xref: /linux/arch/arm64/include/asm/kvm_mmu.h (revision 8e07e0e3964ca4e23ce7b68e2096fe660a888942)
1 /* SPDX-License-Identifier: GPL-2.0-only */
2 /*
3  * Copyright (C) 2012,2013 - ARM Ltd
4  * Author: Marc Zyngier <marc.zyngier@arm.com>
5  */
6 
7 #ifndef __ARM64_KVM_MMU_H__
8 #define __ARM64_KVM_MMU_H__
9 
10 #include <asm/page.h>
11 #include <asm/memory.h>
12 #include <asm/mmu.h>
13 #include <asm/cpufeature.h>
14 
15 /*
16  * As ARMv8.0 only has the TTBR0_EL2 register, we cannot express
17  * "negative" addresses. This makes it impossible to directly share
18  * mappings with the kernel.
19  *
20  * Instead, give the HYP mode its own VA region at a fixed offset from
21  * the kernel by just masking the top bits (which are all ones for a
22  * kernel address). We need to find out how many bits to mask.
23  *
24  * We want to build a set of page tables that cover both parts of the
25  * idmap (the trampoline page used to initialize EL2), and our normal
26  * runtime VA space, at the same time.
27  *
28  * Given that the kernel uses VA_BITS for its entire address space,
29  * and that half of that space (VA_BITS - 1) is used for the linear
30  * mapping, we can also limit the EL2 space to (VA_BITS - 1).
31  *
32  * The main question is "Within the VA_BITS space, does EL2 use the
33  * top or the bottom half of that space to shadow the kernel's linear
34  * mapping?". As we need to idmap the trampoline page, this is
35  * determined by the range in which this page lives.
36  *
37  * If the page is in the bottom half, we have to use the top half. If
38  * the page is in the top half, we have to use the bottom half:
39  *
40  * T = __pa_symbol(__hyp_idmap_text_start)
41  * if (T & BIT(VA_BITS - 1))
42  *	HYP_VA_MIN = 0  //idmap in upper half
43  * else
44  *	HYP_VA_MIN = 1 << (VA_BITS - 1)
45  * HYP_VA_MAX = HYP_VA_MIN + (1 << (VA_BITS - 1)) - 1
46  *
47  * When using VHE, there are no separate hyp mappings and all KVM
48  * functionality is already mapped as part of the main kernel
49  * mappings, and none of this applies in that case.
50  */
51 
52 #ifdef __ASSEMBLY__
53 
54 #include <asm/alternative.h>
55 
56 /*
57  * Convert a kernel VA into a HYP VA.
58  * reg: VA to be converted.
59  *
60  * The actual code generation takes place in kvm_update_va_mask, and
61  * the instructions below are only there to reserve the space and
62  * perform the register allocation (kvm_update_va_mask uses the
63  * specific registers encoded in the instructions).
64  */
65 .macro kern_hyp_va	reg
66 #ifndef __KVM_VHE_HYPERVISOR__
67 alternative_cb ARM64_ALWAYS_SYSTEM, kvm_update_va_mask
68 	and     \reg, \reg, #1		/* mask with va_mask */
69 	ror	\reg, \reg, #1		/* rotate to the first tag bit */
70 	add	\reg, \reg, #0		/* insert the low 12 bits of the tag */
71 	add	\reg, \reg, #0, lsl 12	/* insert the top 12 bits of the tag */
72 	ror	\reg, \reg, #63		/* rotate back */
73 alternative_cb_end
74 #endif
75 .endm
76 
77 /*
78  * Convert a hypervisor VA to a PA
79  * reg: hypervisor address to be converted in place
80  * tmp: temporary register
81  */
82 .macro hyp_pa reg, tmp
83 	ldr_l	\tmp, hyp_physvirt_offset
84 	add	\reg, \reg, \tmp
85 .endm
86 
87 /*
88  * Convert a hypervisor VA to a kernel image address
89  * reg: hypervisor address to be converted in place
90  * tmp: temporary register
91  *
92  * The actual code generation takes place in kvm_get_kimage_voffset, and
93  * the instructions below are only there to reserve the space and
94  * perform the register allocation (kvm_get_kimage_voffset uses the
95  * specific registers encoded in the instructions).
96  */
97 .macro hyp_kimg_va reg, tmp
98 	/* Convert hyp VA -> PA. */
99 	hyp_pa	\reg, \tmp
100 
101 	/* Load kimage_voffset. */
102 alternative_cb ARM64_ALWAYS_SYSTEM, kvm_get_kimage_voffset
103 	movz	\tmp, #0
104 	movk	\tmp, #0, lsl #16
105 	movk	\tmp, #0, lsl #32
106 	movk	\tmp, #0, lsl #48
107 alternative_cb_end
108 
109 	/* Convert PA -> kimg VA. */
110 	add	\reg, \reg, \tmp
111 .endm
112 
113 #else
114 
115 #include <linux/pgtable.h>
116 #include <asm/pgalloc.h>
117 #include <asm/cache.h>
118 #include <asm/cacheflush.h>
119 #include <asm/mmu_context.h>
120 #include <asm/kvm_emulate.h>
121 #include <asm/kvm_host.h>
122 
123 void kvm_update_va_mask(struct alt_instr *alt,
124 			__le32 *origptr, __le32 *updptr, int nr_inst);
125 void kvm_compute_layout(void);
126 void kvm_apply_hyp_relocations(void);
127 
128 #define __hyp_pa(x) (((phys_addr_t)(x)) + hyp_physvirt_offset)
129 
130 static __always_inline unsigned long __kern_hyp_va(unsigned long v)
131 {
132 #ifndef __KVM_VHE_HYPERVISOR__
133 	asm volatile(ALTERNATIVE_CB("and %0, %0, #1\n"
134 				    "ror %0, %0, #1\n"
135 				    "add %0, %0, #0\n"
136 				    "add %0, %0, #0, lsl 12\n"
137 				    "ror %0, %0, #63\n",
138 				    ARM64_ALWAYS_SYSTEM,
139 				    kvm_update_va_mask)
140 		     : "+r" (v));
141 #endif
142 	return v;
143 }
144 
145 #define kern_hyp_va(v) 	((typeof(v))(__kern_hyp_va((unsigned long)(v))))
146 
147 /*
148  * We currently support using a VM-specified IPA size. For backward
149  * compatibility, the default IPA size is fixed to 40bits.
150  */
151 #define KVM_PHYS_SHIFT	(40)
152 
153 #define kvm_phys_shift(mmu)		VTCR_EL2_IPA((mmu)->vtcr)
154 #define kvm_phys_size(mmu)		(_AC(1, ULL) << kvm_phys_shift(mmu))
155 #define kvm_phys_mask(mmu)		(kvm_phys_size(mmu) - _AC(1, ULL))
156 
157 #include <asm/kvm_pgtable.h>
158 #include <asm/stage2_pgtable.h>
159 
160 int kvm_share_hyp(void *from, void *to);
161 void kvm_unshare_hyp(void *from, void *to);
162 int create_hyp_mappings(void *from, void *to, enum kvm_pgtable_prot prot);
163 int __create_hyp_mappings(unsigned long start, unsigned long size,
164 			  unsigned long phys, enum kvm_pgtable_prot prot);
165 int hyp_alloc_private_va_range(size_t size, unsigned long *haddr);
166 int create_hyp_io_mappings(phys_addr_t phys_addr, size_t size,
167 			   void __iomem **kaddr,
168 			   void __iomem **haddr);
169 int create_hyp_exec_mappings(phys_addr_t phys_addr, size_t size,
170 			     void **haddr);
171 int create_hyp_stack(phys_addr_t phys_addr, unsigned long *haddr);
172 void __init free_hyp_pgds(void);
173 
174 void stage2_unmap_vm(struct kvm *kvm);
175 int kvm_init_stage2_mmu(struct kvm *kvm, struct kvm_s2_mmu *mmu, unsigned long type);
176 void kvm_uninit_stage2_mmu(struct kvm *kvm);
177 void kvm_free_stage2_pgd(struct kvm_s2_mmu *mmu);
178 int kvm_phys_addr_ioremap(struct kvm *kvm, phys_addr_t guest_ipa,
179 			  phys_addr_t pa, unsigned long size, bool writable);
180 
181 int kvm_handle_guest_abort(struct kvm_vcpu *vcpu);
182 
183 phys_addr_t kvm_mmu_get_httbr(void);
184 phys_addr_t kvm_get_idmap_vector(void);
185 int __init kvm_mmu_init(u32 *hyp_va_bits);
186 
187 static inline void *__kvm_vector_slot2addr(void *base,
188 					   enum arm64_hyp_spectre_vector slot)
189 {
190 	int idx = slot - (slot != HYP_VECTOR_DIRECT);
191 
192 	return base + (idx * SZ_2K);
193 }
194 
195 struct kvm;
196 
197 #define kvm_flush_dcache_to_poc(a,l)	\
198 	dcache_clean_inval_poc((unsigned long)(a), (unsigned long)(a)+(l))
199 
200 static inline bool vcpu_has_cache_enabled(struct kvm_vcpu *vcpu)
201 {
202 	u64 cache_bits = SCTLR_ELx_M | SCTLR_ELx_C;
203 	int reg;
204 
205 	if (vcpu_is_el2(vcpu))
206 		reg = SCTLR_EL2;
207 	else
208 		reg = SCTLR_EL1;
209 
210 	return (vcpu_read_sys_reg(vcpu, reg) & cache_bits) == cache_bits;
211 }
212 
213 static inline void __clean_dcache_guest_page(void *va, size_t size)
214 {
215 	/*
216 	 * With FWB, we ensure that the guest always accesses memory using
217 	 * cacheable attributes, and we don't have to clean to PoC when
218 	 * faulting in pages. Furthermore, FWB implies IDC, so cleaning to
219 	 * PoU is not required either in this case.
220 	 */
221 	if (cpus_have_final_cap(ARM64_HAS_STAGE2_FWB))
222 		return;
223 
224 	kvm_flush_dcache_to_poc(va, size);
225 }
226 
227 static inline size_t __invalidate_icache_max_range(void)
228 {
229 	u8 iminline;
230 	u64 ctr;
231 
232 	asm volatile(ALTERNATIVE_CB("movz %0, #0\n"
233 				    "movk %0, #0, lsl #16\n"
234 				    "movk %0, #0, lsl #32\n"
235 				    "movk %0, #0, lsl #48\n",
236 				    ARM64_ALWAYS_SYSTEM,
237 				    kvm_compute_final_ctr_el0)
238 		     : "=r" (ctr));
239 
240 	iminline = SYS_FIELD_GET(CTR_EL0, IminLine, ctr) + 2;
241 	return MAX_DVM_OPS << iminline;
242 }
243 
244 static inline void __invalidate_icache_guest_page(void *va, size_t size)
245 {
246 	/*
247 	 * VPIPT I-cache maintenance must be done from EL2. See comment in the
248 	 * nVHE flavor of __kvm_tlb_flush_vmid_ipa().
249 	 */
250 	if (icache_is_vpipt() && read_sysreg(CurrentEL) != CurrentEL_EL2)
251 		return;
252 
253 	/*
254 	 * Blow the whole I-cache if it is aliasing (i.e. VIPT) or the
255 	 * invalidation range exceeds our arbitrary limit on invadations by
256 	 * cache line.
257 	 */
258 	if (icache_is_aliasing() || size > __invalidate_icache_max_range())
259 		icache_inval_all_pou();
260 	else
261 		icache_inval_pou((unsigned long)va, (unsigned long)va + size);
262 }
263 
264 void kvm_set_way_flush(struct kvm_vcpu *vcpu);
265 void kvm_toggle_cache(struct kvm_vcpu *vcpu, bool was_enabled);
266 
267 static inline unsigned int kvm_get_vmid_bits(void)
268 {
269 	int reg = read_sanitised_ftr_reg(SYS_ID_AA64MMFR1_EL1);
270 
271 	return get_vmid_bits(reg);
272 }
273 
274 /*
275  * We are not in the kvm->srcu critical section most of the time, so we take
276  * the SRCU read lock here. Since we copy the data from the user page, we
277  * can immediately drop the lock again.
278  */
279 static inline int kvm_read_guest_lock(struct kvm *kvm,
280 				      gpa_t gpa, void *data, unsigned long len)
281 {
282 	int srcu_idx = srcu_read_lock(&kvm->srcu);
283 	int ret = kvm_read_guest(kvm, gpa, data, len);
284 
285 	srcu_read_unlock(&kvm->srcu, srcu_idx);
286 
287 	return ret;
288 }
289 
290 static inline int kvm_write_guest_lock(struct kvm *kvm, gpa_t gpa,
291 				       const void *data, unsigned long len)
292 {
293 	int srcu_idx = srcu_read_lock(&kvm->srcu);
294 	int ret = kvm_write_guest(kvm, gpa, data, len);
295 
296 	srcu_read_unlock(&kvm->srcu, srcu_idx);
297 
298 	return ret;
299 }
300 
301 #define kvm_phys_to_vttbr(addr)		phys_to_ttbr(addr)
302 
303 /*
304  * When this is (directly or indirectly) used on the TLB invalidation
305  * path, we rely on a previously issued DSB so that page table updates
306  * and VMID reads are correctly ordered.
307  */
308 static __always_inline u64 kvm_get_vttbr(struct kvm_s2_mmu *mmu)
309 {
310 	struct kvm_vmid *vmid = &mmu->vmid;
311 	u64 vmid_field, baddr;
312 	u64 cnp = system_supports_cnp() ? VTTBR_CNP_BIT : 0;
313 
314 	baddr = mmu->pgd_phys;
315 	vmid_field = atomic64_read(&vmid->id) << VTTBR_VMID_SHIFT;
316 	vmid_field &= VTTBR_VMID_MASK(kvm_arm_vmid_bits);
317 	return kvm_phys_to_vttbr(baddr) | vmid_field | cnp;
318 }
319 
320 /*
321  * Must be called from hyp code running at EL2 with an updated VTTBR
322  * and interrupts disabled.
323  */
324 static __always_inline void __load_stage2(struct kvm_s2_mmu *mmu,
325 					  struct kvm_arch *arch)
326 {
327 	write_sysreg(mmu->vtcr, vtcr_el2);
328 	write_sysreg(kvm_get_vttbr(mmu), vttbr_el2);
329 
330 	/*
331 	 * ARM errata 1165522 and 1530923 require the actual execution of the
332 	 * above before we can switch to the EL1/EL0 translation regime used by
333 	 * the guest.
334 	 */
335 	asm(ALTERNATIVE("nop", "isb", ARM64_WORKAROUND_SPECULATIVE_AT));
336 }
337 
338 static inline struct kvm *kvm_s2_mmu_to_kvm(struct kvm_s2_mmu *mmu)
339 {
340 	return container_of(mmu->arch, struct kvm, arch);
341 }
342 #endif /* __ASSEMBLY__ */
343 #endif /* __ARM64_KVM_MMU_H__ */
344