xref: /linux/arch/x86/kernel/cpu/amd.c (revision 6e9128ff9d8113ef208e5ec82573b96ead100072)
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 
380 #define ZEN_MODEL_STEP_UCODE(fam, model, step, ucode) \
381 	X86_MATCH_VFM_STEPS(VFM_MAKE(X86_VENDOR_AMD, fam, model), \
382 			    step, step, ucode)
383 
384 static const struct x86_cpu_id amd_tsa_microcode[] = {
385 	ZEN_MODEL_STEP_UCODE(0x19, 0x01, 0x1, 0x0a0011d7),
386 	ZEN_MODEL_STEP_UCODE(0x19, 0x01, 0x2, 0x0a00123b),
387 	ZEN_MODEL_STEP_UCODE(0x19, 0x08, 0x2, 0x0a00820d),
388 	ZEN_MODEL_STEP_UCODE(0x19, 0x11, 0x1, 0x0a10114c),
389 	ZEN_MODEL_STEP_UCODE(0x19, 0x11, 0x2, 0x0a10124c),
390 	ZEN_MODEL_STEP_UCODE(0x19, 0x18, 0x1, 0x0a108109),
391 	ZEN_MODEL_STEP_UCODE(0x19, 0x21, 0x0, 0x0a20102e),
392 	ZEN_MODEL_STEP_UCODE(0x19, 0x21, 0x2, 0x0a201211),
393 	ZEN_MODEL_STEP_UCODE(0x19, 0x44, 0x1, 0x0a404108),
394 	ZEN_MODEL_STEP_UCODE(0x19, 0x50, 0x0, 0x0a500012),
395 	ZEN_MODEL_STEP_UCODE(0x19, 0x61, 0x2, 0x0a60120a),
396 	ZEN_MODEL_STEP_UCODE(0x19, 0x74, 0x1, 0x0a704108),
397 	ZEN_MODEL_STEP_UCODE(0x19, 0x75, 0x2, 0x0a705208),
398 	ZEN_MODEL_STEP_UCODE(0x19, 0x78, 0x0, 0x0a708008),
399 	ZEN_MODEL_STEP_UCODE(0x19, 0x7c, 0x0, 0x0a70c008),
400 	ZEN_MODEL_STEP_UCODE(0x19, 0xa0, 0x2, 0x0aa00216),
401 	{},
402 };
403 
tsa_init(struct cpuinfo_x86 * c)404 static void tsa_init(struct cpuinfo_x86 *c)
405 {
406 	if (cpu_has(c, X86_FEATURE_HYPERVISOR))
407 		return;
408 
409 	if (cpu_has(c, X86_FEATURE_ZEN3) ||
410 	    cpu_has(c, X86_FEATURE_ZEN4)) {
411 		if (x86_match_min_microcode_rev(amd_tsa_microcode))
412 			setup_force_cpu_cap(X86_FEATURE_VERW_CLEAR);
413 		else
414 			pr_debug("%s: current revision: 0x%x\n", __func__, c->microcode);
415 	} else {
416 		setup_force_cpu_cap(X86_FEATURE_TSA_SQ_NO);
417 		setup_force_cpu_cap(X86_FEATURE_TSA_L1_NO);
418 	}
419 }
420 
bsp_init_amd(struct cpuinfo_x86 * c)421 static void bsp_init_amd(struct cpuinfo_x86 *c)
422 {
423 	if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) {
424 
425 		if (c->x86 > 0x10 ||
426 		    (c->x86 == 0x10 && c->x86_model >= 0x2)) {
427 			u64 val;
428 
429 			rdmsrq(MSR_K7_HWCR, val);
430 			if (!(val & BIT(24)))
431 				pr_warn(FW_BUG "TSC doesn't count with P0 frequency!\n");
432 		}
433 	}
434 
435 	if (c->x86 == 0x15) {
436 		unsigned long upperbit;
437 		u32 cpuid, assoc;
438 
439 		cpuid	 = cpuid_edx(0x80000005);
440 		assoc	 = cpuid >> 16 & 0xff;
441 		upperbit = ((cpuid >> 24) << 10) / assoc;
442 
443 		va_align.mask	  = (upperbit - 1) & PAGE_MASK;
444 		va_align.flags    = ALIGN_VA_32 | ALIGN_VA_64;
445 
446 		/* A random value per boot for bit slice [12:upper_bit) */
447 		va_align.bits = get_random_u32() & va_align.mask;
448 	}
449 
450 	if (cpu_has(c, X86_FEATURE_MWAITX))
451 		use_mwaitx_delay();
452 
453 	if (!boot_cpu_has(X86_FEATURE_AMD_SSBD) &&
454 	    !boot_cpu_has(X86_FEATURE_VIRT_SSBD) &&
455 	    c->x86 >= 0x15 && c->x86 <= 0x17) {
456 		unsigned int bit;
457 
458 		switch (c->x86) {
459 		case 0x15: bit = 54; break;
460 		case 0x16: bit = 33; break;
461 		case 0x17: bit = 10; break;
462 		default: return;
463 		}
464 		/*
465 		 * Try to cache the base value so further operations can
466 		 * avoid RMW. If that faults, do not enable SSBD.
467 		 */
468 		if (!rdmsrq_safe(MSR_AMD64_LS_CFG, &x86_amd_ls_cfg_base)) {
469 			setup_force_cpu_cap(X86_FEATURE_LS_CFG_SSBD);
470 			setup_force_cpu_cap(X86_FEATURE_SSBD);
471 			x86_amd_ls_cfg_ssbd_mask = 1ULL << bit;
472 		}
473 	}
474 
475 	resctrl_cpu_detect(c);
476 
477 	/* Figure out Zen generations: */
478 	switch (c->x86) {
479 	case 0x17:
480 		switch (c->x86_model) {
481 		case 0x00 ... 0x2f:
482 		case 0x50 ... 0x5f:
483 			setup_force_cpu_cap(X86_FEATURE_ZEN1);
484 			break;
485 		case 0x30 ... 0x4f:
486 		case 0x60 ... 0x7f:
487 		case 0x90 ... 0x91:
488 		case 0xa0 ... 0xaf:
489 			setup_force_cpu_cap(X86_FEATURE_ZEN2);
490 			break;
491 		default:
492 			goto warn;
493 		}
494 		break;
495 
496 	case 0x19:
497 		switch (c->x86_model) {
498 		case 0x00 ... 0x0f:
499 		case 0x20 ... 0x5f:
500 			setup_force_cpu_cap(X86_FEATURE_ZEN3);
501 			break;
502 		case 0x10 ... 0x1f:
503 		case 0x60 ... 0xaf:
504 			setup_force_cpu_cap(X86_FEATURE_ZEN4);
505 			break;
506 		default:
507 			goto warn;
508 		}
509 		break;
510 
511 	case 0x1a:
512 		switch (c->x86_model) {
513 		case 0x00 ... 0x2f:
514 		case 0x40 ... 0x4f:
515 		case 0x60 ... 0x7f:
516 			setup_force_cpu_cap(X86_FEATURE_ZEN5);
517 			break;
518 		case 0x50 ... 0x5f:
519 		case 0x90 ... 0xaf:
520 		case 0xc0 ... 0xcf:
521 			setup_force_cpu_cap(X86_FEATURE_ZEN6);
522 			break;
523 		default:
524 			goto warn;
525 		}
526 		break;
527 
528 	default:
529 		break;
530 	}
531 
532 	bsp_determine_snp(c);
533 
534 	tsa_init(c);
535 
536 	return;
537 
538 warn:
539 	WARN_ONCE(1, "Family 0x%x, model: 0x%x??\n", c->x86, c->x86_model);
540 }
541 
early_detect_mem_encrypt(struct cpuinfo_x86 * c)542 static void early_detect_mem_encrypt(struct cpuinfo_x86 *c)
543 {
544 	u64 msr;
545 
546 	/*
547 	 * BIOS support is required for SME and SEV.
548 	 *   For SME: If BIOS has enabled SME then adjust x86_phys_bits by
549 	 *	      the SME physical address space reduction value.
550 	 *	      If BIOS has not enabled SME then don't advertise the
551 	 *	      SME feature (set in scattered.c).
552 	 *	      If the kernel has not enabled SME via any means then
553 	 *	      don't advertise the SME feature.
554 	 *   For SEV: If BIOS has not enabled SEV then don't advertise SEV and
555 	 *	      any additional functionality based on it.
556 	 *
557 	 *   In all cases, since support for SME and SEV requires long mode,
558 	 *   don't advertise the feature under CONFIG_X86_32.
559 	 */
560 	if (cpu_has(c, X86_FEATURE_SME) || cpu_has(c, X86_FEATURE_SEV)) {
561 		/* Check if memory encryption is enabled */
562 		rdmsrq(MSR_AMD64_SYSCFG, msr);
563 		if (!(msr & MSR_AMD64_SYSCFG_MEM_ENCRYPT))
564 			goto clear_all;
565 
566 		/*
567 		 * Always adjust physical address bits. Even though this
568 		 * will be a value above 32-bits this is still done for
569 		 * CONFIG_X86_32 so that accurate values are reported.
570 		 */
571 		c->x86_phys_bits -= (cpuid_ebx(0x8000001f) >> 6) & 0x3f;
572 
573 		if (IS_ENABLED(CONFIG_X86_32))
574 			goto clear_all;
575 
576 		if (!sme_me_mask)
577 			setup_clear_cpu_cap(X86_FEATURE_SME);
578 
579 		rdmsrq(MSR_K7_HWCR, msr);
580 		if (!(msr & MSR_K7_HWCR_SMMLOCK))
581 			goto clear_sev;
582 
583 		return;
584 
585 clear_all:
586 		setup_clear_cpu_cap(X86_FEATURE_SME);
587 clear_sev:
588 		setup_clear_cpu_cap(X86_FEATURE_SEV);
589 		setup_clear_cpu_cap(X86_FEATURE_SEV_ES);
590 		setup_clear_cpu_cap(X86_FEATURE_SEV_SNP);
591 	}
592 }
593 
early_init_amd(struct cpuinfo_x86 * c)594 static void early_init_amd(struct cpuinfo_x86 *c)
595 {
596 	u32 dummy;
597 
598 	if (c->x86 >= 0xf)
599 		set_cpu_cap(c, X86_FEATURE_K8);
600 
601 	rdmsr_safe(MSR_AMD64_PATCH_LEVEL, &c->microcode, &dummy);
602 
603 	/*
604 	 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
605 	 * with P/T states and does not stop in deep C-states
606 	 */
607 	if (c->x86_power & (1 << 8)) {
608 		set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
609 		set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
610 	}
611 
612 	/* Bit 12 of 8000_0007 edx is accumulated power mechanism. */
613 	if (c->x86_power & BIT(12))
614 		set_cpu_cap(c, X86_FEATURE_ACC_POWER);
615 
616 	/* Bit 14 indicates the Runtime Average Power Limit interface. */
617 	if (c->x86_power & BIT(14))
618 		set_cpu_cap(c, X86_FEATURE_RAPL);
619 
620 #ifdef CONFIG_X86_64
621 	set_cpu_cap(c, X86_FEATURE_SYSCALL32);
622 #else
623 	/*  Set MTRR capability flag if appropriate */
624 	if (c->x86 == 5)
625 		if (c->x86_model == 13 || c->x86_model == 9 ||
626 		    (c->x86_model == 8 && c->x86_stepping >= 8))
627 			set_cpu_cap(c, X86_FEATURE_K6_MTRR);
628 #endif
629 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
630 	/*
631 	 * ApicID can always be treated as an 8-bit value for AMD APIC versions
632 	 * >= 0x10, but even old K8s came out of reset with version 0x10. So, we
633 	 * can safely set X86_FEATURE_EXTD_APICID unconditionally for families
634 	 * after 16h.
635 	 */
636 	if (boot_cpu_has(X86_FEATURE_APIC)) {
637 		if (c->x86 > 0x16)
638 			set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
639 		else if (c->x86 >= 0xf) {
640 			/* check CPU config space for extended APIC ID */
641 			unsigned int val;
642 
643 			val = read_pci_config(0, 24, 0, 0x68);
644 			if ((val >> 17 & 0x3) == 0x3)
645 				set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
646 		}
647 	}
648 #endif
649 
650 	/*
651 	 * This is only needed to tell the kernel whether to use VMCALL
652 	 * and VMMCALL.  VMMCALL is never executed except under virt, so
653 	 * we can set it unconditionally.
654 	 */
655 	set_cpu_cap(c, X86_FEATURE_VMMCALL);
656 
657 	/* F16h erratum 793, CVE-2013-6885 */
658 	if (c->x86 == 0x16 && c->x86_model <= 0xf)
659 		msr_set_bit(MSR_AMD64_LS_CFG, 15);
660 
661 	early_detect_mem_encrypt(c);
662 
663 	if (!cpu_has(c, X86_FEATURE_HYPERVISOR) && !cpu_has(c, X86_FEATURE_IBPB_BRTYPE)) {
664 		if (c->x86 == 0x17 && boot_cpu_has(X86_FEATURE_AMD_IBPB))
665 			setup_force_cpu_cap(X86_FEATURE_IBPB_BRTYPE);
666 		else if (c->x86 >= 0x19 && !wrmsrq_safe(MSR_IA32_PRED_CMD, PRED_CMD_SBPB)) {
667 			setup_force_cpu_cap(X86_FEATURE_IBPB_BRTYPE);
668 			setup_force_cpu_cap(X86_FEATURE_SBPB);
669 		}
670 	}
671 }
672 
init_amd_k8(struct cpuinfo_x86 * c)673 static void init_amd_k8(struct cpuinfo_x86 *c)
674 {
675 	u32 level;
676 	u64 value;
677 
678 	/* On C+ stepping K8 rep microcode works well for copy/memset */
679 	level = cpuid_eax(1);
680 	if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58)
681 		set_cpu_cap(c, X86_FEATURE_REP_GOOD);
682 
683 	/*
684 	 * Some BIOSes incorrectly force this feature, but only K8 revision D
685 	 * (model = 0x14) and later actually support it.
686 	 * (AMD Erratum #110, docId: 25759).
687 	 */
688 	if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM) && !cpu_has(c, X86_FEATURE_HYPERVISOR)) {
689 		clear_cpu_cap(c, X86_FEATURE_LAHF_LM);
690 		if (!rdmsrq_amd_safe(0xc001100d, &value)) {
691 			value &= ~BIT_64(32);
692 			wrmsrq_amd_safe(0xc001100d, value);
693 		}
694 	}
695 
696 	if (!c->x86_model_id[0])
697 		strscpy(c->x86_model_id, "Hammer");
698 
699 #ifdef CONFIG_SMP
700 	/*
701 	 * Disable TLB flush filter by setting HWCR.FFDIS on K8
702 	 * bit 6 of msr C001_0015
703 	 *
704 	 * Errata 63 for SH-B3 steppings
705 	 * Errata 122 for all steppings (F+ have it disabled by default)
706 	 */
707 	msr_set_bit(MSR_K7_HWCR, 6);
708 #endif
709 	set_cpu_bug(c, X86_BUG_SWAPGS_FENCE);
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 > 0x41 ||
718 	    (c->x86_model == 0x41 && c->x86_stepping >= 0x2))
719 		setup_force_cpu_bug(X86_BUG_AMD_E400);
720 }
721 
init_amd_gh(struct cpuinfo_x86 * c)722 static void init_amd_gh(struct cpuinfo_x86 *c)
723 {
724 #ifdef CONFIG_MMCONF_FAM10H
725 	/* do this for boot cpu */
726 	if (c == &boot_cpu_data)
727 		check_enable_amd_mmconf_dmi();
728 
729 	fam10h_check_enable_mmcfg();
730 #endif
731 
732 	/*
733 	 * Disable GART TLB Walk Errors on Fam10h. We do this here because this
734 	 * is always needed when GART is enabled, even in a kernel which has no
735 	 * MCE support built in. BIOS should disable GartTlbWlk Errors already.
736 	 * If it doesn't, we do it here as suggested by the BKDG.
737 	 *
738 	 * Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=33012
739 	 */
740 	msr_set_bit(MSR_AMD64_MCx_MASK(4), 10);
741 
742 	/*
743 	 * On family 10h BIOS may not have properly enabled WC+ support, causing
744 	 * it to be converted to CD memtype. This may result in performance
745 	 * degradation for certain nested-paging guests. Prevent this conversion
746 	 * by clearing bit 24 in MSR_AMD64_BU_CFG2.
747 	 *
748 	 * NOTE: we want to use the _safe accessors so as not to #GP kvm
749 	 * guests on older kvm hosts.
750 	 */
751 	msr_clear_bit(MSR_AMD64_BU_CFG2, 24);
752 
753 	set_cpu_bug(c, X86_BUG_AMD_TLB_MMATCH);
754 
755 	/*
756 	 * Check models and steppings affected by erratum 400. This is
757 	 * used to select the proper idle routine and to enable the
758 	 * check whether the machine is affected in arch_post_acpi_subsys_init()
759 	 * which sets the X86_BUG_AMD_APIC_C1E bug depending on the MSR check.
760 	 */
761 	if (c->x86_model > 0x2 ||
762 	    (c->x86_model == 0x2 && c->x86_stepping >= 0x1))
763 		setup_force_cpu_bug(X86_BUG_AMD_E400);
764 }
765 
init_amd_ln(struct cpuinfo_x86 * c)766 static void init_amd_ln(struct cpuinfo_x86 *c)
767 {
768 	/*
769 	 * Apply erratum 665 fix unconditionally so machines without a BIOS
770 	 * fix work.
771 	 */
772 	msr_set_bit(MSR_AMD64_DE_CFG, 31);
773 }
774 
775 static bool rdrand_force;
776 
rdrand_cmdline(char * str)777 static int __init rdrand_cmdline(char *str)
778 {
779 	if (!str)
780 		return -EINVAL;
781 
782 	if (!strcmp(str, "force"))
783 		rdrand_force = true;
784 	else
785 		return -EINVAL;
786 
787 	return 0;
788 }
789 early_param("rdrand", rdrand_cmdline);
790 
clear_rdrand_cpuid_bit(struct cpuinfo_x86 * c)791 static void clear_rdrand_cpuid_bit(struct cpuinfo_x86 *c)
792 {
793 	/*
794 	 * Saving of the MSR used to hide the RDRAND support during
795 	 * suspend/resume is done by arch/x86/power/cpu.c, which is
796 	 * dependent on CONFIG_PM_SLEEP.
797 	 */
798 	if (!IS_ENABLED(CONFIG_PM_SLEEP))
799 		return;
800 
801 	/*
802 	 * The self-test can clear X86_FEATURE_RDRAND, so check for
803 	 * RDRAND support using the CPUID function directly.
804 	 */
805 	if (!(cpuid_ecx(1) & BIT(30)) || rdrand_force)
806 		return;
807 
808 	msr_clear_bit(MSR_AMD64_CPUID_FN_1, 62);
809 
810 	/*
811 	 * Verify that the CPUID change has occurred in case the kernel is
812 	 * running virtualized and the hypervisor doesn't support the MSR.
813 	 */
814 	if (cpuid_ecx(1) & BIT(30)) {
815 		pr_info_once("BIOS may not properly restore RDRAND after suspend, but hypervisor does not support hiding RDRAND via CPUID.\n");
816 		return;
817 	}
818 
819 	clear_cpu_cap(c, X86_FEATURE_RDRAND);
820 	pr_info_once("BIOS may not properly restore RDRAND after suspend, hiding RDRAND via CPUID. Use rdrand=force to reenable.\n");
821 }
822 
init_amd_jg(struct cpuinfo_x86 * c)823 static void init_amd_jg(struct cpuinfo_x86 *c)
824 {
825 	/*
826 	 * Some BIOS implementations do not restore proper RDRAND support
827 	 * across suspend and resume. Check on whether to hide the RDRAND
828 	 * instruction support via CPUID.
829 	 */
830 	clear_rdrand_cpuid_bit(c);
831 }
832 
init_amd_bd(struct cpuinfo_x86 * c)833 static void init_amd_bd(struct cpuinfo_x86 *c)
834 {
835 	u64 value;
836 
837 	/*
838 	 * The way access filter has a performance penalty on some workloads.
839 	 * Disable it on the affected CPUs.
840 	 */
841 	if ((c->x86_model >= 0x02) && (c->x86_model < 0x20)) {
842 		if (!rdmsrq_safe(MSR_F15H_IC_CFG, &value) && !(value & 0x1E)) {
843 			value |= 0x1E;
844 			wrmsrq_safe(MSR_F15H_IC_CFG, value);
845 		}
846 	}
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 const struct x86_cpu_id erratum_1386_microcode[] = {
857 	X86_MATCH_VFM_STEPS(VFM_MAKE(X86_VENDOR_AMD, 0x17, 0x01), 0x2, 0x2, 0x0800126e),
858 	X86_MATCH_VFM_STEPS(VFM_MAKE(X86_VENDOR_AMD, 0x17, 0x31), 0x0, 0x0, 0x08301052),
859 	{}
860 };
861 
fix_erratum_1386(struct cpuinfo_x86 * c)862 static void fix_erratum_1386(struct cpuinfo_x86 *c)
863 {
864 	/*
865 	 * Work around Erratum 1386.  The XSAVES instruction malfunctions in
866 	 * certain circumstances on Zen1/2 uarch, and not all parts have had
867 	 * updated microcode at the time of writing (March 2023).
868 	 *
869 	 * Affected parts all have no supervisor XSAVE states, meaning that
870 	 * the XSAVEC instruction (which works fine) is equivalent.
871 	 *
872 	 * Clear the feature flag only on microcode revisions which
873 	 * don't have the fix.
874 	 */
875 	if (x86_match_min_microcode_rev(erratum_1386_microcode))
876 		return;
877 
878 	clear_cpu_cap(c, X86_FEATURE_XSAVES);
879 }
880 
init_spectral_chicken(struct cpuinfo_x86 * c)881 void init_spectral_chicken(struct cpuinfo_x86 *c)
882 {
883 #ifdef CONFIG_MITIGATION_UNRET_ENTRY
884 	u64 value;
885 
886 	/*
887 	 * On Zen2 we offer this chicken (bit) on the altar of Speculation.
888 	 *
889 	 * This suppresses speculation from the middle of a basic block, i.e. it
890 	 * suppresses non-branch predictions.
891 	 */
892 	if (!cpu_has(c, X86_FEATURE_HYPERVISOR)) {
893 		if (!rdmsrq_safe(MSR_ZEN2_SPECTRAL_CHICKEN, &value)) {
894 			value |= MSR_ZEN2_SPECTRAL_CHICKEN_BIT;
895 			wrmsrq_safe(MSR_ZEN2_SPECTRAL_CHICKEN, value);
896 		}
897 	}
898 #endif
899 }
900 
init_amd_zen_common(void)901 static void init_amd_zen_common(void)
902 {
903 	setup_force_cpu_cap(X86_FEATURE_ZEN);
904 #ifdef CONFIG_NUMA
905 	node_reclaim_distance = 32;
906 #endif
907 }
908 
init_amd_zen1(struct cpuinfo_x86 * c)909 static void init_amd_zen1(struct cpuinfo_x86 *c)
910 {
911 	fix_erratum_1386(c);
912 
913 	/* Fix up CPUID bits, but only if not virtualised. */
914 	if (!cpu_has(c, X86_FEATURE_HYPERVISOR)) {
915 
916 		/* Erratum 1076: CPB feature bit not being set in CPUID. */
917 		if (!cpu_has(c, X86_FEATURE_CPB))
918 			set_cpu_cap(c, X86_FEATURE_CPB);
919 	}
920 
921 	pr_notice_once("AMD Zen1 DIV0 bug detected. Disable SMT for full protection.\n");
922 	setup_force_cpu_bug(X86_BUG_DIV0);
923 
924 	/*
925 	 * Turn off the Instructions Retired free counter on machines that are
926 	 * susceptible to erratum #1054 "Instructions Retired Performance
927 	 * Counter May Be Inaccurate".
928 	 */
929 	if (c->x86_model < 0x30) {
930 		msr_clear_bit(MSR_K7_HWCR, MSR_K7_HWCR_IRPERF_EN_BIT);
931 		clear_cpu_cap(c, X86_FEATURE_IRPERF);
932 	}
933 }
934 
cpu_has_zenbleed_microcode(void)935 static bool cpu_has_zenbleed_microcode(void)
936 {
937 	u32 good_rev = 0;
938 
939 	switch (boot_cpu_data.x86_model) {
940 	case 0x30 ... 0x3f: good_rev = 0x0830107b; break;
941 	case 0x60 ... 0x67: good_rev = 0x0860010c; break;
942 	case 0x68 ... 0x6f: good_rev = 0x08608107; break;
943 	case 0x70 ... 0x7f: good_rev = 0x08701033; break;
944 	case 0xa0 ... 0xaf: good_rev = 0x08a00009; break;
945 
946 	default:
947 		return false;
948 	}
949 
950 	if (boot_cpu_data.microcode < good_rev)
951 		return false;
952 
953 	return true;
954 }
955 
zen2_zenbleed_check(struct cpuinfo_x86 * c)956 static void zen2_zenbleed_check(struct cpuinfo_x86 *c)
957 {
958 	if (cpu_has(c, X86_FEATURE_HYPERVISOR))
959 		return;
960 
961 	if (!cpu_has(c, X86_FEATURE_AVX))
962 		return;
963 
964 	if (!cpu_has_zenbleed_microcode()) {
965 		pr_notice_once("Zenbleed: please update your microcode for the most optimal fix\n");
966 		msr_set_bit(MSR_AMD64_DE_CFG, MSR_AMD64_DE_CFG_ZEN2_FP_BACKUP_FIX_BIT);
967 	} else {
968 		msr_clear_bit(MSR_AMD64_DE_CFG, MSR_AMD64_DE_CFG_ZEN2_FP_BACKUP_FIX_BIT);
969 	}
970 }
971 
init_amd_zen2(struct cpuinfo_x86 * c)972 static void init_amd_zen2(struct cpuinfo_x86 *c)
973 {
974 	init_spectral_chicken(c);
975 	fix_erratum_1386(c);
976 	zen2_zenbleed_check(c);
977 }
978 
init_amd_zen3(struct cpuinfo_x86 * c)979 static void init_amd_zen3(struct cpuinfo_x86 *c)
980 {
981 	if (!cpu_has(c, X86_FEATURE_HYPERVISOR)) {
982 		/*
983 		 * Zen3 (Fam19 model < 0x10) parts are not susceptible to
984 		 * Branch Type Confusion, but predate the allocation of the
985 		 * BTC_NO bit.
986 		 */
987 		if (!cpu_has(c, X86_FEATURE_BTC_NO))
988 			set_cpu_cap(c, X86_FEATURE_BTC_NO);
989 	}
990 }
991 
init_amd_zen4(struct cpuinfo_x86 * c)992 static void init_amd_zen4(struct cpuinfo_x86 *c)
993 {
994 	if (!cpu_has(c, X86_FEATURE_HYPERVISOR))
995 		msr_set_bit(MSR_ZEN4_BP_CFG, MSR_ZEN4_BP_CFG_SHARED_BTB_FIX_BIT);
996 
997 	/*
998 	 * These Zen4 SoCs advertise support for virtualized VMLOAD/VMSAVE
999 	 * in some BIOS versions but they can lead to random host reboots.
1000 	 */
1001 	switch (c->x86_model) {
1002 	case 0x18 ... 0x1f:
1003 	case 0x60 ... 0x7f:
1004 		clear_cpu_cap(c, X86_FEATURE_V_VMSAVE_VMLOAD);
1005 		break;
1006 	}
1007 }
1008 
init_amd_zen5(struct cpuinfo_x86 * c)1009 static void init_amd_zen5(struct cpuinfo_x86 *c)
1010 {
1011 }
1012 
init_amd(struct cpuinfo_x86 * c)1013 static void init_amd(struct cpuinfo_x86 *c)
1014 {
1015 	u64 vm_cr;
1016 
1017 	early_init_amd(c);
1018 
1019 	/*
1020 	 * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
1021 	 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
1022 	 */
1023 	clear_cpu_cap(c, 0*32+31);
1024 
1025 	if (c->x86 >= 0x10)
1026 		set_cpu_cap(c, X86_FEATURE_REP_GOOD);
1027 
1028 	/* AMD FSRM also implies FSRS */
1029 	if (cpu_has(c, X86_FEATURE_FSRM))
1030 		set_cpu_cap(c, X86_FEATURE_FSRS);
1031 
1032 	/* K6s reports MCEs but don't actually have all the MSRs */
1033 	if (c->x86 < 6)
1034 		clear_cpu_cap(c, X86_FEATURE_MCE);
1035 
1036 	switch (c->x86) {
1037 	case 4:    init_amd_k5(c); break;
1038 	case 5:    init_amd_k6(c); break;
1039 	case 6:	   init_amd_k7(c); break;
1040 	case 0xf:  init_amd_k8(c); break;
1041 	case 0x10: init_amd_gh(c); break;
1042 	case 0x12: init_amd_ln(c); break;
1043 	case 0x15: init_amd_bd(c); break;
1044 	case 0x16: init_amd_jg(c); break;
1045 	}
1046 
1047 	/*
1048 	 * Save up on some future enablement work and do common Zen
1049 	 * settings.
1050 	 */
1051 	if (c->x86 >= 0x17)
1052 		init_amd_zen_common();
1053 
1054 	if (boot_cpu_has(X86_FEATURE_ZEN1))
1055 		init_amd_zen1(c);
1056 	else if (boot_cpu_has(X86_FEATURE_ZEN2))
1057 		init_amd_zen2(c);
1058 	else if (boot_cpu_has(X86_FEATURE_ZEN3))
1059 		init_amd_zen3(c);
1060 	else if (boot_cpu_has(X86_FEATURE_ZEN4))
1061 		init_amd_zen4(c);
1062 	else if (boot_cpu_has(X86_FEATURE_ZEN5))
1063 		init_amd_zen5(c);
1064 
1065 	/*
1066 	 * Enable workaround for FXSAVE leak on CPUs
1067 	 * without a XSaveErPtr feature
1068 	 */
1069 	if ((c->x86 >= 6) && (!cpu_has(c, X86_FEATURE_XSAVEERPTR)))
1070 		set_cpu_bug(c, X86_BUG_FXSAVE_LEAK);
1071 
1072 	cpu_detect_cache_sizes(c);
1073 
1074 	srat_detect_node(c);
1075 
1076 	init_amd_cacheinfo(c);
1077 
1078 	if (cpu_has(c, X86_FEATURE_SVM)) {
1079 		rdmsrq(MSR_VM_CR, vm_cr);
1080 		if (vm_cr & SVM_VM_CR_SVM_DIS_MASK) {
1081 			pr_notice_once("SVM disabled (by BIOS) in MSR_VM_CR\n");
1082 			clear_cpu_cap(c, X86_FEATURE_SVM);
1083 		}
1084 	}
1085 
1086 	if (!cpu_has(c, X86_FEATURE_LFENCE_RDTSC) && cpu_has(c, X86_FEATURE_XMM2)) {
1087 		/*
1088 		 * Use LFENCE for execution serialization.  On families which
1089 		 * don't have that MSR, LFENCE is already serializing.
1090 		 * msr_set_bit() uses the safe accessors, too, even if the MSR
1091 		 * is not present.
1092 		 */
1093 		msr_set_bit(MSR_AMD64_DE_CFG,
1094 			    MSR_AMD64_DE_CFG_LFENCE_SERIALIZE_BIT);
1095 
1096 		/* A serializing LFENCE stops RDTSC speculation */
1097 		set_cpu_cap(c, X86_FEATURE_LFENCE_RDTSC);
1098 	}
1099 
1100 	/*
1101 	 * Family 0x12 and above processors have APIC timer
1102 	 * running in deep C states.
1103 	 */
1104 	if (c->x86 > 0x11)
1105 		set_cpu_cap(c, X86_FEATURE_ARAT);
1106 
1107 	/* 3DNow or LM implies PREFETCHW */
1108 	if (!cpu_has(c, X86_FEATURE_3DNOWPREFETCH))
1109 		if (cpu_has(c, X86_FEATURE_3DNOW) || cpu_has(c, X86_FEATURE_LM))
1110 			set_cpu_cap(c, X86_FEATURE_3DNOWPREFETCH);
1111 
1112 	/* AMD CPUs don't reset SS attributes on SYSRET, Xen does. */
1113 	if (!cpu_feature_enabled(X86_FEATURE_XENPV))
1114 		set_cpu_bug(c, X86_BUG_SYSRET_SS_ATTRS);
1115 
1116 	/* Enable the Instructions Retired free counter */
1117 	if (cpu_has(c, X86_FEATURE_IRPERF))
1118 		msr_set_bit(MSR_K7_HWCR, MSR_K7_HWCR_IRPERF_EN_BIT);
1119 
1120 	check_null_seg_clears_base(c);
1121 
1122 	/*
1123 	 * Make sure EFER[AIBRSE - Automatic IBRS Enable] is set. The APs are brought up
1124 	 * using the trampoline code and as part of it, MSR_EFER gets prepared there in
1125 	 * order to be replicated onto them. Regardless, set it here again, if not set,
1126 	 * to protect against any future refactoring/code reorganization which might
1127 	 * miss setting this important bit.
1128 	 */
1129 	if (spectre_v2_in_eibrs_mode(spectre_v2_enabled) &&
1130 	    cpu_has(c, X86_FEATURE_AUTOIBRS))
1131 		WARN_ON_ONCE(msr_set_bit(MSR_EFER, _EFER_AUTOIBRS) < 0);
1132 
1133 	/* AMD CPUs don't need fencing after x2APIC/TSC_DEADLINE MSR writes. */
1134 	clear_cpu_cap(c, X86_FEATURE_APIC_MSRS_FENCE);
1135 
1136 	/* Enable Translation Cache Extension */
1137 	if (cpu_has(c, X86_FEATURE_TCE))
1138 		msr_set_bit(MSR_EFER, _EFER_TCE);
1139 }
1140 
1141 #ifdef CONFIG_X86_32
amd_size_cache(struct cpuinfo_x86 * c,unsigned int size)1142 static unsigned int amd_size_cache(struct cpuinfo_x86 *c, unsigned int size)
1143 {
1144 	/* AMD errata T13 (order #21922) */
1145 	if (c->x86 == 6) {
1146 		/* Duron Rev A0 */
1147 		if (c->x86_model == 3 && c->x86_stepping == 0)
1148 			size = 64;
1149 		/* Tbird rev A1/A2 */
1150 		if (c->x86_model == 4 &&
1151 			(c->x86_stepping == 0 || c->x86_stepping == 1))
1152 			size = 256;
1153 	}
1154 	return size;
1155 }
1156 #endif
1157 
cpu_detect_tlb_amd(struct cpuinfo_x86 * c)1158 static void cpu_detect_tlb_amd(struct cpuinfo_x86 *c)
1159 {
1160 	u32 ebx, eax, ecx, edx;
1161 	u16 mask = 0xfff;
1162 
1163 	if (c->x86 < 0xf)
1164 		return;
1165 
1166 	if (c->extended_cpuid_level < 0x80000006)
1167 		return;
1168 
1169 	cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
1170 
1171 	tlb_lld_4k = (ebx >> 16) & mask;
1172 	tlb_lli_4k = ebx & mask;
1173 
1174 	/*
1175 	 * K8 doesn't have 2M/4M entries in the L2 TLB so read out the L1 TLB
1176 	 * characteristics from the CPUID function 0x80000005 instead.
1177 	 */
1178 	if (c->x86 == 0xf) {
1179 		cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
1180 		mask = 0xff;
1181 	}
1182 
1183 	/* Handle DTLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
1184 	if (!((eax >> 16) & mask))
1185 		tlb_lld_2m = (cpuid_eax(0x80000005) >> 16) & 0xff;
1186 	else
1187 		tlb_lld_2m = (eax >> 16) & mask;
1188 
1189 	/* a 4M entry uses two 2M entries */
1190 	tlb_lld_4m = tlb_lld_2m >> 1;
1191 
1192 	/* Handle ITLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
1193 	if (!(eax & mask)) {
1194 		/* Erratum 658 */
1195 		if (c->x86 == 0x15 && c->x86_model <= 0x1f) {
1196 			tlb_lli_2m = 1024;
1197 		} else {
1198 			cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
1199 			tlb_lli_2m = eax & 0xff;
1200 		}
1201 	} else
1202 		tlb_lli_2m = eax & mask;
1203 
1204 	tlb_lli_4m = tlb_lli_2m >> 1;
1205 
1206 	/* Max number of pages INVLPGB can invalidate in one shot */
1207 	if (cpu_has(c, X86_FEATURE_INVLPGB))
1208 		invlpgb_count_max = (cpuid_edx(0x80000008) & 0xffff) + 1;
1209 }
1210 
1211 static const struct cpu_dev amd_cpu_dev = {
1212 	.c_vendor	= "AMD",
1213 	.c_ident	= { "AuthenticAMD" },
1214 #ifdef CONFIG_X86_32
1215 	.legacy_models = {
1216 		{ .family = 4, .model_names =
1217 		  {
1218 			  [3] = "486 DX/2",
1219 			  [7] = "486 DX/2-WB",
1220 			  [8] = "486 DX/4",
1221 			  [9] = "486 DX/4-WB",
1222 			  [14] = "Am5x86-WT",
1223 			  [15] = "Am5x86-WB"
1224 		  }
1225 		},
1226 	},
1227 	.legacy_cache_size = amd_size_cache,
1228 #endif
1229 	.c_early_init   = early_init_amd,
1230 	.c_detect_tlb	= cpu_detect_tlb_amd,
1231 	.c_bsp_init	= bsp_init_amd,
1232 	.c_init		= init_amd,
1233 	.c_x86_vendor	= X86_VENDOR_AMD,
1234 };
1235 
1236 cpu_dev_register(amd_cpu_dev);
1237 
1238 static DEFINE_PER_CPU_READ_MOSTLY(unsigned long[4], amd_dr_addr_mask);
1239 
1240 static unsigned int amd_msr_dr_addr_masks[] = {
1241 	MSR_F16H_DR0_ADDR_MASK,
1242 	MSR_F16H_DR1_ADDR_MASK,
1243 	MSR_F16H_DR1_ADDR_MASK + 1,
1244 	MSR_F16H_DR1_ADDR_MASK + 2
1245 };
1246 
amd_set_dr_addr_mask(unsigned long mask,unsigned int dr)1247 void amd_set_dr_addr_mask(unsigned long mask, unsigned int dr)
1248 {
1249 	int cpu = smp_processor_id();
1250 
1251 	if (!cpu_feature_enabled(X86_FEATURE_BPEXT))
1252 		return;
1253 
1254 	if (WARN_ON_ONCE(dr >= ARRAY_SIZE(amd_msr_dr_addr_masks)))
1255 		return;
1256 
1257 	if (per_cpu(amd_dr_addr_mask, cpu)[dr] == mask)
1258 		return;
1259 
1260 	wrmsrq(amd_msr_dr_addr_masks[dr], mask);
1261 	per_cpu(amd_dr_addr_mask, cpu)[dr] = mask;
1262 }
1263 
amd_get_dr_addr_mask(unsigned int dr)1264 unsigned long amd_get_dr_addr_mask(unsigned int dr)
1265 {
1266 	if (!cpu_feature_enabled(X86_FEATURE_BPEXT))
1267 		return 0;
1268 
1269 	if (WARN_ON_ONCE(dr >= ARRAY_SIZE(amd_msr_dr_addr_masks)))
1270 		return 0;
1271 
1272 	return per_cpu(amd_dr_addr_mask[dr], smp_processor_id());
1273 }
1274 EXPORT_SYMBOL_GPL(amd_get_dr_addr_mask);
1275 
zenbleed_check_cpu(void * unused)1276 static void zenbleed_check_cpu(void *unused)
1277 {
1278 	struct cpuinfo_x86 *c = &cpu_data(smp_processor_id());
1279 
1280 	zen2_zenbleed_check(c);
1281 }
1282 
amd_check_microcode(void)1283 void amd_check_microcode(void)
1284 {
1285 	if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD)
1286 		return;
1287 
1288 	if (cpu_feature_enabled(X86_FEATURE_ZEN2))
1289 		on_each_cpu(zenbleed_check_cpu, NULL, 1);
1290 }
1291 
1292 static const char * const s5_reset_reason_txt[] = {
1293 	[0]  = "thermal pin BP_THERMTRIP_L was tripped",
1294 	[1]  = "power button was pressed for 4 seconds",
1295 	[2]  = "shutdown pin was tripped",
1296 	[4]  = "remote ASF power off command was received",
1297 	[9]  = "internal CPU thermal limit was tripped",
1298 	[16] = "system reset pin BP_SYS_RST_L was tripped",
1299 	[17] = "software issued PCI reset",
1300 	[18] = "software wrote 0x4 to reset control register 0xCF9",
1301 	[19] = "software wrote 0x6 to reset control register 0xCF9",
1302 	[20] = "software wrote 0xE to reset control register 0xCF9",
1303 	[21] = "ACPI power state transition occurred",
1304 	[22] = "keyboard reset pin KB_RST_L was tripped",
1305 	[23] = "internal CPU shutdown event occurred",
1306 	[24] = "system failed to boot before failed boot timer expired",
1307 	[25] = "hardware watchdog timer expired",
1308 	[26] = "remote ASF reset command was received",
1309 	[27] = "an uncorrected error caused a data fabric sync flood event",
1310 	[29] = "FCH and MP1 failed warm reset handshake",
1311 	[30] = "a parity error occurred",
1312 	[31] = "a software sync flood event occurred",
1313 };
1314 
print_s5_reset_status_mmio(void)1315 static __init int print_s5_reset_status_mmio(void)
1316 {
1317 	unsigned long value;
1318 	void __iomem *addr;
1319 	int i;
1320 
1321 	if (!cpu_feature_enabled(X86_FEATURE_ZEN))
1322 		return 0;
1323 
1324 	addr = ioremap(FCH_PM_BASE + FCH_PM_S5_RESET_STATUS, sizeof(value));
1325 	if (!addr)
1326 		return 0;
1327 
1328 	value = ioread32(addr);
1329 	iounmap(addr);
1330 
1331 	for (i = 0; i < ARRAY_SIZE(s5_reset_reason_txt); i++) {
1332 		if (!(value & BIT(i)))
1333 			continue;
1334 
1335 		if (s5_reset_reason_txt[i]) {
1336 			pr_info("x86/amd: Previous system reset reason [0x%08lx]: %s\n",
1337 				value, s5_reset_reason_txt[i]);
1338 		}
1339 	}
1340 
1341 	return 0;
1342 }
1343 late_initcall(print_s5_reset_status_mmio);
1344