1 /* 2 * AMD Memory Encryption Support 3 * 4 * Copyright (C) 2016 Advanced Micro Devices, Inc. 5 * 6 * Author: Tom Lendacky <thomas.lendacky@amd.com> 7 * 8 * This program is free software; you can redistribute it and/or modify 9 * it under the terms of the GNU General Public License version 2 as 10 * published by the Free Software Foundation. 11 */ 12 13 #define DISABLE_BRANCH_PROFILING 14 15 /* 16 * Since we're dealing with identity mappings, physical and virtual 17 * addresses are the same, so override these defines which are ultimately 18 * used by the headers in misc.h. 19 */ 20 #define __pa(x) ((unsigned long)(x)) 21 #define __va(x) ((void *)((unsigned long)(x))) 22 23 /* 24 * Special hack: we have to be careful, because no indirections are 25 * allowed here, and paravirt_ops is a kind of one. As it will only run in 26 * baremetal anyway, we just keep it from happening. (This list needs to 27 * be extended when new paravirt and debugging variants are added.) 28 */ 29 #undef CONFIG_PARAVIRT 30 #undef CONFIG_PARAVIRT_XXL 31 #undef CONFIG_PARAVIRT_SPINLOCKS 32 33 #include <linux/kernel.h> 34 #include <linux/mm.h> 35 #include <linux/mem_encrypt.h> 36 37 #include <asm/setup.h> 38 #include <asm/sections.h> 39 #include <asm/cmdline.h> 40 41 #include "mm_internal.h" 42 43 #define PGD_FLAGS _KERNPG_TABLE_NOENC 44 #define P4D_FLAGS _KERNPG_TABLE_NOENC 45 #define PUD_FLAGS _KERNPG_TABLE_NOENC 46 #define PMD_FLAGS _KERNPG_TABLE_NOENC 47 48 #define PMD_FLAGS_LARGE (__PAGE_KERNEL_LARGE_EXEC & ~_PAGE_GLOBAL) 49 50 #define PMD_FLAGS_DEC PMD_FLAGS_LARGE 51 #define PMD_FLAGS_DEC_WP ((PMD_FLAGS_DEC & ~_PAGE_CACHE_MASK) | \ 52 (_PAGE_PAT | _PAGE_PWT)) 53 54 #define PMD_FLAGS_ENC (PMD_FLAGS_LARGE | _PAGE_ENC) 55 56 #define PTE_FLAGS (__PAGE_KERNEL_EXEC & ~_PAGE_GLOBAL) 57 58 #define PTE_FLAGS_DEC PTE_FLAGS 59 #define PTE_FLAGS_DEC_WP ((PTE_FLAGS_DEC & ~_PAGE_CACHE_MASK) | \ 60 (_PAGE_PAT | _PAGE_PWT)) 61 62 #define PTE_FLAGS_ENC (PTE_FLAGS | _PAGE_ENC) 63 64 struct sme_populate_pgd_data { 65 void *pgtable_area; 66 pgd_t *pgd; 67 68 pmdval_t pmd_flags; 69 pteval_t pte_flags; 70 unsigned long paddr; 71 72 unsigned long vaddr; 73 unsigned long vaddr_end; 74 }; 75 76 static char sme_cmdline_arg[] __initdata = "mem_encrypt"; 77 static char sme_cmdline_on[] __initdata = "on"; 78 static char sme_cmdline_off[] __initdata = "off"; 79 80 static void __init sme_clear_pgd(struct sme_populate_pgd_data *ppd) 81 { 82 unsigned long pgd_start, pgd_end, pgd_size; 83 pgd_t *pgd_p; 84 85 pgd_start = ppd->vaddr & PGDIR_MASK; 86 pgd_end = ppd->vaddr_end & PGDIR_MASK; 87 88 pgd_size = (((pgd_end - pgd_start) / PGDIR_SIZE) + 1) * sizeof(pgd_t); 89 90 pgd_p = ppd->pgd + pgd_index(ppd->vaddr); 91 92 memset(pgd_p, 0, pgd_size); 93 } 94 95 static pud_t __init *sme_prepare_pgd(struct sme_populate_pgd_data *ppd) 96 { 97 pgd_t *pgd; 98 p4d_t *p4d; 99 pud_t *pud; 100 pmd_t *pmd; 101 102 pgd = ppd->pgd + pgd_index(ppd->vaddr); 103 if (pgd_none(*pgd)) { 104 p4d = ppd->pgtable_area; 105 memset(p4d, 0, sizeof(*p4d) * PTRS_PER_P4D); 106 ppd->pgtable_area += sizeof(*p4d) * PTRS_PER_P4D; 107 set_pgd(pgd, __pgd(PGD_FLAGS | __pa(p4d))); 108 } 109 110 p4d = p4d_offset(pgd, ppd->vaddr); 111 if (p4d_none(*p4d)) { 112 pud = ppd->pgtable_area; 113 memset(pud, 0, sizeof(*pud) * PTRS_PER_PUD); 114 ppd->pgtable_area += sizeof(*pud) * PTRS_PER_PUD; 115 set_p4d(p4d, __p4d(P4D_FLAGS | __pa(pud))); 116 } 117 118 pud = pud_offset(p4d, ppd->vaddr); 119 if (pud_none(*pud)) { 120 pmd = ppd->pgtable_area; 121 memset(pmd, 0, sizeof(*pmd) * PTRS_PER_PMD); 122 ppd->pgtable_area += sizeof(*pmd) * PTRS_PER_PMD; 123 set_pud(pud, __pud(PUD_FLAGS | __pa(pmd))); 124 } 125 126 if (pud_large(*pud)) 127 return NULL; 128 129 return pud; 130 } 131 132 static void __init sme_populate_pgd_large(struct sme_populate_pgd_data *ppd) 133 { 134 pud_t *pud; 135 pmd_t *pmd; 136 137 pud = sme_prepare_pgd(ppd); 138 if (!pud) 139 return; 140 141 pmd = pmd_offset(pud, ppd->vaddr); 142 if (pmd_large(*pmd)) 143 return; 144 145 set_pmd(pmd, __pmd(ppd->paddr | ppd->pmd_flags)); 146 } 147 148 static void __init sme_populate_pgd(struct sme_populate_pgd_data *ppd) 149 { 150 pud_t *pud; 151 pmd_t *pmd; 152 pte_t *pte; 153 154 pud = sme_prepare_pgd(ppd); 155 if (!pud) 156 return; 157 158 pmd = pmd_offset(pud, ppd->vaddr); 159 if (pmd_none(*pmd)) { 160 pte = ppd->pgtable_area; 161 memset(pte, 0, sizeof(pte) * PTRS_PER_PTE); 162 ppd->pgtable_area += sizeof(pte) * PTRS_PER_PTE; 163 set_pmd(pmd, __pmd(PMD_FLAGS | __pa(pte))); 164 } 165 166 if (pmd_large(*pmd)) 167 return; 168 169 pte = pte_offset_map(pmd, ppd->vaddr); 170 if (pte_none(*pte)) 171 set_pte(pte, __pte(ppd->paddr | ppd->pte_flags)); 172 } 173 174 static void __init __sme_map_range_pmd(struct sme_populate_pgd_data *ppd) 175 { 176 while (ppd->vaddr < ppd->vaddr_end) { 177 sme_populate_pgd_large(ppd); 178 179 ppd->vaddr += PMD_PAGE_SIZE; 180 ppd->paddr += PMD_PAGE_SIZE; 181 } 182 } 183 184 static void __init __sme_map_range_pte(struct sme_populate_pgd_data *ppd) 185 { 186 while (ppd->vaddr < ppd->vaddr_end) { 187 sme_populate_pgd(ppd); 188 189 ppd->vaddr += PAGE_SIZE; 190 ppd->paddr += PAGE_SIZE; 191 } 192 } 193 194 static void __init __sme_map_range(struct sme_populate_pgd_data *ppd, 195 pmdval_t pmd_flags, pteval_t pte_flags) 196 { 197 unsigned long vaddr_end; 198 199 ppd->pmd_flags = pmd_flags; 200 ppd->pte_flags = pte_flags; 201 202 /* Save original end value since we modify the struct value */ 203 vaddr_end = ppd->vaddr_end; 204 205 /* If start is not 2MB aligned, create PTE entries */ 206 ppd->vaddr_end = ALIGN(ppd->vaddr, PMD_PAGE_SIZE); 207 __sme_map_range_pte(ppd); 208 209 /* Create PMD entries */ 210 ppd->vaddr_end = vaddr_end & PMD_PAGE_MASK; 211 __sme_map_range_pmd(ppd); 212 213 /* If end is not 2MB aligned, create PTE entries */ 214 ppd->vaddr_end = vaddr_end; 215 __sme_map_range_pte(ppd); 216 } 217 218 static void __init sme_map_range_encrypted(struct sme_populate_pgd_data *ppd) 219 { 220 __sme_map_range(ppd, PMD_FLAGS_ENC, PTE_FLAGS_ENC); 221 } 222 223 static void __init sme_map_range_decrypted(struct sme_populate_pgd_data *ppd) 224 { 225 __sme_map_range(ppd, PMD_FLAGS_DEC, PTE_FLAGS_DEC); 226 } 227 228 static void __init sme_map_range_decrypted_wp(struct sme_populate_pgd_data *ppd) 229 { 230 __sme_map_range(ppd, PMD_FLAGS_DEC_WP, PTE_FLAGS_DEC_WP); 231 } 232 233 static unsigned long __init sme_pgtable_calc(unsigned long len) 234 { 235 unsigned long entries = 0, tables = 0; 236 237 /* 238 * Perform a relatively simplistic calculation of the pagetable 239 * entries that are needed. Those mappings will be covered mostly 240 * by 2MB PMD entries so we can conservatively calculate the required 241 * number of P4D, PUD and PMD structures needed to perform the 242 * mappings. For mappings that are not 2MB aligned, PTE mappings 243 * would be needed for the start and end portion of the address range 244 * that fall outside of the 2MB alignment. This results in, at most, 245 * two extra pages to hold PTE entries for each range that is mapped. 246 * Incrementing the count for each covers the case where the addresses 247 * cross entries. 248 */ 249 250 /* PGDIR_SIZE is equal to P4D_SIZE on 4-level machine. */ 251 if (PTRS_PER_P4D > 1) 252 entries += (DIV_ROUND_UP(len, PGDIR_SIZE) + 1) * sizeof(p4d_t) * PTRS_PER_P4D; 253 entries += (DIV_ROUND_UP(len, P4D_SIZE) + 1) * sizeof(pud_t) * PTRS_PER_PUD; 254 entries += (DIV_ROUND_UP(len, PUD_SIZE) + 1) * sizeof(pmd_t) * PTRS_PER_PMD; 255 entries += 2 * sizeof(pte_t) * PTRS_PER_PTE; 256 257 /* 258 * Now calculate the added pagetable structures needed to populate 259 * the new pagetables. 260 */ 261 262 if (PTRS_PER_P4D > 1) 263 tables += DIV_ROUND_UP(entries, PGDIR_SIZE) * sizeof(p4d_t) * PTRS_PER_P4D; 264 tables += DIV_ROUND_UP(entries, P4D_SIZE) * sizeof(pud_t) * PTRS_PER_PUD; 265 tables += DIV_ROUND_UP(entries, PUD_SIZE) * sizeof(pmd_t) * PTRS_PER_PMD; 266 267 return entries + tables; 268 } 269 270 void __init sme_encrypt_kernel(struct boot_params *bp) 271 { 272 unsigned long workarea_start, workarea_end, workarea_len; 273 unsigned long execute_start, execute_end, execute_len; 274 unsigned long kernel_start, kernel_end, kernel_len; 275 unsigned long initrd_start, initrd_end, initrd_len; 276 struct sme_populate_pgd_data ppd; 277 unsigned long pgtable_area_len; 278 unsigned long decrypted_base; 279 280 if (!sme_active()) 281 return; 282 283 /* 284 * Prepare for encrypting the kernel and initrd by building new 285 * pagetables with the necessary attributes needed to encrypt the 286 * kernel in place. 287 * 288 * One range of virtual addresses will map the memory occupied 289 * by the kernel and initrd as encrypted. 290 * 291 * Another range of virtual addresses will map the memory occupied 292 * by the kernel and initrd as decrypted and write-protected. 293 * 294 * The use of write-protect attribute will prevent any of the 295 * memory from being cached. 296 */ 297 298 /* Physical addresses gives us the identity mapped virtual addresses */ 299 kernel_start = __pa_symbol(_text); 300 kernel_end = ALIGN(__pa_symbol(_end), PMD_PAGE_SIZE); 301 kernel_len = kernel_end - kernel_start; 302 303 initrd_start = 0; 304 initrd_end = 0; 305 initrd_len = 0; 306 #ifdef CONFIG_BLK_DEV_INITRD 307 initrd_len = (unsigned long)bp->hdr.ramdisk_size | 308 ((unsigned long)bp->ext_ramdisk_size << 32); 309 if (initrd_len) { 310 initrd_start = (unsigned long)bp->hdr.ramdisk_image | 311 ((unsigned long)bp->ext_ramdisk_image << 32); 312 initrd_end = PAGE_ALIGN(initrd_start + initrd_len); 313 initrd_len = initrd_end - initrd_start; 314 } 315 #endif 316 317 /* Set the encryption workarea to be immediately after the kernel */ 318 workarea_start = kernel_end; 319 320 /* 321 * Calculate required number of workarea bytes needed: 322 * executable encryption area size: 323 * stack page (PAGE_SIZE) 324 * encryption routine page (PAGE_SIZE) 325 * intermediate copy buffer (PMD_PAGE_SIZE) 326 * pagetable structures for the encryption of the kernel 327 * pagetable structures for workarea (in case not currently mapped) 328 */ 329 execute_start = workarea_start; 330 execute_end = execute_start + (PAGE_SIZE * 2) + PMD_PAGE_SIZE; 331 execute_len = execute_end - execute_start; 332 333 /* 334 * One PGD for both encrypted and decrypted mappings and a set of 335 * PUDs and PMDs for each of the encrypted and decrypted mappings. 336 */ 337 pgtable_area_len = sizeof(pgd_t) * PTRS_PER_PGD; 338 pgtable_area_len += sme_pgtable_calc(execute_end - kernel_start) * 2; 339 if (initrd_len) 340 pgtable_area_len += sme_pgtable_calc(initrd_len) * 2; 341 342 /* PUDs and PMDs needed in the current pagetables for the workarea */ 343 pgtable_area_len += sme_pgtable_calc(execute_len + pgtable_area_len); 344 345 /* 346 * The total workarea includes the executable encryption area and 347 * the pagetable area. The start of the workarea is already 2MB 348 * aligned, align the end of the workarea on a 2MB boundary so that 349 * we don't try to create/allocate PTE entries from the workarea 350 * before it is mapped. 351 */ 352 workarea_len = execute_len + pgtable_area_len; 353 workarea_end = ALIGN(workarea_start + workarea_len, PMD_PAGE_SIZE); 354 355 /* 356 * Set the address to the start of where newly created pagetable 357 * structures (PGDs, PUDs and PMDs) will be allocated. New pagetable 358 * structures are created when the workarea is added to the current 359 * pagetables and when the new encrypted and decrypted kernel 360 * mappings are populated. 361 */ 362 ppd.pgtable_area = (void *)execute_end; 363 364 /* 365 * Make sure the current pagetable structure has entries for 366 * addressing the workarea. 367 */ 368 ppd.pgd = (pgd_t *)native_read_cr3_pa(); 369 ppd.paddr = workarea_start; 370 ppd.vaddr = workarea_start; 371 ppd.vaddr_end = workarea_end; 372 sme_map_range_decrypted(&ppd); 373 374 /* Flush the TLB - no globals so cr3 is enough */ 375 native_write_cr3(__native_read_cr3()); 376 377 /* 378 * A new pagetable structure is being built to allow for the kernel 379 * and initrd to be encrypted. It starts with an empty PGD that will 380 * then be populated with new PUDs and PMDs as the encrypted and 381 * decrypted kernel mappings are created. 382 */ 383 ppd.pgd = ppd.pgtable_area; 384 memset(ppd.pgd, 0, sizeof(pgd_t) * PTRS_PER_PGD); 385 ppd.pgtable_area += sizeof(pgd_t) * PTRS_PER_PGD; 386 387 /* 388 * A different PGD index/entry must be used to get different 389 * pagetable entries for the decrypted mapping. Choose the next 390 * PGD index and convert it to a virtual address to be used as 391 * the base of the mapping. 392 */ 393 decrypted_base = (pgd_index(workarea_end) + 1) & (PTRS_PER_PGD - 1); 394 if (initrd_len) { 395 unsigned long check_base; 396 397 check_base = (pgd_index(initrd_end) + 1) & (PTRS_PER_PGD - 1); 398 decrypted_base = max(decrypted_base, check_base); 399 } 400 decrypted_base <<= PGDIR_SHIFT; 401 402 /* Add encrypted kernel (identity) mappings */ 403 ppd.paddr = kernel_start; 404 ppd.vaddr = kernel_start; 405 ppd.vaddr_end = kernel_end; 406 sme_map_range_encrypted(&ppd); 407 408 /* Add decrypted, write-protected kernel (non-identity) mappings */ 409 ppd.paddr = kernel_start; 410 ppd.vaddr = kernel_start + decrypted_base; 411 ppd.vaddr_end = kernel_end + decrypted_base; 412 sme_map_range_decrypted_wp(&ppd); 413 414 if (initrd_len) { 415 /* Add encrypted initrd (identity) mappings */ 416 ppd.paddr = initrd_start; 417 ppd.vaddr = initrd_start; 418 ppd.vaddr_end = initrd_end; 419 sme_map_range_encrypted(&ppd); 420 /* 421 * Add decrypted, write-protected initrd (non-identity) mappings 422 */ 423 ppd.paddr = initrd_start; 424 ppd.vaddr = initrd_start + decrypted_base; 425 ppd.vaddr_end = initrd_end + decrypted_base; 426 sme_map_range_decrypted_wp(&ppd); 427 } 428 429 /* Add decrypted workarea mappings to both kernel mappings */ 430 ppd.paddr = workarea_start; 431 ppd.vaddr = workarea_start; 432 ppd.vaddr_end = workarea_end; 433 sme_map_range_decrypted(&ppd); 434 435 ppd.paddr = workarea_start; 436 ppd.vaddr = workarea_start + decrypted_base; 437 ppd.vaddr_end = workarea_end + decrypted_base; 438 sme_map_range_decrypted(&ppd); 439 440 /* Perform the encryption */ 441 sme_encrypt_execute(kernel_start, kernel_start + decrypted_base, 442 kernel_len, workarea_start, (unsigned long)ppd.pgd); 443 444 if (initrd_len) 445 sme_encrypt_execute(initrd_start, initrd_start + decrypted_base, 446 initrd_len, workarea_start, 447 (unsigned long)ppd.pgd); 448 449 /* 450 * At this point we are running encrypted. Remove the mappings for 451 * the decrypted areas - all that is needed for this is to remove 452 * the PGD entry/entries. 453 */ 454 ppd.vaddr = kernel_start + decrypted_base; 455 ppd.vaddr_end = kernel_end + decrypted_base; 456 sme_clear_pgd(&ppd); 457 458 if (initrd_len) { 459 ppd.vaddr = initrd_start + decrypted_base; 460 ppd.vaddr_end = initrd_end + decrypted_base; 461 sme_clear_pgd(&ppd); 462 } 463 464 ppd.vaddr = workarea_start + decrypted_base; 465 ppd.vaddr_end = workarea_end + decrypted_base; 466 sme_clear_pgd(&ppd); 467 468 /* Flush the TLB - no globals so cr3 is enough */ 469 native_write_cr3(__native_read_cr3()); 470 } 471 472 void __init sme_enable(struct boot_params *bp) 473 { 474 const char *cmdline_ptr, *cmdline_arg, *cmdline_on, *cmdline_off; 475 unsigned int eax, ebx, ecx, edx; 476 unsigned long feature_mask; 477 bool active_by_default; 478 unsigned long me_mask; 479 char buffer[16]; 480 u64 msr; 481 482 /* Check for the SME/SEV support leaf */ 483 eax = 0x80000000; 484 ecx = 0; 485 native_cpuid(&eax, &ebx, &ecx, &edx); 486 if (eax < 0x8000001f) 487 return; 488 489 #define AMD_SME_BIT BIT(0) 490 #define AMD_SEV_BIT BIT(1) 491 /* 492 * Set the feature mask (SME or SEV) based on whether we are 493 * running under a hypervisor. 494 */ 495 eax = 1; 496 ecx = 0; 497 native_cpuid(&eax, &ebx, &ecx, &edx); 498 feature_mask = (ecx & BIT(31)) ? AMD_SEV_BIT : AMD_SME_BIT; 499 500 /* 501 * Check for the SME/SEV feature: 502 * CPUID Fn8000_001F[EAX] 503 * - Bit 0 - Secure Memory Encryption support 504 * - Bit 1 - Secure Encrypted Virtualization support 505 * CPUID Fn8000_001F[EBX] 506 * - Bits 5:0 - Pagetable bit position used to indicate encryption 507 */ 508 eax = 0x8000001f; 509 ecx = 0; 510 native_cpuid(&eax, &ebx, &ecx, &edx); 511 if (!(eax & feature_mask)) 512 return; 513 514 me_mask = 1UL << (ebx & 0x3f); 515 516 /* Check if memory encryption is enabled */ 517 if (feature_mask == AMD_SME_BIT) { 518 /* For SME, check the SYSCFG MSR */ 519 msr = __rdmsr(MSR_K8_SYSCFG); 520 if (!(msr & MSR_K8_SYSCFG_MEM_ENCRYPT)) 521 return; 522 } else { 523 /* For SEV, check the SEV MSR */ 524 msr = __rdmsr(MSR_AMD64_SEV); 525 if (!(msr & MSR_AMD64_SEV_ENABLED)) 526 return; 527 528 /* SEV state cannot be controlled by a command line option */ 529 sme_me_mask = me_mask; 530 sev_enabled = true; 531 physical_mask &= ~sme_me_mask; 532 return; 533 } 534 535 /* 536 * Fixups have not been applied to phys_base yet and we're running 537 * identity mapped, so we must obtain the address to the SME command 538 * line argument data using rip-relative addressing. 539 */ 540 asm ("lea sme_cmdline_arg(%%rip), %0" 541 : "=r" (cmdline_arg) 542 : "p" (sme_cmdline_arg)); 543 asm ("lea sme_cmdline_on(%%rip), %0" 544 : "=r" (cmdline_on) 545 : "p" (sme_cmdline_on)); 546 asm ("lea sme_cmdline_off(%%rip), %0" 547 : "=r" (cmdline_off) 548 : "p" (sme_cmdline_off)); 549 550 if (IS_ENABLED(CONFIG_AMD_MEM_ENCRYPT_ACTIVE_BY_DEFAULT)) 551 active_by_default = true; 552 else 553 active_by_default = false; 554 555 cmdline_ptr = (const char *)((u64)bp->hdr.cmd_line_ptr | 556 ((u64)bp->ext_cmd_line_ptr << 32)); 557 558 cmdline_find_option(cmdline_ptr, cmdline_arg, buffer, sizeof(buffer)); 559 560 if (!strncmp(buffer, cmdline_on, sizeof(buffer))) 561 sme_me_mask = me_mask; 562 else if (!strncmp(buffer, cmdline_off, sizeof(buffer))) 563 sme_me_mask = 0; 564 else 565 sme_me_mask = active_by_default ? me_mask : 0; 566 567 physical_mask &= ~sme_me_mask; 568 } 569