xref: /linux/arch/x86/kernel/cpu/amd.c (revision 37744feebc086908fd89760650f458ab19071750)
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/pci-direct.h>
19 #include <asm/delay.h>
20 #include <asm/debugreg.h>
21 
22 #ifdef CONFIG_X86_64
23 # include <asm/mmconfig.h>
24 # include <asm/set_memory.h>
25 #endif
26 
27 #include "cpu.h"
28 
29 static const int amd_erratum_383[];
30 static const int amd_erratum_400[];
31 static const int amd_erratum_1054[];
32 static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum);
33 
34 /*
35  * nodes_per_socket: Stores the number of nodes per socket.
36  * Refer to Fam15h Models 00-0fh BKDG - CPUID Fn8000_001E_ECX
37  * Node Identifiers[10:8]
38  */
39 static u32 nodes_per_socket = 1;
40 
41 static inline int rdmsrl_amd_safe(unsigned msr, unsigned long long *p)
42 {
43 	u32 gprs[8] = { 0 };
44 	int err;
45 
46 	WARN_ONCE((boot_cpu_data.x86 != 0xf),
47 		  "%s should only be used on K8!\n", __func__);
48 
49 	gprs[1] = msr;
50 	gprs[7] = 0x9c5a203a;
51 
52 	err = rdmsr_safe_regs(gprs);
53 
54 	*p = gprs[0] | ((u64)gprs[2] << 32);
55 
56 	return err;
57 }
58 
59 static inline int wrmsrl_amd_safe(unsigned msr, unsigned long long val)
60 {
61 	u32 gprs[8] = { 0 };
62 
63 	WARN_ONCE((boot_cpu_data.x86 != 0xf),
64 		  "%s should only be used on K8!\n", __func__);
65 
66 	gprs[0] = (u32)val;
67 	gprs[1] = msr;
68 	gprs[2] = val >> 32;
69 	gprs[7] = 0x9c5a203a;
70 
71 	return wrmsr_safe_regs(gprs);
72 }
73 
74 /*
75  *	B step AMD K6 before B 9730xxxx have hardware bugs that can cause
76  *	misexecution of code under Linux. Owners of such processors should
77  *	contact AMD for precise details and a CPU swap.
78  *
79  *	See	http://www.multimania.com/poulot/k6bug.html
80  *	and	section 2.6.2 of "AMD-K6 Processor Revision Guide - Model 6"
81  *		(Publication # 21266  Issue Date: August 1998)
82  *
83  *	The following test is erm.. interesting. AMD neglected to up
84  *	the chip setting when fixing the bug but they also tweaked some
85  *	performance at the same time..
86  */
87 
88 #ifdef CONFIG_X86_32
89 extern __visible void vide(void);
90 __asm__(".text\n"
91 	".globl vide\n"
92 	".type vide, @function\n"
93 	".align 4\n"
94 	"vide: ret\n");
95 #endif
96 
97 static void init_amd_k5(struct cpuinfo_x86 *c)
98 {
99 #ifdef CONFIG_X86_32
100 /*
101  * General Systems BIOSen alias the cpu frequency registers
102  * of the Elan at 0x000df000. Unfortunately, one of the Linux
103  * drivers subsequently pokes it, and changes the CPU speed.
104  * Workaround : Remove the unneeded alias.
105  */
106 #define CBAR		(0xfffc) /* Configuration Base Address  (32-bit) */
107 #define CBAR_ENB	(0x80000000)
108 #define CBAR_KEY	(0X000000CB)
109 	if (c->x86_model == 9 || c->x86_model == 10) {
110 		if (inl(CBAR) & CBAR_ENB)
111 			outl(0 | CBAR_KEY, CBAR);
112 	}
113 #endif
114 }
115 
116 static void init_amd_k6(struct cpuinfo_x86 *c)
117 {
118 #ifdef CONFIG_X86_32
119 	u32 l, h;
120 	int mbytes = get_num_physpages() >> (20-PAGE_SHIFT);
121 
122 	if (c->x86_model < 6) {
123 		/* Based on AMD doc 20734R - June 2000 */
124 		if (c->x86_model == 0) {
125 			clear_cpu_cap(c, X86_FEATURE_APIC);
126 			set_cpu_cap(c, X86_FEATURE_PGE);
127 		}
128 		return;
129 	}
130 
131 	if (c->x86_model == 6 && c->x86_stepping == 1) {
132 		const int K6_BUG_LOOP = 1000000;
133 		int n;
134 		void (*f_vide)(void);
135 		u64 d, d2;
136 
137 		pr_info("AMD K6 stepping B detected - ");
138 
139 		/*
140 		 * It looks like AMD fixed the 2.6.2 bug and improved indirect
141 		 * calls at the same time.
142 		 */
143 
144 		n = K6_BUG_LOOP;
145 		f_vide = vide;
146 		OPTIMIZER_HIDE_VAR(f_vide);
147 		d = rdtsc();
148 		while (n--)
149 			f_vide();
150 		d2 = rdtsc();
151 		d = d2-d;
152 
153 		if (d > 20*K6_BUG_LOOP)
154 			pr_cont("system stability may be impaired when more than 32 MB are used.\n");
155 		else
156 			pr_cont("probably OK (after B9730xxxx).\n");
157 	}
158 
159 	/* K6 with old style WHCR */
160 	if (c->x86_model < 8 ||
161 	   (c->x86_model == 8 && c->x86_stepping < 8)) {
162 		/* We can only write allocate on the low 508Mb */
163 		if (mbytes > 508)
164 			mbytes = 508;
165 
166 		rdmsr(MSR_K6_WHCR, l, h);
167 		if ((l&0x0000FFFF) == 0) {
168 			unsigned long flags;
169 			l = (1<<0)|((mbytes/4)<<1);
170 			local_irq_save(flags);
171 			wbinvd();
172 			wrmsr(MSR_K6_WHCR, l, h);
173 			local_irq_restore(flags);
174 			pr_info("Enabling old style K6 write allocation for %d Mb\n",
175 				mbytes);
176 		}
177 		return;
178 	}
179 
180 	if ((c->x86_model == 8 && c->x86_stepping > 7) ||
181 	     c->x86_model == 9 || c->x86_model == 13) {
182 		/* The more serious chips .. */
183 
184 		if (mbytes > 4092)
185 			mbytes = 4092;
186 
187 		rdmsr(MSR_K6_WHCR, l, h);
188 		if ((l&0xFFFF0000) == 0) {
189 			unsigned long flags;
190 			l = ((mbytes>>2)<<22)|(1<<16);
191 			local_irq_save(flags);
192 			wbinvd();
193 			wrmsr(MSR_K6_WHCR, l, h);
194 			local_irq_restore(flags);
195 			pr_info("Enabling new style K6 write allocation for %d Mb\n",
196 				mbytes);
197 		}
198 
199 		return;
200 	}
201 
202 	if (c->x86_model == 10) {
203 		/* AMD Geode LX is model 10 */
204 		/* placeholder for any needed mods */
205 		return;
206 	}
207 #endif
208 }
209 
210 static void init_amd_k7(struct cpuinfo_x86 *c)
211 {
212 #ifdef CONFIG_X86_32
213 	u32 l, h;
214 
215 	/*
216 	 * Bit 15 of Athlon specific MSR 15, needs to be 0
217 	 * to enable SSE on Palomino/Morgan/Barton CPU's.
218 	 * If the BIOS didn't enable it already, enable it here.
219 	 */
220 	if (c->x86_model >= 6 && c->x86_model <= 10) {
221 		if (!cpu_has(c, X86_FEATURE_XMM)) {
222 			pr_info("Enabling disabled K7/SSE Support.\n");
223 			msr_clear_bit(MSR_K7_HWCR, 15);
224 			set_cpu_cap(c, X86_FEATURE_XMM);
225 		}
226 	}
227 
228 	/*
229 	 * It's been determined by AMD that Athlons since model 8 stepping 1
230 	 * are more robust with CLK_CTL set to 200xxxxx instead of 600xxxxx
231 	 * As per AMD technical note 27212 0.2
232 	 */
233 	if ((c->x86_model == 8 && c->x86_stepping >= 1) || (c->x86_model > 8)) {
234 		rdmsr(MSR_K7_CLK_CTL, l, h);
235 		if ((l & 0xfff00000) != 0x20000000) {
236 			pr_info("CPU: CLK_CTL MSR was %x. Reprogramming to %x\n",
237 				l, ((l & 0x000fffff)|0x20000000));
238 			wrmsr(MSR_K7_CLK_CTL, (l & 0x000fffff)|0x20000000, h);
239 		}
240 	}
241 
242 	/* calling is from identify_secondary_cpu() ? */
243 	if (!c->cpu_index)
244 		return;
245 
246 	/*
247 	 * Certain Athlons might work (for various values of 'work') in SMP
248 	 * but they are not certified as MP capable.
249 	 */
250 	/* Athlon 660/661 is valid. */
251 	if ((c->x86_model == 6) && ((c->x86_stepping == 0) ||
252 	    (c->x86_stepping == 1)))
253 		return;
254 
255 	/* Duron 670 is valid */
256 	if ((c->x86_model == 7) && (c->x86_stepping == 0))
257 		return;
258 
259 	/*
260 	 * Athlon 662, Duron 671, and Athlon >model 7 have capability
261 	 * bit. It's worth noting that the A5 stepping (662) of some
262 	 * Athlon XP's have the MP bit set.
263 	 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for
264 	 * more.
265 	 */
266 	if (((c->x86_model == 6) && (c->x86_stepping >= 2)) ||
267 	    ((c->x86_model == 7) && (c->x86_stepping >= 1)) ||
268 	     (c->x86_model > 7))
269 		if (cpu_has(c, X86_FEATURE_MP))
270 			return;
271 
272 	/* If we get here, not a certified SMP capable AMD system. */
273 
274 	/*
275 	 * Don't taint if we are running SMP kernel on a single non-MP
276 	 * approved Athlon
277 	 */
278 	WARN_ONCE(1, "WARNING: This combination of AMD"
279 		" processors is not suitable for SMP.\n");
280 	add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_NOW_UNRELIABLE);
281 #endif
282 }
283 
284 #ifdef CONFIG_NUMA
285 /*
286  * To workaround broken NUMA config.  Read the comment in
287  * srat_detect_node().
288  */
289 static int nearby_node(int apicid)
290 {
291 	int i, node;
292 
293 	for (i = apicid - 1; i >= 0; i--) {
294 		node = __apicid_to_node[i];
295 		if (node != NUMA_NO_NODE && node_online(node))
296 			return node;
297 	}
298 	for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) {
299 		node = __apicid_to_node[i];
300 		if (node != NUMA_NO_NODE && node_online(node))
301 			return node;
302 	}
303 	return first_node(node_online_map); /* Shouldn't happen */
304 }
305 #endif
306 
307 /*
308  * Fix up cpu_core_id for pre-F17h systems to be in the
309  * [0 .. cores_per_node - 1] range. Not really needed but
310  * kept so as not to break existing setups.
311  */
312 static void legacy_fixup_core_id(struct cpuinfo_x86 *c)
313 {
314 	u32 cus_per_node;
315 
316 	if (c->x86 >= 0x17)
317 		return;
318 
319 	cus_per_node = c->x86_max_cores / nodes_per_socket;
320 	c->cpu_core_id %= cus_per_node;
321 }
322 
323 /*
324  * Fixup core topology information for
325  * (1) AMD multi-node processors
326  *     Assumption: Number of cores in each internal node is the same.
327  * (2) AMD processors supporting compute units
328  */
329 static void amd_get_topology(struct cpuinfo_x86 *c)
330 {
331 	u8 node_id;
332 	int cpu = smp_processor_id();
333 
334 	/* get information required for multi-node processors */
335 	if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
336 		int err;
337 		u32 eax, ebx, ecx, edx;
338 
339 		cpuid(0x8000001e, &eax, &ebx, &ecx, &edx);
340 
341 		node_id  = ecx & 0xff;
342 
343 		if (c->x86 == 0x15)
344 			c->cu_id = ebx & 0xff;
345 
346 		if (c->x86 >= 0x17) {
347 			c->cpu_core_id = ebx & 0xff;
348 
349 			if (smp_num_siblings > 1)
350 				c->x86_max_cores /= smp_num_siblings;
351 		}
352 
353 		/*
354 		 * In case leaf B is available, use it to derive
355 		 * topology information.
356 		 */
357 		err = detect_extended_topology(c);
358 		if (!err)
359 			c->x86_coreid_bits = get_count_order(c->x86_max_cores);
360 
361 		cacheinfo_amd_init_llc_id(c, cpu, node_id);
362 
363 	} else if (cpu_has(c, X86_FEATURE_NODEID_MSR)) {
364 		u64 value;
365 
366 		rdmsrl(MSR_FAM10H_NODE_ID, value);
367 		node_id = value & 7;
368 
369 		per_cpu(cpu_llc_id, cpu) = node_id;
370 	} else
371 		return;
372 
373 	if (nodes_per_socket > 1) {
374 		set_cpu_cap(c, X86_FEATURE_AMD_DCM);
375 		legacy_fixup_core_id(c);
376 	}
377 }
378 
379 /*
380  * On a AMD dual core setup the lower bits of the APIC id distinguish the cores.
381  * Assumes number of cores is a power of two.
382  */
383 static void amd_detect_cmp(struct cpuinfo_x86 *c)
384 {
385 	unsigned bits;
386 	int cpu = smp_processor_id();
387 
388 	bits = c->x86_coreid_bits;
389 	/* Low order bits define the core id (index of core in socket) */
390 	c->cpu_core_id = c->initial_apicid & ((1 << bits)-1);
391 	/* Convert the initial APIC ID into the socket ID */
392 	c->phys_proc_id = c->initial_apicid >> bits;
393 	/* use socket ID also for last level cache */
394 	per_cpu(cpu_llc_id, cpu) = c->phys_proc_id;
395 }
396 
397 static void amd_detect_ppin(struct cpuinfo_x86 *c)
398 {
399 	unsigned long long val;
400 
401 	if (!cpu_has(c, X86_FEATURE_AMD_PPIN))
402 		return;
403 
404 	/* When PPIN is defined in CPUID, still need to check PPIN_CTL MSR */
405 	if (rdmsrl_safe(MSR_AMD_PPIN_CTL, &val))
406 		goto clear_ppin;
407 
408 	/* PPIN is locked in disabled mode, clear feature bit */
409 	if ((val & 3UL) == 1UL)
410 		goto clear_ppin;
411 
412 	/* If PPIN is disabled, try to enable it */
413 	if (!(val & 2UL)) {
414 		wrmsrl_safe(MSR_AMD_PPIN_CTL,  val | 2UL);
415 		rdmsrl_safe(MSR_AMD_PPIN_CTL, &val);
416 	}
417 
418 	/* If PPIN_EN bit is 1, return from here; otherwise fall through */
419 	if (val & 2UL)
420 		return;
421 
422 clear_ppin:
423 	clear_cpu_cap(c, X86_FEATURE_AMD_PPIN);
424 }
425 
426 u16 amd_get_nb_id(int cpu)
427 {
428 	return per_cpu(cpu_llc_id, cpu);
429 }
430 EXPORT_SYMBOL_GPL(amd_get_nb_id);
431 
432 u32 amd_get_nodes_per_socket(void)
433 {
434 	return nodes_per_socket;
435 }
436 EXPORT_SYMBOL_GPL(amd_get_nodes_per_socket);
437 
438 static void srat_detect_node(struct cpuinfo_x86 *c)
439 {
440 #ifdef CONFIG_NUMA
441 	int cpu = smp_processor_id();
442 	int node;
443 	unsigned apicid = c->apicid;
444 
445 	node = numa_cpu_node(cpu);
446 	if (node == NUMA_NO_NODE)
447 		node = per_cpu(cpu_llc_id, cpu);
448 
449 	/*
450 	 * On multi-fabric platform (e.g. Numascale NumaChip) a
451 	 * platform-specific handler needs to be called to fixup some
452 	 * IDs of the CPU.
453 	 */
454 	if (x86_cpuinit.fixup_cpu_id)
455 		x86_cpuinit.fixup_cpu_id(c, node);
456 
457 	if (!node_online(node)) {
458 		/*
459 		 * Two possibilities here:
460 		 *
461 		 * - The CPU is missing memory and no node was created.  In
462 		 *   that case try picking one from a nearby CPU.
463 		 *
464 		 * - The APIC IDs differ from the HyperTransport node IDs
465 		 *   which the K8 northbridge parsing fills in.  Assume
466 		 *   they are all increased by a constant offset, but in
467 		 *   the same order as the HT nodeids.  If that doesn't
468 		 *   result in a usable node fall back to the path for the
469 		 *   previous case.
470 		 *
471 		 * This workaround operates directly on the mapping between
472 		 * APIC ID and NUMA node, assuming certain relationship
473 		 * between APIC ID, HT node ID and NUMA topology.  As going
474 		 * through CPU mapping may alter the outcome, directly
475 		 * access __apicid_to_node[].
476 		 */
477 		int ht_nodeid = c->initial_apicid;
478 
479 		if (__apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
480 			node = __apicid_to_node[ht_nodeid];
481 		/* Pick a nearby node */
482 		if (!node_online(node))
483 			node = nearby_node(apicid);
484 	}
485 	numa_set_node(cpu, node);
486 #endif
487 }
488 
489 static void early_init_amd_mc(struct cpuinfo_x86 *c)
490 {
491 #ifdef CONFIG_SMP
492 	unsigned bits, ecx;
493 
494 	/* Multi core CPU? */
495 	if (c->extended_cpuid_level < 0x80000008)
496 		return;
497 
498 	ecx = cpuid_ecx(0x80000008);
499 
500 	c->x86_max_cores = (ecx & 0xff) + 1;
501 
502 	/* CPU telling us the core id bits shift? */
503 	bits = (ecx >> 12) & 0xF;
504 
505 	/* Otherwise recompute */
506 	if (bits == 0) {
507 		while ((1 << bits) < c->x86_max_cores)
508 			bits++;
509 	}
510 
511 	c->x86_coreid_bits = bits;
512 #endif
513 }
514 
515 static void bsp_init_amd(struct cpuinfo_x86 *c)
516 {
517 
518 #ifdef CONFIG_X86_64
519 	if (c->x86 >= 0xf) {
520 		unsigned long long tseg;
521 
522 		/*
523 		 * Split up direct mapping around the TSEG SMM area.
524 		 * Don't do it for gbpages because there seems very little
525 		 * benefit in doing so.
526 		 */
527 		if (!rdmsrl_safe(MSR_K8_TSEG_ADDR, &tseg)) {
528 			unsigned long pfn = tseg >> PAGE_SHIFT;
529 
530 			pr_debug("tseg: %010llx\n", tseg);
531 			if (pfn_range_is_mapped(pfn, pfn + 1))
532 				set_memory_4k((unsigned long)__va(tseg), 1);
533 		}
534 	}
535 #endif
536 
537 	if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) {
538 
539 		if (c->x86 > 0x10 ||
540 		    (c->x86 == 0x10 && c->x86_model >= 0x2)) {
541 			u64 val;
542 
543 			rdmsrl(MSR_K7_HWCR, val);
544 			if (!(val & BIT(24)))
545 				pr_warn(FW_BUG "TSC doesn't count with P0 frequency!\n");
546 		}
547 	}
548 
549 	if (c->x86 == 0x15) {
550 		unsigned long upperbit;
551 		u32 cpuid, assoc;
552 
553 		cpuid	 = cpuid_edx(0x80000005);
554 		assoc	 = cpuid >> 16 & 0xff;
555 		upperbit = ((cpuid >> 24) << 10) / assoc;
556 
557 		va_align.mask	  = (upperbit - 1) & PAGE_MASK;
558 		va_align.flags    = ALIGN_VA_32 | ALIGN_VA_64;
559 
560 		/* A random value per boot for bit slice [12:upper_bit) */
561 		va_align.bits = get_random_int() & va_align.mask;
562 	}
563 
564 	if (cpu_has(c, X86_FEATURE_MWAITX))
565 		use_mwaitx_delay();
566 
567 	if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
568 		u32 ecx;
569 
570 		ecx = cpuid_ecx(0x8000001e);
571 		nodes_per_socket = ((ecx >> 8) & 7) + 1;
572 	} else if (boot_cpu_has(X86_FEATURE_NODEID_MSR)) {
573 		u64 value;
574 
575 		rdmsrl(MSR_FAM10H_NODE_ID, value);
576 		nodes_per_socket = ((value >> 3) & 7) + 1;
577 	}
578 
579 	if (!boot_cpu_has(X86_FEATURE_AMD_SSBD) &&
580 	    !boot_cpu_has(X86_FEATURE_VIRT_SSBD) &&
581 	    c->x86 >= 0x15 && c->x86 <= 0x17) {
582 		unsigned int bit;
583 
584 		switch (c->x86) {
585 		case 0x15: bit = 54; break;
586 		case 0x16: bit = 33; break;
587 		case 0x17: bit = 10; break;
588 		default: return;
589 		}
590 		/*
591 		 * Try to cache the base value so further operations can
592 		 * avoid RMW. If that faults, do not enable SSBD.
593 		 */
594 		if (!rdmsrl_safe(MSR_AMD64_LS_CFG, &x86_amd_ls_cfg_base)) {
595 			setup_force_cpu_cap(X86_FEATURE_LS_CFG_SSBD);
596 			setup_force_cpu_cap(X86_FEATURE_SSBD);
597 			x86_amd_ls_cfg_ssbd_mask = 1ULL << bit;
598 		}
599 	}
600 }
601 
602 static void early_detect_mem_encrypt(struct cpuinfo_x86 *c)
603 {
604 	u64 msr;
605 
606 	/*
607 	 * BIOS support is required for SME and SEV.
608 	 *   For SME: If BIOS has enabled SME then adjust x86_phys_bits by
609 	 *	      the SME physical address space reduction value.
610 	 *	      If BIOS has not enabled SME then don't advertise the
611 	 *	      SME feature (set in scattered.c).
612 	 *   For SEV: If BIOS has not enabled SEV then don't advertise the
613 	 *            SEV feature (set in scattered.c).
614 	 *
615 	 *   In all cases, since support for SME and SEV requires long mode,
616 	 *   don't advertise the feature under CONFIG_X86_32.
617 	 */
618 	if (cpu_has(c, X86_FEATURE_SME) || cpu_has(c, X86_FEATURE_SEV)) {
619 		/* Check if memory encryption is enabled */
620 		rdmsrl(MSR_K8_SYSCFG, msr);
621 		if (!(msr & MSR_K8_SYSCFG_MEM_ENCRYPT))
622 			goto clear_all;
623 
624 		/*
625 		 * Always adjust physical address bits. Even though this
626 		 * will be a value above 32-bits this is still done for
627 		 * CONFIG_X86_32 so that accurate values are reported.
628 		 */
629 		c->x86_phys_bits -= (cpuid_ebx(0x8000001f) >> 6) & 0x3f;
630 
631 		if (IS_ENABLED(CONFIG_X86_32))
632 			goto clear_all;
633 
634 		rdmsrl(MSR_K7_HWCR, msr);
635 		if (!(msr & MSR_K7_HWCR_SMMLOCK))
636 			goto clear_sev;
637 
638 		return;
639 
640 clear_all:
641 		setup_clear_cpu_cap(X86_FEATURE_SME);
642 clear_sev:
643 		setup_clear_cpu_cap(X86_FEATURE_SEV);
644 	}
645 }
646 
647 static void early_init_amd(struct cpuinfo_x86 *c)
648 {
649 	u64 value;
650 	u32 dummy;
651 
652 	early_init_amd_mc(c);
653 
654 #ifdef CONFIG_X86_32
655 	if (c->x86 == 6)
656 		set_cpu_cap(c, X86_FEATURE_K7);
657 #endif
658 
659 	if (c->x86 >= 0xf)
660 		set_cpu_cap(c, X86_FEATURE_K8);
661 
662 	rdmsr_safe(MSR_AMD64_PATCH_LEVEL, &c->microcode, &dummy);
663 
664 	/*
665 	 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
666 	 * with P/T states and does not stop in deep C-states
667 	 */
668 	if (c->x86_power & (1 << 8)) {
669 		set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
670 		set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
671 	}
672 
673 	/* Bit 12 of 8000_0007 edx is accumulated power mechanism. */
674 	if (c->x86_power & BIT(12))
675 		set_cpu_cap(c, X86_FEATURE_ACC_POWER);
676 
677 #ifdef CONFIG_X86_64
678 	set_cpu_cap(c, X86_FEATURE_SYSCALL32);
679 #else
680 	/*  Set MTRR capability flag if appropriate */
681 	if (c->x86 == 5)
682 		if (c->x86_model == 13 || c->x86_model == 9 ||
683 		    (c->x86_model == 8 && c->x86_stepping >= 8))
684 			set_cpu_cap(c, X86_FEATURE_K6_MTRR);
685 #endif
686 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
687 	/*
688 	 * ApicID can always be treated as an 8-bit value for AMD APIC versions
689 	 * >= 0x10, but even old K8s came out of reset with version 0x10. So, we
690 	 * can safely set X86_FEATURE_EXTD_APICID unconditionally for families
691 	 * after 16h.
692 	 */
693 	if (boot_cpu_has(X86_FEATURE_APIC)) {
694 		if (c->x86 > 0x16)
695 			set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
696 		else if (c->x86 >= 0xf) {
697 			/* check CPU config space for extended APIC ID */
698 			unsigned int val;
699 
700 			val = read_pci_config(0, 24, 0, 0x68);
701 			if ((val >> 17 & 0x3) == 0x3)
702 				set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
703 		}
704 	}
705 #endif
706 
707 	/*
708 	 * This is only needed to tell the kernel whether to use VMCALL
709 	 * and VMMCALL.  VMMCALL is never executed except under virt, so
710 	 * we can set it unconditionally.
711 	 */
712 	set_cpu_cap(c, X86_FEATURE_VMMCALL);
713 
714 	/* F16h erratum 793, CVE-2013-6885 */
715 	if (c->x86 == 0x16 && c->x86_model <= 0xf)
716 		msr_set_bit(MSR_AMD64_LS_CFG, 15);
717 
718 	/*
719 	 * Check whether the machine is affected by erratum 400. This is
720 	 * used to select the proper idle routine and to enable the check
721 	 * whether the machine is affected in arch_post_acpi_init(), which
722 	 * sets the X86_BUG_AMD_APIC_C1E bug depending on the MSR check.
723 	 */
724 	if (cpu_has_amd_erratum(c, amd_erratum_400))
725 		set_cpu_bug(c, X86_BUG_AMD_E400);
726 
727 	early_detect_mem_encrypt(c);
728 
729 	/* Re-enable TopologyExtensions if switched off by BIOS */
730 	if (c->x86 == 0x15 &&
731 	    (c->x86_model >= 0x10 && c->x86_model <= 0x6f) &&
732 	    !cpu_has(c, X86_FEATURE_TOPOEXT)) {
733 
734 		if (msr_set_bit(0xc0011005, 54) > 0) {
735 			rdmsrl(0xc0011005, value);
736 			if (value & BIT_64(54)) {
737 				set_cpu_cap(c, X86_FEATURE_TOPOEXT);
738 				pr_info_once(FW_INFO "CPU: Re-enabling disabled Topology Extensions Support.\n");
739 			}
740 		}
741 	}
742 
743 	if (cpu_has(c, X86_FEATURE_TOPOEXT))
744 		smp_num_siblings = ((cpuid_ebx(0x8000001e) >> 8) & 0xff) + 1;
745 }
746 
747 static void init_amd_k8(struct cpuinfo_x86 *c)
748 {
749 	u32 level;
750 	u64 value;
751 
752 	/* On C+ stepping K8 rep microcode works well for copy/memset */
753 	level = cpuid_eax(1);
754 	if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58)
755 		set_cpu_cap(c, X86_FEATURE_REP_GOOD);
756 
757 	/*
758 	 * Some BIOSes incorrectly force this feature, but only K8 revision D
759 	 * (model = 0x14) and later actually support it.
760 	 * (AMD Erratum #110, docId: 25759).
761 	 */
762 	if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM)) {
763 		clear_cpu_cap(c, X86_FEATURE_LAHF_LM);
764 		if (!rdmsrl_amd_safe(0xc001100d, &value)) {
765 			value &= ~BIT_64(32);
766 			wrmsrl_amd_safe(0xc001100d, value);
767 		}
768 	}
769 
770 	if (!c->x86_model_id[0])
771 		strcpy(c->x86_model_id, "Hammer");
772 
773 #ifdef CONFIG_SMP
774 	/*
775 	 * Disable TLB flush filter by setting HWCR.FFDIS on K8
776 	 * bit 6 of msr C001_0015
777 	 *
778 	 * Errata 63 for SH-B3 steppings
779 	 * Errata 122 for all steppings (F+ have it disabled by default)
780 	 */
781 	msr_set_bit(MSR_K7_HWCR, 6);
782 #endif
783 	set_cpu_bug(c, X86_BUG_SWAPGS_FENCE);
784 }
785 
786 static void init_amd_gh(struct cpuinfo_x86 *c)
787 {
788 #ifdef CONFIG_MMCONF_FAM10H
789 	/* do this for boot cpu */
790 	if (c == &boot_cpu_data)
791 		check_enable_amd_mmconf_dmi();
792 
793 	fam10h_check_enable_mmcfg();
794 #endif
795 
796 	/*
797 	 * Disable GART TLB Walk Errors on Fam10h. We do this here because this
798 	 * is always needed when GART is enabled, even in a kernel which has no
799 	 * MCE support built in. BIOS should disable GartTlbWlk Errors already.
800 	 * If it doesn't, we do it here as suggested by the BKDG.
801 	 *
802 	 * Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=33012
803 	 */
804 	msr_set_bit(MSR_AMD64_MCx_MASK(4), 10);
805 
806 	/*
807 	 * On family 10h BIOS may not have properly enabled WC+ support, causing
808 	 * it to be converted to CD memtype. This may result in performance
809 	 * degradation for certain nested-paging guests. Prevent this conversion
810 	 * by clearing bit 24 in MSR_AMD64_BU_CFG2.
811 	 *
812 	 * NOTE: we want to use the _safe accessors so as not to #GP kvm
813 	 * guests on older kvm hosts.
814 	 */
815 	msr_clear_bit(MSR_AMD64_BU_CFG2, 24);
816 
817 	if (cpu_has_amd_erratum(c, amd_erratum_383))
818 		set_cpu_bug(c, X86_BUG_AMD_TLB_MMATCH);
819 }
820 
821 #define MSR_AMD64_DE_CFG	0xC0011029
822 
823 static void init_amd_ln(struct cpuinfo_x86 *c)
824 {
825 	/*
826 	 * Apply erratum 665 fix unconditionally so machines without a BIOS
827 	 * fix work.
828 	 */
829 	msr_set_bit(MSR_AMD64_DE_CFG, 31);
830 }
831 
832 static bool rdrand_force;
833 
834 static int __init rdrand_cmdline(char *str)
835 {
836 	if (!str)
837 		return -EINVAL;
838 
839 	if (!strcmp(str, "force"))
840 		rdrand_force = true;
841 	else
842 		return -EINVAL;
843 
844 	return 0;
845 }
846 early_param("rdrand", rdrand_cmdline);
847 
848 static void clear_rdrand_cpuid_bit(struct cpuinfo_x86 *c)
849 {
850 	/*
851 	 * Saving of the MSR used to hide the RDRAND support during
852 	 * suspend/resume is done by arch/x86/power/cpu.c, which is
853 	 * dependent on CONFIG_PM_SLEEP.
854 	 */
855 	if (!IS_ENABLED(CONFIG_PM_SLEEP))
856 		return;
857 
858 	/*
859 	 * The nordrand option can clear X86_FEATURE_RDRAND, so check for
860 	 * RDRAND support using the CPUID function directly.
861 	 */
862 	if (!(cpuid_ecx(1) & BIT(30)) || rdrand_force)
863 		return;
864 
865 	msr_clear_bit(MSR_AMD64_CPUID_FN_1, 62);
866 
867 	/*
868 	 * Verify that the CPUID change has occurred in case the kernel is
869 	 * running virtualized and the hypervisor doesn't support the MSR.
870 	 */
871 	if (cpuid_ecx(1) & BIT(30)) {
872 		pr_info_once("BIOS may not properly restore RDRAND after suspend, but hypervisor does not support hiding RDRAND via CPUID.\n");
873 		return;
874 	}
875 
876 	clear_cpu_cap(c, X86_FEATURE_RDRAND);
877 	pr_info_once("BIOS may not properly restore RDRAND after suspend, hiding RDRAND via CPUID. Use rdrand=force to reenable.\n");
878 }
879 
880 static void init_amd_jg(struct cpuinfo_x86 *c)
881 {
882 	/*
883 	 * Some BIOS implementations do not restore proper RDRAND support
884 	 * across suspend and resume. Check on whether to hide the RDRAND
885 	 * instruction support via CPUID.
886 	 */
887 	clear_rdrand_cpuid_bit(c);
888 }
889 
890 static void init_amd_bd(struct cpuinfo_x86 *c)
891 {
892 	u64 value;
893 
894 	/*
895 	 * The way access filter has a performance penalty on some workloads.
896 	 * Disable it on the affected CPUs.
897 	 */
898 	if ((c->x86_model >= 0x02) && (c->x86_model < 0x20)) {
899 		if (!rdmsrl_safe(MSR_F15H_IC_CFG, &value) && !(value & 0x1E)) {
900 			value |= 0x1E;
901 			wrmsrl_safe(MSR_F15H_IC_CFG, value);
902 		}
903 	}
904 
905 	/*
906 	 * Some BIOS implementations do not restore proper RDRAND support
907 	 * across suspend and resume. Check on whether to hide the RDRAND
908 	 * instruction support via CPUID.
909 	 */
910 	clear_rdrand_cpuid_bit(c);
911 }
912 
913 static void init_amd_zn(struct cpuinfo_x86 *c)
914 {
915 	set_cpu_cap(c, X86_FEATURE_ZEN);
916 
917 #ifdef CONFIG_NUMA
918 	node_reclaim_distance = 32;
919 #endif
920 
921 	/*
922 	 * Fix erratum 1076: CPB feature bit not being set in CPUID.
923 	 * Always set it, except when running under a hypervisor.
924 	 */
925 	if (!cpu_has(c, X86_FEATURE_HYPERVISOR) && !cpu_has(c, X86_FEATURE_CPB))
926 		set_cpu_cap(c, X86_FEATURE_CPB);
927 }
928 
929 static void init_amd(struct cpuinfo_x86 *c)
930 {
931 	early_init_amd(c);
932 
933 	/*
934 	 * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
935 	 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
936 	 */
937 	clear_cpu_cap(c, 0*32+31);
938 
939 	if (c->x86 >= 0x10)
940 		set_cpu_cap(c, X86_FEATURE_REP_GOOD);
941 
942 	/* get apicid instead of initial apic id from cpuid */
943 	c->apicid = hard_smp_processor_id();
944 
945 	/* K6s reports MCEs but don't actually have all the MSRs */
946 	if (c->x86 < 6)
947 		clear_cpu_cap(c, X86_FEATURE_MCE);
948 
949 	switch (c->x86) {
950 	case 4:    init_amd_k5(c); break;
951 	case 5:    init_amd_k6(c); break;
952 	case 6:	   init_amd_k7(c); break;
953 	case 0xf:  init_amd_k8(c); break;
954 	case 0x10: init_amd_gh(c); break;
955 	case 0x12: init_amd_ln(c); break;
956 	case 0x15: init_amd_bd(c); break;
957 	case 0x16: init_amd_jg(c); break;
958 	case 0x17: fallthrough;
959 	case 0x19: init_amd_zn(c); break;
960 	}
961 
962 	/*
963 	 * Enable workaround for FXSAVE leak on CPUs
964 	 * without a XSaveErPtr feature
965 	 */
966 	if ((c->x86 >= 6) && (!cpu_has(c, X86_FEATURE_XSAVEERPTR)))
967 		set_cpu_bug(c, X86_BUG_FXSAVE_LEAK);
968 
969 	cpu_detect_cache_sizes(c);
970 
971 	amd_detect_cmp(c);
972 	amd_get_topology(c);
973 	srat_detect_node(c);
974 	amd_detect_ppin(c);
975 
976 	init_amd_cacheinfo(c);
977 
978 	if (cpu_has(c, X86_FEATURE_XMM2)) {
979 		/*
980 		 * Use LFENCE for execution serialization.  On families which
981 		 * don't have that MSR, LFENCE is already serializing.
982 		 * msr_set_bit() uses the safe accessors, too, even if the MSR
983 		 * is not present.
984 		 */
985 		msr_set_bit(MSR_F10H_DECFG,
986 			    MSR_F10H_DECFG_LFENCE_SERIALIZE_BIT);
987 
988 		/* A serializing LFENCE stops RDTSC speculation */
989 		set_cpu_cap(c, X86_FEATURE_LFENCE_RDTSC);
990 	}
991 
992 	/*
993 	 * Family 0x12 and above processors have APIC timer
994 	 * running in deep C states.
995 	 */
996 	if (c->x86 > 0x11)
997 		set_cpu_cap(c, X86_FEATURE_ARAT);
998 
999 	/* 3DNow or LM implies PREFETCHW */
1000 	if (!cpu_has(c, X86_FEATURE_3DNOWPREFETCH))
1001 		if (cpu_has(c, X86_FEATURE_3DNOW) || cpu_has(c, X86_FEATURE_LM))
1002 			set_cpu_cap(c, X86_FEATURE_3DNOWPREFETCH);
1003 
1004 	/* AMD CPUs don't reset SS attributes on SYSRET, Xen does. */
1005 	if (!cpu_has(c, X86_FEATURE_XENPV))
1006 		set_cpu_bug(c, X86_BUG_SYSRET_SS_ATTRS);
1007 
1008 	/*
1009 	 * Turn on the Instructions Retired free counter on machines not
1010 	 * susceptible to erratum #1054 "Instructions Retired Performance
1011 	 * Counter May Be Inaccurate".
1012 	 */
1013 	if (cpu_has(c, X86_FEATURE_IRPERF) &&
1014 	    !cpu_has_amd_erratum(c, amd_erratum_1054))
1015 		msr_set_bit(MSR_K7_HWCR, MSR_K7_HWCR_IRPERF_EN_BIT);
1016 }
1017 
1018 #ifdef CONFIG_X86_32
1019 static unsigned int amd_size_cache(struct cpuinfo_x86 *c, unsigned int size)
1020 {
1021 	/* AMD errata T13 (order #21922) */
1022 	if (c->x86 == 6) {
1023 		/* Duron Rev A0 */
1024 		if (c->x86_model == 3 && c->x86_stepping == 0)
1025 			size = 64;
1026 		/* Tbird rev A1/A2 */
1027 		if (c->x86_model == 4 &&
1028 			(c->x86_stepping == 0 || c->x86_stepping == 1))
1029 			size = 256;
1030 	}
1031 	return size;
1032 }
1033 #endif
1034 
1035 static void cpu_detect_tlb_amd(struct cpuinfo_x86 *c)
1036 {
1037 	u32 ebx, eax, ecx, edx;
1038 	u16 mask = 0xfff;
1039 
1040 	if (c->x86 < 0xf)
1041 		return;
1042 
1043 	if (c->extended_cpuid_level < 0x80000006)
1044 		return;
1045 
1046 	cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
1047 
1048 	tlb_lld_4k[ENTRIES] = (ebx >> 16) & mask;
1049 	tlb_lli_4k[ENTRIES] = ebx & mask;
1050 
1051 	/*
1052 	 * K8 doesn't have 2M/4M entries in the L2 TLB so read out the L1 TLB
1053 	 * characteristics from the CPUID function 0x80000005 instead.
1054 	 */
1055 	if (c->x86 == 0xf) {
1056 		cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
1057 		mask = 0xff;
1058 	}
1059 
1060 	/* Handle DTLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
1061 	if (!((eax >> 16) & mask))
1062 		tlb_lld_2m[ENTRIES] = (cpuid_eax(0x80000005) >> 16) & 0xff;
1063 	else
1064 		tlb_lld_2m[ENTRIES] = (eax >> 16) & mask;
1065 
1066 	/* a 4M entry uses two 2M entries */
1067 	tlb_lld_4m[ENTRIES] = tlb_lld_2m[ENTRIES] >> 1;
1068 
1069 	/* Handle ITLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
1070 	if (!(eax & mask)) {
1071 		/* Erratum 658 */
1072 		if (c->x86 == 0x15 && c->x86_model <= 0x1f) {
1073 			tlb_lli_2m[ENTRIES] = 1024;
1074 		} else {
1075 			cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
1076 			tlb_lli_2m[ENTRIES] = eax & 0xff;
1077 		}
1078 	} else
1079 		tlb_lli_2m[ENTRIES] = eax & mask;
1080 
1081 	tlb_lli_4m[ENTRIES] = tlb_lli_2m[ENTRIES] >> 1;
1082 }
1083 
1084 static const struct cpu_dev amd_cpu_dev = {
1085 	.c_vendor	= "AMD",
1086 	.c_ident	= { "AuthenticAMD" },
1087 #ifdef CONFIG_X86_32
1088 	.legacy_models = {
1089 		{ .family = 4, .model_names =
1090 		  {
1091 			  [3] = "486 DX/2",
1092 			  [7] = "486 DX/2-WB",
1093 			  [8] = "486 DX/4",
1094 			  [9] = "486 DX/4-WB",
1095 			  [14] = "Am5x86-WT",
1096 			  [15] = "Am5x86-WB"
1097 		  }
1098 		},
1099 	},
1100 	.legacy_cache_size = amd_size_cache,
1101 #endif
1102 	.c_early_init   = early_init_amd,
1103 	.c_detect_tlb	= cpu_detect_tlb_amd,
1104 	.c_bsp_init	= bsp_init_amd,
1105 	.c_init		= init_amd,
1106 	.c_x86_vendor	= X86_VENDOR_AMD,
1107 };
1108 
1109 cpu_dev_register(amd_cpu_dev);
1110 
1111 /*
1112  * AMD errata checking
1113  *
1114  * Errata are defined as arrays of ints using the AMD_LEGACY_ERRATUM() or
1115  * AMD_OSVW_ERRATUM() macros. The latter is intended for newer errata that
1116  * have an OSVW id assigned, which it takes as first argument. Both take a
1117  * variable number of family-specific model-stepping ranges created by
1118  * AMD_MODEL_RANGE().
1119  *
1120  * Example:
1121  *
1122  * const int amd_erratum_319[] =
1123  *	AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0x4, 0x2),
1124  *			   AMD_MODEL_RANGE(0x10, 0x8, 0x0, 0x8, 0x0),
1125  *			   AMD_MODEL_RANGE(0x10, 0x9, 0x0, 0x9, 0x0));
1126  */
1127 
1128 #define AMD_LEGACY_ERRATUM(...)		{ -1, __VA_ARGS__, 0 }
1129 #define AMD_OSVW_ERRATUM(osvw_id, ...)	{ osvw_id, __VA_ARGS__, 0 }
1130 #define AMD_MODEL_RANGE(f, m_start, s_start, m_end, s_end) \
1131 	((f << 24) | (m_start << 16) | (s_start << 12) | (m_end << 4) | (s_end))
1132 #define AMD_MODEL_RANGE_FAMILY(range)	(((range) >> 24) & 0xff)
1133 #define AMD_MODEL_RANGE_START(range)	(((range) >> 12) & 0xfff)
1134 #define AMD_MODEL_RANGE_END(range)	((range) & 0xfff)
1135 
1136 static const int amd_erratum_400[] =
1137 	AMD_OSVW_ERRATUM(1, AMD_MODEL_RANGE(0xf, 0x41, 0x2, 0xff, 0xf),
1138 			    AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0xff, 0xf));
1139 
1140 static const int amd_erratum_383[] =
1141 	AMD_OSVW_ERRATUM(3, AMD_MODEL_RANGE(0x10, 0, 0, 0xff, 0xf));
1142 
1143 /* #1054: Instructions Retired Performance Counter May Be Inaccurate */
1144 static const int amd_erratum_1054[] =
1145 	AMD_OSVW_ERRATUM(0, AMD_MODEL_RANGE(0x17, 0, 0, 0x2f, 0xf));
1146 
1147 
1148 static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum)
1149 {
1150 	int osvw_id = *erratum++;
1151 	u32 range;
1152 	u32 ms;
1153 
1154 	if (osvw_id >= 0 && osvw_id < 65536 &&
1155 	    cpu_has(cpu, X86_FEATURE_OSVW)) {
1156 		u64 osvw_len;
1157 
1158 		rdmsrl(MSR_AMD64_OSVW_ID_LENGTH, osvw_len);
1159 		if (osvw_id < osvw_len) {
1160 			u64 osvw_bits;
1161 
1162 			rdmsrl(MSR_AMD64_OSVW_STATUS + (osvw_id >> 6),
1163 			    osvw_bits);
1164 			return osvw_bits & (1ULL << (osvw_id & 0x3f));
1165 		}
1166 	}
1167 
1168 	/* OSVW unavailable or ID unknown, match family-model-stepping range */
1169 	ms = (cpu->x86_model << 4) | cpu->x86_stepping;
1170 	while ((range = *erratum++))
1171 		if ((cpu->x86 == AMD_MODEL_RANGE_FAMILY(range)) &&
1172 		    (ms >= AMD_MODEL_RANGE_START(range)) &&
1173 		    (ms <= AMD_MODEL_RANGE_END(range)))
1174 			return true;
1175 
1176 	return false;
1177 }
1178 
1179 void set_dr_addr_mask(unsigned long mask, int dr)
1180 {
1181 	if (!boot_cpu_has(X86_FEATURE_BPEXT))
1182 		return;
1183 
1184 	switch (dr) {
1185 	case 0:
1186 		wrmsr(MSR_F16H_DR0_ADDR_MASK, mask, 0);
1187 		break;
1188 	case 1:
1189 	case 2:
1190 	case 3:
1191 		wrmsr(MSR_F16H_DR1_ADDR_MASK - 1 + dr, mask, 0);
1192 		break;
1193 	default:
1194 		break;
1195 	}
1196 }
1197