xref: /openbmc/linux/arch/x86/kernel/cpu/amd.c (revision f79e4d5f92a129a1159c973735007d4ddc8541f3)
1 #include <linux/export.h>
2 #include <linux/bitops.h>
3 #include <linux/elf.h>
4 #include <linux/mm.h>
5 
6 #include <linux/io.h>
7 #include <linux/sched.h>
8 #include <linux/sched/clock.h>
9 #include <linux/random.h>
10 #include <asm/processor.h>
11 #include <asm/apic.h>
12 #include <asm/cacheinfo.h>
13 #include <asm/cpu.h>
14 #include <asm/spec-ctrl.h>
15 #include <asm/smp.h>
16 #include <asm/pci-direct.h>
17 #include <asm/delay.h>
18 
19 #ifdef CONFIG_X86_64
20 # include <asm/mmconfig.h>
21 # include <asm/set_memory.h>
22 #endif
23 
24 #include "cpu.h"
25 
26 static const int amd_erratum_383[];
27 static const int amd_erratum_400[];
28 static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum);
29 
30 /*
31  * nodes_per_socket: Stores the number of nodes per socket.
32  * Refer to Fam15h Models 00-0fh BKDG - CPUID Fn8000_001E_ECX
33  * Node Identifiers[10:8]
34  */
35 static u32 nodes_per_socket = 1;
36 
37 static inline int rdmsrl_amd_safe(unsigned msr, unsigned long long *p)
38 {
39 	u32 gprs[8] = { 0 };
40 	int err;
41 
42 	WARN_ONCE((boot_cpu_data.x86 != 0xf),
43 		  "%s should only be used on K8!\n", __func__);
44 
45 	gprs[1] = msr;
46 	gprs[7] = 0x9c5a203a;
47 
48 	err = rdmsr_safe_regs(gprs);
49 
50 	*p = gprs[0] | ((u64)gprs[2] << 32);
51 
52 	return err;
53 }
54 
55 static inline int wrmsrl_amd_safe(unsigned msr, unsigned long long val)
56 {
57 	u32 gprs[8] = { 0 };
58 
59 	WARN_ONCE((boot_cpu_data.x86 != 0xf),
60 		  "%s should only be used on K8!\n", __func__);
61 
62 	gprs[0] = (u32)val;
63 	gprs[1] = msr;
64 	gprs[2] = val >> 32;
65 	gprs[7] = 0x9c5a203a;
66 
67 	return wrmsr_safe_regs(gprs);
68 }
69 
70 /*
71  *	B step AMD K6 before B 9730xxxx have hardware bugs that can cause
72  *	misexecution of code under Linux. Owners of such processors should
73  *	contact AMD for precise details and a CPU swap.
74  *
75  *	See	http://www.multimania.com/poulot/k6bug.html
76  *	and	section 2.6.2 of "AMD-K6 Processor Revision Guide - Model 6"
77  *		(Publication # 21266  Issue Date: August 1998)
78  *
79  *	The following test is erm.. interesting. AMD neglected to up
80  *	the chip setting when fixing the bug but they also tweaked some
81  *	performance at the same time..
82  */
83 
84 extern __visible void vide(void);
85 __asm__(".globl vide\n"
86 	".type vide, @function\n"
87 	".align 4\n"
88 	"vide: ret\n");
89 
90 static void init_amd_k5(struct cpuinfo_x86 *c)
91 {
92 #ifdef CONFIG_X86_32
93 /*
94  * General Systems BIOSen alias the cpu frequency registers
95  * of the Elan at 0x000df000. Unfortunately, one of the Linux
96  * drivers subsequently pokes it, and changes the CPU speed.
97  * Workaround : Remove the unneeded alias.
98  */
99 #define CBAR		(0xfffc) /* Configuration Base Address  (32-bit) */
100 #define CBAR_ENB	(0x80000000)
101 #define CBAR_KEY	(0X000000CB)
102 	if (c->x86_model == 9 || c->x86_model == 10) {
103 		if (inl(CBAR) & CBAR_ENB)
104 			outl(0 | CBAR_KEY, CBAR);
105 	}
106 #endif
107 }
108 
109 static void init_amd_k6(struct cpuinfo_x86 *c)
110 {
111 #ifdef CONFIG_X86_32
112 	u32 l, h;
113 	int mbytes = get_num_physpages() >> (20-PAGE_SHIFT);
114 
115 	if (c->x86_model < 6) {
116 		/* Based on AMD doc 20734R - June 2000 */
117 		if (c->x86_model == 0) {
118 			clear_cpu_cap(c, X86_FEATURE_APIC);
119 			set_cpu_cap(c, X86_FEATURE_PGE);
120 		}
121 		return;
122 	}
123 
124 	if (c->x86_model == 6 && c->x86_stepping == 1) {
125 		const int K6_BUG_LOOP = 1000000;
126 		int n;
127 		void (*f_vide)(void);
128 		u64 d, d2;
129 
130 		pr_info("AMD K6 stepping B detected - ");
131 
132 		/*
133 		 * It looks like AMD fixed the 2.6.2 bug and improved indirect
134 		 * calls at the same time.
135 		 */
136 
137 		n = K6_BUG_LOOP;
138 		f_vide = vide;
139 		OPTIMIZER_HIDE_VAR(f_vide);
140 		d = rdtsc();
141 		while (n--)
142 			f_vide();
143 		d2 = rdtsc();
144 		d = d2-d;
145 
146 		if (d > 20*K6_BUG_LOOP)
147 			pr_cont("system stability may be impaired when more than 32 MB are used.\n");
148 		else
149 			pr_cont("probably OK (after B9730xxxx).\n");
150 	}
151 
152 	/* K6 with old style WHCR */
153 	if (c->x86_model < 8 ||
154 	   (c->x86_model == 8 && c->x86_stepping < 8)) {
155 		/* We can only write allocate on the low 508Mb */
156 		if (mbytes > 508)
157 			mbytes = 508;
158 
159 		rdmsr(MSR_K6_WHCR, l, h);
160 		if ((l&0x0000FFFF) == 0) {
161 			unsigned long flags;
162 			l = (1<<0)|((mbytes/4)<<1);
163 			local_irq_save(flags);
164 			wbinvd();
165 			wrmsr(MSR_K6_WHCR, l, h);
166 			local_irq_restore(flags);
167 			pr_info("Enabling old style K6 write allocation for %d Mb\n",
168 				mbytes);
169 		}
170 		return;
171 	}
172 
173 	if ((c->x86_model == 8 && c->x86_stepping > 7) ||
174 	     c->x86_model == 9 || c->x86_model == 13) {
175 		/* The more serious chips .. */
176 
177 		if (mbytes > 4092)
178 			mbytes = 4092;
179 
180 		rdmsr(MSR_K6_WHCR, l, h);
181 		if ((l&0xFFFF0000) == 0) {
182 			unsigned long flags;
183 			l = ((mbytes>>2)<<22)|(1<<16);
184 			local_irq_save(flags);
185 			wbinvd();
186 			wrmsr(MSR_K6_WHCR, l, h);
187 			local_irq_restore(flags);
188 			pr_info("Enabling new style K6 write allocation for %d Mb\n",
189 				mbytes);
190 		}
191 
192 		return;
193 	}
194 
195 	if (c->x86_model == 10) {
196 		/* AMD Geode LX is model 10 */
197 		/* placeholder for any needed mods */
198 		return;
199 	}
200 #endif
201 }
202 
203 static void init_amd_k7(struct cpuinfo_x86 *c)
204 {
205 #ifdef CONFIG_X86_32
206 	u32 l, h;
207 
208 	/*
209 	 * Bit 15 of Athlon specific MSR 15, needs to be 0
210 	 * to enable SSE on Palomino/Morgan/Barton CPU's.
211 	 * If the BIOS didn't enable it already, enable it here.
212 	 */
213 	if (c->x86_model >= 6 && c->x86_model <= 10) {
214 		if (!cpu_has(c, X86_FEATURE_XMM)) {
215 			pr_info("Enabling disabled K7/SSE Support.\n");
216 			msr_clear_bit(MSR_K7_HWCR, 15);
217 			set_cpu_cap(c, X86_FEATURE_XMM);
218 		}
219 	}
220 
221 	/*
222 	 * It's been determined by AMD that Athlons since model 8 stepping 1
223 	 * are more robust with CLK_CTL set to 200xxxxx instead of 600xxxxx
224 	 * As per AMD technical note 27212 0.2
225 	 */
226 	if ((c->x86_model == 8 && c->x86_stepping >= 1) || (c->x86_model > 8)) {
227 		rdmsr(MSR_K7_CLK_CTL, l, h);
228 		if ((l & 0xfff00000) != 0x20000000) {
229 			pr_info("CPU: CLK_CTL MSR was %x. Reprogramming to %x\n",
230 				l, ((l & 0x000fffff)|0x20000000));
231 			wrmsr(MSR_K7_CLK_CTL, (l & 0x000fffff)|0x20000000, h);
232 		}
233 	}
234 
235 	set_cpu_cap(c, X86_FEATURE_K7);
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  */
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 
302 /*
303  * Fix up cpu_core_id for pre-F17h systems to be in the
304  * [0 .. cores_per_node - 1] range. Not really needed but
305  * kept so as not to break existing setups.
306  */
307 static void legacy_fixup_core_id(struct cpuinfo_x86 *c)
308 {
309 	u32 cus_per_node;
310 
311 	if (c->x86 >= 0x17)
312 		return;
313 
314 	cus_per_node = c->x86_max_cores / nodes_per_socket;
315 	c->cpu_core_id %= cus_per_node;
316 }
317 
318 /*
319  * Fixup core topology information for
320  * (1) AMD multi-node processors
321  *     Assumption: Number of cores in each internal node is the same.
322  * (2) AMD processors supporting compute units
323  */
324 static void amd_get_topology(struct cpuinfo_x86 *c)
325 {
326 	u8 node_id;
327 	int cpu = smp_processor_id();
328 
329 	/* get information required for multi-node processors */
330 	if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
331 		int err;
332 		u32 eax, ebx, ecx, edx;
333 
334 		cpuid(0x8000001e, &eax, &ebx, &ecx, &edx);
335 
336 		node_id  = ecx & 0xff;
337 		smp_num_siblings = ((ebx >> 8) & 0xff) + 1;
338 
339 		if (c->x86 == 0x15)
340 			c->cu_id = ebx & 0xff;
341 
342 		if (c->x86 >= 0x17) {
343 			c->cpu_core_id = ebx & 0xff;
344 
345 			if (smp_num_siblings > 1)
346 				c->x86_max_cores /= smp_num_siblings;
347 		}
348 
349 		/*
350 		 * In case leaf B is available, use it to derive
351 		 * topology information.
352 		 */
353 		err = detect_extended_topology(c);
354 		if (!err)
355 			c->x86_coreid_bits = get_count_order(c->x86_max_cores);
356 
357 		cacheinfo_amd_init_llc_id(c, cpu, node_id);
358 
359 	} else if (cpu_has(c, X86_FEATURE_NODEID_MSR)) {
360 		u64 value;
361 
362 		rdmsrl(MSR_FAM10H_NODE_ID, value);
363 		node_id = value & 7;
364 
365 		per_cpu(cpu_llc_id, cpu) = node_id;
366 	} else
367 		return;
368 
369 	if (nodes_per_socket > 1) {
370 		set_cpu_cap(c, X86_FEATURE_AMD_DCM);
371 		legacy_fixup_core_id(c);
372 	}
373 }
374 
375 /*
376  * On a AMD dual core setup the lower bits of the APIC id distinguish the cores.
377  * Assumes number of cores is a power of two.
378  */
379 static void amd_detect_cmp(struct cpuinfo_x86 *c)
380 {
381 	unsigned bits;
382 	int cpu = smp_processor_id();
383 
384 	bits = c->x86_coreid_bits;
385 	/* Low order bits define the core id (index of core in socket) */
386 	c->cpu_core_id = c->initial_apicid & ((1 << bits)-1);
387 	/* Convert the initial APIC ID into the socket ID */
388 	c->phys_proc_id = c->initial_apicid >> bits;
389 	/* use socket ID also for last level cache */
390 	per_cpu(cpu_llc_id, cpu) = c->phys_proc_id;
391 }
392 
393 u16 amd_get_nb_id(int cpu)
394 {
395 	return per_cpu(cpu_llc_id, cpu);
396 }
397 EXPORT_SYMBOL_GPL(amd_get_nb_id);
398 
399 u32 amd_get_nodes_per_socket(void)
400 {
401 	return nodes_per_socket;
402 }
403 EXPORT_SYMBOL_GPL(amd_get_nodes_per_socket);
404 
405 static void srat_detect_node(struct cpuinfo_x86 *c)
406 {
407 #ifdef CONFIG_NUMA
408 	int cpu = smp_processor_id();
409 	int node;
410 	unsigned apicid = c->apicid;
411 
412 	node = numa_cpu_node(cpu);
413 	if (node == NUMA_NO_NODE)
414 		node = per_cpu(cpu_llc_id, cpu);
415 
416 	/*
417 	 * On multi-fabric platform (e.g. Numascale NumaChip) a
418 	 * platform-specific handler needs to be called to fixup some
419 	 * IDs of the CPU.
420 	 */
421 	if (x86_cpuinit.fixup_cpu_id)
422 		x86_cpuinit.fixup_cpu_id(c, node);
423 
424 	if (!node_online(node)) {
425 		/*
426 		 * Two possibilities here:
427 		 *
428 		 * - The CPU is missing memory and no node was created.  In
429 		 *   that case try picking one from a nearby CPU.
430 		 *
431 		 * - The APIC IDs differ from the HyperTransport node IDs
432 		 *   which the K8 northbridge parsing fills in.  Assume
433 		 *   they are all increased by a constant offset, but in
434 		 *   the same order as the HT nodeids.  If that doesn't
435 		 *   result in a usable node fall back to the path for the
436 		 *   previous case.
437 		 *
438 		 * This workaround operates directly on the mapping between
439 		 * APIC ID and NUMA node, assuming certain relationship
440 		 * between APIC ID, HT node ID and NUMA topology.  As going
441 		 * through CPU mapping may alter the outcome, directly
442 		 * access __apicid_to_node[].
443 		 */
444 		int ht_nodeid = c->initial_apicid;
445 
446 		if (__apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
447 			node = __apicid_to_node[ht_nodeid];
448 		/* Pick a nearby node */
449 		if (!node_online(node))
450 			node = nearby_node(apicid);
451 	}
452 	numa_set_node(cpu, node);
453 #endif
454 }
455 
456 static void early_init_amd_mc(struct cpuinfo_x86 *c)
457 {
458 #ifdef CONFIG_SMP
459 	unsigned bits, ecx;
460 
461 	/* Multi core CPU? */
462 	if (c->extended_cpuid_level < 0x80000008)
463 		return;
464 
465 	ecx = cpuid_ecx(0x80000008);
466 
467 	c->x86_max_cores = (ecx & 0xff) + 1;
468 
469 	/* CPU telling us the core id bits shift? */
470 	bits = (ecx >> 12) & 0xF;
471 
472 	/* Otherwise recompute */
473 	if (bits == 0) {
474 		while ((1 << bits) < c->x86_max_cores)
475 			bits++;
476 	}
477 
478 	c->x86_coreid_bits = bits;
479 #endif
480 }
481 
482 static void bsp_init_amd(struct cpuinfo_x86 *c)
483 {
484 
485 #ifdef CONFIG_X86_64
486 	if (c->x86 >= 0xf) {
487 		unsigned long long tseg;
488 
489 		/*
490 		 * Split up direct mapping around the TSEG SMM area.
491 		 * Don't do it for gbpages because there seems very little
492 		 * benefit in doing so.
493 		 */
494 		if (!rdmsrl_safe(MSR_K8_TSEG_ADDR, &tseg)) {
495 			unsigned long pfn = tseg >> PAGE_SHIFT;
496 
497 			pr_debug("tseg: %010llx\n", tseg);
498 			if (pfn_range_is_mapped(pfn, pfn + 1))
499 				set_memory_4k((unsigned long)__va(tseg), 1);
500 		}
501 	}
502 #endif
503 
504 	if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) {
505 
506 		if (c->x86 > 0x10 ||
507 		    (c->x86 == 0x10 && c->x86_model >= 0x2)) {
508 			u64 val;
509 
510 			rdmsrl(MSR_K7_HWCR, val);
511 			if (!(val & BIT(24)))
512 				pr_warn(FW_BUG "TSC doesn't count with P0 frequency!\n");
513 		}
514 	}
515 
516 	if (c->x86 == 0x15) {
517 		unsigned long upperbit;
518 		u32 cpuid, assoc;
519 
520 		cpuid	 = cpuid_edx(0x80000005);
521 		assoc	 = cpuid >> 16 & 0xff;
522 		upperbit = ((cpuid >> 24) << 10) / assoc;
523 
524 		va_align.mask	  = (upperbit - 1) & PAGE_MASK;
525 		va_align.flags    = ALIGN_VA_32 | ALIGN_VA_64;
526 
527 		/* A random value per boot for bit slice [12:upper_bit) */
528 		va_align.bits = get_random_int() & va_align.mask;
529 	}
530 
531 	if (cpu_has(c, X86_FEATURE_MWAITX))
532 		use_mwaitx_delay();
533 
534 	if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
535 		u32 ecx;
536 
537 		ecx = cpuid_ecx(0x8000001e);
538 		nodes_per_socket = ((ecx >> 8) & 7) + 1;
539 	} else if (boot_cpu_has(X86_FEATURE_NODEID_MSR)) {
540 		u64 value;
541 
542 		rdmsrl(MSR_FAM10H_NODE_ID, value);
543 		nodes_per_socket = ((value >> 3) & 7) + 1;
544 	}
545 
546 	if (!boot_cpu_has(X86_FEATURE_AMD_SSBD) &&
547 	    !boot_cpu_has(X86_FEATURE_VIRT_SSBD) &&
548 	    c->x86 >= 0x15 && c->x86 <= 0x17) {
549 		unsigned int bit;
550 
551 		switch (c->x86) {
552 		case 0x15: bit = 54; break;
553 		case 0x16: bit = 33; break;
554 		case 0x17: bit = 10; break;
555 		default: return;
556 		}
557 		/*
558 		 * Try to cache the base value so further operations can
559 		 * avoid RMW. If that faults, do not enable SSBD.
560 		 */
561 		if (!rdmsrl_safe(MSR_AMD64_LS_CFG, &x86_amd_ls_cfg_base)) {
562 			setup_force_cpu_cap(X86_FEATURE_LS_CFG_SSBD);
563 			setup_force_cpu_cap(X86_FEATURE_SSBD);
564 			x86_amd_ls_cfg_ssbd_mask = 1ULL << bit;
565 		}
566 	}
567 }
568 
569 static void early_detect_mem_encrypt(struct cpuinfo_x86 *c)
570 {
571 	u64 msr;
572 
573 	/*
574 	 * BIOS support is required for SME and SEV.
575 	 *   For SME: If BIOS has enabled SME then adjust x86_phys_bits by
576 	 *	      the SME physical address space reduction value.
577 	 *	      If BIOS has not enabled SME then don't advertise the
578 	 *	      SME feature (set in scattered.c).
579 	 *   For SEV: If BIOS has not enabled SEV then don't advertise the
580 	 *            SEV feature (set in scattered.c).
581 	 *
582 	 *   In all cases, since support for SME and SEV requires long mode,
583 	 *   don't advertise the feature under CONFIG_X86_32.
584 	 */
585 	if (cpu_has(c, X86_FEATURE_SME) || cpu_has(c, X86_FEATURE_SEV)) {
586 		/* Check if memory encryption is enabled */
587 		rdmsrl(MSR_K8_SYSCFG, msr);
588 		if (!(msr & MSR_K8_SYSCFG_MEM_ENCRYPT))
589 			goto clear_all;
590 
591 		/*
592 		 * Always adjust physical address bits. Even though this
593 		 * will be a value above 32-bits this is still done for
594 		 * CONFIG_X86_32 so that accurate values are reported.
595 		 */
596 		c->x86_phys_bits -= (cpuid_ebx(0x8000001f) >> 6) & 0x3f;
597 
598 		if (IS_ENABLED(CONFIG_X86_32))
599 			goto clear_all;
600 
601 		rdmsrl(MSR_K7_HWCR, msr);
602 		if (!(msr & MSR_K7_HWCR_SMMLOCK))
603 			goto clear_sev;
604 
605 		return;
606 
607 clear_all:
608 		clear_cpu_cap(c, X86_FEATURE_SME);
609 clear_sev:
610 		clear_cpu_cap(c, X86_FEATURE_SEV);
611 	}
612 }
613 
614 static void early_init_amd(struct cpuinfo_x86 *c)
615 {
616 	u32 dummy;
617 
618 	early_init_amd_mc(c);
619 
620 	rdmsr_safe(MSR_AMD64_PATCH_LEVEL, &c->microcode, &dummy);
621 
622 	/*
623 	 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
624 	 * with P/T states and does not stop in deep C-states
625 	 */
626 	if (c->x86_power & (1 << 8)) {
627 		set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
628 		set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
629 	}
630 
631 	/* Bit 12 of 8000_0007 edx is accumulated power mechanism. */
632 	if (c->x86_power & BIT(12))
633 		set_cpu_cap(c, X86_FEATURE_ACC_POWER);
634 
635 #ifdef CONFIG_X86_64
636 	set_cpu_cap(c, X86_FEATURE_SYSCALL32);
637 #else
638 	/*  Set MTRR capability flag if appropriate */
639 	if (c->x86 == 5)
640 		if (c->x86_model == 13 || c->x86_model == 9 ||
641 		    (c->x86_model == 8 && c->x86_stepping >= 8))
642 			set_cpu_cap(c, X86_FEATURE_K6_MTRR);
643 #endif
644 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
645 	/*
646 	 * ApicID can always be treated as an 8-bit value for AMD APIC versions
647 	 * >= 0x10, but even old K8s came out of reset with version 0x10. So, we
648 	 * can safely set X86_FEATURE_EXTD_APICID unconditionally for families
649 	 * after 16h.
650 	 */
651 	if (boot_cpu_has(X86_FEATURE_APIC)) {
652 		if (c->x86 > 0x16)
653 			set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
654 		else if (c->x86 >= 0xf) {
655 			/* check CPU config space for extended APIC ID */
656 			unsigned int val;
657 
658 			val = read_pci_config(0, 24, 0, 0x68);
659 			if ((val >> 17 & 0x3) == 0x3)
660 				set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
661 		}
662 	}
663 #endif
664 
665 	/*
666 	 * This is only needed to tell the kernel whether to use VMCALL
667 	 * and VMMCALL.  VMMCALL is never executed except under virt, so
668 	 * we can set it unconditionally.
669 	 */
670 	set_cpu_cap(c, X86_FEATURE_VMMCALL);
671 
672 	/* F16h erratum 793, CVE-2013-6885 */
673 	if (c->x86 == 0x16 && c->x86_model <= 0xf)
674 		msr_set_bit(MSR_AMD64_LS_CFG, 15);
675 
676 	/*
677 	 * Check whether the machine is affected by erratum 400. This is
678 	 * used to select the proper idle routine and to enable the check
679 	 * whether the machine is affected in arch_post_acpi_init(), which
680 	 * sets the X86_BUG_AMD_APIC_C1E bug depending on the MSR check.
681 	 */
682 	if (cpu_has_amd_erratum(c, amd_erratum_400))
683 		set_cpu_bug(c, X86_BUG_AMD_E400);
684 
685 	early_detect_mem_encrypt(c);
686 }
687 
688 static void init_amd_k8(struct cpuinfo_x86 *c)
689 {
690 	u32 level;
691 	u64 value;
692 
693 	/* On C+ stepping K8 rep microcode works well for copy/memset */
694 	level = cpuid_eax(1);
695 	if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58)
696 		set_cpu_cap(c, X86_FEATURE_REP_GOOD);
697 
698 	/*
699 	 * Some BIOSes incorrectly force this feature, but only K8 revision D
700 	 * (model = 0x14) and later actually support it.
701 	 * (AMD Erratum #110, docId: 25759).
702 	 */
703 	if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM)) {
704 		clear_cpu_cap(c, X86_FEATURE_LAHF_LM);
705 		if (!rdmsrl_amd_safe(0xc001100d, &value)) {
706 			value &= ~BIT_64(32);
707 			wrmsrl_amd_safe(0xc001100d, value);
708 		}
709 	}
710 
711 	if (!c->x86_model_id[0])
712 		strcpy(c->x86_model_id, "Hammer");
713 
714 #ifdef CONFIG_SMP
715 	/*
716 	 * Disable TLB flush filter by setting HWCR.FFDIS on K8
717 	 * bit 6 of msr C001_0015
718 	 *
719 	 * Errata 63 for SH-B3 steppings
720 	 * Errata 122 for all steppings (F+ have it disabled by default)
721 	 */
722 	msr_set_bit(MSR_K7_HWCR, 6);
723 #endif
724 	set_cpu_bug(c, X86_BUG_SWAPGS_FENCE);
725 }
726 
727 static void init_amd_gh(struct cpuinfo_x86 *c)
728 {
729 #ifdef CONFIG_MMCONF_FAM10H
730 	/* do this for boot cpu */
731 	if (c == &boot_cpu_data)
732 		check_enable_amd_mmconf_dmi();
733 
734 	fam10h_check_enable_mmcfg();
735 #endif
736 
737 	/*
738 	 * Disable GART TLB Walk Errors on Fam10h. We do this here because this
739 	 * is always needed when GART is enabled, even in a kernel which has no
740 	 * MCE support built in. BIOS should disable GartTlbWlk Errors already.
741 	 * If it doesn't, we do it here as suggested by the BKDG.
742 	 *
743 	 * Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=33012
744 	 */
745 	msr_set_bit(MSR_AMD64_MCx_MASK(4), 10);
746 
747 	/*
748 	 * On family 10h BIOS may not have properly enabled WC+ support, causing
749 	 * it to be converted to CD memtype. This may result in performance
750 	 * degradation for certain nested-paging guests. Prevent this conversion
751 	 * by clearing bit 24 in MSR_AMD64_BU_CFG2.
752 	 *
753 	 * NOTE: we want to use the _safe accessors so as not to #GP kvm
754 	 * guests on older kvm hosts.
755 	 */
756 	msr_clear_bit(MSR_AMD64_BU_CFG2, 24);
757 
758 	if (cpu_has_amd_erratum(c, amd_erratum_383))
759 		set_cpu_bug(c, X86_BUG_AMD_TLB_MMATCH);
760 }
761 
762 #define MSR_AMD64_DE_CFG	0xC0011029
763 
764 static void init_amd_ln(struct cpuinfo_x86 *c)
765 {
766 	/*
767 	 * Apply erratum 665 fix unconditionally so machines without a BIOS
768 	 * fix work.
769 	 */
770 	msr_set_bit(MSR_AMD64_DE_CFG, 31);
771 }
772 
773 static void init_amd_bd(struct cpuinfo_x86 *c)
774 {
775 	u64 value;
776 
777 	/* re-enable TopologyExtensions if switched off by BIOS */
778 	if ((c->x86_model >= 0x10) && (c->x86_model <= 0x6f) &&
779 	    !cpu_has(c, X86_FEATURE_TOPOEXT)) {
780 
781 		if (msr_set_bit(0xc0011005, 54) > 0) {
782 			rdmsrl(0xc0011005, value);
783 			if (value & BIT_64(54)) {
784 				set_cpu_cap(c, X86_FEATURE_TOPOEXT);
785 				pr_info_once(FW_INFO "CPU: Re-enabling disabled Topology Extensions Support.\n");
786 			}
787 		}
788 	}
789 
790 	/*
791 	 * The way access filter has a performance penalty on some workloads.
792 	 * Disable it on the affected CPUs.
793 	 */
794 	if ((c->x86_model >= 0x02) && (c->x86_model < 0x20)) {
795 		if (!rdmsrl_safe(MSR_F15H_IC_CFG, &value) && !(value & 0x1E)) {
796 			value |= 0x1E;
797 			wrmsrl_safe(MSR_F15H_IC_CFG, value);
798 		}
799 	}
800 }
801 
802 static void init_amd_zn(struct cpuinfo_x86 *c)
803 {
804 	set_cpu_cap(c, X86_FEATURE_ZEN);
805 	/*
806 	 * Fix erratum 1076: CPB feature bit not being set in CPUID. It affects
807 	 * all up to and including B1.
808 	 */
809 	if (c->x86_model <= 1 && c->x86_stepping <= 1)
810 		set_cpu_cap(c, X86_FEATURE_CPB);
811 }
812 
813 static void init_amd(struct cpuinfo_x86 *c)
814 {
815 	early_init_amd(c);
816 
817 	/*
818 	 * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
819 	 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
820 	 */
821 	clear_cpu_cap(c, 0*32+31);
822 
823 	if (c->x86 >= 0x10)
824 		set_cpu_cap(c, X86_FEATURE_REP_GOOD);
825 
826 	/* get apicid instead of initial apic id from cpuid */
827 	c->apicid = hard_smp_processor_id();
828 
829 	/* K6s reports MCEs but don't actually have all the MSRs */
830 	if (c->x86 < 6)
831 		clear_cpu_cap(c, X86_FEATURE_MCE);
832 
833 	switch (c->x86) {
834 	case 4:    init_amd_k5(c); break;
835 	case 5:    init_amd_k6(c); break;
836 	case 6:	   init_amd_k7(c); break;
837 	case 0xf:  init_amd_k8(c); break;
838 	case 0x10: init_amd_gh(c); break;
839 	case 0x12: init_amd_ln(c); break;
840 	case 0x15: init_amd_bd(c); break;
841 	case 0x17: init_amd_zn(c); break;
842 	}
843 
844 	/*
845 	 * Enable workaround for FXSAVE leak on CPUs
846 	 * without a XSaveErPtr feature
847 	 */
848 	if ((c->x86 >= 6) && (!cpu_has(c, X86_FEATURE_XSAVEERPTR)))
849 		set_cpu_bug(c, X86_BUG_FXSAVE_LEAK);
850 
851 	cpu_detect_cache_sizes(c);
852 
853 	/* Multi core CPU? */
854 	if (c->extended_cpuid_level >= 0x80000008) {
855 		amd_detect_cmp(c);
856 		amd_get_topology(c);
857 		srat_detect_node(c);
858 	}
859 
860 #ifdef CONFIG_X86_32
861 	detect_ht(c);
862 #endif
863 
864 	init_amd_cacheinfo(c);
865 
866 	if (c->x86 >= 0xf)
867 		set_cpu_cap(c, X86_FEATURE_K8);
868 
869 	if (cpu_has(c, X86_FEATURE_XMM2)) {
870 		unsigned long long val;
871 		int ret;
872 
873 		/*
874 		 * A serializing LFENCE has less overhead than MFENCE, so
875 		 * use it for execution serialization.  On families which
876 		 * don't have that MSR, LFENCE is already serializing.
877 		 * msr_set_bit() uses the safe accessors, too, even if the MSR
878 		 * is not present.
879 		 */
880 		msr_set_bit(MSR_F10H_DECFG,
881 			    MSR_F10H_DECFG_LFENCE_SERIALIZE_BIT);
882 
883 		/*
884 		 * Verify that the MSR write was successful (could be running
885 		 * under a hypervisor) and only then assume that LFENCE is
886 		 * serializing.
887 		 */
888 		ret = rdmsrl_safe(MSR_F10H_DECFG, &val);
889 		if (!ret && (val & MSR_F10H_DECFG_LFENCE_SERIALIZE)) {
890 			/* A serializing LFENCE stops RDTSC speculation */
891 			set_cpu_cap(c, X86_FEATURE_LFENCE_RDTSC);
892 		} else {
893 			/* MFENCE stops RDTSC speculation */
894 			set_cpu_cap(c, X86_FEATURE_MFENCE_RDTSC);
895 		}
896 	}
897 
898 	/*
899 	 * Family 0x12 and above processors have APIC timer
900 	 * running in deep C states.
901 	 */
902 	if (c->x86 > 0x11)
903 		set_cpu_cap(c, X86_FEATURE_ARAT);
904 
905 	/* 3DNow or LM implies PREFETCHW */
906 	if (!cpu_has(c, X86_FEATURE_3DNOWPREFETCH))
907 		if (cpu_has(c, X86_FEATURE_3DNOW) || cpu_has(c, X86_FEATURE_LM))
908 			set_cpu_cap(c, X86_FEATURE_3DNOWPREFETCH);
909 
910 	/* AMD CPUs don't reset SS attributes on SYSRET, Xen does. */
911 	if (!cpu_has(c, X86_FEATURE_XENPV))
912 		set_cpu_bug(c, X86_BUG_SYSRET_SS_ATTRS);
913 }
914 
915 #ifdef CONFIG_X86_32
916 static unsigned int amd_size_cache(struct cpuinfo_x86 *c, unsigned int size)
917 {
918 	/* AMD errata T13 (order #21922) */
919 	if ((c->x86 == 6)) {
920 		/* Duron Rev A0 */
921 		if (c->x86_model == 3 && c->x86_stepping == 0)
922 			size = 64;
923 		/* Tbird rev A1/A2 */
924 		if (c->x86_model == 4 &&
925 			(c->x86_stepping == 0 || c->x86_stepping == 1))
926 			size = 256;
927 	}
928 	return size;
929 }
930 #endif
931 
932 static void cpu_detect_tlb_amd(struct cpuinfo_x86 *c)
933 {
934 	u32 ebx, eax, ecx, edx;
935 	u16 mask = 0xfff;
936 
937 	if (c->x86 < 0xf)
938 		return;
939 
940 	if (c->extended_cpuid_level < 0x80000006)
941 		return;
942 
943 	cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
944 
945 	tlb_lld_4k[ENTRIES] = (ebx >> 16) & mask;
946 	tlb_lli_4k[ENTRIES] = ebx & mask;
947 
948 	/*
949 	 * K8 doesn't have 2M/4M entries in the L2 TLB so read out the L1 TLB
950 	 * characteristics from the CPUID function 0x80000005 instead.
951 	 */
952 	if (c->x86 == 0xf) {
953 		cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
954 		mask = 0xff;
955 	}
956 
957 	/* Handle DTLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
958 	if (!((eax >> 16) & mask))
959 		tlb_lld_2m[ENTRIES] = (cpuid_eax(0x80000005) >> 16) & 0xff;
960 	else
961 		tlb_lld_2m[ENTRIES] = (eax >> 16) & mask;
962 
963 	/* a 4M entry uses two 2M entries */
964 	tlb_lld_4m[ENTRIES] = tlb_lld_2m[ENTRIES] >> 1;
965 
966 	/* Handle ITLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
967 	if (!(eax & mask)) {
968 		/* Erratum 658 */
969 		if (c->x86 == 0x15 && c->x86_model <= 0x1f) {
970 			tlb_lli_2m[ENTRIES] = 1024;
971 		} else {
972 			cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
973 			tlb_lli_2m[ENTRIES] = eax & 0xff;
974 		}
975 	} else
976 		tlb_lli_2m[ENTRIES] = eax & mask;
977 
978 	tlb_lli_4m[ENTRIES] = tlb_lli_2m[ENTRIES] >> 1;
979 }
980 
981 static const struct cpu_dev amd_cpu_dev = {
982 	.c_vendor	= "AMD",
983 	.c_ident	= { "AuthenticAMD" },
984 #ifdef CONFIG_X86_32
985 	.legacy_models = {
986 		{ .family = 4, .model_names =
987 		  {
988 			  [3] = "486 DX/2",
989 			  [7] = "486 DX/2-WB",
990 			  [8] = "486 DX/4",
991 			  [9] = "486 DX/4-WB",
992 			  [14] = "Am5x86-WT",
993 			  [15] = "Am5x86-WB"
994 		  }
995 		},
996 	},
997 	.legacy_cache_size = amd_size_cache,
998 #endif
999 	.c_early_init   = early_init_amd,
1000 	.c_detect_tlb	= cpu_detect_tlb_amd,
1001 	.c_bsp_init	= bsp_init_amd,
1002 	.c_init		= init_amd,
1003 	.c_x86_vendor	= X86_VENDOR_AMD,
1004 };
1005 
1006 cpu_dev_register(amd_cpu_dev);
1007 
1008 /*
1009  * AMD errata checking
1010  *
1011  * Errata are defined as arrays of ints using the AMD_LEGACY_ERRATUM() or
1012  * AMD_OSVW_ERRATUM() macros. The latter is intended for newer errata that
1013  * have an OSVW id assigned, which it takes as first argument. Both take a
1014  * variable number of family-specific model-stepping ranges created by
1015  * AMD_MODEL_RANGE().
1016  *
1017  * Example:
1018  *
1019  * const int amd_erratum_319[] =
1020  *	AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0x4, 0x2),
1021  *			   AMD_MODEL_RANGE(0x10, 0x8, 0x0, 0x8, 0x0),
1022  *			   AMD_MODEL_RANGE(0x10, 0x9, 0x0, 0x9, 0x0));
1023  */
1024 
1025 #define AMD_LEGACY_ERRATUM(...)		{ -1, __VA_ARGS__, 0 }
1026 #define AMD_OSVW_ERRATUM(osvw_id, ...)	{ osvw_id, __VA_ARGS__, 0 }
1027 #define AMD_MODEL_RANGE(f, m_start, s_start, m_end, s_end) \
1028 	((f << 24) | (m_start << 16) | (s_start << 12) | (m_end << 4) | (s_end))
1029 #define AMD_MODEL_RANGE_FAMILY(range)	(((range) >> 24) & 0xff)
1030 #define AMD_MODEL_RANGE_START(range)	(((range) >> 12) & 0xfff)
1031 #define AMD_MODEL_RANGE_END(range)	((range) & 0xfff)
1032 
1033 static const int amd_erratum_400[] =
1034 	AMD_OSVW_ERRATUM(1, AMD_MODEL_RANGE(0xf, 0x41, 0x2, 0xff, 0xf),
1035 			    AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0xff, 0xf));
1036 
1037 static const int amd_erratum_383[] =
1038 	AMD_OSVW_ERRATUM(3, AMD_MODEL_RANGE(0x10, 0, 0, 0xff, 0xf));
1039 
1040 
1041 static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum)
1042 {
1043 	int osvw_id = *erratum++;
1044 	u32 range;
1045 	u32 ms;
1046 
1047 	if (osvw_id >= 0 && osvw_id < 65536 &&
1048 	    cpu_has(cpu, X86_FEATURE_OSVW)) {
1049 		u64 osvw_len;
1050 
1051 		rdmsrl(MSR_AMD64_OSVW_ID_LENGTH, osvw_len);
1052 		if (osvw_id < osvw_len) {
1053 			u64 osvw_bits;
1054 
1055 			rdmsrl(MSR_AMD64_OSVW_STATUS + (osvw_id >> 6),
1056 			    osvw_bits);
1057 			return osvw_bits & (1ULL << (osvw_id & 0x3f));
1058 		}
1059 	}
1060 
1061 	/* OSVW unavailable or ID unknown, match family-model-stepping range */
1062 	ms = (cpu->x86_model << 4) | cpu->x86_stepping;
1063 	while ((range = *erratum++))
1064 		if ((cpu->x86 == AMD_MODEL_RANGE_FAMILY(range)) &&
1065 		    (ms >= AMD_MODEL_RANGE_START(range)) &&
1066 		    (ms <= AMD_MODEL_RANGE_END(range)))
1067 			return true;
1068 
1069 	return false;
1070 }
1071 
1072 void set_dr_addr_mask(unsigned long mask, int dr)
1073 {
1074 	if (!boot_cpu_has(X86_FEATURE_BPEXT))
1075 		return;
1076 
1077 	switch (dr) {
1078 	case 0:
1079 		wrmsr(MSR_F16H_DR0_ADDR_MASK, mask, 0);
1080 		break;
1081 	case 1:
1082 	case 2:
1083 	case 3:
1084 		wrmsr(MSR_F16H_DR1_ADDR_MASK - 1 + dr, mask, 0);
1085 		break;
1086 	default:
1087 		break;
1088 	}
1089 }
1090