xref: /linux/arch/arm64/kvm/hyp/nvhe/tlb.c (revision 3d0fe49454652117522f60bfbefb978ba0e5300b)
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 	u64		tcr;
15 };
16 
17 static void __tlb_switch_to_guest(struct kvm_s2_mmu *mmu,
18 				  struct tlb_inv_context *cxt,
19 				  bool nsh)
20 {
21 	/*
22 	 * We have two requirements:
23 	 *
24 	 * - ensure that the page table updates are visible to all
25 	 *   CPUs, for which a dsb(DOMAIN-st) is what we need, DOMAIN
26 	 *   being either ish or nsh, depending on the invalidation
27 	 *   type.
28 	 *
29 	 * - complete any speculative page table walk started before
30 	 *   we trapped to EL2 so that we can mess with the MM
31 	 *   registers out of context, for which dsb(nsh) is enough
32 	 *
33 	 * The composition of these two barriers is a dsb(DOMAIN), and
34 	 * the 'nsh' parameter tracks the distinction between
35 	 * Inner-Shareable and Non-Shareable, as specified by the
36 	 * callers.
37 	 */
38 	if (nsh)
39 		dsb(nsh);
40 	else
41 		dsb(ish);
42 
43 	if (cpus_have_final_cap(ARM64_WORKAROUND_SPECULATIVE_AT)) {
44 		u64 val;
45 
46 		/*
47 		 * For CPUs that are affected by ARM 1319367, we need to
48 		 * avoid a host Stage-1 walk while we have the guest's
49 		 * VMID set in the VTTBR in order to invalidate TLBs.
50 		 * We're guaranteed that the S1 MMU is enabled, so we can
51 		 * simply set the EPD bits to avoid any further TLB fill.
52 		 */
53 		val = cxt->tcr = read_sysreg_el1(SYS_TCR);
54 		val |= TCR_EPD1_MASK | TCR_EPD0_MASK;
55 		write_sysreg_el1(val, SYS_TCR);
56 		isb();
57 	}
58 
59 	/*
60 	 * __load_stage2() includes an ISB only when the AT
61 	 * workaround is applied. Take care of the opposite condition,
62 	 * ensuring that we always have an ISB, but not two ISBs back
63 	 * to back.
64 	 */
65 	__load_stage2(mmu, kern_hyp_va(mmu->arch));
66 	asm(ALTERNATIVE("isb", "nop", ARM64_WORKAROUND_SPECULATIVE_AT));
67 }
68 
69 static void __tlb_switch_to_host(struct tlb_inv_context *cxt)
70 {
71 	__load_host_stage2();
72 
73 	if (cpus_have_final_cap(ARM64_WORKAROUND_SPECULATIVE_AT)) {
74 		/* Ensure write of the host VMID */
75 		isb();
76 		/* Restore the host's TCR_EL1 */
77 		write_sysreg_el1(cxt->tcr, SYS_TCR);
78 	}
79 }
80 
81 void __kvm_tlb_flush_vmid_ipa(struct kvm_s2_mmu *mmu,
82 			      phys_addr_t ipa, int level)
83 {
84 	struct tlb_inv_context cxt;
85 
86 	/* Switch to requested VMID */
87 	__tlb_switch_to_guest(mmu, &cxt, false);
88 
89 	/*
90 	 * We could do so much better if we had the VA as well.
91 	 * Instead, we invalidate Stage-2 for this IPA, and the
92 	 * whole of Stage-1. Weep...
93 	 */
94 	ipa >>= 12;
95 	__tlbi_level(ipas2e1is, ipa, level);
96 
97 	/*
98 	 * We have to ensure completion of the invalidation at Stage-2,
99 	 * since a table walk on another CPU could refill a TLB with a
100 	 * complete (S1 + S2) walk based on the old Stage-2 mapping if
101 	 * the Stage-1 invalidation happened first.
102 	 */
103 	dsb(ish);
104 	__tlbi(vmalle1is);
105 	dsb(ish);
106 	isb();
107 
108 	/*
109 	 * If the host is running at EL1 and we have a VPIPT I-cache,
110 	 * then we must perform I-cache maintenance at EL2 in order for
111 	 * it to have an effect on the guest. Since the guest cannot hit
112 	 * I-cache lines allocated with a different VMID, we don't need
113 	 * to worry about junk out of guest reset (we nuke the I-cache on
114 	 * VMID rollover), but we do need to be careful when remapping
115 	 * executable pages for the same guest. This can happen when KSM
116 	 * takes a CoW fault on an executable page, copies the page into
117 	 * a page that was previously mapped in the guest and then needs
118 	 * to invalidate the guest view of the I-cache for that page
119 	 * from EL1. To solve this, we invalidate the entire I-cache when
120 	 * unmapping a page from a guest if we have a VPIPT I-cache but
121 	 * the host is running at EL1. As above, we could do better if
122 	 * we had the VA.
123 	 *
124 	 * The moral of this story is: if you have a VPIPT I-cache, then
125 	 * you should be running with VHE enabled.
126 	 */
127 	if (icache_is_vpipt())
128 		icache_inval_all_pou();
129 
130 	__tlb_switch_to_host(&cxt);
131 }
132 
133 void __kvm_tlb_flush_vmid_ipa_nsh(struct kvm_s2_mmu *mmu,
134 				  phys_addr_t ipa, int level)
135 {
136 	struct tlb_inv_context cxt;
137 
138 	/* Switch to requested VMID */
139 	__tlb_switch_to_guest(mmu, &cxt, true);
140 
141 	/*
142 	 * We could do so much better if we had the VA as well.
143 	 * Instead, we invalidate Stage-2 for this IPA, and the
144 	 * whole of Stage-1. Weep...
145 	 */
146 	ipa >>= 12;
147 	__tlbi_level(ipas2e1, ipa, level);
148 
149 	/*
150 	 * We have to ensure completion of the invalidation at Stage-2,
151 	 * since a table walk on another CPU could refill a TLB with a
152 	 * complete (S1 + S2) walk based on the old Stage-2 mapping if
153 	 * the Stage-1 invalidation happened first.
154 	 */
155 	dsb(nsh);
156 	__tlbi(vmalle1);
157 	dsb(nsh);
158 	isb();
159 
160 	/*
161 	 * If the host is running at EL1 and we have a VPIPT I-cache,
162 	 * then we must perform I-cache maintenance at EL2 in order for
163 	 * it to have an effect on the guest. Since the guest cannot hit
164 	 * I-cache lines allocated with a different VMID, we don't need
165 	 * to worry about junk out of guest reset (we nuke the I-cache on
166 	 * VMID rollover), but we do need to be careful when remapping
167 	 * executable pages for the same guest. This can happen when KSM
168 	 * takes a CoW fault on an executable page, copies the page into
169 	 * a page that was previously mapped in the guest and then needs
170 	 * to invalidate the guest view of the I-cache for that page
171 	 * from EL1. To solve this, we invalidate the entire I-cache when
172 	 * unmapping a page from a guest if we have a VPIPT I-cache but
173 	 * the host is running at EL1. As above, we could do better if
174 	 * we had the VA.
175 	 *
176 	 * The moral of this story is: if you have a VPIPT I-cache, then
177 	 * you should be running with VHE enabled.
178 	 */
179 	if (icache_is_vpipt())
180 		icache_inval_all_pou();
181 
182 	__tlb_switch_to_host(&cxt);
183 }
184 
185 void __kvm_tlb_flush_vmid_range(struct kvm_s2_mmu *mmu,
186 				phys_addr_t start, unsigned long pages)
187 {
188 	struct tlb_inv_context cxt;
189 	unsigned long stride;
190 
191 	/*
192 	 * Since the range of addresses may not be mapped at
193 	 * the same level, assume the worst case as PAGE_SIZE
194 	 */
195 	stride = PAGE_SIZE;
196 	start = round_down(start, stride);
197 
198 	/* Switch to requested VMID */
199 	__tlb_switch_to_guest(mmu, &cxt, false);
200 
201 	__flush_s2_tlb_range_op(ipas2e1is, start, pages, stride, 0);
202 
203 	dsb(ish);
204 	__tlbi(vmalle1is);
205 	dsb(ish);
206 	isb();
207 
208 	/* See the comment in __kvm_tlb_flush_vmid_ipa() */
209 	if (icache_is_vpipt())
210 		icache_inval_all_pou();
211 
212 	__tlb_switch_to_host(&cxt);
213 }
214 
215 void __kvm_tlb_flush_vmid(struct kvm_s2_mmu *mmu)
216 {
217 	struct tlb_inv_context cxt;
218 
219 	/* Switch to requested VMID */
220 	__tlb_switch_to_guest(mmu, &cxt, false);
221 
222 	__tlbi(vmalls12e1is);
223 	dsb(ish);
224 	isb();
225 
226 	__tlb_switch_to_host(&cxt);
227 }
228 
229 void __kvm_flush_cpu_context(struct kvm_s2_mmu *mmu)
230 {
231 	struct tlb_inv_context cxt;
232 
233 	/* Switch to requested VMID */
234 	__tlb_switch_to_guest(mmu, &cxt, false);
235 
236 	__tlbi(vmalle1);
237 	asm volatile("ic iallu");
238 	dsb(nsh);
239 	isb();
240 
241 	__tlb_switch_to_host(&cxt);
242 }
243 
244 void __kvm_flush_vm_context(void)
245 {
246 	/* Same remark as in __tlb_switch_to_guest() */
247 	dsb(ish);
248 	__tlbi(alle1is);
249 
250 	/*
251 	 * VIPT and PIPT caches are not affected by VMID, so no maintenance
252 	 * is necessary across a VMID rollover.
253 	 *
254 	 * VPIPT caches constrain lookup and maintenance to the active VMID,
255 	 * so we need to invalidate lines with a stale VMID to avoid an ABA
256 	 * race after multiple rollovers.
257 	 *
258 	 */
259 	if (icache_is_vpipt())
260 		asm volatile("ic ialluis");
261 
262 	dsb(ish);
263 }
264