xref: /linux/arch/x86/kernel/cpu/amd.c (revision 5bdef865eb358b6f3760e25e591ae115e9eeddef)
1 #include <linux/init.h>
2 #include <linux/bitops.h>
3 #include <linux/mm.h>
4 
5 #include <asm/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("system stability may be impaired when more than 32 MB are used.\n");
91 		else
92 			printk("probably OK (after B9730xxxx).\n");
93 		printk(KERN_INFO "Please see http://membres.lycos.fr/poulot/k6bug.html\n");
94 	}
95 
96 	/* K6 with old style WHCR */
97 	if (c->x86_model < 8 ||
98 	   (c->x86_model == 8 && c->x86_mask < 8)) {
99 		/* We can only write allocate on the low 508Mb */
100 		if (mbytes > 508)
101 			mbytes = 508;
102 
103 		rdmsr(MSR_K6_WHCR, l, h);
104 		if ((l&0x0000FFFF) == 0) {
105 			unsigned long flags;
106 			l = (1<<0)|((mbytes/4)<<1);
107 			local_irq_save(flags);
108 			wbinvd();
109 			wrmsr(MSR_K6_WHCR, l, h);
110 			local_irq_restore(flags);
111 			printk(KERN_INFO "Enabling old style K6 write allocation for %d Mb\n",
112 				mbytes);
113 		}
114 		return;
115 	}
116 
117 	if ((c->x86_model == 8 && c->x86_mask > 7) ||
118 	     c->x86_model == 9 || c->x86_model == 13) {
119 		/* The more serious chips .. */
120 
121 		if (mbytes > 4092)
122 			mbytes = 4092;
123 
124 		rdmsr(MSR_K6_WHCR, l, h);
125 		if ((l&0xFFFF0000) == 0) {
126 			unsigned long flags;
127 			l = ((mbytes>>2)<<22)|(1<<16);
128 			local_irq_save(flags);
129 			wbinvd();
130 			wrmsr(MSR_K6_WHCR, l, h);
131 			local_irq_restore(flags);
132 			printk(KERN_INFO "Enabling new style K6 write allocation for %d Mb\n",
133 				mbytes);
134 		}
135 
136 		return;
137 	}
138 
139 	if (c->x86_model == 10) {
140 		/* AMD Geode LX is model 10 */
141 		/* placeholder for any needed mods */
142 		return;
143 	}
144 }
145 
146 static void __cpuinit amd_k7_smp_check(struct cpuinfo_x86 *c)
147 {
148 #ifdef CONFIG_SMP
149 	/* calling is from identify_secondary_cpu() ? */
150 	if (c->cpu_index == boot_cpu_id)
151 		return;
152 
153 	/*
154 	 * Certain Athlons might work (for various values of 'work') in SMP
155 	 * but they are not certified as MP capable.
156 	 */
157 	/* Athlon 660/661 is valid. */
158 	if ((c->x86_model == 6) && ((c->x86_mask == 0) ||
159 	    (c->x86_mask == 1)))
160 		goto valid_k7;
161 
162 	/* Duron 670 is valid */
163 	if ((c->x86_model == 7) && (c->x86_mask == 0))
164 		goto valid_k7;
165 
166 	/*
167 	 * Athlon 662, Duron 671, and Athlon >model 7 have capability
168 	 * bit. It's worth noting that the A5 stepping (662) of some
169 	 * Athlon XP's have the MP bit set.
170 	 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for
171 	 * more.
172 	 */
173 	if (((c->x86_model == 6) && (c->x86_mask >= 2)) ||
174 	    ((c->x86_model == 7) && (c->x86_mask >= 1)) ||
175 	     (c->x86_model > 7))
176 		if (cpu_has_mp)
177 			goto valid_k7;
178 
179 	/* If we get here, not a certified SMP capable AMD system. */
180 
181 	/*
182 	 * Don't taint if we are running SMP kernel on a single non-MP
183 	 * approved Athlon
184 	 */
185 	WARN_ONCE(1, "WARNING: This combination of AMD"
186 		"processors is not suitable for SMP.\n");
187 	if (!test_taint(TAINT_UNSAFE_SMP))
188 		add_taint(TAINT_UNSAFE_SMP);
189 
190 valid_k7:
191 	;
192 #endif
193 }
194 
195 static void __cpuinit init_amd_k7(struct cpuinfo_x86 *c)
196 {
197 	u32 l, h;
198 
199 	/*
200 	 * Bit 15 of Athlon specific MSR 15, needs to be 0
201 	 * to enable SSE on Palomino/Morgan/Barton CPU's.
202 	 * If the BIOS didn't enable it already, enable it here.
203 	 */
204 	if (c->x86_model >= 6 && c->x86_model <= 10) {
205 		if (!cpu_has(c, X86_FEATURE_XMM)) {
206 			printk(KERN_INFO "Enabling disabled K7/SSE Support.\n");
207 			rdmsr(MSR_K7_HWCR, l, h);
208 			l &= ~0x00008000;
209 			wrmsr(MSR_K7_HWCR, l, h);
210 			set_cpu_cap(c, X86_FEATURE_XMM);
211 		}
212 	}
213 
214 	/*
215 	 * It's been determined by AMD that Athlons since model 8 stepping 1
216 	 * are more robust with CLK_CTL set to 200xxxxx instead of 600xxxxx
217 	 * As per AMD technical note 27212 0.2
218 	 */
219 	if ((c->x86_model == 8 && c->x86_mask >= 1) || (c->x86_model > 8)) {
220 		rdmsr(MSR_K7_CLK_CTL, l, h);
221 		if ((l & 0xfff00000) != 0x20000000) {
222 			printk ("CPU: CLK_CTL MSR was %x. Reprogramming to %x\n", l,
223 				((l & 0x000fffff)|0x20000000));
224 			wrmsr(MSR_K7_CLK_CTL, (l & 0x000fffff)|0x20000000, h);
225 		}
226 	}
227 
228 	set_cpu_cap(c, X86_FEATURE_K7);
229 
230 	amd_k7_smp_check(c);
231 }
232 #endif
233 
234 #if defined(CONFIG_NUMA) && defined(CONFIG_X86_64)
235 static int __cpuinit nearby_node(int apicid)
236 {
237 	int i, node;
238 
239 	for (i = apicid - 1; i >= 0; i--) {
240 		node = apicid_to_node[i];
241 		if (node != NUMA_NO_NODE && node_online(node))
242 			return node;
243 	}
244 	for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) {
245 		node = apicid_to_node[i];
246 		if (node != NUMA_NO_NODE && node_online(node))
247 			return node;
248 	}
249 	return first_node(node_online_map); /* Shouldn't happen */
250 }
251 #endif
252 
253 /*
254  * On a AMD dual core setup the lower bits of the APIC id distingush the cores.
255  * Assumes number of cores is a power of two.
256  */
257 static void __cpuinit amd_detect_cmp(struct cpuinfo_x86 *c)
258 {
259 #ifdef CONFIG_X86_HT
260 	unsigned bits;
261 	int cpu = smp_processor_id();
262 
263 	bits = c->x86_coreid_bits;
264 	/* Low order bits define the core id (index of core in socket) */
265 	c->cpu_core_id = c->initial_apicid & ((1 << bits)-1);
266 	/* Convert the initial APIC ID into the socket ID */
267 	c->phys_proc_id = c->initial_apicid >> bits;
268 	/* use socket ID also for last level cache */
269 	per_cpu(cpu_llc_id, cpu) = c->phys_proc_id;
270 #endif
271 }
272 
273 static void __cpuinit srat_detect_node(struct cpuinfo_x86 *c)
274 {
275 #if defined(CONFIG_NUMA) && defined(CONFIG_X86_64)
276 	int cpu = smp_processor_id();
277 	int node;
278 	unsigned apicid = cpu_has_apic ? hard_smp_processor_id() : c->apicid;
279 
280 	node = c->phys_proc_id;
281 	if (apicid_to_node[apicid] != NUMA_NO_NODE)
282 		node = apicid_to_node[apicid];
283 	if (!node_online(node)) {
284 		/* Two possibilities here:
285 		   - The CPU is missing memory and no node was created.
286 		   In that case try picking one from a nearby CPU
287 		   - The APIC IDs differ from the HyperTransport node IDs
288 		   which the K8 northbridge parsing fills in.
289 		   Assume they are all increased by a constant offset,
290 		   but in the same order as the HT nodeids.
291 		   If that doesn't result in a usable node fall back to the
292 		   path for the previous case.  */
293 
294 		int ht_nodeid = c->initial_apicid;
295 
296 		if (ht_nodeid >= 0 &&
297 		    apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
298 			node = apicid_to_node[ht_nodeid];
299 		/* Pick a nearby node */
300 		if (!node_online(node))
301 			node = nearby_node(apicid);
302 	}
303 	numa_set_node(cpu, node);
304 
305 	printk(KERN_INFO "CPU %d/0x%x -> Node %d\n", cpu, apicid, node);
306 #endif
307 }
308 
309 static void __cpuinit early_init_amd_mc(struct cpuinfo_x86 *c)
310 {
311 #ifdef CONFIG_X86_HT
312 	unsigned bits, ecx;
313 
314 	/* Multi core CPU? */
315 	if (c->extended_cpuid_level < 0x80000008)
316 		return;
317 
318 	ecx = cpuid_ecx(0x80000008);
319 
320 	c->x86_max_cores = (ecx & 0xff) + 1;
321 
322 	/* CPU telling us the core id bits shift? */
323 	bits = (ecx >> 12) & 0xF;
324 
325 	/* Otherwise recompute */
326 	if (bits == 0) {
327 		while ((1 << bits) < c->x86_max_cores)
328 			bits++;
329 	}
330 
331 	c->x86_coreid_bits = bits;
332 #endif
333 }
334 
335 static void __cpuinit early_init_amd(struct cpuinfo_x86 *c)
336 {
337 	early_init_amd_mc(c);
338 
339 	/*
340 	 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
341 	 * with P/T states and does not stop in deep C-states
342 	 */
343 	if (c->x86_power & (1 << 8)) {
344 		set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
345 		set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
346 	}
347 
348 #ifdef CONFIG_X86_64
349 	set_cpu_cap(c, X86_FEATURE_SYSCALL32);
350 #else
351 	/*  Set MTRR capability flag if appropriate */
352 	if (c->x86 == 5)
353 		if (c->x86_model == 13 || c->x86_model == 9 ||
354 		    (c->x86_model == 8 && c->x86_mask >= 8))
355 			set_cpu_cap(c, X86_FEATURE_K6_MTRR);
356 #endif
357 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
358 	/* check CPU config space for extended APIC ID */
359 	if (c->x86 >= 0xf) {
360 		unsigned int val;
361 		val = read_pci_config(0, 24, 0, 0x68);
362 		if ((val & ((1 << 17) | (1 << 18))) == ((1 << 17) | (1 << 18)))
363 			set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
364 	}
365 #endif
366 }
367 
368 static void __cpuinit init_amd(struct cpuinfo_x86 *c)
369 {
370 #ifdef CONFIG_SMP
371 	unsigned long long value;
372 
373 	/*
374 	 * Disable TLB flush filter by setting HWCR.FFDIS on K8
375 	 * bit 6 of msr C001_0015
376 	 *
377 	 * Errata 63 for SH-B3 steppings
378 	 * Errata 122 for all steppings (F+ have it disabled by default)
379 	 */
380 	if (c->x86 == 0xf) {
381 		rdmsrl(MSR_K7_HWCR, value);
382 		value |= 1 << 6;
383 		wrmsrl(MSR_K7_HWCR, value);
384 	}
385 #endif
386 
387 	early_init_amd(c);
388 
389 	/*
390 	 * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
391 	 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
392 	 */
393 	clear_cpu_cap(c, 0*32+31);
394 
395 #ifdef CONFIG_X86_64
396 	/* On C+ stepping K8 rep microcode works well for copy/memset */
397 	if (c->x86 == 0xf) {
398 		u32 level;
399 
400 		level = cpuid_eax(1);
401 		if((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58)
402 			set_cpu_cap(c, X86_FEATURE_REP_GOOD);
403 	}
404 	if (c->x86 == 0x10 || c->x86 == 0x11)
405 		set_cpu_cap(c, X86_FEATURE_REP_GOOD);
406 #else
407 
408 	/*
409 	 *	FIXME: We should handle the K5 here. Set up the write
410 	 *	range and also turn on MSR 83 bits 4 and 31 (write alloc,
411 	 *	no bus pipeline)
412 	 */
413 
414 	switch (c->x86) {
415 	case 4:
416 		init_amd_k5(c);
417 		break;
418 	case 5:
419 		init_amd_k6(c);
420 		break;
421 	case 6: /* An Athlon/Duron */
422 		init_amd_k7(c);
423 		break;
424 	}
425 
426 	/* K6s reports MCEs but don't actually have all the MSRs */
427 	if (c->x86 < 6)
428 		clear_cpu_cap(c, X86_FEATURE_MCE);
429 #endif
430 
431 	/* Enable workaround for FXSAVE leak */
432 	if (c->x86 >= 6)
433 		set_cpu_cap(c, X86_FEATURE_FXSAVE_LEAK);
434 
435 	if (!c->x86_model_id[0]) {
436 		switch (c->x86) {
437 		case 0xf:
438 			/* Should distinguish Models here, but this is only
439 			   a fallback anyways. */
440 			strcpy(c->x86_model_id, "Hammer");
441 			break;
442 		}
443 	}
444 
445 	display_cacheinfo(c);
446 
447 	/* Multi core CPU? */
448 	if (c->extended_cpuid_level >= 0x80000008) {
449 		amd_detect_cmp(c);
450 		srat_detect_node(c);
451 	}
452 
453 #ifdef CONFIG_X86_32
454 	detect_ht(c);
455 #endif
456 
457 	if (c->extended_cpuid_level >= 0x80000006) {
458 		if ((c->x86 >= 0x0f) && (cpuid_edx(0x80000006) & 0xf000))
459 			num_cache_leaves = 4;
460 		else
461 			num_cache_leaves = 3;
462 	}
463 
464 	if (c->x86 >= 0xf && c->x86 <= 0x11)
465 		set_cpu_cap(c, X86_FEATURE_K8);
466 
467 	if (cpu_has_xmm2) {
468 		/* MFENCE stops RDTSC speculation */
469 		set_cpu_cap(c, X86_FEATURE_MFENCE_RDTSC);
470 	}
471 
472 #ifdef CONFIG_X86_64
473 	if (c->x86 == 0x10) {
474 		/* do this for boot cpu */
475 		if (c == &boot_cpu_data)
476 			check_enable_amd_mmconf_dmi();
477 
478 		fam10h_check_enable_mmcfg();
479 	}
480 
481 	if (c == &boot_cpu_data && c->x86 >= 0xf && c->x86 <= 0x11) {
482 		unsigned long long tseg;
483 
484 		/*
485 		 * Split up direct mapping around the TSEG SMM area.
486 		 * Don't do it for gbpages because there seems very little
487 		 * benefit in doing so.
488 		 */
489 		if (!rdmsrl_safe(MSR_K8_TSEG_ADDR, &tseg)) {
490 		    printk(KERN_DEBUG "tseg: %010llx\n", tseg);
491 		    if ((tseg>>PMD_SHIFT) <
492 				(max_low_pfn_mapped>>(PMD_SHIFT-PAGE_SHIFT)) ||
493 			((tseg>>PMD_SHIFT) <
494 				(max_pfn_mapped>>(PMD_SHIFT-PAGE_SHIFT)) &&
495 			 (tseg>>PMD_SHIFT) >= (1ULL<<(32 - PMD_SHIFT))))
496 			set_memory_4k((unsigned long)__va(tseg), 1);
497 		}
498 	}
499 #endif
500 }
501 
502 #ifdef CONFIG_X86_32
503 static unsigned int __cpuinit amd_size_cache(struct cpuinfo_x86 *c, unsigned int size)
504 {
505 	/* AMD errata T13 (order #21922) */
506 	if ((c->x86 == 6)) {
507 		if (c->x86_model == 3 && c->x86_mask == 0)	/* Duron Rev A0 */
508 			size = 64;
509 		if (c->x86_model == 4 &&
510 		    (c->x86_mask == 0 || c->x86_mask == 1))	/* Tbird rev A1/A2 */
511 			size = 256;
512 	}
513 	return size;
514 }
515 #endif
516 
517 static const struct cpu_dev __cpuinitconst amd_cpu_dev = {
518 	.c_vendor	= "AMD",
519 	.c_ident	= { "AuthenticAMD" },
520 #ifdef CONFIG_X86_32
521 	.c_models = {
522 		{ .vendor = X86_VENDOR_AMD, .family = 4, .model_names =
523 		  {
524 			  [3] = "486 DX/2",
525 			  [7] = "486 DX/2-WB",
526 			  [8] = "486 DX/4",
527 			  [9] = "486 DX/4-WB",
528 			  [14] = "Am5x86-WT",
529 			  [15] = "Am5x86-WB"
530 		  }
531 		},
532 	},
533 	.c_size_cache	= amd_size_cache,
534 #endif
535 	.c_early_init   = early_init_amd,
536 	.c_init		= init_amd,
537 	.c_x86_vendor	= X86_VENDOR_AMD,
538 };
539 
540 cpu_dev_register(amd_cpu_dev);
541