1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * AMD Memory Encryption Support 4 * 5 * Copyright (C) 2016 Advanced Micro Devices, Inc. 6 * 7 * Author: Tom Lendacky <thomas.lendacky@amd.com> 8 */ 9 10 #define DISABLE_BRANCH_PROFILING 11 12 /* 13 * Since we're dealing with identity mappings, physical and virtual 14 * addresses are the same, so override these defines which are ultimately 15 * used by the headers in misc.h. 16 */ 17 #define __pa(x) ((unsigned long)(x)) 18 #define __va(x) ((void *)((unsigned long)(x))) 19 20 /* 21 * Special hack: we have to be careful, because no indirections are 22 * allowed here, and paravirt_ops is a kind of one. As it will only run in 23 * baremetal anyway, we just keep it from happening. (This list needs to 24 * be extended when new paravirt and debugging variants are added.) 25 */ 26 #undef CONFIG_PARAVIRT 27 #undef CONFIG_PARAVIRT_XXL 28 #undef CONFIG_PARAVIRT_SPINLOCKS 29 30 #include <linux/kernel.h> 31 #include <linux/mm.h> 32 #include <linux/mem_encrypt.h> 33 34 #include <asm/setup.h> 35 #include <asm/sections.h> 36 #include <asm/cmdline.h> 37 38 #include "mm_internal.h" 39 40 #define PGD_FLAGS _KERNPG_TABLE_NOENC 41 #define P4D_FLAGS _KERNPG_TABLE_NOENC 42 #define PUD_FLAGS _KERNPG_TABLE_NOENC 43 #define PMD_FLAGS _KERNPG_TABLE_NOENC 44 45 #define PMD_FLAGS_LARGE (__PAGE_KERNEL_LARGE_EXEC & ~_PAGE_GLOBAL) 46 47 #define PMD_FLAGS_DEC PMD_FLAGS_LARGE 48 #define PMD_FLAGS_DEC_WP ((PMD_FLAGS_DEC & ~_PAGE_LARGE_CACHE_MASK) | \ 49 (_PAGE_PAT_LARGE | _PAGE_PWT)) 50 51 #define PMD_FLAGS_ENC (PMD_FLAGS_LARGE | _PAGE_ENC) 52 53 #define PTE_FLAGS (__PAGE_KERNEL_EXEC & ~_PAGE_GLOBAL) 54 55 #define PTE_FLAGS_DEC PTE_FLAGS 56 #define PTE_FLAGS_DEC_WP ((PTE_FLAGS_DEC & ~_PAGE_CACHE_MASK) | \ 57 (_PAGE_PAT | _PAGE_PWT)) 58 59 #define PTE_FLAGS_ENC (PTE_FLAGS | _PAGE_ENC) 60 61 struct sme_populate_pgd_data { 62 void *pgtable_area; 63 pgd_t *pgd; 64 65 pmdval_t pmd_flags; 66 pteval_t pte_flags; 67 unsigned long paddr; 68 69 unsigned long vaddr; 70 unsigned long vaddr_end; 71 }; 72 73 /* 74 * This work area lives in the .init.scratch section, which lives outside of 75 * the kernel proper. It is sized to hold the intermediate copy buffer and 76 * more than enough pagetable pages. 77 * 78 * By using this section, the kernel can be encrypted in place and it 79 * avoids any possibility of boot parameters or initramfs images being 80 * placed such that the in-place encryption logic overwrites them. This 81 * section is 2MB aligned to allow for simple pagetable setup using only 82 * PMD entries (see vmlinux.lds.S). 83 */ 84 static char sme_workarea[2 * PMD_PAGE_SIZE] __section(".init.scratch"); 85 86 static char sme_cmdline_arg[] __initdata = "mem_encrypt"; 87 static char sme_cmdline_on[] __initdata = "on"; 88 static char sme_cmdline_off[] __initdata = "off"; 89 90 static void __init sme_clear_pgd(struct sme_populate_pgd_data *ppd) 91 { 92 unsigned long pgd_start, pgd_end, pgd_size; 93 pgd_t *pgd_p; 94 95 pgd_start = ppd->vaddr & PGDIR_MASK; 96 pgd_end = ppd->vaddr_end & PGDIR_MASK; 97 98 pgd_size = (((pgd_end - pgd_start) / PGDIR_SIZE) + 1) * sizeof(pgd_t); 99 100 pgd_p = ppd->pgd + pgd_index(ppd->vaddr); 101 102 memset(pgd_p, 0, pgd_size); 103 } 104 105 static pud_t __init *sme_prepare_pgd(struct sme_populate_pgd_data *ppd) 106 { 107 pgd_t *pgd; 108 p4d_t *p4d; 109 pud_t *pud; 110 pmd_t *pmd; 111 112 pgd = ppd->pgd + pgd_index(ppd->vaddr); 113 if (pgd_none(*pgd)) { 114 p4d = ppd->pgtable_area; 115 memset(p4d, 0, sizeof(*p4d) * PTRS_PER_P4D); 116 ppd->pgtable_area += sizeof(*p4d) * PTRS_PER_P4D; 117 set_pgd(pgd, __pgd(PGD_FLAGS | __pa(p4d))); 118 } 119 120 p4d = p4d_offset(pgd, ppd->vaddr); 121 if (p4d_none(*p4d)) { 122 pud = ppd->pgtable_area; 123 memset(pud, 0, sizeof(*pud) * PTRS_PER_PUD); 124 ppd->pgtable_area += sizeof(*pud) * PTRS_PER_PUD; 125 set_p4d(p4d, __p4d(P4D_FLAGS | __pa(pud))); 126 } 127 128 pud = pud_offset(p4d, ppd->vaddr); 129 if (pud_none(*pud)) { 130 pmd = ppd->pgtable_area; 131 memset(pmd, 0, sizeof(*pmd) * PTRS_PER_PMD); 132 ppd->pgtable_area += sizeof(*pmd) * PTRS_PER_PMD; 133 set_pud(pud, __pud(PUD_FLAGS | __pa(pmd))); 134 } 135 136 if (pud_large(*pud)) 137 return NULL; 138 139 return pud; 140 } 141 142 static void __init sme_populate_pgd_large(struct sme_populate_pgd_data *ppd) 143 { 144 pud_t *pud; 145 pmd_t *pmd; 146 147 pud = sme_prepare_pgd(ppd); 148 if (!pud) 149 return; 150 151 pmd = pmd_offset(pud, ppd->vaddr); 152 if (pmd_large(*pmd)) 153 return; 154 155 set_pmd(pmd, __pmd(ppd->paddr | ppd->pmd_flags)); 156 } 157 158 static void __init sme_populate_pgd(struct sme_populate_pgd_data *ppd) 159 { 160 pud_t *pud; 161 pmd_t *pmd; 162 pte_t *pte; 163 164 pud = sme_prepare_pgd(ppd); 165 if (!pud) 166 return; 167 168 pmd = pmd_offset(pud, ppd->vaddr); 169 if (pmd_none(*pmd)) { 170 pte = ppd->pgtable_area; 171 memset(pte, 0, sizeof(*pte) * PTRS_PER_PTE); 172 ppd->pgtable_area += sizeof(*pte) * PTRS_PER_PTE; 173 set_pmd(pmd, __pmd(PMD_FLAGS | __pa(pte))); 174 } 175 176 if (pmd_large(*pmd)) 177 return; 178 179 pte = pte_offset_map(pmd, ppd->vaddr); 180 if (pte_none(*pte)) 181 set_pte(pte, __pte(ppd->paddr | ppd->pte_flags)); 182 } 183 184 static void __init __sme_map_range_pmd(struct sme_populate_pgd_data *ppd) 185 { 186 while (ppd->vaddr < ppd->vaddr_end) { 187 sme_populate_pgd_large(ppd); 188 189 ppd->vaddr += PMD_PAGE_SIZE; 190 ppd->paddr += PMD_PAGE_SIZE; 191 } 192 } 193 194 static void __init __sme_map_range_pte(struct sme_populate_pgd_data *ppd) 195 { 196 while (ppd->vaddr < ppd->vaddr_end) { 197 sme_populate_pgd(ppd); 198 199 ppd->vaddr += PAGE_SIZE; 200 ppd->paddr += PAGE_SIZE; 201 } 202 } 203 204 static void __init __sme_map_range(struct sme_populate_pgd_data *ppd, 205 pmdval_t pmd_flags, pteval_t pte_flags) 206 { 207 unsigned long vaddr_end; 208 209 ppd->pmd_flags = pmd_flags; 210 ppd->pte_flags = pte_flags; 211 212 /* Save original end value since we modify the struct value */ 213 vaddr_end = ppd->vaddr_end; 214 215 /* If start is not 2MB aligned, create PTE entries */ 216 ppd->vaddr_end = ALIGN(ppd->vaddr, PMD_PAGE_SIZE); 217 __sme_map_range_pte(ppd); 218 219 /* Create PMD entries */ 220 ppd->vaddr_end = vaddr_end & PMD_PAGE_MASK; 221 __sme_map_range_pmd(ppd); 222 223 /* If end is not 2MB aligned, create PTE entries */ 224 ppd->vaddr_end = vaddr_end; 225 __sme_map_range_pte(ppd); 226 } 227 228 static void __init sme_map_range_encrypted(struct sme_populate_pgd_data *ppd) 229 { 230 __sme_map_range(ppd, PMD_FLAGS_ENC, PTE_FLAGS_ENC); 231 } 232 233 static void __init sme_map_range_decrypted(struct sme_populate_pgd_data *ppd) 234 { 235 __sme_map_range(ppd, PMD_FLAGS_DEC, PTE_FLAGS_DEC); 236 } 237 238 static void __init sme_map_range_decrypted_wp(struct sme_populate_pgd_data *ppd) 239 { 240 __sme_map_range(ppd, PMD_FLAGS_DEC_WP, PTE_FLAGS_DEC_WP); 241 } 242 243 static unsigned long __init sme_pgtable_calc(unsigned long len) 244 { 245 unsigned long entries = 0, tables = 0; 246 247 /* 248 * Perform a relatively simplistic calculation of the pagetable 249 * entries that are needed. Those mappings will be covered mostly 250 * by 2MB PMD entries so we can conservatively calculate the required 251 * number of P4D, PUD and PMD structures needed to perform the 252 * mappings. For mappings that are not 2MB aligned, PTE mappings 253 * would be needed for the start and end portion of the address range 254 * that fall outside of the 2MB alignment. This results in, at most, 255 * two extra pages to hold PTE entries for each range that is mapped. 256 * Incrementing the count for each covers the case where the addresses 257 * cross entries. 258 */ 259 260 /* PGDIR_SIZE is equal to P4D_SIZE on 4-level machine. */ 261 if (PTRS_PER_P4D > 1) 262 entries += (DIV_ROUND_UP(len, PGDIR_SIZE) + 1) * sizeof(p4d_t) * PTRS_PER_P4D; 263 entries += (DIV_ROUND_UP(len, P4D_SIZE) + 1) * sizeof(pud_t) * PTRS_PER_PUD; 264 entries += (DIV_ROUND_UP(len, PUD_SIZE) + 1) * sizeof(pmd_t) * PTRS_PER_PMD; 265 entries += 2 * sizeof(pte_t) * PTRS_PER_PTE; 266 267 /* 268 * Now calculate the added pagetable structures needed to populate 269 * the new pagetables. 270 */ 271 272 if (PTRS_PER_P4D > 1) 273 tables += DIV_ROUND_UP(entries, PGDIR_SIZE) * sizeof(p4d_t) * PTRS_PER_P4D; 274 tables += DIV_ROUND_UP(entries, P4D_SIZE) * sizeof(pud_t) * PTRS_PER_PUD; 275 tables += DIV_ROUND_UP(entries, PUD_SIZE) * sizeof(pmd_t) * PTRS_PER_PMD; 276 277 return entries + tables; 278 } 279 280 void __init sme_encrypt_kernel(struct boot_params *bp) 281 { 282 unsigned long workarea_start, workarea_end, workarea_len; 283 unsigned long execute_start, execute_end, execute_len; 284 unsigned long kernel_start, kernel_end, kernel_len; 285 unsigned long initrd_start, initrd_end, initrd_len; 286 struct sme_populate_pgd_data ppd; 287 unsigned long pgtable_area_len; 288 unsigned long decrypted_base; 289 290 if (!sme_active()) 291 return; 292 293 /* 294 * Prepare for encrypting the kernel and initrd by building new 295 * pagetables with the necessary attributes needed to encrypt the 296 * kernel in place. 297 * 298 * One range of virtual addresses will map the memory occupied 299 * by the kernel and initrd as encrypted. 300 * 301 * Another range of virtual addresses will map the memory occupied 302 * by the kernel and initrd as decrypted and write-protected. 303 * 304 * The use of write-protect attribute will prevent any of the 305 * memory from being cached. 306 */ 307 308 /* Physical addresses gives us the identity mapped virtual addresses */ 309 kernel_start = __pa_symbol(_text); 310 kernel_end = ALIGN(__pa_symbol(_end), PMD_PAGE_SIZE); 311 kernel_len = kernel_end - kernel_start; 312 313 initrd_start = 0; 314 initrd_end = 0; 315 initrd_len = 0; 316 #ifdef CONFIG_BLK_DEV_INITRD 317 initrd_len = (unsigned long)bp->hdr.ramdisk_size | 318 ((unsigned long)bp->ext_ramdisk_size << 32); 319 if (initrd_len) { 320 initrd_start = (unsigned long)bp->hdr.ramdisk_image | 321 ((unsigned long)bp->ext_ramdisk_image << 32); 322 initrd_end = PAGE_ALIGN(initrd_start + initrd_len); 323 initrd_len = initrd_end - initrd_start; 324 } 325 #endif 326 327 /* 328 * We're running identity mapped, so we must obtain the address to the 329 * SME encryption workarea using rip-relative addressing. 330 */ 331 asm ("lea sme_workarea(%%rip), %0" 332 : "=r" (workarea_start) 333 : "p" (sme_workarea)); 334 335 /* 336 * Calculate required number of workarea bytes needed: 337 * executable encryption area size: 338 * stack page (PAGE_SIZE) 339 * encryption routine page (PAGE_SIZE) 340 * intermediate copy buffer (PMD_PAGE_SIZE) 341 * pagetable structures for the encryption of the kernel 342 * pagetable structures for workarea (in case not currently mapped) 343 */ 344 execute_start = workarea_start; 345 execute_end = execute_start + (PAGE_SIZE * 2) + PMD_PAGE_SIZE; 346 execute_len = execute_end - execute_start; 347 348 /* 349 * One PGD for both encrypted and decrypted mappings and a set of 350 * PUDs and PMDs for each of the encrypted and decrypted mappings. 351 */ 352 pgtable_area_len = sizeof(pgd_t) * PTRS_PER_PGD; 353 pgtable_area_len += sme_pgtable_calc(execute_end - kernel_start) * 2; 354 if (initrd_len) 355 pgtable_area_len += sme_pgtable_calc(initrd_len) * 2; 356 357 /* PUDs and PMDs needed in the current pagetables for the workarea */ 358 pgtable_area_len += sme_pgtable_calc(execute_len + pgtable_area_len); 359 360 /* 361 * The total workarea includes the executable encryption area and 362 * the pagetable area. The start of the workarea is already 2MB 363 * aligned, align the end of the workarea on a 2MB boundary so that 364 * we don't try to create/allocate PTE entries from the workarea 365 * before it is mapped. 366 */ 367 workarea_len = execute_len + pgtable_area_len; 368 workarea_end = ALIGN(workarea_start + workarea_len, PMD_PAGE_SIZE); 369 370 /* 371 * Set the address to the start of where newly created pagetable 372 * structures (PGDs, PUDs and PMDs) will be allocated. New pagetable 373 * structures are created when the workarea is added to the current 374 * pagetables and when the new encrypted and decrypted kernel 375 * mappings are populated. 376 */ 377 ppd.pgtable_area = (void *)execute_end; 378 379 /* 380 * Make sure the current pagetable structure has entries for 381 * addressing the workarea. 382 */ 383 ppd.pgd = (pgd_t *)native_read_cr3_pa(); 384 ppd.paddr = workarea_start; 385 ppd.vaddr = workarea_start; 386 ppd.vaddr_end = workarea_end; 387 sme_map_range_decrypted(&ppd); 388 389 /* Flush the TLB - no globals so cr3 is enough */ 390 native_write_cr3(__native_read_cr3()); 391 392 /* 393 * A new pagetable structure is being built to allow for the kernel 394 * and initrd to be encrypted. It starts with an empty PGD that will 395 * then be populated with new PUDs and PMDs as the encrypted and 396 * decrypted kernel mappings are created. 397 */ 398 ppd.pgd = ppd.pgtable_area; 399 memset(ppd.pgd, 0, sizeof(pgd_t) * PTRS_PER_PGD); 400 ppd.pgtable_area += sizeof(pgd_t) * PTRS_PER_PGD; 401 402 /* 403 * A different PGD index/entry must be used to get different 404 * pagetable entries for the decrypted mapping. Choose the next 405 * PGD index and convert it to a virtual address to be used as 406 * the base of the mapping. 407 */ 408 decrypted_base = (pgd_index(workarea_end) + 1) & (PTRS_PER_PGD - 1); 409 if (initrd_len) { 410 unsigned long check_base; 411 412 check_base = (pgd_index(initrd_end) + 1) & (PTRS_PER_PGD - 1); 413 decrypted_base = max(decrypted_base, check_base); 414 } 415 decrypted_base <<= PGDIR_SHIFT; 416 417 /* Add encrypted kernel (identity) mappings */ 418 ppd.paddr = kernel_start; 419 ppd.vaddr = kernel_start; 420 ppd.vaddr_end = kernel_end; 421 sme_map_range_encrypted(&ppd); 422 423 /* Add decrypted, write-protected kernel (non-identity) mappings */ 424 ppd.paddr = kernel_start; 425 ppd.vaddr = kernel_start + decrypted_base; 426 ppd.vaddr_end = kernel_end + decrypted_base; 427 sme_map_range_decrypted_wp(&ppd); 428 429 if (initrd_len) { 430 /* Add encrypted initrd (identity) mappings */ 431 ppd.paddr = initrd_start; 432 ppd.vaddr = initrd_start; 433 ppd.vaddr_end = initrd_end; 434 sme_map_range_encrypted(&ppd); 435 /* 436 * Add decrypted, write-protected initrd (non-identity) mappings 437 */ 438 ppd.paddr = initrd_start; 439 ppd.vaddr = initrd_start + decrypted_base; 440 ppd.vaddr_end = initrd_end + decrypted_base; 441 sme_map_range_decrypted_wp(&ppd); 442 } 443 444 /* Add decrypted workarea mappings to both kernel mappings */ 445 ppd.paddr = workarea_start; 446 ppd.vaddr = workarea_start; 447 ppd.vaddr_end = workarea_end; 448 sme_map_range_decrypted(&ppd); 449 450 ppd.paddr = workarea_start; 451 ppd.vaddr = workarea_start + decrypted_base; 452 ppd.vaddr_end = workarea_end + decrypted_base; 453 sme_map_range_decrypted(&ppd); 454 455 /* Perform the encryption */ 456 sme_encrypt_execute(kernel_start, kernel_start + decrypted_base, 457 kernel_len, workarea_start, (unsigned long)ppd.pgd); 458 459 if (initrd_len) 460 sme_encrypt_execute(initrd_start, initrd_start + decrypted_base, 461 initrd_len, workarea_start, 462 (unsigned long)ppd.pgd); 463 464 /* 465 * At this point we are running encrypted. Remove the mappings for 466 * the decrypted areas - all that is needed for this is to remove 467 * the PGD entry/entries. 468 */ 469 ppd.vaddr = kernel_start + decrypted_base; 470 ppd.vaddr_end = kernel_end + decrypted_base; 471 sme_clear_pgd(&ppd); 472 473 if (initrd_len) { 474 ppd.vaddr = initrd_start + decrypted_base; 475 ppd.vaddr_end = initrd_end + decrypted_base; 476 sme_clear_pgd(&ppd); 477 } 478 479 ppd.vaddr = workarea_start + decrypted_base; 480 ppd.vaddr_end = workarea_end + decrypted_base; 481 sme_clear_pgd(&ppd); 482 483 /* Flush the TLB - no globals so cr3 is enough */ 484 native_write_cr3(__native_read_cr3()); 485 } 486 487 void __init sme_enable(struct boot_params *bp) 488 { 489 const char *cmdline_ptr, *cmdline_arg, *cmdline_on, *cmdline_off; 490 unsigned int eax, ebx, ecx, edx; 491 unsigned long feature_mask; 492 bool active_by_default; 493 unsigned long me_mask; 494 char buffer[16]; 495 u64 msr; 496 497 /* Check for the SME/SEV support leaf */ 498 eax = 0x80000000; 499 ecx = 0; 500 native_cpuid(&eax, &ebx, &ecx, &edx); 501 if (eax < 0x8000001f) 502 return; 503 504 #define AMD_SME_BIT BIT(0) 505 #define AMD_SEV_BIT BIT(1) 506 507 /* 508 * Check for the SME/SEV feature: 509 * CPUID Fn8000_001F[EAX] 510 * - Bit 0 - Secure Memory Encryption support 511 * - Bit 1 - Secure Encrypted Virtualization support 512 * CPUID Fn8000_001F[EBX] 513 * - Bits 5:0 - Pagetable bit position used to indicate encryption 514 */ 515 eax = 0x8000001f; 516 ecx = 0; 517 native_cpuid(&eax, &ebx, &ecx, &edx); 518 /* Check whether SEV or SME is supported */ 519 if (!(eax & (AMD_SEV_BIT | AMD_SME_BIT))) 520 return; 521 522 me_mask = 1UL << (ebx & 0x3f); 523 524 /* Check the SEV MSR whether SEV or SME is enabled */ 525 sev_status = __rdmsr(MSR_AMD64_SEV); 526 feature_mask = (sev_status & MSR_AMD64_SEV_ENABLED) ? AMD_SEV_BIT : AMD_SME_BIT; 527 528 /* Check if memory encryption is enabled */ 529 if (feature_mask == AMD_SME_BIT) { 530 /* 531 * No SME if Hypervisor bit is set. This check is here to 532 * prevent a guest from trying to enable SME. For running as a 533 * KVM guest the MSR_AMD64_SYSCFG will be sufficient, but there 534 * might be other hypervisors which emulate that MSR as non-zero 535 * or even pass it through to the guest. 536 * A malicious hypervisor can still trick a guest into this 537 * path, but there is no way to protect against that. 538 */ 539 eax = 1; 540 ecx = 0; 541 native_cpuid(&eax, &ebx, &ecx, &edx); 542 if (ecx & BIT(31)) 543 return; 544 545 /* For SME, check the SYSCFG MSR */ 546 msr = __rdmsr(MSR_AMD64_SYSCFG); 547 if (!(msr & MSR_AMD64_SYSCFG_MEM_ENCRYPT)) 548 return; 549 } else { 550 /* SEV state cannot be controlled by a command line option */ 551 sme_me_mask = me_mask; 552 physical_mask &= ~sme_me_mask; 553 return; 554 } 555 556 /* 557 * Fixups have not been applied to phys_base yet and we're running 558 * identity mapped, so we must obtain the address to the SME command 559 * line argument data using rip-relative addressing. 560 */ 561 asm ("lea sme_cmdline_arg(%%rip), %0" 562 : "=r" (cmdline_arg) 563 : "p" (sme_cmdline_arg)); 564 asm ("lea sme_cmdline_on(%%rip), %0" 565 : "=r" (cmdline_on) 566 : "p" (sme_cmdline_on)); 567 asm ("lea sme_cmdline_off(%%rip), %0" 568 : "=r" (cmdline_off) 569 : "p" (sme_cmdline_off)); 570 571 if (IS_ENABLED(CONFIG_AMD_MEM_ENCRYPT_ACTIVE_BY_DEFAULT)) 572 active_by_default = true; 573 else 574 active_by_default = false; 575 576 cmdline_ptr = (const char *)((u64)bp->hdr.cmd_line_ptr | 577 ((u64)bp->ext_cmd_line_ptr << 32)); 578 579 cmdline_find_option(cmdline_ptr, cmdline_arg, buffer, sizeof(buffer)); 580 581 if (!strncmp(buffer, cmdline_on, sizeof(buffer))) 582 sme_me_mask = me_mask; 583 else if (!strncmp(buffer, cmdline_off, sizeof(buffer))) 584 sme_me_mask = 0; 585 else 586 sme_me_mask = active_by_default ? me_mask : 0; 587 588 physical_mask &= ~sme_me_mask; 589 } 590