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))
49 static void kvm_tlb_flush_vmid_ipa(struct kvm
*kvm
, phys_addr_t ipa
)
52 * This function also gets called when dealing with HYP page
53 * tables. As HYP doesn't have an associated struct kvm (and
54 * the HYP page tables are fairly static), we don't do
58 kvm_call_hyp(__kvm_tlb_flush_vmid_ipa
, kvm
, ipa
);
61 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache
*cache
,
66 BUG_ON(max
> KVM_NR_MEM_OBJS
);
67 if (cache
->nobjs
>= min
)
69 while (cache
->nobjs
< max
) {
70 page
= (void *)__get_free_page(PGALLOC_GFP
);
73 cache
->objects
[cache
->nobjs
++] = page
;
78 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache
*mc
)
81 free_page((unsigned long)mc
->objects
[--mc
->nobjs
]);
84 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache
*mc
)
88 BUG_ON(!mc
|| !mc
->nobjs
);
89 p
= mc
->objects
[--mc
->nobjs
];
93 static void clear_pgd_entry(struct kvm
*kvm
, pgd_t
*pgd
, phys_addr_t addr
)
95 pud_t
*pud_table __maybe_unused
= pud_offset(pgd
, 0);
97 kvm_tlb_flush_vmid_ipa(kvm
, addr
);
98 pud_free(NULL
, pud_table
);
99 put_page(virt_to_page(pgd
));
102 static void clear_pud_entry(struct kvm
*kvm
, pud_t
*pud
, phys_addr_t addr
)
104 pmd_t
*pmd_table
= pmd_offset(pud
, 0);
105 VM_BUG_ON(pud_huge(*pud
));
107 kvm_tlb_flush_vmid_ipa(kvm
, addr
);
108 pmd_free(NULL
, pmd_table
);
109 put_page(virt_to_page(pud
));
112 static void clear_pmd_entry(struct kvm
*kvm
, pmd_t
*pmd
, phys_addr_t addr
)
114 pte_t
*pte_table
= pte_offset_kernel(pmd
, 0);
115 VM_BUG_ON(kvm_pmd_huge(*pmd
));
117 kvm_tlb_flush_vmid_ipa(kvm
, addr
);
118 pte_free_kernel(NULL
, pte_table
);
119 put_page(virt_to_page(pmd
));
122 static void unmap_ptes(struct kvm
*kvm
, pmd_t
*pmd
,
123 phys_addr_t addr
, phys_addr_t end
)
125 phys_addr_t start_addr
= addr
;
126 pte_t
*pte
, *start_pte
;
128 start_pte
= pte
= pte_offset_kernel(pmd
, addr
);
130 if (!pte_none(*pte
)) {
131 kvm_set_pte(pte
, __pte(0));
132 put_page(virt_to_page(pte
));
133 kvm_tlb_flush_vmid_ipa(kvm
, addr
);
135 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
137 if (kvm_pte_table_empty(kvm
, start_pte
))
138 clear_pmd_entry(kvm
, pmd
, start_addr
);
141 static void unmap_pmds(struct kvm
*kvm
, pud_t
*pud
,
142 phys_addr_t addr
, phys_addr_t end
)
144 phys_addr_t next
, start_addr
= addr
;
145 pmd_t
*pmd
, *start_pmd
;
147 start_pmd
= pmd
= pmd_offset(pud
, addr
);
149 next
= kvm_pmd_addr_end(addr
, end
);
150 if (!pmd_none(*pmd
)) {
151 if (kvm_pmd_huge(*pmd
)) {
153 kvm_tlb_flush_vmid_ipa(kvm
, addr
);
154 put_page(virt_to_page(pmd
));
156 unmap_ptes(kvm
, pmd
, addr
, next
);
159 } while (pmd
++, addr
= next
, addr
!= end
);
161 if (kvm_pmd_table_empty(kvm
, start_pmd
))
162 clear_pud_entry(kvm
, pud
, start_addr
);
165 static void unmap_puds(struct kvm
*kvm
, pgd_t
*pgd
,
166 phys_addr_t addr
, phys_addr_t end
)
168 phys_addr_t next
, start_addr
= addr
;
169 pud_t
*pud
, *start_pud
;
171 start_pud
= pud
= pud_offset(pgd
, addr
);
173 next
= kvm_pud_addr_end(addr
, end
);
174 if (!pud_none(*pud
)) {
175 if (pud_huge(*pud
)) {
177 kvm_tlb_flush_vmid_ipa(kvm
, addr
);
178 put_page(virt_to_page(pud
));
180 unmap_pmds(kvm
, pud
, addr
, next
);
183 } while (pud
++, addr
= next
, addr
!= end
);
185 if (kvm_pud_table_empty(kvm
, start_pud
))
186 clear_pgd_entry(kvm
, pgd
, start_addr
);
190 static void unmap_range(struct kvm
*kvm
, pgd_t
*pgdp
,
191 phys_addr_t start
, u64 size
)
194 phys_addr_t addr
= start
, end
= start
+ size
;
197 pgd
= pgdp
+ pgd_index(addr
);
199 next
= kvm_pgd_addr_end(addr
, end
);
201 unmap_puds(kvm
, pgd
, addr
, next
);
202 } while (pgd
++, addr
= next
, addr
!= end
);
205 static void stage2_flush_ptes(struct kvm
*kvm
, pmd_t
*pmd
,
206 phys_addr_t addr
, phys_addr_t end
)
210 pte
= pte_offset_kernel(pmd
, addr
);
212 if (!pte_none(*pte
)) {
213 hva_t hva
= gfn_to_hva(kvm
, addr
>> PAGE_SHIFT
);
214 kvm_flush_dcache_to_poc((void*)hva
, PAGE_SIZE
);
216 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
219 static void stage2_flush_pmds(struct kvm
*kvm
, pud_t
*pud
,
220 phys_addr_t addr
, phys_addr_t end
)
225 pmd
= pmd_offset(pud
, addr
);
227 next
= kvm_pmd_addr_end(addr
, end
);
228 if (!pmd_none(*pmd
)) {
229 if (kvm_pmd_huge(*pmd
)) {
230 hva_t hva
= gfn_to_hva(kvm
, addr
>> PAGE_SHIFT
);
231 kvm_flush_dcache_to_poc((void*)hva
, PMD_SIZE
);
233 stage2_flush_ptes(kvm
, pmd
, addr
, next
);
236 } while (pmd
++, addr
= next
, addr
!= end
);
239 static void stage2_flush_puds(struct kvm
*kvm
, pgd_t
*pgd
,
240 phys_addr_t addr
, phys_addr_t end
)
245 pud
= pud_offset(pgd
, addr
);
247 next
= kvm_pud_addr_end(addr
, end
);
248 if (!pud_none(*pud
)) {
249 if (pud_huge(*pud
)) {
250 hva_t hva
= gfn_to_hva(kvm
, addr
>> PAGE_SHIFT
);
251 kvm_flush_dcache_to_poc((void*)hva
, PUD_SIZE
);
253 stage2_flush_pmds(kvm
, pud
, addr
, next
);
256 } while (pud
++, addr
= next
, addr
!= end
);
259 static void stage2_flush_memslot(struct kvm
*kvm
,
260 struct kvm_memory_slot
*memslot
)
262 phys_addr_t addr
= memslot
->base_gfn
<< PAGE_SHIFT
;
263 phys_addr_t end
= addr
+ PAGE_SIZE
* memslot
->npages
;
267 pgd
= kvm
->arch
.pgd
+ pgd_index(addr
);
269 next
= kvm_pgd_addr_end(addr
, end
);
270 stage2_flush_puds(kvm
, pgd
, addr
, next
);
271 } while (pgd
++, addr
= next
, addr
!= end
);
275 * stage2_flush_vm - Invalidate cache for pages mapped in stage 2
276 * @kvm: The struct kvm pointer
278 * Go through the stage 2 page tables and invalidate any cache lines
279 * backing memory already mapped to the VM.
281 void stage2_flush_vm(struct kvm
*kvm
)
283 struct kvm_memslots
*slots
;
284 struct kvm_memory_slot
*memslot
;
287 idx
= srcu_read_lock(&kvm
->srcu
);
288 spin_lock(&kvm
->mmu_lock
);
290 slots
= kvm_memslots(kvm
);
291 kvm_for_each_memslot(memslot
, slots
)
292 stage2_flush_memslot(kvm
, memslot
);
294 spin_unlock(&kvm
->mmu_lock
);
295 srcu_read_unlock(&kvm
->srcu
, idx
);
299 * free_boot_hyp_pgd - free HYP boot page tables
301 * Free the HYP boot page tables. The bounce page is also freed.
303 void free_boot_hyp_pgd(void)
305 mutex_lock(&kvm_hyp_pgd_mutex
);
308 unmap_range(NULL
, boot_hyp_pgd
, hyp_idmap_start
, PAGE_SIZE
);
309 unmap_range(NULL
, boot_hyp_pgd
, TRAMPOLINE_VA
, PAGE_SIZE
);
310 free_pages((unsigned long)boot_hyp_pgd
, hyp_pgd_order
);
315 unmap_range(NULL
, hyp_pgd
, TRAMPOLINE_VA
, PAGE_SIZE
);
317 free_page((unsigned long)init_bounce_page
);
318 init_bounce_page
= NULL
;
320 mutex_unlock(&kvm_hyp_pgd_mutex
);
324 * free_hyp_pgds - free Hyp-mode page tables
326 * Assumes hyp_pgd is a page table used strictly in Hyp-mode and
327 * therefore contains either mappings in the kernel memory area (above
328 * PAGE_OFFSET), or device mappings in the vmalloc range (from
329 * VMALLOC_START to VMALLOC_END).
331 * boot_hyp_pgd should only map two pages for the init code.
333 void free_hyp_pgds(void)
339 mutex_lock(&kvm_hyp_pgd_mutex
);
342 for (addr
= PAGE_OFFSET
; virt_addr_valid(addr
); addr
+= PGDIR_SIZE
)
343 unmap_range(NULL
, hyp_pgd
, KERN_TO_HYP(addr
), PGDIR_SIZE
);
344 for (addr
= VMALLOC_START
; is_vmalloc_addr((void*)addr
); addr
+= PGDIR_SIZE
)
345 unmap_range(NULL
, hyp_pgd
, KERN_TO_HYP(addr
), PGDIR_SIZE
);
347 free_pages((unsigned long)hyp_pgd
, hyp_pgd_order
);
351 mutex_unlock(&kvm_hyp_pgd_mutex
);
354 static void create_hyp_pte_mappings(pmd_t
*pmd
, unsigned long start
,
355 unsigned long end
, unsigned long pfn
,
363 pte
= pte_offset_kernel(pmd
, addr
);
364 kvm_set_pte(pte
, pfn_pte(pfn
, prot
));
365 get_page(virt_to_page(pte
));
366 kvm_flush_dcache_to_poc(pte
, sizeof(*pte
));
368 } while (addr
+= PAGE_SIZE
, addr
!= end
);
371 static int create_hyp_pmd_mappings(pud_t
*pud
, unsigned long start
,
372 unsigned long end
, unsigned long pfn
,
377 unsigned long addr
, next
;
381 pmd
= pmd_offset(pud
, addr
);
383 BUG_ON(pmd_sect(*pmd
));
385 if (pmd_none(*pmd
)) {
386 pte
= pte_alloc_one_kernel(NULL
, addr
);
388 kvm_err("Cannot allocate Hyp pte\n");
391 pmd_populate_kernel(NULL
, pmd
, pte
);
392 get_page(virt_to_page(pmd
));
393 kvm_flush_dcache_to_poc(pmd
, sizeof(*pmd
));
396 next
= pmd_addr_end(addr
, end
);
398 create_hyp_pte_mappings(pmd
, addr
, next
, pfn
, prot
);
399 pfn
+= (next
- addr
) >> PAGE_SHIFT
;
400 } while (addr
= next
, addr
!= end
);
405 static int create_hyp_pud_mappings(pgd_t
*pgd
, unsigned long start
,
406 unsigned long end
, unsigned long pfn
,
411 unsigned long addr
, next
;
416 pud
= pud_offset(pgd
, addr
);
418 if (pud_none_or_clear_bad(pud
)) {
419 pmd
= pmd_alloc_one(NULL
, addr
);
421 kvm_err("Cannot allocate Hyp pmd\n");
424 pud_populate(NULL
, pud
, pmd
);
425 get_page(virt_to_page(pud
));
426 kvm_flush_dcache_to_poc(pud
, sizeof(*pud
));
429 next
= pud_addr_end(addr
, end
);
430 ret
= create_hyp_pmd_mappings(pud
, addr
, next
, pfn
, prot
);
433 pfn
+= (next
- addr
) >> PAGE_SHIFT
;
434 } while (addr
= next
, addr
!= end
);
439 static int __create_hyp_mappings(pgd_t
*pgdp
,
440 unsigned long start
, unsigned long end
,
441 unsigned long pfn
, pgprot_t prot
)
445 unsigned long addr
, next
;
448 mutex_lock(&kvm_hyp_pgd_mutex
);
449 addr
= start
& PAGE_MASK
;
450 end
= PAGE_ALIGN(end
);
452 pgd
= pgdp
+ pgd_index(addr
);
454 if (pgd_none(*pgd
)) {
455 pud
= pud_alloc_one(NULL
, addr
);
457 kvm_err("Cannot allocate Hyp pud\n");
461 pgd_populate(NULL
, pgd
, pud
);
462 get_page(virt_to_page(pgd
));
463 kvm_flush_dcache_to_poc(pgd
, sizeof(*pgd
));
466 next
= pgd_addr_end(addr
, end
);
467 err
= create_hyp_pud_mappings(pgd
, addr
, next
, pfn
, prot
);
470 pfn
+= (next
- addr
) >> PAGE_SHIFT
;
471 } while (addr
= next
, addr
!= end
);
473 mutex_unlock(&kvm_hyp_pgd_mutex
);
477 static phys_addr_t
kvm_kaddr_to_phys(void *kaddr
)
479 if (!is_vmalloc_addr(kaddr
)) {
480 BUG_ON(!virt_addr_valid(kaddr
));
483 return page_to_phys(vmalloc_to_page(kaddr
)) +
484 offset_in_page(kaddr
);
489 * create_hyp_mappings - duplicate a kernel virtual address range in Hyp mode
490 * @from: The virtual kernel start address of the range
491 * @to: The virtual kernel end address of the range (exclusive)
493 * The same virtual address as the kernel virtual address is also used
494 * in Hyp-mode mapping (modulo HYP_PAGE_OFFSET) to the same underlying
497 int create_hyp_mappings(void *from
, void *to
)
499 phys_addr_t phys_addr
;
500 unsigned long virt_addr
;
501 unsigned long start
= KERN_TO_HYP((unsigned long)from
);
502 unsigned long end
= KERN_TO_HYP((unsigned long)to
);
504 start
= start
& PAGE_MASK
;
505 end
= PAGE_ALIGN(end
);
507 for (virt_addr
= start
; virt_addr
< end
; virt_addr
+= PAGE_SIZE
) {
510 phys_addr
= kvm_kaddr_to_phys(from
+ virt_addr
- start
);
511 err
= __create_hyp_mappings(hyp_pgd
, virt_addr
,
512 virt_addr
+ PAGE_SIZE
,
513 __phys_to_pfn(phys_addr
),
523 * create_hyp_io_mappings - duplicate a kernel IO mapping into Hyp mode
524 * @from: The kernel start VA of the range
525 * @to: The kernel end VA of the range (exclusive)
526 * @phys_addr: The physical start address which gets mapped
528 * The resulting HYP VA is the same as the kernel VA, modulo
531 int create_hyp_io_mappings(void *from
, void *to
, phys_addr_t phys_addr
)
533 unsigned long start
= KERN_TO_HYP((unsigned long)from
);
534 unsigned long end
= KERN_TO_HYP((unsigned long)to
);
536 /* Check for a valid kernel IO mapping */
537 if (!is_vmalloc_addr(from
) || !is_vmalloc_addr(to
- 1))
540 return __create_hyp_mappings(hyp_pgd
, start
, end
,
541 __phys_to_pfn(phys_addr
), PAGE_HYP_DEVICE
);
545 * kvm_alloc_stage2_pgd - allocate level-1 table for stage-2 translation.
546 * @kvm: The KVM struct pointer for the VM.
548 * Allocates the 1st level table only of size defined by S2_PGD_ORDER (can
549 * support either full 40-bit input addresses or limited to 32-bit input
550 * addresses). Clears the allocated pages.
552 * Note we don't need locking here as this is only called when the VM is
553 * created, which can only be done once.
555 int kvm_alloc_stage2_pgd(struct kvm
*kvm
)
560 if (kvm
->arch
.pgd
!= NULL
) {
561 kvm_err("kvm_arch already initialized?\n");
565 if (KVM_PREALLOC_LEVEL
> 0) {
567 * Allocate fake pgd for the page table manipulation macros to
568 * work. This is not used by the hardware and we have no
569 * alignment requirement for this allocation.
571 pgd
= (pgd_t
*)kmalloc(PTRS_PER_S2_PGD
* sizeof(pgd_t
),
572 GFP_KERNEL
| __GFP_ZERO
);
575 * Allocate actual first-level Stage-2 page table used by the
576 * hardware for Stage-2 page table walks.
578 pgd
= (pgd_t
*)__get_free_pages(GFP_KERNEL
| __GFP_ZERO
, S2_PGD_ORDER
);
584 ret
= kvm_prealloc_hwpgd(kvm
, pgd
);
592 if (KVM_PREALLOC_LEVEL
> 0)
595 free_pages((unsigned long)pgd
, S2_PGD_ORDER
);
600 * unmap_stage2_range -- Clear stage2 page table entries to unmap a range
601 * @kvm: The VM pointer
602 * @start: The intermediate physical base address of the range to unmap
603 * @size: The size of the area to unmap
605 * Clear a range of stage-2 mappings, lowering the various ref-counts. Must
606 * be called while holding mmu_lock (unless for freeing the stage2 pgd before
607 * destroying the VM), otherwise another faulting VCPU may come in and mess
608 * with things behind our backs.
610 static void unmap_stage2_range(struct kvm
*kvm
, phys_addr_t start
, u64 size
)
612 unmap_range(kvm
, kvm
->arch
.pgd
, start
, size
);
615 static void stage2_unmap_memslot(struct kvm
*kvm
,
616 struct kvm_memory_slot
*memslot
)
618 hva_t hva
= memslot
->userspace_addr
;
619 phys_addr_t addr
= memslot
->base_gfn
<< PAGE_SHIFT
;
620 phys_addr_t size
= PAGE_SIZE
* memslot
->npages
;
621 hva_t reg_end
= hva
+ size
;
624 * A memory region could potentially cover multiple VMAs, and any holes
625 * between them, so iterate over all of them to find out if we should
628 * +--------------------------------------------+
629 * +---------------+----------------+ +----------------+
630 * | : VMA 1 | VMA 2 | | VMA 3 : |
631 * +---------------+----------------+ +----------------+
633 * +--------------------------------------------+
636 struct vm_area_struct
*vma
= find_vma(current
->mm
, hva
);
637 hva_t vm_start
, vm_end
;
639 if (!vma
|| vma
->vm_start
>= reg_end
)
643 * Take the intersection of this VMA with the memory region
645 vm_start
= max(hva
, vma
->vm_start
);
646 vm_end
= min(reg_end
, vma
->vm_end
);
648 if (!(vma
->vm_flags
& VM_PFNMAP
)) {
649 gpa_t gpa
= addr
+ (vm_start
- memslot
->userspace_addr
);
650 unmap_stage2_range(kvm
, gpa
, vm_end
- vm_start
);
653 } while (hva
< reg_end
);
657 * stage2_unmap_vm - Unmap Stage-2 RAM mappings
658 * @kvm: The struct kvm pointer
660 * Go through the memregions and unmap any reguler RAM
661 * backing memory already mapped to the VM.
663 void stage2_unmap_vm(struct kvm
*kvm
)
665 struct kvm_memslots
*slots
;
666 struct kvm_memory_slot
*memslot
;
669 idx
= srcu_read_lock(&kvm
->srcu
);
670 spin_lock(&kvm
->mmu_lock
);
672 slots
= kvm_memslots(kvm
);
673 kvm_for_each_memslot(memslot
, slots
)
674 stage2_unmap_memslot(kvm
, memslot
);
676 spin_unlock(&kvm
->mmu_lock
);
677 srcu_read_unlock(&kvm
->srcu
, idx
);
681 * kvm_free_stage2_pgd - free all stage-2 tables
682 * @kvm: The KVM struct pointer for the VM.
684 * Walks the level-1 page table pointed to by kvm->arch.pgd and frees all
685 * underlying level-2 and level-3 tables before freeing the actual level-1 table
686 * and setting the struct pointer to NULL.
688 * Note we don't need locking here as this is only called when the VM is
689 * destroyed, which can only be done once.
691 void kvm_free_stage2_pgd(struct kvm
*kvm
)
693 if (kvm
->arch
.pgd
== NULL
)
696 unmap_stage2_range(kvm
, 0, KVM_PHYS_SIZE
);
698 if (KVM_PREALLOC_LEVEL
> 0)
699 kfree(kvm
->arch
.pgd
);
701 free_pages((unsigned long)kvm
->arch
.pgd
, S2_PGD_ORDER
);
702 kvm
->arch
.pgd
= NULL
;
705 static pud_t
*stage2_get_pud(struct kvm
*kvm
, struct kvm_mmu_memory_cache
*cache
,
711 pgd
= kvm
->arch
.pgd
+ pgd_index(addr
);
712 if (WARN_ON(pgd_none(*pgd
))) {
715 pud
= mmu_memory_cache_alloc(cache
);
716 pgd_populate(NULL
, pgd
, pud
);
717 get_page(virt_to_page(pgd
));
720 return pud_offset(pgd
, addr
);
723 static pmd_t
*stage2_get_pmd(struct kvm
*kvm
, struct kvm_mmu_memory_cache
*cache
,
729 pud
= stage2_get_pud(kvm
, cache
, addr
);
730 if (pud_none(*pud
)) {
733 pmd
= mmu_memory_cache_alloc(cache
);
734 pud_populate(NULL
, pud
, pmd
);
735 get_page(virt_to_page(pud
));
738 return pmd_offset(pud
, addr
);
741 static int stage2_set_pmd_huge(struct kvm
*kvm
, struct kvm_mmu_memory_cache
742 *cache
, phys_addr_t addr
, const pmd_t
*new_pmd
)
746 pmd
= stage2_get_pmd(kvm
, cache
, addr
);
750 * Mapping in huge pages should only happen through a fault. If a
751 * page is merged into a transparent huge page, the individual
752 * subpages of that huge page should be unmapped through MMU
753 * notifiers before we get here.
755 * Merging of CompoundPages is not supported; they should become
756 * splitting first, unmapped, merged, and mapped back in on-demand.
758 VM_BUG_ON(pmd_present(*pmd
) && pmd_pfn(*pmd
) != pmd_pfn(*new_pmd
));
761 kvm_set_pmd(pmd
, *new_pmd
);
762 if (pmd_present(old_pmd
))
763 kvm_tlb_flush_vmid_ipa(kvm
, addr
);
765 get_page(virt_to_page(pmd
));
769 static int stage2_set_pte(struct kvm
*kvm
, struct kvm_mmu_memory_cache
*cache
,
770 phys_addr_t addr
, const pte_t
*new_pte
, bool iomap
)
775 /* Create stage-2 page table mapping - Levels 0 and 1 */
776 pmd
= stage2_get_pmd(kvm
, cache
, addr
);
779 * Ignore calls from kvm_set_spte_hva for unallocated
785 /* Create stage-2 page mappings - Level 2 */
786 if (pmd_none(*pmd
)) {
788 return 0; /* ignore calls from kvm_set_spte_hva */
789 pte
= mmu_memory_cache_alloc(cache
);
791 pmd_populate_kernel(NULL
, pmd
, pte
);
792 get_page(virt_to_page(pmd
));
795 pte
= pte_offset_kernel(pmd
, addr
);
797 if (iomap
&& pte_present(*pte
))
800 /* Create 2nd stage page table mapping - Level 3 */
802 kvm_set_pte(pte
, *new_pte
);
803 if (pte_present(old_pte
))
804 kvm_tlb_flush_vmid_ipa(kvm
, addr
);
806 get_page(virt_to_page(pte
));
812 * kvm_phys_addr_ioremap - map a device range to guest IPA
814 * @kvm: The KVM pointer
815 * @guest_ipa: The IPA at which to insert the mapping
816 * @pa: The physical address of the device
817 * @size: The size of the mapping
819 int kvm_phys_addr_ioremap(struct kvm
*kvm
, phys_addr_t guest_ipa
,
820 phys_addr_t pa
, unsigned long size
, bool writable
)
822 phys_addr_t addr
, end
;
825 struct kvm_mmu_memory_cache cache
= { 0, };
827 end
= (guest_ipa
+ size
+ PAGE_SIZE
- 1) & PAGE_MASK
;
828 pfn
= __phys_to_pfn(pa
);
830 for (addr
= guest_ipa
; addr
< end
; addr
+= PAGE_SIZE
) {
831 pte_t pte
= pfn_pte(pfn
, PAGE_S2_DEVICE
);
834 kvm_set_s2pte_writable(&pte
);
836 ret
= mmu_topup_memory_cache(&cache
, KVM_MMU_CACHE_MIN_PAGES
,
840 spin_lock(&kvm
->mmu_lock
);
841 ret
= stage2_set_pte(kvm
, &cache
, addr
, &pte
, true);
842 spin_unlock(&kvm
->mmu_lock
);
850 mmu_free_memory_cache(&cache
);
854 static bool transparent_hugepage_adjust(pfn_t
*pfnp
, phys_addr_t
*ipap
)
857 gfn_t gfn
= *ipap
>> PAGE_SHIFT
;
859 if (PageTransCompound(pfn_to_page(pfn
))) {
862 * The address we faulted on is backed by a transparent huge
863 * page. However, because we map the compound huge page and
864 * not the individual tail page, we need to transfer the
865 * refcount to the head page. We have to be careful that the
866 * THP doesn't start to split while we are adjusting the
869 * We are sure this doesn't happen, because mmu_notifier_retry
870 * was successful and we are holding the mmu_lock, so if this
871 * THP is trying to split, it will be blocked in the mmu
872 * notifier before touching any of the pages, specifically
873 * before being able to call __split_huge_page_refcount().
875 * We can therefore safely transfer the refcount from PG_tail
876 * to PG_head and switch the pfn from a tail page to the head
879 mask
= PTRS_PER_PMD
- 1;
880 VM_BUG_ON((gfn
& mask
) != (pfn
& mask
));
883 kvm_release_pfn_clean(pfn
);
895 static bool kvm_is_write_fault(struct kvm_vcpu
*vcpu
)
897 if (kvm_vcpu_trap_is_iabt(vcpu
))
900 return kvm_vcpu_dabt_iswrite(vcpu
);
903 static bool kvm_is_device_pfn(unsigned long pfn
)
905 return !pfn_valid(pfn
);
908 static int user_mem_abort(struct kvm_vcpu
*vcpu
, phys_addr_t fault_ipa
,
909 struct kvm_memory_slot
*memslot
, unsigned long hva
,
910 unsigned long fault_status
)
913 bool write_fault
, writable
, hugetlb
= false, force_pte
= false;
914 unsigned long mmu_seq
;
915 gfn_t gfn
= fault_ipa
>> PAGE_SHIFT
;
916 struct kvm
*kvm
= vcpu
->kvm
;
917 struct kvm_mmu_memory_cache
*memcache
= &vcpu
->arch
.mmu_page_cache
;
918 struct vm_area_struct
*vma
;
920 pgprot_t mem_type
= PAGE_S2
;
921 bool fault_ipa_uncached
;
923 write_fault
= kvm_is_write_fault(vcpu
);
924 if (fault_status
== FSC_PERM
&& !write_fault
) {
925 kvm_err("Unexpected L2 read permission error\n");
929 /* Let's check if we will get back a huge page backed by hugetlbfs */
930 down_read(¤t
->mm
->mmap_sem
);
931 vma
= find_vma_intersection(current
->mm
, hva
, hva
+ 1);
932 if (unlikely(!vma
)) {
933 kvm_err("Failed to find VMA for hva 0x%lx\n", hva
);
934 up_read(¤t
->mm
->mmap_sem
);
938 if (is_vm_hugetlb_page(vma
)) {
940 gfn
= (fault_ipa
& PMD_MASK
) >> PAGE_SHIFT
;
943 * Pages belonging to memslots that don't have the same
944 * alignment for userspace and IPA cannot be mapped using
945 * block descriptors even if the pages belong to a THP for
946 * the process, because the stage-2 block descriptor will
947 * cover more than a single THP and we loose atomicity for
948 * unmapping, updates, and splits of the THP or other pages
949 * in the stage-2 block range.
951 if ((memslot
->userspace_addr
& ~PMD_MASK
) !=
952 ((memslot
->base_gfn
<< PAGE_SHIFT
) & ~PMD_MASK
))
955 up_read(¤t
->mm
->mmap_sem
);
957 /* We need minimum second+third level pages */
958 ret
= mmu_topup_memory_cache(memcache
, KVM_MMU_CACHE_MIN_PAGES
,
963 mmu_seq
= vcpu
->kvm
->mmu_notifier_seq
;
965 * Ensure the read of mmu_notifier_seq happens before we call
966 * gfn_to_pfn_prot (which calls get_user_pages), so that we don't risk
967 * the page we just got a reference to gets unmapped before we have a
968 * chance to grab the mmu_lock, which ensure that if the page gets
969 * unmapped afterwards, the call to kvm_unmap_hva will take it away
970 * from us again properly. This smp_rmb() interacts with the smp_wmb()
971 * in kvm_mmu_notifier_invalidate_<page|range_end>.
975 pfn
= gfn_to_pfn_prot(kvm
, gfn
, write_fault
, &writable
);
976 if (is_error_pfn(pfn
))
979 if (kvm_is_device_pfn(pfn
))
980 mem_type
= PAGE_S2_DEVICE
;
982 spin_lock(&kvm
->mmu_lock
);
983 if (mmu_notifier_retry(kvm
, mmu_seq
))
985 if (!hugetlb
&& !force_pte
)
986 hugetlb
= transparent_hugepage_adjust(&pfn
, &fault_ipa
);
988 fault_ipa_uncached
= memslot
->flags
& KVM_MEMSLOT_INCOHERENT
;
991 pmd_t new_pmd
= pfn_pmd(pfn
, mem_type
);
992 new_pmd
= pmd_mkhuge(new_pmd
);
994 kvm_set_s2pmd_writable(&new_pmd
);
995 kvm_set_pfn_dirty(pfn
);
997 coherent_cache_guest_page(vcpu
, hva
& PMD_MASK
, PMD_SIZE
,
999 ret
= stage2_set_pmd_huge(kvm
, memcache
, fault_ipa
, &new_pmd
);
1001 pte_t new_pte
= pfn_pte(pfn
, mem_type
);
1003 kvm_set_s2pte_writable(&new_pte
);
1004 kvm_set_pfn_dirty(pfn
);
1006 coherent_cache_guest_page(vcpu
, hva
, PAGE_SIZE
,
1007 fault_ipa_uncached
);
1008 ret
= stage2_set_pte(kvm
, memcache
, fault_ipa
, &new_pte
,
1009 pgprot_val(mem_type
) == pgprot_val(PAGE_S2_DEVICE
));
1014 spin_unlock(&kvm
->mmu_lock
);
1015 kvm_release_pfn_clean(pfn
);
1020 * kvm_handle_guest_abort - handles all 2nd stage aborts
1021 * @vcpu: the VCPU pointer
1022 * @run: the kvm_run structure
1024 * Any abort that gets to the host is almost guaranteed to be caused by a
1025 * missing second stage translation table entry, which can mean that either the
1026 * guest simply needs more memory and we must allocate an appropriate page or it
1027 * can mean that the guest tried to access I/O memory, which is emulated by user
1028 * space. The distinction is based on the IPA causing the fault and whether this
1029 * memory region has been registered as standard RAM by user space.
1031 int kvm_handle_guest_abort(struct kvm_vcpu
*vcpu
, struct kvm_run
*run
)
1033 unsigned long fault_status
;
1034 phys_addr_t fault_ipa
;
1035 struct kvm_memory_slot
*memslot
;
1037 bool is_iabt
, write_fault
, writable
;
1041 is_iabt
= kvm_vcpu_trap_is_iabt(vcpu
);
1042 fault_ipa
= kvm_vcpu_get_fault_ipa(vcpu
);
1044 trace_kvm_guest_fault(*vcpu_pc(vcpu
), kvm_vcpu_get_hsr(vcpu
),
1045 kvm_vcpu_get_hfar(vcpu
), fault_ipa
);
1047 /* Check the stage-2 fault is trans. fault or write fault */
1048 fault_status
= kvm_vcpu_trap_get_fault_type(vcpu
);
1049 if (fault_status
!= FSC_FAULT
&& fault_status
!= FSC_PERM
) {
1050 kvm_err("Unsupported FSC: EC=%#x xFSC=%#lx ESR_EL2=%#lx\n",
1051 kvm_vcpu_trap_get_class(vcpu
),
1052 (unsigned long)kvm_vcpu_trap_get_fault(vcpu
),
1053 (unsigned long)kvm_vcpu_get_hsr(vcpu
));
1057 idx
= srcu_read_lock(&vcpu
->kvm
->srcu
);
1059 gfn
= fault_ipa
>> PAGE_SHIFT
;
1060 memslot
= gfn_to_memslot(vcpu
->kvm
, gfn
);
1061 hva
= gfn_to_hva_memslot_prot(memslot
, gfn
, &writable
);
1062 write_fault
= kvm_is_write_fault(vcpu
);
1063 if (kvm_is_error_hva(hva
) || (write_fault
&& !writable
)) {
1065 /* Prefetch Abort on I/O address */
1066 kvm_inject_pabt(vcpu
, kvm_vcpu_get_hfar(vcpu
));
1072 * The IPA is reported as [MAX:12], so we need to
1073 * complement it with the bottom 12 bits from the
1074 * faulting VA. This is always 12 bits, irrespective
1077 fault_ipa
|= kvm_vcpu_get_hfar(vcpu
) & ((1 << 12) - 1);
1078 ret
= io_mem_abort(vcpu
, run
, fault_ipa
);
1082 /* Userspace should not be able to register out-of-bounds IPAs */
1083 VM_BUG_ON(fault_ipa
>= KVM_PHYS_SIZE
);
1085 ret
= user_mem_abort(vcpu
, fault_ipa
, memslot
, hva
, fault_status
);
1089 srcu_read_unlock(&vcpu
->kvm
->srcu
, idx
);
1093 static void handle_hva_to_gpa(struct kvm
*kvm
,
1094 unsigned long start
,
1096 void (*handler
)(struct kvm
*kvm
,
1097 gpa_t gpa
, void *data
),
1100 struct kvm_memslots
*slots
;
1101 struct kvm_memory_slot
*memslot
;
1103 slots
= kvm_memslots(kvm
);
1105 /* we only care about the pages that the guest sees */
1106 kvm_for_each_memslot(memslot
, slots
) {
1107 unsigned long hva_start
, hva_end
;
1110 hva_start
= max(start
, memslot
->userspace_addr
);
1111 hva_end
= min(end
, memslot
->userspace_addr
+
1112 (memslot
->npages
<< PAGE_SHIFT
));
1113 if (hva_start
>= hva_end
)
1117 * {gfn(page) | page intersects with [hva_start, hva_end)} =
1118 * {gfn_start, gfn_start+1, ..., gfn_end-1}.
1120 gfn
= hva_to_gfn_memslot(hva_start
, memslot
);
1121 gfn_end
= hva_to_gfn_memslot(hva_end
+ PAGE_SIZE
- 1, memslot
);
1123 for (; gfn
< gfn_end
; ++gfn
) {
1124 gpa_t gpa
= gfn
<< PAGE_SHIFT
;
1125 handler(kvm
, gpa
, data
);
1130 static void kvm_unmap_hva_handler(struct kvm
*kvm
, gpa_t gpa
, void *data
)
1132 unmap_stage2_range(kvm
, gpa
, PAGE_SIZE
);
1135 int kvm_unmap_hva(struct kvm
*kvm
, unsigned long hva
)
1137 unsigned long end
= hva
+ PAGE_SIZE
;
1142 trace_kvm_unmap_hva(hva
);
1143 handle_hva_to_gpa(kvm
, hva
, end
, &kvm_unmap_hva_handler
, NULL
);
1147 int kvm_unmap_hva_range(struct kvm
*kvm
,
1148 unsigned long start
, unsigned long end
)
1153 trace_kvm_unmap_hva_range(start
, end
);
1154 handle_hva_to_gpa(kvm
, start
, end
, &kvm_unmap_hva_handler
, NULL
);
1158 static void kvm_set_spte_handler(struct kvm
*kvm
, gpa_t gpa
, void *data
)
1160 pte_t
*pte
= (pte_t
*)data
;
1162 stage2_set_pte(kvm
, NULL
, gpa
, pte
, false);
1166 void kvm_set_spte_hva(struct kvm
*kvm
, unsigned long hva
, pte_t pte
)
1168 unsigned long end
= hva
+ PAGE_SIZE
;
1174 trace_kvm_set_spte_hva(hva
);
1175 stage2_pte
= pfn_pte(pte_pfn(pte
), PAGE_S2
);
1176 handle_hva_to_gpa(kvm
, hva
, end
, &kvm_set_spte_handler
, &stage2_pte
);
1179 void kvm_mmu_free_memory_caches(struct kvm_vcpu
*vcpu
)
1181 mmu_free_memory_cache(&vcpu
->arch
.mmu_page_cache
);
1184 phys_addr_t
kvm_mmu_get_httbr(void)
1186 return virt_to_phys(hyp_pgd
);
1189 phys_addr_t
kvm_mmu_get_boot_httbr(void)
1191 return virt_to_phys(boot_hyp_pgd
);
1194 phys_addr_t
kvm_get_idmap_vector(void)
1196 return hyp_idmap_vector
;
1199 int kvm_mmu_init(void)
1203 hyp_idmap_start
= kvm_virt_to_phys(__hyp_idmap_text_start
);
1204 hyp_idmap_end
= kvm_virt_to_phys(__hyp_idmap_text_end
);
1205 hyp_idmap_vector
= kvm_virt_to_phys(__kvm_hyp_init
);
1207 if ((hyp_idmap_start
^ hyp_idmap_end
) & PAGE_MASK
) {
1209 * Our init code is crossing a page boundary. Allocate
1210 * a bounce page, copy the code over and use that.
1212 size_t len
= __hyp_idmap_text_end
- __hyp_idmap_text_start
;
1213 phys_addr_t phys_base
;
1215 init_bounce_page
= (void *)__get_free_page(GFP_KERNEL
);
1216 if (!init_bounce_page
) {
1217 kvm_err("Couldn't allocate HYP init bounce page\n");
1222 memcpy(init_bounce_page
, __hyp_idmap_text_start
, len
);
1224 * Warning: the code we just copied to the bounce page
1225 * must be flushed to the point of coherency.
1226 * Otherwise, the data may be sitting in L2, and HYP
1227 * mode won't be able to observe it as it runs with
1228 * caches off at that point.
1230 kvm_flush_dcache_to_poc(init_bounce_page
, len
);
1232 phys_base
= kvm_virt_to_phys(init_bounce_page
);
1233 hyp_idmap_vector
+= phys_base
- hyp_idmap_start
;
1234 hyp_idmap_start
= phys_base
;
1235 hyp_idmap_end
= phys_base
+ len
;
1237 kvm_info("Using HYP init bounce page @%lx\n",
1238 (unsigned long)phys_base
);
1241 hyp_pgd
= (pgd_t
*)__get_free_pages(GFP_KERNEL
| __GFP_ZERO
, hyp_pgd_order
);
1242 boot_hyp_pgd
= (pgd_t
*)__get_free_pages(GFP_KERNEL
| __GFP_ZERO
, hyp_pgd_order
);
1244 if (!hyp_pgd
|| !boot_hyp_pgd
) {
1245 kvm_err("Hyp mode PGD not allocated\n");
1250 /* Create the idmap in the boot page tables */
1251 err
= __create_hyp_mappings(boot_hyp_pgd
,
1252 hyp_idmap_start
, hyp_idmap_end
,
1253 __phys_to_pfn(hyp_idmap_start
),
1257 kvm_err("Failed to idmap %lx-%lx\n",
1258 hyp_idmap_start
, hyp_idmap_end
);
1262 /* Map the very same page at the trampoline VA */
1263 err
= __create_hyp_mappings(boot_hyp_pgd
,
1264 TRAMPOLINE_VA
, TRAMPOLINE_VA
+ PAGE_SIZE
,
1265 __phys_to_pfn(hyp_idmap_start
),
1268 kvm_err("Failed to map trampoline @%lx into boot HYP pgd\n",
1273 /* Map the same page again into the runtime page tables */
1274 err
= __create_hyp_mappings(hyp_pgd
,
1275 TRAMPOLINE_VA
, TRAMPOLINE_VA
+ PAGE_SIZE
,
1276 __phys_to_pfn(hyp_idmap_start
),
1279 kvm_err("Failed to map trampoline @%lx into runtime HYP pgd\n",
1290 void kvm_arch_commit_memory_region(struct kvm
*kvm
,
1291 struct kvm_userspace_memory_region
*mem
,
1292 const struct kvm_memory_slot
*old
,
1293 enum kvm_mr_change change
)
1297 int kvm_arch_prepare_memory_region(struct kvm
*kvm
,
1298 struct kvm_memory_slot
*memslot
,
1299 struct kvm_userspace_memory_region
*mem
,
1300 enum kvm_mr_change change
)
1302 hva_t hva
= mem
->userspace_addr
;
1303 hva_t reg_end
= hva
+ mem
->memory_size
;
1304 bool writable
= !(mem
->flags
& KVM_MEM_READONLY
);
1307 if (change
!= KVM_MR_CREATE
&& change
!= KVM_MR_MOVE
)
1311 * Prevent userspace from creating a memory region outside of the IPA
1312 * space addressable by the KVM guest IPA space.
1314 if (memslot
->base_gfn
+ memslot
->npages
>=
1315 (KVM_PHYS_SIZE
>> PAGE_SHIFT
))
1319 * A memory region could potentially cover multiple VMAs, and any holes
1320 * between them, so iterate over all of them to find out if we can map
1321 * any of them right now.
1323 * +--------------------------------------------+
1324 * +---------------+----------------+ +----------------+
1325 * | : VMA 1 | VMA 2 | | VMA 3 : |
1326 * +---------------+----------------+ +----------------+
1328 * +--------------------------------------------+
1331 struct vm_area_struct
*vma
= find_vma(current
->mm
, hva
);
1332 hva_t vm_start
, vm_end
;
1334 if (!vma
|| vma
->vm_start
>= reg_end
)
1338 * Mapping a read-only VMA is only allowed if the
1339 * memory region is configured as read-only.
1341 if (writable
&& !(vma
->vm_flags
& VM_WRITE
)) {
1347 * Take the intersection of this VMA with the memory region
1349 vm_start
= max(hva
, vma
->vm_start
);
1350 vm_end
= min(reg_end
, vma
->vm_end
);
1352 if (vma
->vm_flags
& VM_PFNMAP
) {
1353 gpa_t gpa
= mem
->guest_phys_addr
+
1354 (vm_start
- mem
->userspace_addr
);
1355 phys_addr_t pa
= (vma
->vm_pgoff
<< PAGE_SHIFT
) +
1356 vm_start
- vma
->vm_start
;
1358 ret
= kvm_phys_addr_ioremap(kvm
, gpa
, pa
,
1365 } while (hva
< reg_end
);
1367 spin_lock(&kvm
->mmu_lock
);
1369 unmap_stage2_range(kvm
, mem
->guest_phys_addr
, mem
->memory_size
);
1371 stage2_flush_memslot(kvm
, memslot
);
1372 spin_unlock(&kvm
->mmu_lock
);
1376 void kvm_arch_free_memslot(struct kvm
*kvm
, struct kvm_memory_slot
*free
,
1377 struct kvm_memory_slot
*dont
)
1381 int kvm_arch_create_memslot(struct kvm
*kvm
, struct kvm_memory_slot
*slot
,
1382 unsigned long npages
)
1385 * Readonly memslots are not incoherent with the caches by definition,
1386 * but in practice, they are used mostly to emulate ROMs or NOR flashes
1387 * that the guest may consider devices and hence map as uncached.
1388 * To prevent incoherency issues in these cases, tag all readonly
1389 * regions as incoherent.
1391 if (slot
->flags
& KVM_MEM_READONLY
)
1392 slot
->flags
|= KVM_MEMSLOT_INCOHERENT
;
1396 void kvm_arch_memslots_updated(struct kvm
*kvm
)
1400 void kvm_arch_flush_shadow_all(struct kvm
*kvm
)
1404 void kvm_arch_flush_shadow_memslot(struct kvm
*kvm
,
1405 struct kvm_memory_slot
*slot
)
1407 gpa_t gpa
= slot
->base_gfn
<< PAGE_SHIFT
;
1408 phys_addr_t size
= slot
->npages
<< PAGE_SHIFT
;
1410 spin_lock(&kvm
->mmu_lock
);
1411 unmap_stage2_range(kvm
, gpa
, size
);
1412 spin_unlock(&kvm
->mmu_lock
);