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