xref: /linux/arch/arm64/kvm/hyp/vhe/tlb.c (revision fc2d791a43d3880496d1c729b8bd74d2c19cb4e7)
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 <linux/irqflags.h>
8 
9 #include <asm/kvm_hyp.h>
10 #include <asm/kvm_mmu.h>
11 #include <asm/tlbflush.h>
12 
13 struct tlb_inv_context {
14 	struct kvm_s2_mmu	*mmu;
15 	unsigned long		flags;
16 	u64			tcr;
17 	u64			sctlr;
18 };
19 
20 static void enter_vmid_context(struct kvm_s2_mmu *mmu,
21 			       struct tlb_inv_context *cxt)
22 {
23 	struct kvm_vcpu *vcpu = kvm_get_running_vcpu();
24 	u64 val;
25 
26 	local_irq_save(cxt->flags);
27 
28 	if (vcpu && mmu != vcpu->arch.hw_mmu)
29 		cxt->mmu = vcpu->arch.hw_mmu;
30 	else
31 		cxt->mmu = NULL;
32 
33 	if (cpus_have_final_cap(ARM64_WORKAROUND_SPECULATIVE_AT)) {
34 		/*
35 		 * For CPUs that are affected by ARM errata 1165522 or 1530923,
36 		 * we cannot trust stage-1 to be in a correct state at that
37 		 * point. Since we do not want to force a full load of the
38 		 * vcpu state, we prevent the EL1 page-table walker to
39 		 * allocate new TLBs. This is done by setting the EPD bits
40 		 * in the TCR_EL1 register. We also need to prevent it to
41 		 * allocate IPA->PA walks, so we enable the S1 MMU...
42 		 */
43 		val = cxt->tcr = read_sysreg_el1(SYS_TCR);
44 		val |= TCR_EPD1_MASK | TCR_EPD0_MASK;
45 		write_sysreg_el1(val, SYS_TCR);
46 		val = cxt->sctlr = read_sysreg_el1(SYS_SCTLR);
47 		val |= SCTLR_ELx_M;
48 		write_sysreg_el1(val, SYS_SCTLR);
49 	}
50 
51 	/*
52 	 * With VHE enabled, we have HCR_EL2.{E2H,TGE} = {1,1}, and
53 	 * most TLB operations target EL2/EL0. In order to affect the
54 	 * guest TLBs (EL1/EL0), we need to change one of these two
55 	 * bits. Changing E2H is impossible (goodbye TTBR1_EL2), so
56 	 * let's flip TGE before executing the TLB operation.
57 	 *
58 	 * ARM erratum 1165522 requires some special handling (again),
59 	 * as we need to make sure both stages of translation are in
60 	 * place before clearing TGE. __load_stage2() already
61 	 * has an ISB in order to deal with this.
62 	 */
63 	__load_stage2(mmu, mmu->arch);
64 	val = read_sysreg(hcr_el2);
65 	val &= ~HCR_TGE;
66 	write_sysreg_hcr(val);
67 	isb();
68 }
69 
70 static void exit_vmid_context(struct tlb_inv_context *cxt)
71 {
72 	/*
73 	 * We're done with the TLB operation, let's restore the host's
74 	 * view of HCR_EL2.
75 	 */
76 	write_sysreg_hcr(HCR_HOST_VHE_FLAGS);
77 	isb();
78 
79 	/* ... and the stage-2 MMU context that we switched away from */
80 	if (cxt->mmu)
81 		__load_stage2(cxt->mmu, cxt->mmu->arch);
82 
83 	if (cpus_have_final_cap(ARM64_WORKAROUND_SPECULATIVE_AT)) {
84 		/* Restore the registers to what they were */
85 		write_sysreg_el1(cxt->tcr, SYS_TCR);
86 		write_sysreg_el1(cxt->sctlr, SYS_SCTLR);
87 	}
88 
89 	local_irq_restore(cxt->flags);
90 }
91 
92 void __kvm_tlb_flush_vmid_ipa(struct kvm_s2_mmu *mmu,
93 			      phys_addr_t ipa, int level)
94 {
95 	struct tlb_inv_context cxt;
96 
97 	dsb(ishst);
98 
99 	/* Switch to requested VMID */
100 	enter_vmid_context(mmu, &cxt);
101 
102 	/*
103 	 * We could do so much better if we had the VA as well.
104 	 * Instead, we invalidate Stage-2 for this IPA, and the
105 	 * whole of Stage-1. Weep...
106 	 */
107 	__tlbi_level(ipas2e1is, ipa, level);
108 
109 	/*
110 	 * We have to ensure completion of the invalidation at Stage-2,
111 	 * since a table walk on another CPU could refill a TLB with a
112 	 * complete (S1 + S2) walk based on the old Stage-2 mapping if
113 	 * the Stage-1 invalidation happened first.
114 	 */
115 	dsb(ish);
116 	__tlbi(vmalle1is);
117 	__tlbi_sync_s1ish_hyp();
118 	isb();
119 
120 	exit_vmid_context(&cxt);
121 }
122 
123 void __kvm_tlb_flush_vmid_ipa_nsh(struct kvm_s2_mmu *mmu,
124 				  phys_addr_t ipa, int level)
125 {
126 	struct tlb_inv_context cxt;
127 
128 	dsb(nshst);
129 
130 	/* Switch to requested VMID */
131 	enter_vmid_context(mmu, &cxt);
132 
133 	/*
134 	 * We could do so much better if we had the VA as well.
135 	 * Instead, we invalidate Stage-2 for this IPA, and the
136 	 * whole of Stage-1. Weep...
137 	 */
138 	__tlbi_level(ipas2e1, ipa, level);
139 
140 	/*
141 	 * We have to ensure completion of the invalidation at Stage-2,
142 	 * since a table walk on another CPU could refill a TLB with a
143 	 * complete (S1 + S2) walk based on the old Stage-2 mapping if
144 	 * the Stage-1 invalidation happened first.
145 	 */
146 	dsb(nsh);
147 	__tlbi(vmalle1);
148 	dsb(nsh);
149 	isb();
150 
151 	exit_vmid_context(&cxt);
152 }
153 
154 void __kvm_tlb_flush_vmid_range(struct kvm_s2_mmu *mmu,
155 				phys_addr_t start, unsigned long pages)
156 {
157 	struct tlb_inv_context cxt;
158 	unsigned long stride;
159 
160 	/*
161 	 * Since the range of addresses may not be mapped at
162 	 * the same level, assume the worst case as PAGE_SIZE
163 	 */
164 	stride = PAGE_SIZE;
165 	start = round_down(start, stride);
166 
167 	dsb(ishst);
168 
169 	/* Switch to requested VMID */
170 	enter_vmid_context(mmu, &cxt);
171 
172 	__flush_s2_tlb_range_op(ipas2e1is, start, pages, stride,
173 				TLBI_TTL_UNKNOWN);
174 
175 	dsb(ish);
176 	__tlbi(vmalle1is);
177 	__tlbi_sync_s1ish_hyp();
178 	isb();
179 
180 	exit_vmid_context(&cxt);
181 }
182 
183 void __kvm_tlb_flush_vmid(struct kvm_s2_mmu *mmu)
184 {
185 	struct tlb_inv_context cxt;
186 
187 	dsb(ishst);
188 
189 	/* Switch to requested VMID */
190 	enter_vmid_context(mmu, &cxt);
191 
192 	__tlbi(vmalls12e1is);
193 	__tlbi_sync_s1ish_hyp();
194 	isb();
195 
196 	exit_vmid_context(&cxt);
197 }
198 
199 void __kvm_flush_cpu_context(struct kvm_s2_mmu *mmu)
200 {
201 	struct tlb_inv_context cxt;
202 
203 	/* Switch to requested VMID */
204 	enter_vmid_context(mmu, &cxt);
205 
206 	__tlbi(vmalle1);
207 	asm volatile("ic iallu");
208 	dsb(nsh);
209 	isb();
210 
211 	exit_vmid_context(&cxt);
212 }
213 
214 void __kvm_flush_vm_context(void)
215 {
216 	dsb(ishst);
217 	__tlbi(alle1is);
218 	__tlbi_sync_s1ish_hyp();
219 }
220 
221 /*
222  * TLB invalidation emulation for NV. For any given instruction, we
223  * perform the following transformtions:
224  *
225  * - a TLBI targeting EL2 S1 is remapped to EL1 S1
226  * - a non-shareable TLBI is upgraded to being inner-shareable
227  * - an outer-shareable TLBI is also mapped to inner-shareable
228  * - an nXS TLBI is upgraded to XS
229  */
230 int __kvm_tlbi_s1e2(struct kvm_s2_mmu *mmu, u64 va, u64 sys_encoding)
231 {
232 	struct tlb_inv_context cxt;
233 	int ret = 0;
234 
235 	/*
236 	 * The guest will have provided its own DSB ISHST before trapping.
237 	 * If it hasn't, that's its own problem, and we won't paper over it
238 	 * (plus, there is plenty of extra synchronisation before we even
239 	 * get here...).
240 	 */
241 
242 	if (mmu)
243 		enter_vmid_context(mmu, &cxt);
244 
245 	switch (sys_encoding) {
246 	case OP_TLBI_ALLE2:
247 	case OP_TLBI_ALLE2IS:
248 	case OP_TLBI_ALLE2OS:
249 	case OP_TLBI_VMALLE1:
250 	case OP_TLBI_VMALLE1IS:
251 	case OP_TLBI_VMALLE1OS:
252 	case OP_TLBI_ALLE2NXS:
253 	case OP_TLBI_ALLE2ISNXS:
254 	case OP_TLBI_ALLE2OSNXS:
255 	case OP_TLBI_VMALLE1NXS:
256 	case OP_TLBI_VMALLE1ISNXS:
257 	case OP_TLBI_VMALLE1OSNXS:
258 		__tlbi(vmalle1is);
259 		break;
260 	case OP_TLBI_VAE2:
261 	case OP_TLBI_VAE2IS:
262 	case OP_TLBI_VAE2OS:
263 	case OP_TLBI_VAE1:
264 	case OP_TLBI_VAE1IS:
265 	case OP_TLBI_VAE1OS:
266 	case OP_TLBI_VAE2NXS:
267 	case OP_TLBI_VAE2ISNXS:
268 	case OP_TLBI_VAE2OSNXS:
269 	case OP_TLBI_VAE1NXS:
270 	case OP_TLBI_VAE1ISNXS:
271 	case OP_TLBI_VAE1OSNXS:
272 		__tlbi(vae1is, va);
273 		break;
274 	case OP_TLBI_VALE2:
275 	case OP_TLBI_VALE2IS:
276 	case OP_TLBI_VALE2OS:
277 	case OP_TLBI_VALE1:
278 	case OP_TLBI_VALE1IS:
279 	case OP_TLBI_VALE1OS:
280 	case OP_TLBI_VALE2NXS:
281 	case OP_TLBI_VALE2ISNXS:
282 	case OP_TLBI_VALE2OSNXS:
283 	case OP_TLBI_VALE1NXS:
284 	case OP_TLBI_VALE1ISNXS:
285 	case OP_TLBI_VALE1OSNXS:
286 		__tlbi(vale1is, va);
287 		break;
288 	case OP_TLBI_ASIDE1:
289 	case OP_TLBI_ASIDE1IS:
290 	case OP_TLBI_ASIDE1OS:
291 	case OP_TLBI_ASIDE1NXS:
292 	case OP_TLBI_ASIDE1ISNXS:
293 	case OP_TLBI_ASIDE1OSNXS:
294 		__tlbi(aside1is, va);
295 		break;
296 	case OP_TLBI_VAAE1:
297 	case OP_TLBI_VAAE1IS:
298 	case OP_TLBI_VAAE1OS:
299 	case OP_TLBI_VAAE1NXS:
300 	case OP_TLBI_VAAE1ISNXS:
301 	case OP_TLBI_VAAE1OSNXS:
302 		__tlbi(vaae1is, va);
303 		break;
304 	case OP_TLBI_VAALE1:
305 	case OP_TLBI_VAALE1IS:
306 	case OP_TLBI_VAALE1OS:
307 	case OP_TLBI_VAALE1NXS:
308 	case OP_TLBI_VAALE1ISNXS:
309 	case OP_TLBI_VAALE1OSNXS:
310 		__tlbi(vaale1is, va);
311 		break;
312 	case OP_TLBI_RVAE2:
313 	case OP_TLBI_RVAE2IS:
314 	case OP_TLBI_RVAE2OS:
315 	case OP_TLBI_RVAE1:
316 	case OP_TLBI_RVAE1IS:
317 	case OP_TLBI_RVAE1OS:
318 	case OP_TLBI_RVAE2NXS:
319 	case OP_TLBI_RVAE2ISNXS:
320 	case OP_TLBI_RVAE2OSNXS:
321 	case OP_TLBI_RVAE1NXS:
322 	case OP_TLBI_RVAE1ISNXS:
323 	case OP_TLBI_RVAE1OSNXS:
324 		__tlbi(rvae1is, va);
325 		break;
326 	case OP_TLBI_RVALE2:
327 	case OP_TLBI_RVALE2IS:
328 	case OP_TLBI_RVALE2OS:
329 	case OP_TLBI_RVALE1:
330 	case OP_TLBI_RVALE1IS:
331 	case OP_TLBI_RVALE1OS:
332 	case OP_TLBI_RVALE2NXS:
333 	case OP_TLBI_RVALE2ISNXS:
334 	case OP_TLBI_RVALE2OSNXS:
335 	case OP_TLBI_RVALE1NXS:
336 	case OP_TLBI_RVALE1ISNXS:
337 	case OP_TLBI_RVALE1OSNXS:
338 		__tlbi(rvale1is, va);
339 		break;
340 	case OP_TLBI_RVAAE1:
341 	case OP_TLBI_RVAAE1IS:
342 	case OP_TLBI_RVAAE1OS:
343 	case OP_TLBI_RVAAE1NXS:
344 	case OP_TLBI_RVAAE1ISNXS:
345 	case OP_TLBI_RVAAE1OSNXS:
346 		__tlbi(rvaae1is, va);
347 		break;
348 	case OP_TLBI_RVAALE1:
349 	case OP_TLBI_RVAALE1IS:
350 	case OP_TLBI_RVAALE1OS:
351 	case OP_TLBI_RVAALE1NXS:
352 	case OP_TLBI_RVAALE1ISNXS:
353 	case OP_TLBI_RVAALE1OSNXS:
354 		__tlbi(rvaale1is, va);
355 		break;
356 	default:
357 		ret = -EINVAL;
358 	}
359 	__tlbi_sync_s1ish_hyp();
360 	isb();
361 
362 	if (mmu)
363 		exit_vmid_context(&cxt);
364 
365 	return ret;
366 }
367