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
__set_clr_pte_enc(pte_t * kpte,int level,bool enc)202 static void __init __set_clr_pte_enc(pte_t *kpte, int level, bool enc)
203 {
204 pgprot_t old_prot, new_prot;
205 unsigned long pfn, pa, size;
206 pte_t new_pte;
207
208 switch (level) {
209 case PG_LEVEL_4K:
210 pfn = pte_pfn(*kpte);
211 old_prot = pte_pgprot(*kpte);
212 break;
213 case PG_LEVEL_2M:
214 pfn = pmd_pfn(*(pmd_t *)kpte);
215 old_prot = pmd_pgprot(*(pmd_t *)kpte);
216 break;
217 case PG_LEVEL_1G:
218 pfn = pud_pfn(*(pud_t *)kpte);
219 old_prot = pud_pgprot(*(pud_t *)kpte);
220 break;
221 default:
222 return;
223 }
224
225 new_prot = old_prot;
226 if (enc)
227 pgprot_val(new_prot) |= _PAGE_ENC;
228 else
229 pgprot_val(new_prot) &= ~_PAGE_ENC;
230
231 /* If prot is same then do nothing. */
232 if (pgprot_val(old_prot) == pgprot_val(new_prot))
233 return;
234
235 pa = pfn << PAGE_SHIFT;
236 size = page_level_size(level);
237
238 /*
239 * We are going to perform in-place en-/decryption and change the
240 * physical page attribute from C=1 to C=0 or vice versa. Flush the
241 * caches to ensure that data gets accessed with the correct C-bit.
242 */
243 clflush_cache_range(__va(pa), size);
244
245 /* Encrypt/decrypt the contents in-place */
246 if (enc)
247 sme_early_encrypt(pa, size);
248 else
249 sme_early_decrypt(pa, size);
250
251 /* Change the page encryption mask. */
252 new_pte = pfn_pte(pfn, new_prot);
253 set_pte_atomic(kpte, new_pte);
254 }
255
early_set_memory_enc_dec(unsigned long vaddr,unsigned long size,bool enc)256 static int __init early_set_memory_enc_dec(unsigned long vaddr,
257 unsigned long size, bool enc)
258 {
259 unsigned long vaddr_end, vaddr_next;
260 unsigned long psize, pmask;
261 int split_page_size_mask;
262 int level, ret;
263 pte_t *kpte;
264
265 vaddr_next = vaddr;
266 vaddr_end = vaddr + size;
267
268 for (; vaddr < vaddr_end; vaddr = vaddr_next) {
269 kpte = lookup_address(vaddr, &level);
270 if (!kpte || pte_none(*kpte)) {
271 ret = 1;
272 goto out;
273 }
274
275 if (level == PG_LEVEL_4K) {
276 __set_clr_pte_enc(kpte, level, enc);
277 vaddr_next = (vaddr & PAGE_MASK) + PAGE_SIZE;
278 continue;
279 }
280
281 psize = page_level_size(level);
282 pmask = page_level_mask(level);
283
284 /*
285 * Check whether we can change the large page in one go.
286 * We request a split when the address is not aligned and
287 * the number of pages to set/clear encryption bit is smaller
288 * than the number of pages in the large page.
289 */
290 if (vaddr == (vaddr & pmask) &&
291 ((vaddr_end - vaddr) >= psize)) {
292 __set_clr_pte_enc(kpte, level, enc);
293 vaddr_next = (vaddr & pmask) + psize;
294 continue;
295 }
296
297 /*
298 * The virtual address is part of a larger page, create the next
299 * level page table mapping (4K or 2M). If it is part of a 2M
300 * page then we request a split of the large page into 4K
301 * chunks. A 1GB large page is split into 2M pages, resp.
302 */
303 if (level == PG_LEVEL_2M)
304 split_page_size_mask = 0;
305 else
306 split_page_size_mask = 1 << PG_LEVEL_2M;
307
308 /*
309 * kernel_physical_mapping_change() does not flush the TLBs, so
310 * a TLB flush is required after we exit from the for loop.
311 */
312 kernel_physical_mapping_change(__pa(vaddr & pmask),
313 __pa((vaddr_end & pmask) + psize),
314 split_page_size_mask);
315 }
316
317 ret = 0;
318
319 out:
320 __flush_tlb_all();
321 return ret;
322 }
323
early_set_memory_decrypted(unsigned long vaddr,unsigned long size)324 int __init early_set_memory_decrypted(unsigned long vaddr, unsigned long size)
325 {
326 return early_set_memory_enc_dec(vaddr, size, false);
327 }
328
early_set_memory_encrypted(unsigned long vaddr,unsigned long size)329 int __init early_set_memory_encrypted(unsigned long vaddr, unsigned long size)
330 {
331 return early_set_memory_enc_dec(vaddr, size, true);
332 }
333
334 /*
335 * SME and SEV are very similar but they are not the same, so there are
336 * times that the kernel will need to distinguish between SME and SEV. The
337 * sme_active() and sev_active() functions are used for this. When a
338 * distinction isn't needed, the mem_encrypt_active() function can be used.
339 *
340 * The trampoline code is a good example for this requirement. Before
341 * paging is activated, SME will access all memory as decrypted, but SEV
342 * will access all memory as encrypted. So, when APs are being brought
343 * up under SME the trampoline area cannot be encrypted, whereas under SEV
344 * the trampoline area must be encrypted.
345 */
sme_active(void)346 bool sme_active(void)
347 {
348 return sme_me_mask && !sev_enabled;
349 }
350
sev_active(void)351 bool sev_active(void)
352 {
353 return sev_status & MSR_AMD64_SEV_ENABLED;
354 }
355 EXPORT_SYMBOL_GPL(sev_active);
356
357 /* Needs to be called from non-instrumentable code */
sev_es_active(void)358 bool noinstr sev_es_active(void)
359 {
360 return sev_status & MSR_AMD64_SEV_ES_ENABLED;
361 }
362
363 /* Override for DMA direct allocation check - ARCH_HAS_FORCE_DMA_UNENCRYPTED */
force_dma_unencrypted(struct device * dev)364 bool force_dma_unencrypted(struct device *dev)
365 {
366 /*
367 * For SEV, all DMA must be to unencrypted addresses.
368 */
369 if (sev_active())
370 return true;
371
372 /*
373 * For SME, all DMA must be to unencrypted addresses if the
374 * device does not support DMA to addresses that include the
375 * encryption mask.
376 */
377 if (sme_active()) {
378 u64 dma_enc_mask = DMA_BIT_MASK(__ffs64(sme_me_mask));
379 u64 dma_dev_mask = min_not_zero(dev->coherent_dma_mask,
380 dev->bus_dma_limit);
381
382 if (dma_dev_mask <= dma_enc_mask)
383 return true;
384 }
385
386 return false;
387 }
388
mem_encrypt_free_decrypted_mem(void)389 void __init mem_encrypt_free_decrypted_mem(void)
390 {
391 unsigned long vaddr, vaddr_end, npages;
392 int r;
393
394 vaddr = (unsigned long)__start_bss_decrypted_unused;
395 vaddr_end = (unsigned long)__end_bss_decrypted;
396 npages = (vaddr_end - vaddr) >> PAGE_SHIFT;
397
398 /*
399 * The unused memory range was mapped decrypted, change the encryption
400 * attribute from decrypted to encrypted before freeing it.
401 */
402 if (mem_encrypt_active()) {
403 r = set_memory_encrypted(vaddr, npages);
404 if (r) {
405 pr_warn("failed to free unused decrypted pages\n");
406 return;
407 }
408 }
409
410 free_init_pages("unused decrypted", vaddr, vaddr_end);
411 }
412
print_mem_encrypt_feature_info(void)413 static void print_mem_encrypt_feature_info(void)
414 {
415 pr_info("AMD Memory Encryption Features active:");
416
417 /* Secure Memory Encryption */
418 if (sme_active()) {
419 /*
420 * SME is mutually exclusive with any of the SEV
421 * features below.
422 */
423 pr_cont(" SME\n");
424 return;
425 }
426
427 /* Secure Encrypted Virtualization */
428 if (sev_active())
429 pr_cont(" SEV");
430
431 /* Encrypted Register State */
432 if (sev_es_active())
433 pr_cont(" SEV-ES");
434
435 pr_cont("\n");
436 }
437
438 /* Architecture __weak replacement functions */
mem_encrypt_init(void)439 void __init mem_encrypt_init(void)
440 {
441 if (!sme_me_mask)
442 return;
443
444 /* Call into SWIOTLB to update the SWIOTLB DMA buffers */
445 swiotlb_update_mem_attributes();
446
447 /*
448 * With SEV, we need to unroll the rep string I/O instructions.
449 */
450 if (sev_active())
451 static_branch_enable(&sev_enable_key);
452
453 print_mem_encrypt_feature_info();
454 }
455
456