2 * Copyright (C) 2012 - Virtual Open Systems and Columbia University
3 * Author: Christoffer Dall <c.dall@virtualopensystems.com>
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License, version 2, as
7 * published by the Free Software Foundation.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
19 #include <linux/mman.h>
20 #include <linux/kvm_host.h>
22 #include <linux/hugetlb.h>
23 #include <trace/events/kvm.h>
24 #include <asm/pgalloc.h>
25 #include <asm/cacheflush.h>
26 #include <asm/kvm_arm.h>
27 #include <asm/kvm_mmu.h>
28 #include <asm/kvm_mmio.h>
29 #include <asm/kvm_asm.h>
30 #include <asm/kvm_emulate.h>
34 extern char __hyp_idmap_text_start[], __hyp_idmap_text_end[];
36 static pgd_t *boot_hyp_pgd;
37 static pgd_t *hyp_pgd;
38 static DEFINE_MUTEX(kvm_hyp_pgd_mutex);
40 static void *init_bounce_page;
41 static unsigned long hyp_idmap_start;
42 static unsigned long hyp_idmap_end;
43 static phys_addr_t hyp_idmap_vector;
45 #define hyp_pgd_order get_order(PTRS_PER_PGD * sizeof(pgd_t))
47 #define kvm_pmd_huge(_x) (pmd_huge(_x) || pmd_trans_huge(_x))
48 #define kvm_pud_huge(_x) pud_huge(_x)
50 #define KVM_S2PTE_FLAG_IS_IOMAP (1UL << 0)
51 #define KVM_S2_FLAG_LOGGING_ACTIVE (1UL << 1)
53 static bool memslot_is_logging(struct kvm_memory_slot *memslot)
55 return memslot->dirty_bitmap && !(memslot->flags & KVM_MEM_READONLY);
59 * kvm_flush_remote_tlbs() - flush all VM TLB entries for v7/8
60 * @kvm: pointer to kvm structure.
62 * Interface to HYP function to flush all VM TLB entries
64 void kvm_flush_remote_tlbs(struct kvm *kvm)
66 kvm_call_hyp(__kvm_tlb_flush_vmid, kvm);
69 static void kvm_tlb_flush_vmid_ipa(struct kvm *kvm, phys_addr_t ipa)
72 * This function also gets called when dealing with HYP page
73 * tables. As HYP doesn't have an associated struct kvm (and
74 * the HYP page tables are fairly static), we don't do
78 kvm_call_hyp(__kvm_tlb_flush_vmid_ipa, kvm, ipa);
82 * D-Cache management functions. They take the page table entries by
83 * value, as they are flushing the cache using the kernel mapping (or
86 static void kvm_flush_dcache_pte(pte_t pte)
88 __kvm_flush_dcache_pte(pte);
91 static void kvm_flush_dcache_pmd(pmd_t pmd)
93 __kvm_flush_dcache_pmd(pmd);
96 static void kvm_flush_dcache_pud(pud_t pud)
98 __kvm_flush_dcache_pud(pud);
102 * stage2_dissolve_pmd() - clear and flush huge PMD entry
103 * @kvm: pointer to kvm structure.
105 * @pmd: pmd pointer for IPA
107 * Function clears a PMD entry, flushes addr 1st and 2nd stage TLBs. Marks all
108 * pages in the range dirty.
110 static void stage2_dissolve_pmd(struct kvm *kvm, phys_addr_t addr, pmd_t *pmd)
112 if (!kvm_pmd_huge(*pmd))
116 kvm_tlb_flush_vmid_ipa(kvm, addr);
117 put_page(virt_to_page(pmd));
120 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
125 BUG_ON(max > KVM_NR_MEM_OBJS);
126 if (cache->nobjs >= min)
128 while (cache->nobjs < max) {
129 page = (void *)__get_free_page(PGALLOC_GFP);
132 cache->objects[cache->nobjs++] = page;
137 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
140 free_page((unsigned long)mc->objects[--mc->nobjs]);
143 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc)
147 BUG_ON(!mc || !mc->nobjs);
148 p = mc->objects[--mc->nobjs];
152 static void clear_pgd_entry(struct kvm *kvm, pgd_t *pgd, phys_addr_t addr)
154 pud_t *pud_table __maybe_unused = pud_offset(pgd, 0);
156 kvm_tlb_flush_vmid_ipa(kvm, addr);
157 pud_free(NULL, pud_table);
158 put_page(virt_to_page(pgd));
161 static void clear_pud_entry(struct kvm *kvm, pud_t *pud, phys_addr_t addr)
163 pmd_t *pmd_table = pmd_offset(pud, 0);
164 VM_BUG_ON(pud_huge(*pud));
166 kvm_tlb_flush_vmid_ipa(kvm, addr);
167 pmd_free(NULL, pmd_table);
168 put_page(virt_to_page(pud));
171 static void clear_pmd_entry(struct kvm *kvm, pmd_t *pmd, phys_addr_t addr)
173 pte_t *pte_table = pte_offset_kernel(pmd, 0);
174 VM_BUG_ON(kvm_pmd_huge(*pmd));
176 kvm_tlb_flush_vmid_ipa(kvm, addr);
177 pte_free_kernel(NULL, pte_table);
178 put_page(virt_to_page(pmd));
182 * Unmapping vs dcache management:
184 * If a guest maps certain memory pages as uncached, all writes will
185 * bypass the data cache and go directly to RAM. However, the CPUs
186 * can still speculate reads (not writes) and fill cache lines with
189 * Those cache lines will be *clean* cache lines though, so a
190 * clean+invalidate operation is equivalent to an invalidate
191 * operation, because no cache lines are marked dirty.
193 * Those clean cache lines could be filled prior to an uncached write
194 * by the guest, and the cache coherent IO subsystem would therefore
195 * end up writing old data to disk.
197 * This is why right after unmapping a page/section and invalidating
198 * the corresponding TLBs, we call kvm_flush_dcache_p*() to make sure
199 * the IO subsystem will never hit in the cache.
201 static void unmap_ptes(struct kvm *kvm, pmd_t *pmd,
202 phys_addr_t addr, phys_addr_t end)
204 phys_addr_t start_addr = addr;
205 pte_t *pte, *start_pte;
207 start_pte = pte = pte_offset_kernel(pmd, addr);
209 if (!pte_none(*pte)) {
210 pte_t old_pte = *pte;
212 kvm_set_pte(pte, __pte(0));
213 kvm_tlb_flush_vmid_ipa(kvm, addr);
215 /* No need to invalidate the cache for device mappings */
216 if ((pte_val(old_pte) & PAGE_S2_DEVICE) != PAGE_S2_DEVICE)
217 kvm_flush_dcache_pte(old_pte);
219 put_page(virt_to_page(pte));
221 } while (pte++, addr += PAGE_SIZE, addr != end);
223 if (kvm_pte_table_empty(kvm, start_pte))
224 clear_pmd_entry(kvm, pmd, start_addr);
227 static void unmap_pmds(struct kvm *kvm, pud_t *pud,
228 phys_addr_t addr, phys_addr_t end)
230 phys_addr_t next, start_addr = addr;
231 pmd_t *pmd, *start_pmd;
233 start_pmd = pmd = pmd_offset(pud, addr);
235 next = kvm_pmd_addr_end(addr, end);
236 if (!pmd_none(*pmd)) {
237 if (kvm_pmd_huge(*pmd)) {
238 pmd_t old_pmd = *pmd;
241 kvm_tlb_flush_vmid_ipa(kvm, addr);
243 kvm_flush_dcache_pmd(old_pmd);
245 put_page(virt_to_page(pmd));
247 unmap_ptes(kvm, pmd, addr, next);
250 } while (pmd++, addr = next, addr != end);
252 if (kvm_pmd_table_empty(kvm, start_pmd))
253 clear_pud_entry(kvm, pud, start_addr);
256 static void unmap_puds(struct kvm *kvm, pgd_t *pgd,
257 phys_addr_t addr, phys_addr_t end)
259 phys_addr_t next, start_addr = addr;
260 pud_t *pud, *start_pud;
262 start_pud = pud = pud_offset(pgd, addr);
264 next = kvm_pud_addr_end(addr, end);
265 if (!pud_none(*pud)) {
266 if (pud_huge(*pud)) {
267 pud_t old_pud = *pud;
270 kvm_tlb_flush_vmid_ipa(kvm, addr);
272 kvm_flush_dcache_pud(old_pud);
274 put_page(virt_to_page(pud));
276 unmap_pmds(kvm, pud, addr, next);
279 } while (pud++, addr = next, addr != end);
281 if (kvm_pud_table_empty(kvm, start_pud))
282 clear_pgd_entry(kvm, pgd, start_addr);
286 static void unmap_range(struct kvm *kvm, pgd_t *pgdp,
287 phys_addr_t start, u64 size)
290 phys_addr_t addr = start, end = start + size;
293 pgd = pgdp + kvm_pgd_index(addr);
295 next = kvm_pgd_addr_end(addr, end);
297 unmap_puds(kvm, pgd, addr, next);
298 } while (pgd++, addr = next, addr != end);
301 static void stage2_flush_ptes(struct kvm *kvm, pmd_t *pmd,
302 phys_addr_t addr, phys_addr_t end)
306 pte = pte_offset_kernel(pmd, addr);
308 if (!pte_none(*pte) &&
309 (pte_val(*pte) & PAGE_S2_DEVICE) != PAGE_S2_DEVICE)
310 kvm_flush_dcache_pte(*pte);
311 } while (pte++, addr += PAGE_SIZE, addr != end);
314 static void stage2_flush_pmds(struct kvm *kvm, pud_t *pud,
315 phys_addr_t addr, phys_addr_t end)
320 pmd = pmd_offset(pud, addr);
322 next = kvm_pmd_addr_end(addr, end);
323 if (!pmd_none(*pmd)) {
324 if (kvm_pmd_huge(*pmd))
325 kvm_flush_dcache_pmd(*pmd);
327 stage2_flush_ptes(kvm, pmd, addr, next);
329 } while (pmd++, addr = next, addr != end);
332 static void stage2_flush_puds(struct kvm *kvm, pgd_t *pgd,
333 phys_addr_t addr, phys_addr_t end)
338 pud = pud_offset(pgd, addr);
340 next = kvm_pud_addr_end(addr, end);
341 if (!pud_none(*pud)) {
343 kvm_flush_dcache_pud(*pud);
345 stage2_flush_pmds(kvm, pud, addr, next);
347 } while (pud++, addr = next, addr != end);
350 static void stage2_flush_memslot(struct kvm *kvm,
351 struct kvm_memory_slot *memslot)
353 phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
354 phys_addr_t end = addr + PAGE_SIZE * memslot->npages;
358 pgd = kvm->arch.pgd + kvm_pgd_index(addr);
360 next = kvm_pgd_addr_end(addr, end);
361 stage2_flush_puds(kvm, pgd, addr, next);
362 } while (pgd++, addr = next, addr != end);
366 * stage2_flush_vm - Invalidate cache for pages mapped in stage 2
367 * @kvm: The struct kvm pointer
369 * Go through the stage 2 page tables and invalidate any cache lines
370 * backing memory already mapped to the VM.
372 static void stage2_flush_vm(struct kvm *kvm)
374 struct kvm_memslots *slots;
375 struct kvm_memory_slot *memslot;
378 idx = srcu_read_lock(&kvm->srcu);
379 spin_lock(&kvm->mmu_lock);
381 slots = kvm_memslots(kvm);
382 kvm_for_each_memslot(memslot, slots)
383 stage2_flush_memslot(kvm, memslot);
385 spin_unlock(&kvm->mmu_lock);
386 srcu_read_unlock(&kvm->srcu, idx);
390 * free_boot_hyp_pgd - free HYP boot page tables
392 * Free the HYP boot page tables. The bounce page is also freed.
394 void free_boot_hyp_pgd(void)
396 mutex_lock(&kvm_hyp_pgd_mutex);
399 unmap_range(NULL, boot_hyp_pgd, hyp_idmap_start, PAGE_SIZE);
400 unmap_range(NULL, boot_hyp_pgd, TRAMPOLINE_VA, PAGE_SIZE);
401 free_pages((unsigned long)boot_hyp_pgd, hyp_pgd_order);
406 unmap_range(NULL, hyp_pgd, TRAMPOLINE_VA, PAGE_SIZE);
408 free_page((unsigned long)init_bounce_page);
409 init_bounce_page = NULL;
411 mutex_unlock(&kvm_hyp_pgd_mutex);
415 * free_hyp_pgds - free Hyp-mode page tables
417 * Assumes hyp_pgd is a page table used strictly in Hyp-mode and
418 * therefore contains either mappings in the kernel memory area (above
419 * PAGE_OFFSET), or device mappings in the vmalloc range (from
420 * VMALLOC_START to VMALLOC_END).
422 * boot_hyp_pgd should only map two pages for the init code.
424 void free_hyp_pgds(void)
430 mutex_lock(&kvm_hyp_pgd_mutex);
433 for (addr = PAGE_OFFSET; virt_addr_valid(addr); addr += PGDIR_SIZE)
434 unmap_range(NULL, hyp_pgd, KERN_TO_HYP(addr), PGDIR_SIZE);
435 for (addr = VMALLOC_START; is_vmalloc_addr((void*)addr); addr += PGDIR_SIZE)
436 unmap_range(NULL, hyp_pgd, KERN_TO_HYP(addr), PGDIR_SIZE);
438 free_pages((unsigned long)hyp_pgd, hyp_pgd_order);
442 mutex_unlock(&kvm_hyp_pgd_mutex);
445 static void create_hyp_pte_mappings(pmd_t *pmd, unsigned long start,
446 unsigned long end, unsigned long pfn,
454 pte = pte_offset_kernel(pmd, addr);
455 kvm_set_pte(pte, pfn_pte(pfn, prot));
456 get_page(virt_to_page(pte));
457 kvm_flush_dcache_to_poc(pte, sizeof(*pte));
459 } while (addr += PAGE_SIZE, addr != end);
462 static int create_hyp_pmd_mappings(pud_t *pud, unsigned long start,
463 unsigned long end, unsigned long pfn,
468 unsigned long addr, next;
472 pmd = pmd_offset(pud, addr);
474 BUG_ON(pmd_sect(*pmd));
476 if (pmd_none(*pmd)) {
477 pte = pte_alloc_one_kernel(NULL, addr);
479 kvm_err("Cannot allocate Hyp pte\n");
482 pmd_populate_kernel(NULL, pmd, pte);
483 get_page(virt_to_page(pmd));
484 kvm_flush_dcache_to_poc(pmd, sizeof(*pmd));
487 next = pmd_addr_end(addr, end);
489 create_hyp_pte_mappings(pmd, addr, next, pfn, prot);
490 pfn += (next - addr) >> PAGE_SHIFT;
491 } while (addr = next, addr != end);
496 static int create_hyp_pud_mappings(pgd_t *pgd, unsigned long start,
497 unsigned long end, unsigned long pfn,
502 unsigned long addr, next;
507 pud = pud_offset(pgd, addr);
509 if (pud_none_or_clear_bad(pud)) {
510 pmd = pmd_alloc_one(NULL, addr);
512 kvm_err("Cannot allocate Hyp pmd\n");
515 pud_populate(NULL, pud, pmd);
516 get_page(virt_to_page(pud));
517 kvm_flush_dcache_to_poc(pud, sizeof(*pud));
520 next = pud_addr_end(addr, end);
521 ret = create_hyp_pmd_mappings(pud, addr, next, pfn, prot);
524 pfn += (next - addr) >> PAGE_SHIFT;
525 } while (addr = next, addr != end);
530 static int __create_hyp_mappings(pgd_t *pgdp,
531 unsigned long start, unsigned long end,
532 unsigned long pfn, pgprot_t prot)
536 unsigned long addr, next;
539 mutex_lock(&kvm_hyp_pgd_mutex);
540 addr = start & PAGE_MASK;
541 end = PAGE_ALIGN(end);
543 pgd = pgdp + pgd_index(addr);
545 if (pgd_none(*pgd)) {
546 pud = pud_alloc_one(NULL, addr);
548 kvm_err("Cannot allocate Hyp pud\n");
552 pgd_populate(NULL, pgd, pud);
553 get_page(virt_to_page(pgd));
554 kvm_flush_dcache_to_poc(pgd, sizeof(*pgd));
557 next = pgd_addr_end(addr, end);
558 err = create_hyp_pud_mappings(pgd, addr, next, pfn, prot);
561 pfn += (next - addr) >> PAGE_SHIFT;
562 } while (addr = next, addr != end);
564 mutex_unlock(&kvm_hyp_pgd_mutex);
568 static phys_addr_t kvm_kaddr_to_phys(void *kaddr)
570 if (!is_vmalloc_addr(kaddr)) {
571 BUG_ON(!virt_addr_valid(kaddr));
574 return page_to_phys(vmalloc_to_page(kaddr)) +
575 offset_in_page(kaddr);
580 * create_hyp_mappings - duplicate a kernel virtual address range in Hyp mode
581 * @from: The virtual kernel start address of the range
582 * @to: The virtual kernel end address of the range (exclusive)
584 * The same virtual address as the kernel virtual address is also used
585 * in Hyp-mode mapping (modulo HYP_PAGE_OFFSET) to the same underlying
588 int create_hyp_mappings(void *from, void *to)
590 phys_addr_t phys_addr;
591 unsigned long virt_addr;
592 unsigned long start = KERN_TO_HYP((unsigned long)from);
593 unsigned long end = KERN_TO_HYP((unsigned long)to);
595 start = start & PAGE_MASK;
596 end = PAGE_ALIGN(end);
598 for (virt_addr = start; virt_addr < end; virt_addr += PAGE_SIZE) {
601 phys_addr = kvm_kaddr_to_phys(from + virt_addr - start);
602 err = __create_hyp_mappings(hyp_pgd, virt_addr,
603 virt_addr + PAGE_SIZE,
604 __phys_to_pfn(phys_addr),
614 * create_hyp_io_mappings - duplicate a kernel IO mapping into Hyp mode
615 * @from: The kernel start VA of the range
616 * @to: The kernel end VA of the range (exclusive)
617 * @phys_addr: The physical start address which gets mapped
619 * The resulting HYP VA is the same as the kernel VA, modulo
622 int create_hyp_io_mappings(void *from, void *to, phys_addr_t phys_addr)
624 unsigned long start = KERN_TO_HYP((unsigned long)from);
625 unsigned long end = KERN_TO_HYP((unsigned long)to);
627 /* Check for a valid kernel IO mapping */
628 if (!is_vmalloc_addr(from) || !is_vmalloc_addr(to - 1))
631 return __create_hyp_mappings(hyp_pgd, start, end,
632 __phys_to_pfn(phys_addr), PAGE_HYP_DEVICE);
635 /* Free the HW pgd, one page at a time */
636 static void kvm_free_hwpgd(void *hwpgd)
638 free_pages_exact(hwpgd, kvm_get_hwpgd_size());
641 /* Allocate the HW PGD, making sure that each page gets its own refcount */
642 static void *kvm_alloc_hwpgd(void)
644 unsigned int size = kvm_get_hwpgd_size();
646 return alloc_pages_exact(size, GFP_KERNEL | __GFP_ZERO);
650 * kvm_alloc_stage2_pgd - allocate level-1 table for stage-2 translation.
651 * @kvm: The KVM struct pointer for the VM.
653 * Allocates the 1st level table only of size defined by S2_PGD_ORDER (can
654 * support either full 40-bit input addresses or limited to 32-bit input
655 * addresses). Clears the allocated pages.
657 * Note we don't need locking here as this is only called when the VM is
658 * created, which can only be done once.
660 int kvm_alloc_stage2_pgd(struct kvm *kvm)
665 if (kvm->arch.pgd != NULL) {
666 kvm_err("kvm_arch already initialized?\n");
670 hwpgd = kvm_alloc_hwpgd();
674 /* When the kernel uses more levels of page tables than the
675 * guest, we allocate a fake PGD and pre-populate it to point
676 * to the next-level page table, which will be the real
677 * initial page table pointed to by the VTTBR.
679 * When KVM_PREALLOC_LEVEL==2, we allocate a single page for
680 * the PMD and the kernel will use folded pud.
681 * When KVM_PREALLOC_LEVEL==1, we allocate 2 consecutive PUD
684 if (KVM_PREALLOC_LEVEL > 0) {
688 * Allocate fake pgd for the page table manipulation macros to
689 * work. This is not used by the hardware and we have no
690 * alignment requirement for this allocation.
692 pgd = (pgd_t *)kmalloc(PTRS_PER_S2_PGD * sizeof(pgd_t),
693 GFP_KERNEL | __GFP_ZERO);
696 kvm_free_hwpgd(hwpgd);
700 /* Plug the HW PGD into the fake one. */
701 for (i = 0; i < PTRS_PER_S2_PGD; i++) {
702 if (KVM_PREALLOC_LEVEL == 1)
703 pgd_populate(NULL, pgd + i,
704 (pud_t *)hwpgd + i * PTRS_PER_PUD);
705 else if (KVM_PREALLOC_LEVEL == 2)
706 pud_populate(NULL, pud_offset(pgd, 0) + i,
707 (pmd_t *)hwpgd + i * PTRS_PER_PMD);
711 * Allocate actual first-level Stage-2 page table used by the
712 * hardware for Stage-2 page table walks.
714 pgd = (pgd_t *)hwpgd;
723 * unmap_stage2_range -- Clear stage2 page table entries to unmap a range
724 * @kvm: The VM pointer
725 * @start: The intermediate physical base address of the range to unmap
726 * @size: The size of the area to unmap
728 * Clear a range of stage-2 mappings, lowering the various ref-counts. Must
729 * be called while holding mmu_lock (unless for freeing the stage2 pgd before
730 * destroying the VM), otherwise another faulting VCPU may come in and mess
731 * with things behind our backs.
733 static void unmap_stage2_range(struct kvm *kvm, phys_addr_t start, u64 size)
735 unmap_range(kvm, kvm->arch.pgd, start, size);
738 static void stage2_unmap_memslot(struct kvm *kvm,
739 struct kvm_memory_slot *memslot)
741 hva_t hva = memslot->userspace_addr;
742 phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
743 phys_addr_t size = PAGE_SIZE * memslot->npages;
744 hva_t reg_end = hva + size;
747 * A memory region could potentially cover multiple VMAs, and any holes
748 * between them, so iterate over all of them to find out if we should
751 * +--------------------------------------------+
752 * +---------------+----------------+ +----------------+
753 * | : VMA 1 | VMA 2 | | VMA 3 : |
754 * +---------------+----------------+ +----------------+
756 * +--------------------------------------------+
759 struct vm_area_struct *vma = find_vma(current->mm, hva);
760 hva_t vm_start, vm_end;
762 if (!vma || vma->vm_start >= reg_end)
766 * Take the intersection of this VMA with the memory region
768 vm_start = max(hva, vma->vm_start);
769 vm_end = min(reg_end, vma->vm_end);
771 if (!(vma->vm_flags & VM_PFNMAP)) {
772 gpa_t gpa = addr + (vm_start - memslot->userspace_addr);
773 unmap_stage2_range(kvm, gpa, vm_end - vm_start);
776 } while (hva < reg_end);
780 * stage2_unmap_vm - Unmap Stage-2 RAM mappings
781 * @kvm: The struct kvm pointer
783 * Go through the memregions and unmap any reguler RAM
784 * backing memory already mapped to the VM.
786 void stage2_unmap_vm(struct kvm *kvm)
788 struct kvm_memslots *slots;
789 struct kvm_memory_slot *memslot;
792 idx = srcu_read_lock(&kvm->srcu);
793 spin_lock(&kvm->mmu_lock);
795 slots = kvm_memslots(kvm);
796 kvm_for_each_memslot(memslot, slots)
797 stage2_unmap_memslot(kvm, memslot);
799 spin_unlock(&kvm->mmu_lock);
800 srcu_read_unlock(&kvm->srcu, idx);
804 * kvm_free_stage2_pgd - free all stage-2 tables
805 * @kvm: The KVM struct pointer for the VM.
807 * Walks the level-1 page table pointed to by kvm->arch.pgd and frees all
808 * underlying level-2 and level-3 tables before freeing the actual level-1 table
809 * and setting the struct pointer to NULL.
811 * Note we don't need locking here as this is only called when the VM is
812 * destroyed, which can only be done once.
814 void kvm_free_stage2_pgd(struct kvm *kvm)
816 if (kvm->arch.pgd == NULL)
819 unmap_stage2_range(kvm, 0, KVM_PHYS_SIZE);
820 kvm_free_hwpgd(kvm_get_hwpgd(kvm));
821 if (KVM_PREALLOC_LEVEL > 0)
822 kfree(kvm->arch.pgd);
824 kvm->arch.pgd = NULL;
827 static pud_t *stage2_get_pud(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
833 pgd = kvm->arch.pgd + kvm_pgd_index(addr);
834 if (WARN_ON(pgd_none(*pgd))) {
837 pud = mmu_memory_cache_alloc(cache);
838 pgd_populate(NULL, pgd, pud);
839 get_page(virt_to_page(pgd));
842 return pud_offset(pgd, addr);
845 static pmd_t *stage2_get_pmd(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
851 pud = stage2_get_pud(kvm, cache, addr);
852 if (pud_none(*pud)) {
855 pmd = mmu_memory_cache_alloc(cache);
856 pud_populate(NULL, pud, pmd);
857 get_page(virt_to_page(pud));
860 return pmd_offset(pud, addr);
863 static int stage2_set_pmd_huge(struct kvm *kvm, struct kvm_mmu_memory_cache
864 *cache, phys_addr_t addr, const pmd_t *new_pmd)
868 pmd = stage2_get_pmd(kvm, cache, addr);
872 * Mapping in huge pages should only happen through a fault. If a
873 * page is merged into a transparent huge page, the individual
874 * subpages of that huge page should be unmapped through MMU
875 * notifiers before we get here.
877 * Merging of CompoundPages is not supported; they should become
878 * splitting first, unmapped, merged, and mapped back in on-demand.
880 VM_BUG_ON(pmd_present(*pmd) && pmd_pfn(*pmd) != pmd_pfn(*new_pmd));
883 kvm_set_pmd(pmd, *new_pmd);
884 if (pmd_present(old_pmd))
885 kvm_tlb_flush_vmid_ipa(kvm, addr);
887 get_page(virt_to_page(pmd));
891 static int stage2_set_pte(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
892 phys_addr_t addr, const pte_t *new_pte,
897 bool iomap = flags & KVM_S2PTE_FLAG_IS_IOMAP;
898 bool logging_active = flags & KVM_S2_FLAG_LOGGING_ACTIVE;
900 VM_BUG_ON(logging_active && !cache);
902 /* Create stage-2 page table mapping - Levels 0 and 1 */
903 pmd = stage2_get_pmd(kvm, cache, addr);
906 * Ignore calls from kvm_set_spte_hva for unallocated
913 * While dirty page logging - dissolve huge PMD, then continue on to
917 stage2_dissolve_pmd(kvm, addr, pmd);
919 /* Create stage-2 page mappings - Level 2 */
920 if (pmd_none(*pmd)) {
922 return 0; /* ignore calls from kvm_set_spte_hva */
923 pte = mmu_memory_cache_alloc(cache);
925 pmd_populate_kernel(NULL, pmd, pte);
926 get_page(virt_to_page(pmd));
929 pte = pte_offset_kernel(pmd, addr);
931 if (iomap && pte_present(*pte))
934 /* Create 2nd stage page table mapping - Level 3 */
936 kvm_set_pte(pte, *new_pte);
937 if (pte_present(old_pte))
938 kvm_tlb_flush_vmid_ipa(kvm, addr);
940 get_page(virt_to_page(pte));
946 * kvm_phys_addr_ioremap - map a device range to guest IPA
948 * @kvm: The KVM pointer
949 * @guest_ipa: The IPA at which to insert the mapping
950 * @pa: The physical address of the device
951 * @size: The size of the mapping
953 int kvm_phys_addr_ioremap(struct kvm *kvm, phys_addr_t guest_ipa,
954 phys_addr_t pa, unsigned long size, bool writable)
956 phys_addr_t addr, end;
959 struct kvm_mmu_memory_cache cache = { 0, };
961 end = (guest_ipa + size + PAGE_SIZE - 1) & PAGE_MASK;
962 pfn = __phys_to_pfn(pa);
964 for (addr = guest_ipa; addr < end; addr += PAGE_SIZE) {
965 pte_t pte = pfn_pte(pfn, PAGE_S2_DEVICE);
968 kvm_set_s2pte_writable(&pte);
970 ret = mmu_topup_memory_cache(&cache, KVM_MMU_CACHE_MIN_PAGES,
974 spin_lock(&kvm->mmu_lock);
975 ret = stage2_set_pte(kvm, &cache, addr, &pte,
976 KVM_S2PTE_FLAG_IS_IOMAP);
977 spin_unlock(&kvm->mmu_lock);
985 mmu_free_memory_cache(&cache);
989 static bool transparent_hugepage_adjust(pfn_t *pfnp, phys_addr_t *ipap)
992 gfn_t gfn = *ipap >> PAGE_SHIFT;
994 if (PageTransCompound(pfn_to_page(pfn))) {
997 * The address we faulted on is backed by a transparent huge
998 * page. However, because we map the compound huge page and
999 * not the individual tail page, we need to transfer the
1000 * refcount to the head page. We have to be careful that the
1001 * THP doesn't start to split while we are adjusting the
1004 * We are sure this doesn't happen, because mmu_notifier_retry
1005 * was successful and we are holding the mmu_lock, so if this
1006 * THP is trying to split, it will be blocked in the mmu
1007 * notifier before touching any of the pages, specifically
1008 * before being able to call __split_huge_page_refcount().
1010 * We can therefore safely transfer the refcount from PG_tail
1011 * to PG_head and switch the pfn from a tail page to the head
1014 mask = PTRS_PER_PMD - 1;
1015 VM_BUG_ON((gfn & mask) != (pfn & mask));
1018 kvm_release_pfn_clean(pfn);
1030 static bool kvm_is_write_fault(struct kvm_vcpu *vcpu)
1032 if (kvm_vcpu_trap_is_iabt(vcpu))
1035 return kvm_vcpu_dabt_iswrite(vcpu);
1038 static bool kvm_is_device_pfn(unsigned long pfn)
1040 return !pfn_valid(pfn);
1044 * stage2_wp_ptes - write protect PMD range
1045 * @pmd: pointer to pmd entry
1046 * @addr: range start address
1047 * @end: range end address
1049 static void stage2_wp_ptes(pmd_t *pmd, phys_addr_t addr, phys_addr_t end)
1053 pte = pte_offset_kernel(pmd, addr);
1055 if (!pte_none(*pte)) {
1056 if (!kvm_s2pte_readonly(pte))
1057 kvm_set_s2pte_readonly(pte);
1059 } while (pte++, addr += PAGE_SIZE, addr != end);
1063 * stage2_wp_pmds - write protect PUD range
1064 * @pud: pointer to pud entry
1065 * @addr: range start address
1066 * @end: range end address
1068 static void stage2_wp_pmds(pud_t *pud, phys_addr_t addr, phys_addr_t end)
1073 pmd = pmd_offset(pud, addr);
1076 next = kvm_pmd_addr_end(addr, end);
1077 if (!pmd_none(*pmd)) {
1078 if (kvm_pmd_huge(*pmd)) {
1079 if (!kvm_s2pmd_readonly(pmd))
1080 kvm_set_s2pmd_readonly(pmd);
1082 stage2_wp_ptes(pmd, addr, next);
1085 } while (pmd++, addr = next, addr != end);
1089 * stage2_wp_puds - write protect PGD range
1090 * @pgd: pointer to pgd entry
1091 * @addr: range start address
1092 * @end: range end address
1094 * Process PUD entries, for a huge PUD we cause a panic.
1096 static void stage2_wp_puds(pgd_t *pgd, phys_addr_t addr, phys_addr_t end)
1101 pud = pud_offset(pgd, addr);
1103 next = kvm_pud_addr_end(addr, end);
1104 if (!pud_none(*pud)) {
1105 /* TODO:PUD not supported, revisit later if supported */
1106 BUG_ON(kvm_pud_huge(*pud));
1107 stage2_wp_pmds(pud, addr, next);
1109 } while (pud++, addr = next, addr != end);
1113 * stage2_wp_range() - write protect stage2 memory region range
1114 * @kvm: The KVM pointer
1115 * @addr: Start address of range
1116 * @end: End address of range
1118 static void stage2_wp_range(struct kvm *kvm, phys_addr_t addr, phys_addr_t end)
1123 pgd = kvm->arch.pgd + kvm_pgd_index(addr);
1126 * Release kvm_mmu_lock periodically if the memory region is
1127 * large. Otherwise, we may see kernel panics with
1128 * CONFIG_DETECT_HUNG_TASK, CONFIG_LOCKUP_DETECTOR,
1129 * CONFIG_LOCKDEP. Additionally, holding the lock too long
1130 * will also starve other vCPUs.
1132 if (need_resched() || spin_needbreak(&kvm->mmu_lock))
1133 cond_resched_lock(&kvm->mmu_lock);
1135 next = kvm_pgd_addr_end(addr, end);
1136 if (pgd_present(*pgd))
1137 stage2_wp_puds(pgd, addr, next);
1138 } while (pgd++, addr = next, addr != end);
1142 * kvm_mmu_wp_memory_region() - write protect stage 2 entries for memory slot
1143 * @kvm: The KVM pointer
1144 * @slot: The memory slot to write protect
1146 * Called to start logging dirty pages after memory region
1147 * KVM_MEM_LOG_DIRTY_PAGES operation is called. After this function returns
1148 * all present PMD and PTEs are write protected in the memory region.
1149 * Afterwards read of dirty page log can be called.
1151 * Acquires kvm_mmu_lock. Called with kvm->slots_lock mutex acquired,
1152 * serializing operations for VM memory regions.
1154 void kvm_mmu_wp_memory_region(struct kvm *kvm, int slot)
1156 struct kvm_memory_slot *memslot = id_to_memslot(kvm->memslots, slot);
1157 phys_addr_t start = memslot->base_gfn << PAGE_SHIFT;
1158 phys_addr_t end = (memslot->base_gfn + memslot->npages) << PAGE_SHIFT;
1160 spin_lock(&kvm->mmu_lock);
1161 stage2_wp_range(kvm, start, end);
1162 spin_unlock(&kvm->mmu_lock);
1163 kvm_flush_remote_tlbs(kvm);
1167 * kvm_mmu_write_protect_pt_masked() - write protect dirty pages
1168 * @kvm: The KVM pointer
1169 * @slot: The memory slot associated with mask
1170 * @gfn_offset: The gfn offset in memory slot
1171 * @mask: The mask of dirty pages at offset 'gfn_offset' in this memory
1172 * slot to be write protected
1174 * Walks bits set in mask write protects the associated pte's. Caller must
1175 * acquire kvm_mmu_lock.
1177 static void kvm_mmu_write_protect_pt_masked(struct kvm *kvm,
1178 struct kvm_memory_slot *slot,
1179 gfn_t gfn_offset, unsigned long mask)
1181 phys_addr_t base_gfn = slot->base_gfn + gfn_offset;
1182 phys_addr_t start = (base_gfn + __ffs(mask)) << PAGE_SHIFT;
1183 phys_addr_t end = (base_gfn + __fls(mask) + 1) << PAGE_SHIFT;
1185 stage2_wp_range(kvm, start, end);
1189 * kvm_arch_mmu_enable_log_dirty_pt_masked - enable dirty logging for selected
1192 * It calls kvm_mmu_write_protect_pt_masked to write protect selected pages to
1193 * enable dirty logging for them.
1195 void kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm *kvm,
1196 struct kvm_memory_slot *slot,
1197 gfn_t gfn_offset, unsigned long mask)
1199 kvm_mmu_write_protect_pt_masked(kvm, slot, gfn_offset, mask);
1202 static void coherent_cache_guest_page(struct kvm_vcpu *vcpu, pfn_t pfn,
1203 unsigned long size, bool uncached)
1205 __coherent_cache_guest_page(vcpu, pfn, size, uncached);
1208 static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
1209 struct kvm_memory_slot *memslot, unsigned long hva,
1210 unsigned long fault_status)
1213 bool write_fault, writable, hugetlb = false, force_pte = false;
1214 unsigned long mmu_seq;
1215 gfn_t gfn = fault_ipa >> PAGE_SHIFT;
1216 struct kvm *kvm = vcpu->kvm;
1217 struct kvm_mmu_memory_cache *memcache = &vcpu->arch.mmu_page_cache;
1218 struct vm_area_struct *vma;
1220 pgprot_t mem_type = PAGE_S2;
1221 bool fault_ipa_uncached;
1222 bool logging_active = memslot_is_logging(memslot);
1223 unsigned long flags = 0;
1225 write_fault = kvm_is_write_fault(vcpu);
1226 if (fault_status == FSC_PERM && !write_fault) {
1227 kvm_err("Unexpected L2 read permission error\n");
1231 /* Let's check if we will get back a huge page backed by hugetlbfs */
1232 down_read(¤t->mm->mmap_sem);
1233 vma = find_vma_intersection(current->mm, hva, hva + 1);
1234 if (unlikely(!vma)) {
1235 kvm_err("Failed to find VMA for hva 0x%lx\n", hva);
1236 up_read(¤t->mm->mmap_sem);
1240 if (is_vm_hugetlb_page(vma) && !logging_active) {
1242 gfn = (fault_ipa & PMD_MASK) >> PAGE_SHIFT;
1245 * Pages belonging to memslots that don't have the same
1246 * alignment for userspace and IPA cannot be mapped using
1247 * block descriptors even if the pages belong to a THP for
1248 * the process, because the stage-2 block descriptor will
1249 * cover more than a single THP and we loose atomicity for
1250 * unmapping, updates, and splits of the THP or other pages
1251 * in the stage-2 block range.
1253 if ((memslot->userspace_addr & ~PMD_MASK) !=
1254 ((memslot->base_gfn << PAGE_SHIFT) & ~PMD_MASK))
1257 up_read(¤t->mm->mmap_sem);
1259 /* We need minimum second+third level pages */
1260 ret = mmu_topup_memory_cache(memcache, KVM_MMU_CACHE_MIN_PAGES,
1265 mmu_seq = vcpu->kvm->mmu_notifier_seq;
1267 * Ensure the read of mmu_notifier_seq happens before we call
1268 * gfn_to_pfn_prot (which calls get_user_pages), so that we don't risk
1269 * the page we just got a reference to gets unmapped before we have a
1270 * chance to grab the mmu_lock, which ensure that if the page gets
1271 * unmapped afterwards, the call to kvm_unmap_hva will take it away
1272 * from us again properly. This smp_rmb() interacts with the smp_wmb()
1273 * in kvm_mmu_notifier_invalidate_<page|range_end>.
1277 pfn = gfn_to_pfn_prot(kvm, gfn, write_fault, &writable);
1278 if (is_error_pfn(pfn))
1281 if (kvm_is_device_pfn(pfn)) {
1282 mem_type = PAGE_S2_DEVICE;
1283 flags |= KVM_S2PTE_FLAG_IS_IOMAP;
1284 } else if (logging_active) {
1286 * Faults on pages in a memslot with logging enabled
1287 * should not be mapped with huge pages (it introduces churn
1288 * and performance degradation), so force a pte mapping.
1291 flags |= KVM_S2_FLAG_LOGGING_ACTIVE;
1294 * Only actually map the page as writable if this was a write
1301 spin_lock(&kvm->mmu_lock);
1302 if (mmu_notifier_retry(kvm, mmu_seq))
1305 if (!hugetlb && !force_pte)
1306 hugetlb = transparent_hugepage_adjust(&pfn, &fault_ipa);
1308 fault_ipa_uncached = memslot->flags & KVM_MEMSLOT_INCOHERENT;
1311 pmd_t new_pmd = pfn_pmd(pfn, mem_type);
1312 new_pmd = pmd_mkhuge(new_pmd);
1314 kvm_set_s2pmd_writable(&new_pmd);
1315 kvm_set_pfn_dirty(pfn);
1317 coherent_cache_guest_page(vcpu, pfn, PMD_SIZE, fault_ipa_uncached);
1318 ret = stage2_set_pmd_huge(kvm, memcache, fault_ipa, &new_pmd);
1320 pte_t new_pte = pfn_pte(pfn, mem_type);
1323 kvm_set_s2pte_writable(&new_pte);
1324 kvm_set_pfn_dirty(pfn);
1325 mark_page_dirty(kvm, gfn);
1327 coherent_cache_guest_page(vcpu, pfn, PAGE_SIZE, fault_ipa_uncached);
1328 ret = stage2_set_pte(kvm, memcache, fault_ipa, &new_pte, flags);
1332 spin_unlock(&kvm->mmu_lock);
1333 kvm_set_pfn_accessed(pfn);
1334 kvm_release_pfn_clean(pfn);
1339 * Resolve the access fault by making the page young again.
1340 * Note that because the faulting entry is guaranteed not to be
1341 * cached in the TLB, we don't need to invalidate anything.
1343 static void handle_access_fault(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa)
1348 bool pfn_valid = false;
1350 trace_kvm_access_fault(fault_ipa);
1352 spin_lock(&vcpu->kvm->mmu_lock);
1354 pmd = stage2_get_pmd(vcpu->kvm, NULL, fault_ipa);
1355 if (!pmd || pmd_none(*pmd)) /* Nothing there */
1358 if (kvm_pmd_huge(*pmd)) { /* THP, HugeTLB */
1359 *pmd = pmd_mkyoung(*pmd);
1360 pfn = pmd_pfn(*pmd);
1365 pte = pte_offset_kernel(pmd, fault_ipa);
1366 if (pte_none(*pte)) /* Nothing there either */
1369 *pte = pte_mkyoung(*pte); /* Just a page... */
1370 pfn = pte_pfn(*pte);
1373 spin_unlock(&vcpu->kvm->mmu_lock);
1375 kvm_set_pfn_accessed(pfn);
1379 * kvm_handle_guest_abort - handles all 2nd stage aborts
1380 * @vcpu: the VCPU pointer
1381 * @run: the kvm_run structure
1383 * Any abort that gets to the host is almost guaranteed to be caused by a
1384 * missing second stage translation table entry, which can mean that either the
1385 * guest simply needs more memory and we must allocate an appropriate page or it
1386 * can mean that the guest tried to access I/O memory, which is emulated by user
1387 * space. The distinction is based on the IPA causing the fault and whether this
1388 * memory region has been registered as standard RAM by user space.
1390 int kvm_handle_guest_abort(struct kvm_vcpu *vcpu, struct kvm_run *run)
1392 unsigned long fault_status;
1393 phys_addr_t fault_ipa;
1394 struct kvm_memory_slot *memslot;
1396 bool is_iabt, write_fault, writable;
1400 is_iabt = kvm_vcpu_trap_is_iabt(vcpu);
1401 fault_ipa = kvm_vcpu_get_fault_ipa(vcpu);
1403 trace_kvm_guest_fault(*vcpu_pc(vcpu), kvm_vcpu_get_hsr(vcpu),
1404 kvm_vcpu_get_hfar(vcpu), fault_ipa);
1406 /* Check the stage-2 fault is trans. fault or write fault */
1407 fault_status = kvm_vcpu_trap_get_fault_type(vcpu);
1408 if (fault_status != FSC_FAULT && fault_status != FSC_PERM &&
1409 fault_status != FSC_ACCESS) {
1410 kvm_err("Unsupported FSC: EC=%#x xFSC=%#lx ESR_EL2=%#lx\n",
1411 kvm_vcpu_trap_get_class(vcpu),
1412 (unsigned long)kvm_vcpu_trap_get_fault(vcpu),
1413 (unsigned long)kvm_vcpu_get_hsr(vcpu));
1417 idx = srcu_read_lock(&vcpu->kvm->srcu);
1419 gfn = fault_ipa >> PAGE_SHIFT;
1420 memslot = gfn_to_memslot(vcpu->kvm, gfn);
1421 hva = gfn_to_hva_memslot_prot(memslot, gfn, &writable);
1422 write_fault = kvm_is_write_fault(vcpu);
1423 if (kvm_is_error_hva(hva) || (write_fault && !writable)) {
1425 /* Prefetch Abort on I/O address */
1426 kvm_inject_pabt(vcpu, kvm_vcpu_get_hfar(vcpu));
1432 * The IPA is reported as [MAX:12], so we need to
1433 * complement it with the bottom 12 bits from the
1434 * faulting VA. This is always 12 bits, irrespective
1437 fault_ipa |= kvm_vcpu_get_hfar(vcpu) & ((1 << 12) - 1);
1438 ret = io_mem_abort(vcpu, run, fault_ipa);
1442 /* Userspace should not be able to register out-of-bounds IPAs */
1443 VM_BUG_ON(fault_ipa >= KVM_PHYS_SIZE);
1445 if (fault_status == FSC_ACCESS) {
1446 handle_access_fault(vcpu, fault_ipa);
1451 ret = user_mem_abort(vcpu, fault_ipa, memslot, hva, fault_status);
1455 srcu_read_unlock(&vcpu->kvm->srcu, idx);
1459 static int handle_hva_to_gpa(struct kvm *kvm,
1460 unsigned long start,
1462 int (*handler)(struct kvm *kvm,
1463 gpa_t gpa, void *data),
1466 struct kvm_memslots *slots;
1467 struct kvm_memory_slot *memslot;
1470 slots = kvm_memslots(kvm);
1472 /* we only care about the pages that the guest sees */
1473 kvm_for_each_memslot(memslot, slots) {
1474 unsigned long hva_start, hva_end;
1477 hva_start = max(start, memslot->userspace_addr);
1478 hva_end = min(end, memslot->userspace_addr +
1479 (memslot->npages << PAGE_SHIFT));
1480 if (hva_start >= hva_end)
1484 * {gfn(page) | page intersects with [hva_start, hva_end)} =
1485 * {gfn_start, gfn_start+1, ..., gfn_end-1}.
1487 gfn = hva_to_gfn_memslot(hva_start, memslot);
1488 gfn_end = hva_to_gfn_memslot(hva_end + PAGE_SIZE - 1, memslot);
1490 for (; gfn < gfn_end; ++gfn) {
1491 gpa_t gpa = gfn << PAGE_SHIFT;
1492 ret |= handler(kvm, gpa, data);
1499 static int kvm_unmap_hva_handler(struct kvm *kvm, gpa_t gpa, void *data)
1501 unmap_stage2_range(kvm, gpa, PAGE_SIZE);
1505 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
1507 unsigned long end = hva + PAGE_SIZE;
1512 trace_kvm_unmap_hva(hva);
1513 handle_hva_to_gpa(kvm, hva, end, &kvm_unmap_hva_handler, NULL);
1517 int kvm_unmap_hva_range(struct kvm *kvm,
1518 unsigned long start, unsigned long end)
1523 trace_kvm_unmap_hva_range(start, end);
1524 handle_hva_to_gpa(kvm, start, end, &kvm_unmap_hva_handler, NULL);
1528 static int kvm_set_spte_handler(struct kvm *kvm, gpa_t gpa, void *data)
1530 pte_t *pte = (pte_t *)data;
1533 * We can always call stage2_set_pte with KVM_S2PTE_FLAG_LOGGING_ACTIVE
1534 * flag clear because MMU notifiers will have unmapped a huge PMD before
1535 * calling ->change_pte() (which in turn calls kvm_set_spte_hva()) and
1536 * therefore stage2_set_pte() never needs to clear out a huge PMD
1537 * through this calling path.
1539 stage2_set_pte(kvm, NULL, gpa, pte, 0);
1544 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
1546 unsigned long end = hva + PAGE_SIZE;
1552 trace_kvm_set_spte_hva(hva);
1553 stage2_pte = pfn_pte(pte_pfn(pte), PAGE_S2);
1554 handle_hva_to_gpa(kvm, hva, end, &kvm_set_spte_handler, &stage2_pte);
1557 static int kvm_age_hva_handler(struct kvm *kvm, gpa_t gpa, void *data)
1562 pmd = stage2_get_pmd(kvm, NULL, gpa);
1563 if (!pmd || pmd_none(*pmd)) /* Nothing there */
1566 if (kvm_pmd_huge(*pmd)) { /* THP, HugeTLB */
1567 if (pmd_young(*pmd)) {
1568 *pmd = pmd_mkold(*pmd);
1575 pte = pte_offset_kernel(pmd, gpa);
1579 if (pte_young(*pte)) {
1580 *pte = pte_mkold(*pte); /* Just a page... */
1587 static int kvm_test_age_hva_handler(struct kvm *kvm, gpa_t gpa, void *data)
1592 pmd = stage2_get_pmd(kvm, NULL, gpa);
1593 if (!pmd || pmd_none(*pmd)) /* Nothing there */
1596 if (kvm_pmd_huge(*pmd)) /* THP, HugeTLB */
1597 return pmd_young(*pmd);
1599 pte = pte_offset_kernel(pmd, gpa);
1600 if (!pte_none(*pte)) /* Just a page... */
1601 return pte_young(*pte);
1606 int kvm_age_hva(struct kvm *kvm, unsigned long start, unsigned long end)
1608 trace_kvm_age_hva(start, end);
1609 return handle_hva_to_gpa(kvm, start, end, kvm_age_hva_handler, NULL);
1612 int kvm_test_age_hva(struct kvm *kvm, unsigned long hva)
1614 trace_kvm_test_age_hva(hva);
1615 return handle_hva_to_gpa(kvm, hva, hva, kvm_test_age_hva_handler, NULL);
1618 void kvm_mmu_free_memory_caches(struct kvm_vcpu *vcpu)
1620 mmu_free_memory_cache(&vcpu->arch.mmu_page_cache);
1623 phys_addr_t kvm_mmu_get_httbr(void)
1625 return virt_to_phys(hyp_pgd);
1628 phys_addr_t kvm_mmu_get_boot_httbr(void)
1630 return virt_to_phys(boot_hyp_pgd);
1633 phys_addr_t kvm_get_idmap_vector(void)
1635 return hyp_idmap_vector;
1638 int kvm_mmu_init(void)
1642 hyp_idmap_start = kvm_virt_to_phys(__hyp_idmap_text_start);
1643 hyp_idmap_end = kvm_virt_to_phys(__hyp_idmap_text_end);
1644 hyp_idmap_vector = kvm_virt_to_phys(__kvm_hyp_init);
1646 if ((hyp_idmap_start ^ hyp_idmap_end) & PAGE_MASK) {
1648 * Our init code is crossing a page boundary. Allocate
1649 * a bounce page, copy the code over and use that.
1651 size_t len = __hyp_idmap_text_end - __hyp_idmap_text_start;
1652 phys_addr_t phys_base;
1654 init_bounce_page = (void *)__get_free_page(GFP_KERNEL);
1655 if (!init_bounce_page) {
1656 kvm_err("Couldn't allocate HYP init bounce page\n");
1661 memcpy(init_bounce_page, __hyp_idmap_text_start, len);
1663 * Warning: the code we just copied to the bounce page
1664 * must be flushed to the point of coherency.
1665 * Otherwise, the data may be sitting in L2, and HYP
1666 * mode won't be able to observe it as it runs with
1667 * caches off at that point.
1669 kvm_flush_dcache_to_poc(init_bounce_page, len);
1671 phys_base = kvm_virt_to_phys(init_bounce_page);
1672 hyp_idmap_vector += phys_base - hyp_idmap_start;
1673 hyp_idmap_start = phys_base;
1674 hyp_idmap_end = phys_base + len;
1676 kvm_info("Using HYP init bounce page @%lx\n",
1677 (unsigned long)phys_base);
1680 hyp_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, hyp_pgd_order);
1681 boot_hyp_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, hyp_pgd_order);
1683 if (!hyp_pgd || !boot_hyp_pgd) {
1684 kvm_err("Hyp mode PGD not allocated\n");
1689 /* Create the idmap in the boot page tables */
1690 err = __create_hyp_mappings(boot_hyp_pgd,
1691 hyp_idmap_start, hyp_idmap_end,
1692 __phys_to_pfn(hyp_idmap_start),
1696 kvm_err("Failed to idmap %lx-%lx\n",
1697 hyp_idmap_start, hyp_idmap_end);
1701 /* Map the very same page at the trampoline VA */
1702 err = __create_hyp_mappings(boot_hyp_pgd,
1703 TRAMPOLINE_VA, TRAMPOLINE_VA + PAGE_SIZE,
1704 __phys_to_pfn(hyp_idmap_start),
1707 kvm_err("Failed to map trampoline @%lx into boot HYP pgd\n",
1712 /* Map the same page again into the runtime page tables */
1713 err = __create_hyp_mappings(hyp_pgd,
1714 TRAMPOLINE_VA, TRAMPOLINE_VA + PAGE_SIZE,
1715 __phys_to_pfn(hyp_idmap_start),
1718 kvm_err("Failed to map trampoline @%lx into runtime HYP pgd\n",
1729 void kvm_arch_commit_memory_region(struct kvm *kvm,
1730 struct kvm_userspace_memory_region *mem,
1731 const struct kvm_memory_slot *old,
1732 enum kvm_mr_change change)
1735 * At this point memslot has been committed and there is an
1736 * allocated dirty_bitmap[], dirty pages will be be tracked while the
1737 * memory slot is write protected.
1739 if (change != KVM_MR_DELETE && mem->flags & KVM_MEM_LOG_DIRTY_PAGES)
1740 kvm_mmu_wp_memory_region(kvm, mem->slot);
1743 int kvm_arch_prepare_memory_region(struct kvm *kvm,
1744 struct kvm_memory_slot *memslot,
1745 struct kvm_userspace_memory_region *mem,
1746 enum kvm_mr_change change)
1748 hva_t hva = mem->userspace_addr;
1749 hva_t reg_end = hva + mem->memory_size;
1750 bool writable = !(mem->flags & KVM_MEM_READONLY);
1753 if (change != KVM_MR_CREATE && change != KVM_MR_MOVE &&
1754 change != KVM_MR_FLAGS_ONLY)
1758 * Prevent userspace from creating a memory region outside of the IPA
1759 * space addressable by the KVM guest IPA space.
1761 if (memslot->base_gfn + memslot->npages >=
1762 (KVM_PHYS_SIZE >> PAGE_SHIFT))
1766 * A memory region could potentially cover multiple VMAs, and any holes
1767 * between them, so iterate over all of them to find out if we can map
1768 * any of them right now.
1770 * +--------------------------------------------+
1771 * +---------------+----------------+ +----------------+
1772 * | : VMA 1 | VMA 2 | | VMA 3 : |
1773 * +---------------+----------------+ +----------------+
1775 * +--------------------------------------------+
1778 struct vm_area_struct *vma = find_vma(current->mm, hva);
1779 hva_t vm_start, vm_end;
1781 if (!vma || vma->vm_start >= reg_end)
1785 * Mapping a read-only VMA is only allowed if the
1786 * memory region is configured as read-only.
1788 if (writable && !(vma->vm_flags & VM_WRITE)) {
1794 * Take the intersection of this VMA with the memory region
1796 vm_start = max(hva, vma->vm_start);
1797 vm_end = min(reg_end, vma->vm_end);
1799 if (vma->vm_flags & VM_PFNMAP) {
1800 gpa_t gpa = mem->guest_phys_addr +
1801 (vm_start - mem->userspace_addr);
1802 phys_addr_t pa = (vma->vm_pgoff << PAGE_SHIFT) +
1803 vm_start - vma->vm_start;
1805 /* IO region dirty page logging not allowed */
1806 if (memslot->flags & KVM_MEM_LOG_DIRTY_PAGES)
1809 ret = kvm_phys_addr_ioremap(kvm, gpa, pa,
1816 } while (hva < reg_end);
1818 if (change == KVM_MR_FLAGS_ONLY)
1821 spin_lock(&kvm->mmu_lock);
1823 unmap_stage2_range(kvm, mem->guest_phys_addr, mem->memory_size);
1825 stage2_flush_memslot(kvm, memslot);
1826 spin_unlock(&kvm->mmu_lock);
1830 void kvm_arch_free_memslot(struct kvm *kvm, struct kvm_memory_slot *free,
1831 struct kvm_memory_slot *dont)
1835 int kvm_arch_create_memslot(struct kvm *kvm, struct kvm_memory_slot *slot,
1836 unsigned long npages)
1839 * Readonly memslots are not incoherent with the caches by definition,
1840 * but in practice, they are used mostly to emulate ROMs or NOR flashes
1841 * that the guest may consider devices and hence map as uncached.
1842 * To prevent incoherency issues in these cases, tag all readonly
1843 * regions as incoherent.
1845 if (slot->flags & KVM_MEM_READONLY)
1846 slot->flags |= KVM_MEMSLOT_INCOHERENT;
1850 void kvm_arch_memslots_updated(struct kvm *kvm)
1854 void kvm_arch_flush_shadow_all(struct kvm *kvm)
1858 void kvm_arch_flush_shadow_memslot(struct kvm *kvm,
1859 struct kvm_memory_slot *slot)
1861 gpa_t gpa = slot->base_gfn << PAGE_SHIFT;
1862 phys_addr_t size = slot->npages << PAGE_SHIFT;
1864 spin_lock(&kvm->mmu_lock);
1865 unmap_stage2_range(kvm, gpa, size);
1866 spin_unlock(&kvm->mmu_lock);
1870 * See note at ARMv7 ARM B1.14.4 (TL;DR: S/W ops are not easily virtualized).
1873 * - S/W ops are local to a CPU (not broadcast)
1874 * - We have line migration behind our back (speculation)
1875 * - System caches don't support S/W at all (damn!)
1877 * In the face of the above, the best we can do is to try and convert
1878 * S/W ops to VA ops. Because the guest is not allowed to infer the
1879 * S/W to PA mapping, it can only use S/W to nuke the whole cache,
1880 * which is a rather good thing for us.
1882 * Also, it is only used when turning caches on/off ("The expected
1883 * usage of the cache maintenance instructions that operate by set/way
1884 * is associated with the cache maintenance instructions associated
1885 * with the powerdown and powerup of caches, if this is required by
1886 * the implementation.").
1888 * We use the following policy:
1890 * - If we trap a S/W operation, we enable VM trapping to detect
1891 * caches being turned on/off, and do a full clean.
1893 * - We flush the caches on both caches being turned on and off.
1895 * - Once the caches are enabled, we stop trapping VM ops.
1897 void kvm_set_way_flush(struct kvm_vcpu *vcpu)
1899 unsigned long hcr = vcpu_get_hcr(vcpu);
1902 * If this is the first time we do a S/W operation
1903 * (i.e. HCR_TVM not set) flush the whole memory, and set the
1906 * Otherwise, rely on the VM trapping to wait for the MMU +
1907 * Caches to be turned off. At that point, we'll be able to
1908 * clean the caches again.
1910 if (!(hcr & HCR_TVM)) {
1911 trace_kvm_set_way_flush(*vcpu_pc(vcpu),
1912 vcpu_has_cache_enabled(vcpu));
1913 stage2_flush_vm(vcpu->kvm);
1914 vcpu_set_hcr(vcpu, hcr | HCR_TVM);
1918 void kvm_toggle_cache(struct kvm_vcpu *vcpu, bool was_enabled)
1920 bool now_enabled = vcpu_has_cache_enabled(vcpu);
1923 * If switching the MMU+caches on, need to invalidate the caches.
1924 * If switching it off, need to clean the caches.
1925 * Clean + invalidate does the trick always.
1927 if (now_enabled != was_enabled)
1928 stage2_flush_vm(vcpu->kvm);
1930 /* Caches are now on, stop trapping VM ops (until a S/W op) */
1932 vcpu_set_hcr(vcpu, vcpu_get_hcr(vcpu) & ~HCR_TVM);
1934 trace_kvm_toggle_cache(*vcpu_pc(vcpu), was_enabled, now_enabled);