xref: /linux/arch/x86/kernel/cpu/intel.c (revision d39d0ed196aa1685bb24771e92f78633c66ac9cb)
1 #include <linux/init.h>
2 #include <linux/kernel.h>
3 
4 #include <linux/string.h>
5 #include <linux/bitops.h>
6 #include <linux/smp.h>
7 #include <linux/sched.h>
8 #include <linux/thread_info.h>
9 #include <linux/module.h>
10 #include <linux/uaccess.h>
11 
12 #include <asm/processor.h>
13 #include <asm/pgtable.h>
14 #include <asm/msr.h>
15 #include <asm/bugs.h>
16 #include <asm/cpu.h>
17 
18 #ifdef CONFIG_X86_64
19 #include <linux/topology.h>
20 #include <asm/numa_64.h>
21 #endif
22 
23 #include "cpu.h"
24 
25 #ifdef CONFIG_X86_LOCAL_APIC
26 #include <asm/mpspec.h>
27 #include <asm/apic.h>
28 #endif
29 
30 static void __cpuinit early_init_intel(struct cpuinfo_x86 *c)
31 {
32 	/* Unmask CPUID levels if masked: */
33 	if (c->x86 > 6 || (c->x86 == 6 && c->x86_model >= 0xd)) {
34 		u64 misc_enable;
35 
36 		rdmsrl(MSR_IA32_MISC_ENABLE, misc_enable);
37 
38 		if (misc_enable & MSR_IA32_MISC_ENABLE_LIMIT_CPUID) {
39 			misc_enable &= ~MSR_IA32_MISC_ENABLE_LIMIT_CPUID;
40 			wrmsrl(MSR_IA32_MISC_ENABLE, misc_enable);
41 			c->cpuid_level = cpuid_eax(0);
42 		}
43 	}
44 
45 	if ((c->x86 == 0xf && c->x86_model >= 0x03) ||
46 		(c->x86 == 0x6 && c->x86_model >= 0x0e))
47 		set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
48 
49 	/*
50 	 * Atom erratum AAE44/AAF40/AAG38/AAH41:
51 	 *
52 	 * A race condition between speculative fetches and invalidating
53 	 * a large page.  This is worked around in microcode, but we
54 	 * need the microcode to have already been loaded... so if it is
55 	 * not, recommend a BIOS update and disable large pages.
56 	 */
57 	if (c->x86 == 6 && c->x86_model == 0x1c && c->x86_mask <= 2) {
58 		u32 ucode, junk;
59 
60 		wrmsr(MSR_IA32_UCODE_REV, 0, 0);
61 		sync_core();
62 		rdmsr(MSR_IA32_UCODE_REV, junk, ucode);
63 
64 		if (ucode < 0x20e) {
65 			printk(KERN_WARNING "Atom PSE erratum detected, BIOS microcode update recommended\n");
66 			clear_cpu_cap(c, X86_FEATURE_PSE);
67 		}
68 	}
69 
70 #ifdef CONFIG_X86_64
71 	set_cpu_cap(c, X86_FEATURE_SYSENTER32);
72 #else
73 	/* Netburst reports 64 bytes clflush size, but does IO in 128 bytes */
74 	if (c->x86 == 15 && c->x86_cache_alignment == 64)
75 		c->x86_cache_alignment = 128;
76 #endif
77 
78 	/* CPUID workaround for 0F33/0F34 CPU */
79 	if (c->x86 == 0xF && c->x86_model == 0x3
80 	    && (c->x86_mask == 0x3 || c->x86_mask == 0x4))
81 		c->x86_phys_bits = 36;
82 
83 	/*
84 	 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
85 	 * with P/T states and does not stop in deep C-states.
86 	 *
87 	 * It is also reliable across cores and sockets. (but not across
88 	 * cabinets - we turn it off in that case explicitly.)
89 	 */
90 	if (c->x86_power & (1 << 8)) {
91 		set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
92 		set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
93 		if (!check_tsc_unstable())
94 			sched_clock_stable = 1;
95 	}
96 
97 	/*
98 	 * There is a known erratum on Pentium III and Core Solo
99 	 * and Core Duo CPUs.
100 	 * " Page with PAT set to WC while associated MTRR is UC
101 	 *   may consolidate to UC "
102 	 * Because of this erratum, it is better to stick with
103 	 * setting WC in MTRR rather than using PAT on these CPUs.
104 	 *
105 	 * Enable PAT WC only on P4, Core 2 or later CPUs.
106 	 */
107 	if (c->x86 == 6 && c->x86_model < 15)
108 		clear_cpu_cap(c, X86_FEATURE_PAT);
109 
110 #ifdef CONFIG_KMEMCHECK
111 	/*
112 	 * P4s have a "fast strings" feature which causes single-
113 	 * stepping REP instructions to only generate a #DB on
114 	 * cache-line boundaries.
115 	 *
116 	 * Ingo Molnar reported a Pentium D (model 6) and a Xeon
117 	 * (model 2) with the same problem.
118 	 */
119 	if (c->x86 == 15) {
120 		u64 misc_enable;
121 
122 		rdmsrl(MSR_IA32_MISC_ENABLE, misc_enable);
123 
124 		if (misc_enable & MSR_IA32_MISC_ENABLE_FAST_STRING) {
125 			printk(KERN_INFO "kmemcheck: Disabling fast string operations\n");
126 
127 			misc_enable &= ~MSR_IA32_MISC_ENABLE_FAST_STRING;
128 			wrmsrl(MSR_IA32_MISC_ENABLE, misc_enable);
129 		}
130 	}
131 #endif
132 }
133 
134 #ifdef CONFIG_X86_32
135 /*
136  *	Early probe support logic for ppro memory erratum #50
137  *
138  *	This is called before we do cpu ident work
139  */
140 
141 int __cpuinit ppro_with_ram_bug(void)
142 {
143 	/* Uses data from early_cpu_detect now */
144 	if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
145 	    boot_cpu_data.x86 == 6 &&
146 	    boot_cpu_data.x86_model == 1 &&
147 	    boot_cpu_data.x86_mask < 8) {
148 		printk(KERN_INFO "Pentium Pro with Errata#50 detected. Taking evasive action.\n");
149 		return 1;
150 	}
151 	return 0;
152 }
153 
154 #ifdef CONFIG_X86_F00F_BUG
155 static void __cpuinit trap_init_f00f_bug(void)
156 {
157 	__set_fixmap(FIX_F00F_IDT, __pa(&idt_table), PAGE_KERNEL_RO);
158 
159 	/*
160 	 * Update the IDT descriptor and reload the IDT so that
161 	 * it uses the read-only mapped virtual address.
162 	 */
163 	idt_descr.address = fix_to_virt(FIX_F00F_IDT);
164 	load_idt(&idt_descr);
165 }
166 #endif
167 
168 static void __cpuinit intel_smp_check(struct cpuinfo_x86 *c)
169 {
170 #ifdef CONFIG_SMP
171 	/* calling is from identify_secondary_cpu() ? */
172 	if (c->cpu_index == boot_cpu_id)
173 		return;
174 
175 	/*
176 	 * Mask B, Pentium, but not Pentium MMX
177 	 */
178 	if (c->x86 == 5 &&
179 	    c->x86_mask >= 1 && c->x86_mask <= 4 &&
180 	    c->x86_model <= 3) {
181 		/*
182 		 * Remember we have B step Pentia with bugs
183 		 */
184 		WARN_ONCE(1, "WARNING: SMP operation may be unreliable"
185 				    "with B stepping processors.\n");
186 	}
187 #endif
188 }
189 
190 static void __cpuinit intel_workarounds(struct cpuinfo_x86 *c)
191 {
192 	unsigned long lo, hi;
193 
194 #ifdef CONFIG_X86_F00F_BUG
195 	/*
196 	 * All current models of Pentium and Pentium with MMX technology CPUs
197 	 * have the F0 0F bug, which lets nonprivileged users lock up the
198 	 * system.
199 	 * Note that the workaround only should be initialized once...
200 	 */
201 	c->f00f_bug = 0;
202 	if (!paravirt_enabled() && c->x86 == 5) {
203 		static int f00f_workaround_enabled;
204 
205 		c->f00f_bug = 1;
206 		if (!f00f_workaround_enabled) {
207 			trap_init_f00f_bug();
208 			printk(KERN_NOTICE "Intel Pentium with F0 0F bug - workaround enabled.\n");
209 			f00f_workaround_enabled = 1;
210 		}
211 	}
212 #endif
213 
214 	/*
215 	 * SEP CPUID bug: Pentium Pro reports SEP but doesn't have it until
216 	 * model 3 mask 3
217 	 */
218 	if ((c->x86<<8 | c->x86_model<<4 | c->x86_mask) < 0x633)
219 		clear_cpu_cap(c, X86_FEATURE_SEP);
220 
221 	/*
222 	 * P4 Xeon errata 037 workaround.
223 	 * Hardware prefetcher may cause stale data to be loaded into the cache.
224 	 */
225 	if ((c->x86 == 15) && (c->x86_model == 1) && (c->x86_mask == 1)) {
226 		rdmsr(MSR_IA32_MISC_ENABLE, lo, hi);
227 		if ((lo & MSR_IA32_MISC_ENABLE_PREFETCH_DISABLE) == 0) {
228 			printk (KERN_INFO "CPU: C0 stepping P4 Xeon detected.\n");
229 			printk (KERN_INFO "CPU: Disabling hardware prefetching (Errata 037)\n");
230 			lo |= MSR_IA32_MISC_ENABLE_PREFETCH_DISABLE;
231 			wrmsr(MSR_IA32_MISC_ENABLE, lo, hi);
232 		}
233 	}
234 
235 	/*
236 	 * See if we have a good local APIC by checking for buggy Pentia,
237 	 * i.e. all B steppings and the C2 stepping of P54C when using their
238 	 * integrated APIC (see 11AP erratum in "Pentium Processor
239 	 * Specification Update").
240 	 */
241 	if (cpu_has_apic && (c->x86<<8 | c->x86_model<<4) == 0x520 &&
242 	    (c->x86_mask < 0x6 || c->x86_mask == 0xb))
243 		set_cpu_cap(c, X86_FEATURE_11AP);
244 
245 
246 #ifdef CONFIG_X86_INTEL_USERCOPY
247 	/*
248 	 * Set up the preferred alignment for movsl bulk memory moves
249 	 */
250 	switch (c->x86) {
251 	case 4:		/* 486: untested */
252 		break;
253 	case 5:		/* Old Pentia: untested */
254 		break;
255 	case 6:		/* PII/PIII only like movsl with 8-byte alignment */
256 		movsl_mask.mask = 7;
257 		break;
258 	case 15:	/* P4 is OK down to 8-byte alignment */
259 		movsl_mask.mask = 7;
260 		break;
261 	}
262 #endif
263 
264 #ifdef CONFIG_X86_NUMAQ
265 	numaq_tsc_disable();
266 #endif
267 
268 	intel_smp_check(c);
269 }
270 #else
271 static void __cpuinit intel_workarounds(struct cpuinfo_x86 *c)
272 {
273 }
274 #endif
275 
276 static void __cpuinit srat_detect_node(struct cpuinfo_x86 *c)
277 {
278 #if defined(CONFIG_NUMA) && defined(CONFIG_X86_64)
279 	unsigned node;
280 	int cpu = smp_processor_id();
281 	int apicid = cpu_has_apic ? hard_smp_processor_id() : c->apicid;
282 
283 	/* Don't do the funky fallback heuristics the AMD version employs
284 	   for now. */
285 	node = apicid_to_node[apicid];
286 	if (node == NUMA_NO_NODE)
287 		node = first_node(node_online_map);
288 	else if (!node_online(node)) {
289 		/* reuse the value from init_cpu_to_node() */
290 		node = cpu_to_node(cpu);
291 	}
292 	numa_set_node(cpu, node);
293 #endif
294 }
295 
296 /*
297  * find out the number of processor cores on the die
298  */
299 static int __cpuinit intel_num_cpu_cores(struct cpuinfo_x86 *c)
300 {
301 	unsigned int eax, ebx, ecx, edx;
302 
303 	if (c->cpuid_level < 4)
304 		return 1;
305 
306 	/* Intel has a non-standard dependency on %ecx for this CPUID level. */
307 	cpuid_count(4, 0, &eax, &ebx, &ecx, &edx);
308 	if (eax & 0x1f)
309 		return (eax >> 26) + 1;
310 	else
311 		return 1;
312 }
313 
314 static void __cpuinit detect_vmx_virtcap(struct cpuinfo_x86 *c)
315 {
316 	/* Intel VMX MSR indicated features */
317 #define X86_VMX_FEATURE_PROC_CTLS_TPR_SHADOW	0x00200000
318 #define X86_VMX_FEATURE_PROC_CTLS_VNMI		0x00400000
319 #define X86_VMX_FEATURE_PROC_CTLS_2ND_CTLS	0x80000000
320 #define X86_VMX_FEATURE_PROC_CTLS2_VIRT_APIC	0x00000001
321 #define X86_VMX_FEATURE_PROC_CTLS2_EPT		0x00000002
322 #define X86_VMX_FEATURE_PROC_CTLS2_VPID		0x00000020
323 
324 	u32 vmx_msr_low, vmx_msr_high, msr_ctl, msr_ctl2;
325 
326 	clear_cpu_cap(c, X86_FEATURE_TPR_SHADOW);
327 	clear_cpu_cap(c, X86_FEATURE_VNMI);
328 	clear_cpu_cap(c, X86_FEATURE_FLEXPRIORITY);
329 	clear_cpu_cap(c, X86_FEATURE_EPT);
330 	clear_cpu_cap(c, X86_FEATURE_VPID);
331 
332 	rdmsr(MSR_IA32_VMX_PROCBASED_CTLS, vmx_msr_low, vmx_msr_high);
333 	msr_ctl = vmx_msr_high | vmx_msr_low;
334 	if (msr_ctl & X86_VMX_FEATURE_PROC_CTLS_TPR_SHADOW)
335 		set_cpu_cap(c, X86_FEATURE_TPR_SHADOW);
336 	if (msr_ctl & X86_VMX_FEATURE_PROC_CTLS_VNMI)
337 		set_cpu_cap(c, X86_FEATURE_VNMI);
338 	if (msr_ctl & X86_VMX_FEATURE_PROC_CTLS_2ND_CTLS) {
339 		rdmsr(MSR_IA32_VMX_PROCBASED_CTLS2,
340 		      vmx_msr_low, vmx_msr_high);
341 		msr_ctl2 = vmx_msr_high | vmx_msr_low;
342 		if ((msr_ctl2 & X86_VMX_FEATURE_PROC_CTLS2_VIRT_APIC) &&
343 		    (msr_ctl & X86_VMX_FEATURE_PROC_CTLS_TPR_SHADOW))
344 			set_cpu_cap(c, X86_FEATURE_FLEXPRIORITY);
345 		if (msr_ctl2 & X86_VMX_FEATURE_PROC_CTLS2_EPT)
346 			set_cpu_cap(c, X86_FEATURE_EPT);
347 		if (msr_ctl2 & X86_VMX_FEATURE_PROC_CTLS2_VPID)
348 			set_cpu_cap(c, X86_FEATURE_VPID);
349 	}
350 }
351 
352 static void __cpuinit init_intel(struct cpuinfo_x86 *c)
353 {
354 	unsigned int l2 = 0;
355 
356 	early_init_intel(c);
357 
358 	intel_workarounds(c);
359 
360 	/*
361 	 * Detect the extended topology information if available. This
362 	 * will reinitialise the initial_apicid which will be used
363 	 * in init_intel_cacheinfo()
364 	 */
365 	detect_extended_topology(c);
366 
367 	l2 = init_intel_cacheinfo(c);
368 	if (c->cpuid_level > 9) {
369 		unsigned eax = cpuid_eax(10);
370 		/* Check for version and the number of counters */
371 		if ((eax & 0xff) && (((eax>>8) & 0xff) > 1))
372 			set_cpu_cap(c, X86_FEATURE_ARCH_PERFMON);
373 	}
374 
375 	if (cpu_has_xmm2)
376 		set_cpu_cap(c, X86_FEATURE_LFENCE_RDTSC);
377 	if (cpu_has_ds) {
378 		unsigned int l1;
379 		rdmsr(MSR_IA32_MISC_ENABLE, l1, l2);
380 		if (!(l1 & (1<<11)))
381 			set_cpu_cap(c, X86_FEATURE_BTS);
382 		if (!(l1 & (1<<12)))
383 			set_cpu_cap(c, X86_FEATURE_PEBS);
384 	}
385 
386 	if (c->x86 == 6 && c->x86_model == 29 && cpu_has_clflush)
387 		set_cpu_cap(c, X86_FEATURE_CLFLUSH_MONITOR);
388 
389 #ifdef CONFIG_X86_64
390 	if (c->x86 == 15)
391 		c->x86_cache_alignment = c->x86_clflush_size * 2;
392 	if (c->x86 == 6)
393 		set_cpu_cap(c, X86_FEATURE_REP_GOOD);
394 #else
395 	/*
396 	 * Names for the Pentium II/Celeron processors
397 	 * detectable only by also checking the cache size.
398 	 * Dixon is NOT a Celeron.
399 	 */
400 	if (c->x86 == 6) {
401 		char *p = NULL;
402 
403 		switch (c->x86_model) {
404 		case 5:
405 			if (c->x86_mask == 0) {
406 				if (l2 == 0)
407 					p = "Celeron (Covington)";
408 				else if (l2 == 256)
409 					p = "Mobile Pentium II (Dixon)";
410 			}
411 			break;
412 
413 		case 6:
414 			if (l2 == 128)
415 				p = "Celeron (Mendocino)";
416 			else if (c->x86_mask == 0 || c->x86_mask == 5)
417 				p = "Celeron-A";
418 			break;
419 
420 		case 8:
421 			if (l2 == 128)
422 				p = "Celeron (Coppermine)";
423 			break;
424 		}
425 
426 		if (p)
427 			strcpy(c->x86_model_id, p);
428 	}
429 
430 	if (c->x86 == 15)
431 		set_cpu_cap(c, X86_FEATURE_P4);
432 	if (c->x86 == 6)
433 		set_cpu_cap(c, X86_FEATURE_P3);
434 #endif
435 
436 	if (!cpu_has(c, X86_FEATURE_XTOPOLOGY)) {
437 		/*
438 		 * let's use the legacy cpuid vector 0x1 and 0x4 for topology
439 		 * detection.
440 		 */
441 		c->x86_max_cores = intel_num_cpu_cores(c);
442 #ifdef CONFIG_X86_32
443 		detect_ht(c);
444 #endif
445 	}
446 
447 	/* Work around errata */
448 	srat_detect_node(c);
449 
450 	if (cpu_has(c, X86_FEATURE_VMX))
451 		detect_vmx_virtcap(c);
452 }
453 
454 #ifdef CONFIG_X86_32
455 static unsigned int __cpuinit intel_size_cache(struct cpuinfo_x86 *c, unsigned int size)
456 {
457 	/*
458 	 * Intel PIII Tualatin. This comes in two flavours.
459 	 * One has 256kb of cache, the other 512. We have no way
460 	 * to determine which, so we use a boottime override
461 	 * for the 512kb model, and assume 256 otherwise.
462 	 */
463 	if ((c->x86 == 6) && (c->x86_model == 11) && (size == 0))
464 		size = 256;
465 	return size;
466 }
467 #endif
468 
469 static const struct cpu_dev __cpuinitconst intel_cpu_dev = {
470 	.c_vendor	= "Intel",
471 	.c_ident	= { "GenuineIntel" },
472 #ifdef CONFIG_X86_32
473 	.c_models = {
474 		{ .vendor = X86_VENDOR_INTEL, .family = 4, .model_names =
475 		  {
476 			  [0] = "486 DX-25/33",
477 			  [1] = "486 DX-50",
478 			  [2] = "486 SX",
479 			  [3] = "486 DX/2",
480 			  [4] = "486 SL",
481 			  [5] = "486 SX/2",
482 			  [7] = "486 DX/2-WB",
483 			  [8] = "486 DX/4",
484 			  [9] = "486 DX/4-WB"
485 		  }
486 		},
487 		{ .vendor = X86_VENDOR_INTEL, .family = 5, .model_names =
488 		  {
489 			  [0] = "Pentium 60/66 A-step",
490 			  [1] = "Pentium 60/66",
491 			  [2] = "Pentium 75 - 200",
492 			  [3] = "OverDrive PODP5V83",
493 			  [4] = "Pentium MMX",
494 			  [7] = "Mobile Pentium 75 - 200",
495 			  [8] = "Mobile Pentium MMX"
496 		  }
497 		},
498 		{ .vendor = X86_VENDOR_INTEL, .family = 6, .model_names =
499 		  {
500 			  [0] = "Pentium Pro A-step",
501 			  [1] = "Pentium Pro",
502 			  [3] = "Pentium II (Klamath)",
503 			  [4] = "Pentium II (Deschutes)",
504 			  [5] = "Pentium II (Deschutes)",
505 			  [6] = "Mobile Pentium II",
506 			  [7] = "Pentium III (Katmai)",
507 			  [8] = "Pentium III (Coppermine)",
508 			  [10] = "Pentium III (Cascades)",
509 			  [11] = "Pentium III (Tualatin)",
510 		  }
511 		},
512 		{ .vendor = X86_VENDOR_INTEL, .family = 15, .model_names =
513 		  {
514 			  [0] = "Pentium 4 (Unknown)",
515 			  [1] = "Pentium 4 (Willamette)",
516 			  [2] = "Pentium 4 (Northwood)",
517 			  [4] = "Pentium 4 (Foster)",
518 			  [5] = "Pentium 4 (Foster)",
519 		  }
520 		},
521 	},
522 	.c_size_cache	= intel_size_cache,
523 #endif
524 	.c_early_init   = early_init_intel,
525 	.c_init		= init_intel,
526 	.c_x86_vendor	= X86_VENDOR_INTEL,
527 };
528 
529 cpu_dev_register(intel_cpu_dev);
530 
531