xref: /titanic_52/usr/src/uts/i86pc/os/cpuid.c (revision 586d07d095794b3debc231dad6a567b554248e19)
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  *  Some undocumented ways of patching the results of the cpuid
198  *  instruction to permit running Solaris 10 on future cpus that
199  *  we don't currently support.  Could be set to non-zero values
200  *  via settings in eeprom.
201  */
202 
203 uint32_t cpuid_feature_ecx_include;
204 uint32_t cpuid_feature_ecx_exclude;
205 uint32_t cpuid_feature_edx_include;
206 uint32_t cpuid_feature_edx_exclude;
207 
208 uint_t
209 cpuid_pass1(cpu_t *cpu)
210 {
211 	uint32_t mask_ecx, mask_edx;
212 	uint_t feature = X86_CPUID;
213 	struct cpuid_info *cpi;
214 	struct cpuid_regs *cp;
215 	int xcpuid;
216 
217 	/*
218 	 * By convention, cpu0 is the boot cpu, which is called
219 	 * before memory allocation is available.  Other cpus are
220 	 * initialized when memory becomes available.
221 	 */
222 	if (cpu->cpu_id == 0)
223 		cpu->cpu_m.mcpu_cpi = cpi = &cpuid_info0;
224 	else
225 		cpu->cpu_m.mcpu_cpi = cpi =
226 		    kmem_zalloc(sizeof (*cpi), KM_SLEEP);
227 
228 	cp = &cpi->cpi_std[0];
229 	cp->cp_eax = 0;
230 	cpi->cpi_maxeax = __cpuid_insn(cp);
231 	{
232 		uint32_t *iptr = (uint32_t *)cpi->cpi_vendorstr;
233 		*iptr++ = cp->cp_ebx;
234 		*iptr++ = cp->cp_edx;
235 		*iptr++ = cp->cp_ecx;
236 		*(char *)&cpi->cpi_vendorstr[12] = '\0';
237 	}
238 
239 	/*
240 	 * Map the vendor string to a type code
241 	 */
242 	if (strcmp(cpi->cpi_vendorstr, "GenuineIntel") == 0)
243 		cpi->cpi_vendor = X86_VENDOR_Intel;
244 	else if (strcmp(cpi->cpi_vendorstr, "AuthenticAMD") == 0)
245 		cpi->cpi_vendor = X86_VENDOR_AMD;
246 	else if (strcmp(cpi->cpi_vendorstr, "GenuineTMx86") == 0)
247 		cpi->cpi_vendor = X86_VENDOR_TM;
248 	else if (strcmp(cpi->cpi_vendorstr, CyrixInstead) == 0)
249 		/*
250 		 * CyrixInstead is a variable used by the Cyrix detection code
251 		 * in locore.
252 		 */
253 		cpi->cpi_vendor = X86_VENDOR_Cyrix;
254 	else if (strcmp(cpi->cpi_vendorstr, "UMC UMC UMC ") == 0)
255 		cpi->cpi_vendor = X86_VENDOR_UMC;
256 	else if (strcmp(cpi->cpi_vendorstr, "NexGenDriven") == 0)
257 		cpi->cpi_vendor = X86_VENDOR_NexGen;
258 	else if (strcmp(cpi->cpi_vendorstr, "CentaurHauls") == 0)
259 		cpi->cpi_vendor = X86_VENDOR_Centaur;
260 	else if (strcmp(cpi->cpi_vendorstr, "RiseRiseRise") == 0)
261 		cpi->cpi_vendor = X86_VENDOR_Rise;
262 	else if (strcmp(cpi->cpi_vendorstr, "SiS SiS SiS ") == 0)
263 		cpi->cpi_vendor = X86_VENDOR_SiS;
264 	else if (strcmp(cpi->cpi_vendorstr, "Geode by NSC") == 0)
265 		cpi->cpi_vendor = X86_VENDOR_NSC;
266 	else
267 		cpi->cpi_vendor = X86_VENDOR_IntelClone;
268 
269 	x86_vendor = cpi->cpi_vendor; /* for compatibility */
270 
271 	/*
272 	 * Limit the range in case of weird hardware
273 	 */
274 	if (cpi->cpi_maxeax > CPI_MAXEAX_MAX)
275 		cpi->cpi_maxeax = CPI_MAXEAX_MAX;
276 	if (cpi->cpi_maxeax < 1)
277 		goto pass1_done;
278 
279 	cp = &cpi->cpi_std[1];
280 	cp->cp_eax = 1;
281 	(void) __cpuid_insn(cp);
282 
283 	/*
284 	 * Extract identifying constants for easy access.
285 	 */
286 	cpi->cpi_model = CPI_MODEL(cpi);
287 	cpi->cpi_family = CPI_FAMILY(cpi);
288 
289 	if (cpi->cpi_family == 0xf) {
290 		cpi->cpi_family += CPI_FAMILY_XTD(cpi);
291 		cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4;
292 	}
293 
294 	cpi->cpi_step = CPI_STEP(cpi);
295 	cpi->cpi_brandid = CPI_BRANDID(cpi);
296 
297 	/*
298 	 * *default* assumptions:
299 	 * - believe %edx feature word
300 	 * - ignore %ecx feature word
301 	 * - 32-bit virtual and physical addressing
302 	 */
303 	mask_edx = 0xffffffff;
304 	mask_ecx = 0;
305 
306 	cpi->cpi_pabits = cpi->cpi_vabits = 32;
307 
308 	switch (cpi->cpi_vendor) {
309 	case X86_VENDOR_Intel:
310 		if (cpi->cpi_family == 5)
311 			x86_type = X86_TYPE_P5;
312 		else if (cpi->cpi_family == 6) {
313 			x86_type = X86_TYPE_P6;
314 			pentiumpro_bug4046376 = 1;
315 			pentiumpro_bug4064495 = 1;
316 			/*
317 			 * Clear the SEP bit when it was set erroneously
318 			 */
319 			if (cpi->cpi_model < 3 && cpi->cpi_step < 3)
320 				cp->cp_edx &= ~CPUID_INTC_EDX_SEP;
321 		} else if (cpi->cpi_family == 0xf) {
322 			x86_type = X86_TYPE_P4;
323 			/*
324 			 * We don't currently depend on any of the %ecx
325 			 * features until Prescott, so we'll only check
326 			 * this from P4 onwards.  We might want to revisit
327 			 * that idea later.
328 			 */
329 			mask_ecx = 0xffffffff;
330 		} else if (cpi->cpi_family > 0xf)
331 			mask_ecx = 0xffffffff;
332 		break;
333 	case X86_VENDOR_IntelClone:
334 	default:
335 		break;
336 	case X86_VENDOR_AMD:
337 #if defined(OPTERON_ERRATUM_108)
338 		if (cpi->cpi_family == 0xf && cpi->cpi_model == 0xe) {
339 			cp->cp_eax = (0xf0f & cp->cp_eax) | 0xc0;
340 			cpi->cpi_model = 0xc;
341 		} else
342 #endif
343 		if (cpi->cpi_family == 5) {
344 			/*
345 			 * AMD K5 and K6
346 			 *
347 			 * These CPUs have an incomplete implementation
348 			 * of MCA/MCE which we mask away.
349 			 */
350 			mask_edx &= ~(CPUID_INTC_EDX_MCE | CPUID_INTC_EDX_MCA);
351 
352 			/*
353 			 * Model 0 uses the wrong (APIC) bit
354 			 * to indicate PGE.  Fix it here.
355 			 */
356 			if (cpi->cpi_model == 0) {
357 				if (cp->cp_edx & 0x200) {
358 					cp->cp_edx &= ~0x200;
359 					cp->cp_edx |= CPUID_INTC_EDX_PGE;
360 				}
361 			}
362 
363 			/*
364 			 * Early models had problems w/ MMX; disable.
365 			 */
366 			if (cpi->cpi_model < 6)
367 				mask_edx &= ~CPUID_INTC_EDX_MMX;
368 		}
369 
370 		/*
371 		 * For newer families, SSE3 and CX16, at least, are valid;
372 		 * enable all
373 		 */
374 		if (cpi->cpi_family >= 0xf)
375 			mask_ecx = 0xffffffff;
376 		break;
377 	case X86_VENDOR_TM:
378 		/*
379 		 * workaround the NT workaround in CMS 4.1
380 		 */
381 		if (cpi->cpi_family == 5 && cpi->cpi_model == 4 &&
382 		    (cpi->cpi_step == 2 || cpi->cpi_step == 3))
383 			cp->cp_edx |= CPUID_INTC_EDX_CX8;
384 		break;
385 	case X86_VENDOR_Centaur:
386 		/*
387 		 * workaround the NT workarounds again
388 		 */
389 		if (cpi->cpi_family == 6)
390 			cp->cp_edx |= CPUID_INTC_EDX_CX8;
391 		break;
392 	case X86_VENDOR_Cyrix:
393 		/*
394 		 * We rely heavily on the probing in locore
395 		 * to actually figure out what parts, if any,
396 		 * of the Cyrix cpuid instruction to believe.
397 		 */
398 		switch (x86_type) {
399 		case X86_TYPE_CYRIX_486:
400 			mask_edx = 0;
401 			break;
402 		case X86_TYPE_CYRIX_6x86:
403 			mask_edx = 0;
404 			break;
405 		case X86_TYPE_CYRIX_6x86L:
406 			mask_edx =
407 			    CPUID_INTC_EDX_DE |
408 			    CPUID_INTC_EDX_CX8;
409 			break;
410 		case X86_TYPE_CYRIX_6x86MX:
411 			mask_edx =
412 			    CPUID_INTC_EDX_DE |
413 			    CPUID_INTC_EDX_MSR |
414 			    CPUID_INTC_EDX_CX8 |
415 			    CPUID_INTC_EDX_PGE |
416 			    CPUID_INTC_EDX_CMOV |
417 			    CPUID_INTC_EDX_MMX;
418 			break;
419 		case X86_TYPE_CYRIX_GXm:
420 			mask_edx =
421 			    CPUID_INTC_EDX_MSR |
422 			    CPUID_INTC_EDX_CX8 |
423 			    CPUID_INTC_EDX_CMOV |
424 			    CPUID_INTC_EDX_MMX;
425 			break;
426 		case X86_TYPE_CYRIX_MediaGX:
427 			break;
428 		case X86_TYPE_CYRIX_MII:
429 		case X86_TYPE_VIA_CYRIX_III:
430 			mask_edx =
431 			    CPUID_INTC_EDX_DE |
432 			    CPUID_INTC_EDX_TSC |
433 			    CPUID_INTC_EDX_MSR |
434 			    CPUID_INTC_EDX_CX8 |
435 			    CPUID_INTC_EDX_PGE |
436 			    CPUID_INTC_EDX_CMOV |
437 			    CPUID_INTC_EDX_MMX;
438 			break;
439 		default:
440 			break;
441 		}
442 		break;
443 	}
444 
445 	/*
446 	 * Now we've figured out the masks that determine
447 	 * which bits we choose to believe, apply the masks
448 	 * to the feature words, then map the kernel's view
449 	 * of these feature words into its feature word.
450 	 */
451 	cp->cp_edx &= mask_edx;
452 	cp->cp_ecx &= mask_ecx;
453 
454 	/*
455 	 * fold in fix ups
456 	 */
457 
458 	cp->cp_edx |= cpuid_feature_edx_include;
459 	cp->cp_edx &= ~cpuid_feature_edx_exclude;
460 
461 
462 	cp->cp_ecx |= cpuid_feature_ecx_include;
463 	cp->cp_ecx &= ~cpuid_feature_ecx_exclude;
464 
465 	if (cp->cp_edx & CPUID_INTC_EDX_PSE)
466 		feature |= X86_LARGEPAGE;
467 	if (cp->cp_edx & CPUID_INTC_EDX_TSC)
468 		feature |= X86_TSC;
469 	if (cp->cp_edx & CPUID_INTC_EDX_MSR)
470 		feature |= X86_MSR;
471 	if (cp->cp_edx & CPUID_INTC_EDX_MTRR)
472 		feature |= X86_MTRR;
473 	if (cp->cp_edx & CPUID_INTC_EDX_PGE)
474 		feature |= X86_PGE;
475 	if (cp->cp_edx & CPUID_INTC_EDX_CMOV)
476 		feature |= X86_CMOV;
477 	if (cp->cp_edx & CPUID_INTC_EDX_MMX)
478 		feature |= X86_MMX;
479 	if ((cp->cp_edx & CPUID_INTC_EDX_MCE) != 0 &&
480 	    (cp->cp_edx & CPUID_INTC_EDX_MCA) != 0)
481 		feature |= X86_MCA;
482 	if (cp->cp_edx & CPUID_INTC_EDX_PAE)
483 		feature |= X86_PAE;
484 	if (cp->cp_edx & CPUID_INTC_EDX_CX8)
485 		feature |= X86_CX8;
486 	/*
487 	 * Once this bit was thought questionable, but it looks like it's
488 	 * back, as of Application Note 485 March 2005 (24161829.pdf)
489 	 */
490 	if (cp->cp_ecx & CPUID_INTC_ECX_CX16)
491 		feature |= X86_CX16;
492 	if (cp->cp_edx & CPUID_INTC_EDX_PAT)
493 		feature |= X86_PAT;
494 	if (cp->cp_edx & CPUID_INTC_EDX_SEP)
495 		feature |= X86_SEP;
496 	if (cp->cp_edx & CPUID_INTC_EDX_FXSR) {
497 		/*
498 		 * In our implementation, fxsave/fxrstor
499 		 * are prerequisites before we'll even
500 		 * try and do SSE things.
501 		 */
502 		if (cp->cp_edx & CPUID_INTC_EDX_SSE)
503 			feature |= X86_SSE;
504 		if (cp->cp_edx & CPUID_INTC_EDX_SSE2)
505 			feature |= X86_SSE2;
506 		if (cp->cp_ecx & CPUID_INTC_ECX_SSE3)
507 			feature |= X86_SSE3;
508 	}
509 	if (cp->cp_edx & CPUID_INTC_EDX_DE)
510 		cr4_value |= CR4_DE;
511 
512 	if (feature & X86_PAE)
513 		cpi->cpi_pabits = 36;
514 
515 	/*
516 	 * Hyperthreading configuration is slightly tricky on Intel
517 	 * and pure clones, and even trickier on AMD.
518 	 *
519 	 * (AMD chose to set the HTT bit on their CMP processors,
520 	 * even though they're not actually hyperthreaded.  Thus it
521 	 * takes a bit more work to figure out what's really going
522 	 * on ... see the handling of the CMP_LEGACY bit below)
523 	 */
524 	if (cp->cp_edx & CPUID_INTC_EDX_HTT) {
525 		cpi->cpi_ncpu_per_chip = CPI_CPU_COUNT(cpi);
526 		if (cpi->cpi_ncpu_per_chip > 1)
527 			feature |= X86_HTT;
528 	} else {
529 		cpi->cpi_ncpu_per_chip = 1;
530 	}
531 
532 	/*
533 	 * Work on the "extended" feature information, doing
534 	 * some basic initialization for cpuid_pass2()
535 	 */
536 	xcpuid = 0;
537 	switch (cpi->cpi_vendor) {
538 	case X86_VENDOR_Intel:
539 		if (cpi->cpi_family >= 0xf)
540 			xcpuid++;
541 		break;
542 	case X86_VENDOR_AMD:
543 		if (cpi->cpi_family > 5 ||
544 		    (cpi->cpi_family == 5 && cpi->cpi_model >= 1))
545 			xcpuid++;
546 		break;
547 	case X86_VENDOR_Cyrix:
548 		/*
549 		 * Only these Cyrix CPUs are -known- to support
550 		 * extended cpuid operations.
551 		 */
552 		if (x86_type == X86_TYPE_VIA_CYRIX_III ||
553 		    x86_type == X86_TYPE_CYRIX_GXm)
554 			xcpuid++;
555 		break;
556 	case X86_VENDOR_Centaur:
557 	case X86_VENDOR_TM:
558 	default:
559 		xcpuid++;
560 		break;
561 	}
562 
563 	if (xcpuid) {
564 		cp = &cpi->cpi_extd[0];
565 		cp->cp_eax = 0x80000000;
566 		cpi->cpi_xmaxeax = __cpuid_insn(cp);
567 	}
568 
569 	if (cpi->cpi_xmaxeax & 0x80000000) {
570 
571 		if (cpi->cpi_xmaxeax > CPI_XMAXEAX_MAX)
572 			cpi->cpi_xmaxeax = CPI_XMAXEAX_MAX;
573 
574 		switch (cpi->cpi_vendor) {
575 		case X86_VENDOR_Intel:
576 		case X86_VENDOR_AMD:
577 			if (cpi->cpi_xmaxeax < 0x80000001)
578 				break;
579 			cp = &cpi->cpi_extd[1];
580 			cp->cp_eax = 0x80000001;
581 			(void) __cpuid_insn(cp);
582 			if (cpi->cpi_vendor == X86_VENDOR_AMD &&
583 			    cpi->cpi_family == 5 &&
584 			    cpi->cpi_model == 6 &&
585 			    cpi->cpi_step == 6) {
586 				/*
587 				 * K6 model 6 uses bit 10 to indicate SYSC
588 				 * Later models use bit 11. Fix it here.
589 				 */
590 				if (cp->cp_edx & 0x400) {
591 					cp->cp_edx &= ~0x400;
592 					cp->cp_edx |= CPUID_AMD_EDX_SYSC;
593 				}
594 			}
595 
596 			/*
597 			 * Compute the additions to the kernel's feature word.
598 			 */
599 			if (cp->cp_edx & CPUID_AMD_EDX_NX)
600 				feature |= X86_NX;
601 
602 			/*
603 			 * If both the HTT and CMP_LEGACY bits are set,
604 			 * then we're not actually HyperThreaded.  Read
605 			 * "AMD CPUID Specification" for more details.
606 			 */
607 			if (cpi->cpi_vendor == X86_VENDOR_AMD &&
608 			    (feature & X86_HTT) &&
609 			    (cp->cp_ecx & CPUID_AMD_ECX_CMP_LEGACY)) {
610 				feature &= ~X86_HTT;
611 				feature |= X86_CMP;
612 			}
613 #if defined(_LP64)
614 			/*
615 			 * It's really tricky to support syscall/sysret in
616 			 * the i386 kernel; we rely on sysenter/sysexit
617 			 * instead.  In the amd64 kernel, things are -way-
618 			 * better.
619 			 */
620 			if (cp->cp_edx & CPUID_AMD_EDX_SYSC)
621 				feature |= X86_ASYSC;
622 
623 			/*
624 			 * While we're thinking about system calls, note
625 			 * that AMD processors don't support sysenter
626 			 * in long mode at all, so don't try to program them.
627 			 */
628 			if (x86_vendor == X86_VENDOR_AMD)
629 				feature &= ~X86_SEP;
630 #endif
631 			break;
632 		default:
633 			break;
634 		}
635 
636 		/*
637 		 * Get CPUID data about processor cores and hyperthreads.
638 		 */
639 		switch (cpi->cpi_vendor) {
640 		case X86_VENDOR_Intel:
641 			if (cpi->cpi_maxeax >= 4) {
642 				cp = &cpi->cpi_std[4];
643 				cp->cp_eax = 4;
644 				cp->cp_ecx = 0;
645 				(void) __cpuid_insn(cp);
646 			}
647 			/*FALLTHROUGH*/
648 		case X86_VENDOR_AMD:
649 			if (cpi->cpi_xmaxeax < 0x80000008)
650 				break;
651 			cp = &cpi->cpi_extd[8];
652 			cp->cp_eax = 0x80000008;
653 			(void) __cpuid_insn(cp);
654 			/*
655 			 * Virtual and physical address limits from
656 			 * cpuid override previously guessed values.
657 			 */
658 			cpi->cpi_pabits = BITX(cp->cp_eax, 7, 0);
659 			cpi->cpi_vabits = BITX(cp->cp_eax, 15, 8);
660 			break;
661 		default:
662 			break;
663 		}
664 
665 		switch (cpi->cpi_vendor) {
666 		case X86_VENDOR_Intel:
667 			if (cpi->cpi_maxeax < 4) {
668 				cpi->cpi_ncore_per_chip = 1;
669 				break;
670 			} else {
671 				cpi->cpi_ncore_per_chip =
672 				    BITX((cpi)->cpi_std[4].cp_eax, 31, 26) + 1;
673 			}
674 			break;
675 		case X86_VENDOR_AMD:
676 			if (cpi->cpi_xmaxeax < 0x80000008) {
677 				cpi->cpi_ncore_per_chip = 1;
678 				break;
679 			} else {
680 				cpi->cpi_ncore_per_chip =
681 				    BITX((cpi)->cpi_extd[8].cp_ecx, 7, 0) + 1;
682 			}
683 			break;
684 		default:
685 			cpi->cpi_ncore_per_chip = 1;
686 			break;
687 		}
688 
689 	}
690 
691 	/*
692 	 * If more than one core, then this processor is CMP.
693 	 */
694 	if (cpi->cpi_ncore_per_chip > 1)
695 		feature |= X86_CMP;
696 	/*
697 	 * If the number of cores is the same as the number
698 	 * of CPUs, then we cannot have HyperThreading.
699 	 */
700 	if (cpi->cpi_ncpu_per_chip == cpi->cpi_ncore_per_chip)
701 		feature &= ~X86_HTT;
702 
703 	if ((feature & (X86_HTT | X86_CMP)) == 0) {
704 		/*
705 		 * Single-core single-threaded processors.
706 		 */
707 		cpi->cpi_chipid = -1;
708 		cpi->cpi_clogid = 0;
709 		cpi->cpi_coreid = cpu->cpu_id;
710 	} else if (cpi->cpi_ncpu_per_chip > 1) {
711 		uint_t i;
712 		uint_t chipid_shift = 0;
713 		uint_t coreid_shift = 0;
714 		uint_t apic_id = CPI_APIC_ID(cpi);
715 
716 		for (i = 1; i < cpi->cpi_ncpu_per_chip; i <<= 1)
717 			chipid_shift++;
718 		cpi->cpi_chipid = apic_id >> chipid_shift;
719 		cpi->cpi_clogid = apic_id & ((1 << chipid_shift) - 1);
720 
721 		if (cpi->cpi_vendor == X86_VENDOR_Intel) {
722 			if (feature & X86_CMP) {
723 				/*
724 				 * Multi-core (and possibly multi-threaded)
725 				 * processors.
726 				 */
727 				uint_t ncpu_per_core;
728 				if (cpi->cpi_ncore_per_chip == 1)
729 					ncpu_per_core = cpi->cpi_ncpu_per_chip;
730 				else if (cpi->cpi_ncore_per_chip > 1)
731 					ncpu_per_core = cpi->cpi_ncpu_per_chip /
732 					    cpi->cpi_ncore_per_chip;
733 				/*
734 				 * 8bit APIC IDs on dual core Pentiums
735 				 * look like this:
736 				 *
737 				 * +-----------------------+------+------+
738 				 * | Physical Package ID   |  MC  |  HT  |
739 				 * +-----------------------+------+------+
740 				 * <------- chipid -------->
741 				 * <------- coreid --------------->
742 				 *			   <--- clogid -->
743 				 *
744 				 * Where the number of bits necessary to
745 				 * represent MC and HT fields together equals
746 				 * to the minimum number of bits necessary to
747 				 * store the value of cpi->cpi_ncpu_per_chip.
748 				 * Of those bits, the MC part uses the number
749 				 * of bits necessary to store the value of
750 				 * cpi->cpi_ncore_per_chip.
751 				 */
752 				for (i = 1; i < ncpu_per_core; i <<= 1)
753 					coreid_shift++;
754 				cpi->cpi_coreid = apic_id >> coreid_shift;
755 			} else if (feature & X86_HTT) {
756 				/*
757 				 * Single-core multi-threaded processors.
758 				 */
759 				cpi->cpi_coreid = cpi->cpi_chipid;
760 			}
761 		} else if (cpi->cpi_vendor == X86_VENDOR_AMD) {
762 			/*
763 			 * AMD currently only has dual-core processors with
764 			 * single-threaded cores.  If they ever release
765 			 * multi-threaded processors, then this code
766 			 * will have to be updated.
767 			 */
768 			cpi->cpi_coreid = cpu->cpu_id;
769 		} else {
770 			/*
771 			 * All other processors are currently
772 			 * assumed to have single cores.
773 			 */
774 			cpi->cpi_coreid = cpi->cpi_chipid;
775 		}
776 	}
777 
778 pass1_done:
779 	cpi->cpi_pass = 1;
780 	return (feature);
781 }
782 
783 /*
784  * Make copies of the cpuid table entries we depend on, in
785  * part for ease of parsing now, in part so that we have only
786  * one place to correct any of it, in part for ease of
787  * later export to userland, and in part so we can look at
788  * this stuff in a crash dump.
789  */
790 
791 /*ARGSUSED*/
792 void
793 cpuid_pass2(cpu_t *cpu)
794 {
795 	uint_t n, nmax;
796 	int i;
797 	struct cpuid_regs *cp;
798 	uint8_t *dp;
799 	uint32_t *iptr;
800 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
801 
802 	ASSERT(cpi->cpi_pass == 1);
803 
804 	if (cpi->cpi_maxeax < 1)
805 		goto pass2_done;
806 
807 	if ((nmax = cpi->cpi_maxeax + 1) > NMAX_CPI_STD)
808 		nmax = NMAX_CPI_STD;
809 	/*
810 	 * (We already handled n == 0 and n == 1 in pass 1)
811 	 */
812 	for (n = 2, cp = &cpi->cpi_std[2]; n < nmax; n++, cp++) {
813 		cp->cp_eax = n;
814 		(void) __cpuid_insn(cp);
815 		switch (n) {
816 		case 2:
817 			/*
818 			 * "the lower 8 bits of the %eax register
819 			 * contain a value that identifies the number
820 			 * of times the cpuid [instruction] has to be
821 			 * executed to obtain a complete image of the
822 			 * processor's caching systems."
823 			 *
824 			 * How *do* they make this stuff up?
825 			 */
826 			cpi->cpi_ncache = sizeof (*cp) *
827 			    BITX(cp->cp_eax, 7, 0);
828 			if (cpi->cpi_ncache == 0)
829 				break;
830 			cpi->cpi_ncache--;	/* skip count byte */
831 
832 			/*
833 			 * Well, for now, rather than attempt to implement
834 			 * this slightly dubious algorithm, we just look
835 			 * at the first 15 ..
836 			 */
837 			if (cpi->cpi_ncache > (sizeof (*cp) - 1))
838 				cpi->cpi_ncache = sizeof (*cp) - 1;
839 
840 			dp = cpi->cpi_cacheinfo;
841 			if (BITX(cp->cp_eax, 31, 31) == 0) {
842 				uint8_t *p = (void *)&cp->cp_eax;
843 				for (i = 1; i < 3; i++)
844 					if (p[i] != 0)
845 						*dp++ = p[i];
846 			}
847 			if (BITX(cp->cp_ebx, 31, 31) == 0) {
848 				uint8_t *p = (void *)&cp->cp_ebx;
849 				for (i = 0; i < 4; i++)
850 					if (p[i] != 0)
851 						*dp++ = p[i];
852 			}
853 			if (BITX(cp->cp_ecx, 31, 31) == 0) {
854 				uint8_t *p = (void *)&cp->cp_ecx;
855 				for (i = 0; i < 4; i++)
856 					if (p[i] != 0)
857 						*dp++ = p[i];
858 			}
859 			if (BITX(cp->cp_edx, 31, 31) == 0) {
860 				uint8_t *p = (void *)&cp->cp_edx;
861 				for (i = 0; i < 4; i++)
862 					if (p[i] != 0)
863 						*dp++ = p[i];
864 			}
865 			break;
866 		case 3:	/* Processor serial number, if PSN supported */
867 		case 4:	/* Deterministic cache parameters */
868 		case 5:	/* Monitor/Mwait parameters */
869 		default:
870 			break;
871 		}
872 	}
873 
874 	if ((cpi->cpi_xmaxeax & 0x80000000) == 0)
875 		goto pass2_done;
876 
877 	if ((nmax = cpi->cpi_xmaxeax - 0x80000000 + 1) > NMAX_CPI_EXTD)
878 		nmax = NMAX_CPI_EXTD;
879 	/*
880 	 * Copy the extended properties, fixing them as we go.
881 	 * (We already handled n == 0 and n == 1 in pass 1)
882 	 */
883 	iptr = (void *)cpi->cpi_brandstr;
884 	for (n = 2, cp = &cpi->cpi_extd[2]; n < nmax; cp++, n++) {
885 		cp->cp_eax = 0x80000000 + n;
886 		(void) __cpuid_insn(cp);
887 		switch (n) {
888 		case 2:
889 		case 3:
890 		case 4:
891 			/*
892 			 * Extract the brand string
893 			 */
894 			*iptr++ = cp->cp_eax;
895 			*iptr++ = cp->cp_ebx;
896 			*iptr++ = cp->cp_ecx;
897 			*iptr++ = cp->cp_edx;
898 			break;
899 		case 5:
900 			switch (cpi->cpi_vendor) {
901 			case X86_VENDOR_AMD:
902 				/*
903 				 * The Athlon and Duron were the first
904 				 * parts to report the sizes of the
905 				 * TLB for large pages. Before then,
906 				 * we don't trust the data.
907 				 */
908 				if (cpi->cpi_family < 6 ||
909 				    (cpi->cpi_family == 6 &&
910 				    cpi->cpi_model < 1))
911 					cp->cp_eax = 0;
912 				break;
913 			default:
914 				break;
915 			}
916 			break;
917 		case 6:
918 			switch (cpi->cpi_vendor) {
919 			case X86_VENDOR_AMD:
920 				/*
921 				 * The Athlon and Duron were the first
922 				 * AMD parts with L2 TLB's.
923 				 * Before then, don't trust the data.
924 				 */
925 				if (cpi->cpi_family < 6 ||
926 				    cpi->cpi_family == 6 &&
927 				    cpi->cpi_model < 1)
928 					cp->cp_eax = cp->cp_ebx = 0;
929 				/*
930 				 * AMD Duron rev A0 reports L2
931 				 * cache size incorrectly as 1K
932 				 * when it is really 64K
933 				 */
934 				if (cpi->cpi_family == 6 &&
935 				    cpi->cpi_model == 3 &&
936 				    cpi->cpi_step == 0) {
937 					cp->cp_ecx &= 0xffff;
938 					cp->cp_ecx |= 0x400000;
939 				}
940 				break;
941 			case X86_VENDOR_Cyrix:	/* VIA C3 */
942 				/*
943 				 * VIA C3 processors are a bit messed
944 				 * up w.r.t. encoding cache sizes in %ecx
945 				 */
946 				if (cpi->cpi_family != 6)
947 					break;
948 				/*
949 				 * model 7 and 8 were incorrectly encoded
950 				 *
951 				 * xxx is model 8 really broken?
952 				 */
953 				if (cpi->cpi_model == 7 ||
954 				    cpi->cpi_model == 8)
955 					cp->cp_ecx =
956 					    BITX(cp->cp_ecx, 31, 24) << 16 |
957 					    BITX(cp->cp_ecx, 23, 16) << 12 |
958 					    BITX(cp->cp_ecx, 15, 8) << 8 |
959 					    BITX(cp->cp_ecx, 7, 0);
960 				/*
961 				 * model 9 stepping 1 has wrong associativity
962 				 */
963 				if (cpi->cpi_model == 9 && cpi->cpi_step == 1)
964 					cp->cp_ecx |= 8 << 12;
965 				break;
966 			case X86_VENDOR_Intel:
967 				/*
968 				 * Extended L2 Cache features function.
969 				 * First appeared on Prescott.
970 				 */
971 			default:
972 				break;
973 			}
974 			break;
975 		default:
976 			break;
977 		}
978 	}
979 
980 pass2_done:
981 	cpi->cpi_pass = 2;
982 }
983 
984 static const char *
985 intel_cpubrand(const struct cpuid_info *cpi)
986 {
987 	int i;
988 
989 	if ((x86_feature & X86_CPUID) == 0 ||
990 	    cpi->cpi_maxeax < 1 || cpi->cpi_family < 5)
991 		return ("i486");
992 
993 	switch (cpi->cpi_family) {
994 	case 5:
995 		return ("Intel Pentium(r)");
996 	case 6:
997 		switch (cpi->cpi_model) {
998 			uint_t celeron, xeon;
999 			const struct cpuid_regs *cp;
1000 		case 0:
1001 		case 1:
1002 		case 2:
1003 			return ("Intel Pentium(r) Pro");
1004 		case 3:
1005 		case 4:
1006 			return ("Intel Pentium(r) II");
1007 		case 6:
1008 			return ("Intel Celeron(r)");
1009 		case 5:
1010 		case 7:
1011 			celeron = xeon = 0;
1012 			cp = &cpi->cpi_std[2];	/* cache info */
1013 
1014 			for (i = 1; i < 3; i++) {
1015 				uint_t tmp;
1016 
1017 				tmp = (cp->cp_eax >> (8 * i)) & 0xff;
1018 				if (tmp == 0x40)
1019 					celeron++;
1020 				if (tmp >= 0x44 && tmp <= 0x45)
1021 					xeon++;
1022 			}
1023 
1024 			for (i = 0; i < 2; i++) {
1025 				uint_t tmp;
1026 
1027 				tmp = (cp->cp_ebx >> (8 * i)) & 0xff;
1028 				if (tmp == 0x40)
1029 					celeron++;
1030 				else if (tmp >= 0x44 && tmp <= 0x45)
1031 					xeon++;
1032 			}
1033 
1034 			for (i = 0; i < 4; i++) {
1035 				uint_t tmp;
1036 
1037 				tmp = (cp->cp_ecx >> (8 * i)) & 0xff;
1038 				if (tmp == 0x40)
1039 					celeron++;
1040 				else if (tmp >= 0x44 && tmp <= 0x45)
1041 					xeon++;
1042 			}
1043 
1044 			for (i = 0; i < 4; i++) {
1045 				uint_t tmp;
1046 
1047 				tmp = (cp->cp_edx >> (8 * i)) & 0xff;
1048 				if (tmp == 0x40)
1049 					celeron++;
1050 				else if (tmp >= 0x44 && tmp <= 0x45)
1051 					xeon++;
1052 			}
1053 
1054 			if (celeron)
1055 				return ("Intel Celeron(r)");
1056 			if (xeon)
1057 				return (cpi->cpi_model == 5 ?
1058 				    "Intel Pentium(r) II Xeon(tm)" :
1059 				    "Intel Pentium(r) III Xeon(tm)");
1060 			return (cpi->cpi_model == 5 ?
1061 			    "Intel Pentium(r) II or Pentium(r) II Xeon(tm)" :
1062 			    "Intel Pentium(r) III or Pentium(r) III Xeon(tm)");
1063 		default:
1064 			break;
1065 		}
1066 	default:
1067 		break;
1068 	}
1069 
1070 	if (cpi->cpi_family <= 0xf && cpi->cpi_model <= 0xf &&
1071 	    cpi->cpi_brandid != 0) {
1072 		static const struct {
1073 			uint_t bt_bid;
1074 			const char *bt_str;
1075 		} brand_tbl[] = {
1076 			{ 0x1,	"Intel(r) Celeron(r)" },
1077 			{ 0x2,	"Intel(r) Pentium(r) III" },
1078 			{ 0x3,	"Intel(r) Pentium(r) III Xeon(tm)" },
1079 			{ 0x4,	"Intel(r) Pentium(r) III" },
1080 			{ 0x6,	"Mobile Intel(r) Pentium(r) III" },
1081 			{ 0x7,	"Mobile Intel(r) Celeron(r)" },
1082 			{ 0x8,	"Intel(r) Pentium(r) 4" },
1083 			{ 0x9,	"Intel(r) Pentium(r) 4" },
1084 			{ 0xa,	"Intel(r) Celeron(r)" },
1085 			{ 0xb,	"Intel(r) Xeon(tm)" },
1086 			{ 0xc,	"Intel(r) Xeon(tm) MP" },
1087 			{ 0xe,	"Mobile Intel(r) Pentium(r) 4" },
1088 			{ 0xf,	"Mobile Intel(r) Celeron(r)" }
1089 		};
1090 		uint_t btblmax = sizeof (brand_tbl) / sizeof (brand_tbl[0]);
1091 		uint_t sgn;
1092 
1093 		sgn = (cpi->cpi_family << 8) |
1094 		    (cpi->cpi_model << 4) | cpi->cpi_step;
1095 
1096 		for (i = 0; i < btblmax; i++)
1097 			if (brand_tbl[i].bt_bid == cpi->cpi_brandid)
1098 				break;
1099 		if (i < btblmax) {
1100 			if (sgn == 0x6b1 && cpi->cpi_brandid == 3)
1101 				return ("Intel(r) Celeron(r)");
1102 			if (sgn < 0xf13 && cpi->cpi_brandid == 0xb)
1103 				return ("Intel(r) Xeon(tm) MP");
1104 			if (sgn < 0xf13 && cpi->cpi_brandid == 0xe)
1105 				return ("Intel(r) Xeon(tm)");
1106 			return (brand_tbl[i].bt_str);
1107 		}
1108 	}
1109 
1110 	return (NULL);
1111 }
1112 
1113 static const char *
1114 amd_cpubrand(const struct cpuid_info *cpi)
1115 {
1116 	if ((x86_feature & X86_CPUID) == 0 ||
1117 	    cpi->cpi_maxeax < 1 || cpi->cpi_family < 5)
1118 		return ("i486 compatible");
1119 
1120 	switch (cpi->cpi_family) {
1121 	case 5:
1122 		switch (cpi->cpi_model) {
1123 		case 0:
1124 		case 1:
1125 		case 2:
1126 		case 3:
1127 		case 4:
1128 		case 5:
1129 			return ("AMD-K5(r)");
1130 		case 6:
1131 		case 7:
1132 			return ("AMD-K6(r)");
1133 		case 8:
1134 			return ("AMD-K6(r)-2");
1135 		case 9:
1136 			return ("AMD-K6(r)-III");
1137 		default:
1138 			return ("AMD (family 5)");
1139 		}
1140 	case 6:
1141 		switch (cpi->cpi_model) {
1142 		case 1:
1143 			return ("AMD-K7(tm)");
1144 		case 0:
1145 		case 2:
1146 		case 4:
1147 			return ("AMD Athlon(tm)");
1148 		case 3:
1149 		case 7:
1150 			return ("AMD Duron(tm)");
1151 		case 6:
1152 		case 8:
1153 		case 10:
1154 			/*
1155 			 * Use the L2 cache size to distinguish
1156 			 */
1157 			return ((cpi->cpi_extd[6].cp_ecx >> 16) >= 256 ?
1158 			    "AMD Athlon(tm)" : "AMD Duron(tm)");
1159 		default:
1160 			return ("AMD (family 6)");
1161 		}
1162 	default:
1163 		break;
1164 	}
1165 
1166 	if (cpi->cpi_family == 0xf && cpi->cpi_model == 5 &&
1167 	    cpi->cpi_brandid != 0) {
1168 		switch (BITX(cpi->cpi_brandid, 7, 5)) {
1169 		case 3:
1170 			return ("AMD Opteron(tm) UP 1xx");
1171 		case 4:
1172 			return ("AMD Opteron(tm) DP 2xx");
1173 		case 5:
1174 			return ("AMD Opteron(tm) MP 8xx");
1175 		default:
1176 			return ("AMD Opteron(tm)");
1177 		}
1178 	}
1179 
1180 	return (NULL);
1181 }
1182 
1183 static const char *
1184 cyrix_cpubrand(struct cpuid_info *cpi, uint_t type)
1185 {
1186 	if ((x86_feature & X86_CPUID) == 0 ||
1187 	    cpi->cpi_maxeax < 1 || cpi->cpi_family < 5 ||
1188 	    type == X86_TYPE_CYRIX_486)
1189 		return ("i486 compatible");
1190 
1191 	switch (type) {
1192 	case X86_TYPE_CYRIX_6x86:
1193 		return ("Cyrix 6x86");
1194 	case X86_TYPE_CYRIX_6x86L:
1195 		return ("Cyrix 6x86L");
1196 	case X86_TYPE_CYRIX_6x86MX:
1197 		return ("Cyrix 6x86MX");
1198 	case X86_TYPE_CYRIX_GXm:
1199 		return ("Cyrix GXm");
1200 	case X86_TYPE_CYRIX_MediaGX:
1201 		return ("Cyrix MediaGX");
1202 	case X86_TYPE_CYRIX_MII:
1203 		return ("Cyrix M2");
1204 	case X86_TYPE_VIA_CYRIX_III:
1205 		return ("VIA Cyrix M3");
1206 	default:
1207 		/*
1208 		 * Have another wild guess ..
1209 		 */
1210 		if (cpi->cpi_family == 4 && cpi->cpi_model == 9)
1211 			return ("Cyrix 5x86");
1212 		else if (cpi->cpi_family == 5) {
1213 			switch (cpi->cpi_model) {
1214 			case 2:
1215 				return ("Cyrix 6x86");	/* Cyrix M1 */
1216 			case 4:
1217 				return ("Cyrix MediaGX");
1218 			default:
1219 				break;
1220 			}
1221 		} else if (cpi->cpi_family == 6) {
1222 			switch (cpi->cpi_model) {
1223 			case 0:
1224 				return ("Cyrix 6x86MX"); /* Cyrix M2? */
1225 			case 5:
1226 			case 6:
1227 			case 7:
1228 			case 8:
1229 			case 9:
1230 				return ("VIA C3");
1231 			default:
1232 				break;
1233 			}
1234 		}
1235 		break;
1236 	}
1237 	return (NULL);
1238 }
1239 
1240 /*
1241  * This only gets called in the case that the CPU extended
1242  * feature brand string (0x80000002, 0x80000003, 0x80000004)
1243  * aren't available, or contain null bytes for some reason.
1244  */
1245 static void
1246 fabricate_brandstr(struct cpuid_info *cpi)
1247 {
1248 	const char *brand = NULL;
1249 
1250 	switch (cpi->cpi_vendor) {
1251 	case X86_VENDOR_Intel:
1252 		brand = intel_cpubrand(cpi);
1253 		break;
1254 	case X86_VENDOR_AMD:
1255 		brand = amd_cpubrand(cpi);
1256 		break;
1257 	case X86_VENDOR_Cyrix:
1258 		brand = cyrix_cpubrand(cpi, x86_type);
1259 		break;
1260 	case X86_VENDOR_NexGen:
1261 		if (cpi->cpi_family == 5 && cpi->cpi_model == 0)
1262 			brand = "NexGen Nx586";
1263 		break;
1264 	case X86_VENDOR_Centaur:
1265 		if (cpi->cpi_family == 5)
1266 			switch (cpi->cpi_model) {
1267 			case 4:
1268 				brand = "Centaur C6";
1269 				break;
1270 			case 8:
1271 				brand = "Centaur C2";
1272 				break;
1273 			case 9:
1274 				brand = "Centaur C3";
1275 				break;
1276 			default:
1277 				break;
1278 			}
1279 		break;
1280 	case X86_VENDOR_Rise:
1281 		if (cpi->cpi_family == 5 &&
1282 		    (cpi->cpi_model == 0 || cpi->cpi_model == 2))
1283 			brand = "Rise mP6";
1284 		break;
1285 	case X86_VENDOR_SiS:
1286 		if (cpi->cpi_family == 5 && cpi->cpi_model == 0)
1287 			brand = "SiS 55x";
1288 		break;
1289 	case X86_VENDOR_TM:
1290 		if (cpi->cpi_family == 5 && cpi->cpi_model == 4)
1291 			brand = "Transmeta Crusoe TM3x00 or TM5x00";
1292 		break;
1293 	case X86_VENDOR_NSC:
1294 	case X86_VENDOR_UMC:
1295 	default:
1296 		break;
1297 	}
1298 	if (brand) {
1299 		(void) strcpy((char *)cpi->cpi_brandstr, brand);
1300 		return;
1301 	}
1302 
1303 	/*
1304 	 * If all else fails ...
1305 	 */
1306 	(void) snprintf(cpi->cpi_brandstr, sizeof (cpi->cpi_brandstr),
1307 	    "%s %d.%d.%d", cpi->cpi_vendorstr, cpi->cpi_family,
1308 	    cpi->cpi_model, cpi->cpi_step);
1309 }
1310 
1311 /*
1312  * This routine is called just after kernel memory allocation
1313  * becomes available on cpu0, and as part of mp_startup() on
1314  * the other cpus.
1315  *
1316  * Fixup the brand string.
1317  */
1318 /*ARGSUSED*/
1319 void
1320 cpuid_pass3(cpu_t *cpu)
1321 {
1322 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
1323 
1324 	ASSERT(cpi->cpi_pass == 2);
1325 
1326 	if ((cpi->cpi_xmaxeax & 0x80000000) == 0) {
1327 		fabricate_brandstr(cpi);
1328 		goto pass3_done;
1329 	}
1330 
1331 	/*
1332 	 * If we successfully extracted a brand string from the cpuid
1333 	 * instruction, clean it up by removing leading spaces and
1334 	 * similar junk.
1335 	 */
1336 	if (cpi->cpi_brandstr[0]) {
1337 		size_t maxlen = sizeof (cpi->cpi_brandstr);
1338 		char *src, *dst;
1339 
1340 		dst = src = (char *)cpi->cpi_brandstr;
1341 		src[maxlen - 1] = '\0';
1342 		/*
1343 		 * strip leading spaces
1344 		 */
1345 		while (*src == ' ')
1346 			src++;
1347 		/*
1348 		 * Remove any 'Genuine' or "Authentic" prefixes
1349 		 */
1350 		if (strncmp(src, "Genuine ", 8) == 0)
1351 			src += 8;
1352 		if (strncmp(src, "Authentic ", 10) == 0)
1353 			src += 10;
1354 
1355 		/*
1356 		 * Now do an in-place copy.
1357 		 * Map (R) to (r) and (TM) to (tm).
1358 		 * The era of teletypes is long gone, and there's
1359 		 * -really- no need to shout.
1360 		 */
1361 		while (*src != '\0') {
1362 			if (src[0] == '(') {
1363 				if (strncmp(src + 1, "R)", 2) == 0) {
1364 					(void) strncpy(dst, "(r)", 3);
1365 					src += 3;
1366 					dst += 3;
1367 					continue;
1368 				}
1369 				if (strncmp(src + 1, "TM)", 3) == 0) {
1370 					(void) strncpy(dst, "(tm)", 4);
1371 					src += 4;
1372 					dst += 4;
1373 					continue;
1374 				}
1375 			}
1376 			*dst++ = *src++;
1377 		}
1378 		*dst = '\0';
1379 
1380 		/*
1381 		 * Finally, remove any trailing spaces
1382 		 */
1383 		while (--dst > cpi->cpi_brandstr)
1384 			if (*dst == ' ')
1385 				*dst = '\0';
1386 			else
1387 				break;
1388 	} else
1389 		fabricate_brandstr(cpi);
1390 
1391 pass3_done:
1392 	cpi->cpi_pass = 3;
1393 }
1394 
1395 /*
1396  * This routine is called out of bind_hwcap() much later in the life
1397  * of the kernel (post_startup()).  The job of this routine is to resolve
1398  * the hardware feature support and kernel support for those features into
1399  * what we're actually going to tell applications via the aux vector.
1400  */
1401 uint_t
1402 cpuid_pass4(cpu_t *cpu)
1403 {
1404 	struct cpuid_info *cpi;
1405 	uint_t hwcap_flags = 0;
1406 
1407 	if (cpu == NULL)
1408 		cpu = CPU;
1409 	cpi = cpu->cpu_m.mcpu_cpi;
1410 
1411 	ASSERT(cpi->cpi_pass == 3);
1412 
1413 	if (cpi->cpi_maxeax >= 1) {
1414 		uint32_t *edx = &cpi->cpi_support[STD_EDX_FEATURES];
1415 		uint32_t *ecx = &cpi->cpi_support[STD_ECX_FEATURES];
1416 
1417 		*edx = CPI_FEATURES_EDX(cpi);
1418 		*ecx = CPI_FEATURES_ECX(cpi);
1419 
1420 		/*
1421 		 * [these require explicit kernel support]
1422 		 */
1423 		if ((x86_feature & X86_SEP) == 0)
1424 			*edx &= ~CPUID_INTC_EDX_SEP;
1425 
1426 		if ((x86_feature & X86_SSE) == 0)
1427 			*edx &= ~(CPUID_INTC_EDX_FXSR|CPUID_INTC_EDX_SSE);
1428 		if ((x86_feature & X86_SSE2) == 0)
1429 			*edx &= ~CPUID_INTC_EDX_SSE2;
1430 
1431 		if ((x86_feature & X86_HTT) == 0)
1432 			*edx &= ~CPUID_INTC_EDX_HTT;
1433 
1434 		if ((x86_feature & X86_SSE3) == 0)
1435 			*ecx &= ~CPUID_INTC_ECX_SSE3;
1436 
1437 		/*
1438 		 * [no explicit support required beyond x87 fp context]
1439 		 */
1440 		if (!fpu_exists)
1441 			*edx &= ~(CPUID_INTC_EDX_FPU | CPUID_INTC_EDX_MMX);
1442 
1443 		/*
1444 		 * Now map the supported feature vector to things that we
1445 		 * think userland will care about.
1446 		 */
1447 		if (*edx & CPUID_INTC_EDX_SEP)
1448 			hwcap_flags |= AV_386_SEP;
1449 		if (*edx & CPUID_INTC_EDX_SSE)
1450 			hwcap_flags |= AV_386_FXSR | AV_386_SSE;
1451 		if (*edx & CPUID_INTC_EDX_SSE2)
1452 			hwcap_flags |= AV_386_SSE2;
1453 		if (*ecx & CPUID_INTC_ECX_SSE3)
1454 			hwcap_flags |= AV_386_SSE3;
1455 
1456 		if (*edx & CPUID_INTC_EDX_FPU)
1457 			hwcap_flags |= AV_386_FPU;
1458 		if (*edx & CPUID_INTC_EDX_MMX)
1459 			hwcap_flags |= AV_386_MMX;
1460 
1461 		if (*edx & CPUID_INTC_EDX_TSC)
1462 			hwcap_flags |= AV_386_TSC;
1463 		if (*edx & CPUID_INTC_EDX_CX8)
1464 			hwcap_flags |= AV_386_CX8;
1465 		if (*edx & CPUID_INTC_EDX_CMOV)
1466 			hwcap_flags |= AV_386_CMOV;
1467 		if (*ecx & CPUID_INTC_ECX_MON)
1468 			hwcap_flags |= AV_386_MON;
1469 #if defined(CPUID_INTC_ECX_CX16)
1470 		if (*ecx & CPUID_INTC_ECX_CX16)
1471 			hwcap_flags |= AV_386_CX16;
1472 #endif
1473 	}
1474 
1475 	if (x86_feature & X86_HTT)
1476 		hwcap_flags |= AV_386_PAUSE;
1477 
1478 	if (cpi->cpi_xmaxeax < 0x80000001)
1479 		goto pass4_done;
1480 
1481 	switch (cpi->cpi_vendor) {
1482 		struct cpuid_regs cp;
1483 		uint32_t *edx;
1484 
1485 	case X86_VENDOR_Intel:	/* sigh */
1486 	case X86_VENDOR_AMD:
1487 		edx = &cpi->cpi_support[AMD_EDX_FEATURES];
1488 
1489 		*edx = CPI_FEATURES_XTD_EDX(cpi);
1490 
1491 		/*
1492 		 * [no explicit support required beyond
1493 		 * x87 fp context and exception handlers]
1494 		 */
1495 		if (!fpu_exists)
1496 			*edx &= ~(CPUID_AMD_EDX_MMXamd |
1497 			    CPUID_AMD_EDX_3DNow | CPUID_AMD_EDX_3DNowx);
1498 
1499 		if ((x86_feature & X86_ASYSC) == 0)
1500 			*edx &= ~CPUID_AMD_EDX_SYSC;
1501 		if ((x86_feature & X86_NX) == 0)
1502 			*edx &= ~CPUID_AMD_EDX_NX;
1503 #if !defined(_LP64)
1504 		*edx &= ~CPUID_AMD_EDX_LM;
1505 #endif
1506 		/*
1507 		 * Now map the supported feature vector to
1508 		 * things that we think userland will care about.
1509 		 */
1510 		if (*edx & CPUID_AMD_EDX_SYSC)
1511 			hwcap_flags |= AV_386_AMD_SYSC;
1512 		if (*edx & CPUID_AMD_EDX_MMXamd)
1513 			hwcap_flags |= AV_386_AMD_MMX;
1514 		if (*edx & CPUID_AMD_EDX_3DNow)
1515 			hwcap_flags |= AV_386_AMD_3DNow;
1516 		if (*edx & CPUID_AMD_EDX_3DNowx)
1517 			hwcap_flags |= AV_386_AMD_3DNowx;
1518 		break;
1519 
1520 	case X86_VENDOR_TM:
1521 		cp.cp_eax = 0x80860001;
1522 		(void) __cpuid_insn(&cp);
1523 		cpi->cpi_support[TM_EDX_FEATURES] = cp.cp_edx;
1524 		break;
1525 
1526 	default:
1527 		break;
1528 	}
1529 
1530 pass4_done:
1531 	cpi->cpi_pass = 4;
1532 	return (hwcap_flags);
1533 }
1534 
1535 
1536 /*
1537  * Simulate the cpuid instruction using the data we previously
1538  * captured about this CPU.  We try our best to return the truth
1539  * about the hardware, independently of kernel support.
1540  */
1541 uint32_t
1542 cpuid_insn(cpu_t *cpu, struct cpuid_regs *cp)
1543 {
1544 	struct cpuid_info *cpi;
1545 	struct cpuid_regs *xcp;
1546 
1547 	if (cpu == NULL)
1548 		cpu = CPU;
1549 	cpi = cpu->cpu_m.mcpu_cpi;
1550 
1551 	ASSERT(cpuid_checkpass(cpu, 3));
1552 
1553 	/*
1554 	 * CPUID data is cached in two separate places: cpi_std for standard
1555 	 * CPUID functions, and cpi_extd for extended CPUID functions.
1556 	 */
1557 	if (cp->cp_eax <= cpi->cpi_maxeax && cp->cp_eax < NMAX_CPI_STD)
1558 		xcp = &cpi->cpi_std[cp->cp_eax];
1559 	else if (cp->cp_eax >= 0x80000000 && cp->cp_eax <= cpi->cpi_xmaxeax &&
1560 	    cp->cp_eax < 0x80000000 + NMAX_CPI_EXTD)
1561 		xcp = &cpi->cpi_extd[cp->cp_eax - 0x80000000];
1562 	else
1563 		/*
1564 		 * The caller is asking for data from an input parameter which
1565 		 * the kernel has not cached.  In this case we go fetch from
1566 		 * the hardware and return the data directly to the user.
1567 		 */
1568 		return (__cpuid_insn(cp));
1569 
1570 	cp->cp_eax = xcp->cp_eax;
1571 	cp->cp_ebx = xcp->cp_ebx;
1572 	cp->cp_ecx = xcp->cp_ecx;
1573 	cp->cp_edx = xcp->cp_edx;
1574 	return (cp->cp_eax);
1575 }
1576 
1577 int
1578 cpuid_checkpass(cpu_t *cpu, int pass)
1579 {
1580 	return (cpu != NULL && cpu->cpu_m.mcpu_cpi != NULL &&
1581 	    cpu->cpu_m.mcpu_cpi->cpi_pass >= pass);
1582 }
1583 
1584 int
1585 cpuid_getbrandstr(cpu_t *cpu, char *s, size_t n)
1586 {
1587 	ASSERT(cpuid_checkpass(cpu, 3));
1588 
1589 	return (snprintf(s, n, "%s", cpu->cpu_m.mcpu_cpi->cpi_brandstr));
1590 }
1591 
1592 int
1593 cpuid_is_cmt(cpu_t *cpu)
1594 {
1595 	if (cpu == NULL)
1596 		cpu = CPU;
1597 
1598 	ASSERT(cpuid_checkpass(cpu, 1));
1599 
1600 	return (cpu->cpu_m.mcpu_cpi->cpi_chipid >= 0);
1601 }
1602 
1603 /*
1604  * AMD and Intel both implement the 64-bit variant of the syscall
1605  * instruction (syscallq), so if there's -any- support for syscall,
1606  * cpuid currently says "yes, we support this".
1607  *
1608  * However, Intel decided to -not- implement the 32-bit variant of the
1609  * syscall instruction, so we provide a predicate to allow our caller
1610  * to test that subtlety here.
1611  */
1612 /*ARGSUSED*/
1613 int
1614 cpuid_syscall32_insn(cpu_t *cpu)
1615 {
1616 	ASSERT(cpuid_checkpass((cpu == NULL ? CPU : cpu), 1));
1617 
1618 	if (x86_feature & X86_ASYSC)
1619 		return (x86_vendor != X86_VENDOR_Intel);
1620 	return (0);
1621 }
1622 
1623 int
1624 cpuid_getidstr(cpu_t *cpu, char *s, size_t n)
1625 {
1626 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
1627 
1628 	static const char fmt[] =
1629 	    "x86 (%s family %d model %d step %d clock %d MHz)";
1630 	static const char fmt_ht[] =
1631 	    "x86 (chipid 0x%x %s family %d model %d step %d clock %d MHz)";
1632 
1633 	ASSERT(cpuid_checkpass(cpu, 1));
1634 
1635 	if (cpuid_is_cmt(cpu))
1636 		return (snprintf(s, n, fmt_ht, cpi->cpi_chipid,
1637 		    cpi->cpi_vendorstr, cpi->cpi_family, cpi->cpi_model,
1638 		    cpi->cpi_step, cpu->cpu_type_info.pi_clock));
1639 	return (snprintf(s, n, fmt,
1640 	    cpi->cpi_vendorstr, cpi->cpi_family, cpi->cpi_model,
1641 	    cpi->cpi_step, cpu->cpu_type_info.pi_clock));
1642 }
1643 
1644 const char *
1645 cpuid_getvendorstr(cpu_t *cpu)
1646 {
1647 	ASSERT(cpuid_checkpass(cpu, 1));
1648 	return ((const char *)cpu->cpu_m.mcpu_cpi->cpi_vendorstr);
1649 }
1650 
1651 uint_t
1652 cpuid_getvendor(cpu_t *cpu)
1653 {
1654 	ASSERT(cpuid_checkpass(cpu, 1));
1655 	return (cpu->cpu_m.mcpu_cpi->cpi_vendor);
1656 }
1657 
1658 uint_t
1659 cpuid_getfamily(cpu_t *cpu)
1660 {
1661 	ASSERT(cpuid_checkpass(cpu, 1));
1662 	return (cpu->cpu_m.mcpu_cpi->cpi_family);
1663 }
1664 
1665 uint_t
1666 cpuid_getmodel(cpu_t *cpu)
1667 {
1668 	ASSERT(cpuid_checkpass(cpu, 1));
1669 	return (cpu->cpu_m.mcpu_cpi->cpi_model);
1670 }
1671 
1672 uint_t
1673 cpuid_get_ncpu_per_chip(cpu_t *cpu)
1674 {
1675 	ASSERT(cpuid_checkpass(cpu, 1));
1676 	return (cpu->cpu_m.mcpu_cpi->cpi_ncpu_per_chip);
1677 }
1678 
1679 uint_t
1680 cpuid_get_ncore_per_chip(cpu_t *cpu)
1681 {
1682 	ASSERT(cpuid_checkpass(cpu, 1));
1683 	return (cpu->cpu_m.mcpu_cpi->cpi_ncore_per_chip);
1684 }
1685 
1686 uint_t
1687 cpuid_getstep(cpu_t *cpu)
1688 {
1689 	ASSERT(cpuid_checkpass(cpu, 1));
1690 	return (cpu->cpu_m.mcpu_cpi->cpi_step);
1691 }
1692 
1693 chipid_t
1694 chip_plat_get_chipid(cpu_t *cpu)
1695 {
1696 	ASSERT(cpuid_checkpass(cpu, 1));
1697 
1698 	if (cpuid_is_cmt(cpu))
1699 		return (cpu->cpu_m.mcpu_cpi->cpi_chipid);
1700 	return (cpu->cpu_id);
1701 }
1702 
1703 id_t
1704 chip_plat_get_coreid(cpu_t *cpu)
1705 {
1706 	ASSERT(cpuid_checkpass(cpu, 1));
1707 	return (cpu->cpu_m.mcpu_cpi->cpi_coreid);
1708 }
1709 
1710 int
1711 chip_plat_get_clogid(cpu_t *cpu)
1712 {
1713 	ASSERT(cpuid_checkpass(cpu, 1));
1714 	return (cpu->cpu_m.mcpu_cpi->cpi_clogid);
1715 }
1716 
1717 void
1718 cpuid_get_addrsize(cpu_t *cpu, uint_t *pabits, uint_t *vabits)
1719 {
1720 	struct cpuid_info *cpi;
1721 
1722 	if (cpu == NULL)
1723 		cpu = CPU;
1724 	cpi = cpu->cpu_m.mcpu_cpi;
1725 
1726 	ASSERT(cpuid_checkpass(cpu, 1));
1727 
1728 	if (pabits)
1729 		*pabits = cpi->cpi_pabits;
1730 	if (vabits)
1731 		*vabits = cpi->cpi_vabits;
1732 }
1733 
1734 /*
1735  * Returns the number of data TLB entries for a corresponding
1736  * pagesize.  If it can't be computed, or isn't known, the
1737  * routine returns zero.  If you ask about an architecturally
1738  * impossible pagesize, the routine will panic (so that the
1739  * hat implementor knows that things are inconsistent.)
1740  */
1741 uint_t
1742 cpuid_get_dtlb_nent(cpu_t *cpu, size_t pagesize)
1743 {
1744 	struct cpuid_info *cpi;
1745 	uint_t dtlb_nent = 0;
1746 
1747 	if (cpu == NULL)
1748 		cpu = CPU;
1749 	cpi = cpu->cpu_m.mcpu_cpi;
1750 
1751 	ASSERT(cpuid_checkpass(cpu, 1));
1752 
1753 	/*
1754 	 * Check the L2 TLB info
1755 	 */
1756 	if (cpi->cpi_xmaxeax >= 0x80000006) {
1757 		struct cpuid_regs *cp = &cpi->cpi_extd[6];
1758 
1759 		switch (pagesize) {
1760 
1761 		case 4 * 1024:
1762 			/*
1763 			 * All zero in the top 16 bits of the register
1764 			 * indicates a unified TLB. Size is in low 16 bits.
1765 			 */
1766 			if ((cp->cp_ebx & 0xffff0000) == 0)
1767 				dtlb_nent = cp->cp_ebx & 0x0000ffff;
1768 			else
1769 				dtlb_nent = BITX(cp->cp_ebx, 27, 16);
1770 			break;
1771 
1772 		case 2 * 1024 * 1024:
1773 			if ((cp->cp_eax & 0xffff0000) == 0)
1774 				dtlb_nent = cp->cp_eax & 0x0000ffff;
1775 			else
1776 				dtlb_nent = BITX(cp->cp_eax, 27, 16);
1777 			break;
1778 
1779 		default:
1780 			panic("unknown L2 pagesize");
1781 			/*NOTREACHED*/
1782 		}
1783 	}
1784 
1785 	if (dtlb_nent != 0)
1786 		return (dtlb_nent);
1787 
1788 	/*
1789 	 * No L2 TLB support for this size, try L1.
1790 	 */
1791 	if (cpi->cpi_xmaxeax >= 0x80000005) {
1792 		struct cpuid_regs *cp = &cpi->cpi_extd[5];
1793 
1794 		switch (pagesize) {
1795 		case 4 * 1024:
1796 			dtlb_nent = BITX(cp->cp_ebx, 23, 16);
1797 			break;
1798 		case 2 * 1024 * 1024:
1799 			dtlb_nent = BITX(cp->cp_eax, 23, 16);
1800 			break;
1801 		default:
1802 			panic("unknown L1 d-TLB pagesize");
1803 			/*NOTREACHED*/
1804 		}
1805 	}
1806 
1807 	return (dtlb_nent);
1808 }
1809 
1810 /*
1811  * Return 0 if the erratum is not present or not applicable, positive
1812  * if it is, and negative if the status of the erratum is unknown.
1813  *
1814  * See "Revision Guide for AMD Athlon(tm) 64 and AMD Opteron(tm)
1815  * Processors" #25759, Rev 3.57, August 2005
1816  */
1817 int
1818 cpuid_opteron_erratum(cpu_t *cpu, uint_t erratum)
1819 {
1820 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
1821 	uint_t eax;
1822 
1823 	if (cpi->cpi_vendor != X86_VENDOR_AMD)
1824 		return (0);
1825 
1826 	eax = cpi->cpi_std[1].cp_eax;
1827 
1828 #define	SH_B0(eax)	(eax == 0xf40 || eax == 0xf50)
1829 #define	SH_B3(eax) 	(eax == 0xf51)
1830 #define	B(eax)		(SH_B0(eax) || SH_B3(eax))
1831 
1832 #define	SH_C0(eax)	(eax == 0xf48 || eax == 0xf58)
1833 
1834 #define	SH_CG(eax)	(eax == 0xf4a || eax == 0xf5a || eax == 0xf7a)
1835 #define	DH_CG(eax)	(eax == 0xfc0 || eax == 0xfe0 || eax == 0xff0)
1836 #define	CH_CG(eax)	(eax == 0xf82 || eax == 0xfb2)
1837 #define	CG(eax)		(SH_CG(eax) || DH_CG(eax) || CH_CG(eax))
1838 
1839 #define	SH_D0(eax)	(eax == 0x10f40 || eax == 0x10f50 || eax == 0x10f70)
1840 #define	DH_D0(eax)	(eax == 0x10fc0 || eax == 0x10ff0)
1841 #define	CH_D0(eax)	(eax == 0x10f80 || eax == 0x10fb0)
1842 #define	D0(eax)		(SH_D0(eax) || DH_D0(eax) || CH_D0(eax))
1843 
1844 #define	SH_E0(eax)	(eax == 0x20f50 || eax == 0x20f40 || eax == 0x20f70)
1845 #define	JH_E1(eax)	(eax == 0x20f10)	/* JH8_E0 had 0x20f30 */
1846 #define	DH_E3(eax)	(eax == 0x20fc0 || eax == 0x20ff0)
1847 #define	SH_E4(eax)	(eax == 0x20f51 || eax == 0x20f71)
1848 #define	BH_E4(eax)	(eax == 0x20fb1)
1849 #define	SH_E5(eax)	(eax == 0x20f42)
1850 #define	DH_E6(eax)	(eax == 0x20ff2 || eax == 0x20fc2)
1851 #define	JH_E6(eax)	(eax == 0x20f12 || eax == 0x20f32)
1852 #define	EX(eax)		(SH_E0(eax) || JH_E1(eax) || DH_E3(eax) || \
1853 			    SH_E4(eax) || BH_E4(eax) || SH_E5(eax) || \
1854 			    DH_E6(eax) || JH_E6(eax))
1855 
1856 	switch (erratum) {
1857 	case 1:
1858 		return (1);
1859 	case 51:	/* what does the asterisk mean? */
1860 		return (B(eax) || SH_C0(eax) || CG(eax));
1861 	case 52:
1862 		return (B(eax));
1863 	case 57:
1864 		return (1);
1865 	case 58:
1866 		return (B(eax));
1867 	case 60:
1868 		return (1);
1869 	case 61:
1870 	case 62:
1871 	case 63:
1872 	case 64:
1873 	case 65:
1874 	case 66:
1875 	case 68:
1876 	case 69:
1877 	case 70:
1878 	case 71:
1879 		return (B(eax));
1880 	case 72:
1881 		return (SH_B0(eax));
1882 	case 74:
1883 		return (B(eax));
1884 	case 75:
1885 		return (1);
1886 	case 76:
1887 		return (B(eax));
1888 	case 77:
1889 		return (1);
1890 	case 78:
1891 		return (B(eax) || SH_C0(eax));
1892 	case 79:
1893 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
1894 	case 80:
1895 	case 81:
1896 	case 82:
1897 		return (B(eax));
1898 	case 83:
1899 		return (B(eax) || SH_C0(eax) || CG(eax));
1900 	case 85:
1901 		return (1);
1902 	case 86:
1903 		return (SH_C0(eax) || CG(eax));
1904 	case 88:
1905 #if !defined(__amd64)
1906 		return (0);
1907 #else
1908 		return (B(eax) || SH_C0(eax));
1909 #endif
1910 	case 89:
1911 		return (1);
1912 	case 90:
1913 		return (B(eax) || SH_C0(eax) || CG(eax));
1914 	case 91:
1915 	case 92:
1916 		return (B(eax) || SH_C0(eax));
1917 	case 93:
1918 		return (SH_C0(eax));
1919 	case 94:
1920 		return (B(eax) || SH_C0(eax) || CG(eax));
1921 	case 95:
1922 #if !defined(__amd64)
1923 		return (0);
1924 #else
1925 		return (B(eax) || SH_C0(eax));
1926 #endif
1927 	case 96:
1928 		return (B(eax) || SH_C0(eax) || CG(eax));
1929 	case 97:
1930 	case 98:
1931 		return (SH_C0(eax) || CG(eax));
1932 	case 99:
1933 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
1934 	case 100:
1935 		return (B(eax) || SH_C0(eax));
1936 	case 101:
1937 	case 103:
1938 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
1939 	case 104:
1940 		return (SH_C0(eax) || CG(eax) || D0(eax));
1941 	case 105:
1942 	case 106:
1943 	case 107:
1944 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
1945 	case 108:
1946 		return (DH_CG(eax));
1947 	case 109:
1948 		return (SH_C0(eax) || CG(eax) || D0(eax));
1949 	case 110:
1950 		return (D0(eax) || EX(eax));
1951 	case 111:
1952 		return (CG(eax));
1953 	case 112:
1954 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
1955 	case 113:
1956 		return (eax == 0x20fc0);
1957 	case 114:
1958 		return (SH_E0(eax) || JH_E1(eax) || DH_E3(eax));
1959 	case 115:
1960 		return (SH_E0(eax) || JH_E1(eax));
1961 	case 116:
1962 		return (SH_E0(eax) || JH_E1(eax) || DH_E3(eax));
1963 	case 117:
1964 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
1965 	case 118:
1966 		return (SH_E0(eax) || JH_E1(eax) || SH_E4(eax) || BH_E4(eax) ||
1967 		    JH_E6(eax));
1968 	case 121:
1969 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
1970 	case 122:
1971 		return (SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
1972 	case 123:
1973 		return (JH_E1(eax) || BH_E4(eax) || JH_E6(eax));
1974 	case 131:
1975 		return (1);
1976 	case 6336786:
1977 		/*
1978 		 * Test for AdvPowerMgmtInfo.TscPStateInvariant
1979 		 * if this is a K8 family processor
1980 		 */
1981 		if (CPI_FAMILY(cpi) == 0xf) {
1982 			struct cpuid_regs regs;
1983 			regs.cp_eax = 0x80000007;
1984 			(void) __cpuid_insn(&regs);
1985 			return (!(regs.cp_edx & 0x100));
1986 		}
1987 		return (0);
1988 	case 6323525:
1989 		return (((((eax >> 12) & 0xff00) + (eax & 0xf00)) |
1990 		    (((eax >> 4) & 0xf) | ((eax >> 12) & 0xf0))) < 0xf40);
1991 
1992 	default:
1993 		return (-1);
1994 	}
1995 }
1996 
1997 static const char assoc_str[] = "associativity";
1998 static const char line_str[] = "line-size";
1999 static const char size_str[] = "size";
2000 
2001 static void
2002 add_cache_prop(dev_info_t *devi, const char *label, const char *type,
2003     uint32_t val)
2004 {
2005 	char buf[128];
2006 
2007 	/*
2008 	 * ndi_prop_update_int() is used because it is desirable for
2009 	 * DDI_PROP_HW_DEF and DDI_PROP_DONTSLEEP to be set.
2010 	 */
2011 	if (snprintf(buf, sizeof (buf), "%s-%s", label, type) < sizeof (buf))
2012 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, devi, buf, val);
2013 }
2014 
2015 /*
2016  * Intel-style cache/tlb description
2017  *
2018  * Standard cpuid level 2 gives a randomly ordered
2019  * selection of tags that index into a table that describes
2020  * cache and tlb properties.
2021  */
2022 
2023 static const char l1_icache_str[] = "l1-icache";
2024 static const char l1_dcache_str[] = "l1-dcache";
2025 static const char l2_cache_str[] = "l2-cache";
2026 static const char itlb4k_str[] = "itlb-4K";
2027 static const char dtlb4k_str[] = "dtlb-4K";
2028 static const char itlb4M_str[] = "itlb-4M";
2029 static const char dtlb4M_str[] = "dtlb-4M";
2030 static const char itlb424_str[] = "itlb-4K-2M-4M";
2031 static const char dtlb44_str[] = "dtlb-4K-4M";
2032 static const char sl1_dcache_str[] = "sectored-l1-dcache";
2033 static const char sl2_cache_str[] = "sectored-l2-cache";
2034 static const char itrace_str[] = "itrace-cache";
2035 static const char sl3_cache_str[] = "sectored-l3-cache";
2036 
2037 static const struct cachetab {
2038 	uint8_t 	ct_code;
2039 	uint8_t		ct_assoc;
2040 	uint16_t 	ct_line_size;
2041 	size_t		ct_size;
2042 	const char	*ct_label;
2043 } intel_ctab[] = {
2044 	/* maintain descending order! */
2045 	{ 0xb3, 4, 0, 128, dtlb4k_str },
2046 	{ 0xb0, 4, 0, 128, itlb4k_str },
2047 	{ 0x87, 8, 64, 1024*1024, l2_cache_str},
2048 	{ 0x86, 4, 64, 512*1024, l2_cache_str},
2049 	{ 0x85, 8, 32, 2*1024*1024, l2_cache_str},
2050 	{ 0x84, 8, 32, 1024*1024, l2_cache_str},
2051 	{ 0x83, 8, 32, 512*1024, l2_cache_str},
2052 	{ 0x82, 8, 32, 256*1024, l2_cache_str},
2053 	{ 0x81, 8, 32, 128*1024, l2_cache_str},		/* suspect! */
2054 	{ 0x7f, 2, 64, 512*1024, l2_cache_str},
2055 	{ 0x7d, 8, 64, 2*1024*1024, sl2_cache_str},
2056 	{ 0x7c, 8, 64, 1024*1024, sl2_cache_str},
2057 	{ 0x7b, 8, 64, 512*1024, sl2_cache_str},
2058 	{ 0x7a, 8, 64, 256*1024, sl2_cache_str},
2059 	{ 0x79, 8, 64, 128*1024, sl2_cache_str},
2060 	{ 0x78, 8, 64, 1024*1024, l2_cache_str},
2061 	{ 0x72, 8, 0, 32*1024, itrace_str},
2062 	{ 0x71, 8, 0, 16*1024, itrace_str},
2063 	{ 0x70, 8, 0, 12*1024, itrace_str},
2064 	{ 0x68, 4, 64, 32*1024, sl1_dcache_str},
2065 	{ 0x67, 4, 64, 16*1024, sl1_dcache_str},
2066 	{ 0x66, 4, 64, 8*1024, sl1_dcache_str},
2067 	{ 0x60, 8, 64, 16*1024, sl1_dcache_str},
2068 	{ 0x5d, 0, 0, 256, dtlb44_str},
2069 	{ 0x5c, 0, 0, 128, dtlb44_str},
2070 	{ 0x5b, 0, 0, 64, dtlb44_str},
2071 	{ 0x52, 0, 0, 256, itlb424_str},
2072 	{ 0x51, 0, 0, 128, itlb424_str},
2073 	{ 0x50, 0, 0, 64, itlb424_str},
2074 	{ 0x45, 4, 32, 2*1024*1024, l2_cache_str},
2075 	{ 0x44, 4, 32, 1024*1024, l2_cache_str},
2076 	{ 0x43, 4, 32, 512*1024, l2_cache_str},
2077 	{ 0x42, 4, 32, 256*1024, l2_cache_str},
2078 	{ 0x41, 4, 32, 128*1024, l2_cache_str},
2079 	{ 0x3c, 4, 64, 256*1024, sl2_cache_str},
2080 	{ 0x3b, 2, 64, 128*1024, sl2_cache_str},
2081 	{ 0x39, 4, 64, 128*1024, sl2_cache_str},
2082 	{ 0x30, 8, 64, 32*1024, l1_icache_str},
2083 	{ 0x2c, 8, 64, 32*1024, l1_dcache_str},
2084 	{ 0x29, 8, 64, 4096*1024, sl3_cache_str},
2085 	{ 0x25, 8, 64, 2048*1024, sl3_cache_str},
2086 	{ 0x23, 8, 64, 1024*1024, sl3_cache_str},
2087 	{ 0x22, 4, 64, 512*1024, sl3_cache_str},
2088 	{ 0x0c, 4, 32, 16*1024, l1_dcache_str},
2089 	{ 0x0a, 2, 32, 8*1024, l1_dcache_str},
2090 	{ 0x08, 4, 32, 16*1024, l1_icache_str},
2091 	{ 0x06, 4, 32, 8*1024, l1_icache_str},
2092 	{ 0x04, 4, 0, 8, dtlb4M_str},
2093 	{ 0x03, 4, 0, 64, dtlb4k_str},
2094 	{ 0x02, 4, 0, 2, itlb4M_str},
2095 	{ 0x01, 4, 0, 32, itlb4k_str},
2096 	{ 0 }
2097 };
2098 
2099 static const struct cachetab cyrix_ctab[] = {
2100 	{ 0x70, 4, 0, 32, "tlb-4K" },
2101 	{ 0x80, 4, 16, 16*1024, "l1-cache" },
2102 	{ 0 }
2103 };
2104 
2105 /*
2106  * Search a cache table for a matching entry
2107  */
2108 static const struct cachetab *
2109 find_cacheent(const struct cachetab *ct, uint_t code)
2110 {
2111 	if (code != 0) {
2112 		for (; ct->ct_code != 0; ct++)
2113 			if (ct->ct_code <= code)
2114 				break;
2115 		if (ct->ct_code == code)
2116 			return (ct);
2117 	}
2118 	return (NULL);
2119 }
2120 
2121 /*
2122  * Walk the cacheinfo descriptor, applying 'func' to every valid element
2123  * The walk is terminated if the walker returns non-zero.
2124  */
2125 static void
2126 intel_walk_cacheinfo(struct cpuid_info *cpi,
2127     void *arg, int (*func)(void *, const struct cachetab *))
2128 {
2129 	const struct cachetab *ct;
2130 	uint8_t *dp;
2131 	int i;
2132 
2133 	if ((dp = cpi->cpi_cacheinfo) == NULL)
2134 		return;
2135 	for (i = 0; i < cpi->cpi_ncache; i++, dp++)
2136 		if ((ct = find_cacheent(intel_ctab, *dp)) != NULL) {
2137 			if (func(arg, ct) != 0)
2138 				break;
2139 		}
2140 }
2141 
2142 /*
2143  * (Like the Intel one, except for Cyrix CPUs)
2144  */
2145 static void
2146 cyrix_walk_cacheinfo(struct cpuid_info *cpi,
2147     void *arg, int (*func)(void *, const struct cachetab *))
2148 {
2149 	const struct cachetab *ct;
2150 	uint8_t *dp;
2151 	int i;
2152 
2153 	if ((dp = cpi->cpi_cacheinfo) == NULL)
2154 		return;
2155 	for (i = 0; i < cpi->cpi_ncache; i++, dp++) {
2156 		/*
2157 		 * Search Cyrix-specific descriptor table first ..
2158 		 */
2159 		if ((ct = find_cacheent(cyrix_ctab, *dp)) != NULL) {
2160 			if (func(arg, ct) != 0)
2161 				break;
2162 			continue;
2163 		}
2164 		/*
2165 		 * .. else fall back to the Intel one
2166 		 */
2167 		if ((ct = find_cacheent(intel_ctab, *dp)) != NULL) {
2168 			if (func(arg, ct) != 0)
2169 				break;
2170 			continue;
2171 		}
2172 	}
2173 }
2174 
2175 /*
2176  * A cacheinfo walker that adds associativity, line-size, and size properties
2177  * to the devinfo node it is passed as an argument.
2178  */
2179 static int
2180 add_cacheent_props(void *arg, const struct cachetab *ct)
2181 {
2182 	dev_info_t *devi = arg;
2183 
2184 	add_cache_prop(devi, ct->ct_label, assoc_str, ct->ct_assoc);
2185 	if (ct->ct_line_size != 0)
2186 		add_cache_prop(devi, ct->ct_label, line_str,
2187 		    ct->ct_line_size);
2188 	add_cache_prop(devi, ct->ct_label, size_str, ct->ct_size);
2189 	return (0);
2190 }
2191 
2192 static const char fully_assoc[] = "fully-associative?";
2193 
2194 /*
2195  * AMD style cache/tlb description
2196  *
2197  * Extended functions 5 and 6 directly describe properties of
2198  * tlbs and various cache levels.
2199  */
2200 static void
2201 add_amd_assoc(dev_info_t *devi, const char *label, uint_t assoc)
2202 {
2203 	switch (assoc) {
2204 	case 0:	/* reserved; ignore */
2205 		break;
2206 	default:
2207 		add_cache_prop(devi, label, assoc_str, assoc);
2208 		break;
2209 	case 0xff:
2210 		add_cache_prop(devi, label, fully_assoc, 1);
2211 		break;
2212 	}
2213 }
2214 
2215 static void
2216 add_amd_tlb(dev_info_t *devi, const char *label, uint_t assoc, uint_t size)
2217 {
2218 	if (size == 0)
2219 		return;
2220 	add_cache_prop(devi, label, size_str, size);
2221 	add_amd_assoc(devi, label, assoc);
2222 }
2223 
2224 static void
2225 add_amd_cache(dev_info_t *devi, const char *label,
2226     uint_t size, uint_t assoc, uint_t lines_per_tag, uint_t line_size)
2227 {
2228 	if (size == 0 || line_size == 0)
2229 		return;
2230 	add_amd_assoc(devi, label, assoc);
2231 	/*
2232 	 * Most AMD parts have a sectored cache. Multiple cache lines are
2233 	 * associated with each tag. A sector consists of all cache lines
2234 	 * associated with a tag. For example, the AMD K6-III has a sector
2235 	 * size of 2 cache lines per tag.
2236 	 */
2237 	if (lines_per_tag != 0)
2238 		add_cache_prop(devi, label, "lines-per-tag", lines_per_tag);
2239 	add_cache_prop(devi, label, line_str, line_size);
2240 	add_cache_prop(devi, label, size_str, size * 1024);
2241 }
2242 
2243 static void
2244 add_amd_l2_assoc(dev_info_t *devi, const char *label, uint_t assoc)
2245 {
2246 	switch (assoc) {
2247 	case 0:	/* off */
2248 		break;
2249 	case 1:
2250 	case 2:
2251 	case 4:
2252 		add_cache_prop(devi, label, assoc_str, assoc);
2253 		break;
2254 	case 6:
2255 		add_cache_prop(devi, label, assoc_str, 8);
2256 		break;
2257 	case 8:
2258 		add_cache_prop(devi, label, assoc_str, 16);
2259 		break;
2260 	case 0xf:
2261 		add_cache_prop(devi, label, fully_assoc, 1);
2262 		break;
2263 	default: /* reserved; ignore */
2264 		break;
2265 	}
2266 }
2267 
2268 static void
2269 add_amd_l2_tlb(dev_info_t *devi, const char *label, uint_t assoc, uint_t size)
2270 {
2271 	if (size == 0 || assoc == 0)
2272 		return;
2273 	add_amd_l2_assoc(devi, label, assoc);
2274 	add_cache_prop(devi, label, size_str, size);
2275 }
2276 
2277 static void
2278 add_amd_l2_cache(dev_info_t *devi, const char *label,
2279     uint_t size, uint_t assoc, uint_t lines_per_tag, uint_t line_size)
2280 {
2281 	if (size == 0 || assoc == 0 || line_size == 0)
2282 		return;
2283 	add_amd_l2_assoc(devi, label, assoc);
2284 	if (lines_per_tag != 0)
2285 		add_cache_prop(devi, label, "lines-per-tag", lines_per_tag);
2286 	add_cache_prop(devi, label, line_str, line_size);
2287 	add_cache_prop(devi, label, size_str, size * 1024);
2288 }
2289 
2290 static void
2291 amd_cache_info(struct cpuid_info *cpi, dev_info_t *devi)
2292 {
2293 	struct cpuid_regs *cp;
2294 
2295 	if (cpi->cpi_xmaxeax < 0x80000005)
2296 		return;
2297 	cp = &cpi->cpi_extd[5];
2298 
2299 	/*
2300 	 * 4M/2M L1 TLB configuration
2301 	 *
2302 	 * We report the size for 2M pages because AMD uses two
2303 	 * TLB entries for one 4M page.
2304 	 */
2305 	add_amd_tlb(devi, "dtlb-2M",
2306 	    BITX(cp->cp_eax, 31, 24), BITX(cp->cp_eax, 23, 16));
2307 	add_amd_tlb(devi, "itlb-2M",
2308 	    BITX(cp->cp_eax, 15, 8), BITX(cp->cp_eax, 7, 0));
2309 
2310 	/*
2311 	 * 4K L1 TLB configuration
2312 	 */
2313 
2314 	switch (cpi->cpi_vendor) {
2315 		uint_t nentries;
2316 	case X86_VENDOR_TM:
2317 		if (cpi->cpi_family >= 5) {
2318 			/*
2319 			 * Crusoe processors have 256 TLB entries, but
2320 			 * cpuid data format constrains them to only
2321 			 * reporting 255 of them.
2322 			 */
2323 			if ((nentries = BITX(cp->cp_ebx, 23, 16)) == 255)
2324 				nentries = 256;
2325 			/*
2326 			 * Crusoe processors also have a unified TLB
2327 			 */
2328 			add_amd_tlb(devi, "tlb-4K", BITX(cp->cp_ebx, 31, 24),
2329 			    nentries);
2330 			break;
2331 		}
2332 		/*FALLTHROUGH*/
2333 	default:
2334 		add_amd_tlb(devi, itlb4k_str,
2335 		    BITX(cp->cp_ebx, 31, 24), BITX(cp->cp_ebx, 23, 16));
2336 		add_amd_tlb(devi, dtlb4k_str,
2337 		    BITX(cp->cp_ebx, 15, 8), BITX(cp->cp_ebx, 7, 0));
2338 		break;
2339 	}
2340 
2341 	/*
2342 	 * data L1 cache configuration
2343 	 */
2344 
2345 	add_amd_cache(devi, l1_dcache_str,
2346 	    BITX(cp->cp_ecx, 31, 24), BITX(cp->cp_ecx, 23, 16),
2347 	    BITX(cp->cp_ecx, 15, 8), BITX(cp->cp_ecx, 7, 0));
2348 
2349 	/*
2350 	 * code L1 cache configuration
2351 	 */
2352 
2353 	add_amd_cache(devi, l1_icache_str,
2354 	    BITX(cp->cp_edx, 31, 24), BITX(cp->cp_edx, 23, 16),
2355 	    BITX(cp->cp_edx, 15, 8), BITX(cp->cp_edx, 7, 0));
2356 
2357 	if (cpi->cpi_xmaxeax < 0x80000006)
2358 		return;
2359 	cp = &cpi->cpi_extd[6];
2360 
2361 	/* Check for a unified L2 TLB for large pages */
2362 
2363 	if (BITX(cp->cp_eax, 31, 16) == 0)
2364 		add_amd_l2_tlb(devi, "l2-tlb-2M",
2365 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
2366 	else {
2367 		add_amd_l2_tlb(devi, "l2-dtlb-2M",
2368 		    BITX(cp->cp_eax, 31, 28), BITX(cp->cp_eax, 27, 16));
2369 		add_amd_l2_tlb(devi, "l2-itlb-2M",
2370 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
2371 	}
2372 
2373 	/* Check for a unified L2 TLB for 4K pages */
2374 
2375 	if (BITX(cp->cp_ebx, 31, 16) == 0) {
2376 		add_amd_l2_tlb(devi, "l2-tlb-4K",
2377 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
2378 	} else {
2379 		add_amd_l2_tlb(devi, "l2-dtlb-4K",
2380 		    BITX(cp->cp_eax, 31, 28), BITX(cp->cp_eax, 27, 16));
2381 		add_amd_l2_tlb(devi, "l2-itlb-4K",
2382 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
2383 	}
2384 
2385 	add_amd_l2_cache(devi, l2_cache_str,
2386 	    BITX(cp->cp_ecx, 31, 16), BITX(cp->cp_ecx, 15, 12),
2387 	    BITX(cp->cp_ecx, 11, 8), BITX(cp->cp_ecx, 7, 0));
2388 }
2389 
2390 /*
2391  * There are two basic ways that the x86 world describes it cache
2392  * and tlb architecture - Intel's way and AMD's way.
2393  *
2394  * Return which flavor of cache architecture we should use
2395  */
2396 static int
2397 x86_which_cacheinfo(struct cpuid_info *cpi)
2398 {
2399 	switch (cpi->cpi_vendor) {
2400 	case X86_VENDOR_Intel:
2401 		if (cpi->cpi_maxeax >= 2)
2402 			return (X86_VENDOR_Intel);
2403 		break;
2404 	case X86_VENDOR_AMD:
2405 		/*
2406 		 * The K5 model 1 was the first part from AMD that reported
2407 		 * cache sizes via extended cpuid functions.
2408 		 */
2409 		if (cpi->cpi_family > 5 ||
2410 		    (cpi->cpi_family == 5 && cpi->cpi_model >= 1))
2411 			return (X86_VENDOR_AMD);
2412 		break;
2413 	case X86_VENDOR_TM:
2414 		if (cpi->cpi_family >= 5)
2415 			return (X86_VENDOR_AMD);
2416 		/*FALLTHROUGH*/
2417 	default:
2418 		/*
2419 		 * If they have extended CPU data for 0x80000005
2420 		 * then we assume they have AMD-format cache
2421 		 * information.
2422 		 *
2423 		 * If not, and the vendor happens to be Cyrix,
2424 		 * then try our-Cyrix specific handler.
2425 		 *
2426 		 * If we're not Cyrix, then assume we're using Intel's
2427 		 * table-driven format instead.
2428 		 */
2429 		if (cpi->cpi_xmaxeax >= 0x80000005)
2430 			return (X86_VENDOR_AMD);
2431 		else if (cpi->cpi_vendor == X86_VENDOR_Cyrix)
2432 			return (X86_VENDOR_Cyrix);
2433 		else if (cpi->cpi_maxeax >= 2)
2434 			return (X86_VENDOR_Intel);
2435 		break;
2436 	}
2437 	return (-1);
2438 }
2439 
2440 /*
2441  * create a node for the given cpu under the prom root node.
2442  * Also, create a cpu node in the device tree.
2443  */
2444 static dev_info_t *cpu_nex_devi = NULL;
2445 static kmutex_t cpu_node_lock;
2446 
2447 /*
2448  * Called from post_startup() and mp_startup()
2449  */
2450 void
2451 add_cpunode2devtree(processorid_t cpu_id, struct cpuid_info *cpi)
2452 {
2453 	dev_info_t *cpu_devi;
2454 	int create;
2455 
2456 	mutex_enter(&cpu_node_lock);
2457 
2458 	/*
2459 	 * create a nexus node for all cpus identified as 'cpu_id' under
2460 	 * the root node.
2461 	 */
2462 	if (cpu_nex_devi == NULL) {
2463 		if (ndi_devi_alloc(ddi_root_node(), "cpus",
2464 		    (pnode_t)DEVI_SID_NODEID, &cpu_nex_devi) != NDI_SUCCESS) {
2465 			mutex_exit(&cpu_node_lock);
2466 			return;
2467 		}
2468 		(void) ndi_devi_online(cpu_nex_devi, 0);
2469 	}
2470 
2471 	/*
2472 	 * create a child node for cpu identified as 'cpu_id'
2473 	 */
2474 	cpu_devi = ddi_add_child(cpu_nex_devi, "cpu", DEVI_SID_NODEID,
2475 		cpu_id);
2476 	if (cpu_devi == NULL) {
2477 		mutex_exit(&cpu_node_lock);
2478 		return;
2479 	}
2480 
2481 	/* device_type */
2482 
2483 	(void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi,
2484 	    "device_type", "cpu");
2485 
2486 	/* reg */
2487 
2488 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2489 	    "reg", cpu_id);
2490 
2491 	/* cpu-mhz, and clock-frequency */
2492 
2493 	if (cpu_freq > 0) {
2494 		long long mul;
2495 
2496 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2497 		    "cpu-mhz", cpu_freq);
2498 
2499 		if ((mul = cpu_freq * 1000000LL) <= INT_MAX)
2500 			(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2501 			    "clock-frequency", (int)mul);
2502 	}
2503 
2504 	(void) ndi_devi_online(cpu_devi, 0);
2505 
2506 	if ((x86_feature & X86_CPUID) == 0) {
2507 		mutex_exit(&cpu_node_lock);
2508 		return;
2509 	}
2510 
2511 	/* vendor-id */
2512 
2513 	(void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi,
2514 		"vendor-id", cpi->cpi_vendorstr);
2515 
2516 	if (cpi->cpi_maxeax == 0) {
2517 		mutex_exit(&cpu_node_lock);
2518 		return;
2519 	}
2520 
2521 	/*
2522 	 * family, model, and step
2523 	 */
2524 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2525 		"family", CPI_FAMILY(cpi));
2526 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2527 		"cpu-model", CPI_MODEL(cpi));
2528 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2529 		"stepping-id", CPI_STEP(cpi));
2530 
2531 	/* type */
2532 
2533 	switch (cpi->cpi_vendor) {
2534 	case X86_VENDOR_Intel:
2535 		create = 1;
2536 		break;
2537 	default:
2538 		create = 0;
2539 		break;
2540 	}
2541 	if (create)
2542 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2543 			"type", CPI_TYPE(cpi));
2544 
2545 	/* ext-family */
2546 
2547 	switch (cpi->cpi_vendor) {
2548 	case X86_VENDOR_Intel:
2549 	case X86_VENDOR_AMD:
2550 		create = cpi->cpi_family >= 0xf;
2551 		break;
2552 	default:
2553 		create = 0;
2554 		break;
2555 	}
2556 	if (create)
2557 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2558 		    "ext-family", CPI_FAMILY_XTD(cpi));
2559 
2560 	/* ext-model */
2561 
2562 	switch (cpi->cpi_vendor) {
2563 	case X86_VENDOR_Intel:
2564 	case X86_VENDOR_AMD:
2565 		create = CPI_FAMILY(cpi) == 0xf;
2566 		break;
2567 	default:
2568 		create = 0;
2569 		break;
2570 	}
2571 	if (create)
2572 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2573 			"ext-model", CPI_MODEL_XTD(cpi));
2574 
2575 	/* generation */
2576 
2577 	switch (cpi->cpi_vendor) {
2578 	case X86_VENDOR_AMD:
2579 		/*
2580 		 * AMD K5 model 1 was the first part to support this
2581 		 */
2582 		create = cpi->cpi_xmaxeax >= 0x80000001;
2583 		break;
2584 	default:
2585 		create = 0;
2586 		break;
2587 	}
2588 	if (create)
2589 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2590 		    "generation", BITX((cpi)->cpi_extd[1].cp_eax, 11, 8));
2591 
2592 	/* brand-id */
2593 
2594 	switch (cpi->cpi_vendor) {
2595 	case X86_VENDOR_Intel:
2596 		/*
2597 		 * brand id first appeared on Pentium III Xeon model 8,
2598 		 * and Celeron model 8 processors and Opteron
2599 		 */
2600 		create = cpi->cpi_family > 6 ||
2601 		    (cpi->cpi_family == 6 && cpi->cpi_model >= 8);
2602 		break;
2603 	case X86_VENDOR_AMD:
2604 		create = cpi->cpi_family >= 0xf;
2605 		break;
2606 	default:
2607 		create = 0;
2608 		break;
2609 	}
2610 	if (create && cpi->cpi_brandid != 0) {
2611 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2612 		    "brand-id", cpi->cpi_brandid);
2613 	}
2614 
2615 	/* chunks, and apic-id */
2616 
2617 	switch (cpi->cpi_vendor) {
2618 	case X86_VENDOR_Intel:
2619 	case X86_VENDOR_AMD:
2620 		/*
2621 		 * first available on Pentium IV and Opteron (K8)
2622 		 */
2623 		create = cpi->cpi_family >= 0xf;
2624 		break;
2625 	default:
2626 		create = 0;
2627 		break;
2628 	}
2629 	if (create) {
2630 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2631 			"chunks", CPI_CHUNKS(cpi));
2632 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2633 			"apic-id", CPI_APIC_ID(cpi));
2634 		if (cpi->cpi_chipid >= 0) {
2635 			(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2636 			    "chip#", cpi->cpi_chipid);
2637 			(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2638 			    "clog#", cpi->cpi_clogid);
2639 		}
2640 	}
2641 
2642 	/* cpuid-features */
2643 
2644 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2645 	    "cpuid-features", CPI_FEATURES_EDX(cpi));
2646 
2647 
2648 	/* cpuid-features-ecx */
2649 
2650 	switch (cpi->cpi_vendor) {
2651 	case X86_VENDOR_Intel:
2652 		create = cpi->cpi_family >= 0xf;
2653 		break;
2654 	default:
2655 		create = 0;
2656 		break;
2657 	}
2658 	if (create)
2659 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2660 		    "cpuid-features-ecx", CPI_FEATURES_ECX(cpi));
2661 
2662 	/* ext-cpuid-features */
2663 
2664 	switch (cpi->cpi_vendor) {
2665 	case X86_VENDOR_AMD:
2666 	case X86_VENDOR_Cyrix:
2667 	case X86_VENDOR_TM:
2668 	case X86_VENDOR_Centaur:
2669 		/*
2670 		 * The extended cpuid features are not relevant on
2671 		 * Intel but are available from the AMD K5 model 1
2672 		 * and most Cyrix GXm and later.
2673 		 */
2674 		create = cpi->cpi_xmaxeax >= 0x80000001;
2675 		break;
2676 	default:
2677 		create = 0;
2678 		break;
2679 	}
2680 	if (create)
2681 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2682 			"ext-cpuid-features", CPI_FEATURES_XTD_EDX(cpi));
2683 
2684 	/*
2685 	 * Brand String first appeared in Intel Pentium IV, AMD K5
2686 	 * model 1, and Cyrix GXm.  On earlier models we try and
2687 	 * simulate something similar .. so this string should always
2688 	 * same -something- about the processor, however lame.
2689 	 */
2690 	(void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi,
2691 	    "brand-string", cpi->cpi_brandstr);
2692 
2693 	/*
2694 	 * Finally, cache and tlb information
2695 	 */
2696 	switch (x86_which_cacheinfo(cpi)) {
2697 	case X86_VENDOR_Intel:
2698 		intel_walk_cacheinfo(cpi, cpu_devi, add_cacheent_props);
2699 		break;
2700 	case X86_VENDOR_Cyrix:
2701 		cyrix_walk_cacheinfo(cpi, cpu_devi, add_cacheent_props);
2702 		break;
2703 	case X86_VENDOR_AMD:
2704 		amd_cache_info(cpi, cpu_devi);
2705 		break;
2706 	default:
2707 		break;
2708 	}
2709 
2710 	mutex_exit(&cpu_node_lock);
2711 }
2712 
2713 struct l2info {
2714 	int *l2i_csz;
2715 	int *l2i_lsz;
2716 	int *l2i_assoc;
2717 	int l2i_ret;
2718 };
2719 
2720 /*
2721  * A cacheinfo walker that fetches the size, line-size and associativity
2722  * of the L2 cache
2723  */
2724 static int
2725 intel_l2cinfo(void *arg, const struct cachetab *ct)
2726 {
2727 	struct l2info *l2i = arg;
2728 	int *ip;
2729 
2730 	if (ct->ct_label != l2_cache_str &&
2731 	    ct->ct_label != sl2_cache_str)
2732 		return (0);	/* not an L2 -- keep walking */
2733 
2734 	if ((ip = l2i->l2i_csz) != NULL)
2735 		*ip = ct->ct_size;
2736 	if ((ip = l2i->l2i_lsz) != NULL)
2737 		*ip = ct->ct_line_size;
2738 	if ((ip = l2i->l2i_assoc) != NULL)
2739 		*ip = ct->ct_assoc;
2740 	l2i->l2i_ret = ct->ct_size;
2741 	return (1);		/* was an L2 -- terminate walk */
2742 }
2743 
2744 static void
2745 amd_l2cacheinfo(struct cpuid_info *cpi, struct l2info *l2i)
2746 {
2747 	struct cpuid_regs *cp;
2748 	uint_t size, assoc;
2749 	int *ip;
2750 
2751 	if (cpi->cpi_xmaxeax < 0x80000006)
2752 		return;
2753 	cp = &cpi->cpi_extd[6];
2754 
2755 	if ((assoc = BITX(cp->cp_ecx, 15, 12)) != 0 &&
2756 	    (size = BITX(cp->cp_ecx, 31, 16)) != 0) {
2757 		uint_t cachesz = size * 1024;
2758 
2759 
2760 		if ((ip = l2i->l2i_csz) != NULL)
2761 			*ip = cachesz;
2762 		if ((ip = l2i->l2i_lsz) != NULL)
2763 			*ip = BITX(cp->cp_ecx, 7, 0);
2764 		if ((ip = l2i->l2i_assoc) != NULL)
2765 			*ip = assoc;
2766 		l2i->l2i_ret = cachesz;
2767 	}
2768 }
2769 
2770 int
2771 getl2cacheinfo(cpu_t *cpu, int *csz, int *lsz, int *assoc)
2772 {
2773 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2774 	struct l2info __l2info, *l2i = &__l2info;
2775 
2776 	l2i->l2i_csz = csz;
2777 	l2i->l2i_lsz = lsz;
2778 	l2i->l2i_assoc = assoc;
2779 	l2i->l2i_ret = -1;
2780 
2781 	switch (x86_which_cacheinfo(cpi)) {
2782 	case X86_VENDOR_Intel:
2783 		intel_walk_cacheinfo(cpi, l2i, intel_l2cinfo);
2784 		break;
2785 	case X86_VENDOR_Cyrix:
2786 		cyrix_walk_cacheinfo(cpi, l2i, intel_l2cinfo);
2787 		break;
2788 	case X86_VENDOR_AMD:
2789 		amd_l2cacheinfo(cpi, l2i);
2790 		break;
2791 	default:
2792 		break;
2793 	}
2794 	return (l2i->l2i_ret);
2795 }
2796