xref: /kernel/linux/linux-6.6/arch/arm64/kvm/hyp/nvhe/tlb.c (revision 62306a36)
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
13struct tlb_inv_context {
14	u64		tcr;
15};
16
17static 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
69static 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
81void __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
133void __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
185void __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
215void __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
229void __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
244void __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