xref: /linux/arch/x86/kernel/cpu/amd.c (revision 59e6295fac26b8e85c1ea859cdd89fa1e47519d7)
1 // SPDX-License-Identifier: GPL-2.0-only
2 #include <linux/export.h>
3 #include <linux/bitops.h>
4 #include <linux/dmi.h>
5 #include <linux/elf.h>
6 #include <linux/mm.h>
7 #include <linux/kvm_types.h>
8 #include <linux/io.h>
9 #include <linux/sched.h>
10 #include <linux/sched/clock.h>
11 #include <linux/random.h>
12 #include <linux/topology.h>
13 #include <linux/platform_data/x86/amd-fch.h>
14 #include <asm/processor.h>
15 #include <asm/apic.h>
16 #include <asm/cacheinfo.h>
17 #include <asm/cpu.h>
18 #include <asm/cpu_device_id.h>
19 #include <asm/cpuid/api.h>
20 #include <asm/spec-ctrl.h>
21 #include <asm/smp.h>
22 #include <asm/numa.h>
23 #include <asm/pci-direct.h>
24 #include <asm/delay.h>
25 #include <asm/debugreg.h>
26 #include <asm/resctrl.h>
27 #include <asm/msr.h>
28 #include <asm/sev.h>
29 
30 #ifdef CONFIG_X86_64
31 # include <asm/mmconfig.h>
32 #endif
33 
34 #include "cpu.h"
35 
36 u16 invlpgb_count_max __ro_after_init = 1;
37 
38 static inline int rdmsrq_amd_safe(unsigned msr, u64 *p)
39 {
40 	u32 gprs[8] = { 0 };
41 	int err;
42 
43 	WARN_ONCE((boot_cpu_data.x86 != 0xf),
44 		  "%s should only be used on K8!\n", __func__);
45 
46 	gprs[1] = msr;
47 	gprs[7] = 0x9c5a203a;
48 
49 	err = rdmsr_safe_regs(gprs);
50 
51 	*p = gprs[0] | ((u64)gprs[2] << 32);
52 
53 	return err;
54 }
55 
56 static inline int wrmsrq_amd_safe(unsigned msr, u64 val)
57 {
58 	u32 gprs[8] = { 0 };
59 
60 	WARN_ONCE((boot_cpu_data.x86 != 0xf),
61 		  "%s should only be used on K8!\n", __func__);
62 
63 	gprs[0] = (u32)val;
64 	gprs[1] = msr;
65 	gprs[2] = val >> 32;
66 	gprs[7] = 0x9c5a203a;
67 
68 	return wrmsr_safe_regs(gprs);
69 }
70 
71 /*
72  *	B step AMD K6 before B 9730xxxx have hardware bugs that can cause
73  *	misexecution of code under Linux. Owners of such processors should
74  *	contact AMD for precise details and a CPU swap.
75  *
76  *	See	http://www.multimania.com/poulot/k6bug.html
77  *	and	section 2.6.2 of "AMD-K6 Processor Revision Guide - Model 6"
78  *		(Publication # 21266  Issue Date: August 1998)
79  *
80  *	The following test is erm.. interesting. AMD neglected to up
81  *	the chip setting when fixing the bug but they also tweaked some
82  *	performance at the same time..
83  */
84 
85 #ifdef CONFIG_X86_32
86 extern __visible void vide(void);
87 __asm__(".text\n"
88 	".globl vide\n"
89 	".type vide, @function\n"
90 	".align 4\n"
91 	"vide: ret\n");
92 #endif
93 
94 static void init_amd_k5(struct cpuinfo_x86 *c)
95 {
96 #ifdef CONFIG_X86_32
97 /*
98  * General Systems BIOSen alias the cpu frequency registers
99  * of the Elan at 0x000df000. Unfortunately, one of the Linux
100  * drivers subsequently pokes it, and changes the CPU speed.
101  * Workaround : Remove the unneeded alias.
102  */
103 #define CBAR		(0xfffc) /* Configuration Base Address  (32-bit) */
104 #define CBAR_ENB	(0x80000000)
105 #define CBAR_KEY	(0X000000CB)
106 	if (c->x86_model == 9 || c->x86_model == 10) {
107 		if (inl(CBAR) & CBAR_ENB)
108 			outl(0 | CBAR_KEY, CBAR);
109 	}
110 #endif
111 }
112 
113 static void init_amd_k6(struct cpuinfo_x86 *c)
114 {
115 #ifdef CONFIG_X86_32
116 	struct msr val;
117 	int mbytes = get_num_physpages() >> (20-PAGE_SHIFT);
118 
119 	if (c->x86_model < 6) {
120 		/* Based on AMD doc 20734R - June 2000 */
121 		if (c->x86_model == 0) {
122 			clear_cpu_cap(c, X86_FEATURE_APIC);
123 			set_cpu_cap(c, X86_FEATURE_PGE);
124 		}
125 		return;
126 	}
127 
128 	if (c->x86_model == 6 && c->x86_stepping == 1) {
129 		const int K6_BUG_LOOP = 1000000;
130 		int n;
131 		void (*f_vide)(void);
132 		u64 d, d2;
133 
134 		pr_info("AMD K6 stepping B detected - ");
135 
136 		/*
137 		 * It looks like AMD fixed the 2.6.2 bug and improved indirect
138 		 * calls at the same time.
139 		 */
140 
141 		n = K6_BUG_LOOP;
142 		f_vide = vide;
143 		OPTIMIZER_HIDE_VAR(f_vide);
144 		d = rdtsc();
145 		while (n--)
146 			f_vide();
147 		d2 = rdtsc();
148 		d = d2-d;
149 
150 		if (d > 20*K6_BUG_LOOP)
151 			pr_cont("system stability may be impaired when more than 32 MB are used.\n");
152 		else
153 			pr_cont("probably OK (after B9730xxxx).\n");
154 	}
155 
156 	/* K6 with old style WHCR */
157 	if (c->x86_model < 8 ||
158 	   (c->x86_model == 8 && c->x86_stepping < 8)) {
159 		/* We can only write allocate on the low 508Mb */
160 		if (mbytes > 508)
161 			mbytes = 508;
162 
163 		rdmsrq(MSR_K6_WHCR, val.q);
164 		if ((val.l & 0x0000FFFF) == 0) {
165 			unsigned long flags;
166 			val.l = (1 << 0) | ((mbytes / 4) << 1);
167 			local_irq_save(flags);
168 			wbinvd();
169 			wrmsrq(MSR_K6_WHCR, val.q);
170 			local_irq_restore(flags);
171 			pr_info("Enabling old style K6 write allocation for %d Mb\n",
172 				mbytes);
173 		}
174 		return;
175 	}
176 
177 	if ((c->x86_model == 8 && c->x86_stepping > 7) ||
178 	     c->x86_model == 9 || c->x86_model == 13) {
179 		/* The more serious chips .. */
180 
181 		if (mbytes > 4092)
182 			mbytes = 4092;
183 
184 		rdmsrq(MSR_K6_WHCR, val.q);
185 		if ((val.l & 0xFFFF0000) == 0) {
186 			unsigned long flags;
187 			val.l = ((mbytes >> 2) << 22) | (1 << 16);
188 			local_irq_save(flags);
189 			wbinvd();
190 			wrmsrq(MSR_K6_WHCR, val.q);
191 			local_irq_restore(flags);
192 			pr_info("Enabling new style K6 write allocation for %d Mb\n",
193 				mbytes);
194 		}
195 
196 		return;
197 	}
198 
199 	if (c->x86_model == 10) {
200 		/* AMD Geode LX is model 10 */
201 		/* placeholder for any needed mods */
202 		return;
203 	}
204 #endif
205 }
206 
207 static void init_amd_k7(struct cpuinfo_x86 *c)
208 {
209 #ifdef CONFIG_X86_32
210 	struct msr val;
211 
212 	/*
213 	 * Bit 15 of Athlon specific MSR 15, needs to be 0
214 	 * to enable SSE on Palomino/Morgan/Barton CPU's.
215 	 * If the BIOS didn't enable it already, enable it here.
216 	 */
217 	if (c->x86_model >= 6 && c->x86_model <= 10) {
218 		if (!cpu_has(c, X86_FEATURE_XMM)) {
219 			pr_info("Enabling disabled K7/SSE Support.\n");
220 			msr_clear_bit(MSR_K7_HWCR, 15);
221 			set_cpu_cap(c, X86_FEATURE_XMM);
222 		}
223 	}
224 
225 	/*
226 	 * It's been determined by AMD that Athlons since model 8 stepping 1
227 	 * are more robust with CLK_CTL set to 200xxxxx instead of 600xxxxx
228 	 * As per AMD technical note 27212 0.2
229 	 */
230 	if ((c->x86_model == 8 && c->x86_stepping >= 1) || (c->x86_model > 8)) {
231 		rdmsrq(MSR_K7_CLK_CTL, val.q);
232 		if ((val.l & 0xfff00000) != 0x20000000) {
233 			pr_info("CPU: CLK_CTL MSR was %x. Reprogramming to %x\n",
234 				val.l, ((val.l & 0x000fffff) | 0x20000000));
235 			val.l = (val.l & 0x000fffff) | 0x20000000;
236 			wrmsrq(MSR_K7_CLK_CTL, val.q);
237 		}
238 	}
239 
240 	/* calling is from identify_secondary_cpu() ? */
241 	if (!c->cpu_index)
242 		return;
243 
244 	/*
245 	 * Certain Athlons might work (for various values of 'work') in SMP
246 	 * but they are not certified as MP capable.
247 	 */
248 	/* Athlon 660/661 is valid. */
249 	if ((c->x86_model == 6) && ((c->x86_stepping == 0) ||
250 	    (c->x86_stepping == 1)))
251 		return;
252 
253 	/* Duron 670 is valid */
254 	if ((c->x86_model == 7) && (c->x86_stepping == 0))
255 		return;
256 
257 	/*
258 	 * Athlon 662, Duron 671, and Athlon >model 7 have capability
259 	 * bit. It's worth noting that the A5 stepping (662) of some
260 	 * Athlon XP's have the MP bit set.
261 	 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for
262 	 * more.
263 	 */
264 	if (((c->x86_model == 6) && (c->x86_stepping >= 2)) ||
265 	    ((c->x86_model == 7) && (c->x86_stepping >= 1)) ||
266 	     (c->x86_model > 7))
267 		if (cpu_has(c, X86_FEATURE_MP))
268 			return;
269 
270 	/* If we get here, not a certified SMP capable AMD system. */
271 
272 	/*
273 	 * Don't taint if we are running SMP kernel on a single non-MP
274 	 * approved Athlon
275 	 */
276 	WARN_ONCE(1, "WARNING: This combination of AMD"
277 		" processors is not suitable for SMP.\n");
278 	add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_NOW_UNRELIABLE);
279 #endif
280 }
281 
282 #ifdef CONFIG_NUMA
283 /*
284  * To workaround broken NUMA config.  Read the comment in
285  * srat_detect_node().
286  */
287 static int nearby_node(int apicid)
288 {
289 	int i, node;
290 
291 	for (i = apicid - 1; i >= 0; i--) {
292 		node = __apicid_to_node[i];
293 		if (node != NUMA_NO_NODE && node_online(node))
294 			return node;
295 	}
296 	for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) {
297 		node = __apicid_to_node[i];
298 		if (node != NUMA_NO_NODE && node_online(node))
299 			return node;
300 	}
301 	return first_node(node_online_map); /* Shouldn't happen */
302 }
303 #endif
304 
305 static void srat_detect_node(struct cpuinfo_x86 *c)
306 {
307 #ifdef CONFIG_NUMA
308 	int cpu = smp_processor_id();
309 	int node;
310 	unsigned apicid = c->topo.apicid;
311 
312 	node = numa_cpu_node(cpu);
313 	if (node == NUMA_NO_NODE)
314 		node = per_cpu_llc_id(cpu);
315 
316 	/*
317 	 * On multi-fabric platform (e.g. Numascale NumaChip) a
318 	 * platform-specific handler needs to be called to fixup some
319 	 * IDs of the CPU.
320 	 */
321 	if (x86_cpuinit.fixup_cpu_id)
322 		x86_cpuinit.fixup_cpu_id(c, node);
323 
324 	if (!node_online(node)) {
325 		/*
326 		 * Two possibilities here:
327 		 *
328 		 * - The CPU is missing memory and no node was created.  In
329 		 *   that case try picking one from a nearby CPU.
330 		 *
331 		 * - The APIC IDs differ from the HyperTransport node IDs
332 		 *   which the K8 northbridge parsing fills in.  Assume
333 		 *   they are all increased by a constant offset, but in
334 		 *   the same order as the HT nodeids.  If that doesn't
335 		 *   result in a usable node fall back to the path for the
336 		 *   previous case.
337 		 *
338 		 * This workaround operates directly on the mapping between
339 		 * APIC ID and NUMA node, assuming certain relationship
340 		 * between APIC ID, HT node ID and NUMA topology.  As going
341 		 * through CPU mapping may alter the outcome, directly
342 		 * access __apicid_to_node[].
343 		 */
344 		int ht_nodeid = c->topo.initial_apicid;
345 
346 		if (__apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
347 			node = __apicid_to_node[ht_nodeid];
348 		/* Pick a nearby node */
349 		if (!node_online(node))
350 			node = nearby_node(apicid);
351 	}
352 	numa_set_node(cpu, node);
353 #endif
354 }
355 
356 static void bsp_determine_snp(struct cpuinfo_x86 *c)
357 {
358 #ifdef CONFIG_ARCH_HAS_CC_PLATFORM
359 	cc_vendor = CC_VENDOR_AMD;
360 
361 	if (cpu_has(c, X86_FEATURE_SEV_SNP)) {
362 		/*
363 		 * RMP table entry format is not architectural and is defined by the
364 		 * per-processor PPR. Restrict SNP support on the known CPU models
365 		 * for which the RMP table entry format is currently defined or for
366 		 * processors which support the architecturally defined RMPREAD
367 		 * instruction.
368 		 */
369 		if (!cpu_has(c, X86_FEATURE_HYPERVISOR) &&
370 		    (cpu_feature_enabled(X86_FEATURE_ZEN3) ||
371 		     cpu_feature_enabled(X86_FEATURE_ZEN4) ||
372 		     cpu_feature_enabled(X86_FEATURE_RMPREAD)) &&
373 		    snp_probe_rmptable_info()) {
374 			cc_platform_set(CC_ATTR_HOST_SEV_SNP);
375 		} else {
376 			setup_clear_cpu_cap(X86_FEATURE_SEV_SNP);
377 			cc_platform_clear(CC_ATTR_HOST_SEV_SNP);
378 		}
379 	}
380 #endif
381 }
382 
383 #define ZEN_MODEL_STEP_UCODE(fam, model, step, ucode) \
384 	X86_MATCH_VFM_STEPS(VFM_MAKE(X86_VENDOR_AMD, fam, model), \
385 			    step, step, ucode)
386 
387 static const struct x86_cpu_id amd_tsa_microcode[] = {
388 	ZEN_MODEL_STEP_UCODE(0x19, 0x01, 0x1, 0x0a0011d7),
389 	ZEN_MODEL_STEP_UCODE(0x19, 0x01, 0x2, 0x0a00123b),
390 	ZEN_MODEL_STEP_UCODE(0x19, 0x08, 0x2, 0x0a00820d),
391 	ZEN_MODEL_STEP_UCODE(0x19, 0x11, 0x1, 0x0a10114c),
392 	ZEN_MODEL_STEP_UCODE(0x19, 0x11, 0x2, 0x0a10124c),
393 	ZEN_MODEL_STEP_UCODE(0x19, 0x18, 0x1, 0x0a108109),
394 	ZEN_MODEL_STEP_UCODE(0x19, 0x21, 0x0, 0x0a20102e),
395 	ZEN_MODEL_STEP_UCODE(0x19, 0x21, 0x2, 0x0a201211),
396 	ZEN_MODEL_STEP_UCODE(0x19, 0x44, 0x1, 0x0a404108),
397 	ZEN_MODEL_STEP_UCODE(0x19, 0x50, 0x0, 0x0a500012),
398 	ZEN_MODEL_STEP_UCODE(0x19, 0x61, 0x2, 0x0a60120a),
399 	ZEN_MODEL_STEP_UCODE(0x19, 0x74, 0x1, 0x0a704108),
400 	ZEN_MODEL_STEP_UCODE(0x19, 0x75, 0x2, 0x0a705208),
401 	ZEN_MODEL_STEP_UCODE(0x19, 0x78, 0x0, 0x0a708008),
402 	ZEN_MODEL_STEP_UCODE(0x19, 0x7c, 0x0, 0x0a70c008),
403 	ZEN_MODEL_STEP_UCODE(0x19, 0xa0, 0x2, 0x0aa00216),
404 	{},
405 };
406 
407 static void tsa_init(struct cpuinfo_x86 *c)
408 {
409 	if (cpu_has(c, X86_FEATURE_HYPERVISOR))
410 		return;
411 
412 	if (cpu_has(c, X86_FEATURE_ZEN3) ||
413 	    cpu_has(c, X86_FEATURE_ZEN4)) {
414 		if (x86_match_min_microcode_rev(amd_tsa_microcode))
415 			setup_force_cpu_cap(X86_FEATURE_VERW_CLEAR);
416 		else
417 			pr_debug("%s: current revision: 0x%x\n", __func__, c->microcode);
418 	} else {
419 		setup_force_cpu_cap(X86_FEATURE_TSA_SQ_NO);
420 		setup_force_cpu_cap(X86_FEATURE_TSA_L1_NO);
421 	}
422 }
423 
424 static void bsp_init_amd(struct cpuinfo_x86 *c)
425 {
426 	if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) {
427 
428 		if (c->x86 > 0x10 ||
429 		    (c->x86 == 0x10 && c->x86_model >= 0x2)) {
430 			u64 val;
431 
432 			rdmsrq(MSR_K7_HWCR, val);
433 			if (!(val & BIT(24)))
434 				pr_warn(FW_BUG "TSC doesn't count with P0 frequency!\n");
435 		}
436 	}
437 
438 	if (c->x86 == 0x15) {
439 		unsigned long upperbit;
440 		u32 cpuid, assoc;
441 
442 		cpuid	 = cpuid_edx(0x80000005);
443 		assoc	 = cpuid >> 16 & 0xff;
444 		upperbit = ((cpuid >> 24) << 10) / assoc;
445 
446 		va_align.mask	  = (upperbit - 1) & PAGE_MASK;
447 		va_align.flags    = ALIGN_VA_32 | ALIGN_VA_64;
448 
449 		/* A random value per boot for bit slice [12:upper_bit) */
450 		va_align.bits = get_random_u32() & va_align.mask;
451 	}
452 
453 	if (cpu_has(c, X86_FEATURE_MWAITX))
454 		use_mwaitx_delay();
455 
456 	if (!boot_cpu_has(X86_FEATURE_AMD_SSBD) &&
457 	    !boot_cpu_has(X86_FEATURE_VIRT_SSBD) &&
458 	    c->x86 >= 0x15 && c->x86 <= 0x17) {
459 		unsigned int bit;
460 
461 		switch (c->x86) {
462 		case 0x15: bit = 54; break;
463 		case 0x16: bit = 33; break;
464 		case 0x17: bit = 10; break;
465 		default: return;
466 		}
467 		/*
468 		 * Try to cache the base value so further operations can
469 		 * avoid RMW. If that faults, do not enable SSBD.
470 		 */
471 		if (!rdmsrq_safe(MSR_AMD64_LS_CFG, &x86_amd_ls_cfg_base)) {
472 			setup_force_cpu_cap(X86_FEATURE_LS_CFG_SSBD);
473 			setup_force_cpu_cap(X86_FEATURE_SSBD);
474 			x86_amd_ls_cfg_ssbd_mask = 1ULL << bit;
475 		}
476 	}
477 
478 	resctrl_cpu_detect(c);
479 
480 	/* Figure out Zen generations: */
481 	switch (c->x86) {
482 	case 0x17:
483 		switch (c->x86_model) {
484 		case 0x00 ... 0x2f:
485 		case 0x50 ... 0x5f:
486 			setup_force_cpu_cap(X86_FEATURE_ZEN1);
487 			break;
488 		case 0x30 ... 0x4f:
489 		case 0x60 ... 0x7f:
490 		case 0x90 ... 0x91:
491 		case 0xa0 ... 0xaf:
492 			setup_force_cpu_cap(X86_FEATURE_ZEN2);
493 			break;
494 		default:
495 			goto warn;
496 		}
497 		break;
498 
499 	case 0x19:
500 		switch (c->x86_model) {
501 		case 0x00 ... 0x0f:
502 		case 0x20 ... 0x5f:
503 			setup_force_cpu_cap(X86_FEATURE_ZEN3);
504 			break;
505 		case 0x10 ... 0x1f:
506 		case 0x60 ... 0xaf:
507 			setup_force_cpu_cap(X86_FEATURE_ZEN4);
508 			break;
509 		default:
510 			goto warn;
511 		}
512 		break;
513 
514 	case 0x1a:
515 		switch (c->x86_model) {
516 		case 0x00 ... 0x2f:
517 		case 0x40 ... 0x4f:
518 		case 0x60 ... 0x7f:
519 		case 0xd0 ... 0xd7:
520 			setup_force_cpu_cap(X86_FEATURE_ZEN5);
521 			break;
522 		case 0x50 ... 0x5f:
523 		case 0x80 ... 0xaf:
524 		case 0xc0 ... 0xcf:
525 		case 0xd8 ... 0xef:
526 			setup_force_cpu_cap(X86_FEATURE_ZEN6);
527 			break;
528 		default:
529 			goto warn;
530 		}
531 		break;
532 
533 	default:
534 		break;
535 	}
536 
537 	bsp_determine_snp(c);
538 	tsa_init(c);
539 
540 	if (cpu_has(c, X86_FEATURE_GP_ON_USER_CPUID))
541 		setup_force_cpu_cap(X86_FEATURE_CPUID_FAULT);
542 
543 	return;
544 
545 warn:
546 	WARN_ONCE(1, "Family 0x%x, model: 0x%x??\n", c->x86, c->x86_model);
547 }
548 
549 static void early_detect_mem_encrypt(struct cpuinfo_x86 *c)
550 {
551 	u64 msr;
552 
553 	/*
554 	 * Mark using WBINVD is needed during kexec on processors that
555 	 * support SME. This provides support for performing a successful
556 	 * kexec when going from SME inactive to SME active (or vice-versa).
557 	 *
558 	 * The cache must be cleared so that if there are entries with the
559 	 * same physical address, both with and without the encryption bit,
560 	 * they don't race each other when flushed and potentially end up
561 	 * with the wrong entry being committed to memory.
562 	 *
563 	 * Test the CPUID bit directly because with mem_encrypt=off the
564 	 * BSP will clear the X86_FEATURE_SME bit and the APs will not
565 	 * see it set after that.
566 	 */
567 	if (c->extended_cpuid_level >= 0x8000001f && (cpuid_eax(0x8000001f) & BIT(0)))
568 		__this_cpu_write(cache_state_incoherent, true);
569 
570 	/*
571 	 * BIOS support is required for SME and SEV.
572 	 *   For SME: If BIOS has enabled SME then adjust x86_phys_bits by
573 	 *	      the SME physical address space reduction value.
574 	 *	      If BIOS has not enabled SME then don't advertise the
575 	 *	      SME feature (set in scattered.c).
576 	 *	      If the kernel has not enabled SME via any means then
577 	 *	      don't advertise the SME feature.
578 	 *   For SEV: If BIOS has not enabled SEV then don't advertise SEV and
579 	 *	      any additional functionality based on it.
580 	 *
581 	 *   In all cases, since support for SME and SEV requires long mode,
582 	 *   don't advertise the feature under CONFIG_X86_32.
583 	 */
584 	if (cpu_has(c, X86_FEATURE_SME) || cpu_has(c, X86_FEATURE_SEV)) {
585 		/* Check if memory encryption is enabled */
586 		rdmsrq(MSR_AMD64_SYSCFG, msr);
587 		if (!(msr & MSR_AMD64_SYSCFG_MEM_ENCRYPT))
588 			goto clear_all;
589 
590 		/*
591 		 * Always adjust physical address bits. Even though this
592 		 * will be a value above 32-bits this is still done for
593 		 * CONFIG_X86_32 so that accurate values are reported.
594 		 */
595 		c->x86_phys_bits -= (cpuid_ebx(0x8000001f) >> 6) & 0x3f;
596 
597 		if (IS_ENABLED(CONFIG_X86_32))
598 			goto clear_all;
599 
600 		if (!sme_me_mask)
601 			setup_clear_cpu_cap(X86_FEATURE_SME);
602 
603 		rdmsrq(MSR_K7_HWCR, msr);
604 		if (!(msr & MSR_K7_HWCR_SMMLOCK))
605 			goto clear_sev;
606 
607 		return;
608 
609 clear_all:
610 		setup_clear_cpu_cap(X86_FEATURE_SME);
611 clear_sev:
612 		setup_clear_cpu_cap(X86_FEATURE_SEV);
613 		setup_clear_cpu_cap(X86_FEATURE_SEV_ES);
614 		setup_clear_cpu_cap(X86_FEATURE_SEV_SNP);
615 	}
616 }
617 
618 static void early_init_amd(struct cpuinfo_x86 *c)
619 {
620 	u64 val;
621 
622 	if (c->x86 >= 0xf)
623 		set_cpu_cap(c, X86_FEATURE_K8);
624 
625 	rdmsrq_safe(MSR_AMD64_PATCH_LEVEL, &val);
626 	c->microcode = (u32)val;
627 
628 	/*
629 	 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
630 	 * with P/T states and does not stop in deep C-states
631 	 */
632 	if (c->x86_power & (1 << 8)) {
633 		set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
634 		set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
635 	}
636 
637 	/* Bit 12 of 8000_0007 edx is accumulated power mechanism. */
638 	if (c->x86_power & BIT(12))
639 		set_cpu_cap(c, X86_FEATURE_ACC_POWER);
640 
641 	/* Bit 14 indicates the Runtime Average Power Limit interface. */
642 	if (c->x86_power & BIT(14))
643 		set_cpu_cap(c, X86_FEATURE_RAPL);
644 
645 #ifdef CONFIG_X86_64
646 	set_cpu_cap(c, X86_FEATURE_SYSCALL32);
647 #else
648 	/*  Set MTRR capability flag if appropriate */
649 	if (c->x86 == 5)
650 		if (c->x86_model == 13 || c->x86_model == 9 ||
651 		    (c->x86_model == 8 && c->x86_stepping >= 8))
652 			set_cpu_cap(c, X86_FEATURE_K6_MTRR);
653 #endif
654 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
655 	/*
656 	 * ApicID can always be treated as an 8-bit value for AMD APIC versions
657 	 * >= 0x10, but even old K8s came out of reset with version 0x10. So, we
658 	 * can safely set X86_FEATURE_EXTD_APICID unconditionally for families
659 	 * after 16h.
660 	 */
661 	if (boot_cpu_has(X86_FEATURE_APIC)) {
662 		if (c->x86 > 0x16)
663 			set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
664 		else if (c->x86 >= 0xf) {
665 			/* check CPU config space for extended APIC ID */
666 			unsigned int val;
667 
668 			val = read_pci_config(0, 24, 0, 0x68);
669 			if ((val >> 17 & 0x3) == 0x3)
670 				set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
671 		}
672 	}
673 #endif
674 
675 	/*
676 	 * This is only needed to tell the kernel whether to use VMCALL
677 	 * and VMMCALL.  VMMCALL is never executed except under virt, so
678 	 * we can set it unconditionally.
679 	 */
680 	set_cpu_cap(c, X86_FEATURE_VMMCALL);
681 
682 	/* F16h erratum 793, CVE-2013-6885 */
683 	if (c->x86 == 0x16 && c->x86_model <= 0xf)
684 		msr_set_bit(MSR_AMD64_LS_CFG, 15);
685 
686 	early_detect_mem_encrypt(c);
687 
688 	if (!cpu_has(c, X86_FEATURE_HYPERVISOR) && !cpu_has(c, X86_FEATURE_IBPB_BRTYPE)) {
689 		if (c->x86 == 0x17 && boot_cpu_has(X86_FEATURE_AMD_IBPB))
690 			setup_force_cpu_cap(X86_FEATURE_IBPB_BRTYPE);
691 		else if (c->x86 >= 0x19 && !wrmsrq_safe(MSR_IA32_PRED_CMD, PRED_CMD_SBPB)) {
692 			setup_force_cpu_cap(X86_FEATURE_IBPB_BRTYPE);
693 			setup_force_cpu_cap(X86_FEATURE_SBPB);
694 		}
695 	}
696 }
697 
698 static void init_amd_k8(struct cpuinfo_x86 *c)
699 {
700 	u32 level;
701 	u64 value;
702 
703 	/* On C+ stepping K8 rep microcode works well for copy/memset */
704 	level = cpuid_eax(1);
705 	if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58)
706 		set_cpu_cap(c, X86_FEATURE_REP_GOOD);
707 
708 	/*
709 	 * Some BIOSes incorrectly force this feature, but only K8 revision D
710 	 * (model = 0x14) and later actually support it.
711 	 * (AMD Erratum #110, docId: 25759).
712 	 */
713 	if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM) && !cpu_has(c, X86_FEATURE_HYPERVISOR)) {
714 		clear_cpu_cap(c, X86_FEATURE_LAHF_LM);
715 		if (!rdmsrq_amd_safe(0xc001100d, &value)) {
716 			value &= ~BIT_64(32);
717 			wrmsrq_amd_safe(0xc001100d, value);
718 		}
719 	}
720 
721 	if (!c->x86_model_id[0])
722 		strscpy(c->x86_model_id, "Hammer");
723 
724 #ifdef CONFIG_SMP
725 	/*
726 	 * Disable TLB flush filter by setting HWCR.FFDIS on K8
727 	 * bit 6 of msr C001_0015
728 	 *
729 	 * Errata 63 for SH-B3 steppings
730 	 * Errata 122 for all steppings (F+ have it disabled by default)
731 	 */
732 	msr_set_bit(MSR_K7_HWCR, 6);
733 #endif
734 	set_cpu_bug(c, X86_BUG_SWAPGS_FENCE);
735 
736 	/*
737 	 * Check models and steppings affected by erratum 400. This is
738 	 * used to select the proper idle routine and to enable the
739 	 * check whether the machine is affected in arch_post_acpi_subsys_init()
740 	 * which sets the X86_BUG_AMD_APIC_C1E bug depending on the MSR check.
741 	 */
742 	if (c->x86_model > 0x41 ||
743 	    (c->x86_model == 0x41 && c->x86_stepping >= 0x2))
744 		setup_force_cpu_bug(X86_BUG_AMD_E400);
745 }
746 
747 static void init_amd_gh(struct cpuinfo_x86 *c)
748 {
749 #ifdef CONFIG_MMCONF_FAM10H
750 	/* do this for boot cpu */
751 	if (c == &boot_cpu_data)
752 		check_enable_amd_mmconf_dmi();
753 
754 	fam10h_check_enable_mmcfg();
755 #endif
756 
757 	/*
758 	 * Disable GART TLB Walk Errors on Fam10h. We do this here because this
759 	 * is always needed when GART is enabled, even in a kernel which has no
760 	 * MCE support built in. BIOS should disable GartTlbWlk Errors already.
761 	 * If it doesn't, we do it here as suggested by the BKDG.
762 	 *
763 	 * Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=33012
764 	 */
765 	msr_set_bit(MSR_AMD64_MCx_MASK(4), 10);
766 
767 	/*
768 	 * On family 10h BIOS may not have properly enabled WC+ support, causing
769 	 * it to be converted to CD memtype. This may result in performance
770 	 * degradation for certain nested-paging guests. Prevent this conversion
771 	 * by clearing bit 24 in MSR_AMD64_BU_CFG2.
772 	 *
773 	 * NOTE: we want to use the _safe accessors so as not to #GP kvm
774 	 * guests on older kvm hosts.
775 	 */
776 	msr_clear_bit(MSR_AMD64_BU_CFG2, 24);
777 
778 	set_cpu_bug(c, X86_BUG_AMD_TLB_MMATCH);
779 
780 	/*
781 	 * Check models and steppings affected by erratum 400. This is
782 	 * used to select the proper idle routine and to enable the
783 	 * check whether the machine is affected in arch_post_acpi_subsys_init()
784 	 * which sets the X86_BUG_AMD_APIC_C1E bug depending on the MSR check.
785 	 */
786 	if (c->x86_model > 0x2 ||
787 	    (c->x86_model == 0x2 && c->x86_stepping >= 0x1))
788 		setup_force_cpu_bug(X86_BUG_AMD_E400);
789 }
790 
791 static void init_amd_ln(struct cpuinfo_x86 *c)
792 {
793 	/*
794 	 * Apply erratum 665 fix unconditionally so machines without a BIOS
795 	 * fix work.
796 	 */
797 	msr_set_bit(MSR_AMD64_DE_CFG, 31);
798 }
799 
800 static bool rdrand_force;
801 
802 static int __init rdrand_cmdline(char *str)
803 {
804 	if (!str)
805 		return -EINVAL;
806 
807 	if (!strcmp(str, "force"))
808 		rdrand_force = true;
809 	else
810 		return -EINVAL;
811 
812 	return 0;
813 }
814 early_param("rdrand", rdrand_cmdline);
815 
816 static void clear_rdrand_cpuid_bit(struct cpuinfo_x86 *c)
817 {
818 	/*
819 	 * Saving of the MSR used to hide the RDRAND support during
820 	 * suspend/resume is done by arch/x86/power/cpu.c, which is
821 	 * dependent on CONFIG_PM_SLEEP.
822 	 */
823 	if (!IS_ENABLED(CONFIG_PM_SLEEP))
824 		return;
825 
826 	/*
827 	 * The self-test can clear X86_FEATURE_RDRAND, so check for
828 	 * RDRAND support using the CPUID function directly.
829 	 */
830 	if (!(cpuid_ecx(1) & BIT(30)) || rdrand_force)
831 		return;
832 
833 	msr_clear_bit(MSR_AMD64_CPUID_FN_1, 62);
834 
835 	/*
836 	 * Verify that the CPUID change has occurred in case the kernel is
837 	 * running virtualized and the hypervisor doesn't support the MSR.
838 	 */
839 	if (cpuid_ecx(1) & BIT(30)) {
840 		pr_info_once("BIOS may not properly restore RDRAND after suspend, but hypervisor does not support hiding RDRAND via CPUID.\n");
841 		return;
842 	}
843 
844 	clear_cpu_cap(c, X86_FEATURE_RDRAND);
845 	pr_info_once("BIOS may not properly restore RDRAND after suspend, hiding RDRAND via CPUID. Use rdrand=force to reenable.\n");
846 }
847 
848 static void init_amd_jg(struct cpuinfo_x86 *c)
849 {
850 	/*
851 	 * Some BIOS implementations do not restore proper RDRAND support
852 	 * across suspend and resume. Check on whether to hide the RDRAND
853 	 * instruction support via CPUID.
854 	 */
855 	clear_rdrand_cpuid_bit(c);
856 }
857 
858 static void init_amd_bd(struct cpuinfo_x86 *c)
859 {
860 	u64 value;
861 
862 	/*
863 	 * The way access filter has a performance penalty on some workloads.
864 	 * Disable it on the affected CPUs.
865 	 */
866 	if ((c->x86_model >= 0x02) && (c->x86_model < 0x20)) {
867 		if (!rdmsrq_safe(MSR_F15H_IC_CFG, &value) && !(value & 0x1E)) {
868 			value |= 0x1E;
869 			wrmsrq_safe(MSR_F15H_IC_CFG, value);
870 		}
871 	}
872 
873 	/*
874 	 * Some BIOS implementations do not restore proper RDRAND support
875 	 * across suspend and resume. Check on whether to hide the RDRAND
876 	 * instruction support via CPUID.
877 	 */
878 	clear_rdrand_cpuid_bit(c);
879 }
880 
881 static const struct x86_cpu_id erratum_1386_microcode[] = {
882 	ZEN_MODEL_STEP_UCODE(0x17, 0x01, 0x2, 0x0800126e),
883 	ZEN_MODEL_STEP_UCODE(0x17, 0x31, 0x0, 0x08301052),
884 	{}
885 };
886 
887 static void fix_erratum_1386(struct cpuinfo_x86 *c)
888 {
889 	/*
890 	 * Work around Erratum 1386.  The XSAVES instruction malfunctions in
891 	 * certain circumstances on Zen1/2 uarch, and not all parts have had
892 	 * updated microcode at the time of writing (March 2023).
893 	 *
894 	 * Affected parts all have no supervisor XSAVE states, meaning that
895 	 * the XSAVEC instruction (which works fine) is equivalent.
896 	 *
897 	 * Clear the feature flag only on microcode revisions which
898 	 * don't have the fix.
899 	 */
900 	if (x86_match_min_microcode_rev(erratum_1386_microcode))
901 		return;
902 
903 	clear_cpu_cap(c, X86_FEATURE_XSAVES);
904 }
905 
906 void init_spectral_chicken(struct cpuinfo_x86 *c)
907 {
908 #ifdef CONFIG_MITIGATION_UNRET_ENTRY
909 	/*
910 	 * On Zen2 we offer this chicken (bit) on the altar of Speculation.
911 	 *
912 	 * This suppresses speculation from the middle of a basic block, i.e. it
913 	 * suppresses non-branch predictions.
914 	 */
915 	if (!cpu_has(c, X86_FEATURE_HYPERVISOR))
916 		msr_set_bit(MSR_ZEN2_SPECTRAL_CHICKEN, MSR_ZEN2_SPECTRAL_CHICKEN_BIT);
917 #endif
918 }
919 
920 static void init_amd_zen_common(void)
921 {
922 	setup_force_cpu_cap(X86_FEATURE_ZEN);
923 #ifdef CONFIG_NUMA
924 	node_reclaim_distance = 32;
925 #endif
926 }
927 
928 static void init_amd_zen1(struct cpuinfo_x86 *c)
929 {
930 	fix_erratum_1386(c);
931 
932 	/* Fix up CPUID bits, but only if not virtualised. */
933 	if (!cpu_has(c, X86_FEATURE_HYPERVISOR)) {
934 
935 		/* Erratum 1076: CPB feature bit not being set in CPUID. */
936 		if (!cpu_has(c, X86_FEATURE_CPB))
937 			set_cpu_cap(c, X86_FEATURE_CPB);
938 	}
939 
940 	pr_notice_once("AMD Zen1 DIV0 bug detected. Disable SMT for full protection.\n");
941 	setup_force_cpu_bug(X86_BUG_DIV0);
942 
943 	/*
944 	 * Turn off the Instructions Retired free counter on machines that are
945 	 * susceptible to erratum #1054 "Instructions Retired Performance
946 	 * Counter May Be Inaccurate".
947 	 */
948 	if (c->x86_model < 0x30) {
949 		msr_clear_bit(MSR_K7_HWCR, MSR_K7_HWCR_IRPERF_EN_BIT);
950 		clear_cpu_cap(c, X86_FEATURE_IRPERF);
951 	}
952 
953 	pr_notice_once("AMD Zen1 FPDSS bug detected, enabling mitigation.\n");
954 	msr_set_bit(MSR_AMD64_FP_CFG, MSR_AMD64_FP_CFG_ZEN1_DENORM_FIX_BIT);
955 }
956 
957 static const struct x86_cpu_id amd_zenbleed_microcode[] = {
958 	ZEN_MODEL_STEP_UCODE(0x17, 0x31, 0x0, 0x0830107b),
959 	ZEN_MODEL_STEP_UCODE(0x17, 0x60, 0x1, 0x0860010c),
960 	ZEN_MODEL_STEP_UCODE(0x17, 0x68, 0x1, 0x08608107),
961 	ZEN_MODEL_STEP_UCODE(0x17, 0x71, 0x0, 0x08701033),
962 	ZEN_MODEL_STEP_UCODE(0x17, 0xa0, 0x0, 0x08a00009),
963 	{}
964 };
965 
966 static void zen2_zenbleed_check(struct cpuinfo_x86 *c)
967 {
968 	if (cpu_has(c, X86_FEATURE_HYPERVISOR))
969 		return;
970 
971 	if (!cpu_has(c, X86_FEATURE_AVX))
972 		return;
973 
974 	if (!x86_match_min_microcode_rev(amd_zenbleed_microcode)) {
975 		pr_notice_once("Zenbleed: please update your microcode for the most optimal fix\n");
976 		msr_set_bit(MSR_AMD64_DE_CFG, MSR_AMD64_DE_CFG_ZEN2_FP_BACKUP_FIX_BIT);
977 	} else {
978 		msr_clear_bit(MSR_AMD64_DE_CFG, MSR_AMD64_DE_CFG_ZEN2_FP_BACKUP_FIX_BIT);
979 	}
980 }
981 
982 static void init_amd_zen2(struct cpuinfo_x86 *c)
983 {
984 	init_spectral_chicken(c);
985 	fix_erratum_1386(c);
986 	zen2_zenbleed_check(c);
987 
988 	/* Disable RDSEED on AMD Cyan Skillfish because of an error. */
989 	if (c->x86_model == 0x47 && c->x86_stepping == 0x0) {
990 		clear_cpu_cap(c, X86_FEATURE_RDSEED);
991 		msr_clear_bit(MSR_AMD64_CPUID_FN_7, 18);
992 		pr_emerg("RDSEED is not reliable on this platform; disabling.\n");
993 	}
994 
995 	/* Correct misconfigured CPUID on some clients. */
996 	clear_cpu_cap(c, X86_FEATURE_INVLPGB);
997 
998 	if (!cpu_has(c, X86_FEATURE_HYPERVISOR))
999 		msr_set_bit(MSR_ZEN4_BP_CFG, MSR_ZEN2_BP_CFG_BUG_FIX_BIT);
1000 }
1001 
1002 static void init_amd_zen3(struct cpuinfo_x86 *c)
1003 {
1004 	if (!cpu_has(c, X86_FEATURE_HYPERVISOR)) {
1005 		/*
1006 		 * Zen3 (Fam19 model < 0x10) parts are not susceptible to
1007 		 * Branch Type Confusion, but predate the allocation of the
1008 		 * BTC_NO bit.
1009 		 */
1010 		if (!cpu_has(c, X86_FEATURE_BTC_NO))
1011 			set_cpu_cap(c, X86_FEATURE_BTC_NO);
1012 	}
1013 }
1014 
1015 static void init_amd_zen4(struct cpuinfo_x86 *c)
1016 {
1017 	if (!cpu_has(c, X86_FEATURE_HYPERVISOR))
1018 		msr_set_bit(MSR_ZEN4_BP_CFG, MSR_ZEN4_BP_CFG_SHARED_BTB_FIX_BIT);
1019 
1020 	/*
1021 	 * These Zen4 SoCs advertise support for virtualized VMLOAD/VMSAVE
1022 	 * in some BIOS versions but they can lead to random host reboots.
1023 	 */
1024 	switch (c->x86_model) {
1025 	case 0x18 ... 0x1f:
1026 	case 0x60 ... 0x7f:
1027 		clear_cpu_cap(c, X86_FEATURE_V_VMSAVE_VMLOAD);
1028 		break;
1029 	}
1030 }
1031 
1032 static const struct x86_cpu_id zen5_rdseed_microcode[] = {
1033 	ZEN_MODEL_STEP_UCODE(0x1a, 0x02, 0x1, 0x0b00215a),
1034 	ZEN_MODEL_STEP_UCODE(0x1a, 0x08, 0x1, 0x0b008121),
1035 	ZEN_MODEL_STEP_UCODE(0x1a, 0x11, 0x0, 0x0b101054),
1036 	ZEN_MODEL_STEP_UCODE(0x1a, 0x24, 0x0, 0x0b204037),
1037 	ZEN_MODEL_STEP_UCODE(0x1a, 0x44, 0x0, 0x0b404035),
1038 	ZEN_MODEL_STEP_UCODE(0x1a, 0x44, 0x1, 0x0b404108),
1039 	ZEN_MODEL_STEP_UCODE(0x1a, 0x60, 0x0, 0x0b600037),
1040 	ZEN_MODEL_STEP_UCODE(0x1a, 0x68, 0x0, 0x0b608038),
1041 	ZEN_MODEL_STEP_UCODE(0x1a, 0x70, 0x0, 0x0b700037),
1042 	{},
1043 };
1044 
1045 static void init_amd_zen5(struct cpuinfo_x86 *c)
1046 {
1047 	if (!x86_match_min_microcode_rev(zen5_rdseed_microcode)) {
1048 		clear_cpu_cap(c, X86_FEATURE_RDSEED);
1049 		msr_clear_bit(MSR_AMD64_CPUID_FN_7, 18);
1050 		pr_emerg_once("RDSEED32 is broken. Disabling the corresponding CPUID bit.\n");
1051 	}
1052 }
1053 
1054 static void init_amd(struct cpuinfo_x86 *c)
1055 {
1056 	u64 vm_cr;
1057 
1058 	early_init_amd(c);
1059 
1060 	if (c->x86 >= 0x10)
1061 		set_cpu_cap(c, X86_FEATURE_REP_GOOD);
1062 
1063 	/* AMD FSRM also implies FSRS */
1064 	if (cpu_has(c, X86_FEATURE_FSRM))
1065 		set_cpu_cap(c, X86_FEATURE_FSRS);
1066 
1067 	/* K6s reports MCEs but don't actually have all the MSRs */
1068 	if (c->x86 < 6)
1069 		clear_cpu_cap(c, X86_FEATURE_MCE);
1070 
1071 	switch (c->x86) {
1072 	case 4:    init_amd_k5(c); break;
1073 	case 5:    init_amd_k6(c); break;
1074 	case 6:	   init_amd_k7(c); break;
1075 	case 0xf:  init_amd_k8(c); break;
1076 	case 0x10: init_amd_gh(c); break;
1077 	case 0x12: init_amd_ln(c); break;
1078 	case 0x15: init_amd_bd(c); break;
1079 	case 0x16: init_amd_jg(c); break;
1080 	}
1081 
1082 	/*
1083 	 * Save up on some future enablement work and do common Zen
1084 	 * settings.
1085 	 */
1086 	if (c->x86 >= 0x17)
1087 		init_amd_zen_common();
1088 
1089 	if (boot_cpu_has(X86_FEATURE_ZEN1))
1090 		init_amd_zen1(c);
1091 	else if (boot_cpu_has(X86_FEATURE_ZEN2))
1092 		init_amd_zen2(c);
1093 	else if (boot_cpu_has(X86_FEATURE_ZEN3))
1094 		init_amd_zen3(c);
1095 	else if (boot_cpu_has(X86_FEATURE_ZEN4))
1096 		init_amd_zen4(c);
1097 	else if (boot_cpu_has(X86_FEATURE_ZEN5))
1098 		init_amd_zen5(c);
1099 
1100 	/*
1101 	 * Enable workaround for FXSAVE leak on CPUs
1102 	 * without a XSaveErPtr feature
1103 	 */
1104 	if ((c->x86 >= 6) && (!cpu_has(c, X86_FEATURE_XSAVEERPTR)))
1105 		set_cpu_bug(c, X86_BUG_FXSAVE_LEAK);
1106 
1107 	cpu_detect_cache_sizes(c);
1108 
1109 	srat_detect_node(c);
1110 
1111 	init_amd_cacheinfo(c);
1112 
1113 	if (cpu_has(c, X86_FEATURE_SVM)) {
1114 		rdmsrq(MSR_VM_CR, vm_cr);
1115 		if (vm_cr & SVM_VM_CR_SVM_DIS_MASK) {
1116 			pr_notice_once("SVM disabled (by BIOS) in MSR_VM_CR\n");
1117 			clear_cpu_cap(c, X86_FEATURE_SVM);
1118 		}
1119 	}
1120 
1121 	if (!cpu_has(c, X86_FEATURE_LFENCE_RDTSC) && cpu_has(c, X86_FEATURE_XMM2)) {
1122 		/*
1123 		 * Use LFENCE for execution serialization.  On families which
1124 		 * don't have that MSR, LFENCE is already serializing.
1125 		 * msr_set_bit() uses the safe accessors, too, even if the MSR
1126 		 * is not present.
1127 		 */
1128 		msr_set_bit(MSR_AMD64_DE_CFG,
1129 			    MSR_AMD64_DE_CFG_LFENCE_SERIALIZE_BIT);
1130 
1131 		/* A serializing LFENCE stops RDTSC speculation */
1132 		set_cpu_cap(c, X86_FEATURE_LFENCE_RDTSC);
1133 	}
1134 
1135 	/*
1136 	 * Family 0x12 and above processors have APIC timer
1137 	 * running in deep C states.
1138 	 */
1139 	if (c->x86 > 0x11)
1140 		set_cpu_cap(c, X86_FEATURE_ARAT);
1141 
1142 	/* 3DNow or LM implies PREFETCHW */
1143 	if (!cpu_has(c, X86_FEATURE_3DNOWPREFETCH))
1144 		if (cpu_has(c, X86_FEATURE_3DNOW) || cpu_has(c, X86_FEATURE_LM))
1145 			set_cpu_cap(c, X86_FEATURE_3DNOWPREFETCH);
1146 
1147 	/* AMD CPUs don't reset SS attributes on SYSRET, Xen does. */
1148 	if (!cpu_feature_enabled(X86_FEATURE_XENPV))
1149 		set_cpu_bug(c, X86_BUG_SYSRET_SS_ATTRS);
1150 
1151 	/* Enable the Instructions Retired free counter */
1152 	if (cpu_has(c, X86_FEATURE_IRPERF))
1153 		msr_set_bit(MSR_K7_HWCR, MSR_K7_HWCR_IRPERF_EN_BIT);
1154 
1155 	check_null_seg_clears_base(c);
1156 
1157 	/*
1158 	 * Make sure EFER[AIBRSE - Automatic IBRS Enable] is set. The APs are brought up
1159 	 * using the trampoline code and as part of it, MSR_EFER gets prepared there in
1160 	 * order to be replicated onto them. Regardless, set it here again, if not set,
1161 	 * to protect against any future refactoring/code reorganization which might
1162 	 * miss setting this important bit.
1163 	 */
1164 	if (spectre_v2_in_eibrs_mode(spectre_v2_enabled) &&
1165 	    cpu_has(c, X86_FEATURE_AUTOIBRS))
1166 		WARN_ON_ONCE(msr_set_bit(MSR_EFER, _EFER_AUTOIBRS) < 0);
1167 
1168 	/* AMD CPUs don't need fencing after x2APIC/TSC_DEADLINE MSR writes. */
1169 	clear_cpu_cap(c, X86_FEATURE_APIC_MSRS_FENCE);
1170 
1171 	/* Enable Translation Cache Extension */
1172 	if (cpu_has(c, X86_FEATURE_TCE))
1173 		msr_set_bit(MSR_EFER, _EFER_TCE);
1174 }
1175 
1176 #ifdef CONFIG_X86_32
1177 static unsigned int amd_size_cache(struct cpuinfo_x86 *c, unsigned int size)
1178 {
1179 	/* AMD errata T13 (order #21922) */
1180 	if (c->x86 == 6) {
1181 		/* Duron Rev A0 */
1182 		if (c->x86_model == 3 && c->x86_stepping == 0)
1183 			size = 64;
1184 		/* Tbird rev A1/A2 */
1185 		if (c->x86_model == 4 &&
1186 			(c->x86_stepping == 0 || c->x86_stepping == 1))
1187 			size = 256;
1188 	}
1189 	return size;
1190 }
1191 #endif
1192 
1193 static void cpu_detect_tlb_amd(struct cpuinfo_x86 *c)
1194 {
1195 	u32 ebx, eax, ecx, edx;
1196 	u16 mask = 0xfff;
1197 
1198 	if (c->x86 < 0xf)
1199 		return;
1200 
1201 	if (c->extended_cpuid_level < 0x80000006)
1202 		return;
1203 
1204 	cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
1205 
1206 	tlb_lld_4k = (ebx >> 16) & mask;
1207 	tlb_lli_4k = ebx & mask;
1208 
1209 	/*
1210 	 * K8 doesn't have 2M/4M entries in the L2 TLB so read out the L1 TLB
1211 	 * characteristics from the CPUID function 0x80000005 instead.
1212 	 */
1213 	if (c->x86 == 0xf) {
1214 		cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
1215 		mask = 0xff;
1216 	}
1217 
1218 	/* Handle DTLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
1219 	if (!((eax >> 16) & mask))
1220 		tlb_lld_2m = (cpuid_eax(0x80000005) >> 16) & 0xff;
1221 	else
1222 		tlb_lld_2m = (eax >> 16) & mask;
1223 
1224 	/* a 4M entry uses two 2M entries */
1225 	tlb_lld_4m = tlb_lld_2m >> 1;
1226 
1227 	/* Handle ITLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
1228 	if (!(eax & mask)) {
1229 		/* Erratum 658 */
1230 		if (c->x86 == 0x15 && c->x86_model <= 0x1f) {
1231 			tlb_lli_2m = 1024;
1232 		} else {
1233 			cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
1234 			tlb_lli_2m = eax & 0xff;
1235 		}
1236 	} else
1237 		tlb_lli_2m = eax & mask;
1238 
1239 	tlb_lli_4m = tlb_lli_2m >> 1;
1240 
1241 	/* Max number of pages INVLPGB can invalidate in one shot */
1242 	if (cpu_has(c, X86_FEATURE_INVLPGB))
1243 		invlpgb_count_max = (cpuid_edx(0x80000008) & 0xffff) + 1;
1244 }
1245 
1246 static const struct cpu_dev amd_cpu_dev = {
1247 	.c_vendor	= "AMD",
1248 	.c_ident	= { "AuthenticAMD" },
1249 #ifdef CONFIG_X86_32
1250 	.legacy_models = {
1251 		{ .family = 4, .model_names =
1252 		  {
1253 			  [3] = "486 DX/2",
1254 			  [7] = "486 DX/2-WB",
1255 			  [8] = "486 DX/4",
1256 			  [9] = "486 DX/4-WB",
1257 			  [14] = "Am5x86-WT",
1258 			  [15] = "Am5x86-WB"
1259 		  }
1260 		},
1261 	},
1262 	.legacy_cache_size = amd_size_cache,
1263 #endif
1264 	.c_early_init   = early_init_amd,
1265 	.c_detect_tlb	= cpu_detect_tlb_amd,
1266 	.c_bsp_init	= bsp_init_amd,
1267 	.c_init		= init_amd,
1268 	.c_x86_vendor	= X86_VENDOR_AMD,
1269 };
1270 
1271 cpu_dev_register(amd_cpu_dev);
1272 
1273 static DEFINE_PER_CPU_READ_MOSTLY(unsigned long[4], amd_dr_addr_mask);
1274 
1275 static unsigned int amd_msr_dr_addr_masks[] = {
1276 	MSR_F16H_DR0_ADDR_MASK,
1277 	MSR_F16H_DR1_ADDR_MASK,
1278 	MSR_F16H_DR1_ADDR_MASK + 1,
1279 	MSR_F16H_DR1_ADDR_MASK + 2
1280 };
1281 
1282 void amd_set_dr_addr_mask(unsigned long mask, unsigned int dr)
1283 {
1284 	int cpu = smp_processor_id();
1285 
1286 	if (!cpu_feature_enabled(X86_FEATURE_BPEXT))
1287 		return;
1288 
1289 	if (WARN_ON_ONCE(dr >= ARRAY_SIZE(amd_msr_dr_addr_masks)))
1290 		return;
1291 
1292 	if (per_cpu(amd_dr_addr_mask, cpu)[dr] == mask)
1293 		return;
1294 
1295 	wrmsrq(amd_msr_dr_addr_masks[dr], mask);
1296 	per_cpu(amd_dr_addr_mask, cpu)[dr] = mask;
1297 }
1298 
1299 unsigned long amd_get_dr_addr_mask(unsigned int dr)
1300 {
1301 	if (!cpu_feature_enabled(X86_FEATURE_BPEXT))
1302 		return 0;
1303 
1304 	if (WARN_ON_ONCE(dr >= ARRAY_SIZE(amd_msr_dr_addr_masks)))
1305 		return 0;
1306 
1307 	return per_cpu(amd_dr_addr_mask[dr], smp_processor_id());
1308 }
1309 EXPORT_SYMBOL_FOR_KVM(amd_get_dr_addr_mask);
1310 
1311 static void zenbleed_check_cpu(void *unused)
1312 {
1313 	struct cpuinfo_x86 *c = &cpu_data(smp_processor_id());
1314 
1315 	zen2_zenbleed_check(c);
1316 }
1317 
1318 void amd_check_microcode(void)
1319 {
1320 	if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD)
1321 		return;
1322 
1323 	if (cpu_feature_enabled(X86_FEATURE_ZEN2))
1324 		on_each_cpu(zenbleed_check_cpu, NULL, 1);
1325 }
1326 
1327 static const char * const s5_reset_reason_txt[] = {
1328 	[0]  = "thermal pin BP_THERMTRIP_L was tripped",
1329 	[1]  = "power button was pressed for 4 seconds",
1330 	[2]  = "shutdown pin was tripped",
1331 	[4]  = "remote ASF power off command was received",
1332 	[9]  = "internal CPU thermal limit was tripped",
1333 	[16] = "system reset pin BP_SYS_RST_L was tripped",
1334 	[17] = "software issued PCI reset",
1335 	[18] = "software wrote 0x4 to reset control register 0xCF9",
1336 	[19] = "software wrote 0x6 to reset control register 0xCF9",
1337 	[20] = "software wrote 0xE to reset control register 0xCF9",
1338 	[21] = "ACPI power state transition occurred",
1339 	[22] = "keyboard reset pin KB_RST_L was tripped",
1340 	[23] = "internal CPU shutdown event occurred",
1341 	[24] = "system failed to boot before failed boot timer expired",
1342 	[25] = "hardware watchdog timer expired",
1343 	[26] = "remote ASF reset command was received",
1344 	[27] = "an uncorrected error caused a data fabric sync flood event",
1345 	[29] = "FCH and MP1 failed warm reset handshake",
1346 	[30] = "a parity error occurred",
1347 	[31] = "a software sync flood event occurred",
1348 };
1349 
1350 static __init int print_s5_reset_status_mmio(void)
1351 {
1352 	void __iomem *addr;
1353 	u32 value;
1354 	int i;
1355 
1356 	if (!cpu_feature_enabled(X86_FEATURE_ZEN))
1357 		return 0;
1358 
1359 	addr = ioremap(FCH_PM_BASE + FCH_PM_S5_RESET_STATUS, sizeof(value));
1360 	if (!addr)
1361 		return 0;
1362 
1363 	value = ioread32(addr);
1364 
1365 	/* Value with "all bits set" is an error response and should be ignored. */
1366 	if (value == U32_MAX) {
1367 		iounmap(addr);
1368 		return 0;
1369 	}
1370 
1371 	/*
1372 	 * Clear all reason bits so they won't be retained if the next reset
1373 	 * does not update the register. Besides, some bits are never cleared by
1374 	 * hardware so it's software's responsibility to clear them.
1375 	 *
1376 	 * Writing the value back effectively clears all reason bits as they are
1377 	 * write-1-to-clear.
1378 	 */
1379 	iowrite32(value, addr);
1380 	iounmap(addr);
1381 
1382 	for (i = 0; i < ARRAY_SIZE(s5_reset_reason_txt); i++) {
1383 		if (!(value & BIT(i)))
1384 			continue;
1385 
1386 		if (s5_reset_reason_txt[i]) {
1387 			pr_info("x86/amd: Previous system reset reason [0x%08x]: %s\n",
1388 				value, s5_reset_reason_txt[i]);
1389 		}
1390 	}
1391 
1392 	return 0;
1393 }
1394 late_initcall(print_s5_reset_status_mmio);
1395 
1396 static void __init dmi_scan_additional(const struct dmi_header *d, void *p)
1397 {
1398 	struct dmi_a_info *info = (struct dmi_a_info *)d;
1399 	void *next, *end;
1400 
1401 	if (!IS_ENABLED(CONFIG_DMI))
1402 		return;
1403 
1404 	if (info->header.type != DMI_ENTRY_ADDITIONAL ||
1405 	    info->header.length < DMI_A_INFO_MIN_SIZE ||
1406 	    info->count < 1)
1407 		return;
1408 
1409 	next = (void *)(info + 1);
1410 	end  = (void *)info + info->header.length;
1411 
1412 	do {
1413 		struct dmi_a_info_entry *entry;
1414 		const char *string_ptr;
1415 
1416 		entry = (struct dmi_a_info_entry *)next;
1417 
1418 		/*
1419 		 * Not much can be done to validate data. At least the entry
1420 		 * length shouldn't be 0.
1421 		 */
1422 		if (!entry->length)
1423 			return;
1424 
1425 		string_ptr = dmi_string_nosave(&info->header, entry->str_num);
1426 
1427 		/* Sample string: AGESA!V9 StrixKrackanPI-FP8 1.1.0.0c */
1428 		if (!strncmp(string_ptr, "AGESA", 5)) {
1429 			pr_info("AGESA: %s\n", string_ptr);
1430 			break;
1431 		}
1432 
1433 		next += entry->length;
1434 	} while (end - next >= DMI_A_INFO_ENT_MIN_SIZE);
1435 }
1436 
1437 static __init int print_dmi_agesa(void)
1438 {
1439 	dmi_walk(dmi_scan_additional, NULL);
1440 	return 0;
1441 }
1442 late_initcall(print_dmi_agesa);
1443