xref: /openbmc/linux/arch/x86/kernel/cpu/amd.c (revision 29c37341)
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 <asm/processor.h>
13 #include <asm/apic.h>
14 #include <asm/cacheinfo.h>
15 #include <asm/cpu.h>
16 #include <asm/spec-ctrl.h>
17 #include <asm/smp.h>
18 #include <asm/numa.h>
19 #include <asm/pci-direct.h>
20 #include <asm/delay.h>
21 #include <asm/debugreg.h>
22 #include <asm/resctrl.h>
23 
24 #ifdef CONFIG_X86_64
25 # include <asm/mmconfig.h>
26 # include <asm/set_memory.h>
27 #endif
28 
29 #include "cpu.h"
30 
31 static const int amd_erratum_383[];
32 static const int amd_erratum_400[];
33 static const int amd_erratum_1054[];
34 static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum);
35 
36 /*
37  * nodes_per_socket: Stores the number of nodes per socket.
38  * Refer to Fam15h Models 00-0fh BKDG - CPUID Fn8000_001E_ECX
39  * Node Identifiers[10:8]
40  */
41 static u32 nodes_per_socket = 1;
42 
43 static inline int rdmsrl_amd_safe(unsigned msr, unsigned long long *p)
44 {
45 	u32 gprs[8] = { 0 };
46 	int err;
47 
48 	WARN_ONCE((boot_cpu_data.x86 != 0xf),
49 		  "%s should only be used on K8!\n", __func__);
50 
51 	gprs[1] = msr;
52 	gprs[7] = 0x9c5a203a;
53 
54 	err = rdmsr_safe_regs(gprs);
55 
56 	*p = gprs[0] | ((u64)gprs[2] << 32);
57 
58 	return err;
59 }
60 
61 static inline int wrmsrl_amd_safe(unsigned msr, unsigned long long val)
62 {
63 	u32 gprs[8] = { 0 };
64 
65 	WARN_ONCE((boot_cpu_data.x86 != 0xf),
66 		  "%s should only be used on K8!\n", __func__);
67 
68 	gprs[0] = (u32)val;
69 	gprs[1] = msr;
70 	gprs[2] = val >> 32;
71 	gprs[7] = 0x9c5a203a;
72 
73 	return wrmsr_safe_regs(gprs);
74 }
75 
76 /*
77  *	B step AMD K6 before B 9730xxxx have hardware bugs that can cause
78  *	misexecution of code under Linux. Owners of such processors should
79  *	contact AMD for precise details and a CPU swap.
80  *
81  *	See	http://www.multimania.com/poulot/k6bug.html
82  *	and	section 2.6.2 of "AMD-K6 Processor Revision Guide - Model 6"
83  *		(Publication # 21266  Issue Date: August 1998)
84  *
85  *	The following test is erm.. interesting. AMD neglected to up
86  *	the chip setting when fixing the bug but they also tweaked some
87  *	performance at the same time..
88  */
89 
90 #ifdef CONFIG_X86_32
91 extern __visible void vide(void);
92 __asm__(".text\n"
93 	".globl vide\n"
94 	".type vide, @function\n"
95 	".align 4\n"
96 	"vide: ret\n");
97 #endif
98 
99 static void init_amd_k5(struct cpuinfo_x86 *c)
100 {
101 #ifdef CONFIG_X86_32
102 /*
103  * General Systems BIOSen alias the cpu frequency registers
104  * of the Elan at 0x000df000. Unfortunately, one of the Linux
105  * drivers subsequently pokes it, and changes the CPU speed.
106  * Workaround : Remove the unneeded alias.
107  */
108 #define CBAR		(0xfffc) /* Configuration Base Address  (32-bit) */
109 #define CBAR_ENB	(0x80000000)
110 #define CBAR_KEY	(0X000000CB)
111 	if (c->x86_model == 9 || c->x86_model == 10) {
112 		if (inl(CBAR) & CBAR_ENB)
113 			outl(0 | CBAR_KEY, CBAR);
114 	}
115 #endif
116 }
117 
118 static void init_amd_k6(struct cpuinfo_x86 *c)
119 {
120 #ifdef CONFIG_X86_32
121 	u32 l, h;
122 	int mbytes = get_num_physpages() >> (20-PAGE_SHIFT);
123 
124 	if (c->x86_model < 6) {
125 		/* Based on AMD doc 20734R - June 2000 */
126 		if (c->x86_model == 0) {
127 			clear_cpu_cap(c, X86_FEATURE_APIC);
128 			set_cpu_cap(c, X86_FEATURE_PGE);
129 		}
130 		return;
131 	}
132 
133 	if (c->x86_model == 6 && c->x86_stepping == 1) {
134 		const int K6_BUG_LOOP = 1000000;
135 		int n;
136 		void (*f_vide)(void);
137 		u64 d, d2;
138 
139 		pr_info("AMD K6 stepping B detected - ");
140 
141 		/*
142 		 * It looks like AMD fixed the 2.6.2 bug and improved indirect
143 		 * calls at the same time.
144 		 */
145 
146 		n = K6_BUG_LOOP;
147 		f_vide = vide;
148 		OPTIMIZER_HIDE_VAR(f_vide);
149 		d = rdtsc();
150 		while (n--)
151 			f_vide();
152 		d2 = rdtsc();
153 		d = d2-d;
154 
155 		if (d > 20*K6_BUG_LOOP)
156 			pr_cont("system stability may be impaired when more than 32 MB are used.\n");
157 		else
158 			pr_cont("probably OK (after B9730xxxx).\n");
159 	}
160 
161 	/* K6 with old style WHCR */
162 	if (c->x86_model < 8 ||
163 	   (c->x86_model == 8 && c->x86_stepping < 8)) {
164 		/* We can only write allocate on the low 508Mb */
165 		if (mbytes > 508)
166 			mbytes = 508;
167 
168 		rdmsr(MSR_K6_WHCR, l, h);
169 		if ((l&0x0000FFFF) == 0) {
170 			unsigned long flags;
171 			l = (1<<0)|((mbytes/4)<<1);
172 			local_irq_save(flags);
173 			wbinvd();
174 			wrmsr(MSR_K6_WHCR, l, h);
175 			local_irq_restore(flags);
176 			pr_info("Enabling old style K6 write allocation for %d Mb\n",
177 				mbytes);
178 		}
179 		return;
180 	}
181 
182 	if ((c->x86_model == 8 && c->x86_stepping > 7) ||
183 	     c->x86_model == 9 || c->x86_model == 13) {
184 		/* The more serious chips .. */
185 
186 		if (mbytes > 4092)
187 			mbytes = 4092;
188 
189 		rdmsr(MSR_K6_WHCR, l, h);
190 		if ((l&0xFFFF0000) == 0) {
191 			unsigned long flags;
192 			l = ((mbytes>>2)<<22)|(1<<16);
193 			local_irq_save(flags);
194 			wbinvd();
195 			wrmsr(MSR_K6_WHCR, l, h);
196 			local_irq_restore(flags);
197 			pr_info("Enabling new style K6 write allocation for %d Mb\n",
198 				mbytes);
199 		}
200 
201 		return;
202 	}
203 
204 	if (c->x86_model == 10) {
205 		/* AMD Geode LX is model 10 */
206 		/* placeholder for any needed mods */
207 		return;
208 	}
209 #endif
210 }
211 
212 static void init_amd_k7(struct cpuinfo_x86 *c)
213 {
214 #ifdef CONFIG_X86_32
215 	u32 l, h;
216 
217 	/*
218 	 * Bit 15 of Athlon specific MSR 15, needs to be 0
219 	 * to enable SSE on Palomino/Morgan/Barton CPU's.
220 	 * If the BIOS didn't enable it already, enable it here.
221 	 */
222 	if (c->x86_model >= 6 && c->x86_model <= 10) {
223 		if (!cpu_has(c, X86_FEATURE_XMM)) {
224 			pr_info("Enabling disabled K7/SSE Support.\n");
225 			msr_clear_bit(MSR_K7_HWCR, 15);
226 			set_cpu_cap(c, X86_FEATURE_XMM);
227 		}
228 	}
229 
230 	/*
231 	 * It's been determined by AMD that Athlons since model 8 stepping 1
232 	 * are more robust with CLK_CTL set to 200xxxxx instead of 600xxxxx
233 	 * As per AMD technical note 27212 0.2
234 	 */
235 	if ((c->x86_model == 8 && c->x86_stepping >= 1) || (c->x86_model > 8)) {
236 		rdmsr(MSR_K7_CLK_CTL, l, h);
237 		if ((l & 0xfff00000) != 0x20000000) {
238 			pr_info("CPU: CLK_CTL MSR was %x. Reprogramming to %x\n",
239 				l, ((l & 0x000fffff)|0x20000000));
240 			wrmsr(MSR_K7_CLK_CTL, (l & 0x000fffff)|0x20000000, h);
241 		}
242 	}
243 
244 	/* calling is from identify_secondary_cpu() ? */
245 	if (!c->cpu_index)
246 		return;
247 
248 	/*
249 	 * Certain Athlons might work (for various values of 'work') in SMP
250 	 * but they are not certified as MP capable.
251 	 */
252 	/* Athlon 660/661 is valid. */
253 	if ((c->x86_model == 6) && ((c->x86_stepping == 0) ||
254 	    (c->x86_stepping == 1)))
255 		return;
256 
257 	/* Duron 670 is valid */
258 	if ((c->x86_model == 7) && (c->x86_stepping == 0))
259 		return;
260 
261 	/*
262 	 * Athlon 662, Duron 671, and Athlon >model 7 have capability
263 	 * bit. It's worth noting that the A5 stepping (662) of some
264 	 * Athlon XP's have the MP bit set.
265 	 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for
266 	 * more.
267 	 */
268 	if (((c->x86_model == 6) && (c->x86_stepping >= 2)) ||
269 	    ((c->x86_model == 7) && (c->x86_stepping >= 1)) ||
270 	     (c->x86_model > 7))
271 		if (cpu_has(c, X86_FEATURE_MP))
272 			return;
273 
274 	/* If we get here, not a certified SMP capable AMD system. */
275 
276 	/*
277 	 * Don't taint if we are running SMP kernel on a single non-MP
278 	 * approved Athlon
279 	 */
280 	WARN_ONCE(1, "WARNING: This combination of AMD"
281 		" processors is not suitable for SMP.\n");
282 	add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_NOW_UNRELIABLE);
283 #endif
284 }
285 
286 #ifdef CONFIG_NUMA
287 /*
288  * To workaround broken NUMA config.  Read the comment in
289  * srat_detect_node().
290  */
291 static int nearby_node(int apicid)
292 {
293 	int i, node;
294 
295 	for (i = apicid - 1; i >= 0; i--) {
296 		node = __apicid_to_node[i];
297 		if (node != NUMA_NO_NODE && node_online(node))
298 			return node;
299 	}
300 	for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) {
301 		node = __apicid_to_node[i];
302 		if (node != NUMA_NO_NODE && node_online(node))
303 			return node;
304 	}
305 	return first_node(node_online_map); /* Shouldn't happen */
306 }
307 #endif
308 
309 /*
310  * Fix up cpu_core_id for pre-F17h systems to be in the
311  * [0 .. cores_per_node - 1] range. Not really needed but
312  * kept so as not to break existing setups.
313  */
314 static void legacy_fixup_core_id(struct cpuinfo_x86 *c)
315 {
316 	u32 cus_per_node;
317 
318 	if (c->x86 >= 0x17)
319 		return;
320 
321 	cus_per_node = c->x86_max_cores / nodes_per_socket;
322 	c->cpu_core_id %= cus_per_node;
323 }
324 
325 /*
326  * Fixup core topology information for
327  * (1) AMD multi-node processors
328  *     Assumption: Number of cores in each internal node is the same.
329  * (2) AMD processors supporting compute units
330  */
331 static void amd_get_topology(struct cpuinfo_x86 *c)
332 {
333 	u8 node_id;
334 	int cpu = smp_processor_id();
335 
336 	/* get information required for multi-node processors */
337 	if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
338 		int err;
339 		u32 eax, ebx, ecx, edx;
340 
341 		cpuid(0x8000001e, &eax, &ebx, &ecx, &edx);
342 
343 		node_id  = ecx & 0xff;
344 
345 		if (c->x86 == 0x15)
346 			c->cu_id = ebx & 0xff;
347 
348 		if (c->x86 >= 0x17) {
349 			c->cpu_core_id = ebx & 0xff;
350 
351 			if (smp_num_siblings > 1)
352 				c->x86_max_cores /= smp_num_siblings;
353 		}
354 
355 		/*
356 		 * In case leaf B is available, use it to derive
357 		 * topology information.
358 		 */
359 		err = detect_extended_topology(c);
360 		if (!err)
361 			c->x86_coreid_bits = get_count_order(c->x86_max_cores);
362 
363 		cacheinfo_amd_init_llc_id(c, cpu, node_id);
364 
365 	} else if (cpu_has(c, X86_FEATURE_NODEID_MSR)) {
366 		u64 value;
367 
368 		rdmsrl(MSR_FAM10H_NODE_ID, value);
369 		node_id = value & 7;
370 
371 		per_cpu(cpu_llc_id, cpu) = node_id;
372 	} else
373 		return;
374 
375 	if (nodes_per_socket > 1) {
376 		set_cpu_cap(c, X86_FEATURE_AMD_DCM);
377 		legacy_fixup_core_id(c);
378 	}
379 }
380 
381 /*
382  * On a AMD dual core setup the lower bits of the APIC id distinguish the cores.
383  * Assumes number of cores is a power of two.
384  */
385 static void amd_detect_cmp(struct cpuinfo_x86 *c)
386 {
387 	unsigned bits;
388 	int cpu = smp_processor_id();
389 
390 	bits = c->x86_coreid_bits;
391 	/* Low order bits define the core id (index of core in socket) */
392 	c->cpu_core_id = c->initial_apicid & ((1 << bits)-1);
393 	/* Convert the initial APIC ID into the socket ID */
394 	c->phys_proc_id = c->initial_apicid >> bits;
395 	/* use socket ID also for last level cache */
396 	per_cpu(cpu_llc_id, cpu) = c->phys_proc_id;
397 }
398 
399 static void amd_detect_ppin(struct cpuinfo_x86 *c)
400 {
401 	unsigned long long val;
402 
403 	if (!cpu_has(c, X86_FEATURE_AMD_PPIN))
404 		return;
405 
406 	/* When PPIN is defined in CPUID, still need to check PPIN_CTL MSR */
407 	if (rdmsrl_safe(MSR_AMD_PPIN_CTL, &val))
408 		goto clear_ppin;
409 
410 	/* PPIN is locked in disabled mode, clear feature bit */
411 	if ((val & 3UL) == 1UL)
412 		goto clear_ppin;
413 
414 	/* If PPIN is disabled, try to enable it */
415 	if (!(val & 2UL)) {
416 		wrmsrl_safe(MSR_AMD_PPIN_CTL,  val | 2UL);
417 		rdmsrl_safe(MSR_AMD_PPIN_CTL, &val);
418 	}
419 
420 	/* If PPIN_EN bit is 1, return from here; otherwise fall through */
421 	if (val & 2UL)
422 		return;
423 
424 clear_ppin:
425 	clear_cpu_cap(c, X86_FEATURE_AMD_PPIN);
426 }
427 
428 u16 amd_get_nb_id(int cpu)
429 {
430 	return per_cpu(cpu_llc_id, cpu);
431 }
432 EXPORT_SYMBOL_GPL(amd_get_nb_id);
433 
434 u32 amd_get_nodes_per_socket(void)
435 {
436 	return nodes_per_socket;
437 }
438 EXPORT_SYMBOL_GPL(amd_get_nodes_per_socket);
439 
440 static void srat_detect_node(struct cpuinfo_x86 *c)
441 {
442 #ifdef CONFIG_NUMA
443 	int cpu = smp_processor_id();
444 	int node;
445 	unsigned apicid = c->apicid;
446 
447 	node = numa_cpu_node(cpu);
448 	if (node == NUMA_NO_NODE)
449 		node = per_cpu(cpu_llc_id, cpu);
450 
451 	/*
452 	 * On multi-fabric platform (e.g. Numascale NumaChip) a
453 	 * platform-specific handler needs to be called to fixup some
454 	 * IDs of the CPU.
455 	 */
456 	if (x86_cpuinit.fixup_cpu_id)
457 		x86_cpuinit.fixup_cpu_id(c, node);
458 
459 	if (!node_online(node)) {
460 		/*
461 		 * Two possibilities here:
462 		 *
463 		 * - The CPU is missing memory and no node was created.  In
464 		 *   that case try picking one from a nearby CPU.
465 		 *
466 		 * - The APIC IDs differ from the HyperTransport node IDs
467 		 *   which the K8 northbridge parsing fills in.  Assume
468 		 *   they are all increased by a constant offset, but in
469 		 *   the same order as the HT nodeids.  If that doesn't
470 		 *   result in a usable node fall back to the path for the
471 		 *   previous case.
472 		 *
473 		 * This workaround operates directly on the mapping between
474 		 * APIC ID and NUMA node, assuming certain relationship
475 		 * between APIC ID, HT node ID and NUMA topology.  As going
476 		 * through CPU mapping may alter the outcome, directly
477 		 * access __apicid_to_node[].
478 		 */
479 		int ht_nodeid = c->initial_apicid;
480 
481 		if (__apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
482 			node = __apicid_to_node[ht_nodeid];
483 		/* Pick a nearby node */
484 		if (!node_online(node))
485 			node = nearby_node(apicid);
486 	}
487 	numa_set_node(cpu, node);
488 #endif
489 }
490 
491 static void early_init_amd_mc(struct cpuinfo_x86 *c)
492 {
493 #ifdef CONFIG_SMP
494 	unsigned bits, ecx;
495 
496 	/* Multi core CPU? */
497 	if (c->extended_cpuid_level < 0x80000008)
498 		return;
499 
500 	ecx = cpuid_ecx(0x80000008);
501 
502 	c->x86_max_cores = (ecx & 0xff) + 1;
503 
504 	/* CPU telling us the core id bits shift? */
505 	bits = (ecx >> 12) & 0xF;
506 
507 	/* Otherwise recompute */
508 	if (bits == 0) {
509 		while ((1 << bits) < c->x86_max_cores)
510 			bits++;
511 	}
512 
513 	c->x86_coreid_bits = bits;
514 #endif
515 }
516 
517 static void bsp_init_amd(struct cpuinfo_x86 *c)
518 {
519 
520 #ifdef CONFIG_X86_64
521 	if (c->x86 >= 0xf) {
522 		unsigned long long tseg;
523 
524 		/*
525 		 * Split up direct mapping around the TSEG SMM area.
526 		 * Don't do it for gbpages because there seems very little
527 		 * benefit in doing so.
528 		 */
529 		if (!rdmsrl_safe(MSR_K8_TSEG_ADDR, &tseg)) {
530 			unsigned long pfn = tseg >> PAGE_SHIFT;
531 
532 			pr_debug("tseg: %010llx\n", tseg);
533 			if (pfn_range_is_mapped(pfn, pfn + 1))
534 				set_memory_4k((unsigned long)__va(tseg), 1);
535 		}
536 	}
537 #endif
538 
539 	if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) {
540 
541 		if (c->x86 > 0x10 ||
542 		    (c->x86 == 0x10 && c->x86_model >= 0x2)) {
543 			u64 val;
544 
545 			rdmsrl(MSR_K7_HWCR, val);
546 			if (!(val & BIT(24)))
547 				pr_warn(FW_BUG "TSC doesn't count with P0 frequency!\n");
548 		}
549 	}
550 
551 	if (c->x86 == 0x15) {
552 		unsigned long upperbit;
553 		u32 cpuid, assoc;
554 
555 		cpuid	 = cpuid_edx(0x80000005);
556 		assoc	 = cpuid >> 16 & 0xff;
557 		upperbit = ((cpuid >> 24) << 10) / assoc;
558 
559 		va_align.mask	  = (upperbit - 1) & PAGE_MASK;
560 		va_align.flags    = ALIGN_VA_32 | ALIGN_VA_64;
561 
562 		/* A random value per boot for bit slice [12:upper_bit) */
563 		va_align.bits = get_random_int() & va_align.mask;
564 	}
565 
566 	if (cpu_has(c, X86_FEATURE_MWAITX))
567 		use_mwaitx_delay();
568 
569 	if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
570 		u32 ecx;
571 
572 		ecx = cpuid_ecx(0x8000001e);
573 		nodes_per_socket = ((ecx >> 8) & 7) + 1;
574 	} else if (boot_cpu_has(X86_FEATURE_NODEID_MSR)) {
575 		u64 value;
576 
577 		rdmsrl(MSR_FAM10H_NODE_ID, value);
578 		nodes_per_socket = ((value >> 3) & 7) + 1;
579 	}
580 
581 	if (!boot_cpu_has(X86_FEATURE_AMD_SSBD) &&
582 	    !boot_cpu_has(X86_FEATURE_VIRT_SSBD) &&
583 	    c->x86 >= 0x15 && c->x86 <= 0x17) {
584 		unsigned int bit;
585 
586 		switch (c->x86) {
587 		case 0x15: bit = 54; break;
588 		case 0x16: bit = 33; break;
589 		case 0x17: bit = 10; break;
590 		default: return;
591 		}
592 		/*
593 		 * Try to cache the base value so further operations can
594 		 * avoid RMW. If that faults, do not enable SSBD.
595 		 */
596 		if (!rdmsrl_safe(MSR_AMD64_LS_CFG, &x86_amd_ls_cfg_base)) {
597 			setup_force_cpu_cap(X86_FEATURE_LS_CFG_SSBD);
598 			setup_force_cpu_cap(X86_FEATURE_SSBD);
599 			x86_amd_ls_cfg_ssbd_mask = 1ULL << bit;
600 		}
601 	}
602 
603 	resctrl_cpu_detect(c);
604 }
605 
606 static void early_detect_mem_encrypt(struct cpuinfo_x86 *c)
607 {
608 	u64 msr;
609 
610 	/*
611 	 * BIOS support is required for SME and SEV.
612 	 *   For SME: If BIOS has enabled SME then adjust x86_phys_bits by
613 	 *	      the SME physical address space reduction value.
614 	 *	      If BIOS has not enabled SME then don't advertise the
615 	 *	      SME feature (set in scattered.c).
616 	 *   For SEV: If BIOS has not enabled SEV then don't advertise the
617 	 *            SEV feature (set in scattered.c).
618 	 *
619 	 *   In all cases, since support for SME and SEV requires long mode,
620 	 *   don't advertise the feature under CONFIG_X86_32.
621 	 */
622 	if (cpu_has(c, X86_FEATURE_SME) || cpu_has(c, X86_FEATURE_SEV)) {
623 		/* Check if memory encryption is enabled */
624 		rdmsrl(MSR_K8_SYSCFG, msr);
625 		if (!(msr & MSR_K8_SYSCFG_MEM_ENCRYPT))
626 			goto clear_all;
627 
628 		/*
629 		 * Always adjust physical address bits. Even though this
630 		 * will be a value above 32-bits this is still done for
631 		 * CONFIG_X86_32 so that accurate values are reported.
632 		 */
633 		c->x86_phys_bits -= (cpuid_ebx(0x8000001f) >> 6) & 0x3f;
634 
635 		if (IS_ENABLED(CONFIG_X86_32))
636 			goto clear_all;
637 
638 		rdmsrl(MSR_K7_HWCR, msr);
639 		if (!(msr & MSR_K7_HWCR_SMMLOCK))
640 			goto clear_sev;
641 
642 		return;
643 
644 clear_all:
645 		setup_clear_cpu_cap(X86_FEATURE_SME);
646 clear_sev:
647 		setup_clear_cpu_cap(X86_FEATURE_SEV);
648 	}
649 }
650 
651 static void early_init_amd(struct cpuinfo_x86 *c)
652 {
653 	u64 value;
654 	u32 dummy;
655 
656 	early_init_amd_mc(c);
657 
658 #ifdef CONFIG_X86_32
659 	if (c->x86 == 6)
660 		set_cpu_cap(c, X86_FEATURE_K7);
661 #endif
662 
663 	if (c->x86 >= 0xf)
664 		set_cpu_cap(c, X86_FEATURE_K8);
665 
666 	rdmsr_safe(MSR_AMD64_PATCH_LEVEL, &c->microcode, &dummy);
667 
668 	/*
669 	 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
670 	 * with P/T states and does not stop in deep C-states
671 	 */
672 	if (c->x86_power & (1 << 8)) {
673 		set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
674 		set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
675 	}
676 
677 	/* Bit 12 of 8000_0007 edx is accumulated power mechanism. */
678 	if (c->x86_power & BIT(12))
679 		set_cpu_cap(c, X86_FEATURE_ACC_POWER);
680 
681 #ifdef CONFIG_X86_64
682 	set_cpu_cap(c, X86_FEATURE_SYSCALL32);
683 #else
684 	/*  Set MTRR capability flag if appropriate */
685 	if (c->x86 == 5)
686 		if (c->x86_model == 13 || c->x86_model == 9 ||
687 		    (c->x86_model == 8 && c->x86_stepping >= 8))
688 			set_cpu_cap(c, X86_FEATURE_K6_MTRR);
689 #endif
690 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
691 	/*
692 	 * ApicID can always be treated as an 8-bit value for AMD APIC versions
693 	 * >= 0x10, but even old K8s came out of reset with version 0x10. So, we
694 	 * can safely set X86_FEATURE_EXTD_APICID unconditionally for families
695 	 * after 16h.
696 	 */
697 	if (boot_cpu_has(X86_FEATURE_APIC)) {
698 		if (c->x86 > 0x16)
699 			set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
700 		else if (c->x86 >= 0xf) {
701 			/* check CPU config space for extended APIC ID */
702 			unsigned int val;
703 
704 			val = read_pci_config(0, 24, 0, 0x68);
705 			if ((val >> 17 & 0x3) == 0x3)
706 				set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
707 		}
708 	}
709 #endif
710 
711 	/*
712 	 * This is only needed to tell the kernel whether to use VMCALL
713 	 * and VMMCALL.  VMMCALL is never executed except under virt, so
714 	 * we can set it unconditionally.
715 	 */
716 	set_cpu_cap(c, X86_FEATURE_VMMCALL);
717 
718 	/* F16h erratum 793, CVE-2013-6885 */
719 	if (c->x86 == 0x16 && c->x86_model <= 0xf)
720 		msr_set_bit(MSR_AMD64_LS_CFG, 15);
721 
722 	/*
723 	 * Check whether the machine is affected by erratum 400. This is
724 	 * used to select the proper idle routine and to enable the check
725 	 * whether the machine is affected in arch_post_acpi_init(), which
726 	 * sets the X86_BUG_AMD_APIC_C1E bug depending on the MSR check.
727 	 */
728 	if (cpu_has_amd_erratum(c, amd_erratum_400))
729 		set_cpu_bug(c, X86_BUG_AMD_E400);
730 
731 	early_detect_mem_encrypt(c);
732 
733 	/* Re-enable TopologyExtensions if switched off by BIOS */
734 	if (c->x86 == 0x15 &&
735 	    (c->x86_model >= 0x10 && c->x86_model <= 0x6f) &&
736 	    !cpu_has(c, X86_FEATURE_TOPOEXT)) {
737 
738 		if (msr_set_bit(0xc0011005, 54) > 0) {
739 			rdmsrl(0xc0011005, value);
740 			if (value & BIT_64(54)) {
741 				set_cpu_cap(c, X86_FEATURE_TOPOEXT);
742 				pr_info_once(FW_INFO "CPU: Re-enabling disabled Topology Extensions Support.\n");
743 			}
744 		}
745 	}
746 
747 	if (cpu_has(c, X86_FEATURE_TOPOEXT))
748 		smp_num_siblings = ((cpuid_ebx(0x8000001e) >> 8) & 0xff) + 1;
749 }
750 
751 static void init_amd_k8(struct cpuinfo_x86 *c)
752 {
753 	u32 level;
754 	u64 value;
755 
756 	/* On C+ stepping K8 rep microcode works well for copy/memset */
757 	level = cpuid_eax(1);
758 	if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58)
759 		set_cpu_cap(c, X86_FEATURE_REP_GOOD);
760 
761 	/*
762 	 * Some BIOSes incorrectly force this feature, but only K8 revision D
763 	 * (model = 0x14) and later actually support it.
764 	 * (AMD Erratum #110, docId: 25759).
765 	 */
766 	if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM)) {
767 		clear_cpu_cap(c, X86_FEATURE_LAHF_LM);
768 		if (!rdmsrl_amd_safe(0xc001100d, &value)) {
769 			value &= ~BIT_64(32);
770 			wrmsrl_amd_safe(0xc001100d, value);
771 		}
772 	}
773 
774 	if (!c->x86_model_id[0])
775 		strcpy(c->x86_model_id, "Hammer");
776 
777 #ifdef CONFIG_SMP
778 	/*
779 	 * Disable TLB flush filter by setting HWCR.FFDIS on K8
780 	 * bit 6 of msr C001_0015
781 	 *
782 	 * Errata 63 for SH-B3 steppings
783 	 * Errata 122 for all steppings (F+ have it disabled by default)
784 	 */
785 	msr_set_bit(MSR_K7_HWCR, 6);
786 #endif
787 	set_cpu_bug(c, X86_BUG_SWAPGS_FENCE);
788 }
789 
790 static void init_amd_gh(struct cpuinfo_x86 *c)
791 {
792 #ifdef CONFIG_MMCONF_FAM10H
793 	/* do this for boot cpu */
794 	if (c == &boot_cpu_data)
795 		check_enable_amd_mmconf_dmi();
796 
797 	fam10h_check_enable_mmcfg();
798 #endif
799 
800 	/*
801 	 * Disable GART TLB Walk Errors on Fam10h. We do this here because this
802 	 * is always needed when GART is enabled, even in a kernel which has no
803 	 * MCE support built in. BIOS should disable GartTlbWlk Errors already.
804 	 * If it doesn't, we do it here as suggested by the BKDG.
805 	 *
806 	 * Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=33012
807 	 */
808 	msr_set_bit(MSR_AMD64_MCx_MASK(4), 10);
809 
810 	/*
811 	 * On family 10h BIOS may not have properly enabled WC+ support, causing
812 	 * it to be converted to CD memtype. This may result in performance
813 	 * degradation for certain nested-paging guests. Prevent this conversion
814 	 * by clearing bit 24 in MSR_AMD64_BU_CFG2.
815 	 *
816 	 * NOTE: we want to use the _safe accessors so as not to #GP kvm
817 	 * guests on older kvm hosts.
818 	 */
819 	msr_clear_bit(MSR_AMD64_BU_CFG2, 24);
820 
821 	if (cpu_has_amd_erratum(c, amd_erratum_383))
822 		set_cpu_bug(c, X86_BUG_AMD_TLB_MMATCH);
823 }
824 
825 #define MSR_AMD64_DE_CFG	0xC0011029
826 
827 static void init_amd_ln(struct cpuinfo_x86 *c)
828 {
829 	/*
830 	 * Apply erratum 665 fix unconditionally so machines without a BIOS
831 	 * fix work.
832 	 */
833 	msr_set_bit(MSR_AMD64_DE_CFG, 31);
834 }
835 
836 static bool rdrand_force;
837 
838 static int __init rdrand_cmdline(char *str)
839 {
840 	if (!str)
841 		return -EINVAL;
842 
843 	if (!strcmp(str, "force"))
844 		rdrand_force = true;
845 	else
846 		return -EINVAL;
847 
848 	return 0;
849 }
850 early_param("rdrand", rdrand_cmdline);
851 
852 static void clear_rdrand_cpuid_bit(struct cpuinfo_x86 *c)
853 {
854 	/*
855 	 * Saving of the MSR used to hide the RDRAND support during
856 	 * suspend/resume is done by arch/x86/power/cpu.c, which is
857 	 * dependent on CONFIG_PM_SLEEP.
858 	 */
859 	if (!IS_ENABLED(CONFIG_PM_SLEEP))
860 		return;
861 
862 	/*
863 	 * The nordrand option can clear X86_FEATURE_RDRAND, so check for
864 	 * RDRAND support using the CPUID function directly.
865 	 */
866 	if (!(cpuid_ecx(1) & BIT(30)) || rdrand_force)
867 		return;
868 
869 	msr_clear_bit(MSR_AMD64_CPUID_FN_1, 62);
870 
871 	/*
872 	 * Verify that the CPUID change has occurred in case the kernel is
873 	 * running virtualized and the hypervisor doesn't support the MSR.
874 	 */
875 	if (cpuid_ecx(1) & BIT(30)) {
876 		pr_info_once("BIOS may not properly restore RDRAND after suspend, but hypervisor does not support hiding RDRAND via CPUID.\n");
877 		return;
878 	}
879 
880 	clear_cpu_cap(c, X86_FEATURE_RDRAND);
881 	pr_info_once("BIOS may not properly restore RDRAND after suspend, hiding RDRAND via CPUID. Use rdrand=force to reenable.\n");
882 }
883 
884 static void init_amd_jg(struct cpuinfo_x86 *c)
885 {
886 	/*
887 	 * Some BIOS implementations do not restore proper RDRAND support
888 	 * across suspend and resume. Check on whether to hide the RDRAND
889 	 * instruction support via CPUID.
890 	 */
891 	clear_rdrand_cpuid_bit(c);
892 }
893 
894 static void init_amd_bd(struct cpuinfo_x86 *c)
895 {
896 	u64 value;
897 
898 	/*
899 	 * The way access filter has a performance penalty on some workloads.
900 	 * Disable it on the affected CPUs.
901 	 */
902 	if ((c->x86_model >= 0x02) && (c->x86_model < 0x20)) {
903 		if (!rdmsrl_safe(MSR_F15H_IC_CFG, &value) && !(value & 0x1E)) {
904 			value |= 0x1E;
905 			wrmsrl_safe(MSR_F15H_IC_CFG, value);
906 		}
907 	}
908 
909 	/*
910 	 * Some BIOS implementations do not restore proper RDRAND support
911 	 * across suspend and resume. Check on whether to hide the RDRAND
912 	 * instruction support via CPUID.
913 	 */
914 	clear_rdrand_cpuid_bit(c);
915 }
916 
917 static void init_amd_zn(struct cpuinfo_x86 *c)
918 {
919 	set_cpu_cap(c, X86_FEATURE_ZEN);
920 
921 #ifdef CONFIG_NUMA
922 	node_reclaim_distance = 32;
923 #endif
924 
925 	/*
926 	 * Fix erratum 1076: CPB feature bit not being set in CPUID.
927 	 * Always set it, except when running under a hypervisor.
928 	 */
929 	if (!cpu_has(c, X86_FEATURE_HYPERVISOR) && !cpu_has(c, X86_FEATURE_CPB))
930 		set_cpu_cap(c, X86_FEATURE_CPB);
931 }
932 
933 static void init_amd(struct cpuinfo_x86 *c)
934 {
935 	early_init_amd(c);
936 
937 	/*
938 	 * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
939 	 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
940 	 */
941 	clear_cpu_cap(c, 0*32+31);
942 
943 	if (c->x86 >= 0x10)
944 		set_cpu_cap(c, X86_FEATURE_REP_GOOD);
945 
946 	/* get apicid instead of initial apic id from cpuid */
947 	c->apicid = hard_smp_processor_id();
948 
949 	/* K6s reports MCEs but don't actually have all the MSRs */
950 	if (c->x86 < 6)
951 		clear_cpu_cap(c, X86_FEATURE_MCE);
952 
953 	switch (c->x86) {
954 	case 4:    init_amd_k5(c); break;
955 	case 5:    init_amd_k6(c); break;
956 	case 6:	   init_amd_k7(c); break;
957 	case 0xf:  init_amd_k8(c); break;
958 	case 0x10: init_amd_gh(c); break;
959 	case 0x12: init_amd_ln(c); break;
960 	case 0x15: init_amd_bd(c); break;
961 	case 0x16: init_amd_jg(c); break;
962 	case 0x17: fallthrough;
963 	case 0x19: init_amd_zn(c); break;
964 	}
965 
966 	/*
967 	 * Enable workaround for FXSAVE leak on CPUs
968 	 * without a XSaveErPtr feature
969 	 */
970 	if ((c->x86 >= 6) && (!cpu_has(c, X86_FEATURE_XSAVEERPTR)))
971 		set_cpu_bug(c, X86_BUG_FXSAVE_LEAK);
972 
973 	cpu_detect_cache_sizes(c);
974 
975 	amd_detect_cmp(c);
976 	amd_get_topology(c);
977 	srat_detect_node(c);
978 	amd_detect_ppin(c);
979 
980 	init_amd_cacheinfo(c);
981 
982 	if (cpu_has(c, X86_FEATURE_XMM2)) {
983 		/*
984 		 * Use LFENCE for execution serialization.  On families which
985 		 * don't have that MSR, LFENCE is already serializing.
986 		 * msr_set_bit() uses the safe accessors, too, even if the MSR
987 		 * is not present.
988 		 */
989 		msr_set_bit(MSR_F10H_DECFG,
990 			    MSR_F10H_DECFG_LFENCE_SERIALIZE_BIT);
991 
992 		/* A serializing LFENCE stops RDTSC speculation */
993 		set_cpu_cap(c, X86_FEATURE_LFENCE_RDTSC);
994 	}
995 
996 	/*
997 	 * Family 0x12 and above processors have APIC timer
998 	 * running in deep C states.
999 	 */
1000 	if (c->x86 > 0x11)
1001 		set_cpu_cap(c, X86_FEATURE_ARAT);
1002 
1003 	/* 3DNow or LM implies PREFETCHW */
1004 	if (!cpu_has(c, X86_FEATURE_3DNOWPREFETCH))
1005 		if (cpu_has(c, X86_FEATURE_3DNOW) || cpu_has(c, X86_FEATURE_LM))
1006 			set_cpu_cap(c, X86_FEATURE_3DNOWPREFETCH);
1007 
1008 	/* AMD CPUs don't reset SS attributes on SYSRET, Xen does. */
1009 	if (!cpu_has(c, X86_FEATURE_XENPV))
1010 		set_cpu_bug(c, X86_BUG_SYSRET_SS_ATTRS);
1011 
1012 	/*
1013 	 * Turn on the Instructions Retired free counter on machines not
1014 	 * susceptible to erratum #1054 "Instructions Retired Performance
1015 	 * Counter May Be Inaccurate".
1016 	 */
1017 	if (cpu_has(c, X86_FEATURE_IRPERF) &&
1018 	    !cpu_has_amd_erratum(c, amd_erratum_1054))
1019 		msr_set_bit(MSR_K7_HWCR, MSR_K7_HWCR_IRPERF_EN_BIT);
1020 }
1021 
1022 #ifdef CONFIG_X86_32
1023 static unsigned int amd_size_cache(struct cpuinfo_x86 *c, unsigned int size)
1024 {
1025 	/* AMD errata T13 (order #21922) */
1026 	if (c->x86 == 6) {
1027 		/* Duron Rev A0 */
1028 		if (c->x86_model == 3 && c->x86_stepping == 0)
1029 			size = 64;
1030 		/* Tbird rev A1/A2 */
1031 		if (c->x86_model == 4 &&
1032 			(c->x86_stepping == 0 || c->x86_stepping == 1))
1033 			size = 256;
1034 	}
1035 	return size;
1036 }
1037 #endif
1038 
1039 static void cpu_detect_tlb_amd(struct cpuinfo_x86 *c)
1040 {
1041 	u32 ebx, eax, ecx, edx;
1042 	u16 mask = 0xfff;
1043 
1044 	if (c->x86 < 0xf)
1045 		return;
1046 
1047 	if (c->extended_cpuid_level < 0x80000006)
1048 		return;
1049 
1050 	cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
1051 
1052 	tlb_lld_4k[ENTRIES] = (ebx >> 16) & mask;
1053 	tlb_lli_4k[ENTRIES] = ebx & mask;
1054 
1055 	/*
1056 	 * K8 doesn't have 2M/4M entries in the L2 TLB so read out the L1 TLB
1057 	 * characteristics from the CPUID function 0x80000005 instead.
1058 	 */
1059 	if (c->x86 == 0xf) {
1060 		cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
1061 		mask = 0xff;
1062 	}
1063 
1064 	/* Handle DTLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
1065 	if (!((eax >> 16) & mask))
1066 		tlb_lld_2m[ENTRIES] = (cpuid_eax(0x80000005) >> 16) & 0xff;
1067 	else
1068 		tlb_lld_2m[ENTRIES] = (eax >> 16) & mask;
1069 
1070 	/* a 4M entry uses two 2M entries */
1071 	tlb_lld_4m[ENTRIES] = tlb_lld_2m[ENTRIES] >> 1;
1072 
1073 	/* Handle ITLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
1074 	if (!(eax & mask)) {
1075 		/* Erratum 658 */
1076 		if (c->x86 == 0x15 && c->x86_model <= 0x1f) {
1077 			tlb_lli_2m[ENTRIES] = 1024;
1078 		} else {
1079 			cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
1080 			tlb_lli_2m[ENTRIES] = eax & 0xff;
1081 		}
1082 	} else
1083 		tlb_lli_2m[ENTRIES] = eax & mask;
1084 
1085 	tlb_lli_4m[ENTRIES] = tlb_lli_2m[ENTRIES] >> 1;
1086 }
1087 
1088 static const struct cpu_dev amd_cpu_dev = {
1089 	.c_vendor	= "AMD",
1090 	.c_ident	= { "AuthenticAMD" },
1091 #ifdef CONFIG_X86_32
1092 	.legacy_models = {
1093 		{ .family = 4, .model_names =
1094 		  {
1095 			  [3] = "486 DX/2",
1096 			  [7] = "486 DX/2-WB",
1097 			  [8] = "486 DX/4",
1098 			  [9] = "486 DX/4-WB",
1099 			  [14] = "Am5x86-WT",
1100 			  [15] = "Am5x86-WB"
1101 		  }
1102 		},
1103 	},
1104 	.legacy_cache_size = amd_size_cache,
1105 #endif
1106 	.c_early_init   = early_init_amd,
1107 	.c_detect_tlb	= cpu_detect_tlb_amd,
1108 	.c_bsp_init	= bsp_init_amd,
1109 	.c_init		= init_amd,
1110 	.c_x86_vendor	= X86_VENDOR_AMD,
1111 };
1112 
1113 cpu_dev_register(amd_cpu_dev);
1114 
1115 /*
1116  * AMD errata checking
1117  *
1118  * Errata are defined as arrays of ints using the AMD_LEGACY_ERRATUM() or
1119  * AMD_OSVW_ERRATUM() macros. The latter is intended for newer errata that
1120  * have an OSVW id assigned, which it takes as first argument. Both take a
1121  * variable number of family-specific model-stepping ranges created by
1122  * AMD_MODEL_RANGE().
1123  *
1124  * Example:
1125  *
1126  * const int amd_erratum_319[] =
1127  *	AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0x4, 0x2),
1128  *			   AMD_MODEL_RANGE(0x10, 0x8, 0x0, 0x8, 0x0),
1129  *			   AMD_MODEL_RANGE(0x10, 0x9, 0x0, 0x9, 0x0));
1130  */
1131 
1132 #define AMD_LEGACY_ERRATUM(...)		{ -1, __VA_ARGS__, 0 }
1133 #define AMD_OSVW_ERRATUM(osvw_id, ...)	{ osvw_id, __VA_ARGS__, 0 }
1134 #define AMD_MODEL_RANGE(f, m_start, s_start, m_end, s_end) \
1135 	((f << 24) | (m_start << 16) | (s_start << 12) | (m_end << 4) | (s_end))
1136 #define AMD_MODEL_RANGE_FAMILY(range)	(((range) >> 24) & 0xff)
1137 #define AMD_MODEL_RANGE_START(range)	(((range) >> 12) & 0xfff)
1138 #define AMD_MODEL_RANGE_END(range)	((range) & 0xfff)
1139 
1140 static const int amd_erratum_400[] =
1141 	AMD_OSVW_ERRATUM(1, AMD_MODEL_RANGE(0xf, 0x41, 0x2, 0xff, 0xf),
1142 			    AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0xff, 0xf));
1143 
1144 static const int amd_erratum_383[] =
1145 	AMD_OSVW_ERRATUM(3, AMD_MODEL_RANGE(0x10, 0, 0, 0xff, 0xf));
1146 
1147 /* #1054: Instructions Retired Performance Counter May Be Inaccurate */
1148 static const int amd_erratum_1054[] =
1149 	AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x17, 0, 0, 0x2f, 0xf));
1150 
1151 static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum)
1152 {
1153 	int osvw_id = *erratum++;
1154 	u32 range;
1155 	u32 ms;
1156 
1157 	if (osvw_id >= 0 && osvw_id < 65536 &&
1158 	    cpu_has(cpu, X86_FEATURE_OSVW)) {
1159 		u64 osvw_len;
1160 
1161 		rdmsrl(MSR_AMD64_OSVW_ID_LENGTH, osvw_len);
1162 		if (osvw_id < osvw_len) {
1163 			u64 osvw_bits;
1164 
1165 			rdmsrl(MSR_AMD64_OSVW_STATUS + (osvw_id >> 6),
1166 			    osvw_bits);
1167 			return osvw_bits & (1ULL << (osvw_id & 0x3f));
1168 		}
1169 	}
1170 
1171 	/* OSVW unavailable or ID unknown, match family-model-stepping range */
1172 	ms = (cpu->x86_model << 4) | cpu->x86_stepping;
1173 	while ((range = *erratum++))
1174 		if ((cpu->x86 == AMD_MODEL_RANGE_FAMILY(range)) &&
1175 		    (ms >= AMD_MODEL_RANGE_START(range)) &&
1176 		    (ms <= AMD_MODEL_RANGE_END(range)))
1177 			return true;
1178 
1179 	return false;
1180 }
1181 
1182 void set_dr_addr_mask(unsigned long mask, int dr)
1183 {
1184 	if (!boot_cpu_has(X86_FEATURE_BPEXT))
1185 		return;
1186 
1187 	switch (dr) {
1188 	case 0:
1189 		wrmsr(MSR_F16H_DR0_ADDR_MASK, mask, 0);
1190 		break;
1191 	case 1:
1192 	case 2:
1193 	case 3:
1194 		wrmsr(MSR_F16H_DR1_ADDR_MASK - 1 + dr, mask, 0);
1195 		break;
1196 	default:
1197 		break;
1198 	}
1199 }
1200