xref: /titanic_50/usr/src/uts/i86pc/os/cpuid.c (revision c73a93f2f62aa040c7fc1755230b787b0ad03967)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2006 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #pragma ident	"%Z%%M%	%I%	%E% SMI"
27 
28 /*
29  * Various routines to handle identification
30  * and classification of x86 processors.
31  */
32 
33 #include <sys/types.h>
34 #include <sys/archsystm.h>
35 #include <sys/x86_archext.h>
36 #include <sys/kmem.h>
37 #include <sys/systm.h>
38 #include <sys/cmn_err.h>
39 #include <sys/sunddi.h>
40 #include <sys/sunndi.h>
41 #include <sys/cpuvar.h>
42 #include <sys/processor.h>
43 #include <sys/chip.h>
44 #include <sys/fp.h>
45 #include <sys/controlregs.h>
46 #include <sys/auxv_386.h>
47 #include <sys/bitmap.h>
48 #include <sys/controlregs.h>
49 #include <sys/memnode.h>
50 
51 /*
52  * Pass 0 of cpuid feature analysis happens in locore. It contains special code
53  * to recognize Cyrix processors that are not cpuid-compliant, and to deal with
54  * them accordingly. For most modern processors, feature detection occurs here
55  * in pass 1.
56  *
57  * Pass 1 of cpuid feature analysis happens just at the beginning of mlsetup()
58  * for the boot CPU and does the basic analysis that the early kernel needs.
59  * x86_feature is set based on the return value of cpuid_pass1() of the boot
60  * CPU.
61  *
62  * Pass 1 includes:
63  *
64  *	o Determining vendor/model/family/stepping and setting x86_type and
65  *	  x86_vendor accordingly.
66  *	o Processing the feature flags returned by the cpuid instruction while
67  *	  applying any workarounds or tricks for the specific processor.
68  *	o Mapping the feature flags into Solaris feature bits (X86_*).
69  *	o Processing extended feature flags if supported by the processor,
70  *	  again while applying specific processor knowledge.
71  *	o Determining the CMT characteristics of the system.
72  *
73  * Pass 1 is done on non-boot CPUs during their initialization and the results
74  * are used only as a meager attempt at ensuring that all processors within the
75  * system support the same features.
76  *
77  * Pass 2 of cpuid feature analysis happens just at the beginning
78  * of startup().  It just copies in and corrects the remainder
79  * of the cpuid data we depend on: standard cpuid functions that we didn't
80  * need for pass1 feature analysis, and extended cpuid functions beyond the
81  * simple feature processing done in pass1.
82  *
83  * Pass 3 of cpuid analysis is invoked after basic kernel services; in
84  * particular kernel memory allocation has been made available. It creates a
85  * readable brand string based on the data collected in the first two passes.
86  *
87  * Pass 4 of cpuid analysis is invoked after post_startup() when all
88  * the support infrastructure for various hardware features has been
89  * initialized. It determines which processor features will be reported
90  * to userland via the aux vector.
91  *
92  * All passes are executed on all CPUs, but only the boot CPU determines what
93  * features the kernel will use.
94  *
95  * Much of the worst junk in this file is for the support of processors
96  * that didn't really implement the cpuid instruction properly.
97  *
98  * NOTE: The accessor functions (cpuid_get*) are aware of, and ASSERT upon,
99  * the pass numbers.  Accordingly, changes to the pass code may require changes
100  * to the accessor code.
101  */
102 
103 uint_t x86_feature = 0;
104 uint_t x86_vendor = X86_VENDOR_IntelClone;
105 uint_t x86_type = X86_TYPE_OTHER;
106 
107 ulong_t cr4_value;
108 uint_t pentiumpro_bug4046376;
109 uint_t pentiumpro_bug4064495;
110 
111 uint_t enable486;
112 
113 /*
114  * This set of strings are for processors rumored to support the cpuid
115  * instruction, and is used by locore.s to figure out how to set x86_vendor
116  */
117 const char CyrixInstead[] = "CyrixInstead";
118 
119 /*
120  * These constants determine how many of the elements of the
121  * cpuid we cache in the cpuid_info data structure; the
122  * remaining elements are accessible via the cpuid instruction.
123  */
124 
125 #define	NMAX_CPI_STD	6		/* eax = 0 .. 5 */
126 #define	NMAX_CPI_EXTD	9		/* eax = 0x80000000 .. 0x80000008 */
127 
128 struct cpuid_info {
129 	uint_t cpi_pass;		/* last pass completed */
130 	/*
131 	 * standard function information
132 	 */
133 	uint_t cpi_maxeax;		/* fn 0: %eax */
134 	char cpi_vendorstr[13];		/* fn 0: %ebx:%ecx:%edx */
135 	uint_t cpi_vendor;		/* enum of cpi_vendorstr */
136 
137 	uint_t cpi_family;		/* fn 1: extended family */
138 	uint_t cpi_model;		/* fn 1: extended model */
139 	uint_t cpi_step;		/* fn 1: stepping */
140 	chipid_t cpi_chipid;		/* fn 1: %ebx: chip # on ht cpus */
141 	uint_t cpi_brandid;		/* fn 1: %ebx: brand ID */
142 	int cpi_clogid;			/* fn 1: %ebx: thread # */
143 	uint_t cpi_ncpu_per_chip;	/* fn 1: %ebx: logical cpu count */
144 	uint8_t cpi_cacheinfo[16];	/* fn 2: intel-style cache desc */
145 	uint_t cpi_ncache;		/* fn 2: number of elements */
146 	struct cpuid_regs cpi_std[NMAX_CPI_STD];	/* 0 .. 5 */
147 	/*
148 	 * extended function information
149 	 */
150 	uint_t cpi_xmaxeax;		/* fn 0x80000000: %eax */
151 	char cpi_brandstr[49];		/* fn 0x8000000[234] */
152 	uint8_t cpi_pabits;		/* fn 0x80000006: %eax */
153 	uint8_t cpi_vabits;		/* fn 0x80000006: %eax */
154 	struct cpuid_regs cpi_extd[NMAX_CPI_EXTD]; /* 0x8000000[0-8] */
155 	id_t cpi_coreid;
156 	uint_t cpi_ncore_per_chip;	/* AMD: fn 0x80000008: %ecx[7-0] */
157 					/* Intel: fn 4: %eax[31-26] */
158 	/*
159 	 * supported feature information
160 	 */
161 	uint32_t cpi_support[4];
162 #define	STD_EDX_FEATURES	0
163 #define	AMD_EDX_FEATURES	1
164 #define	TM_EDX_FEATURES		2
165 #define	STD_ECX_FEATURES	3
166 
167 };
168 
169 
170 static struct cpuid_info cpuid_info0;
171 
172 /*
173  * These bit fields are defined by the Intel Application Note AP-485
174  * "Intel Processor Identification and the CPUID Instruction"
175  */
176 #define	CPI_FAMILY_XTD(cpi)	BITX((cpi)->cpi_std[1].cp_eax, 27, 20)
177 #define	CPI_MODEL_XTD(cpi)	BITX((cpi)->cpi_std[1].cp_eax, 19, 16)
178 #define	CPI_TYPE(cpi)		BITX((cpi)->cpi_std[1].cp_eax, 13, 12)
179 #define	CPI_FAMILY(cpi)		BITX((cpi)->cpi_std[1].cp_eax, 11, 8)
180 #define	CPI_STEP(cpi)		BITX((cpi)->cpi_std[1].cp_eax, 3, 0)
181 #define	CPI_MODEL(cpi)		BITX((cpi)->cpi_std[1].cp_eax, 7, 4)
182 
183 #define	CPI_FEATURES_EDX(cpi)		((cpi)->cpi_std[1].cp_edx)
184 #define	CPI_FEATURES_ECX(cpi)		((cpi)->cpi_std[1].cp_ecx)
185 #define	CPI_FEATURES_XTD_EDX(cpi)	((cpi)->cpi_extd[1].cp_edx)
186 #define	CPI_FEATURES_XTD_ECX(cpi)	((cpi)->cpi_extd[1].cp_ecx)
187 
188 #define	CPI_BRANDID(cpi)	BITX((cpi)->cpi_std[1].cp_ebx, 7, 0)
189 #define	CPI_CHUNKS(cpi)		BITX((cpi)->cpi_std[1].cp_ebx, 15, 7)
190 #define	CPI_CPU_COUNT(cpi)	BITX((cpi)->cpi_std[1].cp_ebx, 23, 16)
191 #define	CPI_APIC_ID(cpi)	BITX((cpi)->cpi_std[1].cp_ebx, 31, 24)
192 
193 #define	CPI_MAXEAX_MAX		0x100		/* sanity control */
194 #define	CPI_XMAXEAX_MAX		0x80000100
195 
196 /*
197  * A couple of shorthand macros to identify "later" P6-family chips
198  * like the Pentium M and Core.  First, the "older" P6-based stuff
199  * (loosely defined as "pre-Pentium-4"):
200  * P6, PII, Mobile PII, PII Xeon, PIII, Mobile PIII, PIII Xeon
201  */
202 
203 #define	IS_LEGACY_P6(cpi) (			\
204 	cpi->cpi_family == 6 && 		\
205 		(cpi->cpi_model == 1 ||		\
206 		cpi->cpi_model == 3 ||		\
207 		cpi->cpi_model == 5 ||		\
208 		cpi->cpi_model == 6 ||		\
209 		cpi->cpi_model == 7 ||		\
210 		cpi->cpi_model == 8 ||		\
211 		cpi->cpi_model == 0xA ||	\
212 		cpi->cpi_model == 0xB)		\
213 )
214 
215 /* A "new F6" is everything with family 6 that's not the above */
216 #define	IS_NEW_F6(cpi) ((cpi->cpi_family == 6) && !IS_LEGACY_P6(cpi))
217 
218 /*
219  *  Some undocumented ways of patching the results of the cpuid
220  *  instruction to permit running Solaris 10 on future cpus that
221  *  we don't currently support.  Could be set to non-zero values
222  *  via settings in eeprom.
223  */
224 
225 uint32_t cpuid_feature_ecx_include;
226 uint32_t cpuid_feature_ecx_exclude;
227 uint32_t cpuid_feature_edx_include;
228 uint32_t cpuid_feature_edx_exclude;
229 
230 uint_t
231 cpuid_pass1(cpu_t *cpu)
232 {
233 	uint32_t mask_ecx, mask_edx;
234 	uint_t feature = X86_CPUID;
235 	struct cpuid_info *cpi;
236 	struct cpuid_regs *cp;
237 	int xcpuid;
238 
239 	/*
240 	 * By convention, cpu0 is the boot cpu, which is called
241 	 * before memory allocation is available.  Other cpus are
242 	 * initialized when memory becomes available.
243 	 */
244 	if (cpu->cpu_id == 0)
245 		cpu->cpu_m.mcpu_cpi = cpi = &cpuid_info0;
246 	else
247 		cpu->cpu_m.mcpu_cpi = cpi =
248 		    kmem_zalloc(sizeof (*cpi), KM_SLEEP);
249 
250 	cp = &cpi->cpi_std[0];
251 	cp->cp_eax = 0;
252 	cpi->cpi_maxeax = __cpuid_insn(cp);
253 	{
254 		uint32_t *iptr = (uint32_t *)cpi->cpi_vendorstr;
255 		*iptr++ = cp->cp_ebx;
256 		*iptr++ = cp->cp_edx;
257 		*iptr++ = cp->cp_ecx;
258 		*(char *)&cpi->cpi_vendorstr[12] = '\0';
259 	}
260 
261 	/*
262 	 * Map the vendor string to a type code
263 	 */
264 	if (strcmp(cpi->cpi_vendorstr, "GenuineIntel") == 0)
265 		cpi->cpi_vendor = X86_VENDOR_Intel;
266 	else if (strcmp(cpi->cpi_vendorstr, "AuthenticAMD") == 0)
267 		cpi->cpi_vendor = X86_VENDOR_AMD;
268 	else if (strcmp(cpi->cpi_vendorstr, "GenuineTMx86") == 0)
269 		cpi->cpi_vendor = X86_VENDOR_TM;
270 	else if (strcmp(cpi->cpi_vendorstr, CyrixInstead) == 0)
271 		/*
272 		 * CyrixInstead is a variable used by the Cyrix detection code
273 		 * in locore.
274 		 */
275 		cpi->cpi_vendor = X86_VENDOR_Cyrix;
276 	else if (strcmp(cpi->cpi_vendorstr, "UMC UMC UMC ") == 0)
277 		cpi->cpi_vendor = X86_VENDOR_UMC;
278 	else if (strcmp(cpi->cpi_vendorstr, "NexGenDriven") == 0)
279 		cpi->cpi_vendor = X86_VENDOR_NexGen;
280 	else if (strcmp(cpi->cpi_vendorstr, "CentaurHauls") == 0)
281 		cpi->cpi_vendor = X86_VENDOR_Centaur;
282 	else if (strcmp(cpi->cpi_vendorstr, "RiseRiseRise") == 0)
283 		cpi->cpi_vendor = X86_VENDOR_Rise;
284 	else if (strcmp(cpi->cpi_vendorstr, "SiS SiS SiS ") == 0)
285 		cpi->cpi_vendor = X86_VENDOR_SiS;
286 	else if (strcmp(cpi->cpi_vendorstr, "Geode by NSC") == 0)
287 		cpi->cpi_vendor = X86_VENDOR_NSC;
288 	else
289 		cpi->cpi_vendor = X86_VENDOR_IntelClone;
290 
291 	x86_vendor = cpi->cpi_vendor; /* for compatibility */
292 
293 	/*
294 	 * Limit the range in case of weird hardware
295 	 */
296 	if (cpi->cpi_maxeax > CPI_MAXEAX_MAX)
297 		cpi->cpi_maxeax = CPI_MAXEAX_MAX;
298 	if (cpi->cpi_maxeax < 1)
299 		goto pass1_done;
300 
301 	cp = &cpi->cpi_std[1];
302 	cp->cp_eax = 1;
303 	(void) __cpuid_insn(cp);
304 
305 	/*
306 	 * Extract identifying constants for easy access.
307 	 */
308 	cpi->cpi_model = CPI_MODEL(cpi);
309 	cpi->cpi_family = CPI_FAMILY(cpi);
310 
311 	if (cpi->cpi_family == 0xf)
312 		cpi->cpi_family += CPI_FAMILY_XTD(cpi);
313 
314 	/*
315 	 * Beware: AMD uses "extended model" iff *FAMILY* == 0xf.
316 	 * Intel, and presumably everyone else, uses model == 0xf, as
317 	 * one would expect (max value means possible overflow).  Sigh.
318 	 */
319 
320 	switch (cpi->cpi_vendor) {
321 	case X86_VENDOR_AMD:
322 		if (cpi->cpi_family == 0xf)
323 			cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4;
324 		break;
325 	default:
326 		if (cpi->cpi_model == 0xf)
327 			cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4;
328 		break;
329 	}
330 
331 	cpi->cpi_step = CPI_STEP(cpi);
332 	cpi->cpi_brandid = CPI_BRANDID(cpi);
333 
334 	/*
335 	 * *default* assumptions:
336 	 * - believe %edx feature word
337 	 * - ignore %ecx feature word
338 	 * - 32-bit virtual and physical addressing
339 	 */
340 	mask_edx = 0xffffffff;
341 	mask_ecx = 0;
342 
343 	cpi->cpi_pabits = cpi->cpi_vabits = 32;
344 
345 	switch (cpi->cpi_vendor) {
346 	case X86_VENDOR_Intel:
347 		if (cpi->cpi_family == 5)
348 			x86_type = X86_TYPE_P5;
349 		else if (IS_LEGACY_P6(cpi)) {
350 			x86_type = X86_TYPE_P6;
351 			pentiumpro_bug4046376 = 1;
352 			pentiumpro_bug4064495 = 1;
353 			/*
354 			 * Clear the SEP bit when it was set erroneously
355 			 */
356 			if (cpi->cpi_model < 3 && cpi->cpi_step < 3)
357 				cp->cp_edx &= ~CPUID_INTC_EDX_SEP;
358 		} else if (IS_NEW_F6(cpi) || cpi->cpi_family == 0xf) {
359 			x86_type = X86_TYPE_P4;
360 			/*
361 			 * We don't currently depend on any of the %ecx
362 			 * features until Prescott, so we'll only check
363 			 * this from P4 onwards.  We might want to revisit
364 			 * that idea later.
365 			 */
366 			mask_ecx = 0xffffffff;
367 		} else if (cpi->cpi_family > 0xf)
368 			mask_ecx = 0xffffffff;
369 		break;
370 	case X86_VENDOR_IntelClone:
371 	default:
372 		break;
373 	case X86_VENDOR_AMD:
374 #if defined(OPTERON_ERRATUM_108)
375 		if (cpi->cpi_family == 0xf && cpi->cpi_model == 0xe) {
376 			cp->cp_eax = (0xf0f & cp->cp_eax) | 0xc0;
377 			cpi->cpi_model = 0xc;
378 		} else
379 #endif
380 		if (cpi->cpi_family == 5) {
381 			/*
382 			 * AMD K5 and K6
383 			 *
384 			 * These CPUs have an incomplete implementation
385 			 * of MCA/MCE which we mask away.
386 			 */
387 			mask_edx &= ~(CPUID_INTC_EDX_MCE | CPUID_INTC_EDX_MCA);
388 
389 			/*
390 			 * Model 0 uses the wrong (APIC) bit
391 			 * to indicate PGE.  Fix it here.
392 			 */
393 			if (cpi->cpi_model == 0) {
394 				if (cp->cp_edx & 0x200) {
395 					cp->cp_edx &= ~0x200;
396 					cp->cp_edx |= CPUID_INTC_EDX_PGE;
397 				}
398 			}
399 
400 			/*
401 			 * Early models had problems w/ MMX; disable.
402 			 */
403 			if (cpi->cpi_model < 6)
404 				mask_edx &= ~CPUID_INTC_EDX_MMX;
405 		}
406 
407 		/*
408 		 * For newer families, SSE3 and CX16, at least, are valid;
409 		 * enable all
410 		 */
411 		if (cpi->cpi_family >= 0xf)
412 			mask_ecx = 0xffffffff;
413 		break;
414 	case X86_VENDOR_TM:
415 		/*
416 		 * workaround the NT workaround in CMS 4.1
417 		 */
418 		if (cpi->cpi_family == 5 && cpi->cpi_model == 4 &&
419 		    (cpi->cpi_step == 2 || cpi->cpi_step == 3))
420 			cp->cp_edx |= CPUID_INTC_EDX_CX8;
421 		break;
422 	case X86_VENDOR_Centaur:
423 		/*
424 		 * workaround the NT workarounds again
425 		 */
426 		if (cpi->cpi_family == 6)
427 			cp->cp_edx |= CPUID_INTC_EDX_CX8;
428 		break;
429 	case X86_VENDOR_Cyrix:
430 		/*
431 		 * We rely heavily on the probing in locore
432 		 * to actually figure out what parts, if any,
433 		 * of the Cyrix cpuid instruction to believe.
434 		 */
435 		switch (x86_type) {
436 		case X86_TYPE_CYRIX_486:
437 			mask_edx = 0;
438 			break;
439 		case X86_TYPE_CYRIX_6x86:
440 			mask_edx = 0;
441 			break;
442 		case X86_TYPE_CYRIX_6x86L:
443 			mask_edx =
444 			    CPUID_INTC_EDX_DE |
445 			    CPUID_INTC_EDX_CX8;
446 			break;
447 		case X86_TYPE_CYRIX_6x86MX:
448 			mask_edx =
449 			    CPUID_INTC_EDX_DE |
450 			    CPUID_INTC_EDX_MSR |
451 			    CPUID_INTC_EDX_CX8 |
452 			    CPUID_INTC_EDX_PGE |
453 			    CPUID_INTC_EDX_CMOV |
454 			    CPUID_INTC_EDX_MMX;
455 			break;
456 		case X86_TYPE_CYRIX_GXm:
457 			mask_edx =
458 			    CPUID_INTC_EDX_MSR |
459 			    CPUID_INTC_EDX_CX8 |
460 			    CPUID_INTC_EDX_CMOV |
461 			    CPUID_INTC_EDX_MMX;
462 			break;
463 		case X86_TYPE_CYRIX_MediaGX:
464 			break;
465 		case X86_TYPE_CYRIX_MII:
466 		case X86_TYPE_VIA_CYRIX_III:
467 			mask_edx =
468 			    CPUID_INTC_EDX_DE |
469 			    CPUID_INTC_EDX_TSC |
470 			    CPUID_INTC_EDX_MSR |
471 			    CPUID_INTC_EDX_CX8 |
472 			    CPUID_INTC_EDX_PGE |
473 			    CPUID_INTC_EDX_CMOV |
474 			    CPUID_INTC_EDX_MMX;
475 			break;
476 		default:
477 			break;
478 		}
479 		break;
480 	}
481 
482 	/*
483 	 * Now we've figured out the masks that determine
484 	 * which bits we choose to believe, apply the masks
485 	 * to the feature words, then map the kernel's view
486 	 * of these feature words into its feature word.
487 	 */
488 	cp->cp_edx &= mask_edx;
489 	cp->cp_ecx &= mask_ecx;
490 
491 	/*
492 	 * fold in fix ups
493 	 */
494 
495 	cp->cp_edx |= cpuid_feature_edx_include;
496 	cp->cp_edx &= ~cpuid_feature_edx_exclude;
497 
498 
499 	cp->cp_ecx |= cpuid_feature_ecx_include;
500 	cp->cp_ecx &= ~cpuid_feature_ecx_exclude;
501 
502 	if (cp->cp_edx & CPUID_INTC_EDX_PSE)
503 		feature |= X86_LARGEPAGE;
504 	if (cp->cp_edx & CPUID_INTC_EDX_TSC)
505 		feature |= X86_TSC;
506 	if (cp->cp_edx & CPUID_INTC_EDX_MSR)
507 		feature |= X86_MSR;
508 	if (cp->cp_edx & CPUID_INTC_EDX_MTRR)
509 		feature |= X86_MTRR;
510 	if (cp->cp_edx & CPUID_INTC_EDX_PGE)
511 		feature |= X86_PGE;
512 	if (cp->cp_edx & CPUID_INTC_EDX_CMOV)
513 		feature |= X86_CMOV;
514 	if (cp->cp_edx & CPUID_INTC_EDX_MMX)
515 		feature |= X86_MMX;
516 	if ((cp->cp_edx & CPUID_INTC_EDX_MCE) != 0 &&
517 	    (cp->cp_edx & CPUID_INTC_EDX_MCA) != 0)
518 		feature |= X86_MCA;
519 	if (cp->cp_edx & CPUID_INTC_EDX_PAE)
520 		feature |= X86_PAE;
521 	if (cp->cp_edx & CPUID_INTC_EDX_CX8)
522 		feature |= X86_CX8;
523 	/*
524 	 * Once this bit was thought questionable, but it looks like it's
525 	 * back, as of Application Note 485 March 2005 (24161829.pdf)
526 	 */
527 	if (cp->cp_ecx & CPUID_INTC_ECX_CX16)
528 		feature |= X86_CX16;
529 	if (cp->cp_edx & CPUID_INTC_EDX_PAT)
530 		feature |= X86_PAT;
531 	if (cp->cp_edx & CPUID_INTC_EDX_SEP)
532 		feature |= X86_SEP;
533 	if (cp->cp_edx & CPUID_INTC_EDX_FXSR) {
534 		/*
535 		 * In our implementation, fxsave/fxrstor
536 		 * are prerequisites before we'll even
537 		 * try and do SSE things.
538 		 */
539 		if (cp->cp_edx & CPUID_INTC_EDX_SSE)
540 			feature |= X86_SSE;
541 		if (cp->cp_edx & CPUID_INTC_EDX_SSE2)
542 			feature |= X86_SSE2;
543 		if (cp->cp_ecx & CPUID_INTC_ECX_SSE3)
544 			feature |= X86_SSE3;
545 	}
546 	if (cp->cp_edx & CPUID_INTC_EDX_DE)
547 		cr4_value |= CR4_DE;
548 
549 	if (feature & X86_PAE)
550 		cpi->cpi_pabits = 36;
551 
552 	/*
553 	 * Hyperthreading configuration is slightly tricky on Intel
554 	 * and pure clones, and even trickier on AMD.
555 	 *
556 	 * (AMD chose to set the HTT bit on their CMP processors,
557 	 * even though they're not actually hyperthreaded.  Thus it
558 	 * takes a bit more work to figure out what's really going
559 	 * on ... see the handling of the CMP_LEGACY bit below)
560 	 */
561 	if (cp->cp_edx & CPUID_INTC_EDX_HTT) {
562 		cpi->cpi_ncpu_per_chip = CPI_CPU_COUNT(cpi);
563 		if (cpi->cpi_ncpu_per_chip > 1)
564 			feature |= X86_HTT;
565 	} else {
566 		cpi->cpi_ncpu_per_chip = 1;
567 	}
568 
569 	/*
570 	 * Work on the "extended" feature information, doing
571 	 * some basic initialization for cpuid_pass2()
572 	 */
573 	xcpuid = 0;
574 	switch (cpi->cpi_vendor) {
575 	case X86_VENDOR_Intel:
576 		if (IS_NEW_F6(cpi) || cpi->cpi_family >= 0xf)
577 			xcpuid++;
578 		break;
579 	case X86_VENDOR_AMD:
580 		if (cpi->cpi_family > 5 ||
581 		    (cpi->cpi_family == 5 && cpi->cpi_model >= 1))
582 			xcpuid++;
583 		break;
584 	case X86_VENDOR_Cyrix:
585 		/*
586 		 * Only these Cyrix CPUs are -known- to support
587 		 * extended cpuid operations.
588 		 */
589 		if (x86_type == X86_TYPE_VIA_CYRIX_III ||
590 		    x86_type == X86_TYPE_CYRIX_GXm)
591 			xcpuid++;
592 		break;
593 	case X86_VENDOR_Centaur:
594 	case X86_VENDOR_TM:
595 	default:
596 		xcpuid++;
597 		break;
598 	}
599 
600 	if (xcpuid) {
601 		cp = &cpi->cpi_extd[0];
602 		cp->cp_eax = 0x80000000;
603 		cpi->cpi_xmaxeax = __cpuid_insn(cp);
604 	}
605 
606 	if (cpi->cpi_xmaxeax & 0x80000000) {
607 
608 		if (cpi->cpi_xmaxeax > CPI_XMAXEAX_MAX)
609 			cpi->cpi_xmaxeax = CPI_XMAXEAX_MAX;
610 
611 		switch (cpi->cpi_vendor) {
612 		case X86_VENDOR_Intel:
613 		case X86_VENDOR_AMD:
614 			if (cpi->cpi_xmaxeax < 0x80000001)
615 				break;
616 			cp = &cpi->cpi_extd[1];
617 			cp->cp_eax = 0x80000001;
618 			(void) __cpuid_insn(cp);
619 			if (cpi->cpi_vendor == X86_VENDOR_AMD &&
620 			    cpi->cpi_family == 5 &&
621 			    cpi->cpi_model == 6 &&
622 			    cpi->cpi_step == 6) {
623 				/*
624 				 * K6 model 6 uses bit 10 to indicate SYSC
625 				 * Later models use bit 11. Fix it here.
626 				 */
627 				if (cp->cp_edx & 0x400) {
628 					cp->cp_edx &= ~0x400;
629 					cp->cp_edx |= CPUID_AMD_EDX_SYSC;
630 				}
631 			}
632 
633 			/*
634 			 * Compute the additions to the kernel's feature word.
635 			 */
636 			if (cp->cp_edx & CPUID_AMD_EDX_NX)
637 				feature |= X86_NX;
638 
639 			/*
640 			 * If both the HTT and CMP_LEGACY bits are set,
641 			 * then we're not actually HyperThreaded.  Read
642 			 * "AMD CPUID Specification" for more details.
643 			 */
644 			if (cpi->cpi_vendor == X86_VENDOR_AMD &&
645 			    (feature & X86_HTT) &&
646 			    (cp->cp_ecx & CPUID_AMD_ECX_CMP_LEGACY)) {
647 				feature &= ~X86_HTT;
648 				feature |= X86_CMP;
649 			}
650 #if defined(_LP64)
651 			/*
652 			 * It's really tricky to support syscall/sysret in
653 			 * the i386 kernel; we rely on sysenter/sysexit
654 			 * instead.  In the amd64 kernel, things are -way-
655 			 * better.
656 			 */
657 			if (cp->cp_edx & CPUID_AMD_EDX_SYSC)
658 				feature |= X86_ASYSC;
659 
660 			/*
661 			 * While we're thinking about system calls, note
662 			 * that AMD processors don't support sysenter
663 			 * in long mode at all, so don't try to program them.
664 			 */
665 			if (x86_vendor == X86_VENDOR_AMD)
666 				feature &= ~X86_SEP;
667 #endif
668 			break;
669 		default:
670 			break;
671 		}
672 
673 		/*
674 		 * Get CPUID data about processor cores and hyperthreads.
675 		 */
676 		switch (cpi->cpi_vendor) {
677 		case X86_VENDOR_Intel:
678 			if (cpi->cpi_maxeax >= 4) {
679 				cp = &cpi->cpi_std[4];
680 				cp->cp_eax = 4;
681 				cp->cp_ecx = 0;
682 				(void) __cpuid_insn(cp);
683 			}
684 			/*FALLTHROUGH*/
685 		case X86_VENDOR_AMD:
686 			if (cpi->cpi_xmaxeax < 0x80000008)
687 				break;
688 			cp = &cpi->cpi_extd[8];
689 			cp->cp_eax = 0x80000008;
690 			(void) __cpuid_insn(cp);
691 			/*
692 			 * Virtual and physical address limits from
693 			 * cpuid override previously guessed values.
694 			 */
695 			cpi->cpi_pabits = BITX(cp->cp_eax, 7, 0);
696 			cpi->cpi_vabits = BITX(cp->cp_eax, 15, 8);
697 			break;
698 		default:
699 			break;
700 		}
701 
702 		switch (cpi->cpi_vendor) {
703 		case X86_VENDOR_Intel:
704 			if (cpi->cpi_maxeax < 4) {
705 				cpi->cpi_ncore_per_chip = 1;
706 				break;
707 			} else {
708 				cpi->cpi_ncore_per_chip =
709 				    BITX((cpi)->cpi_std[4].cp_eax, 31, 26) + 1;
710 			}
711 			break;
712 		case X86_VENDOR_AMD:
713 			if (cpi->cpi_xmaxeax < 0x80000008) {
714 				cpi->cpi_ncore_per_chip = 1;
715 				break;
716 			} else {
717 				cpi->cpi_ncore_per_chip =
718 				    BITX((cpi)->cpi_extd[8].cp_ecx, 7, 0) + 1;
719 			}
720 			break;
721 		default:
722 			cpi->cpi_ncore_per_chip = 1;
723 			break;
724 		}
725 
726 	}
727 
728 	/*
729 	 * If more than one core, then this processor is CMP.
730 	 */
731 	if (cpi->cpi_ncore_per_chip > 1)
732 		feature |= X86_CMP;
733 	/*
734 	 * If the number of cores is the same as the number
735 	 * of CPUs, then we cannot have HyperThreading.
736 	 */
737 	if (cpi->cpi_ncpu_per_chip == cpi->cpi_ncore_per_chip)
738 		feature &= ~X86_HTT;
739 
740 	if ((feature & (X86_HTT | X86_CMP)) == 0) {
741 		/*
742 		 * Single-core single-threaded processors.
743 		 */
744 		cpi->cpi_chipid = -1;
745 		cpi->cpi_clogid = 0;
746 		cpi->cpi_coreid = cpu->cpu_id;
747 	} else if (cpi->cpi_ncpu_per_chip > 1) {
748 		uint_t i;
749 		uint_t chipid_shift = 0;
750 		uint_t coreid_shift = 0;
751 		uint_t apic_id = CPI_APIC_ID(cpi);
752 
753 		for (i = 1; i < cpi->cpi_ncpu_per_chip; i <<= 1)
754 			chipid_shift++;
755 		cpi->cpi_chipid = apic_id >> chipid_shift;
756 		cpi->cpi_clogid = apic_id & ((1 << chipid_shift) - 1);
757 
758 		if (cpi->cpi_vendor == X86_VENDOR_Intel) {
759 			if (feature & X86_CMP) {
760 				/*
761 				 * Multi-core (and possibly multi-threaded)
762 				 * processors.
763 				 */
764 				uint_t ncpu_per_core;
765 				if (cpi->cpi_ncore_per_chip == 1)
766 					ncpu_per_core = cpi->cpi_ncpu_per_chip;
767 				else if (cpi->cpi_ncore_per_chip > 1)
768 					ncpu_per_core = cpi->cpi_ncpu_per_chip /
769 					    cpi->cpi_ncore_per_chip;
770 				/*
771 				 * 8bit APIC IDs on dual core Pentiums
772 				 * look like this:
773 				 *
774 				 * +-----------------------+------+------+
775 				 * | Physical Package ID   |  MC  |  HT  |
776 				 * +-----------------------+------+------+
777 				 * <------- chipid -------->
778 				 * <------- coreid --------------->
779 				 *			   <--- clogid -->
780 				 *
781 				 * Where the number of bits necessary to
782 				 * represent MC and HT fields together equals
783 				 * to the minimum number of bits necessary to
784 				 * store the value of cpi->cpi_ncpu_per_chip.
785 				 * Of those bits, the MC part uses the number
786 				 * of bits necessary to store the value of
787 				 * cpi->cpi_ncore_per_chip.
788 				 */
789 				for (i = 1; i < ncpu_per_core; i <<= 1)
790 					coreid_shift++;
791 				cpi->cpi_coreid = apic_id >> coreid_shift;
792 			} else if (feature & X86_HTT) {
793 				/*
794 				 * Single-core multi-threaded processors.
795 				 */
796 				cpi->cpi_coreid = cpi->cpi_chipid;
797 			}
798 		} else if (cpi->cpi_vendor == X86_VENDOR_AMD) {
799 			/*
800 			 * AMD currently only has dual-core processors with
801 			 * single-threaded cores.  If they ever release
802 			 * multi-threaded processors, then this code
803 			 * will have to be updated.
804 			 */
805 			cpi->cpi_coreid = cpu->cpu_id;
806 		} else {
807 			/*
808 			 * All other processors are currently
809 			 * assumed to have single cores.
810 			 */
811 			cpi->cpi_coreid = cpi->cpi_chipid;
812 		}
813 	}
814 
815 pass1_done:
816 	cpi->cpi_pass = 1;
817 	return (feature);
818 }
819 
820 /*
821  * Make copies of the cpuid table entries we depend on, in
822  * part for ease of parsing now, in part so that we have only
823  * one place to correct any of it, in part for ease of
824  * later export to userland, and in part so we can look at
825  * this stuff in a crash dump.
826  */
827 
828 /*ARGSUSED*/
829 void
830 cpuid_pass2(cpu_t *cpu)
831 {
832 	uint_t n, nmax;
833 	int i;
834 	struct cpuid_regs *cp;
835 	uint8_t *dp;
836 	uint32_t *iptr;
837 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
838 
839 	ASSERT(cpi->cpi_pass == 1);
840 
841 	if (cpi->cpi_maxeax < 1)
842 		goto pass2_done;
843 
844 	if ((nmax = cpi->cpi_maxeax + 1) > NMAX_CPI_STD)
845 		nmax = NMAX_CPI_STD;
846 	/*
847 	 * (We already handled n == 0 and n == 1 in pass 1)
848 	 */
849 	for (n = 2, cp = &cpi->cpi_std[2]; n < nmax; n++, cp++) {
850 		cp->cp_eax = n;
851 		(void) __cpuid_insn(cp);
852 		switch (n) {
853 		case 2:
854 			/*
855 			 * "the lower 8 bits of the %eax register
856 			 * contain a value that identifies the number
857 			 * of times the cpuid [instruction] has to be
858 			 * executed to obtain a complete image of the
859 			 * processor's caching systems."
860 			 *
861 			 * How *do* they make this stuff up?
862 			 */
863 			cpi->cpi_ncache = sizeof (*cp) *
864 			    BITX(cp->cp_eax, 7, 0);
865 			if (cpi->cpi_ncache == 0)
866 				break;
867 			cpi->cpi_ncache--;	/* skip count byte */
868 
869 			/*
870 			 * Well, for now, rather than attempt to implement
871 			 * this slightly dubious algorithm, we just look
872 			 * at the first 15 ..
873 			 */
874 			if (cpi->cpi_ncache > (sizeof (*cp) - 1))
875 				cpi->cpi_ncache = sizeof (*cp) - 1;
876 
877 			dp = cpi->cpi_cacheinfo;
878 			if (BITX(cp->cp_eax, 31, 31) == 0) {
879 				uint8_t *p = (void *)&cp->cp_eax;
880 				for (i = 1; i < 3; i++)
881 					if (p[i] != 0)
882 						*dp++ = p[i];
883 			}
884 			if (BITX(cp->cp_ebx, 31, 31) == 0) {
885 				uint8_t *p = (void *)&cp->cp_ebx;
886 				for (i = 0; i < 4; i++)
887 					if (p[i] != 0)
888 						*dp++ = p[i];
889 			}
890 			if (BITX(cp->cp_ecx, 31, 31) == 0) {
891 				uint8_t *p = (void *)&cp->cp_ecx;
892 				for (i = 0; i < 4; i++)
893 					if (p[i] != 0)
894 						*dp++ = p[i];
895 			}
896 			if (BITX(cp->cp_edx, 31, 31) == 0) {
897 				uint8_t *p = (void *)&cp->cp_edx;
898 				for (i = 0; i < 4; i++)
899 					if (p[i] != 0)
900 						*dp++ = p[i];
901 			}
902 			break;
903 		case 3:	/* Processor serial number, if PSN supported */
904 		case 4:	/* Deterministic cache parameters */
905 		case 5:	/* Monitor/Mwait parameters */
906 		default:
907 			break;
908 		}
909 	}
910 
911 	if ((cpi->cpi_xmaxeax & 0x80000000) == 0)
912 		goto pass2_done;
913 
914 	if ((nmax = cpi->cpi_xmaxeax - 0x80000000 + 1) > NMAX_CPI_EXTD)
915 		nmax = NMAX_CPI_EXTD;
916 	/*
917 	 * Copy the extended properties, fixing them as we go.
918 	 * (We already handled n == 0 and n == 1 in pass 1)
919 	 */
920 	iptr = (void *)cpi->cpi_brandstr;
921 	for (n = 2, cp = &cpi->cpi_extd[2]; n < nmax; cp++, n++) {
922 		cp->cp_eax = 0x80000000 + n;
923 		(void) __cpuid_insn(cp);
924 		switch (n) {
925 		case 2:
926 		case 3:
927 		case 4:
928 			/*
929 			 * Extract the brand string
930 			 */
931 			*iptr++ = cp->cp_eax;
932 			*iptr++ = cp->cp_ebx;
933 			*iptr++ = cp->cp_ecx;
934 			*iptr++ = cp->cp_edx;
935 			break;
936 		case 5:
937 			switch (cpi->cpi_vendor) {
938 			case X86_VENDOR_AMD:
939 				/*
940 				 * The Athlon and Duron were the first
941 				 * parts to report the sizes of the
942 				 * TLB for large pages. Before then,
943 				 * we don't trust the data.
944 				 */
945 				if (cpi->cpi_family < 6 ||
946 				    (cpi->cpi_family == 6 &&
947 				    cpi->cpi_model < 1))
948 					cp->cp_eax = 0;
949 				break;
950 			default:
951 				break;
952 			}
953 			break;
954 		case 6:
955 			switch (cpi->cpi_vendor) {
956 			case X86_VENDOR_AMD:
957 				/*
958 				 * The Athlon and Duron were the first
959 				 * AMD parts with L2 TLB's.
960 				 * Before then, don't trust the data.
961 				 */
962 				if (cpi->cpi_family < 6 ||
963 				    cpi->cpi_family == 6 &&
964 				    cpi->cpi_model < 1)
965 					cp->cp_eax = cp->cp_ebx = 0;
966 				/*
967 				 * AMD Duron rev A0 reports L2
968 				 * cache size incorrectly as 1K
969 				 * when it is really 64K
970 				 */
971 				if (cpi->cpi_family == 6 &&
972 				    cpi->cpi_model == 3 &&
973 				    cpi->cpi_step == 0) {
974 					cp->cp_ecx &= 0xffff;
975 					cp->cp_ecx |= 0x400000;
976 				}
977 				break;
978 			case X86_VENDOR_Cyrix:	/* VIA C3 */
979 				/*
980 				 * VIA C3 processors are a bit messed
981 				 * up w.r.t. encoding cache sizes in %ecx
982 				 */
983 				if (cpi->cpi_family != 6)
984 					break;
985 				/*
986 				 * model 7 and 8 were incorrectly encoded
987 				 *
988 				 * xxx is model 8 really broken?
989 				 */
990 				if (cpi->cpi_model == 7 ||
991 				    cpi->cpi_model == 8)
992 					cp->cp_ecx =
993 					    BITX(cp->cp_ecx, 31, 24) << 16 |
994 					    BITX(cp->cp_ecx, 23, 16) << 12 |
995 					    BITX(cp->cp_ecx, 15, 8) << 8 |
996 					    BITX(cp->cp_ecx, 7, 0);
997 				/*
998 				 * model 9 stepping 1 has wrong associativity
999 				 */
1000 				if (cpi->cpi_model == 9 && cpi->cpi_step == 1)
1001 					cp->cp_ecx |= 8 << 12;
1002 				break;
1003 			case X86_VENDOR_Intel:
1004 				/*
1005 				 * Extended L2 Cache features function.
1006 				 * First appeared on Prescott.
1007 				 */
1008 			default:
1009 				break;
1010 			}
1011 			break;
1012 		default:
1013 			break;
1014 		}
1015 	}
1016 
1017 pass2_done:
1018 	cpi->cpi_pass = 2;
1019 }
1020 
1021 static const char *
1022 intel_cpubrand(const struct cpuid_info *cpi)
1023 {
1024 	int i;
1025 
1026 	if ((x86_feature & X86_CPUID) == 0 ||
1027 	    cpi->cpi_maxeax < 1 || cpi->cpi_family < 5)
1028 		return ("i486");
1029 
1030 	switch (cpi->cpi_family) {
1031 	case 5:
1032 		return ("Intel Pentium(r)");
1033 	case 6:
1034 		switch (cpi->cpi_model) {
1035 			uint_t celeron, xeon;
1036 			const struct cpuid_regs *cp;
1037 		case 0:
1038 		case 1:
1039 		case 2:
1040 			return ("Intel Pentium(r) Pro");
1041 		case 3:
1042 		case 4:
1043 			return ("Intel Pentium(r) II");
1044 		case 6:
1045 			return ("Intel Celeron(r)");
1046 		case 5:
1047 		case 7:
1048 			celeron = xeon = 0;
1049 			cp = &cpi->cpi_std[2];	/* cache info */
1050 
1051 			for (i = 1; i < 3; i++) {
1052 				uint_t tmp;
1053 
1054 				tmp = (cp->cp_eax >> (8 * i)) & 0xff;
1055 				if (tmp == 0x40)
1056 					celeron++;
1057 				if (tmp >= 0x44 && tmp <= 0x45)
1058 					xeon++;
1059 			}
1060 
1061 			for (i = 0; i < 2; i++) {
1062 				uint_t tmp;
1063 
1064 				tmp = (cp->cp_ebx >> (8 * i)) & 0xff;
1065 				if (tmp == 0x40)
1066 					celeron++;
1067 				else if (tmp >= 0x44 && tmp <= 0x45)
1068 					xeon++;
1069 			}
1070 
1071 			for (i = 0; i < 4; i++) {
1072 				uint_t tmp;
1073 
1074 				tmp = (cp->cp_ecx >> (8 * i)) & 0xff;
1075 				if (tmp == 0x40)
1076 					celeron++;
1077 				else if (tmp >= 0x44 && tmp <= 0x45)
1078 					xeon++;
1079 			}
1080 
1081 			for (i = 0; i < 4; i++) {
1082 				uint_t tmp;
1083 
1084 				tmp = (cp->cp_edx >> (8 * i)) & 0xff;
1085 				if (tmp == 0x40)
1086 					celeron++;
1087 				else if (tmp >= 0x44 && tmp <= 0x45)
1088 					xeon++;
1089 			}
1090 
1091 			if (celeron)
1092 				return ("Intel Celeron(r)");
1093 			if (xeon)
1094 				return (cpi->cpi_model == 5 ?
1095 				    "Intel Pentium(r) II Xeon(tm)" :
1096 				    "Intel Pentium(r) III Xeon(tm)");
1097 			return (cpi->cpi_model == 5 ?
1098 			    "Intel Pentium(r) II or Pentium(r) II Xeon(tm)" :
1099 			    "Intel Pentium(r) III or Pentium(r) III Xeon(tm)");
1100 		default:
1101 			break;
1102 		}
1103 	default:
1104 		break;
1105 	}
1106 
1107 	/* BrandID is present if the field is nonzero */
1108 	if (cpi->cpi_brandid != 0) {
1109 		static const struct {
1110 			uint_t bt_bid;
1111 			const char *bt_str;
1112 		} brand_tbl[] = {
1113 			{ 0x1,	"Intel(r) Celeron(r)" },
1114 			{ 0x2,	"Intel(r) Pentium(r) III" },
1115 			{ 0x3,	"Intel(r) Pentium(r) III Xeon(tm)" },
1116 			{ 0x4,	"Intel(r) Pentium(r) III" },
1117 			{ 0x6,	"Mobile Intel(r) Pentium(r) III" },
1118 			{ 0x7,	"Mobile Intel(r) Celeron(r)" },
1119 			{ 0x8,	"Intel(r) Pentium(r) 4" },
1120 			{ 0x9,	"Intel(r) Pentium(r) 4" },
1121 			{ 0xa,	"Intel(r) Celeron(r)" },
1122 			{ 0xb,	"Intel(r) Xeon(tm)" },
1123 			{ 0xc,	"Intel(r) Xeon(tm) MP" },
1124 			{ 0xe,	"Mobile Intel(r) Pentium(r) 4" },
1125 			{ 0xf,	"Mobile Intel(r) Celeron(r)" },
1126 			{ 0x11, "Mobile Genuine Intel(r)" },
1127 			{ 0x12, "Intel(r) Celeron(r) M" },
1128 			{ 0x13, "Mobile Intel(r) Celeron(r)" },
1129 			{ 0x14, "Intel(r) Celeron(r)" },
1130 			{ 0x15, "Mobile Genuine Intel(r)" },
1131 			{ 0x16,	"Intel(r) Pentium(r) M" },
1132 			{ 0x17, "Mobile Intel(r) Celeron(r)" }
1133 		};
1134 		uint_t btblmax = sizeof (brand_tbl) / sizeof (brand_tbl[0]);
1135 		uint_t sgn;
1136 
1137 		sgn = (cpi->cpi_family << 8) |
1138 		    (cpi->cpi_model << 4) | cpi->cpi_step;
1139 
1140 		for (i = 0; i < btblmax; i++)
1141 			if (brand_tbl[i].bt_bid == cpi->cpi_brandid)
1142 				break;
1143 		if (i < btblmax) {
1144 			if (sgn == 0x6b1 && cpi->cpi_brandid == 3)
1145 				return ("Intel(r) Celeron(r)");
1146 			if (sgn < 0xf13 && cpi->cpi_brandid == 0xb)
1147 				return ("Intel(r) Xeon(tm) MP");
1148 			if (sgn < 0xf13 && cpi->cpi_brandid == 0xe)
1149 				return ("Intel(r) Xeon(tm)");
1150 			return (brand_tbl[i].bt_str);
1151 		}
1152 	}
1153 
1154 	return (NULL);
1155 }
1156 
1157 static const char *
1158 amd_cpubrand(const struct cpuid_info *cpi)
1159 {
1160 	if ((x86_feature & X86_CPUID) == 0 ||
1161 	    cpi->cpi_maxeax < 1 || cpi->cpi_family < 5)
1162 		return ("i486 compatible");
1163 
1164 	switch (cpi->cpi_family) {
1165 	case 5:
1166 		switch (cpi->cpi_model) {
1167 		case 0:
1168 		case 1:
1169 		case 2:
1170 		case 3:
1171 		case 4:
1172 		case 5:
1173 			return ("AMD-K5(r)");
1174 		case 6:
1175 		case 7:
1176 			return ("AMD-K6(r)");
1177 		case 8:
1178 			return ("AMD-K6(r)-2");
1179 		case 9:
1180 			return ("AMD-K6(r)-III");
1181 		default:
1182 			return ("AMD (family 5)");
1183 		}
1184 	case 6:
1185 		switch (cpi->cpi_model) {
1186 		case 1:
1187 			return ("AMD-K7(tm)");
1188 		case 0:
1189 		case 2:
1190 		case 4:
1191 			return ("AMD Athlon(tm)");
1192 		case 3:
1193 		case 7:
1194 			return ("AMD Duron(tm)");
1195 		case 6:
1196 		case 8:
1197 		case 10:
1198 			/*
1199 			 * Use the L2 cache size to distinguish
1200 			 */
1201 			return ((cpi->cpi_extd[6].cp_ecx >> 16) >= 256 ?
1202 			    "AMD Athlon(tm)" : "AMD Duron(tm)");
1203 		default:
1204 			return ("AMD (family 6)");
1205 		}
1206 	default:
1207 		break;
1208 	}
1209 
1210 	if (cpi->cpi_family == 0xf && cpi->cpi_model == 5 &&
1211 	    cpi->cpi_brandid != 0) {
1212 		switch (BITX(cpi->cpi_brandid, 7, 5)) {
1213 		case 3:
1214 			return ("AMD Opteron(tm) UP 1xx");
1215 		case 4:
1216 			return ("AMD Opteron(tm) DP 2xx");
1217 		case 5:
1218 			return ("AMD Opteron(tm) MP 8xx");
1219 		default:
1220 			return ("AMD Opteron(tm)");
1221 		}
1222 	}
1223 
1224 	return (NULL);
1225 }
1226 
1227 static const char *
1228 cyrix_cpubrand(struct cpuid_info *cpi, uint_t type)
1229 {
1230 	if ((x86_feature & X86_CPUID) == 0 ||
1231 	    cpi->cpi_maxeax < 1 || cpi->cpi_family < 5 ||
1232 	    type == X86_TYPE_CYRIX_486)
1233 		return ("i486 compatible");
1234 
1235 	switch (type) {
1236 	case X86_TYPE_CYRIX_6x86:
1237 		return ("Cyrix 6x86");
1238 	case X86_TYPE_CYRIX_6x86L:
1239 		return ("Cyrix 6x86L");
1240 	case X86_TYPE_CYRIX_6x86MX:
1241 		return ("Cyrix 6x86MX");
1242 	case X86_TYPE_CYRIX_GXm:
1243 		return ("Cyrix GXm");
1244 	case X86_TYPE_CYRIX_MediaGX:
1245 		return ("Cyrix MediaGX");
1246 	case X86_TYPE_CYRIX_MII:
1247 		return ("Cyrix M2");
1248 	case X86_TYPE_VIA_CYRIX_III:
1249 		return ("VIA Cyrix M3");
1250 	default:
1251 		/*
1252 		 * Have another wild guess ..
1253 		 */
1254 		if (cpi->cpi_family == 4 && cpi->cpi_model == 9)
1255 			return ("Cyrix 5x86");
1256 		else if (cpi->cpi_family == 5) {
1257 			switch (cpi->cpi_model) {
1258 			case 2:
1259 				return ("Cyrix 6x86");	/* Cyrix M1 */
1260 			case 4:
1261 				return ("Cyrix MediaGX");
1262 			default:
1263 				break;
1264 			}
1265 		} else if (cpi->cpi_family == 6) {
1266 			switch (cpi->cpi_model) {
1267 			case 0:
1268 				return ("Cyrix 6x86MX"); /* Cyrix M2? */
1269 			case 5:
1270 			case 6:
1271 			case 7:
1272 			case 8:
1273 			case 9:
1274 				return ("VIA C3");
1275 			default:
1276 				break;
1277 			}
1278 		}
1279 		break;
1280 	}
1281 	return (NULL);
1282 }
1283 
1284 /*
1285  * This only gets called in the case that the CPU extended
1286  * feature brand string (0x80000002, 0x80000003, 0x80000004)
1287  * aren't available, or contain null bytes for some reason.
1288  */
1289 static void
1290 fabricate_brandstr(struct cpuid_info *cpi)
1291 {
1292 	const char *brand = NULL;
1293 
1294 	switch (cpi->cpi_vendor) {
1295 	case X86_VENDOR_Intel:
1296 		brand = intel_cpubrand(cpi);
1297 		break;
1298 	case X86_VENDOR_AMD:
1299 		brand = amd_cpubrand(cpi);
1300 		break;
1301 	case X86_VENDOR_Cyrix:
1302 		brand = cyrix_cpubrand(cpi, x86_type);
1303 		break;
1304 	case X86_VENDOR_NexGen:
1305 		if (cpi->cpi_family == 5 && cpi->cpi_model == 0)
1306 			brand = "NexGen Nx586";
1307 		break;
1308 	case X86_VENDOR_Centaur:
1309 		if (cpi->cpi_family == 5)
1310 			switch (cpi->cpi_model) {
1311 			case 4:
1312 				brand = "Centaur C6";
1313 				break;
1314 			case 8:
1315 				brand = "Centaur C2";
1316 				break;
1317 			case 9:
1318 				brand = "Centaur C3";
1319 				break;
1320 			default:
1321 				break;
1322 			}
1323 		break;
1324 	case X86_VENDOR_Rise:
1325 		if (cpi->cpi_family == 5 &&
1326 		    (cpi->cpi_model == 0 || cpi->cpi_model == 2))
1327 			brand = "Rise mP6";
1328 		break;
1329 	case X86_VENDOR_SiS:
1330 		if (cpi->cpi_family == 5 && cpi->cpi_model == 0)
1331 			brand = "SiS 55x";
1332 		break;
1333 	case X86_VENDOR_TM:
1334 		if (cpi->cpi_family == 5 && cpi->cpi_model == 4)
1335 			brand = "Transmeta Crusoe TM3x00 or TM5x00";
1336 		break;
1337 	case X86_VENDOR_NSC:
1338 	case X86_VENDOR_UMC:
1339 	default:
1340 		break;
1341 	}
1342 	if (brand) {
1343 		(void) strcpy((char *)cpi->cpi_brandstr, brand);
1344 		return;
1345 	}
1346 
1347 	/*
1348 	 * If all else fails ...
1349 	 */
1350 	(void) snprintf(cpi->cpi_brandstr, sizeof (cpi->cpi_brandstr),
1351 	    "%s %d.%d.%d", cpi->cpi_vendorstr, cpi->cpi_family,
1352 	    cpi->cpi_model, cpi->cpi_step);
1353 }
1354 
1355 /*
1356  * This routine is called just after kernel memory allocation
1357  * becomes available on cpu0, and as part of mp_startup() on
1358  * the other cpus.
1359  *
1360  * Fixup the brand string.
1361  */
1362 /*ARGSUSED*/
1363 void
1364 cpuid_pass3(cpu_t *cpu)
1365 {
1366 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
1367 
1368 	ASSERT(cpi->cpi_pass == 2);
1369 
1370 	if ((cpi->cpi_xmaxeax & 0x80000000) == 0) {
1371 		fabricate_brandstr(cpi);
1372 		goto pass3_done;
1373 	}
1374 
1375 	/*
1376 	 * If we successfully extracted a brand string from the cpuid
1377 	 * instruction, clean it up by removing leading spaces and
1378 	 * similar junk.
1379 	 */
1380 	if (cpi->cpi_brandstr[0]) {
1381 		size_t maxlen = sizeof (cpi->cpi_brandstr);
1382 		char *src, *dst;
1383 
1384 		dst = src = (char *)cpi->cpi_brandstr;
1385 		src[maxlen - 1] = '\0';
1386 		/*
1387 		 * strip leading spaces
1388 		 */
1389 		while (*src == ' ')
1390 			src++;
1391 		/*
1392 		 * Remove any 'Genuine' or "Authentic" prefixes
1393 		 */
1394 		if (strncmp(src, "Genuine ", 8) == 0)
1395 			src += 8;
1396 		if (strncmp(src, "Authentic ", 10) == 0)
1397 			src += 10;
1398 
1399 		/*
1400 		 * Now do an in-place copy.
1401 		 * Map (R) to (r) and (TM) to (tm).
1402 		 * The era of teletypes is long gone, and there's
1403 		 * -really- no need to shout.
1404 		 */
1405 		while (*src != '\0') {
1406 			if (src[0] == '(') {
1407 				if (strncmp(src + 1, "R)", 2) == 0) {
1408 					(void) strncpy(dst, "(r)", 3);
1409 					src += 3;
1410 					dst += 3;
1411 					continue;
1412 				}
1413 				if (strncmp(src + 1, "TM)", 3) == 0) {
1414 					(void) strncpy(dst, "(tm)", 4);
1415 					src += 4;
1416 					dst += 4;
1417 					continue;
1418 				}
1419 			}
1420 			*dst++ = *src++;
1421 		}
1422 		*dst = '\0';
1423 
1424 		/*
1425 		 * Finally, remove any trailing spaces
1426 		 */
1427 		while (--dst > cpi->cpi_brandstr)
1428 			if (*dst == ' ')
1429 				*dst = '\0';
1430 			else
1431 				break;
1432 	} else
1433 		fabricate_brandstr(cpi);
1434 
1435 pass3_done:
1436 	cpi->cpi_pass = 3;
1437 }
1438 
1439 /*
1440  * This routine is called out of bind_hwcap() much later in the life
1441  * of the kernel (post_startup()).  The job of this routine is to resolve
1442  * the hardware feature support and kernel support for those features into
1443  * what we're actually going to tell applications via the aux vector.
1444  */
1445 uint_t
1446 cpuid_pass4(cpu_t *cpu)
1447 {
1448 	struct cpuid_info *cpi;
1449 	uint_t hwcap_flags = 0;
1450 
1451 	if (cpu == NULL)
1452 		cpu = CPU;
1453 	cpi = cpu->cpu_m.mcpu_cpi;
1454 
1455 	ASSERT(cpi->cpi_pass == 3);
1456 
1457 	if (cpi->cpi_maxeax >= 1) {
1458 		uint32_t *edx = &cpi->cpi_support[STD_EDX_FEATURES];
1459 		uint32_t *ecx = &cpi->cpi_support[STD_ECX_FEATURES];
1460 
1461 		*edx = CPI_FEATURES_EDX(cpi);
1462 		*ecx = CPI_FEATURES_ECX(cpi);
1463 
1464 		/*
1465 		 * [these require explicit kernel support]
1466 		 */
1467 		if ((x86_feature & X86_SEP) == 0)
1468 			*edx &= ~CPUID_INTC_EDX_SEP;
1469 
1470 		if ((x86_feature & X86_SSE) == 0)
1471 			*edx &= ~(CPUID_INTC_EDX_FXSR|CPUID_INTC_EDX_SSE);
1472 		if ((x86_feature & X86_SSE2) == 0)
1473 			*edx &= ~CPUID_INTC_EDX_SSE2;
1474 
1475 		if ((x86_feature & X86_HTT) == 0)
1476 			*edx &= ~CPUID_INTC_EDX_HTT;
1477 
1478 		if ((x86_feature & X86_SSE3) == 0)
1479 			*ecx &= ~CPUID_INTC_ECX_SSE3;
1480 
1481 		/*
1482 		 * [no explicit support required beyond x87 fp context]
1483 		 */
1484 		if (!fpu_exists)
1485 			*edx &= ~(CPUID_INTC_EDX_FPU | CPUID_INTC_EDX_MMX);
1486 
1487 		/*
1488 		 * Now map the supported feature vector to things that we
1489 		 * think userland will care about.
1490 		 */
1491 		if (*edx & CPUID_INTC_EDX_SEP)
1492 			hwcap_flags |= AV_386_SEP;
1493 		if (*edx & CPUID_INTC_EDX_SSE)
1494 			hwcap_flags |= AV_386_FXSR | AV_386_SSE;
1495 		if (*edx & CPUID_INTC_EDX_SSE2)
1496 			hwcap_flags |= AV_386_SSE2;
1497 		if (*ecx & CPUID_INTC_ECX_SSE3)
1498 			hwcap_flags |= AV_386_SSE3;
1499 
1500 		if (*edx & CPUID_INTC_EDX_FPU)
1501 			hwcap_flags |= AV_386_FPU;
1502 		if (*edx & CPUID_INTC_EDX_MMX)
1503 			hwcap_flags |= AV_386_MMX;
1504 
1505 		if (*edx & CPUID_INTC_EDX_TSC)
1506 			hwcap_flags |= AV_386_TSC;
1507 		if (*edx & CPUID_INTC_EDX_CX8)
1508 			hwcap_flags |= AV_386_CX8;
1509 		if (*edx & CPUID_INTC_EDX_CMOV)
1510 			hwcap_flags |= AV_386_CMOV;
1511 		if (*ecx & CPUID_INTC_ECX_MON)
1512 			hwcap_flags |= AV_386_MON;
1513 #if defined(CPUID_INTC_ECX_CX16)
1514 		if (*ecx & CPUID_INTC_ECX_CX16)
1515 			hwcap_flags |= AV_386_CX16;
1516 #endif
1517 	}
1518 
1519 	if (x86_feature & X86_HTT)
1520 		hwcap_flags |= AV_386_PAUSE;
1521 
1522 	if (cpi->cpi_xmaxeax < 0x80000001)
1523 		goto pass4_done;
1524 
1525 	switch (cpi->cpi_vendor) {
1526 		struct cpuid_regs cp;
1527 		uint32_t *edx;
1528 
1529 	case X86_VENDOR_Intel:	/* sigh */
1530 	case X86_VENDOR_AMD:
1531 		edx = &cpi->cpi_support[AMD_EDX_FEATURES];
1532 
1533 		*edx = CPI_FEATURES_XTD_EDX(cpi);
1534 
1535 		/*
1536 		 * [no explicit support required beyond
1537 		 * x87 fp context and exception handlers]
1538 		 */
1539 		if (!fpu_exists)
1540 			*edx &= ~(CPUID_AMD_EDX_MMXamd |
1541 			    CPUID_AMD_EDX_3DNow | CPUID_AMD_EDX_3DNowx);
1542 
1543 		if ((x86_feature & X86_ASYSC) == 0)
1544 			*edx &= ~CPUID_AMD_EDX_SYSC;
1545 		if ((x86_feature & X86_NX) == 0)
1546 			*edx &= ~CPUID_AMD_EDX_NX;
1547 #if !defined(_LP64)
1548 		*edx &= ~CPUID_AMD_EDX_LM;
1549 #endif
1550 		/*
1551 		 * Now map the supported feature vector to
1552 		 * things that we think userland will care about.
1553 		 */
1554 		if (*edx & CPUID_AMD_EDX_SYSC)
1555 			hwcap_flags |= AV_386_AMD_SYSC;
1556 		if (*edx & CPUID_AMD_EDX_MMXamd)
1557 			hwcap_flags |= AV_386_AMD_MMX;
1558 		if (*edx & CPUID_AMD_EDX_3DNow)
1559 			hwcap_flags |= AV_386_AMD_3DNow;
1560 		if (*edx & CPUID_AMD_EDX_3DNowx)
1561 			hwcap_flags |= AV_386_AMD_3DNowx;
1562 		break;
1563 
1564 	case X86_VENDOR_TM:
1565 		cp.cp_eax = 0x80860001;
1566 		(void) __cpuid_insn(&cp);
1567 		cpi->cpi_support[TM_EDX_FEATURES] = cp.cp_edx;
1568 		break;
1569 
1570 	default:
1571 		break;
1572 	}
1573 
1574 pass4_done:
1575 	cpi->cpi_pass = 4;
1576 	return (hwcap_flags);
1577 }
1578 
1579 
1580 /*
1581  * Simulate the cpuid instruction using the data we previously
1582  * captured about this CPU.  We try our best to return the truth
1583  * about the hardware, independently of kernel support.
1584  */
1585 uint32_t
1586 cpuid_insn(cpu_t *cpu, struct cpuid_regs *cp)
1587 {
1588 	struct cpuid_info *cpi;
1589 	struct cpuid_regs *xcp;
1590 
1591 	if (cpu == NULL)
1592 		cpu = CPU;
1593 	cpi = cpu->cpu_m.mcpu_cpi;
1594 
1595 	ASSERT(cpuid_checkpass(cpu, 3));
1596 
1597 	/*
1598 	 * CPUID data is cached in two separate places: cpi_std for standard
1599 	 * CPUID functions, and cpi_extd for extended CPUID functions.
1600 	 */
1601 	if (cp->cp_eax <= cpi->cpi_maxeax && cp->cp_eax < NMAX_CPI_STD)
1602 		xcp = &cpi->cpi_std[cp->cp_eax];
1603 	else if (cp->cp_eax >= 0x80000000 && cp->cp_eax <= cpi->cpi_xmaxeax &&
1604 	    cp->cp_eax < 0x80000000 + NMAX_CPI_EXTD)
1605 		xcp = &cpi->cpi_extd[cp->cp_eax - 0x80000000];
1606 	else
1607 		/*
1608 		 * The caller is asking for data from an input parameter which
1609 		 * the kernel has not cached.  In this case we go fetch from
1610 		 * the hardware and return the data directly to the user.
1611 		 */
1612 		return (__cpuid_insn(cp));
1613 
1614 	cp->cp_eax = xcp->cp_eax;
1615 	cp->cp_ebx = xcp->cp_ebx;
1616 	cp->cp_ecx = xcp->cp_ecx;
1617 	cp->cp_edx = xcp->cp_edx;
1618 	return (cp->cp_eax);
1619 }
1620 
1621 int
1622 cpuid_checkpass(cpu_t *cpu, int pass)
1623 {
1624 	return (cpu != NULL && cpu->cpu_m.mcpu_cpi != NULL &&
1625 	    cpu->cpu_m.mcpu_cpi->cpi_pass >= pass);
1626 }
1627 
1628 int
1629 cpuid_getbrandstr(cpu_t *cpu, char *s, size_t n)
1630 {
1631 	ASSERT(cpuid_checkpass(cpu, 3));
1632 
1633 	return (snprintf(s, n, "%s", cpu->cpu_m.mcpu_cpi->cpi_brandstr));
1634 }
1635 
1636 int
1637 cpuid_is_cmt(cpu_t *cpu)
1638 {
1639 	if (cpu == NULL)
1640 		cpu = CPU;
1641 
1642 	ASSERT(cpuid_checkpass(cpu, 1));
1643 
1644 	return (cpu->cpu_m.mcpu_cpi->cpi_chipid >= 0);
1645 }
1646 
1647 /*
1648  * AMD and Intel both implement the 64-bit variant of the syscall
1649  * instruction (syscallq), so if there's -any- support for syscall,
1650  * cpuid currently says "yes, we support this".
1651  *
1652  * However, Intel decided to -not- implement the 32-bit variant of the
1653  * syscall instruction, so we provide a predicate to allow our caller
1654  * to test that subtlety here.
1655  */
1656 /*ARGSUSED*/
1657 int
1658 cpuid_syscall32_insn(cpu_t *cpu)
1659 {
1660 	ASSERT(cpuid_checkpass((cpu == NULL ? CPU : cpu), 1));
1661 
1662 	if (x86_feature & X86_ASYSC)
1663 		return (x86_vendor != X86_VENDOR_Intel);
1664 	return (0);
1665 }
1666 
1667 int
1668 cpuid_getidstr(cpu_t *cpu, char *s, size_t n)
1669 {
1670 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
1671 
1672 	static const char fmt[] =
1673 	    "x86 (%s family %d model %d step %d clock %d MHz)";
1674 	static const char fmt_ht[] =
1675 	    "x86 (chipid 0x%x %s family %d model %d step %d clock %d MHz)";
1676 
1677 	ASSERT(cpuid_checkpass(cpu, 1));
1678 
1679 	if (cpuid_is_cmt(cpu))
1680 		return (snprintf(s, n, fmt_ht, cpi->cpi_chipid,
1681 		    cpi->cpi_vendorstr, cpi->cpi_family, cpi->cpi_model,
1682 		    cpi->cpi_step, cpu->cpu_type_info.pi_clock));
1683 	return (snprintf(s, n, fmt,
1684 	    cpi->cpi_vendorstr, cpi->cpi_family, cpi->cpi_model,
1685 	    cpi->cpi_step, cpu->cpu_type_info.pi_clock));
1686 }
1687 
1688 const char *
1689 cpuid_getvendorstr(cpu_t *cpu)
1690 {
1691 	ASSERT(cpuid_checkpass(cpu, 1));
1692 	return ((const char *)cpu->cpu_m.mcpu_cpi->cpi_vendorstr);
1693 }
1694 
1695 uint_t
1696 cpuid_getvendor(cpu_t *cpu)
1697 {
1698 	ASSERT(cpuid_checkpass(cpu, 1));
1699 	return (cpu->cpu_m.mcpu_cpi->cpi_vendor);
1700 }
1701 
1702 uint_t
1703 cpuid_getfamily(cpu_t *cpu)
1704 {
1705 	ASSERT(cpuid_checkpass(cpu, 1));
1706 	return (cpu->cpu_m.mcpu_cpi->cpi_family);
1707 }
1708 
1709 uint_t
1710 cpuid_getmodel(cpu_t *cpu)
1711 {
1712 	ASSERT(cpuid_checkpass(cpu, 1));
1713 	return (cpu->cpu_m.mcpu_cpi->cpi_model);
1714 }
1715 
1716 uint_t
1717 cpuid_get_ncpu_per_chip(cpu_t *cpu)
1718 {
1719 	ASSERT(cpuid_checkpass(cpu, 1));
1720 	return (cpu->cpu_m.mcpu_cpi->cpi_ncpu_per_chip);
1721 }
1722 
1723 uint_t
1724 cpuid_get_ncore_per_chip(cpu_t *cpu)
1725 {
1726 	ASSERT(cpuid_checkpass(cpu, 1));
1727 	return (cpu->cpu_m.mcpu_cpi->cpi_ncore_per_chip);
1728 }
1729 
1730 uint_t
1731 cpuid_getstep(cpu_t *cpu)
1732 {
1733 	ASSERT(cpuid_checkpass(cpu, 1));
1734 	return (cpu->cpu_m.mcpu_cpi->cpi_step);
1735 }
1736 
1737 chipid_t
1738 chip_plat_get_chipid(cpu_t *cpu)
1739 {
1740 	ASSERT(cpuid_checkpass(cpu, 1));
1741 
1742 	if (cpuid_is_cmt(cpu))
1743 		return (cpu->cpu_m.mcpu_cpi->cpi_chipid);
1744 	return (cpu->cpu_id);
1745 }
1746 
1747 id_t
1748 chip_plat_get_coreid(cpu_t *cpu)
1749 {
1750 	ASSERT(cpuid_checkpass(cpu, 1));
1751 	return (cpu->cpu_m.mcpu_cpi->cpi_coreid);
1752 }
1753 
1754 int
1755 chip_plat_get_clogid(cpu_t *cpu)
1756 {
1757 	ASSERT(cpuid_checkpass(cpu, 1));
1758 	return (cpu->cpu_m.mcpu_cpi->cpi_clogid);
1759 }
1760 
1761 void
1762 cpuid_get_addrsize(cpu_t *cpu, uint_t *pabits, uint_t *vabits)
1763 {
1764 	struct cpuid_info *cpi;
1765 
1766 	if (cpu == NULL)
1767 		cpu = CPU;
1768 	cpi = cpu->cpu_m.mcpu_cpi;
1769 
1770 	ASSERT(cpuid_checkpass(cpu, 1));
1771 
1772 	if (pabits)
1773 		*pabits = cpi->cpi_pabits;
1774 	if (vabits)
1775 		*vabits = cpi->cpi_vabits;
1776 }
1777 
1778 /*
1779  * Returns the number of data TLB entries for a corresponding
1780  * pagesize.  If it can't be computed, or isn't known, the
1781  * routine returns zero.  If you ask about an architecturally
1782  * impossible pagesize, the routine will panic (so that the
1783  * hat implementor knows that things are inconsistent.)
1784  */
1785 uint_t
1786 cpuid_get_dtlb_nent(cpu_t *cpu, size_t pagesize)
1787 {
1788 	struct cpuid_info *cpi;
1789 	uint_t dtlb_nent = 0;
1790 
1791 	if (cpu == NULL)
1792 		cpu = CPU;
1793 	cpi = cpu->cpu_m.mcpu_cpi;
1794 
1795 	ASSERT(cpuid_checkpass(cpu, 1));
1796 
1797 	/*
1798 	 * Check the L2 TLB info
1799 	 */
1800 	if (cpi->cpi_xmaxeax >= 0x80000006) {
1801 		struct cpuid_regs *cp = &cpi->cpi_extd[6];
1802 
1803 		switch (pagesize) {
1804 
1805 		case 4 * 1024:
1806 			/*
1807 			 * All zero in the top 16 bits of the register
1808 			 * indicates a unified TLB. Size is in low 16 bits.
1809 			 */
1810 			if ((cp->cp_ebx & 0xffff0000) == 0)
1811 				dtlb_nent = cp->cp_ebx & 0x0000ffff;
1812 			else
1813 				dtlb_nent = BITX(cp->cp_ebx, 27, 16);
1814 			break;
1815 
1816 		case 2 * 1024 * 1024:
1817 			if ((cp->cp_eax & 0xffff0000) == 0)
1818 				dtlb_nent = cp->cp_eax & 0x0000ffff;
1819 			else
1820 				dtlb_nent = BITX(cp->cp_eax, 27, 16);
1821 			break;
1822 
1823 		default:
1824 			panic("unknown L2 pagesize");
1825 			/*NOTREACHED*/
1826 		}
1827 	}
1828 
1829 	if (dtlb_nent != 0)
1830 		return (dtlb_nent);
1831 
1832 	/*
1833 	 * No L2 TLB support for this size, try L1.
1834 	 */
1835 	if (cpi->cpi_xmaxeax >= 0x80000005) {
1836 		struct cpuid_regs *cp = &cpi->cpi_extd[5];
1837 
1838 		switch (pagesize) {
1839 		case 4 * 1024:
1840 			dtlb_nent = BITX(cp->cp_ebx, 23, 16);
1841 			break;
1842 		case 2 * 1024 * 1024:
1843 			dtlb_nent = BITX(cp->cp_eax, 23, 16);
1844 			break;
1845 		default:
1846 			panic("unknown L1 d-TLB pagesize");
1847 			/*NOTREACHED*/
1848 		}
1849 	}
1850 
1851 	return (dtlb_nent);
1852 }
1853 
1854 /*
1855  * Return 0 if the erratum is not present or not applicable, positive
1856  * if it is, and negative if the status of the erratum is unknown.
1857  *
1858  * See "Revision Guide for AMD Athlon(tm) 64 and AMD Opteron(tm)
1859  * Processors" #25759, Rev 3.57, August 2005
1860  */
1861 int
1862 cpuid_opteron_erratum(cpu_t *cpu, uint_t erratum)
1863 {
1864 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
1865 	uint_t eax;
1866 
1867 	if (cpi->cpi_vendor != X86_VENDOR_AMD)
1868 		return (0);
1869 
1870 	eax = cpi->cpi_std[1].cp_eax;
1871 
1872 #define	SH_B0(eax)	(eax == 0xf40 || eax == 0xf50)
1873 #define	SH_B3(eax) 	(eax == 0xf51)
1874 #define	B(eax)		(SH_B0(eax) || SH_B3(eax))
1875 
1876 #define	SH_C0(eax)	(eax == 0xf48 || eax == 0xf58)
1877 
1878 #define	SH_CG(eax)	(eax == 0xf4a || eax == 0xf5a || eax == 0xf7a)
1879 #define	DH_CG(eax)	(eax == 0xfc0 || eax == 0xfe0 || eax == 0xff0)
1880 #define	CH_CG(eax)	(eax == 0xf82 || eax == 0xfb2)
1881 #define	CG(eax)		(SH_CG(eax) || DH_CG(eax) || CH_CG(eax))
1882 
1883 #define	SH_D0(eax)	(eax == 0x10f40 || eax == 0x10f50 || eax == 0x10f70)
1884 #define	DH_D0(eax)	(eax == 0x10fc0 || eax == 0x10ff0)
1885 #define	CH_D0(eax)	(eax == 0x10f80 || eax == 0x10fb0)
1886 #define	D0(eax)		(SH_D0(eax) || DH_D0(eax) || CH_D0(eax))
1887 
1888 #define	SH_E0(eax)	(eax == 0x20f50 || eax == 0x20f40 || eax == 0x20f70)
1889 #define	JH_E1(eax)	(eax == 0x20f10)	/* JH8_E0 had 0x20f30 */
1890 #define	DH_E3(eax)	(eax == 0x20fc0 || eax == 0x20ff0)
1891 #define	SH_E4(eax)	(eax == 0x20f51 || eax == 0x20f71)
1892 #define	BH_E4(eax)	(eax == 0x20fb1)
1893 #define	SH_E5(eax)	(eax == 0x20f42)
1894 #define	DH_E6(eax)	(eax == 0x20ff2 || eax == 0x20fc2)
1895 #define	JH_E6(eax)	(eax == 0x20f12 || eax == 0x20f32)
1896 #define	EX(eax)		(SH_E0(eax) || JH_E1(eax) || DH_E3(eax) || \
1897 			    SH_E4(eax) || BH_E4(eax) || SH_E5(eax) || \
1898 			    DH_E6(eax) || JH_E6(eax))
1899 
1900 	switch (erratum) {
1901 	case 1:
1902 		return (1);
1903 	case 51:	/* what does the asterisk mean? */
1904 		return (B(eax) || SH_C0(eax) || CG(eax));
1905 	case 52:
1906 		return (B(eax));
1907 	case 57:
1908 		return (1);
1909 	case 58:
1910 		return (B(eax));
1911 	case 60:
1912 		return (1);
1913 	case 61:
1914 	case 62:
1915 	case 63:
1916 	case 64:
1917 	case 65:
1918 	case 66:
1919 	case 68:
1920 	case 69:
1921 	case 70:
1922 	case 71:
1923 		return (B(eax));
1924 	case 72:
1925 		return (SH_B0(eax));
1926 	case 74:
1927 		return (B(eax));
1928 	case 75:
1929 		return (1);
1930 	case 76:
1931 		return (B(eax));
1932 	case 77:
1933 		return (1);
1934 	case 78:
1935 		return (B(eax) || SH_C0(eax));
1936 	case 79:
1937 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
1938 	case 80:
1939 	case 81:
1940 	case 82:
1941 		return (B(eax));
1942 	case 83:
1943 		return (B(eax) || SH_C0(eax) || CG(eax));
1944 	case 85:
1945 		return (1);
1946 	case 86:
1947 		return (SH_C0(eax) || CG(eax));
1948 	case 88:
1949 #if !defined(__amd64)
1950 		return (0);
1951 #else
1952 		return (B(eax) || SH_C0(eax));
1953 #endif
1954 	case 89:
1955 		return (1);
1956 	case 90:
1957 		return (B(eax) || SH_C0(eax) || CG(eax));
1958 	case 91:
1959 	case 92:
1960 		return (B(eax) || SH_C0(eax));
1961 	case 93:
1962 		return (SH_C0(eax));
1963 	case 94:
1964 		return (B(eax) || SH_C0(eax) || CG(eax));
1965 	case 95:
1966 #if !defined(__amd64)
1967 		return (0);
1968 #else
1969 		return (B(eax) || SH_C0(eax));
1970 #endif
1971 	case 96:
1972 		return (B(eax) || SH_C0(eax) || CG(eax));
1973 	case 97:
1974 	case 98:
1975 		return (SH_C0(eax) || CG(eax));
1976 	case 99:
1977 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
1978 	case 100:
1979 		return (B(eax) || SH_C0(eax));
1980 	case 101:
1981 	case 103:
1982 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
1983 	case 104:
1984 		return (SH_C0(eax) || CG(eax) || D0(eax));
1985 	case 105:
1986 	case 106:
1987 	case 107:
1988 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
1989 	case 108:
1990 		return (DH_CG(eax));
1991 	case 109:
1992 		return (SH_C0(eax) || CG(eax) || D0(eax));
1993 	case 110:
1994 		return (D0(eax) || EX(eax));
1995 	case 111:
1996 		return (CG(eax));
1997 	case 112:
1998 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
1999 	case 113:
2000 		return (eax == 0x20fc0);
2001 	case 114:
2002 		return (SH_E0(eax) || JH_E1(eax) || DH_E3(eax));
2003 	case 115:
2004 		return (SH_E0(eax) || JH_E1(eax));
2005 	case 116:
2006 		return (SH_E0(eax) || JH_E1(eax) || DH_E3(eax));
2007 	case 117:
2008 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
2009 	case 118:
2010 		return (SH_E0(eax) || JH_E1(eax) || SH_E4(eax) || BH_E4(eax) ||
2011 		    JH_E6(eax));
2012 	case 121:
2013 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
2014 	case 122:
2015 		return (SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
2016 	case 123:
2017 		return (JH_E1(eax) || BH_E4(eax) || JH_E6(eax));
2018 	case 131:
2019 		return (1);
2020 	case 6336786:
2021 		/*
2022 		 * Test for AdvPowerMgmtInfo.TscPStateInvariant
2023 		 * if this is a K8 family processor
2024 		 */
2025 		if (CPI_FAMILY(cpi) == 0xf) {
2026 			struct cpuid_regs regs;
2027 			regs.cp_eax = 0x80000007;
2028 			(void) __cpuid_insn(&regs);
2029 			return (!(regs.cp_edx & 0x100));
2030 		}
2031 		return (0);
2032 	case 6323525:
2033 		return (((((eax >> 12) & 0xff00) + (eax & 0xf00)) |
2034 		    (((eax >> 4) & 0xf) | ((eax >> 12) & 0xf0))) < 0xf40);
2035 
2036 	default:
2037 		return (-1);
2038 	}
2039 }
2040 
2041 static const char assoc_str[] = "associativity";
2042 static const char line_str[] = "line-size";
2043 static const char size_str[] = "size";
2044 
2045 static void
2046 add_cache_prop(dev_info_t *devi, const char *label, const char *type,
2047     uint32_t val)
2048 {
2049 	char buf[128];
2050 
2051 	/*
2052 	 * ndi_prop_update_int() is used because it is desirable for
2053 	 * DDI_PROP_HW_DEF and DDI_PROP_DONTSLEEP to be set.
2054 	 */
2055 	if (snprintf(buf, sizeof (buf), "%s-%s", label, type) < sizeof (buf))
2056 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, devi, buf, val);
2057 }
2058 
2059 /*
2060  * Intel-style cache/tlb description
2061  *
2062  * Standard cpuid level 2 gives a randomly ordered
2063  * selection of tags that index into a table that describes
2064  * cache and tlb properties.
2065  */
2066 
2067 static const char l1_icache_str[] = "l1-icache";
2068 static const char l1_dcache_str[] = "l1-dcache";
2069 static const char l2_cache_str[] = "l2-cache";
2070 static const char itlb4k_str[] = "itlb-4K";
2071 static const char dtlb4k_str[] = "dtlb-4K";
2072 static const char itlb4M_str[] = "itlb-4M";
2073 static const char dtlb4M_str[] = "dtlb-4M";
2074 static const char itlb424_str[] = "itlb-4K-2M-4M";
2075 static const char dtlb44_str[] = "dtlb-4K-4M";
2076 static const char sl1_dcache_str[] = "sectored-l1-dcache";
2077 static const char sl2_cache_str[] = "sectored-l2-cache";
2078 static const char itrace_str[] = "itrace-cache";
2079 static const char sl3_cache_str[] = "sectored-l3-cache";
2080 
2081 static const struct cachetab {
2082 	uint8_t 	ct_code;
2083 	uint8_t		ct_assoc;
2084 	uint16_t 	ct_line_size;
2085 	size_t		ct_size;
2086 	const char	*ct_label;
2087 } intel_ctab[] = {
2088 	/* maintain descending order! */
2089 	{ 0xb3, 4, 0, 128, dtlb4k_str },
2090 	{ 0xb0, 4, 0, 128, itlb4k_str },
2091 	{ 0x87, 8, 64, 1024*1024, l2_cache_str},
2092 	{ 0x86, 4, 64, 512*1024, l2_cache_str},
2093 	{ 0x85, 8, 32, 2*1024*1024, l2_cache_str},
2094 	{ 0x84, 8, 32, 1024*1024, l2_cache_str},
2095 	{ 0x83, 8, 32, 512*1024, l2_cache_str},
2096 	{ 0x82, 8, 32, 256*1024, l2_cache_str},
2097 	{ 0x81, 8, 32, 128*1024, l2_cache_str},		/* suspect! */
2098 	{ 0x7f, 2, 64, 512*1024, l2_cache_str},
2099 	{ 0x7d, 8, 64, 2*1024*1024, sl2_cache_str},
2100 	{ 0x7c, 8, 64, 1024*1024, sl2_cache_str},
2101 	{ 0x7b, 8, 64, 512*1024, sl2_cache_str},
2102 	{ 0x7a, 8, 64, 256*1024, sl2_cache_str},
2103 	{ 0x79, 8, 64, 128*1024, sl2_cache_str},
2104 	{ 0x78, 8, 64, 1024*1024, l2_cache_str},
2105 	{ 0x72, 8, 0, 32*1024, itrace_str},
2106 	{ 0x71, 8, 0, 16*1024, itrace_str},
2107 	{ 0x70, 8, 0, 12*1024, itrace_str},
2108 	{ 0x68, 4, 64, 32*1024, sl1_dcache_str},
2109 	{ 0x67, 4, 64, 16*1024, sl1_dcache_str},
2110 	{ 0x66, 4, 64, 8*1024, sl1_dcache_str},
2111 	{ 0x60, 8, 64, 16*1024, sl1_dcache_str},
2112 	{ 0x5d, 0, 0, 256, dtlb44_str},
2113 	{ 0x5c, 0, 0, 128, dtlb44_str},
2114 	{ 0x5b, 0, 0, 64, dtlb44_str},
2115 	{ 0x52, 0, 0, 256, itlb424_str},
2116 	{ 0x51, 0, 0, 128, itlb424_str},
2117 	{ 0x50, 0, 0, 64, itlb424_str},
2118 	{ 0x45, 4, 32, 2*1024*1024, l2_cache_str},
2119 	{ 0x44, 4, 32, 1024*1024, l2_cache_str},
2120 	{ 0x43, 4, 32, 512*1024, l2_cache_str},
2121 	{ 0x42, 4, 32, 256*1024, l2_cache_str},
2122 	{ 0x41, 4, 32, 128*1024, l2_cache_str},
2123 	{ 0x3c, 4, 64, 256*1024, sl2_cache_str},
2124 	{ 0x3b, 2, 64, 128*1024, sl2_cache_str},
2125 	{ 0x39, 4, 64, 128*1024, sl2_cache_str},
2126 	{ 0x30, 8, 64, 32*1024, l1_icache_str},
2127 	{ 0x2c, 8, 64, 32*1024, l1_dcache_str},
2128 	{ 0x29, 8, 64, 4096*1024, sl3_cache_str},
2129 	{ 0x25, 8, 64, 2048*1024, sl3_cache_str},
2130 	{ 0x23, 8, 64, 1024*1024, sl3_cache_str},
2131 	{ 0x22, 4, 64, 512*1024, sl3_cache_str},
2132 	{ 0x0c, 4, 32, 16*1024, l1_dcache_str},
2133 	{ 0x0a, 2, 32, 8*1024, l1_dcache_str},
2134 	{ 0x08, 4, 32, 16*1024, l1_icache_str},
2135 	{ 0x06, 4, 32, 8*1024, l1_icache_str},
2136 	{ 0x04, 4, 0, 8, dtlb4M_str},
2137 	{ 0x03, 4, 0, 64, dtlb4k_str},
2138 	{ 0x02, 4, 0, 2, itlb4M_str},
2139 	{ 0x01, 4, 0, 32, itlb4k_str},
2140 	{ 0 }
2141 };
2142 
2143 static const struct cachetab cyrix_ctab[] = {
2144 	{ 0x70, 4, 0, 32, "tlb-4K" },
2145 	{ 0x80, 4, 16, 16*1024, "l1-cache" },
2146 	{ 0 }
2147 };
2148 
2149 /*
2150  * Search a cache table for a matching entry
2151  */
2152 static const struct cachetab *
2153 find_cacheent(const struct cachetab *ct, uint_t code)
2154 {
2155 	if (code != 0) {
2156 		for (; ct->ct_code != 0; ct++)
2157 			if (ct->ct_code <= code)
2158 				break;
2159 		if (ct->ct_code == code)
2160 			return (ct);
2161 	}
2162 	return (NULL);
2163 }
2164 
2165 /*
2166  * Walk the cacheinfo descriptor, applying 'func' to every valid element
2167  * The walk is terminated if the walker returns non-zero.
2168  */
2169 static void
2170 intel_walk_cacheinfo(struct cpuid_info *cpi,
2171     void *arg, int (*func)(void *, const struct cachetab *))
2172 {
2173 	const struct cachetab *ct;
2174 	uint8_t *dp;
2175 	int i;
2176 
2177 	if ((dp = cpi->cpi_cacheinfo) == NULL)
2178 		return;
2179 	for (i = 0; i < cpi->cpi_ncache; i++, dp++)
2180 		if ((ct = find_cacheent(intel_ctab, *dp)) != NULL) {
2181 			if (func(arg, ct) != 0)
2182 				break;
2183 		}
2184 }
2185 
2186 /*
2187  * (Like the Intel one, except for Cyrix CPUs)
2188  */
2189 static void
2190 cyrix_walk_cacheinfo(struct cpuid_info *cpi,
2191     void *arg, int (*func)(void *, const struct cachetab *))
2192 {
2193 	const struct cachetab *ct;
2194 	uint8_t *dp;
2195 	int i;
2196 
2197 	if ((dp = cpi->cpi_cacheinfo) == NULL)
2198 		return;
2199 	for (i = 0; i < cpi->cpi_ncache; i++, dp++) {
2200 		/*
2201 		 * Search Cyrix-specific descriptor table first ..
2202 		 */
2203 		if ((ct = find_cacheent(cyrix_ctab, *dp)) != NULL) {
2204 			if (func(arg, ct) != 0)
2205 				break;
2206 			continue;
2207 		}
2208 		/*
2209 		 * .. else fall back to the Intel one
2210 		 */
2211 		if ((ct = find_cacheent(intel_ctab, *dp)) != NULL) {
2212 			if (func(arg, ct) != 0)
2213 				break;
2214 			continue;
2215 		}
2216 	}
2217 }
2218 
2219 /*
2220  * A cacheinfo walker that adds associativity, line-size, and size properties
2221  * to the devinfo node it is passed as an argument.
2222  */
2223 static int
2224 add_cacheent_props(void *arg, const struct cachetab *ct)
2225 {
2226 	dev_info_t *devi = arg;
2227 
2228 	add_cache_prop(devi, ct->ct_label, assoc_str, ct->ct_assoc);
2229 	if (ct->ct_line_size != 0)
2230 		add_cache_prop(devi, ct->ct_label, line_str,
2231 		    ct->ct_line_size);
2232 	add_cache_prop(devi, ct->ct_label, size_str, ct->ct_size);
2233 	return (0);
2234 }
2235 
2236 static const char fully_assoc[] = "fully-associative?";
2237 
2238 /*
2239  * AMD style cache/tlb description
2240  *
2241  * Extended functions 5 and 6 directly describe properties of
2242  * tlbs and various cache levels.
2243  */
2244 static void
2245 add_amd_assoc(dev_info_t *devi, const char *label, uint_t assoc)
2246 {
2247 	switch (assoc) {
2248 	case 0:	/* reserved; ignore */
2249 		break;
2250 	default:
2251 		add_cache_prop(devi, label, assoc_str, assoc);
2252 		break;
2253 	case 0xff:
2254 		add_cache_prop(devi, label, fully_assoc, 1);
2255 		break;
2256 	}
2257 }
2258 
2259 static void
2260 add_amd_tlb(dev_info_t *devi, const char *label, uint_t assoc, uint_t size)
2261 {
2262 	if (size == 0)
2263 		return;
2264 	add_cache_prop(devi, label, size_str, size);
2265 	add_amd_assoc(devi, label, assoc);
2266 }
2267 
2268 static void
2269 add_amd_cache(dev_info_t *devi, const char *label,
2270     uint_t size, uint_t assoc, uint_t lines_per_tag, uint_t line_size)
2271 {
2272 	if (size == 0 || line_size == 0)
2273 		return;
2274 	add_amd_assoc(devi, label, assoc);
2275 	/*
2276 	 * Most AMD parts have a sectored cache. Multiple cache lines are
2277 	 * associated with each tag. A sector consists of all cache lines
2278 	 * associated with a tag. For example, the AMD K6-III has a sector
2279 	 * size of 2 cache lines per tag.
2280 	 */
2281 	if (lines_per_tag != 0)
2282 		add_cache_prop(devi, label, "lines-per-tag", lines_per_tag);
2283 	add_cache_prop(devi, label, line_str, line_size);
2284 	add_cache_prop(devi, label, size_str, size * 1024);
2285 }
2286 
2287 static void
2288 add_amd_l2_assoc(dev_info_t *devi, const char *label, uint_t assoc)
2289 {
2290 	switch (assoc) {
2291 	case 0:	/* off */
2292 		break;
2293 	case 1:
2294 	case 2:
2295 	case 4:
2296 		add_cache_prop(devi, label, assoc_str, assoc);
2297 		break;
2298 	case 6:
2299 		add_cache_prop(devi, label, assoc_str, 8);
2300 		break;
2301 	case 8:
2302 		add_cache_prop(devi, label, assoc_str, 16);
2303 		break;
2304 	case 0xf:
2305 		add_cache_prop(devi, label, fully_assoc, 1);
2306 		break;
2307 	default: /* reserved; ignore */
2308 		break;
2309 	}
2310 }
2311 
2312 static void
2313 add_amd_l2_tlb(dev_info_t *devi, const char *label, uint_t assoc, uint_t size)
2314 {
2315 	if (size == 0 || assoc == 0)
2316 		return;
2317 	add_amd_l2_assoc(devi, label, assoc);
2318 	add_cache_prop(devi, label, size_str, size);
2319 }
2320 
2321 static void
2322 add_amd_l2_cache(dev_info_t *devi, const char *label,
2323     uint_t size, uint_t assoc, uint_t lines_per_tag, uint_t line_size)
2324 {
2325 	if (size == 0 || assoc == 0 || line_size == 0)
2326 		return;
2327 	add_amd_l2_assoc(devi, label, assoc);
2328 	if (lines_per_tag != 0)
2329 		add_cache_prop(devi, label, "lines-per-tag", lines_per_tag);
2330 	add_cache_prop(devi, label, line_str, line_size);
2331 	add_cache_prop(devi, label, size_str, size * 1024);
2332 }
2333 
2334 static void
2335 amd_cache_info(struct cpuid_info *cpi, dev_info_t *devi)
2336 {
2337 	struct cpuid_regs *cp;
2338 
2339 	if (cpi->cpi_xmaxeax < 0x80000005)
2340 		return;
2341 	cp = &cpi->cpi_extd[5];
2342 
2343 	/*
2344 	 * 4M/2M L1 TLB configuration
2345 	 *
2346 	 * We report the size for 2M pages because AMD uses two
2347 	 * TLB entries for one 4M page.
2348 	 */
2349 	add_amd_tlb(devi, "dtlb-2M",
2350 	    BITX(cp->cp_eax, 31, 24), BITX(cp->cp_eax, 23, 16));
2351 	add_amd_tlb(devi, "itlb-2M",
2352 	    BITX(cp->cp_eax, 15, 8), BITX(cp->cp_eax, 7, 0));
2353 
2354 	/*
2355 	 * 4K L1 TLB configuration
2356 	 */
2357 
2358 	switch (cpi->cpi_vendor) {
2359 		uint_t nentries;
2360 	case X86_VENDOR_TM:
2361 		if (cpi->cpi_family >= 5) {
2362 			/*
2363 			 * Crusoe processors have 256 TLB entries, but
2364 			 * cpuid data format constrains them to only
2365 			 * reporting 255 of them.
2366 			 */
2367 			if ((nentries = BITX(cp->cp_ebx, 23, 16)) == 255)
2368 				nentries = 256;
2369 			/*
2370 			 * Crusoe processors also have a unified TLB
2371 			 */
2372 			add_amd_tlb(devi, "tlb-4K", BITX(cp->cp_ebx, 31, 24),
2373 			    nentries);
2374 			break;
2375 		}
2376 		/*FALLTHROUGH*/
2377 	default:
2378 		add_amd_tlb(devi, itlb4k_str,
2379 		    BITX(cp->cp_ebx, 31, 24), BITX(cp->cp_ebx, 23, 16));
2380 		add_amd_tlb(devi, dtlb4k_str,
2381 		    BITX(cp->cp_ebx, 15, 8), BITX(cp->cp_ebx, 7, 0));
2382 		break;
2383 	}
2384 
2385 	/*
2386 	 * data L1 cache configuration
2387 	 */
2388 
2389 	add_amd_cache(devi, l1_dcache_str,
2390 	    BITX(cp->cp_ecx, 31, 24), BITX(cp->cp_ecx, 23, 16),
2391 	    BITX(cp->cp_ecx, 15, 8), BITX(cp->cp_ecx, 7, 0));
2392 
2393 	/*
2394 	 * code L1 cache configuration
2395 	 */
2396 
2397 	add_amd_cache(devi, l1_icache_str,
2398 	    BITX(cp->cp_edx, 31, 24), BITX(cp->cp_edx, 23, 16),
2399 	    BITX(cp->cp_edx, 15, 8), BITX(cp->cp_edx, 7, 0));
2400 
2401 	if (cpi->cpi_xmaxeax < 0x80000006)
2402 		return;
2403 	cp = &cpi->cpi_extd[6];
2404 
2405 	/* Check for a unified L2 TLB for large pages */
2406 
2407 	if (BITX(cp->cp_eax, 31, 16) == 0)
2408 		add_amd_l2_tlb(devi, "l2-tlb-2M",
2409 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
2410 	else {
2411 		add_amd_l2_tlb(devi, "l2-dtlb-2M",
2412 		    BITX(cp->cp_eax, 31, 28), BITX(cp->cp_eax, 27, 16));
2413 		add_amd_l2_tlb(devi, "l2-itlb-2M",
2414 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
2415 	}
2416 
2417 	/* Check for a unified L2 TLB for 4K pages */
2418 
2419 	if (BITX(cp->cp_ebx, 31, 16) == 0) {
2420 		add_amd_l2_tlb(devi, "l2-tlb-4K",
2421 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
2422 	} else {
2423 		add_amd_l2_tlb(devi, "l2-dtlb-4K",
2424 		    BITX(cp->cp_eax, 31, 28), BITX(cp->cp_eax, 27, 16));
2425 		add_amd_l2_tlb(devi, "l2-itlb-4K",
2426 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
2427 	}
2428 
2429 	add_amd_l2_cache(devi, l2_cache_str,
2430 	    BITX(cp->cp_ecx, 31, 16), BITX(cp->cp_ecx, 15, 12),
2431 	    BITX(cp->cp_ecx, 11, 8), BITX(cp->cp_ecx, 7, 0));
2432 }
2433 
2434 /*
2435  * There are two basic ways that the x86 world describes it cache
2436  * and tlb architecture - Intel's way and AMD's way.
2437  *
2438  * Return which flavor of cache architecture we should use
2439  */
2440 static int
2441 x86_which_cacheinfo(struct cpuid_info *cpi)
2442 {
2443 	switch (cpi->cpi_vendor) {
2444 	case X86_VENDOR_Intel:
2445 		if (cpi->cpi_maxeax >= 2)
2446 			return (X86_VENDOR_Intel);
2447 		break;
2448 	case X86_VENDOR_AMD:
2449 		/*
2450 		 * The K5 model 1 was the first part from AMD that reported
2451 		 * cache sizes via extended cpuid functions.
2452 		 */
2453 		if (cpi->cpi_family > 5 ||
2454 		    (cpi->cpi_family == 5 && cpi->cpi_model >= 1))
2455 			return (X86_VENDOR_AMD);
2456 		break;
2457 	case X86_VENDOR_TM:
2458 		if (cpi->cpi_family >= 5)
2459 			return (X86_VENDOR_AMD);
2460 		/*FALLTHROUGH*/
2461 	default:
2462 		/*
2463 		 * If they have extended CPU data for 0x80000005
2464 		 * then we assume they have AMD-format cache
2465 		 * information.
2466 		 *
2467 		 * If not, and the vendor happens to be Cyrix,
2468 		 * then try our-Cyrix specific handler.
2469 		 *
2470 		 * If we're not Cyrix, then assume we're using Intel's
2471 		 * table-driven format instead.
2472 		 */
2473 		if (cpi->cpi_xmaxeax >= 0x80000005)
2474 			return (X86_VENDOR_AMD);
2475 		else if (cpi->cpi_vendor == X86_VENDOR_Cyrix)
2476 			return (X86_VENDOR_Cyrix);
2477 		else if (cpi->cpi_maxeax >= 2)
2478 			return (X86_VENDOR_Intel);
2479 		break;
2480 	}
2481 	return (-1);
2482 }
2483 
2484 /*
2485  * create a node for the given cpu under the prom root node.
2486  * Also, create a cpu node in the device tree.
2487  */
2488 static dev_info_t *cpu_nex_devi = NULL;
2489 static kmutex_t cpu_node_lock;
2490 
2491 /*
2492  * Called from post_startup() and mp_startup()
2493  */
2494 void
2495 add_cpunode2devtree(processorid_t cpu_id, struct cpuid_info *cpi)
2496 {
2497 	dev_info_t *cpu_devi;
2498 	int create;
2499 
2500 	mutex_enter(&cpu_node_lock);
2501 
2502 	/*
2503 	 * create a nexus node for all cpus identified as 'cpu_id' under
2504 	 * the root node.
2505 	 */
2506 	if (cpu_nex_devi == NULL) {
2507 		if (ndi_devi_alloc(ddi_root_node(), "cpus",
2508 		    (pnode_t)DEVI_SID_NODEID, &cpu_nex_devi) != NDI_SUCCESS) {
2509 			mutex_exit(&cpu_node_lock);
2510 			return;
2511 		}
2512 		(void) ndi_devi_online(cpu_nex_devi, 0);
2513 	}
2514 
2515 	/*
2516 	 * create a child node for cpu identified as 'cpu_id'
2517 	 */
2518 	cpu_devi = ddi_add_child(cpu_nex_devi, "cpu", DEVI_SID_NODEID,
2519 		cpu_id);
2520 	if (cpu_devi == NULL) {
2521 		mutex_exit(&cpu_node_lock);
2522 		return;
2523 	}
2524 
2525 	/* device_type */
2526 
2527 	(void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi,
2528 	    "device_type", "cpu");
2529 
2530 	/* reg */
2531 
2532 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2533 	    "reg", cpu_id);
2534 
2535 	/* cpu-mhz, and clock-frequency */
2536 
2537 	if (cpu_freq > 0) {
2538 		long long mul;
2539 
2540 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2541 		    "cpu-mhz", cpu_freq);
2542 
2543 		if ((mul = cpu_freq * 1000000LL) <= INT_MAX)
2544 			(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2545 			    "clock-frequency", (int)mul);
2546 	}
2547 
2548 	(void) ndi_devi_online(cpu_devi, 0);
2549 
2550 	if ((x86_feature & X86_CPUID) == 0) {
2551 		mutex_exit(&cpu_node_lock);
2552 		return;
2553 	}
2554 
2555 	/* vendor-id */
2556 
2557 	(void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi,
2558 		"vendor-id", cpi->cpi_vendorstr);
2559 
2560 	if (cpi->cpi_maxeax == 0) {
2561 		mutex_exit(&cpu_node_lock);
2562 		return;
2563 	}
2564 
2565 	/*
2566 	 * family, model, and step
2567 	 */
2568 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2569 		"family", CPI_FAMILY(cpi));
2570 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2571 		"cpu-model", CPI_MODEL(cpi));
2572 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2573 		"stepping-id", CPI_STEP(cpi));
2574 
2575 	/* type */
2576 
2577 	switch (cpi->cpi_vendor) {
2578 	case X86_VENDOR_Intel:
2579 		create = 1;
2580 		break;
2581 	default:
2582 		create = 0;
2583 		break;
2584 	}
2585 	if (create)
2586 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2587 			"type", CPI_TYPE(cpi));
2588 
2589 	/* ext-family */
2590 
2591 	switch (cpi->cpi_vendor) {
2592 	case X86_VENDOR_Intel:
2593 	case X86_VENDOR_AMD:
2594 		create = cpi->cpi_family >= 0xf;
2595 		break;
2596 	default:
2597 		create = 0;
2598 		break;
2599 	}
2600 	if (create)
2601 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2602 		    "ext-family", CPI_FAMILY_XTD(cpi));
2603 
2604 	/* ext-model */
2605 
2606 	switch (cpi->cpi_vendor) {
2607 	case X86_VENDOR_Intel:
2608 		create = CPI_MODEL(cpi) == 0xf;
2609 		break;
2610 	case X86_VENDOR_AMD:
2611 		create = CPI_FAMILY(cpi) == 0xf;
2612 		break;
2613 	default:
2614 		create = 0;
2615 		break;
2616 	}
2617 	if (create)
2618 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2619 			"ext-model", CPI_MODEL_XTD(cpi));
2620 
2621 	/* generation */
2622 
2623 	switch (cpi->cpi_vendor) {
2624 	case X86_VENDOR_AMD:
2625 		/*
2626 		 * AMD K5 model 1 was the first part to support this
2627 		 */
2628 		create = cpi->cpi_xmaxeax >= 0x80000001;
2629 		break;
2630 	default:
2631 		create = 0;
2632 		break;
2633 	}
2634 	if (create)
2635 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2636 		    "generation", BITX((cpi)->cpi_extd[1].cp_eax, 11, 8));
2637 
2638 	/* brand-id */
2639 
2640 	switch (cpi->cpi_vendor) {
2641 	case X86_VENDOR_Intel:
2642 		/*
2643 		 * brand id first appeared on Pentium III Xeon model 8,
2644 		 * and Celeron model 8 processors and Opteron
2645 		 */
2646 		create = cpi->cpi_family > 6 ||
2647 		    (cpi->cpi_family == 6 && cpi->cpi_model >= 8);
2648 		break;
2649 	case X86_VENDOR_AMD:
2650 		create = cpi->cpi_family >= 0xf;
2651 		break;
2652 	default:
2653 		create = 0;
2654 		break;
2655 	}
2656 	if (create && cpi->cpi_brandid != 0) {
2657 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2658 		    "brand-id", cpi->cpi_brandid);
2659 	}
2660 
2661 	/* chunks, and apic-id */
2662 
2663 	switch (cpi->cpi_vendor) {
2664 		/*
2665 		 * first available on Pentium IV and Opteron (K8)
2666 		 */
2667 	case X86_VENDOR_Intel:
2668 		create = IS_NEW_F6(cpi) || cpi->cpi_family >= 0xf;
2669 		break;
2670 	case X86_VENDOR_AMD:
2671 		create = cpi->cpi_family >= 0xf;
2672 		break;
2673 	default:
2674 		create = 0;
2675 		break;
2676 	}
2677 	if (create) {
2678 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2679 			"chunks", CPI_CHUNKS(cpi));
2680 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2681 			"apic-id", CPI_APIC_ID(cpi));
2682 		if (cpi->cpi_chipid >= 0) {
2683 			(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2684 			    "chip#", cpi->cpi_chipid);
2685 			(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2686 			    "clog#", cpi->cpi_clogid);
2687 		}
2688 	}
2689 
2690 	/* cpuid-features */
2691 
2692 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2693 	    "cpuid-features", CPI_FEATURES_EDX(cpi));
2694 
2695 
2696 	/* cpuid-features-ecx */
2697 
2698 	switch (cpi->cpi_vendor) {
2699 	case X86_VENDOR_Intel:
2700 		create = IS_NEW_F6(cpi) || cpi->cpi_family >= 0xf;
2701 		break;
2702 	default:
2703 		create = 0;
2704 		break;
2705 	}
2706 	if (create)
2707 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2708 		    "cpuid-features-ecx", CPI_FEATURES_ECX(cpi));
2709 
2710 	/* ext-cpuid-features */
2711 
2712 	switch (cpi->cpi_vendor) {
2713 	case X86_VENDOR_Intel:
2714 	case X86_VENDOR_AMD:
2715 	case X86_VENDOR_Cyrix:
2716 	case X86_VENDOR_TM:
2717 	case X86_VENDOR_Centaur:
2718 		create = cpi->cpi_xmaxeax >= 0x80000001;
2719 		break;
2720 	default:
2721 		create = 0;
2722 		break;
2723 	}
2724 	if (create) {
2725 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2726 			"ext-cpuid-features", CPI_FEATURES_XTD_EDX(cpi));
2727 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2728 			"ext-cpuid-features-ecx", CPI_FEATURES_XTD_ECX(cpi));
2729 	}
2730 
2731 	/*
2732 	 * Brand String first appeared in Intel Pentium IV, AMD K5
2733 	 * model 1, and Cyrix GXm.  On earlier models we try and
2734 	 * simulate something similar .. so this string should always
2735 	 * same -something- about the processor, however lame.
2736 	 */
2737 	(void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi,
2738 	    "brand-string", cpi->cpi_brandstr);
2739 
2740 	/*
2741 	 * Finally, cache and tlb information
2742 	 */
2743 	switch (x86_which_cacheinfo(cpi)) {
2744 	case X86_VENDOR_Intel:
2745 		intel_walk_cacheinfo(cpi, cpu_devi, add_cacheent_props);
2746 		break;
2747 	case X86_VENDOR_Cyrix:
2748 		cyrix_walk_cacheinfo(cpi, cpu_devi, add_cacheent_props);
2749 		break;
2750 	case X86_VENDOR_AMD:
2751 		amd_cache_info(cpi, cpu_devi);
2752 		break;
2753 	default:
2754 		break;
2755 	}
2756 
2757 	mutex_exit(&cpu_node_lock);
2758 }
2759 
2760 struct l2info {
2761 	int *l2i_csz;
2762 	int *l2i_lsz;
2763 	int *l2i_assoc;
2764 	int l2i_ret;
2765 };
2766 
2767 /*
2768  * A cacheinfo walker that fetches the size, line-size and associativity
2769  * of the L2 cache
2770  */
2771 static int
2772 intel_l2cinfo(void *arg, const struct cachetab *ct)
2773 {
2774 	struct l2info *l2i = arg;
2775 	int *ip;
2776 
2777 	if (ct->ct_label != l2_cache_str &&
2778 	    ct->ct_label != sl2_cache_str)
2779 		return (0);	/* not an L2 -- keep walking */
2780 
2781 	if ((ip = l2i->l2i_csz) != NULL)
2782 		*ip = ct->ct_size;
2783 	if ((ip = l2i->l2i_lsz) != NULL)
2784 		*ip = ct->ct_line_size;
2785 	if ((ip = l2i->l2i_assoc) != NULL)
2786 		*ip = ct->ct_assoc;
2787 	l2i->l2i_ret = ct->ct_size;
2788 	return (1);		/* was an L2 -- terminate walk */
2789 }
2790 
2791 static void
2792 amd_l2cacheinfo(struct cpuid_info *cpi, struct l2info *l2i)
2793 {
2794 	struct cpuid_regs *cp;
2795 	uint_t size, assoc;
2796 	int *ip;
2797 
2798 	if (cpi->cpi_xmaxeax < 0x80000006)
2799 		return;
2800 	cp = &cpi->cpi_extd[6];
2801 
2802 	if ((assoc = BITX(cp->cp_ecx, 15, 12)) != 0 &&
2803 	    (size = BITX(cp->cp_ecx, 31, 16)) != 0) {
2804 		uint_t cachesz = size * 1024;
2805 
2806 
2807 		if ((ip = l2i->l2i_csz) != NULL)
2808 			*ip = cachesz;
2809 		if ((ip = l2i->l2i_lsz) != NULL)
2810 			*ip = BITX(cp->cp_ecx, 7, 0);
2811 		if ((ip = l2i->l2i_assoc) != NULL)
2812 			*ip = assoc;
2813 		l2i->l2i_ret = cachesz;
2814 	}
2815 }
2816 
2817 int
2818 getl2cacheinfo(cpu_t *cpu, int *csz, int *lsz, int *assoc)
2819 {
2820 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2821 	struct l2info __l2info, *l2i = &__l2info;
2822 
2823 	l2i->l2i_csz = csz;
2824 	l2i->l2i_lsz = lsz;
2825 	l2i->l2i_assoc = assoc;
2826 	l2i->l2i_ret = -1;
2827 
2828 	switch (x86_which_cacheinfo(cpi)) {
2829 	case X86_VENDOR_Intel:
2830 		intel_walk_cacheinfo(cpi, l2i, intel_l2cinfo);
2831 		break;
2832 	case X86_VENDOR_Cyrix:
2833 		cyrix_walk_cacheinfo(cpi, l2i, intel_l2cinfo);
2834 		break;
2835 	case X86_VENDOR_AMD:
2836 		amd_l2cacheinfo(cpi, l2i);
2837 		break;
2838 	default:
2839 		break;
2840 	}
2841 	return (l2i->l2i_ret);
2842 }
2843