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/*
24 * nodes_per_socket: Stores the number of nodes per socket.
25 * Refer to Fam15h Models 00-0fh BKDG - CPUID Fn8000_001E_ECX
26 * Node Identifiers[10:8]
27 */
28static u32 nodes_per_socket = 1;
29
30static inline int rdmsrl_amd_safe(unsigned msr, unsigned long long *p)
31{
32	u32 gprs[8] = { 0 };
33	int err;
34
35	WARN_ONCE((boot_cpu_data.x86 != 0xf),
36		  "%s should only be used on K8!\n", __func__);
37
38	gprs[1] = msr;
39	gprs[7] = 0x9c5a203a;
40
41	err = rdmsr_safe_regs(gprs);
42
43	*p = gprs[0] | ((u64)gprs[2] << 32);
44
45	return err;
46}
47
48static inline int wrmsrl_amd_safe(unsigned msr, unsigned long long val)
49{
50	u32 gprs[8] = { 0 };
51
52	WARN_ONCE((boot_cpu_data.x86 != 0xf),
53		  "%s should only be used on K8!\n", __func__);
54
55	gprs[0] = (u32)val;
56	gprs[1] = msr;
57	gprs[2] = val >> 32;
58	gprs[7] = 0x9c5a203a;
59
60	return wrmsr_safe_regs(gprs);
61}
62
63/*
64 *	B step AMD K6 before B 9730xxxx have hardware bugs that can cause
65 *	misexecution of code under Linux. Owners of such processors should
66 *	contact AMD for precise details and a CPU swap.
67 *
68 *	See	http://www.multimania.com/poulot/k6bug.html
69 *	and	section 2.6.2 of "AMD-K6 Processor Revision Guide - Model 6"
70 *		(Publication # 21266  Issue Date: August 1998)
71 *
72 *	The following test is erm.. interesting. AMD neglected to up
73 *	the chip setting when fixing the bug but they also tweaked some
74 *	performance at the same time..
75 */
76
77extern __visible void vide(void);
78__asm__(".globl vide\n\t.align 4\nvide: ret");
79
80static void init_amd_k5(struct cpuinfo_x86 *c)
81{
82#ifdef CONFIG_X86_32
83/*
84 * General Systems BIOSen alias the cpu frequency registers
85 * of the Elan at 0x000df000. Unfortuantly, one of the Linux
86 * drivers subsequently pokes it, and changes the CPU speed.
87 * Workaround : Remove the unneeded alias.
88 */
89#define CBAR		(0xfffc) /* Configuration Base Address  (32-bit) */
90#define CBAR_ENB	(0x80000000)
91#define CBAR_KEY	(0X000000CB)
92	if (c->x86_model == 9 || c->x86_model == 10) {
93		if (inl(CBAR) & CBAR_ENB)
94			outl(0 | CBAR_KEY, CBAR);
95	}
96#endif
97}
98
99static void init_amd_k6(struct cpuinfo_x86 *c)
100{
101#ifdef CONFIG_X86_32
102	u32 l, h;
103	int mbytes = get_num_physpages() >> (20-PAGE_SHIFT);
104
105	if (c->x86_model < 6) {
106		/* Based on AMD doc 20734R - June 2000 */
107		if (c->x86_model == 0) {
108			clear_cpu_cap(c, X86_FEATURE_APIC);
109			set_cpu_cap(c, X86_FEATURE_PGE);
110		}
111		return;
112	}
113
114	if (c->x86_model == 6 && c->x86_mask == 1) {
115		const int K6_BUG_LOOP = 1000000;
116		int n;
117		void (*f_vide)(void);
118		u64 d, d2;
119
120		printk(KERN_INFO "AMD K6 stepping B detected - ");
121
122		/*
123		 * It looks like AMD fixed the 2.6.2 bug and improved indirect
124		 * calls at the same time.
125		 */
126
127		n = K6_BUG_LOOP;
128		f_vide = vide;
129		d = rdtsc();
130		while (n--)
131			f_vide();
132		d2 = rdtsc();
133		d = d2-d;
134
135		if (d > 20*K6_BUG_LOOP)
136			printk(KERN_CONT
137				"system stability may be impaired when more than 32 MB are used.\n");
138		else
139			printk(KERN_CONT "probably OK (after B9730xxxx).\n");
140	}
141
142	/* K6 with old style WHCR */
143	if (c->x86_model < 8 ||
144	   (c->x86_model == 8 && c->x86_mask < 8)) {
145		/* We can only write allocate on the low 508Mb */
146		if (mbytes > 508)
147			mbytes = 508;
148
149		rdmsr(MSR_K6_WHCR, l, h);
150		if ((l&0x0000FFFF) == 0) {
151			unsigned long flags;
152			l = (1<<0)|((mbytes/4)<<1);
153			local_irq_save(flags);
154			wbinvd();
155			wrmsr(MSR_K6_WHCR, l, h);
156			local_irq_restore(flags);
157			printk(KERN_INFO "Enabling old style K6 write allocation for %d Mb\n",
158				mbytes);
159		}
160		return;
161	}
162
163	if ((c->x86_model == 8 && c->x86_mask > 7) ||
164	     c->x86_model == 9 || c->x86_model == 13) {
165		/* The more serious chips .. */
166
167		if (mbytes > 4092)
168			mbytes = 4092;
169
170		rdmsr(MSR_K6_WHCR, l, h);
171		if ((l&0xFFFF0000) == 0) {
172			unsigned long flags;
173			l = ((mbytes>>2)<<22)|(1<<16);
174			local_irq_save(flags);
175			wbinvd();
176			wrmsr(MSR_K6_WHCR, l, h);
177			local_irq_restore(flags);
178			printk(KERN_INFO "Enabling new style K6 write allocation for %d Mb\n",
179				mbytes);
180		}
181
182		return;
183	}
184
185	if (c->x86_model == 10) {
186		/* AMD Geode LX is model 10 */
187		/* placeholder for any needed mods */
188		return;
189	}
190#endif
191}
192
193static void init_amd_k7(struct cpuinfo_x86 *c)
194{
195#ifdef CONFIG_X86_32
196	u32 l, h;
197
198	/*
199	 * Bit 15 of Athlon specific MSR 15, needs to be 0
200	 * to enable SSE on Palomino/Morgan/Barton CPU's.
201	 * If the BIOS didn't enable it already, enable it here.
202	 */
203	if (c->x86_model >= 6 && c->x86_model <= 10) {
204		if (!cpu_has(c, X86_FEATURE_XMM)) {
205			printk(KERN_INFO "Enabling disabled K7/SSE Support.\n");
206			msr_clear_bit(MSR_K7_HWCR, 15);
207			set_cpu_cap(c, X86_FEATURE_XMM);
208		}
209	}
210
211	/*
212	 * It's been determined by AMD that Athlons since model 8 stepping 1
213	 * are more robust with CLK_CTL set to 200xxxxx instead of 600xxxxx
214	 * As per AMD technical note 27212 0.2
215	 */
216	if ((c->x86_model == 8 && c->x86_mask >= 1) || (c->x86_model > 8)) {
217		rdmsr(MSR_K7_CLK_CTL, l, h);
218		if ((l & 0xfff00000) != 0x20000000) {
219			printk(KERN_INFO
220			    "CPU: CLK_CTL MSR was %x. Reprogramming to %x\n",
221					l, ((l & 0x000fffff)|0x20000000));
222			wrmsr(MSR_K7_CLK_CTL, (l & 0x000fffff)|0x20000000, h);
223		}
224	}
225
226	set_cpu_cap(c, X86_FEATURE_K7);
227
228	/* calling is from identify_secondary_cpu() ? */
229	if (!c->cpu_index)
230		return;
231
232	/*
233	 * Certain Athlons might work (for various values of 'work') in SMP
234	 * but they are not certified as MP capable.
235	 */
236	/* Athlon 660/661 is valid. */
237	if ((c->x86_model == 6) && ((c->x86_mask == 0) ||
238	    (c->x86_mask == 1)))
239		return;
240
241	/* Duron 670 is valid */
242	if ((c->x86_model == 7) && (c->x86_mask == 0))
243		return;
244
245	/*
246	 * Athlon 662, Duron 671, and Athlon >model 7 have capability
247	 * bit. It's worth noting that the A5 stepping (662) of some
248	 * Athlon XP's have the MP bit set.
249	 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for
250	 * more.
251	 */
252	if (((c->x86_model == 6) && (c->x86_mask >= 2)) ||
253	    ((c->x86_model == 7) && (c->x86_mask >= 1)) ||
254	     (c->x86_model > 7))
255		if (cpu_has(c, X86_FEATURE_MP))
256			return;
257
258	/* If we get here, not a certified SMP capable AMD system. */
259
260	/*
261	 * Don't taint if we are running SMP kernel on a single non-MP
262	 * approved Athlon
263	 */
264	WARN_ONCE(1, "WARNING: This combination of AMD"
265		" processors is not suitable for SMP.\n");
266	add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_NOW_UNRELIABLE);
267#endif
268}
269
270#ifdef CONFIG_NUMA
271/*
272 * To workaround broken NUMA config.  Read the comment in
273 * srat_detect_node().
274 */
275static int nearby_node(int apicid)
276{
277	int i, node;
278
279	for (i = apicid - 1; i >= 0; i--) {
280		node = __apicid_to_node[i];
281		if (node != NUMA_NO_NODE && node_online(node))
282			return node;
283	}
284	for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) {
285		node = __apicid_to_node[i];
286		if (node != NUMA_NO_NODE && node_online(node))
287			return node;
288	}
289	return first_node(node_online_map); /* Shouldn't happen */
290}
291#endif
292
293/*
294 * Fixup core topology information for
295 * (1) AMD multi-node processors
296 *     Assumption: Number of cores in each internal node is the same.
297 * (2) AMD processors supporting compute units
298 */
299#ifdef CONFIG_SMP
300static void amd_get_topology(struct cpuinfo_x86 *c)
301{
302	u32 cores_per_cu = 1;
303	u8 node_id;
304	int cpu = smp_processor_id();
305
306	/* get information required for multi-node processors */
307	if (cpu_has_topoext) {
308		u32 eax, ebx, ecx, edx;
309
310		cpuid(0x8000001e, &eax, &ebx, &ecx, &edx);
311		nodes_per_socket = ((ecx >> 8) & 7) + 1;
312		node_id = ecx & 7;
313
314		/* get compute unit information */
315		smp_num_siblings = ((ebx >> 8) & 3) + 1;
316		c->compute_unit_id = ebx & 0xff;
317		cores_per_cu += ((ebx >> 8) & 3);
318	} else if (cpu_has(c, X86_FEATURE_NODEID_MSR)) {
319		u64 value;
320
321		rdmsrl(MSR_FAM10H_NODE_ID, value);
322		nodes_per_socket = ((value >> 3) & 7) + 1;
323		node_id = value & 7;
324	} else
325		return;
326
327	/* fixup multi-node processor information */
328	if (nodes_per_socket > 1) {
329		u32 cores_per_node;
330		u32 cus_per_node;
331
332		set_cpu_cap(c, X86_FEATURE_AMD_DCM);
333		cores_per_node = c->x86_max_cores / nodes_per_socket;
334		cus_per_node = cores_per_node / cores_per_cu;
335
336		/* store NodeID, use llc_shared_map to store sibling info */
337		per_cpu(cpu_llc_id, cpu) = node_id;
338
339		/* core id has to be in the [0 .. cores_per_node - 1] range */
340		c->cpu_core_id %= cores_per_node;
341		c->compute_unit_id %= cus_per_node;
342	}
343}
344#endif
345
346/*
347 * On a AMD dual core setup the lower bits of the APIC id distinguish the cores.
348 * Assumes number of cores is a power of two.
349 */
350static void amd_detect_cmp(struct cpuinfo_x86 *c)
351{
352#ifdef CONFIG_SMP
353	unsigned bits;
354	int cpu = smp_processor_id();
355	unsigned int socket_id, core_complex_id;
356
357	bits = c->x86_coreid_bits;
358	/* Low order bits define the core id (index of core in socket) */
359	c->cpu_core_id = c->initial_apicid & ((1 << bits)-1);
360	/* Convert the initial APIC ID into the socket ID */
361	c->phys_proc_id = c->initial_apicid >> bits;
362	/* use socket ID also for last level cache */
363	per_cpu(cpu_llc_id, cpu) = c->phys_proc_id;
364	amd_get_topology(c);
365
366	/*
367	 * Fix percpu cpu_llc_id here as LLC topology is different
368	 * for Fam17h systems.
369	 */
370	 if (c->x86 != 0x17 || !cpuid_edx(0x80000006))
371		return;
372
373	socket_id	= (c->apicid >> bits) - 1;
374	core_complex_id	= (c->apicid & ((1 << bits) - 1)) >> 3;
375
376	per_cpu(cpu_llc_id, cpu) = (socket_id << 3) | core_complex_id;
377#endif
378}
379
380u16 amd_get_nb_id(int cpu)
381{
382	u16 id = 0;
383#ifdef CONFIG_SMP
384	id = per_cpu(cpu_llc_id, cpu);
385#endif
386	return id;
387}
388EXPORT_SYMBOL_GPL(amd_get_nb_id);
389
390u32 amd_get_nodes_per_socket(void)
391{
392	return nodes_per_socket;
393}
394EXPORT_SYMBOL_GPL(amd_get_nodes_per_socket);
395
396static void srat_detect_node(struct cpuinfo_x86 *c)
397{
398#ifdef CONFIG_NUMA
399	int cpu = smp_processor_id();
400	int node;
401	unsigned apicid = c->apicid;
402
403	node = numa_cpu_node(cpu);
404	if (node == NUMA_NO_NODE)
405		node = per_cpu(cpu_llc_id, cpu);
406
407	/*
408	 * On multi-fabric platform (e.g. Numascale NumaChip) a
409	 * platform-specific handler needs to be called to fixup some
410	 * IDs of the CPU.
411	 */
412	if (x86_cpuinit.fixup_cpu_id)
413		x86_cpuinit.fixup_cpu_id(c, node);
414
415	if (!node_online(node)) {
416		/*
417		 * Two possibilities here:
418		 *
419		 * - The CPU is missing memory and no node was created.  In
420		 *   that case try picking one from a nearby CPU.
421		 *
422		 * - The APIC IDs differ from the HyperTransport node IDs
423		 *   which the K8 northbridge parsing fills in.  Assume
424		 *   they are all increased by a constant offset, but in
425		 *   the same order as the HT nodeids.  If that doesn't
426		 *   result in a usable node fall back to the path for the
427		 *   previous case.
428		 *
429		 * This workaround operates directly on the mapping between
430		 * APIC ID and NUMA node, assuming certain relationship
431		 * between APIC ID, HT node ID and NUMA topology.  As going
432		 * through CPU mapping may alter the outcome, directly
433		 * access __apicid_to_node[].
434		 */
435		int ht_nodeid = c->initial_apicid;
436
437		if (ht_nodeid >= 0 &&
438		    __apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
439			node = __apicid_to_node[ht_nodeid];
440		/* Pick a nearby node */
441		if (!node_online(node))
442			node = nearby_node(apicid);
443	}
444	numa_set_node(cpu, node);
445#endif
446}
447
448static void early_init_amd_mc(struct cpuinfo_x86 *c)
449{
450#ifdef CONFIG_SMP
451	unsigned bits, ecx;
452
453	/* Multi core CPU? */
454	if (c->extended_cpuid_level < 0x80000008)
455		return;
456
457	ecx = cpuid_ecx(0x80000008);
458
459	c->x86_max_cores = (ecx & 0xff) + 1;
460
461	/* CPU telling us the core id bits shift? */
462	bits = (ecx >> 12) & 0xF;
463
464	/* Otherwise recompute */
465	if (bits == 0) {
466		while ((1 << bits) < c->x86_max_cores)
467			bits++;
468	}
469
470	c->x86_coreid_bits = bits;
471#endif
472}
473
474static void bsp_init_amd(struct cpuinfo_x86 *c)
475{
476
477#ifdef CONFIG_X86_64
478	if (c->x86 >= 0xf) {
479		unsigned long long tseg;
480
481		/*
482		 * Split up direct mapping around the TSEG SMM area.
483		 * Don't do it for gbpages because there seems very little
484		 * benefit in doing so.
485		 */
486		if (!rdmsrl_safe(MSR_K8_TSEG_ADDR, &tseg)) {
487			unsigned long pfn = tseg >> PAGE_SHIFT;
488
489			printk(KERN_DEBUG "tseg: %010llx\n", tseg);
490			if (pfn_range_is_mapped(pfn, pfn + 1))
491				set_memory_4k((unsigned long)__va(tseg), 1);
492		}
493	}
494#endif
495
496	if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) {
497
498		if (c->x86 > 0x10 ||
499		    (c->x86 == 0x10 && c->x86_model >= 0x2)) {
500			u64 val;
501
502			rdmsrl(MSR_K7_HWCR, val);
503			if (!(val & BIT(24)))
504				printk(KERN_WARNING FW_BUG "TSC doesn't count "
505					"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
528static void early_init_amd(struct cpuinfo_x86 *c)
529{
530	early_init_amd_mc(c);
531
532	/*
533	 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
534	 * with P/T states and does not stop in deep C-states
535	 */
536	if (c->x86_power & (1 << 8)) {
537		set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
538		set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
539		if (!check_tsc_unstable())
540			set_sched_clock_stable();
541	}
542
543#ifdef CONFIG_X86_64
544	set_cpu_cap(c, X86_FEATURE_SYSCALL32);
545#else
546	/*  Set MTRR capability flag if appropriate */
547	if (c->x86 == 5)
548		if (c->x86_model == 13 || c->x86_model == 9 ||
549		    (c->x86_model == 8 && c->x86_mask >= 8))
550			set_cpu_cap(c, X86_FEATURE_K6_MTRR);
551#endif
552#if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
553	/*
554	 * ApicID can always be treated as an 8-bit value for AMD APIC versions
555	 * >= 0x10, but even old K8s came out of reset with version 0x10. So, we
556	 * can safely set X86_FEATURE_EXTD_APICID unconditionally for families
557	 * after 16h.
558	 */
559	if (cpu_has_apic && c->x86 > 0x16) {
560		set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
561	} else if (cpu_has_apic && c->x86 >= 0xf) {
562		/* check CPU config space for extended APIC ID */
563		unsigned int val;
564		val = read_pci_config(0, 24, 0, 0x68);
565		if ((val & ((1 << 17) | (1 << 18))) == ((1 << 17) | (1 << 18)))
566			set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
567	}
568#endif
569
570	/*
571	 * This is only needed to tell the kernel whether to use VMCALL
572	 * and VMMCALL.  VMMCALL is never executed except under virt, so
573	 * we can set it unconditionally.
574	 */
575	set_cpu_cap(c, X86_FEATURE_VMMCALL);
576
577	/* F16h erratum 793, CVE-2013-6885 */
578	if (c->x86 == 0x16 && c->x86_model <= 0xf)
579		msr_set_bit(MSR_AMD64_LS_CFG, 15);
580}
581
582static const int amd_erratum_383[];
583static const int amd_erratum_400[];
584static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum);
585
586static void init_amd_k8(struct cpuinfo_x86 *c)
587{
588	u32 level;
589	u64 value;
590
591	/* On C+ stepping K8 rep microcode works well for copy/memset */
592	level = cpuid_eax(1);
593	if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58)
594		set_cpu_cap(c, X86_FEATURE_REP_GOOD);
595
596	/*
597	 * Some BIOSes incorrectly force this feature, but only K8 revision D
598	 * (model = 0x14) and later actually support it.
599	 * (AMD Erratum #110, docId: 25759).
600	 */
601	if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM)) {
602		clear_cpu_cap(c, X86_FEATURE_LAHF_LM);
603		if (!rdmsrl_amd_safe(0xc001100d, &value)) {
604			value &= ~BIT_64(32);
605			wrmsrl_amd_safe(0xc001100d, value);
606		}
607	}
608
609	if (!c->x86_model_id[0])
610		strcpy(c->x86_model_id, "Hammer");
611
612#ifdef CONFIG_SMP
613	/*
614	 * Disable TLB flush filter by setting HWCR.FFDIS on K8
615	 * bit 6 of msr C001_0015
616	 *
617	 * Errata 63 for SH-B3 steppings
618	 * Errata 122 for all steppings (F+ have it disabled by default)
619	 */
620	msr_set_bit(MSR_K7_HWCR, 6);
621#endif
622}
623
624static void init_amd_gh(struct cpuinfo_x86 *c)
625{
626#ifdef CONFIG_X86_64
627	/* do this for boot cpu */
628	if (c == &boot_cpu_data)
629		check_enable_amd_mmconf_dmi();
630
631	fam10h_check_enable_mmcfg();
632#endif
633
634	/*
635	 * Disable GART TLB Walk Errors on Fam10h. We do this here because this
636	 * is always needed when GART is enabled, even in a kernel which has no
637	 * MCE support built in. BIOS should disable GartTlbWlk Errors already.
638	 * If it doesn't, we do it here as suggested by the BKDG.
639	 *
640	 * Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=33012
641	 */
642	msr_set_bit(MSR_AMD64_MCx_MASK(4), 10);
643
644	/*
645	 * On family 10h BIOS may not have properly enabled WC+ support, causing
646	 * it to be converted to CD memtype. This may result in performance
647	 * degradation for certain nested-paging guests. Prevent this conversion
648	 * by clearing bit 24 in MSR_AMD64_BU_CFG2.
649	 *
650	 * NOTE: we want to use the _safe accessors so as not to #GP kvm
651	 * guests on older kvm hosts.
652	 */
653	msr_clear_bit(MSR_AMD64_BU_CFG2, 24);
654
655	if (cpu_has_amd_erratum(c, amd_erratum_383))
656		set_cpu_bug(c, X86_BUG_AMD_TLB_MMATCH);
657}
658
659static void init_amd_bd(struct cpuinfo_x86 *c)
660{
661	u64 value;
662
663	/* re-enable TopologyExtensions if switched off by BIOS */
664	if ((c->x86_model >= 0x10) && (c->x86_model <= 0x1f) &&
665	    !cpu_has(c, X86_FEATURE_TOPOEXT)) {
666
667		if (msr_set_bit(0xc0011005, 54) > 0) {
668			rdmsrl(0xc0011005, value);
669			if (value & BIT_64(54)) {
670				set_cpu_cap(c, X86_FEATURE_TOPOEXT);
671				pr_info(FW_INFO "CPU: Re-enabling disabled Topology Extensions Support.\n");
672			}
673		}
674	}
675
676	/*
677	 * The way access filter has a performance penalty on some workloads.
678	 * Disable it on the affected CPUs.
679	 */
680	if ((c->x86_model >= 0x02) && (c->x86_model < 0x20)) {
681		if (!rdmsrl_safe(0xc0011021, &value) && !(value & 0x1E)) {
682			value |= 0x1E;
683			wrmsrl_safe(0xc0011021, value);
684		}
685	}
686}
687
688static void init_amd(struct cpuinfo_x86 *c)
689{
690	u32 dummy;
691
692	early_init_amd(c);
693
694	/*
695	 * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
696	 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
697	 */
698	clear_cpu_cap(c, 0*32+31);
699
700	if (c->x86 >= 0x10)
701		set_cpu_cap(c, X86_FEATURE_REP_GOOD);
702
703	/* get apicid instead of initial apic id from cpuid */
704	c->apicid = hard_smp_processor_id();
705
706	/* K6s reports MCEs but don't actually have all the MSRs */
707	if (c->x86 < 6)
708		clear_cpu_cap(c, X86_FEATURE_MCE);
709
710	switch (c->x86) {
711	case 4:    init_amd_k5(c); break;
712	case 5:    init_amd_k6(c); break;
713	case 6:	   init_amd_k7(c); break;
714	case 0xf:  init_amd_k8(c); break;
715	case 0x10: init_amd_gh(c); break;
716	case 0x15: init_amd_bd(c); break;
717	}
718
719	/* Enable workaround for FXSAVE leak */
720	if (c->x86 >= 6)
721		set_cpu_bug(c, X86_BUG_FXSAVE_LEAK);
722
723	cpu_detect_cache_sizes(c);
724
725	/* Multi core CPU? */
726	if (c->extended_cpuid_level >= 0x80000008) {
727		amd_detect_cmp(c);
728		srat_detect_node(c);
729	}
730
731#ifdef CONFIG_X86_32
732	detect_ht(c);
733#endif
734
735	init_amd_cacheinfo(c);
736
737	if (c->x86 >= 0xf)
738		set_cpu_cap(c, X86_FEATURE_K8);
739
740	if (cpu_has_xmm2) {
741		/* MFENCE stops RDTSC speculation */
742		set_cpu_cap(c, X86_FEATURE_MFENCE_RDTSC);
743	}
744
745	/*
746	 * Family 0x12 and above processors have APIC timer
747	 * running in deep C states.
748	 */
749	if (c->x86 > 0x11)
750		set_cpu_cap(c, X86_FEATURE_ARAT);
751
752	if (cpu_has_amd_erratum(c, amd_erratum_400))
753		set_cpu_bug(c, X86_BUG_AMD_APIC_C1E);
754
755	rdmsr_safe(MSR_AMD64_PATCH_LEVEL, &c->microcode, &dummy);
756
757	/* 3DNow or LM implies PREFETCHW */
758	if (!cpu_has(c, X86_FEATURE_3DNOWPREFETCH))
759		if (cpu_has(c, X86_FEATURE_3DNOW) || cpu_has(c, X86_FEATURE_LM))
760			set_cpu_cap(c, X86_FEATURE_3DNOWPREFETCH);
761
762	/* AMD CPUs don't reset SS attributes on SYSRET */
763	set_cpu_bug(c, X86_BUG_SYSRET_SS_ATTRS);
764}
765
766#ifdef CONFIG_X86_32
767static unsigned int amd_size_cache(struct cpuinfo_x86 *c, unsigned int size)
768{
769	/* AMD errata T13 (order #21922) */
770	if ((c->x86 == 6)) {
771		/* Duron Rev A0 */
772		if (c->x86_model == 3 && c->x86_mask == 0)
773			size = 64;
774		/* Tbird rev A1/A2 */
775		if (c->x86_model == 4 &&
776			(c->x86_mask == 0 || c->x86_mask == 1))
777			size = 256;
778	}
779	return size;
780}
781#endif
782
783static void cpu_detect_tlb_amd(struct cpuinfo_x86 *c)
784{
785	u32 ebx, eax, ecx, edx;
786	u16 mask = 0xfff;
787
788	if (c->x86 < 0xf)
789		return;
790
791	if (c->extended_cpuid_level < 0x80000006)
792		return;
793
794	cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
795
796	tlb_lld_4k[ENTRIES] = (ebx >> 16) & mask;
797	tlb_lli_4k[ENTRIES] = ebx & mask;
798
799	/*
800	 * K8 doesn't have 2M/4M entries in the L2 TLB so read out the L1 TLB
801	 * characteristics from the CPUID function 0x80000005 instead.
802	 */
803	if (c->x86 == 0xf) {
804		cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
805		mask = 0xff;
806	}
807
808	/* Handle DTLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
809	if (!((eax >> 16) & mask))
810		tlb_lld_2m[ENTRIES] = (cpuid_eax(0x80000005) >> 16) & 0xff;
811	else
812		tlb_lld_2m[ENTRIES] = (eax >> 16) & mask;
813
814	/* a 4M entry uses two 2M entries */
815	tlb_lld_4m[ENTRIES] = tlb_lld_2m[ENTRIES] >> 1;
816
817	/* Handle ITLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
818	if (!(eax & mask)) {
819		/* Erratum 658 */
820		if (c->x86 == 0x15 && c->x86_model <= 0x1f) {
821			tlb_lli_2m[ENTRIES] = 1024;
822		} else {
823			cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
824			tlb_lli_2m[ENTRIES] = eax & 0xff;
825		}
826	} else
827		tlb_lli_2m[ENTRIES] = eax & mask;
828
829	tlb_lli_4m[ENTRIES] = tlb_lli_2m[ENTRIES] >> 1;
830}
831
832static const struct cpu_dev amd_cpu_dev = {
833	.c_vendor	= "AMD",
834	.c_ident	= { "AuthenticAMD" },
835#ifdef CONFIG_X86_32
836	.legacy_models = {
837		{ .family = 4, .model_names =
838		  {
839			  [3] = "486 DX/2",
840			  [7] = "486 DX/2-WB",
841			  [8] = "486 DX/4",
842			  [9] = "486 DX/4-WB",
843			  [14] = "Am5x86-WT",
844			  [15] = "Am5x86-WB"
845		  }
846		},
847	},
848	.legacy_cache_size = amd_size_cache,
849#endif
850	.c_early_init   = early_init_amd,
851	.c_detect_tlb	= cpu_detect_tlb_amd,
852	.c_bsp_init	= bsp_init_amd,
853	.c_init		= init_amd,
854	.c_x86_vendor	= X86_VENDOR_AMD,
855};
856
857cpu_dev_register(amd_cpu_dev);
858
859/*
860 * AMD errata checking
861 *
862 * Errata are defined as arrays of ints using the AMD_LEGACY_ERRATUM() or
863 * AMD_OSVW_ERRATUM() macros. The latter is intended for newer errata that
864 * have an OSVW id assigned, which it takes as first argument. Both take a
865 * variable number of family-specific model-stepping ranges created by
866 * AMD_MODEL_RANGE().
867 *
868 * Example:
869 *
870 * const int amd_erratum_319[] =
871 *	AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0x4, 0x2),
872 *			   AMD_MODEL_RANGE(0x10, 0x8, 0x0, 0x8, 0x0),
873 *			   AMD_MODEL_RANGE(0x10, 0x9, 0x0, 0x9, 0x0));
874 */
875
876#define AMD_LEGACY_ERRATUM(...)		{ -1, __VA_ARGS__, 0 }
877#define AMD_OSVW_ERRATUM(osvw_id, ...)	{ osvw_id, __VA_ARGS__, 0 }
878#define AMD_MODEL_RANGE(f, m_start, s_start, m_end, s_end) \
879	((f << 24) | (m_start << 16) | (s_start << 12) | (m_end << 4) | (s_end))
880#define AMD_MODEL_RANGE_FAMILY(range)	(((range) >> 24) & 0xff)
881#define AMD_MODEL_RANGE_START(range)	(((range) >> 12) & 0xfff)
882#define AMD_MODEL_RANGE_END(range)	((range) & 0xfff)
883
884static const int amd_erratum_400[] =
885	AMD_OSVW_ERRATUM(1, AMD_MODEL_RANGE(0xf, 0x41, 0x2, 0xff, 0xf),
886			    AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0xff, 0xf));
887
888static const int amd_erratum_383[] =
889	AMD_OSVW_ERRATUM(3, AMD_MODEL_RANGE(0x10, 0, 0, 0xff, 0xf));
890
891
892static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum)
893{
894	int osvw_id = *erratum++;
895	u32 range;
896	u32 ms;
897
898	if (osvw_id >= 0 && osvw_id < 65536 &&
899	    cpu_has(cpu, X86_FEATURE_OSVW)) {
900		u64 osvw_len;
901
902		rdmsrl(MSR_AMD64_OSVW_ID_LENGTH, osvw_len);
903		if (osvw_id < osvw_len) {
904			u64 osvw_bits;
905
906			rdmsrl(MSR_AMD64_OSVW_STATUS + (osvw_id >> 6),
907			    osvw_bits);
908			return osvw_bits & (1ULL << (osvw_id & 0x3f));
909		}
910	}
911
912	/* OSVW unavailable or ID unknown, match family-model-stepping range */
913	ms = (cpu->x86_model << 4) | cpu->x86_mask;
914	while ((range = *erratum++))
915		if ((cpu->x86 == AMD_MODEL_RANGE_FAMILY(range)) &&
916		    (ms >= AMD_MODEL_RANGE_START(range)) &&
917		    (ms <= AMD_MODEL_RANGE_END(range)))
918			return true;
919
920	return false;
921}
922
923void set_dr_addr_mask(unsigned long mask, int dr)
924{
925	if (!cpu_has_bpext)
926		return;
927
928	switch (dr) {
929	case 0:
930		wrmsr(MSR_F16H_DR0_ADDR_MASK, mask, 0);
931		break;
932	case 1:
933	case 2:
934	case 3:
935		wrmsr(MSR_F16H_DR1_ADDR_MASK - 1 + dr, mask, 0);
936		break;
937	default:
938		break;
939	}
940}
941