xref: /openbmc/linux/arch/x86/kernel/cpu/amd.c (revision fd589a8f)
1 #include <linux/init.h>
2 #include <linux/bitops.h>
3 #include <linux/mm.h>
4 
5 #include <linux/io.h>
6 #include <asm/processor.h>
7 #include <asm/apic.h>
8 #include <asm/cpu.h>
9 #include <asm/pci-direct.h>
10 
11 #ifdef CONFIG_X86_64
12 # include <asm/numa_64.h>
13 # include <asm/mmconfig.h>
14 # include <asm/cacheflush.h>
15 #endif
16 
17 #include "cpu.h"
18 
19 #ifdef CONFIG_X86_32
20 /*
21  *	B step AMD K6 before B 9730xxxx have hardware bugs that can cause
22  *	misexecution of code under Linux. Owners of such processors should
23  *	contact AMD for precise details and a CPU swap.
24  *
25  *	See	http://www.multimania.com/poulot/k6bug.html
26  *		http://www.amd.com/K6/k6docs/revgd.html
27  *
28  *	The following test is erm.. interesting. AMD neglected to up
29  *	the chip setting when fixing the bug but they also tweaked some
30  *	performance at the same time..
31  */
32 
33 extern void vide(void);
34 __asm__(".align 4\nvide: ret");
35 
36 static void __cpuinit init_amd_k5(struct cpuinfo_x86 *c)
37 {
38 /*
39  * General Systems BIOSen alias the cpu frequency registers
40  * of the Elan at 0x000df000. Unfortuantly, one of the Linux
41  * drivers subsequently pokes it, and changes the CPU speed.
42  * Workaround : Remove the unneeded alias.
43  */
44 #define CBAR		(0xfffc) /* Configuration Base Address  (32-bit) */
45 #define CBAR_ENB	(0x80000000)
46 #define CBAR_KEY	(0X000000CB)
47 	if (c->x86_model == 9 || c->x86_model == 10) {
48 		if (inl(CBAR) & CBAR_ENB)
49 			outl(0 | CBAR_KEY, CBAR);
50 	}
51 }
52 
53 
54 static void __cpuinit init_amd_k6(struct cpuinfo_x86 *c)
55 {
56 	u32 l, h;
57 	int mbytes = num_physpages >> (20-PAGE_SHIFT);
58 
59 	if (c->x86_model < 6) {
60 		/* Based on AMD doc 20734R - June 2000 */
61 		if (c->x86_model == 0) {
62 			clear_cpu_cap(c, X86_FEATURE_APIC);
63 			set_cpu_cap(c, X86_FEATURE_PGE);
64 		}
65 		return;
66 	}
67 
68 	if (c->x86_model == 6 && c->x86_mask == 1) {
69 		const int K6_BUG_LOOP = 1000000;
70 		int n;
71 		void (*f_vide)(void);
72 		unsigned long d, d2;
73 
74 		printk(KERN_INFO "AMD K6 stepping B detected - ");
75 
76 		/*
77 		 * It looks like AMD fixed the 2.6.2 bug and improved indirect
78 		 * calls at the same time.
79 		 */
80 
81 		n = K6_BUG_LOOP;
82 		f_vide = vide;
83 		rdtscl(d);
84 		while (n--)
85 			f_vide();
86 		rdtscl(d2);
87 		d = d2-d;
88 
89 		if (d > 20*K6_BUG_LOOP)
90 			printk(KERN_CONT
91 				"system stability may be impaired when more than 32 MB are used.\n");
92 		else
93 			printk(KERN_CONT "probably OK (after B9730xxxx).\n");
94 		printk(KERN_INFO "Please see http://membres.lycos.fr/poulot/k6bug.html\n");
95 	}
96 
97 	/* K6 with old style WHCR */
98 	if (c->x86_model < 8 ||
99 	   (c->x86_model == 8 && c->x86_mask < 8)) {
100 		/* We can only write allocate on the low 508Mb */
101 		if (mbytes > 508)
102 			mbytes = 508;
103 
104 		rdmsr(MSR_K6_WHCR, l, h);
105 		if ((l&0x0000FFFF) == 0) {
106 			unsigned long flags;
107 			l = (1<<0)|((mbytes/4)<<1);
108 			local_irq_save(flags);
109 			wbinvd();
110 			wrmsr(MSR_K6_WHCR, l, h);
111 			local_irq_restore(flags);
112 			printk(KERN_INFO "Enabling old style K6 write allocation for %d Mb\n",
113 				mbytes);
114 		}
115 		return;
116 	}
117 
118 	if ((c->x86_model == 8 && c->x86_mask > 7) ||
119 	     c->x86_model == 9 || c->x86_model == 13) {
120 		/* The more serious chips .. */
121 
122 		if (mbytes > 4092)
123 			mbytes = 4092;
124 
125 		rdmsr(MSR_K6_WHCR, l, h);
126 		if ((l&0xFFFF0000) == 0) {
127 			unsigned long flags;
128 			l = ((mbytes>>2)<<22)|(1<<16);
129 			local_irq_save(flags);
130 			wbinvd();
131 			wrmsr(MSR_K6_WHCR, l, h);
132 			local_irq_restore(flags);
133 			printk(KERN_INFO "Enabling new style K6 write allocation for %d Mb\n",
134 				mbytes);
135 		}
136 
137 		return;
138 	}
139 
140 	if (c->x86_model == 10) {
141 		/* AMD Geode LX is model 10 */
142 		/* placeholder for any needed mods */
143 		return;
144 	}
145 }
146 
147 static void __cpuinit amd_k7_smp_check(struct cpuinfo_x86 *c)
148 {
149 #ifdef CONFIG_SMP
150 	/* calling is from identify_secondary_cpu() ? */
151 	if (c->cpu_index == boot_cpu_id)
152 		return;
153 
154 	/*
155 	 * Certain Athlons might work (for various values of 'work') in SMP
156 	 * but they are not certified as MP capable.
157 	 */
158 	/* Athlon 660/661 is valid. */
159 	if ((c->x86_model == 6) && ((c->x86_mask == 0) ||
160 	    (c->x86_mask == 1)))
161 		goto valid_k7;
162 
163 	/* Duron 670 is valid */
164 	if ((c->x86_model == 7) && (c->x86_mask == 0))
165 		goto valid_k7;
166 
167 	/*
168 	 * Athlon 662, Duron 671, and Athlon >model 7 have capability
169 	 * bit. It's worth noting that the A5 stepping (662) of some
170 	 * Athlon XP's have the MP bit set.
171 	 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for
172 	 * more.
173 	 */
174 	if (((c->x86_model == 6) && (c->x86_mask >= 2)) ||
175 	    ((c->x86_model == 7) && (c->x86_mask >= 1)) ||
176 	     (c->x86_model > 7))
177 		if (cpu_has_mp)
178 			goto valid_k7;
179 
180 	/* If we get here, not a certified SMP capable AMD system. */
181 
182 	/*
183 	 * Don't taint if we are running SMP kernel on a single non-MP
184 	 * approved Athlon
185 	 */
186 	WARN_ONCE(1, "WARNING: This combination of AMD"
187 		"processors is not suitable for SMP.\n");
188 	if (!test_taint(TAINT_UNSAFE_SMP))
189 		add_taint(TAINT_UNSAFE_SMP);
190 
191 valid_k7:
192 	;
193 #endif
194 }
195 
196 static void __cpuinit init_amd_k7(struct cpuinfo_x86 *c)
197 {
198 	u32 l, h;
199 
200 	/*
201 	 * Bit 15 of Athlon specific MSR 15, needs to be 0
202 	 * to enable SSE on Palomino/Morgan/Barton CPU's.
203 	 * If the BIOS didn't enable it already, enable it here.
204 	 */
205 	if (c->x86_model >= 6 && c->x86_model <= 10) {
206 		if (!cpu_has(c, X86_FEATURE_XMM)) {
207 			printk(KERN_INFO "Enabling disabled K7/SSE Support.\n");
208 			rdmsr(MSR_K7_HWCR, l, h);
209 			l &= ~0x00008000;
210 			wrmsr(MSR_K7_HWCR, l, h);
211 			set_cpu_cap(c, X86_FEATURE_XMM);
212 		}
213 	}
214 
215 	/*
216 	 * It's been determined by AMD that Athlons since model 8 stepping 1
217 	 * are more robust with CLK_CTL set to 200xxxxx instead of 600xxxxx
218 	 * As per AMD technical note 27212 0.2
219 	 */
220 	if ((c->x86_model == 8 && c->x86_mask >= 1) || (c->x86_model > 8)) {
221 		rdmsr(MSR_K7_CLK_CTL, l, h);
222 		if ((l & 0xfff00000) != 0x20000000) {
223 			printk(KERN_INFO
224 			    "CPU: CLK_CTL MSR was %x. Reprogramming to %x\n",
225 					l, ((l & 0x000fffff)|0x20000000));
226 			wrmsr(MSR_K7_CLK_CTL, (l & 0x000fffff)|0x20000000, h);
227 		}
228 	}
229 
230 	set_cpu_cap(c, X86_FEATURE_K7);
231 
232 	amd_k7_smp_check(c);
233 }
234 #endif
235 
236 #if defined(CONFIG_NUMA) && defined(CONFIG_X86_64)
237 static int __cpuinit nearby_node(int apicid)
238 {
239 	int i, node;
240 
241 	for (i = apicid - 1; i >= 0; i--) {
242 		node = apicid_to_node[i];
243 		if (node != NUMA_NO_NODE && node_online(node))
244 			return node;
245 	}
246 	for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) {
247 		node = apicid_to_node[i];
248 		if (node != NUMA_NO_NODE && node_online(node))
249 			return node;
250 	}
251 	return first_node(node_online_map); /* Shouldn't happen */
252 }
253 #endif
254 
255 /*
256  * Fixup core topology information for AMD multi-node processors.
257  * Assumption 1: Number of cores in each internal node is the same.
258  * Assumption 2: Mixed systems with both single-node and dual-node
259  *               processors are not supported.
260  */
261 #ifdef CONFIG_X86_HT
262 static void __cpuinit amd_fixup_dcm(struct cpuinfo_x86 *c)
263 {
264 #ifdef CONFIG_PCI
265 	u32 t, cpn;
266 	u8 n, n_id;
267 	int cpu = smp_processor_id();
268 
269 	/* fixup topology information only once for a core */
270 	if (cpu_has(c, X86_FEATURE_AMD_DCM))
271 		return;
272 
273 	/* check for multi-node processor on boot cpu */
274 	t = read_pci_config(0, 24, 3, 0xe8);
275 	if (!(t & (1 << 29)))
276 		return;
277 
278 	set_cpu_cap(c, X86_FEATURE_AMD_DCM);
279 
280 	/* cores per node: each internal node has half the number of cores */
281 	cpn = c->x86_max_cores >> 1;
282 
283 	/* even-numbered NB_id of this dual-node processor */
284 	n = c->phys_proc_id << 1;
285 
286 	/*
287 	 * determine internal node id and assign cores fifty-fifty to
288 	 * each node of the dual-node processor
289 	 */
290 	t = read_pci_config(0, 24 + n, 3, 0xe8);
291 	n = (t>>30) & 0x3;
292 	if (n == 0) {
293 		if (c->cpu_core_id < cpn)
294 			n_id = 0;
295 		else
296 			n_id = 1;
297 	} else {
298 		if (c->cpu_core_id < cpn)
299 			n_id = 1;
300 		else
301 			n_id = 0;
302 	}
303 
304 	/* compute entire NodeID, use llc_shared_map to store sibling info */
305 	per_cpu(cpu_llc_id, cpu) = (c->phys_proc_id << 1) + n_id;
306 
307 	/* fixup core id to be in range from 0 to cpn */
308 	c->cpu_core_id = c->cpu_core_id % cpn;
309 #endif
310 }
311 #endif
312 
313 /*
314  * On a AMD dual core setup the lower bits of the APIC id distingush the cores.
315  * Assumes number of cores is a power of two.
316  */
317 static void __cpuinit amd_detect_cmp(struct cpuinfo_x86 *c)
318 {
319 #ifdef CONFIG_X86_HT
320 	unsigned bits;
321 	int cpu = smp_processor_id();
322 
323 	bits = c->x86_coreid_bits;
324 	/* Low order bits define the core id (index of core in socket) */
325 	c->cpu_core_id = c->initial_apicid & ((1 << bits)-1);
326 	/* Convert the initial APIC ID into the socket ID */
327 	c->phys_proc_id = c->initial_apicid >> bits;
328 	/* use socket ID also for last level cache */
329 	per_cpu(cpu_llc_id, cpu) = c->phys_proc_id;
330 	/* fixup topology information on multi-node processors */
331 	if ((c->x86 == 0x10) && (c->x86_model == 9))
332 		amd_fixup_dcm(c);
333 #endif
334 }
335 
336 int amd_get_nb_id(int cpu)
337 {
338 	int id = 0;
339 #ifdef CONFIG_SMP
340 	id = per_cpu(cpu_llc_id, cpu);
341 #endif
342 	return id;
343 }
344 EXPORT_SYMBOL_GPL(amd_get_nb_id);
345 
346 static void __cpuinit srat_detect_node(struct cpuinfo_x86 *c)
347 {
348 #if defined(CONFIG_NUMA) && defined(CONFIG_X86_64)
349 	int cpu = smp_processor_id();
350 	int node;
351 	unsigned apicid = c->apicid;
352 
353 	node = per_cpu(cpu_llc_id, cpu);
354 
355 	if (apicid_to_node[apicid] != NUMA_NO_NODE)
356 		node = apicid_to_node[apicid];
357 	if (!node_online(node)) {
358 		/* Two possibilities here:
359 		   - The CPU is missing memory and no node was created.
360 		   In that case try picking one from a nearby CPU
361 		   - The APIC IDs differ from the HyperTransport node IDs
362 		   which the K8 northbridge parsing fills in.
363 		   Assume they are all increased by a constant offset,
364 		   but in the same order as the HT nodeids.
365 		   If that doesn't result in a usable node fall back to the
366 		   path for the previous case.  */
367 
368 		int ht_nodeid = c->initial_apicid;
369 
370 		if (ht_nodeid >= 0 &&
371 		    apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
372 			node = apicid_to_node[ht_nodeid];
373 		/* Pick a nearby node */
374 		if (!node_online(node))
375 			node = nearby_node(apicid);
376 	}
377 	numa_set_node(cpu, node);
378 
379 	printk(KERN_INFO "CPU %d/0x%x -> Node %d\n", cpu, apicid, node);
380 #endif
381 }
382 
383 static void __cpuinit early_init_amd_mc(struct cpuinfo_x86 *c)
384 {
385 #ifdef CONFIG_X86_HT
386 	unsigned bits, ecx;
387 
388 	/* Multi core CPU? */
389 	if (c->extended_cpuid_level < 0x80000008)
390 		return;
391 
392 	ecx = cpuid_ecx(0x80000008);
393 
394 	c->x86_max_cores = (ecx & 0xff) + 1;
395 
396 	/* CPU telling us the core id bits shift? */
397 	bits = (ecx >> 12) & 0xF;
398 
399 	/* Otherwise recompute */
400 	if (bits == 0) {
401 		while ((1 << bits) < c->x86_max_cores)
402 			bits++;
403 	}
404 
405 	c->x86_coreid_bits = bits;
406 #endif
407 }
408 
409 static void __cpuinit early_init_amd(struct cpuinfo_x86 *c)
410 {
411 	early_init_amd_mc(c);
412 
413 	/*
414 	 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
415 	 * with P/T states and does not stop in deep C-states
416 	 */
417 	if (c->x86_power & (1 << 8)) {
418 		set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
419 		set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
420 	}
421 
422 #ifdef CONFIG_X86_64
423 	set_cpu_cap(c, X86_FEATURE_SYSCALL32);
424 #else
425 	/*  Set MTRR capability flag if appropriate */
426 	if (c->x86 == 5)
427 		if (c->x86_model == 13 || c->x86_model == 9 ||
428 		    (c->x86_model == 8 && c->x86_mask >= 8))
429 			set_cpu_cap(c, X86_FEATURE_K6_MTRR);
430 #endif
431 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
432 	/* check CPU config space for extended APIC ID */
433 	if (cpu_has_apic && c->x86 >= 0xf) {
434 		unsigned int val;
435 		val = read_pci_config(0, 24, 0, 0x68);
436 		if ((val & ((1 << 17) | (1 << 18))) == ((1 << 17) | (1 << 18)))
437 			set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
438 	}
439 #endif
440 }
441 
442 static void __cpuinit init_amd(struct cpuinfo_x86 *c)
443 {
444 #ifdef CONFIG_SMP
445 	unsigned long long value;
446 
447 	/*
448 	 * Disable TLB flush filter by setting HWCR.FFDIS on K8
449 	 * bit 6 of msr C001_0015
450 	 *
451 	 * Errata 63 for SH-B3 steppings
452 	 * Errata 122 for all steppings (F+ have it disabled by default)
453 	 */
454 	if (c->x86 == 0xf) {
455 		rdmsrl(MSR_K7_HWCR, value);
456 		value |= 1 << 6;
457 		wrmsrl(MSR_K7_HWCR, value);
458 	}
459 #endif
460 
461 	early_init_amd(c);
462 
463 	/*
464 	 * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
465 	 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
466 	 */
467 	clear_cpu_cap(c, 0*32+31);
468 
469 #ifdef CONFIG_X86_64
470 	/* On C+ stepping K8 rep microcode works well for copy/memset */
471 	if (c->x86 == 0xf) {
472 		u32 level;
473 
474 		level = cpuid_eax(1);
475 		if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58)
476 			set_cpu_cap(c, X86_FEATURE_REP_GOOD);
477 
478 		/*
479 		 * Some BIOSes incorrectly force this feature, but only K8
480 		 * revision D (model = 0x14) and later actually support it.
481 		 * (AMD Erratum #110, docId: 25759).
482 		 */
483 		if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM)) {
484 			u64 val;
485 
486 			clear_cpu_cap(c, X86_FEATURE_LAHF_LM);
487 			if (!rdmsrl_amd_safe(0xc001100d, &val)) {
488 				val &= ~(1ULL << 32);
489 				wrmsrl_amd_safe(0xc001100d, val);
490 			}
491 		}
492 
493 	}
494 	if (c->x86 == 0x10 || c->x86 == 0x11)
495 		set_cpu_cap(c, X86_FEATURE_REP_GOOD);
496 
497 	/* get apicid instead of initial apic id from cpuid */
498 	c->apicid = hard_smp_processor_id();
499 #else
500 
501 	/*
502 	 *	FIXME: We should handle the K5 here. Set up the write
503 	 *	range and also turn on MSR 83 bits 4 and 31 (write alloc,
504 	 *	no bus pipeline)
505 	 */
506 
507 	switch (c->x86) {
508 	case 4:
509 		init_amd_k5(c);
510 		break;
511 	case 5:
512 		init_amd_k6(c);
513 		break;
514 	case 6: /* An Athlon/Duron */
515 		init_amd_k7(c);
516 		break;
517 	}
518 
519 	/* K6s reports MCEs but don't actually have all the MSRs */
520 	if (c->x86 < 6)
521 		clear_cpu_cap(c, X86_FEATURE_MCE);
522 #endif
523 
524 	/* Enable workaround for FXSAVE leak */
525 	if (c->x86 >= 6)
526 		set_cpu_cap(c, X86_FEATURE_FXSAVE_LEAK);
527 
528 	if (!c->x86_model_id[0]) {
529 		switch (c->x86) {
530 		case 0xf:
531 			/* Should distinguish Models here, but this is only
532 			   a fallback anyways. */
533 			strcpy(c->x86_model_id, "Hammer");
534 			break;
535 		}
536 	}
537 
538 	display_cacheinfo(c);
539 
540 	/* Multi core CPU? */
541 	if (c->extended_cpuid_level >= 0x80000008) {
542 		amd_detect_cmp(c);
543 		srat_detect_node(c);
544 	}
545 
546 #ifdef CONFIG_X86_32
547 	detect_ht(c);
548 #endif
549 
550 	if (c->extended_cpuid_level >= 0x80000006) {
551 		if ((c->x86 >= 0x0f) && (cpuid_edx(0x80000006) & 0xf000))
552 			num_cache_leaves = 4;
553 		else
554 			num_cache_leaves = 3;
555 	}
556 
557 	if (c->x86 >= 0xf && c->x86 <= 0x11)
558 		set_cpu_cap(c, X86_FEATURE_K8);
559 
560 	if (cpu_has_xmm2) {
561 		/* MFENCE stops RDTSC speculation */
562 		set_cpu_cap(c, X86_FEATURE_MFENCE_RDTSC);
563 	}
564 
565 #ifdef CONFIG_X86_64
566 	if (c->x86 == 0x10) {
567 		/* do this for boot cpu */
568 		if (c == &boot_cpu_data)
569 			check_enable_amd_mmconf_dmi();
570 
571 		fam10h_check_enable_mmcfg();
572 	}
573 
574 	if (c == &boot_cpu_data && c->x86 >= 0xf && c->x86 <= 0x11) {
575 		unsigned long long tseg;
576 
577 		/*
578 		 * Split up direct mapping around the TSEG SMM area.
579 		 * Don't do it for gbpages because there seems very little
580 		 * benefit in doing so.
581 		 */
582 		if (!rdmsrl_safe(MSR_K8_TSEG_ADDR, &tseg)) {
583 			printk(KERN_DEBUG "tseg: %010llx\n", tseg);
584 			if ((tseg>>PMD_SHIFT) <
585 				(max_low_pfn_mapped>>(PMD_SHIFT-PAGE_SHIFT)) ||
586 				((tseg>>PMD_SHIFT) <
587 				(max_pfn_mapped>>(PMD_SHIFT-PAGE_SHIFT)) &&
588 				(tseg>>PMD_SHIFT) >= (1ULL<<(32 - PMD_SHIFT))))
589 				set_memory_4k((unsigned long)__va(tseg), 1);
590 		}
591 	}
592 #endif
593 }
594 
595 #ifdef CONFIG_X86_32
596 static unsigned int __cpuinit amd_size_cache(struct cpuinfo_x86 *c,
597 							unsigned int size)
598 {
599 	/* AMD errata T13 (order #21922) */
600 	if ((c->x86 == 6)) {
601 		/* Duron Rev A0 */
602 		if (c->x86_model == 3 && c->x86_mask == 0)
603 			size = 64;
604 		/* Tbird rev A1/A2 */
605 		if (c->x86_model == 4 &&
606 			(c->x86_mask == 0 || c->x86_mask == 1))
607 			size = 256;
608 	}
609 	return size;
610 }
611 #endif
612 
613 static const struct cpu_dev __cpuinitconst amd_cpu_dev = {
614 	.c_vendor	= "AMD",
615 	.c_ident	= { "AuthenticAMD" },
616 #ifdef CONFIG_X86_32
617 	.c_models = {
618 		{ .vendor = X86_VENDOR_AMD, .family = 4, .model_names =
619 		  {
620 			  [3] = "486 DX/2",
621 			  [7] = "486 DX/2-WB",
622 			  [8] = "486 DX/4",
623 			  [9] = "486 DX/4-WB",
624 			  [14] = "Am5x86-WT",
625 			  [15] = "Am5x86-WB"
626 		  }
627 		},
628 	},
629 	.c_size_cache	= amd_size_cache,
630 #endif
631 	.c_early_init   = early_init_amd,
632 	.c_init		= init_amd,
633 	.c_x86_vendor	= X86_VENDOR_AMD,
634 };
635 
636 cpu_dev_register(amd_cpu_dev);
637