xref: /openbmc/linux/arch/x86/kvm/mmu/spte.c (revision 55eb9a6c)
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 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_value;
37 u64 __read_mostly shadow_me_mask;
38 u64 __read_mostly shadow_acc_track_mask;
39 
40 u64 __read_mostly shadow_nonpresent_or_rsvd_mask;
41 u64 __read_mostly shadow_nonpresent_or_rsvd_lower_gfn_mask;
42 
43 u8 __read_mostly shadow_phys_bits;
44 
45 static u64 generation_mmio_spte_mask(u64 gen)
46 {
47 	u64 mask;
48 
49 	WARN_ON(gen & ~MMIO_SPTE_GEN_MASK);
50 
51 	mask = (gen << MMIO_SPTE_GEN_LOW_SHIFT) & MMIO_SPTE_GEN_LOW_MASK;
52 	mask |= (gen << MMIO_SPTE_GEN_HIGH_SHIFT) & MMIO_SPTE_GEN_HIGH_MASK;
53 	return mask;
54 }
55 
56 u64 make_mmio_spte(struct kvm_vcpu *vcpu, u64 gfn, unsigned int access)
57 {
58 	u64 gen = kvm_vcpu_memslots(vcpu)->generation & MMIO_SPTE_GEN_MASK;
59 	u64 spte = generation_mmio_spte_mask(gen);
60 	u64 gpa = gfn << PAGE_SHIFT;
61 
62 	WARN_ON_ONCE(!shadow_mmio_value);
63 
64 	access &= shadow_mmio_access_mask;
65 	spte |= shadow_mmio_value | access;
66 	spte |= gpa | shadow_nonpresent_or_rsvd_mask;
67 	spte |= (gpa & shadow_nonpresent_or_rsvd_mask)
68 		<< SHADOW_NONPRESENT_OR_RSVD_MASK_LEN;
69 
70 	return spte;
71 }
72 
73 static bool kvm_is_mmio_pfn(kvm_pfn_t pfn)
74 {
75 	if (pfn_valid(pfn))
76 		return !is_zero_pfn(pfn) && PageReserved(pfn_to_page(pfn)) &&
77 			/*
78 			 * Some reserved pages, such as those from NVDIMM
79 			 * DAX devices, are not for MMIO, and can be mapped
80 			 * with cached memory type for better performance.
81 			 * However, the above check misconceives those pages
82 			 * as MMIO, and results in KVM mapping them with UC
83 			 * memory type, which would hurt the performance.
84 			 * Therefore, we check the host memory type in addition
85 			 * and only treat UC/UC-/WC pages as MMIO.
86 			 */
87 			(!pat_enabled() || pat_pfn_immune_to_uc_mtrr(pfn));
88 
89 	return !e820__mapped_raw_any(pfn_to_hpa(pfn),
90 				     pfn_to_hpa(pfn + 1) - 1,
91 				     E820_TYPE_RAM);
92 }
93 
94 /*
95  * Returns true if the SPTE has bits that may be set without holding mmu_lock.
96  * The caller is responsible for checking if the SPTE is shadow-present, and
97  * for determining whether or not the caller cares about non-leaf SPTEs.
98  */
99 bool spte_has_volatile_bits(u64 spte)
100 {
101 	/*
102 	 * Always atomically update spte if it can be updated
103 	 * out of mmu-lock, it can ensure dirty bit is not lost,
104 	 * also, it can help us to get a stable is_writable_pte()
105 	 * to ensure tlb flush is not missed.
106 	 */
107 	if (!is_writable_pte(spte) && is_mmu_writable_spte(spte))
108 		return true;
109 
110 	if (is_access_track_spte(spte))
111 		return true;
112 
113 	if (spte_ad_enabled(spte)) {
114 		if (!(spte & shadow_accessed_mask) ||
115 		    (is_writable_pte(spte) && !(spte & shadow_dirty_mask)))
116 			return true;
117 	}
118 
119 	return false;
120 }
121 
122 bool make_spte(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
123 	       const struct kvm_memory_slot *slot,
124 	       unsigned int pte_access, gfn_t gfn, kvm_pfn_t pfn,
125 	       u64 old_spte, bool prefetch, bool can_unsync,
126 	       bool host_writable, u64 *new_spte)
127 {
128 	int level = sp->role.level;
129 	u64 spte = SPTE_MMU_PRESENT_MASK;
130 	bool wrprot = false;
131 
132 	if (sp->role.ad_disabled)
133 		spte |= SPTE_TDP_AD_DISABLED_MASK;
134 	else if (kvm_mmu_page_ad_need_write_protect(sp))
135 		spte |= SPTE_TDP_AD_WRPROT_ONLY_MASK;
136 
137 	/*
138 	 * For the EPT case, shadow_present_mask is 0 if hardware
139 	 * supports exec-only page table entries.  In that case,
140 	 * ACC_USER_MASK and shadow_user_mask are used to represent
141 	 * read access.  See FNAME(gpte_access) in paging_tmpl.h.
142 	 */
143 	spte |= shadow_present_mask;
144 	if (!prefetch)
145 		spte |= spte_shadow_accessed_mask(spte);
146 
147 	if (level > PG_LEVEL_4K && (pte_access & ACC_EXEC_MASK) &&
148 	    is_nx_huge_page_enabled()) {
149 		pte_access &= ~ACC_EXEC_MASK;
150 	}
151 
152 	if (pte_access & ACC_EXEC_MASK)
153 		spte |= shadow_x_mask;
154 	else
155 		spte |= shadow_nx_mask;
156 
157 	if (pte_access & ACC_USER_MASK)
158 		spte |= shadow_user_mask;
159 
160 	if (level > PG_LEVEL_4K)
161 		spte |= PT_PAGE_SIZE_MASK;
162 	if (tdp_enabled)
163 		spte |= static_call(kvm_x86_get_mt_mask)(vcpu, gfn,
164 			kvm_is_mmio_pfn(pfn));
165 
166 	if (host_writable)
167 		spte |= shadow_host_writable_mask;
168 	else
169 		pte_access &= ~ACC_WRITE_MASK;
170 
171 	if (shadow_me_value && !kvm_is_mmio_pfn(pfn))
172 		spte |= shadow_me_value;
173 
174 	spte |= (u64)pfn << PAGE_SHIFT;
175 
176 	if (pte_access & ACC_WRITE_MASK) {
177 		spte |= PT_WRITABLE_MASK | shadow_mmu_writable_mask;
178 
179 		/*
180 		 * Optimization: for pte sync, if spte was writable the hash
181 		 * lookup is unnecessary (and expensive). Write protection
182 		 * is responsibility of kvm_mmu_get_page / kvm_mmu_sync_roots.
183 		 * Same reasoning can be applied to dirty page accounting.
184 		 */
185 		if (is_writable_pte(old_spte))
186 			goto out;
187 
188 		/*
189 		 * Unsync shadow pages that are reachable by the new, writable
190 		 * SPTE.  Write-protect the SPTE if the page can't be unsync'd,
191 		 * e.g. it's write-tracked (upper-level SPs) or has one or more
192 		 * shadow pages and unsync'ing pages is not allowed.
193 		 */
194 		if (mmu_try_to_unsync_pages(vcpu->kvm, slot, gfn, can_unsync, prefetch)) {
195 			pgprintk("%s: found shadow page for %llx, marking ro\n",
196 				 __func__, gfn);
197 			wrprot = true;
198 			pte_access &= ~ACC_WRITE_MASK;
199 			spte &= ~(PT_WRITABLE_MASK | shadow_mmu_writable_mask);
200 		}
201 	}
202 
203 	if (pte_access & ACC_WRITE_MASK)
204 		spte |= spte_shadow_dirty_mask(spte);
205 
206 out:
207 	if (prefetch)
208 		spte = mark_spte_for_access_track(spte);
209 
210 	WARN_ONCE(is_rsvd_spte(&vcpu->arch.mmu->shadow_zero_check, spte, level),
211 		  "spte = 0x%llx, level = %d, rsvd bits = 0x%llx", spte, level,
212 		  get_rsvd_bits(&vcpu->arch.mmu->shadow_zero_check, spte, level));
213 
214 	if ((spte & PT_WRITABLE_MASK) && kvm_slot_dirty_track_enabled(slot)) {
215 		/* Enforced by kvm_mmu_hugepage_adjust. */
216 		WARN_ON(level > PG_LEVEL_4K);
217 		mark_page_dirty_in_slot(vcpu->kvm, slot, gfn);
218 	}
219 
220 	*new_spte = spte;
221 	return wrprot;
222 }
223 
224 static u64 make_spte_executable(u64 spte)
225 {
226 	bool is_access_track = is_access_track_spte(spte);
227 
228 	if (is_access_track)
229 		spte = restore_acc_track_spte(spte);
230 
231 	spte &= ~shadow_nx_mask;
232 	spte |= shadow_x_mask;
233 
234 	if (is_access_track)
235 		spte = mark_spte_for_access_track(spte);
236 
237 	return spte;
238 }
239 
240 /*
241  * Construct an SPTE that maps a sub-page of the given huge page SPTE where
242  * `index` identifies which sub-page.
243  *
244  * This is used during huge page splitting to build the SPTEs that make up the
245  * new page table.
246  */
247 u64 make_huge_page_split_spte(u64 huge_spte, int huge_level, int index)
248 {
249 	u64 child_spte;
250 	int child_level;
251 
252 	if (WARN_ON_ONCE(!is_shadow_present_pte(huge_spte)))
253 		return 0;
254 
255 	if (WARN_ON_ONCE(!is_large_pte(huge_spte)))
256 		return 0;
257 
258 	child_spte = huge_spte;
259 	child_level = huge_level - 1;
260 
261 	/*
262 	 * The child_spte already has the base address of the huge page being
263 	 * split. So we just have to OR in the offset to the page at the next
264 	 * lower level for the given index.
265 	 */
266 	child_spte |= (index * KVM_PAGES_PER_HPAGE(child_level)) << PAGE_SHIFT;
267 
268 	if (child_level == PG_LEVEL_4K) {
269 		child_spte &= ~PT_PAGE_SIZE_MASK;
270 
271 		/*
272 		 * When splitting to a 4K page, mark the page executable as the
273 		 * NX hugepage mitigation no longer applies.
274 		 */
275 		if (is_nx_huge_page_enabled())
276 			child_spte = make_spte_executable(child_spte);
277 	}
278 
279 	return child_spte;
280 }
281 
282 
283 u64 make_nonleaf_spte(u64 *child_pt, bool ad_disabled)
284 {
285 	u64 spte = SPTE_MMU_PRESENT_MASK;
286 
287 	spte |= __pa(child_pt) | shadow_present_mask | PT_WRITABLE_MASK |
288 		shadow_user_mask | shadow_x_mask | shadow_me_value;
289 
290 	if (ad_disabled)
291 		spte |= SPTE_TDP_AD_DISABLED_MASK;
292 	else
293 		spte |= shadow_accessed_mask;
294 
295 	return spte;
296 }
297 
298 u64 kvm_mmu_changed_pte_notifier_make_spte(u64 old_spte, kvm_pfn_t new_pfn)
299 {
300 	u64 new_spte;
301 
302 	new_spte = old_spte & ~PT64_BASE_ADDR_MASK;
303 	new_spte |= (u64)new_pfn << PAGE_SHIFT;
304 
305 	new_spte &= ~PT_WRITABLE_MASK;
306 	new_spte &= ~shadow_host_writable_mask;
307 	new_spte &= ~shadow_mmu_writable_mask;
308 
309 	new_spte = mark_spte_for_access_track(new_spte);
310 
311 	return new_spte;
312 }
313 
314 u64 mark_spte_for_access_track(u64 spte)
315 {
316 	if (spte_ad_enabled(spte))
317 		return spte & ~shadow_accessed_mask;
318 
319 	if (is_access_track_spte(spte))
320 		return spte;
321 
322 	check_spte_writable_invariants(spte);
323 
324 	WARN_ONCE(spte & (SHADOW_ACC_TRACK_SAVED_BITS_MASK <<
325 			  SHADOW_ACC_TRACK_SAVED_BITS_SHIFT),
326 		  "kvm: Access Tracking saved bit locations are not zero\n");
327 
328 	spte |= (spte & SHADOW_ACC_TRACK_SAVED_BITS_MASK) <<
329 		SHADOW_ACC_TRACK_SAVED_BITS_SHIFT;
330 	spte &= ~shadow_acc_track_mask;
331 
332 	return spte;
333 }
334 
335 void kvm_mmu_set_mmio_spte_mask(u64 mmio_value, u64 mmio_mask, u64 access_mask)
336 {
337 	BUG_ON((u64)(unsigned)access_mask != access_mask);
338 	WARN_ON(mmio_value & shadow_nonpresent_or_rsvd_lower_gfn_mask);
339 
340 	if (!enable_mmio_caching)
341 		mmio_value = 0;
342 
343 	/*
344 	 * Disable MMIO caching if the MMIO value collides with the bits that
345 	 * are used to hold the relocated GFN when the L1TF mitigation is
346 	 * enabled.  This should never fire as there is no known hardware that
347 	 * can trigger this condition, e.g. SME/SEV CPUs that require a custom
348 	 * MMIO value are not susceptible to L1TF.
349 	 */
350 	if (WARN_ON(mmio_value & (shadow_nonpresent_or_rsvd_mask <<
351 				  SHADOW_NONPRESENT_OR_RSVD_MASK_LEN)))
352 		mmio_value = 0;
353 
354 	/*
355 	 * The masked MMIO value must obviously match itself and a removed SPTE
356 	 * must not get a false positive.  Removed SPTEs and MMIO SPTEs should
357 	 * never collide as MMIO must set some RWX bits, and removed SPTEs must
358 	 * not set any RWX bits.
359 	 */
360 	if (WARN_ON((mmio_value & mmio_mask) != mmio_value) ||
361 	    WARN_ON(mmio_value && (REMOVED_SPTE & mmio_mask) == mmio_value))
362 		mmio_value = 0;
363 
364 	if (!mmio_value)
365 		enable_mmio_caching = false;
366 
367 	shadow_mmio_value = mmio_value;
368 	shadow_mmio_mask  = mmio_mask;
369 	shadow_mmio_access_mask = access_mask;
370 }
371 EXPORT_SYMBOL_GPL(kvm_mmu_set_mmio_spte_mask);
372 
373 void kvm_mmu_set_me_spte_mask(u64 me_value, u64 me_mask)
374 {
375 	/* shadow_me_value must be a subset of shadow_me_mask */
376 	if (WARN_ON(me_value & ~me_mask))
377 		me_value = me_mask = 0;
378 
379 	shadow_me_value = me_value;
380 	shadow_me_mask = me_mask;
381 }
382 EXPORT_SYMBOL_GPL(kvm_mmu_set_me_spte_mask);
383 
384 void kvm_mmu_set_ept_masks(bool has_ad_bits, bool has_exec_only)
385 {
386 	shadow_user_mask	= VMX_EPT_READABLE_MASK;
387 	shadow_accessed_mask	= has_ad_bits ? VMX_EPT_ACCESS_BIT : 0ull;
388 	shadow_dirty_mask	= has_ad_bits ? VMX_EPT_DIRTY_BIT : 0ull;
389 	shadow_nx_mask		= 0ull;
390 	shadow_x_mask		= VMX_EPT_EXECUTABLE_MASK;
391 	shadow_present_mask	= has_exec_only ? 0ull : VMX_EPT_READABLE_MASK;
392 	shadow_acc_track_mask	= VMX_EPT_RWX_MASK;
393 	shadow_host_writable_mask = EPT_SPTE_HOST_WRITABLE;
394 	shadow_mmu_writable_mask  = EPT_SPTE_MMU_WRITABLE;
395 
396 	/*
397 	 * EPT Misconfigurations are generated if the value of bits 2:0
398 	 * of an EPT paging-structure entry is 110b (write/execute).
399 	 */
400 	kvm_mmu_set_mmio_spte_mask(VMX_EPT_MISCONFIG_WX_VALUE,
401 				   VMX_EPT_RWX_MASK, 0);
402 }
403 EXPORT_SYMBOL_GPL(kvm_mmu_set_ept_masks);
404 
405 void kvm_mmu_reset_all_pte_masks(void)
406 {
407 	u8 low_phys_bits;
408 	u64 mask;
409 
410 	shadow_phys_bits = kvm_get_shadow_phys_bits();
411 
412 	/*
413 	 * If the CPU has 46 or less physical address bits, then set an
414 	 * appropriate mask to guard against L1TF attacks. Otherwise, it is
415 	 * assumed that the CPU is not vulnerable to L1TF.
416 	 *
417 	 * Some Intel CPUs address the L1 cache using more PA bits than are
418 	 * reported by CPUID. Use the PA width of the L1 cache when possible
419 	 * to achieve more effective mitigation, e.g. if system RAM overlaps
420 	 * the most significant bits of legal physical address space.
421 	 */
422 	shadow_nonpresent_or_rsvd_mask = 0;
423 	low_phys_bits = boot_cpu_data.x86_phys_bits;
424 	if (boot_cpu_has_bug(X86_BUG_L1TF) &&
425 	    !WARN_ON_ONCE(boot_cpu_data.x86_cache_bits >=
426 			  52 - SHADOW_NONPRESENT_OR_RSVD_MASK_LEN)) {
427 		low_phys_bits = boot_cpu_data.x86_cache_bits
428 			- SHADOW_NONPRESENT_OR_RSVD_MASK_LEN;
429 		shadow_nonpresent_or_rsvd_mask =
430 			rsvd_bits(low_phys_bits, boot_cpu_data.x86_cache_bits - 1);
431 	}
432 
433 	shadow_nonpresent_or_rsvd_lower_gfn_mask =
434 		GENMASK_ULL(low_phys_bits - 1, PAGE_SHIFT);
435 
436 	shadow_user_mask	= PT_USER_MASK;
437 	shadow_accessed_mask	= PT_ACCESSED_MASK;
438 	shadow_dirty_mask	= PT_DIRTY_MASK;
439 	shadow_nx_mask		= PT64_NX_MASK;
440 	shadow_x_mask		= 0;
441 	shadow_present_mask	= PT_PRESENT_MASK;
442 	shadow_acc_track_mask	= 0;
443 	shadow_me_mask		= 0;
444 	shadow_me_value		= 0;
445 
446 	shadow_host_writable_mask = DEFAULT_SPTE_HOST_WRITABLE;
447 	shadow_mmu_writable_mask  = DEFAULT_SPTE_MMU_WRITABLE;
448 
449 	/*
450 	 * Set a reserved PA bit in MMIO SPTEs to generate page faults with
451 	 * PFEC.RSVD=1 on MMIO accesses.  64-bit PTEs (PAE, x86-64, and EPT
452 	 * paging) support a maximum of 52 bits of PA, i.e. if the CPU supports
453 	 * 52-bit physical addresses then there are no reserved PA bits in the
454 	 * PTEs and so the reserved PA approach must be disabled.
455 	 */
456 	if (shadow_phys_bits < 52)
457 		mask = BIT_ULL(51) | PT_PRESENT_MASK;
458 	else
459 		mask = 0;
460 
461 	kvm_mmu_set_mmio_spte_mask(mask, mask, ACC_WRITE_MASK | ACC_USER_MASK);
462 }
463