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>
32 #include <asm/system_misc.h>
36 static pgd_t *boot_hyp_pgd;
37 static pgd_t *hyp_pgd;
38 static pgd_t *merged_hyp_pgd;
39 static DEFINE_MUTEX(kvm_hyp_pgd_mutex);
41 static unsigned long hyp_idmap_start;
42 static unsigned long hyp_idmap_end;
43 static phys_addr_t hyp_idmap_vector;
45 #define S2_PGD_SIZE (PTRS_PER_S2_PGD * sizeof(pgd_t))
46 #define hyp_pgd_order get_order(PTRS_PER_PGD * sizeof(pgd_t))
48 #define KVM_S2PTE_FLAG_IS_IOMAP (1UL << 0)
49 #define KVM_S2_FLAG_LOGGING_ACTIVE (1UL << 1)
51 static bool memslot_is_logging(struct kvm_memory_slot *memslot)
53 return memslot->dirty_bitmap && !(memslot->flags & KVM_MEM_READONLY);
57 * kvm_flush_remote_tlbs() - flush all VM TLB entries for v7/8
58 * @kvm: pointer to kvm structure.
60 * Interface to HYP function to flush all VM TLB entries
62 void kvm_flush_remote_tlbs(struct kvm *kvm)
64 kvm_call_hyp(__kvm_tlb_flush_vmid, kvm);
67 static void kvm_tlb_flush_vmid_ipa(struct kvm *kvm, phys_addr_t ipa)
69 kvm_call_hyp(__kvm_tlb_flush_vmid_ipa, kvm, ipa);
73 * D-Cache management functions. They take the page table entries by
74 * value, as they are flushing the cache using the kernel mapping (or
77 static void kvm_flush_dcache_pte(pte_t pte)
79 __kvm_flush_dcache_pte(pte);
82 static void kvm_flush_dcache_pmd(pmd_t pmd)
84 __kvm_flush_dcache_pmd(pmd);
87 static void kvm_flush_dcache_pud(pud_t pud)
89 __kvm_flush_dcache_pud(pud);
92 static bool kvm_is_device_pfn(unsigned long pfn)
94 return !pfn_valid(pfn);
98 * stage2_dissolve_pmd() - clear and flush huge PMD entry
99 * @kvm: pointer to kvm structure.
101 * @pmd: pmd pointer for IPA
103 * Function clears a PMD entry, flushes addr 1st and 2nd stage TLBs. Marks all
104 * pages in the range dirty.
106 static void stage2_dissolve_pmd(struct kvm *kvm, phys_addr_t addr, pmd_t *pmd)
108 if (!pmd_thp_or_huge(*pmd))
112 kvm_tlb_flush_vmid_ipa(kvm, addr);
113 put_page(virt_to_page(pmd));
116 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
121 BUG_ON(max > KVM_NR_MEM_OBJS);
122 if (cache->nobjs >= min)
124 while (cache->nobjs < max) {
125 page = (void *)__get_free_page(PGALLOC_GFP);
128 cache->objects[cache->nobjs++] = page;
133 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
136 free_page((unsigned long)mc->objects[--mc->nobjs]);
139 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc)
143 BUG_ON(!mc || !mc->nobjs);
144 p = mc->objects[--mc->nobjs];
148 static void clear_stage2_pgd_entry(struct kvm *kvm, pgd_t *pgd, phys_addr_t addr)
150 pud_t *pud_table __maybe_unused = stage2_pud_offset(pgd, 0UL);
151 stage2_pgd_clear(pgd);
152 kvm_tlb_flush_vmid_ipa(kvm, addr);
153 stage2_pud_free(pud_table);
154 put_page(virt_to_page(pgd));
157 static void clear_stage2_pud_entry(struct kvm *kvm, pud_t *pud, phys_addr_t addr)
159 pmd_t *pmd_table __maybe_unused = stage2_pmd_offset(pud, 0);
160 VM_BUG_ON(stage2_pud_huge(*pud));
161 stage2_pud_clear(pud);
162 kvm_tlb_flush_vmid_ipa(kvm, addr);
163 stage2_pmd_free(pmd_table);
164 put_page(virt_to_page(pud));
167 static void clear_stage2_pmd_entry(struct kvm *kvm, pmd_t *pmd, phys_addr_t addr)
169 pte_t *pte_table = pte_offset_kernel(pmd, 0);
170 VM_BUG_ON(pmd_thp_or_huge(*pmd));
172 kvm_tlb_flush_vmid_ipa(kvm, addr);
173 pte_free_kernel(NULL, pte_table);
174 put_page(virt_to_page(pmd));
178 * Unmapping vs dcache management:
180 * If a guest maps certain memory pages as uncached, all writes will
181 * bypass the data cache and go directly to RAM. However, the CPUs
182 * can still speculate reads (not writes) and fill cache lines with
185 * Those cache lines will be *clean* cache lines though, so a
186 * clean+invalidate operation is equivalent to an invalidate
187 * operation, because no cache lines are marked dirty.
189 * Those clean cache lines could be filled prior to an uncached write
190 * by the guest, and the cache coherent IO subsystem would therefore
191 * end up writing old data to disk.
193 * This is why right after unmapping a page/section and invalidating
194 * the corresponding TLBs, we call kvm_flush_dcache_p*() to make sure
195 * the IO subsystem will never hit in the cache.
197 static void unmap_stage2_ptes(struct kvm *kvm, pmd_t *pmd,
198 phys_addr_t addr, phys_addr_t end)
200 phys_addr_t start_addr = addr;
201 pte_t *pte, *start_pte;
203 start_pte = pte = pte_offset_kernel(pmd, addr);
205 if (!pte_none(*pte)) {
206 pte_t old_pte = *pte;
208 kvm_set_pte(pte, __pte(0));
209 kvm_tlb_flush_vmid_ipa(kvm, addr);
211 /* No need to invalidate the cache for device mappings */
212 if (!kvm_is_device_pfn(pte_pfn(old_pte)))
213 kvm_flush_dcache_pte(old_pte);
215 put_page(virt_to_page(pte));
217 } while (pte++, addr += PAGE_SIZE, addr != end);
219 if (stage2_pte_table_empty(start_pte))
220 clear_stage2_pmd_entry(kvm, pmd, start_addr);
223 static void unmap_stage2_pmds(struct kvm *kvm, pud_t *pud,
224 phys_addr_t addr, phys_addr_t end)
226 phys_addr_t next, start_addr = addr;
227 pmd_t *pmd, *start_pmd;
229 start_pmd = pmd = stage2_pmd_offset(pud, addr);
231 next = stage2_pmd_addr_end(addr, end);
232 if (!pmd_none(*pmd)) {
233 if (pmd_thp_or_huge(*pmd)) {
234 pmd_t old_pmd = *pmd;
237 kvm_tlb_flush_vmid_ipa(kvm, addr);
239 kvm_flush_dcache_pmd(old_pmd);
241 put_page(virt_to_page(pmd));
243 unmap_stage2_ptes(kvm, pmd, addr, next);
246 } while (pmd++, addr = next, addr != end);
248 if (stage2_pmd_table_empty(start_pmd))
249 clear_stage2_pud_entry(kvm, pud, start_addr);
252 static void unmap_stage2_puds(struct kvm *kvm, pgd_t *pgd,
253 phys_addr_t addr, phys_addr_t end)
255 phys_addr_t next, start_addr = addr;
256 pud_t *pud, *start_pud;
258 start_pud = pud = stage2_pud_offset(pgd, addr);
260 next = stage2_pud_addr_end(addr, end);
261 if (!stage2_pud_none(*pud)) {
262 if (stage2_pud_huge(*pud)) {
263 pud_t old_pud = *pud;
265 stage2_pud_clear(pud);
266 kvm_tlb_flush_vmid_ipa(kvm, addr);
267 kvm_flush_dcache_pud(old_pud);
268 put_page(virt_to_page(pud));
270 unmap_stage2_pmds(kvm, pud, addr, next);
273 } while (pud++, addr = next, addr != end);
275 if (stage2_pud_table_empty(start_pud))
276 clear_stage2_pgd_entry(kvm, pgd, start_addr);
280 * unmap_stage2_range -- Clear stage2 page table entries to unmap a range
281 * @kvm: The VM pointer
282 * @start: The intermediate physical base address of the range to unmap
283 * @size: The size of the area to unmap
285 * Clear a range of stage-2 mappings, lowering the various ref-counts. Must
286 * be called while holding mmu_lock (unless for freeing the stage2 pgd before
287 * destroying the VM), otherwise another faulting VCPU may come in and mess
288 * with things behind our backs.
290 static void unmap_stage2_range(struct kvm *kvm, phys_addr_t start, u64 size)
293 phys_addr_t addr = start, end = start + size;
296 assert_spin_locked(&kvm->mmu_lock);
297 pgd = kvm->arch.pgd + stage2_pgd_index(addr);
300 * Make sure the page table is still active, as another thread
301 * could have possibly freed the page table, while we released
304 if (!READ_ONCE(kvm->arch.pgd))
306 next = stage2_pgd_addr_end(addr, end);
307 if (!stage2_pgd_none(*pgd))
308 unmap_stage2_puds(kvm, pgd, addr, next);
310 * If the range is too large, release the kvm->mmu_lock
311 * to prevent starvation and lockup detector warnings.
314 cond_resched_lock(&kvm->mmu_lock);
315 } while (pgd++, addr = next, addr != end);
318 static void stage2_flush_ptes(struct kvm *kvm, pmd_t *pmd,
319 phys_addr_t addr, phys_addr_t end)
323 pte = pte_offset_kernel(pmd, addr);
325 if (!pte_none(*pte) && !kvm_is_device_pfn(pte_pfn(*pte)))
326 kvm_flush_dcache_pte(*pte);
327 } while (pte++, addr += PAGE_SIZE, addr != end);
330 static void stage2_flush_pmds(struct kvm *kvm, pud_t *pud,
331 phys_addr_t addr, phys_addr_t end)
336 pmd = stage2_pmd_offset(pud, addr);
338 next = stage2_pmd_addr_end(addr, end);
339 if (!pmd_none(*pmd)) {
340 if (pmd_thp_or_huge(*pmd))
341 kvm_flush_dcache_pmd(*pmd);
343 stage2_flush_ptes(kvm, pmd, addr, next);
345 } while (pmd++, addr = next, addr != end);
348 static void stage2_flush_puds(struct kvm *kvm, pgd_t *pgd,
349 phys_addr_t addr, phys_addr_t end)
354 pud = stage2_pud_offset(pgd, addr);
356 next = stage2_pud_addr_end(addr, end);
357 if (!stage2_pud_none(*pud)) {
358 if (stage2_pud_huge(*pud))
359 kvm_flush_dcache_pud(*pud);
361 stage2_flush_pmds(kvm, pud, addr, next);
363 } while (pud++, addr = next, addr != end);
366 static void stage2_flush_memslot(struct kvm *kvm,
367 struct kvm_memory_slot *memslot)
369 phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
370 phys_addr_t end = addr + PAGE_SIZE * memslot->npages;
374 pgd = kvm->arch.pgd + stage2_pgd_index(addr);
376 next = stage2_pgd_addr_end(addr, end);
377 stage2_flush_puds(kvm, pgd, addr, next);
378 } while (pgd++, addr = next, addr != end);
382 * stage2_flush_vm - Invalidate cache for pages mapped in stage 2
383 * @kvm: The struct kvm pointer
385 * Go through the stage 2 page tables and invalidate any cache lines
386 * backing memory already mapped to the VM.
388 static void stage2_flush_vm(struct kvm *kvm)
390 struct kvm_memslots *slots;
391 struct kvm_memory_slot *memslot;
394 idx = srcu_read_lock(&kvm->srcu);
395 spin_lock(&kvm->mmu_lock);
397 slots = kvm_memslots(kvm);
398 kvm_for_each_memslot(memslot, slots)
399 stage2_flush_memslot(kvm, memslot);
401 spin_unlock(&kvm->mmu_lock);
402 srcu_read_unlock(&kvm->srcu, idx);
405 static void clear_hyp_pgd_entry(pgd_t *pgd)
407 pud_t *pud_table __maybe_unused = pud_offset(pgd, 0UL);
409 pud_free(NULL, pud_table);
410 put_page(virt_to_page(pgd));
413 static void clear_hyp_pud_entry(pud_t *pud)
415 pmd_t *pmd_table __maybe_unused = pmd_offset(pud, 0);
416 VM_BUG_ON(pud_huge(*pud));
418 pmd_free(NULL, pmd_table);
419 put_page(virt_to_page(pud));
422 static void clear_hyp_pmd_entry(pmd_t *pmd)
424 pte_t *pte_table = pte_offset_kernel(pmd, 0);
425 VM_BUG_ON(pmd_thp_or_huge(*pmd));
427 pte_free_kernel(NULL, pte_table);
428 put_page(virt_to_page(pmd));
431 static void unmap_hyp_ptes(pmd_t *pmd, phys_addr_t addr, phys_addr_t end)
433 pte_t *pte, *start_pte;
435 start_pte = pte = pte_offset_kernel(pmd, addr);
437 if (!pte_none(*pte)) {
438 kvm_set_pte(pte, __pte(0));
439 put_page(virt_to_page(pte));
441 } while (pte++, addr += PAGE_SIZE, addr != end);
443 if (hyp_pte_table_empty(start_pte))
444 clear_hyp_pmd_entry(pmd);
447 static void unmap_hyp_pmds(pud_t *pud, phys_addr_t addr, phys_addr_t end)
450 pmd_t *pmd, *start_pmd;
452 start_pmd = pmd = pmd_offset(pud, addr);
454 next = pmd_addr_end(addr, end);
455 /* Hyp doesn't use huge pmds */
457 unmap_hyp_ptes(pmd, addr, next);
458 } while (pmd++, addr = next, addr != end);
460 if (hyp_pmd_table_empty(start_pmd))
461 clear_hyp_pud_entry(pud);
464 static void unmap_hyp_puds(pgd_t *pgd, phys_addr_t addr, phys_addr_t end)
467 pud_t *pud, *start_pud;
469 start_pud = pud = pud_offset(pgd, addr);
471 next = pud_addr_end(addr, end);
472 /* Hyp doesn't use huge puds */
474 unmap_hyp_pmds(pud, addr, next);
475 } while (pud++, addr = next, addr != end);
477 if (hyp_pud_table_empty(start_pud))
478 clear_hyp_pgd_entry(pgd);
481 static void unmap_hyp_range(pgd_t *pgdp, phys_addr_t start, u64 size)
484 phys_addr_t addr = start, end = start + size;
488 * We don't unmap anything from HYP, except at the hyp tear down.
489 * Hence, we don't have to invalidate the TLBs here.
491 pgd = pgdp + pgd_index(addr);
493 next = pgd_addr_end(addr, end);
495 unmap_hyp_puds(pgd, addr, next);
496 } while (pgd++, addr = next, addr != end);
500 * free_hyp_pgds - free Hyp-mode page tables
502 * Assumes hyp_pgd is a page table used strictly in Hyp-mode and
503 * therefore contains either mappings in the kernel memory area (above
504 * PAGE_OFFSET), or device mappings in the vmalloc range (from
505 * VMALLOC_START to VMALLOC_END).
507 * boot_hyp_pgd should only map two pages for the init code.
509 void free_hyp_pgds(void)
513 mutex_lock(&kvm_hyp_pgd_mutex);
516 unmap_hyp_range(boot_hyp_pgd, hyp_idmap_start, PAGE_SIZE);
517 free_pages((unsigned long)boot_hyp_pgd, hyp_pgd_order);
522 unmap_hyp_range(hyp_pgd, hyp_idmap_start, PAGE_SIZE);
523 for (addr = PAGE_OFFSET; virt_addr_valid(addr); addr += PGDIR_SIZE)
524 unmap_hyp_range(hyp_pgd, kern_hyp_va(addr), PGDIR_SIZE);
525 for (addr = VMALLOC_START; is_vmalloc_addr((void*)addr); addr += PGDIR_SIZE)
526 unmap_hyp_range(hyp_pgd, kern_hyp_va(addr), PGDIR_SIZE);
528 free_pages((unsigned long)hyp_pgd, hyp_pgd_order);
531 if (merged_hyp_pgd) {
532 clear_page(merged_hyp_pgd);
533 free_page((unsigned long)merged_hyp_pgd);
534 merged_hyp_pgd = NULL;
537 mutex_unlock(&kvm_hyp_pgd_mutex);
540 static void create_hyp_pte_mappings(pmd_t *pmd, unsigned long start,
541 unsigned long end, unsigned long pfn,
549 pte = pte_offset_kernel(pmd, addr);
550 kvm_set_pte(pte, pfn_pte(pfn, prot));
551 get_page(virt_to_page(pte));
552 kvm_flush_dcache_to_poc(pte, sizeof(*pte));
554 } while (addr += PAGE_SIZE, addr != end);
557 static int create_hyp_pmd_mappings(pud_t *pud, unsigned long start,
558 unsigned long end, unsigned long pfn,
563 unsigned long addr, next;
567 pmd = pmd_offset(pud, addr);
569 BUG_ON(pmd_sect(*pmd));
571 if (pmd_none(*pmd)) {
572 pte = pte_alloc_one_kernel(NULL, addr);
574 kvm_err("Cannot allocate Hyp pte\n");
577 pmd_populate_kernel(NULL, pmd, pte);
578 get_page(virt_to_page(pmd));
579 kvm_flush_dcache_to_poc(pmd, sizeof(*pmd));
582 next = pmd_addr_end(addr, end);
584 create_hyp_pte_mappings(pmd, addr, next, pfn, prot);
585 pfn += (next - addr) >> PAGE_SHIFT;
586 } while (addr = next, addr != end);
591 static int create_hyp_pud_mappings(pgd_t *pgd, unsigned long start,
592 unsigned long end, unsigned long pfn,
597 unsigned long addr, next;
602 pud = pud_offset(pgd, addr);
604 if (pud_none_or_clear_bad(pud)) {
605 pmd = pmd_alloc_one(NULL, addr);
607 kvm_err("Cannot allocate Hyp pmd\n");
610 pud_populate(NULL, pud, pmd);
611 get_page(virt_to_page(pud));
612 kvm_flush_dcache_to_poc(pud, sizeof(*pud));
615 next = pud_addr_end(addr, end);
616 ret = create_hyp_pmd_mappings(pud, addr, next, pfn, prot);
619 pfn += (next - addr) >> PAGE_SHIFT;
620 } while (addr = next, addr != end);
625 static int __create_hyp_mappings(pgd_t *pgdp,
626 unsigned long start, unsigned long end,
627 unsigned long pfn, pgprot_t prot)
631 unsigned long addr, next;
634 mutex_lock(&kvm_hyp_pgd_mutex);
635 addr = start & PAGE_MASK;
636 end = PAGE_ALIGN(end);
638 pgd = pgdp + pgd_index(addr);
640 if (pgd_none(*pgd)) {
641 pud = pud_alloc_one(NULL, addr);
643 kvm_err("Cannot allocate Hyp pud\n");
647 pgd_populate(NULL, pgd, pud);
648 get_page(virt_to_page(pgd));
649 kvm_flush_dcache_to_poc(pgd, sizeof(*pgd));
652 next = pgd_addr_end(addr, end);
653 err = create_hyp_pud_mappings(pgd, addr, next, pfn, prot);
656 pfn += (next - addr) >> PAGE_SHIFT;
657 } while (addr = next, addr != end);
659 mutex_unlock(&kvm_hyp_pgd_mutex);
663 static phys_addr_t kvm_kaddr_to_phys(void *kaddr)
665 if (!is_vmalloc_addr(kaddr)) {
666 BUG_ON(!virt_addr_valid(kaddr));
669 return page_to_phys(vmalloc_to_page(kaddr)) +
670 offset_in_page(kaddr);
675 * create_hyp_mappings - duplicate a kernel virtual address range in Hyp mode
676 * @from: The virtual kernel start address of the range
677 * @to: The virtual kernel end address of the range (exclusive)
678 * @prot: The protection to be applied to this range
680 * The same virtual address as the kernel virtual address is also used
681 * in Hyp-mode mapping (modulo HYP_PAGE_OFFSET) to the same underlying
684 int create_hyp_mappings(void *from, void *to, pgprot_t prot)
686 phys_addr_t phys_addr;
687 unsigned long virt_addr;
688 unsigned long start = kern_hyp_va((unsigned long)from);
689 unsigned long end = kern_hyp_va((unsigned long)to);
691 if (is_kernel_in_hyp_mode())
694 start = start & PAGE_MASK;
695 end = PAGE_ALIGN(end);
697 for (virt_addr = start; virt_addr < end; virt_addr += PAGE_SIZE) {
700 phys_addr = kvm_kaddr_to_phys(from + virt_addr - start);
701 err = __create_hyp_mappings(hyp_pgd, virt_addr,
702 virt_addr + PAGE_SIZE,
703 __phys_to_pfn(phys_addr),
713 * create_hyp_io_mappings - duplicate a kernel IO mapping into Hyp mode
714 * @from: The kernel start VA of the range
715 * @to: The kernel end VA of the range (exclusive)
716 * @phys_addr: The physical start address which gets mapped
718 * The resulting HYP VA is the same as the kernel VA, modulo
721 int create_hyp_io_mappings(void *from, void *to, phys_addr_t phys_addr)
723 unsigned long start = kern_hyp_va((unsigned long)from);
724 unsigned long end = kern_hyp_va((unsigned long)to);
726 if (is_kernel_in_hyp_mode())
729 /* Check for a valid kernel IO mapping */
730 if (!is_vmalloc_addr(from) || !is_vmalloc_addr(to - 1))
733 return __create_hyp_mappings(hyp_pgd, start, end,
734 __phys_to_pfn(phys_addr), PAGE_HYP_DEVICE);
738 * kvm_alloc_stage2_pgd - allocate level-1 table for stage-2 translation.
739 * @kvm: The KVM struct pointer for the VM.
741 * Allocates only the stage-2 HW PGD level table(s) (can support either full
742 * 40-bit input addresses or limited to 32-bit input addresses). Clears the
745 * Note we don't need locking here as this is only called when the VM is
746 * created, which can only be done once.
748 int kvm_alloc_stage2_pgd(struct kvm *kvm)
752 if (kvm->arch.pgd != NULL) {
753 kvm_err("kvm_arch already initialized?\n");
757 /* Allocate the HW PGD, making sure that each page gets its own refcount */
758 pgd = alloc_pages_exact(S2_PGD_SIZE, GFP_KERNEL | __GFP_ZERO);
766 static void stage2_unmap_memslot(struct kvm *kvm,
767 struct kvm_memory_slot *memslot)
769 hva_t hva = memslot->userspace_addr;
770 phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
771 phys_addr_t size = PAGE_SIZE * memslot->npages;
772 hva_t reg_end = hva + size;
775 * A memory region could potentially cover multiple VMAs, and any holes
776 * between them, so iterate over all of them to find out if we should
779 * +--------------------------------------------+
780 * +---------------+----------------+ +----------------+
781 * | : VMA 1 | VMA 2 | | VMA 3 : |
782 * +---------------+----------------+ +----------------+
784 * +--------------------------------------------+
787 struct vm_area_struct *vma = find_vma(current->mm, hva);
788 hva_t vm_start, vm_end;
790 if (!vma || vma->vm_start >= reg_end)
794 * Take the intersection of this VMA with the memory region
796 vm_start = max(hva, vma->vm_start);
797 vm_end = min(reg_end, vma->vm_end);
799 if (!(vma->vm_flags & VM_PFNMAP)) {
800 gpa_t gpa = addr + (vm_start - memslot->userspace_addr);
801 unmap_stage2_range(kvm, gpa, vm_end - vm_start);
804 } while (hva < reg_end);
808 * stage2_unmap_vm - Unmap Stage-2 RAM mappings
809 * @kvm: The struct kvm pointer
811 * Go through the memregions and unmap any reguler RAM
812 * backing memory already mapped to the VM.
814 void stage2_unmap_vm(struct kvm *kvm)
816 struct kvm_memslots *slots;
817 struct kvm_memory_slot *memslot;
820 idx = srcu_read_lock(&kvm->srcu);
821 down_read(¤t->mm->mmap_sem);
822 spin_lock(&kvm->mmu_lock);
824 slots = kvm_memslots(kvm);
825 kvm_for_each_memslot(memslot, slots)
826 stage2_unmap_memslot(kvm, memslot);
828 spin_unlock(&kvm->mmu_lock);
829 up_read(¤t->mm->mmap_sem);
830 srcu_read_unlock(&kvm->srcu, idx);
834 * kvm_free_stage2_pgd - free all stage-2 tables
835 * @kvm: The KVM struct pointer for the VM.
837 * Walks the level-1 page table pointed to by kvm->arch.pgd and frees all
838 * underlying level-2 and level-3 tables before freeing the actual level-1 table
839 * and setting the struct pointer to NULL.
841 void kvm_free_stage2_pgd(struct kvm *kvm)
845 spin_lock(&kvm->mmu_lock);
847 unmap_stage2_range(kvm, 0, KVM_PHYS_SIZE);
848 pgd = READ_ONCE(kvm->arch.pgd);
849 kvm->arch.pgd = NULL;
851 spin_unlock(&kvm->mmu_lock);
853 /* Free the HW pgd, one page at a time */
855 free_pages_exact(pgd, S2_PGD_SIZE);
858 static pud_t *stage2_get_pud(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
864 pgd = kvm->arch.pgd + stage2_pgd_index(addr);
865 if (WARN_ON(stage2_pgd_none(*pgd))) {
868 pud = mmu_memory_cache_alloc(cache);
869 stage2_pgd_populate(pgd, pud);
870 get_page(virt_to_page(pgd));
873 return stage2_pud_offset(pgd, addr);
876 static pmd_t *stage2_get_pmd(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
882 pud = stage2_get_pud(kvm, cache, addr);
886 if (stage2_pud_none(*pud)) {
889 pmd = mmu_memory_cache_alloc(cache);
890 stage2_pud_populate(pud, pmd);
891 get_page(virt_to_page(pud));
894 return stage2_pmd_offset(pud, addr);
897 static int stage2_set_pmd_huge(struct kvm *kvm, struct kvm_mmu_memory_cache
898 *cache, phys_addr_t addr, const pmd_t *new_pmd)
902 pmd = stage2_get_pmd(kvm, cache, addr);
906 * Mapping in huge pages should only happen through a fault. If a
907 * page is merged into a transparent huge page, the individual
908 * subpages of that huge page should be unmapped through MMU
909 * notifiers before we get here.
911 * Merging of CompoundPages is not supported; they should become
912 * splitting first, unmapped, merged, and mapped back in on-demand.
914 VM_BUG_ON(pmd_present(*pmd) && pmd_pfn(*pmd) != pmd_pfn(*new_pmd));
917 if (pmd_present(old_pmd)) {
919 kvm_tlb_flush_vmid_ipa(kvm, addr);
921 get_page(virt_to_page(pmd));
924 kvm_set_pmd(pmd, *new_pmd);
928 static int stage2_set_pte(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
929 phys_addr_t addr, const pte_t *new_pte,
934 bool iomap = flags & KVM_S2PTE_FLAG_IS_IOMAP;
935 bool logging_active = flags & KVM_S2_FLAG_LOGGING_ACTIVE;
937 VM_BUG_ON(logging_active && !cache);
939 /* Create stage-2 page table mapping - Levels 0 and 1 */
940 pmd = stage2_get_pmd(kvm, cache, addr);
943 * Ignore calls from kvm_set_spte_hva for unallocated
950 * While dirty page logging - dissolve huge PMD, then continue on to
954 stage2_dissolve_pmd(kvm, addr, pmd);
956 /* Create stage-2 page mappings - Level 2 */
957 if (pmd_none(*pmd)) {
959 return 0; /* ignore calls from kvm_set_spte_hva */
960 pte = mmu_memory_cache_alloc(cache);
961 pmd_populate_kernel(NULL, pmd, pte);
962 get_page(virt_to_page(pmd));
965 pte = pte_offset_kernel(pmd, addr);
967 if (iomap && pte_present(*pte))
970 /* Create 2nd stage page table mapping - Level 3 */
972 if (pte_present(old_pte)) {
973 kvm_set_pte(pte, __pte(0));
974 kvm_tlb_flush_vmid_ipa(kvm, addr);
976 get_page(virt_to_page(pte));
979 kvm_set_pte(pte, *new_pte);
983 #ifndef __HAVE_ARCH_PTEP_TEST_AND_CLEAR_YOUNG
984 static int stage2_ptep_test_and_clear_young(pte_t *pte)
986 if (pte_young(*pte)) {
987 *pte = pte_mkold(*pte);
993 static int stage2_ptep_test_and_clear_young(pte_t *pte)
995 return __ptep_test_and_clear_young(pte);
999 static int stage2_pmdp_test_and_clear_young(pmd_t *pmd)
1001 return stage2_ptep_test_and_clear_young((pte_t *)pmd);
1005 * kvm_phys_addr_ioremap - map a device range to guest IPA
1007 * @kvm: The KVM pointer
1008 * @guest_ipa: The IPA at which to insert the mapping
1009 * @pa: The physical address of the device
1010 * @size: The size of the mapping
1012 int kvm_phys_addr_ioremap(struct kvm *kvm, phys_addr_t guest_ipa,
1013 phys_addr_t pa, unsigned long size, bool writable)
1015 phys_addr_t addr, end;
1018 struct kvm_mmu_memory_cache cache = { 0, };
1020 end = (guest_ipa + size + PAGE_SIZE - 1) & PAGE_MASK;
1021 pfn = __phys_to_pfn(pa);
1023 for (addr = guest_ipa; addr < end; addr += PAGE_SIZE) {
1024 pte_t pte = pfn_pte(pfn, PAGE_S2_DEVICE);
1027 pte = kvm_s2pte_mkwrite(pte);
1029 ret = mmu_topup_memory_cache(&cache, KVM_MMU_CACHE_MIN_PAGES,
1033 spin_lock(&kvm->mmu_lock);
1034 ret = stage2_set_pte(kvm, &cache, addr, &pte,
1035 KVM_S2PTE_FLAG_IS_IOMAP);
1036 spin_unlock(&kvm->mmu_lock);
1044 mmu_free_memory_cache(&cache);
1048 static bool transparent_hugepage_adjust(kvm_pfn_t *pfnp, phys_addr_t *ipap)
1050 kvm_pfn_t pfn = *pfnp;
1051 gfn_t gfn = *ipap >> PAGE_SHIFT;
1053 if (PageTransCompoundMap(pfn_to_page(pfn))) {
1056 * The address we faulted on is backed by a transparent huge
1057 * page. However, because we map the compound huge page and
1058 * not the individual tail page, we need to transfer the
1059 * refcount to the head page. We have to be careful that the
1060 * THP doesn't start to split while we are adjusting the
1063 * We are sure this doesn't happen, because mmu_notifier_retry
1064 * was successful and we are holding the mmu_lock, so if this
1065 * THP is trying to split, it will be blocked in the mmu
1066 * notifier before touching any of the pages, specifically
1067 * before being able to call __split_huge_page_refcount().
1069 * We can therefore safely transfer the refcount from PG_tail
1070 * to PG_head and switch the pfn from a tail page to the head
1073 mask = PTRS_PER_PMD - 1;
1074 VM_BUG_ON((gfn & mask) != (pfn & mask));
1077 kvm_release_pfn_clean(pfn);
1089 static bool kvm_is_write_fault(struct kvm_vcpu *vcpu)
1091 if (kvm_vcpu_trap_is_iabt(vcpu))
1094 return kvm_vcpu_dabt_iswrite(vcpu);
1098 * stage2_wp_ptes - write protect PMD range
1099 * @pmd: pointer to pmd entry
1100 * @addr: range start address
1101 * @end: range end address
1103 static void stage2_wp_ptes(pmd_t *pmd, phys_addr_t addr, phys_addr_t end)
1107 pte = pte_offset_kernel(pmd, addr);
1109 if (!pte_none(*pte)) {
1110 if (!kvm_s2pte_readonly(pte))
1111 kvm_set_s2pte_readonly(pte);
1113 } while (pte++, addr += PAGE_SIZE, addr != end);
1117 * stage2_wp_pmds - write protect PUD range
1118 * @pud: pointer to pud entry
1119 * @addr: range start address
1120 * @end: range end address
1122 static void stage2_wp_pmds(pud_t *pud, phys_addr_t addr, phys_addr_t end)
1127 pmd = stage2_pmd_offset(pud, addr);
1130 next = stage2_pmd_addr_end(addr, end);
1131 if (!pmd_none(*pmd)) {
1132 if (pmd_thp_or_huge(*pmd)) {
1133 if (!kvm_s2pmd_readonly(pmd))
1134 kvm_set_s2pmd_readonly(pmd);
1136 stage2_wp_ptes(pmd, addr, next);
1139 } while (pmd++, addr = next, addr != end);
1143 * stage2_wp_puds - write protect PGD range
1144 * @pgd: pointer to pgd entry
1145 * @addr: range start address
1146 * @end: range end address
1148 * Process PUD entries, for a huge PUD we cause a panic.
1150 static void stage2_wp_puds(pgd_t *pgd, phys_addr_t addr, phys_addr_t end)
1155 pud = stage2_pud_offset(pgd, addr);
1157 next = stage2_pud_addr_end(addr, end);
1158 if (!stage2_pud_none(*pud)) {
1159 /* TODO:PUD not supported, revisit later if supported */
1160 BUG_ON(stage2_pud_huge(*pud));
1161 stage2_wp_pmds(pud, addr, next);
1163 } while (pud++, addr = next, addr != end);
1167 * stage2_wp_range() - write protect stage2 memory region range
1168 * @kvm: The KVM pointer
1169 * @addr: Start address of range
1170 * @end: End address of range
1172 static void stage2_wp_range(struct kvm *kvm, phys_addr_t addr, phys_addr_t end)
1177 pgd = kvm->arch.pgd + stage2_pgd_index(addr);
1180 * Release kvm_mmu_lock periodically if the memory region is
1181 * large. Otherwise, we may see kernel panics with
1182 * CONFIG_DETECT_HUNG_TASK, CONFIG_LOCKUP_DETECTOR,
1183 * CONFIG_LOCKDEP. Additionally, holding the lock too long
1184 * will also starve other vCPUs. We have to also make sure
1185 * that the page tables are not freed while we released
1188 cond_resched_lock(&kvm->mmu_lock);
1189 if (!READ_ONCE(kvm->arch.pgd))
1191 next = stage2_pgd_addr_end(addr, end);
1192 if (stage2_pgd_present(*pgd))
1193 stage2_wp_puds(pgd, addr, next);
1194 } while (pgd++, addr = next, addr != end);
1198 * kvm_mmu_wp_memory_region() - write protect stage 2 entries for memory slot
1199 * @kvm: The KVM pointer
1200 * @slot: The memory slot to write protect
1202 * Called to start logging dirty pages after memory region
1203 * KVM_MEM_LOG_DIRTY_PAGES operation is called. After this function returns
1204 * all present PMD and PTEs are write protected in the memory region.
1205 * Afterwards read of dirty page log can be called.
1207 * Acquires kvm_mmu_lock. Called with kvm->slots_lock mutex acquired,
1208 * serializing operations for VM memory regions.
1210 void kvm_mmu_wp_memory_region(struct kvm *kvm, int slot)
1212 struct kvm_memslots *slots = kvm_memslots(kvm);
1213 struct kvm_memory_slot *memslot = id_to_memslot(slots, slot);
1214 phys_addr_t start = memslot->base_gfn << PAGE_SHIFT;
1215 phys_addr_t end = (memslot->base_gfn + memslot->npages) << PAGE_SHIFT;
1217 spin_lock(&kvm->mmu_lock);
1218 stage2_wp_range(kvm, start, end);
1219 spin_unlock(&kvm->mmu_lock);
1220 kvm_flush_remote_tlbs(kvm);
1224 * kvm_mmu_write_protect_pt_masked() - write protect dirty pages
1225 * @kvm: The KVM pointer
1226 * @slot: The memory slot associated with mask
1227 * @gfn_offset: The gfn offset in memory slot
1228 * @mask: The mask of dirty pages at offset 'gfn_offset' in this memory
1229 * slot to be write protected
1231 * Walks bits set in mask write protects the associated pte's. Caller must
1232 * acquire kvm_mmu_lock.
1234 static void kvm_mmu_write_protect_pt_masked(struct kvm *kvm,
1235 struct kvm_memory_slot *slot,
1236 gfn_t gfn_offset, unsigned long mask)
1238 phys_addr_t base_gfn = slot->base_gfn + gfn_offset;
1239 phys_addr_t start = (base_gfn + __ffs(mask)) << PAGE_SHIFT;
1240 phys_addr_t end = (base_gfn + __fls(mask) + 1) << PAGE_SHIFT;
1242 stage2_wp_range(kvm, start, end);
1246 * kvm_arch_mmu_enable_log_dirty_pt_masked - enable dirty logging for selected
1249 * It calls kvm_mmu_write_protect_pt_masked to write protect selected pages to
1250 * enable dirty logging for them.
1252 void kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm *kvm,
1253 struct kvm_memory_slot *slot,
1254 gfn_t gfn_offset, unsigned long mask)
1256 kvm_mmu_write_protect_pt_masked(kvm, slot, gfn_offset, mask);
1259 static void coherent_cache_guest_page(struct kvm_vcpu *vcpu, kvm_pfn_t pfn,
1262 __coherent_cache_guest_page(vcpu, pfn, size);
1265 static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
1266 struct kvm_memory_slot *memslot, unsigned long hva,
1267 unsigned long fault_status)
1270 bool write_fault, writable, hugetlb = false, force_pte = false;
1271 unsigned long mmu_seq;
1272 gfn_t gfn = fault_ipa >> PAGE_SHIFT;
1273 struct kvm *kvm = vcpu->kvm;
1274 struct kvm_mmu_memory_cache *memcache = &vcpu->arch.mmu_page_cache;
1275 struct vm_area_struct *vma;
1277 pgprot_t mem_type = PAGE_S2;
1278 bool logging_active = memslot_is_logging(memslot);
1279 unsigned long flags = 0;
1281 write_fault = kvm_is_write_fault(vcpu);
1282 if (fault_status == FSC_PERM && !write_fault) {
1283 kvm_err("Unexpected L2 read permission error\n");
1287 /* Let's check if we will get back a huge page backed by hugetlbfs */
1288 down_read(¤t->mm->mmap_sem);
1289 vma = find_vma_intersection(current->mm, hva, hva + 1);
1290 if (unlikely(!vma)) {
1291 kvm_err("Failed to find VMA for hva 0x%lx\n", hva);
1292 up_read(¤t->mm->mmap_sem);
1296 if (is_vm_hugetlb_page(vma) && !logging_active) {
1298 gfn = (fault_ipa & PMD_MASK) >> PAGE_SHIFT;
1301 * Pages belonging to memslots that don't have the same
1302 * alignment for userspace and IPA cannot be mapped using
1303 * block descriptors even if the pages belong to a THP for
1304 * the process, because the stage-2 block descriptor will
1305 * cover more than a single THP and we loose atomicity for
1306 * unmapping, updates, and splits of the THP or other pages
1307 * in the stage-2 block range.
1309 if ((memslot->userspace_addr & ~PMD_MASK) !=
1310 ((memslot->base_gfn << PAGE_SHIFT) & ~PMD_MASK))
1313 up_read(¤t->mm->mmap_sem);
1315 /* We need minimum second+third level pages */
1316 ret = mmu_topup_memory_cache(memcache, KVM_MMU_CACHE_MIN_PAGES,
1321 mmu_seq = vcpu->kvm->mmu_notifier_seq;
1323 * Ensure the read of mmu_notifier_seq happens before we call
1324 * gfn_to_pfn_prot (which calls get_user_pages), so that we don't risk
1325 * the page we just got a reference to gets unmapped before we have a
1326 * chance to grab the mmu_lock, which ensure that if the page gets
1327 * unmapped afterwards, the call to kvm_unmap_hva will take it away
1328 * from us again properly. This smp_rmb() interacts with the smp_wmb()
1329 * in kvm_mmu_notifier_invalidate_<page|range_end>.
1333 pfn = gfn_to_pfn_prot(kvm, gfn, write_fault, &writable);
1334 if (is_error_noslot_pfn(pfn))
1337 if (kvm_is_device_pfn(pfn)) {
1338 mem_type = PAGE_S2_DEVICE;
1339 flags |= KVM_S2PTE_FLAG_IS_IOMAP;
1340 } else if (logging_active) {
1342 * Faults on pages in a memslot with logging enabled
1343 * should not be mapped with huge pages (it introduces churn
1344 * and performance degradation), so force a pte mapping.
1347 flags |= KVM_S2_FLAG_LOGGING_ACTIVE;
1350 * Only actually map the page as writable if this was a write
1357 spin_lock(&kvm->mmu_lock);
1358 if (mmu_notifier_retry(kvm, mmu_seq))
1361 if (!hugetlb && !force_pte)
1362 hugetlb = transparent_hugepage_adjust(&pfn, &fault_ipa);
1365 pmd_t new_pmd = pfn_pmd(pfn, mem_type);
1366 new_pmd = pmd_mkhuge(new_pmd);
1368 new_pmd = kvm_s2pmd_mkwrite(new_pmd);
1369 kvm_set_pfn_dirty(pfn);
1371 coherent_cache_guest_page(vcpu, pfn, PMD_SIZE);
1372 ret = stage2_set_pmd_huge(kvm, memcache, fault_ipa, &new_pmd);
1374 pte_t new_pte = pfn_pte(pfn, mem_type);
1377 new_pte = kvm_s2pte_mkwrite(new_pte);
1378 kvm_set_pfn_dirty(pfn);
1379 mark_page_dirty(kvm, gfn);
1381 coherent_cache_guest_page(vcpu, pfn, PAGE_SIZE);
1382 ret = stage2_set_pte(kvm, memcache, fault_ipa, &new_pte, flags);
1386 spin_unlock(&kvm->mmu_lock);
1387 kvm_set_pfn_accessed(pfn);
1388 kvm_release_pfn_clean(pfn);
1393 * Resolve the access fault by making the page young again.
1394 * Note that because the faulting entry is guaranteed not to be
1395 * cached in the TLB, we don't need to invalidate anything.
1396 * Only the HW Access Flag updates are supported for Stage 2 (no DBM),
1397 * so there is no need for atomic (pte|pmd)_mkyoung operations.
1399 static void handle_access_fault(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa)
1404 bool pfn_valid = false;
1406 trace_kvm_access_fault(fault_ipa);
1408 spin_lock(&vcpu->kvm->mmu_lock);
1410 pmd = stage2_get_pmd(vcpu->kvm, NULL, fault_ipa);
1411 if (!pmd || pmd_none(*pmd)) /* Nothing there */
1414 if (pmd_thp_or_huge(*pmd)) { /* THP, HugeTLB */
1415 *pmd = pmd_mkyoung(*pmd);
1416 pfn = pmd_pfn(*pmd);
1421 pte = pte_offset_kernel(pmd, fault_ipa);
1422 if (pte_none(*pte)) /* Nothing there either */
1425 *pte = pte_mkyoung(*pte); /* Just a page... */
1426 pfn = pte_pfn(*pte);
1429 spin_unlock(&vcpu->kvm->mmu_lock);
1431 kvm_set_pfn_accessed(pfn);
1434 static bool is_abort_sea(unsigned long fault_status)
1436 switch (fault_status) {
1454 * kvm_handle_guest_abort - handles all 2nd stage aborts
1455 * @vcpu: the VCPU pointer
1456 * @run: the kvm_run structure
1458 * Any abort that gets to the host is almost guaranteed to be caused by a
1459 * missing second stage translation table entry, which can mean that either the
1460 * guest simply needs more memory and we must allocate an appropriate page or it
1461 * can mean that the guest tried to access I/O memory, which is emulated by user
1462 * space. The distinction is based on the IPA causing the fault and whether this
1463 * memory region has been registered as standard RAM by user space.
1465 int kvm_handle_guest_abort(struct kvm_vcpu *vcpu, struct kvm_run *run)
1467 unsigned long fault_status;
1468 phys_addr_t fault_ipa;
1469 struct kvm_memory_slot *memslot;
1471 bool is_iabt, write_fault, writable;
1475 fault_status = kvm_vcpu_trap_get_fault_type(vcpu);
1477 fault_ipa = kvm_vcpu_get_fault_ipa(vcpu);
1480 * The host kernel will handle the synchronous external abort. There
1481 * is no need to pass the error into the guest.
1483 if (is_abort_sea(fault_status)) {
1484 if (!handle_guest_sea(fault_ipa, kvm_vcpu_get_hsr(vcpu)))
1488 is_iabt = kvm_vcpu_trap_is_iabt(vcpu);
1489 if (unlikely(!is_iabt && kvm_vcpu_dabt_isextabt(vcpu))) {
1490 kvm_inject_vabt(vcpu);
1494 trace_kvm_guest_fault(*vcpu_pc(vcpu), kvm_vcpu_get_hsr(vcpu),
1495 kvm_vcpu_get_hfar(vcpu), fault_ipa);
1497 /* Check the stage-2 fault is trans. fault or write fault */
1498 if (fault_status != FSC_FAULT && fault_status != FSC_PERM &&
1499 fault_status != FSC_ACCESS) {
1500 kvm_err("Unsupported FSC: EC=%#x xFSC=%#lx ESR_EL2=%#lx\n",
1501 kvm_vcpu_trap_get_class(vcpu),
1502 (unsigned long)kvm_vcpu_trap_get_fault(vcpu),
1503 (unsigned long)kvm_vcpu_get_hsr(vcpu));
1507 idx = srcu_read_lock(&vcpu->kvm->srcu);
1509 gfn = fault_ipa >> PAGE_SHIFT;
1510 memslot = gfn_to_memslot(vcpu->kvm, gfn);
1511 hva = gfn_to_hva_memslot_prot(memslot, gfn, &writable);
1512 write_fault = kvm_is_write_fault(vcpu);
1513 if (kvm_is_error_hva(hva) || (write_fault && !writable)) {
1515 /* Prefetch Abort on I/O address */
1516 kvm_inject_pabt(vcpu, kvm_vcpu_get_hfar(vcpu));
1522 * Check for a cache maintenance operation. Since we
1523 * ended-up here, we know it is outside of any memory
1524 * slot. But we can't find out if that is for a device,
1525 * or if the guest is just being stupid. The only thing
1526 * we know for sure is that this range cannot be cached.
1528 * So let's assume that the guest is just being
1529 * cautious, and skip the instruction.
1531 if (kvm_vcpu_dabt_is_cm(vcpu)) {
1532 kvm_skip_instr(vcpu, kvm_vcpu_trap_il_is32bit(vcpu));
1538 * The IPA is reported as [MAX:12], so we need to
1539 * complement it with the bottom 12 bits from the
1540 * faulting VA. This is always 12 bits, irrespective
1543 fault_ipa |= kvm_vcpu_get_hfar(vcpu) & ((1 << 12) - 1);
1544 ret = io_mem_abort(vcpu, run, fault_ipa);
1548 /* Userspace should not be able to register out-of-bounds IPAs */
1549 VM_BUG_ON(fault_ipa >= KVM_PHYS_SIZE);
1551 if (fault_status == FSC_ACCESS) {
1552 handle_access_fault(vcpu, fault_ipa);
1557 ret = user_mem_abort(vcpu, fault_ipa, memslot, hva, fault_status);
1561 srcu_read_unlock(&vcpu->kvm->srcu, idx);
1565 static int handle_hva_to_gpa(struct kvm *kvm,
1566 unsigned long start,
1568 int (*handler)(struct kvm *kvm,
1569 gpa_t gpa, u64 size,
1573 struct kvm_memslots *slots;
1574 struct kvm_memory_slot *memslot;
1577 slots = kvm_memslots(kvm);
1579 /* we only care about the pages that the guest sees */
1580 kvm_for_each_memslot(memslot, slots) {
1581 unsigned long hva_start, hva_end;
1584 hva_start = max(start, memslot->userspace_addr);
1585 hva_end = min(end, memslot->userspace_addr +
1586 (memslot->npages << PAGE_SHIFT));
1587 if (hva_start >= hva_end)
1590 gpa = hva_to_gfn_memslot(hva_start, memslot) << PAGE_SHIFT;
1591 ret |= handler(kvm, gpa, (u64)(hva_end - hva_start), data);
1597 static int kvm_unmap_hva_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
1599 unmap_stage2_range(kvm, gpa, size);
1603 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
1605 unsigned long end = hva + PAGE_SIZE;
1610 trace_kvm_unmap_hva(hva);
1611 handle_hva_to_gpa(kvm, hva, end, &kvm_unmap_hva_handler, NULL);
1615 int kvm_unmap_hva_range(struct kvm *kvm,
1616 unsigned long start, unsigned long end)
1621 trace_kvm_unmap_hva_range(start, end);
1622 handle_hva_to_gpa(kvm, start, end, &kvm_unmap_hva_handler, NULL);
1626 static int kvm_set_spte_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
1628 pte_t *pte = (pte_t *)data;
1630 WARN_ON(size != PAGE_SIZE);
1632 * We can always call stage2_set_pte with KVM_S2PTE_FLAG_LOGGING_ACTIVE
1633 * flag clear because MMU notifiers will have unmapped a huge PMD before
1634 * calling ->change_pte() (which in turn calls kvm_set_spte_hva()) and
1635 * therefore stage2_set_pte() never needs to clear out a huge PMD
1636 * through this calling path.
1638 stage2_set_pte(kvm, NULL, gpa, pte, 0);
1643 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
1645 unsigned long end = hva + PAGE_SIZE;
1651 trace_kvm_set_spte_hva(hva);
1652 stage2_pte = pfn_pte(pte_pfn(pte), PAGE_S2);
1653 handle_hva_to_gpa(kvm, hva, end, &kvm_set_spte_handler, &stage2_pte);
1656 static int kvm_age_hva_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
1661 WARN_ON(size != PAGE_SIZE && size != PMD_SIZE);
1662 pmd = stage2_get_pmd(kvm, NULL, gpa);
1663 if (!pmd || pmd_none(*pmd)) /* Nothing there */
1666 if (pmd_thp_or_huge(*pmd)) /* THP, HugeTLB */
1667 return stage2_pmdp_test_and_clear_young(pmd);
1669 pte = pte_offset_kernel(pmd, gpa);
1673 return stage2_ptep_test_and_clear_young(pte);
1676 static int kvm_test_age_hva_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
1681 WARN_ON(size != PAGE_SIZE && size != PMD_SIZE);
1682 pmd = stage2_get_pmd(kvm, NULL, gpa);
1683 if (!pmd || pmd_none(*pmd)) /* Nothing there */
1686 if (pmd_thp_or_huge(*pmd)) /* THP, HugeTLB */
1687 return pmd_young(*pmd);
1689 pte = pte_offset_kernel(pmd, gpa);
1690 if (!pte_none(*pte)) /* Just a page... */
1691 return pte_young(*pte);
1696 int kvm_age_hva(struct kvm *kvm, unsigned long start, unsigned long end)
1698 trace_kvm_age_hva(start, end);
1699 return handle_hva_to_gpa(kvm, start, end, kvm_age_hva_handler, NULL);
1702 int kvm_test_age_hva(struct kvm *kvm, unsigned long hva)
1704 trace_kvm_test_age_hva(hva);
1705 return handle_hva_to_gpa(kvm, hva, hva, kvm_test_age_hva_handler, NULL);
1708 void kvm_mmu_free_memory_caches(struct kvm_vcpu *vcpu)
1710 mmu_free_memory_cache(&vcpu->arch.mmu_page_cache);
1713 phys_addr_t kvm_mmu_get_httbr(void)
1715 if (__kvm_cpu_uses_extended_idmap())
1716 return virt_to_phys(merged_hyp_pgd);
1718 return virt_to_phys(hyp_pgd);
1721 phys_addr_t kvm_get_idmap_vector(void)
1723 return hyp_idmap_vector;
1726 static int kvm_map_idmap_text(pgd_t *pgd)
1730 /* Create the idmap in the boot page tables */
1731 err = __create_hyp_mappings(pgd,
1732 hyp_idmap_start, hyp_idmap_end,
1733 __phys_to_pfn(hyp_idmap_start),
1736 kvm_err("Failed to idmap %lx-%lx\n",
1737 hyp_idmap_start, hyp_idmap_end);
1742 int kvm_mmu_init(void)
1746 hyp_idmap_start = kvm_virt_to_phys(__hyp_idmap_text_start);
1747 hyp_idmap_end = kvm_virt_to_phys(__hyp_idmap_text_end);
1748 hyp_idmap_vector = kvm_virt_to_phys(__kvm_hyp_init);
1751 * We rely on the linker script to ensure at build time that the HYP
1752 * init code does not cross a page boundary.
1754 BUG_ON((hyp_idmap_start ^ (hyp_idmap_end - 1)) & PAGE_MASK);
1756 kvm_info("IDMAP page: %lx\n", hyp_idmap_start);
1757 kvm_info("HYP VA range: %lx:%lx\n",
1758 kern_hyp_va(PAGE_OFFSET), kern_hyp_va(~0UL));
1760 if (hyp_idmap_start >= kern_hyp_va(PAGE_OFFSET) &&
1761 hyp_idmap_start < kern_hyp_va(~0UL) &&
1762 hyp_idmap_start != (unsigned long)__hyp_idmap_text_start) {
1764 * The idmap page is intersecting with the VA space,
1765 * it is not safe to continue further.
1767 kvm_err("IDMAP intersecting with HYP VA, unable to continue\n");
1772 hyp_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, hyp_pgd_order);
1774 kvm_err("Hyp mode PGD not allocated\n");
1779 if (__kvm_cpu_uses_extended_idmap()) {
1780 boot_hyp_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
1782 if (!boot_hyp_pgd) {
1783 kvm_err("Hyp boot PGD not allocated\n");
1788 err = kvm_map_idmap_text(boot_hyp_pgd);
1792 merged_hyp_pgd = (pgd_t *)__get_free_page(GFP_KERNEL | __GFP_ZERO);
1793 if (!merged_hyp_pgd) {
1794 kvm_err("Failed to allocate extra HYP pgd\n");
1797 __kvm_extend_hypmap(boot_hyp_pgd, hyp_pgd, merged_hyp_pgd,
1800 err = kvm_map_idmap_text(hyp_pgd);
1811 void kvm_arch_commit_memory_region(struct kvm *kvm,
1812 const struct kvm_userspace_memory_region *mem,
1813 const struct kvm_memory_slot *old,
1814 const struct kvm_memory_slot *new,
1815 enum kvm_mr_change change)
1818 * At this point memslot has been committed and there is an
1819 * allocated dirty_bitmap[], dirty pages will be be tracked while the
1820 * memory slot is write protected.
1822 if (change != KVM_MR_DELETE && mem->flags & KVM_MEM_LOG_DIRTY_PAGES)
1823 kvm_mmu_wp_memory_region(kvm, mem->slot);
1826 int kvm_arch_prepare_memory_region(struct kvm *kvm,
1827 struct kvm_memory_slot *memslot,
1828 const struct kvm_userspace_memory_region *mem,
1829 enum kvm_mr_change change)
1831 hva_t hva = mem->userspace_addr;
1832 hva_t reg_end = hva + mem->memory_size;
1833 bool writable = !(mem->flags & KVM_MEM_READONLY);
1836 if (change != KVM_MR_CREATE && change != KVM_MR_MOVE &&
1837 change != KVM_MR_FLAGS_ONLY)
1841 * Prevent userspace from creating a memory region outside of the IPA
1842 * space addressable by the KVM guest IPA space.
1844 if (memslot->base_gfn + memslot->npages >=
1845 (KVM_PHYS_SIZE >> PAGE_SHIFT))
1848 down_read(¤t->mm->mmap_sem);
1850 * A memory region could potentially cover multiple VMAs, and any holes
1851 * between them, so iterate over all of them to find out if we can map
1852 * any of them right now.
1854 * +--------------------------------------------+
1855 * +---------------+----------------+ +----------------+
1856 * | : VMA 1 | VMA 2 | | VMA 3 : |
1857 * +---------------+----------------+ +----------------+
1859 * +--------------------------------------------+
1862 struct vm_area_struct *vma = find_vma(current->mm, hva);
1863 hva_t vm_start, vm_end;
1865 if (!vma || vma->vm_start >= reg_end)
1869 * Mapping a read-only VMA is only allowed if the
1870 * memory region is configured as read-only.
1872 if (writable && !(vma->vm_flags & VM_WRITE)) {
1878 * Take the intersection of this VMA with the memory region
1880 vm_start = max(hva, vma->vm_start);
1881 vm_end = min(reg_end, vma->vm_end);
1883 if (vma->vm_flags & VM_PFNMAP) {
1884 gpa_t gpa = mem->guest_phys_addr +
1885 (vm_start - mem->userspace_addr);
1888 pa = (phys_addr_t)vma->vm_pgoff << PAGE_SHIFT;
1889 pa += vm_start - vma->vm_start;
1891 /* IO region dirty page logging not allowed */
1892 if (memslot->flags & KVM_MEM_LOG_DIRTY_PAGES) {
1897 ret = kvm_phys_addr_ioremap(kvm, gpa, pa,
1904 } while (hva < reg_end);
1906 if (change == KVM_MR_FLAGS_ONLY)
1909 spin_lock(&kvm->mmu_lock);
1911 unmap_stage2_range(kvm, mem->guest_phys_addr, mem->memory_size);
1913 stage2_flush_memslot(kvm, memslot);
1914 spin_unlock(&kvm->mmu_lock);
1916 up_read(¤t->mm->mmap_sem);
1920 void kvm_arch_free_memslot(struct kvm *kvm, struct kvm_memory_slot *free,
1921 struct kvm_memory_slot *dont)
1925 int kvm_arch_create_memslot(struct kvm *kvm, struct kvm_memory_slot *slot,
1926 unsigned long npages)
1931 void kvm_arch_memslots_updated(struct kvm *kvm, struct kvm_memslots *slots)
1935 void kvm_arch_flush_shadow_all(struct kvm *kvm)
1937 kvm_free_stage2_pgd(kvm);
1940 void kvm_arch_flush_shadow_memslot(struct kvm *kvm,
1941 struct kvm_memory_slot *slot)
1943 gpa_t gpa = slot->base_gfn << PAGE_SHIFT;
1944 phys_addr_t size = slot->npages << PAGE_SHIFT;
1946 spin_lock(&kvm->mmu_lock);
1947 unmap_stage2_range(kvm, gpa, size);
1948 spin_unlock(&kvm->mmu_lock);
1952 * See note at ARMv7 ARM B1.14.4 (TL;DR: S/W ops are not easily virtualized).
1955 * - S/W ops are local to a CPU (not broadcast)
1956 * - We have line migration behind our back (speculation)
1957 * - System caches don't support S/W at all (damn!)
1959 * In the face of the above, the best we can do is to try and convert
1960 * S/W ops to VA ops. Because the guest is not allowed to infer the
1961 * S/W to PA mapping, it can only use S/W to nuke the whole cache,
1962 * which is a rather good thing for us.
1964 * Also, it is only used when turning caches on/off ("The expected
1965 * usage of the cache maintenance instructions that operate by set/way
1966 * is associated with the cache maintenance instructions associated
1967 * with the powerdown and powerup of caches, if this is required by
1968 * the implementation.").
1970 * We use the following policy:
1972 * - If we trap a S/W operation, we enable VM trapping to detect
1973 * caches being turned on/off, and do a full clean.
1975 * - We flush the caches on both caches being turned on and off.
1977 * - Once the caches are enabled, we stop trapping VM ops.
1979 void kvm_set_way_flush(struct kvm_vcpu *vcpu)
1981 unsigned long hcr = vcpu_get_hcr(vcpu);
1984 * If this is the first time we do a S/W operation
1985 * (i.e. HCR_TVM not set) flush the whole memory, and set the
1988 * Otherwise, rely on the VM trapping to wait for the MMU +
1989 * Caches to be turned off. At that point, we'll be able to
1990 * clean the caches again.
1992 if (!(hcr & HCR_TVM)) {
1993 trace_kvm_set_way_flush(*vcpu_pc(vcpu),
1994 vcpu_has_cache_enabled(vcpu));
1995 stage2_flush_vm(vcpu->kvm);
1996 vcpu_set_hcr(vcpu, hcr | HCR_TVM);
2000 void kvm_toggle_cache(struct kvm_vcpu *vcpu, bool was_enabled)
2002 bool now_enabled = vcpu_has_cache_enabled(vcpu);
2005 * If switching the MMU+caches on, need to invalidate the caches.
2006 * If switching it off, need to clean the caches.
2007 * Clean + invalidate does the trick always.
2009 if (now_enabled != was_enabled)
2010 stage2_flush_vm(vcpu->kvm);
2012 /* Caches are now on, stop trapping VM ops (until a S/W op) */
2014 vcpu_set_hcr(vcpu, vcpu_get_hcr(vcpu) & ~HCR_TVM);
2016 trace_kvm_toggle_cache(*vcpu_pc(vcpu), was_enabled, now_enabled);