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 <asm/idmap.h>
23 #include <asm/pgalloc.h>
24 #include <asm/cacheflush.h>
25 #include <asm/kvm_arm.h>
26 #include <asm/kvm_mmu.h>
27 #include <asm/kvm_asm.h>
28 #include <asm/kvm_emulate.h>
29 #include <asm/mach/map.h>
30 #include <trace/events/kvm.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 void kvm_set_pte(pte_t *pte, pte_t new_pte)
45 pte_val(*pte) = new_pte;
47 * flush_pmd_entry just takes a void pointer and cleans the necessary
48 * cache entries, so we can reuse the function for ptes.
53 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
58 BUG_ON(max > KVM_NR_MEM_OBJS);
59 if (cache->nobjs >= min)
61 while (cache->nobjs < max) {
62 page = (void *)__get_free_page(PGALLOC_GFP);
65 cache->objects[cache->nobjs++] = page;
70 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
73 free_page((unsigned long)mc->objects[--mc->nobjs]);
76 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc)
80 BUG_ON(!mc || !mc->nobjs);
81 p = mc->objects[--mc->nobjs];
85 static void free_ptes(pmd_t *pmd, unsigned long addr)
90 for (i = 0; i < PTRS_PER_PMD; i++, addr += PMD_SIZE) {
91 if (!pmd_none(*pmd) && pmd_table(*pmd)) {
92 pte = pte_offset_kernel(pmd, addr);
93 pte_free_kernel(NULL, pte);
100 * free_hyp_pmds - free a Hyp-mode level-2 tables and child level-3 tables
102 * Assumes this is a page table used strictly in Hyp-mode and therefore contains
103 * only mappings in the kernel memory area, which is above PAGE_OFFSET.
105 void free_hyp_pmds(void)
112 mutex_lock(&kvm_hyp_pgd_mutex);
113 for (addr = PAGE_OFFSET; addr != 0; addr += PGDIR_SIZE) {
114 pgd = hyp_pgd + pgd_index(addr);
115 pud = pud_offset(pgd, addr);
119 BUG_ON(pud_bad(*pud));
121 pmd = pmd_offset(pud, addr);
122 free_ptes(pmd, addr);
126 mutex_unlock(&kvm_hyp_pgd_mutex);
129 static void create_hyp_pte_mappings(pmd_t *pmd, unsigned long start,
136 for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) {
137 pte = pte_offset_kernel(pmd, addr);
138 BUG_ON(!virt_addr_valid(addr));
139 page = virt_to_page(addr);
140 kvm_set_pte(pte, mk_pte(page, PAGE_HYP));
144 static void create_hyp_io_pte_mappings(pmd_t *pmd, unsigned long start,
146 unsigned long *pfn_base)
151 for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) {
152 pte = pte_offset_kernel(pmd, addr);
153 BUG_ON(pfn_valid(*pfn_base));
154 kvm_set_pte(pte, pfn_pte(*pfn_base, PAGE_HYP_DEVICE));
159 static int create_hyp_pmd_mappings(pud_t *pud, unsigned long start,
160 unsigned long end, unsigned long *pfn_base)
164 unsigned long addr, next;
166 for (addr = start; addr < end; addr = next) {
167 pmd = pmd_offset(pud, addr);
169 BUG_ON(pmd_sect(*pmd));
171 if (pmd_none(*pmd)) {
172 pte = pte_alloc_one_kernel(NULL, addr);
174 kvm_err("Cannot allocate Hyp pte\n");
177 pmd_populate_kernel(NULL, pmd, pte);
180 next = pmd_addr_end(addr, end);
183 * If pfn_base is NULL, we map kernel pages into HYP with the
184 * virtual address. Otherwise, this is considered an I/O
185 * mapping and we map the physical region starting at
186 * *pfn_base to [start, end[.
189 create_hyp_pte_mappings(pmd, addr, next);
191 create_hyp_io_pte_mappings(pmd, addr, next, pfn_base);
197 static int __create_hyp_mappings(void *from, void *to, unsigned long *pfn_base)
199 unsigned long start = (unsigned long)from;
200 unsigned long end = (unsigned long)to;
204 unsigned long addr, next;
208 if (start < PAGE_OFFSET)
211 mutex_lock(&kvm_hyp_pgd_mutex);
212 for (addr = start; addr < end; addr = next) {
213 pgd = hyp_pgd + pgd_index(addr);
214 pud = pud_offset(pgd, addr);
216 if (pud_none_or_clear_bad(pud)) {
217 pmd = pmd_alloc_one(NULL, addr);
219 kvm_err("Cannot allocate Hyp pmd\n");
223 pud_populate(NULL, pud, pmd);
226 next = pgd_addr_end(addr, end);
227 err = create_hyp_pmd_mappings(pud, addr, next, pfn_base);
232 mutex_unlock(&kvm_hyp_pgd_mutex);
237 * create_hyp_mappings - map a kernel virtual address range in Hyp mode
238 * @from: The virtual kernel start address of the range
239 * @to: The virtual kernel end address of the range (exclusive)
241 * The same virtual address as the kernel virtual address is also used in
242 * Hyp-mode mapping to the same underlying physical pages.
244 * Note: Wrapping around zero in the "to" address is not supported.
246 int create_hyp_mappings(void *from, void *to)
248 return __create_hyp_mappings(from, to, NULL);
252 * create_hyp_io_mappings - map a physical IO range in Hyp mode
253 * @from: The virtual HYP start address of the range
254 * @to: The virtual HYP end address of the range (exclusive)
255 * @addr: The physical start address which gets mapped
257 int create_hyp_io_mappings(void *from, void *to, phys_addr_t addr)
259 unsigned long pfn = __phys_to_pfn(addr);
260 return __create_hyp_mappings(from, to, &pfn);
264 * kvm_alloc_stage2_pgd - allocate level-1 table for stage-2 translation.
265 * @kvm: The KVM struct pointer for the VM.
267 * Allocates the 1st level table only of size defined by S2_PGD_ORDER (can
268 * support either full 40-bit input addresses or limited to 32-bit input
269 * addresses). Clears the allocated pages.
271 * Note we don't need locking here as this is only called when the VM is
272 * created, which can only be done once.
274 int kvm_alloc_stage2_pgd(struct kvm *kvm)
278 if (kvm->arch.pgd != NULL) {
279 kvm_err("kvm_arch already initialized?\n");
283 pgd = (pgd_t *)__get_free_pages(GFP_KERNEL, S2_PGD_ORDER);
287 /* stage-2 pgd must be aligned to its size */
288 VM_BUG_ON((unsigned long)pgd & (S2_PGD_SIZE - 1));
290 memset(pgd, 0, PTRS_PER_S2_PGD * sizeof(pgd_t));
291 clean_dcache_area(pgd, PTRS_PER_S2_PGD * sizeof(pgd_t));
297 static void clear_pud_entry(pud_t *pud)
299 pmd_t *pmd_table = pmd_offset(pud, 0);
301 pmd_free(NULL, pmd_table);
302 put_page(virt_to_page(pud));
305 static void clear_pmd_entry(pmd_t *pmd)
307 pte_t *pte_table = pte_offset_kernel(pmd, 0);
309 pte_free_kernel(NULL, pte_table);
310 put_page(virt_to_page(pmd));
313 static bool pmd_empty(pmd_t *pmd)
315 struct page *pmd_page = virt_to_page(pmd);
316 return page_count(pmd_page) == 1;
319 static void clear_pte_entry(pte_t *pte)
321 if (pte_present(*pte)) {
322 kvm_set_pte(pte, __pte(0));
323 put_page(virt_to_page(pte));
327 static bool pte_empty(pte_t *pte)
329 struct page *pte_page = virt_to_page(pte);
330 return page_count(pte_page) == 1;
334 * unmap_stage2_range -- Clear stage2 page table entries to unmap a range
335 * @kvm: The VM pointer
336 * @start: The intermediate physical base address of the range to unmap
337 * @size: The size of the area to unmap
339 * Clear a range of stage-2 mappings, lowering the various ref-counts. Must
340 * be called while holding mmu_lock (unless for freeing the stage2 pgd before
341 * destroying the VM), otherwise another faulting VCPU may come in and mess
342 * with things behind our backs.
344 static void unmap_stage2_range(struct kvm *kvm, phys_addr_t start, u64 size)
350 phys_addr_t addr = start, end = start + size;
354 pgd = kvm->arch.pgd + pgd_index(addr);
355 pud = pud_offset(pgd, addr);
356 if (pud_none(*pud)) {
361 pmd = pmd_offset(pud, addr);
362 if (pmd_none(*pmd)) {
367 pte = pte_offset_kernel(pmd, addr);
368 clear_pte_entry(pte);
371 /* If we emptied the pte, walk back up the ladder */
372 if (pte_empty(pte)) {
373 clear_pmd_entry(pmd);
375 if (pmd_empty(pmd)) {
376 clear_pud_entry(pud);
386 * kvm_free_stage2_pgd - free all stage-2 tables
387 * @kvm: The KVM struct pointer for the VM.
389 * Walks the level-1 page table pointed to by kvm->arch.pgd and frees all
390 * underlying level-2 and level-3 tables before freeing the actual level-1 table
391 * and setting the struct pointer to NULL.
393 * Note we don't need locking here as this is only called when the VM is
394 * destroyed, which can only be done once.
396 void kvm_free_stage2_pgd(struct kvm *kvm)
398 if (kvm->arch.pgd == NULL)
401 unmap_stage2_range(kvm, 0, KVM_PHYS_SIZE);
402 free_pages((unsigned long)kvm->arch.pgd, S2_PGD_ORDER);
403 kvm->arch.pgd = NULL;
407 static int stage2_set_pte(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
408 phys_addr_t addr, const pte_t *new_pte, bool iomap)
415 /* Create 2nd stage page table mapping - Level 1 */
416 pgd = kvm->arch.pgd + pgd_index(addr);
417 pud = pud_offset(pgd, addr);
418 if (pud_none(*pud)) {
420 return 0; /* ignore calls from kvm_set_spte_hva */
421 pmd = mmu_memory_cache_alloc(cache);
422 pud_populate(NULL, pud, pmd);
423 pmd += pmd_index(addr);
424 get_page(virt_to_page(pud));
426 pmd = pmd_offset(pud, addr);
428 /* Create 2nd stage page table mapping - Level 2 */
429 if (pmd_none(*pmd)) {
431 return 0; /* ignore calls from kvm_set_spte_hva */
432 pte = mmu_memory_cache_alloc(cache);
433 clean_pte_table(pte);
434 pmd_populate_kernel(NULL, pmd, pte);
435 pte += pte_index(addr);
436 get_page(virt_to_page(pmd));
438 pte = pte_offset_kernel(pmd, addr);
440 if (iomap && pte_present(*pte))
443 /* Create 2nd stage page table mapping - Level 3 */
445 kvm_set_pte(pte, *new_pte);
446 if (pte_present(old_pte))
447 kvm_tlb_flush_vmid(kvm);
449 get_page(virt_to_page(pte));
455 * kvm_phys_addr_ioremap - map a device range to guest IPA
457 * @kvm: The KVM pointer
458 * @guest_ipa: The IPA at which to insert the mapping
459 * @pa: The physical address of the device
460 * @size: The size of the mapping
462 int kvm_phys_addr_ioremap(struct kvm *kvm, phys_addr_t guest_ipa,
463 phys_addr_t pa, unsigned long size)
465 phys_addr_t addr, end;
468 struct kvm_mmu_memory_cache cache = { 0, };
470 end = (guest_ipa + size + PAGE_SIZE - 1) & PAGE_MASK;
471 pfn = __phys_to_pfn(pa);
473 for (addr = guest_ipa; addr < end; addr += PAGE_SIZE) {
474 pte_t pte = pfn_pte(pfn, PAGE_S2_DEVICE | L_PTE_S2_RDWR);
476 ret = mmu_topup_memory_cache(&cache, 2, 2);
479 spin_lock(&kvm->mmu_lock);
480 ret = stage2_set_pte(kvm, &cache, addr, &pte, true);
481 spin_unlock(&kvm->mmu_lock);
489 mmu_free_memory_cache(&cache);
493 static void coherent_icache_guest_page(struct kvm *kvm, gfn_t gfn)
496 * If we are going to insert an instruction page and the icache is
497 * either VIPT or PIPT, there is a potential problem where the host
498 * (or another VM) may have used the same page as this guest, and we
499 * read incorrect data from the icache. If we're using a PIPT cache,
500 * we can invalidate just that page, but if we are using a VIPT cache
501 * we need to invalidate the entire icache - damn shame - as written
502 * in the ARM ARM (DDI 0406C.b - Page B3-1393).
504 * VIVT caches are tagged using both the ASID and the VMID and doesn't
505 * need any kind of flushing (DDI 0406C.b - Page B3-1392).
507 if (icache_is_pipt()) {
508 unsigned long hva = gfn_to_hva(kvm, gfn);
509 __cpuc_coherent_user_range(hva, hva + PAGE_SIZE);
510 } else if (!icache_is_vivt_asid_tagged()) {
511 /* any kind of VIPT cache */
512 __flush_icache_all();
516 static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
517 gfn_t gfn, struct kvm_memory_slot *memslot,
518 unsigned long fault_status)
523 bool write_fault, writable;
524 unsigned long mmu_seq;
525 struct kvm_mmu_memory_cache *memcache = &vcpu->arch.mmu_page_cache;
527 write_fault = kvm_is_write_fault(vcpu->arch.hsr);
528 if (fault_status == FSC_PERM && !write_fault) {
529 kvm_err("Unexpected L2 read permission error\n");
533 /* We need minimum second+third level pages */
534 ret = mmu_topup_memory_cache(memcache, 2, KVM_NR_MEM_OBJS);
538 mmu_seq = vcpu->kvm->mmu_notifier_seq;
540 * Ensure the read of mmu_notifier_seq happens before we call
541 * gfn_to_pfn_prot (which calls get_user_pages), so that we don't risk
542 * the page we just got a reference to gets unmapped before we have a
543 * chance to grab the mmu_lock, which ensure that if the page gets
544 * unmapped afterwards, the call to kvm_unmap_hva will take it away
545 * from us again properly. This smp_rmb() interacts with the smp_wmb()
546 * in kvm_mmu_notifier_invalidate_<page|range_end>.
550 pfn = gfn_to_pfn_prot(vcpu->kvm, gfn, write_fault, &writable);
551 if (is_error_pfn(pfn))
554 new_pte = pfn_pte(pfn, PAGE_S2);
555 coherent_icache_guest_page(vcpu->kvm, gfn);
557 spin_lock(&vcpu->kvm->mmu_lock);
558 if (mmu_notifier_retry(vcpu->kvm, mmu_seq))
561 pte_val(new_pte) |= L_PTE_S2_RDWR;
562 kvm_set_pfn_dirty(pfn);
564 stage2_set_pte(vcpu->kvm, memcache, fault_ipa, &new_pte, false);
567 spin_unlock(&vcpu->kvm->mmu_lock);
568 kvm_release_pfn_clean(pfn);
573 * kvm_handle_guest_abort - handles all 2nd stage aborts
574 * @vcpu: the VCPU pointer
575 * @run: the kvm_run structure
577 * Any abort that gets to the host is almost guaranteed to be caused by a
578 * missing second stage translation table entry, which can mean that either the
579 * guest simply needs more memory and we must allocate an appropriate page or it
580 * can mean that the guest tried to access I/O memory, which is emulated by user
581 * space. The distinction is based on the IPA causing the fault and whether this
582 * memory region has been registered as standard RAM by user space.
584 int kvm_handle_guest_abort(struct kvm_vcpu *vcpu, struct kvm_run *run)
586 unsigned long hsr_ec;
587 unsigned long fault_status;
588 phys_addr_t fault_ipa;
589 struct kvm_memory_slot *memslot;
594 hsr_ec = vcpu->arch.hsr >> HSR_EC_SHIFT;
595 is_iabt = (hsr_ec == HSR_EC_IABT);
596 fault_ipa = ((phys_addr_t)vcpu->arch.hpfar & HPFAR_MASK) << 8;
598 trace_kvm_guest_fault(*vcpu_pc(vcpu), vcpu->arch.hsr,
599 vcpu->arch.hxfar, fault_ipa);
601 /* Check the stage-2 fault is trans. fault or write fault */
602 fault_status = (vcpu->arch.hsr & HSR_FSC_TYPE);
603 if (fault_status != FSC_FAULT && fault_status != FSC_PERM) {
604 kvm_err("Unsupported fault status: EC=%#lx DFCS=%#lx\n",
605 hsr_ec, fault_status);
609 idx = srcu_read_lock(&vcpu->kvm->srcu);
611 gfn = fault_ipa >> PAGE_SHIFT;
612 if (!kvm_is_visible_gfn(vcpu->kvm, gfn)) {
614 /* Prefetch Abort on I/O address */
615 kvm_inject_pabt(vcpu, vcpu->arch.hxfar);
620 if (fault_status != FSC_FAULT) {
621 kvm_err("Unsupported fault status on io memory: %#lx\n",
627 kvm_pr_unimpl("I/O address abort...");
632 memslot = gfn_to_memslot(vcpu->kvm, gfn);
633 if (!memslot->user_alloc) {
634 kvm_err("non user-alloc memslots not supported\n");
639 ret = user_mem_abort(vcpu, fault_ipa, gfn, memslot, fault_status);
643 srcu_read_unlock(&vcpu->kvm->srcu, idx);
647 static void handle_hva_to_gpa(struct kvm *kvm,
650 void (*handler)(struct kvm *kvm,
651 gpa_t gpa, void *data),
654 struct kvm_memslots *slots;
655 struct kvm_memory_slot *memslot;
657 slots = kvm_memslots(kvm);
659 /* we only care about the pages that the guest sees */
660 kvm_for_each_memslot(memslot, slots) {
661 unsigned long hva_start, hva_end;
664 hva_start = max(start, memslot->userspace_addr);
665 hva_end = min(end, memslot->userspace_addr +
666 (memslot->npages << PAGE_SHIFT));
667 if (hva_start >= hva_end)
671 * {gfn(page) | page intersects with [hva_start, hva_end)} =
672 * {gfn_start, gfn_start+1, ..., gfn_end-1}.
674 gfn = hva_to_gfn_memslot(hva_start, memslot);
675 gfn_end = hva_to_gfn_memslot(hva_end + PAGE_SIZE - 1, memslot);
677 for (; gfn < gfn_end; ++gfn) {
678 gpa_t gpa = gfn << PAGE_SHIFT;
679 handler(kvm, gpa, data);
684 static void kvm_unmap_hva_handler(struct kvm *kvm, gpa_t gpa, void *data)
686 unmap_stage2_range(kvm, gpa, PAGE_SIZE);
687 kvm_tlb_flush_vmid(kvm);
690 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
692 unsigned long end = hva + PAGE_SIZE;
697 trace_kvm_unmap_hva(hva);
698 handle_hva_to_gpa(kvm, hva, end, &kvm_unmap_hva_handler, NULL);
702 int kvm_unmap_hva_range(struct kvm *kvm,
703 unsigned long start, unsigned long end)
708 trace_kvm_unmap_hva_range(start, end);
709 handle_hva_to_gpa(kvm, start, end, &kvm_unmap_hva_handler, NULL);
713 static void kvm_set_spte_handler(struct kvm *kvm, gpa_t gpa, void *data)
715 pte_t *pte = (pte_t *)data;
717 stage2_set_pte(kvm, NULL, gpa, pte, false);
721 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
723 unsigned long end = hva + PAGE_SIZE;
729 trace_kvm_set_spte_hva(hva);
730 stage2_pte = pfn_pte(pte_pfn(pte), PAGE_S2);
731 handle_hva_to_gpa(kvm, hva, end, &kvm_set_spte_handler, &stage2_pte);
734 void kvm_mmu_free_memory_caches(struct kvm_vcpu *vcpu)
736 mmu_free_memory_cache(&vcpu->arch.mmu_page_cache);
739 phys_addr_t kvm_mmu_get_httbr(void)
741 VM_BUG_ON(!virt_addr_valid(hyp_pgd));
742 return virt_to_phys(hyp_pgd);
745 int kvm_mmu_init(void)
748 kvm_err("Hyp mode PGD not allocated\n");
756 * kvm_clear_idmap - remove all idmaps from the hyp pgd
758 * Free the underlying pmds for all pgds in range and clear the pgds (but
759 * don't free them) afterwards.
761 void kvm_clear_hyp_idmap(void)
763 unsigned long addr, end;
765 pgd_t *pgd = hyp_pgd;
769 addr = virt_to_phys(__hyp_idmap_text_start);
770 end = virt_to_phys(__hyp_idmap_text_end);
772 pgd += pgd_index(addr);
774 next = pgd_addr_end(addr, end);
775 if (pgd_none_or_clear_bad(pgd))
777 pud = pud_offset(pgd, addr);
778 pmd = pmd_offset(pud, addr);
781 clean_pmd_entry(pmd);
782 pmd_free(NULL, (pmd_t *)((unsigned long)pmd & PAGE_MASK));
783 } while (pgd++, addr = next, addr < end);