xref: /openbmc/linux/arch/x86/kernel/cpu/amd.c (revision 1fa0a7dc)
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 #include <asm/debugreg.h>
19 
20 #ifdef CONFIG_X86_64
21 # include <asm/mmconfig.h>
22 # include <asm/set_memory.h>
23 #endif
24 
25 #include "cpu.h"
26 
27 static const int amd_erratum_383[];
28 static const int amd_erratum_400[];
29 static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum);
30 
31 /*
32  * nodes_per_socket: Stores the number of nodes per socket.
33  * Refer to Fam15h Models 00-0fh BKDG - CPUID Fn8000_001E_ECX
34  * Node Identifiers[10:8]
35  */
36 static u32 nodes_per_socket = 1;
37 
38 static inline int rdmsrl_amd_safe(unsigned msr, unsigned long long *p)
39 {
40 	u32 gprs[8] = { 0 };
41 	int err;
42 
43 	WARN_ONCE((boot_cpu_data.x86 != 0xf),
44 		  "%s should only be used on K8!\n", __func__);
45 
46 	gprs[1] = msr;
47 	gprs[7] = 0x9c5a203a;
48 
49 	err = rdmsr_safe_regs(gprs);
50 
51 	*p = gprs[0] | ((u64)gprs[2] << 32);
52 
53 	return err;
54 }
55 
56 static inline int wrmsrl_amd_safe(unsigned msr, unsigned long long val)
57 {
58 	u32 gprs[8] = { 0 };
59 
60 	WARN_ONCE((boot_cpu_data.x86 != 0xf),
61 		  "%s should only be used on K8!\n", __func__);
62 
63 	gprs[0] = (u32)val;
64 	gprs[1] = msr;
65 	gprs[2] = val >> 32;
66 	gprs[7] = 0x9c5a203a;
67 
68 	return wrmsr_safe_regs(gprs);
69 }
70 
71 /*
72  *	B step AMD K6 before B 9730xxxx have hardware bugs that can cause
73  *	misexecution of code under Linux. Owners of such processors should
74  *	contact AMD for precise details and a CPU swap.
75  *
76  *	See	http://www.multimania.com/poulot/k6bug.html
77  *	and	section 2.6.2 of "AMD-K6 Processor Revision Guide - Model 6"
78  *		(Publication # 21266  Issue Date: August 1998)
79  *
80  *	The following test is erm.. interesting. AMD neglected to up
81  *	the chip setting when fixing the bug but they also tweaked some
82  *	performance at the same time..
83  */
84 
85 #ifdef CONFIG_X86_32
86 extern __visible void vide(void);
87 __asm__(".text\n"
88 	".globl vide\n"
89 	".type vide, @function\n"
90 	".align 4\n"
91 	"vide: ret\n");
92 #endif
93 
94 static void init_amd_k5(struct cpuinfo_x86 *c)
95 {
96 #ifdef CONFIG_X86_32
97 /*
98  * General Systems BIOSen alias the cpu frequency registers
99  * of the Elan at 0x000df000. Unfortunately, one of the Linux
100  * drivers subsequently pokes it, and changes the CPU speed.
101  * Workaround : Remove the unneeded alias.
102  */
103 #define CBAR		(0xfffc) /* Configuration Base Address  (32-bit) */
104 #define CBAR_ENB	(0x80000000)
105 #define CBAR_KEY	(0X000000CB)
106 	if (c->x86_model == 9 || c->x86_model == 10) {
107 		if (inl(CBAR) & CBAR_ENB)
108 			outl(0 | CBAR_KEY, CBAR);
109 	}
110 #endif
111 }
112 
113 static void init_amd_k6(struct cpuinfo_x86 *c)
114 {
115 #ifdef CONFIG_X86_32
116 	u32 l, h;
117 	int mbytes = get_num_physpages() >> (20-PAGE_SHIFT);
118 
119 	if (c->x86_model < 6) {
120 		/* Based on AMD doc 20734R - June 2000 */
121 		if (c->x86_model == 0) {
122 			clear_cpu_cap(c, X86_FEATURE_APIC);
123 			set_cpu_cap(c, X86_FEATURE_PGE);
124 		}
125 		return;
126 	}
127 
128 	if (c->x86_model == 6 && c->x86_stepping == 1) {
129 		const int K6_BUG_LOOP = 1000000;
130 		int n;
131 		void (*f_vide)(void);
132 		u64 d, d2;
133 
134 		pr_info("AMD K6 stepping B detected - ");
135 
136 		/*
137 		 * It looks like AMD fixed the 2.6.2 bug and improved indirect
138 		 * calls at the same time.
139 		 */
140 
141 		n = K6_BUG_LOOP;
142 		f_vide = vide;
143 		OPTIMIZER_HIDE_VAR(f_vide);
144 		d = rdtsc();
145 		while (n--)
146 			f_vide();
147 		d2 = rdtsc();
148 		d = d2-d;
149 
150 		if (d > 20*K6_BUG_LOOP)
151 			pr_cont("system stability may be impaired when more than 32 MB are used.\n");
152 		else
153 			pr_cont("probably OK (after B9730xxxx).\n");
154 	}
155 
156 	/* K6 with old style WHCR */
157 	if (c->x86_model < 8 ||
158 	   (c->x86_model == 8 && c->x86_stepping < 8)) {
159 		/* We can only write allocate on the low 508Mb */
160 		if (mbytes > 508)
161 			mbytes = 508;
162 
163 		rdmsr(MSR_K6_WHCR, l, h);
164 		if ((l&0x0000FFFF) == 0) {
165 			unsigned long flags;
166 			l = (1<<0)|((mbytes/4)<<1);
167 			local_irq_save(flags);
168 			wbinvd();
169 			wrmsr(MSR_K6_WHCR, l, h);
170 			local_irq_restore(flags);
171 			pr_info("Enabling old style K6 write allocation for %d Mb\n",
172 				mbytes);
173 		}
174 		return;
175 	}
176 
177 	if ((c->x86_model == 8 && c->x86_stepping > 7) ||
178 	     c->x86_model == 9 || c->x86_model == 13) {
179 		/* The more serious chips .. */
180 
181 		if (mbytes > 4092)
182 			mbytes = 4092;
183 
184 		rdmsr(MSR_K6_WHCR, l, h);
185 		if ((l&0xFFFF0000) == 0) {
186 			unsigned long flags;
187 			l = ((mbytes>>2)<<22)|(1<<16);
188 			local_irq_save(flags);
189 			wbinvd();
190 			wrmsr(MSR_K6_WHCR, l, h);
191 			local_irq_restore(flags);
192 			pr_info("Enabling new style K6 write allocation for %d Mb\n",
193 				mbytes);
194 		}
195 
196 		return;
197 	}
198 
199 	if (c->x86_model == 10) {
200 		/* AMD Geode LX is model 10 */
201 		/* placeholder for any needed mods */
202 		return;
203 	}
204 #endif
205 }
206 
207 static void init_amd_k7(struct cpuinfo_x86 *c)
208 {
209 #ifdef CONFIG_X86_32
210 	u32 l, h;
211 
212 	/*
213 	 * Bit 15 of Athlon specific MSR 15, needs to be 0
214 	 * to enable SSE on Palomino/Morgan/Barton CPU's.
215 	 * If the BIOS didn't enable it already, enable it here.
216 	 */
217 	if (c->x86_model >= 6 && c->x86_model <= 10) {
218 		if (!cpu_has(c, X86_FEATURE_XMM)) {
219 			pr_info("Enabling disabled K7/SSE Support.\n");
220 			msr_clear_bit(MSR_K7_HWCR, 15);
221 			set_cpu_cap(c, X86_FEATURE_XMM);
222 		}
223 	}
224 
225 	/*
226 	 * It's been determined by AMD that Athlons since model 8 stepping 1
227 	 * are more robust with CLK_CTL set to 200xxxxx instead of 600xxxxx
228 	 * As per AMD technical note 27212 0.2
229 	 */
230 	if ((c->x86_model == 8 && c->x86_stepping >= 1) || (c->x86_model > 8)) {
231 		rdmsr(MSR_K7_CLK_CTL, l, h);
232 		if ((l & 0xfff00000) != 0x20000000) {
233 			pr_info("CPU: CLK_CTL MSR was %x. Reprogramming to %x\n",
234 				l, ((l & 0x000fffff)|0x20000000));
235 			wrmsr(MSR_K7_CLK_CTL, (l & 0x000fffff)|0x20000000, h);
236 		}
237 	}
238 
239 	/* calling is from identify_secondary_cpu() ? */
240 	if (!c->cpu_index)
241 		return;
242 
243 	/*
244 	 * Certain Athlons might work (for various values of 'work') in SMP
245 	 * but they are not certified as MP capable.
246 	 */
247 	/* Athlon 660/661 is valid. */
248 	if ((c->x86_model == 6) && ((c->x86_stepping == 0) ||
249 	    (c->x86_stepping == 1)))
250 		return;
251 
252 	/* Duron 670 is valid */
253 	if ((c->x86_model == 7) && (c->x86_stepping == 0))
254 		return;
255 
256 	/*
257 	 * Athlon 662, Duron 671, and Athlon >model 7 have capability
258 	 * bit. It's worth noting that the A5 stepping (662) of some
259 	 * Athlon XP's have the MP bit set.
260 	 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for
261 	 * more.
262 	 */
263 	if (((c->x86_model == 6) && (c->x86_stepping >= 2)) ||
264 	    ((c->x86_model == 7) && (c->x86_stepping >= 1)) ||
265 	     (c->x86_model > 7))
266 		if (cpu_has(c, X86_FEATURE_MP))
267 			return;
268 
269 	/* If we get here, not a certified SMP capable AMD system. */
270 
271 	/*
272 	 * Don't taint if we are running SMP kernel on a single non-MP
273 	 * approved Athlon
274 	 */
275 	WARN_ONCE(1, "WARNING: This combination of AMD"
276 		" processors is not suitable for SMP.\n");
277 	add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_NOW_UNRELIABLE);
278 #endif
279 }
280 
281 #ifdef CONFIG_NUMA
282 /*
283  * To workaround broken NUMA config.  Read the comment in
284  * srat_detect_node().
285  */
286 static int nearby_node(int apicid)
287 {
288 	int i, node;
289 
290 	for (i = apicid - 1; i >= 0; i--) {
291 		node = __apicid_to_node[i];
292 		if (node != NUMA_NO_NODE && node_online(node))
293 			return node;
294 	}
295 	for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) {
296 		node = __apicid_to_node[i];
297 		if (node != NUMA_NO_NODE && node_online(node))
298 			return node;
299 	}
300 	return first_node(node_online_map); /* Shouldn't happen */
301 }
302 #endif
303 
304 /*
305  * Fix up cpu_core_id for pre-F17h systems to be in the
306  * [0 .. cores_per_node - 1] range. Not really needed but
307  * kept so as not to break existing setups.
308  */
309 static void legacy_fixup_core_id(struct cpuinfo_x86 *c)
310 {
311 	u32 cus_per_node;
312 
313 	if (c->x86 >= 0x17)
314 		return;
315 
316 	cus_per_node = c->x86_max_cores / nodes_per_socket;
317 	c->cpu_core_id %= cus_per_node;
318 }
319 
320 
321 static void amd_get_topology_early(struct cpuinfo_x86 *c)
322 {
323 	if (cpu_has(c, X86_FEATURE_TOPOEXT))
324 		smp_num_siblings = ((cpuid_ebx(0x8000001e) >> 8) & 0xff) + 1;
325 }
326 
327 /*
328  * Fixup core topology information for
329  * (1) AMD multi-node processors
330  *     Assumption: Number of cores in each internal node is the same.
331  * (2) AMD processors supporting compute units
332  */
333 static void amd_get_topology(struct cpuinfo_x86 *c)
334 {
335 	u8 node_id;
336 	int cpu = smp_processor_id();
337 
338 	/* get information required for multi-node processors */
339 	if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
340 		int err;
341 		u32 eax, ebx, ecx, edx;
342 
343 		cpuid(0x8000001e, &eax, &ebx, &ecx, &edx);
344 
345 		node_id  = ecx & 0xff;
346 
347 		if (c->x86 == 0x15)
348 			c->cu_id = ebx & 0xff;
349 
350 		if (c->x86 >= 0x17) {
351 			c->cpu_core_id = ebx & 0xff;
352 
353 			if (smp_num_siblings > 1)
354 				c->x86_max_cores /= smp_num_siblings;
355 		}
356 
357 		/*
358 		 * In case leaf B is available, use it to derive
359 		 * topology information.
360 		 */
361 		err = detect_extended_topology(c);
362 		if (!err)
363 			c->x86_coreid_bits = get_count_order(c->x86_max_cores);
364 
365 		cacheinfo_amd_init_llc_id(c, cpu, node_id);
366 
367 	} else if (cpu_has(c, X86_FEATURE_NODEID_MSR)) {
368 		u64 value;
369 
370 		rdmsrl(MSR_FAM10H_NODE_ID, value);
371 		node_id = value & 7;
372 
373 		per_cpu(cpu_llc_id, cpu) = node_id;
374 	} else
375 		return;
376 
377 	if (nodes_per_socket > 1) {
378 		set_cpu_cap(c, X86_FEATURE_AMD_DCM);
379 		legacy_fixup_core_id(c);
380 	}
381 }
382 
383 /*
384  * On a AMD dual core setup the lower bits of the APIC id distinguish the cores.
385  * Assumes number of cores is a power of two.
386  */
387 static void amd_detect_cmp(struct cpuinfo_x86 *c)
388 {
389 	unsigned bits;
390 	int cpu = smp_processor_id();
391 
392 	bits = c->x86_coreid_bits;
393 	/* Low order bits define the core id (index of core in socket) */
394 	c->cpu_core_id = c->initial_apicid & ((1 << bits)-1);
395 	/* Convert the initial APIC ID into the socket ID */
396 	c->phys_proc_id = c->initial_apicid >> bits;
397 	/* use socket ID also for last level cache */
398 	per_cpu(cpu_llc_id, cpu) = c->phys_proc_id;
399 }
400 
401 u16 amd_get_nb_id(int cpu)
402 {
403 	return per_cpu(cpu_llc_id, cpu);
404 }
405 EXPORT_SYMBOL_GPL(amd_get_nb_id);
406 
407 u32 amd_get_nodes_per_socket(void)
408 {
409 	return nodes_per_socket;
410 }
411 EXPORT_SYMBOL_GPL(amd_get_nodes_per_socket);
412 
413 static void srat_detect_node(struct cpuinfo_x86 *c)
414 {
415 #ifdef CONFIG_NUMA
416 	int cpu = smp_processor_id();
417 	int node;
418 	unsigned apicid = c->apicid;
419 
420 	node = numa_cpu_node(cpu);
421 	if (node == NUMA_NO_NODE)
422 		node = per_cpu(cpu_llc_id, cpu);
423 
424 	/*
425 	 * On multi-fabric platform (e.g. Numascale NumaChip) a
426 	 * platform-specific handler needs to be called to fixup some
427 	 * IDs of the CPU.
428 	 */
429 	if (x86_cpuinit.fixup_cpu_id)
430 		x86_cpuinit.fixup_cpu_id(c, node);
431 
432 	if (!node_online(node)) {
433 		/*
434 		 * Two possibilities here:
435 		 *
436 		 * - The CPU is missing memory and no node was created.  In
437 		 *   that case try picking one from a nearby CPU.
438 		 *
439 		 * - The APIC IDs differ from the HyperTransport node IDs
440 		 *   which the K8 northbridge parsing fills in.  Assume
441 		 *   they are all increased by a constant offset, but in
442 		 *   the same order as the HT nodeids.  If that doesn't
443 		 *   result in a usable node fall back to the path for the
444 		 *   previous case.
445 		 *
446 		 * This workaround operates directly on the mapping between
447 		 * APIC ID and NUMA node, assuming certain relationship
448 		 * between APIC ID, HT node ID and NUMA topology.  As going
449 		 * through CPU mapping may alter the outcome, directly
450 		 * access __apicid_to_node[].
451 		 */
452 		int ht_nodeid = c->initial_apicid;
453 
454 		if (__apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
455 			node = __apicid_to_node[ht_nodeid];
456 		/* Pick a nearby node */
457 		if (!node_online(node))
458 			node = nearby_node(apicid);
459 	}
460 	numa_set_node(cpu, node);
461 #endif
462 }
463 
464 static void early_init_amd_mc(struct cpuinfo_x86 *c)
465 {
466 #ifdef CONFIG_SMP
467 	unsigned bits, ecx;
468 
469 	/* Multi core CPU? */
470 	if (c->extended_cpuid_level < 0x80000008)
471 		return;
472 
473 	ecx = cpuid_ecx(0x80000008);
474 
475 	c->x86_max_cores = (ecx & 0xff) + 1;
476 
477 	/* CPU telling us the core id bits shift? */
478 	bits = (ecx >> 12) & 0xF;
479 
480 	/* Otherwise recompute */
481 	if (bits == 0) {
482 		while ((1 << bits) < c->x86_max_cores)
483 			bits++;
484 	}
485 
486 	c->x86_coreid_bits = bits;
487 #endif
488 }
489 
490 static void bsp_init_amd(struct cpuinfo_x86 *c)
491 {
492 
493 #ifdef CONFIG_X86_64
494 	if (c->x86 >= 0xf) {
495 		unsigned long long tseg;
496 
497 		/*
498 		 * Split up direct mapping around the TSEG SMM area.
499 		 * Don't do it for gbpages because there seems very little
500 		 * benefit in doing so.
501 		 */
502 		if (!rdmsrl_safe(MSR_K8_TSEG_ADDR, &tseg)) {
503 			unsigned long pfn = tseg >> PAGE_SHIFT;
504 
505 			pr_debug("tseg: %010llx\n", tseg);
506 			if (pfn_range_is_mapped(pfn, pfn + 1))
507 				set_memory_4k((unsigned long)__va(tseg), 1);
508 		}
509 	}
510 #endif
511 
512 	if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) {
513 
514 		if (c->x86 > 0x10 ||
515 		    (c->x86 == 0x10 && c->x86_model >= 0x2)) {
516 			u64 val;
517 
518 			rdmsrl(MSR_K7_HWCR, val);
519 			if (!(val & BIT(24)))
520 				pr_warn(FW_BUG "TSC doesn't count with P0 frequency!\n");
521 		}
522 	}
523 
524 	if (c->x86 == 0x15) {
525 		unsigned long upperbit;
526 		u32 cpuid, assoc;
527 
528 		cpuid	 = cpuid_edx(0x80000005);
529 		assoc	 = cpuid >> 16 & 0xff;
530 		upperbit = ((cpuid >> 24) << 10) / assoc;
531 
532 		va_align.mask	  = (upperbit - 1) & PAGE_MASK;
533 		va_align.flags    = ALIGN_VA_32 | ALIGN_VA_64;
534 
535 		/* A random value per boot for bit slice [12:upper_bit) */
536 		va_align.bits = get_random_int() & va_align.mask;
537 	}
538 
539 	if (cpu_has(c, X86_FEATURE_MWAITX))
540 		use_mwaitx_delay();
541 
542 	if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
543 		u32 ecx;
544 
545 		ecx = cpuid_ecx(0x8000001e);
546 		nodes_per_socket = ((ecx >> 8) & 7) + 1;
547 	} else if (boot_cpu_has(X86_FEATURE_NODEID_MSR)) {
548 		u64 value;
549 
550 		rdmsrl(MSR_FAM10H_NODE_ID, value);
551 		nodes_per_socket = ((value >> 3) & 7) + 1;
552 	}
553 
554 	if (!boot_cpu_has(X86_FEATURE_AMD_SSBD) &&
555 	    !boot_cpu_has(X86_FEATURE_VIRT_SSBD) &&
556 	    c->x86 >= 0x15 && c->x86 <= 0x17) {
557 		unsigned int bit;
558 
559 		switch (c->x86) {
560 		case 0x15: bit = 54; break;
561 		case 0x16: bit = 33; break;
562 		case 0x17: bit = 10; break;
563 		default: return;
564 		}
565 		/*
566 		 * Try to cache the base value so further operations can
567 		 * avoid RMW. If that faults, do not enable SSBD.
568 		 */
569 		if (!rdmsrl_safe(MSR_AMD64_LS_CFG, &x86_amd_ls_cfg_base)) {
570 			setup_force_cpu_cap(X86_FEATURE_LS_CFG_SSBD);
571 			setup_force_cpu_cap(X86_FEATURE_SSBD);
572 			x86_amd_ls_cfg_ssbd_mask = 1ULL << bit;
573 		}
574 	}
575 }
576 
577 static void early_detect_mem_encrypt(struct cpuinfo_x86 *c)
578 {
579 	u64 msr;
580 
581 	/*
582 	 * BIOS support is required for SME and SEV.
583 	 *   For SME: If BIOS has enabled SME then adjust x86_phys_bits by
584 	 *	      the SME physical address space reduction value.
585 	 *	      If BIOS has not enabled SME then don't advertise the
586 	 *	      SME feature (set in scattered.c).
587 	 *   For SEV: If BIOS has not enabled SEV then don't advertise the
588 	 *            SEV feature (set in scattered.c).
589 	 *
590 	 *   In all cases, since support for SME and SEV requires long mode,
591 	 *   don't advertise the feature under CONFIG_X86_32.
592 	 */
593 	if (cpu_has(c, X86_FEATURE_SME) || cpu_has(c, X86_FEATURE_SEV)) {
594 		/* Check if memory encryption is enabled */
595 		rdmsrl(MSR_K8_SYSCFG, msr);
596 		if (!(msr & MSR_K8_SYSCFG_MEM_ENCRYPT))
597 			goto clear_all;
598 
599 		/*
600 		 * Always adjust physical address bits. Even though this
601 		 * will be a value above 32-bits this is still done for
602 		 * CONFIG_X86_32 so that accurate values are reported.
603 		 */
604 		c->x86_phys_bits -= (cpuid_ebx(0x8000001f) >> 6) & 0x3f;
605 
606 		if (IS_ENABLED(CONFIG_X86_32))
607 			goto clear_all;
608 
609 		rdmsrl(MSR_K7_HWCR, msr);
610 		if (!(msr & MSR_K7_HWCR_SMMLOCK))
611 			goto clear_sev;
612 
613 		return;
614 
615 clear_all:
616 		clear_cpu_cap(c, X86_FEATURE_SME);
617 clear_sev:
618 		clear_cpu_cap(c, X86_FEATURE_SEV);
619 	}
620 }
621 
622 static void early_init_amd(struct cpuinfo_x86 *c)
623 {
624 	u64 value;
625 	u32 dummy;
626 
627 	early_init_amd_mc(c);
628 
629 #ifdef CONFIG_X86_32
630 	if (c->x86 == 6)
631 		set_cpu_cap(c, X86_FEATURE_K7);
632 #endif
633 
634 	if (c->x86 >= 0xf)
635 		set_cpu_cap(c, X86_FEATURE_K8);
636 
637 	rdmsr_safe(MSR_AMD64_PATCH_LEVEL, &c->microcode, &dummy);
638 
639 	/*
640 	 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
641 	 * with P/T states and does not stop in deep C-states
642 	 */
643 	if (c->x86_power & (1 << 8)) {
644 		set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
645 		set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
646 	}
647 
648 	/* Bit 12 of 8000_0007 edx is accumulated power mechanism. */
649 	if (c->x86_power & BIT(12))
650 		set_cpu_cap(c, X86_FEATURE_ACC_POWER);
651 
652 #ifdef CONFIG_X86_64
653 	set_cpu_cap(c, X86_FEATURE_SYSCALL32);
654 #else
655 	/*  Set MTRR capability flag if appropriate */
656 	if (c->x86 == 5)
657 		if (c->x86_model == 13 || c->x86_model == 9 ||
658 		    (c->x86_model == 8 && c->x86_stepping >= 8))
659 			set_cpu_cap(c, X86_FEATURE_K6_MTRR);
660 #endif
661 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
662 	/*
663 	 * ApicID can always be treated as an 8-bit value for AMD APIC versions
664 	 * >= 0x10, but even old K8s came out of reset with version 0x10. So, we
665 	 * can safely set X86_FEATURE_EXTD_APICID unconditionally for families
666 	 * after 16h.
667 	 */
668 	if (boot_cpu_has(X86_FEATURE_APIC)) {
669 		if (c->x86 > 0x16)
670 			set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
671 		else if (c->x86 >= 0xf) {
672 			/* check CPU config space for extended APIC ID */
673 			unsigned int val;
674 
675 			val = read_pci_config(0, 24, 0, 0x68);
676 			if ((val >> 17 & 0x3) == 0x3)
677 				set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
678 		}
679 	}
680 #endif
681 
682 	/*
683 	 * This is only needed to tell the kernel whether to use VMCALL
684 	 * and VMMCALL.  VMMCALL is never executed except under virt, so
685 	 * we can set it unconditionally.
686 	 */
687 	set_cpu_cap(c, X86_FEATURE_VMMCALL);
688 
689 	/* F16h erratum 793, CVE-2013-6885 */
690 	if (c->x86 == 0x16 && c->x86_model <= 0xf)
691 		msr_set_bit(MSR_AMD64_LS_CFG, 15);
692 
693 	/*
694 	 * Check whether the machine is affected by erratum 400. This is
695 	 * used to select the proper idle routine and to enable the check
696 	 * whether the machine is affected in arch_post_acpi_init(), which
697 	 * sets the X86_BUG_AMD_APIC_C1E bug depending on the MSR check.
698 	 */
699 	if (cpu_has_amd_erratum(c, amd_erratum_400))
700 		set_cpu_bug(c, X86_BUG_AMD_E400);
701 
702 	early_detect_mem_encrypt(c);
703 
704 	/* Re-enable TopologyExtensions if switched off by BIOS */
705 	if (c->x86 == 0x15 &&
706 	    (c->x86_model >= 0x10 && c->x86_model <= 0x6f) &&
707 	    !cpu_has(c, X86_FEATURE_TOPOEXT)) {
708 
709 		if (msr_set_bit(0xc0011005, 54) > 0) {
710 			rdmsrl(0xc0011005, value);
711 			if (value & BIT_64(54)) {
712 				set_cpu_cap(c, X86_FEATURE_TOPOEXT);
713 				pr_info_once(FW_INFO "CPU: Re-enabling disabled Topology Extensions Support.\n");
714 			}
715 		}
716 	}
717 
718 	amd_get_topology_early(c);
719 }
720 
721 static void init_amd_k8(struct cpuinfo_x86 *c)
722 {
723 	u32 level;
724 	u64 value;
725 
726 	/* On C+ stepping K8 rep microcode works well for copy/memset */
727 	level = cpuid_eax(1);
728 	if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58)
729 		set_cpu_cap(c, X86_FEATURE_REP_GOOD);
730 
731 	/*
732 	 * Some BIOSes incorrectly force this feature, but only K8 revision D
733 	 * (model = 0x14) and later actually support it.
734 	 * (AMD Erratum #110, docId: 25759).
735 	 */
736 	if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM)) {
737 		clear_cpu_cap(c, X86_FEATURE_LAHF_LM);
738 		if (!rdmsrl_amd_safe(0xc001100d, &value)) {
739 			value &= ~BIT_64(32);
740 			wrmsrl_amd_safe(0xc001100d, value);
741 		}
742 	}
743 
744 	if (!c->x86_model_id[0])
745 		strcpy(c->x86_model_id, "Hammer");
746 
747 #ifdef CONFIG_SMP
748 	/*
749 	 * Disable TLB flush filter by setting HWCR.FFDIS on K8
750 	 * bit 6 of msr C001_0015
751 	 *
752 	 * Errata 63 for SH-B3 steppings
753 	 * Errata 122 for all steppings (F+ have it disabled by default)
754 	 */
755 	msr_set_bit(MSR_K7_HWCR, 6);
756 #endif
757 	set_cpu_bug(c, X86_BUG_SWAPGS_FENCE);
758 }
759 
760 static void init_amd_gh(struct cpuinfo_x86 *c)
761 {
762 #ifdef CONFIG_MMCONF_FAM10H
763 	/* do this for boot cpu */
764 	if (c == &boot_cpu_data)
765 		check_enable_amd_mmconf_dmi();
766 
767 	fam10h_check_enable_mmcfg();
768 #endif
769 
770 	/*
771 	 * Disable GART TLB Walk Errors on Fam10h. We do this here because this
772 	 * is always needed when GART is enabled, even in a kernel which has no
773 	 * MCE support built in. BIOS should disable GartTlbWlk Errors already.
774 	 * If it doesn't, we do it here as suggested by the BKDG.
775 	 *
776 	 * Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=33012
777 	 */
778 	msr_set_bit(MSR_AMD64_MCx_MASK(4), 10);
779 
780 	/*
781 	 * On family 10h BIOS may not have properly enabled WC+ support, causing
782 	 * it to be converted to CD memtype. This may result in performance
783 	 * degradation for certain nested-paging guests. Prevent this conversion
784 	 * by clearing bit 24 in MSR_AMD64_BU_CFG2.
785 	 *
786 	 * NOTE: we want to use the _safe accessors so as not to #GP kvm
787 	 * guests on older kvm hosts.
788 	 */
789 	msr_clear_bit(MSR_AMD64_BU_CFG2, 24);
790 
791 	if (cpu_has_amd_erratum(c, amd_erratum_383))
792 		set_cpu_bug(c, X86_BUG_AMD_TLB_MMATCH);
793 }
794 
795 #define MSR_AMD64_DE_CFG	0xC0011029
796 
797 static void init_amd_ln(struct cpuinfo_x86 *c)
798 {
799 	/*
800 	 * Apply erratum 665 fix unconditionally so machines without a BIOS
801 	 * fix work.
802 	 */
803 	msr_set_bit(MSR_AMD64_DE_CFG, 31);
804 }
805 
806 static void init_amd_bd(struct cpuinfo_x86 *c)
807 {
808 	u64 value;
809 
810 	/*
811 	 * The way access filter has a performance penalty on some workloads.
812 	 * Disable it on the affected CPUs.
813 	 */
814 	if ((c->x86_model >= 0x02) && (c->x86_model < 0x20)) {
815 		if (!rdmsrl_safe(MSR_F15H_IC_CFG, &value) && !(value & 0x1E)) {
816 			value |= 0x1E;
817 			wrmsrl_safe(MSR_F15H_IC_CFG, value);
818 		}
819 	}
820 }
821 
822 static void init_amd_zn(struct cpuinfo_x86 *c)
823 {
824 	set_cpu_cap(c, X86_FEATURE_ZEN);
825 
826 	/* Fix erratum 1076: CPB feature bit not being set in CPUID. */
827 	if (!cpu_has(c, X86_FEATURE_CPB))
828 		set_cpu_cap(c, X86_FEATURE_CPB);
829 }
830 
831 static void init_amd(struct cpuinfo_x86 *c)
832 {
833 	early_init_amd(c);
834 
835 	/*
836 	 * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
837 	 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
838 	 */
839 	clear_cpu_cap(c, 0*32+31);
840 
841 	if (c->x86 >= 0x10)
842 		set_cpu_cap(c, X86_FEATURE_REP_GOOD);
843 
844 	/* get apicid instead of initial apic id from cpuid */
845 	c->apicid = hard_smp_processor_id();
846 
847 	/* K6s reports MCEs but don't actually have all the MSRs */
848 	if (c->x86 < 6)
849 		clear_cpu_cap(c, X86_FEATURE_MCE);
850 
851 	switch (c->x86) {
852 	case 4:    init_amd_k5(c); break;
853 	case 5:    init_amd_k6(c); break;
854 	case 6:	   init_amd_k7(c); break;
855 	case 0xf:  init_amd_k8(c); break;
856 	case 0x10: init_amd_gh(c); break;
857 	case 0x12: init_amd_ln(c); break;
858 	case 0x15: init_amd_bd(c); break;
859 	case 0x17: init_amd_zn(c); break;
860 	}
861 
862 	/*
863 	 * Enable workaround for FXSAVE leak on CPUs
864 	 * without a XSaveErPtr feature
865 	 */
866 	if ((c->x86 >= 6) && (!cpu_has(c, X86_FEATURE_XSAVEERPTR)))
867 		set_cpu_bug(c, X86_BUG_FXSAVE_LEAK);
868 
869 	cpu_detect_cache_sizes(c);
870 
871 	amd_detect_cmp(c);
872 	amd_get_topology(c);
873 	srat_detect_node(c);
874 
875 	init_amd_cacheinfo(c);
876 
877 	if (cpu_has(c, X86_FEATURE_XMM2)) {
878 		unsigned long long val;
879 		int ret;
880 
881 		/*
882 		 * A serializing LFENCE has less overhead than MFENCE, so
883 		 * use it for execution serialization.  On families which
884 		 * don't have that MSR, LFENCE is already serializing.
885 		 * msr_set_bit() uses the safe accessors, too, even if the MSR
886 		 * is not present.
887 		 */
888 		msr_set_bit(MSR_F10H_DECFG,
889 			    MSR_F10H_DECFG_LFENCE_SERIALIZE_BIT);
890 
891 		/*
892 		 * Verify that the MSR write was successful (could be running
893 		 * under a hypervisor) and only then assume that LFENCE is
894 		 * serializing.
895 		 */
896 		ret = rdmsrl_safe(MSR_F10H_DECFG, &val);
897 		if (!ret && (val & MSR_F10H_DECFG_LFENCE_SERIALIZE)) {
898 			/* A serializing LFENCE stops RDTSC speculation */
899 			set_cpu_cap(c, X86_FEATURE_LFENCE_RDTSC);
900 		} else {
901 			/* MFENCE stops RDTSC speculation */
902 			set_cpu_cap(c, X86_FEATURE_MFENCE_RDTSC);
903 		}
904 	}
905 
906 	/*
907 	 * Family 0x12 and above processors have APIC timer
908 	 * running in deep C states.
909 	 */
910 	if (c->x86 > 0x11)
911 		set_cpu_cap(c, X86_FEATURE_ARAT);
912 
913 	/* 3DNow or LM implies PREFETCHW */
914 	if (!cpu_has(c, X86_FEATURE_3DNOWPREFETCH))
915 		if (cpu_has(c, X86_FEATURE_3DNOW) || cpu_has(c, X86_FEATURE_LM))
916 			set_cpu_cap(c, X86_FEATURE_3DNOWPREFETCH);
917 
918 	/* AMD CPUs don't reset SS attributes on SYSRET, Xen does. */
919 	if (!cpu_has(c, X86_FEATURE_XENPV))
920 		set_cpu_bug(c, X86_BUG_SYSRET_SS_ATTRS);
921 }
922 
923 #ifdef CONFIG_X86_32
924 static unsigned int amd_size_cache(struct cpuinfo_x86 *c, unsigned int size)
925 {
926 	/* AMD errata T13 (order #21922) */
927 	if (c->x86 == 6) {
928 		/* Duron Rev A0 */
929 		if (c->x86_model == 3 && c->x86_stepping == 0)
930 			size = 64;
931 		/* Tbird rev A1/A2 */
932 		if (c->x86_model == 4 &&
933 			(c->x86_stepping == 0 || c->x86_stepping == 1))
934 			size = 256;
935 	}
936 	return size;
937 }
938 #endif
939 
940 static void cpu_detect_tlb_amd(struct cpuinfo_x86 *c)
941 {
942 	u32 ebx, eax, ecx, edx;
943 	u16 mask = 0xfff;
944 
945 	if (c->x86 < 0xf)
946 		return;
947 
948 	if (c->extended_cpuid_level < 0x80000006)
949 		return;
950 
951 	cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
952 
953 	tlb_lld_4k[ENTRIES] = (ebx >> 16) & mask;
954 	tlb_lli_4k[ENTRIES] = ebx & mask;
955 
956 	/*
957 	 * K8 doesn't have 2M/4M entries in the L2 TLB so read out the L1 TLB
958 	 * characteristics from the CPUID function 0x80000005 instead.
959 	 */
960 	if (c->x86 == 0xf) {
961 		cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
962 		mask = 0xff;
963 	}
964 
965 	/* Handle DTLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
966 	if (!((eax >> 16) & mask))
967 		tlb_lld_2m[ENTRIES] = (cpuid_eax(0x80000005) >> 16) & 0xff;
968 	else
969 		tlb_lld_2m[ENTRIES] = (eax >> 16) & mask;
970 
971 	/* a 4M entry uses two 2M entries */
972 	tlb_lld_4m[ENTRIES] = tlb_lld_2m[ENTRIES] >> 1;
973 
974 	/* Handle ITLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
975 	if (!(eax & mask)) {
976 		/* Erratum 658 */
977 		if (c->x86 == 0x15 && c->x86_model <= 0x1f) {
978 			tlb_lli_2m[ENTRIES] = 1024;
979 		} else {
980 			cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
981 			tlb_lli_2m[ENTRIES] = eax & 0xff;
982 		}
983 	} else
984 		tlb_lli_2m[ENTRIES] = eax & mask;
985 
986 	tlb_lli_4m[ENTRIES] = tlb_lli_2m[ENTRIES] >> 1;
987 }
988 
989 static const struct cpu_dev amd_cpu_dev = {
990 	.c_vendor	= "AMD",
991 	.c_ident	= { "AuthenticAMD" },
992 #ifdef CONFIG_X86_32
993 	.legacy_models = {
994 		{ .family = 4, .model_names =
995 		  {
996 			  [3] = "486 DX/2",
997 			  [7] = "486 DX/2-WB",
998 			  [8] = "486 DX/4",
999 			  [9] = "486 DX/4-WB",
1000 			  [14] = "Am5x86-WT",
1001 			  [15] = "Am5x86-WB"
1002 		  }
1003 		},
1004 	},
1005 	.legacy_cache_size = amd_size_cache,
1006 #endif
1007 	.c_early_init   = early_init_amd,
1008 	.c_detect_tlb	= cpu_detect_tlb_amd,
1009 	.c_bsp_init	= bsp_init_amd,
1010 	.c_init		= init_amd,
1011 	.c_x86_vendor	= X86_VENDOR_AMD,
1012 };
1013 
1014 cpu_dev_register(amd_cpu_dev);
1015 
1016 /*
1017  * AMD errata checking
1018  *
1019  * Errata are defined as arrays of ints using the AMD_LEGACY_ERRATUM() or
1020  * AMD_OSVW_ERRATUM() macros. The latter is intended for newer errata that
1021  * have an OSVW id assigned, which it takes as first argument. Both take a
1022  * variable number of family-specific model-stepping ranges created by
1023  * AMD_MODEL_RANGE().
1024  *
1025  * Example:
1026  *
1027  * const int amd_erratum_319[] =
1028  *	AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0x4, 0x2),
1029  *			   AMD_MODEL_RANGE(0x10, 0x8, 0x0, 0x8, 0x0),
1030  *			   AMD_MODEL_RANGE(0x10, 0x9, 0x0, 0x9, 0x0));
1031  */
1032 
1033 #define AMD_LEGACY_ERRATUM(...)		{ -1, __VA_ARGS__, 0 }
1034 #define AMD_OSVW_ERRATUM(osvw_id, ...)	{ osvw_id, __VA_ARGS__, 0 }
1035 #define AMD_MODEL_RANGE(f, m_start, s_start, m_end, s_end) \
1036 	((f << 24) | (m_start << 16) | (s_start << 12) | (m_end << 4) | (s_end))
1037 #define AMD_MODEL_RANGE_FAMILY(range)	(((range) >> 24) & 0xff)
1038 #define AMD_MODEL_RANGE_START(range)	(((range) >> 12) & 0xfff)
1039 #define AMD_MODEL_RANGE_END(range)	((range) & 0xfff)
1040 
1041 static const int amd_erratum_400[] =
1042 	AMD_OSVW_ERRATUM(1, AMD_MODEL_RANGE(0xf, 0x41, 0x2, 0xff, 0xf),
1043 			    AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0xff, 0xf));
1044 
1045 static const int amd_erratum_383[] =
1046 	AMD_OSVW_ERRATUM(3, AMD_MODEL_RANGE(0x10, 0, 0, 0xff, 0xf));
1047 
1048 
1049 static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum)
1050 {
1051 	int osvw_id = *erratum++;
1052 	u32 range;
1053 	u32 ms;
1054 
1055 	if (osvw_id >= 0 && osvw_id < 65536 &&
1056 	    cpu_has(cpu, X86_FEATURE_OSVW)) {
1057 		u64 osvw_len;
1058 
1059 		rdmsrl(MSR_AMD64_OSVW_ID_LENGTH, osvw_len);
1060 		if (osvw_id < osvw_len) {
1061 			u64 osvw_bits;
1062 
1063 			rdmsrl(MSR_AMD64_OSVW_STATUS + (osvw_id >> 6),
1064 			    osvw_bits);
1065 			return osvw_bits & (1ULL << (osvw_id & 0x3f));
1066 		}
1067 	}
1068 
1069 	/* OSVW unavailable or ID unknown, match family-model-stepping range */
1070 	ms = (cpu->x86_model << 4) | cpu->x86_stepping;
1071 	while ((range = *erratum++))
1072 		if ((cpu->x86 == AMD_MODEL_RANGE_FAMILY(range)) &&
1073 		    (ms >= AMD_MODEL_RANGE_START(range)) &&
1074 		    (ms <= AMD_MODEL_RANGE_END(range)))
1075 			return true;
1076 
1077 	return false;
1078 }
1079 
1080 void set_dr_addr_mask(unsigned long mask, int dr)
1081 {
1082 	if (!boot_cpu_has(X86_FEATURE_BPEXT))
1083 		return;
1084 
1085 	switch (dr) {
1086 	case 0:
1087 		wrmsr(MSR_F16H_DR0_ADDR_MASK, mask, 0);
1088 		break;
1089 	case 1:
1090 	case 2:
1091 	case 3:
1092 		wrmsr(MSR_F16H_DR1_ADDR_MASK - 1 + dr, mask, 0);
1093 		break;
1094 	default:
1095 		break;
1096 	}
1097 }
1098