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 hypervisor VA to a PA 58 * reg: hypervisor address to be converted in place 59 * tmp: temporary register 60 */ 61.macro hyp_pa reg, tmp 62 ldr_l \tmp, hyp_physvirt_offset 63 add \reg, \reg, \tmp 64.endm 65 66/* 67 * Convert a hypervisor VA to a kernel image address 68 * reg: hypervisor address to be converted in place 69 * tmp: temporary register 70 * 71 * The actual code generation takes place in kvm_get_kimage_voffset, and 72 * the instructions below are only there to reserve the space and 73 * perform the register allocation (kvm_get_kimage_voffset uses the 74 * specific registers encoded in the instructions). 75 */ 76.macro hyp_kimg_va reg, tmp 77 /* Convert hyp VA -> PA. */ 78 hyp_pa \reg, \tmp 79 80 /* Load kimage_voffset. */ 81alternative_cb ARM64_ALWAYS_SYSTEM, kvm_get_kimage_voffset 82 movz \tmp, #0 83 movk \tmp, #0, lsl #16 84 movk \tmp, #0, lsl #32 85 movk \tmp, #0, lsl #48 86alternative_cb_end 87 88 /* Convert PA -> kimg VA. */ 89 add \reg, \reg, \tmp 90.endm 91 92#else 93 94#include <linux/pgtable.h> 95#include <asm/pgalloc.h> 96#include <asm/cache.h> 97#include <asm/cacheflush.h> 98#include <asm/mmu_context.h> 99#include <asm/kvm_emulate.h> 100#include <asm/kvm_host.h> 101 102void kvm_update_va_mask(struct alt_instr *alt, 103 __le32 *origptr, __le32 *updptr, int nr_inst); 104void kvm_compute_layout(void); 105void kvm_apply_hyp_relocations(void); 106 107#define __hyp_pa(x) (((phys_addr_t)(x)) + hyp_physvirt_offset) 108 109/* 110 * Convert a kernel VA into a HYP VA. 111 * 112 * Can be called from hyp or non-hyp context. 113 * 114 * The actual code generation takes place in kvm_update_va_mask(), and 115 * the instructions below are only there to reserve the space and 116 * perform the register allocation (kvm_update_va_mask() uses the 117 * specific registers encoded in the instructions). 118 */ 119static __always_inline unsigned long __kern_hyp_va(unsigned long v) 120{ 121/* 122 * This #ifndef is an optimisation for when this is called from VHE hyp 123 * context. When called from a VHE non-hyp context, kvm_update_va_mask() will 124 * replace the instructions with `nop`s. 125 */ 126#ifndef __KVM_VHE_HYPERVISOR__ 127 asm volatile(ALTERNATIVE_CB("and %0, %0, #1\n" /* mask with va_mask */ 128 "ror %0, %0, #1\n" /* rotate to the first tag bit */ 129 "add %0, %0, #0\n" /* insert the low 12 bits of the tag */ 130 "add %0, %0, #0, lsl 12\n" /* insert the top 12 bits of the tag */ 131 "ror %0, %0, #63\n", /* rotate back */ 132 ARM64_ALWAYS_SYSTEM, 133 kvm_update_va_mask) 134 : "+r" (v)); 135#endif 136 return v; 137} 138 139#define kern_hyp_va(v) ((typeof(v))(__kern_hyp_va((unsigned long)(v)))) 140 141/* 142 * We currently support using a VM-specified IPA size. For backward 143 * compatibility, the default IPA size is fixed to 40bits. 144 */ 145#define KVM_PHYS_SHIFT (40) 146 147#define kvm_phys_shift(mmu) VTCR_EL2_IPA((mmu)->vtcr) 148#define kvm_phys_size(mmu) (_AC(1, ULL) << kvm_phys_shift(mmu)) 149#define kvm_phys_mask(mmu) (kvm_phys_size(mmu) - _AC(1, ULL)) 150 151#include <asm/kvm_pgtable.h> 152#include <asm/stage2_pgtable.h> 153 154int kvm_share_hyp(void *from, void *to); 155void kvm_unshare_hyp(void *from, void *to); 156int create_hyp_mappings(void *from, void *to, enum kvm_pgtable_prot prot); 157int __create_hyp_mappings(unsigned long start, unsigned long size, 158 unsigned long phys, enum kvm_pgtable_prot prot); 159int hyp_alloc_private_va_range(size_t size, unsigned long *haddr); 160int create_hyp_io_mappings(phys_addr_t phys_addr, size_t size, 161 void __iomem **kaddr, 162 void __iomem **haddr); 163int create_hyp_exec_mappings(phys_addr_t phys_addr, size_t size, 164 void **haddr); 165int create_hyp_stack(phys_addr_t phys_addr, unsigned long *haddr); 166void __init free_hyp_pgds(void); 167 168void stage2_unmap_vm(struct kvm *kvm); 169int kvm_init_stage2_mmu(struct kvm *kvm, struct kvm_s2_mmu *mmu, unsigned long type); 170void kvm_uninit_stage2_mmu(struct kvm *kvm); 171void kvm_free_stage2_pgd(struct kvm_s2_mmu *mmu); 172int kvm_phys_addr_ioremap(struct kvm *kvm, phys_addr_t guest_ipa, 173 phys_addr_t pa, unsigned long size, bool writable); 174 175int kvm_handle_guest_abort(struct kvm_vcpu *vcpu); 176 177phys_addr_t kvm_mmu_get_httbr(void); 178phys_addr_t kvm_get_idmap_vector(void); 179int __init kvm_mmu_init(u32 *hyp_va_bits); 180 181static inline void *__kvm_vector_slot2addr(void *base, 182 enum arm64_hyp_spectre_vector slot) 183{ 184 int idx = slot - (slot != HYP_VECTOR_DIRECT); 185 186 return base + (idx * SZ_2K); 187} 188 189struct kvm; 190 191#define kvm_flush_dcache_to_poc(a,l) \ 192 dcache_clean_inval_poc((unsigned long)(a), (unsigned long)(a)+(l)) 193 194static inline bool vcpu_has_cache_enabled(struct kvm_vcpu *vcpu) 195{ 196 u64 cache_bits = SCTLR_ELx_M | SCTLR_ELx_C; 197 int reg; 198 199 if (vcpu_is_el2(vcpu)) 200 reg = SCTLR_EL2; 201 else 202 reg = SCTLR_EL1; 203 204 return (vcpu_read_sys_reg(vcpu, reg) & cache_bits) == cache_bits; 205} 206 207static inline void __clean_dcache_guest_page(void *va, size_t size) 208{ 209 /* 210 * With FWB, we ensure that the guest always accesses memory using 211 * cacheable attributes, and we don't have to clean to PoC when 212 * faulting in pages. Furthermore, FWB implies IDC, so cleaning to 213 * PoU is not required either in this case. 214 */ 215 if (cpus_have_final_cap(ARM64_HAS_STAGE2_FWB)) 216 return; 217 218 kvm_flush_dcache_to_poc(va, size); 219} 220 221static inline size_t __invalidate_icache_max_range(void) 222{ 223 u8 iminline; 224 u64 ctr; 225 226 asm volatile(ALTERNATIVE_CB("movz %0, #0\n" 227 "movk %0, #0, lsl #16\n" 228 "movk %0, #0, lsl #32\n" 229 "movk %0, #0, lsl #48\n", 230 ARM64_ALWAYS_SYSTEM, 231 kvm_compute_final_ctr_el0) 232 : "=r" (ctr)); 233 234 iminline = SYS_FIELD_GET(CTR_EL0, IminLine, ctr) + 2; 235 return MAX_DVM_OPS << iminline; 236} 237 238static inline void __invalidate_icache_guest_page(void *va, size_t size) 239{ 240 /* 241 * Blow the whole I-cache if it is aliasing (i.e. VIPT) or the 242 * invalidation range exceeds our arbitrary limit on invadations by 243 * cache line. 244 */ 245 if (icache_is_aliasing() || size > __invalidate_icache_max_range()) 246 icache_inval_all_pou(); 247 else 248 icache_inval_pou((unsigned long)va, (unsigned long)va + size); 249} 250 251void kvm_set_way_flush(struct kvm_vcpu *vcpu); 252void kvm_toggle_cache(struct kvm_vcpu *vcpu, bool was_enabled); 253 254static inline unsigned int kvm_get_vmid_bits(void) 255{ 256 int reg = read_sanitised_ftr_reg(SYS_ID_AA64MMFR1_EL1); 257 258 return get_vmid_bits(reg); 259} 260 261/* 262 * We are not in the kvm->srcu critical section most of the time, so we take 263 * the SRCU read lock here. Since we copy the data from the user page, we 264 * can immediately drop the lock again. 265 */ 266static inline int kvm_read_guest_lock(struct kvm *kvm, 267 gpa_t gpa, void *data, unsigned long len) 268{ 269 int srcu_idx = srcu_read_lock(&kvm->srcu); 270 int ret = kvm_read_guest(kvm, gpa, data, len); 271 272 srcu_read_unlock(&kvm->srcu, srcu_idx); 273 274 return ret; 275} 276 277static inline int kvm_write_guest_lock(struct kvm *kvm, gpa_t gpa, 278 const void *data, unsigned long len) 279{ 280 int srcu_idx = srcu_read_lock(&kvm->srcu); 281 int ret = kvm_write_guest(kvm, gpa, data, len); 282 283 srcu_read_unlock(&kvm->srcu, srcu_idx); 284 285 return ret; 286} 287 288#define kvm_phys_to_vttbr(addr) phys_to_ttbr(addr) 289 290/* 291 * When this is (directly or indirectly) used on the TLB invalidation 292 * path, we rely on a previously issued DSB so that page table updates 293 * and VMID reads are correctly ordered. 294 */ 295static __always_inline u64 kvm_get_vttbr(struct kvm_s2_mmu *mmu) 296{ 297 struct kvm_vmid *vmid = &mmu->vmid; 298 u64 vmid_field, baddr; 299 u64 cnp = system_supports_cnp() ? VTTBR_CNP_BIT : 0; 300 301 baddr = mmu->pgd_phys; 302 vmid_field = atomic64_read(&vmid->id) << VTTBR_VMID_SHIFT; 303 vmid_field &= VTTBR_VMID_MASK(kvm_arm_vmid_bits); 304 return kvm_phys_to_vttbr(baddr) | vmid_field | cnp; 305} 306 307/* 308 * Must be called from hyp code running at EL2 with an updated VTTBR 309 * and interrupts disabled. 310 */ 311static __always_inline void __load_stage2(struct kvm_s2_mmu *mmu, 312 struct kvm_arch *arch) 313{ 314 write_sysreg(mmu->vtcr, vtcr_el2); 315 write_sysreg(kvm_get_vttbr(mmu), vttbr_el2); 316 317 /* 318 * ARM errata 1165522 and 1530923 require the actual execution of the 319 * above before we can switch to the EL1/EL0 translation regime used by 320 * the guest. 321 */ 322 asm(ALTERNATIVE("nop", "isb", ARM64_WORKAROUND_SPECULATIVE_AT)); 323} 324 325static inline struct kvm *kvm_s2_mmu_to_kvm(struct kvm_s2_mmu *mmu) 326{ 327 return container_of(mmu->arch, struct kvm, arch); 328} 329#endif /* __ASSEMBLY__ */ 330#endif /* __ARM64_KVM_MMU_H__ */ 331