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);
980 #ifndef __HAVE_ARCH_PTEP_TEST_AND_CLEAR_YOUNG
981 static int stage2_ptep_test_and_clear_young(pte_t *pte)
983 if (pte_young(*pte)) {
984 *pte = pte_mkold(*pte);
990 static int stage2_ptep_test_and_clear_young(pte_t *pte)
992 return __ptep_test_and_clear_young(pte);
996 static int stage2_pmdp_test_and_clear_young(pmd_t *pmd)
998 return stage2_ptep_test_and_clear_young((pte_t *)pmd);
1002 * kvm_phys_addr_ioremap - map a device range to guest IPA
1004 * @kvm: The KVM pointer
1005 * @guest_ipa: The IPA at which to insert the mapping
1006 * @pa: The physical address of the device
1007 * @size: The size of the mapping
1009 int kvm_phys_addr_ioremap(struct kvm *kvm, phys_addr_t guest_ipa,
1010 phys_addr_t pa, unsigned long size, bool writable)
1012 phys_addr_t addr, end;
1015 struct kvm_mmu_memory_cache cache = { 0, };
1017 end = (guest_ipa + size + PAGE_SIZE - 1) & PAGE_MASK;
1018 pfn = __phys_to_pfn(pa);
1020 for (addr = guest_ipa; addr < end; addr += PAGE_SIZE) {
1021 pte_t pte = pfn_pte(pfn, PAGE_S2_DEVICE);
1024 pte = kvm_s2pte_mkwrite(pte);
1026 ret = mmu_topup_memory_cache(&cache, KVM_MMU_CACHE_MIN_PAGES,
1030 spin_lock(&kvm->mmu_lock);
1031 ret = stage2_set_pte(kvm, &cache, addr, &pte,
1032 KVM_S2PTE_FLAG_IS_IOMAP);
1033 spin_unlock(&kvm->mmu_lock);
1041 mmu_free_memory_cache(&cache);
1045 static bool transparent_hugepage_adjust(kvm_pfn_t *pfnp, phys_addr_t *ipap)
1047 kvm_pfn_t pfn = *pfnp;
1048 gfn_t gfn = *ipap >> PAGE_SHIFT;
1050 if (PageTransCompoundMap(pfn_to_page(pfn))) {
1053 * The address we faulted on is backed by a transparent huge
1054 * page. However, because we map the compound huge page and
1055 * not the individual tail page, we need to transfer the
1056 * refcount to the head page. We have to be careful that the
1057 * THP doesn't start to split while we are adjusting the
1060 * We are sure this doesn't happen, because mmu_notifier_retry
1061 * was successful and we are holding the mmu_lock, so if this
1062 * THP is trying to split, it will be blocked in the mmu
1063 * notifier before touching any of the pages, specifically
1064 * before being able to call __split_huge_page_refcount().
1066 * We can therefore safely transfer the refcount from PG_tail
1067 * to PG_head and switch the pfn from a tail page to the head
1070 mask = PTRS_PER_PMD - 1;
1071 VM_BUG_ON((gfn & mask) != (pfn & mask));
1074 kvm_release_pfn_clean(pfn);
1086 static bool kvm_is_write_fault(struct kvm_vcpu *vcpu)
1088 if (kvm_vcpu_trap_is_iabt(vcpu))
1091 return kvm_vcpu_dabt_iswrite(vcpu);
1095 * stage2_wp_ptes - write protect PMD range
1096 * @pmd: pointer to pmd entry
1097 * @addr: range start address
1098 * @end: range end address
1100 static void stage2_wp_ptes(pmd_t *pmd, phys_addr_t addr, phys_addr_t end)
1104 pte = pte_offset_kernel(pmd, addr);
1106 if (!pte_none(*pte)) {
1107 if (!kvm_s2pte_readonly(pte))
1108 kvm_set_s2pte_readonly(pte);
1110 } while (pte++, addr += PAGE_SIZE, addr != end);
1114 * stage2_wp_pmds - write protect PUD range
1115 * @pud: pointer to pud entry
1116 * @addr: range start address
1117 * @end: range end address
1119 static void stage2_wp_pmds(pud_t *pud, phys_addr_t addr, phys_addr_t end)
1124 pmd = stage2_pmd_offset(pud, addr);
1127 next = stage2_pmd_addr_end(addr, end);
1128 if (!pmd_none(*pmd)) {
1129 if (pmd_thp_or_huge(*pmd)) {
1130 if (!kvm_s2pmd_readonly(pmd))
1131 kvm_set_s2pmd_readonly(pmd);
1133 stage2_wp_ptes(pmd, addr, next);
1136 } while (pmd++, addr = next, addr != end);
1140 * stage2_wp_puds - write protect PGD range
1141 * @pgd: pointer to pgd entry
1142 * @addr: range start address
1143 * @end: range end address
1145 * Process PUD entries, for a huge PUD we cause a panic.
1147 static void stage2_wp_puds(pgd_t *pgd, phys_addr_t addr, phys_addr_t end)
1152 pud = stage2_pud_offset(pgd, addr);
1154 next = stage2_pud_addr_end(addr, end);
1155 if (!stage2_pud_none(*pud)) {
1156 /* TODO:PUD not supported, revisit later if supported */
1157 BUG_ON(stage2_pud_huge(*pud));
1158 stage2_wp_pmds(pud, addr, next);
1160 } while (pud++, addr = next, addr != end);
1164 * stage2_wp_range() - write protect stage2 memory region range
1165 * @kvm: The KVM pointer
1166 * @addr: Start address of range
1167 * @end: End address of range
1169 static void stage2_wp_range(struct kvm *kvm, phys_addr_t addr, phys_addr_t end)
1174 pgd = kvm->arch.pgd + stage2_pgd_index(addr);
1177 * Release kvm_mmu_lock periodically if the memory region is
1178 * large. Otherwise, we may see kernel panics with
1179 * CONFIG_DETECT_HUNG_TASK, CONFIG_LOCKUP_DETECTOR,
1180 * CONFIG_LOCKDEP. Additionally, holding the lock too long
1181 * will also starve other vCPUs.
1183 if (need_resched() || spin_needbreak(&kvm->mmu_lock))
1184 cond_resched_lock(&kvm->mmu_lock);
1186 next = stage2_pgd_addr_end(addr, end);
1187 if (stage2_pgd_present(*pgd))
1188 stage2_wp_puds(pgd, addr, next);
1189 } while (pgd++, addr = next, addr != end);
1193 * kvm_mmu_wp_memory_region() - write protect stage 2 entries for memory slot
1194 * @kvm: The KVM pointer
1195 * @slot: The memory slot to write protect
1197 * Called to start logging dirty pages after memory region
1198 * KVM_MEM_LOG_DIRTY_PAGES operation is called. After this function returns
1199 * all present PMD and PTEs are write protected in the memory region.
1200 * Afterwards read of dirty page log can be called.
1202 * Acquires kvm_mmu_lock. Called with kvm->slots_lock mutex acquired,
1203 * serializing operations for VM memory regions.
1205 void kvm_mmu_wp_memory_region(struct kvm *kvm, int slot)
1207 struct kvm_memslots *slots = kvm_memslots(kvm);
1208 struct kvm_memory_slot *memslot = id_to_memslot(slots, slot);
1209 phys_addr_t start = memslot->base_gfn << PAGE_SHIFT;
1210 phys_addr_t end = (memslot->base_gfn + memslot->npages) << PAGE_SHIFT;
1212 spin_lock(&kvm->mmu_lock);
1213 stage2_wp_range(kvm, start, end);
1214 spin_unlock(&kvm->mmu_lock);
1215 kvm_flush_remote_tlbs(kvm);
1219 * kvm_mmu_write_protect_pt_masked() - write protect dirty pages
1220 * @kvm: The KVM pointer
1221 * @slot: The memory slot associated with mask
1222 * @gfn_offset: The gfn offset in memory slot
1223 * @mask: The mask of dirty pages at offset 'gfn_offset' in this memory
1224 * slot to be write protected
1226 * Walks bits set in mask write protects the associated pte's. Caller must
1227 * acquire kvm_mmu_lock.
1229 static void kvm_mmu_write_protect_pt_masked(struct kvm *kvm,
1230 struct kvm_memory_slot *slot,
1231 gfn_t gfn_offset, unsigned long mask)
1233 phys_addr_t base_gfn = slot->base_gfn + gfn_offset;
1234 phys_addr_t start = (base_gfn + __ffs(mask)) << PAGE_SHIFT;
1235 phys_addr_t end = (base_gfn + __fls(mask) + 1) << PAGE_SHIFT;
1237 stage2_wp_range(kvm, start, end);
1241 * kvm_arch_mmu_enable_log_dirty_pt_masked - enable dirty logging for selected
1244 * It calls kvm_mmu_write_protect_pt_masked to write protect selected pages to
1245 * enable dirty logging for them.
1247 void kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm *kvm,
1248 struct kvm_memory_slot *slot,
1249 gfn_t gfn_offset, unsigned long mask)
1251 kvm_mmu_write_protect_pt_masked(kvm, slot, gfn_offset, mask);
1254 static void coherent_cache_guest_page(struct kvm_vcpu *vcpu, kvm_pfn_t pfn,
1255 unsigned long size, bool uncached)
1257 __coherent_cache_guest_page(vcpu, pfn, size, uncached);
1260 static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
1261 struct kvm_memory_slot *memslot, unsigned long hva,
1262 unsigned long fault_status)
1265 bool write_fault, writable, hugetlb = false, force_pte = false;
1266 unsigned long mmu_seq;
1267 gfn_t gfn = fault_ipa >> PAGE_SHIFT;
1268 struct kvm *kvm = vcpu->kvm;
1269 struct kvm_mmu_memory_cache *memcache = &vcpu->arch.mmu_page_cache;
1270 struct vm_area_struct *vma;
1272 pgprot_t mem_type = PAGE_S2;
1273 bool fault_ipa_uncached;
1274 bool logging_active = memslot_is_logging(memslot);
1275 unsigned long flags = 0;
1277 write_fault = kvm_is_write_fault(vcpu);
1278 if (fault_status == FSC_PERM && !write_fault) {
1279 kvm_err("Unexpected L2 read permission error\n");
1283 /* Let's check if we will get back a huge page backed by hugetlbfs */
1284 down_read(¤t->mm->mmap_sem);
1285 vma = find_vma_intersection(current->mm, hva, hva + 1);
1286 if (unlikely(!vma)) {
1287 kvm_err("Failed to find VMA for hva 0x%lx\n", hva);
1288 up_read(¤t->mm->mmap_sem);
1292 if (is_vm_hugetlb_page(vma) && !logging_active) {
1294 gfn = (fault_ipa & PMD_MASK) >> PAGE_SHIFT;
1297 * Pages belonging to memslots that don't have the same
1298 * alignment for userspace and IPA cannot be mapped using
1299 * block descriptors even if the pages belong to a THP for
1300 * the process, because the stage-2 block descriptor will
1301 * cover more than a single THP and we loose atomicity for
1302 * unmapping, updates, and splits of the THP or other pages
1303 * in the stage-2 block range.
1305 if ((memslot->userspace_addr & ~PMD_MASK) !=
1306 ((memslot->base_gfn << PAGE_SHIFT) & ~PMD_MASK))
1309 up_read(¤t->mm->mmap_sem);
1311 /* We need minimum second+third level pages */
1312 ret = mmu_topup_memory_cache(memcache, KVM_MMU_CACHE_MIN_PAGES,
1317 mmu_seq = vcpu->kvm->mmu_notifier_seq;
1319 * Ensure the read of mmu_notifier_seq happens before we call
1320 * gfn_to_pfn_prot (which calls get_user_pages), so that we don't risk
1321 * the page we just got a reference to gets unmapped before we have a
1322 * chance to grab the mmu_lock, which ensure that if the page gets
1323 * unmapped afterwards, the call to kvm_unmap_hva will take it away
1324 * from us again properly. This smp_rmb() interacts with the smp_wmb()
1325 * in kvm_mmu_notifier_invalidate_<page|range_end>.
1329 pfn = gfn_to_pfn_prot(kvm, gfn, write_fault, &writable);
1330 if (is_error_pfn(pfn))
1333 if (kvm_is_device_pfn(pfn)) {
1334 mem_type = PAGE_S2_DEVICE;
1335 flags |= KVM_S2PTE_FLAG_IS_IOMAP;
1336 } else if (logging_active) {
1338 * Faults on pages in a memslot with logging enabled
1339 * should not be mapped with huge pages (it introduces churn
1340 * and performance degradation), so force a pte mapping.
1343 flags |= KVM_S2_FLAG_LOGGING_ACTIVE;
1346 * Only actually map the page as writable if this was a write
1353 spin_lock(&kvm->mmu_lock);
1354 if (mmu_notifier_retry(kvm, mmu_seq))
1357 if (!hugetlb && !force_pte)
1358 hugetlb = transparent_hugepage_adjust(&pfn, &fault_ipa);
1360 fault_ipa_uncached = memslot->flags & KVM_MEMSLOT_INCOHERENT;
1363 pmd_t new_pmd = pfn_pmd(pfn, mem_type);
1364 new_pmd = pmd_mkhuge(new_pmd);
1366 new_pmd = kvm_s2pmd_mkwrite(new_pmd);
1367 kvm_set_pfn_dirty(pfn);
1369 coherent_cache_guest_page(vcpu, pfn, PMD_SIZE, fault_ipa_uncached);
1370 ret = stage2_set_pmd_huge(kvm, memcache, fault_ipa, &new_pmd);
1372 pte_t new_pte = pfn_pte(pfn, mem_type);
1375 new_pte = kvm_s2pte_mkwrite(new_pte);
1376 kvm_set_pfn_dirty(pfn);
1377 mark_page_dirty(kvm, gfn);
1379 coherent_cache_guest_page(vcpu, pfn, PAGE_SIZE, fault_ipa_uncached);
1380 ret = stage2_set_pte(kvm, memcache, fault_ipa, &new_pte, flags);
1384 spin_unlock(&kvm->mmu_lock);
1385 kvm_set_pfn_accessed(pfn);
1386 kvm_release_pfn_clean(pfn);
1391 * Resolve the access fault by making the page young again.
1392 * Note that because the faulting entry is guaranteed not to be
1393 * cached in the TLB, we don't need to invalidate anything.
1394 * Only the HW Access Flag updates are supported for Stage 2 (no DBM),
1395 * so there is no need for atomic (pte|pmd)_mkyoung operations.
1397 static void handle_access_fault(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa)
1402 bool pfn_valid = false;
1404 trace_kvm_access_fault(fault_ipa);
1406 spin_lock(&vcpu->kvm->mmu_lock);
1408 pmd = stage2_get_pmd(vcpu->kvm, NULL, fault_ipa);
1409 if (!pmd || pmd_none(*pmd)) /* Nothing there */
1412 if (pmd_thp_or_huge(*pmd)) { /* THP, HugeTLB */
1413 *pmd = pmd_mkyoung(*pmd);
1414 pfn = pmd_pfn(*pmd);
1419 pte = pte_offset_kernel(pmd, fault_ipa);
1420 if (pte_none(*pte)) /* Nothing there either */
1423 *pte = pte_mkyoung(*pte); /* Just a page... */
1424 pfn = pte_pfn(*pte);
1427 spin_unlock(&vcpu->kvm->mmu_lock);
1429 kvm_set_pfn_accessed(pfn);
1433 * kvm_handle_guest_abort - handles all 2nd stage aborts
1434 * @vcpu: the VCPU pointer
1435 * @run: the kvm_run structure
1437 * Any abort that gets to the host is almost guaranteed to be caused by a
1438 * missing second stage translation table entry, which can mean that either the
1439 * guest simply needs more memory and we must allocate an appropriate page or it
1440 * can mean that the guest tried to access I/O memory, which is emulated by user
1441 * space. The distinction is based on the IPA causing the fault and whether this
1442 * memory region has been registered as standard RAM by user space.
1444 int kvm_handle_guest_abort(struct kvm_vcpu *vcpu, struct kvm_run *run)
1446 unsigned long fault_status;
1447 phys_addr_t fault_ipa;
1448 struct kvm_memory_slot *memslot;
1450 bool is_iabt, write_fault, writable;
1454 is_iabt = kvm_vcpu_trap_is_iabt(vcpu);
1455 fault_ipa = kvm_vcpu_get_fault_ipa(vcpu);
1457 trace_kvm_guest_fault(*vcpu_pc(vcpu), kvm_vcpu_get_hsr(vcpu),
1458 kvm_vcpu_get_hfar(vcpu), fault_ipa);
1460 /* Check the stage-2 fault is trans. fault or write fault */
1461 fault_status = kvm_vcpu_trap_get_fault_type(vcpu);
1462 if (fault_status != FSC_FAULT && fault_status != FSC_PERM &&
1463 fault_status != FSC_ACCESS) {
1464 kvm_err("Unsupported FSC: EC=%#x xFSC=%#lx ESR_EL2=%#lx\n",
1465 kvm_vcpu_trap_get_class(vcpu),
1466 (unsigned long)kvm_vcpu_trap_get_fault(vcpu),
1467 (unsigned long)kvm_vcpu_get_hsr(vcpu));
1471 idx = srcu_read_lock(&vcpu->kvm->srcu);
1473 gfn = fault_ipa >> PAGE_SHIFT;
1474 memslot = gfn_to_memslot(vcpu->kvm, gfn);
1475 hva = gfn_to_hva_memslot_prot(memslot, gfn, &writable);
1476 write_fault = kvm_is_write_fault(vcpu);
1477 if (kvm_is_error_hva(hva) || (write_fault && !writable)) {
1479 /* Prefetch Abort on I/O address */
1480 kvm_inject_pabt(vcpu, kvm_vcpu_get_hfar(vcpu));
1486 * Check for a cache maintenance operation. Since we
1487 * ended-up here, we know it is outside of any memory
1488 * slot. But we can't find out if that is for a device,
1489 * or if the guest is just being stupid. The only thing
1490 * we know for sure is that this range cannot be cached.
1492 * So let's assume that the guest is just being
1493 * cautious, and skip the instruction.
1495 if (kvm_vcpu_dabt_is_cm(vcpu)) {
1496 kvm_skip_instr(vcpu, kvm_vcpu_trap_il_is32bit(vcpu));
1502 * The IPA is reported as [MAX:12], so we need to
1503 * complement it with the bottom 12 bits from the
1504 * faulting VA. This is always 12 bits, irrespective
1507 fault_ipa |= kvm_vcpu_get_hfar(vcpu) & ((1 << 12) - 1);
1508 ret = io_mem_abort(vcpu, run, fault_ipa);
1512 /* Userspace should not be able to register out-of-bounds IPAs */
1513 VM_BUG_ON(fault_ipa >= KVM_PHYS_SIZE);
1515 if (fault_status == FSC_ACCESS) {
1516 handle_access_fault(vcpu, fault_ipa);
1521 ret = user_mem_abort(vcpu, fault_ipa, memslot, hva, fault_status);
1525 srcu_read_unlock(&vcpu->kvm->srcu, idx);
1529 static int handle_hva_to_gpa(struct kvm *kvm,
1530 unsigned long start,
1532 int (*handler)(struct kvm *kvm,
1533 gpa_t gpa, void *data),
1536 struct kvm_memslots *slots;
1537 struct kvm_memory_slot *memslot;
1540 slots = kvm_memslots(kvm);
1542 /* we only care about the pages that the guest sees */
1543 kvm_for_each_memslot(memslot, slots) {
1544 unsigned long hva_start, hva_end;
1547 hva_start = max(start, memslot->userspace_addr);
1548 hva_end = min(end, memslot->userspace_addr +
1549 (memslot->npages << PAGE_SHIFT));
1550 if (hva_start >= hva_end)
1554 * {gfn(page) | page intersects with [hva_start, hva_end)} =
1555 * {gfn_start, gfn_start+1, ..., gfn_end-1}.
1557 gfn = hva_to_gfn_memslot(hva_start, memslot);
1558 gfn_end = hva_to_gfn_memslot(hva_end + PAGE_SIZE - 1, memslot);
1560 for (; gfn < gfn_end; ++gfn) {
1561 gpa_t gpa = gfn << PAGE_SHIFT;
1562 ret |= handler(kvm, gpa, data);
1569 static int kvm_unmap_hva_handler(struct kvm *kvm, gpa_t gpa, void *data)
1571 unmap_stage2_range(kvm, gpa, PAGE_SIZE);
1575 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
1577 unsigned long end = hva + PAGE_SIZE;
1582 trace_kvm_unmap_hva(hva);
1583 handle_hva_to_gpa(kvm, hva, end, &kvm_unmap_hva_handler, NULL);
1587 int kvm_unmap_hva_range(struct kvm *kvm,
1588 unsigned long start, unsigned long end)
1593 trace_kvm_unmap_hva_range(start, end);
1594 handle_hva_to_gpa(kvm, start, end, &kvm_unmap_hva_handler, NULL);
1598 static int kvm_set_spte_handler(struct kvm *kvm, gpa_t gpa, void *data)
1600 pte_t *pte = (pte_t *)data;
1603 * We can always call stage2_set_pte with KVM_S2PTE_FLAG_LOGGING_ACTIVE
1604 * flag clear because MMU notifiers will have unmapped a huge PMD before
1605 * calling ->change_pte() (which in turn calls kvm_set_spte_hva()) and
1606 * therefore stage2_set_pte() never needs to clear out a huge PMD
1607 * through this calling path.
1609 stage2_set_pte(kvm, NULL, gpa, pte, 0);
1614 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
1616 unsigned long end = hva + PAGE_SIZE;
1622 trace_kvm_set_spte_hva(hva);
1623 stage2_pte = pfn_pte(pte_pfn(pte), PAGE_S2);
1624 handle_hva_to_gpa(kvm, hva, end, &kvm_set_spte_handler, &stage2_pte);
1627 static int kvm_age_hva_handler(struct kvm *kvm, gpa_t gpa, void *data)
1632 pmd = stage2_get_pmd(kvm, NULL, gpa);
1633 if (!pmd || pmd_none(*pmd)) /* Nothing there */
1636 if (pmd_thp_or_huge(*pmd)) /* THP, HugeTLB */
1637 return stage2_pmdp_test_and_clear_young(pmd);
1639 pte = pte_offset_kernel(pmd, gpa);
1643 return stage2_ptep_test_and_clear_young(pte);
1646 static int kvm_test_age_hva_handler(struct kvm *kvm, gpa_t gpa, void *data)
1651 pmd = stage2_get_pmd(kvm, NULL, gpa);
1652 if (!pmd || pmd_none(*pmd)) /* Nothing there */
1655 if (pmd_thp_or_huge(*pmd)) /* THP, HugeTLB */
1656 return pmd_young(*pmd);
1658 pte = pte_offset_kernel(pmd, gpa);
1659 if (!pte_none(*pte)) /* Just a page... */
1660 return pte_young(*pte);
1665 int kvm_age_hva(struct kvm *kvm, unsigned long start, unsigned long end)
1667 trace_kvm_age_hva(start, end);
1668 return handle_hva_to_gpa(kvm, start, end, kvm_age_hva_handler, NULL);
1671 int kvm_test_age_hva(struct kvm *kvm, unsigned long hva)
1673 trace_kvm_test_age_hva(hva);
1674 return handle_hva_to_gpa(kvm, hva, hva, kvm_test_age_hva_handler, NULL);
1677 void kvm_mmu_free_memory_caches(struct kvm_vcpu *vcpu)
1679 mmu_free_memory_cache(&vcpu->arch.mmu_page_cache);
1682 phys_addr_t kvm_mmu_get_httbr(void)
1684 if (__kvm_cpu_uses_extended_idmap())
1685 return virt_to_phys(merged_hyp_pgd);
1687 return virt_to_phys(hyp_pgd);
1690 phys_addr_t kvm_mmu_get_boot_httbr(void)
1692 if (__kvm_cpu_uses_extended_idmap())
1693 return virt_to_phys(merged_hyp_pgd);
1695 return virt_to_phys(boot_hyp_pgd);
1698 phys_addr_t kvm_get_idmap_vector(void)
1700 return hyp_idmap_vector;
1703 phys_addr_t kvm_get_idmap_start(void)
1705 return hyp_idmap_start;
1708 int kvm_mmu_init(void)
1712 hyp_idmap_start = kvm_virt_to_phys(__hyp_idmap_text_start);
1713 hyp_idmap_end = kvm_virt_to_phys(__hyp_idmap_text_end);
1714 hyp_idmap_vector = kvm_virt_to_phys(__kvm_hyp_init);
1717 * We rely on the linker script to ensure at build time that the HYP
1718 * init code does not cross a page boundary.
1720 BUG_ON((hyp_idmap_start ^ (hyp_idmap_end - 1)) & PAGE_MASK);
1722 hyp_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, hyp_pgd_order);
1723 boot_hyp_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, hyp_pgd_order);
1725 if (!hyp_pgd || !boot_hyp_pgd) {
1726 kvm_err("Hyp mode PGD not allocated\n");
1731 /* Create the idmap in the boot page tables */
1732 err = __create_hyp_mappings(boot_hyp_pgd,
1733 hyp_idmap_start, hyp_idmap_end,
1734 __phys_to_pfn(hyp_idmap_start),
1738 kvm_err("Failed to idmap %lx-%lx\n",
1739 hyp_idmap_start, hyp_idmap_end);
1743 if (__kvm_cpu_uses_extended_idmap()) {
1744 merged_hyp_pgd = (pgd_t *)__get_free_page(GFP_KERNEL | __GFP_ZERO);
1745 if (!merged_hyp_pgd) {
1746 kvm_err("Failed to allocate extra HYP pgd\n");
1749 __kvm_extend_hypmap(boot_hyp_pgd, hyp_pgd, merged_hyp_pgd,
1754 /* Map the very same page at the trampoline VA */
1755 err = __create_hyp_mappings(boot_hyp_pgd,
1756 TRAMPOLINE_VA, TRAMPOLINE_VA + PAGE_SIZE,
1757 __phys_to_pfn(hyp_idmap_start),
1760 kvm_err("Failed to map trampoline @%lx into boot HYP pgd\n",
1765 /* Map the same page again into the runtime page tables */
1766 err = __create_hyp_mappings(hyp_pgd,
1767 TRAMPOLINE_VA, TRAMPOLINE_VA + PAGE_SIZE,
1768 __phys_to_pfn(hyp_idmap_start),
1771 kvm_err("Failed to map trampoline @%lx into runtime HYP pgd\n",
1782 void kvm_arch_commit_memory_region(struct kvm *kvm,
1783 const struct kvm_userspace_memory_region *mem,
1784 const struct kvm_memory_slot *old,
1785 const struct kvm_memory_slot *new,
1786 enum kvm_mr_change change)
1789 * At this point memslot has been committed and there is an
1790 * allocated dirty_bitmap[], dirty pages will be be tracked while the
1791 * memory slot is write protected.
1793 if (change != KVM_MR_DELETE && mem->flags & KVM_MEM_LOG_DIRTY_PAGES)
1794 kvm_mmu_wp_memory_region(kvm, mem->slot);
1797 int kvm_arch_prepare_memory_region(struct kvm *kvm,
1798 struct kvm_memory_slot *memslot,
1799 const struct kvm_userspace_memory_region *mem,
1800 enum kvm_mr_change change)
1802 hva_t hva = mem->userspace_addr;
1803 hva_t reg_end = hva + mem->memory_size;
1804 bool writable = !(mem->flags & KVM_MEM_READONLY);
1807 if (change != KVM_MR_CREATE && change != KVM_MR_MOVE &&
1808 change != KVM_MR_FLAGS_ONLY)
1812 * Prevent userspace from creating a memory region outside of the IPA
1813 * space addressable by the KVM guest IPA space.
1815 if (memslot->base_gfn + memslot->npages >=
1816 (KVM_PHYS_SIZE >> PAGE_SHIFT))
1820 * A memory region could potentially cover multiple VMAs, and any holes
1821 * between them, so iterate over all of them to find out if we can map
1822 * any of them right now.
1824 * +--------------------------------------------+
1825 * +---------------+----------------+ +----------------+
1826 * | : VMA 1 | VMA 2 | | VMA 3 : |
1827 * +---------------+----------------+ +----------------+
1829 * +--------------------------------------------+
1832 struct vm_area_struct *vma = find_vma(current->mm, hva);
1833 hva_t vm_start, vm_end;
1835 if (!vma || vma->vm_start >= reg_end)
1839 * Mapping a read-only VMA is only allowed if the
1840 * memory region is configured as read-only.
1842 if (writable && !(vma->vm_flags & VM_WRITE)) {
1848 * Take the intersection of this VMA with the memory region
1850 vm_start = max(hva, vma->vm_start);
1851 vm_end = min(reg_end, vma->vm_end);
1853 if (vma->vm_flags & VM_PFNMAP) {
1854 gpa_t gpa = mem->guest_phys_addr +
1855 (vm_start - mem->userspace_addr);
1858 pa = (phys_addr_t)vma->vm_pgoff << PAGE_SHIFT;
1859 pa += vm_start - vma->vm_start;
1861 /* IO region dirty page logging not allowed */
1862 if (memslot->flags & KVM_MEM_LOG_DIRTY_PAGES)
1865 ret = kvm_phys_addr_ioremap(kvm, gpa, pa,
1872 } while (hva < reg_end);
1874 if (change == KVM_MR_FLAGS_ONLY)
1877 spin_lock(&kvm->mmu_lock);
1879 unmap_stage2_range(kvm, mem->guest_phys_addr, mem->memory_size);
1881 stage2_flush_memslot(kvm, memslot);
1882 spin_unlock(&kvm->mmu_lock);
1886 void kvm_arch_free_memslot(struct kvm *kvm, struct kvm_memory_slot *free,
1887 struct kvm_memory_slot *dont)
1891 int kvm_arch_create_memslot(struct kvm *kvm, struct kvm_memory_slot *slot,
1892 unsigned long npages)
1895 * Readonly memslots are not incoherent with the caches by definition,
1896 * but in practice, they are used mostly to emulate ROMs or NOR flashes
1897 * that the guest may consider devices and hence map as uncached.
1898 * To prevent incoherency issues in these cases, tag all readonly
1899 * regions as incoherent.
1901 if (slot->flags & KVM_MEM_READONLY)
1902 slot->flags |= KVM_MEMSLOT_INCOHERENT;
1906 void kvm_arch_memslots_updated(struct kvm *kvm, struct kvm_memslots *slots)
1910 void kvm_arch_flush_shadow_all(struct kvm *kvm)
1914 void kvm_arch_flush_shadow_memslot(struct kvm *kvm,
1915 struct kvm_memory_slot *slot)
1917 gpa_t gpa = slot->base_gfn << PAGE_SHIFT;
1918 phys_addr_t size = slot->npages << PAGE_SHIFT;
1920 spin_lock(&kvm->mmu_lock);
1921 unmap_stage2_range(kvm, gpa, size);
1922 spin_unlock(&kvm->mmu_lock);
1926 * See note at ARMv7 ARM B1.14.4 (TL;DR: S/W ops are not easily virtualized).
1929 * - S/W ops are local to a CPU (not broadcast)
1930 * - We have line migration behind our back (speculation)
1931 * - System caches don't support S/W at all (damn!)
1933 * In the face of the above, the best we can do is to try and convert
1934 * S/W ops to VA ops. Because the guest is not allowed to infer the
1935 * S/W to PA mapping, it can only use S/W to nuke the whole cache,
1936 * which is a rather good thing for us.
1938 * Also, it is only used when turning caches on/off ("The expected
1939 * usage of the cache maintenance instructions that operate by set/way
1940 * is associated with the cache maintenance instructions associated
1941 * with the powerdown and powerup of caches, if this is required by
1942 * the implementation.").
1944 * We use the following policy:
1946 * - If we trap a S/W operation, we enable VM trapping to detect
1947 * caches being turned on/off, and do a full clean.
1949 * - We flush the caches on both caches being turned on and off.
1951 * - Once the caches are enabled, we stop trapping VM ops.
1953 void kvm_set_way_flush(struct kvm_vcpu *vcpu)
1955 unsigned long hcr = vcpu_get_hcr(vcpu);
1958 * If this is the first time we do a S/W operation
1959 * (i.e. HCR_TVM not set) flush the whole memory, and set the
1962 * Otherwise, rely on the VM trapping to wait for the MMU +
1963 * Caches to be turned off. At that point, we'll be able to
1964 * clean the caches again.
1966 if (!(hcr & HCR_TVM)) {
1967 trace_kvm_set_way_flush(*vcpu_pc(vcpu),
1968 vcpu_has_cache_enabled(vcpu));
1969 stage2_flush_vm(vcpu->kvm);
1970 vcpu_set_hcr(vcpu, hcr | HCR_TVM);
1974 void kvm_toggle_cache(struct kvm_vcpu *vcpu, bool was_enabled)
1976 bool now_enabled = vcpu_has_cache_enabled(vcpu);
1979 * If switching the MMU+caches on, need to invalidate the caches.
1980 * If switching it off, need to clean the caches.
1981 * Clean + invalidate does the trick always.
1983 if (now_enabled != was_enabled)
1984 stage2_flush_vm(vcpu->kvm);
1986 /* Caches are now on, stop trapping VM ops (until a S/W op) */
1988 vcpu_set_hcr(vcpu, vcpu_get_hcr(vcpu) & ~HCR_TVM);
1990 trace_kvm_toggle_cache(*vcpu_pc(vcpu), was_enabled, now_enabled);