linux/arch/x86/mm/mem_encrypt.c
<<
>>
Prefs
   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#include <linux/linkage.h>
  13#include <linux/init.h>
  14#include <linux/mm.h>
  15#include <linux/dma-direct.h>
  16#include <linux/swiotlb.h>
  17#include <linux/mem_encrypt.h>
  18#include <linux/device.h>
  19#include <linux/kernel.h>
  20#include <linux/bitops.h>
  21#include <linux/dma-mapping.h>
  22#include <linux/virtio_config.h>
  23
  24#include <asm/tlbflush.h>
  25#include <asm/fixmap.h>
  26#include <asm/setup.h>
  27#include <asm/bootparam.h>
  28#include <asm/set_memory.h>
  29#include <asm/cacheflush.h>
  30#include <asm/processor-flags.h>
  31#include <asm/msr.h>
  32#include <asm/cmdline.h>
  33
  34#include "mm_internal.h"
  35
  36/*
  37 * Since SME related variables are set early in the boot process they must
  38 * reside in the .data section so as not to be zeroed out when the .bss
  39 * section is later cleared.
  40 */
  41u64 sme_me_mask __section(".data") = 0;
  42u64 sev_status __section(".data") = 0;
  43u64 sev_check_data __section(".data") = 0;
  44EXPORT_SYMBOL(sme_me_mask);
  45DEFINE_STATIC_KEY_FALSE(sev_enable_key);
  46EXPORT_SYMBOL_GPL(sev_enable_key);
  47
  48/* Buffer used for early in-place encryption by BSP, no locking needed */
  49static char sme_early_buffer[PAGE_SIZE] __initdata __aligned(PAGE_SIZE);
  50
  51/*
  52 * This routine does not change the underlying encryption setting of the
  53 * page(s) that map this memory. It assumes that eventually the memory is
  54 * meant to be accessed as either encrypted or decrypted but the contents
  55 * are currently not in the desired state.
  56 *
  57 * This routine follows the steps outlined in the AMD64 Architecture
  58 * Programmer's Manual Volume 2, Section 7.10.8 Encrypt-in-Place.
  59 */
  60static void __init __sme_early_enc_dec(resource_size_t paddr,
  61                                       unsigned long size, bool enc)
  62{
  63        void *src, *dst;
  64        size_t len;
  65
  66        if (!sme_me_mask)
  67                return;
  68
  69        wbinvd();
  70
  71        /*
  72         * There are limited number of early mapping slots, so map (at most)
  73         * one page at time.
  74         */
  75        while (size) {
  76                len = min_t(size_t, sizeof(sme_early_buffer), size);
  77
  78                /*
  79                 * Create mappings for the current and desired format of
  80                 * the memory. Use a write-protected mapping for the source.
  81                 */
  82                src = enc ? early_memremap_decrypted_wp(paddr, len) :
  83                            early_memremap_encrypted_wp(paddr, len);
  84
  85                dst = enc ? early_memremap_encrypted(paddr, len) :
  86                            early_memremap_decrypted(paddr, len);
  87
  88                /*
  89                 * If a mapping can't be obtained to perform the operation,
  90                 * then eventual access of that area in the desired mode
  91                 * will cause a crash.
  92                 */
  93                BUG_ON(!src || !dst);
  94
  95                /*
  96                 * Use a temporary buffer, of cache-line multiple size, to
  97                 * avoid data corruption as documented in the APM.
  98                 */
  99                memcpy(sme_early_buffer, src, len);
 100                memcpy(dst, sme_early_buffer, len);
 101
 102                early_memunmap(dst, len);
 103                early_memunmap(src, len);
 104
 105                paddr += len;
 106                size -= len;
 107        }
 108}
 109
 110void __init sme_early_encrypt(resource_size_t paddr, unsigned long size)
 111{
 112        __sme_early_enc_dec(paddr, size, true);
 113}
 114
 115void __init sme_early_decrypt(resource_size_t paddr, unsigned long size)
 116{
 117        __sme_early_enc_dec(paddr, size, false);
 118}
 119
 120static void __init __sme_early_map_unmap_mem(void *vaddr, unsigned long size,
 121                                             bool map)
 122{
 123        unsigned long paddr = (unsigned long)vaddr - __PAGE_OFFSET;
 124        pmdval_t pmd_flags, pmd;
 125
 126        /* Use early_pmd_flags but remove the encryption mask */
 127        pmd_flags = __sme_clr(early_pmd_flags);
 128
 129        do {
 130                pmd = map ? (paddr & PMD_MASK) + pmd_flags : 0;
 131                __early_make_pgtable((unsigned long)vaddr, pmd);
 132
 133                vaddr += PMD_SIZE;
 134                paddr += PMD_SIZE;
 135                size = (size <= PMD_SIZE) ? 0 : size - PMD_SIZE;
 136        } while (size);
 137
 138        flush_tlb_local();
 139}
 140
 141void __init sme_unmap_bootdata(char *real_mode_data)
 142{
 143        struct boot_params *boot_data;
 144        unsigned long cmdline_paddr;
 145
 146        if (!sme_active())
 147                return;
 148
 149        /* Get the command line address before unmapping the real_mode_data */
 150        boot_data = (struct boot_params *)real_mode_data;
 151        cmdline_paddr = boot_data->hdr.cmd_line_ptr | ((u64)boot_data->ext_cmd_line_ptr << 32);
 152
 153        __sme_early_map_unmap_mem(real_mode_data, sizeof(boot_params), false);
 154
 155        if (!cmdline_paddr)
 156                return;
 157
 158        __sme_early_map_unmap_mem(__va(cmdline_paddr), COMMAND_LINE_SIZE, false);
 159}
 160
 161void __init sme_map_bootdata(char *real_mode_data)
 162{
 163        struct boot_params *boot_data;
 164        unsigned long cmdline_paddr;
 165
 166        if (!sme_active())
 167                return;
 168
 169        __sme_early_map_unmap_mem(real_mode_data, sizeof(boot_params), true);
 170
 171        /* Get the command line address after mapping the real_mode_data */
 172        boot_data = (struct boot_params *)real_mode_data;
 173        cmdline_paddr = boot_data->hdr.cmd_line_ptr | ((u64)boot_data->ext_cmd_line_ptr << 32);
 174
 175        if (!cmdline_paddr)
 176                return;
 177
 178        __sme_early_map_unmap_mem(__va(cmdline_paddr), COMMAND_LINE_SIZE, true);
 179}
 180
 181void __init sme_early_init(void)
 182{
 183        unsigned int i;
 184
 185        if (!sme_me_mask)
 186                return;
 187
 188        early_pmd_flags = __sme_set(early_pmd_flags);
 189
 190        __supported_pte_mask = __sme_set(__supported_pte_mask);
 191
 192        /* Update the protection map with memory encryption mask */
 193        for (i = 0; i < ARRAY_SIZE(protection_map); i++)
 194                protection_map[i] = pgprot_encrypted(protection_map[i]);
 195
 196        if (sev_active())
 197                swiotlb_force = SWIOTLB_FORCE;
 198}
 199
 200void __init sev_setup_arch(void)
 201{
 202        phys_addr_t total_mem = memblock_phys_mem_size();
 203        unsigned long size;
 204
 205        if (!sev_active())
 206                return;
 207
 208        /*
 209         * For SEV, all DMA has to occur via shared/unencrypted pages.
 210         * SEV uses SWIOTLB to make this happen without changing device
 211         * drivers. However, depending on the workload being run, the
 212         * default 64MB of SWIOTLB may not be enough and SWIOTLB may
 213         * run out of buffers for DMA, resulting in I/O errors and/or
 214         * performance degradation especially with high I/O workloads.
 215         *
 216         * Adjust the default size of SWIOTLB for SEV guests using
 217         * a percentage of guest memory for SWIOTLB buffers.
 218         * Also, as the SWIOTLB bounce buffer memory is allocated
 219         * from low memory, ensure that the adjusted size is within
 220         * the limits of low available memory.
 221         *
 222         * The percentage of guest memory used here for SWIOTLB buffers
 223         * is more of an approximation of the static adjustment which
 224         * 64MB for <1G, and ~128M to 256M for 1G-to-4G, i.e., the 6%
 225         */
 226        size = total_mem * 6 / 100;
 227        size = clamp_val(size, IO_TLB_DEFAULT_SIZE, SZ_1G);
 228        swiotlb_adjust_size(size);
 229}
 230
 231static void __init __set_clr_pte_enc(pte_t *kpte, int level, bool enc)
 232{
 233        pgprot_t old_prot, new_prot;
 234        unsigned long pfn, pa, size;
 235        pte_t new_pte;
 236
 237        switch (level) {
 238        case PG_LEVEL_4K:
 239                pfn = pte_pfn(*kpte);
 240                old_prot = pte_pgprot(*kpte);
 241                break;
 242        case PG_LEVEL_2M:
 243                pfn = pmd_pfn(*(pmd_t *)kpte);
 244                old_prot = pmd_pgprot(*(pmd_t *)kpte);
 245                break;
 246        case PG_LEVEL_1G:
 247                pfn = pud_pfn(*(pud_t *)kpte);
 248                old_prot = pud_pgprot(*(pud_t *)kpte);
 249                break;
 250        default:
 251                return;
 252        }
 253
 254        new_prot = old_prot;
 255        if (enc)
 256                pgprot_val(new_prot) |= _PAGE_ENC;
 257        else
 258                pgprot_val(new_prot) &= ~_PAGE_ENC;
 259
 260        /* If prot is same then do nothing. */
 261        if (pgprot_val(old_prot) == pgprot_val(new_prot))
 262                return;
 263
 264        pa = pfn << PAGE_SHIFT;
 265        size = page_level_size(level);
 266
 267        /*
 268         * We are going to perform in-place en-/decryption and change the
 269         * physical page attribute from C=1 to C=0 or vice versa. Flush the
 270         * caches to ensure that data gets accessed with the correct C-bit.
 271         */
 272        clflush_cache_range(__va(pa), size);
 273
 274        /* Encrypt/decrypt the contents in-place */
 275        if (enc)
 276                sme_early_encrypt(pa, size);
 277        else
 278                sme_early_decrypt(pa, size);
 279
 280        /* Change the page encryption mask. */
 281        new_pte = pfn_pte(pfn, new_prot);
 282        set_pte_atomic(kpte, new_pte);
 283}
 284
 285static int __init early_set_memory_enc_dec(unsigned long vaddr,
 286                                           unsigned long size, bool enc)
 287{
 288        unsigned long vaddr_end, vaddr_next;
 289        unsigned long psize, pmask;
 290        int split_page_size_mask;
 291        int level, ret;
 292        pte_t *kpte;
 293
 294        vaddr_next = vaddr;
 295        vaddr_end = vaddr + size;
 296
 297        for (; vaddr < vaddr_end; vaddr = vaddr_next) {
 298                kpte = lookup_address(vaddr, &level);
 299                if (!kpte || pte_none(*kpte)) {
 300                        ret = 1;
 301                        goto out;
 302                }
 303
 304                if (level == PG_LEVEL_4K) {
 305                        __set_clr_pte_enc(kpte, level, enc);
 306                        vaddr_next = (vaddr & PAGE_MASK) + PAGE_SIZE;
 307                        continue;
 308                }
 309
 310                psize = page_level_size(level);
 311                pmask = page_level_mask(level);
 312
 313                /*
 314                 * Check whether we can change the large page in one go.
 315                 * We request a split when the address is not aligned and
 316                 * the number of pages to set/clear encryption bit is smaller
 317                 * than the number of pages in the large page.
 318                 */
 319                if (vaddr == (vaddr & pmask) &&
 320                    ((vaddr_end - vaddr) >= psize)) {
 321                        __set_clr_pte_enc(kpte, level, enc);
 322                        vaddr_next = (vaddr & pmask) + psize;
 323                        continue;
 324                }
 325
 326                /*
 327                 * The virtual address is part of a larger page, create the next
 328                 * level page table mapping (4K or 2M). If it is part of a 2M
 329                 * page then we request a split of the large page into 4K
 330                 * chunks. A 1GB large page is split into 2M pages, resp.
 331                 */
 332                if (level == PG_LEVEL_2M)
 333                        split_page_size_mask = 0;
 334                else
 335                        split_page_size_mask = 1 << PG_LEVEL_2M;
 336
 337                /*
 338                 * kernel_physical_mapping_change() does not flush the TLBs, so
 339                 * a TLB flush is required after we exit from the for loop.
 340                 */
 341                kernel_physical_mapping_change(__pa(vaddr & pmask),
 342                                               __pa((vaddr_end & pmask) + psize),
 343                                               split_page_size_mask);
 344        }
 345
 346        ret = 0;
 347
 348out:
 349        __flush_tlb_all();
 350        return ret;
 351}
 352
 353int __init early_set_memory_decrypted(unsigned long vaddr, unsigned long size)
 354{
 355        return early_set_memory_enc_dec(vaddr, size, false);
 356}
 357
 358int __init early_set_memory_encrypted(unsigned long vaddr, unsigned long size)
 359{
 360        return early_set_memory_enc_dec(vaddr, size, true);
 361}
 362
 363/*
 364 * SME and SEV are very similar but they are not the same, so there are
 365 * times that the kernel will need to distinguish between SME and SEV. The
 366 * sme_active() and sev_active() functions are used for this.  When a
 367 * distinction isn't needed, the mem_encrypt_active() function can be used.
 368 *
 369 * The trampoline code is a good example for this requirement.  Before
 370 * paging is activated, SME will access all memory as decrypted, but SEV
 371 * will access all memory as encrypted.  So, when APs are being brought
 372 * up under SME the trampoline area cannot be encrypted, whereas under SEV
 373 * the trampoline area must be encrypted.
 374 */
 375bool sev_active(void)
 376{
 377        return sev_status & MSR_AMD64_SEV_ENABLED;
 378}
 379
 380bool sme_active(void)
 381{
 382        return sme_me_mask && !sev_active();
 383}
 384EXPORT_SYMBOL_GPL(sev_active);
 385
 386/* Needs to be called from non-instrumentable code */
 387bool noinstr sev_es_active(void)
 388{
 389        return sev_status & MSR_AMD64_SEV_ES_ENABLED;
 390}
 391
 392/* Override for DMA direct allocation check - ARCH_HAS_FORCE_DMA_UNENCRYPTED */
 393bool force_dma_unencrypted(struct device *dev)
 394{
 395        /*
 396         * For SEV, all DMA must be to unencrypted addresses.
 397         */
 398        if (sev_active())
 399                return true;
 400
 401        /*
 402         * For SME, all DMA must be to unencrypted addresses if the
 403         * device does not support DMA to addresses that include the
 404         * encryption mask.
 405         */
 406        if (sme_active()) {
 407                u64 dma_enc_mask = DMA_BIT_MASK(__ffs64(sme_me_mask));
 408                u64 dma_dev_mask = min_not_zero(dev->coherent_dma_mask,
 409                                                dev->bus_dma_limit);
 410
 411                if (dma_dev_mask <= dma_enc_mask)
 412                        return true;
 413        }
 414
 415        return false;
 416}
 417
 418void __init mem_encrypt_free_decrypted_mem(void)
 419{
 420        unsigned long vaddr, vaddr_end, npages;
 421        int r;
 422
 423        vaddr = (unsigned long)__start_bss_decrypted_unused;
 424        vaddr_end = (unsigned long)__end_bss_decrypted;
 425        npages = (vaddr_end - vaddr) >> PAGE_SHIFT;
 426
 427        /*
 428         * The unused memory range was mapped decrypted, change the encryption
 429         * attribute from decrypted to encrypted before freeing it.
 430         */
 431        if (mem_encrypt_active()) {
 432                r = set_memory_encrypted(vaddr, npages);
 433                if (r) {
 434                        pr_warn("failed to free unused decrypted pages\n");
 435                        return;
 436                }
 437        }
 438
 439        free_init_pages("unused decrypted", vaddr, vaddr_end);
 440}
 441
 442static void print_mem_encrypt_feature_info(void)
 443{
 444        pr_info("AMD Memory Encryption Features active:");
 445
 446        /* Secure Memory Encryption */
 447        if (sme_active()) {
 448                /*
 449                 * SME is mutually exclusive with any of the SEV
 450                 * features below.
 451                 */
 452                pr_cont(" SME\n");
 453                return;
 454        }
 455
 456        /* Secure Encrypted Virtualization */
 457        if (sev_active())
 458                pr_cont(" SEV");
 459
 460        /* Encrypted Register State */
 461        if (sev_es_active())
 462                pr_cont(" SEV-ES");
 463
 464        pr_cont("\n");
 465}
 466
 467/* Architecture __weak replacement functions */
 468void __init mem_encrypt_init(void)
 469{
 470        if (!sme_me_mask)
 471                return;
 472
 473        /* Call into SWIOTLB to update the SWIOTLB DMA buffers */
 474        swiotlb_update_mem_attributes();
 475
 476        /*
 477         * With SEV, we need to unroll the rep string I/O instructions,
 478         * but SEV-ES supports them through the #VC handler.
 479         */
 480        if (sev_active() && !sev_es_active())
 481                static_branch_enable(&sev_enable_key);
 482
 483        print_mem_encrypt_feature_info();
 484}
 485
 486int arch_has_restricted_virtio_memory_access(void)
 487{
 488        return sev_active();
 489}
 490EXPORT_SYMBOL_GPL(arch_has_restricted_virtio_memory_access);
 491
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.