xref: /openbmc/linux/arch/x86/kernel/cpu/amd.c (revision 8730046c)
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 
313 		node_id = cpuid_ecx(0x8000001e) & 7;
314 
315 		/*
316 		 * We may have multiple LLCs if L3 caches exist, so check if we
317 		 * have an L3 cache by looking at the L3 cache CPUID leaf.
318 		 */
319 		if (cpuid_edx(0x80000006)) {
320 			if (c->x86 == 0x17) {
321 				/*
322 				 * LLC is at the core complex level.
323 				 * Core complex id is ApicId[3].
324 				 */
325 				per_cpu(cpu_llc_id, cpu) = c->apicid >> 3;
326 			} else {
327 				/* LLC is at the node level. */
328 				per_cpu(cpu_llc_id, cpu) = node_id;
329 			}
330 		}
331 	} else if (cpu_has(c, X86_FEATURE_NODEID_MSR)) {
332 		u64 value;
333 
334 		rdmsrl(MSR_FAM10H_NODE_ID, value);
335 		node_id = value & 7;
336 
337 		per_cpu(cpu_llc_id, cpu) = node_id;
338 	} else
339 		return;
340 
341 	/* fixup multi-node processor information */
342 	if (nodes_per_socket > 1) {
343 		u32 cus_per_node;
344 
345 		set_cpu_cap(c, X86_FEATURE_AMD_DCM);
346 		cus_per_node = c->x86_max_cores / nodes_per_socket;
347 
348 		/* core id has to be in the [0 .. cores_per_node - 1] range */
349 		c->cpu_core_id %= cus_per_node;
350 	}
351 }
352 #endif
353 
354 /*
355  * On a AMD dual core setup the lower bits of the APIC id distinguish the cores.
356  * Assumes number of cores is a power of two.
357  */
358 static void amd_detect_cmp(struct cpuinfo_x86 *c)
359 {
360 #ifdef CONFIG_SMP
361 	unsigned bits;
362 	int cpu = smp_processor_id();
363 
364 	bits = c->x86_coreid_bits;
365 	/* Low order bits define the core id (index of core in socket) */
366 	c->cpu_core_id = c->initial_apicid & ((1 << bits)-1);
367 	/* Convert the initial APIC ID into the socket ID */
368 	c->phys_proc_id = c->initial_apicid >> bits;
369 	/* use socket ID also for last level cache */
370 	per_cpu(cpu_llc_id, cpu) = c->phys_proc_id;
371 	amd_get_topology(c);
372 #endif
373 }
374 
375 u16 amd_get_nb_id(int cpu)
376 {
377 	u16 id = 0;
378 #ifdef CONFIG_SMP
379 	id = per_cpu(cpu_llc_id, cpu);
380 #endif
381 	return id;
382 }
383 EXPORT_SYMBOL_GPL(amd_get_nb_id);
384 
385 u32 amd_get_nodes_per_socket(void)
386 {
387 	return nodes_per_socket;
388 }
389 EXPORT_SYMBOL_GPL(amd_get_nodes_per_socket);
390 
391 static void srat_detect_node(struct cpuinfo_x86 *c)
392 {
393 #ifdef CONFIG_NUMA
394 	int cpu = smp_processor_id();
395 	int node;
396 	unsigned apicid = c->apicid;
397 
398 	node = numa_cpu_node(cpu);
399 	if (node == NUMA_NO_NODE)
400 		node = per_cpu(cpu_llc_id, cpu);
401 
402 	/*
403 	 * On multi-fabric platform (e.g. Numascale NumaChip) a
404 	 * platform-specific handler needs to be called to fixup some
405 	 * IDs of the CPU.
406 	 */
407 	if (x86_cpuinit.fixup_cpu_id)
408 		x86_cpuinit.fixup_cpu_id(c, node);
409 
410 	if (!node_online(node)) {
411 		/*
412 		 * Two possibilities here:
413 		 *
414 		 * - The CPU is missing memory and no node was created.  In
415 		 *   that case try picking one from a nearby CPU.
416 		 *
417 		 * - The APIC IDs differ from the HyperTransport node IDs
418 		 *   which the K8 northbridge parsing fills in.  Assume
419 		 *   they are all increased by a constant offset, but in
420 		 *   the same order as the HT nodeids.  If that doesn't
421 		 *   result in a usable node fall back to the path for the
422 		 *   previous case.
423 		 *
424 		 * This workaround operates directly on the mapping between
425 		 * APIC ID and NUMA node, assuming certain relationship
426 		 * between APIC ID, HT node ID and NUMA topology.  As going
427 		 * through CPU mapping may alter the outcome, directly
428 		 * access __apicid_to_node[].
429 		 */
430 		int ht_nodeid = c->initial_apicid;
431 
432 		if (__apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
433 			node = __apicid_to_node[ht_nodeid];
434 		/* Pick a nearby node */
435 		if (!node_online(node))
436 			node = nearby_node(apicid);
437 	}
438 	numa_set_node(cpu, node);
439 #endif
440 }
441 
442 static void early_init_amd_mc(struct cpuinfo_x86 *c)
443 {
444 #ifdef CONFIG_SMP
445 	unsigned bits, ecx;
446 
447 	/* Multi core CPU? */
448 	if (c->extended_cpuid_level < 0x80000008)
449 		return;
450 
451 	ecx = cpuid_ecx(0x80000008);
452 
453 	c->x86_max_cores = (ecx & 0xff) + 1;
454 
455 	/* CPU telling us the core id bits shift? */
456 	bits = (ecx >> 12) & 0xF;
457 
458 	/* Otherwise recompute */
459 	if (bits == 0) {
460 		while ((1 << bits) < c->x86_max_cores)
461 			bits++;
462 	}
463 
464 	c->x86_coreid_bits = bits;
465 #endif
466 }
467 
468 static void bsp_init_amd(struct cpuinfo_x86 *c)
469 {
470 
471 #ifdef CONFIG_X86_64
472 	if (c->x86 >= 0xf) {
473 		unsigned long long tseg;
474 
475 		/*
476 		 * Split up direct mapping around the TSEG SMM area.
477 		 * Don't do it for gbpages because there seems very little
478 		 * benefit in doing so.
479 		 */
480 		if (!rdmsrl_safe(MSR_K8_TSEG_ADDR, &tseg)) {
481 			unsigned long pfn = tseg >> PAGE_SHIFT;
482 
483 			pr_debug("tseg: %010llx\n", tseg);
484 			if (pfn_range_is_mapped(pfn, pfn + 1))
485 				set_memory_4k((unsigned long)__va(tseg), 1);
486 		}
487 	}
488 #endif
489 
490 	if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) {
491 
492 		if (c->x86 > 0x10 ||
493 		    (c->x86 == 0x10 && c->x86_model >= 0x2)) {
494 			u64 val;
495 
496 			rdmsrl(MSR_K7_HWCR, val);
497 			if (!(val & BIT(24)))
498 				pr_warn(FW_BUG "TSC doesn't count with P0 frequency!\n");
499 		}
500 	}
501 
502 	if (c->x86 == 0x15) {
503 		unsigned long upperbit;
504 		u32 cpuid, assoc;
505 
506 		cpuid	 = cpuid_edx(0x80000005);
507 		assoc	 = cpuid >> 16 & 0xff;
508 		upperbit = ((cpuid >> 24) << 10) / assoc;
509 
510 		va_align.mask	  = (upperbit - 1) & PAGE_MASK;
511 		va_align.flags    = ALIGN_VA_32 | ALIGN_VA_64;
512 
513 		/* A random value per boot for bit slice [12:upper_bit) */
514 		va_align.bits = get_random_int() & va_align.mask;
515 	}
516 
517 	if (cpu_has(c, X86_FEATURE_MWAITX))
518 		use_mwaitx_delay();
519 
520 	if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
521 		u32 ecx;
522 
523 		ecx = cpuid_ecx(0x8000001e);
524 		nodes_per_socket = ((ecx >> 8) & 7) + 1;
525 	} else if (boot_cpu_has(X86_FEATURE_NODEID_MSR)) {
526 		u64 value;
527 
528 		rdmsrl(MSR_FAM10H_NODE_ID, value);
529 		nodes_per_socket = ((value >> 3) & 7) + 1;
530 	}
531 }
532 
533 static void early_init_amd(struct cpuinfo_x86 *c)
534 {
535 	early_init_amd_mc(c);
536 
537 	/*
538 	 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
539 	 * with P/T states and does not stop in deep C-states
540 	 */
541 	if (c->x86_power & (1 << 8)) {
542 		set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
543 		set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
544 		if (!check_tsc_unstable())
545 			set_sched_clock_stable();
546 	}
547 
548 	/* Bit 12 of 8000_0007 edx is accumulated power mechanism. */
549 	if (c->x86_power & BIT(12))
550 		set_cpu_cap(c, X86_FEATURE_ACC_POWER);
551 
552 #ifdef CONFIG_X86_64
553 	set_cpu_cap(c, X86_FEATURE_SYSCALL32);
554 #else
555 	/*  Set MTRR capability flag if appropriate */
556 	if (c->x86 == 5)
557 		if (c->x86_model == 13 || c->x86_model == 9 ||
558 		    (c->x86_model == 8 && c->x86_mask >= 8))
559 			set_cpu_cap(c, X86_FEATURE_K6_MTRR);
560 #endif
561 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
562 	/*
563 	 * ApicID can always be treated as an 8-bit value for AMD APIC versions
564 	 * >= 0x10, but even old K8s came out of reset with version 0x10. So, we
565 	 * can safely set X86_FEATURE_EXTD_APICID unconditionally for families
566 	 * after 16h.
567 	 */
568 	if (boot_cpu_has(X86_FEATURE_APIC)) {
569 		if (c->x86 > 0x16)
570 			set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
571 		else if (c->x86 >= 0xf) {
572 			/* check CPU config space for extended APIC ID */
573 			unsigned int val;
574 
575 			val = read_pci_config(0, 24, 0, 0x68);
576 			if ((val >> 17 & 0x3) == 0x3)
577 				set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
578 		}
579 	}
580 #endif
581 
582 	/*
583 	 * This is only needed to tell the kernel whether to use VMCALL
584 	 * and VMMCALL.  VMMCALL is never executed except under virt, so
585 	 * we can set it unconditionally.
586 	 */
587 	set_cpu_cap(c, X86_FEATURE_VMMCALL);
588 
589 	/* F16h erratum 793, CVE-2013-6885 */
590 	if (c->x86 == 0x16 && c->x86_model <= 0xf)
591 		msr_set_bit(MSR_AMD64_LS_CFG, 15);
592 
593 	/*
594 	 * Check whether the machine is affected by erratum 400. This is
595 	 * used to select the proper idle routine and to enable the check
596 	 * whether the machine is affected in arch_post_acpi_init(), which
597 	 * sets the X86_BUG_AMD_APIC_C1E bug depending on the MSR check.
598 	 */
599 	if (cpu_has_amd_erratum(c, amd_erratum_400))
600 		set_cpu_bug(c, X86_BUG_AMD_E400);
601 }
602 
603 static void init_amd_k8(struct cpuinfo_x86 *c)
604 {
605 	u32 level;
606 	u64 value;
607 
608 	/* On C+ stepping K8 rep microcode works well for copy/memset */
609 	level = cpuid_eax(1);
610 	if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58)
611 		set_cpu_cap(c, X86_FEATURE_REP_GOOD);
612 
613 	/*
614 	 * Some BIOSes incorrectly force this feature, but only K8 revision D
615 	 * (model = 0x14) and later actually support it.
616 	 * (AMD Erratum #110, docId: 25759).
617 	 */
618 	if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM)) {
619 		clear_cpu_cap(c, X86_FEATURE_LAHF_LM);
620 		if (!rdmsrl_amd_safe(0xc001100d, &value)) {
621 			value &= ~BIT_64(32);
622 			wrmsrl_amd_safe(0xc001100d, value);
623 		}
624 	}
625 
626 	if (!c->x86_model_id[0])
627 		strcpy(c->x86_model_id, "Hammer");
628 
629 #ifdef CONFIG_SMP
630 	/*
631 	 * Disable TLB flush filter by setting HWCR.FFDIS on K8
632 	 * bit 6 of msr C001_0015
633 	 *
634 	 * Errata 63 for SH-B3 steppings
635 	 * Errata 122 for all steppings (F+ have it disabled by default)
636 	 */
637 	msr_set_bit(MSR_K7_HWCR, 6);
638 #endif
639 	set_cpu_bug(c, X86_BUG_SWAPGS_FENCE);
640 }
641 
642 static void init_amd_gh(struct cpuinfo_x86 *c)
643 {
644 #ifdef CONFIG_X86_64
645 	/* do this for boot cpu */
646 	if (c == &boot_cpu_data)
647 		check_enable_amd_mmconf_dmi();
648 
649 	fam10h_check_enable_mmcfg();
650 #endif
651 
652 	/*
653 	 * Disable GART TLB Walk Errors on Fam10h. We do this here because this
654 	 * is always needed when GART is enabled, even in a kernel which has no
655 	 * MCE support built in. BIOS should disable GartTlbWlk Errors already.
656 	 * If it doesn't, we do it here as suggested by the BKDG.
657 	 *
658 	 * Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=33012
659 	 */
660 	msr_set_bit(MSR_AMD64_MCx_MASK(4), 10);
661 
662 	/*
663 	 * On family 10h BIOS may not have properly enabled WC+ support, causing
664 	 * it to be converted to CD memtype. This may result in performance
665 	 * degradation for certain nested-paging guests. Prevent this conversion
666 	 * by clearing bit 24 in MSR_AMD64_BU_CFG2.
667 	 *
668 	 * NOTE: we want to use the _safe accessors so as not to #GP kvm
669 	 * guests on older kvm hosts.
670 	 */
671 	msr_clear_bit(MSR_AMD64_BU_CFG2, 24);
672 
673 	if (cpu_has_amd_erratum(c, amd_erratum_383))
674 		set_cpu_bug(c, X86_BUG_AMD_TLB_MMATCH);
675 }
676 
677 #define MSR_AMD64_DE_CFG	0xC0011029
678 
679 static void init_amd_ln(struct cpuinfo_x86 *c)
680 {
681 	/*
682 	 * Apply erratum 665 fix unconditionally so machines without a BIOS
683 	 * fix work.
684 	 */
685 	msr_set_bit(MSR_AMD64_DE_CFG, 31);
686 }
687 
688 static void init_amd_bd(struct cpuinfo_x86 *c)
689 {
690 	u64 value;
691 
692 	/* re-enable TopologyExtensions if switched off by BIOS */
693 	if ((c->x86_model >= 0x10) && (c->x86_model <= 0x6f) &&
694 	    !cpu_has(c, X86_FEATURE_TOPOEXT)) {
695 
696 		if (msr_set_bit(0xc0011005, 54) > 0) {
697 			rdmsrl(0xc0011005, value);
698 			if (value & BIT_64(54)) {
699 				set_cpu_cap(c, X86_FEATURE_TOPOEXT);
700 				pr_info_once(FW_INFO "CPU: Re-enabling disabled Topology Extensions Support.\n");
701 			}
702 		}
703 	}
704 
705 	/*
706 	 * The way access filter has a performance penalty on some workloads.
707 	 * Disable it on the affected CPUs.
708 	 */
709 	if ((c->x86_model >= 0x02) && (c->x86_model < 0x20)) {
710 		if (!rdmsrl_safe(MSR_F15H_IC_CFG, &value) && !(value & 0x1E)) {
711 			value |= 0x1E;
712 			wrmsrl_safe(MSR_F15H_IC_CFG, value);
713 		}
714 	}
715 }
716 
717 static void init_amd(struct cpuinfo_x86 *c)
718 {
719 	u32 dummy;
720 
721 	early_init_amd(c);
722 
723 	/*
724 	 * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
725 	 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
726 	 */
727 	clear_cpu_cap(c, 0*32+31);
728 
729 	if (c->x86 >= 0x10)
730 		set_cpu_cap(c, X86_FEATURE_REP_GOOD);
731 
732 	/* get apicid instead of initial apic id from cpuid */
733 	c->apicid = hard_smp_processor_id();
734 
735 	/* K6s reports MCEs but don't actually have all the MSRs */
736 	if (c->x86 < 6)
737 		clear_cpu_cap(c, X86_FEATURE_MCE);
738 
739 	switch (c->x86) {
740 	case 4:    init_amd_k5(c); break;
741 	case 5:    init_amd_k6(c); break;
742 	case 6:	   init_amd_k7(c); break;
743 	case 0xf:  init_amd_k8(c); break;
744 	case 0x10: init_amd_gh(c); break;
745 	case 0x12: init_amd_ln(c); break;
746 	case 0x15: init_amd_bd(c); break;
747 	}
748 
749 	/* Enable workaround for FXSAVE leak */
750 	if (c->x86 >= 6)
751 		set_cpu_bug(c, X86_BUG_FXSAVE_LEAK);
752 
753 	cpu_detect_cache_sizes(c);
754 
755 	/* Multi core CPU? */
756 	if (c->extended_cpuid_level >= 0x80000008) {
757 		amd_detect_cmp(c);
758 		srat_detect_node(c);
759 	}
760 
761 #ifdef CONFIG_X86_32
762 	detect_ht(c);
763 #endif
764 
765 	init_amd_cacheinfo(c);
766 
767 	if (c->x86 >= 0xf)
768 		set_cpu_cap(c, X86_FEATURE_K8);
769 
770 	if (cpu_has(c, X86_FEATURE_XMM2)) {
771 		/* MFENCE stops RDTSC speculation */
772 		set_cpu_cap(c, X86_FEATURE_MFENCE_RDTSC);
773 	}
774 
775 	/*
776 	 * Family 0x12 and above processors have APIC timer
777 	 * running in deep C states.
778 	 */
779 	if (c->x86 > 0x11)
780 		set_cpu_cap(c, X86_FEATURE_ARAT);
781 
782 	rdmsr_safe(MSR_AMD64_PATCH_LEVEL, &c->microcode, &dummy);
783 
784 	/* 3DNow or LM implies PREFETCHW */
785 	if (!cpu_has(c, X86_FEATURE_3DNOWPREFETCH))
786 		if (cpu_has(c, X86_FEATURE_3DNOW) || cpu_has(c, X86_FEATURE_LM))
787 			set_cpu_cap(c, X86_FEATURE_3DNOWPREFETCH);
788 
789 	/* AMD CPUs don't reset SS attributes on SYSRET */
790 	set_cpu_bug(c, X86_BUG_SYSRET_SS_ATTRS);
791 }
792 
793 #ifdef CONFIG_X86_32
794 static unsigned int amd_size_cache(struct cpuinfo_x86 *c, unsigned int size)
795 {
796 	/* AMD errata T13 (order #21922) */
797 	if ((c->x86 == 6)) {
798 		/* Duron Rev A0 */
799 		if (c->x86_model == 3 && c->x86_mask == 0)
800 			size = 64;
801 		/* Tbird rev A1/A2 */
802 		if (c->x86_model == 4 &&
803 			(c->x86_mask == 0 || c->x86_mask == 1))
804 			size = 256;
805 	}
806 	return size;
807 }
808 #endif
809 
810 static void cpu_detect_tlb_amd(struct cpuinfo_x86 *c)
811 {
812 	u32 ebx, eax, ecx, edx;
813 	u16 mask = 0xfff;
814 
815 	if (c->x86 < 0xf)
816 		return;
817 
818 	if (c->extended_cpuid_level < 0x80000006)
819 		return;
820 
821 	cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
822 
823 	tlb_lld_4k[ENTRIES] = (ebx >> 16) & mask;
824 	tlb_lli_4k[ENTRIES] = ebx & mask;
825 
826 	/*
827 	 * K8 doesn't have 2M/4M entries in the L2 TLB so read out the L1 TLB
828 	 * characteristics from the CPUID function 0x80000005 instead.
829 	 */
830 	if (c->x86 == 0xf) {
831 		cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
832 		mask = 0xff;
833 	}
834 
835 	/* Handle DTLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
836 	if (!((eax >> 16) & mask))
837 		tlb_lld_2m[ENTRIES] = (cpuid_eax(0x80000005) >> 16) & 0xff;
838 	else
839 		tlb_lld_2m[ENTRIES] = (eax >> 16) & mask;
840 
841 	/* a 4M entry uses two 2M entries */
842 	tlb_lld_4m[ENTRIES] = tlb_lld_2m[ENTRIES] >> 1;
843 
844 	/* Handle ITLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
845 	if (!(eax & mask)) {
846 		/* Erratum 658 */
847 		if (c->x86 == 0x15 && c->x86_model <= 0x1f) {
848 			tlb_lli_2m[ENTRIES] = 1024;
849 		} else {
850 			cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
851 			tlb_lli_2m[ENTRIES] = eax & 0xff;
852 		}
853 	} else
854 		tlb_lli_2m[ENTRIES] = eax & mask;
855 
856 	tlb_lli_4m[ENTRIES] = tlb_lli_2m[ENTRIES] >> 1;
857 }
858 
859 static const struct cpu_dev amd_cpu_dev = {
860 	.c_vendor	= "AMD",
861 	.c_ident	= { "AuthenticAMD" },
862 #ifdef CONFIG_X86_32
863 	.legacy_models = {
864 		{ .family = 4, .model_names =
865 		  {
866 			  [3] = "486 DX/2",
867 			  [7] = "486 DX/2-WB",
868 			  [8] = "486 DX/4",
869 			  [9] = "486 DX/4-WB",
870 			  [14] = "Am5x86-WT",
871 			  [15] = "Am5x86-WB"
872 		  }
873 		},
874 	},
875 	.legacy_cache_size = amd_size_cache,
876 #endif
877 	.c_early_init   = early_init_amd,
878 	.c_detect_tlb	= cpu_detect_tlb_amd,
879 	.c_bsp_init	= bsp_init_amd,
880 	.c_init		= init_amd,
881 	.c_x86_vendor	= X86_VENDOR_AMD,
882 };
883 
884 cpu_dev_register(amd_cpu_dev);
885 
886 /*
887  * AMD errata checking
888  *
889  * Errata are defined as arrays of ints using the AMD_LEGACY_ERRATUM() or
890  * AMD_OSVW_ERRATUM() macros. The latter is intended for newer errata that
891  * have an OSVW id assigned, which it takes as first argument. Both take a
892  * variable number of family-specific model-stepping ranges created by
893  * AMD_MODEL_RANGE().
894  *
895  * Example:
896  *
897  * const int amd_erratum_319[] =
898  *	AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0x4, 0x2),
899  *			   AMD_MODEL_RANGE(0x10, 0x8, 0x0, 0x8, 0x0),
900  *			   AMD_MODEL_RANGE(0x10, 0x9, 0x0, 0x9, 0x0));
901  */
902 
903 #define AMD_LEGACY_ERRATUM(...)		{ -1, __VA_ARGS__, 0 }
904 #define AMD_OSVW_ERRATUM(osvw_id, ...)	{ osvw_id, __VA_ARGS__, 0 }
905 #define AMD_MODEL_RANGE(f, m_start, s_start, m_end, s_end) \
906 	((f << 24) | (m_start << 16) | (s_start << 12) | (m_end << 4) | (s_end))
907 #define AMD_MODEL_RANGE_FAMILY(range)	(((range) >> 24) & 0xff)
908 #define AMD_MODEL_RANGE_START(range)	(((range) >> 12) & 0xfff)
909 #define AMD_MODEL_RANGE_END(range)	((range) & 0xfff)
910 
911 static const int amd_erratum_400[] =
912 	AMD_OSVW_ERRATUM(1, AMD_MODEL_RANGE(0xf, 0x41, 0x2, 0xff, 0xf),
913 			    AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0xff, 0xf));
914 
915 static const int amd_erratum_383[] =
916 	AMD_OSVW_ERRATUM(3, AMD_MODEL_RANGE(0x10, 0, 0, 0xff, 0xf));
917 
918 
919 static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum)
920 {
921 	int osvw_id = *erratum++;
922 	u32 range;
923 	u32 ms;
924 
925 	if (osvw_id >= 0 && osvw_id < 65536 &&
926 	    cpu_has(cpu, X86_FEATURE_OSVW)) {
927 		u64 osvw_len;
928 
929 		rdmsrl(MSR_AMD64_OSVW_ID_LENGTH, osvw_len);
930 		if (osvw_id < osvw_len) {
931 			u64 osvw_bits;
932 
933 			rdmsrl(MSR_AMD64_OSVW_STATUS + (osvw_id >> 6),
934 			    osvw_bits);
935 			return osvw_bits & (1ULL << (osvw_id & 0x3f));
936 		}
937 	}
938 
939 	/* OSVW unavailable or ID unknown, match family-model-stepping range */
940 	ms = (cpu->x86_model << 4) | cpu->x86_mask;
941 	while ((range = *erratum++))
942 		if ((cpu->x86 == AMD_MODEL_RANGE_FAMILY(range)) &&
943 		    (ms >= AMD_MODEL_RANGE_START(range)) &&
944 		    (ms <= AMD_MODEL_RANGE_END(range)))
945 			return true;
946 
947 	return false;
948 }
949 
950 void set_dr_addr_mask(unsigned long mask, int dr)
951 {
952 	if (!boot_cpu_has(X86_FEATURE_BPEXT))
953 		return;
954 
955 	switch (dr) {
956 	case 0:
957 		wrmsr(MSR_F16H_DR0_ADDR_MASK, mask, 0);
958 		break;
959 	case 1:
960 	case 2:
961 	case 3:
962 		wrmsr(MSR_F16H_DR1_ADDR_MASK - 1 + dr, mask, 0);
963 		break;
964 	default:
965 		break;
966 	}
967 }
968