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>
34 extern char __hyp_idmap_text_start[], __hyp_idmap_text_end[];
36 static pgd_t *boot_hyp_pgd;
37 static pgd_t *hyp_pgd;
38 static DEFINE_MUTEX(kvm_hyp_pgd_mutex);
40 static void *init_bounce_page;
41 static unsigned long hyp_idmap_start;
42 static unsigned long hyp_idmap_end;
43 static phys_addr_t hyp_idmap_vector;
45 #define hyp_pgd_order get_order(PTRS_PER_PGD * sizeof(pgd_t))
47 #define kvm_pmd_huge(_x) (pmd_huge(_x) || pmd_trans_huge(_x))
49 static void kvm_tlb_flush_vmid_ipa(struct kvm *kvm, phys_addr_t ipa)
52 * This function also gets called when dealing with HYP page
53 * tables. As HYP doesn't have an associated struct kvm (and
54 * the HYP page tables are fairly static), we don't do
58 kvm_call_hyp(__kvm_tlb_flush_vmid_ipa, kvm, ipa);
62 * D-Cache management functions. They take the page table entries by
63 * value, as they are flushing the cache using the kernel mapping (or
66 static void kvm_flush_dcache_pte(pte_t pte)
68 __kvm_flush_dcache_pte(pte);
71 static void kvm_flush_dcache_pmd(pmd_t pmd)
73 __kvm_flush_dcache_pmd(pmd);
76 static void kvm_flush_dcache_pud(pud_t pud)
78 __kvm_flush_dcache_pud(pud);
81 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
86 BUG_ON(max > KVM_NR_MEM_OBJS);
87 if (cache->nobjs >= min)
89 while (cache->nobjs < max) {
90 page = (void *)__get_free_page(PGALLOC_GFP);
93 cache->objects[cache->nobjs++] = page;
98 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
101 free_page((unsigned long)mc->objects[--mc->nobjs]);
104 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc)
108 BUG_ON(!mc || !mc->nobjs);
109 p = mc->objects[--mc->nobjs];
113 static void clear_pgd_entry(struct kvm *kvm, pgd_t *pgd, phys_addr_t addr)
115 pud_t *pud_table __maybe_unused = pud_offset(pgd, 0);
117 kvm_tlb_flush_vmid_ipa(kvm, addr);
118 pud_free(NULL, pud_table);
119 put_page(virt_to_page(pgd));
122 static void clear_pud_entry(struct kvm *kvm, pud_t *pud, phys_addr_t addr)
124 pmd_t *pmd_table = pmd_offset(pud, 0);
125 VM_BUG_ON(pud_huge(*pud));
127 kvm_tlb_flush_vmid_ipa(kvm, addr);
128 pmd_free(NULL, pmd_table);
129 put_page(virt_to_page(pud));
132 static void clear_pmd_entry(struct kvm *kvm, pmd_t *pmd, phys_addr_t addr)
134 pte_t *pte_table = pte_offset_kernel(pmd, 0);
135 VM_BUG_ON(kvm_pmd_huge(*pmd));
137 kvm_tlb_flush_vmid_ipa(kvm, addr);
138 pte_free_kernel(NULL, pte_table);
139 put_page(virt_to_page(pmd));
143 * Unmapping vs dcache management:
145 * If a guest maps certain memory pages as uncached, all writes will
146 * bypass the data cache and go directly to RAM. However, the CPUs
147 * can still speculate reads (not writes) and fill cache lines with
150 * Those cache lines will be *clean* cache lines though, so a
151 * clean+invalidate operation is equivalent to an invalidate
152 * operation, because no cache lines are marked dirty.
154 * Those clean cache lines could be filled prior to an uncached write
155 * by the guest, and the cache coherent IO subsystem would therefore
156 * end up writing old data to disk.
158 * This is why right after unmapping a page/section and invalidating
159 * the corresponding TLBs, we call kvm_flush_dcache_p*() to make sure
160 * the IO subsystem will never hit in the cache.
162 static void unmap_ptes(struct kvm *kvm, pmd_t *pmd,
163 phys_addr_t addr, phys_addr_t end)
165 phys_addr_t start_addr = addr;
166 pte_t *pte, *start_pte;
168 start_pte = pte = pte_offset_kernel(pmd, addr);
170 if (!pte_none(*pte)) {
171 pte_t old_pte = *pte;
173 kvm_set_pte(pte, __pte(0));
174 kvm_tlb_flush_vmid_ipa(kvm, addr);
176 /* No need to invalidate the cache for device mappings */
177 if ((pte_val(old_pte) & PAGE_S2_DEVICE) != PAGE_S2_DEVICE)
178 kvm_flush_dcache_pte(old_pte);
180 put_page(virt_to_page(pte));
182 } while (pte++, addr += PAGE_SIZE, addr != end);
184 if (kvm_pte_table_empty(kvm, start_pte))
185 clear_pmd_entry(kvm, pmd, start_addr);
188 static void unmap_pmds(struct kvm *kvm, pud_t *pud,
189 phys_addr_t addr, phys_addr_t end)
191 phys_addr_t next, start_addr = addr;
192 pmd_t *pmd, *start_pmd;
194 start_pmd = pmd = pmd_offset(pud, addr);
196 next = kvm_pmd_addr_end(addr, end);
197 if (!pmd_none(*pmd)) {
198 if (kvm_pmd_huge(*pmd)) {
199 pmd_t old_pmd = *pmd;
202 kvm_tlb_flush_vmid_ipa(kvm, addr);
204 kvm_flush_dcache_pmd(old_pmd);
206 put_page(virt_to_page(pmd));
208 unmap_ptes(kvm, pmd, addr, next);
211 } while (pmd++, addr = next, addr != end);
213 if (kvm_pmd_table_empty(kvm, start_pmd))
214 clear_pud_entry(kvm, pud, start_addr);
217 static void unmap_puds(struct kvm *kvm, pgd_t *pgd,
218 phys_addr_t addr, phys_addr_t end)
220 phys_addr_t next, start_addr = addr;
221 pud_t *pud, *start_pud;
223 start_pud = pud = pud_offset(pgd, addr);
225 next = kvm_pud_addr_end(addr, end);
226 if (!pud_none(*pud)) {
227 if (pud_huge(*pud)) {
228 pud_t old_pud = *pud;
231 kvm_tlb_flush_vmid_ipa(kvm, addr);
233 kvm_flush_dcache_pud(old_pud);
235 put_page(virt_to_page(pud));
237 unmap_pmds(kvm, pud, addr, next);
240 } while (pud++, addr = next, addr != end);
242 if (kvm_pud_table_empty(kvm, start_pud))
243 clear_pgd_entry(kvm, pgd, start_addr);
247 static void unmap_range(struct kvm *kvm, pgd_t *pgdp,
248 phys_addr_t start, u64 size)
251 phys_addr_t addr = start, end = start + size;
254 pgd = pgdp + pgd_index(addr);
256 next = kvm_pgd_addr_end(addr, end);
258 unmap_puds(kvm, pgd, addr, next);
259 } while (pgd++, addr = next, addr != end);
262 static void stage2_flush_ptes(struct kvm *kvm, pmd_t *pmd,
263 phys_addr_t addr, phys_addr_t end)
267 pte = pte_offset_kernel(pmd, addr);
269 if (!pte_none(*pte) &&
270 (pte_val(*pte) & PAGE_S2_DEVICE) != PAGE_S2_DEVICE)
271 kvm_flush_dcache_pte(*pte);
272 } while (pte++, addr += PAGE_SIZE, addr != end);
275 static void stage2_flush_pmds(struct kvm *kvm, pud_t *pud,
276 phys_addr_t addr, phys_addr_t end)
281 pmd = pmd_offset(pud, addr);
283 next = kvm_pmd_addr_end(addr, end);
284 if (!pmd_none(*pmd)) {
285 if (kvm_pmd_huge(*pmd))
286 kvm_flush_dcache_pmd(*pmd);
288 stage2_flush_ptes(kvm, pmd, addr, next);
290 } while (pmd++, addr = next, addr != end);
293 static void stage2_flush_puds(struct kvm *kvm, pgd_t *pgd,
294 phys_addr_t addr, phys_addr_t end)
299 pud = pud_offset(pgd, addr);
301 next = kvm_pud_addr_end(addr, end);
302 if (!pud_none(*pud)) {
304 kvm_flush_dcache_pud(*pud);
306 stage2_flush_pmds(kvm, pud, addr, next);
308 } while (pud++, addr = next, addr != end);
311 static void stage2_flush_memslot(struct kvm *kvm,
312 struct kvm_memory_slot *memslot)
314 phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
315 phys_addr_t end = addr + PAGE_SIZE * memslot->npages;
319 pgd = kvm->arch.pgd + pgd_index(addr);
321 next = kvm_pgd_addr_end(addr, end);
322 stage2_flush_puds(kvm, pgd, addr, next);
323 } while (pgd++, addr = next, addr != end);
327 * stage2_flush_vm - Invalidate cache for pages mapped in stage 2
328 * @kvm: The struct kvm pointer
330 * Go through the stage 2 page tables and invalidate any cache lines
331 * backing memory already mapped to the VM.
333 static void stage2_flush_vm(struct kvm *kvm)
335 struct kvm_memslots *slots;
336 struct kvm_memory_slot *memslot;
339 idx = srcu_read_lock(&kvm->srcu);
340 spin_lock(&kvm->mmu_lock);
342 slots = kvm_memslots(kvm);
343 kvm_for_each_memslot(memslot, slots)
344 stage2_flush_memslot(kvm, memslot);
346 spin_unlock(&kvm->mmu_lock);
347 srcu_read_unlock(&kvm->srcu, idx);
351 * free_boot_hyp_pgd - free HYP boot page tables
353 * Free the HYP boot page tables. The bounce page is also freed.
355 void free_boot_hyp_pgd(void)
357 mutex_lock(&kvm_hyp_pgd_mutex);
360 unmap_range(NULL, boot_hyp_pgd, hyp_idmap_start, PAGE_SIZE);
361 unmap_range(NULL, boot_hyp_pgd, TRAMPOLINE_VA, PAGE_SIZE);
362 free_pages((unsigned long)boot_hyp_pgd, hyp_pgd_order);
367 unmap_range(NULL, hyp_pgd, TRAMPOLINE_VA, PAGE_SIZE);
369 free_page((unsigned long)init_bounce_page);
370 init_bounce_page = NULL;
372 mutex_unlock(&kvm_hyp_pgd_mutex);
376 * free_hyp_pgds - free Hyp-mode page tables
378 * Assumes hyp_pgd is a page table used strictly in Hyp-mode and
379 * therefore contains either mappings in the kernel memory area (above
380 * PAGE_OFFSET), or device mappings in the vmalloc range (from
381 * VMALLOC_START to VMALLOC_END).
383 * boot_hyp_pgd should only map two pages for the init code.
385 void free_hyp_pgds(void)
391 mutex_lock(&kvm_hyp_pgd_mutex);
394 for (addr = PAGE_OFFSET; virt_addr_valid(addr); addr += PGDIR_SIZE)
395 unmap_range(NULL, hyp_pgd, KERN_TO_HYP(addr), PGDIR_SIZE);
396 for (addr = VMALLOC_START; is_vmalloc_addr((void*)addr); addr += PGDIR_SIZE)
397 unmap_range(NULL, hyp_pgd, KERN_TO_HYP(addr), PGDIR_SIZE);
399 free_pages((unsigned long)hyp_pgd, hyp_pgd_order);
403 mutex_unlock(&kvm_hyp_pgd_mutex);
406 static void create_hyp_pte_mappings(pmd_t *pmd, unsigned long start,
407 unsigned long end, unsigned long pfn,
415 pte = pte_offset_kernel(pmd, addr);
416 kvm_set_pte(pte, pfn_pte(pfn, prot));
417 get_page(virt_to_page(pte));
418 kvm_flush_dcache_to_poc(pte, sizeof(*pte));
420 } while (addr += PAGE_SIZE, addr != end);
423 static int create_hyp_pmd_mappings(pud_t *pud, unsigned long start,
424 unsigned long end, unsigned long pfn,
429 unsigned long addr, next;
433 pmd = pmd_offset(pud, addr);
435 BUG_ON(pmd_sect(*pmd));
437 if (pmd_none(*pmd)) {
438 pte = pte_alloc_one_kernel(NULL, addr);
440 kvm_err("Cannot allocate Hyp pte\n");
443 pmd_populate_kernel(NULL, pmd, pte);
444 get_page(virt_to_page(pmd));
445 kvm_flush_dcache_to_poc(pmd, sizeof(*pmd));
448 next = pmd_addr_end(addr, end);
450 create_hyp_pte_mappings(pmd, addr, next, pfn, prot);
451 pfn += (next - addr) >> PAGE_SHIFT;
452 } while (addr = next, addr != end);
457 static int create_hyp_pud_mappings(pgd_t *pgd, unsigned long start,
458 unsigned long end, unsigned long pfn,
463 unsigned long addr, next;
468 pud = pud_offset(pgd, addr);
470 if (pud_none_or_clear_bad(pud)) {
471 pmd = pmd_alloc_one(NULL, addr);
473 kvm_err("Cannot allocate Hyp pmd\n");
476 pud_populate(NULL, pud, pmd);
477 get_page(virt_to_page(pud));
478 kvm_flush_dcache_to_poc(pud, sizeof(*pud));
481 next = pud_addr_end(addr, end);
482 ret = create_hyp_pmd_mappings(pud, addr, next, pfn, prot);
485 pfn += (next - addr) >> PAGE_SHIFT;
486 } while (addr = next, addr != end);
491 static int __create_hyp_mappings(pgd_t *pgdp,
492 unsigned long start, unsigned long end,
493 unsigned long pfn, pgprot_t prot)
497 unsigned long addr, next;
500 mutex_lock(&kvm_hyp_pgd_mutex);
501 addr = start & PAGE_MASK;
502 end = PAGE_ALIGN(end);
504 pgd = pgdp + pgd_index(addr);
506 if (pgd_none(*pgd)) {
507 pud = pud_alloc_one(NULL, addr);
509 kvm_err("Cannot allocate Hyp pud\n");
513 pgd_populate(NULL, pgd, pud);
514 get_page(virt_to_page(pgd));
515 kvm_flush_dcache_to_poc(pgd, sizeof(*pgd));
518 next = pgd_addr_end(addr, end);
519 err = create_hyp_pud_mappings(pgd, addr, next, pfn, prot);
522 pfn += (next - addr) >> PAGE_SHIFT;
523 } while (addr = next, addr != end);
525 mutex_unlock(&kvm_hyp_pgd_mutex);
529 static phys_addr_t kvm_kaddr_to_phys(void *kaddr)
531 if (!is_vmalloc_addr(kaddr)) {
532 BUG_ON(!virt_addr_valid(kaddr));
535 return page_to_phys(vmalloc_to_page(kaddr)) +
536 offset_in_page(kaddr);
541 * create_hyp_mappings - duplicate a kernel virtual address range in Hyp mode
542 * @from: The virtual kernel start address of the range
543 * @to: The virtual kernel end address of the range (exclusive)
545 * The same virtual address as the kernel virtual address is also used
546 * in Hyp-mode mapping (modulo HYP_PAGE_OFFSET) to the same underlying
549 int create_hyp_mappings(void *from, void *to)
551 phys_addr_t phys_addr;
552 unsigned long virt_addr;
553 unsigned long start = KERN_TO_HYP((unsigned long)from);
554 unsigned long end = KERN_TO_HYP((unsigned long)to);
556 start = start & PAGE_MASK;
557 end = PAGE_ALIGN(end);
559 for (virt_addr = start; virt_addr < end; virt_addr += PAGE_SIZE) {
562 phys_addr = kvm_kaddr_to_phys(from + virt_addr - start);
563 err = __create_hyp_mappings(hyp_pgd, virt_addr,
564 virt_addr + PAGE_SIZE,
565 __phys_to_pfn(phys_addr),
575 * create_hyp_io_mappings - duplicate a kernel IO mapping into Hyp mode
576 * @from: The kernel start VA of the range
577 * @to: The kernel end VA of the range (exclusive)
578 * @phys_addr: The physical start address which gets mapped
580 * The resulting HYP VA is the same as the kernel VA, modulo
583 int create_hyp_io_mappings(void *from, void *to, phys_addr_t phys_addr)
585 unsigned long start = KERN_TO_HYP((unsigned long)from);
586 unsigned long end = KERN_TO_HYP((unsigned long)to);
588 /* Check for a valid kernel IO mapping */
589 if (!is_vmalloc_addr(from) || !is_vmalloc_addr(to - 1))
592 return __create_hyp_mappings(hyp_pgd, start, end,
593 __phys_to_pfn(phys_addr), PAGE_HYP_DEVICE);
597 * kvm_alloc_stage2_pgd - allocate level-1 table for stage-2 translation.
598 * @kvm: The KVM struct pointer for the VM.
600 * Allocates the 1st level table only of size defined by S2_PGD_ORDER (can
601 * support either full 40-bit input addresses or limited to 32-bit input
602 * addresses). Clears the allocated pages.
604 * Note we don't need locking here as this is only called when the VM is
605 * created, which can only be done once.
607 int kvm_alloc_stage2_pgd(struct kvm *kvm)
612 if (kvm->arch.pgd != NULL) {
613 kvm_err("kvm_arch already initialized?\n");
617 if (KVM_PREALLOC_LEVEL > 0) {
619 * Allocate fake pgd for the page table manipulation macros to
620 * work. This is not used by the hardware and we have no
621 * alignment requirement for this allocation.
623 pgd = (pgd_t *)kmalloc(PTRS_PER_S2_PGD * sizeof(pgd_t),
624 GFP_KERNEL | __GFP_ZERO);
627 * Allocate actual first-level Stage-2 page table used by the
628 * hardware for Stage-2 page table walks.
630 pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, S2_PGD_ORDER);
636 ret = kvm_prealloc_hwpgd(kvm, pgd);
644 if (KVM_PREALLOC_LEVEL > 0)
647 free_pages((unsigned long)pgd, S2_PGD_ORDER);
652 * unmap_stage2_range -- Clear stage2 page table entries to unmap a range
653 * @kvm: The VM pointer
654 * @start: The intermediate physical base address of the range to unmap
655 * @size: The size of the area to unmap
657 * Clear a range of stage-2 mappings, lowering the various ref-counts. Must
658 * be called while holding mmu_lock (unless for freeing the stage2 pgd before
659 * destroying the VM), otherwise another faulting VCPU may come in and mess
660 * with things behind our backs.
662 static void unmap_stage2_range(struct kvm *kvm, phys_addr_t start, u64 size)
664 unmap_range(kvm, kvm->arch.pgd, start, size);
667 static void stage2_unmap_memslot(struct kvm *kvm,
668 struct kvm_memory_slot *memslot)
670 hva_t hva = memslot->userspace_addr;
671 phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
672 phys_addr_t size = PAGE_SIZE * memslot->npages;
673 hva_t reg_end = hva + size;
676 * A memory region could potentially cover multiple VMAs, and any holes
677 * between them, so iterate over all of them to find out if we should
680 * +--------------------------------------------+
681 * +---------------+----------------+ +----------------+
682 * | : VMA 1 | VMA 2 | | VMA 3 : |
683 * +---------------+----------------+ +----------------+
685 * +--------------------------------------------+
688 struct vm_area_struct *vma = find_vma(current->mm, hva);
689 hva_t vm_start, vm_end;
691 if (!vma || vma->vm_start >= reg_end)
695 * Take the intersection of this VMA with the memory region
697 vm_start = max(hva, vma->vm_start);
698 vm_end = min(reg_end, vma->vm_end);
700 if (!(vma->vm_flags & VM_PFNMAP)) {
701 gpa_t gpa = addr + (vm_start - memslot->userspace_addr);
702 unmap_stage2_range(kvm, gpa, vm_end - vm_start);
705 } while (hva < reg_end);
709 * stage2_unmap_vm - Unmap Stage-2 RAM mappings
710 * @kvm: The struct kvm pointer
712 * Go through the memregions and unmap any reguler RAM
713 * backing memory already mapped to the VM.
715 void stage2_unmap_vm(struct kvm *kvm)
717 struct kvm_memslots *slots;
718 struct kvm_memory_slot *memslot;
721 idx = srcu_read_lock(&kvm->srcu);
722 spin_lock(&kvm->mmu_lock);
724 slots = kvm_memslots(kvm);
725 kvm_for_each_memslot(memslot, slots)
726 stage2_unmap_memslot(kvm, memslot);
728 spin_unlock(&kvm->mmu_lock);
729 srcu_read_unlock(&kvm->srcu, idx);
733 * kvm_free_stage2_pgd - free all stage-2 tables
734 * @kvm: The KVM struct pointer for the VM.
736 * Walks the level-1 page table pointed to by kvm->arch.pgd and frees all
737 * underlying level-2 and level-3 tables before freeing the actual level-1 table
738 * and setting the struct pointer to NULL.
740 * Note we don't need locking here as this is only called when the VM is
741 * destroyed, which can only be done once.
743 void kvm_free_stage2_pgd(struct kvm *kvm)
745 if (kvm->arch.pgd == NULL)
748 unmap_stage2_range(kvm, 0, KVM_PHYS_SIZE);
750 if (KVM_PREALLOC_LEVEL > 0)
751 kfree(kvm->arch.pgd);
753 free_pages((unsigned long)kvm->arch.pgd, S2_PGD_ORDER);
754 kvm->arch.pgd = NULL;
757 static pud_t *stage2_get_pud(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
763 pgd = kvm->arch.pgd + pgd_index(addr);
764 if (WARN_ON(pgd_none(*pgd))) {
767 pud = mmu_memory_cache_alloc(cache);
768 pgd_populate(NULL, pgd, pud);
769 get_page(virt_to_page(pgd));
772 return pud_offset(pgd, addr);
775 static pmd_t *stage2_get_pmd(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
781 pud = stage2_get_pud(kvm, cache, addr);
782 if (pud_none(*pud)) {
785 pmd = mmu_memory_cache_alloc(cache);
786 pud_populate(NULL, pud, pmd);
787 get_page(virt_to_page(pud));
790 return pmd_offset(pud, addr);
793 static int stage2_set_pmd_huge(struct kvm *kvm, struct kvm_mmu_memory_cache
794 *cache, phys_addr_t addr, const pmd_t *new_pmd)
798 pmd = stage2_get_pmd(kvm, cache, addr);
802 * Mapping in huge pages should only happen through a fault. If a
803 * page is merged into a transparent huge page, the individual
804 * subpages of that huge page should be unmapped through MMU
805 * notifiers before we get here.
807 * Merging of CompoundPages is not supported; they should become
808 * splitting first, unmapped, merged, and mapped back in on-demand.
810 VM_BUG_ON(pmd_present(*pmd) && pmd_pfn(*pmd) != pmd_pfn(*new_pmd));
813 kvm_set_pmd(pmd, *new_pmd);
814 if (pmd_present(old_pmd))
815 kvm_tlb_flush_vmid_ipa(kvm, addr);
817 get_page(virt_to_page(pmd));
821 static int stage2_set_pte(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
822 phys_addr_t addr, const pte_t *new_pte, bool iomap)
827 /* Create stage-2 page table mapping - Levels 0 and 1 */
828 pmd = stage2_get_pmd(kvm, cache, addr);
831 * Ignore calls from kvm_set_spte_hva for unallocated
837 /* Create stage-2 page mappings - Level 2 */
838 if (pmd_none(*pmd)) {
840 return 0; /* ignore calls from kvm_set_spte_hva */
841 pte = mmu_memory_cache_alloc(cache);
843 pmd_populate_kernel(NULL, pmd, pte);
844 get_page(virt_to_page(pmd));
847 pte = pte_offset_kernel(pmd, addr);
849 if (iomap && pte_present(*pte))
852 /* Create 2nd stage page table mapping - Level 3 */
854 kvm_set_pte(pte, *new_pte);
855 if (pte_present(old_pte))
856 kvm_tlb_flush_vmid_ipa(kvm, addr);
858 get_page(virt_to_page(pte));
864 * kvm_phys_addr_ioremap - map a device range to guest IPA
866 * @kvm: The KVM pointer
867 * @guest_ipa: The IPA at which to insert the mapping
868 * @pa: The physical address of the device
869 * @size: The size of the mapping
871 int kvm_phys_addr_ioremap(struct kvm *kvm, phys_addr_t guest_ipa,
872 phys_addr_t pa, unsigned long size, bool writable)
874 phys_addr_t addr, end;
877 struct kvm_mmu_memory_cache cache = { 0, };
879 end = (guest_ipa + size + PAGE_SIZE - 1) & PAGE_MASK;
880 pfn = __phys_to_pfn(pa);
882 for (addr = guest_ipa; addr < end; addr += PAGE_SIZE) {
883 pte_t pte = pfn_pte(pfn, PAGE_S2_DEVICE);
886 kvm_set_s2pte_writable(&pte);
888 ret = mmu_topup_memory_cache(&cache, KVM_MMU_CACHE_MIN_PAGES,
892 spin_lock(&kvm->mmu_lock);
893 ret = stage2_set_pte(kvm, &cache, addr, &pte, true);
894 spin_unlock(&kvm->mmu_lock);
902 mmu_free_memory_cache(&cache);
906 static bool transparent_hugepage_adjust(pfn_t *pfnp, phys_addr_t *ipap)
909 gfn_t gfn = *ipap >> PAGE_SHIFT;
911 if (PageTransCompound(pfn_to_page(pfn))) {
914 * The address we faulted on is backed by a transparent huge
915 * page. However, because we map the compound huge page and
916 * not the individual tail page, we need to transfer the
917 * refcount to the head page. We have to be careful that the
918 * THP doesn't start to split while we are adjusting the
921 * We are sure this doesn't happen, because mmu_notifier_retry
922 * was successful and we are holding the mmu_lock, so if this
923 * THP is trying to split, it will be blocked in the mmu
924 * notifier before touching any of the pages, specifically
925 * before being able to call __split_huge_page_refcount().
927 * We can therefore safely transfer the refcount from PG_tail
928 * to PG_head and switch the pfn from a tail page to the head
931 mask = PTRS_PER_PMD - 1;
932 VM_BUG_ON((gfn & mask) != (pfn & mask));
935 kvm_release_pfn_clean(pfn);
947 static bool kvm_is_write_fault(struct kvm_vcpu *vcpu)
949 if (kvm_vcpu_trap_is_iabt(vcpu))
952 return kvm_vcpu_dabt_iswrite(vcpu);
955 static bool kvm_is_device_pfn(unsigned long pfn)
957 return !pfn_valid(pfn);
960 static void coherent_cache_guest_page(struct kvm_vcpu *vcpu, pfn_t pfn,
961 unsigned long size, bool uncached)
963 __coherent_cache_guest_page(vcpu, pfn, size, uncached);
966 static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
967 struct kvm_memory_slot *memslot, unsigned long hva,
968 unsigned long fault_status)
971 bool write_fault, writable, hugetlb = false, force_pte = false;
972 unsigned long mmu_seq;
973 gfn_t gfn = fault_ipa >> PAGE_SHIFT;
974 struct kvm *kvm = vcpu->kvm;
975 struct kvm_mmu_memory_cache *memcache = &vcpu->arch.mmu_page_cache;
976 struct vm_area_struct *vma;
978 pgprot_t mem_type = PAGE_S2;
979 bool fault_ipa_uncached;
981 write_fault = kvm_is_write_fault(vcpu);
982 if (fault_status == FSC_PERM && !write_fault) {
983 kvm_err("Unexpected L2 read permission error\n");
987 /* Let's check if we will get back a huge page backed by hugetlbfs */
988 down_read(¤t->mm->mmap_sem);
989 vma = find_vma_intersection(current->mm, hva, hva + 1);
990 if (unlikely(!vma)) {
991 kvm_err("Failed to find VMA for hva 0x%lx\n", hva);
992 up_read(¤t->mm->mmap_sem);
996 if (is_vm_hugetlb_page(vma)) {
998 gfn = (fault_ipa & PMD_MASK) >> PAGE_SHIFT;
1001 * Pages belonging to memslots that don't have the same
1002 * alignment for userspace and IPA cannot be mapped using
1003 * block descriptors even if the pages belong to a THP for
1004 * the process, because the stage-2 block descriptor will
1005 * cover more than a single THP and we loose atomicity for
1006 * unmapping, updates, and splits of the THP or other pages
1007 * in the stage-2 block range.
1009 if ((memslot->userspace_addr & ~PMD_MASK) !=
1010 ((memslot->base_gfn << PAGE_SHIFT) & ~PMD_MASK))
1013 up_read(¤t->mm->mmap_sem);
1015 /* We need minimum second+third level pages */
1016 ret = mmu_topup_memory_cache(memcache, KVM_MMU_CACHE_MIN_PAGES,
1021 mmu_seq = vcpu->kvm->mmu_notifier_seq;
1023 * Ensure the read of mmu_notifier_seq happens before we call
1024 * gfn_to_pfn_prot (which calls get_user_pages), so that we don't risk
1025 * the page we just got a reference to gets unmapped before we have a
1026 * chance to grab the mmu_lock, which ensure that if the page gets
1027 * unmapped afterwards, the call to kvm_unmap_hva will take it away
1028 * from us again properly. This smp_rmb() interacts with the smp_wmb()
1029 * in kvm_mmu_notifier_invalidate_<page|range_end>.
1033 pfn = gfn_to_pfn_prot(kvm, gfn, write_fault, &writable);
1034 if (is_error_pfn(pfn))
1037 if (kvm_is_device_pfn(pfn))
1038 mem_type = PAGE_S2_DEVICE;
1040 spin_lock(&kvm->mmu_lock);
1041 if (mmu_notifier_retry(kvm, mmu_seq))
1043 if (!hugetlb && !force_pte)
1044 hugetlb = transparent_hugepage_adjust(&pfn, &fault_ipa);
1046 fault_ipa_uncached = memslot->flags & KVM_MEMSLOT_INCOHERENT;
1049 pmd_t new_pmd = pfn_pmd(pfn, mem_type);
1050 new_pmd = pmd_mkhuge(new_pmd);
1052 kvm_set_s2pmd_writable(&new_pmd);
1053 kvm_set_pfn_dirty(pfn);
1055 coherent_cache_guest_page(vcpu, pfn, PMD_SIZE, fault_ipa_uncached);
1056 ret = stage2_set_pmd_huge(kvm, memcache, fault_ipa, &new_pmd);
1058 pte_t new_pte = pfn_pte(pfn, mem_type);
1060 kvm_set_s2pte_writable(&new_pte);
1061 kvm_set_pfn_dirty(pfn);
1063 coherent_cache_guest_page(vcpu, pfn, PAGE_SIZE, fault_ipa_uncached);
1064 ret = stage2_set_pte(kvm, memcache, fault_ipa, &new_pte,
1065 pgprot_val(mem_type) == pgprot_val(PAGE_S2_DEVICE));
1070 spin_unlock(&kvm->mmu_lock);
1071 kvm_release_pfn_clean(pfn);
1076 * kvm_handle_guest_abort - handles all 2nd stage aborts
1077 * @vcpu: the VCPU pointer
1078 * @run: the kvm_run structure
1080 * Any abort that gets to the host is almost guaranteed to be caused by a
1081 * missing second stage translation table entry, which can mean that either the
1082 * guest simply needs more memory and we must allocate an appropriate page or it
1083 * can mean that the guest tried to access I/O memory, which is emulated by user
1084 * space. The distinction is based on the IPA causing the fault and whether this
1085 * memory region has been registered as standard RAM by user space.
1087 int kvm_handle_guest_abort(struct kvm_vcpu *vcpu, struct kvm_run *run)
1089 unsigned long fault_status;
1090 phys_addr_t fault_ipa;
1091 struct kvm_memory_slot *memslot;
1093 bool is_iabt, write_fault, writable;
1097 is_iabt = kvm_vcpu_trap_is_iabt(vcpu);
1098 fault_ipa = kvm_vcpu_get_fault_ipa(vcpu);
1100 trace_kvm_guest_fault(*vcpu_pc(vcpu), kvm_vcpu_get_hsr(vcpu),
1101 kvm_vcpu_get_hfar(vcpu), fault_ipa);
1103 /* Check the stage-2 fault is trans. fault or write fault */
1104 fault_status = kvm_vcpu_trap_get_fault_type(vcpu);
1105 if (fault_status != FSC_FAULT && fault_status != FSC_PERM) {
1106 kvm_err("Unsupported FSC: EC=%#x xFSC=%#lx ESR_EL2=%#lx\n",
1107 kvm_vcpu_trap_get_class(vcpu),
1108 (unsigned long)kvm_vcpu_trap_get_fault(vcpu),
1109 (unsigned long)kvm_vcpu_get_hsr(vcpu));
1113 idx = srcu_read_lock(&vcpu->kvm->srcu);
1115 gfn = fault_ipa >> PAGE_SHIFT;
1116 memslot = gfn_to_memslot(vcpu->kvm, gfn);
1117 hva = gfn_to_hva_memslot_prot(memslot, gfn, &writable);
1118 write_fault = kvm_is_write_fault(vcpu);
1119 if (kvm_is_error_hva(hva) || (write_fault && !writable)) {
1121 /* Prefetch Abort on I/O address */
1122 kvm_inject_pabt(vcpu, kvm_vcpu_get_hfar(vcpu));
1128 * The IPA is reported as [MAX:12], so we need to
1129 * complement it with the bottom 12 bits from the
1130 * faulting VA. This is always 12 bits, irrespective
1133 fault_ipa |= kvm_vcpu_get_hfar(vcpu) & ((1 << 12) - 1);
1134 ret = io_mem_abort(vcpu, run, fault_ipa);
1138 /* Userspace should not be able to register out-of-bounds IPAs */
1139 VM_BUG_ON(fault_ipa >= KVM_PHYS_SIZE);
1141 ret = user_mem_abort(vcpu, fault_ipa, memslot, hva, fault_status);
1145 srcu_read_unlock(&vcpu->kvm->srcu, idx);
1149 static void handle_hva_to_gpa(struct kvm *kvm,
1150 unsigned long start,
1152 void (*handler)(struct kvm *kvm,
1153 gpa_t gpa, void *data),
1156 struct kvm_memslots *slots;
1157 struct kvm_memory_slot *memslot;
1159 slots = kvm_memslots(kvm);
1161 /* we only care about the pages that the guest sees */
1162 kvm_for_each_memslot(memslot, slots) {
1163 unsigned long hva_start, hva_end;
1166 hva_start = max(start, memslot->userspace_addr);
1167 hva_end = min(end, memslot->userspace_addr +
1168 (memslot->npages << PAGE_SHIFT));
1169 if (hva_start >= hva_end)
1173 * {gfn(page) | page intersects with [hva_start, hva_end)} =
1174 * {gfn_start, gfn_start+1, ..., gfn_end-1}.
1176 gfn = hva_to_gfn_memslot(hva_start, memslot);
1177 gfn_end = hva_to_gfn_memslot(hva_end + PAGE_SIZE - 1, memslot);
1179 for (; gfn < gfn_end; ++gfn) {
1180 gpa_t gpa = gfn << PAGE_SHIFT;
1181 handler(kvm, gpa, data);
1186 static void kvm_unmap_hva_handler(struct kvm *kvm, gpa_t gpa, void *data)
1188 unmap_stage2_range(kvm, gpa, PAGE_SIZE);
1191 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
1193 unsigned long end = hva + PAGE_SIZE;
1198 trace_kvm_unmap_hva(hva);
1199 handle_hva_to_gpa(kvm, hva, end, &kvm_unmap_hva_handler, NULL);
1203 int kvm_unmap_hva_range(struct kvm *kvm,
1204 unsigned long start, unsigned long end)
1209 trace_kvm_unmap_hva_range(start, end);
1210 handle_hva_to_gpa(kvm, start, end, &kvm_unmap_hva_handler, NULL);
1214 static void kvm_set_spte_handler(struct kvm *kvm, gpa_t gpa, void *data)
1216 pte_t *pte = (pte_t *)data;
1218 stage2_set_pte(kvm, NULL, gpa, pte, false);
1222 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
1224 unsigned long end = hva + PAGE_SIZE;
1230 trace_kvm_set_spte_hva(hva);
1231 stage2_pte = pfn_pte(pte_pfn(pte), PAGE_S2);
1232 handle_hva_to_gpa(kvm, hva, end, &kvm_set_spte_handler, &stage2_pte);
1235 void kvm_mmu_free_memory_caches(struct kvm_vcpu *vcpu)
1237 mmu_free_memory_cache(&vcpu->arch.mmu_page_cache);
1240 phys_addr_t kvm_mmu_get_httbr(void)
1242 return virt_to_phys(hyp_pgd);
1245 phys_addr_t kvm_mmu_get_boot_httbr(void)
1247 return virt_to_phys(boot_hyp_pgd);
1250 phys_addr_t kvm_get_idmap_vector(void)
1252 return hyp_idmap_vector;
1255 int kvm_mmu_init(void)
1259 hyp_idmap_start = kvm_virt_to_phys(__hyp_idmap_text_start);
1260 hyp_idmap_end = kvm_virt_to_phys(__hyp_idmap_text_end);
1261 hyp_idmap_vector = kvm_virt_to_phys(__kvm_hyp_init);
1263 if ((hyp_idmap_start ^ hyp_idmap_end) & PAGE_MASK) {
1265 * Our init code is crossing a page boundary. Allocate
1266 * a bounce page, copy the code over and use that.
1268 size_t len = __hyp_idmap_text_end - __hyp_idmap_text_start;
1269 phys_addr_t phys_base;
1271 init_bounce_page = (void *)__get_free_page(GFP_KERNEL);
1272 if (!init_bounce_page) {
1273 kvm_err("Couldn't allocate HYP init bounce page\n");
1278 memcpy(init_bounce_page, __hyp_idmap_text_start, len);
1280 * Warning: the code we just copied to the bounce page
1281 * must be flushed to the point of coherency.
1282 * Otherwise, the data may be sitting in L2, and HYP
1283 * mode won't be able to observe it as it runs with
1284 * caches off at that point.
1286 kvm_flush_dcache_to_poc(init_bounce_page, len);
1288 phys_base = kvm_virt_to_phys(init_bounce_page);
1289 hyp_idmap_vector += phys_base - hyp_idmap_start;
1290 hyp_idmap_start = phys_base;
1291 hyp_idmap_end = phys_base + len;
1293 kvm_info("Using HYP init bounce page @%lx\n",
1294 (unsigned long)phys_base);
1297 hyp_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, hyp_pgd_order);
1298 boot_hyp_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, hyp_pgd_order);
1300 if (!hyp_pgd || !boot_hyp_pgd) {
1301 kvm_err("Hyp mode PGD not allocated\n");
1306 /* Create the idmap in the boot page tables */
1307 err = __create_hyp_mappings(boot_hyp_pgd,
1308 hyp_idmap_start, hyp_idmap_end,
1309 __phys_to_pfn(hyp_idmap_start),
1313 kvm_err("Failed to idmap %lx-%lx\n",
1314 hyp_idmap_start, hyp_idmap_end);
1318 /* Map the very same page at the trampoline VA */
1319 err = __create_hyp_mappings(boot_hyp_pgd,
1320 TRAMPOLINE_VA, TRAMPOLINE_VA + PAGE_SIZE,
1321 __phys_to_pfn(hyp_idmap_start),
1324 kvm_err("Failed to map trampoline @%lx into boot HYP pgd\n",
1329 /* Map the same page again into the runtime page tables */
1330 err = __create_hyp_mappings(hyp_pgd,
1331 TRAMPOLINE_VA, TRAMPOLINE_VA + PAGE_SIZE,
1332 __phys_to_pfn(hyp_idmap_start),
1335 kvm_err("Failed to map trampoline @%lx into runtime HYP pgd\n",
1346 void kvm_arch_commit_memory_region(struct kvm *kvm,
1347 struct kvm_userspace_memory_region *mem,
1348 const struct kvm_memory_slot *old,
1349 enum kvm_mr_change change)
1353 int kvm_arch_prepare_memory_region(struct kvm *kvm,
1354 struct kvm_memory_slot *memslot,
1355 struct kvm_userspace_memory_region *mem,
1356 enum kvm_mr_change change)
1358 hva_t hva = mem->userspace_addr;
1359 hva_t reg_end = hva + mem->memory_size;
1360 bool writable = !(mem->flags & KVM_MEM_READONLY);
1363 if (change != KVM_MR_CREATE && change != KVM_MR_MOVE)
1367 * Prevent userspace from creating a memory region outside of the IPA
1368 * space addressable by the KVM guest IPA space.
1370 if (memslot->base_gfn + memslot->npages >=
1371 (KVM_PHYS_SIZE >> PAGE_SHIFT))
1375 * A memory region could potentially cover multiple VMAs, and any holes
1376 * between them, so iterate over all of them to find out if we can map
1377 * any of them right now.
1379 * +--------------------------------------------+
1380 * +---------------+----------------+ +----------------+
1381 * | : VMA 1 | VMA 2 | | VMA 3 : |
1382 * +---------------+----------------+ +----------------+
1384 * +--------------------------------------------+
1387 struct vm_area_struct *vma = find_vma(current->mm, hva);
1388 hva_t vm_start, vm_end;
1390 if (!vma || vma->vm_start >= reg_end)
1394 * Mapping a read-only VMA is only allowed if the
1395 * memory region is configured as read-only.
1397 if (writable && !(vma->vm_flags & VM_WRITE)) {
1403 * Take the intersection of this VMA with the memory region
1405 vm_start = max(hva, vma->vm_start);
1406 vm_end = min(reg_end, vma->vm_end);
1408 if (vma->vm_flags & VM_PFNMAP) {
1409 gpa_t gpa = mem->guest_phys_addr +
1410 (vm_start - mem->userspace_addr);
1411 phys_addr_t pa = (vma->vm_pgoff << PAGE_SHIFT) +
1412 vm_start - vma->vm_start;
1414 ret = kvm_phys_addr_ioremap(kvm, gpa, pa,
1421 } while (hva < reg_end);
1423 spin_lock(&kvm->mmu_lock);
1425 unmap_stage2_range(kvm, mem->guest_phys_addr, mem->memory_size);
1427 stage2_flush_memslot(kvm, memslot);
1428 spin_unlock(&kvm->mmu_lock);
1432 void kvm_arch_free_memslot(struct kvm *kvm, struct kvm_memory_slot *free,
1433 struct kvm_memory_slot *dont)
1437 int kvm_arch_create_memslot(struct kvm *kvm, struct kvm_memory_slot *slot,
1438 unsigned long npages)
1441 * Readonly memslots are not incoherent with the caches by definition,
1442 * but in practice, they are used mostly to emulate ROMs or NOR flashes
1443 * that the guest may consider devices and hence map as uncached.
1444 * To prevent incoherency issues in these cases, tag all readonly
1445 * regions as incoherent.
1447 if (slot->flags & KVM_MEM_READONLY)
1448 slot->flags |= KVM_MEMSLOT_INCOHERENT;
1452 void kvm_arch_memslots_updated(struct kvm *kvm)
1456 void kvm_arch_flush_shadow_all(struct kvm *kvm)
1460 void kvm_arch_flush_shadow_memslot(struct kvm *kvm,
1461 struct kvm_memory_slot *slot)
1463 gpa_t gpa = slot->base_gfn << PAGE_SHIFT;
1464 phys_addr_t size = slot->npages << PAGE_SHIFT;
1466 spin_lock(&kvm->mmu_lock);
1467 unmap_stage2_range(kvm, gpa, size);
1468 spin_unlock(&kvm->mmu_lock);
1472 * See note at ARMv7 ARM B1.14.4 (TL;DR: S/W ops are not easily virtualized).
1475 * - S/W ops are local to a CPU (not broadcast)
1476 * - We have line migration behind our back (speculation)
1477 * - System caches don't support S/W at all (damn!)
1479 * In the face of the above, the best we can do is to try and convert
1480 * S/W ops to VA ops. Because the guest is not allowed to infer the
1481 * S/W to PA mapping, it can only use S/W to nuke the whole cache,
1482 * which is a rather good thing for us.
1484 * Also, it is only used when turning caches on/off ("The expected
1485 * usage of the cache maintenance instructions that operate by set/way
1486 * is associated with the cache maintenance instructions associated
1487 * with the powerdown and powerup of caches, if this is required by
1488 * the implementation.").
1490 * We use the following policy:
1492 * - If we trap a S/W operation, we enable VM trapping to detect
1493 * caches being turned on/off, and do a full clean.
1495 * - We flush the caches on both caches being turned on and off.
1497 * - Once the caches are enabled, we stop trapping VM ops.
1499 void kvm_set_way_flush(struct kvm_vcpu *vcpu)
1501 unsigned long hcr = vcpu_get_hcr(vcpu);
1504 * If this is the first time we do a S/W operation
1505 * (i.e. HCR_TVM not set) flush the whole memory, and set the
1508 * Otherwise, rely on the VM trapping to wait for the MMU +
1509 * Caches to be turned off. At that point, we'll be able to
1510 * clean the caches again.
1512 if (!(hcr & HCR_TVM)) {
1513 trace_kvm_set_way_flush(*vcpu_pc(vcpu),
1514 vcpu_has_cache_enabled(vcpu));
1515 stage2_flush_vm(vcpu->kvm);
1516 vcpu_set_hcr(vcpu, hcr | HCR_TVM);
1520 void kvm_toggle_cache(struct kvm_vcpu *vcpu, bool was_enabled)
1522 bool now_enabled = vcpu_has_cache_enabled(vcpu);
1525 * If switching the MMU+caches on, need to invalidate the caches.
1526 * If switching it off, need to clean the caches.
1527 * Clean + invalidate does the trick always.
1529 if (now_enabled != was_enabled)
1530 stage2_flush_vm(vcpu->kvm);
1532 /* Caches are now on, stop trapping VM ops (until a S/W op) */
1534 vcpu_set_hcr(vcpu, vcpu_get_hcr(vcpu) & ~HCR_TVM);
1536 trace_kvm_toggle_cache(*vcpu_pc(vcpu), was_enabled, now_enabled);