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 <trace/events/kvm.h>
24 #include <asm/pgalloc.h>
25 #include <asm/cacheflush.h>
26 #include <asm/kvm_arm.h>
27 #include <asm/kvm_mmu.h>
28 #include <asm/kvm_mmio.h>
29 #include <asm/kvm_asm.h>
30 #include <asm/kvm_emulate.h>
35 static pgd_t *boot_hyp_pgd;
36 static pgd_t *hyp_pgd;
37 static pgd_t *merged_hyp_pgd;
38 static DEFINE_MUTEX(kvm_hyp_pgd_mutex);
40 static unsigned long hyp_idmap_start;
41 static unsigned long hyp_idmap_end;
42 static phys_addr_t hyp_idmap_vector;
44 #define S2_PGD_SIZE (PTRS_PER_S2_PGD * sizeof(pgd_t))
45 #define hyp_pgd_order get_order(PTRS_PER_PGD * sizeof(pgd_t))
47 #define KVM_S2PTE_FLAG_IS_IOMAP (1UL << 0)
48 #define KVM_S2_FLAG_LOGGING_ACTIVE (1UL << 1)
50 static bool memslot_is_logging(struct kvm_memory_slot *memslot)
52 return memslot->dirty_bitmap && !(memslot->flags & KVM_MEM_READONLY);
56 * kvm_flush_remote_tlbs() - flush all VM TLB entries for v7/8
57 * @kvm: pointer to kvm structure.
59 * Interface to HYP function to flush all VM TLB entries
61 void kvm_flush_remote_tlbs(struct kvm *kvm)
63 kvm_call_hyp(__kvm_tlb_flush_vmid, kvm);
66 static void kvm_tlb_flush_vmid_ipa(struct kvm *kvm, phys_addr_t ipa)
68 kvm_call_hyp(__kvm_tlb_flush_vmid_ipa, kvm, ipa);
72 * D-Cache management functions. They take the page table entries by
73 * value, as they are flushing the cache using the kernel mapping (or
76 static void kvm_flush_dcache_pte(pte_t pte)
78 __kvm_flush_dcache_pte(pte);
81 static void kvm_flush_dcache_pmd(pmd_t pmd)
83 __kvm_flush_dcache_pmd(pmd);
86 static void kvm_flush_dcache_pud(pud_t pud)
88 __kvm_flush_dcache_pud(pud);
91 static bool kvm_is_device_pfn(unsigned long pfn)
93 return !pfn_valid(pfn);
97 * stage2_dissolve_pmd() - clear and flush huge PMD entry
98 * @kvm: pointer to kvm structure.
100 * @pmd: pmd pointer for IPA
102 * Function clears a PMD entry, flushes addr 1st and 2nd stage TLBs. Marks all
103 * pages in the range dirty.
105 static void stage2_dissolve_pmd(struct kvm *kvm, phys_addr_t addr, pmd_t *pmd)
107 if (!pmd_thp_or_huge(*pmd))
111 kvm_tlb_flush_vmid_ipa(kvm, addr);
112 put_page(virt_to_page(pmd));
115 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
120 BUG_ON(max > KVM_NR_MEM_OBJS);
121 if (cache->nobjs >= min)
123 while (cache->nobjs < max) {
124 page = (void *)__get_free_page(PGALLOC_GFP);
127 cache->objects[cache->nobjs++] = page;
132 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
135 free_page((unsigned long)mc->objects[--mc->nobjs]);
138 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc)
142 BUG_ON(!mc || !mc->nobjs);
143 p = mc->objects[--mc->nobjs];
147 static void clear_stage2_pgd_entry(struct kvm *kvm, pgd_t *pgd, phys_addr_t addr)
149 pud_t *pud_table __maybe_unused = stage2_pud_offset(pgd, 0UL);
150 stage2_pgd_clear(pgd);
151 kvm_tlb_flush_vmid_ipa(kvm, addr);
152 stage2_pud_free(pud_table);
153 put_page(virt_to_page(pgd));
156 static void clear_stage2_pud_entry(struct kvm *kvm, pud_t *pud, phys_addr_t addr)
158 pmd_t *pmd_table __maybe_unused = stage2_pmd_offset(pud, 0);
159 VM_BUG_ON(stage2_pud_huge(*pud));
160 stage2_pud_clear(pud);
161 kvm_tlb_flush_vmid_ipa(kvm, addr);
162 stage2_pmd_free(pmd_table);
163 put_page(virt_to_page(pud));
166 static void clear_stage2_pmd_entry(struct kvm *kvm, pmd_t *pmd, phys_addr_t addr)
168 pte_t *pte_table = pte_offset_kernel(pmd, 0);
169 VM_BUG_ON(pmd_thp_or_huge(*pmd));
171 kvm_tlb_flush_vmid_ipa(kvm, addr);
172 pte_free_kernel(NULL, pte_table);
173 put_page(virt_to_page(pmd));
177 * Unmapping vs dcache management:
179 * If a guest maps certain memory pages as uncached, all writes will
180 * bypass the data cache and go directly to RAM. However, the CPUs
181 * can still speculate reads (not writes) and fill cache lines with
184 * Those cache lines will be *clean* cache lines though, so a
185 * clean+invalidate operation is equivalent to an invalidate
186 * operation, because no cache lines are marked dirty.
188 * Those clean cache lines could be filled prior to an uncached write
189 * by the guest, and the cache coherent IO subsystem would therefore
190 * end up writing old data to disk.
192 * This is why right after unmapping a page/section and invalidating
193 * the corresponding TLBs, we call kvm_flush_dcache_p*() to make sure
194 * the IO subsystem will never hit in the cache.
196 static void unmap_stage2_ptes(struct kvm *kvm, pmd_t *pmd,
197 phys_addr_t addr, phys_addr_t end)
199 phys_addr_t start_addr = addr;
200 pte_t *pte, *start_pte;
202 start_pte = pte = pte_offset_kernel(pmd, addr);
204 if (!pte_none(*pte)) {
205 pte_t old_pte = *pte;
207 kvm_set_pte(pte, __pte(0));
208 kvm_tlb_flush_vmid_ipa(kvm, addr);
210 /* No need to invalidate the cache for device mappings */
211 if (!kvm_is_device_pfn(pte_pfn(old_pte)))
212 kvm_flush_dcache_pte(old_pte);
214 put_page(virt_to_page(pte));
216 } while (pte++, addr += PAGE_SIZE, addr != end);
218 if (stage2_pte_table_empty(start_pte))
219 clear_stage2_pmd_entry(kvm, pmd, start_addr);
222 static void unmap_stage2_pmds(struct kvm *kvm, pud_t *pud,
223 phys_addr_t addr, phys_addr_t end)
225 phys_addr_t next, start_addr = addr;
226 pmd_t *pmd, *start_pmd;
228 start_pmd = pmd = stage2_pmd_offset(pud, addr);
230 next = stage2_pmd_addr_end(addr, end);
231 if (!pmd_none(*pmd)) {
232 if (pmd_thp_or_huge(*pmd)) {
233 pmd_t old_pmd = *pmd;
236 kvm_tlb_flush_vmid_ipa(kvm, addr);
238 kvm_flush_dcache_pmd(old_pmd);
240 put_page(virt_to_page(pmd));
242 unmap_stage2_ptes(kvm, pmd, addr, next);
245 } while (pmd++, addr = next, addr != end);
247 if (stage2_pmd_table_empty(start_pmd))
248 clear_stage2_pud_entry(kvm, pud, start_addr);
251 static void unmap_stage2_puds(struct kvm *kvm, pgd_t *pgd,
252 phys_addr_t addr, phys_addr_t end)
254 phys_addr_t next, start_addr = addr;
255 pud_t *pud, *start_pud;
257 start_pud = pud = stage2_pud_offset(pgd, addr);
259 next = stage2_pud_addr_end(addr, end);
260 if (!stage2_pud_none(*pud)) {
261 if (stage2_pud_huge(*pud)) {
262 pud_t old_pud = *pud;
264 stage2_pud_clear(pud);
265 kvm_tlb_flush_vmid_ipa(kvm, addr);
266 kvm_flush_dcache_pud(old_pud);
267 put_page(virt_to_page(pud));
269 unmap_stage2_pmds(kvm, pud, addr, next);
272 } while (pud++, addr = next, addr != end);
274 if (stage2_pud_table_empty(start_pud))
275 clear_stage2_pgd_entry(kvm, pgd, start_addr);
279 * unmap_stage2_range -- Clear stage2 page table entries to unmap a range
280 * @kvm: The VM pointer
281 * @start: The intermediate physical base address of the range to unmap
282 * @size: The size of the area to unmap
284 * Clear a range of stage-2 mappings, lowering the various ref-counts. Must
285 * be called while holding mmu_lock (unless for freeing the stage2 pgd before
286 * destroying the VM), otherwise another faulting VCPU may come in and mess
287 * with things behind our backs.
289 static void unmap_stage2_range(struct kvm *kvm, phys_addr_t start, u64 size)
292 phys_addr_t addr = start, end = start + size;
295 assert_spin_locked(&kvm->mmu_lock);
296 pgd = kvm->arch.pgd + stage2_pgd_index(addr);
299 * Make sure the page table is still active, as another thread
300 * could have possibly freed the page table, while we released
303 if (!READ_ONCE(kvm->arch.pgd))
305 next = stage2_pgd_addr_end(addr, end);
306 if (!stage2_pgd_none(*pgd))
307 unmap_stage2_puds(kvm, pgd, addr, next);
309 * If the range is too large, release the kvm->mmu_lock
310 * to prevent starvation and lockup detector warnings.
313 cond_resched_lock(&kvm->mmu_lock);
314 } while (pgd++, addr = next, addr != end);
317 static void stage2_flush_ptes(struct kvm *kvm, pmd_t *pmd,
318 phys_addr_t addr, phys_addr_t end)
322 pte = pte_offset_kernel(pmd, addr);
324 if (!pte_none(*pte) && !kvm_is_device_pfn(pte_pfn(*pte)))
325 kvm_flush_dcache_pte(*pte);
326 } while (pte++, addr += PAGE_SIZE, addr != end);
329 static void stage2_flush_pmds(struct kvm *kvm, pud_t *pud,
330 phys_addr_t addr, phys_addr_t end)
335 pmd = stage2_pmd_offset(pud, addr);
337 next = stage2_pmd_addr_end(addr, end);
338 if (!pmd_none(*pmd)) {
339 if (pmd_thp_or_huge(*pmd))
340 kvm_flush_dcache_pmd(*pmd);
342 stage2_flush_ptes(kvm, pmd, addr, next);
344 } while (pmd++, addr = next, addr != end);
347 static void stage2_flush_puds(struct kvm *kvm, pgd_t *pgd,
348 phys_addr_t addr, phys_addr_t end)
353 pud = stage2_pud_offset(pgd, addr);
355 next = stage2_pud_addr_end(addr, end);
356 if (!stage2_pud_none(*pud)) {
357 if (stage2_pud_huge(*pud))
358 kvm_flush_dcache_pud(*pud);
360 stage2_flush_pmds(kvm, pud, addr, next);
362 } while (pud++, addr = next, addr != end);
365 static void stage2_flush_memslot(struct kvm *kvm,
366 struct kvm_memory_slot *memslot)
368 phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
369 phys_addr_t end = addr + PAGE_SIZE * memslot->npages;
373 pgd = kvm->arch.pgd + stage2_pgd_index(addr);
375 next = stage2_pgd_addr_end(addr, end);
376 stage2_flush_puds(kvm, pgd, addr, next);
377 } while (pgd++, addr = next, addr != end);
381 * stage2_flush_vm - Invalidate cache for pages mapped in stage 2
382 * @kvm: The struct kvm pointer
384 * Go through the stage 2 page tables and invalidate any cache lines
385 * backing memory already mapped to the VM.
387 static void stage2_flush_vm(struct kvm *kvm)
389 struct kvm_memslots *slots;
390 struct kvm_memory_slot *memslot;
393 idx = srcu_read_lock(&kvm->srcu);
394 spin_lock(&kvm->mmu_lock);
396 slots = kvm_memslots(kvm);
397 kvm_for_each_memslot(memslot, slots)
398 stage2_flush_memslot(kvm, memslot);
400 spin_unlock(&kvm->mmu_lock);
401 srcu_read_unlock(&kvm->srcu, idx);
404 static void clear_hyp_pgd_entry(pgd_t *pgd)
406 pud_t *pud_table __maybe_unused = pud_offset(pgd, 0UL);
408 pud_free(NULL, pud_table);
409 put_page(virt_to_page(pgd));
412 static void clear_hyp_pud_entry(pud_t *pud)
414 pmd_t *pmd_table __maybe_unused = pmd_offset(pud, 0);
415 VM_BUG_ON(pud_huge(*pud));
417 pmd_free(NULL, pmd_table);
418 put_page(virt_to_page(pud));
421 static void clear_hyp_pmd_entry(pmd_t *pmd)
423 pte_t *pte_table = pte_offset_kernel(pmd, 0);
424 VM_BUG_ON(pmd_thp_or_huge(*pmd));
426 pte_free_kernel(NULL, pte_table);
427 put_page(virt_to_page(pmd));
430 static void unmap_hyp_ptes(pmd_t *pmd, phys_addr_t addr, phys_addr_t end)
432 pte_t *pte, *start_pte;
434 start_pte = pte = pte_offset_kernel(pmd, addr);
436 if (!pte_none(*pte)) {
437 kvm_set_pte(pte, __pte(0));
438 put_page(virt_to_page(pte));
440 } while (pte++, addr += PAGE_SIZE, addr != end);
442 if (hyp_pte_table_empty(start_pte))
443 clear_hyp_pmd_entry(pmd);
446 static void unmap_hyp_pmds(pud_t *pud, phys_addr_t addr, phys_addr_t end)
449 pmd_t *pmd, *start_pmd;
451 start_pmd = pmd = pmd_offset(pud, addr);
453 next = pmd_addr_end(addr, end);
454 /* Hyp doesn't use huge pmds */
456 unmap_hyp_ptes(pmd, addr, next);
457 } while (pmd++, addr = next, addr != end);
459 if (hyp_pmd_table_empty(start_pmd))
460 clear_hyp_pud_entry(pud);
463 static void unmap_hyp_puds(pgd_t *pgd, phys_addr_t addr, phys_addr_t end)
466 pud_t *pud, *start_pud;
468 start_pud = pud = pud_offset(pgd, addr);
470 next = pud_addr_end(addr, end);
471 /* Hyp doesn't use huge puds */
473 unmap_hyp_pmds(pud, addr, next);
474 } while (pud++, addr = next, addr != end);
476 if (hyp_pud_table_empty(start_pud))
477 clear_hyp_pgd_entry(pgd);
480 static void unmap_hyp_range(pgd_t *pgdp, phys_addr_t start, u64 size)
483 phys_addr_t addr = start, end = start + size;
487 * We don't unmap anything from HYP, except at the hyp tear down.
488 * Hence, we don't have to invalidate the TLBs here.
490 pgd = pgdp + pgd_index(addr);
492 next = pgd_addr_end(addr, end);
494 unmap_hyp_puds(pgd, addr, next);
495 } while (pgd++, addr = next, addr != end);
499 * free_hyp_pgds - free Hyp-mode page tables
501 * Assumes hyp_pgd is a page table used strictly in Hyp-mode and
502 * therefore contains either mappings in the kernel memory area (above
503 * PAGE_OFFSET), or device mappings in the vmalloc range (from
504 * VMALLOC_START to VMALLOC_END).
506 * boot_hyp_pgd should only map two pages for the init code.
508 void free_hyp_pgds(void)
512 mutex_lock(&kvm_hyp_pgd_mutex);
515 unmap_hyp_range(boot_hyp_pgd, hyp_idmap_start, PAGE_SIZE);
516 free_pages((unsigned long)boot_hyp_pgd, hyp_pgd_order);
521 unmap_hyp_range(hyp_pgd, hyp_idmap_start, PAGE_SIZE);
522 for (addr = PAGE_OFFSET; virt_addr_valid(addr); addr += PGDIR_SIZE)
523 unmap_hyp_range(hyp_pgd, kern_hyp_va(addr), PGDIR_SIZE);
524 for (addr = VMALLOC_START; is_vmalloc_addr((void*)addr); addr += PGDIR_SIZE)
525 unmap_hyp_range(hyp_pgd, kern_hyp_va(addr), PGDIR_SIZE);
527 free_pages((unsigned long)hyp_pgd, hyp_pgd_order);
530 if (merged_hyp_pgd) {
531 clear_page(merged_hyp_pgd);
532 free_page((unsigned long)merged_hyp_pgd);
533 merged_hyp_pgd = NULL;
536 mutex_unlock(&kvm_hyp_pgd_mutex);
539 static void create_hyp_pte_mappings(pmd_t *pmd, unsigned long start,
540 unsigned long end, unsigned long pfn,
548 pte = pte_offset_kernel(pmd, addr);
549 kvm_set_pte(pte, pfn_pte(pfn, prot));
550 get_page(virt_to_page(pte));
551 kvm_flush_dcache_to_poc(pte, sizeof(*pte));
553 } while (addr += PAGE_SIZE, addr != end);
556 static int create_hyp_pmd_mappings(pud_t *pud, unsigned long start,
557 unsigned long end, unsigned long pfn,
562 unsigned long addr, next;
566 pmd = pmd_offset(pud, addr);
568 BUG_ON(pmd_sect(*pmd));
570 if (pmd_none(*pmd)) {
571 pte = pte_alloc_one_kernel(NULL, addr);
573 kvm_err("Cannot allocate Hyp pte\n");
576 pmd_populate_kernel(NULL, pmd, pte);
577 get_page(virt_to_page(pmd));
578 kvm_flush_dcache_to_poc(pmd, sizeof(*pmd));
581 next = pmd_addr_end(addr, end);
583 create_hyp_pte_mappings(pmd, addr, next, pfn, prot);
584 pfn += (next - addr) >> PAGE_SHIFT;
585 } while (addr = next, addr != end);
590 static int create_hyp_pud_mappings(pgd_t *pgd, unsigned long start,
591 unsigned long end, unsigned long pfn,
596 unsigned long addr, next;
601 pud = pud_offset(pgd, addr);
603 if (pud_none_or_clear_bad(pud)) {
604 pmd = pmd_alloc_one(NULL, addr);
606 kvm_err("Cannot allocate Hyp pmd\n");
609 pud_populate(NULL, pud, pmd);
610 get_page(virt_to_page(pud));
611 kvm_flush_dcache_to_poc(pud, sizeof(*pud));
614 next = pud_addr_end(addr, end);
615 ret = create_hyp_pmd_mappings(pud, addr, next, pfn, prot);
618 pfn += (next - addr) >> PAGE_SHIFT;
619 } while (addr = next, addr != end);
624 static int __create_hyp_mappings(pgd_t *pgdp,
625 unsigned long start, unsigned long end,
626 unsigned long pfn, pgprot_t prot)
630 unsigned long addr, next;
633 mutex_lock(&kvm_hyp_pgd_mutex);
634 addr = start & PAGE_MASK;
635 end = PAGE_ALIGN(end);
637 pgd = pgdp + pgd_index(addr);
639 if (pgd_none(*pgd)) {
640 pud = pud_alloc_one(NULL, addr);
642 kvm_err("Cannot allocate Hyp pud\n");
646 pgd_populate(NULL, pgd, pud);
647 get_page(virt_to_page(pgd));
648 kvm_flush_dcache_to_poc(pgd, sizeof(*pgd));
651 next = pgd_addr_end(addr, end);
652 err = create_hyp_pud_mappings(pgd, addr, next, pfn, prot);
655 pfn += (next - addr) >> PAGE_SHIFT;
656 } while (addr = next, addr != end);
658 mutex_unlock(&kvm_hyp_pgd_mutex);
662 static phys_addr_t kvm_kaddr_to_phys(void *kaddr)
664 if (!is_vmalloc_addr(kaddr)) {
665 BUG_ON(!virt_addr_valid(kaddr));
668 return page_to_phys(vmalloc_to_page(kaddr)) +
669 offset_in_page(kaddr);
674 * create_hyp_mappings - duplicate a kernel virtual address range in Hyp mode
675 * @from: The virtual kernel start address of the range
676 * @to: The virtual kernel end address of the range (exclusive)
677 * @prot: The protection to be applied to this range
679 * The same virtual address as the kernel virtual address is also used
680 * in Hyp-mode mapping (modulo HYP_PAGE_OFFSET) to the same underlying
683 int create_hyp_mappings(void *from, void *to, pgprot_t prot)
685 phys_addr_t phys_addr;
686 unsigned long virt_addr;
687 unsigned long start = kern_hyp_va((unsigned long)from);
688 unsigned long end = kern_hyp_va((unsigned long)to);
690 if (is_kernel_in_hyp_mode())
693 start = start & PAGE_MASK;
694 end = PAGE_ALIGN(end);
696 for (virt_addr = start; virt_addr < end; virt_addr += PAGE_SIZE) {
699 phys_addr = kvm_kaddr_to_phys(from + virt_addr - start);
700 err = __create_hyp_mappings(hyp_pgd, virt_addr,
701 virt_addr + PAGE_SIZE,
702 __phys_to_pfn(phys_addr),
712 * create_hyp_io_mappings - duplicate a kernel IO mapping into Hyp mode
713 * @from: The kernel start VA of the range
714 * @to: The kernel end VA of the range (exclusive)
715 * @phys_addr: The physical start address which gets mapped
717 * The resulting HYP VA is the same as the kernel VA, modulo
720 int create_hyp_io_mappings(void *from, void *to, phys_addr_t phys_addr)
722 unsigned long start = kern_hyp_va((unsigned long)from);
723 unsigned long end = kern_hyp_va((unsigned long)to);
725 if (is_kernel_in_hyp_mode())
728 /* Check for a valid kernel IO mapping */
729 if (!is_vmalloc_addr(from) || !is_vmalloc_addr(to - 1))
732 return __create_hyp_mappings(hyp_pgd, start, end,
733 __phys_to_pfn(phys_addr), PAGE_HYP_DEVICE);
737 * kvm_alloc_stage2_pgd - allocate level-1 table for stage-2 translation.
738 * @kvm: The KVM struct pointer for the VM.
740 * Allocates only the stage-2 HW PGD level table(s) (can support either full
741 * 40-bit input addresses or limited to 32-bit input addresses). Clears the
744 * Note we don't need locking here as this is only called when the VM is
745 * created, which can only be done once.
747 int kvm_alloc_stage2_pgd(struct kvm *kvm)
751 if (kvm->arch.pgd != NULL) {
752 kvm_err("kvm_arch already initialized?\n");
756 /* Allocate the HW PGD, making sure that each page gets its own refcount */
757 pgd = alloc_pages_exact(S2_PGD_SIZE, GFP_KERNEL | __GFP_ZERO);
765 static void stage2_unmap_memslot(struct kvm *kvm,
766 struct kvm_memory_slot *memslot)
768 hva_t hva = memslot->userspace_addr;
769 phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
770 phys_addr_t size = PAGE_SIZE * memslot->npages;
771 hva_t reg_end = hva + size;
774 * A memory region could potentially cover multiple VMAs, and any holes
775 * between them, so iterate over all of them to find out if we should
778 * +--------------------------------------------+
779 * +---------------+----------------+ +----------------+
780 * | : VMA 1 | VMA 2 | | VMA 3 : |
781 * +---------------+----------------+ +----------------+
783 * +--------------------------------------------+
786 struct vm_area_struct *vma = find_vma(current->mm, hva);
787 hva_t vm_start, vm_end;
789 if (!vma || vma->vm_start >= reg_end)
793 * Take the intersection of this VMA with the memory region
795 vm_start = max(hva, vma->vm_start);
796 vm_end = min(reg_end, vma->vm_end);
798 if (!(vma->vm_flags & VM_PFNMAP)) {
799 gpa_t gpa = addr + (vm_start - memslot->userspace_addr);
800 unmap_stage2_range(kvm, gpa, vm_end - vm_start);
803 } while (hva < reg_end);
807 * stage2_unmap_vm - Unmap Stage-2 RAM mappings
808 * @kvm: The struct kvm pointer
810 * Go through the memregions and unmap any reguler RAM
811 * backing memory already mapped to the VM.
813 void stage2_unmap_vm(struct kvm *kvm)
815 struct kvm_memslots *slots;
816 struct kvm_memory_slot *memslot;
819 idx = srcu_read_lock(&kvm->srcu);
820 down_read(¤t->mm->mmap_sem);
821 spin_lock(&kvm->mmu_lock);
823 slots = kvm_memslots(kvm);
824 kvm_for_each_memslot(memslot, slots)
825 stage2_unmap_memslot(kvm, memslot);
827 spin_unlock(&kvm->mmu_lock);
828 up_read(¤t->mm->mmap_sem);
829 srcu_read_unlock(&kvm->srcu, idx);
833 * kvm_free_stage2_pgd - free all stage-2 tables
834 * @kvm: The KVM struct pointer for the VM.
836 * Walks the level-1 page table pointed to by kvm->arch.pgd and frees all
837 * underlying level-2 and level-3 tables before freeing the actual level-1 table
838 * and setting the struct pointer to NULL.
840 void kvm_free_stage2_pgd(struct kvm *kvm)
844 spin_lock(&kvm->mmu_lock);
846 unmap_stage2_range(kvm, 0, KVM_PHYS_SIZE);
847 pgd = READ_ONCE(kvm->arch.pgd);
848 kvm->arch.pgd = NULL;
850 spin_unlock(&kvm->mmu_lock);
852 /* Free the HW pgd, one page at a time */
854 free_pages_exact(pgd, S2_PGD_SIZE);
857 static pud_t *stage2_get_pud(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
863 pgd = kvm->arch.pgd + stage2_pgd_index(addr);
864 if (WARN_ON(stage2_pgd_none(*pgd))) {
867 pud = mmu_memory_cache_alloc(cache);
868 stage2_pgd_populate(pgd, pud);
869 get_page(virt_to_page(pgd));
872 return stage2_pud_offset(pgd, addr);
875 static pmd_t *stage2_get_pmd(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
881 pud = stage2_get_pud(kvm, cache, addr);
882 if (stage2_pud_none(*pud)) {
885 pmd = mmu_memory_cache_alloc(cache);
886 stage2_pud_populate(pud, pmd);
887 get_page(virt_to_page(pud));
890 return stage2_pmd_offset(pud, addr);
893 static int stage2_set_pmd_huge(struct kvm *kvm, struct kvm_mmu_memory_cache
894 *cache, phys_addr_t addr, const pmd_t *new_pmd)
898 pmd = stage2_get_pmd(kvm, cache, addr);
902 * Mapping in huge pages should only happen through a fault. If a
903 * page is merged into a transparent huge page, the individual
904 * subpages of that huge page should be unmapped through MMU
905 * notifiers before we get here.
907 * Merging of CompoundPages is not supported; they should become
908 * splitting first, unmapped, merged, and mapped back in on-demand.
910 VM_BUG_ON(pmd_present(*pmd) && pmd_pfn(*pmd) != pmd_pfn(*new_pmd));
913 if (pmd_present(old_pmd)) {
915 kvm_tlb_flush_vmid_ipa(kvm, addr);
917 get_page(virt_to_page(pmd));
920 kvm_set_pmd(pmd, *new_pmd);
924 static int stage2_set_pte(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
925 phys_addr_t addr, const pte_t *new_pte,
930 bool iomap = flags & KVM_S2PTE_FLAG_IS_IOMAP;
931 bool logging_active = flags & KVM_S2_FLAG_LOGGING_ACTIVE;
933 VM_BUG_ON(logging_active && !cache);
935 /* Create stage-2 page table mapping - Levels 0 and 1 */
936 pmd = stage2_get_pmd(kvm, cache, addr);
939 * Ignore calls from kvm_set_spte_hva for unallocated
946 * While dirty page logging - dissolve huge PMD, then continue on to
950 stage2_dissolve_pmd(kvm, addr, pmd);
952 /* Create stage-2 page mappings - Level 2 */
953 if (pmd_none(*pmd)) {
955 return 0; /* ignore calls from kvm_set_spte_hva */
956 pte = mmu_memory_cache_alloc(cache);
957 pmd_populate_kernel(NULL, pmd, pte);
958 get_page(virt_to_page(pmd));
961 pte = pte_offset_kernel(pmd, addr);
963 if (iomap && pte_present(*pte))
966 /* Create 2nd stage page table mapping - Level 3 */
968 if (pte_present(old_pte)) {
969 kvm_set_pte(pte, __pte(0));
970 kvm_tlb_flush_vmid_ipa(kvm, addr);
972 get_page(virt_to_page(pte));
975 kvm_set_pte(pte, *new_pte);
979 #ifndef __HAVE_ARCH_PTEP_TEST_AND_CLEAR_YOUNG
980 static int stage2_ptep_test_and_clear_young(pte_t *pte)
982 if (pte_young(*pte)) {
983 *pte = pte_mkold(*pte);
989 static int stage2_ptep_test_and_clear_young(pte_t *pte)
991 return __ptep_test_and_clear_young(pte);
995 static int stage2_pmdp_test_and_clear_young(pmd_t *pmd)
997 return stage2_ptep_test_and_clear_young((pte_t *)pmd);
1001 * kvm_phys_addr_ioremap - map a device range to guest IPA
1003 * @kvm: The KVM pointer
1004 * @guest_ipa: The IPA at which to insert the mapping
1005 * @pa: The physical address of the device
1006 * @size: The size of the mapping
1008 int kvm_phys_addr_ioremap(struct kvm *kvm, phys_addr_t guest_ipa,
1009 phys_addr_t pa, unsigned long size, bool writable)
1011 phys_addr_t addr, end;
1014 struct kvm_mmu_memory_cache cache = { 0, };
1016 end = (guest_ipa + size + PAGE_SIZE - 1) & PAGE_MASK;
1017 pfn = __phys_to_pfn(pa);
1019 for (addr = guest_ipa; addr < end; addr += PAGE_SIZE) {
1020 pte_t pte = pfn_pte(pfn, PAGE_S2_DEVICE);
1023 pte = kvm_s2pte_mkwrite(pte);
1025 ret = mmu_topup_memory_cache(&cache, KVM_MMU_CACHE_MIN_PAGES,
1029 spin_lock(&kvm->mmu_lock);
1030 ret = stage2_set_pte(kvm, &cache, addr, &pte,
1031 KVM_S2PTE_FLAG_IS_IOMAP);
1032 spin_unlock(&kvm->mmu_lock);
1040 mmu_free_memory_cache(&cache);
1044 static bool transparent_hugepage_adjust(kvm_pfn_t *pfnp, phys_addr_t *ipap)
1046 kvm_pfn_t pfn = *pfnp;
1047 gfn_t gfn = *ipap >> PAGE_SHIFT;
1049 if (PageTransCompoundMap(pfn_to_page(pfn))) {
1052 * The address we faulted on is backed by a transparent huge
1053 * page. However, because we map the compound huge page and
1054 * not the individual tail page, we need to transfer the
1055 * refcount to the head page. We have to be careful that the
1056 * THP doesn't start to split while we are adjusting the
1059 * We are sure this doesn't happen, because mmu_notifier_retry
1060 * was successful and we are holding the mmu_lock, so if this
1061 * THP is trying to split, it will be blocked in the mmu
1062 * notifier before touching any of the pages, specifically
1063 * before being able to call __split_huge_page_refcount().
1065 * We can therefore safely transfer the refcount from PG_tail
1066 * to PG_head and switch the pfn from a tail page to the head
1069 mask = PTRS_PER_PMD - 1;
1070 VM_BUG_ON((gfn & mask) != (pfn & mask));
1073 kvm_release_pfn_clean(pfn);
1085 static bool kvm_is_write_fault(struct kvm_vcpu *vcpu)
1087 if (kvm_vcpu_trap_is_iabt(vcpu))
1090 return kvm_vcpu_dabt_iswrite(vcpu);
1094 * stage2_wp_ptes - write protect PMD range
1095 * @pmd: pointer to pmd entry
1096 * @addr: range start address
1097 * @end: range end address
1099 static void stage2_wp_ptes(pmd_t *pmd, phys_addr_t addr, phys_addr_t end)
1103 pte = pte_offset_kernel(pmd, addr);
1105 if (!pte_none(*pte)) {
1106 if (!kvm_s2pte_readonly(pte))
1107 kvm_set_s2pte_readonly(pte);
1109 } while (pte++, addr += PAGE_SIZE, addr != end);
1113 * stage2_wp_pmds - write protect PUD range
1114 * @pud: pointer to pud entry
1115 * @addr: range start address
1116 * @end: range end address
1118 static void stage2_wp_pmds(pud_t *pud, phys_addr_t addr, phys_addr_t end)
1123 pmd = stage2_pmd_offset(pud, addr);
1126 next = stage2_pmd_addr_end(addr, end);
1127 if (!pmd_none(*pmd)) {
1128 if (pmd_thp_or_huge(*pmd)) {
1129 if (!kvm_s2pmd_readonly(pmd))
1130 kvm_set_s2pmd_readonly(pmd);
1132 stage2_wp_ptes(pmd, addr, next);
1135 } while (pmd++, addr = next, addr != end);
1139 * stage2_wp_puds - write protect PGD range
1140 * @pgd: pointer to pgd entry
1141 * @addr: range start address
1142 * @end: range end address
1144 * Process PUD entries, for a huge PUD we cause a panic.
1146 static void stage2_wp_puds(pgd_t *pgd, phys_addr_t addr, phys_addr_t end)
1151 pud = stage2_pud_offset(pgd, addr);
1153 next = stage2_pud_addr_end(addr, end);
1154 if (!stage2_pud_none(*pud)) {
1155 /* TODO:PUD not supported, revisit later if supported */
1156 BUG_ON(stage2_pud_huge(*pud));
1157 stage2_wp_pmds(pud, addr, next);
1159 } while (pud++, addr = next, addr != end);
1163 * stage2_wp_range() - write protect stage2 memory region range
1164 * @kvm: The KVM pointer
1165 * @addr: Start address of range
1166 * @end: End address of range
1168 static void stage2_wp_range(struct kvm *kvm, phys_addr_t addr, phys_addr_t end)
1173 pgd = kvm->arch.pgd + stage2_pgd_index(addr);
1176 * Release kvm_mmu_lock periodically if the memory region is
1177 * large. Otherwise, we may see kernel panics with
1178 * CONFIG_DETECT_HUNG_TASK, CONFIG_LOCKUP_DETECTOR,
1179 * CONFIG_LOCKDEP. Additionally, holding the lock too long
1180 * will also starve other vCPUs. We have to also make sure
1181 * that the page tables are not freed while we released
1184 cond_resched_lock(&kvm->mmu_lock);
1185 if (!READ_ONCE(kvm->arch.pgd))
1187 next = stage2_pgd_addr_end(addr, end);
1188 if (stage2_pgd_present(*pgd))
1189 stage2_wp_puds(pgd, addr, next);
1190 } while (pgd++, addr = next, addr != end);
1194 * kvm_mmu_wp_memory_region() - write protect stage 2 entries for memory slot
1195 * @kvm: The KVM pointer
1196 * @slot: The memory slot to write protect
1198 * Called to start logging dirty pages after memory region
1199 * KVM_MEM_LOG_DIRTY_PAGES operation is called. After this function returns
1200 * all present PMD and PTEs are write protected in the memory region.
1201 * Afterwards read of dirty page log can be called.
1203 * Acquires kvm_mmu_lock. Called with kvm->slots_lock mutex acquired,
1204 * serializing operations for VM memory regions.
1206 void kvm_mmu_wp_memory_region(struct kvm *kvm, int slot)
1208 struct kvm_memslots *slots = kvm_memslots(kvm);
1209 struct kvm_memory_slot *memslot = id_to_memslot(slots, slot);
1210 phys_addr_t start = memslot->base_gfn << PAGE_SHIFT;
1211 phys_addr_t end = (memslot->base_gfn + memslot->npages) << PAGE_SHIFT;
1213 spin_lock(&kvm->mmu_lock);
1214 stage2_wp_range(kvm, start, end);
1215 spin_unlock(&kvm->mmu_lock);
1216 kvm_flush_remote_tlbs(kvm);
1220 * kvm_mmu_write_protect_pt_masked() - write protect dirty pages
1221 * @kvm: The KVM pointer
1222 * @slot: The memory slot associated with mask
1223 * @gfn_offset: The gfn offset in memory slot
1224 * @mask: The mask of dirty pages at offset 'gfn_offset' in this memory
1225 * slot to be write protected
1227 * Walks bits set in mask write protects the associated pte's. Caller must
1228 * acquire kvm_mmu_lock.
1230 static void kvm_mmu_write_protect_pt_masked(struct kvm *kvm,
1231 struct kvm_memory_slot *slot,
1232 gfn_t gfn_offset, unsigned long mask)
1234 phys_addr_t base_gfn = slot->base_gfn + gfn_offset;
1235 phys_addr_t start = (base_gfn + __ffs(mask)) << PAGE_SHIFT;
1236 phys_addr_t end = (base_gfn + __fls(mask) + 1) << PAGE_SHIFT;
1238 stage2_wp_range(kvm, start, end);
1242 * kvm_arch_mmu_enable_log_dirty_pt_masked - enable dirty logging for selected
1245 * It calls kvm_mmu_write_protect_pt_masked to write protect selected pages to
1246 * enable dirty logging for them.
1248 void kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm *kvm,
1249 struct kvm_memory_slot *slot,
1250 gfn_t gfn_offset, unsigned long mask)
1252 kvm_mmu_write_protect_pt_masked(kvm, slot, gfn_offset, mask);
1255 static void coherent_cache_guest_page(struct kvm_vcpu *vcpu, kvm_pfn_t pfn,
1258 __coherent_cache_guest_page(vcpu, pfn, size);
1261 static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
1262 struct kvm_memory_slot *memslot, unsigned long hva,
1263 unsigned long fault_status)
1266 bool write_fault, writable, hugetlb = false, force_pte = false;
1267 unsigned long mmu_seq;
1268 gfn_t gfn = fault_ipa >> PAGE_SHIFT;
1269 struct kvm *kvm = vcpu->kvm;
1270 struct kvm_mmu_memory_cache *memcache = &vcpu->arch.mmu_page_cache;
1271 struct vm_area_struct *vma;
1273 pgprot_t mem_type = PAGE_S2;
1274 bool logging_active = memslot_is_logging(memslot);
1275 unsigned long flags = 0;
1277 write_fault = kvm_is_write_fault(vcpu);
1278 if (fault_status == FSC_PERM && !write_fault) {
1279 kvm_err("Unexpected L2 read permission error\n");
1283 /* Let's check if we will get back a huge page backed by hugetlbfs */
1284 down_read(¤t->mm->mmap_sem);
1285 vma = find_vma_intersection(current->mm, hva, hva + 1);
1286 if (unlikely(!vma)) {
1287 kvm_err("Failed to find VMA for hva 0x%lx\n", hva);
1288 up_read(¤t->mm->mmap_sem);
1292 if (is_vm_hugetlb_page(vma) && !logging_active) {
1294 gfn = (fault_ipa & PMD_MASK) >> PAGE_SHIFT;
1297 * Pages belonging to memslots that don't have the same
1298 * alignment for userspace and IPA cannot be mapped using
1299 * block descriptors even if the pages belong to a THP for
1300 * the process, because the stage-2 block descriptor will
1301 * cover more than a single THP and we loose atomicity for
1302 * unmapping, updates, and splits of the THP or other pages
1303 * in the stage-2 block range.
1305 if ((memslot->userspace_addr & ~PMD_MASK) !=
1306 ((memslot->base_gfn << PAGE_SHIFT) & ~PMD_MASK))
1309 up_read(¤t->mm->mmap_sem);
1311 /* We need minimum second+third level pages */
1312 ret = mmu_topup_memory_cache(memcache, KVM_MMU_CACHE_MIN_PAGES,
1317 mmu_seq = vcpu->kvm->mmu_notifier_seq;
1319 * Ensure the read of mmu_notifier_seq happens before we call
1320 * gfn_to_pfn_prot (which calls get_user_pages), so that we don't risk
1321 * the page we just got a reference to gets unmapped before we have a
1322 * chance to grab the mmu_lock, which ensure that if the page gets
1323 * unmapped afterwards, the call to kvm_unmap_hva will take it away
1324 * from us again properly. This smp_rmb() interacts with the smp_wmb()
1325 * in kvm_mmu_notifier_invalidate_<page|range_end>.
1329 pfn = gfn_to_pfn_prot(kvm, gfn, write_fault, &writable);
1330 if (is_error_noslot_pfn(pfn))
1333 if (kvm_is_device_pfn(pfn)) {
1334 mem_type = PAGE_S2_DEVICE;
1335 flags |= KVM_S2PTE_FLAG_IS_IOMAP;
1336 } else if (logging_active) {
1338 * Faults on pages in a memslot with logging enabled
1339 * should not be mapped with huge pages (it introduces churn
1340 * and performance degradation), so force a pte mapping.
1343 flags |= KVM_S2_FLAG_LOGGING_ACTIVE;
1346 * Only actually map the page as writable if this was a write
1353 spin_lock(&kvm->mmu_lock);
1354 if (mmu_notifier_retry(kvm, mmu_seq))
1357 if (!hugetlb && !force_pte)
1358 hugetlb = transparent_hugepage_adjust(&pfn, &fault_ipa);
1361 pmd_t new_pmd = pfn_pmd(pfn, mem_type);
1362 new_pmd = pmd_mkhuge(new_pmd);
1364 new_pmd = kvm_s2pmd_mkwrite(new_pmd);
1365 kvm_set_pfn_dirty(pfn);
1367 coherent_cache_guest_page(vcpu, pfn, PMD_SIZE);
1368 ret = stage2_set_pmd_huge(kvm, memcache, fault_ipa, &new_pmd);
1370 pte_t new_pte = pfn_pte(pfn, mem_type);
1373 new_pte = kvm_s2pte_mkwrite(new_pte);
1374 kvm_set_pfn_dirty(pfn);
1375 mark_page_dirty(kvm, gfn);
1377 coherent_cache_guest_page(vcpu, pfn, PAGE_SIZE);
1378 ret = stage2_set_pte(kvm, memcache, fault_ipa, &new_pte, flags);
1382 spin_unlock(&kvm->mmu_lock);
1383 kvm_set_pfn_accessed(pfn);
1384 kvm_release_pfn_clean(pfn);
1389 * Resolve the access fault by making the page young again.
1390 * Note that because the faulting entry is guaranteed not to be
1391 * cached in the TLB, we don't need to invalidate anything.
1392 * Only the HW Access Flag updates are supported for Stage 2 (no DBM),
1393 * so there is no need for atomic (pte|pmd)_mkyoung operations.
1395 static void handle_access_fault(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa)
1400 bool pfn_valid = false;
1402 trace_kvm_access_fault(fault_ipa);
1404 spin_lock(&vcpu->kvm->mmu_lock);
1406 pmd = stage2_get_pmd(vcpu->kvm, NULL, fault_ipa);
1407 if (!pmd || pmd_none(*pmd)) /* Nothing there */
1410 if (pmd_thp_or_huge(*pmd)) { /* THP, HugeTLB */
1411 *pmd = pmd_mkyoung(*pmd);
1412 pfn = pmd_pfn(*pmd);
1417 pte = pte_offset_kernel(pmd, fault_ipa);
1418 if (pte_none(*pte)) /* Nothing there either */
1421 *pte = pte_mkyoung(*pte); /* Just a page... */
1422 pfn = pte_pfn(*pte);
1425 spin_unlock(&vcpu->kvm->mmu_lock);
1427 kvm_set_pfn_accessed(pfn);
1431 * kvm_handle_guest_abort - handles all 2nd stage aborts
1432 * @vcpu: the VCPU pointer
1433 * @run: the kvm_run structure
1435 * Any abort that gets to the host is almost guaranteed to be caused by a
1436 * missing second stage translation table entry, which can mean that either the
1437 * guest simply needs more memory and we must allocate an appropriate page or it
1438 * can mean that the guest tried to access I/O memory, which is emulated by user
1439 * space. The distinction is based on the IPA causing the fault and whether this
1440 * memory region has been registered as standard RAM by user space.
1442 int kvm_handle_guest_abort(struct kvm_vcpu *vcpu, struct kvm_run *run)
1444 unsigned long fault_status;
1445 phys_addr_t fault_ipa;
1446 struct kvm_memory_slot *memslot;
1448 bool is_iabt, write_fault, writable;
1452 is_iabt = kvm_vcpu_trap_is_iabt(vcpu);
1453 if (unlikely(!is_iabt && kvm_vcpu_dabt_isextabt(vcpu))) {
1454 kvm_inject_vabt(vcpu);
1458 fault_ipa = kvm_vcpu_get_fault_ipa(vcpu);
1460 trace_kvm_guest_fault(*vcpu_pc(vcpu), kvm_vcpu_get_hsr(vcpu),
1461 kvm_vcpu_get_hfar(vcpu), fault_ipa);
1463 /* Check the stage-2 fault is trans. fault or write fault */
1464 fault_status = kvm_vcpu_trap_get_fault_type(vcpu);
1465 if (fault_status != FSC_FAULT && fault_status != FSC_PERM &&
1466 fault_status != FSC_ACCESS) {
1467 kvm_err("Unsupported FSC: EC=%#x xFSC=%#lx ESR_EL2=%#lx\n",
1468 kvm_vcpu_trap_get_class(vcpu),
1469 (unsigned long)kvm_vcpu_trap_get_fault(vcpu),
1470 (unsigned long)kvm_vcpu_get_hsr(vcpu));
1474 idx = srcu_read_lock(&vcpu->kvm->srcu);
1476 gfn = fault_ipa >> PAGE_SHIFT;
1477 memslot = gfn_to_memslot(vcpu->kvm, gfn);
1478 hva = gfn_to_hva_memslot_prot(memslot, gfn, &writable);
1479 write_fault = kvm_is_write_fault(vcpu);
1480 if (kvm_is_error_hva(hva) || (write_fault && !writable)) {
1482 /* Prefetch Abort on I/O address */
1483 kvm_inject_pabt(vcpu, kvm_vcpu_get_hfar(vcpu));
1489 * Check for a cache maintenance operation. Since we
1490 * ended-up here, we know it is outside of any memory
1491 * slot. But we can't find out if that is for a device,
1492 * or if the guest is just being stupid. The only thing
1493 * we know for sure is that this range cannot be cached.
1495 * So let's assume that the guest is just being
1496 * cautious, and skip the instruction.
1498 if (kvm_vcpu_dabt_is_cm(vcpu)) {
1499 kvm_skip_instr(vcpu, kvm_vcpu_trap_il_is32bit(vcpu));
1505 * The IPA is reported as [MAX:12], so we need to
1506 * complement it with the bottom 12 bits from the
1507 * faulting VA. This is always 12 bits, irrespective
1510 fault_ipa |= kvm_vcpu_get_hfar(vcpu) & ((1 << 12) - 1);
1511 ret = io_mem_abort(vcpu, run, fault_ipa);
1515 /* Userspace should not be able to register out-of-bounds IPAs */
1516 VM_BUG_ON(fault_ipa >= KVM_PHYS_SIZE);
1518 if (fault_status == FSC_ACCESS) {
1519 handle_access_fault(vcpu, fault_ipa);
1524 ret = user_mem_abort(vcpu, fault_ipa, memslot, hva, fault_status);
1528 srcu_read_unlock(&vcpu->kvm->srcu, idx);
1532 static int handle_hva_to_gpa(struct kvm *kvm,
1533 unsigned long start,
1535 int (*handler)(struct kvm *kvm,
1536 gpa_t gpa, u64 size,
1540 struct kvm_memslots *slots;
1541 struct kvm_memory_slot *memslot;
1544 slots = kvm_memslots(kvm);
1546 /* we only care about the pages that the guest sees */
1547 kvm_for_each_memslot(memslot, slots) {
1548 unsigned long hva_start, hva_end;
1551 hva_start = max(start, memslot->userspace_addr);
1552 hva_end = min(end, memslot->userspace_addr +
1553 (memslot->npages << PAGE_SHIFT));
1554 if (hva_start >= hva_end)
1557 gpa = hva_to_gfn_memslot(hva_start, memslot) << PAGE_SHIFT;
1558 ret |= handler(kvm, gpa, (u64)(hva_end - hva_start), data);
1564 static int kvm_unmap_hva_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
1566 unmap_stage2_range(kvm, gpa, size);
1570 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
1572 unsigned long end = hva + PAGE_SIZE;
1577 trace_kvm_unmap_hva(hva);
1578 handle_hva_to_gpa(kvm, hva, end, &kvm_unmap_hva_handler, NULL);
1582 int kvm_unmap_hva_range(struct kvm *kvm,
1583 unsigned long start, unsigned long end)
1588 trace_kvm_unmap_hva_range(start, end);
1589 handle_hva_to_gpa(kvm, start, end, &kvm_unmap_hva_handler, NULL);
1593 static int kvm_set_spte_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
1595 pte_t *pte = (pte_t *)data;
1597 WARN_ON(size != PAGE_SIZE);
1599 * We can always call stage2_set_pte with KVM_S2PTE_FLAG_LOGGING_ACTIVE
1600 * flag clear because MMU notifiers will have unmapped a huge PMD before
1601 * calling ->change_pte() (which in turn calls kvm_set_spte_hva()) and
1602 * therefore stage2_set_pte() never needs to clear out a huge PMD
1603 * through this calling path.
1605 stage2_set_pte(kvm, NULL, gpa, pte, 0);
1610 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
1612 unsigned long end = hva + PAGE_SIZE;
1618 trace_kvm_set_spte_hva(hva);
1619 stage2_pte = pfn_pte(pte_pfn(pte), PAGE_S2);
1620 handle_hva_to_gpa(kvm, hva, end, &kvm_set_spte_handler, &stage2_pte);
1623 static int kvm_age_hva_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
1628 WARN_ON(size != PAGE_SIZE && size != PMD_SIZE);
1629 pmd = stage2_get_pmd(kvm, NULL, gpa);
1630 if (!pmd || pmd_none(*pmd)) /* Nothing there */
1633 if (pmd_thp_or_huge(*pmd)) /* THP, HugeTLB */
1634 return stage2_pmdp_test_and_clear_young(pmd);
1636 pte = pte_offset_kernel(pmd, gpa);
1640 return stage2_ptep_test_and_clear_young(pte);
1643 static int kvm_test_age_hva_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
1648 WARN_ON(size != PAGE_SIZE && size != PMD_SIZE);
1649 pmd = stage2_get_pmd(kvm, NULL, gpa);
1650 if (!pmd || pmd_none(*pmd)) /* Nothing there */
1653 if (pmd_thp_or_huge(*pmd)) /* THP, HugeTLB */
1654 return pmd_young(*pmd);
1656 pte = pte_offset_kernel(pmd, gpa);
1657 if (!pte_none(*pte)) /* Just a page... */
1658 return pte_young(*pte);
1663 int kvm_age_hva(struct kvm *kvm, unsigned long start, unsigned long end)
1665 trace_kvm_age_hva(start, end);
1666 return handle_hva_to_gpa(kvm, start, end, kvm_age_hva_handler, NULL);
1669 int kvm_test_age_hva(struct kvm *kvm, unsigned long hva)
1671 trace_kvm_test_age_hva(hva);
1672 return handle_hva_to_gpa(kvm, hva, hva, kvm_test_age_hva_handler, NULL);
1675 void kvm_mmu_free_memory_caches(struct kvm_vcpu *vcpu)
1677 mmu_free_memory_cache(&vcpu->arch.mmu_page_cache);
1680 phys_addr_t kvm_mmu_get_httbr(void)
1682 if (__kvm_cpu_uses_extended_idmap())
1683 return virt_to_phys(merged_hyp_pgd);
1685 return virt_to_phys(hyp_pgd);
1688 phys_addr_t kvm_get_idmap_vector(void)
1690 return hyp_idmap_vector;
1693 static int kvm_map_idmap_text(pgd_t *pgd)
1697 /* Create the idmap in the boot page tables */
1698 err = __create_hyp_mappings(pgd,
1699 hyp_idmap_start, hyp_idmap_end,
1700 __phys_to_pfn(hyp_idmap_start),
1703 kvm_err("Failed to idmap %lx-%lx\n",
1704 hyp_idmap_start, hyp_idmap_end);
1709 int kvm_mmu_init(void)
1713 hyp_idmap_start = kvm_virt_to_phys(__hyp_idmap_text_start);
1714 hyp_idmap_end = kvm_virt_to_phys(__hyp_idmap_text_end);
1715 hyp_idmap_vector = kvm_virt_to_phys(__kvm_hyp_init);
1718 * We rely on the linker script to ensure at build time that the HYP
1719 * init code does not cross a page boundary.
1721 BUG_ON((hyp_idmap_start ^ (hyp_idmap_end - 1)) & PAGE_MASK);
1723 kvm_info("IDMAP page: %lx\n", hyp_idmap_start);
1724 kvm_info("HYP VA range: %lx:%lx\n",
1725 kern_hyp_va(PAGE_OFFSET), kern_hyp_va(~0UL));
1727 if (hyp_idmap_start >= kern_hyp_va(PAGE_OFFSET) &&
1728 hyp_idmap_start < kern_hyp_va(~0UL) &&
1729 hyp_idmap_start != (unsigned long)__hyp_idmap_text_start) {
1731 * The idmap page is intersecting with the VA space,
1732 * it is not safe to continue further.
1734 kvm_err("IDMAP intersecting with HYP VA, unable to continue\n");
1739 hyp_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, hyp_pgd_order);
1741 kvm_err("Hyp mode PGD not allocated\n");
1746 if (__kvm_cpu_uses_extended_idmap()) {
1747 boot_hyp_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
1749 if (!boot_hyp_pgd) {
1750 kvm_err("Hyp boot PGD not allocated\n");
1755 err = kvm_map_idmap_text(boot_hyp_pgd);
1759 merged_hyp_pgd = (pgd_t *)__get_free_page(GFP_KERNEL | __GFP_ZERO);
1760 if (!merged_hyp_pgd) {
1761 kvm_err("Failed to allocate extra HYP pgd\n");
1764 __kvm_extend_hypmap(boot_hyp_pgd, hyp_pgd, merged_hyp_pgd,
1767 err = kvm_map_idmap_text(hyp_pgd);
1778 void kvm_arch_commit_memory_region(struct kvm *kvm,
1779 const struct kvm_userspace_memory_region *mem,
1780 const struct kvm_memory_slot *old,
1781 const struct kvm_memory_slot *new,
1782 enum kvm_mr_change change)
1785 * At this point memslot has been committed and there is an
1786 * allocated dirty_bitmap[], dirty pages will be be tracked while the
1787 * memory slot is write protected.
1789 if (change != KVM_MR_DELETE && mem->flags & KVM_MEM_LOG_DIRTY_PAGES)
1790 kvm_mmu_wp_memory_region(kvm, mem->slot);
1793 int kvm_arch_prepare_memory_region(struct kvm *kvm,
1794 struct kvm_memory_slot *memslot,
1795 const struct kvm_userspace_memory_region *mem,
1796 enum kvm_mr_change change)
1798 hva_t hva = mem->userspace_addr;
1799 hva_t reg_end = hva + mem->memory_size;
1800 bool writable = !(mem->flags & KVM_MEM_READONLY);
1803 if (change != KVM_MR_CREATE && change != KVM_MR_MOVE &&
1804 change != KVM_MR_FLAGS_ONLY)
1808 * Prevent userspace from creating a memory region outside of the IPA
1809 * space addressable by the KVM guest IPA space.
1811 if (memslot->base_gfn + memslot->npages >=
1812 (KVM_PHYS_SIZE >> PAGE_SHIFT))
1815 down_read(¤t->mm->mmap_sem);
1817 * A memory region could potentially cover multiple VMAs, and any holes
1818 * between them, so iterate over all of them to find out if we can map
1819 * any of them right now.
1821 * +--------------------------------------------+
1822 * +---------------+----------------+ +----------------+
1823 * | : VMA 1 | VMA 2 | | VMA 3 : |
1824 * +---------------+----------------+ +----------------+
1826 * +--------------------------------------------+
1829 struct vm_area_struct *vma = find_vma(current->mm, hva);
1830 hva_t vm_start, vm_end;
1832 if (!vma || vma->vm_start >= reg_end)
1836 * Mapping a read-only VMA is only allowed if the
1837 * memory region is configured as read-only.
1839 if (writable && !(vma->vm_flags & VM_WRITE)) {
1845 * Take the intersection of this VMA with the memory region
1847 vm_start = max(hva, vma->vm_start);
1848 vm_end = min(reg_end, vma->vm_end);
1850 if (vma->vm_flags & VM_PFNMAP) {
1851 gpa_t gpa = mem->guest_phys_addr +
1852 (vm_start - mem->userspace_addr);
1855 pa = (phys_addr_t)vma->vm_pgoff << PAGE_SHIFT;
1856 pa += vm_start - vma->vm_start;
1858 /* IO region dirty page logging not allowed */
1859 if (memslot->flags & KVM_MEM_LOG_DIRTY_PAGES) {
1864 ret = kvm_phys_addr_ioremap(kvm, gpa, pa,
1871 } while (hva < reg_end);
1873 if (change == KVM_MR_FLAGS_ONLY)
1876 spin_lock(&kvm->mmu_lock);
1878 unmap_stage2_range(kvm, mem->guest_phys_addr, mem->memory_size);
1880 stage2_flush_memslot(kvm, memslot);
1881 spin_unlock(&kvm->mmu_lock);
1883 up_read(¤t->mm->mmap_sem);
1887 void kvm_arch_free_memslot(struct kvm *kvm, struct kvm_memory_slot *free,
1888 struct kvm_memory_slot *dont)
1892 int kvm_arch_create_memslot(struct kvm *kvm, struct kvm_memory_slot *slot,
1893 unsigned long npages)
1898 void kvm_arch_memslots_updated(struct kvm *kvm, struct kvm_memslots *slots)
1902 void kvm_arch_flush_shadow_all(struct kvm *kvm)
1904 kvm_free_stage2_pgd(kvm);
1907 void kvm_arch_flush_shadow_memslot(struct kvm *kvm,
1908 struct kvm_memory_slot *slot)
1910 gpa_t gpa = slot->base_gfn << PAGE_SHIFT;
1911 phys_addr_t size = slot->npages << PAGE_SHIFT;
1913 spin_lock(&kvm->mmu_lock);
1914 unmap_stage2_range(kvm, gpa, size);
1915 spin_unlock(&kvm->mmu_lock);
1919 * See note at ARMv7 ARM B1.14.4 (TL;DR: S/W ops are not easily virtualized).
1922 * - S/W ops are local to a CPU (not broadcast)
1923 * - We have line migration behind our back (speculation)
1924 * - System caches don't support S/W at all (damn!)
1926 * In the face of the above, the best we can do is to try and convert
1927 * S/W ops to VA ops. Because the guest is not allowed to infer the
1928 * S/W to PA mapping, it can only use S/W to nuke the whole cache,
1929 * which is a rather good thing for us.
1931 * Also, it is only used when turning caches on/off ("The expected
1932 * usage of the cache maintenance instructions that operate by set/way
1933 * is associated with the cache maintenance instructions associated
1934 * with the powerdown and powerup of caches, if this is required by
1935 * the implementation.").
1937 * We use the following policy:
1939 * - If we trap a S/W operation, we enable VM trapping to detect
1940 * caches being turned on/off, and do a full clean.
1942 * - We flush the caches on both caches being turned on and off.
1944 * - Once the caches are enabled, we stop trapping VM ops.
1946 void kvm_set_way_flush(struct kvm_vcpu *vcpu)
1948 unsigned long hcr = vcpu_get_hcr(vcpu);
1951 * If this is the first time we do a S/W operation
1952 * (i.e. HCR_TVM not set) flush the whole memory, and set the
1955 * Otherwise, rely on the VM trapping to wait for the MMU +
1956 * Caches to be turned off. At that point, we'll be able to
1957 * clean the caches again.
1959 if (!(hcr & HCR_TVM)) {
1960 trace_kvm_set_way_flush(*vcpu_pc(vcpu),
1961 vcpu_has_cache_enabled(vcpu));
1962 stage2_flush_vm(vcpu->kvm);
1963 vcpu_set_hcr(vcpu, hcr | HCR_TVM);
1967 void kvm_toggle_cache(struct kvm_vcpu *vcpu, bool was_enabled)
1969 bool now_enabled = vcpu_has_cache_enabled(vcpu);
1972 * If switching the MMU+caches on, need to invalidate the caches.
1973 * If switching it off, need to clean the caches.
1974 * Clean + invalidate does the trick always.
1976 if (now_enabled != was_enabled)
1977 stage2_flush_vm(vcpu->kvm);
1979 /* Caches are now on, stop trapping VM ops (until a S/W op) */
1981 vcpu_set_hcr(vcpu, vcpu_get_hcr(vcpu) & ~HCR_TVM);
1983 trace_kvm_toggle_cache(*vcpu_pc(vcpu), was_enabled, now_enabled);