xref: /openbmc/linux/arch/x86/kvm/mmu/spte.c (revision fbb6b31a)
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/memtype.h>
20 #include <asm/vmx.h>
21 
22 static bool __read_mostly enable_mmio_caching = true;
23 module_param_named(mmio_caching, enable_mmio_caching, bool, 0444);
24 
25 u64 __read_mostly shadow_host_writable_mask;
26 u64 __read_mostly shadow_mmu_writable_mask;
27 u64 __read_mostly shadow_nx_mask;
28 u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
29 u64 __read_mostly shadow_user_mask;
30 u64 __read_mostly shadow_accessed_mask;
31 u64 __read_mostly shadow_dirty_mask;
32 u64 __read_mostly shadow_mmio_value;
33 u64 __read_mostly shadow_mmio_mask;
34 u64 __read_mostly shadow_mmio_access_mask;
35 u64 __read_mostly shadow_present_mask;
36 u64 __read_mostly shadow_me_mask;
37 u64 __read_mostly shadow_acc_track_mask;
38 
39 u64 __read_mostly shadow_nonpresent_or_rsvd_mask;
40 u64 __read_mostly shadow_nonpresent_or_rsvd_lower_gfn_mask;
41 
42 u8 __read_mostly shadow_phys_bits;
43 
44 static u64 generation_mmio_spte_mask(u64 gen)
45 {
46 	u64 mask;
47 
48 	WARN_ON(gen & ~MMIO_SPTE_GEN_MASK);
49 
50 	mask = (gen << MMIO_SPTE_GEN_LOW_SHIFT) & MMIO_SPTE_GEN_LOW_MASK;
51 	mask |= (gen << MMIO_SPTE_GEN_HIGH_SHIFT) & MMIO_SPTE_GEN_HIGH_MASK;
52 	return mask;
53 }
54 
55 u64 make_mmio_spte(struct kvm_vcpu *vcpu, u64 gfn, unsigned int access)
56 {
57 	u64 gen = kvm_vcpu_memslots(vcpu)->generation & MMIO_SPTE_GEN_MASK;
58 	u64 spte = generation_mmio_spte_mask(gen);
59 	u64 gpa = gfn << PAGE_SHIFT;
60 
61 	WARN_ON_ONCE(!shadow_mmio_value);
62 
63 	access &= shadow_mmio_access_mask;
64 	spte |= shadow_mmio_value | access;
65 	spte |= gpa | shadow_nonpresent_or_rsvd_mask;
66 	spte |= (gpa & shadow_nonpresent_or_rsvd_mask)
67 		<< SHADOW_NONPRESENT_OR_RSVD_MASK_LEN;
68 
69 	return spte;
70 }
71 
72 static bool kvm_is_mmio_pfn(kvm_pfn_t pfn)
73 {
74 	if (pfn_valid(pfn))
75 		return !is_zero_pfn(pfn) && PageReserved(pfn_to_page(pfn)) &&
76 			/*
77 			 * Some reserved pages, such as those from NVDIMM
78 			 * DAX devices, are not for MMIO, and can be mapped
79 			 * with cached memory type for better performance.
80 			 * However, the above check misconceives those pages
81 			 * as MMIO, and results in KVM mapping them with UC
82 			 * memory type, which would hurt the performance.
83 			 * Therefore, we check the host memory type in addition
84 			 * and only treat UC/UC-/WC pages as MMIO.
85 			 */
86 			(!pat_enabled() || pat_pfn_immune_to_uc_mtrr(pfn));
87 
88 	return !e820__mapped_raw_any(pfn_to_hpa(pfn),
89 				     pfn_to_hpa(pfn + 1) - 1,
90 				     E820_TYPE_RAM);
91 }
92 
93 /*
94  * Returns true if the SPTE has bits that may be set without holding mmu_lock.
95  * The caller is responsible for checking if the SPTE is shadow-present, and
96  * for determining whether or not the caller cares about non-leaf SPTEs.
97  */
98 bool spte_has_volatile_bits(u64 spte)
99 {
100 	/*
101 	 * Always atomically update spte if it can be updated
102 	 * out of mmu-lock, it can ensure dirty bit is not lost,
103 	 * also, it can help us to get a stable is_writable_pte()
104 	 * to ensure tlb flush is not missed.
105 	 */
106 	if (!is_writable_pte(spte) && is_mmu_writable_spte(spte))
107 		return true;
108 
109 	if (is_access_track_spte(spte))
110 		return true;
111 
112 	if (spte_ad_enabled(spte)) {
113 		if (!(spte & shadow_accessed_mask) ||
114 		    (is_writable_pte(spte) && !(spte & shadow_dirty_mask)))
115 			return true;
116 	}
117 
118 	return false;
119 }
120 
121 bool make_spte(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
122 	       const struct kvm_memory_slot *slot,
123 	       unsigned int pte_access, gfn_t gfn, kvm_pfn_t pfn,
124 	       u64 old_spte, bool prefetch, bool can_unsync,
125 	       bool host_writable, u64 *new_spte)
126 {
127 	int level = sp->role.level;
128 	u64 spte = SPTE_MMU_PRESENT_MASK;
129 	bool wrprot = false;
130 
131 	if (sp->role.ad_disabled)
132 		spte |= SPTE_TDP_AD_DISABLED_MASK;
133 	else if (kvm_mmu_page_ad_need_write_protect(sp))
134 		spte |= SPTE_TDP_AD_WRPROT_ONLY_MASK;
135 
136 	/*
137 	 * For the EPT case, shadow_present_mask is 0 if hardware
138 	 * supports exec-only page table entries.  In that case,
139 	 * ACC_USER_MASK and shadow_user_mask are used to represent
140 	 * read access.  See FNAME(gpte_access) in paging_tmpl.h.
141 	 */
142 	spte |= shadow_present_mask;
143 	if (!prefetch)
144 		spte |= spte_shadow_accessed_mask(spte);
145 
146 	if (level > PG_LEVEL_4K && (pte_access & ACC_EXEC_MASK) &&
147 	    is_nx_huge_page_enabled()) {
148 		pte_access &= ~ACC_EXEC_MASK;
149 	}
150 
151 	if (pte_access & ACC_EXEC_MASK)
152 		spte |= shadow_x_mask;
153 	else
154 		spte |= shadow_nx_mask;
155 
156 	if (pte_access & ACC_USER_MASK)
157 		spte |= shadow_user_mask;
158 
159 	if (level > PG_LEVEL_4K)
160 		spte |= PT_PAGE_SIZE_MASK;
161 	if (tdp_enabled)
162 		spte |= static_call(kvm_x86_get_mt_mask)(vcpu, gfn,
163 			kvm_is_mmio_pfn(pfn));
164 
165 	if (host_writable)
166 		spte |= shadow_host_writable_mask;
167 	else
168 		pte_access &= ~ACC_WRITE_MASK;
169 
170 	if (!kvm_is_mmio_pfn(pfn))
171 		spte |= shadow_me_mask;
172 
173 	spte |= (u64)pfn << PAGE_SHIFT;
174 
175 	if (pte_access & ACC_WRITE_MASK) {
176 		spte |= PT_WRITABLE_MASK | shadow_mmu_writable_mask;
177 
178 		/*
179 		 * Optimization: for pte sync, if spte was writable the hash
180 		 * lookup is unnecessary (and expensive). Write protection
181 		 * is responsibility of kvm_mmu_get_page / kvm_mmu_sync_roots.
182 		 * Same reasoning can be applied to dirty page accounting.
183 		 */
184 		if (is_writable_pte(old_spte))
185 			goto out;
186 
187 		/*
188 		 * Unsync shadow pages that are reachable by the new, writable
189 		 * SPTE.  Write-protect the SPTE if the page can't be unsync'd,
190 		 * e.g. it's write-tracked (upper-level SPs) or has one or more
191 		 * shadow pages and unsync'ing pages is not allowed.
192 		 */
193 		if (mmu_try_to_unsync_pages(vcpu->kvm, slot, gfn, can_unsync, prefetch)) {
194 			pgprintk("%s: found shadow page for %llx, marking ro\n",
195 				 __func__, gfn);
196 			wrprot = true;
197 			pte_access &= ~ACC_WRITE_MASK;
198 			spte &= ~(PT_WRITABLE_MASK | shadow_mmu_writable_mask);
199 		}
200 	}
201 
202 	if (pte_access & ACC_WRITE_MASK)
203 		spte |= spte_shadow_dirty_mask(spte);
204 
205 out:
206 	if (prefetch)
207 		spte = mark_spte_for_access_track(spte);
208 
209 	WARN_ONCE(is_rsvd_spte(&vcpu->arch.mmu->shadow_zero_check, spte, level),
210 		  "spte = 0x%llx, level = %d, rsvd bits = 0x%llx", spte, level,
211 		  get_rsvd_bits(&vcpu->arch.mmu->shadow_zero_check, spte, level));
212 
213 	if ((spte & PT_WRITABLE_MASK) && kvm_slot_dirty_track_enabled(slot)) {
214 		/* Enforced by kvm_mmu_hugepage_adjust. */
215 		WARN_ON(level > PG_LEVEL_4K);
216 		mark_page_dirty_in_slot(vcpu->kvm, slot, gfn);
217 	}
218 
219 	*new_spte = spte;
220 	return wrprot;
221 }
222 
223 static u64 make_spte_executable(u64 spte)
224 {
225 	bool is_access_track = is_access_track_spte(spte);
226 
227 	if (is_access_track)
228 		spte = restore_acc_track_spte(spte);
229 
230 	spte &= ~shadow_nx_mask;
231 	spte |= shadow_x_mask;
232 
233 	if (is_access_track)
234 		spte = mark_spte_for_access_track(spte);
235 
236 	return spte;
237 }
238 
239 /*
240  * Construct an SPTE that maps a sub-page of the given huge page SPTE where
241  * `index` identifies which sub-page.
242  *
243  * This is used during huge page splitting to build the SPTEs that make up the
244  * new page table.
245  */
246 u64 make_huge_page_split_spte(u64 huge_spte, int huge_level, int index)
247 {
248 	u64 child_spte;
249 	int child_level;
250 
251 	if (WARN_ON_ONCE(!is_shadow_present_pte(huge_spte)))
252 		return 0;
253 
254 	if (WARN_ON_ONCE(!is_large_pte(huge_spte)))
255 		return 0;
256 
257 	child_spte = huge_spte;
258 	child_level = huge_level - 1;
259 
260 	/*
261 	 * The child_spte already has the base address of the huge page being
262 	 * split. So we just have to OR in the offset to the page at the next
263 	 * lower level for the given index.
264 	 */
265 	child_spte |= (index * KVM_PAGES_PER_HPAGE(child_level)) << PAGE_SHIFT;
266 
267 	if (child_level == PG_LEVEL_4K) {
268 		child_spte &= ~PT_PAGE_SIZE_MASK;
269 
270 		/*
271 		 * When splitting to a 4K page, mark the page executable as the
272 		 * NX hugepage mitigation no longer applies.
273 		 */
274 		if (is_nx_huge_page_enabled())
275 			child_spte = make_spte_executable(child_spte);
276 	}
277 
278 	return child_spte;
279 }
280 
281 
282 u64 make_nonleaf_spte(u64 *child_pt, bool ad_disabled)
283 {
284 	u64 spte = SPTE_MMU_PRESENT_MASK;
285 
286 	spte |= __pa(child_pt) | shadow_present_mask | PT_WRITABLE_MASK |
287 		shadow_user_mask | shadow_x_mask | shadow_me_mask;
288 
289 	if (ad_disabled)
290 		spte |= SPTE_TDP_AD_DISABLED_MASK;
291 	else
292 		spte |= shadow_accessed_mask;
293 
294 	return spte;
295 }
296 
297 u64 kvm_mmu_changed_pte_notifier_make_spte(u64 old_spte, kvm_pfn_t new_pfn)
298 {
299 	u64 new_spte;
300 
301 	new_spte = old_spte & ~PT64_BASE_ADDR_MASK;
302 	new_spte |= (u64)new_pfn << PAGE_SHIFT;
303 
304 	new_spte &= ~PT_WRITABLE_MASK;
305 	new_spte &= ~shadow_host_writable_mask;
306 	new_spte &= ~shadow_mmu_writable_mask;
307 
308 	new_spte = mark_spte_for_access_track(new_spte);
309 
310 	return new_spte;
311 }
312 
313 static u8 kvm_get_shadow_phys_bits(void)
314 {
315 	/*
316 	 * boot_cpu_data.x86_phys_bits is reduced when MKTME or SME are detected
317 	 * in CPU detection code, but the processor treats those reduced bits as
318 	 * 'keyID' thus they are not reserved bits. Therefore KVM needs to look at
319 	 * the physical address bits reported by CPUID.
320 	 */
321 	if (likely(boot_cpu_data.extended_cpuid_level >= 0x80000008))
322 		return cpuid_eax(0x80000008) & 0xff;
323 
324 	/*
325 	 * Quite weird to have VMX or SVM but not MAXPHYADDR; probably a VM with
326 	 * custom CPUID.  Proceed with whatever the kernel found since these features
327 	 * aren't virtualizable (SME/SEV also require CPUIDs higher than 0x80000008).
328 	 */
329 	return boot_cpu_data.x86_phys_bits;
330 }
331 
332 u64 mark_spte_for_access_track(u64 spte)
333 {
334 	if (spte_ad_enabled(spte))
335 		return spte & ~shadow_accessed_mask;
336 
337 	if (is_access_track_spte(spte))
338 		return spte;
339 
340 	check_spte_writable_invariants(spte);
341 
342 	WARN_ONCE(spte & (SHADOW_ACC_TRACK_SAVED_BITS_MASK <<
343 			  SHADOW_ACC_TRACK_SAVED_BITS_SHIFT),
344 		  "kvm: Access Tracking saved bit locations are not zero\n");
345 
346 	spte |= (spte & SHADOW_ACC_TRACK_SAVED_BITS_MASK) <<
347 		SHADOW_ACC_TRACK_SAVED_BITS_SHIFT;
348 	spte &= ~shadow_acc_track_mask;
349 
350 	return spte;
351 }
352 
353 void kvm_mmu_set_mmio_spte_mask(u64 mmio_value, u64 mmio_mask, u64 access_mask)
354 {
355 	BUG_ON((u64)(unsigned)access_mask != access_mask);
356 	WARN_ON(mmio_value & shadow_nonpresent_or_rsvd_lower_gfn_mask);
357 
358 	if (!enable_mmio_caching)
359 		mmio_value = 0;
360 
361 	/*
362 	 * Disable MMIO caching if the MMIO value collides with the bits that
363 	 * are used to hold the relocated GFN when the L1TF mitigation is
364 	 * enabled.  This should never fire as there is no known hardware that
365 	 * can trigger this condition, e.g. SME/SEV CPUs that require a custom
366 	 * MMIO value are not susceptible to L1TF.
367 	 */
368 	if (WARN_ON(mmio_value & (shadow_nonpresent_or_rsvd_mask <<
369 				  SHADOW_NONPRESENT_OR_RSVD_MASK_LEN)))
370 		mmio_value = 0;
371 
372 	/*
373 	 * The masked MMIO value must obviously match itself and a removed SPTE
374 	 * must not get a false positive.  Removed SPTEs and MMIO SPTEs should
375 	 * never collide as MMIO must set some RWX bits, and removed SPTEs must
376 	 * not set any RWX bits.
377 	 */
378 	if (WARN_ON((mmio_value & mmio_mask) != mmio_value) ||
379 	    WARN_ON(mmio_value && (REMOVED_SPTE & mmio_mask) == mmio_value))
380 		mmio_value = 0;
381 
382 	shadow_mmio_value = mmio_value;
383 	shadow_mmio_mask  = mmio_mask;
384 	shadow_mmio_access_mask = access_mask;
385 }
386 EXPORT_SYMBOL_GPL(kvm_mmu_set_mmio_spte_mask);
387 
388 void kvm_mmu_set_ept_masks(bool has_ad_bits, bool has_exec_only)
389 {
390 	shadow_user_mask	= VMX_EPT_READABLE_MASK;
391 	shadow_accessed_mask	= has_ad_bits ? VMX_EPT_ACCESS_BIT : 0ull;
392 	shadow_dirty_mask	= has_ad_bits ? VMX_EPT_DIRTY_BIT : 0ull;
393 	shadow_nx_mask		= 0ull;
394 	shadow_x_mask		= VMX_EPT_EXECUTABLE_MASK;
395 	shadow_present_mask	= has_exec_only ? 0ull : VMX_EPT_READABLE_MASK;
396 	shadow_acc_track_mask	= VMX_EPT_RWX_MASK;
397 	shadow_me_mask		= 0ull;
398 
399 	shadow_host_writable_mask = EPT_SPTE_HOST_WRITABLE;
400 	shadow_mmu_writable_mask  = EPT_SPTE_MMU_WRITABLE;
401 
402 	/*
403 	 * EPT Misconfigurations are generated if the value of bits 2:0
404 	 * of an EPT paging-structure entry is 110b (write/execute).
405 	 */
406 	kvm_mmu_set_mmio_spte_mask(VMX_EPT_MISCONFIG_WX_VALUE,
407 				   VMX_EPT_RWX_MASK, 0);
408 }
409 EXPORT_SYMBOL_GPL(kvm_mmu_set_ept_masks);
410 
411 void kvm_mmu_reset_all_pte_masks(void)
412 {
413 	u8 low_phys_bits;
414 	u64 mask;
415 
416 	shadow_phys_bits = kvm_get_shadow_phys_bits();
417 
418 	/*
419 	 * If the CPU has 46 or less physical address bits, then set an
420 	 * appropriate mask to guard against L1TF attacks. Otherwise, it is
421 	 * assumed that the CPU is not vulnerable to L1TF.
422 	 *
423 	 * Some Intel CPUs address the L1 cache using more PA bits than are
424 	 * reported by CPUID. Use the PA width of the L1 cache when possible
425 	 * to achieve more effective mitigation, e.g. if system RAM overlaps
426 	 * the most significant bits of legal physical address space.
427 	 */
428 	shadow_nonpresent_or_rsvd_mask = 0;
429 	low_phys_bits = boot_cpu_data.x86_phys_bits;
430 	if (boot_cpu_has_bug(X86_BUG_L1TF) &&
431 	    !WARN_ON_ONCE(boot_cpu_data.x86_cache_bits >=
432 			  52 - SHADOW_NONPRESENT_OR_RSVD_MASK_LEN)) {
433 		low_phys_bits = boot_cpu_data.x86_cache_bits
434 			- SHADOW_NONPRESENT_OR_RSVD_MASK_LEN;
435 		shadow_nonpresent_or_rsvd_mask =
436 			rsvd_bits(low_phys_bits, boot_cpu_data.x86_cache_bits - 1);
437 	}
438 
439 	shadow_nonpresent_or_rsvd_lower_gfn_mask =
440 		GENMASK_ULL(low_phys_bits - 1, PAGE_SHIFT);
441 
442 	shadow_user_mask	= PT_USER_MASK;
443 	shadow_accessed_mask	= PT_ACCESSED_MASK;
444 	shadow_dirty_mask	= PT_DIRTY_MASK;
445 	shadow_nx_mask		= PT64_NX_MASK;
446 	shadow_x_mask		= 0;
447 	shadow_present_mask	= PT_PRESENT_MASK;
448 	shadow_acc_track_mask	= 0;
449 	shadow_me_mask		= sme_me_mask;
450 
451 	shadow_host_writable_mask = DEFAULT_SPTE_HOST_WRITABLE;
452 	shadow_mmu_writable_mask  = DEFAULT_SPTE_MMU_WRITABLE;
453 
454 	/*
455 	 * Set a reserved PA bit in MMIO SPTEs to generate page faults with
456 	 * PFEC.RSVD=1 on MMIO accesses.  64-bit PTEs (PAE, x86-64, and EPT
457 	 * paging) support a maximum of 52 bits of PA, i.e. if the CPU supports
458 	 * 52-bit physical addresses then there are no reserved PA bits in the
459 	 * PTEs and so the reserved PA approach must be disabled.
460 	 */
461 	if (shadow_phys_bits < 52)
462 		mask = BIT_ULL(51) | PT_PRESENT_MASK;
463 	else
464 		mask = 0;
465 
466 	kvm_mmu_set_mmio_spte_mask(mask, mask, ACC_WRITE_MASK | ACC_USER_MASK);
467 }
468