| 1 | // SPDX-License-Identifier: GPL-2.0-only |
| 2 | /* |
| 3 | * Copyright (C) 2017 ARM Ltd. |
| 4 | * Author: Marc Zyngier <marc.zyngier@arm.com> |
| 5 | */ |
| 6 | |
| 7 | #include <linux/kvm_host.h> |
| 8 | #include <linux/random.h> |
| 9 | #include <linux/memblock.h> |
| 10 | #include <asm/alternative.h> |
| 11 | #include <asm/debug-monitors.h> |
| 12 | #include <asm/insn.h> |
| 13 | #include <asm/kvm_mmu.h> |
| 14 | #include <asm/memory.h> |
| 15 | |
| 16 | /* |
| 17 | * The LSB of the HYP VA tag |
| 18 | */ |
| 19 | static u8 tag_lsb; |
| 20 | /* |
| 21 | * The HYP VA tag value with the region bit |
| 22 | */ |
| 23 | static u64 tag_val; |
| 24 | static u64 va_mask; |
| 25 | |
| 26 | /* |
| 27 | * Compute HYP VA by using the same computation as kern_hyp_va(). |
| 28 | */ |
| 29 | static u64 __early_kern_hyp_va(u64 addr) |
| 30 | { |
| 31 | addr &= va_mask; |
| 32 | addr |= tag_val << tag_lsb; |
| 33 | return addr; |
| 34 | } |
| 35 | |
| 36 | /* |
| 37 | * Store a hyp VA <-> PA offset into a EL2-owned variable. |
| 38 | */ |
| 39 | static void init_hyp_physvirt_offset(void) |
| 40 | { |
| 41 | u64 kern_va, hyp_va; |
| 42 | |
| 43 | /* Compute the offset from the hyp VA and PA of a random symbol. */ |
| 44 | kern_va = (u64)lm_alias(__hyp_text_start); |
| 45 | hyp_va = __early_kern_hyp_va(addr: kern_va); |
| 46 | hyp_physvirt_offset = (s64)__pa(kern_va) - (s64)hyp_va; |
| 47 | } |
| 48 | |
| 49 | /* |
| 50 | * We want to generate a hyp VA with the following format (with V == |
| 51 | * vabits_actual): |
| 52 | * |
| 53 | * 63 ... V | V-1 | V-2 .. tag_lsb | tag_lsb - 1 .. 0 |
| 54 | * --------------------------------------------------------- |
| 55 | * | 0000000 | hyp_va_msb | random tag | kern linear VA | |
| 56 | * |--------- tag_val -----------|----- va_mask ---| |
| 57 | * |
| 58 | * which does not conflict with the idmap regions. |
| 59 | */ |
| 60 | __init void kvm_compute_layout(void) |
| 61 | { |
| 62 | phys_addr_t idmap_addr = __pa_symbol(__hyp_idmap_text_start); |
| 63 | u64 hyp_va_msb; |
| 64 | |
| 65 | /* Where is my RAM region? */ |
| 66 | hyp_va_msb = idmap_addr & BIT(vabits_actual - 1); |
| 67 | hyp_va_msb ^= BIT(vabits_actual - 1); |
| 68 | |
| 69 | tag_lsb = fls64(x: (u64)phys_to_virt(address: memblock_start_of_DRAM()) ^ |
| 70 | (u64)(high_memory - 1)); |
| 71 | |
| 72 | va_mask = GENMASK_ULL(tag_lsb - 1, 0); |
| 73 | tag_val = hyp_va_msb; |
| 74 | |
| 75 | if (IS_ENABLED(CONFIG_RANDOMIZE_BASE) && tag_lsb != (vabits_actual - 1)) { |
| 76 | /* We have some free bits to insert a random tag. */ |
| 77 | tag_val |= get_random_long() & GENMASK_ULL(vabits_actual - 2, tag_lsb); |
| 78 | } |
| 79 | tag_val >>= tag_lsb; |
| 80 | |
| 81 | init_hyp_physvirt_offset(); |
| 82 | } |
| 83 | |
| 84 | /* |
| 85 | * The .hyp.reloc ELF section contains a list of kimg positions that |
| 86 | * contains kimg VAs but will be accessed only in hyp execution context. |
| 87 | * Convert them to hyp VAs. See gen-hyprel.c for more details. |
| 88 | */ |
| 89 | __init void kvm_apply_hyp_relocations(void) |
| 90 | { |
| 91 | int32_t *rel; |
| 92 | int32_t *begin = (int32_t *)__hyp_reloc_begin; |
| 93 | int32_t *end = (int32_t *)__hyp_reloc_end; |
| 94 | |
| 95 | for (rel = begin; rel < end; ++rel) { |
| 96 | uintptr_t *ptr, kimg_va; |
| 97 | |
| 98 | /* |
| 99 | * Each entry contains a 32-bit relative offset from itself |
| 100 | * to a kimg VA position. |
| 101 | */ |
| 102 | ptr = (uintptr_t *)lm_alias((char *)rel + *rel); |
| 103 | |
| 104 | /* Read the kimg VA value at the relocation address. */ |
| 105 | kimg_va = *ptr; |
| 106 | |
| 107 | /* Convert to hyp VA and store back to the relocation address. */ |
| 108 | *ptr = __early_kern_hyp_va(addr: (uintptr_t)lm_alias(kimg_va)); |
| 109 | } |
| 110 | } |
| 111 | |
| 112 | static u32 compute_instruction(int n, u32 rd, u32 rn) |
| 113 | { |
| 114 | u32 insn = AARCH64_BREAK_FAULT; |
| 115 | |
| 116 | switch (n) { |
| 117 | case 0: |
| 118 | insn = aarch64_insn_gen_logical_immediate(AARCH64_INSN_LOGIC_AND, |
| 119 | AARCH64_INSN_VARIANT_64BIT, |
| 120 | rn, rd, va_mask); |
| 121 | break; |
| 122 | |
| 123 | case 1: |
| 124 | /* ROR is a variant of EXTR with Rm = Rn */ |
| 125 | insn = aarch64_insn_gen_extr(AARCH64_INSN_VARIANT_64BIT, |
| 126 | rn, rn, rd, |
| 127 | tag_lsb); |
| 128 | break; |
| 129 | |
| 130 | case 2: |
| 131 | insn = aarch64_insn_gen_add_sub_imm(rd, rn, |
| 132 | tag_val & GENMASK(11, 0), |
| 133 | AARCH64_INSN_VARIANT_64BIT, |
| 134 | AARCH64_INSN_ADSB_ADD); |
| 135 | break; |
| 136 | |
| 137 | case 3: |
| 138 | insn = aarch64_insn_gen_add_sub_imm(rd, rn, |
| 139 | tag_val & GENMASK(23, 12), |
| 140 | AARCH64_INSN_VARIANT_64BIT, |
| 141 | AARCH64_INSN_ADSB_ADD); |
| 142 | break; |
| 143 | |
| 144 | case 4: |
| 145 | /* ROR is a variant of EXTR with Rm = Rn */ |
| 146 | insn = aarch64_insn_gen_extr(AARCH64_INSN_VARIANT_64BIT, |
| 147 | rn, rn, rd, 64 - tag_lsb); |
| 148 | break; |
| 149 | } |
| 150 | |
| 151 | return insn; |
| 152 | } |
| 153 | |
| 154 | void __init kvm_update_va_mask(struct alt_instr *alt, |
| 155 | __le32 *origptr, __le32 *updptr, int nr_inst) |
| 156 | { |
| 157 | int i; |
| 158 | |
| 159 | BUG_ON(nr_inst != 5); |
| 160 | |
| 161 | for (i = 0; i < nr_inst; i++) { |
| 162 | u32 rd, rn, insn, oinsn; |
| 163 | |
| 164 | /* |
| 165 | * VHE doesn't need any address translation, let's NOP |
| 166 | * everything. |
| 167 | * |
| 168 | * Alternatively, if the tag is zero (because the layout |
| 169 | * dictates it and we don't have any spare bits in the |
| 170 | * address), NOP everything after masking the kernel VA. |
| 171 | */ |
| 172 | if (cpus_have_cap(ARM64_HAS_VIRT_HOST_EXTN) || (!tag_val && i > 0)) { |
| 173 | updptr[i] = cpu_to_le32(aarch64_insn_gen_nop()); |
| 174 | continue; |
| 175 | } |
| 176 | |
| 177 | oinsn = le32_to_cpu(origptr[i]); |
| 178 | rd = aarch64_insn_decode_register(AARCH64_INSN_REGTYPE_RD, oinsn); |
| 179 | rn = aarch64_insn_decode_register(AARCH64_INSN_REGTYPE_RN, oinsn); |
| 180 | |
| 181 | insn = compute_instruction(n: i, rd, rn); |
| 182 | BUG_ON(insn == AARCH64_BREAK_FAULT); |
| 183 | |
| 184 | updptr[i] = cpu_to_le32(insn); |
| 185 | } |
| 186 | } |
| 187 | |
| 188 | void kvm_patch_vector_branch(struct alt_instr *alt, |
| 189 | __le32 *origptr, __le32 *updptr, int nr_inst) |
| 190 | { |
| 191 | u64 addr; |
| 192 | u32 insn; |
| 193 | |
| 194 | BUG_ON(nr_inst != 4); |
| 195 | |
| 196 | if (!cpus_have_cap(ARM64_SPECTRE_V3A) || |
| 197 | WARN_ON_ONCE(cpus_have_cap(ARM64_HAS_VIRT_HOST_EXTN))) |
| 198 | return; |
| 199 | |
| 200 | /* |
| 201 | * Compute HYP VA by using the same computation as kern_hyp_va() |
| 202 | */ |
| 203 | addr = __early_kern_hyp_va((u64)kvm_ksym_ref(__kvm_hyp_vector)); |
| 204 | |
| 205 | /* Use PC[10:7] to branch to the same vector in KVM */ |
| 206 | addr |= ((u64)origptr & GENMASK_ULL(10, 7)); |
| 207 | |
| 208 | /* |
| 209 | * Branch over the preamble in order to avoid the initial store on |
| 210 | * the stack (which we already perform in the hardening vectors). |
| 211 | */ |
| 212 | addr += KVM_VECTOR_PREAMBLE; |
| 213 | |
| 214 | /* movz x0, #(addr & 0xffff) */ |
| 215 | insn = aarch64_insn_gen_movewide(AARCH64_INSN_REG_0, |
| 216 | (u16)addr, |
| 217 | 0, |
| 218 | AARCH64_INSN_VARIANT_64BIT, |
| 219 | AARCH64_INSN_MOVEWIDE_ZERO); |
| 220 | *updptr++ = cpu_to_le32(insn); |
| 221 | |
| 222 | /* movk x0, #((addr >> 16) & 0xffff), lsl #16 */ |
| 223 | insn = aarch64_insn_gen_movewide(AARCH64_INSN_REG_0, |
| 224 | (u16)(addr >> 16), |
| 225 | 16, |
| 226 | AARCH64_INSN_VARIANT_64BIT, |
| 227 | AARCH64_INSN_MOVEWIDE_KEEP); |
| 228 | *updptr++ = cpu_to_le32(insn); |
| 229 | |
| 230 | /* movk x0, #((addr >> 32) & 0xffff), lsl #32 */ |
| 231 | insn = aarch64_insn_gen_movewide(AARCH64_INSN_REG_0, |
| 232 | (u16)(addr >> 32), |
| 233 | 32, |
| 234 | AARCH64_INSN_VARIANT_64BIT, |
| 235 | AARCH64_INSN_MOVEWIDE_KEEP); |
| 236 | *updptr++ = cpu_to_le32(insn); |
| 237 | |
| 238 | /* br x0 */ |
| 239 | insn = aarch64_insn_gen_branch_reg(AARCH64_INSN_REG_0, |
| 240 | AARCH64_INSN_BRANCH_NOLINK); |
| 241 | *updptr++ = cpu_to_le32(insn); |
| 242 | } |
| 243 | |
| 244 | static void generate_mov_q(u64 val, __le32 *origptr, __le32 *updptr, int nr_inst) |
| 245 | { |
| 246 | u32 insn, oinsn, rd; |
| 247 | |
| 248 | BUG_ON(nr_inst != 4); |
| 249 | |
| 250 | /* Compute target register */ |
| 251 | oinsn = le32_to_cpu(*origptr); |
| 252 | rd = aarch64_insn_decode_register(AARCH64_INSN_REGTYPE_RD, oinsn); |
| 253 | |
| 254 | /* movz rd, #(val & 0xffff) */ |
| 255 | insn = aarch64_insn_gen_movewide(rd, |
| 256 | (u16)val, |
| 257 | 0, |
| 258 | AARCH64_INSN_VARIANT_64BIT, |
| 259 | AARCH64_INSN_MOVEWIDE_ZERO); |
| 260 | *updptr++ = cpu_to_le32(insn); |
| 261 | |
| 262 | /* movk rd, #((val >> 16) & 0xffff), lsl #16 */ |
| 263 | insn = aarch64_insn_gen_movewide(rd, |
| 264 | (u16)(val >> 16), |
| 265 | 16, |
| 266 | AARCH64_INSN_VARIANT_64BIT, |
| 267 | AARCH64_INSN_MOVEWIDE_KEEP); |
| 268 | *updptr++ = cpu_to_le32(insn); |
| 269 | |
| 270 | /* movk rd, #((val >> 32) & 0xffff), lsl #32 */ |
| 271 | insn = aarch64_insn_gen_movewide(rd, |
| 272 | (u16)(val >> 32), |
| 273 | 32, |
| 274 | AARCH64_INSN_VARIANT_64BIT, |
| 275 | AARCH64_INSN_MOVEWIDE_KEEP); |
| 276 | *updptr++ = cpu_to_le32(insn); |
| 277 | |
| 278 | /* movk rd, #((val >> 48) & 0xffff), lsl #48 */ |
| 279 | insn = aarch64_insn_gen_movewide(rd, |
| 280 | (u16)(val >> 48), |
| 281 | 48, |
| 282 | AARCH64_INSN_VARIANT_64BIT, |
| 283 | AARCH64_INSN_MOVEWIDE_KEEP); |
| 284 | *updptr++ = cpu_to_le32(insn); |
| 285 | } |
| 286 | |
| 287 | void kvm_get_kimage_voffset(struct alt_instr *alt, |
| 288 | __le32 *origptr, __le32 *updptr, int nr_inst) |
| 289 | { |
| 290 | generate_mov_q(kimage_voffset, origptr, updptr, nr_inst); |
| 291 | } |
| 292 | |
| 293 | void kvm_compute_final_ctr_el0(struct alt_instr *alt, |
| 294 | __le32 *origptr, __le32 *updptr, int nr_inst) |
| 295 | { |
| 296 | generate_mov_q(read_sanitised_ftr_reg(SYS_CTR_EL0), |
| 297 | origptr, updptr, nr_inst); |
| 298 | } |
| 299 | |
| 300 | void kvm_pan_patch_el2_entry(struct alt_instr *alt, |
| 301 | __le32 *origptr, __le32 *updptr, int nr_inst) |
| 302 | { |
| 303 | /* |
| 304 | * If we're running at EL1 without hVHE, then SCTLR_EL2.SPAN means |
| 305 | * nothing to us (it is RES1), and we don't need to set PSTATE.PAN |
| 306 | * to anything useful. |
| 307 | */ |
| 308 | if (!is_kernel_in_hyp_mode() && !cpus_have_cap(ARM64_KVM_HVHE)) |
| 309 | return; |
| 310 | |
| 311 | /* |
| 312 | * Leap of faith: at this point, we must be running VHE one way or |
| 313 | * another, and FEAT_PAN is required to be implemented. If KVM |
| 314 | * explodes at runtime because your system does not abide by this |
| 315 | * requirement, call your favourite HW vendor, they have screwed up. |
| 316 | * |
| 317 | * We don't expect hVHE to access any userspace mapping, so always |
| 318 | * set PSTATE.PAN on enty. Same thing if we have PAN enabled on an |
| 319 | * EL2 kernel. Only force it to 0 if we have not configured PAN in |
| 320 | * the kernel (and you know this is really silly). |
| 321 | */ |
| 322 | if (cpus_have_cap(ARM64_KVM_HVHE) || IS_ENABLED(CONFIG_ARM64_PAN)) |
| 323 | *updptr = cpu_to_le32(ENCODE_PSTATE(1, PAN)); |
| 324 | else |
| 325 | *updptr = cpu_to_le32(ENCODE_PSTATE(0, PAN)); |
| 326 | } |
| 327 | |