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 <linux/sched/signal.h>
24 #include <trace/events/kvm.h>
25 #include <asm/pgalloc.h>
26 #include <asm/cacheflush.h>
27 #include <asm/kvm_arm.h>
28 #include <asm/kvm_mmu.h>
29 #include <asm/kvm_mmio.h>
30 #include <asm/kvm_asm.h>
31 #include <asm/kvm_emulate.h>
33 #include <asm/system_misc.h>
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 static unsigned long io_map_base;
48 #define S2_PGD_SIZE (PTRS_PER_S2_PGD * sizeof(pgd_t))
49 #define hyp_pgd_order get_order(PTRS_PER_PGD * sizeof(pgd_t))
51 #define KVM_S2PTE_FLAG_IS_IOMAP (1UL << 0)
52 #define KVM_S2_FLAG_LOGGING_ACTIVE (1UL << 1)
54 static bool memslot_is_logging(struct kvm_memory_slot *memslot)
56 return memslot->dirty_bitmap && !(memslot->flags & KVM_MEM_READONLY);
60 * kvm_flush_remote_tlbs() - flush all VM TLB entries for v7/8
61 * @kvm: pointer to kvm structure.
63 * Interface to HYP function to flush all VM TLB entries
65 void kvm_flush_remote_tlbs(struct kvm *kvm)
67 kvm_call_hyp(__kvm_tlb_flush_vmid, kvm);
70 static void kvm_tlb_flush_vmid_ipa(struct kvm *kvm, phys_addr_t ipa)
72 kvm_call_hyp(__kvm_tlb_flush_vmid_ipa, kvm, ipa);
76 * D-Cache management functions. They take the page table entries by
77 * value, as they are flushing the cache using the kernel mapping (or
80 static void kvm_flush_dcache_pte(pte_t pte)
82 __kvm_flush_dcache_pte(pte);
85 static void kvm_flush_dcache_pmd(pmd_t pmd)
87 __kvm_flush_dcache_pmd(pmd);
90 static void kvm_flush_dcache_pud(pud_t pud)
92 __kvm_flush_dcache_pud(pud);
95 static bool kvm_is_device_pfn(unsigned long pfn)
97 return !pfn_valid(pfn);
101 * stage2_dissolve_pmd() - clear and flush huge PMD entry
102 * @kvm: pointer to kvm structure.
104 * @pmd: pmd pointer for IPA
106 * Function clears a PMD entry, flushes addr 1st and 2nd stage TLBs. Marks all
107 * pages in the range dirty.
109 static void stage2_dissolve_pmd(struct kvm *kvm, phys_addr_t addr, pmd_t *pmd)
111 if (!pmd_thp_or_huge(*pmd))
115 kvm_tlb_flush_vmid_ipa(kvm, addr);
116 put_page(virt_to_page(pmd));
119 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
124 BUG_ON(max > KVM_NR_MEM_OBJS);
125 if (cache->nobjs >= min)
127 while (cache->nobjs < max) {
128 page = (void *)__get_free_page(PGALLOC_GFP);
131 cache->objects[cache->nobjs++] = page;
136 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
139 free_page((unsigned long)mc->objects[--mc->nobjs]);
142 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc)
146 BUG_ON(!mc || !mc->nobjs);
147 p = mc->objects[--mc->nobjs];
151 static void clear_stage2_pgd_entry(struct kvm *kvm, pgd_t *pgd, phys_addr_t addr)
153 pud_t *pud_table __maybe_unused = stage2_pud_offset(pgd, 0UL);
154 stage2_pgd_clear(pgd);
155 kvm_tlb_flush_vmid_ipa(kvm, addr);
156 stage2_pud_free(pud_table);
157 put_page(virt_to_page(pgd));
160 static void clear_stage2_pud_entry(struct kvm *kvm, pud_t *pud, phys_addr_t addr)
162 pmd_t *pmd_table __maybe_unused = stage2_pmd_offset(pud, 0);
163 VM_BUG_ON(stage2_pud_huge(*pud));
164 stage2_pud_clear(pud);
165 kvm_tlb_flush_vmid_ipa(kvm, addr);
166 stage2_pmd_free(pmd_table);
167 put_page(virt_to_page(pud));
170 static void clear_stage2_pmd_entry(struct kvm *kvm, pmd_t *pmd, phys_addr_t addr)
172 pte_t *pte_table = pte_offset_kernel(pmd, 0);
173 VM_BUG_ON(pmd_thp_or_huge(*pmd));
175 kvm_tlb_flush_vmid_ipa(kvm, addr);
176 pte_free_kernel(NULL, pte_table);
177 put_page(virt_to_page(pmd));
181 * Unmapping vs dcache management:
183 * If a guest maps certain memory pages as uncached, all writes will
184 * bypass the data cache and go directly to RAM. However, the CPUs
185 * can still speculate reads (not writes) and fill cache lines with
188 * Those cache lines will be *clean* cache lines though, so a
189 * clean+invalidate operation is equivalent to an invalidate
190 * operation, because no cache lines are marked dirty.
192 * Those clean cache lines could be filled prior to an uncached write
193 * by the guest, and the cache coherent IO subsystem would therefore
194 * end up writing old data to disk.
196 * This is why right after unmapping a page/section and invalidating
197 * the corresponding TLBs, we call kvm_flush_dcache_p*() to make sure
198 * the IO subsystem will never hit in the cache.
200 static void unmap_stage2_ptes(struct kvm *kvm, pmd_t *pmd,
201 phys_addr_t addr, phys_addr_t end)
203 phys_addr_t start_addr = addr;
204 pte_t *pte, *start_pte;
206 start_pte = pte = pte_offset_kernel(pmd, addr);
208 if (!pte_none(*pte)) {
209 pte_t old_pte = *pte;
211 kvm_set_pte(pte, __pte(0));
212 kvm_tlb_flush_vmid_ipa(kvm, addr);
214 /* No need to invalidate the cache for device mappings */
215 if (!kvm_is_device_pfn(pte_pfn(old_pte)))
216 kvm_flush_dcache_pte(old_pte);
218 put_page(virt_to_page(pte));
220 } while (pte++, addr += PAGE_SIZE, addr != end);
222 if (stage2_pte_table_empty(start_pte))
223 clear_stage2_pmd_entry(kvm, pmd, start_addr);
226 static void unmap_stage2_pmds(struct kvm *kvm, pud_t *pud,
227 phys_addr_t addr, phys_addr_t end)
229 phys_addr_t next, start_addr = addr;
230 pmd_t *pmd, *start_pmd;
232 start_pmd = pmd = stage2_pmd_offset(pud, addr);
234 next = stage2_pmd_addr_end(addr, end);
235 if (!pmd_none(*pmd)) {
236 if (pmd_thp_or_huge(*pmd)) {
237 pmd_t old_pmd = *pmd;
240 kvm_tlb_flush_vmid_ipa(kvm, addr);
242 kvm_flush_dcache_pmd(old_pmd);
244 put_page(virt_to_page(pmd));
246 unmap_stage2_ptes(kvm, pmd, addr, next);
249 } while (pmd++, addr = next, addr != end);
251 if (stage2_pmd_table_empty(start_pmd))
252 clear_stage2_pud_entry(kvm, pud, start_addr);
255 static void unmap_stage2_puds(struct kvm *kvm, pgd_t *pgd,
256 phys_addr_t addr, phys_addr_t end)
258 phys_addr_t next, start_addr = addr;
259 pud_t *pud, *start_pud;
261 start_pud = pud = stage2_pud_offset(pgd, addr);
263 next = stage2_pud_addr_end(addr, end);
264 if (!stage2_pud_none(*pud)) {
265 if (stage2_pud_huge(*pud)) {
266 pud_t old_pud = *pud;
268 stage2_pud_clear(pud);
269 kvm_tlb_flush_vmid_ipa(kvm, addr);
270 kvm_flush_dcache_pud(old_pud);
271 put_page(virt_to_page(pud));
273 unmap_stage2_pmds(kvm, pud, addr, next);
276 } while (pud++, addr = next, addr != end);
278 if (stage2_pud_table_empty(start_pud))
279 clear_stage2_pgd_entry(kvm, pgd, start_addr);
283 * unmap_stage2_range -- Clear stage2 page table entries to unmap a range
284 * @kvm: The VM pointer
285 * @start: The intermediate physical base address of the range to unmap
286 * @size: The size of the area to unmap
288 * Clear a range of stage-2 mappings, lowering the various ref-counts. Must
289 * be called while holding mmu_lock (unless for freeing the stage2 pgd before
290 * destroying the VM), otherwise another faulting VCPU may come in and mess
291 * with things behind our backs.
293 static void unmap_stage2_range(struct kvm *kvm, phys_addr_t start, u64 size)
296 phys_addr_t addr = start, end = start + size;
299 assert_spin_locked(&kvm->mmu_lock);
300 pgd = kvm->arch.pgd + stage2_pgd_index(addr);
303 * Make sure the page table is still active, as another thread
304 * could have possibly freed the page table, while we released
307 if (!READ_ONCE(kvm->arch.pgd))
309 next = stage2_pgd_addr_end(addr, end);
310 if (!stage2_pgd_none(*pgd))
311 unmap_stage2_puds(kvm, pgd, addr, next);
313 * If the range is too large, release the kvm->mmu_lock
314 * to prevent starvation and lockup detector warnings.
317 cond_resched_lock(&kvm->mmu_lock);
318 } while (pgd++, addr = next, addr != end);
321 static void stage2_flush_ptes(struct kvm *kvm, pmd_t *pmd,
322 phys_addr_t addr, phys_addr_t end)
326 pte = pte_offset_kernel(pmd, addr);
328 if (!pte_none(*pte) && !kvm_is_device_pfn(pte_pfn(*pte)))
329 kvm_flush_dcache_pte(*pte);
330 } while (pte++, addr += PAGE_SIZE, addr != end);
333 static void stage2_flush_pmds(struct kvm *kvm, pud_t *pud,
334 phys_addr_t addr, phys_addr_t end)
339 pmd = stage2_pmd_offset(pud, addr);
341 next = stage2_pmd_addr_end(addr, end);
342 if (!pmd_none(*pmd)) {
343 if (pmd_thp_or_huge(*pmd))
344 kvm_flush_dcache_pmd(*pmd);
346 stage2_flush_ptes(kvm, pmd, addr, next);
348 } while (pmd++, addr = next, addr != end);
351 static void stage2_flush_puds(struct kvm *kvm, pgd_t *pgd,
352 phys_addr_t addr, phys_addr_t end)
357 pud = stage2_pud_offset(pgd, addr);
359 next = stage2_pud_addr_end(addr, end);
360 if (!stage2_pud_none(*pud)) {
361 if (stage2_pud_huge(*pud))
362 kvm_flush_dcache_pud(*pud);
364 stage2_flush_pmds(kvm, pud, addr, next);
366 } while (pud++, addr = next, addr != end);
369 static void stage2_flush_memslot(struct kvm *kvm,
370 struct kvm_memory_slot *memslot)
372 phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
373 phys_addr_t end = addr + PAGE_SIZE * memslot->npages;
377 pgd = kvm->arch.pgd + stage2_pgd_index(addr);
379 next = stage2_pgd_addr_end(addr, end);
380 stage2_flush_puds(kvm, pgd, addr, next);
381 } while (pgd++, addr = next, addr != end);
385 * stage2_flush_vm - Invalidate cache for pages mapped in stage 2
386 * @kvm: The struct kvm pointer
388 * Go through the stage 2 page tables and invalidate any cache lines
389 * backing memory already mapped to the VM.
391 static void stage2_flush_vm(struct kvm *kvm)
393 struct kvm_memslots *slots;
394 struct kvm_memory_slot *memslot;
397 idx = srcu_read_lock(&kvm->srcu);
398 spin_lock(&kvm->mmu_lock);
400 slots = kvm_memslots(kvm);
401 kvm_for_each_memslot(memslot, slots)
402 stage2_flush_memslot(kvm, memslot);
404 spin_unlock(&kvm->mmu_lock);
405 srcu_read_unlock(&kvm->srcu, idx);
408 static void clear_hyp_pgd_entry(pgd_t *pgd)
410 pud_t *pud_table __maybe_unused = pud_offset(pgd, 0UL);
412 pud_free(NULL, pud_table);
413 put_page(virt_to_page(pgd));
416 static void clear_hyp_pud_entry(pud_t *pud)
418 pmd_t *pmd_table __maybe_unused = pmd_offset(pud, 0);
419 VM_BUG_ON(pud_huge(*pud));
421 pmd_free(NULL, pmd_table);
422 put_page(virt_to_page(pud));
425 static void clear_hyp_pmd_entry(pmd_t *pmd)
427 pte_t *pte_table = pte_offset_kernel(pmd, 0);
428 VM_BUG_ON(pmd_thp_or_huge(*pmd));
430 pte_free_kernel(NULL, pte_table);
431 put_page(virt_to_page(pmd));
434 static void unmap_hyp_ptes(pmd_t *pmd, phys_addr_t addr, phys_addr_t end)
436 pte_t *pte, *start_pte;
438 start_pte = pte = pte_offset_kernel(pmd, addr);
440 if (!pte_none(*pte)) {
441 kvm_set_pte(pte, __pte(0));
442 put_page(virt_to_page(pte));
444 } while (pte++, addr += PAGE_SIZE, addr != end);
446 if (hyp_pte_table_empty(start_pte))
447 clear_hyp_pmd_entry(pmd);
450 static void unmap_hyp_pmds(pud_t *pud, phys_addr_t addr, phys_addr_t end)
453 pmd_t *pmd, *start_pmd;
455 start_pmd = pmd = pmd_offset(pud, addr);
457 next = pmd_addr_end(addr, end);
458 /* Hyp doesn't use huge pmds */
460 unmap_hyp_ptes(pmd, addr, next);
461 } while (pmd++, addr = next, addr != end);
463 if (hyp_pmd_table_empty(start_pmd))
464 clear_hyp_pud_entry(pud);
467 static void unmap_hyp_puds(pgd_t *pgd, phys_addr_t addr, phys_addr_t end)
470 pud_t *pud, *start_pud;
472 start_pud = pud = pud_offset(pgd, addr);
474 next = pud_addr_end(addr, end);
475 /* Hyp doesn't use huge puds */
477 unmap_hyp_pmds(pud, addr, next);
478 } while (pud++, addr = next, addr != end);
480 if (hyp_pud_table_empty(start_pud))
481 clear_hyp_pgd_entry(pgd);
484 static unsigned int kvm_pgd_index(unsigned long addr, unsigned int ptrs_per_pgd)
486 return (addr >> PGDIR_SHIFT) & (ptrs_per_pgd - 1);
489 static void __unmap_hyp_range(pgd_t *pgdp, unsigned long ptrs_per_pgd,
490 phys_addr_t start, u64 size)
493 phys_addr_t addr = start, end = start + size;
497 * We don't unmap anything from HYP, except at the hyp tear down.
498 * Hence, we don't have to invalidate the TLBs here.
500 pgd = pgdp + kvm_pgd_index(addr, ptrs_per_pgd);
502 next = pgd_addr_end(addr, end);
504 unmap_hyp_puds(pgd, addr, next);
505 } while (pgd++, addr = next, addr != end);
508 static void unmap_hyp_range(pgd_t *pgdp, phys_addr_t start, u64 size)
510 __unmap_hyp_range(pgdp, PTRS_PER_PGD, start, size);
513 static void unmap_hyp_idmap_range(pgd_t *pgdp, phys_addr_t start, u64 size)
515 __unmap_hyp_range(pgdp, __kvm_idmap_ptrs_per_pgd(), start, size);
519 * free_hyp_pgds - free Hyp-mode page tables
521 * Assumes hyp_pgd is a page table used strictly in Hyp-mode and
522 * therefore contains either mappings in the kernel memory area (above
523 * PAGE_OFFSET), or device mappings in the idmap range.
525 * boot_hyp_pgd should only map the idmap range, and is only used in
526 * the extended idmap case.
528 void free_hyp_pgds(void)
532 mutex_lock(&kvm_hyp_pgd_mutex);
534 id_pgd = boot_hyp_pgd ? boot_hyp_pgd : hyp_pgd;
537 /* In case we never called hyp_mmu_init() */
539 io_map_base = hyp_idmap_start;
540 unmap_hyp_idmap_range(id_pgd, io_map_base,
541 hyp_idmap_start + PAGE_SIZE - io_map_base);
545 free_pages((unsigned long)boot_hyp_pgd, hyp_pgd_order);
550 unmap_hyp_range(hyp_pgd, kern_hyp_va(PAGE_OFFSET),
551 (uintptr_t)high_memory - PAGE_OFFSET);
553 free_pages((unsigned long)hyp_pgd, hyp_pgd_order);
556 if (merged_hyp_pgd) {
557 clear_page(merged_hyp_pgd);
558 free_page((unsigned long)merged_hyp_pgd);
559 merged_hyp_pgd = NULL;
562 mutex_unlock(&kvm_hyp_pgd_mutex);
565 static void create_hyp_pte_mappings(pmd_t *pmd, unsigned long start,
566 unsigned long end, unsigned long pfn,
574 pte = pte_offset_kernel(pmd, addr);
575 kvm_set_pte(pte, pfn_pte(pfn, prot));
576 get_page(virt_to_page(pte));
577 kvm_flush_dcache_to_poc(pte, sizeof(*pte));
579 } while (addr += PAGE_SIZE, addr != end);
582 static int create_hyp_pmd_mappings(pud_t *pud, unsigned long start,
583 unsigned long end, unsigned long pfn,
588 unsigned long addr, next;
592 pmd = pmd_offset(pud, addr);
594 BUG_ON(pmd_sect(*pmd));
596 if (pmd_none(*pmd)) {
597 pte = pte_alloc_one_kernel(NULL, addr);
599 kvm_err("Cannot allocate Hyp pte\n");
602 pmd_populate_kernel(NULL, pmd, pte);
603 get_page(virt_to_page(pmd));
604 kvm_flush_dcache_to_poc(pmd, sizeof(*pmd));
607 next = pmd_addr_end(addr, end);
609 create_hyp_pte_mappings(pmd, addr, next, pfn, prot);
610 pfn += (next - addr) >> PAGE_SHIFT;
611 } while (addr = next, addr != end);
616 static int create_hyp_pud_mappings(pgd_t *pgd, unsigned long start,
617 unsigned long end, unsigned long pfn,
622 unsigned long addr, next;
627 pud = pud_offset(pgd, addr);
629 if (pud_none_or_clear_bad(pud)) {
630 pmd = pmd_alloc_one(NULL, addr);
632 kvm_err("Cannot allocate Hyp pmd\n");
635 pud_populate(NULL, pud, pmd);
636 get_page(virt_to_page(pud));
637 kvm_flush_dcache_to_poc(pud, sizeof(*pud));
640 next = pud_addr_end(addr, end);
641 ret = create_hyp_pmd_mappings(pud, addr, next, pfn, prot);
644 pfn += (next - addr) >> PAGE_SHIFT;
645 } while (addr = next, addr != end);
650 static int __create_hyp_mappings(pgd_t *pgdp, unsigned long ptrs_per_pgd,
651 unsigned long start, unsigned long end,
652 unsigned long pfn, pgprot_t prot)
656 unsigned long addr, next;
659 mutex_lock(&kvm_hyp_pgd_mutex);
660 addr = start & PAGE_MASK;
661 end = PAGE_ALIGN(end);
663 pgd = pgdp + kvm_pgd_index(addr, ptrs_per_pgd);
665 if (pgd_none(*pgd)) {
666 pud = pud_alloc_one(NULL, addr);
668 kvm_err("Cannot allocate Hyp pud\n");
672 pgd_populate(NULL, pgd, pud);
673 get_page(virt_to_page(pgd));
674 kvm_flush_dcache_to_poc(pgd, sizeof(*pgd));
677 next = pgd_addr_end(addr, end);
678 err = create_hyp_pud_mappings(pgd, addr, next, pfn, prot);
681 pfn += (next - addr) >> PAGE_SHIFT;
682 } while (addr = next, addr != end);
684 mutex_unlock(&kvm_hyp_pgd_mutex);
688 static phys_addr_t kvm_kaddr_to_phys(void *kaddr)
690 if (!is_vmalloc_addr(kaddr)) {
691 BUG_ON(!virt_addr_valid(kaddr));
694 return page_to_phys(vmalloc_to_page(kaddr)) +
695 offset_in_page(kaddr);
700 * create_hyp_mappings - duplicate a kernel virtual address range in Hyp mode
701 * @from: The virtual kernel start address of the range
702 * @to: The virtual kernel end address of the range (exclusive)
703 * @prot: The protection to be applied to this range
705 * The same virtual address as the kernel virtual address is also used
706 * in Hyp-mode mapping (modulo HYP_PAGE_OFFSET) to the same underlying
709 int create_hyp_mappings(void *from, void *to, pgprot_t prot)
711 phys_addr_t phys_addr;
712 unsigned long virt_addr;
713 unsigned long start = kern_hyp_va((unsigned long)from);
714 unsigned long end = kern_hyp_va((unsigned long)to);
716 if (is_kernel_in_hyp_mode())
719 start = start & PAGE_MASK;
720 end = PAGE_ALIGN(end);
722 for (virt_addr = start; virt_addr < end; virt_addr += PAGE_SIZE) {
725 phys_addr = kvm_kaddr_to_phys(from + virt_addr - start);
726 err = __create_hyp_mappings(hyp_pgd, PTRS_PER_PGD,
727 virt_addr, virt_addr + PAGE_SIZE,
728 __phys_to_pfn(phys_addr),
737 static int __create_hyp_private_mapping(phys_addr_t phys_addr, size_t size,
738 unsigned long *haddr, pgprot_t prot)
740 pgd_t *pgd = hyp_pgd;
744 mutex_lock(&kvm_hyp_pgd_mutex);
747 * This assumes that we we have enough space below the idmap
748 * page to allocate our VAs. If not, the check below will
749 * kick. A potential alternative would be to detect that
750 * overflow and switch to an allocation above the idmap.
752 * The allocated size is always a multiple of PAGE_SIZE.
754 size = PAGE_ALIGN(size + offset_in_page(phys_addr));
755 base = io_map_base - size;
758 * Verify that BIT(VA_BITS - 1) hasn't been flipped by
759 * allocating the new area, as it would indicate we've
760 * overflowed the idmap/IO address range.
762 if ((base ^ io_map_base) & BIT(VA_BITS - 1))
767 mutex_unlock(&kvm_hyp_pgd_mutex);
772 if (__kvm_cpu_uses_extended_idmap())
775 ret = __create_hyp_mappings(pgd, __kvm_idmap_ptrs_per_pgd(),
777 __phys_to_pfn(phys_addr), prot);
781 *haddr = base + offset_in_page(phys_addr);
788 * create_hyp_io_mappings - Map IO into both kernel and HYP
789 * @phys_addr: The physical start address which gets mapped
790 * @size: Size of the region being mapped
791 * @kaddr: Kernel VA for this mapping
792 * @haddr: HYP VA for this mapping
794 int create_hyp_io_mappings(phys_addr_t phys_addr, size_t size,
795 void __iomem **kaddr,
796 void __iomem **haddr)
801 *kaddr = ioremap(phys_addr, size);
805 if (is_kernel_in_hyp_mode()) {
810 ret = __create_hyp_private_mapping(phys_addr, size,
811 &addr, PAGE_HYP_DEVICE);
819 *haddr = (void __iomem *)addr;
824 * create_hyp_exec_mappings - Map an executable range into HYP
825 * @phys_addr: The physical start address which gets mapped
826 * @size: Size of the region being mapped
827 * @haddr: HYP VA for this mapping
829 int create_hyp_exec_mappings(phys_addr_t phys_addr, size_t size,
835 BUG_ON(is_kernel_in_hyp_mode());
837 ret = __create_hyp_private_mapping(phys_addr, size,
838 &addr, PAGE_HYP_EXEC);
844 *haddr = (void *)addr;
849 * kvm_alloc_stage2_pgd - allocate level-1 table for stage-2 translation.
850 * @kvm: The KVM struct pointer for the VM.
852 * Allocates only the stage-2 HW PGD level table(s) (can support either full
853 * 40-bit input addresses or limited to 32-bit input addresses). Clears the
856 * Note we don't need locking here as this is only called when the VM is
857 * created, which can only be done once.
859 int kvm_alloc_stage2_pgd(struct kvm *kvm)
863 if (kvm->arch.pgd != NULL) {
864 kvm_err("kvm_arch already initialized?\n");
868 /* Allocate the HW PGD, making sure that each page gets its own refcount */
869 pgd = alloc_pages_exact(S2_PGD_SIZE, GFP_KERNEL | __GFP_ZERO);
877 static void stage2_unmap_memslot(struct kvm *kvm,
878 struct kvm_memory_slot *memslot)
880 hva_t hva = memslot->userspace_addr;
881 phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
882 phys_addr_t size = PAGE_SIZE * memslot->npages;
883 hva_t reg_end = hva + size;
886 * A memory region could potentially cover multiple VMAs, and any holes
887 * between them, so iterate over all of them to find out if we should
890 * +--------------------------------------------+
891 * +---------------+----------------+ +----------------+
892 * | : VMA 1 | VMA 2 | | VMA 3 : |
893 * +---------------+----------------+ +----------------+
895 * +--------------------------------------------+
898 struct vm_area_struct *vma = find_vma(current->mm, hva);
899 hva_t vm_start, vm_end;
901 if (!vma || vma->vm_start >= reg_end)
905 * Take the intersection of this VMA with the memory region
907 vm_start = max(hva, vma->vm_start);
908 vm_end = min(reg_end, vma->vm_end);
910 if (!(vma->vm_flags & VM_PFNMAP)) {
911 gpa_t gpa = addr + (vm_start - memslot->userspace_addr);
912 unmap_stage2_range(kvm, gpa, vm_end - vm_start);
915 } while (hva < reg_end);
919 * stage2_unmap_vm - Unmap Stage-2 RAM mappings
920 * @kvm: The struct kvm pointer
922 * Go through the memregions and unmap any reguler RAM
923 * backing memory already mapped to the VM.
925 void stage2_unmap_vm(struct kvm *kvm)
927 struct kvm_memslots *slots;
928 struct kvm_memory_slot *memslot;
931 idx = srcu_read_lock(&kvm->srcu);
932 down_read(¤t->mm->mmap_sem);
933 spin_lock(&kvm->mmu_lock);
935 slots = kvm_memslots(kvm);
936 kvm_for_each_memslot(memslot, slots)
937 stage2_unmap_memslot(kvm, memslot);
939 spin_unlock(&kvm->mmu_lock);
940 up_read(¤t->mm->mmap_sem);
941 srcu_read_unlock(&kvm->srcu, idx);
945 * kvm_free_stage2_pgd - free all stage-2 tables
946 * @kvm: The KVM struct pointer for the VM.
948 * Walks the level-1 page table pointed to by kvm->arch.pgd and frees all
949 * underlying level-2 and level-3 tables before freeing the actual level-1 table
950 * and setting the struct pointer to NULL.
952 void kvm_free_stage2_pgd(struct kvm *kvm)
956 spin_lock(&kvm->mmu_lock);
958 unmap_stage2_range(kvm, 0, KVM_PHYS_SIZE);
959 pgd = READ_ONCE(kvm->arch.pgd);
960 kvm->arch.pgd = NULL;
962 spin_unlock(&kvm->mmu_lock);
964 /* Free the HW pgd, one page at a time */
966 free_pages_exact(pgd, S2_PGD_SIZE);
969 static pud_t *stage2_get_pud(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
975 pgd = kvm->arch.pgd + stage2_pgd_index(addr);
976 if (WARN_ON(stage2_pgd_none(*pgd))) {
979 pud = mmu_memory_cache_alloc(cache);
980 stage2_pgd_populate(pgd, pud);
981 get_page(virt_to_page(pgd));
984 return stage2_pud_offset(pgd, addr);
987 static pmd_t *stage2_get_pmd(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
993 pud = stage2_get_pud(kvm, cache, addr);
997 if (stage2_pud_none(*pud)) {
1000 pmd = mmu_memory_cache_alloc(cache);
1001 stage2_pud_populate(pud, pmd);
1002 get_page(virt_to_page(pud));
1005 return stage2_pmd_offset(pud, addr);
1008 static int stage2_set_pmd_huge(struct kvm *kvm, struct kvm_mmu_memory_cache
1009 *cache, phys_addr_t addr, const pmd_t *new_pmd)
1011 pmd_t *pmd, old_pmd;
1013 pmd = stage2_get_pmd(kvm, cache, addr);
1017 * Mapping in huge pages should only happen through a fault. If a
1018 * page is merged into a transparent huge page, the individual
1019 * subpages of that huge page should be unmapped through MMU
1020 * notifiers before we get here.
1022 * Merging of CompoundPages is not supported; they should become
1023 * splitting first, unmapped, merged, and mapped back in on-demand.
1025 VM_BUG_ON(pmd_present(*pmd) && pmd_pfn(*pmd) != pmd_pfn(*new_pmd));
1028 if (pmd_present(old_pmd)) {
1030 kvm_tlb_flush_vmid_ipa(kvm, addr);
1032 get_page(virt_to_page(pmd));
1035 kvm_set_pmd(pmd, *new_pmd);
1039 static bool stage2_is_exec(struct kvm *kvm, phys_addr_t addr)
1044 pmdp = stage2_get_pmd(kvm, NULL, addr);
1045 if (!pmdp || pmd_none(*pmdp) || !pmd_present(*pmdp))
1048 if (pmd_thp_or_huge(*pmdp))
1049 return kvm_s2pmd_exec(pmdp);
1051 ptep = pte_offset_kernel(pmdp, addr);
1052 if (!ptep || pte_none(*ptep) || !pte_present(*ptep))
1055 return kvm_s2pte_exec(ptep);
1058 static int stage2_set_pte(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
1059 phys_addr_t addr, const pte_t *new_pte,
1060 unsigned long flags)
1063 pte_t *pte, old_pte;
1064 bool iomap = flags & KVM_S2PTE_FLAG_IS_IOMAP;
1065 bool logging_active = flags & KVM_S2_FLAG_LOGGING_ACTIVE;
1067 VM_BUG_ON(logging_active && !cache);
1069 /* Create stage-2 page table mapping - Levels 0 and 1 */
1070 pmd = stage2_get_pmd(kvm, cache, addr);
1073 * Ignore calls from kvm_set_spte_hva for unallocated
1080 * While dirty page logging - dissolve huge PMD, then continue on to
1084 stage2_dissolve_pmd(kvm, addr, pmd);
1086 /* Create stage-2 page mappings - Level 2 */
1087 if (pmd_none(*pmd)) {
1089 return 0; /* ignore calls from kvm_set_spte_hva */
1090 pte = mmu_memory_cache_alloc(cache);
1091 pmd_populate_kernel(NULL, pmd, pte);
1092 get_page(virt_to_page(pmd));
1095 pte = pte_offset_kernel(pmd, addr);
1097 if (iomap && pte_present(*pte))
1100 /* Create 2nd stage page table mapping - Level 3 */
1102 if (pte_present(old_pte)) {
1103 kvm_set_pte(pte, __pte(0));
1104 kvm_tlb_flush_vmid_ipa(kvm, addr);
1106 get_page(virt_to_page(pte));
1109 kvm_set_pte(pte, *new_pte);
1113 #ifndef __HAVE_ARCH_PTEP_TEST_AND_CLEAR_YOUNG
1114 static int stage2_ptep_test_and_clear_young(pte_t *pte)
1116 if (pte_young(*pte)) {
1117 *pte = pte_mkold(*pte);
1123 static int stage2_ptep_test_and_clear_young(pte_t *pte)
1125 return __ptep_test_and_clear_young(pte);
1129 static int stage2_pmdp_test_and_clear_young(pmd_t *pmd)
1131 return stage2_ptep_test_and_clear_young((pte_t *)pmd);
1135 * kvm_phys_addr_ioremap - map a device range to guest IPA
1137 * @kvm: The KVM pointer
1138 * @guest_ipa: The IPA at which to insert the mapping
1139 * @pa: The physical address of the device
1140 * @size: The size of the mapping
1142 int kvm_phys_addr_ioremap(struct kvm *kvm, phys_addr_t guest_ipa,
1143 phys_addr_t pa, unsigned long size, bool writable)
1145 phys_addr_t addr, end;
1148 struct kvm_mmu_memory_cache cache = { 0, };
1150 end = (guest_ipa + size + PAGE_SIZE - 1) & PAGE_MASK;
1151 pfn = __phys_to_pfn(pa);
1153 for (addr = guest_ipa; addr < end; addr += PAGE_SIZE) {
1154 pte_t pte = pfn_pte(pfn, PAGE_S2_DEVICE);
1157 pte = kvm_s2pte_mkwrite(pte);
1159 ret = mmu_topup_memory_cache(&cache, KVM_MMU_CACHE_MIN_PAGES,
1163 spin_lock(&kvm->mmu_lock);
1164 ret = stage2_set_pte(kvm, &cache, addr, &pte,
1165 KVM_S2PTE_FLAG_IS_IOMAP);
1166 spin_unlock(&kvm->mmu_lock);
1174 mmu_free_memory_cache(&cache);
1178 static bool transparent_hugepage_adjust(kvm_pfn_t *pfnp, phys_addr_t *ipap)
1180 kvm_pfn_t pfn = *pfnp;
1181 gfn_t gfn = *ipap >> PAGE_SHIFT;
1183 if (PageTransCompoundMap(pfn_to_page(pfn))) {
1186 * The address we faulted on is backed by a transparent huge
1187 * page. However, because we map the compound huge page and
1188 * not the individual tail page, we need to transfer the
1189 * refcount to the head page. We have to be careful that the
1190 * THP doesn't start to split while we are adjusting the
1193 * We are sure this doesn't happen, because mmu_notifier_retry
1194 * was successful and we are holding the mmu_lock, so if this
1195 * THP is trying to split, it will be blocked in the mmu
1196 * notifier before touching any of the pages, specifically
1197 * before being able to call __split_huge_page_refcount().
1199 * We can therefore safely transfer the refcount from PG_tail
1200 * to PG_head and switch the pfn from a tail page to the head
1203 mask = PTRS_PER_PMD - 1;
1204 VM_BUG_ON((gfn & mask) != (pfn & mask));
1207 kvm_release_pfn_clean(pfn);
1219 static bool kvm_is_write_fault(struct kvm_vcpu *vcpu)
1221 if (kvm_vcpu_trap_is_iabt(vcpu))
1224 return kvm_vcpu_dabt_iswrite(vcpu);
1228 * stage2_wp_ptes - write protect PMD range
1229 * @pmd: pointer to pmd entry
1230 * @addr: range start address
1231 * @end: range end address
1233 static void stage2_wp_ptes(pmd_t *pmd, phys_addr_t addr, phys_addr_t end)
1237 pte = pte_offset_kernel(pmd, addr);
1239 if (!pte_none(*pte)) {
1240 if (!kvm_s2pte_readonly(pte))
1241 kvm_set_s2pte_readonly(pte);
1243 } while (pte++, addr += PAGE_SIZE, addr != end);
1247 * stage2_wp_pmds - write protect PUD range
1248 * @pud: pointer to pud entry
1249 * @addr: range start address
1250 * @end: range end address
1252 static void stage2_wp_pmds(pud_t *pud, phys_addr_t addr, phys_addr_t end)
1257 pmd = stage2_pmd_offset(pud, addr);
1260 next = stage2_pmd_addr_end(addr, end);
1261 if (!pmd_none(*pmd)) {
1262 if (pmd_thp_or_huge(*pmd)) {
1263 if (!kvm_s2pmd_readonly(pmd))
1264 kvm_set_s2pmd_readonly(pmd);
1266 stage2_wp_ptes(pmd, addr, next);
1269 } while (pmd++, addr = next, addr != end);
1273 * stage2_wp_puds - write protect PGD range
1274 * @pgd: pointer to pgd entry
1275 * @addr: range start address
1276 * @end: range end address
1278 * Process PUD entries, for a huge PUD we cause a panic.
1280 static void stage2_wp_puds(pgd_t *pgd, phys_addr_t addr, phys_addr_t end)
1285 pud = stage2_pud_offset(pgd, addr);
1287 next = stage2_pud_addr_end(addr, end);
1288 if (!stage2_pud_none(*pud)) {
1289 /* TODO:PUD not supported, revisit later if supported */
1290 BUG_ON(stage2_pud_huge(*pud));
1291 stage2_wp_pmds(pud, addr, next);
1293 } while (pud++, addr = next, addr != end);
1297 * stage2_wp_range() - write protect stage2 memory region range
1298 * @kvm: The KVM pointer
1299 * @addr: Start address of range
1300 * @end: End address of range
1302 static void stage2_wp_range(struct kvm *kvm, phys_addr_t addr, phys_addr_t end)
1307 pgd = kvm->arch.pgd + stage2_pgd_index(addr);
1310 * Release kvm_mmu_lock periodically if the memory region is
1311 * large. Otherwise, we may see kernel panics with
1312 * CONFIG_DETECT_HUNG_TASK, CONFIG_LOCKUP_DETECTOR,
1313 * CONFIG_LOCKDEP. Additionally, holding the lock too long
1314 * will also starve other vCPUs. We have to also make sure
1315 * that the page tables are not freed while we released
1318 cond_resched_lock(&kvm->mmu_lock);
1319 if (!READ_ONCE(kvm->arch.pgd))
1321 next = stage2_pgd_addr_end(addr, end);
1322 if (stage2_pgd_present(*pgd))
1323 stage2_wp_puds(pgd, addr, next);
1324 } while (pgd++, addr = next, addr != end);
1328 * kvm_mmu_wp_memory_region() - write protect stage 2 entries for memory slot
1329 * @kvm: The KVM pointer
1330 * @slot: The memory slot to write protect
1332 * Called to start logging dirty pages after memory region
1333 * KVM_MEM_LOG_DIRTY_PAGES operation is called. After this function returns
1334 * all present PMD and PTEs are write protected in the memory region.
1335 * Afterwards read of dirty page log can be called.
1337 * Acquires kvm_mmu_lock. Called with kvm->slots_lock mutex acquired,
1338 * serializing operations for VM memory regions.
1340 void kvm_mmu_wp_memory_region(struct kvm *kvm, int slot)
1342 struct kvm_memslots *slots = kvm_memslots(kvm);
1343 struct kvm_memory_slot *memslot = id_to_memslot(slots, slot);
1344 phys_addr_t start = memslot->base_gfn << PAGE_SHIFT;
1345 phys_addr_t end = (memslot->base_gfn + memslot->npages) << PAGE_SHIFT;
1347 spin_lock(&kvm->mmu_lock);
1348 stage2_wp_range(kvm, start, end);
1349 spin_unlock(&kvm->mmu_lock);
1350 kvm_flush_remote_tlbs(kvm);
1354 * kvm_mmu_write_protect_pt_masked() - write protect dirty pages
1355 * @kvm: The KVM pointer
1356 * @slot: The memory slot associated with mask
1357 * @gfn_offset: The gfn offset in memory slot
1358 * @mask: The mask of dirty pages at offset 'gfn_offset' in this memory
1359 * slot to be write protected
1361 * Walks bits set in mask write protects the associated pte's. Caller must
1362 * acquire kvm_mmu_lock.
1364 static void kvm_mmu_write_protect_pt_masked(struct kvm *kvm,
1365 struct kvm_memory_slot *slot,
1366 gfn_t gfn_offset, unsigned long mask)
1368 phys_addr_t base_gfn = slot->base_gfn + gfn_offset;
1369 phys_addr_t start = (base_gfn + __ffs(mask)) << PAGE_SHIFT;
1370 phys_addr_t end = (base_gfn + __fls(mask) + 1) << PAGE_SHIFT;
1372 stage2_wp_range(kvm, start, end);
1376 * kvm_arch_mmu_enable_log_dirty_pt_masked - enable dirty logging for selected
1379 * It calls kvm_mmu_write_protect_pt_masked to write protect selected pages to
1380 * enable dirty logging for them.
1382 void kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm *kvm,
1383 struct kvm_memory_slot *slot,
1384 gfn_t gfn_offset, unsigned long mask)
1386 kvm_mmu_write_protect_pt_masked(kvm, slot, gfn_offset, mask);
1389 static void clean_dcache_guest_page(kvm_pfn_t pfn, unsigned long size)
1391 __clean_dcache_guest_page(pfn, size);
1394 static void invalidate_icache_guest_page(kvm_pfn_t pfn, unsigned long size)
1396 __invalidate_icache_guest_page(pfn, size);
1399 static void kvm_send_hwpoison_signal(unsigned long address,
1400 struct vm_area_struct *vma)
1404 clear_siginfo(&info);
1405 info.si_signo = SIGBUS;
1407 info.si_code = BUS_MCEERR_AR;
1408 info.si_addr = (void __user *)address;
1410 if (is_vm_hugetlb_page(vma))
1411 info.si_addr_lsb = huge_page_shift(hstate_vma(vma));
1413 info.si_addr_lsb = PAGE_SHIFT;
1415 send_sig_info(SIGBUS, &info, current);
1418 static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
1419 struct kvm_memory_slot *memslot, unsigned long hva,
1420 unsigned long fault_status)
1423 bool write_fault, exec_fault, writable, hugetlb = false, force_pte = false;
1424 unsigned long mmu_seq;
1425 gfn_t gfn = fault_ipa >> PAGE_SHIFT;
1426 struct kvm *kvm = vcpu->kvm;
1427 struct kvm_mmu_memory_cache *memcache = &vcpu->arch.mmu_page_cache;
1428 struct vm_area_struct *vma;
1430 pgprot_t mem_type = PAGE_S2;
1431 bool logging_active = memslot_is_logging(memslot);
1432 unsigned long flags = 0;
1434 write_fault = kvm_is_write_fault(vcpu);
1435 exec_fault = kvm_vcpu_trap_is_iabt(vcpu);
1436 VM_BUG_ON(write_fault && exec_fault);
1438 if (fault_status == FSC_PERM && !write_fault && !exec_fault) {
1439 kvm_err("Unexpected L2 read permission error\n");
1443 /* Let's check if we will get back a huge page backed by hugetlbfs */
1444 down_read(¤t->mm->mmap_sem);
1445 vma = find_vma_intersection(current->mm, hva, hva + 1);
1446 if (unlikely(!vma)) {
1447 kvm_err("Failed to find VMA for hva 0x%lx\n", hva);
1448 up_read(¤t->mm->mmap_sem);
1452 if (vma_kernel_pagesize(vma) == PMD_SIZE && !logging_active) {
1454 gfn = (fault_ipa & PMD_MASK) >> PAGE_SHIFT;
1457 * Pages belonging to memslots that don't have the same
1458 * alignment for userspace and IPA cannot be mapped using
1459 * block descriptors even if the pages belong to a THP for
1460 * the process, because the stage-2 block descriptor will
1461 * cover more than a single THP and we loose atomicity for
1462 * unmapping, updates, and splits of the THP or other pages
1463 * in the stage-2 block range.
1465 if ((memslot->userspace_addr & ~PMD_MASK) !=
1466 ((memslot->base_gfn << PAGE_SHIFT) & ~PMD_MASK))
1469 up_read(¤t->mm->mmap_sem);
1471 /* We need minimum second+third level pages */
1472 ret = mmu_topup_memory_cache(memcache, KVM_MMU_CACHE_MIN_PAGES,
1477 mmu_seq = vcpu->kvm->mmu_notifier_seq;
1479 * Ensure the read of mmu_notifier_seq happens before we call
1480 * gfn_to_pfn_prot (which calls get_user_pages), so that we don't risk
1481 * the page we just got a reference to gets unmapped before we have a
1482 * chance to grab the mmu_lock, which ensure that if the page gets
1483 * unmapped afterwards, the call to kvm_unmap_hva will take it away
1484 * from us again properly. This smp_rmb() interacts with the smp_wmb()
1485 * in kvm_mmu_notifier_invalidate_<page|range_end>.
1489 pfn = gfn_to_pfn_prot(kvm, gfn, write_fault, &writable);
1490 if (pfn == KVM_PFN_ERR_HWPOISON) {
1491 kvm_send_hwpoison_signal(hva, vma);
1494 if (is_error_noslot_pfn(pfn))
1497 if (kvm_is_device_pfn(pfn)) {
1498 mem_type = PAGE_S2_DEVICE;
1499 flags |= KVM_S2PTE_FLAG_IS_IOMAP;
1500 } else if (logging_active) {
1502 * Faults on pages in a memslot with logging enabled
1503 * should not be mapped with huge pages (it introduces churn
1504 * and performance degradation), so force a pte mapping.
1507 flags |= KVM_S2_FLAG_LOGGING_ACTIVE;
1510 * Only actually map the page as writable if this was a write
1517 spin_lock(&kvm->mmu_lock);
1518 if (mmu_notifier_retry(kvm, mmu_seq))
1521 if (!hugetlb && !force_pte)
1522 hugetlb = transparent_hugepage_adjust(&pfn, &fault_ipa);
1525 pmd_t new_pmd = pfn_pmd(pfn, mem_type);
1526 new_pmd = pmd_mkhuge(new_pmd);
1528 new_pmd = kvm_s2pmd_mkwrite(new_pmd);
1529 kvm_set_pfn_dirty(pfn);
1532 if (fault_status != FSC_PERM)
1533 clean_dcache_guest_page(pfn, PMD_SIZE);
1536 new_pmd = kvm_s2pmd_mkexec(new_pmd);
1537 invalidate_icache_guest_page(pfn, PMD_SIZE);
1538 } else if (fault_status == FSC_PERM) {
1539 /* Preserve execute if XN was already cleared */
1540 if (stage2_is_exec(kvm, fault_ipa))
1541 new_pmd = kvm_s2pmd_mkexec(new_pmd);
1544 ret = stage2_set_pmd_huge(kvm, memcache, fault_ipa, &new_pmd);
1546 pte_t new_pte = pfn_pte(pfn, mem_type);
1549 new_pte = kvm_s2pte_mkwrite(new_pte);
1550 kvm_set_pfn_dirty(pfn);
1551 mark_page_dirty(kvm, gfn);
1554 if (fault_status != FSC_PERM)
1555 clean_dcache_guest_page(pfn, PAGE_SIZE);
1558 new_pte = kvm_s2pte_mkexec(new_pte);
1559 invalidate_icache_guest_page(pfn, PAGE_SIZE);
1560 } else if (fault_status == FSC_PERM) {
1561 /* Preserve execute if XN was already cleared */
1562 if (stage2_is_exec(kvm, fault_ipa))
1563 new_pte = kvm_s2pte_mkexec(new_pte);
1566 ret = stage2_set_pte(kvm, memcache, fault_ipa, &new_pte, flags);
1570 spin_unlock(&kvm->mmu_lock);
1571 kvm_set_pfn_accessed(pfn);
1572 kvm_release_pfn_clean(pfn);
1577 * Resolve the access fault by making the page young again.
1578 * Note that because the faulting entry is guaranteed not to be
1579 * cached in the TLB, we don't need to invalidate anything.
1580 * Only the HW Access Flag updates are supported for Stage 2 (no DBM),
1581 * so there is no need for atomic (pte|pmd)_mkyoung operations.
1583 static void handle_access_fault(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa)
1588 bool pfn_valid = false;
1590 trace_kvm_access_fault(fault_ipa);
1592 spin_lock(&vcpu->kvm->mmu_lock);
1594 pmd = stage2_get_pmd(vcpu->kvm, NULL, fault_ipa);
1595 if (!pmd || pmd_none(*pmd)) /* Nothing there */
1598 if (pmd_thp_or_huge(*pmd)) { /* THP, HugeTLB */
1599 *pmd = pmd_mkyoung(*pmd);
1600 pfn = pmd_pfn(*pmd);
1605 pte = pte_offset_kernel(pmd, fault_ipa);
1606 if (pte_none(*pte)) /* Nothing there either */
1609 *pte = pte_mkyoung(*pte); /* Just a page... */
1610 pfn = pte_pfn(*pte);
1613 spin_unlock(&vcpu->kvm->mmu_lock);
1615 kvm_set_pfn_accessed(pfn);
1619 * kvm_handle_guest_abort - handles all 2nd stage aborts
1620 * @vcpu: the VCPU pointer
1621 * @run: the kvm_run structure
1623 * Any abort that gets to the host is almost guaranteed to be caused by a
1624 * missing second stage translation table entry, which can mean that either the
1625 * guest simply needs more memory and we must allocate an appropriate page or it
1626 * can mean that the guest tried to access I/O memory, which is emulated by user
1627 * space. The distinction is based on the IPA causing the fault and whether this
1628 * memory region has been registered as standard RAM by user space.
1630 int kvm_handle_guest_abort(struct kvm_vcpu *vcpu, struct kvm_run *run)
1632 unsigned long fault_status;
1633 phys_addr_t fault_ipa;
1634 struct kvm_memory_slot *memslot;
1636 bool is_iabt, write_fault, writable;
1640 fault_status = kvm_vcpu_trap_get_fault_type(vcpu);
1642 fault_ipa = kvm_vcpu_get_fault_ipa(vcpu);
1643 is_iabt = kvm_vcpu_trap_is_iabt(vcpu);
1645 /* Synchronous External Abort? */
1646 if (kvm_vcpu_dabt_isextabt(vcpu)) {
1648 * For RAS the host kernel may handle this abort.
1649 * There is no need to pass the error into the guest.
1651 if (!handle_guest_sea(fault_ipa, kvm_vcpu_get_hsr(vcpu)))
1654 if (unlikely(!is_iabt)) {
1655 kvm_inject_vabt(vcpu);
1660 trace_kvm_guest_fault(*vcpu_pc(vcpu), kvm_vcpu_get_hsr(vcpu),
1661 kvm_vcpu_get_hfar(vcpu), fault_ipa);
1663 /* Check the stage-2 fault is trans. fault or write fault */
1664 if (fault_status != FSC_FAULT && fault_status != FSC_PERM &&
1665 fault_status != FSC_ACCESS) {
1666 kvm_err("Unsupported FSC: EC=%#x xFSC=%#lx ESR_EL2=%#lx\n",
1667 kvm_vcpu_trap_get_class(vcpu),
1668 (unsigned long)kvm_vcpu_trap_get_fault(vcpu),
1669 (unsigned long)kvm_vcpu_get_hsr(vcpu));
1673 idx = srcu_read_lock(&vcpu->kvm->srcu);
1675 gfn = fault_ipa >> PAGE_SHIFT;
1676 memslot = gfn_to_memslot(vcpu->kvm, gfn);
1677 hva = gfn_to_hva_memslot_prot(memslot, gfn, &writable);
1678 write_fault = kvm_is_write_fault(vcpu);
1679 if (kvm_is_error_hva(hva) || (write_fault && !writable)) {
1681 /* Prefetch Abort on I/O address */
1682 kvm_inject_pabt(vcpu, kvm_vcpu_get_hfar(vcpu));
1688 * Check for a cache maintenance operation. Since we
1689 * ended-up here, we know it is outside of any memory
1690 * slot. But we can't find out if that is for a device,
1691 * or if the guest is just being stupid. The only thing
1692 * we know for sure is that this range cannot be cached.
1694 * So let's assume that the guest is just being
1695 * cautious, and skip the instruction.
1697 if (kvm_vcpu_dabt_is_cm(vcpu)) {
1698 kvm_skip_instr(vcpu, kvm_vcpu_trap_il_is32bit(vcpu));
1704 * The IPA is reported as [MAX:12], so we need to
1705 * complement it with the bottom 12 bits from the
1706 * faulting VA. This is always 12 bits, irrespective
1709 fault_ipa |= kvm_vcpu_get_hfar(vcpu) & ((1 << 12) - 1);
1710 ret = io_mem_abort(vcpu, run, fault_ipa);
1714 /* Userspace should not be able to register out-of-bounds IPAs */
1715 VM_BUG_ON(fault_ipa >= KVM_PHYS_SIZE);
1717 if (fault_status == FSC_ACCESS) {
1718 handle_access_fault(vcpu, fault_ipa);
1723 ret = user_mem_abort(vcpu, fault_ipa, memslot, hva, fault_status);
1727 srcu_read_unlock(&vcpu->kvm->srcu, idx);
1731 static int handle_hva_to_gpa(struct kvm *kvm,
1732 unsigned long start,
1734 int (*handler)(struct kvm *kvm,
1735 gpa_t gpa, u64 size,
1739 struct kvm_memslots *slots;
1740 struct kvm_memory_slot *memslot;
1743 slots = kvm_memslots(kvm);
1745 /* we only care about the pages that the guest sees */
1746 kvm_for_each_memslot(memslot, slots) {
1747 unsigned long hva_start, hva_end;
1750 hva_start = max(start, memslot->userspace_addr);
1751 hva_end = min(end, memslot->userspace_addr +
1752 (memslot->npages << PAGE_SHIFT));
1753 if (hva_start >= hva_end)
1756 gpa = hva_to_gfn_memslot(hva_start, memslot) << PAGE_SHIFT;
1757 ret |= handler(kvm, gpa, (u64)(hva_end - hva_start), data);
1763 static int kvm_unmap_hva_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
1765 unmap_stage2_range(kvm, gpa, size);
1769 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
1771 unsigned long end = hva + PAGE_SIZE;
1776 trace_kvm_unmap_hva(hva);
1777 handle_hva_to_gpa(kvm, hva, end, &kvm_unmap_hva_handler, NULL);
1781 int kvm_unmap_hva_range(struct kvm *kvm,
1782 unsigned long start, unsigned long end)
1787 trace_kvm_unmap_hva_range(start, end);
1788 handle_hva_to_gpa(kvm, start, end, &kvm_unmap_hva_handler, NULL);
1792 static int kvm_set_spte_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
1794 pte_t *pte = (pte_t *)data;
1796 WARN_ON(size != PAGE_SIZE);
1798 * We can always call stage2_set_pte with KVM_S2PTE_FLAG_LOGGING_ACTIVE
1799 * flag clear because MMU notifiers will have unmapped a huge PMD before
1800 * calling ->change_pte() (which in turn calls kvm_set_spte_hva()) and
1801 * therefore stage2_set_pte() never needs to clear out a huge PMD
1802 * through this calling path.
1804 stage2_set_pte(kvm, NULL, gpa, pte, 0);
1809 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
1811 unsigned long end = hva + PAGE_SIZE;
1817 trace_kvm_set_spte_hva(hva);
1818 stage2_pte = pfn_pte(pte_pfn(pte), PAGE_S2);
1819 handle_hva_to_gpa(kvm, hva, end, &kvm_set_spte_handler, &stage2_pte);
1822 static int kvm_age_hva_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
1827 WARN_ON(size != PAGE_SIZE && size != PMD_SIZE);
1828 pmd = stage2_get_pmd(kvm, NULL, gpa);
1829 if (!pmd || pmd_none(*pmd)) /* Nothing there */
1832 if (pmd_thp_or_huge(*pmd)) /* THP, HugeTLB */
1833 return stage2_pmdp_test_and_clear_young(pmd);
1835 pte = pte_offset_kernel(pmd, gpa);
1839 return stage2_ptep_test_and_clear_young(pte);
1842 static int kvm_test_age_hva_handler(struct kvm *kvm, gpa_t gpa, u64 size, void *data)
1847 WARN_ON(size != PAGE_SIZE && size != PMD_SIZE);
1848 pmd = stage2_get_pmd(kvm, NULL, gpa);
1849 if (!pmd || pmd_none(*pmd)) /* Nothing there */
1852 if (pmd_thp_or_huge(*pmd)) /* THP, HugeTLB */
1853 return pmd_young(*pmd);
1855 pte = pte_offset_kernel(pmd, gpa);
1856 if (!pte_none(*pte)) /* Just a page... */
1857 return pte_young(*pte);
1862 int kvm_age_hva(struct kvm *kvm, unsigned long start, unsigned long end)
1866 trace_kvm_age_hva(start, end);
1867 return handle_hva_to_gpa(kvm, start, end, kvm_age_hva_handler, NULL);
1870 int kvm_test_age_hva(struct kvm *kvm, unsigned long hva)
1874 trace_kvm_test_age_hva(hva);
1875 return handle_hva_to_gpa(kvm, hva, hva, kvm_test_age_hva_handler, NULL);
1878 void kvm_mmu_free_memory_caches(struct kvm_vcpu *vcpu)
1880 mmu_free_memory_cache(&vcpu->arch.mmu_page_cache);
1883 phys_addr_t kvm_mmu_get_httbr(void)
1885 if (__kvm_cpu_uses_extended_idmap())
1886 return virt_to_phys(merged_hyp_pgd);
1888 return virt_to_phys(hyp_pgd);
1891 phys_addr_t kvm_get_idmap_vector(void)
1893 return hyp_idmap_vector;
1896 static int kvm_map_idmap_text(pgd_t *pgd)
1900 /* Create the idmap in the boot page tables */
1901 err = __create_hyp_mappings(pgd, __kvm_idmap_ptrs_per_pgd(),
1902 hyp_idmap_start, hyp_idmap_end,
1903 __phys_to_pfn(hyp_idmap_start),
1906 kvm_err("Failed to idmap %lx-%lx\n",
1907 hyp_idmap_start, hyp_idmap_end);
1912 int kvm_mmu_init(void)
1916 hyp_idmap_start = kvm_virt_to_phys(__hyp_idmap_text_start);
1917 hyp_idmap_start = ALIGN_DOWN(hyp_idmap_start, PAGE_SIZE);
1918 hyp_idmap_end = kvm_virt_to_phys(__hyp_idmap_text_end);
1919 hyp_idmap_end = ALIGN(hyp_idmap_end, PAGE_SIZE);
1920 hyp_idmap_vector = kvm_virt_to_phys(__kvm_hyp_init);
1923 * We rely on the linker script to ensure at build time that the HYP
1924 * init code does not cross a page boundary.
1926 BUG_ON((hyp_idmap_start ^ (hyp_idmap_end - 1)) & PAGE_MASK);
1928 kvm_debug("IDMAP page: %lx\n", hyp_idmap_start);
1929 kvm_debug("HYP VA range: %lx:%lx\n",
1930 kern_hyp_va(PAGE_OFFSET),
1931 kern_hyp_va((unsigned long)high_memory - 1));
1933 if (hyp_idmap_start >= kern_hyp_va(PAGE_OFFSET) &&
1934 hyp_idmap_start < kern_hyp_va((unsigned long)high_memory - 1) &&
1935 hyp_idmap_start != (unsigned long)__hyp_idmap_text_start) {
1937 * The idmap page is intersecting with the VA space,
1938 * it is not safe to continue further.
1940 kvm_err("IDMAP intersecting with HYP VA, unable to continue\n");
1945 hyp_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, hyp_pgd_order);
1947 kvm_err("Hyp mode PGD not allocated\n");
1952 if (__kvm_cpu_uses_extended_idmap()) {
1953 boot_hyp_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
1955 if (!boot_hyp_pgd) {
1956 kvm_err("Hyp boot PGD not allocated\n");
1961 err = kvm_map_idmap_text(boot_hyp_pgd);
1965 merged_hyp_pgd = (pgd_t *)__get_free_page(GFP_KERNEL | __GFP_ZERO);
1966 if (!merged_hyp_pgd) {
1967 kvm_err("Failed to allocate extra HYP pgd\n");
1970 __kvm_extend_hypmap(boot_hyp_pgd, hyp_pgd, merged_hyp_pgd,
1973 err = kvm_map_idmap_text(hyp_pgd);
1978 io_map_base = hyp_idmap_start;
1985 void kvm_arch_commit_memory_region(struct kvm *kvm,
1986 const struct kvm_userspace_memory_region *mem,
1987 const struct kvm_memory_slot *old,
1988 const struct kvm_memory_slot *new,
1989 enum kvm_mr_change change)
1992 * At this point memslot has been committed and there is an
1993 * allocated dirty_bitmap[], dirty pages will be be tracked while the
1994 * memory slot is write protected.
1996 if (change != KVM_MR_DELETE && mem->flags & KVM_MEM_LOG_DIRTY_PAGES)
1997 kvm_mmu_wp_memory_region(kvm, mem->slot);
2000 int kvm_arch_prepare_memory_region(struct kvm *kvm,
2001 struct kvm_memory_slot *memslot,
2002 const struct kvm_userspace_memory_region *mem,
2003 enum kvm_mr_change change)
2005 hva_t hva = mem->userspace_addr;
2006 hva_t reg_end = hva + mem->memory_size;
2007 bool writable = !(mem->flags & KVM_MEM_READONLY);
2010 if (change != KVM_MR_CREATE && change != KVM_MR_MOVE &&
2011 change != KVM_MR_FLAGS_ONLY)
2015 * Prevent userspace from creating a memory region outside of the IPA
2016 * space addressable by the KVM guest IPA space.
2018 if (memslot->base_gfn + memslot->npages >=
2019 (KVM_PHYS_SIZE >> PAGE_SHIFT))
2022 down_read(¤t->mm->mmap_sem);
2024 * A memory region could potentially cover multiple VMAs, and any holes
2025 * between them, so iterate over all of them to find out if we can map
2026 * any of them right now.
2028 * +--------------------------------------------+
2029 * +---------------+----------------+ +----------------+
2030 * | : VMA 1 | VMA 2 | | VMA 3 : |
2031 * +---------------+----------------+ +----------------+
2033 * +--------------------------------------------+
2036 struct vm_area_struct *vma = find_vma(current->mm, hva);
2037 hva_t vm_start, vm_end;
2039 if (!vma || vma->vm_start >= reg_end)
2043 * Mapping a read-only VMA is only allowed if the
2044 * memory region is configured as read-only.
2046 if (writable && !(vma->vm_flags & VM_WRITE)) {
2052 * Take the intersection of this VMA with the memory region
2054 vm_start = max(hva, vma->vm_start);
2055 vm_end = min(reg_end, vma->vm_end);
2057 if (vma->vm_flags & VM_PFNMAP) {
2058 gpa_t gpa = mem->guest_phys_addr +
2059 (vm_start - mem->userspace_addr);
2062 pa = (phys_addr_t)vma->vm_pgoff << PAGE_SHIFT;
2063 pa += vm_start - vma->vm_start;
2065 /* IO region dirty page logging not allowed */
2066 if (memslot->flags & KVM_MEM_LOG_DIRTY_PAGES) {
2071 ret = kvm_phys_addr_ioremap(kvm, gpa, pa,
2078 } while (hva < reg_end);
2080 if (change == KVM_MR_FLAGS_ONLY)
2083 spin_lock(&kvm->mmu_lock);
2085 unmap_stage2_range(kvm, mem->guest_phys_addr, mem->memory_size);
2087 stage2_flush_memslot(kvm, memslot);
2088 spin_unlock(&kvm->mmu_lock);
2090 up_read(¤t->mm->mmap_sem);
2094 void kvm_arch_free_memslot(struct kvm *kvm, struct kvm_memory_slot *free,
2095 struct kvm_memory_slot *dont)
2099 int kvm_arch_create_memslot(struct kvm *kvm, struct kvm_memory_slot *slot,
2100 unsigned long npages)
2105 void kvm_arch_memslots_updated(struct kvm *kvm, struct kvm_memslots *slots)
2109 void kvm_arch_flush_shadow_all(struct kvm *kvm)
2111 kvm_free_stage2_pgd(kvm);
2114 void kvm_arch_flush_shadow_memslot(struct kvm *kvm,
2115 struct kvm_memory_slot *slot)
2117 gpa_t gpa = slot->base_gfn << PAGE_SHIFT;
2118 phys_addr_t size = slot->npages << PAGE_SHIFT;
2120 spin_lock(&kvm->mmu_lock);
2121 unmap_stage2_range(kvm, gpa, size);
2122 spin_unlock(&kvm->mmu_lock);
2126 * See note at ARMv7 ARM B1.14.4 (TL;DR: S/W ops are not easily virtualized).
2129 * - S/W ops are local to a CPU (not broadcast)
2130 * - We have line migration behind our back (speculation)
2131 * - System caches don't support S/W at all (damn!)
2133 * In the face of the above, the best we can do is to try and convert
2134 * S/W ops to VA ops. Because the guest is not allowed to infer the
2135 * S/W to PA mapping, it can only use S/W to nuke the whole cache,
2136 * which is a rather good thing for us.
2138 * Also, it is only used when turning caches on/off ("The expected
2139 * usage of the cache maintenance instructions that operate by set/way
2140 * is associated with the cache maintenance instructions associated
2141 * with the powerdown and powerup of caches, if this is required by
2142 * the implementation.").
2144 * We use the following policy:
2146 * - If we trap a S/W operation, we enable VM trapping to detect
2147 * caches being turned on/off, and do a full clean.
2149 * - We flush the caches on both caches being turned on and off.
2151 * - Once the caches are enabled, we stop trapping VM ops.
2153 void kvm_set_way_flush(struct kvm_vcpu *vcpu)
2155 unsigned long hcr = *vcpu_hcr(vcpu);
2158 * If this is the first time we do a S/W operation
2159 * (i.e. HCR_TVM not set) flush the whole memory, and set the
2162 * Otherwise, rely on the VM trapping to wait for the MMU +
2163 * Caches to be turned off. At that point, we'll be able to
2164 * clean the caches again.
2166 if (!(hcr & HCR_TVM)) {
2167 trace_kvm_set_way_flush(*vcpu_pc(vcpu),
2168 vcpu_has_cache_enabled(vcpu));
2169 stage2_flush_vm(vcpu->kvm);
2170 *vcpu_hcr(vcpu) = hcr | HCR_TVM;
2174 void kvm_toggle_cache(struct kvm_vcpu *vcpu, bool was_enabled)
2176 bool now_enabled = vcpu_has_cache_enabled(vcpu);
2179 * If switching the MMU+caches on, need to invalidate the caches.
2180 * If switching it off, need to clean the caches.
2181 * Clean + invalidate does the trick always.
2183 if (now_enabled != was_enabled)
2184 stage2_flush_vm(vcpu->kvm);
2186 /* Caches are now on, stop trapping VM ops (until a S/W op) */
2188 *vcpu_hcr(vcpu) &= ~HCR_TVM;
2190 trace_kvm_toggle_cache(*vcpu_pc(vcpu), was_enabled, now_enabled);