xref: /linux/arch/x86/kernel/cpu/amd.c (revision 26ba30221c03364d6ed9910be8da4c1fd871b07b)
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 	u32 l, h;
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 		rdmsr(MSR_K6_WHCR, l, h);
164 		if ((l&0x0000FFFF) == 0) {
165 			unsigned long flags;
166 			l = (1<<0)|((mbytes/4)<<1);
167 			local_irq_save(flags);
168 			wbinvd();
169 			wrmsr(MSR_K6_WHCR, l, h);
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 		rdmsr(MSR_K6_WHCR, l, h);
185 		if ((l&0xFFFF0000) == 0) {
186 			unsigned long flags;
187 			l = ((mbytes>>2)<<22)|(1<<16);
188 			local_irq_save(flags);
189 			wbinvd();
190 			wrmsr(MSR_K6_WHCR, l, h);
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 	u32 l, h;
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 		rdmsr(MSR_K7_CLK_CTL, l, h);
232 		if ((l & 0xfff00000) != 0x20000000) {
233 			pr_info("CPU: CLK_CTL MSR was %x. Reprogramming to %x\n",
234 				l, ((l & 0x000fffff)|0x20000000));
235 			wrmsr(MSR_K7_CLK_CTL, (l & 0x000fffff)|0x20000000, h);
236 		}
237 	}
238 
239 	/* calling is from identify_secondary_cpu() ? */
240 	if (!c->cpu_index)
241 		return;
242 
243 	/*
244 	 * Certain Athlons might work (for various values of 'work') in SMP
245 	 * but they are not certified as MP capable.
246 	 */
247 	/* Athlon 660/661 is valid. */
248 	if ((c->x86_model == 6) && ((c->x86_stepping == 0) ||
249 	    (c->x86_stepping == 1)))
250 		return;
251 
252 	/* Duron 670 is valid */
253 	if ((c->x86_model == 7) && (c->x86_stepping == 0))
254 		return;
255 
256 	/*
257 	 * Athlon 662, Duron 671, and Athlon >model 7 have capability
258 	 * bit. It's worth noting that the A5 stepping (662) of some
259 	 * Athlon XP's have the MP bit set.
260 	 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for
261 	 * more.
262 	 */
263 	if (((c->x86_model == 6) && (c->x86_stepping >= 2)) ||
264 	    ((c->x86_model == 7) && (c->x86_stepping >= 1)) ||
265 	     (c->x86_model > 7))
266 		if (cpu_has(c, X86_FEATURE_MP))
267 			return;
268 
269 	/* If we get here, not a certified SMP capable AMD system. */
270 
271 	/*
272 	 * Don't taint if we are running SMP kernel on a single non-MP
273 	 * approved Athlon
274 	 */
275 	WARN_ONCE(1, "WARNING: This combination of AMD"
276 		" processors is not suitable for SMP.\n");
277 	add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_NOW_UNRELIABLE);
278 #endif
279 }
280 
281 #ifdef CONFIG_NUMA
282 /*
283  * To workaround broken NUMA config.  Read the comment in
284  * srat_detect_node().
285  */
286 static int nearby_node(int apicid)
287 {
288 	int i, node;
289 
290 	for (i = apicid - 1; i >= 0; i--) {
291 		node = __apicid_to_node[i];
292 		if (node != NUMA_NO_NODE && node_online(node))
293 			return node;
294 	}
295 	for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) {
296 		node = __apicid_to_node[i];
297 		if (node != NUMA_NO_NODE && node_online(node))
298 			return node;
299 	}
300 	return first_node(node_online_map); /* Shouldn't happen */
301 }
302 #endif
303 
304 static void srat_detect_node(struct cpuinfo_x86 *c)
305 {
306 #ifdef CONFIG_NUMA
307 	int cpu = smp_processor_id();
308 	int node;
309 	unsigned apicid = c->topo.apicid;
310 
311 	node = numa_cpu_node(cpu);
312 	if (node == NUMA_NO_NODE)
313 		node = per_cpu_llc_id(cpu);
314 
315 	/*
316 	 * On multi-fabric platform (e.g. Numascale NumaChip) a
317 	 * platform-specific handler needs to be called to fixup some
318 	 * IDs of the CPU.
319 	 */
320 	if (x86_cpuinit.fixup_cpu_id)
321 		x86_cpuinit.fixup_cpu_id(c, node);
322 
323 	if (!node_online(node)) {
324 		/*
325 		 * Two possibilities here:
326 		 *
327 		 * - The CPU is missing memory and no node was created.  In
328 		 *   that case try picking one from a nearby CPU.
329 		 *
330 		 * - The APIC IDs differ from the HyperTransport node IDs
331 		 *   which the K8 northbridge parsing fills in.  Assume
332 		 *   they are all increased by a constant offset, but in
333 		 *   the same order as the HT nodeids.  If that doesn't
334 		 *   result in a usable node fall back to the path for the
335 		 *   previous case.
336 		 *
337 		 * This workaround operates directly on the mapping between
338 		 * APIC ID and NUMA node, assuming certain relationship
339 		 * between APIC ID, HT node ID and NUMA topology.  As going
340 		 * through CPU mapping may alter the outcome, directly
341 		 * access __apicid_to_node[].
342 		 */
343 		int ht_nodeid = c->topo.initial_apicid;
344 
345 		if (__apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
346 			node = __apicid_to_node[ht_nodeid];
347 		/* Pick a nearby node */
348 		if (!node_online(node))
349 			node = nearby_node(apicid);
350 	}
351 	numa_set_node(cpu, node);
352 #endif
353 }
354 
355 static void bsp_determine_snp(struct cpuinfo_x86 *c)
356 {
357 #ifdef CONFIG_ARCH_HAS_CC_PLATFORM
358 	cc_vendor = CC_VENDOR_AMD;
359 
360 	if (cpu_has(c, X86_FEATURE_SEV_SNP)) {
361 		/*
362 		 * RMP table entry format is not architectural and is defined by the
363 		 * per-processor PPR. Restrict SNP support on the known CPU models
364 		 * for which the RMP table entry format is currently defined or for
365 		 * processors which support the architecturally defined RMPREAD
366 		 * instruction.
367 		 */
368 		if (!cpu_has(c, X86_FEATURE_HYPERVISOR) &&
369 		    (cpu_feature_enabled(X86_FEATURE_ZEN3) ||
370 		     cpu_feature_enabled(X86_FEATURE_ZEN4) ||
371 		     cpu_feature_enabled(X86_FEATURE_RMPREAD)) &&
372 		    snp_probe_rmptable_info()) {
373 			cc_platform_set(CC_ATTR_HOST_SEV_SNP);
374 		} else {
375 			setup_clear_cpu_cap(X86_FEATURE_SEV_SNP);
376 			cc_platform_clear(CC_ATTR_HOST_SEV_SNP);
377 		}
378 	}
379 #endif
380 }
381 
382 #define ZEN_MODEL_STEP_UCODE(fam, model, step, ucode) \
383 	X86_MATCH_VFM_STEPS(VFM_MAKE(X86_VENDOR_AMD, fam, model), \
384 			    step, step, ucode)
385 
386 static const struct x86_cpu_id amd_tsa_microcode[] = {
387 	ZEN_MODEL_STEP_UCODE(0x19, 0x01, 0x1, 0x0a0011d7),
388 	ZEN_MODEL_STEP_UCODE(0x19, 0x01, 0x2, 0x0a00123b),
389 	ZEN_MODEL_STEP_UCODE(0x19, 0x08, 0x2, 0x0a00820d),
390 	ZEN_MODEL_STEP_UCODE(0x19, 0x11, 0x1, 0x0a10114c),
391 	ZEN_MODEL_STEP_UCODE(0x19, 0x11, 0x2, 0x0a10124c),
392 	ZEN_MODEL_STEP_UCODE(0x19, 0x18, 0x1, 0x0a108109),
393 	ZEN_MODEL_STEP_UCODE(0x19, 0x21, 0x0, 0x0a20102e),
394 	ZEN_MODEL_STEP_UCODE(0x19, 0x21, 0x2, 0x0a201211),
395 	ZEN_MODEL_STEP_UCODE(0x19, 0x44, 0x1, 0x0a404108),
396 	ZEN_MODEL_STEP_UCODE(0x19, 0x50, 0x0, 0x0a500012),
397 	ZEN_MODEL_STEP_UCODE(0x19, 0x61, 0x2, 0x0a60120a),
398 	ZEN_MODEL_STEP_UCODE(0x19, 0x74, 0x1, 0x0a704108),
399 	ZEN_MODEL_STEP_UCODE(0x19, 0x75, 0x2, 0x0a705208),
400 	ZEN_MODEL_STEP_UCODE(0x19, 0x78, 0x0, 0x0a708008),
401 	ZEN_MODEL_STEP_UCODE(0x19, 0x7c, 0x0, 0x0a70c008),
402 	ZEN_MODEL_STEP_UCODE(0x19, 0xa0, 0x2, 0x0aa00216),
403 	{},
404 };
405 
406 static void tsa_init(struct cpuinfo_x86 *c)
407 {
408 	if (cpu_has(c, X86_FEATURE_HYPERVISOR))
409 		return;
410 
411 	if (cpu_has(c, X86_FEATURE_ZEN3) ||
412 	    cpu_has(c, X86_FEATURE_ZEN4)) {
413 		if (x86_match_min_microcode_rev(amd_tsa_microcode))
414 			setup_force_cpu_cap(X86_FEATURE_VERW_CLEAR);
415 		else
416 			pr_debug("%s: current revision: 0x%x\n", __func__, c->microcode);
417 	} else {
418 		setup_force_cpu_cap(X86_FEATURE_TSA_SQ_NO);
419 		setup_force_cpu_cap(X86_FEATURE_TSA_L1_NO);
420 	}
421 }
422 
423 static void bsp_init_amd(struct cpuinfo_x86 *c)
424 {
425 	if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) {
426 
427 		if (c->x86 > 0x10 ||
428 		    (c->x86 == 0x10 && c->x86_model >= 0x2)) {
429 			u64 val;
430 
431 			rdmsrq(MSR_K7_HWCR, val);
432 			if (!(val & BIT(24)))
433 				pr_warn(FW_BUG "TSC doesn't count with P0 frequency!\n");
434 		}
435 	}
436 
437 	if (c->x86 == 0x15) {
438 		unsigned long upperbit;
439 		u32 cpuid, assoc;
440 
441 		cpuid	 = cpuid_edx(0x80000005);
442 		assoc	 = cpuid >> 16 & 0xff;
443 		upperbit = ((cpuid >> 24) << 10) / assoc;
444 
445 		va_align.mask	  = (upperbit - 1) & PAGE_MASK;
446 		va_align.flags    = ALIGN_VA_32 | ALIGN_VA_64;
447 
448 		/* A random value per boot for bit slice [12:upper_bit) */
449 		va_align.bits = get_random_u32() & va_align.mask;
450 	}
451 
452 	if (cpu_has(c, X86_FEATURE_MWAITX))
453 		use_mwaitx_delay();
454 
455 	if (!boot_cpu_has(X86_FEATURE_AMD_SSBD) &&
456 	    !boot_cpu_has(X86_FEATURE_VIRT_SSBD) &&
457 	    c->x86 >= 0x15 && c->x86 <= 0x17) {
458 		unsigned int bit;
459 
460 		switch (c->x86) {
461 		case 0x15: bit = 54; break;
462 		case 0x16: bit = 33; break;
463 		case 0x17: bit = 10; break;
464 		default: return;
465 		}
466 		/*
467 		 * Try to cache the base value so further operations can
468 		 * avoid RMW. If that faults, do not enable SSBD.
469 		 */
470 		if (!rdmsrq_safe(MSR_AMD64_LS_CFG, &x86_amd_ls_cfg_base)) {
471 			setup_force_cpu_cap(X86_FEATURE_LS_CFG_SSBD);
472 			setup_force_cpu_cap(X86_FEATURE_SSBD);
473 			x86_amd_ls_cfg_ssbd_mask = 1ULL << bit;
474 		}
475 	}
476 
477 	resctrl_cpu_detect(c);
478 
479 	/* Figure out Zen generations: */
480 	switch (c->x86) {
481 	case 0x17:
482 		switch (c->x86_model) {
483 		case 0x00 ... 0x2f:
484 		case 0x50 ... 0x5f:
485 			setup_force_cpu_cap(X86_FEATURE_ZEN1);
486 			break;
487 		case 0x30 ... 0x4f:
488 		case 0x60 ... 0x7f:
489 		case 0x90 ... 0x91:
490 		case 0xa0 ... 0xaf:
491 			setup_force_cpu_cap(X86_FEATURE_ZEN2);
492 			break;
493 		default:
494 			goto warn;
495 		}
496 		break;
497 
498 	case 0x19:
499 		switch (c->x86_model) {
500 		case 0x00 ... 0x0f:
501 		case 0x20 ... 0x5f:
502 			setup_force_cpu_cap(X86_FEATURE_ZEN3);
503 			break;
504 		case 0x10 ... 0x1f:
505 		case 0x60 ... 0xaf:
506 			setup_force_cpu_cap(X86_FEATURE_ZEN4);
507 			break;
508 		default:
509 			goto warn;
510 		}
511 		break;
512 
513 	case 0x1a:
514 		switch (c->x86_model) {
515 		case 0x00 ... 0x2f:
516 		case 0x40 ... 0x4f:
517 		case 0x60 ... 0x7f:
518 		case 0xd0 ... 0xd7:
519 			setup_force_cpu_cap(X86_FEATURE_ZEN5);
520 			break;
521 		case 0x50 ... 0x5f:
522 		case 0x80 ... 0xaf:
523 		case 0xc0 ... 0xcf:
524 		case 0xd8 ... 0xef:
525 			setup_force_cpu_cap(X86_FEATURE_ZEN6);
526 			break;
527 		default:
528 			goto warn;
529 		}
530 		break;
531 
532 	default:
533 		break;
534 	}
535 
536 	bsp_determine_snp(c);
537 	tsa_init(c);
538 
539 	if (cpu_has(c, X86_FEATURE_GP_ON_USER_CPUID))
540 		setup_force_cpu_cap(X86_FEATURE_CPUID_FAULT);
541 
542 	return;
543 
544 warn:
545 	WARN_ONCE(1, "Family 0x%x, model: 0x%x??\n", c->x86, c->x86_model);
546 }
547 
548 static void early_detect_mem_encrypt(struct cpuinfo_x86 *c)
549 {
550 	u64 msr;
551 
552 	/*
553 	 * Mark using WBINVD is needed during kexec on processors that
554 	 * support SME. This provides support for performing a successful
555 	 * kexec when going from SME inactive to SME active (or vice-versa).
556 	 *
557 	 * The cache must be cleared so that if there are entries with the
558 	 * same physical address, both with and without the encryption bit,
559 	 * they don't race each other when flushed and potentially end up
560 	 * with the wrong entry being committed to memory.
561 	 *
562 	 * Test the CPUID bit directly because with mem_encrypt=off the
563 	 * BSP will clear the X86_FEATURE_SME bit and the APs will not
564 	 * see it set after that.
565 	 */
566 	if (c->extended_cpuid_level >= 0x8000001f && (cpuid_eax(0x8000001f) & BIT(0)))
567 		__this_cpu_write(cache_state_incoherent, true);
568 
569 	/*
570 	 * BIOS support is required for SME and SEV.
571 	 *   For SME: If BIOS has enabled SME then adjust x86_phys_bits by
572 	 *	      the SME physical address space reduction value.
573 	 *	      If BIOS has not enabled SME then don't advertise the
574 	 *	      SME feature (set in scattered.c).
575 	 *	      If the kernel has not enabled SME via any means then
576 	 *	      don't advertise the SME feature.
577 	 *   For SEV: If BIOS has not enabled SEV then don't advertise SEV and
578 	 *	      any additional functionality based on it.
579 	 *
580 	 *   In all cases, since support for SME and SEV requires long mode,
581 	 *   don't advertise the feature under CONFIG_X86_32.
582 	 */
583 	if (cpu_has(c, X86_FEATURE_SME) || cpu_has(c, X86_FEATURE_SEV)) {
584 		/* Check if memory encryption is enabled */
585 		rdmsrq(MSR_AMD64_SYSCFG, msr);
586 		if (!(msr & MSR_AMD64_SYSCFG_MEM_ENCRYPT))
587 			goto clear_all;
588 
589 		/*
590 		 * Always adjust physical address bits. Even though this
591 		 * will be a value above 32-bits this is still done for
592 		 * CONFIG_X86_32 so that accurate values are reported.
593 		 */
594 		c->x86_phys_bits -= (cpuid_ebx(0x8000001f) >> 6) & 0x3f;
595 
596 		if (IS_ENABLED(CONFIG_X86_32))
597 			goto clear_all;
598 
599 		if (!sme_me_mask)
600 			setup_clear_cpu_cap(X86_FEATURE_SME);
601 
602 		rdmsrq(MSR_K7_HWCR, msr);
603 		if (!(msr & MSR_K7_HWCR_SMMLOCK))
604 			goto clear_sev;
605 
606 		return;
607 
608 clear_all:
609 		setup_clear_cpu_cap(X86_FEATURE_SME);
610 clear_sev:
611 		setup_clear_cpu_cap(X86_FEATURE_SEV);
612 		setup_clear_cpu_cap(X86_FEATURE_SEV_ES);
613 		setup_clear_cpu_cap(X86_FEATURE_SEV_SNP);
614 	}
615 }
616 
617 static void early_init_amd(struct cpuinfo_x86 *c)
618 {
619 	u32 dummy;
620 
621 	if (c->x86 >= 0xf)
622 		set_cpu_cap(c, X86_FEATURE_K8);
623 
624 	rdmsr_safe(MSR_AMD64_PATCH_LEVEL, &c->microcode, &dummy);
625 
626 	/*
627 	 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
628 	 * with P/T states and does not stop in deep C-states
629 	 */
630 	if (c->x86_power & (1 << 8)) {
631 		set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
632 		set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
633 	}
634 
635 	/* Bit 12 of 8000_0007 edx is accumulated power mechanism. */
636 	if (c->x86_power & BIT(12))
637 		set_cpu_cap(c, X86_FEATURE_ACC_POWER);
638 
639 	/* Bit 14 indicates the Runtime Average Power Limit interface. */
640 	if (c->x86_power & BIT(14))
641 		set_cpu_cap(c, X86_FEATURE_RAPL);
642 
643 #ifdef CONFIG_X86_64
644 	set_cpu_cap(c, X86_FEATURE_SYSCALL32);
645 #else
646 	/*  Set MTRR capability flag if appropriate */
647 	if (c->x86 == 5)
648 		if (c->x86_model == 13 || c->x86_model == 9 ||
649 		    (c->x86_model == 8 && c->x86_stepping >= 8))
650 			set_cpu_cap(c, X86_FEATURE_K6_MTRR);
651 #endif
652 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
653 	/*
654 	 * ApicID can always be treated as an 8-bit value for AMD APIC versions
655 	 * >= 0x10, but even old K8s came out of reset with version 0x10. So, we
656 	 * can safely set X86_FEATURE_EXTD_APICID unconditionally for families
657 	 * after 16h.
658 	 */
659 	if (boot_cpu_has(X86_FEATURE_APIC)) {
660 		if (c->x86 > 0x16)
661 			set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
662 		else if (c->x86 >= 0xf) {
663 			/* check CPU config space for extended APIC ID */
664 			unsigned int val;
665 
666 			val = read_pci_config(0, 24, 0, 0x68);
667 			if ((val >> 17 & 0x3) == 0x3)
668 				set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
669 		}
670 	}
671 #endif
672 
673 	/*
674 	 * This is only needed to tell the kernel whether to use VMCALL
675 	 * and VMMCALL.  VMMCALL is never executed except under virt, so
676 	 * we can set it unconditionally.
677 	 */
678 	set_cpu_cap(c, X86_FEATURE_VMMCALL);
679 
680 	/* F16h erratum 793, CVE-2013-6885 */
681 	if (c->x86 == 0x16 && c->x86_model <= 0xf)
682 		msr_set_bit(MSR_AMD64_LS_CFG, 15);
683 
684 	early_detect_mem_encrypt(c);
685 
686 	if (!cpu_has(c, X86_FEATURE_HYPERVISOR) && !cpu_has(c, X86_FEATURE_IBPB_BRTYPE)) {
687 		if (c->x86 == 0x17 && boot_cpu_has(X86_FEATURE_AMD_IBPB))
688 			setup_force_cpu_cap(X86_FEATURE_IBPB_BRTYPE);
689 		else if (c->x86 >= 0x19 && !wrmsrq_safe(MSR_IA32_PRED_CMD, PRED_CMD_SBPB)) {
690 			setup_force_cpu_cap(X86_FEATURE_IBPB_BRTYPE);
691 			setup_force_cpu_cap(X86_FEATURE_SBPB);
692 		}
693 	}
694 }
695 
696 static void init_amd_k8(struct cpuinfo_x86 *c)
697 {
698 	u32 level;
699 	u64 value;
700 
701 	/* On C+ stepping K8 rep microcode works well for copy/memset */
702 	level = cpuid_eax(1);
703 	if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58)
704 		set_cpu_cap(c, X86_FEATURE_REP_GOOD);
705 
706 	/*
707 	 * Some BIOSes incorrectly force this feature, but only K8 revision D
708 	 * (model = 0x14) and later actually support it.
709 	 * (AMD Erratum #110, docId: 25759).
710 	 */
711 	if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM) && !cpu_has(c, X86_FEATURE_HYPERVISOR)) {
712 		clear_cpu_cap(c, X86_FEATURE_LAHF_LM);
713 		if (!rdmsrq_amd_safe(0xc001100d, &value)) {
714 			value &= ~BIT_64(32);
715 			wrmsrq_amd_safe(0xc001100d, value);
716 		}
717 	}
718 
719 	if (!c->x86_model_id[0])
720 		strscpy(c->x86_model_id, "Hammer");
721 
722 #ifdef CONFIG_SMP
723 	/*
724 	 * Disable TLB flush filter by setting HWCR.FFDIS on K8
725 	 * bit 6 of msr C001_0015
726 	 *
727 	 * Errata 63 for SH-B3 steppings
728 	 * Errata 122 for all steppings (F+ have it disabled by default)
729 	 */
730 	msr_set_bit(MSR_K7_HWCR, 6);
731 #endif
732 	set_cpu_bug(c, X86_BUG_SWAPGS_FENCE);
733 
734 	/*
735 	 * Check models and steppings affected by erratum 400. This is
736 	 * used to select the proper idle routine and to enable the
737 	 * check whether the machine is affected in arch_post_acpi_subsys_init()
738 	 * which sets the X86_BUG_AMD_APIC_C1E bug depending on the MSR check.
739 	 */
740 	if (c->x86_model > 0x41 ||
741 	    (c->x86_model == 0x41 && c->x86_stepping >= 0x2))
742 		setup_force_cpu_bug(X86_BUG_AMD_E400);
743 }
744 
745 static void init_amd_gh(struct cpuinfo_x86 *c)
746 {
747 #ifdef CONFIG_MMCONF_FAM10H
748 	/* do this for boot cpu */
749 	if (c == &boot_cpu_data)
750 		check_enable_amd_mmconf_dmi();
751 
752 	fam10h_check_enable_mmcfg();
753 #endif
754 
755 	/*
756 	 * Disable GART TLB Walk Errors on Fam10h. We do this here because this
757 	 * is always needed when GART is enabled, even in a kernel which has no
758 	 * MCE support built in. BIOS should disable GartTlbWlk Errors already.
759 	 * If it doesn't, we do it here as suggested by the BKDG.
760 	 *
761 	 * Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=33012
762 	 */
763 	msr_set_bit(MSR_AMD64_MCx_MASK(4), 10);
764 
765 	/*
766 	 * On family 10h BIOS may not have properly enabled WC+ support, causing
767 	 * it to be converted to CD memtype. This may result in performance
768 	 * degradation for certain nested-paging guests. Prevent this conversion
769 	 * by clearing bit 24 in MSR_AMD64_BU_CFG2.
770 	 *
771 	 * NOTE: we want to use the _safe accessors so as not to #GP kvm
772 	 * guests on older kvm hosts.
773 	 */
774 	msr_clear_bit(MSR_AMD64_BU_CFG2, 24);
775 
776 	set_cpu_bug(c, X86_BUG_AMD_TLB_MMATCH);
777 
778 	/*
779 	 * Check models and steppings affected by erratum 400. This is
780 	 * used to select the proper idle routine and to enable the
781 	 * check whether the machine is affected in arch_post_acpi_subsys_init()
782 	 * which sets the X86_BUG_AMD_APIC_C1E bug depending on the MSR check.
783 	 */
784 	if (c->x86_model > 0x2 ||
785 	    (c->x86_model == 0x2 && c->x86_stepping >= 0x1))
786 		setup_force_cpu_bug(X86_BUG_AMD_E400);
787 }
788 
789 static void init_amd_ln(struct cpuinfo_x86 *c)
790 {
791 	/*
792 	 * Apply erratum 665 fix unconditionally so machines without a BIOS
793 	 * fix work.
794 	 */
795 	msr_set_bit(MSR_AMD64_DE_CFG, 31);
796 }
797 
798 static bool rdrand_force;
799 
800 static int __init rdrand_cmdline(char *str)
801 {
802 	if (!str)
803 		return -EINVAL;
804 
805 	if (!strcmp(str, "force"))
806 		rdrand_force = true;
807 	else
808 		return -EINVAL;
809 
810 	return 0;
811 }
812 early_param("rdrand", rdrand_cmdline);
813 
814 static void clear_rdrand_cpuid_bit(struct cpuinfo_x86 *c)
815 {
816 	/*
817 	 * Saving of the MSR used to hide the RDRAND support during
818 	 * suspend/resume is done by arch/x86/power/cpu.c, which is
819 	 * dependent on CONFIG_PM_SLEEP.
820 	 */
821 	if (!IS_ENABLED(CONFIG_PM_SLEEP))
822 		return;
823 
824 	/*
825 	 * The self-test can clear X86_FEATURE_RDRAND, so check for
826 	 * RDRAND support using the CPUID function directly.
827 	 */
828 	if (!(cpuid_ecx(1) & BIT(30)) || rdrand_force)
829 		return;
830 
831 	msr_clear_bit(MSR_AMD64_CPUID_FN_1, 62);
832 
833 	/*
834 	 * Verify that the CPUID change has occurred in case the kernel is
835 	 * running virtualized and the hypervisor doesn't support the MSR.
836 	 */
837 	if (cpuid_ecx(1) & BIT(30)) {
838 		pr_info_once("BIOS may not properly restore RDRAND after suspend, but hypervisor does not support hiding RDRAND via CPUID.\n");
839 		return;
840 	}
841 
842 	clear_cpu_cap(c, X86_FEATURE_RDRAND);
843 	pr_info_once("BIOS may not properly restore RDRAND after suspend, hiding RDRAND via CPUID. Use rdrand=force to reenable.\n");
844 }
845 
846 static void init_amd_jg(struct cpuinfo_x86 *c)
847 {
848 	/*
849 	 * Some BIOS implementations do not restore proper RDRAND support
850 	 * across suspend and resume. Check on whether to hide the RDRAND
851 	 * instruction support via CPUID.
852 	 */
853 	clear_rdrand_cpuid_bit(c);
854 }
855 
856 static void init_amd_bd(struct cpuinfo_x86 *c)
857 {
858 	u64 value;
859 
860 	/*
861 	 * The way access filter has a performance penalty on some workloads.
862 	 * Disable it on the affected CPUs.
863 	 */
864 	if ((c->x86_model >= 0x02) && (c->x86_model < 0x20)) {
865 		if (!rdmsrq_safe(MSR_F15H_IC_CFG, &value) && !(value & 0x1E)) {
866 			value |= 0x1E;
867 			wrmsrq_safe(MSR_F15H_IC_CFG, value);
868 		}
869 	}
870 
871 	/*
872 	 * Some BIOS implementations do not restore proper RDRAND support
873 	 * across suspend and resume. Check on whether to hide the RDRAND
874 	 * instruction support via CPUID.
875 	 */
876 	clear_rdrand_cpuid_bit(c);
877 }
878 
879 static const struct x86_cpu_id erratum_1386_microcode[] = {
880 	ZEN_MODEL_STEP_UCODE(0x17, 0x01, 0x2, 0x0800126e),
881 	ZEN_MODEL_STEP_UCODE(0x17, 0x31, 0x0, 0x08301052),
882 	{}
883 };
884 
885 static void fix_erratum_1386(struct cpuinfo_x86 *c)
886 {
887 	/*
888 	 * Work around Erratum 1386.  The XSAVES instruction malfunctions in
889 	 * certain circumstances on Zen1/2 uarch, and not all parts have had
890 	 * updated microcode at the time of writing (March 2023).
891 	 *
892 	 * Affected parts all have no supervisor XSAVE states, meaning that
893 	 * the XSAVEC instruction (which works fine) is equivalent.
894 	 *
895 	 * Clear the feature flag only on microcode revisions which
896 	 * don't have the fix.
897 	 */
898 	if (x86_match_min_microcode_rev(erratum_1386_microcode))
899 		return;
900 
901 	clear_cpu_cap(c, X86_FEATURE_XSAVES);
902 }
903 
904 void init_spectral_chicken(struct cpuinfo_x86 *c)
905 {
906 #ifdef CONFIG_MITIGATION_UNRET_ENTRY
907 	/*
908 	 * On Zen2 we offer this chicken (bit) on the altar of Speculation.
909 	 *
910 	 * This suppresses speculation from the middle of a basic block, i.e. it
911 	 * suppresses non-branch predictions.
912 	 */
913 	if (!cpu_has(c, X86_FEATURE_HYPERVISOR))
914 		msr_set_bit(MSR_ZEN2_SPECTRAL_CHICKEN, MSR_ZEN2_SPECTRAL_CHICKEN_BIT);
915 #endif
916 }
917 
918 static void init_amd_zen_common(void)
919 {
920 	setup_force_cpu_cap(X86_FEATURE_ZEN);
921 #ifdef CONFIG_NUMA
922 	node_reclaim_distance = 32;
923 #endif
924 }
925 
926 static void init_amd_zen1(struct cpuinfo_x86 *c)
927 {
928 	fix_erratum_1386(c);
929 
930 	/* Fix up CPUID bits, but only if not virtualised. */
931 	if (!cpu_has(c, X86_FEATURE_HYPERVISOR)) {
932 
933 		/* Erratum 1076: CPB feature bit not being set in CPUID. */
934 		if (!cpu_has(c, X86_FEATURE_CPB))
935 			set_cpu_cap(c, X86_FEATURE_CPB);
936 	}
937 
938 	pr_notice_once("AMD Zen1 DIV0 bug detected. Disable SMT for full protection.\n");
939 	setup_force_cpu_bug(X86_BUG_DIV0);
940 
941 	/*
942 	 * Turn off the Instructions Retired free counter on machines that are
943 	 * susceptible to erratum #1054 "Instructions Retired Performance
944 	 * Counter May Be Inaccurate".
945 	 */
946 	if (c->x86_model < 0x30) {
947 		msr_clear_bit(MSR_K7_HWCR, MSR_K7_HWCR_IRPERF_EN_BIT);
948 		clear_cpu_cap(c, X86_FEATURE_IRPERF);
949 	}
950 
951 	pr_notice_once("AMD Zen1 FPDSS bug detected, enabling mitigation.\n");
952 	msr_set_bit(MSR_AMD64_FP_CFG, MSR_AMD64_FP_CFG_ZEN1_DENORM_FIX_BIT);
953 }
954 
955 static const struct x86_cpu_id amd_zenbleed_microcode[] = {
956 	ZEN_MODEL_STEP_UCODE(0x17, 0x31, 0x0, 0x0830107b),
957 	ZEN_MODEL_STEP_UCODE(0x17, 0x60, 0x1, 0x0860010c),
958 	ZEN_MODEL_STEP_UCODE(0x17, 0x68, 0x1, 0x08608107),
959 	ZEN_MODEL_STEP_UCODE(0x17, 0x71, 0x0, 0x08701033),
960 	ZEN_MODEL_STEP_UCODE(0x17, 0xa0, 0x0, 0x08a00009),
961 	{}
962 };
963 
964 static void zen2_zenbleed_check(struct cpuinfo_x86 *c)
965 {
966 	if (cpu_has(c, X86_FEATURE_HYPERVISOR))
967 		return;
968 
969 	if (!cpu_has(c, X86_FEATURE_AVX))
970 		return;
971 
972 	if (!x86_match_min_microcode_rev(amd_zenbleed_microcode)) {
973 		pr_notice_once("Zenbleed: please update your microcode for the most optimal fix\n");
974 		msr_set_bit(MSR_AMD64_DE_CFG, MSR_AMD64_DE_CFG_ZEN2_FP_BACKUP_FIX_BIT);
975 	} else {
976 		msr_clear_bit(MSR_AMD64_DE_CFG, MSR_AMD64_DE_CFG_ZEN2_FP_BACKUP_FIX_BIT);
977 	}
978 }
979 
980 static void init_amd_zen2(struct cpuinfo_x86 *c)
981 {
982 	init_spectral_chicken(c);
983 	fix_erratum_1386(c);
984 	zen2_zenbleed_check(c);
985 
986 	/* Disable RDSEED on AMD Cyan Skillfish because of an error. */
987 	if (c->x86_model == 0x47 && c->x86_stepping == 0x0) {
988 		clear_cpu_cap(c, X86_FEATURE_RDSEED);
989 		msr_clear_bit(MSR_AMD64_CPUID_FN_7, 18);
990 		pr_emerg("RDSEED is not reliable on this platform; disabling.\n");
991 	}
992 
993 	/* Correct misconfigured CPUID on some clients. */
994 	clear_cpu_cap(c, X86_FEATURE_INVLPGB);
995 
996 	if (!cpu_has(c, X86_FEATURE_HYPERVISOR))
997 		msr_set_bit(MSR_ZEN4_BP_CFG, MSR_ZEN2_BP_CFG_BUG_FIX_BIT);
998 }
999 
1000 static void init_amd_zen3(struct cpuinfo_x86 *c)
1001 {
1002 	if (!cpu_has(c, X86_FEATURE_HYPERVISOR)) {
1003 		/*
1004 		 * Zen3 (Fam19 model < 0x10) parts are not susceptible to
1005 		 * Branch Type Confusion, but predate the allocation of the
1006 		 * BTC_NO bit.
1007 		 */
1008 		if (!cpu_has(c, X86_FEATURE_BTC_NO))
1009 			set_cpu_cap(c, X86_FEATURE_BTC_NO);
1010 	}
1011 }
1012 
1013 static void init_amd_zen4(struct cpuinfo_x86 *c)
1014 {
1015 	if (!cpu_has(c, X86_FEATURE_HYPERVISOR))
1016 		msr_set_bit(MSR_ZEN4_BP_CFG, MSR_ZEN4_BP_CFG_SHARED_BTB_FIX_BIT);
1017 
1018 	/*
1019 	 * These Zen4 SoCs advertise support for virtualized VMLOAD/VMSAVE
1020 	 * in some BIOS versions but they can lead to random host reboots.
1021 	 */
1022 	switch (c->x86_model) {
1023 	case 0x18 ... 0x1f:
1024 	case 0x60 ... 0x7f:
1025 		clear_cpu_cap(c, X86_FEATURE_V_VMSAVE_VMLOAD);
1026 		break;
1027 	}
1028 }
1029 
1030 static const struct x86_cpu_id zen5_rdseed_microcode[] = {
1031 	ZEN_MODEL_STEP_UCODE(0x1a, 0x02, 0x1, 0x0b00215a),
1032 	ZEN_MODEL_STEP_UCODE(0x1a, 0x08, 0x1, 0x0b008121),
1033 	ZEN_MODEL_STEP_UCODE(0x1a, 0x11, 0x0, 0x0b101054),
1034 	ZEN_MODEL_STEP_UCODE(0x1a, 0x24, 0x0, 0x0b204037),
1035 	ZEN_MODEL_STEP_UCODE(0x1a, 0x44, 0x0, 0x0b404035),
1036 	ZEN_MODEL_STEP_UCODE(0x1a, 0x44, 0x1, 0x0b404108),
1037 	ZEN_MODEL_STEP_UCODE(0x1a, 0x60, 0x0, 0x0b600037),
1038 	ZEN_MODEL_STEP_UCODE(0x1a, 0x68, 0x0, 0x0b608038),
1039 	ZEN_MODEL_STEP_UCODE(0x1a, 0x70, 0x0, 0x0b700037),
1040 	{},
1041 };
1042 
1043 static void init_amd_zen5(struct cpuinfo_x86 *c)
1044 {
1045 	if (!x86_match_min_microcode_rev(zen5_rdseed_microcode)) {
1046 		clear_cpu_cap(c, X86_FEATURE_RDSEED);
1047 		msr_clear_bit(MSR_AMD64_CPUID_FN_7, 18);
1048 		pr_emerg_once("RDSEED32 is broken. Disabling the corresponding CPUID bit.\n");
1049 	}
1050 }
1051 
1052 static void init_amd(struct cpuinfo_x86 *c)
1053 {
1054 	u64 vm_cr;
1055 
1056 	early_init_amd(c);
1057 
1058 	if (c->x86 >= 0x10)
1059 		set_cpu_cap(c, X86_FEATURE_REP_GOOD);
1060 
1061 	/* AMD FSRM also implies FSRS */
1062 	if (cpu_has(c, X86_FEATURE_FSRM))
1063 		set_cpu_cap(c, X86_FEATURE_FSRS);
1064 
1065 	/* K6s reports MCEs but don't actually have all the MSRs */
1066 	if (c->x86 < 6)
1067 		clear_cpu_cap(c, X86_FEATURE_MCE);
1068 
1069 	switch (c->x86) {
1070 	case 4:    init_amd_k5(c); break;
1071 	case 5:    init_amd_k6(c); break;
1072 	case 6:	   init_amd_k7(c); break;
1073 	case 0xf:  init_amd_k8(c); break;
1074 	case 0x10: init_amd_gh(c); break;
1075 	case 0x12: init_amd_ln(c); break;
1076 	case 0x15: init_amd_bd(c); break;
1077 	case 0x16: init_amd_jg(c); break;
1078 	}
1079 
1080 	/*
1081 	 * Save up on some future enablement work and do common Zen
1082 	 * settings.
1083 	 */
1084 	if (c->x86 >= 0x17)
1085 		init_amd_zen_common();
1086 
1087 	if (boot_cpu_has(X86_FEATURE_ZEN1))
1088 		init_amd_zen1(c);
1089 	else if (boot_cpu_has(X86_FEATURE_ZEN2))
1090 		init_amd_zen2(c);
1091 	else if (boot_cpu_has(X86_FEATURE_ZEN3))
1092 		init_amd_zen3(c);
1093 	else if (boot_cpu_has(X86_FEATURE_ZEN4))
1094 		init_amd_zen4(c);
1095 	else if (boot_cpu_has(X86_FEATURE_ZEN5))
1096 		init_amd_zen5(c);
1097 
1098 	/*
1099 	 * Enable workaround for FXSAVE leak on CPUs
1100 	 * without a XSaveErPtr feature
1101 	 */
1102 	if ((c->x86 >= 6) && (!cpu_has(c, X86_FEATURE_XSAVEERPTR)))
1103 		set_cpu_bug(c, X86_BUG_FXSAVE_LEAK);
1104 
1105 	cpu_detect_cache_sizes(c);
1106 
1107 	srat_detect_node(c);
1108 
1109 	init_amd_cacheinfo(c);
1110 
1111 	if (cpu_has(c, X86_FEATURE_SVM)) {
1112 		rdmsrq(MSR_VM_CR, vm_cr);
1113 		if (vm_cr & SVM_VM_CR_SVM_DIS_MASK) {
1114 			pr_notice_once("SVM disabled (by BIOS) in MSR_VM_CR\n");
1115 			clear_cpu_cap(c, X86_FEATURE_SVM);
1116 		}
1117 	}
1118 
1119 	if (!cpu_has(c, X86_FEATURE_LFENCE_RDTSC) && cpu_has(c, X86_FEATURE_XMM2)) {
1120 		/*
1121 		 * Use LFENCE for execution serialization.  On families which
1122 		 * don't have that MSR, LFENCE is already serializing.
1123 		 * msr_set_bit() uses the safe accessors, too, even if the MSR
1124 		 * is not present.
1125 		 */
1126 		msr_set_bit(MSR_AMD64_DE_CFG,
1127 			    MSR_AMD64_DE_CFG_LFENCE_SERIALIZE_BIT);
1128 
1129 		/* A serializing LFENCE stops RDTSC speculation */
1130 		set_cpu_cap(c, X86_FEATURE_LFENCE_RDTSC);
1131 	}
1132 
1133 	/*
1134 	 * Family 0x12 and above processors have APIC timer
1135 	 * running in deep C states.
1136 	 */
1137 	if (c->x86 > 0x11)
1138 		set_cpu_cap(c, X86_FEATURE_ARAT);
1139 
1140 	/* 3DNow or LM implies PREFETCHW */
1141 	if (!cpu_has(c, X86_FEATURE_3DNOWPREFETCH))
1142 		if (cpu_has(c, X86_FEATURE_3DNOW) || cpu_has(c, X86_FEATURE_LM))
1143 			set_cpu_cap(c, X86_FEATURE_3DNOWPREFETCH);
1144 
1145 	/* AMD CPUs don't reset SS attributes on SYSRET, Xen does. */
1146 	if (!cpu_feature_enabled(X86_FEATURE_XENPV))
1147 		set_cpu_bug(c, X86_BUG_SYSRET_SS_ATTRS);
1148 
1149 	/* Enable the Instructions Retired free counter */
1150 	if (cpu_has(c, X86_FEATURE_IRPERF))
1151 		msr_set_bit(MSR_K7_HWCR, MSR_K7_HWCR_IRPERF_EN_BIT);
1152 
1153 	check_null_seg_clears_base(c);
1154 
1155 	/*
1156 	 * Make sure EFER[AIBRSE - Automatic IBRS Enable] is set. The APs are brought up
1157 	 * using the trampoline code and as part of it, MSR_EFER gets prepared there in
1158 	 * order to be replicated onto them. Regardless, set it here again, if not set,
1159 	 * to protect against any future refactoring/code reorganization which might
1160 	 * miss setting this important bit.
1161 	 */
1162 	if (spectre_v2_in_eibrs_mode(spectre_v2_enabled) &&
1163 	    cpu_has(c, X86_FEATURE_AUTOIBRS))
1164 		WARN_ON_ONCE(msr_set_bit(MSR_EFER, _EFER_AUTOIBRS) < 0);
1165 
1166 	/* AMD CPUs don't need fencing after x2APIC/TSC_DEADLINE MSR writes. */
1167 	clear_cpu_cap(c, X86_FEATURE_APIC_MSRS_FENCE);
1168 
1169 	/* Enable Translation Cache Extension */
1170 	if (cpu_has(c, X86_FEATURE_TCE))
1171 		msr_set_bit(MSR_EFER, _EFER_TCE);
1172 }
1173 
1174 #ifdef CONFIG_X86_32
1175 static unsigned int amd_size_cache(struct cpuinfo_x86 *c, unsigned int size)
1176 {
1177 	/* AMD errata T13 (order #21922) */
1178 	if (c->x86 == 6) {
1179 		/* Duron Rev A0 */
1180 		if (c->x86_model == 3 && c->x86_stepping == 0)
1181 			size = 64;
1182 		/* Tbird rev A1/A2 */
1183 		if (c->x86_model == 4 &&
1184 			(c->x86_stepping == 0 || c->x86_stepping == 1))
1185 			size = 256;
1186 	}
1187 	return size;
1188 }
1189 #endif
1190 
1191 static void cpu_detect_tlb_amd(struct cpuinfo_x86 *c)
1192 {
1193 	u32 ebx, eax, ecx, edx;
1194 	u16 mask = 0xfff;
1195 
1196 	if (c->x86 < 0xf)
1197 		return;
1198 
1199 	if (c->extended_cpuid_level < 0x80000006)
1200 		return;
1201 
1202 	cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
1203 
1204 	tlb_lld_4k = (ebx >> 16) & mask;
1205 	tlb_lli_4k = ebx & mask;
1206 
1207 	/*
1208 	 * K8 doesn't have 2M/4M entries in the L2 TLB so read out the L1 TLB
1209 	 * characteristics from the CPUID function 0x80000005 instead.
1210 	 */
1211 	if (c->x86 == 0xf) {
1212 		cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
1213 		mask = 0xff;
1214 	}
1215 
1216 	/* Handle DTLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
1217 	if (!((eax >> 16) & mask))
1218 		tlb_lld_2m = (cpuid_eax(0x80000005) >> 16) & 0xff;
1219 	else
1220 		tlb_lld_2m = (eax >> 16) & mask;
1221 
1222 	/* a 4M entry uses two 2M entries */
1223 	tlb_lld_4m = tlb_lld_2m >> 1;
1224 
1225 	/* Handle ITLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
1226 	if (!(eax & mask)) {
1227 		/* Erratum 658 */
1228 		if (c->x86 == 0x15 && c->x86_model <= 0x1f) {
1229 			tlb_lli_2m = 1024;
1230 		} else {
1231 			cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
1232 			tlb_lli_2m = eax & 0xff;
1233 		}
1234 	} else
1235 		tlb_lli_2m = eax & mask;
1236 
1237 	tlb_lli_4m = tlb_lli_2m >> 1;
1238 
1239 	/* Max number of pages INVLPGB can invalidate in one shot */
1240 	if (cpu_has(c, X86_FEATURE_INVLPGB))
1241 		invlpgb_count_max = (cpuid_edx(0x80000008) & 0xffff) + 1;
1242 }
1243 
1244 static const struct cpu_dev amd_cpu_dev = {
1245 	.c_vendor	= "AMD",
1246 	.c_ident	= { "AuthenticAMD" },
1247 #ifdef CONFIG_X86_32
1248 	.legacy_models = {
1249 		{ .family = 4, .model_names =
1250 		  {
1251 			  [3] = "486 DX/2",
1252 			  [7] = "486 DX/2-WB",
1253 			  [8] = "486 DX/4",
1254 			  [9] = "486 DX/4-WB",
1255 			  [14] = "Am5x86-WT",
1256 			  [15] = "Am5x86-WB"
1257 		  }
1258 		},
1259 	},
1260 	.legacy_cache_size = amd_size_cache,
1261 #endif
1262 	.c_early_init   = early_init_amd,
1263 	.c_detect_tlb	= cpu_detect_tlb_amd,
1264 	.c_bsp_init	= bsp_init_amd,
1265 	.c_init		= init_amd,
1266 	.c_x86_vendor	= X86_VENDOR_AMD,
1267 };
1268 
1269 cpu_dev_register(amd_cpu_dev);
1270 
1271 static DEFINE_PER_CPU_READ_MOSTLY(unsigned long[4], amd_dr_addr_mask);
1272 
1273 static unsigned int amd_msr_dr_addr_masks[] = {
1274 	MSR_F16H_DR0_ADDR_MASK,
1275 	MSR_F16H_DR1_ADDR_MASK,
1276 	MSR_F16H_DR1_ADDR_MASK + 1,
1277 	MSR_F16H_DR1_ADDR_MASK + 2
1278 };
1279 
1280 void amd_set_dr_addr_mask(unsigned long mask, unsigned int dr)
1281 {
1282 	int cpu = smp_processor_id();
1283 
1284 	if (!cpu_feature_enabled(X86_FEATURE_BPEXT))
1285 		return;
1286 
1287 	if (WARN_ON_ONCE(dr >= ARRAY_SIZE(amd_msr_dr_addr_masks)))
1288 		return;
1289 
1290 	if (per_cpu(amd_dr_addr_mask, cpu)[dr] == mask)
1291 		return;
1292 
1293 	wrmsrq(amd_msr_dr_addr_masks[dr], mask);
1294 	per_cpu(amd_dr_addr_mask, cpu)[dr] = mask;
1295 }
1296 
1297 unsigned long amd_get_dr_addr_mask(unsigned int dr)
1298 {
1299 	if (!cpu_feature_enabled(X86_FEATURE_BPEXT))
1300 		return 0;
1301 
1302 	if (WARN_ON_ONCE(dr >= ARRAY_SIZE(amd_msr_dr_addr_masks)))
1303 		return 0;
1304 
1305 	return per_cpu(amd_dr_addr_mask[dr], smp_processor_id());
1306 }
1307 EXPORT_SYMBOL_FOR_KVM(amd_get_dr_addr_mask);
1308 
1309 static void zenbleed_check_cpu(void *unused)
1310 {
1311 	struct cpuinfo_x86 *c = &cpu_data(smp_processor_id());
1312 
1313 	zen2_zenbleed_check(c);
1314 }
1315 
1316 void amd_check_microcode(void)
1317 {
1318 	if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD)
1319 		return;
1320 
1321 	if (cpu_feature_enabled(X86_FEATURE_ZEN2))
1322 		on_each_cpu(zenbleed_check_cpu, NULL, 1);
1323 }
1324 
1325 static const char * const s5_reset_reason_txt[] = {
1326 	[0]  = "thermal pin BP_THERMTRIP_L was tripped",
1327 	[1]  = "power button was pressed for 4 seconds",
1328 	[2]  = "shutdown pin was tripped",
1329 	[4]  = "remote ASF power off command was received",
1330 	[9]  = "internal CPU thermal limit was tripped",
1331 	[16] = "system reset pin BP_SYS_RST_L was tripped",
1332 	[17] = "software issued PCI reset",
1333 	[18] = "software wrote 0x4 to reset control register 0xCF9",
1334 	[19] = "software wrote 0x6 to reset control register 0xCF9",
1335 	[20] = "software wrote 0xE to reset control register 0xCF9",
1336 	[21] = "ACPI power state transition occurred",
1337 	[22] = "keyboard reset pin KB_RST_L was tripped",
1338 	[23] = "internal CPU shutdown event occurred",
1339 	[24] = "system failed to boot before failed boot timer expired",
1340 	[25] = "hardware watchdog timer expired",
1341 	[26] = "remote ASF reset command was received",
1342 	[27] = "an uncorrected error caused a data fabric sync flood event",
1343 	[29] = "FCH and MP1 failed warm reset handshake",
1344 	[30] = "a parity error occurred",
1345 	[31] = "a software sync flood event occurred",
1346 };
1347 
1348 static __init int print_s5_reset_status_mmio(void)
1349 {
1350 	void __iomem *addr;
1351 	u32 value;
1352 	int i;
1353 
1354 	if (!cpu_feature_enabled(X86_FEATURE_ZEN))
1355 		return 0;
1356 
1357 	addr = ioremap(FCH_PM_BASE + FCH_PM_S5_RESET_STATUS, sizeof(value));
1358 	if (!addr)
1359 		return 0;
1360 
1361 	value = ioread32(addr);
1362 
1363 	/* Value with "all bits set" is an error response and should be ignored. */
1364 	if (value == U32_MAX) {
1365 		iounmap(addr);
1366 		return 0;
1367 	}
1368 
1369 	/*
1370 	 * Clear all reason bits so they won't be retained if the next reset
1371 	 * does not update the register. Besides, some bits are never cleared by
1372 	 * hardware so it's software's responsibility to clear them.
1373 	 *
1374 	 * Writing the value back effectively clears all reason bits as they are
1375 	 * write-1-to-clear.
1376 	 */
1377 	iowrite32(value, addr);
1378 	iounmap(addr);
1379 
1380 	for (i = 0; i < ARRAY_SIZE(s5_reset_reason_txt); i++) {
1381 		if (!(value & BIT(i)))
1382 			continue;
1383 
1384 		if (s5_reset_reason_txt[i]) {
1385 			pr_info("x86/amd: Previous system reset reason [0x%08x]: %s\n",
1386 				value, s5_reset_reason_txt[i]);
1387 		}
1388 	}
1389 
1390 	return 0;
1391 }
1392 late_initcall(print_s5_reset_status_mmio);
1393 
1394 static void __init dmi_scan_additional(const struct dmi_header *d, void *p)
1395 {
1396 	struct dmi_a_info *info = (struct dmi_a_info *)d;
1397 	void *next, *end;
1398 
1399 	if (!IS_ENABLED(CONFIG_DMI))
1400 		return;
1401 
1402 	if (info->header.type != DMI_ENTRY_ADDITIONAL ||
1403 	    info->header.length < DMI_A_INFO_MIN_SIZE ||
1404 	    info->count < 1)
1405 		return;
1406 
1407 	next = (void *)(info + 1);
1408 	end  = (void *)info + info->header.length;
1409 
1410 	do {
1411 		struct dmi_a_info_entry *entry;
1412 		const char *string_ptr;
1413 
1414 		entry = (struct dmi_a_info_entry *)next;
1415 
1416 		/*
1417 		 * Not much can be done to validate data. At least the entry
1418 		 * length shouldn't be 0.
1419 		 */
1420 		if (!entry->length)
1421 			return;
1422 
1423 		string_ptr = dmi_string_nosave(&info->header, entry->str_num);
1424 
1425 		/* Sample string: AGESA!V9 StrixKrackanPI-FP8 1.1.0.0c */
1426 		if (!strncmp(string_ptr, "AGESA", 5)) {
1427 			pr_info("AGESA: %s\n", string_ptr);
1428 			break;
1429 		}
1430 
1431 		next += entry->length;
1432 	} while (end - next >= DMI_A_INFO_ENT_MIN_SIZE);
1433 }
1434 
1435 static __init int print_dmi_agesa(void)
1436 {
1437 	dmi_walk(dmi_scan_additional, NULL);
1438 	return 0;
1439 }
1440 late_initcall(print_dmi_agesa);
1441