSC_PERM)
1622 clean_dcache_guest_page(pfn, PAGE_SIZE);
1625 new_pte = kvm_s2pte_mkexec(new_pte);
1626 invalidate_icache_guest_page(pfn, PAGE_SIZE);
1627 } else if (fault_status == FSC_PERM) {
1628 /* Preserve execute if XN was already cleared */
1629 if (stage2_is_exec(kvm, fault_ipa))
1630 new_pte = kvm_s2pte_mkexec(new_pte);
1633 ret = stage2_set_pte(kvm, memcache, fault_ipa, &new_pte, flags);
1637 spin_unlock(&kvm->mmu_lock);
1638 kvm_set_pfn_accessed(pfn);
1639 kvm_release_pfn_clean(pfn);
1644 * Resolve the access fault by making the page young again.
1645 * Note that because the faulting entry is guaranteed not to be
1646 * cached in the TLB, we don't need to invalidate anything.
1647 * Only the HW Access Flag updates are supported for Stage 2 (no DBM),
1648 * so there is no need for atomic (pte|pmd)_mkyoung operations.
1650 static void handle_access_fault(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa)
1655 bool pfn_valid = false;
1657 trace_kvm_access_fault(fault_ipa);
1659 spin_lock(&vcpu->kvm->mmu_lock);
1661 pmd = stage2_get_pmd(vcpu->kvm, NULL, fault_ipa);
1662 if (!pmd || pmd_none(*pmd)) /* Nothing there */
1665 if (pmd_thp_or_huge(*pmd)) { /* THP, HugeTLB */
1666 *pmd = pmd_mkyoung(*pmd);
1667 pfn = pmd_pfn(*pmd);
1672 pte = pte_offset_kernel(pmd, fault_ipa);
1673 if (pte_none(*pte)) /* Nothing there either */
1676 *pte = pte_mkyoung(*pte); /* Just a page... */
1677 pfn = pte_pfn(*pte);
1680 spin_unlock(&vcpu->kvm->mmu_lock);
1682 kvm_set_pfn_accessed(pfn);
1686 * kvm_handle_guest_abort - handles all 2nd stage aborts
1687 * @vcpu: the VCPU pointer
1688 * @run: the kvm_run structure
1690 * Any abort that gets to the host is almost guaranteed to be caused by a
1691 * missing second stage translation table entry, which can mean that either the
1692 * guest simply needs more memory and we must allocate an appropriate page or it
1693 * can mean that the guest tried to access I/O memory, which is emulated by user
1694 * space. The distinction is based on the IPA causing the fault and whether this
1695 * memory region has been registered as standard RAM by user space.
1697 int kvm_handle_guest_abort(struct kvm_vcpu *vcpu, struct kvm_run *run)
1699 unsigned long fault_status;
1700 phys_addr_t fault_ipa;
1701 struct kvm_memory_slot *memslot;
1703 bool is_iabt, write_fault, writable;
1707 fault_status = kvm_vcpu_trap_get_fault_type(vcpu);
1709 fault_ipa = kvm_vcpu_get_fault_ipa(vcpu);
1710 is_iabt = kvm_vcpu_trap_is_iabt(vcpu);
1712 /* Synchronous External Abort? */
1713 if (kvm_vcpu_dabt_isextabt(vcpu)) {
1715 * For RAS the host kernel may handle this abort.
1716 * There is no need to pass the error into the guest.
1718 if (!handle_guest_sea(fault_ipa, kvm_vcpu_get_hsr(vcpu)))
1721 if (unlikely(!is_iabt)) {
1722 kvm_inject_vabt(vcpu);
1727 trace_kvm_guest_fault(*vcpu_pc(vcpu), kvm_vcpu_get_hsr(vcpu),
1728 kvm_vcpu_get_hfar(vcpu), fault_ipa);
1730 /* Check the stage-2 fault is trans. fault or write fault */
1731 if (fault_status != FSC_FAULT && fault_status != FSC_PERM &&
1732 fault_status != FSC_ACCESS) {
1733 kvm_err("Unsupported FSC: EC=%#x xFSC=%#lx ESR_EL2=%#lx\n",
1734 kvm_vcpu_trap_get_class(vcpu),
1735 (unsigned long)kvm_vcpu_trap_get_fault(vcpu),
1736 (unsigned long)kvm_vcpu_get_hsr(vcpu));
1740 idx = srcu_read_lock(&vcpu->kvm->srcu);
1742 gfn = fault_ipa >> PAGE_SHIFT;
1743 memslot = gfn_to_memslot(vcpu->kvm, gfn);
1744 hva = gfn_to_hva_memslot_prot(memslot, gfn, &writable);
1745 write_fault = kvm_is_write_fault(vcpu);
1746 if (kvm_is_error_hva(hva) || (write_fault && !writable)) {
1748 /* Prefetch Abort on I/O address */
1749 kvm_inject_pabt(vcpu, kvm_vcpu_get_hfar(vcpu));
1755 * Check for a cache maintenance operation. Since we
1756 * ended-up here, we know it is outside of any memory
1757 * slot. But we can't find out if that is for a device,
1758 * or if the guest is just being stupid. The only thing
1759 * we know for sure is that this range cannot be cached.
1761 * So let's assume that the guest is just being
1762 * cautious, and skip the instruction.
1764 if (kvm_vcpu_dabt_is_cm(vcpu)) {
1765 kvm_skip_instr(vcpu, kvm_vcpu_trap_il_is32bit(vcpu));
1771 * The IPA is reported as [MAX:12], so we need to
1772 * complement it with the bottom 12 bits from the
1773 * faulting VA. This is always 12 bits, irrespective
1776 fault_ipa |= kvm_vcpu_get_hfar(vcpu) & ((1 << 12) - 1);
1777 ret = io_mem_abort(vcpu, run, fault_ipa);
1781 /* Userspace should not be able to register out-of-bounds IPAs */
1782 VM_BUG_ON(fault_ipa >= KVM_PHYS_SIZE);
1784 if (fault_status == FSC_ACCESS) {
1785 handle_access_fault(vcpu, fault_ipa);
1790 ret = user_mem_abort(vcpu, fault_ipa, memslot, hva, fault_status);
1794 srcu_read_unlock(&vcpu->kvm->srcu, idx);
1798 static int handle_hva_to_gpa(struct kvm *kvm,
1799 unsigned long start,
1801 int (*handler)(struct kvm *kvm,
1802 gpa_t gpa, u64 size,
1806 struct kvm_memslots *slots;
1807 struct kvm_memory_slot *memslot;
1810 slots = kvm_memslots(kvm);
1812 /* we only care about the pages that the guest sees */
1813 kvm_for_each_memslot(memslot, slots) {
1814 unsigned long hva_start, hva_end;
1817 hva_start = max(start, memslot->userspace_addr);
1818 hva_end = min(end, memslot->userspace_addr +
1819 (memslot->npages << PAGE_SHIFT));
1820 if (hva_start >= hva_end)
1823 gpa = hva_to_gfn_memslot(hva_start, memslot) << PAGE_SHIFT;
1824 ret |= handler(kvm, gpa, (u64)(hva_end - hva_start), data);
1830 static int kvm_unmap_hva_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
1832 bool may_block = *(bool *)data;
1834 __unmap_stage2_range(kvm, gpa, size, may_block);
1838 int kvm_unmap_hva_range(struct kvm *kvm,
1839 unsigned long start, unsigned long end, bool blockable)
1844 trace_kvm_unmap_hva_range(start, end);
1845 handle_hva_to_gpa(kvm, start, end, &kvm_unmap_hva_handler, &blockable);
1849 static int kvm_set_spte_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
1851 pte_t *pte = (pte_t *)data;
1853 WARN_ON(size != PAGE_SIZE);
1855 * We can always call stage2_set_pte with KVM_S2PTE_FLAG_LOGGING_ACTIVE
1856 * flag clear because MMU notifiers will have unmapped a huge PMD before
1857 * calling ->change_pte() (which in turn calls kvm_set_spte_hva()) and
1858 * therefore stage2_set_pte() never needs to clear out a huge PMD
1859 * through this calling path.
1861 stage2_set_pte(kvm, NULL, gpa, pte, 0);
1866 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
1868 unsigned long end = hva + PAGE_SIZE;
1869 kvm_pfn_t pfn = pte_pfn(pte);
1875 trace_kvm_set_spte_hva(hva);
1878 * We've moved a page around, probably through CoW, so let's treat it
1879 * just like a translation fault and clean the cache to the PoC.
1881 clean_dcache_guest_page(pfn, PAGE_SIZE);
1882 stage2_pte = pfn_pte(pfn, PAGE_S2);
1883 handle_hva_to_gpa(kvm, hva, end, &kvm_set_spte_handler, &stage2_pte);
1886 static int kvm_age_hva_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
1891 WARN_ON(size != PAGE_SIZE && size != PMD_SIZE);
1892 pmd = stage2_get_pmd(kvm, NULL, gpa);
1893 if (!pmd || pmd_none(*pmd)) /* Nothing there */
1896 if (pmd_thp_or_huge(*pmd)) /* THP, HugeTLB */
1897 return stage2_pmdp_test_and_clear_young(pmd);
1899 pte = pte_offset_kernel(pmd, gpa);
1903 return stage2_ptep_test_and_clear_young(pte);
1906 static int kvm_test_age_hva_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
1911 WARN_ON(size != PAGE_SIZE && size != PMD_SIZE);
1912 pmd = stage2_get_pmd(kvm, NULL, gpa);
1913 if (!pmd || pmd_none(*pmd)) /* Nothing there */
1916 if (pmd_thp_or_huge(*pmd)) /* THP, HugeTLB */
1917 return pmd_young(*pmd);
1919 pte = pte_offset_kernel(pmd, gpa);
1920 if (!pte_none(*pte)) /* Just a page... */
1921 return pte_young(*pte);
1926 int kvm_age_hva(struct kvm *kvm, unsigned long start, unsigned long end)
1930 trace_kvm_age_hva(start, end);
1931 return handle_hva_to_gpa(kvm, start, end, kvm_age_hva_handler, NULL);
1934 int kvm_test_age_hva(struct kvm *kvm, unsigned long hva)
1938 trace_kvm_test_age_hva(hva);
1939 return handle_hva_to_gpa(kvm, hva, hva + PAGE_SIZE,
1940 kvm_test_age_hva_handler, NULL);
1943 void kvm_mmu_free_memory_caches(struct kvm_vcpu *vcpu)
1945 mmu_free_memory_cache(&vcpu->arch.mmu_page_cache);
1948 phys_addr_t kvm_mmu_get_httbr(void)
1950 if (__kvm_cpu_uses_extended_idmap())
1951 return virt_to_phys(merged_hyp_pgd);
1953 return virt_to_phys(hyp_pgd);
1956 phys_addr_t kvm_get_idmap_vector(void)
1958 return hyp_idmap_vector;
1961 static int kvm_map_idmap_text(pgd_t *pgd)
1965 /* Create the idmap in the boot page tables */
1966 err = __create_hyp_mappings(pgd, __kvm_idmap_ptrs_per_pgd(),
1967 hyp_idmap_start, hyp_idmap_end,
1968 __phys_to_pfn(hyp_idmap_start),
1971 kvm_err("Failed to idmap %lx-%lx\n",
1972 hyp_idmap_start, hyp_idmap_end);
1977 int kvm_mmu_init(void)
1981 hyp_idmap_start = kvm_virt_to_phys(__hyp_idmap_text_start);
1982 hyp_idmap_start = ALIGN_DOWN(hyp_idmap_start, PAGE_SIZE);
1983 hyp_idmap_end = kvm_virt_to_phys(__hyp_idmap_text_end);
1984 hyp_idmap_end = ALIGN(hyp_idmap_end, PAGE_SIZE);
1985 hyp_idmap_vector = kvm_virt_to_phys(__kvm_hyp_init);
1988 * We rely on the linker script to ensure at build time that the HYP
1989 * init code does not cross a page boundary.
1991 BUG_ON((hyp_idmap_start ^ (hyp_idmap_end - 1)) & PAGE_MASK);
1993 kvm_debug("IDMAP page: %lx\n", hyp_idmap_start);
1994 kvm_debug("HYP VA range: %lx:%lx\n",
1995 kern_hyp_va(PAGE_OFFSET),
1996 kern_hyp_va((unsigned long)high_memory - 1));
1998 if (hyp_idmap_start >= kern_hyp_va(PAGE_OFFSET) &&
1999 hyp_idmap_start < kern_hyp_va((unsigned long)high_memory - 1) &&
2000 hyp_idmap_start != (unsigned long)__hyp_idmap_text_start) {
2002 * The idmap page is intersecting with the VA space,
2003 * it is not safe to continue further.
2005 kvm_err("IDMAP intersecting with HYP VA, unable to continue\n");
2010 hyp_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, hyp_pgd_order);
2012 kvm_err("Hyp mode PGD not allocated\n");
2017 if (__kvm_cpu_uses_extended_idmap()) {
2018 boot_hyp_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
2020 if (!boot_hyp_pgd) {
2021 kvm_err("Hyp boot PGD not allocated\n");
2026 err = kvm_map_idmap_text(boot_hyp_pgd);
2030 merged_hyp_pgd = (pgd_t *)__get_free_page(GFP_KERNEL | __GFP_ZERO);
2031 if (!merged_hyp_pgd) {
2032 kvm_err("Failed to allocate extra HYP pgd\n");
2035 __kvm_extend_hypmap(boot_hyp_pgd, hyp_pgd, merged_hyp_pgd,
2038 err = kvm_map_idmap_text(hyp_pgd);
2043 io_map_base = hyp_idmap_start;
2050 void kvm_arch_commit_memory_region(struct kvm *kvm,
2051 const struct kvm_userspace_memory_region *mem,
2052 const struct kvm_memory_slot *old,
2053 const struct kvm_memory_slot *new,
2054 enum kvm_mr_change change)
2057 * At this point memslot has been committed and there is an
2058 * allocated dirty_bitmap[], dirty pages will be be tracked while the
2059 * memory slot is write protected.
2061 if (change != KVM_MR_DELETE && mem->flags & KVM_MEM_LOG_DIRTY_PAGES)
2062 kvm_mmu_wp_memory_region(kvm, mem->slot);
2065 int kvm_arch_prepare_memory_region(struct kvm *kvm,
2066 struct kvm_memory_slot *memslot,
2067 const struct kvm_userspace_memory_region *mem,
2068 enum kvm_mr_change change)
2070 hva_t hva = mem->userspace_addr;
2071 hva_t reg_end = hva + mem->memory_size;
2072 bool writable = !(mem->flags & KVM_MEM_READONLY);
2075 if (change != KVM_MR_CREATE && change != KVM_MR_MOVE &&
2076 change != KVM_MR_FLAGS_ONLY)
2080 * Prevent userspace from creating a memory region outside of the IPA
2081 * space addressable by the KVM guest IPA space.
2083 if (memslot->base_gfn + memslot->npages >
2084 (KVM_PHYS_SIZE >> PAGE_SHIFT))
2087 down_read(¤t->mm->mmap_sem);
2089 * A memory region could potentially cover multiple VMAs, and any holes
2090 * between them, so iterate over all of them to find out if we can map
2091 * any of them right now.
2093 * +--------------------------------------------+
2094 * +---------------+----------------+ +----------------+
2095 * | : VMA 1 | VMA 2 | | VMA 3 : |
2096 * +---------------+----------------+ +----------------+
2098 * +--------------------------------------------+
2101 struct vm_area_struct *vma = find_vma(current->mm, hva);
2102 hva_t vm_start, vm_end;
2104 if (!vma || vma->vm_start >= reg_end)
2108 * Mapping a read-only VMA is only allowed if the
2109 * memory region is configured as read-only.
2111 if (writable && !(vma->vm_flags & VM_WRITE)) {
2117 * Take the intersection of this VMA with the memory region
2119 vm_start = max(hva, vma->vm_start);
2120 vm_end = min(reg_end, vma->vm_end);
2122 if (vma->vm_flags & VM_PFNMAP) {
2123 gpa_t gpa = mem->guest_phys_addr +
2124 (vm_start - mem->userspace_addr);
2127 pa = (phys_addr_t)vma->vm_pgoff << PAGE_SHIFT;
2128 pa += vm_start - vma->vm_start;
2130 /* IO region dirty page logging not allowed */
2131 if (memslot->flags & KVM_MEM_LOG_DIRTY_PAGES) {
2136 ret = kvm_phys_addr_ioremap(kvm, gpa, pa,
2143 } while (hva < reg_end);
2145 if (change == KVM_MR_FLAGS_ONLY)
2148 spin_lock(&kvm->mmu_lock);
2150 unmap_stage2_range(kvm, mem->guest_phys_addr, mem->memory_size);
2152 stage2_flush_memslot(kvm, memslot);
2153 spin_unlock(&kvm->mmu_lock);
2155 up_read(¤t->mm->mmap_sem);
2159 void kvm_arch_free_memslot(struct kvm *kvm, struct kvm_memory_slot *free,
2160 struct kvm_memory_slot *dont)
2164 int kvm_arch_create_memslot(struct kvm *kvm, struct kvm_memory_slot *slot,
2165 unsigned long npages)
2170 void kvm_arch_memslots_updated(struct kvm *kvm, u64 gen)
2174 void kvm_arch_flush_shadow_all(struct kvm *kvm)
2176 kvm_free_stage2_pgd(kvm);
2179 void kvm_arch_flush_shadow_memslot(struct kvm *kvm,
2180 struct kvm_memory_slot *slot)
2182 gpa_t gpa = slot->base_gfn << PAGE_SHIFT;
2183 phys_addr_t size = slot->npages << PAGE_SHIFT;
2185 spin_lock(&kvm->mmu_lock);
2186 unmap_stage2_range(kvm, gpa, size);
2187 spin_unlock(&kvm->mmu_lock);
2191 * See note at ARMv7 ARM B1.14.4 (TL;DR: S/W ops are not easily virtualized).
2194 * - S/W ops are local to a CPU (not broadcast)
2195 * - We have line migration behind our back (speculation)
2196 * - System caches don't support S/W at all (damn!)
2198 * In the face of the above, the best we can do is to try and convert
2199 * S/W ops to VA ops. Because the guest is not allowed to infer the
2200 * S/W to PA mapping, it can only use S/W to nuke the whole cache,
2201 * which is a rather good thing for us.
2203 * Also, it is only used when turning caches on/off ("The expected
2204 * usage of the cache maintenance instructions that operate by set/way
2205 * is associated with the cache maintenance instructions associated
2206 * with the powerdown and powerup of caches, if this is required by
2207 * the implementation.").
2209 * We use the following policy:
2211 * - If we trap a S/W operation, we enable VM trapping to detect
2212 * caches being turned on/off, and do a full clean.
2214 * - We flush the caches on both caches being turned on and off.
2216 * - Once the caches are enabled, we stop trapping VM ops.
2218 void kvm_set_way_flush(struct kvm_vcpu *vcpu)
2220 unsigned long hcr = *vcpu_hcr(vcpu);
2223 * If this is the first time we do a S/W operation
2224 * (i.e. HCR_TVM not set) flush the whole memory, and set the
2227 * Otherwise, rely on the VM trapping to wait for the MMU +
2228 * Caches to be turned off. At that point, we'll be able to
2229 * clean the caches again.
2231 if (!(hcr & HCR_TVM)) {
2232 trace_kvm_set_way_flush(*vcpu_pc(vcpu),
2233 vcpu_has_cache_enabled(vcpu));
2234 stage2_flush_vm(vcpu->kvm);
2235 *vcpu_hcr(vcpu) = hcr | HCR_TVM;
2239 void kvm_toggle_cache(struct kvm_vcpu *vcpu, bool was_enabled)
2241 bool now_enabled = vcpu_has_cache_enabled(vcpu);
2244 * If switching the MMU+caches on, need to invalidate the caches.
2245 * If switching it off, need to clean the caches.
2246 * Clean + invalidate does the trick always.
2248 if (now_enabled != was_enabled)
2249 stage2_flush_vm(vcpu->kvm);
2251 /* Caches are now on, stop trapping VM ops (until a S/W op) */
2253 *vcpu_hcr(vcpu) &= ~HCR_TVM;
2255 trace_kvm_toggle_cache(*vcpu_pc(vcpu), was_enabled, now_enabled);