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