xref: /linux/arch/x86/kernel/cpu/amd.c (revision df46426745ac58618c822ce3f93d2bb25b9a5060)
1 // SPDX-License-Identifier: GPL-2.0-only
2 #include <linux/export.h>
3 #include <linux/bitops.h>
4 #include <linux/elf.h>
5 #include <linux/mm.h>
6 
7 #include <linux/io.h>
8 #include <linux/sched.h>
9 #include <linux/sched/clock.h>
10 #include <linux/random.h>
11 #include <linux/topology.h>
12 #include <linux/platform_data/x86/amd-fch.h>
13 #include <asm/processor.h>
14 #include <asm/apic.h>
15 #include <asm/cacheinfo.h>
16 #include <asm/cpu.h>
17 #include <asm/cpu_device_id.h>
18 #include <asm/spec-ctrl.h>
19 #include <asm/smp.h>
20 #include <asm/numa.h>
21 #include <asm/pci-direct.h>
22 #include <asm/delay.h>
23 #include <asm/debugreg.h>
24 #include <asm/resctrl.h>
25 #include <asm/msr.h>
26 #include <asm/sev.h>
27 
28 #ifdef CONFIG_X86_64
29 # include <asm/mmconfig.h>
30 #endif
31 
32 #include "cpu.h"
33 
34 u16 invlpgb_count_max __ro_after_init = 1;
35 
rdmsrq_amd_safe(unsigned msr,u64 * p)36 static inline int rdmsrq_amd_safe(unsigned msr, u64 *p)
37 {
38 	u32 gprs[8] = { 0 };
39 	int err;
40 
41 	WARN_ONCE((boot_cpu_data.x86 != 0xf),
42 		  "%s should only be used on K8!\n", __func__);
43 
44 	gprs[1] = msr;
45 	gprs[7] = 0x9c5a203a;
46 
47 	err = rdmsr_safe_regs(gprs);
48 
49 	*p = gprs[0] | ((u64)gprs[2] << 32);
50 
51 	return err;
52 }
53 
wrmsrq_amd_safe(unsigned msr,u64 val)54 static inline int wrmsrq_amd_safe(unsigned msr, u64 val)
55 {
56 	u32 gprs[8] = { 0 };
57 
58 	WARN_ONCE((boot_cpu_data.x86 != 0xf),
59 		  "%s should only be used on K8!\n", __func__);
60 
61 	gprs[0] = (u32)val;
62 	gprs[1] = msr;
63 	gprs[2] = val >> 32;
64 	gprs[7] = 0x9c5a203a;
65 
66 	return wrmsr_safe_regs(gprs);
67 }
68 
69 /*
70  *	B step AMD K6 before B 9730xxxx have hardware bugs that can cause
71  *	misexecution of code under Linux. Owners of such processors should
72  *	contact AMD for precise details and a CPU swap.
73  *
74  *	See	http://www.multimania.com/poulot/k6bug.html
75  *	and	section 2.6.2 of "AMD-K6 Processor Revision Guide - Model 6"
76  *		(Publication # 21266  Issue Date: August 1998)
77  *
78  *	The following test is erm.. interesting. AMD neglected to up
79  *	the chip setting when fixing the bug but they also tweaked some
80  *	performance at the same time..
81  */
82 
83 #ifdef CONFIG_X86_32
84 extern __visible void vide(void);
85 __asm__(".text\n"
86 	".globl vide\n"
87 	".type vide, @function\n"
88 	".align 4\n"
89 	"vide: ret\n");
90 #endif
91 
init_amd_k5(struct cpuinfo_x86 * c)92 static void init_amd_k5(struct cpuinfo_x86 *c)
93 {
94 #ifdef CONFIG_X86_32
95 /*
96  * General Systems BIOSen alias the cpu frequency registers
97  * of the Elan at 0x000df000. Unfortunately, one of the Linux
98  * drivers subsequently pokes it, and changes the CPU speed.
99  * Workaround : Remove the unneeded alias.
100  */
101 #define CBAR		(0xfffc) /* Configuration Base Address  (32-bit) */
102 #define CBAR_ENB	(0x80000000)
103 #define CBAR_KEY	(0X000000CB)
104 	if (c->x86_model == 9 || c->x86_model == 10) {
105 		if (inl(CBAR) & CBAR_ENB)
106 			outl(0 | CBAR_KEY, CBAR);
107 	}
108 #endif
109 }
110 
init_amd_k6(struct cpuinfo_x86 * c)111 static void init_amd_k6(struct cpuinfo_x86 *c)
112 {
113 #ifdef CONFIG_X86_32
114 	u32 l, h;
115 	int mbytes = get_num_physpages() >> (20-PAGE_SHIFT);
116 
117 	if (c->x86_model < 6) {
118 		/* Based on AMD doc 20734R - June 2000 */
119 		if (c->x86_model == 0) {
120 			clear_cpu_cap(c, X86_FEATURE_APIC);
121 			set_cpu_cap(c, X86_FEATURE_PGE);
122 		}
123 		return;
124 	}
125 
126 	if (c->x86_model == 6 && c->x86_stepping == 1) {
127 		const int K6_BUG_LOOP = 1000000;
128 		int n;
129 		void (*f_vide)(void);
130 		u64 d, d2;
131 
132 		pr_info("AMD K6 stepping B detected - ");
133 
134 		/*
135 		 * It looks like AMD fixed the 2.6.2 bug and improved indirect
136 		 * calls at the same time.
137 		 */
138 
139 		n = K6_BUG_LOOP;
140 		f_vide = vide;
141 		OPTIMIZER_HIDE_VAR(f_vide);
142 		d = rdtsc();
143 		while (n--)
144 			f_vide();
145 		d2 = rdtsc();
146 		d = d2-d;
147 
148 		if (d > 20*K6_BUG_LOOP)
149 			pr_cont("system stability may be impaired when more than 32 MB are used.\n");
150 		else
151 			pr_cont("probably OK (after B9730xxxx).\n");
152 	}
153 
154 	/* K6 with old style WHCR */
155 	if (c->x86_model < 8 ||
156 	   (c->x86_model == 8 && c->x86_stepping < 8)) {
157 		/* We can only write allocate on the low 508Mb */
158 		if (mbytes > 508)
159 			mbytes = 508;
160 
161 		rdmsr(MSR_K6_WHCR, l, h);
162 		if ((l&0x0000FFFF) == 0) {
163 			unsigned long flags;
164 			l = (1<<0)|((mbytes/4)<<1);
165 			local_irq_save(flags);
166 			wbinvd();
167 			wrmsr(MSR_K6_WHCR, l, h);
168 			local_irq_restore(flags);
169 			pr_info("Enabling old style K6 write allocation for %d Mb\n",
170 				mbytes);
171 		}
172 		return;
173 	}
174 
175 	if ((c->x86_model == 8 && c->x86_stepping > 7) ||
176 	     c->x86_model == 9 || c->x86_model == 13) {
177 		/* The more serious chips .. */
178 
179 		if (mbytes > 4092)
180 			mbytes = 4092;
181 
182 		rdmsr(MSR_K6_WHCR, l, h);
183 		if ((l&0xFFFF0000) == 0) {
184 			unsigned long flags;
185 			l = ((mbytes>>2)<<22)|(1<<16);
186 			local_irq_save(flags);
187 			wbinvd();
188 			wrmsr(MSR_K6_WHCR, l, h);
189 			local_irq_restore(flags);
190 			pr_info("Enabling new style K6 write allocation for %d Mb\n",
191 				mbytes);
192 		}
193 
194 		return;
195 	}
196 
197 	if (c->x86_model == 10) {
198 		/* AMD Geode LX is model 10 */
199 		/* placeholder for any needed mods */
200 		return;
201 	}
202 #endif
203 }
204 
init_amd_k7(struct cpuinfo_x86 * c)205 static void init_amd_k7(struct cpuinfo_x86 *c)
206 {
207 #ifdef CONFIG_X86_32
208 	u32 l, h;
209 
210 	/*
211 	 * Bit 15 of Athlon specific MSR 15, needs to be 0
212 	 * to enable SSE on Palomino/Morgan/Barton CPU's.
213 	 * If the BIOS didn't enable it already, enable it here.
214 	 */
215 	if (c->x86_model >= 6 && c->x86_model <= 10) {
216 		if (!cpu_has(c, X86_FEATURE_XMM)) {
217 			pr_info("Enabling disabled K7/SSE Support.\n");
218 			msr_clear_bit(MSR_K7_HWCR, 15);
219 			set_cpu_cap(c, X86_FEATURE_XMM);
220 		}
221 	}
222 
223 	/*
224 	 * It's been determined by AMD that Athlons since model 8 stepping 1
225 	 * are more robust with CLK_CTL set to 200xxxxx instead of 600xxxxx
226 	 * As per AMD technical note 27212 0.2
227 	 */
228 	if ((c->x86_model == 8 && c->x86_stepping >= 1) || (c->x86_model > 8)) {
229 		rdmsr(MSR_K7_CLK_CTL, l, h);
230 		if ((l & 0xfff00000) != 0x20000000) {
231 			pr_info("CPU: CLK_CTL MSR was %x. Reprogramming to %x\n",
232 				l, ((l & 0x000fffff)|0x20000000));
233 			wrmsr(MSR_K7_CLK_CTL, (l & 0x000fffff)|0x20000000, h);
234 		}
235 	}
236 
237 	/* calling is from identify_secondary_cpu() ? */
238 	if (!c->cpu_index)
239 		return;
240 
241 	/*
242 	 * Certain Athlons might work (for various values of 'work') in SMP
243 	 * but they are not certified as MP capable.
244 	 */
245 	/* Athlon 660/661 is valid. */
246 	if ((c->x86_model == 6) && ((c->x86_stepping == 0) ||
247 	    (c->x86_stepping == 1)))
248 		return;
249 
250 	/* Duron 670 is valid */
251 	if ((c->x86_model == 7) && (c->x86_stepping == 0))
252 		return;
253 
254 	/*
255 	 * Athlon 662, Duron 671, and Athlon >model 7 have capability
256 	 * bit. It's worth noting that the A5 stepping (662) of some
257 	 * Athlon XP's have the MP bit set.
258 	 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for
259 	 * more.
260 	 */
261 	if (((c->x86_model == 6) && (c->x86_stepping >= 2)) ||
262 	    ((c->x86_model == 7) && (c->x86_stepping >= 1)) ||
263 	     (c->x86_model > 7))
264 		if (cpu_has(c, X86_FEATURE_MP))
265 			return;
266 
267 	/* If we get here, not a certified SMP capable AMD system. */
268 
269 	/*
270 	 * Don't taint if we are running SMP kernel on a single non-MP
271 	 * approved Athlon
272 	 */
273 	WARN_ONCE(1, "WARNING: This combination of AMD"
274 		" processors is not suitable for SMP.\n");
275 	add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_NOW_UNRELIABLE);
276 #endif
277 }
278 
279 #ifdef CONFIG_NUMA
280 /*
281  * To workaround broken NUMA config.  Read the comment in
282  * srat_detect_node().
283  */
nearby_node(int apicid)284 static int nearby_node(int apicid)
285 {
286 	int i, node;
287 
288 	for (i = apicid - 1; i >= 0; i--) {
289 		node = __apicid_to_node[i];
290 		if (node != NUMA_NO_NODE && node_online(node))
291 			return node;
292 	}
293 	for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) {
294 		node = __apicid_to_node[i];
295 		if (node != NUMA_NO_NODE && node_online(node))
296 			return node;
297 	}
298 	return first_node(node_online_map); /* Shouldn't happen */
299 }
300 #endif
301 
srat_detect_node(struct cpuinfo_x86 * c)302 static void srat_detect_node(struct cpuinfo_x86 *c)
303 {
304 #ifdef CONFIG_NUMA
305 	int cpu = smp_processor_id();
306 	int node;
307 	unsigned apicid = c->topo.apicid;
308 
309 	node = numa_cpu_node(cpu);
310 	if (node == NUMA_NO_NODE)
311 		node = per_cpu_llc_id(cpu);
312 
313 	/*
314 	 * On multi-fabric platform (e.g. Numascale NumaChip) a
315 	 * platform-specific handler needs to be called to fixup some
316 	 * IDs of the CPU.
317 	 */
318 	if (x86_cpuinit.fixup_cpu_id)
319 		x86_cpuinit.fixup_cpu_id(c, node);
320 
321 	if (!node_online(node)) {
322 		/*
323 		 * Two possibilities here:
324 		 *
325 		 * - The CPU is missing memory and no node was created.  In
326 		 *   that case try picking one from a nearby CPU.
327 		 *
328 		 * - The APIC IDs differ from the HyperTransport node IDs
329 		 *   which the K8 northbridge parsing fills in.  Assume
330 		 *   they are all increased by a constant offset, but in
331 		 *   the same order as the HT nodeids.  If that doesn't
332 		 *   result in a usable node fall back to the path for the
333 		 *   previous case.
334 		 *
335 		 * This workaround operates directly on the mapping between
336 		 * APIC ID and NUMA node, assuming certain relationship
337 		 * between APIC ID, HT node ID and NUMA topology.  As going
338 		 * through CPU mapping may alter the outcome, directly
339 		 * access __apicid_to_node[].
340 		 */
341 		int ht_nodeid = c->topo.initial_apicid;
342 
343 		if (__apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
344 			node = __apicid_to_node[ht_nodeid];
345 		/* Pick a nearby node */
346 		if (!node_online(node))
347 			node = nearby_node(apicid);
348 	}
349 	numa_set_node(cpu, node);
350 #endif
351 }
352 
bsp_determine_snp(struct cpuinfo_x86 * c)353 static void bsp_determine_snp(struct cpuinfo_x86 *c)
354 {
355 #ifdef CONFIG_ARCH_HAS_CC_PLATFORM
356 	cc_vendor = CC_VENDOR_AMD;
357 
358 	if (cpu_has(c, X86_FEATURE_SEV_SNP)) {
359 		/*
360 		 * RMP table entry format is not architectural and is defined by the
361 		 * per-processor PPR. Restrict SNP support on the known CPU models
362 		 * for which the RMP table entry format is currently defined or for
363 		 * processors which support the architecturally defined RMPREAD
364 		 * instruction.
365 		 */
366 		if (!cpu_has(c, X86_FEATURE_HYPERVISOR) &&
367 		    (cpu_feature_enabled(X86_FEATURE_ZEN3) ||
368 		     cpu_feature_enabled(X86_FEATURE_ZEN4) ||
369 		     cpu_feature_enabled(X86_FEATURE_RMPREAD)) &&
370 		    snp_probe_rmptable_info()) {
371 			cc_platform_set(CC_ATTR_HOST_SEV_SNP);
372 		} else {
373 			setup_clear_cpu_cap(X86_FEATURE_SEV_SNP);
374 			cc_platform_clear(CC_ATTR_HOST_SEV_SNP);
375 		}
376 	}
377 #endif
378 }
379 
bsp_init_amd(struct cpuinfo_x86 * c)380 static void bsp_init_amd(struct cpuinfo_x86 *c)
381 {
382 	if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) {
383 
384 		if (c->x86 > 0x10 ||
385 		    (c->x86 == 0x10 && c->x86_model >= 0x2)) {
386 			u64 val;
387 
388 			rdmsrq(MSR_K7_HWCR, val);
389 			if (!(val & BIT(24)))
390 				pr_warn(FW_BUG "TSC doesn't count with P0 frequency!\n");
391 		}
392 	}
393 
394 	if (c->x86 == 0x15) {
395 		unsigned long upperbit;
396 		u32 cpuid, assoc;
397 
398 		cpuid	 = cpuid_edx(0x80000005);
399 		assoc	 = cpuid >> 16 & 0xff;
400 		upperbit = ((cpuid >> 24) << 10) / assoc;
401 
402 		va_align.mask	  = (upperbit - 1) & PAGE_MASK;
403 		va_align.flags    = ALIGN_VA_32 | ALIGN_VA_64;
404 
405 		/* A random value per boot for bit slice [12:upper_bit) */
406 		va_align.bits = get_random_u32() & va_align.mask;
407 	}
408 
409 	if (cpu_has(c, X86_FEATURE_MWAITX))
410 		use_mwaitx_delay();
411 
412 	if (!boot_cpu_has(X86_FEATURE_AMD_SSBD) &&
413 	    !boot_cpu_has(X86_FEATURE_VIRT_SSBD) &&
414 	    c->x86 >= 0x15 && c->x86 <= 0x17) {
415 		unsigned int bit;
416 
417 		switch (c->x86) {
418 		case 0x15: bit = 54; break;
419 		case 0x16: bit = 33; break;
420 		case 0x17: bit = 10; break;
421 		default: return;
422 		}
423 		/*
424 		 * Try to cache the base value so further operations can
425 		 * avoid RMW. If that faults, do not enable SSBD.
426 		 */
427 		if (!rdmsrq_safe(MSR_AMD64_LS_CFG, &x86_amd_ls_cfg_base)) {
428 			setup_force_cpu_cap(X86_FEATURE_LS_CFG_SSBD);
429 			setup_force_cpu_cap(X86_FEATURE_SSBD);
430 			x86_amd_ls_cfg_ssbd_mask = 1ULL << bit;
431 		}
432 	}
433 
434 	resctrl_cpu_detect(c);
435 
436 	/* Figure out Zen generations: */
437 	switch (c->x86) {
438 	case 0x17:
439 		switch (c->x86_model) {
440 		case 0x00 ... 0x2f:
441 		case 0x50 ... 0x5f:
442 			setup_force_cpu_cap(X86_FEATURE_ZEN1);
443 			break;
444 		case 0x30 ... 0x4f:
445 		case 0x60 ... 0x7f:
446 		case 0x90 ... 0x91:
447 		case 0xa0 ... 0xaf:
448 			setup_force_cpu_cap(X86_FEATURE_ZEN2);
449 			break;
450 		default:
451 			goto warn;
452 		}
453 		break;
454 
455 	case 0x19:
456 		switch (c->x86_model) {
457 		case 0x00 ... 0x0f:
458 		case 0x20 ... 0x5f:
459 			setup_force_cpu_cap(X86_FEATURE_ZEN3);
460 			break;
461 		case 0x10 ... 0x1f:
462 		case 0x60 ... 0xaf:
463 			setup_force_cpu_cap(X86_FEATURE_ZEN4);
464 			break;
465 		default:
466 			goto warn;
467 		}
468 		break;
469 
470 	case 0x1a:
471 		switch (c->x86_model) {
472 		case 0x00 ... 0x2f:
473 		case 0x40 ... 0x4f:
474 		case 0x60 ... 0x7f:
475 			setup_force_cpu_cap(X86_FEATURE_ZEN5);
476 			break;
477 		case 0x50 ... 0x5f:
478 		case 0x90 ... 0xaf:
479 		case 0xc0 ... 0xcf:
480 			setup_force_cpu_cap(X86_FEATURE_ZEN6);
481 			break;
482 		default:
483 			goto warn;
484 		}
485 		break;
486 
487 	default:
488 		break;
489 	}
490 
491 	bsp_determine_snp(c);
492 	return;
493 
494 warn:
495 	WARN_ONCE(1, "Family 0x%x, model: 0x%x??\n", c->x86, c->x86_model);
496 }
497 
early_detect_mem_encrypt(struct cpuinfo_x86 * c)498 static void early_detect_mem_encrypt(struct cpuinfo_x86 *c)
499 {
500 	u64 msr;
501 
502 	/*
503 	 * BIOS support is required for SME and SEV.
504 	 *   For SME: If BIOS has enabled SME then adjust x86_phys_bits by
505 	 *	      the SME physical address space reduction value.
506 	 *	      If BIOS has not enabled SME then don't advertise the
507 	 *	      SME feature (set in scattered.c).
508 	 *	      If the kernel has not enabled SME via any means then
509 	 *	      don't advertise the SME feature.
510 	 *   For SEV: If BIOS has not enabled SEV then don't advertise SEV and
511 	 *	      any additional functionality based on it.
512 	 *
513 	 *   In all cases, since support for SME and SEV requires long mode,
514 	 *   don't advertise the feature under CONFIG_X86_32.
515 	 */
516 	if (cpu_has(c, X86_FEATURE_SME) || cpu_has(c, X86_FEATURE_SEV)) {
517 		/* Check if memory encryption is enabled */
518 		rdmsrq(MSR_AMD64_SYSCFG, msr);
519 		if (!(msr & MSR_AMD64_SYSCFG_MEM_ENCRYPT))
520 			goto clear_all;
521 
522 		/*
523 		 * Always adjust physical address bits. Even though this
524 		 * will be a value above 32-bits this is still done for
525 		 * CONFIG_X86_32 so that accurate values are reported.
526 		 */
527 		c->x86_phys_bits -= (cpuid_ebx(0x8000001f) >> 6) & 0x3f;
528 
529 		if (IS_ENABLED(CONFIG_X86_32))
530 			goto clear_all;
531 
532 		if (!sme_me_mask)
533 			setup_clear_cpu_cap(X86_FEATURE_SME);
534 
535 		rdmsrq(MSR_K7_HWCR, msr);
536 		if (!(msr & MSR_K7_HWCR_SMMLOCK))
537 			goto clear_sev;
538 
539 		return;
540 
541 clear_all:
542 		setup_clear_cpu_cap(X86_FEATURE_SME);
543 clear_sev:
544 		setup_clear_cpu_cap(X86_FEATURE_SEV);
545 		setup_clear_cpu_cap(X86_FEATURE_SEV_ES);
546 		setup_clear_cpu_cap(X86_FEATURE_SEV_SNP);
547 	}
548 }
549 
early_init_amd(struct cpuinfo_x86 * c)550 static void early_init_amd(struct cpuinfo_x86 *c)
551 {
552 	u32 dummy;
553 
554 	if (c->x86 >= 0xf)
555 		set_cpu_cap(c, X86_FEATURE_K8);
556 
557 	rdmsr_safe(MSR_AMD64_PATCH_LEVEL, &c->microcode, &dummy);
558 
559 	/*
560 	 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
561 	 * with P/T states and does not stop in deep C-states
562 	 */
563 	if (c->x86_power & (1 << 8)) {
564 		set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
565 		set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
566 	}
567 
568 	/* Bit 12 of 8000_0007 edx is accumulated power mechanism. */
569 	if (c->x86_power & BIT(12))
570 		set_cpu_cap(c, X86_FEATURE_ACC_POWER);
571 
572 	/* Bit 14 indicates the Runtime Average Power Limit interface. */
573 	if (c->x86_power & BIT(14))
574 		set_cpu_cap(c, X86_FEATURE_RAPL);
575 
576 #ifdef CONFIG_X86_64
577 	set_cpu_cap(c, X86_FEATURE_SYSCALL32);
578 #else
579 	/*  Set MTRR capability flag if appropriate */
580 	if (c->x86 == 5)
581 		if (c->x86_model == 13 || c->x86_model == 9 ||
582 		    (c->x86_model == 8 && c->x86_stepping >= 8))
583 			set_cpu_cap(c, X86_FEATURE_K6_MTRR);
584 #endif
585 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
586 	/*
587 	 * ApicID can always be treated as an 8-bit value for AMD APIC versions
588 	 * >= 0x10, but even old K8s came out of reset with version 0x10. So, we
589 	 * can safely set X86_FEATURE_EXTD_APICID unconditionally for families
590 	 * after 16h.
591 	 */
592 	if (boot_cpu_has(X86_FEATURE_APIC)) {
593 		if (c->x86 > 0x16)
594 			set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
595 		else if (c->x86 >= 0xf) {
596 			/* check CPU config space for extended APIC ID */
597 			unsigned int val;
598 
599 			val = read_pci_config(0, 24, 0, 0x68);
600 			if ((val >> 17 & 0x3) == 0x3)
601 				set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
602 		}
603 	}
604 #endif
605 
606 	/*
607 	 * This is only needed to tell the kernel whether to use VMCALL
608 	 * and VMMCALL.  VMMCALL is never executed except under virt, so
609 	 * we can set it unconditionally.
610 	 */
611 	set_cpu_cap(c, X86_FEATURE_VMMCALL);
612 
613 	/* F16h erratum 793, CVE-2013-6885 */
614 	if (c->x86 == 0x16 && c->x86_model <= 0xf)
615 		msr_set_bit(MSR_AMD64_LS_CFG, 15);
616 
617 	early_detect_mem_encrypt(c);
618 
619 	if (!cpu_has(c, X86_FEATURE_HYPERVISOR) && !cpu_has(c, X86_FEATURE_IBPB_BRTYPE)) {
620 		if (c->x86 == 0x17 && boot_cpu_has(X86_FEATURE_AMD_IBPB))
621 			setup_force_cpu_cap(X86_FEATURE_IBPB_BRTYPE);
622 		else if (c->x86 >= 0x19 && !wrmsrq_safe(MSR_IA32_PRED_CMD, PRED_CMD_SBPB)) {
623 			setup_force_cpu_cap(X86_FEATURE_IBPB_BRTYPE);
624 			setup_force_cpu_cap(X86_FEATURE_SBPB);
625 		}
626 	}
627 }
628 
init_amd_k8(struct cpuinfo_x86 * c)629 static void init_amd_k8(struct cpuinfo_x86 *c)
630 {
631 	u32 level;
632 	u64 value;
633 
634 	/* On C+ stepping K8 rep microcode works well for copy/memset */
635 	level = cpuid_eax(1);
636 	if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58)
637 		set_cpu_cap(c, X86_FEATURE_REP_GOOD);
638 
639 	/*
640 	 * Some BIOSes incorrectly force this feature, but only K8 revision D
641 	 * (model = 0x14) and later actually support it.
642 	 * (AMD Erratum #110, docId: 25759).
643 	 */
644 	if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM) && !cpu_has(c, X86_FEATURE_HYPERVISOR)) {
645 		clear_cpu_cap(c, X86_FEATURE_LAHF_LM);
646 		if (!rdmsrq_amd_safe(0xc001100d, &value)) {
647 			value &= ~BIT_64(32);
648 			wrmsrq_amd_safe(0xc001100d, value);
649 		}
650 	}
651 
652 	if (!c->x86_model_id[0])
653 		strscpy(c->x86_model_id, "Hammer");
654 
655 #ifdef CONFIG_SMP
656 	/*
657 	 * Disable TLB flush filter by setting HWCR.FFDIS on K8
658 	 * bit 6 of msr C001_0015
659 	 *
660 	 * Errata 63 for SH-B3 steppings
661 	 * Errata 122 for all steppings (F+ have it disabled by default)
662 	 */
663 	msr_set_bit(MSR_K7_HWCR, 6);
664 #endif
665 	set_cpu_bug(c, X86_BUG_SWAPGS_FENCE);
666 
667 	/*
668 	 * Check models and steppings affected by erratum 400. This is
669 	 * used to select the proper idle routine and to enable the
670 	 * check whether the machine is affected in arch_post_acpi_subsys_init()
671 	 * which sets the X86_BUG_AMD_APIC_C1E bug depending on the MSR check.
672 	 */
673 	if (c->x86_model > 0x41 ||
674 	    (c->x86_model == 0x41 && c->x86_stepping >= 0x2))
675 		setup_force_cpu_bug(X86_BUG_AMD_E400);
676 }
677 
init_amd_gh(struct cpuinfo_x86 * c)678 static void init_amd_gh(struct cpuinfo_x86 *c)
679 {
680 #ifdef CONFIG_MMCONF_FAM10H
681 	/* do this for boot cpu */
682 	if (c == &boot_cpu_data)
683 		check_enable_amd_mmconf_dmi();
684 
685 	fam10h_check_enable_mmcfg();
686 #endif
687 
688 	/*
689 	 * Disable GART TLB Walk Errors on Fam10h. We do this here because this
690 	 * is always needed when GART is enabled, even in a kernel which has no
691 	 * MCE support built in. BIOS should disable GartTlbWlk Errors already.
692 	 * If it doesn't, we do it here as suggested by the BKDG.
693 	 *
694 	 * Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=33012
695 	 */
696 	msr_set_bit(MSR_AMD64_MCx_MASK(4), 10);
697 
698 	/*
699 	 * On family 10h BIOS may not have properly enabled WC+ support, causing
700 	 * it to be converted to CD memtype. This may result in performance
701 	 * degradation for certain nested-paging guests. Prevent this conversion
702 	 * by clearing bit 24 in MSR_AMD64_BU_CFG2.
703 	 *
704 	 * NOTE: we want to use the _safe accessors so as not to #GP kvm
705 	 * guests on older kvm hosts.
706 	 */
707 	msr_clear_bit(MSR_AMD64_BU_CFG2, 24);
708 
709 	set_cpu_bug(c, X86_BUG_AMD_TLB_MMATCH);
710 
711 	/*
712 	 * Check models and steppings affected by erratum 400. This is
713 	 * used to select the proper idle routine and to enable the
714 	 * check whether the machine is affected in arch_post_acpi_subsys_init()
715 	 * which sets the X86_BUG_AMD_APIC_C1E bug depending on the MSR check.
716 	 */
717 	if (c->x86_model > 0x2 ||
718 	    (c->x86_model == 0x2 && c->x86_stepping >= 0x1))
719 		setup_force_cpu_bug(X86_BUG_AMD_E400);
720 }
721 
init_amd_ln(struct cpuinfo_x86 * c)722 static void init_amd_ln(struct cpuinfo_x86 *c)
723 {
724 	/*
725 	 * Apply erratum 665 fix unconditionally so machines without a BIOS
726 	 * fix work.
727 	 */
728 	msr_set_bit(MSR_AMD64_DE_CFG, 31);
729 }
730 
731 static bool rdrand_force;
732 
rdrand_cmdline(char * str)733 static int __init rdrand_cmdline(char *str)
734 {
735 	if (!str)
736 		return -EINVAL;
737 
738 	if (!strcmp(str, "force"))
739 		rdrand_force = true;
740 	else
741 		return -EINVAL;
742 
743 	return 0;
744 }
745 early_param("rdrand", rdrand_cmdline);
746 
clear_rdrand_cpuid_bit(struct cpuinfo_x86 * c)747 static void clear_rdrand_cpuid_bit(struct cpuinfo_x86 *c)
748 {
749 	/*
750 	 * Saving of the MSR used to hide the RDRAND support during
751 	 * suspend/resume is done by arch/x86/power/cpu.c, which is
752 	 * dependent on CONFIG_PM_SLEEP.
753 	 */
754 	if (!IS_ENABLED(CONFIG_PM_SLEEP))
755 		return;
756 
757 	/*
758 	 * The self-test can clear X86_FEATURE_RDRAND, so check for
759 	 * RDRAND support using the CPUID function directly.
760 	 */
761 	if (!(cpuid_ecx(1) & BIT(30)) || rdrand_force)
762 		return;
763 
764 	msr_clear_bit(MSR_AMD64_CPUID_FN_1, 62);
765 
766 	/*
767 	 * Verify that the CPUID change has occurred in case the kernel is
768 	 * running virtualized and the hypervisor doesn't support the MSR.
769 	 */
770 	if (cpuid_ecx(1) & BIT(30)) {
771 		pr_info_once("BIOS may not properly restore RDRAND after suspend, but hypervisor does not support hiding RDRAND via CPUID.\n");
772 		return;
773 	}
774 
775 	clear_cpu_cap(c, X86_FEATURE_RDRAND);
776 	pr_info_once("BIOS may not properly restore RDRAND after suspend, hiding RDRAND via CPUID. Use rdrand=force to reenable.\n");
777 }
778 
init_amd_jg(struct cpuinfo_x86 * c)779 static void init_amd_jg(struct cpuinfo_x86 *c)
780 {
781 	/*
782 	 * Some BIOS implementations do not restore proper RDRAND support
783 	 * across suspend and resume. Check on whether to hide the RDRAND
784 	 * instruction support via CPUID.
785 	 */
786 	clear_rdrand_cpuid_bit(c);
787 }
788 
init_amd_bd(struct cpuinfo_x86 * c)789 static void init_amd_bd(struct cpuinfo_x86 *c)
790 {
791 	u64 value;
792 
793 	/*
794 	 * The way access filter has a performance penalty on some workloads.
795 	 * Disable it on the affected CPUs.
796 	 */
797 	if ((c->x86_model >= 0x02) && (c->x86_model < 0x20)) {
798 		if (!rdmsrq_safe(MSR_F15H_IC_CFG, &value) && !(value & 0x1E)) {
799 			value |= 0x1E;
800 			wrmsrq_safe(MSR_F15H_IC_CFG, value);
801 		}
802 	}
803 
804 	/*
805 	 * Some BIOS implementations do not restore proper RDRAND support
806 	 * across suspend and resume. Check on whether to hide the RDRAND
807 	 * instruction support via CPUID.
808 	 */
809 	clear_rdrand_cpuid_bit(c);
810 }
811 
812 static const struct x86_cpu_id erratum_1386_microcode[] = {
813 	X86_MATCH_VFM_STEPS(VFM_MAKE(X86_VENDOR_AMD, 0x17, 0x01), 0x2, 0x2, 0x0800126e),
814 	X86_MATCH_VFM_STEPS(VFM_MAKE(X86_VENDOR_AMD, 0x17, 0x31), 0x0, 0x0, 0x08301052),
815 	{}
816 };
817 
fix_erratum_1386(struct cpuinfo_x86 * c)818 static void fix_erratum_1386(struct cpuinfo_x86 *c)
819 {
820 	/*
821 	 * Work around Erratum 1386.  The XSAVES instruction malfunctions in
822 	 * certain circumstances on Zen1/2 uarch, and not all parts have had
823 	 * updated microcode at the time of writing (March 2023).
824 	 *
825 	 * Affected parts all have no supervisor XSAVE states, meaning that
826 	 * the XSAVEC instruction (which works fine) is equivalent.
827 	 *
828 	 * Clear the feature flag only on microcode revisions which
829 	 * don't have the fix.
830 	 */
831 	if (x86_match_min_microcode_rev(erratum_1386_microcode))
832 		return;
833 
834 	clear_cpu_cap(c, X86_FEATURE_XSAVES);
835 }
836 
init_spectral_chicken(struct cpuinfo_x86 * c)837 void init_spectral_chicken(struct cpuinfo_x86 *c)
838 {
839 #ifdef CONFIG_MITIGATION_UNRET_ENTRY
840 	u64 value;
841 
842 	/*
843 	 * On Zen2 we offer this chicken (bit) on the altar of Speculation.
844 	 *
845 	 * This suppresses speculation from the middle of a basic block, i.e. it
846 	 * suppresses non-branch predictions.
847 	 */
848 	if (!cpu_has(c, X86_FEATURE_HYPERVISOR)) {
849 		if (!rdmsrq_safe(MSR_ZEN2_SPECTRAL_CHICKEN, &value)) {
850 			value |= MSR_ZEN2_SPECTRAL_CHICKEN_BIT;
851 			wrmsrq_safe(MSR_ZEN2_SPECTRAL_CHICKEN, value);
852 		}
853 	}
854 #endif
855 }
856 
init_amd_zen_common(void)857 static void init_amd_zen_common(void)
858 {
859 	setup_force_cpu_cap(X86_FEATURE_ZEN);
860 #ifdef CONFIG_NUMA
861 	node_reclaim_distance = 32;
862 #endif
863 }
864 
init_amd_zen1(struct cpuinfo_x86 * c)865 static void init_amd_zen1(struct cpuinfo_x86 *c)
866 {
867 	fix_erratum_1386(c);
868 
869 	/* Fix up CPUID bits, but only if not virtualised. */
870 	if (!cpu_has(c, X86_FEATURE_HYPERVISOR)) {
871 
872 		/* Erratum 1076: CPB feature bit not being set in CPUID. */
873 		if (!cpu_has(c, X86_FEATURE_CPB))
874 			set_cpu_cap(c, X86_FEATURE_CPB);
875 	}
876 
877 	pr_notice_once("AMD Zen1 DIV0 bug detected. Disable SMT for full protection.\n");
878 	setup_force_cpu_bug(X86_BUG_DIV0);
879 
880 	/*
881 	 * Turn off the Instructions Retired free counter on machines that are
882 	 * susceptible to erratum #1054 "Instructions Retired Performance
883 	 * Counter May Be Inaccurate".
884 	 */
885 	if (c->x86_model < 0x30) {
886 		msr_clear_bit(MSR_K7_HWCR, MSR_K7_HWCR_IRPERF_EN_BIT);
887 		clear_cpu_cap(c, X86_FEATURE_IRPERF);
888 	}
889 }
890 
cpu_has_zenbleed_microcode(void)891 static bool cpu_has_zenbleed_microcode(void)
892 {
893 	u32 good_rev = 0;
894 
895 	switch (boot_cpu_data.x86_model) {
896 	case 0x30 ... 0x3f: good_rev = 0x0830107b; break;
897 	case 0x60 ... 0x67: good_rev = 0x0860010c; break;
898 	case 0x68 ... 0x6f: good_rev = 0x08608107; break;
899 	case 0x70 ... 0x7f: good_rev = 0x08701033; break;
900 	case 0xa0 ... 0xaf: good_rev = 0x08a00009; break;
901 
902 	default:
903 		return false;
904 	}
905 
906 	if (boot_cpu_data.microcode < good_rev)
907 		return false;
908 
909 	return true;
910 }
911 
zen2_zenbleed_check(struct cpuinfo_x86 * c)912 static void zen2_zenbleed_check(struct cpuinfo_x86 *c)
913 {
914 	if (cpu_has(c, X86_FEATURE_HYPERVISOR))
915 		return;
916 
917 	if (!cpu_has(c, X86_FEATURE_AVX))
918 		return;
919 
920 	if (!cpu_has_zenbleed_microcode()) {
921 		pr_notice_once("Zenbleed: please update your microcode for the most optimal fix\n");
922 		msr_set_bit(MSR_AMD64_DE_CFG, MSR_AMD64_DE_CFG_ZEN2_FP_BACKUP_FIX_BIT);
923 	} else {
924 		msr_clear_bit(MSR_AMD64_DE_CFG, MSR_AMD64_DE_CFG_ZEN2_FP_BACKUP_FIX_BIT);
925 	}
926 }
927 
init_amd_zen2(struct cpuinfo_x86 * c)928 static void init_amd_zen2(struct cpuinfo_x86 *c)
929 {
930 	init_spectral_chicken(c);
931 	fix_erratum_1386(c);
932 	zen2_zenbleed_check(c);
933 }
934 
init_amd_zen3(struct cpuinfo_x86 * c)935 static void init_amd_zen3(struct cpuinfo_x86 *c)
936 {
937 	if (!cpu_has(c, X86_FEATURE_HYPERVISOR)) {
938 		/*
939 		 * Zen3 (Fam19 model < 0x10) parts are not susceptible to
940 		 * Branch Type Confusion, but predate the allocation of the
941 		 * BTC_NO bit.
942 		 */
943 		if (!cpu_has(c, X86_FEATURE_BTC_NO))
944 			set_cpu_cap(c, X86_FEATURE_BTC_NO);
945 	}
946 }
947 
init_amd_zen4(struct cpuinfo_x86 * c)948 static void init_amd_zen4(struct cpuinfo_x86 *c)
949 {
950 	if (!cpu_has(c, X86_FEATURE_HYPERVISOR))
951 		msr_set_bit(MSR_ZEN4_BP_CFG, MSR_ZEN4_BP_CFG_SHARED_BTB_FIX_BIT);
952 
953 	/*
954 	 * These Zen4 SoCs advertise support for virtualized VMLOAD/VMSAVE
955 	 * in some BIOS versions but they can lead to random host reboots.
956 	 */
957 	switch (c->x86_model) {
958 	case 0x18 ... 0x1f:
959 	case 0x60 ... 0x7f:
960 		clear_cpu_cap(c, X86_FEATURE_V_VMSAVE_VMLOAD);
961 		break;
962 	}
963 }
964 
init_amd_zen5(struct cpuinfo_x86 * c)965 static void init_amd_zen5(struct cpuinfo_x86 *c)
966 {
967 }
968 
init_amd(struct cpuinfo_x86 * c)969 static void init_amd(struct cpuinfo_x86 *c)
970 {
971 	u64 vm_cr;
972 
973 	early_init_amd(c);
974 
975 	/*
976 	 * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
977 	 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
978 	 */
979 	clear_cpu_cap(c, 0*32+31);
980 
981 	if (c->x86 >= 0x10)
982 		set_cpu_cap(c, X86_FEATURE_REP_GOOD);
983 
984 	/* AMD FSRM also implies FSRS */
985 	if (cpu_has(c, X86_FEATURE_FSRM))
986 		set_cpu_cap(c, X86_FEATURE_FSRS);
987 
988 	/* K6s reports MCEs but don't actually have all the MSRs */
989 	if (c->x86 < 6)
990 		clear_cpu_cap(c, X86_FEATURE_MCE);
991 
992 	switch (c->x86) {
993 	case 4:    init_amd_k5(c); break;
994 	case 5:    init_amd_k6(c); break;
995 	case 6:	   init_amd_k7(c); break;
996 	case 0xf:  init_amd_k8(c); break;
997 	case 0x10: init_amd_gh(c); break;
998 	case 0x12: init_amd_ln(c); break;
999 	case 0x15: init_amd_bd(c); break;
1000 	case 0x16: init_amd_jg(c); break;
1001 	}
1002 
1003 	/*
1004 	 * Save up on some future enablement work and do common Zen
1005 	 * settings.
1006 	 */
1007 	if (c->x86 >= 0x17)
1008 		init_amd_zen_common();
1009 
1010 	if (boot_cpu_has(X86_FEATURE_ZEN1))
1011 		init_amd_zen1(c);
1012 	else if (boot_cpu_has(X86_FEATURE_ZEN2))
1013 		init_amd_zen2(c);
1014 	else if (boot_cpu_has(X86_FEATURE_ZEN3))
1015 		init_amd_zen3(c);
1016 	else if (boot_cpu_has(X86_FEATURE_ZEN4))
1017 		init_amd_zen4(c);
1018 	else if (boot_cpu_has(X86_FEATURE_ZEN5))
1019 		init_amd_zen5(c);
1020 
1021 	/*
1022 	 * Enable workaround for FXSAVE leak on CPUs
1023 	 * without a XSaveErPtr feature
1024 	 */
1025 	if ((c->x86 >= 6) && (!cpu_has(c, X86_FEATURE_XSAVEERPTR)))
1026 		set_cpu_bug(c, X86_BUG_FXSAVE_LEAK);
1027 
1028 	cpu_detect_cache_sizes(c);
1029 
1030 	srat_detect_node(c);
1031 
1032 	init_amd_cacheinfo(c);
1033 
1034 	if (cpu_has(c, X86_FEATURE_SVM)) {
1035 		rdmsrq(MSR_VM_CR, vm_cr);
1036 		if (vm_cr & SVM_VM_CR_SVM_DIS_MASK) {
1037 			pr_notice_once("SVM disabled (by BIOS) in MSR_VM_CR\n");
1038 			clear_cpu_cap(c, X86_FEATURE_SVM);
1039 		}
1040 	}
1041 
1042 	if (!cpu_has(c, X86_FEATURE_LFENCE_RDTSC) && cpu_has(c, X86_FEATURE_XMM2)) {
1043 		/*
1044 		 * Use LFENCE for execution serialization.  On families which
1045 		 * don't have that MSR, LFENCE is already serializing.
1046 		 * msr_set_bit() uses the safe accessors, too, even if the MSR
1047 		 * is not present.
1048 		 */
1049 		msr_set_bit(MSR_AMD64_DE_CFG,
1050 			    MSR_AMD64_DE_CFG_LFENCE_SERIALIZE_BIT);
1051 
1052 		/* A serializing LFENCE stops RDTSC speculation */
1053 		set_cpu_cap(c, X86_FEATURE_LFENCE_RDTSC);
1054 	}
1055 
1056 	/*
1057 	 * Family 0x12 and above processors have APIC timer
1058 	 * running in deep C states.
1059 	 */
1060 	if (c->x86 > 0x11)
1061 		set_cpu_cap(c, X86_FEATURE_ARAT);
1062 
1063 	/* 3DNow or LM implies PREFETCHW */
1064 	if (!cpu_has(c, X86_FEATURE_3DNOWPREFETCH))
1065 		if (cpu_has(c, X86_FEATURE_3DNOW) || cpu_has(c, X86_FEATURE_LM))
1066 			set_cpu_cap(c, X86_FEATURE_3DNOWPREFETCH);
1067 
1068 	/* AMD CPUs don't reset SS attributes on SYSRET, Xen does. */
1069 	if (!cpu_feature_enabled(X86_FEATURE_XENPV))
1070 		set_cpu_bug(c, X86_BUG_SYSRET_SS_ATTRS);
1071 
1072 	/* Enable the Instructions Retired free counter */
1073 	if (cpu_has(c, X86_FEATURE_IRPERF))
1074 		msr_set_bit(MSR_K7_HWCR, MSR_K7_HWCR_IRPERF_EN_BIT);
1075 
1076 	check_null_seg_clears_base(c);
1077 
1078 	/*
1079 	 * Make sure EFER[AIBRSE - Automatic IBRS Enable] is set. The APs are brought up
1080 	 * using the trampoline code and as part of it, MSR_EFER gets prepared there in
1081 	 * order to be replicated onto them. Regardless, set it here again, if not set,
1082 	 * to protect against any future refactoring/code reorganization which might
1083 	 * miss setting this important bit.
1084 	 */
1085 	if (spectre_v2_in_eibrs_mode(spectre_v2_enabled) &&
1086 	    cpu_has(c, X86_FEATURE_AUTOIBRS))
1087 		WARN_ON_ONCE(msr_set_bit(MSR_EFER, _EFER_AUTOIBRS) < 0);
1088 
1089 	/* AMD CPUs don't need fencing after x2APIC/TSC_DEADLINE MSR writes. */
1090 	clear_cpu_cap(c, X86_FEATURE_APIC_MSRS_FENCE);
1091 
1092 	/* Enable Translation Cache Extension */
1093 	if (cpu_has(c, X86_FEATURE_TCE))
1094 		msr_set_bit(MSR_EFER, _EFER_TCE);
1095 }
1096 
1097 #ifdef CONFIG_X86_32
amd_size_cache(struct cpuinfo_x86 * c,unsigned int size)1098 static unsigned int amd_size_cache(struct cpuinfo_x86 *c, unsigned int size)
1099 {
1100 	/* AMD errata T13 (order #21922) */
1101 	if (c->x86 == 6) {
1102 		/* Duron Rev A0 */
1103 		if (c->x86_model == 3 && c->x86_stepping == 0)
1104 			size = 64;
1105 		/* Tbird rev A1/A2 */
1106 		if (c->x86_model == 4 &&
1107 			(c->x86_stepping == 0 || c->x86_stepping == 1))
1108 			size = 256;
1109 	}
1110 	return size;
1111 }
1112 #endif
1113 
cpu_detect_tlb_amd(struct cpuinfo_x86 * c)1114 static void cpu_detect_tlb_amd(struct cpuinfo_x86 *c)
1115 {
1116 	u32 ebx, eax, ecx, edx;
1117 	u16 mask = 0xfff;
1118 
1119 	if (c->x86 < 0xf)
1120 		return;
1121 
1122 	if (c->extended_cpuid_level < 0x80000006)
1123 		return;
1124 
1125 	cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
1126 
1127 	tlb_lld_4k = (ebx >> 16) & mask;
1128 	tlb_lli_4k = ebx & mask;
1129 
1130 	/*
1131 	 * K8 doesn't have 2M/4M entries in the L2 TLB so read out the L1 TLB
1132 	 * characteristics from the CPUID function 0x80000005 instead.
1133 	 */
1134 	if (c->x86 == 0xf) {
1135 		cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
1136 		mask = 0xff;
1137 	}
1138 
1139 	/* Handle DTLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
1140 	if (!((eax >> 16) & mask))
1141 		tlb_lld_2m = (cpuid_eax(0x80000005) >> 16) & 0xff;
1142 	else
1143 		tlb_lld_2m = (eax >> 16) & mask;
1144 
1145 	/* a 4M entry uses two 2M entries */
1146 	tlb_lld_4m = tlb_lld_2m >> 1;
1147 
1148 	/* Handle ITLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
1149 	if (!(eax & mask)) {
1150 		/* Erratum 658 */
1151 		if (c->x86 == 0x15 && c->x86_model <= 0x1f) {
1152 			tlb_lli_2m = 1024;
1153 		} else {
1154 			cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
1155 			tlb_lli_2m = eax & 0xff;
1156 		}
1157 	} else
1158 		tlb_lli_2m = eax & mask;
1159 
1160 	tlb_lli_4m = tlb_lli_2m >> 1;
1161 
1162 	/* Max number of pages INVLPGB can invalidate in one shot */
1163 	if (cpu_has(c, X86_FEATURE_INVLPGB))
1164 		invlpgb_count_max = (cpuid_edx(0x80000008) & 0xffff) + 1;
1165 }
1166 
1167 static const struct cpu_dev amd_cpu_dev = {
1168 	.c_vendor	= "AMD",
1169 	.c_ident	= { "AuthenticAMD" },
1170 #ifdef CONFIG_X86_32
1171 	.legacy_models = {
1172 		{ .family = 4, .model_names =
1173 		  {
1174 			  [3] = "486 DX/2",
1175 			  [7] = "486 DX/2-WB",
1176 			  [8] = "486 DX/4",
1177 			  [9] = "486 DX/4-WB",
1178 			  [14] = "Am5x86-WT",
1179 			  [15] = "Am5x86-WB"
1180 		  }
1181 		},
1182 	},
1183 	.legacy_cache_size = amd_size_cache,
1184 #endif
1185 	.c_early_init   = early_init_amd,
1186 	.c_detect_tlb	= cpu_detect_tlb_amd,
1187 	.c_bsp_init	= bsp_init_amd,
1188 	.c_init		= init_amd,
1189 	.c_x86_vendor	= X86_VENDOR_AMD,
1190 };
1191 
1192 cpu_dev_register(amd_cpu_dev);
1193 
1194 static DEFINE_PER_CPU_READ_MOSTLY(unsigned long[4], amd_dr_addr_mask);
1195 
1196 static unsigned int amd_msr_dr_addr_masks[] = {
1197 	MSR_F16H_DR0_ADDR_MASK,
1198 	MSR_F16H_DR1_ADDR_MASK,
1199 	MSR_F16H_DR1_ADDR_MASK + 1,
1200 	MSR_F16H_DR1_ADDR_MASK + 2
1201 };
1202 
amd_set_dr_addr_mask(unsigned long mask,unsigned int dr)1203 void amd_set_dr_addr_mask(unsigned long mask, unsigned int dr)
1204 {
1205 	int cpu = smp_processor_id();
1206 
1207 	if (!cpu_feature_enabled(X86_FEATURE_BPEXT))
1208 		return;
1209 
1210 	if (WARN_ON_ONCE(dr >= ARRAY_SIZE(amd_msr_dr_addr_masks)))
1211 		return;
1212 
1213 	if (per_cpu(amd_dr_addr_mask, cpu)[dr] == mask)
1214 		return;
1215 
1216 	wrmsrq(amd_msr_dr_addr_masks[dr], mask);
1217 	per_cpu(amd_dr_addr_mask, cpu)[dr] = mask;
1218 }
1219 
amd_get_dr_addr_mask(unsigned int dr)1220 unsigned long amd_get_dr_addr_mask(unsigned int dr)
1221 {
1222 	if (!cpu_feature_enabled(X86_FEATURE_BPEXT))
1223 		return 0;
1224 
1225 	if (WARN_ON_ONCE(dr >= ARRAY_SIZE(amd_msr_dr_addr_masks)))
1226 		return 0;
1227 
1228 	return per_cpu(amd_dr_addr_mask[dr], smp_processor_id());
1229 }
1230 EXPORT_SYMBOL_GPL(amd_get_dr_addr_mask);
1231 
zenbleed_check_cpu(void * unused)1232 static void zenbleed_check_cpu(void *unused)
1233 {
1234 	struct cpuinfo_x86 *c = &cpu_data(smp_processor_id());
1235 
1236 	zen2_zenbleed_check(c);
1237 }
1238 
amd_check_microcode(void)1239 void amd_check_microcode(void)
1240 {
1241 	if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD)
1242 		return;
1243 
1244 	if (cpu_feature_enabled(X86_FEATURE_ZEN2))
1245 		on_each_cpu(zenbleed_check_cpu, NULL, 1);
1246 }
1247 
1248 static const char * const s5_reset_reason_txt[] = {
1249 	[0]  = "thermal pin BP_THERMTRIP_L was tripped",
1250 	[1]  = "power button was pressed for 4 seconds",
1251 	[2]  = "shutdown pin was tripped",
1252 	[4]  = "remote ASF power off command was received",
1253 	[9]  = "internal CPU thermal limit was tripped",
1254 	[16] = "system reset pin BP_SYS_RST_L was tripped",
1255 	[17] = "software issued PCI reset",
1256 	[18] = "software wrote 0x4 to reset control register 0xCF9",
1257 	[19] = "software wrote 0x6 to reset control register 0xCF9",
1258 	[20] = "software wrote 0xE to reset control register 0xCF9",
1259 	[21] = "ACPI power state transition occurred",
1260 	[22] = "keyboard reset pin KB_RST_L was tripped",
1261 	[23] = "internal CPU shutdown event occurred",
1262 	[24] = "system failed to boot before failed boot timer expired",
1263 	[25] = "hardware watchdog timer expired",
1264 	[26] = "remote ASF reset command was received",
1265 	[27] = "an uncorrected error caused a data fabric sync flood event",
1266 	[29] = "FCH and MP1 failed warm reset handshake",
1267 	[30] = "a parity error occurred",
1268 	[31] = "a software sync flood event occurred",
1269 };
1270 
print_s5_reset_status_mmio(void)1271 static __init int print_s5_reset_status_mmio(void)
1272 {
1273 	unsigned long value;
1274 	void __iomem *addr;
1275 	int i;
1276 
1277 	if (!cpu_feature_enabled(X86_FEATURE_ZEN))
1278 		return 0;
1279 
1280 	addr = ioremap(FCH_PM_BASE + FCH_PM_S5_RESET_STATUS, sizeof(value));
1281 	if (!addr)
1282 		return 0;
1283 
1284 	value = ioread32(addr);
1285 	iounmap(addr);
1286 
1287 	for (i = 0; i < ARRAY_SIZE(s5_reset_reason_txt); i++) {
1288 		if (!(value & BIT(i)))
1289 			continue;
1290 
1291 		if (s5_reset_reason_txt[i]) {
1292 			pr_info("x86/amd: Previous system reset reason [0x%08lx]: %s\n",
1293 				value, s5_reset_reason_txt[i]);
1294 		}
1295 	}
1296 
1297 	return 0;
1298 }
1299 late_initcall(print_s5_reset_status_mmio);
1300