xref: /openbmc/linux/arch/x86/kvm/mmu/spte.c (revision 3d37ef41)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Kernel-based Virtual Machine driver for Linux
4  *
5  * Macros and functions to access KVM PTEs (also known as SPTEs)
6  *
7  * Copyright (C) 2006 Qumranet, Inc.
8  * Copyright 2020 Red Hat, Inc. and/or its affiliates.
9  */
10 
11 
12 #include <linux/kvm_host.h>
13 #include "mmu.h"
14 #include "mmu_internal.h"
15 #include "x86.h"
16 #include "spte.h"
17 
18 #include <asm/e820/api.h>
19 #include <asm/vmx.h>
20 
21 static bool __read_mostly enable_mmio_caching = true;
22 module_param_named(mmio_caching, enable_mmio_caching, bool, 0444);
23 
24 u64 __read_mostly shadow_host_writable_mask;
25 u64 __read_mostly shadow_mmu_writable_mask;
26 u64 __read_mostly shadow_nx_mask;
27 u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
28 u64 __read_mostly shadow_user_mask;
29 u64 __read_mostly shadow_accessed_mask;
30 u64 __read_mostly shadow_dirty_mask;
31 u64 __read_mostly shadow_mmio_value;
32 u64 __read_mostly shadow_mmio_mask;
33 u64 __read_mostly shadow_mmio_access_mask;
34 u64 __read_mostly shadow_present_mask;
35 u64 __read_mostly shadow_me_mask;
36 u64 __read_mostly shadow_acc_track_mask;
37 
38 u64 __read_mostly shadow_nonpresent_or_rsvd_mask;
39 u64 __read_mostly shadow_nonpresent_or_rsvd_lower_gfn_mask;
40 
41 u8 __read_mostly shadow_phys_bits;
42 
43 static u64 generation_mmio_spte_mask(u64 gen)
44 {
45 	u64 mask;
46 
47 	WARN_ON(gen & ~MMIO_SPTE_GEN_MASK);
48 
49 	mask = (gen << MMIO_SPTE_GEN_LOW_SHIFT) & MMIO_SPTE_GEN_LOW_MASK;
50 	mask |= (gen << MMIO_SPTE_GEN_HIGH_SHIFT) & MMIO_SPTE_GEN_HIGH_MASK;
51 	return mask;
52 }
53 
54 u64 make_mmio_spte(struct kvm_vcpu *vcpu, u64 gfn, unsigned int access)
55 {
56 	u64 gen = kvm_vcpu_memslots(vcpu)->generation & MMIO_SPTE_GEN_MASK;
57 	u64 spte = generation_mmio_spte_mask(gen);
58 	u64 gpa = gfn << PAGE_SHIFT;
59 
60 	WARN_ON_ONCE(!shadow_mmio_value);
61 
62 	access &= shadow_mmio_access_mask;
63 	spte |= shadow_mmio_value | access;
64 	spte |= gpa | shadow_nonpresent_or_rsvd_mask;
65 	spte |= (gpa & shadow_nonpresent_or_rsvd_mask)
66 		<< SHADOW_NONPRESENT_OR_RSVD_MASK_LEN;
67 
68 	return spte;
69 }
70 
71 static bool kvm_is_mmio_pfn(kvm_pfn_t pfn)
72 {
73 	if (pfn_valid(pfn))
74 		return !is_zero_pfn(pfn) && PageReserved(pfn_to_page(pfn)) &&
75 			/*
76 			 * Some reserved pages, such as those from NVDIMM
77 			 * DAX devices, are not for MMIO, and can be mapped
78 			 * with cached memory type for better performance.
79 			 * However, the above check misconceives those pages
80 			 * as MMIO, and results in KVM mapping them with UC
81 			 * memory type, which would hurt the performance.
82 			 * Therefore, we check the host memory type in addition
83 			 * and only treat UC/UC-/WC pages as MMIO.
84 			 */
85 			(!pat_enabled() || pat_pfn_immune_to_uc_mtrr(pfn));
86 
87 	return !e820__mapped_raw_any(pfn_to_hpa(pfn),
88 				     pfn_to_hpa(pfn + 1) - 1,
89 				     E820_TYPE_RAM);
90 }
91 
92 int make_spte(struct kvm_vcpu *vcpu, unsigned int pte_access, int level,
93 		     gfn_t gfn, kvm_pfn_t pfn, u64 old_spte, bool speculative,
94 		     bool can_unsync, bool host_writable, bool ad_disabled,
95 		     u64 *new_spte)
96 {
97 	u64 spte = SPTE_MMU_PRESENT_MASK;
98 	int ret = 0;
99 
100 	if (ad_disabled)
101 		spte |= SPTE_TDP_AD_DISABLED_MASK;
102 	else if (kvm_vcpu_ad_need_write_protect(vcpu))
103 		spte |= SPTE_TDP_AD_WRPROT_ONLY_MASK;
104 
105 	/*
106 	 * Bits 62:52 of PAE SPTEs are reserved.  WARN if said bits are set
107 	 * if PAE paging may be employed (shadow paging or any 32-bit KVM).
108 	 */
109 	WARN_ON_ONCE((!tdp_enabled || !IS_ENABLED(CONFIG_X86_64)) &&
110 		     (spte & SPTE_TDP_AD_MASK));
111 
112 	/*
113 	 * For the EPT case, shadow_present_mask is 0 if hardware
114 	 * supports exec-only page table entries.  In that case,
115 	 * ACC_USER_MASK and shadow_user_mask are used to represent
116 	 * read access.  See FNAME(gpte_access) in paging_tmpl.h.
117 	 */
118 	spte |= shadow_present_mask;
119 	if (!speculative)
120 		spte |= spte_shadow_accessed_mask(spte);
121 
122 	if (level > PG_LEVEL_4K && (pte_access & ACC_EXEC_MASK) &&
123 	    is_nx_huge_page_enabled()) {
124 		pte_access &= ~ACC_EXEC_MASK;
125 	}
126 
127 	if (pte_access & ACC_EXEC_MASK)
128 		spte |= shadow_x_mask;
129 	else
130 		spte |= shadow_nx_mask;
131 
132 	if (pte_access & ACC_USER_MASK)
133 		spte |= shadow_user_mask;
134 
135 	if (level > PG_LEVEL_4K)
136 		spte |= PT_PAGE_SIZE_MASK;
137 	if (tdp_enabled)
138 		spte |= static_call(kvm_x86_get_mt_mask)(vcpu, gfn,
139 			kvm_is_mmio_pfn(pfn));
140 
141 	if (host_writable)
142 		spte |= shadow_host_writable_mask;
143 	else
144 		pte_access &= ~ACC_WRITE_MASK;
145 
146 	if (!kvm_is_mmio_pfn(pfn))
147 		spte |= shadow_me_mask;
148 
149 	spte |= (u64)pfn << PAGE_SHIFT;
150 
151 	if (pte_access & ACC_WRITE_MASK) {
152 		spte |= PT_WRITABLE_MASK | shadow_mmu_writable_mask;
153 
154 		/*
155 		 * Optimization: for pte sync, if spte was writable the hash
156 		 * lookup is unnecessary (and expensive). Write protection
157 		 * is responsibility of mmu_get_page / kvm_sync_page.
158 		 * Same reasoning can be applied to dirty page accounting.
159 		 */
160 		if (!can_unsync && is_writable_pte(old_spte))
161 			goto out;
162 
163 		if (mmu_need_write_protect(vcpu, gfn, can_unsync)) {
164 			pgprintk("%s: found shadow page for %llx, marking ro\n",
165 				 __func__, gfn);
166 			ret |= SET_SPTE_WRITE_PROTECTED_PT;
167 			pte_access &= ~ACC_WRITE_MASK;
168 			spte &= ~(PT_WRITABLE_MASK | shadow_mmu_writable_mask);
169 		}
170 	}
171 
172 	if (pte_access & ACC_WRITE_MASK)
173 		spte |= spte_shadow_dirty_mask(spte);
174 
175 	if (speculative)
176 		spte = mark_spte_for_access_track(spte);
177 
178 out:
179 	WARN_ON(is_mmio_spte(spte));
180 	*new_spte = spte;
181 	return ret;
182 }
183 
184 u64 make_nonleaf_spte(u64 *child_pt, bool ad_disabled)
185 {
186 	u64 spte = SPTE_MMU_PRESENT_MASK;
187 
188 	spte |= __pa(child_pt) | shadow_present_mask | PT_WRITABLE_MASK |
189 		shadow_user_mask | shadow_x_mask | shadow_me_mask;
190 
191 	if (ad_disabled)
192 		spte |= SPTE_TDP_AD_DISABLED_MASK;
193 	else
194 		spte |= shadow_accessed_mask;
195 
196 	return spte;
197 }
198 
199 u64 kvm_mmu_changed_pte_notifier_make_spte(u64 old_spte, kvm_pfn_t new_pfn)
200 {
201 	u64 new_spte;
202 
203 	new_spte = old_spte & ~PT64_BASE_ADDR_MASK;
204 	new_spte |= (u64)new_pfn << PAGE_SHIFT;
205 
206 	new_spte &= ~PT_WRITABLE_MASK;
207 	new_spte &= ~shadow_host_writable_mask;
208 
209 	new_spte = mark_spte_for_access_track(new_spte);
210 
211 	return new_spte;
212 }
213 
214 static u8 kvm_get_shadow_phys_bits(void)
215 {
216 	/*
217 	 * boot_cpu_data.x86_phys_bits is reduced when MKTME or SME are detected
218 	 * in CPU detection code, but the processor treats those reduced bits as
219 	 * 'keyID' thus they are not reserved bits. Therefore KVM needs to look at
220 	 * the physical address bits reported by CPUID.
221 	 */
222 	if (likely(boot_cpu_data.extended_cpuid_level >= 0x80000008))
223 		return cpuid_eax(0x80000008) & 0xff;
224 
225 	/*
226 	 * Quite weird to have VMX or SVM but not MAXPHYADDR; probably a VM with
227 	 * custom CPUID.  Proceed with whatever the kernel found since these features
228 	 * aren't virtualizable (SME/SEV also require CPUIDs higher than 0x80000008).
229 	 */
230 	return boot_cpu_data.x86_phys_bits;
231 }
232 
233 u64 mark_spte_for_access_track(u64 spte)
234 {
235 	if (spte_ad_enabled(spte))
236 		return spte & ~shadow_accessed_mask;
237 
238 	if (is_access_track_spte(spte))
239 		return spte;
240 
241 	/*
242 	 * Making an Access Tracking PTE will result in removal of write access
243 	 * from the PTE. So, verify that we will be able to restore the write
244 	 * access in the fast page fault path later on.
245 	 */
246 	WARN_ONCE((spte & PT_WRITABLE_MASK) &&
247 		  !spte_can_locklessly_be_made_writable(spte),
248 		  "kvm: Writable SPTE is not locklessly dirty-trackable\n");
249 
250 	WARN_ONCE(spte & (SHADOW_ACC_TRACK_SAVED_BITS_MASK <<
251 			  SHADOW_ACC_TRACK_SAVED_BITS_SHIFT),
252 		  "kvm: Access Tracking saved bit locations are not zero\n");
253 
254 	spte |= (spte & SHADOW_ACC_TRACK_SAVED_BITS_MASK) <<
255 		SHADOW_ACC_TRACK_SAVED_BITS_SHIFT;
256 	spte &= ~shadow_acc_track_mask;
257 
258 	return spte;
259 }
260 
261 void kvm_mmu_set_mmio_spte_mask(u64 mmio_value, u64 mmio_mask, u64 access_mask)
262 {
263 	BUG_ON((u64)(unsigned)access_mask != access_mask);
264 	WARN_ON(mmio_value & shadow_nonpresent_or_rsvd_lower_gfn_mask);
265 
266 	if (!enable_mmio_caching)
267 		mmio_value = 0;
268 
269 	/*
270 	 * Disable MMIO caching if the MMIO value collides with the bits that
271 	 * are used to hold the relocated GFN when the L1TF mitigation is
272 	 * enabled.  This should never fire as there is no known hardware that
273 	 * can trigger this condition, e.g. SME/SEV CPUs that require a custom
274 	 * MMIO value are not susceptible to L1TF.
275 	 */
276 	if (WARN_ON(mmio_value & (shadow_nonpresent_or_rsvd_mask <<
277 				  SHADOW_NONPRESENT_OR_RSVD_MASK_LEN)))
278 		mmio_value = 0;
279 
280 	/*
281 	 * The masked MMIO value must obviously match itself and a removed SPTE
282 	 * must not get a false positive.  Removed SPTEs and MMIO SPTEs should
283 	 * never collide as MMIO must set some RWX bits, and removed SPTEs must
284 	 * not set any RWX bits.
285 	 */
286 	if (WARN_ON((mmio_value & mmio_mask) != mmio_value) ||
287 	    WARN_ON(mmio_value && (REMOVED_SPTE & mmio_mask) == mmio_value))
288 		mmio_value = 0;
289 
290 	shadow_mmio_value = mmio_value;
291 	shadow_mmio_mask  = mmio_mask;
292 	shadow_mmio_access_mask = access_mask;
293 }
294 EXPORT_SYMBOL_GPL(kvm_mmu_set_mmio_spte_mask);
295 
296 void kvm_mmu_set_ept_masks(bool has_ad_bits, bool has_exec_only)
297 {
298 	shadow_user_mask	= VMX_EPT_READABLE_MASK;
299 	shadow_accessed_mask	= has_ad_bits ? VMX_EPT_ACCESS_BIT : 0ull;
300 	shadow_dirty_mask	= has_ad_bits ? VMX_EPT_DIRTY_BIT : 0ull;
301 	shadow_nx_mask		= 0ull;
302 	shadow_x_mask		= VMX_EPT_EXECUTABLE_MASK;
303 	shadow_present_mask	= has_exec_only ? 0ull : VMX_EPT_READABLE_MASK;
304 	shadow_acc_track_mask	= VMX_EPT_RWX_MASK;
305 	shadow_me_mask		= 0ull;
306 
307 	shadow_host_writable_mask = EPT_SPTE_HOST_WRITABLE;
308 	shadow_mmu_writable_mask  = EPT_SPTE_MMU_WRITABLE;
309 
310 	/*
311 	 * EPT Misconfigurations are generated if the value of bits 2:0
312 	 * of an EPT paging-structure entry is 110b (write/execute).
313 	 */
314 	kvm_mmu_set_mmio_spte_mask(VMX_EPT_MISCONFIG_WX_VALUE,
315 				   VMX_EPT_RWX_MASK, 0);
316 }
317 EXPORT_SYMBOL_GPL(kvm_mmu_set_ept_masks);
318 
319 void kvm_mmu_reset_all_pte_masks(void)
320 {
321 	u8 low_phys_bits;
322 	u64 mask;
323 
324 	shadow_phys_bits = kvm_get_shadow_phys_bits();
325 
326 	/*
327 	 * If the CPU has 46 or less physical address bits, then set an
328 	 * appropriate mask to guard against L1TF attacks. Otherwise, it is
329 	 * assumed that the CPU is not vulnerable to L1TF.
330 	 *
331 	 * Some Intel CPUs address the L1 cache using more PA bits than are
332 	 * reported by CPUID. Use the PA width of the L1 cache when possible
333 	 * to achieve more effective mitigation, e.g. if system RAM overlaps
334 	 * the most significant bits of legal physical address space.
335 	 */
336 	shadow_nonpresent_or_rsvd_mask = 0;
337 	low_phys_bits = boot_cpu_data.x86_phys_bits;
338 	if (boot_cpu_has_bug(X86_BUG_L1TF) &&
339 	    !WARN_ON_ONCE(boot_cpu_data.x86_cache_bits >=
340 			  52 - SHADOW_NONPRESENT_OR_RSVD_MASK_LEN)) {
341 		low_phys_bits = boot_cpu_data.x86_cache_bits
342 			- SHADOW_NONPRESENT_OR_RSVD_MASK_LEN;
343 		shadow_nonpresent_or_rsvd_mask =
344 			rsvd_bits(low_phys_bits, boot_cpu_data.x86_cache_bits - 1);
345 	}
346 
347 	shadow_nonpresent_or_rsvd_lower_gfn_mask =
348 		GENMASK_ULL(low_phys_bits - 1, PAGE_SHIFT);
349 
350 	shadow_user_mask	= PT_USER_MASK;
351 	shadow_accessed_mask	= PT_ACCESSED_MASK;
352 	shadow_dirty_mask	= PT_DIRTY_MASK;
353 	shadow_nx_mask		= PT64_NX_MASK;
354 	shadow_x_mask		= 0;
355 	shadow_present_mask	= PT_PRESENT_MASK;
356 	shadow_acc_track_mask	= 0;
357 	shadow_me_mask		= sme_me_mask;
358 
359 	shadow_host_writable_mask = DEFAULT_SPTE_HOST_WRITEABLE;
360 	shadow_mmu_writable_mask  = DEFAULT_SPTE_MMU_WRITEABLE;
361 
362 	/*
363 	 * Set a reserved PA bit in MMIO SPTEs to generate page faults with
364 	 * PFEC.RSVD=1 on MMIO accesses.  64-bit PTEs (PAE, x86-64, and EPT
365 	 * paging) support a maximum of 52 bits of PA, i.e. if the CPU supports
366 	 * 52-bit physical addresses then there are no reserved PA bits in the
367 	 * PTEs and so the reserved PA approach must be disabled.
368 	 */
369 	if (shadow_phys_bits < 52)
370 		mask = BIT_ULL(51) | PT_PRESENT_MASK;
371 	else
372 		mask = 0;
373 
374 	kvm_mmu_set_mmio_spte_mask(mask, mask, ACC_WRITE_MASK | ACC_USER_MASK);
375 }
376