xref: /linux/arch/x86/boot/startup/sev-shared.c (revision 20e0d8576484c60c8c0c9d5d6665541c37dee327)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * AMD Encrypted Register State Support
4  *
5  * Author: Joerg Roedel <jroedel@suse.de>
6  *
7  * This file is not compiled stand-alone. It contains code shared
8  * between the pre-decompression boot code and the running Linux kernel
9  * and is included directly into both code-bases.
10  */
11 
12 #include <asm/setup_data.h>
13 
14 #ifndef __BOOT_COMPRESSED
15 #define error(v)			pr_err(v)
16 #define has_cpuflag(f)			boot_cpu_has(f)
17 #else
18 #undef WARN
19 #define WARN(condition, format...) (!!(condition))
20 #undef vc_forward_exception
21 #define vc_forward_exception(c)		panic("SNP: Hypervisor requested exception\n")
22 #endif
23 
24 /*
25  * SVSM related information:
26  *   During boot, the page tables are set up as identity mapped and later
27  *   changed to use kernel virtual addresses. Maintain separate virtual and
28  *   physical addresses for the CAA to allow SVSM functions to be used during
29  *   early boot, both with identity mapped virtual addresses and proper kernel
30  *   virtual addresses.
31  */
32 struct svsm_ca *boot_svsm_caa __ro_after_init;
33 u64 boot_svsm_caa_pa __ro_after_init;
34 
35 /*
36  * Since feature negotiation related variables are set early in the boot
37  * process they must reside in the .data section so as not to be zeroed
38  * out when the .bss section is later cleared.
39  *
40  * GHCB protocol version negotiated with the hypervisor.
41  */
42 static u16 ghcb_version __ro_after_init;
43 
44 /* Copy of the SNP firmware's CPUID page. */
45 static struct snp_cpuid_table cpuid_table_copy __ro_after_init;
46 
47 /*
48  * These will be initialized based on CPUID table so that non-present
49  * all-zero leaves (for sparse tables) can be differentiated from
50  * invalid/out-of-range leaves. This is needed since all-zero leaves
51  * still need to be post-processed.
52  */
53 static u32 cpuid_std_range_max __ro_after_init;
54 static u32 cpuid_hyp_range_max __ro_after_init;
55 static u32 cpuid_ext_range_max __ro_after_init;
56 
sev_es_check_cpu_features(void)57 bool __init sev_es_check_cpu_features(void)
58 {
59 	if (!has_cpuflag(X86_FEATURE_RDRAND)) {
60 		error("RDRAND instruction not supported - no trusted source of randomness available\n");
61 		return false;
62 	}
63 
64 	return true;
65 }
66 
67 void __head __noreturn
sev_es_terminate(unsigned int set,unsigned int reason)68 sev_es_terminate(unsigned int set, unsigned int reason)
69 {
70 	u64 val = GHCB_MSR_TERM_REQ;
71 
72 	/* Tell the hypervisor what went wrong. */
73 	val |= GHCB_SEV_TERM_REASON(set, reason);
74 
75 	/* Request Guest Termination from Hypervisor */
76 	sev_es_wr_ghcb_msr(val);
77 	VMGEXIT();
78 
79 	while (true)
80 		asm volatile("hlt\n" : : : "memory");
81 }
82 
83 /*
84  * The hypervisor features are available from GHCB version 2 onward.
85  */
get_hv_features(void)86 u64 get_hv_features(void)
87 {
88 	u64 val;
89 
90 	if (ghcb_version < 2)
91 		return 0;
92 
93 	sev_es_wr_ghcb_msr(GHCB_MSR_HV_FT_REQ);
94 	VMGEXIT();
95 
96 	val = sev_es_rd_ghcb_msr();
97 	if (GHCB_RESP_CODE(val) != GHCB_MSR_HV_FT_RESP)
98 		return 0;
99 
100 	return GHCB_MSR_HV_FT_RESP_VAL(val);
101 }
102 
snp_register_ghcb_early(unsigned long paddr)103 void snp_register_ghcb_early(unsigned long paddr)
104 {
105 	unsigned long pfn = paddr >> PAGE_SHIFT;
106 	u64 val;
107 
108 	sev_es_wr_ghcb_msr(GHCB_MSR_REG_GPA_REQ_VAL(pfn));
109 	VMGEXIT();
110 
111 	val = sev_es_rd_ghcb_msr();
112 
113 	/* If the response GPA is not ours then abort the guest */
114 	if ((GHCB_RESP_CODE(val) != GHCB_MSR_REG_GPA_RESP) ||
115 	    (GHCB_MSR_REG_GPA_RESP_VAL(val) != pfn))
116 		sev_es_terminate(SEV_TERM_SET_LINUX, GHCB_TERM_REGISTER);
117 }
118 
sev_es_negotiate_protocol(void)119 bool sev_es_negotiate_protocol(void)
120 {
121 	u64 val;
122 
123 	/* Do the GHCB protocol version negotiation */
124 	sev_es_wr_ghcb_msr(GHCB_MSR_SEV_INFO_REQ);
125 	VMGEXIT();
126 	val = sev_es_rd_ghcb_msr();
127 
128 	if (GHCB_MSR_INFO(val) != GHCB_MSR_SEV_INFO_RESP)
129 		return false;
130 
131 	if (GHCB_MSR_PROTO_MAX(val) < GHCB_PROTOCOL_MIN ||
132 	    GHCB_MSR_PROTO_MIN(val) > GHCB_PROTOCOL_MAX)
133 		return false;
134 
135 	ghcb_version = min_t(size_t, GHCB_MSR_PROTO_MAX(val), GHCB_PROTOCOL_MAX);
136 
137 	return true;
138 }
139 
verify_exception_info(struct ghcb * ghcb,struct es_em_ctxt * ctxt)140 static enum es_result verify_exception_info(struct ghcb *ghcb, struct es_em_ctxt *ctxt)
141 {
142 	u32 ret;
143 
144 	ret = ghcb->save.sw_exit_info_1 & GENMASK_ULL(31, 0);
145 	if (!ret)
146 		return ES_OK;
147 
148 	if (ret == 1) {
149 		u64 info = ghcb->save.sw_exit_info_2;
150 		unsigned long v = info & SVM_EVTINJ_VEC_MASK;
151 
152 		/* Check if exception information from hypervisor is sane. */
153 		if ((info & SVM_EVTINJ_VALID) &&
154 		    ((v == X86_TRAP_GP) || (v == X86_TRAP_UD)) &&
155 		    ((info & SVM_EVTINJ_TYPE_MASK) == SVM_EVTINJ_TYPE_EXEPT)) {
156 			ctxt->fi.vector = v;
157 
158 			if (info & SVM_EVTINJ_VALID_ERR)
159 				ctxt->fi.error_code = info >> 32;
160 
161 			return ES_EXCEPTION;
162 		}
163 	}
164 
165 	return ES_VMM_ERROR;
166 }
167 
svsm_process_result_codes(struct svsm_call * call)168 static inline int svsm_process_result_codes(struct svsm_call *call)
169 {
170 	switch (call->rax_out) {
171 	case SVSM_SUCCESS:
172 		return 0;
173 	case SVSM_ERR_INCOMPLETE:
174 	case SVSM_ERR_BUSY:
175 		return -EAGAIN;
176 	default:
177 		return -EINVAL;
178 	}
179 }
180 
181 /*
182  * Issue a VMGEXIT to call the SVSM:
183  *   - Load the SVSM register state (RAX, RCX, RDX, R8 and R9)
184  *   - Set the CA call pending field to 1
185  *   - Issue VMGEXIT
186  *   - Save the SVSM return register state (RAX, RCX, RDX, R8 and R9)
187  *   - Perform atomic exchange of the CA call pending field
188  *
189  *   - See the "Secure VM Service Module for SEV-SNP Guests" specification for
190  *     details on the calling convention.
191  *     - The calling convention loosely follows the Microsoft X64 calling
192  *       convention by putting arguments in RCX, RDX, R8 and R9.
193  *     - RAX specifies the SVSM protocol/callid as input and the return code
194  *       as output.
195  */
svsm_issue_call(struct svsm_call * call,u8 * pending)196 static __always_inline void svsm_issue_call(struct svsm_call *call, u8 *pending)
197 {
198 	register unsigned long rax asm("rax") = call->rax;
199 	register unsigned long rcx asm("rcx") = call->rcx;
200 	register unsigned long rdx asm("rdx") = call->rdx;
201 	register unsigned long r8  asm("r8")  = call->r8;
202 	register unsigned long r9  asm("r9")  = call->r9;
203 
204 	call->caa->call_pending = 1;
205 
206 	asm volatile("rep; vmmcall\n\t"
207 		     : "+r" (rax), "+r" (rcx), "+r" (rdx), "+r" (r8), "+r" (r9)
208 		     : : "memory");
209 
210 	*pending = xchg(&call->caa->call_pending, *pending);
211 
212 	call->rax_out = rax;
213 	call->rcx_out = rcx;
214 	call->rdx_out = rdx;
215 	call->r8_out  = r8;
216 	call->r9_out  = r9;
217 }
218 
svsm_perform_msr_protocol(struct svsm_call * call)219 static int svsm_perform_msr_protocol(struct svsm_call *call)
220 {
221 	u8 pending = 0;
222 	u64 val, resp;
223 
224 	/*
225 	 * When using the MSR protocol, be sure to save and restore
226 	 * the current MSR value.
227 	 */
228 	val = sev_es_rd_ghcb_msr();
229 
230 	sev_es_wr_ghcb_msr(GHCB_MSR_VMPL_REQ_LEVEL(0));
231 
232 	svsm_issue_call(call, &pending);
233 
234 	resp = sev_es_rd_ghcb_msr();
235 
236 	sev_es_wr_ghcb_msr(val);
237 
238 	if (pending)
239 		return -EINVAL;
240 
241 	if (GHCB_RESP_CODE(resp) != GHCB_MSR_VMPL_RESP)
242 		return -EINVAL;
243 
244 	if (GHCB_MSR_VMPL_RESP_VAL(resp))
245 		return -EINVAL;
246 
247 	return svsm_process_result_codes(call);
248 }
249 
svsm_perform_ghcb_protocol(struct ghcb * ghcb,struct svsm_call * call)250 static int svsm_perform_ghcb_protocol(struct ghcb *ghcb, struct svsm_call *call)
251 {
252 	struct es_em_ctxt ctxt;
253 	u8 pending = 0;
254 
255 	vc_ghcb_invalidate(ghcb);
256 
257 	/*
258 	 * Fill in protocol and format specifiers. This can be called very early
259 	 * in the boot, so use rip-relative references as needed.
260 	 */
261 	ghcb->protocol_version = ghcb_version;
262 	ghcb->ghcb_usage       = GHCB_DEFAULT_USAGE;
263 
264 	ghcb_set_sw_exit_code(ghcb, SVM_VMGEXIT_SNP_RUN_VMPL);
265 	ghcb_set_sw_exit_info_1(ghcb, 0);
266 	ghcb_set_sw_exit_info_2(ghcb, 0);
267 
268 	sev_es_wr_ghcb_msr(__pa(ghcb));
269 
270 	svsm_issue_call(call, &pending);
271 
272 	if (pending)
273 		return -EINVAL;
274 
275 	switch (verify_exception_info(ghcb, &ctxt)) {
276 	case ES_OK:
277 		break;
278 	case ES_EXCEPTION:
279 		vc_forward_exception(&ctxt);
280 		fallthrough;
281 	default:
282 		return -EINVAL;
283 	}
284 
285 	return svsm_process_result_codes(call);
286 }
287 
sev_es_ghcb_hv_call(struct ghcb * ghcb,struct es_em_ctxt * ctxt,u64 exit_code,u64 exit_info_1,u64 exit_info_2)288 enum es_result sev_es_ghcb_hv_call(struct ghcb *ghcb,
289 				   struct es_em_ctxt *ctxt,
290 				   u64 exit_code, u64 exit_info_1,
291 				   u64 exit_info_2)
292 {
293 	/* Fill in protocol and format specifiers */
294 	ghcb->protocol_version = ghcb_version;
295 	ghcb->ghcb_usage       = GHCB_DEFAULT_USAGE;
296 
297 	ghcb_set_sw_exit_code(ghcb, exit_code);
298 	ghcb_set_sw_exit_info_1(ghcb, exit_info_1);
299 	ghcb_set_sw_exit_info_2(ghcb, exit_info_2);
300 
301 	sev_es_wr_ghcb_msr(__pa(ghcb));
302 	VMGEXIT();
303 
304 	return verify_exception_info(ghcb, ctxt);
305 }
306 
__sev_cpuid_hv(u32 fn,int reg_idx,u32 * reg)307 static int __sev_cpuid_hv(u32 fn, int reg_idx, u32 *reg)
308 {
309 	u64 val;
310 
311 	sev_es_wr_ghcb_msr(GHCB_CPUID_REQ(fn, reg_idx));
312 	VMGEXIT();
313 	val = sev_es_rd_ghcb_msr();
314 	if (GHCB_RESP_CODE(val) != GHCB_MSR_CPUID_RESP)
315 		return -EIO;
316 
317 	*reg = (val >> 32);
318 
319 	return 0;
320 }
321 
__sev_cpuid_hv_msr(struct cpuid_leaf * leaf)322 static int __sev_cpuid_hv_msr(struct cpuid_leaf *leaf)
323 {
324 	int ret;
325 
326 	/*
327 	 * MSR protocol does not support fetching non-zero subfunctions, but is
328 	 * sufficient to handle current early-boot cases. Should that change,
329 	 * make sure to report an error rather than ignoring the index and
330 	 * grabbing random values. If this issue arises in the future, handling
331 	 * can be added here to use GHCB-page protocol for cases that occur late
332 	 * enough in boot that GHCB page is available.
333 	 */
334 	if (cpuid_function_is_indexed(leaf->fn) && leaf->subfn)
335 		return -EINVAL;
336 
337 	ret =         __sev_cpuid_hv(leaf->fn, GHCB_CPUID_REQ_EAX, &leaf->eax);
338 	ret = ret ? : __sev_cpuid_hv(leaf->fn, GHCB_CPUID_REQ_EBX, &leaf->ebx);
339 	ret = ret ? : __sev_cpuid_hv(leaf->fn, GHCB_CPUID_REQ_ECX, &leaf->ecx);
340 	ret = ret ? : __sev_cpuid_hv(leaf->fn, GHCB_CPUID_REQ_EDX, &leaf->edx);
341 
342 	return ret;
343 }
344 
__sev_cpuid_hv_ghcb(struct ghcb * ghcb,struct es_em_ctxt * ctxt,struct cpuid_leaf * leaf)345 static int __sev_cpuid_hv_ghcb(struct ghcb *ghcb, struct es_em_ctxt *ctxt, struct cpuid_leaf *leaf)
346 {
347 	u32 cr4 = native_read_cr4();
348 	int ret;
349 
350 	ghcb_set_rax(ghcb, leaf->fn);
351 	ghcb_set_rcx(ghcb, leaf->subfn);
352 
353 	if (cr4 & X86_CR4_OSXSAVE)
354 		/* Safe to read xcr0 */
355 		ghcb_set_xcr0(ghcb, xgetbv(XCR_XFEATURE_ENABLED_MASK));
356 	else
357 		/* xgetbv will cause #UD - use reset value for xcr0 */
358 		ghcb_set_xcr0(ghcb, 1);
359 
360 	ret = sev_es_ghcb_hv_call(ghcb, ctxt, SVM_EXIT_CPUID, 0, 0);
361 	if (ret != ES_OK)
362 		return ret;
363 
364 	if (!(ghcb_rax_is_valid(ghcb) &&
365 	      ghcb_rbx_is_valid(ghcb) &&
366 	      ghcb_rcx_is_valid(ghcb) &&
367 	      ghcb_rdx_is_valid(ghcb)))
368 		return ES_VMM_ERROR;
369 
370 	leaf->eax = ghcb->save.rax;
371 	leaf->ebx = ghcb->save.rbx;
372 	leaf->ecx = ghcb->save.rcx;
373 	leaf->edx = ghcb->save.rdx;
374 
375 	return ES_OK;
376 }
377 
sev_cpuid_hv(struct ghcb * ghcb,struct es_em_ctxt * ctxt,struct cpuid_leaf * leaf)378 static int sev_cpuid_hv(struct ghcb *ghcb, struct es_em_ctxt *ctxt, struct cpuid_leaf *leaf)
379 {
380 	return ghcb ? __sev_cpuid_hv_ghcb(ghcb, ctxt, leaf)
381 		    : __sev_cpuid_hv_msr(leaf);
382 }
383 
384 /*
385  * This may be called early while still running on the initial identity
386  * mapping. Use RIP-relative addressing to obtain the correct address
387  * while running with the initial identity mapping as well as the
388  * switch-over to kernel virtual addresses later.
389  */
snp_cpuid_get_table(void)390 const struct snp_cpuid_table *snp_cpuid_get_table(void)
391 {
392 	return rip_rel_ptr(&cpuid_table_copy);
393 }
394 
395 /*
396  * The SNP Firmware ABI, Revision 0.9, Section 7.1, details the use of
397  * XCR0_IN and XSS_IN to encode multiple versions of 0xD subfunctions 0
398  * and 1 based on the corresponding features enabled by a particular
399  * combination of XCR0 and XSS registers so that a guest can look up the
400  * version corresponding to the features currently enabled in its XCR0/XSS
401  * registers. The only values that differ between these versions/table
402  * entries is the enabled XSAVE area size advertised via EBX.
403  *
404  * While hypervisors may choose to make use of this support, it is more
405  * robust/secure for a guest to simply find the entry corresponding to the
406  * base/legacy XSAVE area size (XCR0=1 or XCR0=3), and then calculate the
407  * XSAVE area size using subfunctions 2 through 64, as documented in APM
408  * Volume 3, Rev 3.31, Appendix E.3.8, which is what is done here.
409  *
410  * Since base/legacy XSAVE area size is documented as 0x240, use that value
411  * directly rather than relying on the base size in the CPUID table.
412  *
413  * Return: XSAVE area size on success, 0 otherwise.
414  */
snp_cpuid_calc_xsave_size(u64 xfeatures_en,bool compacted)415 static u32 __head snp_cpuid_calc_xsave_size(u64 xfeatures_en, bool compacted)
416 {
417 	const struct snp_cpuid_table *cpuid_table = snp_cpuid_get_table();
418 	u64 xfeatures_found = 0;
419 	u32 xsave_size = 0x240;
420 	int i;
421 
422 	for (i = 0; i < cpuid_table->count; i++) {
423 		const struct snp_cpuid_fn *e = &cpuid_table->fn[i];
424 
425 		if (!(e->eax_in == 0xD && e->ecx_in > 1 && e->ecx_in < 64))
426 			continue;
427 		if (!(xfeatures_en & (BIT_ULL(e->ecx_in))))
428 			continue;
429 		if (xfeatures_found & (BIT_ULL(e->ecx_in)))
430 			continue;
431 
432 		xfeatures_found |= (BIT_ULL(e->ecx_in));
433 
434 		if (compacted)
435 			xsave_size += e->eax;
436 		else
437 			xsave_size = max(xsave_size, e->eax + e->ebx);
438 	}
439 
440 	/*
441 	 * Either the guest set unsupported XCR0/XSS bits, or the corresponding
442 	 * entries in the CPUID table were not present. This is not a valid
443 	 * state to be in.
444 	 */
445 	if (xfeatures_found != (xfeatures_en & GENMASK_ULL(63, 2)))
446 		return 0;
447 
448 	return xsave_size;
449 }
450 
451 static bool __head
snp_cpuid_get_validated_func(struct cpuid_leaf * leaf)452 snp_cpuid_get_validated_func(struct cpuid_leaf *leaf)
453 {
454 	const struct snp_cpuid_table *cpuid_table = snp_cpuid_get_table();
455 	int i;
456 
457 	for (i = 0; i < cpuid_table->count; i++) {
458 		const struct snp_cpuid_fn *e = &cpuid_table->fn[i];
459 
460 		if (e->eax_in != leaf->fn)
461 			continue;
462 
463 		if (cpuid_function_is_indexed(leaf->fn) && e->ecx_in != leaf->subfn)
464 			continue;
465 
466 		/*
467 		 * For 0xD subfunctions 0 and 1, only use the entry corresponding
468 		 * to the base/legacy XSAVE area size (XCR0=1 or XCR0=3, XSS=0).
469 		 * See the comments above snp_cpuid_calc_xsave_size() for more
470 		 * details.
471 		 */
472 		if (e->eax_in == 0xD && (e->ecx_in == 0 || e->ecx_in == 1))
473 			if (!(e->xcr0_in == 1 || e->xcr0_in == 3) || e->xss_in)
474 				continue;
475 
476 		leaf->eax = e->eax;
477 		leaf->ebx = e->ebx;
478 		leaf->ecx = e->ecx;
479 		leaf->edx = e->edx;
480 
481 		return true;
482 	}
483 
484 	return false;
485 }
486 
snp_cpuid_hv(struct ghcb * ghcb,struct es_em_ctxt * ctxt,struct cpuid_leaf * leaf)487 static void snp_cpuid_hv(struct ghcb *ghcb, struct es_em_ctxt *ctxt, struct cpuid_leaf *leaf)
488 {
489 	if (sev_cpuid_hv(ghcb, ctxt, leaf))
490 		sev_es_terminate(SEV_TERM_SET_LINUX, GHCB_TERM_CPUID_HV);
491 }
492 
493 static int __head
snp_cpuid_postprocess(struct ghcb * ghcb,struct es_em_ctxt * ctxt,struct cpuid_leaf * leaf)494 snp_cpuid_postprocess(struct ghcb *ghcb, struct es_em_ctxt *ctxt,
495 		      struct cpuid_leaf *leaf)
496 {
497 	struct cpuid_leaf leaf_hv = *leaf;
498 
499 	switch (leaf->fn) {
500 	case 0x1:
501 		snp_cpuid_hv(ghcb, ctxt, &leaf_hv);
502 
503 		/* initial APIC ID */
504 		leaf->ebx = (leaf_hv.ebx & GENMASK(31, 24)) | (leaf->ebx & GENMASK(23, 0));
505 		/* APIC enabled bit */
506 		leaf->edx = (leaf_hv.edx & BIT(9)) | (leaf->edx & ~BIT(9));
507 
508 		/* OSXSAVE enabled bit */
509 		if (native_read_cr4() & X86_CR4_OSXSAVE)
510 			leaf->ecx |= BIT(27);
511 		break;
512 	case 0x7:
513 		/* OSPKE enabled bit */
514 		leaf->ecx &= ~BIT(4);
515 		if (native_read_cr4() & X86_CR4_PKE)
516 			leaf->ecx |= BIT(4);
517 		break;
518 	case 0xB:
519 		leaf_hv.subfn = 0;
520 		snp_cpuid_hv(ghcb, ctxt, &leaf_hv);
521 
522 		/* extended APIC ID */
523 		leaf->edx = leaf_hv.edx;
524 		break;
525 	case 0xD: {
526 		bool compacted = false;
527 		u64 xcr0 = 1, xss = 0;
528 		u32 xsave_size;
529 
530 		if (leaf->subfn != 0 && leaf->subfn != 1)
531 			return 0;
532 
533 		if (native_read_cr4() & X86_CR4_OSXSAVE)
534 			xcr0 = xgetbv(XCR_XFEATURE_ENABLED_MASK);
535 		if (leaf->subfn == 1) {
536 			/* Get XSS value if XSAVES is enabled. */
537 			if (leaf->eax & BIT(3)) {
538 				unsigned long lo, hi;
539 
540 				asm volatile("rdmsr" : "=a" (lo), "=d" (hi)
541 						     : "c" (MSR_IA32_XSS));
542 				xss = (hi << 32) | lo;
543 			}
544 
545 			/*
546 			 * The PPR and APM aren't clear on what size should be
547 			 * encoded in 0xD:0x1:EBX when compaction is not enabled
548 			 * by either XSAVEC (feature bit 1) or XSAVES (feature
549 			 * bit 3) since SNP-capable hardware has these feature
550 			 * bits fixed as 1. KVM sets it to 0 in this case, but
551 			 * to avoid this becoming an issue it's safer to simply
552 			 * treat this as unsupported for SNP guests.
553 			 */
554 			if (!(leaf->eax & (BIT(1) | BIT(3))))
555 				return -EINVAL;
556 
557 			compacted = true;
558 		}
559 
560 		xsave_size = snp_cpuid_calc_xsave_size(xcr0 | xss, compacted);
561 		if (!xsave_size)
562 			return -EINVAL;
563 
564 		leaf->ebx = xsave_size;
565 		}
566 		break;
567 	case 0x8000001E:
568 		snp_cpuid_hv(ghcb, ctxt, &leaf_hv);
569 
570 		/* extended APIC ID */
571 		leaf->eax = leaf_hv.eax;
572 		/* compute ID */
573 		leaf->ebx = (leaf->ebx & GENMASK(31, 8)) | (leaf_hv.ebx & GENMASK(7, 0));
574 		/* node ID */
575 		leaf->ecx = (leaf->ecx & GENMASK(31, 8)) | (leaf_hv.ecx & GENMASK(7, 0));
576 		break;
577 	default:
578 		/* No fix-ups needed, use values as-is. */
579 		break;
580 	}
581 
582 	return 0;
583 }
584 
585 /*
586  * Returns -EOPNOTSUPP if feature not enabled. Any other non-zero return value
587  * should be treated as fatal by caller.
588  */
589 int __head
snp_cpuid(struct ghcb * ghcb,struct es_em_ctxt * ctxt,struct cpuid_leaf * leaf)590 snp_cpuid(struct ghcb *ghcb, struct es_em_ctxt *ctxt, struct cpuid_leaf *leaf)
591 {
592 	const struct snp_cpuid_table *cpuid_table = snp_cpuid_get_table();
593 
594 	if (!cpuid_table->count)
595 		return -EOPNOTSUPP;
596 
597 	if (!snp_cpuid_get_validated_func(leaf)) {
598 		/*
599 		 * Some hypervisors will avoid keeping track of CPUID entries
600 		 * where all values are zero, since they can be handled the
601 		 * same as out-of-range values (all-zero). This is useful here
602 		 * as well as it allows virtually all guest configurations to
603 		 * work using a single SNP CPUID table.
604 		 *
605 		 * To allow for this, there is a need to distinguish between
606 		 * out-of-range entries and in-range zero entries, since the
607 		 * CPUID table entries are only a template that may need to be
608 		 * augmented with additional values for things like
609 		 * CPU-specific information during post-processing. So if it's
610 		 * not in the table, set the values to zero. Then, if they are
611 		 * within a valid CPUID range, proceed with post-processing
612 		 * using zeros as the initial values. Otherwise, skip
613 		 * post-processing and just return zeros immediately.
614 		 */
615 		leaf->eax = leaf->ebx = leaf->ecx = leaf->edx = 0;
616 
617 		/* Skip post-processing for out-of-range zero leafs. */
618 		if (!(leaf->fn <= cpuid_std_range_max ||
619 		      (leaf->fn >= 0x40000000 && leaf->fn <= cpuid_hyp_range_max) ||
620 		      (leaf->fn >= 0x80000000 && leaf->fn <= cpuid_ext_range_max)))
621 			return 0;
622 	}
623 
624 	return snp_cpuid_postprocess(ghcb, ctxt, leaf);
625 }
626 
627 /*
628  * Boot VC Handler - This is the first VC handler during boot, there is no GHCB
629  * page yet, so it only supports the MSR based communication with the
630  * hypervisor and only the CPUID exit-code.
631  */
do_vc_no_ghcb(struct pt_regs * regs,unsigned long exit_code)632 void __head do_vc_no_ghcb(struct pt_regs *regs, unsigned long exit_code)
633 {
634 	unsigned int subfn = lower_bits(regs->cx, 32);
635 	unsigned int fn = lower_bits(regs->ax, 32);
636 	u16 opcode = *(unsigned short *)regs->ip;
637 	struct cpuid_leaf leaf;
638 	int ret;
639 
640 	/* Only CPUID is supported via MSR protocol */
641 	if (exit_code != SVM_EXIT_CPUID)
642 		goto fail;
643 
644 	/* Is it really a CPUID insn? */
645 	if (opcode != 0xa20f)
646 		goto fail;
647 
648 	leaf.fn = fn;
649 	leaf.subfn = subfn;
650 
651 	ret = snp_cpuid(NULL, NULL, &leaf);
652 	if (!ret)
653 		goto cpuid_done;
654 
655 	if (ret != -EOPNOTSUPP)
656 		goto fail;
657 
658 	if (__sev_cpuid_hv_msr(&leaf))
659 		goto fail;
660 
661 cpuid_done:
662 	regs->ax = leaf.eax;
663 	regs->bx = leaf.ebx;
664 	regs->cx = leaf.ecx;
665 	regs->dx = leaf.edx;
666 
667 	/*
668 	 * This is a VC handler and the #VC is only raised when SEV-ES is
669 	 * active, which means SEV must be active too. Do sanity checks on the
670 	 * CPUID results to make sure the hypervisor does not trick the kernel
671 	 * into the no-sev path. This could map sensitive data unencrypted and
672 	 * make it accessible to the hypervisor.
673 	 *
674 	 * In particular, check for:
675 	 *	- Availability of CPUID leaf 0x8000001f
676 	 *	- SEV CPUID bit.
677 	 *
678 	 * The hypervisor might still report the wrong C-bit position, but this
679 	 * can't be checked here.
680 	 */
681 
682 	if (fn == 0x80000000 && (regs->ax < 0x8000001f))
683 		/* SEV leaf check */
684 		goto fail;
685 	else if ((fn == 0x8000001f && !(regs->ax & BIT(1))))
686 		/* SEV bit */
687 		goto fail;
688 
689 	/* Skip over the CPUID two-byte opcode */
690 	regs->ip += 2;
691 
692 	return;
693 
694 fail:
695 	/* Terminate the guest */
696 	sev_es_terminate(SEV_TERM_SET_GEN, GHCB_SEV_ES_GEN_REQ);
697 }
698 
699 struct cc_setup_data {
700 	struct setup_data header;
701 	u32 cc_blob_address;
702 };
703 
704 /*
705  * Search for a Confidential Computing blob passed in as a setup_data entry
706  * via the Linux Boot Protocol.
707  */
708 static __head
find_cc_blob_setup_data(struct boot_params * bp)709 struct cc_blob_sev_info *find_cc_blob_setup_data(struct boot_params *bp)
710 {
711 	struct cc_setup_data *sd = NULL;
712 	struct setup_data *hdr;
713 
714 	hdr = (struct setup_data *)bp->hdr.setup_data;
715 
716 	while (hdr) {
717 		if (hdr->type == SETUP_CC_BLOB) {
718 			sd = (struct cc_setup_data *)hdr;
719 			return (struct cc_blob_sev_info *)(unsigned long)sd->cc_blob_address;
720 		}
721 		hdr = (struct setup_data *)hdr->next;
722 	}
723 
724 	return NULL;
725 }
726 
727 /*
728  * Initialize the kernel's copy of the SNP CPUID table, and set up the
729  * pointer that will be used to access it.
730  *
731  * Maintaining a direct mapping of the SNP CPUID table used by firmware would
732  * be possible as an alternative, but the approach is brittle since the
733  * mapping needs to be updated in sync with all the changes to virtual memory
734  * layout and related mapping facilities throughout the boot process.
735  */
setup_cpuid_table(const struct cc_blob_sev_info * cc_info)736 static void __head setup_cpuid_table(const struct cc_blob_sev_info *cc_info)
737 {
738 	const struct snp_cpuid_table *cpuid_table_fw, *cpuid_table;
739 	int i;
740 
741 	if (!cc_info || !cc_info->cpuid_phys || cc_info->cpuid_len < PAGE_SIZE)
742 		sev_es_terminate(SEV_TERM_SET_LINUX, GHCB_TERM_CPUID);
743 
744 	cpuid_table_fw = (const struct snp_cpuid_table *)cc_info->cpuid_phys;
745 	if (!cpuid_table_fw->count || cpuid_table_fw->count > SNP_CPUID_COUNT_MAX)
746 		sev_es_terminate(SEV_TERM_SET_LINUX, GHCB_TERM_CPUID);
747 
748 	cpuid_table = snp_cpuid_get_table();
749 	memcpy((void *)cpuid_table, cpuid_table_fw, sizeof(*cpuid_table));
750 
751 	/* Initialize CPUID ranges for range-checking. */
752 	for (i = 0; i < cpuid_table->count; i++) {
753 		const struct snp_cpuid_fn *fn = &cpuid_table->fn[i];
754 
755 		if (fn->eax_in == 0x0)
756 			cpuid_std_range_max = fn->eax;
757 		else if (fn->eax_in == 0x40000000)
758 			cpuid_hyp_range_max = fn->eax;
759 		else if (fn->eax_in == 0x80000000)
760 			cpuid_ext_range_max = fn->eax;
761 	}
762 }
763 
svsm_pval_4k_page(unsigned long paddr,bool validate)764 static void __head svsm_pval_4k_page(unsigned long paddr, bool validate)
765 {
766 	struct svsm_pvalidate_call *pc;
767 	struct svsm_call call = {};
768 	unsigned long flags;
769 	u64 pc_pa;
770 	int ret;
771 
772 	/*
773 	 * This can be called very early in the boot, use native functions in
774 	 * order to avoid paravirt issues.
775 	 */
776 	flags = native_local_irq_save();
777 
778 	call.caa = svsm_get_caa();
779 
780 	pc = (struct svsm_pvalidate_call *)call.caa->svsm_buffer;
781 	pc_pa = svsm_get_caa_pa() + offsetof(struct svsm_ca, svsm_buffer);
782 
783 	pc->num_entries = 1;
784 	pc->cur_index   = 0;
785 	pc->entry[0].page_size = RMP_PG_SIZE_4K;
786 	pc->entry[0].action    = validate;
787 	pc->entry[0].ignore_cf = 0;
788 	pc->entry[0].pfn       = paddr >> PAGE_SHIFT;
789 
790 	/* Protocol 0, Call ID 1 */
791 	call.rax = SVSM_CORE_CALL(SVSM_CORE_PVALIDATE);
792 	call.rcx = pc_pa;
793 
794 	ret = svsm_perform_call_protocol(&call);
795 	if (ret)
796 		sev_es_terminate(SEV_TERM_SET_LINUX, GHCB_TERM_PVALIDATE);
797 
798 	native_local_irq_restore(flags);
799 }
800 
pvalidate_4k_page(unsigned long vaddr,unsigned long paddr,bool validate)801 static void __head pvalidate_4k_page(unsigned long vaddr, unsigned long paddr,
802 				     bool validate)
803 {
804 	int ret;
805 
806 	if (snp_vmpl) {
807 		svsm_pval_4k_page(paddr, validate);
808 	} else {
809 		ret = pvalidate(vaddr, RMP_PG_SIZE_4K, validate);
810 		if (ret)
811 			sev_es_terminate(SEV_TERM_SET_LINUX, GHCB_TERM_PVALIDATE);
812 	}
813 
814 	/*
815 	 * If validating memory (making it private) and affected by the
816 	 * cache-coherency vulnerability, perform the cache eviction mitigation.
817 	 */
818 	if (validate && !has_cpuflag(X86_FEATURE_COHERENCY_SFW_NO))
819 		sev_evict_cache((void *)vaddr, 1);
820 }
821 
822 /*
823  * Maintain the GPA of the SVSM Calling Area (CA) in order to utilize the SVSM
824  * services needed when not running in VMPL0.
825  */
svsm_setup_ca(const struct cc_blob_sev_info * cc_info)826 static bool __head svsm_setup_ca(const struct cc_blob_sev_info *cc_info)
827 {
828 	struct snp_secrets_page *secrets_page;
829 	struct snp_cpuid_table *cpuid_table;
830 	unsigned int i;
831 	u64 caa;
832 
833 	BUILD_BUG_ON(sizeof(*secrets_page) != PAGE_SIZE);
834 
835 	/*
836 	 * Check if running at VMPL0.
837 	 *
838 	 * Use RMPADJUST (see the rmpadjust() function for a description of what
839 	 * the instruction does) to update the VMPL1 permissions of a page. If
840 	 * the guest is running at VMPL0, this will succeed and implies there is
841 	 * no SVSM. If the guest is running at any other VMPL, this will fail.
842 	 * Linux SNP guests only ever run at a single VMPL level so permission mask
843 	 * changes of a lesser-privileged VMPL are a don't-care.
844 	 *
845 	 * Use a rip-relative reference to obtain the proper address, since this
846 	 * routine is running identity mapped when called, both by the decompressor
847 	 * code and the early kernel code.
848 	 */
849 	if (!rmpadjust((unsigned long)rip_rel_ptr(&boot_ghcb_page), RMP_PG_SIZE_4K, 1))
850 		return false;
851 
852 	/*
853 	 * Not running at VMPL0, ensure everything has been properly supplied
854 	 * for running under an SVSM.
855 	 */
856 	if (!cc_info || !cc_info->secrets_phys || cc_info->secrets_len != PAGE_SIZE)
857 		sev_es_terminate(SEV_TERM_SET_LINUX, GHCB_TERM_SECRETS_PAGE);
858 
859 	secrets_page = (struct snp_secrets_page *)cc_info->secrets_phys;
860 	if (!secrets_page->svsm_size)
861 		sev_es_terminate(SEV_TERM_SET_LINUX, GHCB_TERM_NO_SVSM);
862 
863 	if (!secrets_page->svsm_guest_vmpl)
864 		sev_es_terminate(SEV_TERM_SET_LINUX, GHCB_TERM_SVSM_VMPL0);
865 
866 	snp_vmpl = secrets_page->svsm_guest_vmpl;
867 
868 	caa = secrets_page->svsm_caa;
869 
870 	/*
871 	 * An open-coded PAGE_ALIGNED() in order to avoid including
872 	 * kernel-proper headers into the decompressor.
873 	 */
874 	if (caa & (PAGE_SIZE - 1))
875 		sev_es_terminate(SEV_TERM_SET_LINUX, GHCB_TERM_SVSM_CAA);
876 
877 	/*
878 	 * The CA is identity mapped when this routine is called, both by the
879 	 * decompressor code and the early kernel code.
880 	 */
881 	boot_svsm_caa = (struct svsm_ca *)caa;
882 	boot_svsm_caa_pa = caa;
883 
884 	/* Advertise the SVSM presence via CPUID. */
885 	cpuid_table = (struct snp_cpuid_table *)snp_cpuid_get_table();
886 	for (i = 0; i < cpuid_table->count; i++) {
887 		struct snp_cpuid_fn *fn = &cpuid_table->fn[i];
888 
889 		if (fn->eax_in == 0x8000001f)
890 			fn->eax |= BIT(28);
891 	}
892 
893 	return true;
894 }
895