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 static bool __ro_after_init allow_mmio_caching;
24 module_param_named(mmio_caching, enable_mmio_caching, bool, 0444);
25 EXPORT_SYMBOL_GPL(enable_mmio_caching);
26 
27 u64 __read_mostly shadow_host_writable_mask;
28 u64 __read_mostly shadow_mmu_writable_mask;
29 u64 __read_mostly shadow_nx_mask;
30 u64 __read_mostly shadow_x_mask; /* mutual exclusive with nx_mask */
31 u64 __read_mostly shadow_user_mask;
32 u64 __read_mostly shadow_accessed_mask;
33 u64 __read_mostly shadow_dirty_mask;
34 u64 __read_mostly shadow_mmio_value;
35 u64 __read_mostly shadow_mmio_mask;
36 u64 __read_mostly shadow_mmio_access_mask;
37 u64 __read_mostly shadow_present_mask;
38 u64 __read_mostly shadow_me_value;
39 u64 __read_mostly shadow_me_mask;
40 u64 __read_mostly shadow_acc_track_mask;
41 
42 u64 __read_mostly shadow_nonpresent_or_rsvd_mask;
43 u64 __read_mostly shadow_nonpresent_or_rsvd_lower_gfn_mask;
44 
45 u8 __read_mostly shadow_phys_bits;
46 
kvm_mmu_spte_module_init(void)47 void __init kvm_mmu_spte_module_init(void)
48 {
49 	/*
50 	 * Snapshot userspace's desire to allow MMIO caching.  Whether or not
51 	 * KVM can actually enable MMIO caching depends on vendor-specific
52 	 * hardware capabilities and other module params that can't be resolved
53 	 * until the vendor module is loaded, i.e. enable_mmio_caching can and
54 	 * will change when the vendor module is (re)loaded.
55 	 */
56 	allow_mmio_caching = enable_mmio_caching;
57 }
58 
generation_mmio_spte_mask(u64 gen)59 static u64 generation_mmio_spte_mask(u64 gen)
60 {
61 	u64 mask;
62 
63 	WARN_ON(gen & ~MMIO_SPTE_GEN_MASK);
64 
65 	mask = (gen << MMIO_SPTE_GEN_LOW_SHIFT) & MMIO_SPTE_GEN_LOW_MASK;
66 	mask |= (gen << MMIO_SPTE_GEN_HIGH_SHIFT) & MMIO_SPTE_GEN_HIGH_MASK;
67 	return mask;
68 }
69 
make_mmio_spte(struct kvm_vcpu * vcpu,u64 gfn,unsigned int access)70 u64 make_mmio_spte(struct kvm_vcpu *vcpu, u64 gfn, unsigned int access)
71 {
72 	u64 gen = kvm_vcpu_memslots(vcpu)->generation & MMIO_SPTE_GEN_MASK;
73 	u64 spte = generation_mmio_spte_mask(gen);
74 	u64 gpa = gfn << PAGE_SHIFT;
75 
76 	WARN_ON_ONCE(!shadow_mmio_value);
77 
78 	access &= shadow_mmio_access_mask;
79 	spte |= shadow_mmio_value | access;
80 	spte |= gpa | shadow_nonpresent_or_rsvd_mask;
81 	spte |= (gpa & shadow_nonpresent_or_rsvd_mask)
82 		<< SHADOW_NONPRESENT_OR_RSVD_MASK_LEN;
83 
84 	return spte;
85 }
86 
kvm_is_mmio_pfn(kvm_pfn_t pfn)87 static bool kvm_is_mmio_pfn(kvm_pfn_t pfn)
88 {
89 	if (pfn_valid(pfn))
90 		return !is_zero_pfn(pfn) && PageReserved(pfn_to_page(pfn)) &&
91 			/*
92 			 * Some reserved pages, such as those from NVDIMM
93 			 * DAX devices, are not for MMIO, and can be mapped
94 			 * with cached memory type for better performance.
95 			 * However, the above check misconceives those pages
96 			 * as MMIO, and results in KVM mapping them with UC
97 			 * memory type, which would hurt the performance.
98 			 * Therefore, we check the host memory type in addition
99 			 * and only treat UC/UC-/WC pages as MMIO.
100 			 */
101 			(!pat_enabled() || pat_pfn_immune_to_uc_mtrr(pfn));
102 
103 	return !e820__mapped_raw_any(pfn_to_hpa(pfn),
104 				     pfn_to_hpa(pfn + 1) - 1,
105 				     E820_TYPE_RAM);
106 }
107 
108 /*
109  * Returns true if the SPTE has bits that may be set without holding mmu_lock.
110  * The caller is responsible for checking if the SPTE is shadow-present, and
111  * for determining whether or not the caller cares about non-leaf SPTEs.
112  */
spte_has_volatile_bits(u64 spte)113 bool spte_has_volatile_bits(u64 spte)
114 {
115 	/*
116 	 * Always atomically update spte if it can be updated
117 	 * out of mmu-lock, it can ensure dirty bit is not lost,
118 	 * also, it can help us to get a stable is_writable_pte()
119 	 * to ensure tlb flush is not missed.
120 	 */
121 	if (!is_writable_pte(spte) && is_mmu_writable_spte(spte))
122 		return true;
123 
124 	if (is_access_track_spte(spte))
125 		return true;
126 
127 	if (spte_ad_enabled(spte)) {
128 		if (!(spte & shadow_accessed_mask) ||
129 		    (is_writable_pte(spte) && !(spte & shadow_dirty_mask)))
130 			return true;
131 	}
132 
133 	return false;
134 }
135 
make_spte(struct kvm_vcpu * vcpu,struct kvm_mmu_page * sp,const struct kvm_memory_slot * slot,unsigned int pte_access,gfn_t gfn,kvm_pfn_t pfn,u64 old_spte,bool prefetch,bool can_unsync,bool host_writable,u64 * new_spte)136 bool make_spte(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
137 	       const struct kvm_memory_slot *slot,
138 	       unsigned int pte_access, gfn_t gfn, kvm_pfn_t pfn,
139 	       u64 old_spte, bool prefetch, bool can_unsync,
140 	       bool host_writable, u64 *new_spte)
141 {
142 	int level = sp->role.level;
143 	u64 spte = SPTE_MMU_PRESENT_MASK;
144 	bool wrprot = false;
145 
146 	WARN_ON_ONCE(!pte_access && !shadow_present_mask);
147 
148 	if (sp->role.ad_disabled)
149 		spte |= SPTE_TDP_AD_DISABLED_MASK;
150 	else if (kvm_mmu_page_ad_need_write_protect(sp))
151 		spte |= SPTE_TDP_AD_WRPROT_ONLY_MASK;
152 
153 	/*
154 	 * For the EPT case, shadow_present_mask is 0 if hardware
155 	 * supports exec-only page table entries.  In that case,
156 	 * ACC_USER_MASK and shadow_user_mask are used to represent
157 	 * read access.  See FNAME(gpte_access) in paging_tmpl.h.
158 	 */
159 	spte |= shadow_present_mask;
160 	if (!prefetch)
161 		spte |= spte_shadow_accessed_mask(spte);
162 
163 	if (level > PG_LEVEL_4K && (pte_access & ACC_EXEC_MASK) &&
164 	    is_nx_huge_page_enabled()) {
165 		pte_access &= ~ACC_EXEC_MASK;
166 	}
167 
168 	if (pte_access & ACC_EXEC_MASK)
169 		spte |= shadow_x_mask;
170 	else
171 		spte |= shadow_nx_mask;
172 
173 	if (pte_access & ACC_USER_MASK)
174 		spte |= shadow_user_mask;
175 
176 	if (level > PG_LEVEL_4K)
177 		spte |= PT_PAGE_SIZE_MASK;
178 	if (tdp_enabled)
179 		spte |= static_call(kvm_x86_get_mt_mask)(vcpu, gfn,
180 			kvm_is_mmio_pfn(pfn));
181 
182 	if (host_writable)
183 		spte |= shadow_host_writable_mask;
184 	else
185 		pte_access &= ~ACC_WRITE_MASK;
186 
187 	if (shadow_me_value && !kvm_is_mmio_pfn(pfn))
188 		spte |= shadow_me_value;
189 
190 	spte |= (u64)pfn << PAGE_SHIFT;
191 
192 	if (pte_access & ACC_WRITE_MASK) {
193 		spte |= PT_WRITABLE_MASK | shadow_mmu_writable_mask;
194 
195 		/*
196 		 * Optimization: for pte sync, if spte was writable the hash
197 		 * lookup is unnecessary (and expensive). Write protection
198 		 * is responsibility of kvm_mmu_get_page / kvm_mmu_sync_roots.
199 		 * Same reasoning can be applied to dirty page accounting.
200 		 */
201 		if (is_writable_pte(old_spte))
202 			goto out;
203 
204 		/*
205 		 * Unsync shadow pages that are reachable by the new, writable
206 		 * SPTE.  Write-protect the SPTE if the page can't be unsync'd,
207 		 * e.g. it's write-tracked (upper-level SPs) or has one or more
208 		 * shadow pages and unsync'ing pages is not allowed.
209 		 */
210 		if (mmu_try_to_unsync_pages(vcpu->kvm, slot, gfn, can_unsync, prefetch)) {
211 			pgprintk("%s: found shadow page for %llx, marking ro\n",
212 				 __func__, gfn);
213 			wrprot = true;
214 			pte_access &= ~ACC_WRITE_MASK;
215 			spte &= ~(PT_WRITABLE_MASK | shadow_mmu_writable_mask);
216 		}
217 	}
218 
219 	if (pte_access & ACC_WRITE_MASK)
220 		spte |= spte_shadow_dirty_mask(spte);
221 
222 out:
223 	if (prefetch)
224 		spte = mark_spte_for_access_track(spte);
225 
226 	WARN_ONCE(is_rsvd_spte(&vcpu->arch.mmu->shadow_zero_check, spte, level),
227 		  "spte = 0x%llx, level = %d, rsvd bits = 0x%llx", spte, level,
228 		  get_rsvd_bits(&vcpu->arch.mmu->shadow_zero_check, spte, level));
229 
230 	if ((spte & PT_WRITABLE_MASK) && kvm_slot_dirty_track_enabled(slot)) {
231 		/* Enforced by kvm_mmu_hugepage_adjust. */
232 		WARN_ON(level > PG_LEVEL_4K);
233 		mark_page_dirty_in_slot(vcpu->kvm, slot, gfn);
234 	}
235 
236 	*new_spte = spte;
237 	return wrprot;
238 }
239 
make_spte_executable(u64 spte)240 static u64 make_spte_executable(u64 spte)
241 {
242 	bool is_access_track = is_access_track_spte(spte);
243 
244 	if (is_access_track)
245 		spte = restore_acc_track_spte(spte);
246 
247 	spte &= ~shadow_nx_mask;
248 	spte |= shadow_x_mask;
249 
250 	if (is_access_track)
251 		spte = mark_spte_for_access_track(spte);
252 
253 	return spte;
254 }
255 
256 /*
257  * Construct an SPTE that maps a sub-page of the given huge page SPTE where
258  * `index` identifies which sub-page.
259  *
260  * This is used during huge page splitting to build the SPTEs that make up the
261  * new page table.
262  */
make_huge_page_split_spte(u64 huge_spte,int huge_level,int index)263 u64 make_huge_page_split_spte(u64 huge_spte, int huge_level, int index)
264 {
265 	u64 child_spte;
266 	int child_level;
267 
268 	if (WARN_ON_ONCE(!is_shadow_present_pte(huge_spte)))
269 		return 0;
270 
271 	if (WARN_ON_ONCE(!is_large_pte(huge_spte)))
272 		return 0;
273 
274 	child_spte = huge_spte;
275 	child_level = huge_level - 1;
276 
277 	/*
278 	 * The child_spte already has the base address of the huge page being
279 	 * split. So we just have to OR in the offset to the page at the next
280 	 * lower level for the given index.
281 	 */
282 	child_spte |= (index * KVM_PAGES_PER_HPAGE(child_level)) << PAGE_SHIFT;
283 
284 	if (child_level == PG_LEVEL_4K) {
285 		child_spte &= ~PT_PAGE_SIZE_MASK;
286 
287 		/*
288 		 * When splitting to a 4K page, mark the page executable as the
289 		 * NX hugepage mitigation no longer applies.
290 		 */
291 		if (is_nx_huge_page_enabled())
292 			child_spte = make_spte_executable(child_spte);
293 	}
294 
295 	return child_spte;
296 }
297 
298 
make_nonleaf_spte(u64 * child_pt,bool ad_disabled)299 u64 make_nonleaf_spte(u64 *child_pt, bool ad_disabled)
300 {
301 	u64 spte = SPTE_MMU_PRESENT_MASK;
302 
303 	spte |= __pa(child_pt) | shadow_present_mask | PT_WRITABLE_MASK |
304 		shadow_user_mask | shadow_x_mask | shadow_me_value;
305 
306 	if (ad_disabled)
307 		spte |= SPTE_TDP_AD_DISABLED_MASK;
308 	else
309 		spte |= shadow_accessed_mask;
310 
311 	return spte;
312 }
313 
kvm_mmu_changed_pte_notifier_make_spte(u64 old_spte,kvm_pfn_t new_pfn)314 u64 kvm_mmu_changed_pte_notifier_make_spte(u64 old_spte, kvm_pfn_t new_pfn)
315 {
316 	u64 new_spte;
317 
318 	new_spte = old_spte & ~PT64_BASE_ADDR_MASK;
319 	new_spte |= (u64)new_pfn << PAGE_SHIFT;
320 
321 	new_spte &= ~PT_WRITABLE_MASK;
322 	new_spte &= ~shadow_host_writable_mask;
323 	new_spte &= ~shadow_mmu_writable_mask;
324 
325 	new_spte = mark_spte_for_access_track(new_spte);
326 
327 	return new_spte;
328 }
329 
mark_spte_for_access_track(u64 spte)330 u64 mark_spte_for_access_track(u64 spte)
331 {
332 	if (spte_ad_enabled(spte))
333 		return spte & ~shadow_accessed_mask;
334 
335 	if (is_access_track_spte(spte))
336 		return spte;
337 
338 	check_spte_writable_invariants(spte);
339 
340 	WARN_ONCE(spte & (SHADOW_ACC_TRACK_SAVED_BITS_MASK <<
341 			  SHADOW_ACC_TRACK_SAVED_BITS_SHIFT),
342 		  "kvm: Access Tracking saved bit locations are not zero\n");
343 
344 	spte |= (spte & SHADOW_ACC_TRACK_SAVED_BITS_MASK) <<
345 		SHADOW_ACC_TRACK_SAVED_BITS_SHIFT;
346 	spte &= ~shadow_acc_track_mask;
347 
348 	return spte;
349 }
350 
kvm_mmu_set_mmio_spte_mask(u64 mmio_value,u64 mmio_mask,u64 access_mask)351 void kvm_mmu_set_mmio_spte_mask(u64 mmio_value, u64 mmio_mask, u64 access_mask)
352 {
353 	BUG_ON((u64)(unsigned)access_mask != access_mask);
354 	WARN_ON(mmio_value & shadow_nonpresent_or_rsvd_lower_gfn_mask);
355 
356 	/*
357 	 * Reset to the original module param value to honor userspace's desire
358 	 * to (dis)allow MMIO caching.  Update the param itself so that
359 	 * userspace can see whether or not KVM is actually using MMIO caching.
360 	 */
361 	enable_mmio_caching = allow_mmio_caching;
362 	if (!enable_mmio_caching)
363 		mmio_value = 0;
364 
365 	/*
366 	 * Disable MMIO caching if the MMIO value collides with the bits that
367 	 * are used to hold the relocated GFN when the L1TF mitigation is
368 	 * enabled.  This should never fire as there is no known hardware that
369 	 * can trigger this condition, e.g. SME/SEV CPUs that require a custom
370 	 * MMIO value are not susceptible to L1TF.
371 	 */
372 	if (WARN_ON(mmio_value & (shadow_nonpresent_or_rsvd_mask <<
373 				  SHADOW_NONPRESENT_OR_RSVD_MASK_LEN)))
374 		mmio_value = 0;
375 
376 	/*
377 	 * The masked MMIO value must obviously match itself and a removed SPTE
378 	 * must not get a false positive.  Removed SPTEs and MMIO SPTEs should
379 	 * never collide as MMIO must set some RWX bits, and removed SPTEs must
380 	 * not set any RWX bits.
381 	 */
382 	if (WARN_ON((mmio_value & mmio_mask) != mmio_value) ||
383 	    WARN_ON(mmio_value && (REMOVED_SPTE & mmio_mask) == mmio_value))
384 		mmio_value = 0;
385 
386 	if (!mmio_value)
387 		enable_mmio_caching = false;
388 
389 	shadow_mmio_value = mmio_value;
390 	shadow_mmio_mask  = mmio_mask;
391 	shadow_mmio_access_mask = access_mask;
392 }
393 EXPORT_SYMBOL_GPL(kvm_mmu_set_mmio_spte_mask);
394 
kvm_mmu_set_me_spte_mask(u64 me_value,u64 me_mask)395 void kvm_mmu_set_me_spte_mask(u64 me_value, u64 me_mask)
396 {
397 	/* shadow_me_value must be a subset of shadow_me_mask */
398 	if (WARN_ON(me_value & ~me_mask))
399 		me_value = me_mask = 0;
400 
401 	shadow_me_value = me_value;
402 	shadow_me_mask = me_mask;
403 }
404 EXPORT_SYMBOL_GPL(kvm_mmu_set_me_spte_mask);
405 
kvm_mmu_set_ept_masks(bool has_ad_bits,bool has_exec_only)406 void kvm_mmu_set_ept_masks(bool has_ad_bits, bool has_exec_only)
407 {
408 	shadow_user_mask	= VMX_EPT_READABLE_MASK;
409 	shadow_accessed_mask	= has_ad_bits ? VMX_EPT_ACCESS_BIT : 0ull;
410 	shadow_dirty_mask	= has_ad_bits ? VMX_EPT_DIRTY_BIT : 0ull;
411 	shadow_nx_mask		= 0ull;
412 	shadow_x_mask		= VMX_EPT_EXECUTABLE_MASK;
413 	shadow_present_mask	= has_exec_only ? 0ull : VMX_EPT_READABLE_MASK;
414 	shadow_acc_track_mask	= VMX_EPT_RWX_MASK;
415 	shadow_host_writable_mask = EPT_SPTE_HOST_WRITABLE;
416 	shadow_mmu_writable_mask  = EPT_SPTE_MMU_WRITABLE;
417 
418 	/*
419 	 * EPT Misconfigurations are generated if the value of bits 2:0
420 	 * of an EPT paging-structure entry is 110b (write/execute).
421 	 */
422 	kvm_mmu_set_mmio_spte_mask(VMX_EPT_MISCONFIG_WX_VALUE,
423 				   VMX_EPT_RWX_MASK, 0);
424 }
425 EXPORT_SYMBOL_GPL(kvm_mmu_set_ept_masks);
426 
kvm_mmu_reset_all_pte_masks(void)427 void kvm_mmu_reset_all_pte_masks(void)
428 {
429 	u8 low_phys_bits;
430 	u64 mask;
431 
432 	shadow_phys_bits = kvm_get_shadow_phys_bits();
433 
434 	/*
435 	 * If the CPU has 46 or less physical address bits, then set an
436 	 * appropriate mask to guard against L1TF attacks. Otherwise, it is
437 	 * assumed that the CPU is not vulnerable to L1TF.
438 	 *
439 	 * Some Intel CPUs address the L1 cache using more PA bits than are
440 	 * reported by CPUID. Use the PA width of the L1 cache when possible
441 	 * to achieve more effective mitigation, e.g. if system RAM overlaps
442 	 * the most significant bits of legal physical address space.
443 	 */
444 	shadow_nonpresent_or_rsvd_mask = 0;
445 	low_phys_bits = boot_cpu_data.x86_phys_bits;
446 	if (boot_cpu_has_bug(X86_BUG_L1TF) &&
447 	    !WARN_ON_ONCE(boot_cpu_data.x86_cache_bits >=
448 			  52 - SHADOW_NONPRESENT_OR_RSVD_MASK_LEN)) {
449 		low_phys_bits = boot_cpu_data.x86_cache_bits
450 			- SHADOW_NONPRESENT_OR_RSVD_MASK_LEN;
451 		shadow_nonpresent_or_rsvd_mask =
452 			rsvd_bits(low_phys_bits, boot_cpu_data.x86_cache_bits - 1);
453 	}
454 
455 	shadow_nonpresent_or_rsvd_lower_gfn_mask =
456 		GENMASK_ULL(low_phys_bits - 1, PAGE_SHIFT);
457 
458 	shadow_user_mask	= PT_USER_MASK;
459 	shadow_accessed_mask	= PT_ACCESSED_MASK;
460 	shadow_dirty_mask	= PT_DIRTY_MASK;
461 	shadow_nx_mask		= PT64_NX_MASK;
462 	shadow_x_mask		= 0;
463 	shadow_present_mask	= PT_PRESENT_MASK;
464 	shadow_acc_track_mask	= 0;
465 	shadow_me_mask		= 0;
466 	shadow_me_value		= 0;
467 
468 	shadow_host_writable_mask = DEFAULT_SPTE_HOST_WRITABLE;
469 	shadow_mmu_writable_mask  = DEFAULT_SPTE_MMU_WRITABLE;
470 
471 	/*
472 	 * Set a reserved PA bit in MMIO SPTEs to generate page faults with
473 	 * PFEC.RSVD=1 on MMIO accesses.  64-bit PTEs (PAE, x86-64, and EPT
474 	 * paging) support a maximum of 52 bits of PA, i.e. if the CPU supports
475 	 * 52-bit physical addresses then there are no reserved PA bits in the
476 	 * PTEs and so the reserved PA approach must be disabled.
477 	 */
478 	if (shadow_phys_bits < 52)
479 		mask = BIT_ULL(51) | PT_PRESENT_MASK;
480 	else
481 		mask = 0;
482 
483 	kvm_mmu_set_mmio_spte_mask(mask, mask, ACC_WRITE_MASK | ACC_USER_MASK);
484 }
485