1// SPDX-License-Identifier: GPL-2.0-only
  2/*
  3 * mmu_audit.c:
  4 *
  5 * Audit code for KVM MMU
  6 *
  7 * Copyright (C) 2006 Qumranet, Inc.
  8 * Copyright 2010 Red Hat, Inc. and/or its affiliates.
  9 *
 10 * Authors:
 11 *   Yaniv Kamay  <yaniv@qumranet.com>
 12 *   Avi Kivity   <avi@qumranet.com>
 13 *   Marcelo Tosatti <mtosatti@redhat.com>
 14 *   Xiao Guangrong <xiaoguangrong@cn.fujitsu.com>
 15 */
 16
 17#include <linux/ratelimit.h>
 18
 19static char const *audit_point_name[] = {
 20	"pre page fault",
 21	"post page fault",
 22	"pre pte write",
 23	"post pte write",
 24	"pre sync",
 25	"post sync"
 26};
 27
 28#define audit_printk(kvm, fmt, args...)		\
 29	printk(KERN_ERR "audit: (%s) error: "	\
 30		fmt, audit_point_name[kvm->arch.audit_point], ##args)
 31
 32typedef void (*inspect_spte_fn) (struct kvm_vcpu *vcpu, u64 *sptep, int level);
 33
 34static void __mmu_spte_walk(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
 35			    inspect_spte_fn fn, int level)
 36{
 37	int i;
 38
 39	for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
 40		u64 *ent = sp->spt;
 41
 42		fn(vcpu, ent + i, level);
 43
 44		if (is_shadow_present_pte(ent[i]) &&
 45		      !is_last_spte(ent[i], level)) {
 46			struct kvm_mmu_page *child;
 47
 48			child = page_header(ent[i] & PT64_BASE_ADDR_MASK);
 49			__mmu_spte_walk(vcpu, child, fn, level - 1);
 50		}
 51	}
 52}
 53
 54static void mmu_spte_walk(struct kvm_vcpu *vcpu, inspect_spte_fn fn)
 55{
 56	int i;
 57	struct kvm_mmu_page *sp;
 58
 59	if (!VALID_PAGE(vcpu->arch.mmu->root_hpa))
 60		return;
 61
 62	if (vcpu->arch.mmu->root_level >= PT64_ROOT_4LEVEL) {
 63		hpa_t root = vcpu->arch.mmu->root_hpa;
 64
 65		sp = page_header(root);
 66		__mmu_spte_walk(vcpu, sp, fn, vcpu->arch.mmu->root_level);
 67		return;
 68	}
 69
 70	for (i = 0; i < 4; ++i) {
 71		hpa_t root = vcpu->arch.mmu->pae_root[i];
 72
 73		if (root && VALID_PAGE(root)) {
 74			root &= PT64_BASE_ADDR_MASK;
 75			sp = page_header(root);
 76			__mmu_spte_walk(vcpu, sp, fn, 2);
 77		}
 78	}
 79
 80	return;
 81}
 82
 83typedef void (*sp_handler) (struct kvm *kvm, struct kvm_mmu_page *sp);
 84
 85static void walk_all_active_sps(struct kvm *kvm, sp_handler fn)
 86{
 87	struct kvm_mmu_page *sp;
 88
 89	list_for_each_entry(sp, &kvm->arch.active_mmu_pages, link)
 90		fn(kvm, sp);
 91}
 92
 93static void audit_mappings(struct kvm_vcpu *vcpu, u64 *sptep, int level)
 94{
 95	struct kvm_mmu_page *sp;
 96	gfn_t gfn;
 97	kvm_pfn_t pfn;
 98	hpa_t hpa;
 99
100	sp = page_header(__pa(sptep));
101
102	if (sp->unsync) {
103		if (level != PT_PAGE_TABLE_LEVEL) {
104			audit_printk(vcpu->kvm, "unsync sp: %p "
105				     "level = %d\n", sp, level);
106			return;
107		}
108	}
109
110	if (!is_shadow_present_pte(*sptep) || !is_last_spte(*sptep, level))
111		return;
112
113	gfn = kvm_mmu_page_get_gfn(sp, sptep - sp->spt);
114	pfn = kvm_vcpu_gfn_to_pfn_atomic(vcpu, gfn);
115
116	if (is_error_pfn(pfn))
117		return;
118
119	hpa =  pfn << PAGE_SHIFT;
120	if ((*sptep & PT64_BASE_ADDR_MASK) != hpa)
121		audit_printk(vcpu->kvm, "levels %d pfn %llx hpa %llx "
122			     "ent %llxn", vcpu->arch.mmu->root_level, pfn,
123			     hpa, *sptep);
124}
125
126static void inspect_spte_has_rmap(struct kvm *kvm, u64 *sptep)
127{
128	static DEFINE_RATELIMIT_STATE(ratelimit_state, 5 * HZ, 10);
129	struct kvm_rmap_head *rmap_head;
130	struct kvm_mmu_page *rev_sp;
131	struct kvm_memslots *slots;
132	struct kvm_memory_slot *slot;
133	gfn_t gfn;
134
135	rev_sp = page_header(__pa(sptep));
136	gfn = kvm_mmu_page_get_gfn(rev_sp, sptep - rev_sp->spt);
137
138	slots = kvm_memslots_for_spte_role(kvm, rev_sp->role);
139	slot = __gfn_to_memslot(slots, gfn);
140	if (!slot) {
141		if (!__ratelimit(&ratelimit_state))
142			return;
143		audit_printk(kvm, "no memslot for gfn %llx\n", gfn);
144		audit_printk(kvm, "index %ld of sp (gfn=%llx)\n",
145		       (long int)(sptep - rev_sp->spt), rev_sp->gfn);
146		dump_stack();
147		return;
148	}
149
150	rmap_head = __gfn_to_rmap(gfn, rev_sp->role.level, slot);
151	if (!rmap_head->val) {
152		if (!__ratelimit(&ratelimit_state))
153			return;
154		audit_printk(kvm, "no rmap for writable spte %llx\n",
155			     *sptep);
156		dump_stack();
157	}
158}
159
160static void audit_sptes_have_rmaps(struct kvm_vcpu *vcpu, u64 *sptep, int level)
161{
162	if (is_shadow_present_pte(*sptep) && is_last_spte(*sptep, level))
163		inspect_spte_has_rmap(vcpu->kvm, sptep);
164}
165
166static void audit_spte_after_sync(struct kvm_vcpu *vcpu, u64 *sptep, int level)
167{
168	struct kvm_mmu_page *sp = page_header(__pa(sptep));
169
170	if (vcpu->kvm->arch.audit_point == AUDIT_POST_SYNC && sp->unsync)
171		audit_printk(vcpu->kvm, "meet unsync sp(%p) after sync "
172			     "root.\n", sp);
173}
174
175static void check_mappings_rmap(struct kvm *kvm, struct kvm_mmu_page *sp)
176{
177	int i;
178
179	if (sp->role.level != PT_PAGE_TABLE_LEVEL)
180		return;
181
182	for (i = 0; i < PT64_ENT_PER_PAGE; ++i) {
183		if (!is_shadow_present_pte(sp->spt[i]))
184			continue;
185
186		inspect_spte_has_rmap(kvm, sp->spt + i);
187	}
188}
189
190static void audit_write_protection(struct kvm *kvm, struct kvm_mmu_page *sp)
191{
192	struct kvm_rmap_head *rmap_head;
193	u64 *sptep;
194	struct rmap_iterator iter;
195	struct kvm_memslots *slots;
196	struct kvm_memory_slot *slot;
197
198	if (sp->role.direct || sp->unsync || sp->role.invalid)
199		return;
200
201	slots = kvm_memslots_for_spte_role(kvm, sp->role);
202	slot = __gfn_to_memslot(slots, sp->gfn);
203	rmap_head = __gfn_to_rmap(sp->gfn, PT_PAGE_TABLE_LEVEL, slot);
204
205	for_each_rmap_spte(rmap_head, &iter, sptep) {
206		if (is_writable_pte(*sptep))
207			audit_printk(kvm, "shadow page has writable "
208				     "mappings: gfn %llx role %x\n",
209				     sp->gfn, sp->role.word);
210	}
211}
212
213static void audit_sp(struct kvm *kvm, struct kvm_mmu_page *sp)
214{
215	check_mappings_rmap(kvm, sp);
216	audit_write_protection(kvm, sp);
217}
218
219static void audit_all_active_sps(struct kvm *kvm)
220{
221	walk_all_active_sps(kvm, audit_sp);
222}
223
224static void audit_spte(struct kvm_vcpu *vcpu, u64 *sptep, int level)
225{
226	audit_sptes_have_rmaps(vcpu, sptep, level);
227	audit_mappings(vcpu, sptep, level);
228	audit_spte_after_sync(vcpu, sptep, level);
229}
230
231static void audit_vcpu_spte(struct kvm_vcpu *vcpu)
232{
233	mmu_spte_walk(vcpu, audit_spte);
234}
235
236static bool mmu_audit;
237static struct static_key mmu_audit_key;
238
239static void __kvm_mmu_audit(struct kvm_vcpu *vcpu, int point)
240{
241	static DEFINE_RATELIMIT_STATE(ratelimit_state, 5 * HZ, 10);
242
243	if (!__ratelimit(&ratelimit_state))
244		return;
245
246	vcpu->kvm->arch.audit_point = point;
247	audit_all_active_sps(vcpu->kvm);
248	audit_vcpu_spte(vcpu);
249}
250
251static inline void kvm_mmu_audit(struct kvm_vcpu *vcpu, int point)
252{
253	if (static_key_false((&mmu_audit_key)))
254		__kvm_mmu_audit(vcpu, point);
255}
256
257static void mmu_audit_enable(void)
258{
259	if (mmu_audit)
260		return;
261
262	static_key_slow_inc(&mmu_audit_key);
263	mmu_audit = true;
264}
265
266static void mmu_audit_disable(void)
267{
268	if (!mmu_audit)
269		return;
270
271	static_key_slow_dec(&mmu_audit_key);
272	mmu_audit = false;
273}
274
275static int mmu_audit_set(const char *val, const struct kernel_param *kp)
276{
277	int ret;
278	unsigned long enable;
279
280	ret = kstrtoul(val, 10, &enable);
281	if (ret < 0)
282		return -EINVAL;
283
284	switch (enable) {
285	case 0:
286		mmu_audit_disable();
287		break;
288	case 1:
289		mmu_audit_enable();
290		break;
291	default:
292		return -EINVAL;
293	}
294
295	return 0;
296}
297
298static const struct kernel_param_ops audit_param_ops = {
299	.set = mmu_audit_set,
300	.get = param_get_bool,
301};
302
303arch_param_cb(mmu_audit, &audit_param_ops, &mmu_audit, 0644);