xref: /openbmc/linux/arch/x86/kernel/cpu/amd.c (revision f7d84fa7)
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/cpu.h>
13 #include <asm/smp.h>
14 #include <asm/pci-direct.h>
15 #include <asm/delay.h>
16 
17 #ifdef CONFIG_X86_64
18 # include <asm/mmconfig.h>
19 # include <asm/set_memory.h>
20 #endif
21 
22 #include "cpu.h"
23 
24 static const int amd_erratum_383[];
25 static const int amd_erratum_400[];
26 static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum);
27 
28 /*
29  * nodes_per_socket: Stores the number of nodes per socket.
30  * Refer to Fam15h Models 00-0fh BKDG - CPUID Fn8000_001E_ECX
31  * Node Identifiers[10:8]
32  */
33 static u32 nodes_per_socket = 1;
34 
35 static inline int rdmsrl_amd_safe(unsigned msr, unsigned long long *p)
36 {
37 	u32 gprs[8] = { 0 };
38 	int err;
39 
40 	WARN_ONCE((boot_cpu_data.x86 != 0xf),
41 		  "%s should only be used on K8!\n", __func__);
42 
43 	gprs[1] = msr;
44 	gprs[7] = 0x9c5a203a;
45 
46 	err = rdmsr_safe_regs(gprs);
47 
48 	*p = gprs[0] | ((u64)gprs[2] << 32);
49 
50 	return err;
51 }
52 
53 static inline int wrmsrl_amd_safe(unsigned msr, unsigned long long val)
54 {
55 	u32 gprs[8] = { 0 };
56 
57 	WARN_ONCE((boot_cpu_data.x86 != 0xf),
58 		  "%s should only be used on K8!\n", __func__);
59 
60 	gprs[0] = (u32)val;
61 	gprs[1] = msr;
62 	gprs[2] = val >> 32;
63 	gprs[7] = 0x9c5a203a;
64 
65 	return wrmsr_safe_regs(gprs);
66 }
67 
68 /*
69  *	B step AMD K6 before B 9730xxxx have hardware bugs that can cause
70  *	misexecution of code under Linux. Owners of such processors should
71  *	contact AMD for precise details and a CPU swap.
72  *
73  *	See	http://www.multimania.com/poulot/k6bug.html
74  *	and	section 2.6.2 of "AMD-K6 Processor Revision Guide - Model 6"
75  *		(Publication # 21266  Issue Date: August 1998)
76  *
77  *	The following test is erm.. interesting. AMD neglected to up
78  *	the chip setting when fixing the bug but they also tweaked some
79  *	performance at the same time..
80  */
81 
82 extern __visible void vide(void);
83 __asm__(".globl vide\n"
84 	".type vide, @function\n"
85 	".align 4\n"
86 	"vide: ret\n");
87 
88 static void init_amd_k5(struct cpuinfo_x86 *c)
89 {
90 #ifdef CONFIG_X86_32
91 /*
92  * General Systems BIOSen alias the cpu frequency registers
93  * of the Elan at 0x000df000. Unfortunately, one of the Linux
94  * drivers subsequently pokes it, and changes the CPU speed.
95  * Workaround : Remove the unneeded alias.
96  */
97 #define CBAR		(0xfffc) /* Configuration Base Address  (32-bit) */
98 #define CBAR_ENB	(0x80000000)
99 #define CBAR_KEY	(0X000000CB)
100 	if (c->x86_model == 9 || c->x86_model == 10) {
101 		if (inl(CBAR) & CBAR_ENB)
102 			outl(0 | CBAR_KEY, CBAR);
103 	}
104 #endif
105 }
106 
107 static void init_amd_k6(struct cpuinfo_x86 *c)
108 {
109 #ifdef CONFIG_X86_32
110 	u32 l, h;
111 	int mbytes = get_num_physpages() >> (20-PAGE_SHIFT);
112 
113 	if (c->x86_model < 6) {
114 		/* Based on AMD doc 20734R - June 2000 */
115 		if (c->x86_model == 0) {
116 			clear_cpu_cap(c, X86_FEATURE_APIC);
117 			set_cpu_cap(c, X86_FEATURE_PGE);
118 		}
119 		return;
120 	}
121 
122 	if (c->x86_model == 6 && c->x86_mask == 1) {
123 		const int K6_BUG_LOOP = 1000000;
124 		int n;
125 		void (*f_vide)(void);
126 		u64 d, d2;
127 
128 		pr_info("AMD K6 stepping B detected - ");
129 
130 		/*
131 		 * It looks like AMD fixed the 2.6.2 bug and improved indirect
132 		 * calls at the same time.
133 		 */
134 
135 		n = K6_BUG_LOOP;
136 		f_vide = vide;
137 		d = rdtsc();
138 		while (n--)
139 			f_vide();
140 		d2 = rdtsc();
141 		d = d2-d;
142 
143 		if (d > 20*K6_BUG_LOOP)
144 			pr_cont("system stability may be impaired when more than 32 MB are used.\n");
145 		else
146 			pr_cont("probably OK (after B9730xxxx).\n");
147 	}
148 
149 	/* K6 with old style WHCR */
150 	if (c->x86_model < 8 ||
151 	   (c->x86_model == 8 && c->x86_mask < 8)) {
152 		/* We can only write allocate on the low 508Mb */
153 		if (mbytes > 508)
154 			mbytes = 508;
155 
156 		rdmsr(MSR_K6_WHCR, l, h);
157 		if ((l&0x0000FFFF) == 0) {
158 			unsigned long flags;
159 			l = (1<<0)|((mbytes/4)<<1);
160 			local_irq_save(flags);
161 			wbinvd();
162 			wrmsr(MSR_K6_WHCR, l, h);
163 			local_irq_restore(flags);
164 			pr_info("Enabling old style K6 write allocation for %d Mb\n",
165 				mbytes);
166 		}
167 		return;
168 	}
169 
170 	if ((c->x86_model == 8 && c->x86_mask > 7) ||
171 	     c->x86_model == 9 || c->x86_model == 13) {
172 		/* The more serious chips .. */
173 
174 		if (mbytes > 4092)
175 			mbytes = 4092;
176 
177 		rdmsr(MSR_K6_WHCR, l, h);
178 		if ((l&0xFFFF0000) == 0) {
179 			unsigned long flags;
180 			l = ((mbytes>>2)<<22)|(1<<16);
181 			local_irq_save(flags);
182 			wbinvd();
183 			wrmsr(MSR_K6_WHCR, l, h);
184 			local_irq_restore(flags);
185 			pr_info("Enabling new style K6 write allocation for %d Mb\n",
186 				mbytes);
187 		}
188 
189 		return;
190 	}
191 
192 	if (c->x86_model == 10) {
193 		/* AMD Geode LX is model 10 */
194 		/* placeholder for any needed mods */
195 		return;
196 	}
197 #endif
198 }
199 
200 static void init_amd_k7(struct cpuinfo_x86 *c)
201 {
202 #ifdef CONFIG_X86_32
203 	u32 l, h;
204 
205 	/*
206 	 * Bit 15 of Athlon specific MSR 15, needs to be 0
207 	 * to enable SSE on Palomino/Morgan/Barton CPU's.
208 	 * If the BIOS didn't enable it already, enable it here.
209 	 */
210 	if (c->x86_model >= 6 && c->x86_model <= 10) {
211 		if (!cpu_has(c, X86_FEATURE_XMM)) {
212 			pr_info("Enabling disabled K7/SSE Support.\n");
213 			msr_clear_bit(MSR_K7_HWCR, 15);
214 			set_cpu_cap(c, X86_FEATURE_XMM);
215 		}
216 	}
217 
218 	/*
219 	 * It's been determined by AMD that Athlons since model 8 stepping 1
220 	 * are more robust with CLK_CTL set to 200xxxxx instead of 600xxxxx
221 	 * As per AMD technical note 27212 0.2
222 	 */
223 	if ((c->x86_model == 8 && c->x86_mask >= 1) || (c->x86_model > 8)) {
224 		rdmsr(MSR_K7_CLK_CTL, l, h);
225 		if ((l & 0xfff00000) != 0x20000000) {
226 			pr_info("CPU: CLK_CTL MSR was %x. Reprogramming to %x\n",
227 				l, ((l & 0x000fffff)|0x20000000));
228 			wrmsr(MSR_K7_CLK_CTL, (l & 0x000fffff)|0x20000000, h);
229 		}
230 	}
231 
232 	set_cpu_cap(c, X86_FEATURE_K7);
233 
234 	/* calling is from identify_secondary_cpu() ? */
235 	if (!c->cpu_index)
236 		return;
237 
238 	/*
239 	 * Certain Athlons might work (for various values of 'work') in SMP
240 	 * but they are not certified as MP capable.
241 	 */
242 	/* Athlon 660/661 is valid. */
243 	if ((c->x86_model == 6) && ((c->x86_mask == 0) ||
244 	    (c->x86_mask == 1)))
245 		return;
246 
247 	/* Duron 670 is valid */
248 	if ((c->x86_model == 7) && (c->x86_mask == 0))
249 		return;
250 
251 	/*
252 	 * Athlon 662, Duron 671, and Athlon >model 7 have capability
253 	 * bit. It's worth noting that the A5 stepping (662) of some
254 	 * Athlon XP's have the MP bit set.
255 	 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for
256 	 * more.
257 	 */
258 	if (((c->x86_model == 6) && (c->x86_mask >= 2)) ||
259 	    ((c->x86_model == 7) && (c->x86_mask >= 1)) ||
260 	     (c->x86_model > 7))
261 		if (cpu_has(c, X86_FEATURE_MP))
262 			return;
263 
264 	/* If we get here, not a certified SMP capable AMD system. */
265 
266 	/*
267 	 * Don't taint if we are running SMP kernel on a single non-MP
268 	 * approved Athlon
269 	 */
270 	WARN_ONCE(1, "WARNING: This combination of AMD"
271 		" processors is not suitable for SMP.\n");
272 	add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_NOW_UNRELIABLE);
273 #endif
274 }
275 
276 #ifdef CONFIG_NUMA
277 /*
278  * To workaround broken NUMA config.  Read the comment in
279  * srat_detect_node().
280  */
281 static int nearby_node(int apicid)
282 {
283 	int i, node;
284 
285 	for (i = apicid - 1; i >= 0; i--) {
286 		node = __apicid_to_node[i];
287 		if (node != NUMA_NO_NODE && node_online(node))
288 			return node;
289 	}
290 	for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) {
291 		node = __apicid_to_node[i];
292 		if (node != NUMA_NO_NODE && node_online(node))
293 			return node;
294 	}
295 	return first_node(node_online_map); /* Shouldn't happen */
296 }
297 #endif
298 
299 /*
300  * Fixup core topology information for
301  * (1) AMD multi-node processors
302  *     Assumption: Number of cores in each internal node is the same.
303  * (2) AMD processors supporting compute units
304  */
305 #ifdef CONFIG_SMP
306 static void amd_get_topology(struct cpuinfo_x86 *c)
307 {
308 	u8 node_id;
309 	int cpu = smp_processor_id();
310 
311 	/* get information required for multi-node processors */
312 	if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
313 		u32 eax, ebx, ecx, edx;
314 
315 		cpuid(0x8000001e, &eax, &ebx, &ecx, &edx);
316 
317 		node_id  = ecx & 0xff;
318 		smp_num_siblings = ((ebx >> 8) & 0xff) + 1;
319 
320 		if (c->x86 == 0x15)
321 			c->cu_id = ebx & 0xff;
322 
323 		if (c->x86 >= 0x17) {
324 			c->cpu_core_id = ebx & 0xff;
325 
326 			if (smp_num_siblings > 1)
327 				c->x86_max_cores /= smp_num_siblings;
328 		}
329 
330 		/*
331 		 * We may have multiple LLCs if L3 caches exist, so check if we
332 		 * have an L3 cache by looking at the L3 cache CPUID leaf.
333 		 */
334 		if (cpuid_edx(0x80000006)) {
335 			if (c->x86 == 0x17) {
336 				/*
337 				 * LLC is at the core complex level.
338 				 * Core complex id is ApicId[3].
339 				 */
340 				per_cpu(cpu_llc_id, cpu) = c->apicid >> 3;
341 			} else {
342 				/* LLC is at the node level. */
343 				per_cpu(cpu_llc_id, cpu) = node_id;
344 			}
345 		}
346 	} else if (cpu_has(c, X86_FEATURE_NODEID_MSR)) {
347 		u64 value;
348 
349 		rdmsrl(MSR_FAM10H_NODE_ID, value);
350 		node_id = value & 7;
351 
352 		per_cpu(cpu_llc_id, cpu) = node_id;
353 	} else
354 		return;
355 
356 	/* fixup multi-node processor information */
357 	if (nodes_per_socket > 1) {
358 		u32 cus_per_node;
359 
360 		set_cpu_cap(c, X86_FEATURE_AMD_DCM);
361 		cus_per_node = c->x86_max_cores / nodes_per_socket;
362 
363 		/* core id has to be in the [0 .. cores_per_node - 1] range */
364 		c->cpu_core_id %= cus_per_node;
365 	}
366 }
367 #endif
368 
369 /*
370  * On a AMD dual core setup the lower bits of the APIC id distinguish the cores.
371  * Assumes number of cores is a power of two.
372  */
373 static void amd_detect_cmp(struct cpuinfo_x86 *c)
374 {
375 #ifdef CONFIG_SMP
376 	unsigned bits;
377 	int cpu = smp_processor_id();
378 
379 	bits = c->x86_coreid_bits;
380 	/* Low order bits define the core id (index of core in socket) */
381 	c->cpu_core_id = c->initial_apicid & ((1 << bits)-1);
382 	/* Convert the initial APIC ID into the socket ID */
383 	c->phys_proc_id = c->initial_apicid >> bits;
384 	/* use socket ID also for last level cache */
385 	per_cpu(cpu_llc_id, cpu) = c->phys_proc_id;
386 	amd_get_topology(c);
387 #endif
388 }
389 
390 u16 amd_get_nb_id(int cpu)
391 {
392 	u16 id = 0;
393 #ifdef CONFIG_SMP
394 	id = per_cpu(cpu_llc_id, cpu);
395 #endif
396 	return id;
397 }
398 EXPORT_SYMBOL_GPL(amd_get_nb_id);
399 
400 u32 amd_get_nodes_per_socket(void)
401 {
402 	return nodes_per_socket;
403 }
404 EXPORT_SYMBOL_GPL(amd_get_nodes_per_socket);
405 
406 static void srat_detect_node(struct cpuinfo_x86 *c)
407 {
408 #ifdef CONFIG_NUMA
409 	int cpu = smp_processor_id();
410 	int node;
411 	unsigned apicid = c->apicid;
412 
413 	node = numa_cpu_node(cpu);
414 	if (node == NUMA_NO_NODE)
415 		node = per_cpu(cpu_llc_id, cpu);
416 
417 	/*
418 	 * On multi-fabric platform (e.g. Numascale NumaChip) a
419 	 * platform-specific handler needs to be called to fixup some
420 	 * IDs of the CPU.
421 	 */
422 	if (x86_cpuinit.fixup_cpu_id)
423 		x86_cpuinit.fixup_cpu_id(c, node);
424 
425 	if (!node_online(node)) {
426 		/*
427 		 * Two possibilities here:
428 		 *
429 		 * - The CPU is missing memory and no node was created.  In
430 		 *   that case try picking one from a nearby CPU.
431 		 *
432 		 * - The APIC IDs differ from the HyperTransport node IDs
433 		 *   which the K8 northbridge parsing fills in.  Assume
434 		 *   they are all increased by a constant offset, but in
435 		 *   the same order as the HT nodeids.  If that doesn't
436 		 *   result in a usable node fall back to the path for the
437 		 *   previous case.
438 		 *
439 		 * This workaround operates directly on the mapping between
440 		 * APIC ID and NUMA node, assuming certain relationship
441 		 * between APIC ID, HT node ID and NUMA topology.  As going
442 		 * through CPU mapping may alter the outcome, directly
443 		 * access __apicid_to_node[].
444 		 */
445 		int ht_nodeid = c->initial_apicid;
446 
447 		if (__apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
448 			node = __apicid_to_node[ht_nodeid];
449 		/* Pick a nearby node */
450 		if (!node_online(node))
451 			node = nearby_node(apicid);
452 	}
453 	numa_set_node(cpu, node);
454 #endif
455 }
456 
457 static void early_init_amd_mc(struct cpuinfo_x86 *c)
458 {
459 #ifdef CONFIG_SMP
460 	unsigned bits, ecx;
461 
462 	/* Multi core CPU? */
463 	if (c->extended_cpuid_level < 0x80000008)
464 		return;
465 
466 	ecx = cpuid_ecx(0x80000008);
467 
468 	c->x86_max_cores = (ecx & 0xff) + 1;
469 
470 	/* CPU telling us the core id bits shift? */
471 	bits = (ecx >> 12) & 0xF;
472 
473 	/* Otherwise recompute */
474 	if (bits == 0) {
475 		while ((1 << bits) < c->x86_max_cores)
476 			bits++;
477 	}
478 
479 	c->x86_coreid_bits = bits;
480 #endif
481 }
482 
483 static void bsp_init_amd(struct cpuinfo_x86 *c)
484 {
485 
486 #ifdef CONFIG_X86_64
487 	if (c->x86 >= 0xf) {
488 		unsigned long long tseg;
489 
490 		/*
491 		 * Split up direct mapping around the TSEG SMM area.
492 		 * Don't do it for gbpages because there seems very little
493 		 * benefit in doing so.
494 		 */
495 		if (!rdmsrl_safe(MSR_K8_TSEG_ADDR, &tseg)) {
496 			unsigned long pfn = tseg >> PAGE_SHIFT;
497 
498 			pr_debug("tseg: %010llx\n", tseg);
499 			if (pfn_range_is_mapped(pfn, pfn + 1))
500 				set_memory_4k((unsigned long)__va(tseg), 1);
501 		}
502 	}
503 #endif
504 
505 	if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) {
506 
507 		if (c->x86 > 0x10 ||
508 		    (c->x86 == 0x10 && c->x86_model >= 0x2)) {
509 			u64 val;
510 
511 			rdmsrl(MSR_K7_HWCR, val);
512 			if (!(val & BIT(24)))
513 				pr_warn(FW_BUG "TSC doesn't count with P0 frequency!\n");
514 		}
515 	}
516 
517 	if (c->x86 == 0x15) {
518 		unsigned long upperbit;
519 		u32 cpuid, assoc;
520 
521 		cpuid	 = cpuid_edx(0x80000005);
522 		assoc	 = cpuid >> 16 & 0xff;
523 		upperbit = ((cpuid >> 24) << 10) / assoc;
524 
525 		va_align.mask	  = (upperbit - 1) & PAGE_MASK;
526 		va_align.flags    = ALIGN_VA_32 | ALIGN_VA_64;
527 
528 		/* A random value per boot for bit slice [12:upper_bit) */
529 		va_align.bits = get_random_int() & va_align.mask;
530 	}
531 
532 	if (cpu_has(c, X86_FEATURE_MWAITX))
533 		use_mwaitx_delay();
534 
535 	if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
536 		u32 ecx;
537 
538 		ecx = cpuid_ecx(0x8000001e);
539 		nodes_per_socket = ((ecx >> 8) & 7) + 1;
540 	} else if (boot_cpu_has(X86_FEATURE_NODEID_MSR)) {
541 		u64 value;
542 
543 		rdmsrl(MSR_FAM10H_NODE_ID, value);
544 		nodes_per_socket = ((value >> 3) & 7) + 1;
545 	}
546 }
547 
548 static void early_init_amd(struct cpuinfo_x86 *c)
549 {
550 	early_init_amd_mc(c);
551 
552 	/*
553 	 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
554 	 * with P/T states and does not stop in deep C-states
555 	 */
556 	if (c->x86_power & (1 << 8)) {
557 		set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
558 		set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
559 	}
560 
561 	/* Bit 12 of 8000_0007 edx is accumulated power mechanism. */
562 	if (c->x86_power & BIT(12))
563 		set_cpu_cap(c, X86_FEATURE_ACC_POWER);
564 
565 #ifdef CONFIG_X86_64
566 	set_cpu_cap(c, X86_FEATURE_SYSCALL32);
567 #else
568 	/*  Set MTRR capability flag if appropriate */
569 	if (c->x86 == 5)
570 		if (c->x86_model == 13 || c->x86_model == 9 ||
571 		    (c->x86_model == 8 && c->x86_mask >= 8))
572 			set_cpu_cap(c, X86_FEATURE_K6_MTRR);
573 #endif
574 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
575 	/*
576 	 * ApicID can always be treated as an 8-bit value for AMD APIC versions
577 	 * >= 0x10, but even old K8s came out of reset with version 0x10. So, we
578 	 * can safely set X86_FEATURE_EXTD_APICID unconditionally for families
579 	 * after 16h.
580 	 */
581 	if (boot_cpu_has(X86_FEATURE_APIC)) {
582 		if (c->x86 > 0x16)
583 			set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
584 		else if (c->x86 >= 0xf) {
585 			/* check CPU config space for extended APIC ID */
586 			unsigned int val;
587 
588 			val = read_pci_config(0, 24, 0, 0x68);
589 			if ((val >> 17 & 0x3) == 0x3)
590 				set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
591 		}
592 	}
593 #endif
594 
595 	/*
596 	 * This is only needed to tell the kernel whether to use VMCALL
597 	 * and VMMCALL.  VMMCALL is never executed except under virt, so
598 	 * we can set it unconditionally.
599 	 */
600 	set_cpu_cap(c, X86_FEATURE_VMMCALL);
601 
602 	/* F16h erratum 793, CVE-2013-6885 */
603 	if (c->x86 == 0x16 && c->x86_model <= 0xf)
604 		msr_set_bit(MSR_AMD64_LS_CFG, 15);
605 
606 	/*
607 	 * Check whether the machine is affected by erratum 400. This is
608 	 * used to select the proper idle routine and to enable the check
609 	 * whether the machine is affected in arch_post_acpi_init(), which
610 	 * sets the X86_BUG_AMD_APIC_C1E bug depending on the MSR check.
611 	 */
612 	if (cpu_has_amd_erratum(c, amd_erratum_400))
613 		set_cpu_bug(c, X86_BUG_AMD_E400);
614 }
615 
616 static void init_amd_k8(struct cpuinfo_x86 *c)
617 {
618 	u32 level;
619 	u64 value;
620 
621 	/* On C+ stepping K8 rep microcode works well for copy/memset */
622 	level = cpuid_eax(1);
623 	if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58)
624 		set_cpu_cap(c, X86_FEATURE_REP_GOOD);
625 
626 	/*
627 	 * Some BIOSes incorrectly force this feature, but only K8 revision D
628 	 * (model = 0x14) and later actually support it.
629 	 * (AMD Erratum #110, docId: 25759).
630 	 */
631 	if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM)) {
632 		clear_cpu_cap(c, X86_FEATURE_LAHF_LM);
633 		if (!rdmsrl_amd_safe(0xc001100d, &value)) {
634 			value &= ~BIT_64(32);
635 			wrmsrl_amd_safe(0xc001100d, value);
636 		}
637 	}
638 
639 	if (!c->x86_model_id[0])
640 		strcpy(c->x86_model_id, "Hammer");
641 
642 #ifdef CONFIG_SMP
643 	/*
644 	 * Disable TLB flush filter by setting HWCR.FFDIS on K8
645 	 * bit 6 of msr C001_0015
646 	 *
647 	 * Errata 63 for SH-B3 steppings
648 	 * Errata 122 for all steppings (F+ have it disabled by default)
649 	 */
650 	msr_set_bit(MSR_K7_HWCR, 6);
651 #endif
652 	set_cpu_bug(c, X86_BUG_SWAPGS_FENCE);
653 }
654 
655 static void init_amd_gh(struct cpuinfo_x86 *c)
656 {
657 #ifdef CONFIG_X86_64
658 	/* do this for boot cpu */
659 	if (c == &boot_cpu_data)
660 		check_enable_amd_mmconf_dmi();
661 
662 	fam10h_check_enable_mmcfg();
663 #endif
664 
665 	/*
666 	 * Disable GART TLB Walk Errors on Fam10h. We do this here because this
667 	 * is always needed when GART is enabled, even in a kernel which has no
668 	 * MCE support built in. BIOS should disable GartTlbWlk Errors already.
669 	 * If it doesn't, we do it here as suggested by the BKDG.
670 	 *
671 	 * Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=33012
672 	 */
673 	msr_set_bit(MSR_AMD64_MCx_MASK(4), 10);
674 
675 	/*
676 	 * On family 10h BIOS may not have properly enabled WC+ support, causing
677 	 * it to be converted to CD memtype. This may result in performance
678 	 * degradation for certain nested-paging guests. Prevent this conversion
679 	 * by clearing bit 24 in MSR_AMD64_BU_CFG2.
680 	 *
681 	 * NOTE: we want to use the _safe accessors so as not to #GP kvm
682 	 * guests on older kvm hosts.
683 	 */
684 	msr_clear_bit(MSR_AMD64_BU_CFG2, 24);
685 
686 	if (cpu_has_amd_erratum(c, amd_erratum_383))
687 		set_cpu_bug(c, X86_BUG_AMD_TLB_MMATCH);
688 }
689 
690 #define MSR_AMD64_DE_CFG	0xC0011029
691 
692 static void init_amd_ln(struct cpuinfo_x86 *c)
693 {
694 	/*
695 	 * Apply erratum 665 fix unconditionally so machines without a BIOS
696 	 * fix work.
697 	 */
698 	msr_set_bit(MSR_AMD64_DE_CFG, 31);
699 }
700 
701 static void init_amd_bd(struct cpuinfo_x86 *c)
702 {
703 	u64 value;
704 
705 	/* re-enable TopologyExtensions if switched off by BIOS */
706 	if ((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 	/*
719 	 * The way access filter has a performance penalty on some workloads.
720 	 * Disable it on the affected CPUs.
721 	 */
722 	if ((c->x86_model >= 0x02) && (c->x86_model < 0x20)) {
723 		if (!rdmsrl_safe(MSR_F15H_IC_CFG, &value) && !(value & 0x1E)) {
724 			value |= 0x1E;
725 			wrmsrl_safe(MSR_F15H_IC_CFG, value);
726 		}
727 	}
728 }
729 
730 static void init_amd(struct cpuinfo_x86 *c)
731 {
732 	u32 dummy;
733 
734 	early_init_amd(c);
735 
736 	/*
737 	 * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
738 	 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
739 	 */
740 	clear_cpu_cap(c, 0*32+31);
741 
742 	if (c->x86 >= 0x10)
743 		set_cpu_cap(c, X86_FEATURE_REP_GOOD);
744 
745 	/* get apicid instead of initial apic id from cpuid */
746 	c->apicid = hard_smp_processor_id();
747 
748 	/* K6s reports MCEs but don't actually have all the MSRs */
749 	if (c->x86 < 6)
750 		clear_cpu_cap(c, X86_FEATURE_MCE);
751 
752 	switch (c->x86) {
753 	case 4:    init_amd_k5(c); break;
754 	case 5:    init_amd_k6(c); break;
755 	case 6:	   init_amd_k7(c); break;
756 	case 0xf:  init_amd_k8(c); break;
757 	case 0x10: init_amd_gh(c); break;
758 	case 0x12: init_amd_ln(c); break;
759 	case 0x15: init_amd_bd(c); break;
760 	}
761 
762 	/* Enable workaround for FXSAVE leak */
763 	if (c->x86 >= 6)
764 		set_cpu_bug(c, X86_BUG_FXSAVE_LEAK);
765 
766 	cpu_detect_cache_sizes(c);
767 
768 	/* Multi core CPU? */
769 	if (c->extended_cpuid_level >= 0x80000008) {
770 		amd_detect_cmp(c);
771 		srat_detect_node(c);
772 	}
773 
774 #ifdef CONFIG_X86_32
775 	detect_ht(c);
776 #endif
777 
778 	init_amd_cacheinfo(c);
779 
780 	if (c->x86 >= 0xf)
781 		set_cpu_cap(c, X86_FEATURE_K8);
782 
783 	if (cpu_has(c, X86_FEATURE_XMM2)) {
784 		/* MFENCE stops RDTSC speculation */
785 		set_cpu_cap(c, X86_FEATURE_MFENCE_RDTSC);
786 	}
787 
788 	/*
789 	 * Family 0x12 and above processors have APIC timer
790 	 * running in deep C states.
791 	 */
792 	if (c->x86 > 0x11)
793 		set_cpu_cap(c, X86_FEATURE_ARAT);
794 
795 	rdmsr_safe(MSR_AMD64_PATCH_LEVEL, &c->microcode, &dummy);
796 
797 	/* 3DNow or LM implies PREFETCHW */
798 	if (!cpu_has(c, X86_FEATURE_3DNOWPREFETCH))
799 		if (cpu_has(c, X86_FEATURE_3DNOW) || cpu_has(c, X86_FEATURE_LM))
800 			set_cpu_cap(c, X86_FEATURE_3DNOWPREFETCH);
801 
802 	/* AMD CPUs don't reset SS attributes on SYSRET, Xen does. */
803 	if (!cpu_has(c, X86_FEATURE_XENPV))
804 		set_cpu_bug(c, X86_BUG_SYSRET_SS_ATTRS);
805 }
806 
807 #ifdef CONFIG_X86_32
808 static unsigned int amd_size_cache(struct cpuinfo_x86 *c, unsigned int size)
809 {
810 	/* AMD errata T13 (order #21922) */
811 	if ((c->x86 == 6)) {
812 		/* Duron Rev A0 */
813 		if (c->x86_model == 3 && c->x86_mask == 0)
814 			size = 64;
815 		/* Tbird rev A1/A2 */
816 		if (c->x86_model == 4 &&
817 			(c->x86_mask == 0 || c->x86_mask == 1))
818 			size = 256;
819 	}
820 	return size;
821 }
822 #endif
823 
824 static void cpu_detect_tlb_amd(struct cpuinfo_x86 *c)
825 {
826 	u32 ebx, eax, ecx, edx;
827 	u16 mask = 0xfff;
828 
829 	if (c->x86 < 0xf)
830 		return;
831 
832 	if (c->extended_cpuid_level < 0x80000006)
833 		return;
834 
835 	cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
836 
837 	tlb_lld_4k[ENTRIES] = (ebx >> 16) & mask;
838 	tlb_lli_4k[ENTRIES] = ebx & mask;
839 
840 	/*
841 	 * K8 doesn't have 2M/4M entries in the L2 TLB so read out the L1 TLB
842 	 * characteristics from the CPUID function 0x80000005 instead.
843 	 */
844 	if (c->x86 == 0xf) {
845 		cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
846 		mask = 0xff;
847 	}
848 
849 	/* Handle DTLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
850 	if (!((eax >> 16) & mask))
851 		tlb_lld_2m[ENTRIES] = (cpuid_eax(0x80000005) >> 16) & 0xff;
852 	else
853 		tlb_lld_2m[ENTRIES] = (eax >> 16) & mask;
854 
855 	/* a 4M entry uses two 2M entries */
856 	tlb_lld_4m[ENTRIES] = tlb_lld_2m[ENTRIES] >> 1;
857 
858 	/* Handle ITLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
859 	if (!(eax & mask)) {
860 		/* Erratum 658 */
861 		if (c->x86 == 0x15 && c->x86_model <= 0x1f) {
862 			tlb_lli_2m[ENTRIES] = 1024;
863 		} else {
864 			cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
865 			tlb_lli_2m[ENTRIES] = eax & 0xff;
866 		}
867 	} else
868 		tlb_lli_2m[ENTRIES] = eax & mask;
869 
870 	tlb_lli_4m[ENTRIES] = tlb_lli_2m[ENTRIES] >> 1;
871 }
872 
873 static const struct cpu_dev amd_cpu_dev = {
874 	.c_vendor	= "AMD",
875 	.c_ident	= { "AuthenticAMD" },
876 #ifdef CONFIG_X86_32
877 	.legacy_models = {
878 		{ .family = 4, .model_names =
879 		  {
880 			  [3] = "486 DX/2",
881 			  [7] = "486 DX/2-WB",
882 			  [8] = "486 DX/4",
883 			  [9] = "486 DX/4-WB",
884 			  [14] = "Am5x86-WT",
885 			  [15] = "Am5x86-WB"
886 		  }
887 		},
888 	},
889 	.legacy_cache_size = amd_size_cache,
890 #endif
891 	.c_early_init   = early_init_amd,
892 	.c_detect_tlb	= cpu_detect_tlb_amd,
893 	.c_bsp_init	= bsp_init_amd,
894 	.c_init		= init_amd,
895 	.c_x86_vendor	= X86_VENDOR_AMD,
896 };
897 
898 cpu_dev_register(amd_cpu_dev);
899 
900 /*
901  * AMD errata checking
902  *
903  * Errata are defined as arrays of ints using the AMD_LEGACY_ERRATUM() or
904  * AMD_OSVW_ERRATUM() macros. The latter is intended for newer errata that
905  * have an OSVW id assigned, which it takes as first argument. Both take a
906  * variable number of family-specific model-stepping ranges created by
907  * AMD_MODEL_RANGE().
908  *
909  * Example:
910  *
911  * const int amd_erratum_319[] =
912  *	AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0x4, 0x2),
913  *			   AMD_MODEL_RANGE(0x10, 0x8, 0x0, 0x8, 0x0),
914  *			   AMD_MODEL_RANGE(0x10, 0x9, 0x0, 0x9, 0x0));
915  */
916 
917 #define AMD_LEGACY_ERRATUM(...)		{ -1, __VA_ARGS__, 0 }
918 #define AMD_OSVW_ERRATUM(osvw_id, ...)	{ osvw_id, __VA_ARGS__, 0 }
919 #define AMD_MODEL_RANGE(f, m_start, s_start, m_end, s_end) \
920 	((f << 24) | (m_start << 16) | (s_start << 12) | (m_end << 4) | (s_end))
921 #define AMD_MODEL_RANGE_FAMILY(range)	(((range) >> 24) & 0xff)
922 #define AMD_MODEL_RANGE_START(range)	(((range) >> 12) & 0xfff)
923 #define AMD_MODEL_RANGE_END(range)	((range) & 0xfff)
924 
925 static const int amd_erratum_400[] =
926 	AMD_OSVW_ERRATUM(1, AMD_MODEL_RANGE(0xf, 0x41, 0x2, 0xff, 0xf),
927 			    AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0xff, 0xf));
928 
929 static const int amd_erratum_383[] =
930 	AMD_OSVW_ERRATUM(3, AMD_MODEL_RANGE(0x10, 0, 0, 0xff, 0xf));
931 
932 
933 static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum)
934 {
935 	int osvw_id = *erratum++;
936 	u32 range;
937 	u32 ms;
938 
939 	if (osvw_id >= 0 && osvw_id < 65536 &&
940 	    cpu_has(cpu, X86_FEATURE_OSVW)) {
941 		u64 osvw_len;
942 
943 		rdmsrl(MSR_AMD64_OSVW_ID_LENGTH, osvw_len);
944 		if (osvw_id < osvw_len) {
945 			u64 osvw_bits;
946 
947 			rdmsrl(MSR_AMD64_OSVW_STATUS + (osvw_id >> 6),
948 			    osvw_bits);
949 			return osvw_bits & (1ULL << (osvw_id & 0x3f));
950 		}
951 	}
952 
953 	/* OSVW unavailable or ID unknown, match family-model-stepping range */
954 	ms = (cpu->x86_model << 4) | cpu->x86_mask;
955 	while ((range = *erratum++))
956 		if ((cpu->x86 == AMD_MODEL_RANGE_FAMILY(range)) &&
957 		    (ms >= AMD_MODEL_RANGE_START(range)) &&
958 		    (ms <= AMD_MODEL_RANGE_END(range)))
959 			return true;
960 
961 	return false;
962 }
963 
964 void set_dr_addr_mask(unsigned long mask, int dr)
965 {
966 	if (!boot_cpu_has(X86_FEATURE_BPEXT))
967 		return;
968 
969 	switch (dr) {
970 	case 0:
971 		wrmsr(MSR_F16H_DR0_ADDR_MASK, mask, 0);
972 		break;
973 	case 1:
974 	case 2:
975 	case 3:
976 		wrmsr(MSR_F16H_DR1_ADDR_MASK - 1 + dr, mask, 0);
977 		break;
978 	default:
979 		break;
980 	}
981 }
982