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 <trace/events/kvm.h>
23 #include <asm/idmap.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 DEFINE_MUTEX(kvm_hyp_pgd_mutex);
38 static void kvm_tlb_flush_vmid(struct kvm *kvm)
40 kvm_call_hyp(__kvm_tlb_flush_vmid, kvm);
43 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
48 BUG_ON(max > KVM_NR_MEM_OBJS);
49 if (cache->nobjs >= min)
51 while (cache->nobjs < max) {
52 page = (void *)__get_free_page(PGALLOC_GFP);
55 cache->objects[cache->nobjs++] = page;
60 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
63 free_page((unsigned long)mc->objects[--mc->nobjs]);
66 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc)
70 BUG_ON(!mc || !mc->nobjs);
71 p = mc->objects[--mc->nobjs];
75 static void free_ptes(pmd_t *pmd, unsigned long addr)
80 for (i = 0; i < PTRS_PER_PMD; i++, addr += PMD_SIZE) {
81 if (!pmd_none(*pmd) && pmd_table(*pmd)) {
82 pte = pte_offset_kernel(pmd, addr);
83 pte_free_kernel(NULL, pte);
90 * free_hyp_pmds - free a Hyp-mode level-2 tables and child level-3 tables
92 * Assumes this is a page table used strictly in Hyp-mode and therefore contains
93 * only mappings in the kernel memory area, which is above PAGE_OFFSET.
95 void free_hyp_pmds(void)
102 mutex_lock(&kvm_hyp_pgd_mutex);
103 for (addr = PAGE_OFFSET; addr != 0; addr += PGDIR_SIZE) {
104 pgd = hyp_pgd + pgd_index(addr);
105 pud = pud_offset(pgd, addr);
109 BUG_ON(pud_bad(*pud));
111 pmd = pmd_offset(pud, addr);
112 free_ptes(pmd, addr);
116 mutex_unlock(&kvm_hyp_pgd_mutex);
119 static void create_hyp_pte_mappings(pmd_t *pmd, unsigned long start,
126 for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) {
127 pte = pte_offset_kernel(pmd, addr);
128 BUG_ON(!virt_addr_valid(addr));
129 page = virt_to_page(addr);
130 kvm_set_pte(pte, mk_pte(page, PAGE_HYP));
134 static void create_hyp_io_pte_mappings(pmd_t *pmd, unsigned long start,
136 unsigned long *pfn_base)
141 for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) {
142 pte = pte_offset_kernel(pmd, addr);
143 BUG_ON(pfn_valid(*pfn_base));
144 kvm_set_pte(pte, pfn_pte(*pfn_base, PAGE_HYP_DEVICE));
149 static int create_hyp_pmd_mappings(pud_t *pud, unsigned long start,
150 unsigned long end, unsigned long *pfn_base)
154 unsigned long addr, next;
156 for (addr = start; addr < end; addr = next) {
157 pmd = pmd_offset(pud, addr);
159 BUG_ON(pmd_sect(*pmd));
161 if (pmd_none(*pmd)) {
162 pte = pte_alloc_one_kernel(NULL, addr);
164 kvm_err("Cannot allocate Hyp pte\n");
167 pmd_populate_kernel(NULL, pmd, pte);
170 next = pmd_addr_end(addr, end);
173 * If pfn_base is NULL, we map kernel pages into HYP with the
174 * virtual address. Otherwise, this is considered an I/O
175 * mapping and we map the physical region starting at
176 * *pfn_base to [start, end[.
179 create_hyp_pte_mappings(pmd, addr, next);
181 create_hyp_io_pte_mappings(pmd, addr, next, pfn_base);
187 static int __create_hyp_mappings(void *from, void *to, unsigned long *pfn_base)
189 unsigned long start = (unsigned long)from;
190 unsigned long end = (unsigned long)to;
194 unsigned long addr, next;
198 if (start < PAGE_OFFSET)
201 mutex_lock(&kvm_hyp_pgd_mutex);
202 for (addr = start; addr < end; addr = next) {
203 pgd = hyp_pgd + pgd_index(addr);
204 pud = pud_offset(pgd, addr);
206 if (pud_none_or_clear_bad(pud)) {
207 pmd = pmd_alloc_one(NULL, addr);
209 kvm_err("Cannot allocate Hyp pmd\n");
213 pud_populate(NULL, pud, pmd);
216 next = pgd_addr_end(addr, end);
217 err = create_hyp_pmd_mappings(pud, addr, next, pfn_base);
222 mutex_unlock(&kvm_hyp_pgd_mutex);
227 * create_hyp_mappings - map a kernel virtual address range in Hyp mode
228 * @from: The virtual kernel start address of the range
229 * @to: The virtual kernel end address of the range (exclusive)
231 * The same virtual address as the kernel virtual address is also used in
232 * Hyp-mode mapping to the same underlying physical pages.
234 * Note: Wrapping around zero in the "to" address is not supported.
236 int create_hyp_mappings(void *from, void *to)
238 return __create_hyp_mappings(from, to, NULL);
242 * create_hyp_io_mappings - map a physical IO range in Hyp mode
243 * @from: The virtual HYP start address of the range
244 * @to: The virtual HYP end address of the range (exclusive)
245 * @addr: The physical start address which gets mapped
247 int create_hyp_io_mappings(void *from, void *to, phys_addr_t addr)
249 unsigned long pfn = __phys_to_pfn(addr);
250 return __create_hyp_mappings(from, to, &pfn);
254 * kvm_alloc_stage2_pgd - allocate level-1 table for stage-2 translation.
255 * @kvm: The KVM struct pointer for the VM.
257 * Allocates the 1st level table only of size defined by S2_PGD_ORDER (can
258 * support either full 40-bit input addresses or limited to 32-bit input
259 * addresses). Clears the allocated pages.
261 * Note we don't need locking here as this is only called when the VM is
262 * created, which can only be done once.
264 int kvm_alloc_stage2_pgd(struct kvm *kvm)
268 if (kvm->arch.pgd != NULL) {
269 kvm_err("kvm_arch already initialized?\n");
273 pgd = (pgd_t *)__get_free_pages(GFP_KERNEL, S2_PGD_ORDER);
277 /* stage-2 pgd must be aligned to its size */
278 VM_BUG_ON((unsigned long)pgd & (S2_PGD_SIZE - 1));
280 memset(pgd, 0, PTRS_PER_S2_PGD * sizeof(pgd_t));
287 static void clear_pud_entry(pud_t *pud)
289 pmd_t *pmd_table = pmd_offset(pud, 0);
291 pmd_free(NULL, pmd_table);
292 put_page(virt_to_page(pud));
295 static void clear_pmd_entry(pmd_t *pmd)
297 pte_t *pte_table = pte_offset_kernel(pmd, 0);
299 pte_free_kernel(NULL, pte_table);
300 put_page(virt_to_page(pmd));
303 static bool pmd_empty(pmd_t *pmd)
305 struct page *pmd_page = virt_to_page(pmd);
306 return page_count(pmd_page) == 1;
309 static void clear_pte_entry(pte_t *pte)
311 if (pte_present(*pte)) {
312 kvm_set_pte(pte, __pte(0));
313 put_page(virt_to_page(pte));
317 static bool pte_empty(pte_t *pte)
319 struct page *pte_page = virt_to_page(pte);
320 return page_count(pte_page) == 1;
324 * unmap_stage2_range -- Clear stage2 page table entries to unmap a range
325 * @kvm: The VM pointer
326 * @start: The intermediate physical base address of the range to unmap
327 * @size: The size of the area to unmap
329 * Clear a range of stage-2 mappings, lowering the various ref-counts. Must
330 * be called while holding mmu_lock (unless for freeing the stage2 pgd before
331 * destroying the VM), otherwise another faulting VCPU may come in and mess
332 * with things behind our backs.
334 static void unmap_stage2_range(struct kvm *kvm, phys_addr_t start, u64 size)
340 phys_addr_t addr = start, end = start + size;
344 pgd = kvm->arch.pgd + pgd_index(addr);
345 pud = pud_offset(pgd, addr);
346 if (pud_none(*pud)) {
351 pmd = pmd_offset(pud, addr);
352 if (pmd_none(*pmd)) {
357 pte = pte_offset_kernel(pmd, addr);
358 clear_pte_entry(pte);
361 /* If we emptied the pte, walk back up the ladder */
362 if (pte_empty(pte)) {
363 clear_pmd_entry(pmd);
365 if (pmd_empty(pmd)) {
366 clear_pud_entry(pud);
376 * kvm_free_stage2_pgd - free all stage-2 tables
377 * @kvm: The KVM struct pointer for the VM.
379 * Walks the level-1 page table pointed to by kvm->arch.pgd and frees all
380 * underlying level-2 and level-3 tables before freeing the actual level-1 table
381 * and setting the struct pointer to NULL.
383 * Note we don't need locking here as this is only called when the VM is
384 * destroyed, which can only be done once.
386 void kvm_free_stage2_pgd(struct kvm *kvm)
388 if (kvm->arch.pgd == NULL)
391 unmap_stage2_range(kvm, 0, KVM_PHYS_SIZE);
392 free_pages((unsigned long)kvm->arch.pgd, S2_PGD_ORDER);
393 kvm->arch.pgd = NULL;
397 static int stage2_set_pte(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
398 phys_addr_t addr, const pte_t *new_pte, bool iomap)
405 /* Create 2nd stage page table mapping - Level 1 */
406 pgd = kvm->arch.pgd + pgd_index(addr);
407 pud = pud_offset(pgd, addr);
408 if (pud_none(*pud)) {
410 return 0; /* ignore calls from kvm_set_spte_hva */
411 pmd = mmu_memory_cache_alloc(cache);
412 pud_populate(NULL, pud, pmd);
413 get_page(virt_to_page(pud));
416 pmd = pmd_offset(pud, addr);
418 /* Create 2nd stage page table mapping - Level 2 */
419 if (pmd_none(*pmd)) {
421 return 0; /* ignore calls from kvm_set_spte_hva */
422 pte = mmu_memory_cache_alloc(cache);
424 pmd_populate_kernel(NULL, pmd, pte);
425 get_page(virt_to_page(pmd));
428 pte = pte_offset_kernel(pmd, addr);
430 if (iomap && pte_present(*pte))
433 /* Create 2nd stage page table mapping - Level 3 */
435 kvm_set_pte(pte, *new_pte);
436 if (pte_present(old_pte))
437 kvm_tlb_flush_vmid(kvm);
439 get_page(virt_to_page(pte));
445 * kvm_phys_addr_ioremap - map a device range to guest IPA
447 * @kvm: The KVM pointer
448 * @guest_ipa: The IPA at which to insert the mapping
449 * @pa: The physical address of the device
450 * @size: The size of the mapping
452 int kvm_phys_addr_ioremap(struct kvm *kvm, phys_addr_t guest_ipa,
453 phys_addr_t pa, unsigned long size)
455 phys_addr_t addr, end;
458 struct kvm_mmu_memory_cache cache = { 0, };
460 end = (guest_ipa + size + PAGE_SIZE - 1) & PAGE_MASK;
461 pfn = __phys_to_pfn(pa);
463 for (addr = guest_ipa; addr < end; addr += PAGE_SIZE) {
464 pte_t pte = pfn_pte(pfn, PAGE_S2_DEVICE);
465 kvm_set_s2pte_writable(&pte);
467 ret = mmu_topup_memory_cache(&cache, 2, 2);
470 spin_lock(&kvm->mmu_lock);
471 ret = stage2_set_pte(kvm, &cache, addr, &pte, true);
472 spin_unlock(&kvm->mmu_lock);
480 mmu_free_memory_cache(&cache);
484 static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
485 gfn_t gfn, struct kvm_memory_slot *memslot,
486 unsigned long fault_status)
491 bool write_fault, writable;
492 unsigned long mmu_seq;
493 struct kvm_mmu_memory_cache *memcache = &vcpu->arch.mmu_page_cache;
495 write_fault = kvm_is_write_fault(kvm_vcpu_get_hsr(vcpu));
496 if (fault_status == FSC_PERM && !write_fault) {
497 kvm_err("Unexpected L2 read permission error\n");
501 /* We need minimum second+third level pages */
502 ret = mmu_topup_memory_cache(memcache, 2, KVM_NR_MEM_OBJS);
506 mmu_seq = vcpu->kvm->mmu_notifier_seq;
508 * Ensure the read of mmu_notifier_seq happens before we call
509 * gfn_to_pfn_prot (which calls get_user_pages), so that we don't risk
510 * the page we just got a reference to gets unmapped before we have a
511 * chance to grab the mmu_lock, which ensure that if the page gets
512 * unmapped afterwards, the call to kvm_unmap_hva will take it away
513 * from us again properly. This smp_rmb() interacts with the smp_wmb()
514 * in kvm_mmu_notifier_invalidate_<page|range_end>.
518 pfn = gfn_to_pfn_prot(vcpu->kvm, gfn, write_fault, &writable);
519 if (is_error_pfn(pfn))
522 new_pte = pfn_pte(pfn, PAGE_S2);
523 coherent_icache_guest_page(vcpu->kvm, gfn);
525 spin_lock(&vcpu->kvm->mmu_lock);
526 if (mmu_notifier_retry(vcpu->kvm, mmu_seq))
529 kvm_set_s2pte_writable(&new_pte);
530 kvm_set_pfn_dirty(pfn);
532 stage2_set_pte(vcpu->kvm, memcache, fault_ipa, &new_pte, false);
535 spin_unlock(&vcpu->kvm->mmu_lock);
536 kvm_release_pfn_clean(pfn);
541 * kvm_handle_guest_abort - handles all 2nd stage aborts
542 * @vcpu: the VCPU pointer
543 * @run: the kvm_run structure
545 * Any abort that gets to the host is almost guaranteed to be caused by a
546 * missing second stage translation table entry, which can mean that either the
547 * guest simply needs more memory and we must allocate an appropriate page or it
548 * can mean that the guest tried to access I/O memory, which is emulated by user
549 * space. The distinction is based on the IPA causing the fault and whether this
550 * memory region has been registered as standard RAM by user space.
552 int kvm_handle_guest_abort(struct kvm_vcpu *vcpu, struct kvm_run *run)
554 unsigned long fault_status;
555 phys_addr_t fault_ipa;
556 struct kvm_memory_slot *memslot;
561 is_iabt = kvm_vcpu_trap_is_iabt(vcpu);
562 fault_ipa = kvm_vcpu_get_fault_ipa(vcpu);
564 trace_kvm_guest_fault(*vcpu_pc(vcpu), kvm_vcpu_get_hsr(vcpu),
565 kvm_vcpu_get_hfar(vcpu), fault_ipa);
567 /* Check the stage-2 fault is trans. fault or write fault */
568 fault_status = kvm_vcpu_trap_get_fault(vcpu);
569 if (fault_status != FSC_FAULT && fault_status != FSC_PERM) {
570 kvm_err("Unsupported fault status: EC=%#x DFCS=%#lx\n",
571 kvm_vcpu_trap_get_class(vcpu), fault_status);
575 idx = srcu_read_lock(&vcpu->kvm->srcu);
577 gfn = fault_ipa >> PAGE_SHIFT;
578 if (!kvm_is_visible_gfn(vcpu->kvm, gfn)) {
580 /* Prefetch Abort on I/O address */
581 kvm_inject_pabt(vcpu, kvm_vcpu_get_hfar(vcpu));
586 if (fault_status != FSC_FAULT) {
587 kvm_err("Unsupported fault status on io memory: %#lx\n",
593 /* Adjust page offset */
594 fault_ipa |= kvm_vcpu_get_hfar(vcpu) & ~PAGE_MASK;
595 ret = io_mem_abort(vcpu, run, fault_ipa);
599 memslot = gfn_to_memslot(vcpu->kvm, gfn);
601 ret = user_mem_abort(vcpu, fault_ipa, gfn, memslot, fault_status);
605 srcu_read_unlock(&vcpu->kvm->srcu, idx);
609 static void handle_hva_to_gpa(struct kvm *kvm,
612 void (*handler)(struct kvm *kvm,
613 gpa_t gpa, void *data),
616 struct kvm_memslots *slots;
617 struct kvm_memory_slot *memslot;
619 slots = kvm_memslots(kvm);
621 /* we only care about the pages that the guest sees */
622 kvm_for_each_memslot(memslot, slots) {
623 unsigned long hva_start, hva_end;
626 hva_start = max(start, memslot->userspace_addr);
627 hva_end = min(end, memslot->userspace_addr +
628 (memslot->npages << PAGE_SHIFT));
629 if (hva_start >= hva_end)
633 * {gfn(page) | page intersects with [hva_start, hva_end)} =
634 * {gfn_start, gfn_start+1, ..., gfn_end-1}.
636 gfn = hva_to_gfn_memslot(hva_start, memslot);
637 gfn_end = hva_to_gfn_memslot(hva_end + PAGE_SIZE - 1, memslot);
639 for (; gfn < gfn_end; ++gfn) {
640 gpa_t gpa = gfn << PAGE_SHIFT;
641 handler(kvm, gpa, data);
646 static void kvm_unmap_hva_handler(struct kvm *kvm, gpa_t gpa, void *data)
648 unmap_stage2_range(kvm, gpa, PAGE_SIZE);
649 kvm_tlb_flush_vmid(kvm);
652 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
654 unsigned long end = hva + PAGE_SIZE;
659 trace_kvm_unmap_hva(hva);
660 handle_hva_to_gpa(kvm, hva, end, &kvm_unmap_hva_handler, NULL);
664 int kvm_unmap_hva_range(struct kvm *kvm,
665 unsigned long start, unsigned long end)
670 trace_kvm_unmap_hva_range(start, end);
671 handle_hva_to_gpa(kvm, start, end, &kvm_unmap_hva_handler, NULL);
675 static void kvm_set_spte_handler(struct kvm *kvm, gpa_t gpa, void *data)
677 pte_t *pte = (pte_t *)data;
679 stage2_set_pte(kvm, NULL, gpa, pte, false);
683 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
685 unsigned long end = hva + PAGE_SIZE;
691 trace_kvm_set_spte_hva(hva);
692 stage2_pte = pfn_pte(pte_pfn(pte), PAGE_S2);
693 handle_hva_to_gpa(kvm, hva, end, &kvm_set_spte_handler, &stage2_pte);
696 void kvm_mmu_free_memory_caches(struct kvm_vcpu *vcpu)
698 mmu_free_memory_cache(&vcpu->arch.mmu_page_cache);
701 phys_addr_t kvm_mmu_get_httbr(void)
703 VM_BUG_ON(!virt_addr_valid(hyp_pgd));
704 return virt_to_phys(hyp_pgd);
707 int kvm_mmu_init(void)
710 kvm_err("Hyp mode PGD not allocated\n");
718 * kvm_clear_idmap - remove all idmaps from the hyp pgd
720 * Free the underlying pmds for all pgds in range and clear the pgds (but
721 * don't free them) afterwards.
723 void kvm_clear_hyp_idmap(void)
725 unsigned long addr, end;
727 pgd_t *pgd = hyp_pgd;
731 addr = virt_to_phys(__hyp_idmap_text_start);
732 end = virt_to_phys(__hyp_idmap_text_end);
734 pgd += pgd_index(addr);
736 next = pgd_addr_end(addr, end);
737 if (pgd_none_or_clear_bad(pgd))
739 pud = pud_offset(pgd, addr);
740 pmd = pmd_offset(pud, addr);
743 kvm_clean_pmd_entry(pmd);
744 pmd_free(NULL, (pmd_t *)((unsigned long)pmd & PAGE_MASK));
745 } while (pgd++, addr = next, addr < end);