xref: /linux/virt/kvm/pfncache.c (revision 08b7cf134eafca3b38e818d934b00dfe6b5b0fb4)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Kernel-based Virtual Machine driver for Linux
4  *
5  * This module enables kernel and guest-mode vCPU access to guest physical
6  * memory with suitable invalidation mechanisms.
7  *
8  * Copyright © 2021 Amazon.com, Inc. or its affiliates.
9  *
10  * Authors:
11  *   David Woodhouse <dwmw2@infradead.org>
12  */
13 
14 #include <linux/kvm_host.h>
15 #include <linux/kvm.h>
16 #include <linux/highmem.h>
17 #include <linux/module.h>
18 #include <linux/errno.h>
19 
20 #include "kvm_mm.h"
21 
22 /*
23  * MMU notifier 'invalidate_range_start' hook.
24  */
25 void gfn_to_pfn_cache_invalidate_start(struct kvm *kvm, unsigned long start,
26 				       unsigned long end, bool may_block)
27 {
28 	DECLARE_BITMAP(vcpu_bitmap, KVM_MAX_VCPUS);
29 	struct gfn_to_pfn_cache *gpc;
30 	bool evict_vcpus = false;
31 
32 	spin_lock(&kvm->gpc_lock);
33 	list_for_each_entry(gpc, &kvm->gpc_list, list) {
34 		write_lock_irq(&gpc->lock);
35 
36 		/* Only a single page so no need to care about length */
37 		if (gpc->valid && !is_error_noslot_pfn(gpc->pfn) &&
38 		    gpc->uhva >= start && gpc->uhva < end) {
39 			gpc->valid = false;
40 
41 			/*
42 			 * If a guest vCPU could be using the physical address,
43 			 * it needs to be forced out of guest mode.
44 			 */
45 			if (gpc->usage & KVM_GUEST_USES_PFN) {
46 				if (!evict_vcpus) {
47 					evict_vcpus = true;
48 					bitmap_zero(vcpu_bitmap, KVM_MAX_VCPUS);
49 				}
50 				__set_bit(gpc->vcpu->vcpu_idx, vcpu_bitmap);
51 			}
52 		}
53 		write_unlock_irq(&gpc->lock);
54 	}
55 	spin_unlock(&kvm->gpc_lock);
56 
57 	if (evict_vcpus) {
58 		/*
59 		 * KVM needs to ensure the vCPU is fully out of guest context
60 		 * before allowing the invalidation to continue.
61 		 */
62 		unsigned int req = KVM_REQ_OUTSIDE_GUEST_MODE;
63 		bool called;
64 
65 		/*
66 		 * If the OOM reaper is active, then all vCPUs should have
67 		 * been stopped already, so perform the request without
68 		 * KVM_REQUEST_WAIT and be sad if any needed to be IPI'd.
69 		 */
70 		if (!may_block)
71 			req &= ~KVM_REQUEST_WAIT;
72 
73 		called = kvm_make_vcpus_request_mask(kvm, req, vcpu_bitmap);
74 
75 		WARN_ON_ONCE(called && !may_block);
76 	}
77 }
78 
79 bool kvm_gfn_to_pfn_cache_check(struct kvm *kvm, struct gfn_to_pfn_cache *gpc,
80 				gpa_t gpa, unsigned long len)
81 {
82 	struct kvm_memslots *slots = kvm_memslots(kvm);
83 
84 	if ((gpa & ~PAGE_MASK) + len > PAGE_SIZE)
85 		return false;
86 
87 	if (gpc->gpa != gpa || gpc->generation != slots->generation ||
88 	    kvm_is_error_hva(gpc->uhva))
89 		return false;
90 
91 	if (!gpc->valid)
92 		return false;
93 
94 	return true;
95 }
96 EXPORT_SYMBOL_GPL(kvm_gfn_to_pfn_cache_check);
97 
98 static void __release_gpc(struct kvm *kvm, kvm_pfn_t pfn, void *khva, gpa_t gpa)
99 {
100 	/* Unmap the old page if it was mapped before, and release it */
101 	if (!is_error_noslot_pfn(pfn)) {
102 		if (khva) {
103 			if (pfn_valid(pfn))
104 				kunmap(pfn_to_page(pfn));
105 #ifdef CONFIG_HAS_IOMEM
106 			else
107 				memunmap(khva);
108 #endif
109 		}
110 
111 		kvm_release_pfn(pfn, false);
112 	}
113 }
114 
115 static kvm_pfn_t hva_to_pfn_retry(struct kvm *kvm, unsigned long uhva)
116 {
117 	unsigned long mmu_seq;
118 	kvm_pfn_t new_pfn;
119 	int retry;
120 
121 	do {
122 		mmu_seq = kvm->mmu_notifier_seq;
123 		smp_rmb();
124 
125 		/* We always request a writeable mapping */
126 		new_pfn = hva_to_pfn(uhva, false, NULL, true, NULL);
127 		if (is_error_noslot_pfn(new_pfn))
128 			break;
129 
130 		KVM_MMU_READ_LOCK(kvm);
131 		retry = mmu_notifier_retry_hva(kvm, mmu_seq, uhva);
132 		KVM_MMU_READ_UNLOCK(kvm);
133 		if (!retry)
134 			break;
135 
136 		cond_resched();
137 	} while (1);
138 
139 	return new_pfn;
140 }
141 
142 int kvm_gfn_to_pfn_cache_refresh(struct kvm *kvm, struct gfn_to_pfn_cache *gpc,
143 				 gpa_t gpa, unsigned long len)
144 {
145 	struct kvm_memslots *slots = kvm_memslots(kvm);
146 	unsigned long page_offset = gpa & ~PAGE_MASK;
147 	kvm_pfn_t old_pfn, new_pfn;
148 	unsigned long old_uhva;
149 	gpa_t old_gpa;
150 	void *old_khva;
151 	bool old_valid;
152 	int ret = 0;
153 
154 	/*
155 	 * If must fit within a single page. The 'len' argument is
156 	 * only to enforce that.
157 	 */
158 	if (page_offset + len > PAGE_SIZE)
159 		return -EINVAL;
160 
161 	write_lock_irq(&gpc->lock);
162 
163 	old_gpa = gpc->gpa;
164 	old_pfn = gpc->pfn;
165 	old_khva = gpc->khva - offset_in_page(gpc->khva);
166 	old_uhva = gpc->uhva;
167 	old_valid = gpc->valid;
168 
169 	/* If the userspace HVA is invalid, refresh that first */
170 	if (gpc->gpa != gpa || gpc->generation != slots->generation ||
171 	    kvm_is_error_hva(gpc->uhva)) {
172 		gfn_t gfn = gpa_to_gfn(gpa);
173 
174 		gpc->gpa = gpa;
175 		gpc->generation = slots->generation;
176 		gpc->memslot = __gfn_to_memslot(slots, gfn);
177 		gpc->uhva = gfn_to_hva_memslot(gpc->memslot, gfn);
178 
179 		if (kvm_is_error_hva(gpc->uhva)) {
180 			gpc->pfn = KVM_PFN_ERR_FAULT;
181 			ret = -EFAULT;
182 			goto out;
183 		}
184 
185 		gpc->uhva += page_offset;
186 	}
187 
188 	/*
189 	 * If the userspace HVA changed or the PFN was already invalid,
190 	 * drop the lock and do the HVA to PFN lookup again.
191 	 */
192 	if (!old_valid || old_uhva != gpc->uhva) {
193 		unsigned long uhva = gpc->uhva;
194 		void *new_khva = NULL;
195 
196 		/* Placeholders for "hva is valid but not yet mapped" */
197 		gpc->pfn = KVM_PFN_ERR_FAULT;
198 		gpc->khva = NULL;
199 		gpc->valid = true;
200 
201 		write_unlock_irq(&gpc->lock);
202 
203 		new_pfn = hva_to_pfn_retry(kvm, uhva);
204 		if (is_error_noslot_pfn(new_pfn)) {
205 			ret = -EFAULT;
206 			goto map_done;
207 		}
208 
209 		if (gpc->usage & KVM_HOST_USES_PFN) {
210 			if (new_pfn == old_pfn) {
211 				new_khva = old_khva;
212 				old_pfn = KVM_PFN_ERR_FAULT;
213 				old_khva = NULL;
214 			} else if (pfn_valid(new_pfn)) {
215 				new_khva = kmap(pfn_to_page(new_pfn));
216 #ifdef CONFIG_HAS_IOMEM
217 			} else {
218 				new_khva = memremap(pfn_to_hpa(new_pfn), PAGE_SIZE, MEMREMAP_WB);
219 #endif
220 			}
221 			if (new_khva)
222 				new_khva += page_offset;
223 			else
224 				ret = -EFAULT;
225 		}
226 
227 	map_done:
228 		write_lock_irq(&gpc->lock);
229 		if (ret) {
230 			gpc->valid = false;
231 			gpc->pfn = KVM_PFN_ERR_FAULT;
232 			gpc->khva = NULL;
233 		} else {
234 			/* At this point, gpc->valid may already have been cleared */
235 			gpc->pfn = new_pfn;
236 			gpc->khva = new_khva;
237 		}
238 	} else {
239 		/* If the HVA→PFN mapping was already valid, don't unmap it. */
240 		old_pfn = KVM_PFN_ERR_FAULT;
241 		old_khva = NULL;
242 	}
243 
244  out:
245 	write_unlock_irq(&gpc->lock);
246 
247 	__release_gpc(kvm, old_pfn, old_khva, old_gpa);
248 
249 	return ret;
250 }
251 EXPORT_SYMBOL_GPL(kvm_gfn_to_pfn_cache_refresh);
252 
253 void kvm_gfn_to_pfn_cache_unmap(struct kvm *kvm, struct gfn_to_pfn_cache *gpc)
254 {
255 	void *old_khva;
256 	kvm_pfn_t old_pfn;
257 	gpa_t old_gpa;
258 
259 	write_lock_irq(&gpc->lock);
260 
261 	gpc->valid = false;
262 
263 	old_khva = gpc->khva - offset_in_page(gpc->khva);
264 	old_gpa = gpc->gpa;
265 	old_pfn = gpc->pfn;
266 
267 	/*
268 	 * We can leave the GPA → uHVA map cache intact but the PFN
269 	 * lookup will need to be redone even for the same page.
270 	 */
271 	gpc->khva = NULL;
272 	gpc->pfn = KVM_PFN_ERR_FAULT;
273 
274 	write_unlock_irq(&gpc->lock);
275 
276 	__release_gpc(kvm, old_pfn, old_khva, old_gpa);
277 }
278 EXPORT_SYMBOL_GPL(kvm_gfn_to_pfn_cache_unmap);
279 
280 
281 int kvm_gfn_to_pfn_cache_init(struct kvm *kvm, struct gfn_to_pfn_cache *gpc,
282 			      struct kvm_vcpu *vcpu, enum pfn_cache_usage usage,
283 			      gpa_t gpa, unsigned long len)
284 {
285 	WARN_ON_ONCE(!usage || (usage & KVM_GUEST_AND_HOST_USE_PFN) != usage);
286 
287 	if (!gpc->active) {
288 		rwlock_init(&gpc->lock);
289 
290 		gpc->khva = NULL;
291 		gpc->pfn = KVM_PFN_ERR_FAULT;
292 		gpc->uhva = KVM_HVA_ERR_BAD;
293 		gpc->vcpu = vcpu;
294 		gpc->usage = usage;
295 		gpc->valid = false;
296 		gpc->active = true;
297 
298 		spin_lock(&kvm->gpc_lock);
299 		list_add(&gpc->list, &kvm->gpc_list);
300 		spin_unlock(&kvm->gpc_lock);
301 	}
302 	return kvm_gfn_to_pfn_cache_refresh(kvm, gpc, gpa, len);
303 }
304 EXPORT_SYMBOL_GPL(kvm_gfn_to_pfn_cache_init);
305 
306 void kvm_gfn_to_pfn_cache_destroy(struct kvm *kvm, struct gfn_to_pfn_cache *gpc)
307 {
308 	if (gpc->active) {
309 		spin_lock(&kvm->gpc_lock);
310 		list_del(&gpc->list);
311 		spin_unlock(&kvm->gpc_lock);
312 
313 		kvm_gfn_to_pfn_cache_unmap(kvm, gpc);
314 		gpc->active = false;
315 	}
316 }
317 EXPORT_SYMBOL_GPL(kvm_gfn_to_pfn_cache_destroy);
318