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 S2_PGD_SIZE (PTRS_PER_S2_PGD * sizeof(pgd_t))
47 #define hyp_pgd_order get_order(PTRS_PER_PGD * sizeof(pgd_t))
49 #define KVM_S2PTE_FLAG_IS_IOMAP (1UL << 0)
50 #define KVM_S2_FLAG_LOGGING_ACTIVE (1UL << 1)
52 static bool memslot_is_logging(struct kvm_memory_slot *memslot)
54 return memslot->dirty_bitmap && !(memslot->flags & KVM_MEM_READONLY);
58 * kvm_flush_remote_tlbs() - flush all VM TLB entries for v7/8
59 * @kvm: pointer to kvm structure.
61 * Interface to HYP function to flush all VM TLB entries
63 void kvm_flush_remote_tlbs(struct kvm *kvm)
65 kvm_call_hyp(__kvm_tlb_flush_vmid, kvm);
68 static void kvm_tlb_flush_vmid_ipa(struct kvm *kvm, phys_addr_t ipa)
70 kvm_call_hyp(__kvm_tlb_flush_vmid_ipa, kvm, ipa);
74 * D-Cache management functions. They take the page table entries by
75 * value, as they are flushing the cache using the kernel mapping (or
78 static void kvm_flush_dcache_pte(pte_t pte)
80 __kvm_flush_dcache_pte(pte);
83 static void kvm_flush_dcache_pmd(pmd_t pmd)
85 __kvm_flush_dcache_pmd(pmd);
88 static void kvm_flush_dcache_pud(pud_t pud)
90 __kvm_flush_dcache_pud(pud);
93 static bool kvm_is_device_pfn(unsigned long pfn)
95 return !pfn_valid(pfn);
99 * stage2_dissolve_pmd() - clear and flush huge PMD entry
100 * @kvm: pointer to kvm structure.
102 * @pmd: pmd pointer for IPA
104 * Function clears a PMD entry, flushes addr 1st and 2nd stage TLBs. Marks all
105 * pages in the range dirty.
107 static void stage2_dissolve_pmd(struct kvm *kvm, phys_addr_t addr, pmd_t *pmd)
109 if (!pmd_thp_or_huge(*pmd))
113 kvm_tlb_flush_vmid_ipa(kvm, addr);
114 put_page(virt_to_page(pmd));
117 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
122 BUG_ON(max > KVM_NR_MEM_OBJS);
123 if (cache->nobjs >= min)
125 while (cache->nobjs < max) {
126 page = (void *)__get_free_page(PGALLOC_GFP);
129 cache->objects[cache->nobjs++] = page;
134 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
137 free_page((unsigned long)mc->objects[--mc->nobjs]);
140 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc)
144 BUG_ON(!mc || !mc->nobjs);
145 p = mc->objects[--mc->nobjs];
149 static void clear_stage2_pgd_entry(struct kvm *kvm, pgd_t *pgd, phys_addr_t addr)
151 pud_t *pud_table __maybe_unused = stage2_pud_offset(pgd, 0UL);
152 stage2_pgd_clear(pgd);
153 kvm_tlb_flush_vmid_ipa(kvm, addr);
154 stage2_pud_free(pud_table);
155 put_page(virt_to_page(pgd));
158 static void clear_stage2_pud_entry(struct kvm *kvm, pud_t *pud, phys_addr_t addr)
160 pmd_t *pmd_table __maybe_unused = stage2_pmd_offset(pud, 0);
161 VM_BUG_ON(stage2_pud_huge(*pud));
162 stage2_pud_clear(pud);
163 kvm_tlb_flush_vmid_ipa(kvm, addr);
164 stage2_pmd_free(pmd_table);
165 put_page(virt_to_page(pud));
168 static void clear_stage2_pmd_entry(struct kvm *kvm, pmd_t *pmd, phys_addr_t addr)
170 pte_t *pte_table = pte_offset_kernel(pmd, 0);
171 VM_BUG_ON(pmd_thp_or_huge(*pmd));
173 kvm_tlb_flush_vmid_ipa(kvm, addr);
174 pte_free_kernel(NULL, pte_table);
175 put_page(virt_to_page(pmd));
179 * Unmapping vs dcache management:
181 * If a guest maps certain memory pages as uncached, all writes will
182 * bypass the data cache and go directly to RAM. However, the CPUs
183 * can still speculate reads (not writes) and fill cache lines with
186 * Those cache lines will be *clean* cache lines though, so a
187 * clean+invalidate operation is equivalent to an invalidate
188 * operation, because no cache lines are marked dirty.
190 * Those clean cache lines could be filled prior to an uncached write
191 * by the guest, and the cache coherent IO subsystem would therefore
192 * end up writing old data to disk.
194 * This is why right after unmapping a page/section and invalidating
195 * the corresponding TLBs, we call kvm_flush_dcache_p*() to make sure
196 * the IO subsystem will never hit in the cache.
198 static void unmap_stage2_ptes(struct kvm *kvm, pmd_t *pmd,
199 phys_addr_t addr, phys_addr_t end)
201 phys_addr_t start_addr = addr;
202 pte_t *pte, *start_pte;
204 start_pte = pte = pte_offset_kernel(pmd, addr);
206 if (!pte_none(*pte)) {
207 pte_t old_pte = *pte;
209 kvm_set_pte(pte, __pte(0));
210 kvm_tlb_flush_vmid_ipa(kvm, addr);
212 /* No need to invalidate the cache for device mappings */
213 if (!kvm_is_device_pfn(pte_pfn(old_pte)))
214 kvm_flush_dcache_pte(old_pte);
216 put_page(virt_to_page(pte));
218 } while (pte++, addr += PAGE_SIZE, addr != end);
220 if (stage2_pte_table_empty(start_pte))
221 clear_stage2_pmd_entry(kvm, pmd, start_addr);
224 static void unmap_stage2_pmds(struct kvm *kvm, pud_t *pud,
225 phys_addr_t addr, phys_addr_t end)
227 phys_addr_t next, start_addr = addr;
228 pmd_t *pmd, *start_pmd;
230 start_pmd = pmd = stage2_pmd_offset(pud, addr);
232 next = stage2_pmd_addr_end(addr, end);
233 if (!pmd_none(*pmd)) {
234 if (pmd_thp_or_huge(*pmd)) {
235 pmd_t old_pmd = *pmd;
238 kvm_tlb_flush_vmid_ipa(kvm, addr);
240 kvm_flush_dcache_pmd(old_pmd);
242 put_page(virt_to_page(pmd));
244 unmap_stage2_ptes(kvm, pmd, addr, next);
247 } while (pmd++, addr = next, addr != end);
249 if (stage2_pmd_table_empty(start_pmd))
250 clear_stage2_pud_entry(kvm, pud, start_addr);
253 static void unmap_stage2_puds(struct kvm *kvm, pgd_t *pgd,
254 phys_addr_t addr, phys_addr_t end)
256 phys_addr_t next, start_addr = addr;
257 pud_t *pud, *start_pud;
259 start_pud = pud = stage2_pud_offset(pgd, addr);
261 next = stage2_pud_addr_end(addr, end);
262 if (!stage2_pud_none(*pud)) {
263 if (stage2_pud_huge(*pud)) {
264 pud_t old_pud = *pud;
266 stage2_pud_clear(pud);
267 kvm_tlb_flush_vmid_ipa(kvm, addr);
268 kvm_flush_dcache_pud(old_pud);
269 put_page(virt_to_page(pud));
271 unmap_stage2_pmds(kvm, pud, addr, next);
274 } while (pud++, addr = next, addr != end);
276 if (stage2_pud_table_empty(start_pud))
277 clear_stage2_pgd_entry(kvm, pgd, start_addr);
281 * unmap_stage2_range -- Clear stage2 page table entries to unmap a range
282 * @kvm: The VM pointer
283 * @start: The intermediate physical base address of the range to unmap
284 * @size: The size of the area to unmap
286 * Clear a range of stage-2 mappings, lowering the various ref-counts. Must
287 * be called while holding mmu_lock (unless for freeing the stage2 pgd before
288 * destroying the VM), otherwise another faulting VCPU may come in and mess
289 * with things behind our backs.
291 static void unmap_stage2_range(struct kvm *kvm, phys_addr_t start, u64 size)
294 phys_addr_t addr = start, end = start + size;
297 pgd = kvm->arch.pgd + stage2_pgd_index(addr);
299 next = stage2_pgd_addr_end(addr, end);
300 if (!stage2_pgd_none(*pgd))
301 unmap_stage2_puds(kvm, pgd, addr, next);
302 } while (pgd++, addr = next, addr != end);
305 static void stage2_flush_ptes(struct kvm *kvm, pmd_t *pmd,
306 phys_addr_t addr, phys_addr_t end)
310 pte = pte_offset_kernel(pmd, addr);
312 if (!pte_none(*pte) && !kvm_is_device_pfn(pte_pfn(*pte)))
313 kvm_flush_dcache_pte(*pte);
314 } while (pte++, addr += PAGE_SIZE, addr != end);
317 static void stage2_flush_pmds(struct kvm *kvm, pud_t *pud,
318 phys_addr_t addr, phys_addr_t end)
323 pmd = stage2_pmd_offset(pud, addr);
325 next = stage2_pmd_addr_end(addr, end);
326 if (!pmd_none(*pmd)) {
327 if (pmd_thp_or_huge(*pmd))
328 kvm_flush_dcache_pmd(*pmd);
330 stage2_flush_ptes(kvm, pmd, addr, next);
332 } while (pmd++, addr = next, addr != end);
335 static void stage2_flush_puds(struct kvm *kvm, pgd_t *pgd,
336 phys_addr_t addr, phys_addr_t end)
341 pud = stage2_pud_offset(pgd, addr);
343 next = stage2_pud_addr_end(addr, end);
344 if (!stage2_pud_none(*pud)) {
345 if (stage2_pud_huge(*pud))
346 kvm_flush_dcache_pud(*pud);
348 stage2_flush_pmds(kvm, pud, addr, next);
350 } while (pud++, addr = next, addr != end);
353 static void stage2_flush_memslot(struct kvm *kvm,
354 struct kvm_memory_slot *memslot)
356 phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
357 phys_addr_t end = addr + PAGE_SIZE * memslot->npages;
361 pgd = kvm->arch.pgd + stage2_pgd_index(addr);
363 next = stage2_pgd_addr_end(addr, end);
364 stage2_flush_puds(kvm, pgd, addr, next);
365 } while (pgd++, addr = next, addr != end);
369 * stage2_flush_vm - Invalidate cache for pages mapped in stage 2
370 * @kvm: The struct kvm pointer
372 * Go through the stage 2 page tables and invalidate any cache lines
373 * backing memory already mapped to the VM.
375 static void stage2_flush_vm(struct kvm *kvm)
377 struct kvm_memslots *slots;
378 struct kvm_memory_slot *memslot;
381 idx = srcu_read_lock(&kvm->srcu);
382 spin_lock(&kvm->mmu_lock);
384 slots = kvm_memslots(kvm);
385 kvm_for_each_memslot(memslot, slots)
386 stage2_flush_memslot(kvm, memslot);
388 spin_unlock(&kvm->mmu_lock);
389 srcu_read_unlock(&kvm->srcu, idx);
392 static void clear_hyp_pgd_entry(pgd_t *pgd)
394 pud_t *pud_table __maybe_unused = pud_offset(pgd, 0UL);
396 pud_free(NULL, pud_table);
397 put_page(virt_to_page(pgd));
400 static void clear_hyp_pud_entry(pud_t *pud)
402 pmd_t *pmd_table __maybe_unused = pmd_offset(pud, 0);
403 VM_BUG_ON(pud_huge(*pud));
405 pmd_free(NULL, pmd_table);
406 put_page(virt_to_page(pud));
409 static void clear_hyp_pmd_entry(pmd_t *pmd)
411 pte_t *pte_table = pte_offset_kernel(pmd, 0);
412 VM_BUG_ON(pmd_thp_or_huge(*pmd));
414 pte_free_kernel(NULL, pte_table);
415 put_page(virt_to_page(pmd));
418 static void unmap_hyp_ptes(pmd_t *pmd, phys_addr_t addr, phys_addr_t end)
420 pte_t *pte, *start_pte;
422 start_pte = pte = pte_offset_kernel(pmd, addr);
424 if (!pte_none(*pte)) {
425 kvm_set_pte(pte, __pte(0));
426 put_page(virt_to_page(pte));
428 } while (pte++, addr += PAGE_SIZE, addr != end);
430 if (hyp_pte_table_empty(start_pte))
431 clear_hyp_pmd_entry(pmd);
434 static void unmap_hyp_pmds(pud_t *pud, phys_addr_t addr, phys_addr_t end)
437 pmd_t *pmd, *start_pmd;
439 start_pmd = pmd = pmd_offset(pud, addr);
441 next = pmd_addr_end(addr, end);
442 /* Hyp doesn't use huge pmds */
444 unmap_hyp_ptes(pmd, addr, next);
445 } while (pmd++, addr = next, addr != end);
447 if (hyp_pmd_table_empty(start_pmd))
448 clear_hyp_pud_entry(pud);
451 static void unmap_hyp_puds(pgd_t *pgd, phys_addr_t addr, phys_addr_t end)
454 pud_t *pud, *start_pud;
456 start_pud = pud = pud_offset(pgd, addr);
458 next = pud_addr_end(addr, end);
459 /* Hyp doesn't use huge puds */
461 unmap_hyp_pmds(pud, addr, next);
462 } while (pud++, addr = next, addr != end);
464 if (hyp_pud_table_empty(start_pud))
465 clear_hyp_pgd_entry(pgd);
468 static void unmap_hyp_range(pgd_t *pgdp, phys_addr_t start, u64 size)
471 phys_addr_t addr = start, end = start + size;
475 * We don't unmap anything from HYP, except at the hyp tear down.
476 * Hence, we don't have to invalidate the TLBs here.
478 pgd = pgdp + pgd_index(addr);
480 next = pgd_addr_end(addr, end);
482 unmap_hyp_puds(pgd, addr, next);
483 } while (pgd++, addr = next, addr != end);
487 * free_boot_hyp_pgd - free HYP boot page tables
489 * Free the HYP boot page tables. The bounce page is also freed.
491 void free_boot_hyp_pgd(void)
493 mutex_lock(&kvm_hyp_pgd_mutex);
496 unmap_hyp_range(boot_hyp_pgd, hyp_idmap_start, PAGE_SIZE);
497 unmap_hyp_range(boot_hyp_pgd, TRAMPOLINE_VA, PAGE_SIZE);
498 free_pages((unsigned long)boot_hyp_pgd, hyp_pgd_order);
503 unmap_hyp_range(hyp_pgd, TRAMPOLINE_VA, PAGE_SIZE);
505 mutex_unlock(&kvm_hyp_pgd_mutex);
509 * free_hyp_pgds - free Hyp-mode page tables
511 * Assumes hyp_pgd is a page table used strictly in Hyp-mode and
512 * therefore contains either mappings in the kernel memory area (above
513 * PAGE_OFFSET), or device mappings in the vmalloc range (from
514 * VMALLOC_START to VMALLOC_END).
516 * boot_hyp_pgd should only map two pages for the init code.
518 void free_hyp_pgds(void)
524 mutex_lock(&kvm_hyp_pgd_mutex);
527 for (addr = PAGE_OFFSET; virt_addr_valid(addr); addr += PGDIR_SIZE)
528 unmap_hyp_range(hyp_pgd, KERN_TO_HYP(addr), PGDIR_SIZE);
529 for (addr = VMALLOC_START; is_vmalloc_addr((void*)addr); addr += PGDIR_SIZE)
530 unmap_hyp_range(hyp_pgd, KERN_TO_HYP(addr), PGDIR_SIZE);
532 free_pages((unsigned long)hyp_pgd, hyp_pgd_order);
535 if (merged_hyp_pgd) {
536 clear_page(merged_hyp_pgd);
537 free_page((unsigned long)merged_hyp_pgd);
538 merged_hyp_pgd = NULL;
541 mutex_unlock(&kvm_hyp_pgd_mutex);
544 static void create_hyp_pte_mappings(pmd_t *pmd, unsigned long start,
545 unsigned long end, unsigned long pfn,
553 pte = pte_offset_kernel(pmd, addr);
554 kvm_set_pte(pte, pfn_pte(pfn, prot));
555 get_page(virt_to_page(pte));
556 kvm_flush_dcache_to_poc(pte, sizeof(*pte));
558 } while (addr += PAGE_SIZE, addr != end);
561 static int create_hyp_pmd_mappings(pud_t *pud, unsigned long start,
562 unsigned long end, unsigned long pfn,
567 unsigned long addr, next;
571 pmd = pmd_offset(pud, addr);
573 BUG_ON(pmd_sect(*pmd));
575 if (pmd_none(*pmd)) {
576 pte = pte_alloc_one_kernel(NULL, addr);
578 kvm_err("Cannot allocate Hyp pte\n");
581 pmd_populate_kernel(NULL, pmd, pte);
582 get_page(virt_to_page(pmd));
583 kvm_flush_dcache_to_poc(pmd, sizeof(*pmd));
586 next = pmd_addr_end(addr, end);
588 create_hyp_pte_mappings(pmd, addr, next, pfn, prot);
589 pfn += (next - addr) >> PAGE_SHIFT;
590 } while (addr = next, addr != end);
595 static int create_hyp_pud_mappings(pgd_t *pgd, unsigned long start,
596 unsigned long end, unsigned long pfn,
601 unsigned long addr, next;
606 pud = pud_offset(pgd, addr);
608 if (pud_none_or_clear_bad(pud)) {
609 pmd = pmd_alloc_one(NULL, addr);
611 kvm_err("Cannot allocate Hyp pmd\n");
614 pud_populate(NULL, pud, pmd);
615 get_page(virt_to_page(pud));
616 kvm_flush_dcache_to_poc(pud, sizeof(*pud));
619 next = pud_addr_end(addr, end);
620 ret = create_hyp_pmd_mappings(pud, addr, next, pfn, prot);
623 pfn += (next - addr) >> PAGE_SHIFT;
624 } while (addr = next, addr != end);
629 static int __create_hyp_mappings(pgd_t *pgdp,
630 unsigned long start, unsigned long end,
631 unsigned long pfn, pgprot_t prot)
635 unsigned long addr, next;
638 mutex_lock(&kvm_hyp_pgd_mutex);
639 addr = start & PAGE_MASK;
640 end = PAGE_ALIGN(end);
642 pgd = pgdp + pgd_index(addr);
644 if (pgd_none(*pgd)) {
645 pud = pud_alloc_one(NULL, addr);
647 kvm_err("Cannot allocate Hyp pud\n");
651 pgd_populate(NULL, pgd, pud);
652 get_page(virt_to_page(pgd));
653 kvm_flush_dcache_to_poc(pgd, sizeof(*pgd));
656 next = pgd_addr_end(addr, end);
657 err = create_hyp_pud_mappings(pgd, addr, next, pfn, prot);
660 pfn += (next - addr) >> PAGE_SHIFT;
661 } while (addr = next, addr != end);
663 mutex_unlock(&kvm_hyp_pgd_mutex);
667 static phys_addr_t kvm_kaddr_to_phys(void *kaddr)
669 if (!is_vmalloc_addr(kaddr)) {
670 BUG_ON(!virt_addr_valid(kaddr));
673 return page_to_phys(vmalloc_to_page(kaddr)) +
674 offset_in_page(kaddr);
679 * create_hyp_mappings - duplicate a kernel virtual address range in Hyp mode
680 * @from: The virtual kernel start address of the range
681 * @to: The virtual kernel end address of the range (exclusive)
683 * The same virtual address as the kernel virtual address is also used
684 * in Hyp-mode mapping (modulo HYP_PAGE_OFFSET) to the same underlying
687 int create_hyp_mappings(void *from, void *to)
689 phys_addr_t phys_addr;
690 unsigned long virt_addr;
691 unsigned long start = KERN_TO_HYP((unsigned long)from);
692 unsigned long end = KERN_TO_HYP((unsigned long)to);
694 if (is_kernel_in_hyp_mode())
697 start = start & PAGE_MASK;
698 end = PAGE_ALIGN(end);
700 for (virt_addr = start; virt_addr < end; virt_addr += PAGE_SIZE) {
703 phys_addr = kvm_kaddr_to_phys(from + virt_addr - start);
704 err = __create_hyp_mappings(hyp_pgd, virt_addr,
705 virt_addr + PAGE_SIZE,
706 __phys_to_pfn(phys_addr),
716 * create_hyp_io_mappings - duplicate a kernel IO mapping into Hyp mode
717 * @from: The kernel start VA of the range
718 * @to: The kernel end VA of the range (exclusive)
719 * @phys_addr: The physical start address which gets mapped
721 * The resulting HYP VA is the same as the kernel VA, modulo
724 int create_hyp_io_mappings(void *from, void *to, phys_addr_t phys_addr)
726 unsigned long start = KERN_TO_HYP((unsigned long)from);
727 unsigned long end = KERN_TO_HYP((unsigned long)to);
729 if (is_kernel_in_hyp_mode())
732 /* Check for a valid kernel IO mapping */
733 if (!is_vmalloc_addr(from) || !is_vmalloc_addr(to - 1))
736 return __create_hyp_mappings(hyp_pgd, start, end,
737 __phys_to_pfn(phys_addr), PAGE_HYP_DEVICE);
741 * kvm_alloc_stage2_pgd - allocate level-1 table for stage-2 translation.
742 * @kvm: The KVM struct pointer for the VM.
744 * Allocates only the stage-2 HW PGD level table(s) (can support either full
745 * 40-bit input addresses or limited to 32-bit input addresses). Clears the
748 * Note we don't need locking here as this is only called when the VM is
749 * created, which can only be done once.
751 int kvm_alloc_stage2_pgd(struct kvm *kvm)
755 if (kvm->arch.pgd != NULL) {
756 kvm_err("kvm_arch already initialized?\n");
760 /* Allocate the HW PGD, making sure that each page gets its own refcount */
761 pgd = alloc_pages_exact(S2_PGD_SIZE, GFP_KERNEL | __GFP_ZERO);
770 static void stage2_unmap_memslot(struct kvm *kvm,
771 struct kvm_memory_slot *memslot)
773 hva_t hva = memslot->userspace_addr;
774 phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
775 phys_addr_t size = PAGE_SIZE * memslot->npages;
776 hva_t reg_end = hva + size;
779 * A memory region could potentially cover multiple VMAs, and any holes
780 * between them, so iterate over all of them to find out if we should
783 * +--------------------------------------------+
784 * +---------------+----------------+ +----------------+
785 * | : VMA 1 | VMA 2 | | VMA 3 : |
786 * +---------------+----------------+ +----------------+
788 * +--------------------------------------------+
791 struct vm_area_struct *vma = find_vma(current->mm, hva);
792 hva_t vm_start, vm_end;
794 if (!vma || vma->vm_start >= reg_end)
798 * Take the intersection of this VMA with the memory region
800 vm_start = max(hva, vma->vm_start);
801 vm_end = min(reg_end, vma->vm_end);
803 if (!(vma->vm_flags & VM_PFNMAP)) {
804 gpa_t gpa = addr + (vm_start - memslot->userspace_addr);
805 unmap_stage2_range(kvm, gpa, vm_end - vm_start);
808 } while (hva < reg_end);
812 * stage2_unmap_vm - Unmap Stage-2 RAM mappings
813 * @kvm: The struct kvm pointer
815 * Go through the memregions and unmap any reguler RAM
816 * backing memory already mapped to the VM.
818 void stage2_unmap_vm(struct kvm *kvm)
820 struct kvm_memslots *slots;
821 struct kvm_memory_slot *memslot;
824 idx = srcu_read_lock(&kvm->srcu);
825 spin_lock(&kvm->mmu_lock);
827 slots = kvm_memslots(kvm);
828 kvm_for_each_memslot(memslot, slots)
829 stage2_unmap_memslot(kvm, memslot);
831 spin_unlock(&kvm->mmu_lock);
832 srcu_read_unlock(&kvm->srcu, idx);
836 * kvm_free_stage2_pgd - free all stage-2 tables
837 * @kvm: The KVM struct pointer for the VM.
839 * Walks the level-1 page table pointed to by kvm->arch.pgd and frees all
840 * underlying level-2 and level-3 tables before freeing the actual level-1 table
841 * and setting the struct pointer to NULL.
843 * Note we don't need locking here as this is only called when the VM is
844 * destroyed, which can only be done once.
846 void kvm_free_stage2_pgd(struct kvm *kvm)
848 if (kvm->arch.pgd == NULL)
851 unmap_stage2_range(kvm, 0, KVM_PHYS_SIZE);
852 /* Free the HW pgd, one page at a time */
853 free_pages_exact(kvm->arch.pgd, S2_PGD_SIZE);
854 kvm->arch.pgd = NULL;
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);
958 pmd_populate_kernel(NULL, pmd, pte);
959 get_page(virt_to_page(pmd));
962 pte = pte_offset_kernel(pmd, addr);
964 if (iomap && pte_present(*pte))
967 /* Create 2nd stage page table mapping - Level 3 */
969 if (pte_present(old_pte)) {
970 kvm_set_pte(pte, __pte(0));
971 kvm_tlb_flush_vmid_ipa(kvm, addr);
973 get_page(virt_to_page(pte));
976 kvm_set_pte(pte, *new_pte);
981 * kvm_phys_addr_ioremap - map a device range to guest IPA
983 * @kvm: The KVM pointer
984 * @guest_ipa: The IPA at which to insert the mapping
985 * @pa: The physical address of the device
986 * @size: The size of the mapping
988 int kvm_phys_addr_ioremap(struct kvm *kvm, phys_addr_t guest_ipa,
989 phys_addr_t pa, unsigned long size, bool writable)
991 phys_addr_t addr, end;
994 struct kvm_mmu_memory_cache cache = { 0, };
996 end = (guest_ipa + size + PAGE_SIZE - 1) & PAGE_MASK;
997 pfn = __phys_to_pfn(pa);
999 for (addr = guest_ipa; addr < end; addr += PAGE_SIZE) {
1000 pte_t pte = pfn_pte(pfn, PAGE_S2_DEVICE);
1003 kvm_set_s2pte_writable(&pte);
1005 ret = mmu_topup_memory_cache(&cache, KVM_MMU_CACHE_MIN_PAGES,
1009 spin_lock(&kvm->mmu_lock);
1010 ret = stage2_set_pte(kvm, &cache, addr, &pte,
1011 KVM_S2PTE_FLAG_IS_IOMAP);
1012 spin_unlock(&kvm->mmu_lock);
1020 mmu_free_memory_cache(&cache);
1024 static bool transparent_hugepage_adjust(kvm_pfn_t *pfnp, phys_addr_t *ipap)
1026 kvm_pfn_t pfn = *pfnp;
1027 gfn_t gfn = *ipap >> PAGE_SHIFT;
1029 if (PageTransCompound(pfn_to_page(pfn))) {
1032 * The address we faulted on is backed by a transparent huge
1033 * page. However, because we map the compound huge page and
1034 * not the individual tail page, we need to transfer the
1035 * refcount to the head page. We have to be careful that the
1036 * THP doesn't start to split while we are adjusting the
1039 * We are sure this doesn't happen, because mmu_notifier_retry
1040 * was successful and we are holding the mmu_lock, so if this
1041 * THP is trying to split, it will be blocked in the mmu
1042 * notifier before touching any of the pages, specifically
1043 * before being able to call __split_huge_page_refcount().
1045 * We can therefore safely transfer the refcount from PG_tail
1046 * to PG_head and switch the pfn from a tail page to the head
1049 mask = PTRS_PER_PMD - 1;
1050 VM_BUG_ON((gfn & mask) != (pfn & mask));
1053 kvm_release_pfn_clean(pfn);
1065 static bool kvm_is_write_fault(struct kvm_vcpu *vcpu)
1067 if (kvm_vcpu_trap_is_iabt(vcpu))
1070 return kvm_vcpu_dabt_iswrite(vcpu);
1074 * stage2_wp_ptes - write protect PMD range
1075 * @pmd: pointer to pmd entry
1076 * @addr: range start address
1077 * @end: range end address
1079 static void stage2_wp_ptes(pmd_t *pmd, phys_addr_t addr, phys_addr_t end)
1083 pte = pte_offset_kernel(pmd, addr);
1085 if (!pte_none(*pte)) {
1086 if (!kvm_s2pte_readonly(pte))
1087 kvm_set_s2pte_readonly(pte);
1089 } while (pte++, addr += PAGE_SIZE, addr != end);
1093 * stage2_wp_pmds - write protect PUD range
1094 * @pud: pointer to pud entry
1095 * @addr: range start address
1096 * @end: range end address
1098 static void stage2_wp_pmds(pud_t *pud, phys_addr_t addr, phys_addr_t end)
1103 pmd = stage2_pmd_offset(pud, addr);
1106 next = stage2_pmd_addr_end(addr, end);
1107 if (!pmd_none(*pmd)) {
1108 if (pmd_thp_or_huge(*pmd)) {
1109 if (!kvm_s2pmd_readonly(pmd))
1110 kvm_set_s2pmd_readonly(pmd);
1112 stage2_wp_ptes(pmd, addr, next);
1115 } while (pmd++, addr = next, addr != end);
1119 * stage2_wp_puds - write protect PGD range
1120 * @pgd: pointer to pgd entry
1121 * @addr: range start address
1122 * @end: range end address
1124 * Process PUD entries, for a huge PUD we cause a panic.
1126 static void stage2_wp_puds(pgd_t *pgd, phys_addr_t addr, phys_addr_t end)
1131 pud = stage2_pud_offset(pgd, addr);
1133 next = stage2_pud_addr_end(addr, end);
1134 if (!stage2_pud_none(*pud)) {
1135 /* TODO:PUD not supported, revisit later if supported */
1136 BUG_ON(stage2_pud_huge(*pud));
1137 stage2_wp_pmds(pud, addr, next);
1139 } while (pud++, addr = next, addr != end);
1143 * stage2_wp_range() - write protect stage2 memory region range
1144 * @kvm: The KVM pointer
1145 * @addr: Start address of range
1146 * @end: End address of range
1148 static void stage2_wp_range(struct kvm *kvm, phys_addr_t addr, phys_addr_t end)
1153 pgd = kvm->arch.pgd + stage2_pgd_index(addr);
1156 * Release kvm_mmu_lock periodically if the memory region is
1157 * large. Otherwise, we may see kernel panics with
1158 * CONFIG_DETECT_HUNG_TASK, CONFIG_LOCKUP_DETECTOR,
1159 * CONFIG_LOCKDEP. Additionally, holding the lock too long
1160 * will also starve other vCPUs.
1162 if (need_resched() || spin_needbreak(&kvm->mmu_lock))
1163 cond_resched_lock(&kvm->mmu_lock);
1165 next = stage2_pgd_addr_end(addr, end);
1166 if (stage2_pgd_present(*pgd))
1167 stage2_wp_puds(pgd, addr, next);
1168 } while (pgd++, addr = next, addr != end);
1172 * kvm_mmu_wp_memory_region() - write protect stage 2 entries for memory slot
1173 * @kvm: The KVM pointer
1174 * @slot: The memory slot to write protect
1176 * Called to start logging dirty pages after memory region
1177 * KVM_MEM_LOG_DIRTY_PAGES operation is called. After this function returns
1178 * all present PMD and PTEs are write protected in the memory region.
1179 * Afterwards read of dirty page log can be called.
1181 * Acquires kvm_mmu_lock. Called with kvm->slots_lock mutex acquired,
1182 * serializing operations for VM memory regions.
1184 void kvm_mmu_wp_memory_region(struct kvm *kvm, int slot)
1186 struct kvm_memslots *slots = kvm_memslots(kvm);
1187 struct kvm_memory_slot *memslot = id_to_memslot(slots, slot);
1188 phys_addr_t start = memslot->base_gfn << PAGE_SHIFT;
1189 phys_addr_t end = (memslot->base_gfn + memslot->npages) << PAGE_SHIFT;
1191 spin_lock(&kvm->mmu_lock);
1192 stage2_wp_range(kvm, start, end);
1193 spin_unlock(&kvm->mmu_lock);
1194 kvm_flush_remote_tlbs(kvm);
1198 * kvm_mmu_write_protect_pt_masked() - write protect dirty pages
1199 * @kvm: The KVM pointer
1200 * @slot: The memory slot associated with mask
1201 * @gfn_offset: The gfn offset in memory slot
1202 * @mask: The mask of dirty pages at offset 'gfn_offset' in this memory
1203 * slot to be write protected
1205 * Walks bits set in mask write protects the associated pte's. Caller must
1206 * acquire kvm_mmu_lock.
1208 static void kvm_mmu_write_protect_pt_masked(struct kvm *kvm,
1209 struct kvm_memory_slot *slot,
1210 gfn_t gfn_offset, unsigned long mask)
1212 phys_addr_t base_gfn = slot->base_gfn + gfn_offset;
1213 phys_addr_t start = (base_gfn + __ffs(mask)) << PAGE_SHIFT;
1214 phys_addr_t end = (base_gfn + __fls(mask) + 1) << PAGE_SHIFT;
1216 stage2_wp_range(kvm, start, end);
1220 * kvm_arch_mmu_enable_log_dirty_pt_masked - enable dirty logging for selected
1223 * It calls kvm_mmu_write_protect_pt_masked to write protect selected pages to
1224 * enable dirty logging for them.
1226 void kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm *kvm,
1227 struct kvm_memory_slot *slot,
1228 gfn_t gfn_offset, unsigned long mask)
1230 kvm_mmu_write_protect_pt_masked(kvm, slot, gfn_offset, mask);
1233 static void coherent_cache_guest_page(struct kvm_vcpu *vcpu, kvm_pfn_t pfn,
1234 unsigned long size, bool uncached)
1236 __coherent_cache_guest_page(vcpu, pfn, size, uncached);
1239 static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
1240 struct kvm_memory_slot *memslot, unsigned long hva,
1241 unsigned long fault_status)
1244 bool write_fault, writable, hugetlb = false, force_pte = false;
1245 unsigned long mmu_seq;
1246 gfn_t gfn = fault_ipa >> PAGE_SHIFT;
1247 struct kvm *kvm = vcpu->kvm;
1248 struct kvm_mmu_memory_cache *memcache = &vcpu->arch.mmu_page_cache;
1249 struct vm_area_struct *vma;
1251 pgprot_t mem_type = PAGE_S2;
1252 bool fault_ipa_uncached;
1253 bool logging_active = memslot_is_logging(memslot);
1254 unsigned long flags = 0;
1256 write_fault = kvm_is_write_fault(vcpu);
1257 if (fault_status == FSC_PERM && !write_fault) {
1258 kvm_err("Unexpected L2 read permission error\n");
1262 /* Let's check if we will get back a huge page backed by hugetlbfs */
1263 down_read(¤t->mm->mmap_sem);
1264 vma = find_vma_intersection(current->mm, hva, hva + 1);
1265 if (unlikely(!vma)) {
1266 kvm_err("Failed to find VMA for hva 0x%lx\n", hva);
1267 up_read(¤t->mm->mmap_sem);
1271 if (is_vm_hugetlb_page(vma) && !logging_active) {
1273 gfn = (fault_ipa & PMD_MASK) >> PAGE_SHIFT;
1276 * Pages belonging to memslots that don't have the same
1277 * alignment for userspace and IPA cannot be mapped using
1278 * block descriptors even if the pages belong to a THP for
1279 * the process, because the stage-2 block descriptor will
1280 * cover more than a single THP and we loose atomicity for
1281 * unmapping, updates, and splits of the THP or other pages
1282 * in the stage-2 block range.
1284 if ((memslot->userspace_addr & ~PMD_MASK) !=
1285 ((memslot->base_gfn << PAGE_SHIFT) & ~PMD_MASK))
1288 up_read(¤t->mm->mmap_sem);
1290 /* We need minimum second+third level pages */
1291 ret = mmu_topup_memory_cache(memcache, KVM_MMU_CACHE_MIN_PAGES,
1296 mmu_seq = vcpu->kvm->mmu_notifier_seq;
1298 * Ensure the read of mmu_notifier_seq happens before we call
1299 * gfn_to_pfn_prot (which calls get_user_pages), so that we don't risk
1300 * the page we just got a reference to gets unmapped before we have a
1301 * chance to grab the mmu_lock, which ensure that if the page gets
1302 * unmapped afterwards, the call to kvm_unmap_hva will take it away
1303 * from us again properly. This smp_rmb() interacts with the smp_wmb()
1304 * in kvm_mmu_notifier_invalidate_<page|range_end>.
1308 pfn = gfn_to_pfn_prot(kvm, gfn, write_fault, &writable);
1309 if (is_error_pfn(pfn))
1312 if (kvm_is_device_pfn(pfn)) {
1313 mem_type = PAGE_S2_DEVICE;
1314 flags |= KVM_S2PTE_FLAG_IS_IOMAP;
1315 } else if (logging_active) {
1317 * Faults on pages in a memslot with logging enabled
1318 * should not be mapped with huge pages (it introduces churn
1319 * and performance degradation), so force a pte mapping.
1322 flags |= KVM_S2_FLAG_LOGGING_ACTIVE;
1325 * Only actually map the page as writable if this was a write
1332 spin_lock(&kvm->mmu_lock);
1333 if (mmu_notifier_retry(kvm, mmu_seq))
1336 if (!hugetlb && !force_pte)
1337 hugetlb = transparent_hugepage_adjust(&pfn, &fault_ipa);
1339 fault_ipa_uncached = memslot->flags & KVM_MEMSLOT_INCOHERENT;
1342 pmd_t new_pmd = pfn_pmd(pfn, mem_type);
1343 new_pmd = pmd_mkhuge(new_pmd);
1345 kvm_set_s2pmd_writable(&new_pmd);
1346 kvm_set_pfn_dirty(pfn);
1348 coherent_cache_guest_page(vcpu, pfn, PMD_SIZE, fault_ipa_uncached);
1349 ret = stage2_set_pmd_huge(kvm, memcache, fault_ipa, &new_pmd);
1351 pte_t new_pte = pfn_pte(pfn, mem_type);
1354 kvm_set_s2pte_writable(&new_pte);
1355 kvm_set_pfn_dirty(pfn);
1356 mark_page_dirty(kvm, gfn);
1358 coherent_cache_guest_page(vcpu, pfn, PAGE_SIZE, fault_ipa_uncached);
1359 ret = stage2_set_pte(kvm, memcache, fault_ipa, &new_pte, flags);
1363 spin_unlock(&kvm->mmu_lock);
1364 kvm_set_pfn_accessed(pfn);
1365 kvm_release_pfn_clean(pfn);
1370 * Resolve the access fault by making the page young again.
1371 * Note that because the faulting entry is guaranteed not to be
1372 * cached in the TLB, we don't need to invalidate anything.
1374 static void handle_access_fault(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa)
1379 bool pfn_valid = false;
1381 trace_kvm_access_fault(fault_ipa);
1383 spin_lock(&vcpu->kvm->mmu_lock);
1385 pmd = stage2_get_pmd(vcpu->kvm, NULL, fault_ipa);
1386 if (!pmd || pmd_none(*pmd)) /* Nothing there */
1389 if (pmd_thp_or_huge(*pmd)) { /* THP, HugeTLB */
1390 *pmd = pmd_mkyoung(*pmd);
1391 pfn = pmd_pfn(*pmd);
1396 pte = pte_offset_kernel(pmd, fault_ipa);
1397 if (pte_none(*pte)) /* Nothing there either */
1400 *pte = pte_mkyoung(*pte); /* Just a page... */
1401 pfn = pte_pfn(*pte);
1404 spin_unlock(&vcpu->kvm->mmu_lock);
1406 kvm_set_pfn_accessed(pfn);
1410 * kvm_handle_guest_abort - handles all 2nd stage aborts
1411 * @vcpu: the VCPU pointer
1412 * @run: the kvm_run structure
1414 * Any abort that gets to the host is almost guaranteed to be caused by a
1415 * missing second stage translation table entry, which can mean that either the
1416 * guest simply needs more memory and we must allocate an appropriate page or it
1417 * can mean that the guest tried to access I/O memory, which is emulated by user
1418 * space. The distinction is based on the IPA causing the fault and whether this
1419 * memory region has been registered as standard RAM by user space.
1421 int kvm_handle_guest_abort(struct kvm_vcpu *vcpu, struct kvm_run *run)
1423 unsigned long fault_status;
1424 phys_addr_t fault_ipa;
1425 struct kvm_memory_slot *memslot;
1427 bool is_iabt, write_fault, writable;
1431 is_iabt = kvm_vcpu_trap_is_iabt(vcpu);
1432 fault_ipa = kvm_vcpu_get_fault_ipa(vcpu);
1434 trace_kvm_guest_fault(*vcpu_pc(vcpu), kvm_vcpu_get_hsr(vcpu),
1435 kvm_vcpu_get_hfar(vcpu), fault_ipa);
1437 /* Check the stage-2 fault is trans. fault or write fault */
1438 fault_status = kvm_vcpu_trap_get_fault_type(vcpu);
1439 if (fault_status != FSC_FAULT && fault_status != FSC_PERM &&
1440 fault_status != FSC_ACCESS) {
1441 kvm_err("Unsupported FSC: EC=%#x xFSC=%#lx ESR_EL2=%#lx\n",
1442 kvm_vcpu_trap_get_class(vcpu),
1443 (unsigned long)kvm_vcpu_trap_get_fault(vcpu),
1444 (unsigned long)kvm_vcpu_get_hsr(vcpu));
1448 idx = srcu_read_lock(&vcpu->kvm->srcu);
1450 gfn = fault_ipa >> PAGE_SHIFT;
1451 memslot = gfn_to_memslot(vcpu->kvm, gfn);
1452 hva = gfn_to_hva_memslot_prot(memslot, gfn, &writable);
1453 write_fault = kvm_is_write_fault(vcpu);
1454 if (kvm_is_error_hva(hva) || (write_fault && !writable)) {
1456 /* Prefetch Abort on I/O address */
1457 kvm_inject_pabt(vcpu, kvm_vcpu_get_hfar(vcpu));
1463 * Check for a cache maintenance operation. Since we
1464 * ended-up here, we know it is outside of any memory
1465 * slot. But we can't find out if that is for a device,
1466 * or if the guest is just being stupid. The only thing
1467 * we know for sure is that this range cannot be cached.
1469 * So let's assume that the guest is just being
1470 * cautious, and skip the instruction.
1472 if (kvm_vcpu_dabt_is_cm(vcpu)) {
1473 kvm_skip_instr(vcpu, kvm_vcpu_trap_il_is32bit(vcpu));
1479 * The IPA is reported as [MAX:12], so we need to
1480 * complement it with the bottom 12 bits from the
1481 * faulting VA. This is always 12 bits, irrespective
1484 fault_ipa |= kvm_vcpu_get_hfar(vcpu) & ((1 << 12) - 1);
1485 ret = io_mem_abort(vcpu, run, fault_ipa);
1489 /* Userspace should not be able to register out-of-bounds IPAs */
1490 VM_BUG_ON(fault_ipa >= KVM_PHYS_SIZE);
1492 if (fault_status == FSC_ACCESS) {
1493 handle_access_fault(vcpu, fault_ipa);
1498 ret = user_mem_abort(vcpu, fault_ipa, memslot, hva, fault_status);
1502 srcu_read_unlock(&vcpu->kvm->srcu, idx);
1506 static int handle_hva_to_gpa(struct kvm *kvm,
1507 unsigned long start,
1509 int (*handler)(struct kvm *kvm,
1510 gpa_t gpa, void *data),
1513 struct kvm_memslots *slots;
1514 struct kvm_memory_slot *memslot;
1517 slots = kvm_memslots(kvm);
1519 /* we only care about the pages that the guest sees */
1520 kvm_for_each_memslot(memslot, slots) {
1521 unsigned long hva_start, hva_end;
1524 hva_start = max(start, memslot->userspace_addr);
1525 hva_end = min(end, memslot->userspace_addr +
1526 (memslot->npages << PAGE_SHIFT));
1527 if (hva_start >= hva_end)
1531 * {gfn(page) | page intersects with [hva_start, hva_end)} =
1532 * {gfn_start, gfn_start+1, ..., gfn_end-1}.
1534 gfn = hva_to_gfn_memslot(hva_start, memslot);
1535 gfn_end = hva_to_gfn_memslot(hva_end + PAGE_SIZE - 1, memslot);
1537 for (; gfn < gfn_end; ++gfn) {
1538 gpa_t gpa = gfn << PAGE_SHIFT;
1539 ret |= handler(kvm, gpa, data);
1546 static int kvm_unmap_hva_handler(struct kvm *kvm, gpa_t gpa, void *data)
1548 unmap_stage2_range(kvm, gpa, PAGE_SIZE);
1552 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
1554 unsigned long end = hva + PAGE_SIZE;
1559 trace_kvm_unmap_hva(hva);
1560 handle_hva_to_gpa(kvm, hva, end, &kvm_unmap_hva_handler, NULL);
1564 int kvm_unmap_hva_range(struct kvm *kvm,
1565 unsigned long start, unsigned long end)
1570 trace_kvm_unmap_hva_range(start, end);
1571 handle_hva_to_gpa(kvm, start, end, &kvm_unmap_hva_handler, NULL);
1575 static int kvm_set_spte_handler(struct kvm *kvm, gpa_t gpa, void *data)
1577 pte_t *pte = (pte_t *)data;
1580 * We can always call stage2_set_pte with KVM_S2PTE_FLAG_LOGGING_ACTIVE
1581 * flag clear because MMU notifiers will have unmapped a huge PMD before
1582 * calling ->change_pte() (which in turn calls kvm_set_spte_hva()) and
1583 * therefore stage2_set_pte() never needs to clear out a huge PMD
1584 * through this calling path.
1586 stage2_set_pte(kvm, NULL, gpa, pte, 0);
1591 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
1593 unsigned long end = hva + PAGE_SIZE;
1599 trace_kvm_set_spte_hva(hva);
1600 stage2_pte = pfn_pte(pte_pfn(pte), PAGE_S2);
1601 handle_hva_to_gpa(kvm, hva, end, &kvm_set_spte_handler, &stage2_pte);
1604 static int kvm_age_hva_handler(struct kvm *kvm, gpa_t gpa, void *data)
1609 pmd = stage2_get_pmd(kvm, NULL, gpa);
1610 if (!pmd || pmd_none(*pmd)) /* Nothing there */
1613 if (pmd_thp_or_huge(*pmd)) { /* THP, HugeTLB */
1614 if (pmd_young(*pmd)) {
1615 *pmd = pmd_mkold(*pmd);
1622 pte = pte_offset_kernel(pmd, gpa);
1626 if (pte_young(*pte)) {
1627 *pte = pte_mkold(*pte); /* Just a page... */
1634 static int kvm_test_age_hva_handler(struct kvm *kvm, gpa_t gpa, void *data)
1639 pmd = stage2_get_pmd(kvm, NULL, gpa);
1640 if (!pmd || pmd_none(*pmd)) /* Nothing there */
1643 if (pmd_thp_or_huge(*pmd)) /* THP, HugeTLB */
1644 return pmd_young(*pmd);
1646 pte = pte_offset_kernel(pmd, gpa);
1647 if (!pte_none(*pte)) /* Just a page... */
1648 return pte_young(*pte);
1653 int kvm_age_hva(struct kvm *kvm, unsigned long start, unsigned long end)
1655 trace_kvm_age_hva(start, end);
1656 return handle_hva_to_gpa(kvm, start, end, kvm_age_hva_handler, NULL);
1659 int kvm_test_age_hva(struct kvm *kvm, unsigned long hva)
1661 trace_kvm_test_age_hva(hva);
1662 return handle_hva_to_gpa(kvm, hva, hva, kvm_test_age_hva_handler, NULL);
1665 void kvm_mmu_free_memory_caches(struct kvm_vcpu *vcpu)
1667 mmu_free_memory_cache(&vcpu->arch.mmu_page_cache);
1670 phys_addr_t kvm_mmu_get_httbr(void)
1672 if (__kvm_cpu_uses_extended_idmap())
1673 return virt_to_phys(merged_hyp_pgd);
1675 return virt_to_phys(hyp_pgd);
1678 phys_addr_t kvm_mmu_get_boot_httbr(void)
1680 if (__kvm_cpu_uses_extended_idmap())
1681 return virt_to_phys(merged_hyp_pgd);
1683 return virt_to_phys(boot_hyp_pgd);
1686 phys_addr_t kvm_get_idmap_vector(void)
1688 return hyp_idmap_vector;
1691 int kvm_mmu_init(void)
1695 hyp_idmap_start = kvm_virt_to_phys(__hyp_idmap_text_start);
1696 hyp_idmap_end = kvm_virt_to_phys(__hyp_idmap_text_end);
1697 hyp_idmap_vector = kvm_virt_to_phys(__kvm_hyp_init);
1700 * We rely on the linker script to ensure at build time that the HYP
1701 * init code does not cross a page boundary.
1703 BUG_ON((hyp_idmap_start ^ (hyp_idmap_end - 1)) & PAGE_MASK);
1705 hyp_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, hyp_pgd_order);
1706 boot_hyp_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, hyp_pgd_order);
1708 if (!hyp_pgd || !boot_hyp_pgd) {
1709 kvm_err("Hyp mode PGD not allocated\n");
1714 /* Create the idmap in the boot page tables */
1715 err = __create_hyp_mappings(boot_hyp_pgd,
1716 hyp_idmap_start, hyp_idmap_end,
1717 __phys_to_pfn(hyp_idmap_start),
1721 kvm_err("Failed to idmap %lx-%lx\n",
1722 hyp_idmap_start, hyp_idmap_end);
1726 if (__kvm_cpu_uses_extended_idmap()) {
1727 merged_hyp_pgd = (pgd_t *)__get_free_page(GFP_KERNEL | __GFP_ZERO);
1728 if (!merged_hyp_pgd) {
1729 kvm_err("Failed to allocate extra HYP pgd\n");
1732 __kvm_extend_hypmap(boot_hyp_pgd, hyp_pgd, merged_hyp_pgd,
1737 /* Map the very same page at the trampoline VA */
1738 err = __create_hyp_mappings(boot_hyp_pgd,
1739 TRAMPOLINE_VA, TRAMPOLINE_VA + PAGE_SIZE,
1740 __phys_to_pfn(hyp_idmap_start),
1743 kvm_err("Failed to map trampoline @%lx into boot HYP pgd\n",
1748 /* Map the same page again into the runtime page tables */
1749 err = __create_hyp_mappings(hyp_pgd,
1750 TRAMPOLINE_VA, TRAMPOLINE_VA + PAGE_SIZE,
1751 __phys_to_pfn(hyp_idmap_start),
1754 kvm_err("Failed to map trampoline @%lx into runtime HYP pgd\n",
1765 void kvm_arch_commit_memory_region(struct kvm *kvm,
1766 const struct kvm_userspace_memory_region *mem,
1767 const struct kvm_memory_slot *old,
1768 const struct kvm_memory_slot *new,
1769 enum kvm_mr_change change)
1772 * At this point memslot has been committed and there is an
1773 * allocated dirty_bitmap[], dirty pages will be be tracked while the
1774 * memory slot is write protected.
1776 if (change != KVM_MR_DELETE && mem->flags & KVM_MEM_LOG_DIRTY_PAGES)
1777 kvm_mmu_wp_memory_region(kvm, mem->slot);
1780 int kvm_arch_prepare_memory_region(struct kvm *kvm,
1781 struct kvm_memory_slot *memslot,
1782 const struct kvm_userspace_memory_region *mem,
1783 enum kvm_mr_change change)
1785 hva_t hva = mem->userspace_addr;
1786 hva_t reg_end = hva + mem->memory_size;
1787 bool writable = !(mem->flags & KVM_MEM_READONLY);
1790 if (change != KVM_MR_CREATE && change != KVM_MR_MOVE &&
1791 change != KVM_MR_FLAGS_ONLY)
1795 * Prevent userspace from creating a memory region outside of the IPA
1796 * space addressable by the KVM guest IPA space.
1798 if (memslot->base_gfn + memslot->npages >=
1799 (KVM_PHYS_SIZE >> PAGE_SHIFT))
1803 * A memory region could potentially cover multiple VMAs, and any holes
1804 * between them, so iterate over all of them to find out if we can map
1805 * any of them right now.
1807 * +--------------------------------------------+
1808 * +---------------+----------------+ +----------------+
1809 * | : VMA 1 | VMA 2 | | VMA 3 : |
1810 * +---------------+----------------+ +----------------+
1812 * +--------------------------------------------+
1815 struct vm_area_struct *vma = find_vma(current->mm, hva);
1816 hva_t vm_start, vm_end;
1818 if (!vma || vma->vm_start >= reg_end)
1822 * Mapping a read-only VMA is only allowed if the
1823 * memory region is configured as read-only.
1825 if (writable && !(vma->vm_flags & VM_WRITE)) {
1831 * Take the intersection of this VMA with the memory region
1833 vm_start = max(hva, vma->vm_start);
1834 vm_end = min(reg_end, vma->vm_end);
1836 if (vma->vm_flags & VM_PFNMAP) {
1837 gpa_t gpa = mem->guest_phys_addr +
1838 (vm_start - mem->userspace_addr);
1841 pa = (phys_addr_t)vma->vm_pgoff << PAGE_SHIFT;
1842 pa += vm_start - vma->vm_start;
1844 /* IO region dirty page logging not allowed */
1845 if (memslot->flags & KVM_MEM_LOG_DIRTY_PAGES)
1848 ret = kvm_phys_addr_ioremap(kvm, gpa, pa,
1855 } while (hva < reg_end);
1857 if (change == KVM_MR_FLAGS_ONLY)
1860 spin_lock(&kvm->mmu_lock);
1862 unmap_stage2_range(kvm, mem->guest_phys_addr, mem->memory_size);
1864 stage2_flush_memslot(kvm, memslot);
1865 spin_unlock(&kvm->mmu_lock);
1869 void kvm_arch_free_memslot(struct kvm *kvm, struct kvm_memory_slot *free,
1870 struct kvm_memory_slot *dont)
1874 int kvm_arch_create_memslot(struct kvm *kvm, struct kvm_memory_slot *slot,
1875 unsigned long npages)
1878 * Readonly memslots are not incoherent with the caches by definition,
1879 * but in practice, they are used mostly to emulate ROMs or NOR flashes
1880 * that the guest may consider devices and hence map as uncached.
1881 * To prevent incoherency issues in these cases, tag all readonly
1882 * regions as incoherent.
1884 if (slot->flags & KVM_MEM_READONLY)
1885 slot->flags |= KVM_MEMSLOT_INCOHERENT;
1889 void kvm_arch_memslots_updated(struct kvm *kvm, struct kvm_memslots *slots)
1893 void kvm_arch_flush_shadow_all(struct kvm *kvm)
1897 void kvm_arch_flush_shadow_memslot(struct kvm *kvm,
1898 struct kvm_memory_slot *slot)
1900 gpa_t gpa = slot->base_gfn << PAGE_SHIFT;
1901 phys_addr_t size = slot->npages << PAGE_SHIFT;
1903 spin_lock(&kvm->mmu_lock);
1904 unmap_stage2_range(kvm, gpa, size);
1905 spin_unlock(&kvm->mmu_lock);
1909 * See note at ARMv7 ARM B1.14.4 (TL;DR: S/W ops are not easily virtualized).
1912 * - S/W ops are local to a CPU (not broadcast)
1913 * - We have line migration behind our back (speculation)
1914 * - System caches don't support S/W at all (damn!)
1916 * In the face of the above, the best we can do is to try and convert
1917 * S/W ops to VA ops. Because the guest is not allowed to infer the
1918 * S/W to PA mapping, it can only use S/W to nuke the whole cache,
1919 * which is a rather good thing for us.
1921 * Also, it is only used when turning caches on/off ("The expected
1922 * usage of the cache maintenance instructions that operate by set/way
1923 * is associated with the cache maintenance instructions associated
1924 * with the powerdown and powerup of caches, if this is required by
1925 * the implementation.").
1927 * We use the following policy:
1929 * - If we trap a S/W operation, we enable VM trapping to detect
1930 * caches being turned on/off, and do a full clean.
1932 * - We flush the caches on both caches being turned on and off.
1934 * - Once the caches are enabled, we stop trapping VM ops.
1936 void kvm_set_way_flush(struct kvm_vcpu *vcpu)
1938 unsigned long hcr = vcpu_get_hcr(vcpu);
1941 * If this is the first time we do a S/W operation
1942 * (i.e. HCR_TVM not set) flush the whole memory, and set the
1945 * Otherwise, rely on the VM trapping to wait for the MMU +
1946 * Caches to be turned off. At that point, we'll be able to
1947 * clean the caches again.
1949 if (!(hcr & HCR_TVM)) {
1950 trace_kvm_set_way_flush(*vcpu_pc(vcpu),
1951 vcpu_has_cache_enabled(vcpu));
1952 stage2_flush_vm(vcpu->kvm);
1953 vcpu_set_hcr(vcpu, hcr | HCR_TVM);
1957 void kvm_toggle_cache(struct kvm_vcpu *vcpu, bool was_enabled)
1959 bool now_enabled = vcpu_has_cache_enabled(vcpu);
1962 * If switching the MMU+caches on, need to invalidate the caches.
1963 * If switching it off, need to clean the caches.
1964 * Clean + invalidate does the trick always.
1966 if (now_enabled != was_enabled)
1967 stage2_flush_vm(vcpu->kvm);
1969 /* Caches are now on, stop trapping VM ops (until a S/W op) */
1971 vcpu_set_hcr(vcpu, vcpu_get_hcr(vcpu) & ~HCR_TVM);
1973 trace_kvm_toggle_cache(*vcpu_pc(vcpu), was_enabled, now_enabled);