xref: /linux/arch/x86/kernel/cpu/amd.c (revision bcb63314e2c23f1ed622418b65f9409512659c73)
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/random.h>
9 #include <asm/processor.h>
10 #include <asm/apic.h>
11 #include <asm/cpu.h>
12 #include <asm/smp.h>
13 #include <asm/pci-direct.h>
14 #include <asm/delay.h>
15 
16 #ifdef CONFIG_X86_64
17 # include <asm/mmconfig.h>
18 # include <asm/cacheflush.h>
19 #endif
20 
21 #include "cpu.h"
22 
23 static const int amd_erratum_383[];
24 static const int amd_erratum_400[];
25 static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum);
26 
27 /*
28  * nodes_per_socket: Stores the number of nodes per socket.
29  * Refer to Fam15h Models 00-0fh BKDG - CPUID Fn8000_001E_ECX
30  * Node Identifiers[10:8]
31  */
32 static u32 nodes_per_socket = 1;
33 
34 static inline int rdmsrl_amd_safe(unsigned msr, unsigned long long *p)
35 {
36 	u32 gprs[8] = { 0 };
37 	int err;
38 
39 	WARN_ONCE((boot_cpu_data.x86 != 0xf),
40 		  "%s should only be used on K8!\n", __func__);
41 
42 	gprs[1] = msr;
43 	gprs[7] = 0x9c5a203a;
44 
45 	err = rdmsr_safe_regs(gprs);
46 
47 	*p = gprs[0] | ((u64)gprs[2] << 32);
48 
49 	return err;
50 }
51 
52 static inline int wrmsrl_amd_safe(unsigned msr, unsigned long long val)
53 {
54 	u32 gprs[8] = { 0 };
55 
56 	WARN_ONCE((boot_cpu_data.x86 != 0xf),
57 		  "%s should only be used on K8!\n", __func__);
58 
59 	gprs[0] = (u32)val;
60 	gprs[1] = msr;
61 	gprs[2] = val >> 32;
62 	gprs[7] = 0x9c5a203a;
63 
64 	return wrmsr_safe_regs(gprs);
65 }
66 
67 /*
68  *	B step AMD K6 before B 9730xxxx have hardware bugs that can cause
69  *	misexecution of code under Linux. Owners of such processors should
70  *	contact AMD for precise details and a CPU swap.
71  *
72  *	See	http://www.multimania.com/poulot/k6bug.html
73  *	and	section 2.6.2 of "AMD-K6 Processor Revision Guide - Model 6"
74  *		(Publication # 21266  Issue Date: August 1998)
75  *
76  *	The following test is erm.. interesting. AMD neglected to up
77  *	the chip setting when fixing the bug but they also tweaked some
78  *	performance at the same time..
79  */
80 
81 extern __visible void vide(void);
82 __asm__(".globl vide\n"
83 	".type vide, @function\n"
84 	".align 4\n"
85 	"vide: ret\n");
86 
87 static void init_amd_k5(struct cpuinfo_x86 *c)
88 {
89 #ifdef CONFIG_X86_32
90 /*
91  * General Systems BIOSen alias the cpu frequency registers
92  * of the Elan at 0x000df000. Unfortunately, one of the Linux
93  * drivers subsequently pokes it, and changes the CPU speed.
94  * Workaround : Remove the unneeded alias.
95  */
96 #define CBAR		(0xfffc) /* Configuration Base Address  (32-bit) */
97 #define CBAR_ENB	(0x80000000)
98 #define CBAR_KEY	(0X000000CB)
99 	if (c->x86_model == 9 || c->x86_model == 10) {
100 		if (inl(CBAR) & CBAR_ENB)
101 			outl(0 | CBAR_KEY, CBAR);
102 	}
103 #endif
104 }
105 
106 static void init_amd_k6(struct cpuinfo_x86 *c)
107 {
108 #ifdef CONFIG_X86_32
109 	u32 l, h;
110 	int mbytes = get_num_physpages() >> (20-PAGE_SHIFT);
111 
112 	if (c->x86_model < 6) {
113 		/* Based on AMD doc 20734R - June 2000 */
114 		if (c->x86_model == 0) {
115 			clear_cpu_cap(c, X86_FEATURE_APIC);
116 			set_cpu_cap(c, X86_FEATURE_PGE);
117 		}
118 		return;
119 	}
120 
121 	if (c->x86_model == 6 && c->x86_mask == 1) {
122 		const int K6_BUG_LOOP = 1000000;
123 		int n;
124 		void (*f_vide)(void);
125 		u64 d, d2;
126 
127 		pr_info("AMD K6 stepping B detected - ");
128 
129 		/*
130 		 * It looks like AMD fixed the 2.6.2 bug and improved indirect
131 		 * calls at the same time.
132 		 */
133 
134 		n = K6_BUG_LOOP;
135 		f_vide = vide;
136 		d = rdtsc();
137 		while (n--)
138 			f_vide();
139 		d2 = rdtsc();
140 		d = d2-d;
141 
142 		if (d > 20*K6_BUG_LOOP)
143 			pr_cont("system stability may be impaired when more than 32 MB are used.\n");
144 		else
145 			pr_cont("probably OK (after B9730xxxx).\n");
146 	}
147 
148 	/* K6 with old style WHCR */
149 	if (c->x86_model < 8 ||
150 	   (c->x86_model == 8 && c->x86_mask < 8)) {
151 		/* We can only write allocate on the low 508Mb */
152 		if (mbytes > 508)
153 			mbytes = 508;
154 
155 		rdmsr(MSR_K6_WHCR, l, h);
156 		if ((l&0x0000FFFF) == 0) {
157 			unsigned long flags;
158 			l = (1<<0)|((mbytes/4)<<1);
159 			local_irq_save(flags);
160 			wbinvd();
161 			wrmsr(MSR_K6_WHCR, l, h);
162 			local_irq_restore(flags);
163 			pr_info("Enabling old style K6 write allocation for %d Mb\n",
164 				mbytes);
165 		}
166 		return;
167 	}
168 
169 	if ((c->x86_model == 8 && c->x86_mask > 7) ||
170 	     c->x86_model == 9 || c->x86_model == 13) {
171 		/* The more serious chips .. */
172 
173 		if (mbytes > 4092)
174 			mbytes = 4092;
175 
176 		rdmsr(MSR_K6_WHCR, l, h);
177 		if ((l&0xFFFF0000) == 0) {
178 			unsigned long flags;
179 			l = ((mbytes>>2)<<22)|(1<<16);
180 			local_irq_save(flags);
181 			wbinvd();
182 			wrmsr(MSR_K6_WHCR, l, h);
183 			local_irq_restore(flags);
184 			pr_info("Enabling new style K6 write allocation for %d Mb\n",
185 				mbytes);
186 		}
187 
188 		return;
189 	}
190 
191 	if (c->x86_model == 10) {
192 		/* AMD Geode LX is model 10 */
193 		/* placeholder for any needed mods */
194 		return;
195 	}
196 #endif
197 }
198 
199 static void init_amd_k7(struct cpuinfo_x86 *c)
200 {
201 #ifdef CONFIG_X86_32
202 	u32 l, h;
203 
204 	/*
205 	 * Bit 15 of Athlon specific MSR 15, needs to be 0
206 	 * to enable SSE on Palomino/Morgan/Barton CPU's.
207 	 * If the BIOS didn't enable it already, enable it here.
208 	 */
209 	if (c->x86_model >= 6 && c->x86_model <= 10) {
210 		if (!cpu_has(c, X86_FEATURE_XMM)) {
211 			pr_info("Enabling disabled K7/SSE Support.\n");
212 			msr_clear_bit(MSR_K7_HWCR, 15);
213 			set_cpu_cap(c, X86_FEATURE_XMM);
214 		}
215 	}
216 
217 	/*
218 	 * It's been determined by AMD that Athlons since model 8 stepping 1
219 	 * are more robust with CLK_CTL set to 200xxxxx instead of 600xxxxx
220 	 * As per AMD technical note 27212 0.2
221 	 */
222 	if ((c->x86_model == 8 && c->x86_mask >= 1) || (c->x86_model > 8)) {
223 		rdmsr(MSR_K7_CLK_CTL, l, h);
224 		if ((l & 0xfff00000) != 0x20000000) {
225 			pr_info("CPU: CLK_CTL MSR was %x. Reprogramming to %x\n",
226 				l, ((l & 0x000fffff)|0x20000000));
227 			wrmsr(MSR_K7_CLK_CTL, (l & 0x000fffff)|0x20000000, h);
228 		}
229 	}
230 
231 	set_cpu_cap(c, X86_FEATURE_K7);
232 
233 	/* calling is from identify_secondary_cpu() ? */
234 	if (!c->cpu_index)
235 		return;
236 
237 	/*
238 	 * Certain Athlons might work (for various values of 'work') in SMP
239 	 * but they are not certified as MP capable.
240 	 */
241 	/* Athlon 660/661 is valid. */
242 	if ((c->x86_model == 6) && ((c->x86_mask == 0) ||
243 	    (c->x86_mask == 1)))
244 		return;
245 
246 	/* Duron 670 is valid */
247 	if ((c->x86_model == 7) && (c->x86_mask == 0))
248 		return;
249 
250 	/*
251 	 * Athlon 662, Duron 671, and Athlon >model 7 have capability
252 	 * bit. It's worth noting that the A5 stepping (662) of some
253 	 * Athlon XP's have the MP bit set.
254 	 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for
255 	 * more.
256 	 */
257 	if (((c->x86_model == 6) && (c->x86_mask >= 2)) ||
258 	    ((c->x86_model == 7) && (c->x86_mask >= 1)) ||
259 	     (c->x86_model > 7))
260 		if (cpu_has(c, X86_FEATURE_MP))
261 			return;
262 
263 	/* If we get here, not a certified SMP capable AMD system. */
264 
265 	/*
266 	 * Don't taint if we are running SMP kernel on a single non-MP
267 	 * approved Athlon
268 	 */
269 	WARN_ONCE(1, "WARNING: This combination of AMD"
270 		" processors is not suitable for SMP.\n");
271 	add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_NOW_UNRELIABLE);
272 #endif
273 }
274 
275 #ifdef CONFIG_NUMA
276 /*
277  * To workaround broken NUMA config.  Read the comment in
278  * srat_detect_node().
279  */
280 static int nearby_node(int apicid)
281 {
282 	int i, node;
283 
284 	for (i = apicid - 1; i >= 0; i--) {
285 		node = __apicid_to_node[i];
286 		if (node != NUMA_NO_NODE && node_online(node))
287 			return node;
288 	}
289 	for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) {
290 		node = __apicid_to_node[i];
291 		if (node != NUMA_NO_NODE && node_online(node))
292 			return node;
293 	}
294 	return first_node(node_online_map); /* Shouldn't happen */
295 }
296 #endif
297 
298 /*
299  * Fixup core topology information for
300  * (1) AMD multi-node processors
301  *     Assumption: Number of cores in each internal node is the same.
302  * (2) AMD processors supporting compute units
303  */
304 #ifdef CONFIG_SMP
305 static void amd_get_topology(struct cpuinfo_x86 *c)
306 {
307 	u8 node_id;
308 	int cpu = smp_processor_id();
309 
310 	/* get information required for multi-node processors */
311 	if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
312 		u32 eax, ebx, ecx, edx;
313 
314 		cpuid(0x8000001e, &eax, &ebx, &ecx, &edx);
315 		node_id = ecx & 7;
316 
317 		/* get compute unit information */
318 		smp_num_siblings = ((ebx >> 8) & 3) + 1;
319 		c->x86_max_cores /= smp_num_siblings;
320 		c->cpu_core_id = ebx & 0xff;
321 
322 		/*
323 		 * We may have multiple LLCs if L3 caches exist, so check if we
324 		 * have an L3 cache by looking at the L3 cache CPUID leaf.
325 		 */
326 		if (cpuid_edx(0x80000006)) {
327 			if (c->x86 == 0x17) {
328 				/*
329 				 * LLC is at the core complex level.
330 				 * Core complex id is ApicId[3].
331 				 */
332 				per_cpu(cpu_llc_id, cpu) = c->apicid >> 3;
333 			} else {
334 				/* LLC is at the node level. */
335 				per_cpu(cpu_llc_id, cpu) = node_id;
336 			}
337 		}
338 	} else if (cpu_has(c, X86_FEATURE_NODEID_MSR)) {
339 		u64 value;
340 
341 		rdmsrl(MSR_FAM10H_NODE_ID, value);
342 		node_id = value & 7;
343 
344 		per_cpu(cpu_llc_id, cpu) = node_id;
345 	} else
346 		return;
347 
348 	/* fixup multi-node processor information */
349 	if (nodes_per_socket > 1) {
350 		u32 cus_per_node;
351 
352 		set_cpu_cap(c, X86_FEATURE_AMD_DCM);
353 		cus_per_node = c->x86_max_cores / nodes_per_socket;
354 
355 		/* core id has to be in the [0 .. cores_per_node - 1] range */
356 		c->cpu_core_id %= cus_per_node;
357 	}
358 }
359 #endif
360 
361 /*
362  * On a AMD dual core setup the lower bits of the APIC id distinguish the cores.
363  * Assumes number of cores is a power of two.
364  */
365 static void amd_detect_cmp(struct cpuinfo_x86 *c)
366 {
367 #ifdef CONFIG_SMP
368 	unsigned bits;
369 	int cpu = smp_processor_id();
370 
371 	bits = c->x86_coreid_bits;
372 	/* Low order bits define the core id (index of core in socket) */
373 	c->cpu_core_id = c->initial_apicid & ((1 << bits)-1);
374 	/* Convert the initial APIC ID into the socket ID */
375 	c->phys_proc_id = c->initial_apicid >> bits;
376 	/* use socket ID also for last level cache */
377 	per_cpu(cpu_llc_id, cpu) = c->phys_proc_id;
378 	amd_get_topology(c);
379 #endif
380 }
381 
382 u16 amd_get_nb_id(int cpu)
383 {
384 	u16 id = 0;
385 #ifdef CONFIG_SMP
386 	id = per_cpu(cpu_llc_id, cpu);
387 #endif
388 	return id;
389 }
390 EXPORT_SYMBOL_GPL(amd_get_nb_id);
391 
392 u32 amd_get_nodes_per_socket(void)
393 {
394 	return nodes_per_socket;
395 }
396 EXPORT_SYMBOL_GPL(amd_get_nodes_per_socket);
397 
398 static void srat_detect_node(struct cpuinfo_x86 *c)
399 {
400 #ifdef CONFIG_NUMA
401 	int cpu = smp_processor_id();
402 	int node;
403 	unsigned apicid = c->apicid;
404 
405 	node = numa_cpu_node(cpu);
406 	if (node == NUMA_NO_NODE)
407 		node = per_cpu(cpu_llc_id, cpu);
408 
409 	/*
410 	 * On multi-fabric platform (e.g. Numascale NumaChip) a
411 	 * platform-specific handler needs to be called to fixup some
412 	 * IDs of the CPU.
413 	 */
414 	if (x86_cpuinit.fixup_cpu_id)
415 		x86_cpuinit.fixup_cpu_id(c, node);
416 
417 	if (!node_online(node)) {
418 		/*
419 		 * Two possibilities here:
420 		 *
421 		 * - The CPU is missing memory and no node was created.  In
422 		 *   that case try picking one from a nearby CPU.
423 		 *
424 		 * - The APIC IDs differ from the HyperTransport node IDs
425 		 *   which the K8 northbridge parsing fills in.  Assume
426 		 *   they are all increased by a constant offset, but in
427 		 *   the same order as the HT nodeids.  If that doesn't
428 		 *   result in a usable node fall back to the path for the
429 		 *   previous case.
430 		 *
431 		 * This workaround operates directly on the mapping between
432 		 * APIC ID and NUMA node, assuming certain relationship
433 		 * between APIC ID, HT node ID and NUMA topology.  As going
434 		 * through CPU mapping may alter the outcome, directly
435 		 * access __apicid_to_node[].
436 		 */
437 		int ht_nodeid = c->initial_apicid;
438 
439 		if (__apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
440 			node = __apicid_to_node[ht_nodeid];
441 		/* Pick a nearby node */
442 		if (!node_online(node))
443 			node = nearby_node(apicid);
444 	}
445 	numa_set_node(cpu, node);
446 #endif
447 }
448 
449 static void early_init_amd_mc(struct cpuinfo_x86 *c)
450 {
451 #ifdef CONFIG_SMP
452 	unsigned bits, ecx;
453 
454 	/* Multi core CPU? */
455 	if (c->extended_cpuid_level < 0x80000008)
456 		return;
457 
458 	ecx = cpuid_ecx(0x80000008);
459 
460 	c->x86_max_cores = (ecx & 0xff) + 1;
461 
462 	/* CPU telling us the core id bits shift? */
463 	bits = (ecx >> 12) & 0xF;
464 
465 	/* Otherwise recompute */
466 	if (bits == 0) {
467 		while ((1 << bits) < c->x86_max_cores)
468 			bits++;
469 	}
470 
471 	c->x86_coreid_bits = bits;
472 #endif
473 }
474 
475 static void bsp_init_amd(struct cpuinfo_x86 *c)
476 {
477 
478 #ifdef CONFIG_X86_64
479 	if (c->x86 >= 0xf) {
480 		unsigned long long tseg;
481 
482 		/*
483 		 * Split up direct mapping around the TSEG SMM area.
484 		 * Don't do it for gbpages because there seems very little
485 		 * benefit in doing so.
486 		 */
487 		if (!rdmsrl_safe(MSR_K8_TSEG_ADDR, &tseg)) {
488 			unsigned long pfn = tseg >> PAGE_SHIFT;
489 
490 			pr_debug("tseg: %010llx\n", tseg);
491 			if (pfn_range_is_mapped(pfn, pfn + 1))
492 				set_memory_4k((unsigned long)__va(tseg), 1);
493 		}
494 	}
495 #endif
496 
497 	if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) {
498 
499 		if (c->x86 > 0x10 ||
500 		    (c->x86 == 0x10 && c->x86_model >= 0x2)) {
501 			u64 val;
502 
503 			rdmsrl(MSR_K7_HWCR, val);
504 			if (!(val & BIT(24)))
505 				pr_warn(FW_BUG "TSC doesn't count with P0 frequency!\n");
506 		}
507 	}
508 
509 	if (c->x86 == 0x15) {
510 		unsigned long upperbit;
511 		u32 cpuid, assoc;
512 
513 		cpuid	 = cpuid_edx(0x80000005);
514 		assoc	 = cpuid >> 16 & 0xff;
515 		upperbit = ((cpuid >> 24) << 10) / assoc;
516 
517 		va_align.mask	  = (upperbit - 1) & PAGE_MASK;
518 		va_align.flags    = ALIGN_VA_32 | ALIGN_VA_64;
519 
520 		/* A random value per boot for bit slice [12:upper_bit) */
521 		va_align.bits = get_random_int() & va_align.mask;
522 	}
523 
524 	if (cpu_has(c, X86_FEATURE_MWAITX))
525 		use_mwaitx_delay();
526 
527 	if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
528 		u32 ecx;
529 
530 		ecx = cpuid_ecx(0x8000001e);
531 		nodes_per_socket = ((ecx >> 8) & 7) + 1;
532 	} else if (boot_cpu_has(X86_FEATURE_NODEID_MSR)) {
533 		u64 value;
534 
535 		rdmsrl(MSR_FAM10H_NODE_ID, value);
536 		nodes_per_socket = ((value >> 3) & 7) + 1;
537 	}
538 }
539 
540 static void early_init_amd(struct cpuinfo_x86 *c)
541 {
542 	early_init_amd_mc(c);
543 
544 	/*
545 	 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
546 	 * with P/T states and does not stop in deep C-states
547 	 */
548 	if (c->x86_power & (1 << 8)) {
549 		set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
550 		set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
551 		if (!check_tsc_unstable())
552 			set_sched_clock_stable();
553 	}
554 
555 	/* Bit 12 of 8000_0007 edx is accumulated power mechanism. */
556 	if (c->x86_power & BIT(12))
557 		set_cpu_cap(c, X86_FEATURE_ACC_POWER);
558 
559 #ifdef CONFIG_X86_64
560 	set_cpu_cap(c, X86_FEATURE_SYSCALL32);
561 #else
562 	/*  Set MTRR capability flag if appropriate */
563 	if (c->x86 == 5)
564 		if (c->x86_model == 13 || c->x86_model == 9 ||
565 		    (c->x86_model == 8 && c->x86_mask >= 8))
566 			set_cpu_cap(c, X86_FEATURE_K6_MTRR);
567 #endif
568 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
569 	/*
570 	 * ApicID can always be treated as an 8-bit value for AMD APIC versions
571 	 * >= 0x10, but even old K8s came out of reset with version 0x10. So, we
572 	 * can safely set X86_FEATURE_EXTD_APICID unconditionally for families
573 	 * after 16h.
574 	 */
575 	if (boot_cpu_has(X86_FEATURE_APIC)) {
576 		if (c->x86 > 0x16)
577 			set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
578 		else if (c->x86 >= 0xf) {
579 			/* check CPU config space for extended APIC ID */
580 			unsigned int val;
581 
582 			val = read_pci_config(0, 24, 0, 0x68);
583 			if ((val >> 17 & 0x3) == 0x3)
584 				set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
585 		}
586 	}
587 #endif
588 
589 	/*
590 	 * This is only needed to tell the kernel whether to use VMCALL
591 	 * and VMMCALL.  VMMCALL is never executed except under virt, so
592 	 * we can set it unconditionally.
593 	 */
594 	set_cpu_cap(c, X86_FEATURE_VMMCALL);
595 
596 	/* F16h erratum 793, CVE-2013-6885 */
597 	if (c->x86 == 0x16 && c->x86_model <= 0xf)
598 		msr_set_bit(MSR_AMD64_LS_CFG, 15);
599 
600 	/*
601 	 * Check whether the machine is affected by erratum 400. This is
602 	 * used to select the proper idle routine and to enable the check
603 	 * whether the machine is affected in arch_post_acpi_init(), which
604 	 * sets the X86_BUG_AMD_APIC_C1E bug depending on the MSR check.
605 	 */
606 	if (cpu_has_amd_erratum(c, amd_erratum_400))
607 		set_cpu_bug(c, X86_BUG_AMD_E400);
608 }
609 
610 static void init_amd_k8(struct cpuinfo_x86 *c)
611 {
612 	u32 level;
613 	u64 value;
614 
615 	/* On C+ stepping K8 rep microcode works well for copy/memset */
616 	level = cpuid_eax(1);
617 	if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58)
618 		set_cpu_cap(c, X86_FEATURE_REP_GOOD);
619 
620 	/*
621 	 * Some BIOSes incorrectly force this feature, but only K8 revision D
622 	 * (model = 0x14) and later actually support it.
623 	 * (AMD Erratum #110, docId: 25759).
624 	 */
625 	if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM)) {
626 		clear_cpu_cap(c, X86_FEATURE_LAHF_LM);
627 		if (!rdmsrl_amd_safe(0xc001100d, &value)) {
628 			value &= ~BIT_64(32);
629 			wrmsrl_amd_safe(0xc001100d, value);
630 		}
631 	}
632 
633 	if (!c->x86_model_id[0])
634 		strcpy(c->x86_model_id, "Hammer");
635 
636 #ifdef CONFIG_SMP
637 	/*
638 	 * Disable TLB flush filter by setting HWCR.FFDIS on K8
639 	 * bit 6 of msr C001_0015
640 	 *
641 	 * Errata 63 for SH-B3 steppings
642 	 * Errata 122 for all steppings (F+ have it disabled by default)
643 	 */
644 	msr_set_bit(MSR_K7_HWCR, 6);
645 #endif
646 	set_cpu_bug(c, X86_BUG_SWAPGS_FENCE);
647 }
648 
649 static void init_amd_gh(struct cpuinfo_x86 *c)
650 {
651 #ifdef CONFIG_X86_64
652 	/* do this for boot cpu */
653 	if (c == &boot_cpu_data)
654 		check_enable_amd_mmconf_dmi();
655 
656 	fam10h_check_enable_mmcfg();
657 #endif
658 
659 	/*
660 	 * Disable GART TLB Walk Errors on Fam10h. We do this here because this
661 	 * is always needed when GART is enabled, even in a kernel which has no
662 	 * MCE support built in. BIOS should disable GartTlbWlk Errors already.
663 	 * If it doesn't, we do it here as suggested by the BKDG.
664 	 *
665 	 * Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=33012
666 	 */
667 	msr_set_bit(MSR_AMD64_MCx_MASK(4), 10);
668 
669 	/*
670 	 * On family 10h BIOS may not have properly enabled WC+ support, causing
671 	 * it to be converted to CD memtype. This may result in performance
672 	 * degradation for certain nested-paging guests. Prevent this conversion
673 	 * by clearing bit 24 in MSR_AMD64_BU_CFG2.
674 	 *
675 	 * NOTE: we want to use the _safe accessors so as not to #GP kvm
676 	 * guests on older kvm hosts.
677 	 */
678 	msr_clear_bit(MSR_AMD64_BU_CFG2, 24);
679 
680 	if (cpu_has_amd_erratum(c, amd_erratum_383))
681 		set_cpu_bug(c, X86_BUG_AMD_TLB_MMATCH);
682 }
683 
684 #define MSR_AMD64_DE_CFG	0xC0011029
685 
686 static void init_amd_ln(struct cpuinfo_x86 *c)
687 {
688 	/*
689 	 * Apply erratum 665 fix unconditionally so machines without a BIOS
690 	 * fix work.
691 	 */
692 	msr_set_bit(MSR_AMD64_DE_CFG, 31);
693 }
694 
695 static void init_amd_bd(struct cpuinfo_x86 *c)
696 {
697 	u64 value;
698 
699 	/* re-enable TopologyExtensions if switched off by BIOS */
700 	if ((c->x86_model >= 0x10) && (c->x86_model <= 0x6f) &&
701 	    !cpu_has(c, X86_FEATURE_TOPOEXT)) {
702 
703 		if (msr_set_bit(0xc0011005, 54) > 0) {
704 			rdmsrl(0xc0011005, value);
705 			if (value & BIT_64(54)) {
706 				set_cpu_cap(c, X86_FEATURE_TOPOEXT);
707 				pr_info_once(FW_INFO "CPU: Re-enabling disabled Topology Extensions Support.\n");
708 			}
709 		}
710 	}
711 
712 	/*
713 	 * The way access filter has a performance penalty on some workloads.
714 	 * Disable it on the affected CPUs.
715 	 */
716 	if ((c->x86_model >= 0x02) && (c->x86_model < 0x20)) {
717 		if (!rdmsrl_safe(MSR_F15H_IC_CFG, &value) && !(value & 0x1E)) {
718 			value |= 0x1E;
719 			wrmsrl_safe(MSR_F15H_IC_CFG, value);
720 		}
721 	}
722 }
723 
724 static void init_amd(struct cpuinfo_x86 *c)
725 {
726 	u32 dummy;
727 
728 	early_init_amd(c);
729 
730 	/*
731 	 * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
732 	 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
733 	 */
734 	clear_cpu_cap(c, 0*32+31);
735 
736 	if (c->x86 >= 0x10)
737 		set_cpu_cap(c, X86_FEATURE_REP_GOOD);
738 
739 	/* get apicid instead of initial apic id from cpuid */
740 	c->apicid = hard_smp_processor_id();
741 
742 	/* K6s reports MCEs but don't actually have all the MSRs */
743 	if (c->x86 < 6)
744 		clear_cpu_cap(c, X86_FEATURE_MCE);
745 
746 	switch (c->x86) {
747 	case 4:    init_amd_k5(c); break;
748 	case 5:    init_amd_k6(c); break;
749 	case 6:	   init_amd_k7(c); break;
750 	case 0xf:  init_amd_k8(c); break;
751 	case 0x10: init_amd_gh(c); break;
752 	case 0x12: init_amd_ln(c); break;
753 	case 0x15: init_amd_bd(c); break;
754 	}
755 
756 	/* Enable workaround for FXSAVE leak */
757 	if (c->x86 >= 6)
758 		set_cpu_bug(c, X86_BUG_FXSAVE_LEAK);
759 
760 	cpu_detect_cache_sizes(c);
761 
762 	/* Multi core CPU? */
763 	if (c->extended_cpuid_level >= 0x80000008) {
764 		amd_detect_cmp(c);
765 		srat_detect_node(c);
766 	}
767 
768 #ifdef CONFIG_X86_32
769 	detect_ht(c);
770 #endif
771 
772 	init_amd_cacheinfo(c);
773 
774 	if (c->x86 >= 0xf)
775 		set_cpu_cap(c, X86_FEATURE_K8);
776 
777 	if (cpu_has(c, X86_FEATURE_XMM2)) {
778 		/* MFENCE stops RDTSC speculation */
779 		set_cpu_cap(c, X86_FEATURE_MFENCE_RDTSC);
780 	}
781 
782 	/*
783 	 * Family 0x12 and above processors have APIC timer
784 	 * running in deep C states.
785 	 */
786 	if (c->x86 > 0x11)
787 		set_cpu_cap(c, X86_FEATURE_ARAT);
788 
789 	rdmsr_safe(MSR_AMD64_PATCH_LEVEL, &c->microcode, &dummy);
790 
791 	/* 3DNow or LM implies PREFETCHW */
792 	if (!cpu_has(c, X86_FEATURE_3DNOWPREFETCH))
793 		if (cpu_has(c, X86_FEATURE_3DNOW) || cpu_has(c, X86_FEATURE_LM))
794 			set_cpu_cap(c, X86_FEATURE_3DNOWPREFETCH);
795 
796 	/* AMD CPUs don't reset SS attributes on SYSRET */
797 	set_cpu_bug(c, X86_BUG_SYSRET_SS_ATTRS);
798 }
799 
800 #ifdef CONFIG_X86_32
801 static unsigned int amd_size_cache(struct cpuinfo_x86 *c, unsigned int size)
802 {
803 	/* AMD errata T13 (order #21922) */
804 	if ((c->x86 == 6)) {
805 		/* Duron Rev A0 */
806 		if (c->x86_model == 3 && c->x86_mask == 0)
807 			size = 64;
808 		/* Tbird rev A1/A2 */
809 		if (c->x86_model == 4 &&
810 			(c->x86_mask == 0 || c->x86_mask == 1))
811 			size = 256;
812 	}
813 	return size;
814 }
815 #endif
816 
817 static void cpu_detect_tlb_amd(struct cpuinfo_x86 *c)
818 {
819 	u32 ebx, eax, ecx, edx;
820 	u16 mask = 0xfff;
821 
822 	if (c->x86 < 0xf)
823 		return;
824 
825 	if (c->extended_cpuid_level < 0x80000006)
826 		return;
827 
828 	cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
829 
830 	tlb_lld_4k[ENTRIES] = (ebx >> 16) & mask;
831 	tlb_lli_4k[ENTRIES] = ebx & mask;
832 
833 	/*
834 	 * K8 doesn't have 2M/4M entries in the L2 TLB so read out the L1 TLB
835 	 * characteristics from the CPUID function 0x80000005 instead.
836 	 */
837 	if (c->x86 == 0xf) {
838 		cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
839 		mask = 0xff;
840 	}
841 
842 	/* Handle DTLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
843 	if (!((eax >> 16) & mask))
844 		tlb_lld_2m[ENTRIES] = (cpuid_eax(0x80000005) >> 16) & 0xff;
845 	else
846 		tlb_lld_2m[ENTRIES] = (eax >> 16) & mask;
847 
848 	/* a 4M entry uses two 2M entries */
849 	tlb_lld_4m[ENTRIES] = tlb_lld_2m[ENTRIES] >> 1;
850 
851 	/* Handle ITLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
852 	if (!(eax & mask)) {
853 		/* Erratum 658 */
854 		if (c->x86 == 0x15 && c->x86_model <= 0x1f) {
855 			tlb_lli_2m[ENTRIES] = 1024;
856 		} else {
857 			cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
858 			tlb_lli_2m[ENTRIES] = eax & 0xff;
859 		}
860 	} else
861 		tlb_lli_2m[ENTRIES] = eax & mask;
862 
863 	tlb_lli_4m[ENTRIES] = tlb_lli_2m[ENTRIES] >> 1;
864 }
865 
866 static const struct cpu_dev amd_cpu_dev = {
867 	.c_vendor	= "AMD",
868 	.c_ident	= { "AuthenticAMD" },
869 #ifdef CONFIG_X86_32
870 	.legacy_models = {
871 		{ .family = 4, .model_names =
872 		  {
873 			  [3] = "486 DX/2",
874 			  [7] = "486 DX/2-WB",
875 			  [8] = "486 DX/4",
876 			  [9] = "486 DX/4-WB",
877 			  [14] = "Am5x86-WT",
878 			  [15] = "Am5x86-WB"
879 		  }
880 		},
881 	},
882 	.legacy_cache_size = amd_size_cache,
883 #endif
884 	.c_early_init   = early_init_amd,
885 	.c_detect_tlb	= cpu_detect_tlb_amd,
886 	.c_bsp_init	= bsp_init_amd,
887 	.c_init		= init_amd,
888 	.c_x86_vendor	= X86_VENDOR_AMD,
889 };
890 
891 cpu_dev_register(amd_cpu_dev);
892 
893 /*
894  * AMD errata checking
895  *
896  * Errata are defined as arrays of ints using the AMD_LEGACY_ERRATUM() or
897  * AMD_OSVW_ERRATUM() macros. The latter is intended for newer errata that
898  * have an OSVW id assigned, which it takes as first argument. Both take a
899  * variable number of family-specific model-stepping ranges created by
900  * AMD_MODEL_RANGE().
901  *
902  * Example:
903  *
904  * const int amd_erratum_319[] =
905  *	AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0x4, 0x2),
906  *			   AMD_MODEL_RANGE(0x10, 0x8, 0x0, 0x8, 0x0),
907  *			   AMD_MODEL_RANGE(0x10, 0x9, 0x0, 0x9, 0x0));
908  */
909 
910 #define AMD_LEGACY_ERRATUM(...)		{ -1, __VA_ARGS__, 0 }
911 #define AMD_OSVW_ERRATUM(osvw_id, ...)	{ osvw_id, __VA_ARGS__, 0 }
912 #define AMD_MODEL_RANGE(f, m_start, s_start, m_end, s_end) \
913 	((f << 24) | (m_start << 16) | (s_start << 12) | (m_end << 4) | (s_end))
914 #define AMD_MODEL_RANGE_FAMILY(range)	(((range) >> 24) & 0xff)
915 #define AMD_MODEL_RANGE_START(range)	(((range) >> 12) & 0xfff)
916 #define AMD_MODEL_RANGE_END(range)	((range) & 0xfff)
917 
918 static const int amd_erratum_400[] =
919 	AMD_OSVW_ERRATUM(1, AMD_MODEL_RANGE(0xf, 0x41, 0x2, 0xff, 0xf),
920 			    AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0xff, 0xf));
921 
922 static const int amd_erratum_383[] =
923 	AMD_OSVW_ERRATUM(3, AMD_MODEL_RANGE(0x10, 0, 0, 0xff, 0xf));
924 
925 
926 static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum)
927 {
928 	int osvw_id = *erratum++;
929 	u32 range;
930 	u32 ms;
931 
932 	if (osvw_id >= 0 && osvw_id < 65536 &&
933 	    cpu_has(cpu, X86_FEATURE_OSVW)) {
934 		u64 osvw_len;
935 
936 		rdmsrl(MSR_AMD64_OSVW_ID_LENGTH, osvw_len);
937 		if (osvw_id < osvw_len) {
938 			u64 osvw_bits;
939 
940 			rdmsrl(MSR_AMD64_OSVW_STATUS + (osvw_id >> 6),
941 			    osvw_bits);
942 			return osvw_bits & (1ULL << (osvw_id & 0x3f));
943 		}
944 	}
945 
946 	/* OSVW unavailable or ID unknown, match family-model-stepping range */
947 	ms = (cpu->x86_model << 4) | cpu->x86_mask;
948 	while ((range = *erratum++))
949 		if ((cpu->x86 == AMD_MODEL_RANGE_FAMILY(range)) &&
950 		    (ms >= AMD_MODEL_RANGE_START(range)) &&
951 		    (ms <= AMD_MODEL_RANGE_END(range)))
952 			return true;
953 
954 	return false;
955 }
956 
957 void set_dr_addr_mask(unsigned long mask, int dr)
958 {
959 	if (!boot_cpu_has(X86_FEATURE_BPEXT))
960 		return;
961 
962 	switch (dr) {
963 	case 0:
964 		wrmsr(MSR_F16H_DR0_ADDR_MASK, mask, 0);
965 		break;
966 	case 1:
967 	case 2:
968 	case 3:
969 		wrmsr(MSR_F16H_DR1_ADDR_MASK - 1 + dr, mask, 0);
970 		break;
971 	default:
972 		break;
973 	}
974 }
975