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 extern char __hyp_idmap_text_start[], __hyp_idmap_text_end[];
37 static pgd_t *boot_hyp_pgd;
38 static pgd_t *hyp_pgd;
39 static pgd_t *merged_hyp_pgd;
40 static DEFINE_MUTEX(kvm_hyp_pgd_mutex);
42 static unsigned long hyp_idmap_start;
43 static unsigned long hyp_idmap_end;
44 static phys_addr_t hyp_idmap_vector;
46 #define hyp_pgd_order get_order(PTRS_PER_PGD * sizeof(pgd_t))
48 #define kvm_pmd_huge(_x) (pmd_huge(_x) || pmd_trans_huge(_x))
49 #define kvm_pud_huge(_x) pud_huge(_x)
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)
73 * This function also gets called when dealing with HYP page
74 * tables. As HYP doesn't have an associated struct kvm (and
75 * the HYP page tables are fairly static), we don't do
79 kvm_call_hyp(__kvm_tlb_flush_vmid_ipa, kvm, ipa);
83 * D-Cache management functions. They take the page table entries by
84 * value, as they are flushing the cache using the kernel mapping (or
87 static void kvm_flush_dcache_pte(pte_t pte)
89 __kvm_flush_dcache_pte(pte);
92 static void kvm_flush_dcache_pmd(pmd_t pmd)
94 __kvm_flush_dcache_pmd(pmd);
97 static void kvm_flush_dcache_pud(pud_t pud)
99 __kvm_flush_dcache_pud(pud);
102 static bool kvm_is_device_pfn(unsigned long pfn)
104 return !pfn_valid(pfn);
108 * stage2_dissolve_pmd() - clear and flush huge PMD entry
109 * @kvm: pointer to kvm structure.
111 * @pmd: pmd pointer for IPA
113 * Function clears a PMD entry, flushes addr 1st and 2nd stage TLBs. Marks all
114 * pages in the range dirty.
116 static void stage2_dissolve_pmd(struct kvm *kvm, phys_addr_t addr, pmd_t *pmd)
118 if (!kvm_pmd_huge(*pmd))
122 kvm_tlb_flush_vmid_ipa(kvm, addr);
123 put_page(virt_to_page(pmd));
126 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
131 BUG_ON(max > KVM_NR_MEM_OBJS);
132 if (cache->nobjs >= min)
134 while (cache->nobjs < max) {
135 page = (void *)__get_free_page(PGALLOC_GFP);
138 cache->objects[cache->nobjs++] = page;
143 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
146 free_page((unsigned long)mc->objects[--mc->nobjs]);
149 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc)
153 BUG_ON(!mc || !mc->nobjs);
154 p = mc->objects[--mc->nobjs];
158 static void clear_pgd_entry(struct kvm *kvm, pgd_t *pgd, phys_addr_t addr)
160 pud_t *pud_table __maybe_unused = pud_offset(pgd, 0);
162 kvm_tlb_flush_vmid_ipa(kvm, addr);
163 pud_free(NULL, pud_table);
164 put_page(virt_to_page(pgd));
167 static void clear_pud_entry(struct kvm *kvm, pud_t *pud, phys_addr_t addr)
169 pmd_t *pmd_table = pmd_offset(pud, 0);
170 VM_BUG_ON(pud_huge(*pud));
172 kvm_tlb_flush_vmid_ipa(kvm, addr);
173 pmd_free(NULL, pmd_table);
174 put_page(virt_to_page(pud));
177 static void clear_pmd_entry(struct kvm *kvm, pmd_t *pmd, phys_addr_t addr)
179 pte_t *pte_table = pte_offset_kernel(pmd, 0);
180 VM_BUG_ON(kvm_pmd_huge(*pmd));
182 kvm_tlb_flush_vmid_ipa(kvm, addr);
183 pte_free_kernel(NULL, pte_table);
184 put_page(virt_to_page(pmd));
188 * Unmapping vs dcache management:
190 * If a guest maps certain memory pages as uncached, all writes will
191 * bypass the data cache and go directly to RAM. However, the CPUs
192 * can still speculate reads (not writes) and fill cache lines with
195 * Those cache lines will be *clean* cache lines though, so a
196 * clean+invalidate operation is equivalent to an invalidate
197 * operation, because no cache lines are marked dirty.
199 * Those clean cache lines could be filled prior to an uncached write
200 * by the guest, and the cache coherent IO subsystem would therefore
201 * end up writing old data to disk.
203 * This is why right after unmapping a page/section and invalidating
204 * the corresponding TLBs, we call kvm_flush_dcache_p*() to make sure
205 * the IO subsystem will never hit in the cache.
207 static void unmap_ptes(struct kvm *kvm, pmd_t *pmd,
208 phys_addr_t addr, phys_addr_t end)
210 phys_addr_t start_addr = addr;
211 pte_t *pte, *start_pte;
213 start_pte = pte = pte_offset_kernel(pmd, addr);
215 if (!pte_none(*pte)) {
216 pte_t old_pte = *pte;
218 kvm_set_pte(pte, __pte(0));
219 kvm_tlb_flush_vmid_ipa(kvm, addr);
221 /* No need to invalidate the cache for device mappings */
222 if (!kvm_is_device_pfn(pte_pfn(old_pte)))
223 kvm_flush_dcache_pte(old_pte);
225 put_page(virt_to_page(pte));
227 } while (pte++, addr += PAGE_SIZE, addr != end);
229 if (kvm_pte_table_empty(kvm, start_pte))
230 clear_pmd_entry(kvm, pmd, start_addr);
233 static void unmap_pmds(struct kvm *kvm, pud_t *pud,
234 phys_addr_t addr, phys_addr_t end)
236 phys_addr_t next, start_addr = addr;
237 pmd_t *pmd, *start_pmd;
239 start_pmd = pmd = pmd_offset(pud, addr);
241 next = kvm_pmd_addr_end(addr, end);
242 if (!pmd_none(*pmd)) {
243 if (kvm_pmd_huge(*pmd)) {
244 pmd_t old_pmd = *pmd;
247 kvm_tlb_flush_vmid_ipa(kvm, addr);
249 kvm_flush_dcache_pmd(old_pmd);
251 put_page(virt_to_page(pmd));
253 unmap_ptes(kvm, pmd, addr, next);
256 } while (pmd++, addr = next, addr != end);
258 if (kvm_pmd_table_empty(kvm, start_pmd))
259 clear_pud_entry(kvm, pud, start_addr);
262 static void unmap_puds(struct kvm *kvm, pgd_t *pgd,
263 phys_addr_t addr, phys_addr_t end)
265 phys_addr_t next, start_addr = addr;
266 pud_t *pud, *start_pud;
268 start_pud = pud = pud_offset(pgd, addr);
270 next = kvm_pud_addr_end(addr, end);
271 if (!pud_none(*pud)) {
272 if (pud_huge(*pud)) {
273 pud_t old_pud = *pud;
276 kvm_tlb_flush_vmid_ipa(kvm, addr);
278 kvm_flush_dcache_pud(old_pud);
280 put_page(virt_to_page(pud));
282 unmap_pmds(kvm, pud, addr, next);
285 } while (pud++, addr = next, addr != end);
287 if (kvm_pud_table_empty(kvm, start_pud))
288 clear_pgd_entry(kvm, pgd, start_addr);
292 static void unmap_range(struct kvm *kvm, pgd_t *pgdp,
293 phys_addr_t start, u64 size)
296 phys_addr_t addr = start, end = start + size;
299 pgd = pgdp + kvm_pgd_index(addr);
301 next = kvm_pgd_addr_end(addr, end);
303 unmap_puds(kvm, pgd, addr, next);
304 } while (pgd++, addr = next, addr != end);
307 static void stage2_flush_ptes(struct kvm *kvm, pmd_t *pmd,
308 phys_addr_t addr, phys_addr_t end)
312 pte = pte_offset_kernel(pmd, addr);
314 if (!pte_none(*pte) && !kvm_is_device_pfn(pte_pfn(*pte)))
315 kvm_flush_dcache_pte(*pte);
316 } while (pte++, addr += PAGE_SIZE, addr != end);
319 static void stage2_flush_pmds(struct kvm *kvm, pud_t *pud,
320 phys_addr_t addr, phys_addr_t end)
325 pmd = pmd_offset(pud, addr);
327 next = kvm_pmd_addr_end(addr, end);
328 if (!pmd_none(*pmd)) {
329 if (kvm_pmd_huge(*pmd))
330 kvm_flush_dcache_pmd(*pmd);
332 stage2_flush_ptes(kvm, pmd, addr, next);
334 } while (pmd++, addr = next, addr != end);
337 static void stage2_flush_puds(struct kvm *kvm, pgd_t *pgd,
338 phys_addr_t addr, phys_addr_t end)
343 pud = pud_offset(pgd, addr);
345 next = kvm_pud_addr_end(addr, end);
346 if (!pud_none(*pud)) {
348 kvm_flush_dcache_pud(*pud);
350 stage2_flush_pmds(kvm, pud, addr, next);
352 } while (pud++, addr = next, addr != end);
355 static void stage2_flush_memslot(struct kvm *kvm,
356 struct kvm_memory_slot *memslot)
358 phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
359 phys_addr_t end = addr + PAGE_SIZE * memslot->npages;
363 pgd = kvm->arch.pgd + kvm_pgd_index(addr);
365 next = kvm_pgd_addr_end(addr, end);
366 stage2_flush_puds(kvm, pgd, addr, next);
367 } while (pgd++, addr = next, addr != end);
371 * stage2_flush_vm - Invalidate cache for pages mapped in stage 2
372 * @kvm: The struct kvm pointer
374 * Go through the stage 2 page tables and invalidate any cache lines
375 * backing memory already mapped to the VM.
377 static void stage2_flush_vm(struct kvm *kvm)
379 struct kvm_memslots *slots;
380 struct kvm_memory_slot *memslot;
383 idx = srcu_read_lock(&kvm->srcu);
384 spin_lock(&kvm->mmu_lock);
386 slots = kvm_memslots(kvm);
387 kvm_for_each_memslot(memslot, slots)
388 stage2_flush_memslot(kvm, memslot);
390 spin_unlock(&kvm->mmu_lock);
391 srcu_read_unlock(&kvm->srcu, idx);
395 * free_boot_hyp_pgd - free HYP boot page tables
397 * Free the HYP boot page tables. The bounce page is also freed.
399 void free_boot_hyp_pgd(void)
401 mutex_lock(&kvm_hyp_pgd_mutex);
404 unmap_range(NULL, boot_hyp_pgd, hyp_idmap_start, PAGE_SIZE);
405 unmap_range(NULL, boot_hyp_pgd, TRAMPOLINE_VA, PAGE_SIZE);
406 free_pages((unsigned long)boot_hyp_pgd, hyp_pgd_order);
411 unmap_range(NULL, hyp_pgd, TRAMPOLINE_VA, PAGE_SIZE);
413 mutex_unlock(&kvm_hyp_pgd_mutex);
417 * free_hyp_pgds - free Hyp-mode page tables
419 * Assumes hyp_pgd is a page table used strictly in Hyp-mode and
420 * therefore contains either mappings in the kernel memory area (above
421 * PAGE_OFFSET), or device mappings in the vmalloc range (from
422 * VMALLOC_START to VMALLOC_END).
424 * boot_hyp_pgd should only map two pages for the init code.
426 void free_hyp_pgds(void)
432 mutex_lock(&kvm_hyp_pgd_mutex);
435 for (addr = PAGE_OFFSET; virt_addr_valid(addr); addr += PGDIR_SIZE)
436 unmap_range(NULL, hyp_pgd, KERN_TO_HYP(addr), PGDIR_SIZE);
437 for (addr = VMALLOC_START; is_vmalloc_addr((void*)addr); addr += PGDIR_SIZE)
438 unmap_range(NULL, hyp_pgd, KERN_TO_HYP(addr), PGDIR_SIZE);
440 free_pages((unsigned long)hyp_pgd, hyp_pgd_order);
443 if (merged_hyp_pgd) {
444 clear_page(merged_hyp_pgd);
445 free_page((unsigned long)merged_hyp_pgd);
446 merged_hyp_pgd = NULL;
449 mutex_unlock(&kvm_hyp_pgd_mutex);
452 static void create_hyp_pte_mappings(pmd_t *pmd, unsigned long start,
453 unsigned long end, unsigned long pfn,
461 pte = pte_offset_kernel(pmd, addr);
462 kvm_set_pte(pte, pfn_pte(pfn, prot));
463 get_page(virt_to_page(pte));
464 kvm_flush_dcache_to_poc(pte, sizeof(*pte));
466 } while (addr += PAGE_SIZE, addr != end);
469 static int create_hyp_pmd_mappings(pud_t *pud, unsigned long start,
470 unsigned long end, unsigned long pfn,
475 unsigned long addr, next;
479 pmd = pmd_offset(pud, addr);
481 BUG_ON(pmd_sect(*pmd));
483 if (pmd_none(*pmd)) {
484 pte = pte_alloc_one_kernel(NULL, addr);
486 kvm_err("Cannot allocate Hyp pte\n");
489 pmd_populate_kernel(NULL, pmd, pte);
490 get_page(virt_to_page(pmd));
491 kvm_flush_dcache_to_poc(pmd, sizeof(*pmd));
494 next = pmd_addr_end(addr, end);
496 create_hyp_pte_mappings(pmd, addr, next, pfn, prot);
497 pfn += (next - addr) >> PAGE_SHIFT;
498 } while (addr = next, addr != end);
503 static int create_hyp_pud_mappings(pgd_t *pgd, unsigned long start,
504 unsigned long end, unsigned long pfn,
509 unsigned long addr, next;
514 pud = pud_offset(pgd, addr);
516 if (pud_none_or_clear_bad(pud)) {
517 pmd = pmd_alloc_one(NULL, addr);
519 kvm_err("Cannot allocate Hyp pmd\n");
522 pud_populate(NULL, pud, pmd);
523 get_page(virt_to_page(pud));
524 kvm_flush_dcache_to_poc(pud, sizeof(*pud));
527 next = pud_addr_end(addr, end);
528 ret = create_hyp_pmd_mappings(pud, addr, next, pfn, prot);
531 pfn += (next - addr) >> PAGE_SHIFT;
532 } while (addr = next, addr != end);
537 static int __create_hyp_mappings(pgd_t *pgdp,
538 unsigned long start, unsigned long end,
539 unsigned long pfn, pgprot_t prot)
543 unsigned long addr, next;
546 mutex_lock(&kvm_hyp_pgd_mutex);
547 addr = start & PAGE_MASK;
548 end = PAGE_ALIGN(end);
550 pgd = pgdp + pgd_index(addr);
552 if (pgd_none(*pgd)) {
553 pud = pud_alloc_one(NULL, addr);
555 kvm_err("Cannot allocate Hyp pud\n");
559 pgd_populate(NULL, pgd, pud);
560 get_page(virt_to_page(pgd));
561 kvm_flush_dcache_to_poc(pgd, sizeof(*pgd));
564 next = pgd_addr_end(addr, end);
565 err = create_hyp_pud_mappings(pgd, addr, next, pfn, prot);
568 pfn += (next - addr) >> PAGE_SHIFT;
569 } while (addr = next, addr != end);
571 mutex_unlock(&kvm_hyp_pgd_mutex);
575 static phys_addr_t kvm_kaddr_to_phys(void *kaddr)
577 if (!is_vmalloc_addr(kaddr)) {
578 BUG_ON(!virt_addr_valid(kaddr));
581 return page_to_phys(vmalloc_to_page(kaddr)) +
582 offset_in_page(kaddr);
587 * create_hyp_mappings - duplicate a kernel virtual address range in Hyp mode
588 * @from: The virtual kernel start address of the range
589 * @to: The virtual kernel end address of the range (exclusive)
591 * The same virtual address as the kernel virtual address is also used
592 * in Hyp-mode mapping (modulo HYP_PAGE_OFFSET) to the same underlying
595 int create_hyp_mappings(void *from, void *to)
597 phys_addr_t phys_addr;
598 unsigned long virt_addr;
599 unsigned long start = KERN_TO_HYP((unsigned long)from);
600 unsigned long end = KERN_TO_HYP((unsigned long)to);
602 if (is_kernel_in_hyp_mode())
605 start = start & PAGE_MASK;
606 end = PAGE_ALIGN(end);
608 for (virt_addr = start; virt_addr < end; virt_addr += PAGE_SIZE) {
611 phys_addr = kvm_kaddr_to_phys(from + virt_addr - start);
612 err = __create_hyp_mappings(hyp_pgd, virt_addr,
613 virt_addr + PAGE_SIZE,
614 __phys_to_pfn(phys_addr),
624 * create_hyp_io_mappings - duplicate a kernel IO mapping into Hyp mode
625 * @from: The kernel start VA of the range
626 * @to: The kernel end VA of the range (exclusive)
627 * @phys_addr: The physical start address which gets mapped
629 * The resulting HYP VA is the same as the kernel VA, modulo
632 int create_hyp_io_mappings(void *from, void *to, phys_addr_t phys_addr)
634 unsigned long start = KERN_TO_HYP((unsigned long)from);
635 unsigned long end = KERN_TO_HYP((unsigned long)to);
637 if (is_kernel_in_hyp_mode())
640 /* Check for a valid kernel IO mapping */
641 if (!is_vmalloc_addr(from) || !is_vmalloc_addr(to - 1))
644 return __create_hyp_mappings(hyp_pgd, start, end,
645 __phys_to_pfn(phys_addr), PAGE_HYP_DEVICE);
648 /* Free the HW pgd, one page at a time */
649 static void kvm_free_hwpgd(void *hwpgd)
651 free_pages_exact(hwpgd, kvm_get_hwpgd_size());
654 /* Allocate the HW PGD, making sure that each page gets its own refcount */
655 static void *kvm_alloc_hwpgd(void)
657 unsigned int size = kvm_get_hwpgd_size();
659 return alloc_pages_exact(size, GFP_KERNEL | __GFP_ZERO);
663 * kvm_alloc_stage2_pgd - allocate level-1 table for stage-2 translation.
664 * @kvm: The KVM struct pointer for the VM.
666 * Allocates only the stage-2 HW PGD level table(s) (can support either full
667 * 40-bit input addresses or limited to 32-bit input addresses). Clears the
670 * Note we don't need locking here as this is only called when the VM is
671 * created, which can only be done once.
673 int kvm_alloc_stage2_pgd(struct kvm *kvm)
678 if (kvm->arch.pgd != NULL) {
679 kvm_err("kvm_arch already initialized?\n");
683 hwpgd = kvm_alloc_hwpgd();
687 /* When the kernel uses more levels of page tables than the
688 * guest, we allocate a fake PGD and pre-populate it to point
689 * to the next-level page table, which will be the real
690 * initial page table pointed to by the VTTBR.
692 * When KVM_PREALLOC_LEVEL==2, we allocate a single page for
693 * the PMD and the kernel will use folded pud.
694 * When KVM_PREALLOC_LEVEL==1, we allocate 2 consecutive PUD
697 if (KVM_PREALLOC_LEVEL > 0) {
701 * Allocate fake pgd for the page table manipulation macros to
702 * work. This is not used by the hardware and we have no
703 * alignment requirement for this allocation.
705 pgd = kmalloc(PTRS_PER_S2_PGD * sizeof(pgd_t),
706 GFP_KERNEL | __GFP_ZERO);
709 kvm_free_hwpgd(hwpgd);
713 /* Plug the HW PGD into the fake one. */
714 for (i = 0; i < PTRS_PER_S2_PGD; i++) {
715 if (KVM_PREALLOC_LEVEL == 1)
716 pgd_populate(NULL, pgd + i,
717 (pud_t *)hwpgd + i * PTRS_PER_PUD);
718 else if (KVM_PREALLOC_LEVEL == 2)
719 pud_populate(NULL, pud_offset(pgd, 0) + i,
720 (pmd_t *)hwpgd + i * PTRS_PER_PMD);
724 * Allocate actual first-level Stage-2 page table used by the
725 * hardware for Stage-2 page table walks.
727 pgd = (pgd_t *)hwpgd;
736 * unmap_stage2_range -- Clear stage2 page table entries to unmap a range
737 * @kvm: The VM pointer
738 * @start: The intermediate physical base address of the range to unmap
739 * @size: The size of the area to unmap
741 * Clear a range of stage-2 mappings, lowering the various ref-counts. Must
742 * be called while holding mmu_lock (unless for freeing the stage2 pgd before
743 * destroying the VM), otherwise another faulting VCPU may come in and mess
744 * with things behind our backs.
746 static void unmap_stage2_range(struct kvm *kvm, phys_addr_t start, u64 size)
748 unmap_range(kvm, kvm->arch.pgd, start, size);
751 static void stage2_unmap_memslot(struct kvm *kvm,
752 struct kvm_memory_slot *memslot)
754 hva_t hva = memslot->userspace_addr;
755 phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
756 phys_addr_t size = PAGE_SIZE * memslot->npages;
757 hva_t reg_end = hva + size;
760 * A memory region could potentially cover multiple VMAs, and any holes
761 * between them, so iterate over all of them to find out if we should
764 * +--------------------------------------------+
765 * +---------------+----------------+ +----------------+
766 * | : VMA 1 | VMA 2 | | VMA 3 : |
767 * +---------------+----------------+ +----------------+
769 * +--------------------------------------------+
772 struct vm_area_struct *vma = find_vma(current->mm, hva);
773 hva_t vm_start, vm_end;
775 if (!vma || vma->vm_start >= reg_end)
779 * Take the intersection of this VMA with the memory region
781 vm_start = max(hva, vma->vm_start);
782 vm_end = min(reg_end, vma->vm_end);
784 if (!(vma->vm_flags & VM_PFNMAP)) {
785 gpa_t gpa = addr + (vm_start - memslot->userspace_addr);
786 unmap_stage2_range(kvm, gpa, vm_end - vm_start);
789 } while (hva < reg_end);
793 * stage2_unmap_vm - Unmap Stage-2 RAM mappings
794 * @kvm: The struct kvm pointer
796 * Go through the memregions and unmap any reguler RAM
797 * backing memory already mapped to the VM.
799 void stage2_unmap_vm(struct kvm *kvm)
801 struct kvm_memslots *slots;
802 struct kvm_memory_slot *memslot;
805 idx = srcu_read_lock(&kvm->srcu);
806 spin_lock(&kvm->mmu_lock);
808 slots = kvm_memslots(kvm);
809 kvm_for_each_memslot(memslot, slots)
810 stage2_unmap_memslot(kvm, memslot);
812 spin_unlock(&kvm->mmu_lock);
813 srcu_read_unlock(&kvm->srcu, idx);
817 * kvm_free_stage2_pgd - free all stage-2 tables
818 * @kvm: The KVM struct pointer for the VM.
820 * Walks the level-1 page table pointed to by kvm->arch.pgd and frees all
821 * underlying level-2 and level-3 tables before freeing the actual level-1 table
822 * and setting the struct pointer to NULL.
824 * Note we don't need locking here as this is only called when the VM is
825 * destroyed, which can only be done once.
827 void kvm_free_stage2_pgd(struct kvm *kvm)
829 if (kvm->arch.pgd == NULL)
832 unmap_stage2_range(kvm, 0, KVM_PHYS_SIZE);
833 kvm_free_hwpgd(kvm_get_hwpgd(kvm));
834 if (KVM_PREALLOC_LEVEL > 0)
835 kfree(kvm->arch.pgd);
837 kvm->arch.pgd = NULL;
840 static pud_t *stage2_get_pud(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
846 pgd = kvm->arch.pgd + kvm_pgd_index(addr);
847 if (WARN_ON(pgd_none(*pgd))) {
850 pud = mmu_memory_cache_alloc(cache);
851 pgd_populate(NULL, pgd, pud);
852 get_page(virt_to_page(pgd));
855 return pud_offset(pgd, addr);
858 static pmd_t *stage2_get_pmd(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
864 pud = stage2_get_pud(kvm, cache, addr);
865 if (pud_none(*pud)) {
868 pmd = mmu_memory_cache_alloc(cache);
869 pud_populate(NULL, pud, pmd);
870 get_page(virt_to_page(pud));
873 return pmd_offset(pud, addr);
876 static int stage2_set_pmd_huge(struct kvm *kvm, struct kvm_mmu_memory_cache
877 *cache, phys_addr_t addr, const pmd_t *new_pmd)
881 pmd = stage2_get_pmd(kvm, cache, addr);
885 * Mapping in huge pages should only happen through a fault. If a
886 * page is merged into a transparent huge page, the individual
887 * subpages of that huge page should be unmapped through MMU
888 * notifiers before we get here.
890 * Merging of CompoundPages is not supported; they should become
891 * splitting first, unmapped, merged, and mapped back in on-demand.
893 VM_BUG_ON(pmd_present(*pmd) && pmd_pfn(*pmd) != pmd_pfn(*new_pmd));
896 if (pmd_present(old_pmd)) {
898 kvm_tlb_flush_vmid_ipa(kvm, addr);
900 get_page(virt_to_page(pmd));
903 kvm_set_pmd(pmd, *new_pmd);
907 static int stage2_set_pte(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
908 phys_addr_t addr, const pte_t *new_pte,
913 bool iomap = flags & KVM_S2PTE_FLAG_IS_IOMAP;
914 bool logging_active = flags & KVM_S2_FLAG_LOGGING_ACTIVE;
916 VM_BUG_ON(logging_active && !cache);
918 /* Create stage-2 page table mapping - Levels 0 and 1 */
919 pmd = stage2_get_pmd(kvm, cache, addr);
922 * Ignore calls from kvm_set_spte_hva for unallocated
929 * While dirty page logging - dissolve huge PMD, then continue on to
933 stage2_dissolve_pmd(kvm, addr, pmd);
935 /* Create stage-2 page mappings - Level 2 */
936 if (pmd_none(*pmd)) {
938 return 0; /* ignore calls from kvm_set_spte_hva */
939 pte = mmu_memory_cache_alloc(cache);
941 pmd_populate_kernel(NULL, pmd, pte);
942 get_page(virt_to_page(pmd));
945 pte = pte_offset_kernel(pmd, addr);
947 if (iomap && pte_present(*pte))
950 /* Create 2nd stage page table mapping - Level 3 */
952 if (pte_present(old_pte)) {
953 kvm_set_pte(pte, __pte(0));
954 kvm_tlb_flush_vmid_ipa(kvm, addr);
956 get_page(virt_to_page(pte));
959 kvm_set_pte(pte, *new_pte);
964 * kvm_phys_addr_ioremap - map a device range to guest IPA
966 * @kvm: The KVM pointer
967 * @guest_ipa: The IPA at which to insert the mapping
968 * @pa: The physical address of the device
969 * @size: The size of the mapping
971 int kvm_phys_addr_ioremap(struct kvm *kvm, phys_addr_t guest_ipa,
972 phys_addr_t pa, unsigned long size, bool writable)
974 phys_addr_t addr, end;
977 struct kvm_mmu_memory_cache cache = { 0, };
979 end = (guest_ipa + size + PAGE_SIZE - 1) & PAGE_MASK;
980 pfn = __phys_to_pfn(pa);
982 for (addr = guest_ipa; addr < end; addr += PAGE_SIZE) {
983 pte_t pte = pfn_pte(pfn, PAGE_S2_DEVICE);
986 kvm_set_s2pte_writable(&pte);
988 ret = mmu_topup_memory_cache(&cache, KVM_MMU_CACHE_MIN_PAGES,
992 spin_lock(&kvm->mmu_lock);
993 ret = stage2_set_pte(kvm, &cache, addr, &pte,
994 KVM_S2PTE_FLAG_IS_IOMAP);
995 spin_unlock(&kvm->mmu_lock);
1003 mmu_free_memory_cache(&cache);
1007 static bool transparent_hugepage_adjust(pfn_t *pfnp, phys_addr_t *ipap)
1010 gfn_t gfn = *ipap >> PAGE_SHIFT;
1012 if (PageTransCompound(pfn_to_page(pfn))) {
1015 * The address we faulted on is backed by a transparent huge
1016 * page. However, because we map the compound huge page and
1017 * not the individual tail page, we need to transfer the
1018 * refcount to the head page. We have to be careful that the
1019 * THP doesn't start to split while we are adjusting the
1022 * We are sure this doesn't happen, because mmu_notifier_retry
1023 * was successful and we are holding the mmu_lock, so if this
1024 * THP is trying to split, it will be blocked in the mmu
1025 * notifier before touching any of the pages, specifically
1026 * before being able to call __split_huge_page_refcount().
1028 * We can therefore safely transfer the refcount from PG_tail
1029 * to PG_head and switch the pfn from a tail page to the head
1032 mask = PTRS_PER_PMD - 1;
1033 VM_BUG_ON((gfn & mask) != (pfn & mask));
1036 kvm_release_pfn_clean(pfn);
1048 static bool kvm_is_write_fault(struct kvm_vcpu *vcpu)
1050 if (kvm_vcpu_trap_is_iabt(vcpu))
1053 return kvm_vcpu_dabt_iswrite(vcpu);
1057 * stage2_wp_ptes - write protect PMD range
1058 * @pmd: pointer to pmd entry
1059 * @addr: range start address
1060 * @end: range end address
1062 static void stage2_wp_ptes(pmd_t *pmd, phys_addr_t addr, phys_addr_t end)
1066 pte = pte_offset_kernel(pmd, addr);
1068 if (!pte_none(*pte)) {
1069 if (!kvm_s2pte_readonly(pte))
1070 kvm_set_s2pte_readonly(pte);
1072 } while (pte++, addr += PAGE_SIZE, addr != end);
1076 * stage2_wp_pmds - write protect PUD range
1077 * @pud: pointer to pud entry
1078 * @addr: range start address
1079 * @end: range end address
1081 static void stage2_wp_pmds(pud_t *pud, phys_addr_t addr, phys_addr_t end)
1086 pmd = pmd_offset(pud, addr);
1089 next = kvm_pmd_addr_end(addr, end);
1090 if (!pmd_none(*pmd)) {
1091 if (kvm_pmd_huge(*pmd)) {
1092 if (!kvm_s2pmd_readonly(pmd))
1093 kvm_set_s2pmd_readonly(pmd);
1095 stage2_wp_ptes(pmd, addr, next);
1098 } while (pmd++, addr = next, addr != end);
1102 * stage2_wp_puds - write protect PGD range
1103 * @pgd: pointer to pgd entry
1104 * @addr: range start address
1105 * @end: range end address
1107 * Process PUD entries, for a huge PUD we cause a panic.
1109 static void stage2_wp_puds(pgd_t *pgd, phys_addr_t addr, phys_addr_t end)
1114 pud = pud_offset(pgd, addr);
1116 next = kvm_pud_addr_end(addr, end);
1117 if (!pud_none(*pud)) {
1118 /* TODO:PUD not supported, revisit later if supported */
1119 BUG_ON(kvm_pud_huge(*pud));
1120 stage2_wp_pmds(pud, addr, next);
1122 } while (pud++, addr = next, addr != end);
1126 * stage2_wp_range() - write protect stage2 memory region range
1127 * @kvm: The KVM pointer
1128 * @addr: Start address of range
1129 * @end: End address of range
1131 static void stage2_wp_range(struct kvm *kvm, phys_addr_t addr, phys_addr_t end)
1136 pgd = kvm->arch.pgd + kvm_pgd_index(addr);
1139 * Release kvm_mmu_lock periodically if the memory region is
1140 * large. Otherwise, we may see kernel panics with
1141 * CONFIG_DETECT_HUNG_TASK, CONFIG_LOCKUP_DETECTOR,
1142 * CONFIG_LOCKDEP. Additionally, holding the lock too long
1143 * will also starve other vCPUs.
1145 if (need_resched() || spin_needbreak(&kvm->mmu_lock))
1146 cond_resched_lock(&kvm->mmu_lock);
1148 next = kvm_pgd_addr_end(addr, end);
1149 if (pgd_present(*pgd))
1150 stage2_wp_puds(pgd, addr, next);
1151 } while (pgd++, addr = next, addr != end);
1155 * kvm_mmu_wp_memory_region() - write protect stage 2 entries for memory slot
1156 * @kvm: The KVM pointer
1157 * @slot: The memory slot to write protect
1159 * Called to start logging dirty pages after memory region
1160 * KVM_MEM_LOG_DIRTY_PAGES operation is called. After this function returns
1161 * all present PMD and PTEs are write protected in the memory region.
1162 * Afterwards read of dirty page log can be called.
1164 * Acquires kvm_mmu_lock. Called with kvm->slots_lock mutex acquired,
1165 * serializing operations for VM memory regions.
1167 void kvm_mmu_wp_memory_region(struct kvm *kvm, int slot)
1169 struct kvm_memslots *slots = kvm_memslots(kvm);
1170 struct kvm_memory_slot *memslot = id_to_memslot(slots, slot);
1171 phys_addr_t start = memslot->base_gfn << PAGE_SHIFT;
1172 phys_addr_t end = (memslot->base_gfn + memslot->npages) << PAGE_SHIFT;
1174 spin_lock(&kvm->mmu_lock);
1175 stage2_wp_range(kvm, start, end);
1176 spin_unlock(&kvm->mmu_lock);
1177 kvm_flush_remote_tlbs(kvm);
1181 * kvm_mmu_write_protect_pt_masked() - write protect dirty pages
1182 * @kvm: The KVM pointer
1183 * @slot: The memory slot associated with mask
1184 * @gfn_offset: The gfn offset in memory slot
1185 * @mask: The mask of dirty pages at offset 'gfn_offset' in this memory
1186 * slot to be write protected
1188 * Walks bits set in mask write protects the associated pte's. Caller must
1189 * acquire kvm_mmu_lock.
1191 static void kvm_mmu_write_protect_pt_masked(struct kvm *kvm,
1192 struct kvm_memory_slot *slot,
1193 gfn_t gfn_offset, unsigned long mask)
1195 phys_addr_t base_gfn = slot->base_gfn + gfn_offset;
1196 phys_addr_t start = (base_gfn + __ffs(mask)) << PAGE_SHIFT;
1197 phys_addr_t end = (base_gfn + __fls(mask) + 1) << PAGE_SHIFT;
1199 stage2_wp_range(kvm, start, end);
1203 * kvm_arch_mmu_enable_log_dirty_pt_masked - enable dirty logging for selected
1206 * It calls kvm_mmu_write_protect_pt_masked to write protect selected pages to
1207 * enable dirty logging for them.
1209 void kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm *kvm,
1210 struct kvm_memory_slot *slot,
1211 gfn_t gfn_offset, unsigned long mask)
1213 kvm_mmu_write_protect_pt_masked(kvm, slot, gfn_offset, mask);
1216 static void coherent_cache_guest_page(struct kvm_vcpu *vcpu, pfn_t pfn,
1217 unsigned long size, bool uncached)
1219 __coherent_cache_guest_page(vcpu, pfn, size, uncached);
1222 static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
1223 struct kvm_memory_slot *memslot, unsigned long hva,
1224 unsigned long fault_status)
1227 bool write_fault, writable, hugetlb = false, force_pte = false;
1228 unsigned long mmu_seq;
1229 gfn_t gfn = fault_ipa >> PAGE_SHIFT;
1230 struct kvm *kvm = vcpu->kvm;
1231 struct kvm_mmu_memory_cache *memcache = &vcpu->arch.mmu_page_cache;
1232 struct vm_area_struct *vma;
1234 pgprot_t mem_type = PAGE_S2;
1235 bool fault_ipa_uncached;
1236 bool logging_active = memslot_is_logging(memslot);
1237 unsigned long flags = 0;
1239 write_fault = kvm_is_write_fault(vcpu);
1240 if (fault_status == FSC_PERM && !write_fault) {
1241 kvm_err("Unexpected L2 read permission error\n");
1245 /* Let's check if we will get back a huge page backed by hugetlbfs */
1246 down_read(¤t->mm->mmap_sem);
1247 vma = find_vma_intersection(current->mm, hva, hva + 1);
1248 if (unlikely(!vma)) {
1249 kvm_err("Failed to find VMA for hva 0x%lx\n", hva);
1250 up_read(¤t->mm->mmap_sem);
1254 if (is_vm_hugetlb_page(vma) && !logging_active) {
1256 gfn = (fault_ipa & PMD_MASK) >> PAGE_SHIFT;
1259 * Pages belonging to memslots that don't have the same
1260 * alignment for userspace and IPA cannot be mapped using
1261 * block descriptors even if the pages belong to a THP for
1262 * the process, because the stage-2 block descriptor will
1263 * cover more than a single THP and we loose atomicity for
1264 * unmapping, updates, and splits of the THP or other pages
1265 * in the stage-2 block range.
1267 if ((memslot->userspace_addr & ~PMD_MASK) !=
1268 ((memslot->base_gfn << PAGE_SHIFT) & ~PMD_MASK))
1271 up_read(¤t->mm->mmap_sem);
1273 /* We need minimum second+third level pages */
1274 ret = mmu_topup_memory_cache(memcache, KVM_MMU_CACHE_MIN_PAGES,
1279 mmu_seq = vcpu->kvm->mmu_notifier_seq;
1281 * Ensure the read of mmu_notifier_seq happens before we call
1282 * gfn_to_pfn_prot (which calls get_user_pages), so that we don't risk
1283 * the page we just got a reference to gets unmapped before we have a
1284 * chance to grab the mmu_lock, which ensure that if the page gets
1285 * unmapped afterwards, the call to kvm_unmap_hva will take it away
1286 * from us again properly. This smp_rmb() interacts with the smp_wmb()
1287 * in kvm_mmu_notifier_invalidate_<page|range_end>.
1291 pfn = gfn_to_pfn_prot(kvm, gfn, write_fault, &writable);
1292 if (is_error_pfn(pfn))
1295 if (kvm_is_device_pfn(pfn)) {
1296 mem_type = PAGE_S2_DEVICE;
1297 flags |= KVM_S2PTE_FLAG_IS_IOMAP;
1298 } else if (logging_active) {
1300 * Faults on pages in a memslot with logging enabled
1301 * should not be mapped with huge pages (it introduces churn
1302 * and performance degradation), so force a pte mapping.
1305 flags |= KVM_S2_FLAG_LOGGING_ACTIVE;
1308 * Only actually map the page as writable if this was a write
1315 spin_lock(&kvm->mmu_lock);
1316 if (mmu_notifier_retry(kvm, mmu_seq))
1319 if (!hugetlb && !force_pte)
1320 hugetlb = transparent_hugepage_adjust(&pfn, &fault_ipa);
1322 fault_ipa_uncached = memslot->flags & KVM_MEMSLOT_INCOHERENT;
1325 pmd_t new_pmd = pfn_pmd(pfn, mem_type);
1326 new_pmd = pmd_mkhuge(new_pmd);
1328 kvm_set_s2pmd_writable(&new_pmd);
1329 kvm_set_pfn_dirty(pfn);
1331 coherent_cache_guest_page(vcpu, pfn, PMD_SIZE, fault_ipa_uncached);
1332 ret = stage2_set_pmd_huge(kvm, memcache, fault_ipa, &new_pmd);
1334 pte_t new_pte = pfn_pte(pfn, mem_type);
1337 kvm_set_s2pte_writable(&new_pte);
1338 kvm_set_pfn_dirty(pfn);
1339 mark_page_dirty(kvm, gfn);
1341 coherent_cache_guest_page(vcpu, pfn, PAGE_SIZE, fault_ipa_uncached);
1342 ret = stage2_set_pte(kvm, memcache, fault_ipa, &new_pte, flags);
1346 spin_unlock(&kvm->mmu_lock);
1347 kvm_set_pfn_accessed(pfn);
1348 kvm_release_pfn_clean(pfn);
1353 * Resolve the access fault by making the page young again.
1354 * Note that because the faulting entry is guaranteed not to be
1355 * cached in the TLB, we don't need to invalidate anything.
1357 static void handle_access_fault(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa)
1362 bool pfn_valid = false;
1364 trace_kvm_access_fault(fault_ipa);
1366 spin_lock(&vcpu->kvm->mmu_lock);
1368 pmd = stage2_get_pmd(vcpu->kvm, NULL, fault_ipa);
1369 if (!pmd || pmd_none(*pmd)) /* Nothing there */
1372 if (kvm_pmd_huge(*pmd)) { /* THP, HugeTLB */
1373 *pmd = pmd_mkyoung(*pmd);
1374 pfn = pmd_pfn(*pmd);
1379 pte = pte_offset_kernel(pmd, fault_ipa);
1380 if (pte_none(*pte)) /* Nothing there either */
1383 *pte = pte_mkyoung(*pte); /* Just a page... */
1384 pfn = pte_pfn(*pte);
1387 spin_unlock(&vcpu->kvm->mmu_lock);
1389 kvm_set_pfn_accessed(pfn);
1393 * kvm_handle_guest_abort - handles all 2nd stage aborts
1394 * @vcpu: the VCPU pointer
1395 * @run: the kvm_run structure
1397 * Any abort that gets to the host is almost guaranteed to be caused by a
1398 * missing second stage translation table entry, which can mean that either the
1399 * guest simply needs more memory and we must allocate an appropriate page or it
1400 * can mean that the guest tried to access I/O memory, which is emulated by user
1401 * space. The distinction is based on the IPA causing the fault and whether this
1402 * memory region has been registered as standard RAM by user space.
1404 int kvm_handle_guest_abort(struct kvm_vcpu *vcpu, struct kvm_run *run)
1406 unsigned long fault_status;
1407 phys_addr_t fault_ipa;
1408 struct kvm_memory_slot *memslot;
1410 bool is_iabt, write_fault, writable;
1414 is_iabt = kvm_vcpu_trap_is_iabt(vcpu);
1415 fault_ipa = kvm_vcpu_get_fault_ipa(vcpu);
1417 trace_kvm_guest_fault(*vcpu_pc(vcpu), kvm_vcpu_get_hsr(vcpu),
1418 kvm_vcpu_get_hfar(vcpu), fault_ipa);
1420 /* Check the stage-2 fault is trans. fault or write fault */
1421 fault_status = kvm_vcpu_trap_get_fault_type(vcpu);
1422 if (fault_status != FSC_FAULT && fault_status != FSC_PERM &&
1423 fault_status != FSC_ACCESS) {
1424 kvm_err("Unsupported FSC: EC=%#x xFSC=%#lx ESR_EL2=%#lx\n",
1425 kvm_vcpu_trap_get_class(vcpu),
1426 (unsigned long)kvm_vcpu_trap_get_fault(vcpu),
1427 (unsigned long)kvm_vcpu_get_hsr(vcpu));
1431 idx = srcu_read_lock(&vcpu->kvm->srcu);
1433 gfn = fault_ipa >> PAGE_SHIFT;
1434 memslot = gfn_to_memslot(vcpu->kvm, gfn);
1435 hva = gfn_to_hva_memslot_prot(memslot, gfn, &writable);
1436 write_fault = kvm_is_write_fault(vcpu);
1437 if (kvm_is_error_hva(hva) || (write_fault && !writable)) {
1439 /* Prefetch Abort on I/O address */
1440 kvm_inject_pabt(vcpu, kvm_vcpu_get_hfar(vcpu));
1446 * The IPA is reported as [MAX:12], so we need to
1447 * complement it with the bottom 12 bits from the
1448 * faulting VA. This is always 12 bits, irrespective
1451 fault_ipa |= kvm_vcpu_get_hfar(vcpu) & ((1 << 12) - 1);
1452 ret = io_mem_abort(vcpu, run, fault_ipa);
1456 /* Userspace should not be able to register out-of-bounds IPAs */
1457 VM_BUG_ON(fault_ipa >= KVM_PHYS_SIZE);
1459 if (fault_status == FSC_ACCESS) {
1460 handle_access_fault(vcpu, fault_ipa);
1465 ret = user_mem_abort(vcpu, fault_ipa, memslot, hva, fault_status);
1469 srcu_read_unlock(&vcpu->kvm->srcu, idx);
1473 static int handle_hva_to_gpa(struct kvm *kvm,
1474 unsigned long start,
1476 int (*handler)(struct kvm *kvm,
1477 gpa_t gpa, void *data),
1480 struct kvm_memslots *slots;
1481 struct kvm_memory_slot *memslot;
1484 slots = kvm_memslots(kvm);
1486 /* we only care about the pages that the guest sees */
1487 kvm_for_each_memslot(memslot, slots) {
1488 unsigned long hva_start, hva_end;
1491 hva_start = max(start, memslot->userspace_addr);
1492 hva_end = min(end, memslot->userspace_addr +
1493 (memslot->npages << PAGE_SHIFT));
1494 if (hva_start >= hva_end)
1498 * {gfn(page) | page intersects with [hva_start, hva_end)} =
1499 * {gfn_start, gfn_start+1, ..., gfn_end-1}.
1501 gfn = hva_to_gfn_memslot(hva_start, memslot);
1502 gfn_end = hva_to_gfn_memslot(hva_end + PAGE_SIZE - 1, memslot);
1504 for (; gfn < gfn_end; ++gfn) {
1505 gpa_t gpa = gfn << PAGE_SHIFT;
1506 ret |= handler(kvm, gpa, data);
1513 static int kvm_unmap_hva_handler(struct kvm *kvm, gpa_t gpa, void *data)
1515 unmap_stage2_range(kvm, gpa, PAGE_SIZE);
1519 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
1521 unsigned long end = hva + PAGE_SIZE;
1526 trace_kvm_unmap_hva(hva);
1527 handle_hva_to_gpa(kvm, hva, end, &kvm_unmap_hva_handler, NULL);
1531 int kvm_unmap_hva_range(struct kvm *kvm,
1532 unsigned long start, unsigned long end)
1537 trace_kvm_unmap_hva_range(start, end);
1538 handle_hva_to_gpa(kvm, start, end, &kvm_unmap_hva_handler, NULL);
1542 static int kvm_set_spte_handler(struct kvm *kvm, gpa_t gpa, void *data)
1544 pte_t *pte = (pte_t *)data;
1547 * We can always call stage2_set_pte with KVM_S2PTE_FLAG_LOGGING_ACTIVE
1548 * flag clear because MMU notifiers will have unmapped a huge PMD before
1549 * calling ->change_pte() (which in turn calls kvm_set_spte_hva()) and
1550 * therefore stage2_set_pte() never needs to clear out a huge PMD
1551 * through this calling path.
1553 stage2_set_pte(kvm, NULL, gpa, pte, 0);
1558 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
1560 unsigned long end = hva + PAGE_SIZE;
1566 trace_kvm_set_spte_hva(hva);
1567 stage2_pte = pfn_pte(pte_pfn(pte), PAGE_S2);
1568 handle_hva_to_gpa(kvm, hva, end, &kvm_set_spte_handler, &stage2_pte);
1571 static int kvm_age_hva_handler(struct kvm *kvm, gpa_t gpa, void *data)
1576 pmd = stage2_get_pmd(kvm, NULL, gpa);
1577 if (!pmd || pmd_none(*pmd)) /* Nothing there */
1580 if (kvm_pmd_huge(*pmd)) { /* THP, HugeTLB */
1581 if (pmd_young(*pmd)) {
1582 *pmd = pmd_mkold(*pmd);
1589 pte = pte_offset_kernel(pmd, gpa);
1593 if (pte_young(*pte)) {
1594 *pte = pte_mkold(*pte); /* Just a page... */
1601 static int kvm_test_age_hva_handler(struct kvm *kvm, gpa_t gpa, void *data)
1606 pmd = stage2_get_pmd(kvm, NULL, gpa);
1607 if (!pmd || pmd_none(*pmd)) /* Nothing there */
1610 if (kvm_pmd_huge(*pmd)) /* THP, HugeTLB */
1611 return pmd_young(*pmd);
1613 pte = pte_offset_kernel(pmd, gpa);
1614 if (!pte_none(*pte)) /* Just a page... */
1615 return pte_young(*pte);
1620 int kvm_age_hva(struct kvm *kvm, unsigned long start, unsigned long end)
1622 trace_kvm_age_hva(start, end);
1623 return handle_hva_to_gpa(kvm, start, end, kvm_age_hva_handler, NULL);
1626 int kvm_test_age_hva(struct kvm *kvm, unsigned long hva)
1628 trace_kvm_test_age_hva(hva);
1629 return handle_hva_to_gpa(kvm, hva, hva, kvm_test_age_hva_handler, NULL);
1632 void kvm_mmu_free_memory_caches(struct kvm_vcpu *vcpu)
1634 mmu_free_memory_cache(&vcpu->arch.mmu_page_cache);
1637 phys_addr_t kvm_mmu_get_httbr(void)
1639 if (__kvm_cpu_uses_extended_idmap())
1640 return virt_to_phys(merged_hyp_pgd);
1642 return virt_to_phys(hyp_pgd);
1645 phys_addr_t kvm_mmu_get_boot_httbr(void)
1647 if (__kvm_cpu_uses_extended_idmap())
1648 return virt_to_phys(merged_hyp_pgd);
1650 return virt_to_phys(boot_hyp_pgd);
1653 phys_addr_t kvm_get_idmap_vector(void)
1655 return hyp_idmap_vector;
1658 phys_addr_t kvm_get_idmap_start(void)
1660 return hyp_idmap_start;
1663 int kvm_mmu_init(void)
1667 hyp_idmap_start = kvm_virt_to_phys(__hyp_idmap_text_start);
1668 hyp_idmap_end = kvm_virt_to_phys(__hyp_idmap_text_end);
1669 hyp_idmap_vector = kvm_virt_to_phys(__kvm_hyp_init);
1672 * We rely on the linker script to ensure at build time that the HYP
1673 * init code does not cross a page boundary.
1675 BUG_ON((hyp_idmap_start ^ (hyp_idmap_end - 1)) & PAGE_MASK);
1677 hyp_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, hyp_pgd_order);
1678 boot_hyp_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, hyp_pgd_order);
1680 if (!hyp_pgd || !boot_hyp_pgd) {
1681 kvm_err("Hyp mode PGD not allocated\n");
1686 /* Create the idmap in the boot page tables */
1687 err = __create_hyp_mappings(boot_hyp_pgd,
1688 hyp_idmap_start, hyp_idmap_end,
1689 __phys_to_pfn(hyp_idmap_start),
1693 kvm_err("Failed to idmap %lx-%lx\n",
1694 hyp_idmap_start, hyp_idmap_end);
1698 if (__kvm_cpu_uses_extended_idmap()) {
1699 merged_hyp_pgd = (pgd_t *)__get_free_page(GFP_KERNEL | __GFP_ZERO);
1700 if (!merged_hyp_pgd) {
1701 kvm_err("Failed to allocate extra HYP pgd\n");
1704 __kvm_extend_hypmap(boot_hyp_pgd, hyp_pgd, merged_hyp_pgd,
1709 /* Map the very same page at the trampoline VA */
1710 err = __create_hyp_mappings(boot_hyp_pgd,
1711 TRAMPOLINE_VA, TRAMPOLINE_VA + PAGE_SIZE,
1712 __phys_to_pfn(hyp_idmap_start),
1715 kvm_err("Failed to map trampoline @%lx into boot HYP pgd\n",
1720 /* Map the same page again into the runtime page tables */
1721 err = __create_hyp_mappings(hyp_pgd,
1722 TRAMPOLINE_VA, TRAMPOLINE_VA + PAGE_SIZE,
1723 __phys_to_pfn(hyp_idmap_start),
1726 kvm_err("Failed to map trampoline @%lx into runtime HYP pgd\n",
1737 void kvm_arch_commit_memory_region(struct kvm *kvm,
1738 const struct kvm_userspace_memory_region *mem,
1739 const struct kvm_memory_slot *old,
1740 const struct kvm_memory_slot *new,
1741 enum kvm_mr_change change)
1744 * At this point memslot has been committed and there is an
1745 * allocated dirty_bitmap[], dirty pages will be be tracked while the
1746 * memory slot is write protected.
1748 if (change != KVM_MR_DELETE && mem->flags & KVM_MEM_LOG_DIRTY_PAGES)
1749 kvm_mmu_wp_memory_region(kvm, mem->slot);
1752 int kvm_arch_prepare_memory_region(struct kvm *kvm,
1753 struct kvm_memory_slot *memslot,
1754 const struct kvm_userspace_memory_region *mem,
1755 enum kvm_mr_change change)
1757 hva_t hva = mem->userspace_addr;
1758 hva_t reg_end = hva + mem->memory_size;
1759 bool writable = !(mem->flags & KVM_MEM_READONLY);
1762 if (change != KVM_MR_CREATE && change != KVM_MR_MOVE &&
1763 change != KVM_MR_FLAGS_ONLY)
1767 * Prevent userspace from creating a memory region outside of the IPA
1768 * space addressable by the KVM guest IPA space.
1770 if (memslot->base_gfn + memslot->npages >=
1771 (KVM_PHYS_SIZE >> PAGE_SHIFT))
1775 * A memory region could potentially cover multiple VMAs, and any holes
1776 * between them, so iterate over all of them to find out if we can map
1777 * any of them right now.
1779 * +--------------------------------------------+
1780 * +---------------+----------------+ +----------------+
1781 * | : VMA 1 | VMA 2 | | VMA 3 : |
1782 * +---------------+----------------+ +----------------+
1784 * +--------------------------------------------+
1787 struct vm_area_struct *vma = find_vma(current->mm, hva);
1788 hva_t vm_start, vm_end;
1790 if (!vma || vma->vm_start >= reg_end)
1794 * Mapping a read-only VMA is only allowed if the
1795 * memory region is configured as read-only.
1797 if (writable && !(vma->vm_flags & VM_WRITE)) {
1803 * Take the intersection of this VMA with the memory region
1805 vm_start = max(hva, vma->vm_start);
1806 vm_end = min(reg_end, vma->vm_end);
1808 if (vma->vm_flags & VM_PFNMAP) {
1809 gpa_t gpa = mem->guest_phys_addr +
1810 (vm_start - mem->userspace_addr);
1813 pa = (phys_addr_t)vma->vm_pgoff << PAGE_SHIFT;
1814 pa += vm_start - vma->vm_start;
1816 /* IO region dirty page logging not allowed */
1817 if (memslot->flags & KVM_MEM_LOG_DIRTY_PAGES)
1820 ret = kvm_phys_addr_ioremap(kvm, gpa, pa,
1827 } while (hva < reg_end);
1829 if (change == KVM_MR_FLAGS_ONLY)
1832 spin_lock(&kvm->mmu_lock);
1834 unmap_stage2_range(kvm, mem->guest_phys_addr, mem->memory_size);
1836 stage2_flush_memslot(kvm, memslot);
1837 spin_unlock(&kvm->mmu_lock);
1841 void kvm_arch_free_memslot(struct kvm *kvm, struct kvm_memory_slot *free,
1842 struct kvm_memory_slot *dont)
1846 int kvm_arch_create_memslot(struct kvm *kvm, struct kvm_memory_slot *slot,
1847 unsigned long npages)
1850 * Readonly memslots are not incoherent with the caches by definition,
1851 * but in practice, they are used mostly to emulate ROMs or NOR flashes
1852 * that the guest may consider devices and hence map as uncached.
1853 * To prevent incoherency issues in these cases, tag all readonly
1854 * regions as incoherent.
1856 if (slot->flags & KVM_MEM_READONLY)
1857 slot->flags |= KVM_MEMSLOT_INCOHERENT;
1861 void kvm_arch_memslots_updated(struct kvm *kvm, struct kvm_memslots *slots)
1865 void kvm_arch_flush_shadow_all(struct kvm *kvm)
1867 kvm_free_stage2_pgd(kvm);
1870 void kvm_arch_flush_shadow_memslot(struct kvm *kvm,
1871 struct kvm_memory_slot *slot)
1873 gpa_t gpa = slot->base_gfn << PAGE_SHIFT;
1874 phys_addr_t size = slot->npages << PAGE_SHIFT;
1876 spin_lock(&kvm->mmu_lock);
1877 unmap_stage2_range(kvm, gpa, size);
1878 spin_unlock(&kvm->mmu_lock);
1882 * See note at ARMv7 ARM B1.14.4 (TL;DR: S/W ops are not easily virtualized).
1885 * - S/W ops are local to a CPU (not broadcast)
1886 * - We have line migration behind our back (speculation)
1887 * - System caches don't support S/W at all (damn!)
1889 * In the face of the above, the best we can do is to try and convert
1890 * S/W ops to VA ops. Because the guest is not allowed to infer the
1891 * S/W to PA mapping, it can only use S/W to nuke the whole cache,
1892 * which is a rather good thing for us.
1894 * Also, it is only used when turning caches on/off ("The expected
1895 * usage of the cache maintenance instructions that operate by set/way
1896 * is associated with the cache maintenance instructions associated
1897 * with the powerdown and powerup of caches, if this is required by
1898 * the implementation.").
1900 * We use the following policy:
1902 * - If we trap a S/W operation, we enable VM trapping to detect
1903 * caches being turned on/off, and do a full clean.
1905 * - We flush the caches on both caches being turned on and off.
1907 * - Once the caches are enabled, we stop trapping VM ops.
1909 void kvm_set_way_flush(struct kvm_vcpu *vcpu)
1911 unsigned long hcr = vcpu_get_hcr(vcpu);
1914 * If this is the first time we do a S/W operation
1915 * (i.e. HCR_TVM not set) flush the whole memory, and set the
1918 * Otherwise, rely on the VM trapping to wait for the MMU +
1919 * Caches to be turned off. At that point, we'll be able to
1920 * clean the caches again.
1922 if (!(hcr & HCR_TVM)) {
1923 trace_kvm_set_way_flush(*vcpu_pc(vcpu),
1924 vcpu_has_cache_enabled(vcpu));
1925 stage2_flush_vm(vcpu->kvm);
1926 vcpu_set_hcr(vcpu, hcr | HCR_TVM);
1930 void kvm_toggle_cache(struct kvm_vcpu *vcpu, bool was_enabled)
1932 bool now_enabled = vcpu_has_cache_enabled(vcpu);
1935 * If switching the MMU+caches on, need to invalidate the caches.
1936 * If switching it off, need to clean the caches.
1937 * Clean + invalidate does the trick always.
1939 if (now_enabled != was_enabled)
1940 stage2_flush_vm(vcpu->kvm);
1942 /* Caches are now on, stop trapping VM ops (until a S/W op) */
1944 vcpu_set_hcr(vcpu, vcpu_get_hcr(vcpu) & ~HCR_TVM);
1946 trace_kvm_toggle_cache(*vcpu_pc(vcpu), was_enabled, now_enabled);