xref: /linux/arch/arm64/kvm/hyp/nvhe/tlb.c (revision 90d32e92011eaae8e70a9169b4e7acf4ca8f9d3a)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2015 - ARM Ltd
4  * Author: Marc Zyngier <marc.zyngier@arm.com>
5  */
6 
7 #include <asm/kvm_hyp.h>
8 #include <asm/kvm_mmu.h>
9 #include <asm/tlbflush.h>
10 
11 #include <nvhe/mem_protect.h>
12 
13 struct tlb_inv_context {
14 	struct kvm_s2_mmu	*mmu;
15 	u64			tcr;
16 	u64			sctlr;
17 };
18 
19 static void enter_vmid_context(struct kvm_s2_mmu *mmu,
20 			       struct tlb_inv_context *cxt,
21 			       bool nsh)
22 {
23 	struct kvm_s2_mmu *host_s2_mmu = &host_mmu.arch.mmu;
24 	struct kvm_cpu_context *host_ctxt;
25 	struct kvm_vcpu *vcpu;
26 
27 	host_ctxt = &this_cpu_ptr(&kvm_host_data)->host_ctxt;
28 	vcpu = host_ctxt->__hyp_running_vcpu;
29 	cxt->mmu = NULL;
30 
31 	/*
32 	 * We have two requirements:
33 	 *
34 	 * - ensure that the page table updates are visible to all
35 	 *   CPUs, for which a dsb(DOMAIN-st) is what we need, DOMAIN
36 	 *   being either ish or nsh, depending on the invalidation
37 	 *   type.
38 	 *
39 	 * - complete any speculative page table walk started before
40 	 *   we trapped to EL2 so that we can mess with the MM
41 	 *   registers out of context, for which dsb(nsh) is enough
42 	 *
43 	 * The composition of these two barriers is a dsb(DOMAIN), and
44 	 * the 'nsh' parameter tracks the distinction between
45 	 * Inner-Shareable and Non-Shareable, as specified by the
46 	 * callers.
47 	 */
48 	if (nsh)
49 		dsb(nsh);
50 	else
51 		dsb(ish);
52 
53 	/*
54 	 * If we're already in the desired context, then there's nothing to do.
55 	 */
56 	if (vcpu) {
57 		/*
58 		 * We're in guest context. However, for this to work, this needs
59 		 * to be called from within __kvm_vcpu_run(), which ensures that
60 		 * __hyp_running_vcpu is set to the current guest vcpu.
61 		 */
62 		if (mmu == vcpu->arch.hw_mmu || WARN_ON(mmu != host_s2_mmu))
63 			return;
64 
65 		cxt->mmu = vcpu->arch.hw_mmu;
66 	} else {
67 		/* We're in host context. */
68 		if (mmu == host_s2_mmu)
69 			return;
70 
71 		cxt->mmu = host_s2_mmu;
72 	}
73 
74 	if (cpus_have_final_cap(ARM64_WORKAROUND_SPECULATIVE_AT)) {
75 		u64 val;
76 
77 		/*
78 		 * For CPUs that are affected by ARM 1319367, we need to
79 		 * avoid a Stage-1 walk with the old VMID while we have
80 		 * the new VMID set in the VTTBR in order to invalidate TLBs.
81 		 * We're guaranteed that the host S1 MMU is enabled, so
82 		 * we can simply set the EPD bits to avoid any further
83 		 * TLB fill. For guests, we ensure that the S1 MMU is
84 		 * temporarily enabled in the next context.
85 		 */
86 		val = cxt->tcr = read_sysreg_el1(SYS_TCR);
87 		val |= TCR_EPD1_MASK | TCR_EPD0_MASK;
88 		write_sysreg_el1(val, SYS_TCR);
89 		isb();
90 
91 		if (vcpu) {
92 			val = cxt->sctlr = read_sysreg_el1(SYS_SCTLR);
93 			if (!(val & SCTLR_ELx_M)) {
94 				val |= SCTLR_ELx_M;
95 				write_sysreg_el1(val, SYS_SCTLR);
96 				isb();
97 			}
98 		} else {
99 			/* The host S1 MMU is always enabled. */
100 			cxt->sctlr = SCTLR_ELx_M;
101 		}
102 	}
103 
104 	/*
105 	 * __load_stage2() includes an ISB only when the AT
106 	 * workaround is applied. Take care of the opposite condition,
107 	 * ensuring that we always have an ISB, but not two ISBs back
108 	 * to back.
109 	 */
110 	if (vcpu)
111 		__load_host_stage2();
112 	else
113 		__load_stage2(mmu, kern_hyp_va(mmu->arch));
114 
115 	asm(ALTERNATIVE("isb", "nop", ARM64_WORKAROUND_SPECULATIVE_AT));
116 }
117 
118 static void exit_vmid_context(struct tlb_inv_context *cxt)
119 {
120 	struct kvm_s2_mmu *mmu = cxt->mmu;
121 	struct kvm_cpu_context *host_ctxt;
122 	struct kvm_vcpu *vcpu;
123 
124 	host_ctxt = &this_cpu_ptr(&kvm_host_data)->host_ctxt;
125 	vcpu = host_ctxt->__hyp_running_vcpu;
126 
127 	if (!mmu)
128 		return;
129 
130 	if (vcpu)
131 		__load_stage2(mmu, kern_hyp_va(mmu->arch));
132 	else
133 		__load_host_stage2();
134 
135 	if (cpus_have_final_cap(ARM64_WORKAROUND_SPECULATIVE_AT)) {
136 		/* Ensure write of the old VMID */
137 		isb();
138 
139 		if (!(cxt->sctlr & SCTLR_ELx_M)) {
140 			write_sysreg_el1(cxt->sctlr, SYS_SCTLR);
141 			isb();
142 		}
143 
144 		write_sysreg_el1(cxt->tcr, SYS_TCR);
145 	}
146 }
147 
148 void __kvm_tlb_flush_vmid_ipa(struct kvm_s2_mmu *mmu,
149 			      phys_addr_t ipa, int level)
150 {
151 	struct tlb_inv_context cxt;
152 
153 	/* Switch to requested VMID */
154 	enter_vmid_context(mmu, &cxt, false);
155 
156 	/*
157 	 * We could do so much better if we had the VA as well.
158 	 * Instead, we invalidate Stage-2 for this IPA, and the
159 	 * whole of Stage-1. Weep...
160 	 */
161 	ipa >>= 12;
162 	__tlbi_level(ipas2e1is, ipa, level);
163 
164 	/*
165 	 * We have to ensure completion of the invalidation at Stage-2,
166 	 * since a table walk on another CPU could refill a TLB with a
167 	 * complete (S1 + S2) walk based on the old Stage-2 mapping if
168 	 * the Stage-1 invalidation happened first.
169 	 */
170 	dsb(ish);
171 	__tlbi(vmalle1is);
172 	dsb(ish);
173 	isb();
174 
175 	exit_vmid_context(&cxt);
176 }
177 
178 void __kvm_tlb_flush_vmid_ipa_nsh(struct kvm_s2_mmu *mmu,
179 				  phys_addr_t ipa, int level)
180 {
181 	struct tlb_inv_context cxt;
182 
183 	/* Switch to requested VMID */
184 	enter_vmid_context(mmu, &cxt, true);
185 
186 	/*
187 	 * We could do so much better if we had the VA as well.
188 	 * Instead, we invalidate Stage-2 for this IPA, and the
189 	 * whole of Stage-1. Weep...
190 	 */
191 	ipa >>= 12;
192 	__tlbi_level(ipas2e1, ipa, level);
193 
194 	/*
195 	 * We have to ensure completion of the invalidation at Stage-2,
196 	 * since a table walk on another CPU could refill a TLB with a
197 	 * complete (S1 + S2) walk based on the old Stage-2 mapping if
198 	 * the Stage-1 invalidation happened first.
199 	 */
200 	dsb(nsh);
201 	__tlbi(vmalle1);
202 	dsb(nsh);
203 	isb();
204 
205 	exit_vmid_context(&cxt);
206 }
207 
208 void __kvm_tlb_flush_vmid_range(struct kvm_s2_mmu *mmu,
209 				phys_addr_t start, unsigned long pages)
210 {
211 	struct tlb_inv_context cxt;
212 	unsigned long stride;
213 
214 	/*
215 	 * Since the range of addresses may not be mapped at
216 	 * the same level, assume the worst case as PAGE_SIZE
217 	 */
218 	stride = PAGE_SIZE;
219 	start = round_down(start, stride);
220 
221 	/* Switch to requested VMID */
222 	enter_vmid_context(mmu, &cxt, false);
223 
224 	__flush_s2_tlb_range_op(ipas2e1is, start, pages, stride,
225 				TLBI_TTL_UNKNOWN);
226 
227 	dsb(ish);
228 	__tlbi(vmalle1is);
229 	dsb(ish);
230 	isb();
231 
232 	exit_vmid_context(&cxt);
233 }
234 
235 void __kvm_tlb_flush_vmid(struct kvm_s2_mmu *mmu)
236 {
237 	struct tlb_inv_context cxt;
238 
239 	/* Switch to requested VMID */
240 	enter_vmid_context(mmu, &cxt, false);
241 
242 	__tlbi(vmalls12e1is);
243 	dsb(ish);
244 	isb();
245 
246 	exit_vmid_context(&cxt);
247 }
248 
249 void __kvm_flush_cpu_context(struct kvm_s2_mmu *mmu)
250 {
251 	struct tlb_inv_context cxt;
252 
253 	/* Switch to requested VMID */
254 	enter_vmid_context(mmu, &cxt, false);
255 
256 	__tlbi(vmalle1);
257 	asm volatile("ic iallu");
258 	dsb(nsh);
259 	isb();
260 
261 	exit_vmid_context(&cxt);
262 }
263 
264 void __kvm_flush_vm_context(void)
265 {
266 	/* Same remark as in enter_vmid_context() */
267 	dsb(ish);
268 	__tlbi(alle1is);
269 	dsb(ish);
270 }
271