| 1 | // SPDX-License-Identifier: GPL-2.0-only |
| 2 | /* |
| 3 | * Copyright (C) 2020 Google LLC |
| 4 | * Author: Quentin Perret <qperret@google.com> |
| 5 | */ |
| 6 | |
| 7 | #include <linux/kvm_host.h> |
| 8 | #include <asm/kvm_hyp.h> |
| 9 | #include <asm/kvm_mmu.h> |
| 10 | #include <asm/kvm_pgtable.h> |
| 11 | #include <asm/kvm_pkvm.h> |
| 12 | #include <asm/spectre.h> |
| 13 | |
| 14 | #include <nvhe/early_alloc.h> |
| 15 | #include <nvhe/gfp.h> |
| 16 | #include <nvhe/memory.h> |
| 17 | #include <nvhe/mem_protect.h> |
| 18 | #include <nvhe/mm.h> |
| 19 | #include <nvhe/spinlock.h> |
| 20 | |
| 21 | struct kvm_pgtable pkvm_pgtable; |
| 22 | hyp_spinlock_t pkvm_pgd_lock; |
| 23 | |
| 24 | struct memblock_region hyp_memory[HYP_MEMBLOCK_REGIONS]; |
| 25 | unsigned int hyp_memblock_nr; |
| 26 | |
| 27 | static u64 __io_map_base; |
| 28 | |
| 29 | struct hyp_fixmap_slot { |
| 30 | u64 addr; |
| 31 | kvm_pte_t *ptep; |
| 32 | }; |
| 33 | static DEFINE_PER_CPU(struct hyp_fixmap_slot, fixmap_slots); |
| 34 | |
| 35 | static int __pkvm_create_mappings(unsigned long start, unsigned long size, |
| 36 | unsigned long phys, enum kvm_pgtable_prot prot) |
| 37 | { |
| 38 | int err; |
| 39 | |
| 40 | hyp_spin_lock(&pkvm_pgd_lock); |
| 41 | err = kvm_pgtable_hyp_map(&pkvm_pgtable, start, size, phys, prot); |
| 42 | hyp_spin_unlock(&pkvm_pgd_lock); |
| 43 | |
| 44 | return err; |
| 45 | } |
| 46 | |
| 47 | static int __pkvm_alloc_private_va_range(unsigned long start, size_t size) |
| 48 | { |
| 49 | unsigned long cur; |
| 50 | |
| 51 | hyp_assert_lock_held(&pkvm_pgd_lock); |
| 52 | |
| 53 | if (!start || start < __io_map_base) |
| 54 | return -EINVAL; |
| 55 | |
| 56 | /* The allocated size is always a multiple of PAGE_SIZE */ |
| 57 | cur = start + PAGE_ALIGN(size); |
| 58 | |
| 59 | /* Are we overflowing on the vmemmap ? */ |
| 60 | if (cur > __hyp_vmemmap) |
| 61 | return -ENOMEM; |
| 62 | |
| 63 | __io_map_base = cur; |
| 64 | |
| 65 | return 0; |
| 66 | } |
| 67 | |
| 68 | /** |
| 69 | * pkvm_alloc_private_va_range - Allocates a private VA range. |
| 70 | * @size: The size of the VA range to reserve. |
| 71 | * @haddr: The hypervisor virtual start address of the allocation. |
| 72 | * |
| 73 | * The private virtual address (VA) range is allocated above __io_map_base |
| 74 | * and aligned based on the order of @size. |
| 75 | * |
| 76 | * Return: 0 on success or negative error code on failure. |
| 77 | */ |
| 78 | int pkvm_alloc_private_va_range(size_t size, unsigned long *haddr) |
| 79 | { |
| 80 | unsigned long addr; |
| 81 | int ret; |
| 82 | |
| 83 | hyp_spin_lock(&pkvm_pgd_lock); |
| 84 | addr = __io_map_base; |
| 85 | ret = __pkvm_alloc_private_va_range(start: addr, size); |
| 86 | hyp_spin_unlock(&pkvm_pgd_lock); |
| 87 | |
| 88 | *haddr = addr; |
| 89 | |
| 90 | return ret; |
| 91 | } |
| 92 | |
| 93 | int __pkvm_create_private_mapping(phys_addr_t phys, size_t size, |
| 94 | enum kvm_pgtable_prot prot, |
| 95 | unsigned long *haddr) |
| 96 | { |
| 97 | unsigned long addr; |
| 98 | int err; |
| 99 | |
| 100 | size = PAGE_ALIGN(size + offset_in_page(phys)); |
| 101 | err = pkvm_alloc_private_va_range(size, haddr: &addr); |
| 102 | if (err) |
| 103 | return err; |
| 104 | |
| 105 | err = __pkvm_create_mappings(start: addr, size, phys, prot: prot); |
| 106 | if (err) |
| 107 | return err; |
| 108 | |
| 109 | *haddr = addr + offset_in_page(phys); |
| 110 | return err; |
| 111 | } |
| 112 | |
| 113 | int pkvm_create_mappings_locked(void *from, void *to, enum kvm_pgtable_prot prot) |
| 114 | { |
| 115 | unsigned long start = (unsigned long)from; |
| 116 | unsigned long end = (unsigned long)to; |
| 117 | unsigned long virt_addr; |
| 118 | phys_addr_t phys; |
| 119 | |
| 120 | hyp_assert_lock_held(&pkvm_pgd_lock); |
| 121 | |
| 122 | start = start & PAGE_MASK; |
| 123 | end = PAGE_ALIGN(end); |
| 124 | |
| 125 | for (virt_addr = start; virt_addr < end; virt_addr += PAGE_SIZE) { |
| 126 | int err; |
| 127 | |
| 128 | phys = hyp_virt_to_phys((void *)virt_addr); |
| 129 | err = kvm_pgtable_hyp_map(&pkvm_pgtable, virt_addr, PAGE_SIZE, |
| 130 | phys, prot); |
| 131 | if (err) |
| 132 | return err; |
| 133 | } |
| 134 | |
| 135 | return 0; |
| 136 | } |
| 137 | |
| 138 | int pkvm_create_mappings(void *from, void *to, enum kvm_pgtable_prot prot) |
| 139 | { |
| 140 | int ret; |
| 141 | |
| 142 | hyp_spin_lock(&pkvm_pgd_lock); |
| 143 | ret = pkvm_create_mappings_locked(from, to, prot: prot); |
| 144 | hyp_spin_unlock(&pkvm_pgd_lock); |
| 145 | |
| 146 | return ret; |
| 147 | } |
| 148 | |
| 149 | int hyp_back_vmemmap(phys_addr_t back) |
| 150 | { |
| 151 | unsigned long i, start, size, end = 0; |
| 152 | int ret; |
| 153 | |
| 154 | for (i = 0; i < hyp_memblock_nr; i++) { |
| 155 | start = hyp_memory[i].base; |
| 156 | start = ALIGN_DOWN((u64)hyp_phys_to_page(start), PAGE_SIZE); |
| 157 | /* |
| 158 | * The beginning of the hyp_vmemmap region for the current |
| 159 | * memblock may already be backed by the page backing the end |
| 160 | * the previous region, so avoid mapping it twice. |
| 161 | */ |
| 162 | start = max(start, end); |
| 163 | |
| 164 | end = hyp_memory[i].base + hyp_memory[i].size; |
| 165 | end = PAGE_ALIGN((u64)hyp_phys_to_page(end)); |
| 166 | if (start >= end) |
| 167 | continue; |
| 168 | |
| 169 | size = end - start; |
| 170 | ret = __pkvm_create_mappings(start, size, phys: back, prot: PAGE_HYP); |
| 171 | if (ret) |
| 172 | return ret; |
| 173 | |
| 174 | memset(hyp_phys_to_virt(back), 0, size); |
| 175 | back += size; |
| 176 | } |
| 177 | |
| 178 | return 0; |
| 179 | } |
| 180 | |
| 181 | static void *__hyp_bp_vect_base; |
| 182 | int pkvm_cpu_set_vector(enum arm64_hyp_spectre_vector slot) |
| 183 | { |
| 184 | void *vector; |
| 185 | |
| 186 | switch (slot) { |
| 187 | case HYP_VECTOR_DIRECT: { |
| 188 | vector = __kvm_hyp_vector; |
| 189 | break; |
| 190 | } |
| 191 | case HYP_VECTOR_SPECTRE_DIRECT: { |
| 192 | vector = __bp_harden_hyp_vecs; |
| 193 | break; |
| 194 | } |
| 195 | case HYP_VECTOR_INDIRECT: |
| 196 | case HYP_VECTOR_SPECTRE_INDIRECT: { |
| 197 | vector = (void *)__hyp_bp_vect_base; |
| 198 | break; |
| 199 | } |
| 200 | default: |
| 201 | return -EINVAL; |
| 202 | } |
| 203 | |
| 204 | vector = __kvm_vector_slot2addr(vector, slot); |
| 205 | *this_cpu_ptr(&kvm_hyp_vector) = (unsigned long)vector; |
| 206 | |
| 207 | return 0; |
| 208 | } |
| 209 | |
| 210 | int hyp_map_vectors(void) |
| 211 | { |
| 212 | phys_addr_t phys; |
| 213 | unsigned long bp_base; |
| 214 | int ret; |
| 215 | |
| 216 | if (!kvm_system_needs_idmapped_vectors()) { |
| 217 | __hyp_bp_vect_base = __bp_harden_hyp_vecs; |
| 218 | return 0; |
| 219 | } |
| 220 | |
| 221 | phys = __hyp_pa(__bp_harden_hyp_vecs); |
| 222 | ret = __pkvm_create_private_mapping(phys, size: __BP_HARDEN_HYP_VECS_SZ, |
| 223 | prot: PAGE_HYP_EXEC, haddr: &bp_base); |
| 224 | if (ret) |
| 225 | return ret; |
| 226 | |
| 227 | __hyp_bp_vect_base = (void *)bp_base; |
| 228 | |
| 229 | return 0; |
| 230 | } |
| 231 | |
| 232 | static void *fixmap_map_slot(struct hyp_fixmap_slot *slot, phys_addr_t phys) |
| 233 | { |
| 234 | kvm_pte_t pte, *ptep = slot->ptep; |
| 235 | |
| 236 | pte = *ptep; |
| 237 | pte &= ~kvm_phys_to_pte(KVM_PHYS_INVALID); |
| 238 | pte |= kvm_phys_to_pte(phys) | KVM_PTE_VALID; |
| 239 | WRITE_ONCE(*ptep, pte); |
| 240 | dsb(ishst); |
| 241 | |
| 242 | return (void *)slot->addr; |
| 243 | } |
| 244 | |
| 245 | void *hyp_fixmap_map(phys_addr_t phys) |
| 246 | { |
| 247 | return fixmap_map_slot(this_cpu_ptr(&fixmap_slots), phys); |
| 248 | } |
| 249 | |
| 250 | static void fixmap_clear_slot(struct hyp_fixmap_slot *slot) |
| 251 | { |
| 252 | kvm_pte_t *ptep = slot->ptep; |
| 253 | u64 addr = slot->addr; |
| 254 | u32 level; |
| 255 | |
| 256 | if (FIELD_GET(KVM_PTE_TYPE, *ptep) == KVM_PTE_TYPE_PAGE) |
| 257 | level = KVM_PGTABLE_LAST_LEVEL; |
| 258 | else |
| 259 | level = KVM_PGTABLE_LAST_LEVEL - 1; /* create_fixblock() guarantees PMD level */ |
| 260 | |
| 261 | WRITE_ONCE(*ptep, *ptep & ~KVM_PTE_VALID); |
| 262 | |
| 263 | /* |
| 264 | * Irritatingly, the architecture requires that we use inner-shareable |
| 265 | * broadcast TLB invalidation here in case another CPU speculates |
| 266 | * through our fixmap and decides to create an "amalagamation of the |
| 267 | * values held in the TLB" due to the apparent lack of a |
| 268 | * break-before-make sequence. |
| 269 | * |
| 270 | * https://lore.kernel.org/kvm/20221017115209.2099-1-will@kernel.org/T/#mf10dfbaf1eaef9274c581b81c53758918c1d0f03 |
| 271 | */ |
| 272 | dsb(ishst); |
| 273 | __tlbi_level(vale2is, __TLBI_VADDR(addr, 0), level); |
| 274 | dsb(ish); |
| 275 | isb(); |
| 276 | } |
| 277 | |
| 278 | void hyp_fixmap_unmap(void) |
| 279 | { |
| 280 | fixmap_clear_slot(this_cpu_ptr(&fixmap_slots)); |
| 281 | } |
| 282 | |
| 283 | static int __create_fixmap_slot_cb(const struct kvm_pgtable_visit_ctx *ctx, |
| 284 | enum kvm_pgtable_walk_flags visit) |
| 285 | { |
| 286 | struct hyp_fixmap_slot *slot = (struct hyp_fixmap_slot *)ctx->arg; |
| 287 | |
| 288 | if (!kvm_pte_valid(ctx->old) || (ctx->end - ctx->start) != kvm_granule_size(ctx->level)) |
| 289 | return -EINVAL; |
| 290 | |
| 291 | slot->addr = ctx->addr; |
| 292 | slot->ptep = ctx->ptep; |
| 293 | |
| 294 | /* |
| 295 | * Clear the PTE, but keep the page-table page refcount elevated to |
| 296 | * prevent it from ever being freed. This lets us manipulate the PTEs |
| 297 | * by hand safely without ever needing to allocate memory. |
| 298 | */ |
| 299 | fixmap_clear_slot(slot); |
| 300 | |
| 301 | return 0; |
| 302 | } |
| 303 | |
| 304 | static int create_fixmap_slot(u64 addr, u64 cpu) |
| 305 | { |
| 306 | struct kvm_pgtable_walker walker = { |
| 307 | .cb = __create_fixmap_slot_cb, |
| 308 | .flags = KVM_PGTABLE_WALK_LEAF, |
| 309 | .arg = per_cpu_ptr(&fixmap_slots, cpu), |
| 310 | }; |
| 311 | |
| 312 | return kvm_pgtable_walk(&pkvm_pgtable, addr, PAGE_SIZE, &walker); |
| 313 | } |
| 314 | |
| 315 | #if PAGE_SHIFT < 16 |
| 316 | #define HAS_FIXBLOCK |
| 317 | static struct hyp_fixmap_slot hyp_fixblock_slot; |
| 318 | static DEFINE_HYP_SPINLOCK(hyp_fixblock_lock); |
| 319 | #endif |
| 320 | |
| 321 | static int create_fixblock(void) |
| 322 | { |
| 323 | #ifdef HAS_FIXBLOCK |
| 324 | struct kvm_pgtable_walker walker = { |
| 325 | .cb = __create_fixmap_slot_cb, |
| 326 | .flags = KVM_PGTABLE_WALK_LEAF, |
| 327 | .arg = &hyp_fixblock_slot, |
| 328 | }; |
| 329 | unsigned long addr; |
| 330 | phys_addr_t phys; |
| 331 | int ret, i; |
| 332 | |
| 333 | /* Find a RAM phys address, PMD aligned */ |
| 334 | for (i = 0; i < hyp_memblock_nr; i++) { |
| 335 | phys = ALIGN(hyp_memory[i].base, PMD_SIZE); |
| 336 | if (phys + PMD_SIZE < (hyp_memory[i].base + hyp_memory[i].size)) |
| 337 | break; |
| 338 | } |
| 339 | |
| 340 | if (i >= hyp_memblock_nr) |
| 341 | return -EINVAL; |
| 342 | |
| 343 | hyp_spin_lock(&pkvm_pgd_lock); |
| 344 | addr = ALIGN(__io_map_base, PMD_SIZE); |
| 345 | ret = __pkvm_alloc_private_va_range(start: addr, PMD_SIZE); |
| 346 | if (ret) |
| 347 | goto unlock; |
| 348 | |
| 349 | ret = kvm_pgtable_hyp_map(&pkvm_pgtable, addr, PMD_SIZE, phys, PAGE_HYP); |
| 350 | if (ret) |
| 351 | goto unlock; |
| 352 | |
| 353 | ret = kvm_pgtable_walk(&pkvm_pgtable, addr, PMD_SIZE, &walker); |
| 354 | |
| 355 | unlock: |
| 356 | hyp_spin_unlock(&pkvm_pgd_lock); |
| 357 | |
| 358 | return ret; |
| 359 | #else |
| 360 | return 0; |
| 361 | #endif |
| 362 | } |
| 363 | |
| 364 | void *hyp_fixblock_map(phys_addr_t phys, size_t *size) |
| 365 | { |
| 366 | #ifdef HAS_FIXBLOCK |
| 367 | *size = PMD_SIZE; |
| 368 | hyp_spin_lock(&hyp_fixblock_lock); |
| 369 | return fixmap_map_slot(slot: &hyp_fixblock_slot, phys); |
| 370 | #else |
| 371 | *size = PAGE_SIZE; |
| 372 | return hyp_fixmap_map(phys); |
| 373 | #endif |
| 374 | } |
| 375 | |
| 376 | void hyp_fixblock_unmap(void) |
| 377 | { |
| 378 | #ifdef HAS_FIXBLOCK |
| 379 | fixmap_clear_slot(slot: &hyp_fixblock_slot); |
| 380 | hyp_spin_unlock(&hyp_fixblock_lock); |
| 381 | #else |
| 382 | hyp_fixmap_unmap(); |
| 383 | #endif |
| 384 | } |
| 385 | |
| 386 | int hyp_create_fixmap(void) |
| 387 | { |
| 388 | unsigned long addr, i; |
| 389 | int ret; |
| 390 | |
| 391 | for (i = 0; i < hyp_nr_cpus; i++) { |
| 392 | ret = pkvm_alloc_private_va_range(PAGE_SIZE, &addr); |
| 393 | if (ret) |
| 394 | return ret; |
| 395 | |
| 396 | ret = kvm_pgtable_hyp_map(&pkvm_pgtable, addr, PAGE_SIZE, |
| 397 | __hyp_pa(__hyp_bss_start), PAGE_HYP); |
| 398 | if (ret) |
| 399 | return ret; |
| 400 | |
| 401 | ret = create_fixmap_slot(addr, i); |
| 402 | if (ret) |
| 403 | return ret; |
| 404 | } |
| 405 | |
| 406 | return create_fixblock(); |
| 407 | } |
| 408 | |
| 409 | int hyp_create_idmap(u32 hyp_va_bits) |
| 410 | { |
| 411 | unsigned long start, end; |
| 412 | |
| 413 | start = hyp_virt_to_phys((void *)__hyp_idmap_text_start); |
| 414 | start = ALIGN_DOWN(start, PAGE_SIZE); |
| 415 | |
| 416 | end = hyp_virt_to_phys((void *)__hyp_idmap_text_end); |
| 417 | end = ALIGN(end, PAGE_SIZE); |
| 418 | |
| 419 | /* |
| 420 | * One half of the VA space is reserved to linearly map portions of |
| 421 | * memory -- see va_layout.c for more details. The other half of the VA |
| 422 | * space contains the trampoline page, and needs some care. Split that |
| 423 | * second half in two and find the quarter of VA space not conflicting |
| 424 | * with the idmap to place the IOs and the vmemmap. IOs use the lower |
| 425 | * half of the quarter and the vmemmap the upper half. |
| 426 | */ |
| 427 | __io_map_base = start & BIT(hyp_va_bits - 2); |
| 428 | __io_map_base ^= BIT(hyp_va_bits - 2); |
| 429 | __hyp_vmemmap = __io_map_base | BIT(hyp_va_bits - 3); |
| 430 | |
| 431 | return __pkvm_create_mappings(start, end - start, start, PAGE_HYP_EXEC); |
| 432 | } |
| 433 | |
| 434 | int pkvm_create_stack(phys_addr_t phys, unsigned long *haddr) |
| 435 | { |
| 436 | unsigned long addr, prev_base; |
| 437 | size_t size; |
| 438 | int ret; |
| 439 | |
| 440 | hyp_spin_lock(&pkvm_pgd_lock); |
| 441 | |
| 442 | prev_base = __io_map_base; |
| 443 | /* |
| 444 | * Efficient stack verification using the NVHE_STACK_SHIFT bit implies |
| 445 | * an alignment of our allocation on the order of the size. |
| 446 | */ |
| 447 | size = NVHE_STACK_SIZE * 2; |
| 448 | addr = ALIGN(__io_map_base, size); |
| 449 | |
| 450 | ret = __pkvm_alloc_private_va_range(start: addr, size); |
| 451 | if (!ret) { |
| 452 | /* |
| 453 | * Since the stack grows downwards, map the stack to the page |
| 454 | * at the higher address and leave the lower guard page |
| 455 | * unbacked. |
| 456 | * |
| 457 | * Any valid stack address now has the NVHE_STACK_SHIFT bit as 1 |
| 458 | * and addresses corresponding to the guard page have the |
| 459 | * NVHE_STACK_SHIFT bit as 0 - this is used for overflow detection. |
| 460 | */ |
| 461 | ret = kvm_pgtable_hyp_map(&pkvm_pgtable, addr + NVHE_STACK_SIZE, |
| 462 | NVHE_STACK_SIZE, phys, PAGE_HYP); |
| 463 | if (ret) |
| 464 | __io_map_base = prev_base; |
| 465 | } |
| 466 | hyp_spin_unlock(&pkvm_pgd_lock); |
| 467 | |
| 468 | *haddr = addr + size; |
| 469 | |
| 470 | return ret; |
| 471 | } |
| 472 | |
| 473 | static void *admit_host_page(void *arg) |
| 474 | { |
| 475 | struct kvm_hyp_memcache *host_mc = arg; |
| 476 | |
| 477 | if (!host_mc->nr_pages) |
| 478 | return NULL; |
| 479 | |
| 480 | /* |
| 481 | * The host still owns the pages in its memcache, so we need to go |
| 482 | * through a full host-to-hyp donation cycle to change it. Fortunately, |
| 483 | * __pkvm_host_donate_hyp() takes care of races for us, so if it |
| 484 | * succeeds we're good to go. |
| 485 | */ |
| 486 | if (__pkvm_host_donate_hyp(hyp_phys_to_pfn(host_mc->head), 1)) |
| 487 | return NULL; |
| 488 | |
| 489 | return pop_hyp_memcache(host_mc, hyp_phys_to_virt); |
| 490 | } |
| 491 | |
| 492 | /* Refill our local memcache by popping pages from the one provided by the host. */ |
| 493 | int refill_memcache(struct kvm_hyp_memcache *mc, unsigned long min_pages, |
| 494 | struct kvm_hyp_memcache *host_mc) |
| 495 | { |
| 496 | struct kvm_hyp_memcache tmp = *host_mc; |
| 497 | int ret; |
| 498 | |
| 499 | ret = __topup_hyp_memcache(mc, min_pages, admit_host_page, |
| 500 | hyp_virt_to_phys, &tmp); |
| 501 | *host_mc = tmp; |
| 502 | |
| 503 | return ret; |
| 504 | } |
| 505 | |