xref: /openbmc/linux/arch/x86/mm/mem_encrypt.c (revision ca48739e)
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 
23 #include <asm/tlbflush.h>
24 #include <asm/fixmap.h>
25 #include <asm/setup.h>
26 #include <asm/bootparam.h>
27 #include <asm/set_memory.h>
28 #include <asm/cacheflush.h>
29 #include <asm/processor-flags.h>
30 #include <asm/msr.h>
31 #include <asm/cmdline.h>
32 
33 #include "mm_internal.h"
34 
35 /*
36  * Since SME related variables are set early in the boot process they must
37  * reside in the .data section so as not to be zeroed out when the .bss
38  * section is later cleared.
39  */
40 u64 sme_me_mask __section(".data") = 0;
41 u64 sev_status __section(".data") = 0;
42 u64 sev_check_data __section(".data") = 0;
43 EXPORT_SYMBOL(sme_me_mask);
44 DEFINE_STATIC_KEY_FALSE(sev_enable_key);
45 EXPORT_SYMBOL_GPL(sev_enable_key);
46 
47 bool sev_enabled __section(".data");
48 
49 /* Buffer used for early in-place encryption by BSP, no locking needed */
50 static char sme_early_buffer[PAGE_SIZE] __initdata __aligned(PAGE_SIZE);
51 
52 /*
53  * This routine does not change the underlying encryption setting of the
54  * page(s) that map this memory. It assumes that eventually the memory is
55  * meant to be accessed as either encrypted or decrypted but the contents
56  * are currently not in the desired state.
57  *
58  * This routine follows the steps outlined in the AMD64 Architecture
59  * Programmer's Manual Volume 2, Section 7.10.8 Encrypt-in-Place.
60  */
61 static void __init __sme_early_enc_dec(resource_size_t paddr,
62 				       unsigned long size, bool enc)
63 {
64 	void *src, *dst;
65 	size_t len;
66 
67 	if (!sme_me_mask)
68 		return;
69 
70 	wbinvd();
71 
72 	/*
73 	 * There are limited number of early mapping slots, so map (at most)
74 	 * one page at time.
75 	 */
76 	while (size) {
77 		len = min_t(size_t, sizeof(sme_early_buffer), size);
78 
79 		/*
80 		 * Create mappings for the current and desired format of
81 		 * the memory. Use a write-protected mapping for the source.
82 		 */
83 		src = enc ? early_memremap_decrypted_wp(paddr, len) :
84 			    early_memremap_encrypted_wp(paddr, len);
85 
86 		dst = enc ? early_memremap_encrypted(paddr, len) :
87 			    early_memremap_decrypted(paddr, len);
88 
89 		/*
90 		 * If a mapping can't be obtained to perform the operation,
91 		 * then eventual access of that area in the desired mode
92 		 * will cause a crash.
93 		 */
94 		BUG_ON(!src || !dst);
95 
96 		/*
97 		 * Use a temporary buffer, of cache-line multiple size, to
98 		 * avoid data corruption as documented in the APM.
99 		 */
100 		memcpy(sme_early_buffer, src, len);
101 		memcpy(dst, sme_early_buffer, len);
102 
103 		early_memunmap(dst, len);
104 		early_memunmap(src, len);
105 
106 		paddr += len;
107 		size -= len;
108 	}
109 }
110 
111 void __init sme_early_encrypt(resource_size_t paddr, unsigned long size)
112 {
113 	__sme_early_enc_dec(paddr, size, true);
114 }
115 
116 void __init sme_early_decrypt(resource_size_t paddr, unsigned long size)
117 {
118 	__sme_early_enc_dec(paddr, size, false);
119 }
120 
121 static void __init __sme_early_map_unmap_mem(void *vaddr, unsigned long size,
122 					     bool map)
123 {
124 	unsigned long paddr = (unsigned long)vaddr - __PAGE_OFFSET;
125 	pmdval_t pmd_flags, pmd;
126 
127 	/* Use early_pmd_flags but remove the encryption mask */
128 	pmd_flags = __sme_clr(early_pmd_flags);
129 
130 	do {
131 		pmd = map ? (paddr & PMD_MASK) + pmd_flags : 0;
132 		__early_make_pgtable((unsigned long)vaddr, pmd);
133 
134 		vaddr += PMD_SIZE;
135 		paddr += PMD_SIZE;
136 		size = (size <= PMD_SIZE) ? 0 : size - PMD_SIZE;
137 	} while (size);
138 
139 	flush_tlb_local();
140 }
141 
142 void __init sme_unmap_bootdata(char *real_mode_data)
143 {
144 	struct boot_params *boot_data;
145 	unsigned long cmdline_paddr;
146 
147 	if (!sme_active())
148 		return;
149 
150 	/* Get the command line address before unmapping the real_mode_data */
151 	boot_data = (struct boot_params *)real_mode_data;
152 	cmdline_paddr = boot_data->hdr.cmd_line_ptr | ((u64)boot_data->ext_cmd_line_ptr << 32);
153 
154 	__sme_early_map_unmap_mem(real_mode_data, sizeof(boot_params), false);
155 
156 	if (!cmdline_paddr)
157 		return;
158 
159 	__sme_early_map_unmap_mem(__va(cmdline_paddr), COMMAND_LINE_SIZE, false);
160 }
161 
162 void __init sme_map_bootdata(char *real_mode_data)
163 {
164 	struct boot_params *boot_data;
165 	unsigned long cmdline_paddr;
166 
167 	if (!sme_active())
168 		return;
169 
170 	__sme_early_map_unmap_mem(real_mode_data, sizeof(boot_params), true);
171 
172 	/* Get the command line address after mapping the real_mode_data */
173 	boot_data = (struct boot_params *)real_mode_data;
174 	cmdline_paddr = boot_data->hdr.cmd_line_ptr | ((u64)boot_data->ext_cmd_line_ptr << 32);
175 
176 	if (!cmdline_paddr)
177 		return;
178 
179 	__sme_early_map_unmap_mem(__va(cmdline_paddr), COMMAND_LINE_SIZE, true);
180 }
181 
182 void __init sme_early_init(void)
183 {
184 	unsigned int i;
185 
186 	if (!sme_me_mask)
187 		return;
188 
189 	early_pmd_flags = __sme_set(early_pmd_flags);
190 
191 	__supported_pte_mask = __sme_set(__supported_pte_mask);
192 
193 	/* Update the protection map with memory encryption mask */
194 	for (i = 0; i < ARRAY_SIZE(protection_map); i++)
195 		protection_map[i] = pgprot_encrypted(protection_map[i]);
196 
197 	if (sev_active())
198 		swiotlb_force = SWIOTLB_FORCE;
199 }
200 
201 void __init sev_setup_arch(void)
202 {
203 	phys_addr_t total_mem = memblock_phys_mem_size();
204 	unsigned long size;
205 
206 	if (!sev_active())
207 		return;
208 
209 	/*
210 	 * For SEV, all DMA has to occur via shared/unencrypted pages.
211 	 * SEV uses SWIOTLB to make this happen without changing device
212 	 * drivers. However, depending on the workload being run, the
213 	 * default 64MB of SWIOTLB may not be enough and SWIOTLB may
214 	 * run out of buffers for DMA, resulting in I/O errors and/or
215 	 * performance degradation especially with high I/O workloads.
216 	 *
217 	 * Adjust the default size of SWIOTLB for SEV guests using
218 	 * a percentage of guest memory for SWIOTLB buffers.
219 	 * Also, as the SWIOTLB bounce buffer memory is allocated
220 	 * from low memory, ensure that the adjusted size is within
221 	 * the limits of low available memory.
222 	 *
223 	 * The percentage of guest memory used here for SWIOTLB buffers
224 	 * is more of an approximation of the static adjustment which
225 	 * 64MB for <1G, and ~128M to 256M for 1G-to-4G, i.e., the 6%
226 	 */
227 	size = total_mem * 6 / 100;
228 	size = clamp_val(size, IO_TLB_DEFAULT_SIZE, SZ_1G);
229 	swiotlb_adjust_size(size);
230 }
231 
232 static void __init __set_clr_pte_enc(pte_t *kpte, int level, bool enc)
233 {
234 	pgprot_t old_prot, new_prot;
235 	unsigned long pfn, pa, size;
236 	pte_t new_pte;
237 
238 	switch (level) {
239 	case PG_LEVEL_4K:
240 		pfn = pte_pfn(*kpte);
241 		old_prot = pte_pgprot(*kpte);
242 		break;
243 	case PG_LEVEL_2M:
244 		pfn = pmd_pfn(*(pmd_t *)kpte);
245 		old_prot = pmd_pgprot(*(pmd_t *)kpte);
246 		break;
247 	case PG_LEVEL_1G:
248 		pfn = pud_pfn(*(pud_t *)kpte);
249 		old_prot = pud_pgprot(*(pud_t *)kpte);
250 		break;
251 	default:
252 		return;
253 	}
254 
255 	new_prot = old_prot;
256 	if (enc)
257 		pgprot_val(new_prot) |= _PAGE_ENC;
258 	else
259 		pgprot_val(new_prot) &= ~_PAGE_ENC;
260 
261 	/* If prot is same then do nothing. */
262 	if (pgprot_val(old_prot) == pgprot_val(new_prot))
263 		return;
264 
265 	pa = pfn << page_level_shift(level);
266 	size = page_level_size(level);
267 
268 	/*
269 	 * We are going to perform in-place en-/decryption and change the
270 	 * physical page attribute from C=1 to C=0 or vice versa. Flush the
271 	 * caches to ensure that data gets accessed with the correct C-bit.
272 	 */
273 	clflush_cache_range(__va(pa), size);
274 
275 	/* Encrypt/decrypt the contents in-place */
276 	if (enc)
277 		sme_early_encrypt(pa, size);
278 	else
279 		sme_early_decrypt(pa, size);
280 
281 	/* Change the page encryption mask. */
282 	new_pte = pfn_pte(pfn, new_prot);
283 	set_pte_atomic(kpte, new_pte);
284 }
285 
286 static int __init early_set_memory_enc_dec(unsigned long vaddr,
287 					   unsigned long size, bool enc)
288 {
289 	unsigned long vaddr_end, vaddr_next;
290 	unsigned long psize, pmask;
291 	int split_page_size_mask;
292 	int level, ret;
293 	pte_t *kpte;
294 
295 	vaddr_next = vaddr;
296 	vaddr_end = vaddr + size;
297 
298 	for (; vaddr < vaddr_end; vaddr = vaddr_next) {
299 		kpte = lookup_address(vaddr, &level);
300 		if (!kpte || pte_none(*kpte)) {
301 			ret = 1;
302 			goto out;
303 		}
304 
305 		if (level == PG_LEVEL_4K) {
306 			__set_clr_pte_enc(kpte, level, enc);
307 			vaddr_next = (vaddr & PAGE_MASK) + PAGE_SIZE;
308 			continue;
309 		}
310 
311 		psize = page_level_size(level);
312 		pmask = page_level_mask(level);
313 
314 		/*
315 		 * Check whether we can change the large page in one go.
316 		 * We request a split when the address is not aligned and
317 		 * the number of pages to set/clear encryption bit is smaller
318 		 * than the number of pages in the large page.
319 		 */
320 		if (vaddr == (vaddr & pmask) &&
321 		    ((vaddr_end - vaddr) >= psize)) {
322 			__set_clr_pte_enc(kpte, level, enc);
323 			vaddr_next = (vaddr & pmask) + psize;
324 			continue;
325 		}
326 
327 		/*
328 		 * The virtual address is part of a larger page, create the next
329 		 * level page table mapping (4K or 2M). If it is part of a 2M
330 		 * page then we request a split of the large page into 4K
331 		 * chunks. A 1GB large page is split into 2M pages, resp.
332 		 */
333 		if (level == PG_LEVEL_2M)
334 			split_page_size_mask = 0;
335 		else
336 			split_page_size_mask = 1 << PG_LEVEL_2M;
337 
338 		/*
339 		 * kernel_physical_mapping_change() does not flush the TLBs, so
340 		 * a TLB flush is required after we exit from the for loop.
341 		 */
342 		kernel_physical_mapping_change(__pa(vaddr & pmask),
343 					       __pa((vaddr_end & pmask) + psize),
344 					       split_page_size_mask);
345 	}
346 
347 	ret = 0;
348 
349 out:
350 	__flush_tlb_all();
351 	return ret;
352 }
353 
354 int __init early_set_memory_decrypted(unsigned long vaddr, unsigned long size)
355 {
356 	return early_set_memory_enc_dec(vaddr, size, false);
357 }
358 
359 int __init early_set_memory_encrypted(unsigned long vaddr, unsigned long size)
360 {
361 	return early_set_memory_enc_dec(vaddr, size, true);
362 }
363 
364 /*
365  * SME and SEV are very similar but they are not the same, so there are
366  * times that the kernel will need to distinguish between SME and SEV. The
367  * sme_active() and sev_active() functions are used for this.  When a
368  * distinction isn't needed, the mem_encrypt_active() function can be used.
369  *
370  * The trampoline code is a good example for this requirement.  Before
371  * paging is activated, SME will access all memory as decrypted, but SEV
372  * will access all memory as encrypted.  So, when APs are being brought
373  * up under SME the trampoline area cannot be encrypted, whereas under SEV
374  * the trampoline area must be encrypted.
375  */
376 bool sme_active(void)
377 {
378 	return sme_me_mask && !sev_enabled;
379 }
380 
381 bool sev_active(void)
382 {
383 	return sev_status & MSR_AMD64_SEV_ENABLED;
384 }
385 EXPORT_SYMBOL_GPL(sev_active);
386 
387 /* Needs to be called from non-instrumentable code */
388 bool noinstr sev_es_active(void)
389 {
390 	return sev_status & MSR_AMD64_SEV_ES_ENABLED;
391 }
392 
393 /* Override for DMA direct allocation check - ARCH_HAS_FORCE_DMA_UNENCRYPTED */
394 bool force_dma_unencrypted(struct device *dev)
395 {
396 	/*
397 	 * For SEV, all DMA must be to unencrypted addresses.
398 	 */
399 	if (sev_active())
400 		return true;
401 
402 	/*
403 	 * For SME, all DMA must be to unencrypted addresses if the
404 	 * device does not support DMA to addresses that include the
405 	 * encryption mask.
406 	 */
407 	if (sme_active()) {
408 		u64 dma_enc_mask = DMA_BIT_MASK(__ffs64(sme_me_mask));
409 		u64 dma_dev_mask = min_not_zero(dev->coherent_dma_mask,
410 						dev->bus_dma_limit);
411 
412 		if (dma_dev_mask <= dma_enc_mask)
413 			return true;
414 	}
415 
416 	return false;
417 }
418 
419 void __init mem_encrypt_free_decrypted_mem(void)
420 {
421 	unsigned long vaddr, vaddr_end, npages;
422 	int r;
423 
424 	vaddr = (unsigned long)__start_bss_decrypted_unused;
425 	vaddr_end = (unsigned long)__end_bss_decrypted;
426 	npages = (vaddr_end - vaddr) >> PAGE_SHIFT;
427 
428 	/*
429 	 * The unused memory range was mapped decrypted, change the encryption
430 	 * attribute from decrypted to encrypted before freeing it.
431 	 */
432 	if (mem_encrypt_active()) {
433 		r = set_memory_encrypted(vaddr, npages);
434 		if (r) {
435 			pr_warn("failed to free unused decrypted pages\n");
436 			return;
437 		}
438 	}
439 
440 	free_init_pages("unused decrypted", vaddr, vaddr_end);
441 }
442 
443 static void print_mem_encrypt_feature_info(void)
444 {
445 	pr_info("AMD Memory Encryption Features active:");
446 
447 	/* Secure Memory Encryption */
448 	if (sme_active()) {
449 		/*
450 		 * SME is mutually exclusive with any of the SEV
451 		 * features below.
452 		 */
453 		pr_cont(" SME\n");
454 		return;
455 	}
456 
457 	/* Secure Encrypted Virtualization */
458 	if (sev_active())
459 		pr_cont(" SEV");
460 
461 	/* Encrypted Register State */
462 	if (sev_es_active())
463 		pr_cont(" SEV-ES");
464 
465 	pr_cont("\n");
466 }
467 
468 /* Architecture __weak replacement functions */
469 void __init mem_encrypt_init(void)
470 {
471 	if (!sme_me_mask)
472 		return;
473 
474 	/* Call into SWIOTLB to update the SWIOTLB DMA buffers */
475 	swiotlb_update_mem_attributes();
476 
477 	/*
478 	 * With SEV, we need to unroll the rep string I/O instructions,
479 	 * but SEV-ES supports them through the #VC handler.
480 	 */
481 	if (sev_active() && !sev_es_active())
482 		static_branch_enable(&sev_enable_key);
483 
484 	print_mem_encrypt_feature_info();
485 }
486 
487