xref: /titanic_51/usr/src/uts/i86pc/os/cpuid.c (revision 1ae0874509b6811fdde1dfd46f0d93fd09867a3f)
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 	if (cpi->cpi_model == 0xf)
315 		cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4;
316 
317 	cpi->cpi_step = CPI_STEP(cpi);
318 	cpi->cpi_brandid = CPI_BRANDID(cpi);
319 
320 	/*
321 	 * *default* assumptions:
322 	 * - believe %edx feature word
323 	 * - ignore %ecx feature word
324 	 * - 32-bit virtual and physical addressing
325 	 */
326 	mask_edx = 0xffffffff;
327 	mask_ecx = 0;
328 
329 	cpi->cpi_pabits = cpi->cpi_vabits = 32;
330 
331 	switch (cpi->cpi_vendor) {
332 	case X86_VENDOR_Intel:
333 		if (cpi->cpi_family == 5)
334 			x86_type = X86_TYPE_P5;
335 		else if (IS_LEGACY_P6(cpi)) {
336 			x86_type = X86_TYPE_P6;
337 			pentiumpro_bug4046376 = 1;
338 			pentiumpro_bug4064495 = 1;
339 			/*
340 			 * Clear the SEP bit when it was set erroneously
341 			 */
342 			if (cpi->cpi_model < 3 && cpi->cpi_step < 3)
343 				cp->cp_edx &= ~CPUID_INTC_EDX_SEP;
344 		} else if (IS_NEW_F6(cpi) || cpi->cpi_family == 0xf) {
345 			x86_type = X86_TYPE_P4;
346 			/*
347 			 * We don't currently depend on any of the %ecx
348 			 * features until Prescott, so we'll only check
349 			 * this from P4 onwards.  We might want to revisit
350 			 * that idea later.
351 			 */
352 			mask_ecx = 0xffffffff;
353 		} else if (cpi->cpi_family > 0xf)
354 			mask_ecx = 0xffffffff;
355 		break;
356 	case X86_VENDOR_IntelClone:
357 	default:
358 		break;
359 	case X86_VENDOR_AMD:
360 #if defined(OPTERON_ERRATUM_108)
361 		if (cpi->cpi_family == 0xf && cpi->cpi_model == 0xe) {
362 			cp->cp_eax = (0xf0f & cp->cp_eax) | 0xc0;
363 			cpi->cpi_model = 0xc;
364 		} else
365 #endif
366 		if (cpi->cpi_family == 5) {
367 			/*
368 			 * AMD K5 and K6
369 			 *
370 			 * These CPUs have an incomplete implementation
371 			 * of MCA/MCE which we mask away.
372 			 */
373 			mask_edx &= ~(CPUID_INTC_EDX_MCE | CPUID_INTC_EDX_MCA);
374 
375 			/*
376 			 * Model 0 uses the wrong (APIC) bit
377 			 * to indicate PGE.  Fix it here.
378 			 */
379 			if (cpi->cpi_model == 0) {
380 				if (cp->cp_edx & 0x200) {
381 					cp->cp_edx &= ~0x200;
382 					cp->cp_edx |= CPUID_INTC_EDX_PGE;
383 				}
384 			}
385 
386 			/*
387 			 * Early models had problems w/ MMX; disable.
388 			 */
389 			if (cpi->cpi_model < 6)
390 				mask_edx &= ~CPUID_INTC_EDX_MMX;
391 		}
392 
393 		/*
394 		 * For newer families, SSE3 and CX16, at least, are valid;
395 		 * enable all
396 		 */
397 		if (cpi->cpi_family >= 0xf)
398 			mask_ecx = 0xffffffff;
399 		break;
400 	case X86_VENDOR_TM:
401 		/*
402 		 * workaround the NT workaround in CMS 4.1
403 		 */
404 		if (cpi->cpi_family == 5 && cpi->cpi_model == 4 &&
405 		    (cpi->cpi_step == 2 || cpi->cpi_step == 3))
406 			cp->cp_edx |= CPUID_INTC_EDX_CX8;
407 		break;
408 	case X86_VENDOR_Centaur:
409 		/*
410 		 * workaround the NT workarounds again
411 		 */
412 		if (cpi->cpi_family == 6)
413 			cp->cp_edx |= CPUID_INTC_EDX_CX8;
414 		break;
415 	case X86_VENDOR_Cyrix:
416 		/*
417 		 * We rely heavily on the probing in locore
418 		 * to actually figure out what parts, if any,
419 		 * of the Cyrix cpuid instruction to believe.
420 		 */
421 		switch (x86_type) {
422 		case X86_TYPE_CYRIX_486:
423 			mask_edx = 0;
424 			break;
425 		case X86_TYPE_CYRIX_6x86:
426 			mask_edx = 0;
427 			break;
428 		case X86_TYPE_CYRIX_6x86L:
429 			mask_edx =
430 			    CPUID_INTC_EDX_DE |
431 			    CPUID_INTC_EDX_CX8;
432 			break;
433 		case X86_TYPE_CYRIX_6x86MX:
434 			mask_edx =
435 			    CPUID_INTC_EDX_DE |
436 			    CPUID_INTC_EDX_MSR |
437 			    CPUID_INTC_EDX_CX8 |
438 			    CPUID_INTC_EDX_PGE |
439 			    CPUID_INTC_EDX_CMOV |
440 			    CPUID_INTC_EDX_MMX;
441 			break;
442 		case X86_TYPE_CYRIX_GXm:
443 			mask_edx =
444 			    CPUID_INTC_EDX_MSR |
445 			    CPUID_INTC_EDX_CX8 |
446 			    CPUID_INTC_EDX_CMOV |
447 			    CPUID_INTC_EDX_MMX;
448 			break;
449 		case X86_TYPE_CYRIX_MediaGX:
450 			break;
451 		case X86_TYPE_CYRIX_MII:
452 		case X86_TYPE_VIA_CYRIX_III:
453 			mask_edx =
454 			    CPUID_INTC_EDX_DE |
455 			    CPUID_INTC_EDX_TSC |
456 			    CPUID_INTC_EDX_MSR |
457 			    CPUID_INTC_EDX_CX8 |
458 			    CPUID_INTC_EDX_PGE |
459 			    CPUID_INTC_EDX_CMOV |
460 			    CPUID_INTC_EDX_MMX;
461 			break;
462 		default:
463 			break;
464 		}
465 		break;
466 	}
467 
468 	/*
469 	 * Now we've figured out the masks that determine
470 	 * which bits we choose to believe, apply the masks
471 	 * to the feature words, then map the kernel's view
472 	 * of these feature words into its feature word.
473 	 */
474 	cp->cp_edx &= mask_edx;
475 	cp->cp_ecx &= mask_ecx;
476 
477 	/*
478 	 * fold in fix ups
479 	 */
480 
481 	cp->cp_edx |= cpuid_feature_edx_include;
482 	cp->cp_edx &= ~cpuid_feature_edx_exclude;
483 
484 
485 	cp->cp_ecx |= cpuid_feature_ecx_include;
486 	cp->cp_ecx &= ~cpuid_feature_ecx_exclude;
487 
488 	if (cp->cp_edx & CPUID_INTC_EDX_PSE)
489 		feature |= X86_LARGEPAGE;
490 	if (cp->cp_edx & CPUID_INTC_EDX_TSC)
491 		feature |= X86_TSC;
492 	if (cp->cp_edx & CPUID_INTC_EDX_MSR)
493 		feature |= X86_MSR;
494 	if (cp->cp_edx & CPUID_INTC_EDX_MTRR)
495 		feature |= X86_MTRR;
496 	if (cp->cp_edx & CPUID_INTC_EDX_PGE)
497 		feature |= X86_PGE;
498 	if (cp->cp_edx & CPUID_INTC_EDX_CMOV)
499 		feature |= X86_CMOV;
500 	if (cp->cp_edx & CPUID_INTC_EDX_MMX)
501 		feature |= X86_MMX;
502 	if ((cp->cp_edx & CPUID_INTC_EDX_MCE) != 0 &&
503 	    (cp->cp_edx & CPUID_INTC_EDX_MCA) != 0)
504 		feature |= X86_MCA;
505 	if (cp->cp_edx & CPUID_INTC_EDX_PAE)
506 		feature |= X86_PAE;
507 	if (cp->cp_edx & CPUID_INTC_EDX_CX8)
508 		feature |= X86_CX8;
509 	/*
510 	 * Once this bit was thought questionable, but it looks like it's
511 	 * back, as of Application Note 485 March 2005 (24161829.pdf)
512 	 */
513 	if (cp->cp_ecx & CPUID_INTC_ECX_CX16)
514 		feature |= X86_CX16;
515 	if (cp->cp_edx & CPUID_INTC_EDX_PAT)
516 		feature |= X86_PAT;
517 	if (cp->cp_edx & CPUID_INTC_EDX_SEP)
518 		feature |= X86_SEP;
519 	if (cp->cp_edx & CPUID_INTC_EDX_FXSR) {
520 		/*
521 		 * In our implementation, fxsave/fxrstor
522 		 * are prerequisites before we'll even
523 		 * try and do SSE things.
524 		 */
525 		if (cp->cp_edx & CPUID_INTC_EDX_SSE)
526 			feature |= X86_SSE;
527 		if (cp->cp_edx & CPUID_INTC_EDX_SSE2)
528 			feature |= X86_SSE2;
529 		if (cp->cp_ecx & CPUID_INTC_ECX_SSE3)
530 			feature |= X86_SSE3;
531 	}
532 	if (cp->cp_edx & CPUID_INTC_EDX_DE)
533 		cr4_value |= CR4_DE;
534 
535 	if (feature & X86_PAE)
536 		cpi->cpi_pabits = 36;
537 
538 	/*
539 	 * Hyperthreading configuration is slightly tricky on Intel
540 	 * and pure clones, and even trickier on AMD.
541 	 *
542 	 * (AMD chose to set the HTT bit on their CMP processors,
543 	 * even though they're not actually hyperthreaded.  Thus it
544 	 * takes a bit more work to figure out what's really going
545 	 * on ... see the handling of the CMP_LEGACY bit below)
546 	 */
547 	if (cp->cp_edx & CPUID_INTC_EDX_HTT) {
548 		cpi->cpi_ncpu_per_chip = CPI_CPU_COUNT(cpi);
549 		if (cpi->cpi_ncpu_per_chip > 1)
550 			feature |= X86_HTT;
551 	} else {
552 		cpi->cpi_ncpu_per_chip = 1;
553 	}
554 
555 	/*
556 	 * Work on the "extended" feature information, doing
557 	 * some basic initialization for cpuid_pass2()
558 	 */
559 	xcpuid = 0;
560 	switch (cpi->cpi_vendor) {
561 	case X86_VENDOR_Intel:
562 		if (IS_NEW_F6(cpi) || cpi->cpi_family >= 0xf)
563 			xcpuid++;
564 		break;
565 	case X86_VENDOR_AMD:
566 		if (cpi->cpi_family > 5 ||
567 		    (cpi->cpi_family == 5 && cpi->cpi_model >= 1))
568 			xcpuid++;
569 		break;
570 	case X86_VENDOR_Cyrix:
571 		/*
572 		 * Only these Cyrix CPUs are -known- to support
573 		 * extended cpuid operations.
574 		 */
575 		if (x86_type == X86_TYPE_VIA_CYRIX_III ||
576 		    x86_type == X86_TYPE_CYRIX_GXm)
577 			xcpuid++;
578 		break;
579 	case X86_VENDOR_Centaur:
580 	case X86_VENDOR_TM:
581 	default:
582 		xcpuid++;
583 		break;
584 	}
585 
586 	if (xcpuid) {
587 		cp = &cpi->cpi_extd[0];
588 		cp->cp_eax = 0x80000000;
589 		cpi->cpi_xmaxeax = __cpuid_insn(cp);
590 	}
591 
592 	if (cpi->cpi_xmaxeax & 0x80000000) {
593 
594 		if (cpi->cpi_xmaxeax > CPI_XMAXEAX_MAX)
595 			cpi->cpi_xmaxeax = CPI_XMAXEAX_MAX;
596 
597 		switch (cpi->cpi_vendor) {
598 		case X86_VENDOR_Intel:
599 		case X86_VENDOR_AMD:
600 			if (cpi->cpi_xmaxeax < 0x80000001)
601 				break;
602 			cp = &cpi->cpi_extd[1];
603 			cp->cp_eax = 0x80000001;
604 			(void) __cpuid_insn(cp);
605 			if (cpi->cpi_vendor == X86_VENDOR_AMD &&
606 			    cpi->cpi_family == 5 &&
607 			    cpi->cpi_model == 6 &&
608 			    cpi->cpi_step == 6) {
609 				/*
610 				 * K6 model 6 uses bit 10 to indicate SYSC
611 				 * Later models use bit 11. Fix it here.
612 				 */
613 				if (cp->cp_edx & 0x400) {
614 					cp->cp_edx &= ~0x400;
615 					cp->cp_edx |= CPUID_AMD_EDX_SYSC;
616 				}
617 			}
618 
619 			/*
620 			 * Compute the additions to the kernel's feature word.
621 			 */
622 			if (cp->cp_edx & CPUID_AMD_EDX_NX)
623 				feature |= X86_NX;
624 
625 			/*
626 			 * If both the HTT and CMP_LEGACY bits are set,
627 			 * then we're not actually HyperThreaded.  Read
628 			 * "AMD CPUID Specification" for more details.
629 			 */
630 			if (cpi->cpi_vendor == X86_VENDOR_AMD &&
631 			    (feature & X86_HTT) &&
632 			    (cp->cp_ecx & CPUID_AMD_ECX_CMP_LEGACY)) {
633 				feature &= ~X86_HTT;
634 				feature |= X86_CMP;
635 			}
636 #if defined(_LP64)
637 			/*
638 			 * It's really tricky to support syscall/sysret in
639 			 * the i386 kernel; we rely on sysenter/sysexit
640 			 * instead.  In the amd64 kernel, things are -way-
641 			 * better.
642 			 */
643 			if (cp->cp_edx & CPUID_AMD_EDX_SYSC)
644 				feature |= X86_ASYSC;
645 
646 			/*
647 			 * While we're thinking about system calls, note
648 			 * that AMD processors don't support sysenter
649 			 * in long mode at all, so don't try to program them.
650 			 */
651 			if (x86_vendor == X86_VENDOR_AMD)
652 				feature &= ~X86_SEP;
653 #endif
654 			break;
655 		default:
656 			break;
657 		}
658 
659 		/*
660 		 * Get CPUID data about processor cores and hyperthreads.
661 		 */
662 		switch (cpi->cpi_vendor) {
663 		case X86_VENDOR_Intel:
664 			if (cpi->cpi_maxeax >= 4) {
665 				cp = &cpi->cpi_std[4];
666 				cp->cp_eax = 4;
667 				cp->cp_ecx = 0;
668 				(void) __cpuid_insn(cp);
669 			}
670 			/*FALLTHROUGH*/
671 		case X86_VENDOR_AMD:
672 			if (cpi->cpi_xmaxeax < 0x80000008)
673 				break;
674 			cp = &cpi->cpi_extd[8];
675 			cp->cp_eax = 0x80000008;
676 			(void) __cpuid_insn(cp);
677 			/*
678 			 * Virtual and physical address limits from
679 			 * cpuid override previously guessed values.
680 			 */
681 			cpi->cpi_pabits = BITX(cp->cp_eax, 7, 0);
682 			cpi->cpi_vabits = BITX(cp->cp_eax, 15, 8);
683 			break;
684 		default:
685 			break;
686 		}
687 
688 		switch (cpi->cpi_vendor) {
689 		case X86_VENDOR_Intel:
690 			if (cpi->cpi_maxeax < 4) {
691 				cpi->cpi_ncore_per_chip = 1;
692 				break;
693 			} else {
694 				cpi->cpi_ncore_per_chip =
695 				    BITX((cpi)->cpi_std[4].cp_eax, 31, 26) + 1;
696 			}
697 			break;
698 		case X86_VENDOR_AMD:
699 			if (cpi->cpi_xmaxeax < 0x80000008) {
700 				cpi->cpi_ncore_per_chip = 1;
701 				break;
702 			} else {
703 				cpi->cpi_ncore_per_chip =
704 				    BITX((cpi)->cpi_extd[8].cp_ecx, 7, 0) + 1;
705 			}
706 			break;
707 		default:
708 			cpi->cpi_ncore_per_chip = 1;
709 			break;
710 		}
711 
712 	}
713 
714 	/*
715 	 * If more than one core, then this processor is CMP.
716 	 */
717 	if (cpi->cpi_ncore_per_chip > 1)
718 		feature |= X86_CMP;
719 	/*
720 	 * If the number of cores is the same as the number
721 	 * of CPUs, then we cannot have HyperThreading.
722 	 */
723 	if (cpi->cpi_ncpu_per_chip == cpi->cpi_ncore_per_chip)
724 		feature &= ~X86_HTT;
725 
726 	if ((feature & (X86_HTT | X86_CMP)) == 0) {
727 		/*
728 		 * Single-core single-threaded processors.
729 		 */
730 		cpi->cpi_chipid = -1;
731 		cpi->cpi_clogid = 0;
732 		cpi->cpi_coreid = cpu->cpu_id;
733 	} else if (cpi->cpi_ncpu_per_chip > 1) {
734 		uint_t i;
735 		uint_t chipid_shift = 0;
736 		uint_t coreid_shift = 0;
737 		uint_t apic_id = CPI_APIC_ID(cpi);
738 
739 		for (i = 1; i < cpi->cpi_ncpu_per_chip; i <<= 1)
740 			chipid_shift++;
741 		cpi->cpi_chipid = apic_id >> chipid_shift;
742 		cpi->cpi_clogid = apic_id & ((1 << chipid_shift) - 1);
743 
744 		if (cpi->cpi_vendor == X86_VENDOR_Intel) {
745 			if (feature & X86_CMP) {
746 				/*
747 				 * Multi-core (and possibly multi-threaded)
748 				 * processors.
749 				 */
750 				uint_t ncpu_per_core;
751 				if (cpi->cpi_ncore_per_chip == 1)
752 					ncpu_per_core = cpi->cpi_ncpu_per_chip;
753 				else if (cpi->cpi_ncore_per_chip > 1)
754 					ncpu_per_core = cpi->cpi_ncpu_per_chip /
755 					    cpi->cpi_ncore_per_chip;
756 				/*
757 				 * 8bit APIC IDs on dual core Pentiums
758 				 * look like this:
759 				 *
760 				 * +-----------------------+------+------+
761 				 * | Physical Package ID   |  MC  |  HT  |
762 				 * +-----------------------+------+------+
763 				 * <------- chipid -------->
764 				 * <------- coreid --------------->
765 				 *			   <--- clogid -->
766 				 *
767 				 * Where the number of bits necessary to
768 				 * represent MC and HT fields together equals
769 				 * to the minimum number of bits necessary to
770 				 * store the value of cpi->cpi_ncpu_per_chip.
771 				 * Of those bits, the MC part uses the number
772 				 * of bits necessary to store the value of
773 				 * cpi->cpi_ncore_per_chip.
774 				 */
775 				for (i = 1; i < ncpu_per_core; i <<= 1)
776 					coreid_shift++;
777 				cpi->cpi_coreid = apic_id >> coreid_shift;
778 			} else if (feature & X86_HTT) {
779 				/*
780 				 * Single-core multi-threaded processors.
781 				 */
782 				cpi->cpi_coreid = cpi->cpi_chipid;
783 			}
784 		} else if (cpi->cpi_vendor == X86_VENDOR_AMD) {
785 			/*
786 			 * AMD currently only has dual-core processors with
787 			 * single-threaded cores.  If they ever release
788 			 * multi-threaded processors, then this code
789 			 * will have to be updated.
790 			 */
791 			cpi->cpi_coreid = cpu->cpu_id;
792 		} else {
793 			/*
794 			 * All other processors are currently
795 			 * assumed to have single cores.
796 			 */
797 			cpi->cpi_coreid = cpi->cpi_chipid;
798 		}
799 	}
800 
801 pass1_done:
802 	cpi->cpi_pass = 1;
803 	return (feature);
804 }
805 
806 /*
807  * Make copies of the cpuid table entries we depend on, in
808  * part for ease of parsing now, in part so that we have only
809  * one place to correct any of it, in part for ease of
810  * later export to userland, and in part so we can look at
811  * this stuff in a crash dump.
812  */
813 
814 /*ARGSUSED*/
815 void
816 cpuid_pass2(cpu_t *cpu)
817 {
818 	uint_t n, nmax;
819 	int i;
820 	struct cpuid_regs *cp;
821 	uint8_t *dp;
822 	uint32_t *iptr;
823 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
824 
825 	ASSERT(cpi->cpi_pass == 1);
826 
827 	if (cpi->cpi_maxeax < 1)
828 		goto pass2_done;
829 
830 	if ((nmax = cpi->cpi_maxeax + 1) > NMAX_CPI_STD)
831 		nmax = NMAX_CPI_STD;
832 	/*
833 	 * (We already handled n == 0 and n == 1 in pass 1)
834 	 */
835 	for (n = 2, cp = &cpi->cpi_std[2]; n < nmax; n++, cp++) {
836 		cp->cp_eax = n;
837 		(void) __cpuid_insn(cp);
838 		switch (n) {
839 		case 2:
840 			/*
841 			 * "the lower 8 bits of the %eax register
842 			 * contain a value that identifies the number
843 			 * of times the cpuid [instruction] has to be
844 			 * executed to obtain a complete image of the
845 			 * processor's caching systems."
846 			 *
847 			 * How *do* they make this stuff up?
848 			 */
849 			cpi->cpi_ncache = sizeof (*cp) *
850 			    BITX(cp->cp_eax, 7, 0);
851 			if (cpi->cpi_ncache == 0)
852 				break;
853 			cpi->cpi_ncache--;	/* skip count byte */
854 
855 			/*
856 			 * Well, for now, rather than attempt to implement
857 			 * this slightly dubious algorithm, we just look
858 			 * at the first 15 ..
859 			 */
860 			if (cpi->cpi_ncache > (sizeof (*cp) - 1))
861 				cpi->cpi_ncache = sizeof (*cp) - 1;
862 
863 			dp = cpi->cpi_cacheinfo;
864 			if (BITX(cp->cp_eax, 31, 31) == 0) {
865 				uint8_t *p = (void *)&cp->cp_eax;
866 				for (i = 1; i < 3; i++)
867 					if (p[i] != 0)
868 						*dp++ = p[i];
869 			}
870 			if (BITX(cp->cp_ebx, 31, 31) == 0) {
871 				uint8_t *p = (void *)&cp->cp_ebx;
872 				for (i = 0; i < 4; i++)
873 					if (p[i] != 0)
874 						*dp++ = p[i];
875 			}
876 			if (BITX(cp->cp_ecx, 31, 31) == 0) {
877 				uint8_t *p = (void *)&cp->cp_ecx;
878 				for (i = 0; i < 4; i++)
879 					if (p[i] != 0)
880 						*dp++ = p[i];
881 			}
882 			if (BITX(cp->cp_edx, 31, 31) == 0) {
883 				uint8_t *p = (void *)&cp->cp_edx;
884 				for (i = 0; i < 4; i++)
885 					if (p[i] != 0)
886 						*dp++ = p[i];
887 			}
888 			break;
889 		case 3:	/* Processor serial number, if PSN supported */
890 		case 4:	/* Deterministic cache parameters */
891 		case 5:	/* Monitor/Mwait parameters */
892 		default:
893 			break;
894 		}
895 	}
896 
897 	if ((cpi->cpi_xmaxeax & 0x80000000) == 0)
898 		goto pass2_done;
899 
900 	if ((nmax = cpi->cpi_xmaxeax - 0x80000000 + 1) > NMAX_CPI_EXTD)
901 		nmax = NMAX_CPI_EXTD;
902 	/*
903 	 * Copy the extended properties, fixing them as we go.
904 	 * (We already handled n == 0 and n == 1 in pass 1)
905 	 */
906 	iptr = (void *)cpi->cpi_brandstr;
907 	for (n = 2, cp = &cpi->cpi_extd[2]; n < nmax; cp++, n++) {
908 		cp->cp_eax = 0x80000000 + n;
909 		(void) __cpuid_insn(cp);
910 		switch (n) {
911 		case 2:
912 		case 3:
913 		case 4:
914 			/*
915 			 * Extract the brand string
916 			 */
917 			*iptr++ = cp->cp_eax;
918 			*iptr++ = cp->cp_ebx;
919 			*iptr++ = cp->cp_ecx;
920 			*iptr++ = cp->cp_edx;
921 			break;
922 		case 5:
923 			switch (cpi->cpi_vendor) {
924 			case X86_VENDOR_AMD:
925 				/*
926 				 * The Athlon and Duron were the first
927 				 * parts to report the sizes of the
928 				 * TLB for large pages. Before then,
929 				 * we don't trust the data.
930 				 */
931 				if (cpi->cpi_family < 6 ||
932 				    (cpi->cpi_family == 6 &&
933 				    cpi->cpi_model < 1))
934 					cp->cp_eax = 0;
935 				break;
936 			default:
937 				break;
938 			}
939 			break;
940 		case 6:
941 			switch (cpi->cpi_vendor) {
942 			case X86_VENDOR_AMD:
943 				/*
944 				 * The Athlon and Duron were the first
945 				 * AMD parts with L2 TLB's.
946 				 * Before then, don't trust the data.
947 				 */
948 				if (cpi->cpi_family < 6 ||
949 				    cpi->cpi_family == 6 &&
950 				    cpi->cpi_model < 1)
951 					cp->cp_eax = cp->cp_ebx = 0;
952 				/*
953 				 * AMD Duron rev A0 reports L2
954 				 * cache size incorrectly as 1K
955 				 * when it is really 64K
956 				 */
957 				if (cpi->cpi_family == 6 &&
958 				    cpi->cpi_model == 3 &&
959 				    cpi->cpi_step == 0) {
960 					cp->cp_ecx &= 0xffff;
961 					cp->cp_ecx |= 0x400000;
962 				}
963 				break;
964 			case X86_VENDOR_Cyrix:	/* VIA C3 */
965 				/*
966 				 * VIA C3 processors are a bit messed
967 				 * up w.r.t. encoding cache sizes in %ecx
968 				 */
969 				if (cpi->cpi_family != 6)
970 					break;
971 				/*
972 				 * model 7 and 8 were incorrectly encoded
973 				 *
974 				 * xxx is model 8 really broken?
975 				 */
976 				if (cpi->cpi_model == 7 ||
977 				    cpi->cpi_model == 8)
978 					cp->cp_ecx =
979 					    BITX(cp->cp_ecx, 31, 24) << 16 |
980 					    BITX(cp->cp_ecx, 23, 16) << 12 |
981 					    BITX(cp->cp_ecx, 15, 8) << 8 |
982 					    BITX(cp->cp_ecx, 7, 0);
983 				/*
984 				 * model 9 stepping 1 has wrong associativity
985 				 */
986 				if (cpi->cpi_model == 9 && cpi->cpi_step == 1)
987 					cp->cp_ecx |= 8 << 12;
988 				break;
989 			case X86_VENDOR_Intel:
990 				/*
991 				 * Extended L2 Cache features function.
992 				 * First appeared on Prescott.
993 				 */
994 			default:
995 				break;
996 			}
997 			break;
998 		default:
999 			break;
1000 		}
1001 	}
1002 
1003 pass2_done:
1004 	cpi->cpi_pass = 2;
1005 }
1006 
1007 static const char *
1008 intel_cpubrand(const struct cpuid_info *cpi)
1009 {
1010 	int i;
1011 
1012 	if ((x86_feature & X86_CPUID) == 0 ||
1013 	    cpi->cpi_maxeax < 1 || cpi->cpi_family < 5)
1014 		return ("i486");
1015 
1016 	switch (cpi->cpi_family) {
1017 	case 5:
1018 		return ("Intel Pentium(r)");
1019 	case 6:
1020 		switch (cpi->cpi_model) {
1021 			uint_t celeron, xeon;
1022 			const struct cpuid_regs *cp;
1023 		case 0:
1024 		case 1:
1025 		case 2:
1026 			return ("Intel Pentium(r) Pro");
1027 		case 3:
1028 		case 4:
1029 			return ("Intel Pentium(r) II");
1030 		case 6:
1031 			return ("Intel Celeron(r)");
1032 		case 5:
1033 		case 7:
1034 			celeron = xeon = 0;
1035 			cp = &cpi->cpi_std[2];	/* cache info */
1036 
1037 			for (i = 1; i < 3; i++) {
1038 				uint_t tmp;
1039 
1040 				tmp = (cp->cp_eax >> (8 * i)) & 0xff;
1041 				if (tmp == 0x40)
1042 					celeron++;
1043 				if (tmp >= 0x44 && tmp <= 0x45)
1044 					xeon++;
1045 			}
1046 
1047 			for (i = 0; i < 2; i++) {
1048 				uint_t tmp;
1049 
1050 				tmp = (cp->cp_ebx >> (8 * i)) & 0xff;
1051 				if (tmp == 0x40)
1052 					celeron++;
1053 				else if (tmp >= 0x44 && tmp <= 0x45)
1054 					xeon++;
1055 			}
1056 
1057 			for (i = 0; i < 4; i++) {
1058 				uint_t tmp;
1059 
1060 				tmp = (cp->cp_ecx >> (8 * i)) & 0xff;
1061 				if (tmp == 0x40)
1062 					celeron++;
1063 				else if (tmp >= 0x44 && tmp <= 0x45)
1064 					xeon++;
1065 			}
1066 
1067 			for (i = 0; i < 4; i++) {
1068 				uint_t tmp;
1069 
1070 				tmp = (cp->cp_edx >> (8 * i)) & 0xff;
1071 				if (tmp == 0x40)
1072 					celeron++;
1073 				else if (tmp >= 0x44 && tmp <= 0x45)
1074 					xeon++;
1075 			}
1076 
1077 			if (celeron)
1078 				return ("Intel Celeron(r)");
1079 			if (xeon)
1080 				return (cpi->cpi_model == 5 ?
1081 				    "Intel Pentium(r) II Xeon(tm)" :
1082 				    "Intel Pentium(r) III Xeon(tm)");
1083 			return (cpi->cpi_model == 5 ?
1084 			    "Intel Pentium(r) II or Pentium(r) II Xeon(tm)" :
1085 			    "Intel Pentium(r) III or Pentium(r) III Xeon(tm)");
1086 		default:
1087 			break;
1088 		}
1089 	default:
1090 		break;
1091 	}
1092 
1093 	/* BrandID is present if the field is nonzero */
1094 	if (cpi->cpi_brandid != 0) {
1095 		static const struct {
1096 			uint_t bt_bid;
1097 			const char *bt_str;
1098 		} brand_tbl[] = {
1099 			{ 0x1,	"Intel(r) Celeron(r)" },
1100 			{ 0x2,	"Intel(r) Pentium(r) III" },
1101 			{ 0x3,	"Intel(r) Pentium(r) III Xeon(tm)" },
1102 			{ 0x4,	"Intel(r) Pentium(r) III" },
1103 			{ 0x6,	"Mobile Intel(r) Pentium(r) III" },
1104 			{ 0x7,	"Mobile Intel(r) Celeron(r)" },
1105 			{ 0x8,	"Intel(r) Pentium(r) 4" },
1106 			{ 0x9,	"Intel(r) Pentium(r) 4" },
1107 			{ 0xa,	"Intel(r) Celeron(r)" },
1108 			{ 0xb,	"Intel(r) Xeon(tm)" },
1109 			{ 0xc,	"Intel(r) Xeon(tm) MP" },
1110 			{ 0xe,	"Mobile Intel(r) Pentium(r) 4" },
1111 			{ 0xf,	"Mobile Intel(r) Celeron(r)" },
1112 			{ 0x11, "Mobile Genuine Intel(r)" },
1113 			{ 0x12, "Intel(r) Celeron(r) M" },
1114 			{ 0x13, "Mobile Intel(r) Celeron(r)" },
1115 			{ 0x14, "Intel(r) Celeron(r)" },
1116 			{ 0x15, "Mobile Genuine Intel(r)" },
1117 			{ 0x16,	"Intel(r) Pentium(r) M" },
1118 			{ 0x17, "Mobile Intel(r) Celeron(r)" }
1119 		};
1120 		uint_t btblmax = sizeof (brand_tbl) / sizeof (brand_tbl[0]);
1121 		uint_t sgn;
1122 
1123 		sgn = (cpi->cpi_family << 8) |
1124 		    (cpi->cpi_model << 4) | cpi->cpi_step;
1125 
1126 		for (i = 0; i < btblmax; i++)
1127 			if (brand_tbl[i].bt_bid == cpi->cpi_brandid)
1128 				break;
1129 		if (i < btblmax) {
1130 			if (sgn == 0x6b1 && cpi->cpi_brandid == 3)
1131 				return ("Intel(r) Celeron(r)");
1132 			if (sgn < 0xf13 && cpi->cpi_brandid == 0xb)
1133 				return ("Intel(r) Xeon(tm) MP");
1134 			if (sgn < 0xf13 && cpi->cpi_brandid == 0xe)
1135 				return ("Intel(r) Xeon(tm)");
1136 			return (brand_tbl[i].bt_str);
1137 		}
1138 	}
1139 
1140 	return (NULL);
1141 }
1142 
1143 static const char *
1144 amd_cpubrand(const struct cpuid_info *cpi)
1145 {
1146 	if ((x86_feature & X86_CPUID) == 0 ||
1147 	    cpi->cpi_maxeax < 1 || cpi->cpi_family < 5)
1148 		return ("i486 compatible");
1149 
1150 	switch (cpi->cpi_family) {
1151 	case 5:
1152 		switch (cpi->cpi_model) {
1153 		case 0:
1154 		case 1:
1155 		case 2:
1156 		case 3:
1157 		case 4:
1158 		case 5:
1159 			return ("AMD-K5(r)");
1160 		case 6:
1161 		case 7:
1162 			return ("AMD-K6(r)");
1163 		case 8:
1164 			return ("AMD-K6(r)-2");
1165 		case 9:
1166 			return ("AMD-K6(r)-III");
1167 		default:
1168 			return ("AMD (family 5)");
1169 		}
1170 	case 6:
1171 		switch (cpi->cpi_model) {
1172 		case 1:
1173 			return ("AMD-K7(tm)");
1174 		case 0:
1175 		case 2:
1176 		case 4:
1177 			return ("AMD Athlon(tm)");
1178 		case 3:
1179 		case 7:
1180 			return ("AMD Duron(tm)");
1181 		case 6:
1182 		case 8:
1183 		case 10:
1184 			/*
1185 			 * Use the L2 cache size to distinguish
1186 			 */
1187 			return ((cpi->cpi_extd[6].cp_ecx >> 16) >= 256 ?
1188 			    "AMD Athlon(tm)" : "AMD Duron(tm)");
1189 		default:
1190 			return ("AMD (family 6)");
1191 		}
1192 	default:
1193 		break;
1194 	}
1195 
1196 	if (cpi->cpi_family == 0xf && cpi->cpi_model == 5 &&
1197 	    cpi->cpi_brandid != 0) {
1198 		switch (BITX(cpi->cpi_brandid, 7, 5)) {
1199 		case 3:
1200 			return ("AMD Opteron(tm) UP 1xx");
1201 		case 4:
1202 			return ("AMD Opteron(tm) DP 2xx");
1203 		case 5:
1204 			return ("AMD Opteron(tm) MP 8xx");
1205 		default:
1206 			return ("AMD Opteron(tm)");
1207 		}
1208 	}
1209 
1210 	return (NULL);
1211 }
1212 
1213 static const char *
1214 cyrix_cpubrand(struct cpuid_info *cpi, uint_t type)
1215 {
1216 	if ((x86_feature & X86_CPUID) == 0 ||
1217 	    cpi->cpi_maxeax < 1 || cpi->cpi_family < 5 ||
1218 	    type == X86_TYPE_CYRIX_486)
1219 		return ("i486 compatible");
1220 
1221 	switch (type) {
1222 	case X86_TYPE_CYRIX_6x86:
1223 		return ("Cyrix 6x86");
1224 	case X86_TYPE_CYRIX_6x86L:
1225 		return ("Cyrix 6x86L");
1226 	case X86_TYPE_CYRIX_6x86MX:
1227 		return ("Cyrix 6x86MX");
1228 	case X86_TYPE_CYRIX_GXm:
1229 		return ("Cyrix GXm");
1230 	case X86_TYPE_CYRIX_MediaGX:
1231 		return ("Cyrix MediaGX");
1232 	case X86_TYPE_CYRIX_MII:
1233 		return ("Cyrix M2");
1234 	case X86_TYPE_VIA_CYRIX_III:
1235 		return ("VIA Cyrix M3");
1236 	default:
1237 		/*
1238 		 * Have another wild guess ..
1239 		 */
1240 		if (cpi->cpi_family == 4 && cpi->cpi_model == 9)
1241 			return ("Cyrix 5x86");
1242 		else if (cpi->cpi_family == 5) {
1243 			switch (cpi->cpi_model) {
1244 			case 2:
1245 				return ("Cyrix 6x86");	/* Cyrix M1 */
1246 			case 4:
1247 				return ("Cyrix MediaGX");
1248 			default:
1249 				break;
1250 			}
1251 		} else if (cpi->cpi_family == 6) {
1252 			switch (cpi->cpi_model) {
1253 			case 0:
1254 				return ("Cyrix 6x86MX"); /* Cyrix M2? */
1255 			case 5:
1256 			case 6:
1257 			case 7:
1258 			case 8:
1259 			case 9:
1260 				return ("VIA C3");
1261 			default:
1262 				break;
1263 			}
1264 		}
1265 		break;
1266 	}
1267 	return (NULL);
1268 }
1269 
1270 /*
1271  * This only gets called in the case that the CPU extended
1272  * feature brand string (0x80000002, 0x80000003, 0x80000004)
1273  * aren't available, or contain null bytes for some reason.
1274  */
1275 static void
1276 fabricate_brandstr(struct cpuid_info *cpi)
1277 {
1278 	const char *brand = NULL;
1279 
1280 	switch (cpi->cpi_vendor) {
1281 	case X86_VENDOR_Intel:
1282 		brand = intel_cpubrand(cpi);
1283 		break;
1284 	case X86_VENDOR_AMD:
1285 		brand = amd_cpubrand(cpi);
1286 		break;
1287 	case X86_VENDOR_Cyrix:
1288 		brand = cyrix_cpubrand(cpi, x86_type);
1289 		break;
1290 	case X86_VENDOR_NexGen:
1291 		if (cpi->cpi_family == 5 && cpi->cpi_model == 0)
1292 			brand = "NexGen Nx586";
1293 		break;
1294 	case X86_VENDOR_Centaur:
1295 		if (cpi->cpi_family == 5)
1296 			switch (cpi->cpi_model) {
1297 			case 4:
1298 				brand = "Centaur C6";
1299 				break;
1300 			case 8:
1301 				brand = "Centaur C2";
1302 				break;
1303 			case 9:
1304 				brand = "Centaur C3";
1305 				break;
1306 			default:
1307 				break;
1308 			}
1309 		break;
1310 	case X86_VENDOR_Rise:
1311 		if (cpi->cpi_family == 5 &&
1312 		    (cpi->cpi_model == 0 || cpi->cpi_model == 2))
1313 			brand = "Rise mP6";
1314 		break;
1315 	case X86_VENDOR_SiS:
1316 		if (cpi->cpi_family == 5 && cpi->cpi_model == 0)
1317 			brand = "SiS 55x";
1318 		break;
1319 	case X86_VENDOR_TM:
1320 		if (cpi->cpi_family == 5 && cpi->cpi_model == 4)
1321 			brand = "Transmeta Crusoe TM3x00 or TM5x00";
1322 		break;
1323 	case X86_VENDOR_NSC:
1324 	case X86_VENDOR_UMC:
1325 	default:
1326 		break;
1327 	}
1328 	if (brand) {
1329 		(void) strcpy((char *)cpi->cpi_brandstr, brand);
1330 		return;
1331 	}
1332 
1333 	/*
1334 	 * If all else fails ...
1335 	 */
1336 	(void) snprintf(cpi->cpi_brandstr, sizeof (cpi->cpi_brandstr),
1337 	    "%s %d.%d.%d", cpi->cpi_vendorstr, cpi->cpi_family,
1338 	    cpi->cpi_model, cpi->cpi_step);
1339 }
1340 
1341 /*
1342  * This routine is called just after kernel memory allocation
1343  * becomes available on cpu0, and as part of mp_startup() on
1344  * the other cpus.
1345  *
1346  * Fixup the brand string.
1347  */
1348 /*ARGSUSED*/
1349 void
1350 cpuid_pass3(cpu_t *cpu)
1351 {
1352 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
1353 
1354 	ASSERT(cpi->cpi_pass == 2);
1355 
1356 	if ((cpi->cpi_xmaxeax & 0x80000000) == 0) {
1357 		fabricate_brandstr(cpi);
1358 		goto pass3_done;
1359 	}
1360 
1361 	/*
1362 	 * If we successfully extracted a brand string from the cpuid
1363 	 * instruction, clean it up by removing leading spaces and
1364 	 * similar junk.
1365 	 */
1366 	if (cpi->cpi_brandstr[0]) {
1367 		size_t maxlen = sizeof (cpi->cpi_brandstr);
1368 		char *src, *dst;
1369 
1370 		dst = src = (char *)cpi->cpi_brandstr;
1371 		src[maxlen - 1] = '\0';
1372 		/*
1373 		 * strip leading spaces
1374 		 */
1375 		while (*src == ' ')
1376 			src++;
1377 		/*
1378 		 * Remove any 'Genuine' or "Authentic" prefixes
1379 		 */
1380 		if (strncmp(src, "Genuine ", 8) == 0)
1381 			src += 8;
1382 		if (strncmp(src, "Authentic ", 10) == 0)
1383 			src += 10;
1384 
1385 		/*
1386 		 * Now do an in-place copy.
1387 		 * Map (R) to (r) and (TM) to (tm).
1388 		 * The era of teletypes is long gone, and there's
1389 		 * -really- no need to shout.
1390 		 */
1391 		while (*src != '\0') {
1392 			if (src[0] == '(') {
1393 				if (strncmp(src + 1, "R)", 2) == 0) {
1394 					(void) strncpy(dst, "(r)", 3);
1395 					src += 3;
1396 					dst += 3;
1397 					continue;
1398 				}
1399 				if (strncmp(src + 1, "TM)", 3) == 0) {
1400 					(void) strncpy(dst, "(tm)", 4);
1401 					src += 4;
1402 					dst += 4;
1403 					continue;
1404 				}
1405 			}
1406 			*dst++ = *src++;
1407 		}
1408 		*dst = '\0';
1409 
1410 		/*
1411 		 * Finally, remove any trailing spaces
1412 		 */
1413 		while (--dst > cpi->cpi_brandstr)
1414 			if (*dst == ' ')
1415 				*dst = '\0';
1416 			else
1417 				break;
1418 	} else
1419 		fabricate_brandstr(cpi);
1420 
1421 pass3_done:
1422 	cpi->cpi_pass = 3;
1423 }
1424 
1425 /*
1426  * This routine is called out of bind_hwcap() much later in the life
1427  * of the kernel (post_startup()).  The job of this routine is to resolve
1428  * the hardware feature support and kernel support for those features into
1429  * what we're actually going to tell applications via the aux vector.
1430  */
1431 uint_t
1432 cpuid_pass4(cpu_t *cpu)
1433 {
1434 	struct cpuid_info *cpi;
1435 	uint_t hwcap_flags = 0;
1436 
1437 	if (cpu == NULL)
1438 		cpu = CPU;
1439 	cpi = cpu->cpu_m.mcpu_cpi;
1440 
1441 	ASSERT(cpi->cpi_pass == 3);
1442 
1443 	if (cpi->cpi_maxeax >= 1) {
1444 		uint32_t *edx = &cpi->cpi_support[STD_EDX_FEATURES];
1445 		uint32_t *ecx = &cpi->cpi_support[STD_ECX_FEATURES];
1446 
1447 		*edx = CPI_FEATURES_EDX(cpi);
1448 		*ecx = CPI_FEATURES_ECX(cpi);
1449 
1450 		/*
1451 		 * [these require explicit kernel support]
1452 		 */
1453 		if ((x86_feature & X86_SEP) == 0)
1454 			*edx &= ~CPUID_INTC_EDX_SEP;
1455 
1456 		if ((x86_feature & X86_SSE) == 0)
1457 			*edx &= ~(CPUID_INTC_EDX_FXSR|CPUID_INTC_EDX_SSE);
1458 		if ((x86_feature & X86_SSE2) == 0)
1459 			*edx &= ~CPUID_INTC_EDX_SSE2;
1460 
1461 		if ((x86_feature & X86_HTT) == 0)
1462 			*edx &= ~CPUID_INTC_EDX_HTT;
1463 
1464 		if ((x86_feature & X86_SSE3) == 0)
1465 			*ecx &= ~CPUID_INTC_ECX_SSE3;
1466 
1467 		/*
1468 		 * [no explicit support required beyond x87 fp context]
1469 		 */
1470 		if (!fpu_exists)
1471 			*edx &= ~(CPUID_INTC_EDX_FPU | CPUID_INTC_EDX_MMX);
1472 
1473 		/*
1474 		 * Now map the supported feature vector to things that we
1475 		 * think userland will care about.
1476 		 */
1477 		if (*edx & CPUID_INTC_EDX_SEP)
1478 			hwcap_flags |= AV_386_SEP;
1479 		if (*edx & CPUID_INTC_EDX_SSE)
1480 			hwcap_flags |= AV_386_FXSR | AV_386_SSE;
1481 		if (*edx & CPUID_INTC_EDX_SSE2)
1482 			hwcap_flags |= AV_386_SSE2;
1483 		if (*ecx & CPUID_INTC_ECX_SSE3)
1484 			hwcap_flags |= AV_386_SSE3;
1485 
1486 		if (*edx & CPUID_INTC_EDX_FPU)
1487 			hwcap_flags |= AV_386_FPU;
1488 		if (*edx & CPUID_INTC_EDX_MMX)
1489 			hwcap_flags |= AV_386_MMX;
1490 
1491 		if (*edx & CPUID_INTC_EDX_TSC)
1492 			hwcap_flags |= AV_386_TSC;
1493 		if (*edx & CPUID_INTC_EDX_CX8)
1494 			hwcap_flags |= AV_386_CX8;
1495 		if (*edx & CPUID_INTC_EDX_CMOV)
1496 			hwcap_flags |= AV_386_CMOV;
1497 		if (*ecx & CPUID_INTC_ECX_MON)
1498 			hwcap_flags |= AV_386_MON;
1499 #if defined(CPUID_INTC_ECX_CX16)
1500 		if (*ecx & CPUID_INTC_ECX_CX16)
1501 			hwcap_flags |= AV_386_CX16;
1502 #endif
1503 	}
1504 
1505 	if (x86_feature & X86_HTT)
1506 		hwcap_flags |= AV_386_PAUSE;
1507 
1508 	if (cpi->cpi_xmaxeax < 0x80000001)
1509 		goto pass4_done;
1510 
1511 	switch (cpi->cpi_vendor) {
1512 		struct cpuid_regs cp;
1513 		uint32_t *edx;
1514 
1515 	case X86_VENDOR_Intel:	/* sigh */
1516 	case X86_VENDOR_AMD:
1517 		edx = &cpi->cpi_support[AMD_EDX_FEATURES];
1518 
1519 		*edx = CPI_FEATURES_XTD_EDX(cpi);
1520 
1521 		/*
1522 		 * [no explicit support required beyond
1523 		 * x87 fp context and exception handlers]
1524 		 */
1525 		if (!fpu_exists)
1526 			*edx &= ~(CPUID_AMD_EDX_MMXamd |
1527 			    CPUID_AMD_EDX_3DNow | CPUID_AMD_EDX_3DNowx);
1528 
1529 		if ((x86_feature & X86_ASYSC) == 0)
1530 			*edx &= ~CPUID_AMD_EDX_SYSC;
1531 		if ((x86_feature & X86_NX) == 0)
1532 			*edx &= ~CPUID_AMD_EDX_NX;
1533 #if !defined(_LP64)
1534 		*edx &= ~CPUID_AMD_EDX_LM;
1535 #endif
1536 		/*
1537 		 * Now map the supported feature vector to
1538 		 * things that we think userland will care about.
1539 		 */
1540 		if (*edx & CPUID_AMD_EDX_SYSC)
1541 			hwcap_flags |= AV_386_AMD_SYSC;
1542 		if (*edx & CPUID_AMD_EDX_MMXamd)
1543 			hwcap_flags |= AV_386_AMD_MMX;
1544 		if (*edx & CPUID_AMD_EDX_3DNow)
1545 			hwcap_flags |= AV_386_AMD_3DNow;
1546 		if (*edx & CPUID_AMD_EDX_3DNowx)
1547 			hwcap_flags |= AV_386_AMD_3DNowx;
1548 		break;
1549 
1550 	case X86_VENDOR_TM:
1551 		cp.cp_eax = 0x80860001;
1552 		(void) __cpuid_insn(&cp);
1553 		cpi->cpi_support[TM_EDX_FEATURES] = cp.cp_edx;
1554 		break;
1555 
1556 	default:
1557 		break;
1558 	}
1559 
1560 pass4_done:
1561 	cpi->cpi_pass = 4;
1562 	return (hwcap_flags);
1563 }
1564 
1565 
1566 /*
1567  * Simulate the cpuid instruction using the data we previously
1568  * captured about this CPU.  We try our best to return the truth
1569  * about the hardware, independently of kernel support.
1570  */
1571 uint32_t
1572 cpuid_insn(cpu_t *cpu, struct cpuid_regs *cp)
1573 {
1574 	struct cpuid_info *cpi;
1575 	struct cpuid_regs *xcp;
1576 
1577 	if (cpu == NULL)
1578 		cpu = CPU;
1579 	cpi = cpu->cpu_m.mcpu_cpi;
1580 
1581 	ASSERT(cpuid_checkpass(cpu, 3));
1582 
1583 	/*
1584 	 * CPUID data is cached in two separate places: cpi_std for standard
1585 	 * CPUID functions, and cpi_extd for extended CPUID functions.
1586 	 */
1587 	if (cp->cp_eax <= cpi->cpi_maxeax && cp->cp_eax < NMAX_CPI_STD)
1588 		xcp = &cpi->cpi_std[cp->cp_eax];
1589 	else if (cp->cp_eax >= 0x80000000 && cp->cp_eax <= cpi->cpi_xmaxeax &&
1590 	    cp->cp_eax < 0x80000000 + NMAX_CPI_EXTD)
1591 		xcp = &cpi->cpi_extd[cp->cp_eax - 0x80000000];
1592 	else
1593 		/*
1594 		 * The caller is asking for data from an input parameter which
1595 		 * the kernel has not cached.  In this case we go fetch from
1596 		 * the hardware and return the data directly to the user.
1597 		 */
1598 		return (__cpuid_insn(cp));
1599 
1600 	cp->cp_eax = xcp->cp_eax;
1601 	cp->cp_ebx = xcp->cp_ebx;
1602 	cp->cp_ecx = xcp->cp_ecx;
1603 	cp->cp_edx = xcp->cp_edx;
1604 	return (cp->cp_eax);
1605 }
1606 
1607 int
1608 cpuid_checkpass(cpu_t *cpu, int pass)
1609 {
1610 	return (cpu != NULL && cpu->cpu_m.mcpu_cpi != NULL &&
1611 	    cpu->cpu_m.mcpu_cpi->cpi_pass >= pass);
1612 }
1613 
1614 int
1615 cpuid_getbrandstr(cpu_t *cpu, char *s, size_t n)
1616 {
1617 	ASSERT(cpuid_checkpass(cpu, 3));
1618 
1619 	return (snprintf(s, n, "%s", cpu->cpu_m.mcpu_cpi->cpi_brandstr));
1620 }
1621 
1622 int
1623 cpuid_is_cmt(cpu_t *cpu)
1624 {
1625 	if (cpu == NULL)
1626 		cpu = CPU;
1627 
1628 	ASSERT(cpuid_checkpass(cpu, 1));
1629 
1630 	return (cpu->cpu_m.mcpu_cpi->cpi_chipid >= 0);
1631 }
1632 
1633 /*
1634  * AMD and Intel both implement the 64-bit variant of the syscall
1635  * instruction (syscallq), so if there's -any- support for syscall,
1636  * cpuid currently says "yes, we support this".
1637  *
1638  * However, Intel decided to -not- implement the 32-bit variant of the
1639  * syscall instruction, so we provide a predicate to allow our caller
1640  * to test that subtlety here.
1641  */
1642 /*ARGSUSED*/
1643 int
1644 cpuid_syscall32_insn(cpu_t *cpu)
1645 {
1646 	ASSERT(cpuid_checkpass((cpu == NULL ? CPU : cpu), 1));
1647 
1648 	if (x86_feature & X86_ASYSC)
1649 		return (x86_vendor != X86_VENDOR_Intel);
1650 	return (0);
1651 }
1652 
1653 int
1654 cpuid_getidstr(cpu_t *cpu, char *s, size_t n)
1655 {
1656 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
1657 
1658 	static const char fmt[] =
1659 	    "x86 (%s family %d model %d step %d clock %d MHz)";
1660 	static const char fmt_ht[] =
1661 	    "x86 (chipid 0x%x %s family %d model %d step %d clock %d MHz)";
1662 
1663 	ASSERT(cpuid_checkpass(cpu, 1));
1664 
1665 	if (cpuid_is_cmt(cpu))
1666 		return (snprintf(s, n, fmt_ht, cpi->cpi_chipid,
1667 		    cpi->cpi_vendorstr, cpi->cpi_family, cpi->cpi_model,
1668 		    cpi->cpi_step, cpu->cpu_type_info.pi_clock));
1669 	return (snprintf(s, n, fmt,
1670 	    cpi->cpi_vendorstr, cpi->cpi_family, cpi->cpi_model,
1671 	    cpi->cpi_step, cpu->cpu_type_info.pi_clock));
1672 }
1673 
1674 const char *
1675 cpuid_getvendorstr(cpu_t *cpu)
1676 {
1677 	ASSERT(cpuid_checkpass(cpu, 1));
1678 	return ((const char *)cpu->cpu_m.mcpu_cpi->cpi_vendorstr);
1679 }
1680 
1681 uint_t
1682 cpuid_getvendor(cpu_t *cpu)
1683 {
1684 	ASSERT(cpuid_checkpass(cpu, 1));
1685 	return (cpu->cpu_m.mcpu_cpi->cpi_vendor);
1686 }
1687 
1688 uint_t
1689 cpuid_getfamily(cpu_t *cpu)
1690 {
1691 	ASSERT(cpuid_checkpass(cpu, 1));
1692 	return (cpu->cpu_m.mcpu_cpi->cpi_family);
1693 }
1694 
1695 uint_t
1696 cpuid_getmodel(cpu_t *cpu)
1697 {
1698 	ASSERT(cpuid_checkpass(cpu, 1));
1699 	return (cpu->cpu_m.mcpu_cpi->cpi_model);
1700 }
1701 
1702 uint_t
1703 cpuid_get_ncpu_per_chip(cpu_t *cpu)
1704 {
1705 	ASSERT(cpuid_checkpass(cpu, 1));
1706 	return (cpu->cpu_m.mcpu_cpi->cpi_ncpu_per_chip);
1707 }
1708 
1709 uint_t
1710 cpuid_get_ncore_per_chip(cpu_t *cpu)
1711 {
1712 	ASSERT(cpuid_checkpass(cpu, 1));
1713 	return (cpu->cpu_m.mcpu_cpi->cpi_ncore_per_chip);
1714 }
1715 
1716 uint_t
1717 cpuid_getstep(cpu_t *cpu)
1718 {
1719 	ASSERT(cpuid_checkpass(cpu, 1));
1720 	return (cpu->cpu_m.mcpu_cpi->cpi_step);
1721 }
1722 
1723 chipid_t
1724 chip_plat_get_chipid(cpu_t *cpu)
1725 {
1726 	ASSERT(cpuid_checkpass(cpu, 1));
1727 
1728 	if (cpuid_is_cmt(cpu))
1729 		return (cpu->cpu_m.mcpu_cpi->cpi_chipid);
1730 	return (cpu->cpu_id);
1731 }
1732 
1733 id_t
1734 chip_plat_get_coreid(cpu_t *cpu)
1735 {
1736 	ASSERT(cpuid_checkpass(cpu, 1));
1737 	return (cpu->cpu_m.mcpu_cpi->cpi_coreid);
1738 }
1739 
1740 int
1741 chip_plat_get_clogid(cpu_t *cpu)
1742 {
1743 	ASSERT(cpuid_checkpass(cpu, 1));
1744 	return (cpu->cpu_m.mcpu_cpi->cpi_clogid);
1745 }
1746 
1747 void
1748 cpuid_get_addrsize(cpu_t *cpu, uint_t *pabits, uint_t *vabits)
1749 {
1750 	struct cpuid_info *cpi;
1751 
1752 	if (cpu == NULL)
1753 		cpu = CPU;
1754 	cpi = cpu->cpu_m.mcpu_cpi;
1755 
1756 	ASSERT(cpuid_checkpass(cpu, 1));
1757 
1758 	if (pabits)
1759 		*pabits = cpi->cpi_pabits;
1760 	if (vabits)
1761 		*vabits = cpi->cpi_vabits;
1762 }
1763 
1764 /*
1765  * Returns the number of data TLB entries for a corresponding
1766  * pagesize.  If it can't be computed, or isn't known, the
1767  * routine returns zero.  If you ask about an architecturally
1768  * impossible pagesize, the routine will panic (so that the
1769  * hat implementor knows that things are inconsistent.)
1770  */
1771 uint_t
1772 cpuid_get_dtlb_nent(cpu_t *cpu, size_t pagesize)
1773 {
1774 	struct cpuid_info *cpi;
1775 	uint_t dtlb_nent = 0;
1776 
1777 	if (cpu == NULL)
1778 		cpu = CPU;
1779 	cpi = cpu->cpu_m.mcpu_cpi;
1780 
1781 	ASSERT(cpuid_checkpass(cpu, 1));
1782 
1783 	/*
1784 	 * Check the L2 TLB info
1785 	 */
1786 	if (cpi->cpi_xmaxeax >= 0x80000006) {
1787 		struct cpuid_regs *cp = &cpi->cpi_extd[6];
1788 
1789 		switch (pagesize) {
1790 
1791 		case 4 * 1024:
1792 			/*
1793 			 * All zero in the top 16 bits of the register
1794 			 * indicates a unified TLB. Size is in low 16 bits.
1795 			 */
1796 			if ((cp->cp_ebx & 0xffff0000) == 0)
1797 				dtlb_nent = cp->cp_ebx & 0x0000ffff;
1798 			else
1799 				dtlb_nent = BITX(cp->cp_ebx, 27, 16);
1800 			break;
1801 
1802 		case 2 * 1024 * 1024:
1803 			if ((cp->cp_eax & 0xffff0000) == 0)
1804 				dtlb_nent = cp->cp_eax & 0x0000ffff;
1805 			else
1806 				dtlb_nent = BITX(cp->cp_eax, 27, 16);
1807 			break;
1808 
1809 		default:
1810 			panic("unknown L2 pagesize");
1811 			/*NOTREACHED*/
1812 		}
1813 	}
1814 
1815 	if (dtlb_nent != 0)
1816 		return (dtlb_nent);
1817 
1818 	/*
1819 	 * No L2 TLB support for this size, try L1.
1820 	 */
1821 	if (cpi->cpi_xmaxeax >= 0x80000005) {
1822 		struct cpuid_regs *cp = &cpi->cpi_extd[5];
1823 
1824 		switch (pagesize) {
1825 		case 4 * 1024:
1826 			dtlb_nent = BITX(cp->cp_ebx, 23, 16);
1827 			break;
1828 		case 2 * 1024 * 1024:
1829 			dtlb_nent = BITX(cp->cp_eax, 23, 16);
1830 			break;
1831 		default:
1832 			panic("unknown L1 d-TLB pagesize");
1833 			/*NOTREACHED*/
1834 		}
1835 	}
1836 
1837 	return (dtlb_nent);
1838 }
1839 
1840 /*
1841  * Return 0 if the erratum is not present or not applicable, positive
1842  * if it is, and negative if the status of the erratum is unknown.
1843  *
1844  * See "Revision Guide for AMD Athlon(tm) 64 and AMD Opteron(tm)
1845  * Processors" #25759, Rev 3.57, August 2005
1846  */
1847 int
1848 cpuid_opteron_erratum(cpu_t *cpu, uint_t erratum)
1849 {
1850 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
1851 	uint_t eax;
1852 
1853 	if (cpi->cpi_vendor != X86_VENDOR_AMD)
1854 		return (0);
1855 
1856 	eax = cpi->cpi_std[1].cp_eax;
1857 
1858 #define	SH_B0(eax)	(eax == 0xf40 || eax == 0xf50)
1859 #define	SH_B3(eax) 	(eax == 0xf51)
1860 #define	B(eax)		(SH_B0(eax) || SH_B3(eax))
1861 
1862 #define	SH_C0(eax)	(eax == 0xf48 || eax == 0xf58)
1863 
1864 #define	SH_CG(eax)	(eax == 0xf4a || eax == 0xf5a || eax == 0xf7a)
1865 #define	DH_CG(eax)	(eax == 0xfc0 || eax == 0xfe0 || eax == 0xff0)
1866 #define	CH_CG(eax)	(eax == 0xf82 || eax == 0xfb2)
1867 #define	CG(eax)		(SH_CG(eax) || DH_CG(eax) || CH_CG(eax))
1868 
1869 #define	SH_D0(eax)	(eax == 0x10f40 || eax == 0x10f50 || eax == 0x10f70)
1870 #define	DH_D0(eax)	(eax == 0x10fc0 || eax == 0x10ff0)
1871 #define	CH_D0(eax)	(eax == 0x10f80 || eax == 0x10fb0)
1872 #define	D0(eax)		(SH_D0(eax) || DH_D0(eax) || CH_D0(eax))
1873 
1874 #define	SH_E0(eax)	(eax == 0x20f50 || eax == 0x20f40 || eax == 0x20f70)
1875 #define	JH_E1(eax)	(eax == 0x20f10)	/* JH8_E0 had 0x20f30 */
1876 #define	DH_E3(eax)	(eax == 0x20fc0 || eax == 0x20ff0)
1877 #define	SH_E4(eax)	(eax == 0x20f51 || eax == 0x20f71)
1878 #define	BH_E4(eax)	(eax == 0x20fb1)
1879 #define	SH_E5(eax)	(eax == 0x20f42)
1880 #define	DH_E6(eax)	(eax == 0x20ff2 || eax == 0x20fc2)
1881 #define	JH_E6(eax)	(eax == 0x20f12 || eax == 0x20f32)
1882 #define	EX(eax)		(SH_E0(eax) || JH_E1(eax) || DH_E3(eax) || \
1883 			    SH_E4(eax) || BH_E4(eax) || SH_E5(eax) || \
1884 			    DH_E6(eax) || JH_E6(eax))
1885 
1886 	switch (erratum) {
1887 	case 1:
1888 		return (1);
1889 	case 51:	/* what does the asterisk mean? */
1890 		return (B(eax) || SH_C0(eax) || CG(eax));
1891 	case 52:
1892 		return (B(eax));
1893 	case 57:
1894 		return (1);
1895 	case 58:
1896 		return (B(eax));
1897 	case 60:
1898 		return (1);
1899 	case 61:
1900 	case 62:
1901 	case 63:
1902 	case 64:
1903 	case 65:
1904 	case 66:
1905 	case 68:
1906 	case 69:
1907 	case 70:
1908 	case 71:
1909 		return (B(eax));
1910 	case 72:
1911 		return (SH_B0(eax));
1912 	case 74:
1913 		return (B(eax));
1914 	case 75:
1915 		return (1);
1916 	case 76:
1917 		return (B(eax));
1918 	case 77:
1919 		return (1);
1920 	case 78:
1921 		return (B(eax) || SH_C0(eax));
1922 	case 79:
1923 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
1924 	case 80:
1925 	case 81:
1926 	case 82:
1927 		return (B(eax));
1928 	case 83:
1929 		return (B(eax) || SH_C0(eax) || CG(eax));
1930 	case 85:
1931 		return (1);
1932 	case 86:
1933 		return (SH_C0(eax) || CG(eax));
1934 	case 88:
1935 #if !defined(__amd64)
1936 		return (0);
1937 #else
1938 		return (B(eax) || SH_C0(eax));
1939 #endif
1940 	case 89:
1941 		return (1);
1942 	case 90:
1943 		return (B(eax) || SH_C0(eax) || CG(eax));
1944 	case 91:
1945 	case 92:
1946 		return (B(eax) || SH_C0(eax));
1947 	case 93:
1948 		return (SH_C0(eax));
1949 	case 94:
1950 		return (B(eax) || SH_C0(eax) || CG(eax));
1951 	case 95:
1952 #if !defined(__amd64)
1953 		return (0);
1954 #else
1955 		return (B(eax) || SH_C0(eax));
1956 #endif
1957 	case 96:
1958 		return (B(eax) || SH_C0(eax) || CG(eax));
1959 	case 97:
1960 	case 98:
1961 		return (SH_C0(eax) || CG(eax));
1962 	case 99:
1963 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
1964 	case 100:
1965 		return (B(eax) || SH_C0(eax));
1966 	case 101:
1967 	case 103:
1968 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
1969 	case 104:
1970 		return (SH_C0(eax) || CG(eax) || D0(eax));
1971 	case 105:
1972 	case 106:
1973 	case 107:
1974 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
1975 	case 108:
1976 		return (DH_CG(eax));
1977 	case 109:
1978 		return (SH_C0(eax) || CG(eax) || D0(eax));
1979 	case 110:
1980 		return (D0(eax) || EX(eax));
1981 	case 111:
1982 		return (CG(eax));
1983 	case 112:
1984 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
1985 	case 113:
1986 		return (eax == 0x20fc0);
1987 	case 114:
1988 		return (SH_E0(eax) || JH_E1(eax) || DH_E3(eax));
1989 	case 115:
1990 		return (SH_E0(eax) || JH_E1(eax));
1991 	case 116:
1992 		return (SH_E0(eax) || JH_E1(eax) || DH_E3(eax));
1993 	case 117:
1994 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
1995 	case 118:
1996 		return (SH_E0(eax) || JH_E1(eax) || SH_E4(eax) || BH_E4(eax) ||
1997 		    JH_E6(eax));
1998 	case 121:
1999 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
2000 	case 122:
2001 		return (SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
2002 	case 123:
2003 		return (JH_E1(eax) || BH_E4(eax) || JH_E6(eax));
2004 	case 131:
2005 		return (1);
2006 	case 6336786:
2007 		/*
2008 		 * Test for AdvPowerMgmtInfo.TscPStateInvariant
2009 		 * if this is a K8 family processor
2010 		 */
2011 		if (CPI_FAMILY(cpi) == 0xf) {
2012 			struct cpuid_regs regs;
2013 			regs.cp_eax = 0x80000007;
2014 			(void) __cpuid_insn(&regs);
2015 			return (!(regs.cp_edx & 0x100));
2016 		}
2017 		return (0);
2018 	case 6323525:
2019 		return (((((eax >> 12) & 0xff00) + (eax & 0xf00)) |
2020 		    (((eax >> 4) & 0xf) | ((eax >> 12) & 0xf0))) < 0xf40);
2021 
2022 	default:
2023 		return (-1);
2024 	}
2025 }
2026 
2027 static const char assoc_str[] = "associativity";
2028 static const char line_str[] = "line-size";
2029 static const char size_str[] = "size";
2030 
2031 static void
2032 add_cache_prop(dev_info_t *devi, const char *label, const char *type,
2033     uint32_t val)
2034 {
2035 	char buf[128];
2036 
2037 	/*
2038 	 * ndi_prop_update_int() is used because it is desirable for
2039 	 * DDI_PROP_HW_DEF and DDI_PROP_DONTSLEEP to be set.
2040 	 */
2041 	if (snprintf(buf, sizeof (buf), "%s-%s", label, type) < sizeof (buf))
2042 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, devi, buf, val);
2043 }
2044 
2045 /*
2046  * Intel-style cache/tlb description
2047  *
2048  * Standard cpuid level 2 gives a randomly ordered
2049  * selection of tags that index into a table that describes
2050  * cache and tlb properties.
2051  */
2052 
2053 static const char l1_icache_str[] = "l1-icache";
2054 static const char l1_dcache_str[] = "l1-dcache";
2055 static const char l2_cache_str[] = "l2-cache";
2056 static const char itlb4k_str[] = "itlb-4K";
2057 static const char dtlb4k_str[] = "dtlb-4K";
2058 static const char itlb4M_str[] = "itlb-4M";
2059 static const char dtlb4M_str[] = "dtlb-4M";
2060 static const char itlb424_str[] = "itlb-4K-2M-4M";
2061 static const char dtlb44_str[] = "dtlb-4K-4M";
2062 static const char sl1_dcache_str[] = "sectored-l1-dcache";
2063 static const char sl2_cache_str[] = "sectored-l2-cache";
2064 static const char itrace_str[] = "itrace-cache";
2065 static const char sl3_cache_str[] = "sectored-l3-cache";
2066 
2067 static const struct cachetab {
2068 	uint8_t 	ct_code;
2069 	uint8_t		ct_assoc;
2070 	uint16_t 	ct_line_size;
2071 	size_t		ct_size;
2072 	const char	*ct_label;
2073 } intel_ctab[] = {
2074 	/* maintain descending order! */
2075 	{ 0xb3, 4, 0, 128, dtlb4k_str },
2076 	{ 0xb0, 4, 0, 128, itlb4k_str },
2077 	{ 0x87, 8, 64, 1024*1024, l2_cache_str},
2078 	{ 0x86, 4, 64, 512*1024, l2_cache_str},
2079 	{ 0x85, 8, 32, 2*1024*1024, l2_cache_str},
2080 	{ 0x84, 8, 32, 1024*1024, l2_cache_str},
2081 	{ 0x83, 8, 32, 512*1024, l2_cache_str},
2082 	{ 0x82, 8, 32, 256*1024, l2_cache_str},
2083 	{ 0x81, 8, 32, 128*1024, l2_cache_str},		/* suspect! */
2084 	{ 0x7f, 2, 64, 512*1024, l2_cache_str},
2085 	{ 0x7d, 8, 64, 2*1024*1024, sl2_cache_str},
2086 	{ 0x7c, 8, 64, 1024*1024, sl2_cache_str},
2087 	{ 0x7b, 8, 64, 512*1024, sl2_cache_str},
2088 	{ 0x7a, 8, 64, 256*1024, sl2_cache_str},
2089 	{ 0x79, 8, 64, 128*1024, sl2_cache_str},
2090 	{ 0x78, 8, 64, 1024*1024, l2_cache_str},
2091 	{ 0x72, 8, 0, 32*1024, itrace_str},
2092 	{ 0x71, 8, 0, 16*1024, itrace_str},
2093 	{ 0x70, 8, 0, 12*1024, itrace_str},
2094 	{ 0x68, 4, 64, 32*1024, sl1_dcache_str},
2095 	{ 0x67, 4, 64, 16*1024, sl1_dcache_str},
2096 	{ 0x66, 4, 64, 8*1024, sl1_dcache_str},
2097 	{ 0x60, 8, 64, 16*1024, sl1_dcache_str},
2098 	{ 0x5d, 0, 0, 256, dtlb44_str},
2099 	{ 0x5c, 0, 0, 128, dtlb44_str},
2100 	{ 0x5b, 0, 0, 64, dtlb44_str},
2101 	{ 0x52, 0, 0, 256, itlb424_str},
2102 	{ 0x51, 0, 0, 128, itlb424_str},
2103 	{ 0x50, 0, 0, 64, itlb424_str},
2104 	{ 0x45, 4, 32, 2*1024*1024, l2_cache_str},
2105 	{ 0x44, 4, 32, 1024*1024, l2_cache_str},
2106 	{ 0x43, 4, 32, 512*1024, l2_cache_str},
2107 	{ 0x42, 4, 32, 256*1024, l2_cache_str},
2108 	{ 0x41, 4, 32, 128*1024, l2_cache_str},
2109 	{ 0x3c, 4, 64, 256*1024, sl2_cache_str},
2110 	{ 0x3b, 2, 64, 128*1024, sl2_cache_str},
2111 	{ 0x39, 4, 64, 128*1024, sl2_cache_str},
2112 	{ 0x30, 8, 64, 32*1024, l1_icache_str},
2113 	{ 0x2c, 8, 64, 32*1024, l1_dcache_str},
2114 	{ 0x29, 8, 64, 4096*1024, sl3_cache_str},
2115 	{ 0x25, 8, 64, 2048*1024, sl3_cache_str},
2116 	{ 0x23, 8, 64, 1024*1024, sl3_cache_str},
2117 	{ 0x22, 4, 64, 512*1024, sl3_cache_str},
2118 	{ 0x0c, 4, 32, 16*1024, l1_dcache_str},
2119 	{ 0x0a, 2, 32, 8*1024, l1_dcache_str},
2120 	{ 0x08, 4, 32, 16*1024, l1_icache_str},
2121 	{ 0x06, 4, 32, 8*1024, l1_icache_str},
2122 	{ 0x04, 4, 0, 8, dtlb4M_str},
2123 	{ 0x03, 4, 0, 64, dtlb4k_str},
2124 	{ 0x02, 4, 0, 2, itlb4M_str},
2125 	{ 0x01, 4, 0, 32, itlb4k_str},
2126 	{ 0 }
2127 };
2128 
2129 static const struct cachetab cyrix_ctab[] = {
2130 	{ 0x70, 4, 0, 32, "tlb-4K" },
2131 	{ 0x80, 4, 16, 16*1024, "l1-cache" },
2132 	{ 0 }
2133 };
2134 
2135 /*
2136  * Search a cache table for a matching entry
2137  */
2138 static const struct cachetab *
2139 find_cacheent(const struct cachetab *ct, uint_t code)
2140 {
2141 	if (code != 0) {
2142 		for (; ct->ct_code != 0; ct++)
2143 			if (ct->ct_code <= code)
2144 				break;
2145 		if (ct->ct_code == code)
2146 			return (ct);
2147 	}
2148 	return (NULL);
2149 }
2150 
2151 /*
2152  * Walk the cacheinfo descriptor, applying 'func' to every valid element
2153  * The walk is terminated if the walker returns non-zero.
2154  */
2155 static void
2156 intel_walk_cacheinfo(struct cpuid_info *cpi,
2157     void *arg, int (*func)(void *, const struct cachetab *))
2158 {
2159 	const struct cachetab *ct;
2160 	uint8_t *dp;
2161 	int i;
2162 
2163 	if ((dp = cpi->cpi_cacheinfo) == NULL)
2164 		return;
2165 	for (i = 0; i < cpi->cpi_ncache; i++, dp++)
2166 		if ((ct = find_cacheent(intel_ctab, *dp)) != NULL) {
2167 			if (func(arg, ct) != 0)
2168 				break;
2169 		}
2170 }
2171 
2172 /*
2173  * (Like the Intel one, except for Cyrix CPUs)
2174  */
2175 static void
2176 cyrix_walk_cacheinfo(struct cpuid_info *cpi,
2177     void *arg, int (*func)(void *, const struct cachetab *))
2178 {
2179 	const struct cachetab *ct;
2180 	uint8_t *dp;
2181 	int i;
2182 
2183 	if ((dp = cpi->cpi_cacheinfo) == NULL)
2184 		return;
2185 	for (i = 0; i < cpi->cpi_ncache; i++, dp++) {
2186 		/*
2187 		 * Search Cyrix-specific descriptor table first ..
2188 		 */
2189 		if ((ct = find_cacheent(cyrix_ctab, *dp)) != NULL) {
2190 			if (func(arg, ct) != 0)
2191 				break;
2192 			continue;
2193 		}
2194 		/*
2195 		 * .. else fall back to the Intel one
2196 		 */
2197 		if ((ct = find_cacheent(intel_ctab, *dp)) != NULL) {
2198 			if (func(arg, ct) != 0)
2199 				break;
2200 			continue;
2201 		}
2202 	}
2203 }
2204 
2205 /*
2206  * A cacheinfo walker that adds associativity, line-size, and size properties
2207  * to the devinfo node it is passed as an argument.
2208  */
2209 static int
2210 add_cacheent_props(void *arg, const struct cachetab *ct)
2211 {
2212 	dev_info_t *devi = arg;
2213 
2214 	add_cache_prop(devi, ct->ct_label, assoc_str, ct->ct_assoc);
2215 	if (ct->ct_line_size != 0)
2216 		add_cache_prop(devi, ct->ct_label, line_str,
2217 		    ct->ct_line_size);
2218 	add_cache_prop(devi, ct->ct_label, size_str, ct->ct_size);
2219 	return (0);
2220 }
2221 
2222 static const char fully_assoc[] = "fully-associative?";
2223 
2224 /*
2225  * AMD style cache/tlb description
2226  *
2227  * Extended functions 5 and 6 directly describe properties of
2228  * tlbs and various cache levels.
2229  */
2230 static void
2231 add_amd_assoc(dev_info_t *devi, const char *label, uint_t assoc)
2232 {
2233 	switch (assoc) {
2234 	case 0:	/* reserved; ignore */
2235 		break;
2236 	default:
2237 		add_cache_prop(devi, label, assoc_str, assoc);
2238 		break;
2239 	case 0xff:
2240 		add_cache_prop(devi, label, fully_assoc, 1);
2241 		break;
2242 	}
2243 }
2244 
2245 static void
2246 add_amd_tlb(dev_info_t *devi, const char *label, uint_t assoc, uint_t size)
2247 {
2248 	if (size == 0)
2249 		return;
2250 	add_cache_prop(devi, label, size_str, size);
2251 	add_amd_assoc(devi, label, assoc);
2252 }
2253 
2254 static void
2255 add_amd_cache(dev_info_t *devi, const char *label,
2256     uint_t size, uint_t assoc, uint_t lines_per_tag, uint_t line_size)
2257 {
2258 	if (size == 0 || line_size == 0)
2259 		return;
2260 	add_amd_assoc(devi, label, assoc);
2261 	/*
2262 	 * Most AMD parts have a sectored cache. Multiple cache lines are
2263 	 * associated with each tag. A sector consists of all cache lines
2264 	 * associated with a tag. For example, the AMD K6-III has a sector
2265 	 * size of 2 cache lines per tag.
2266 	 */
2267 	if (lines_per_tag != 0)
2268 		add_cache_prop(devi, label, "lines-per-tag", lines_per_tag);
2269 	add_cache_prop(devi, label, line_str, line_size);
2270 	add_cache_prop(devi, label, size_str, size * 1024);
2271 }
2272 
2273 static void
2274 add_amd_l2_assoc(dev_info_t *devi, const char *label, uint_t assoc)
2275 {
2276 	switch (assoc) {
2277 	case 0:	/* off */
2278 		break;
2279 	case 1:
2280 	case 2:
2281 	case 4:
2282 		add_cache_prop(devi, label, assoc_str, assoc);
2283 		break;
2284 	case 6:
2285 		add_cache_prop(devi, label, assoc_str, 8);
2286 		break;
2287 	case 8:
2288 		add_cache_prop(devi, label, assoc_str, 16);
2289 		break;
2290 	case 0xf:
2291 		add_cache_prop(devi, label, fully_assoc, 1);
2292 		break;
2293 	default: /* reserved; ignore */
2294 		break;
2295 	}
2296 }
2297 
2298 static void
2299 add_amd_l2_tlb(dev_info_t *devi, const char *label, uint_t assoc, uint_t size)
2300 {
2301 	if (size == 0 || assoc == 0)
2302 		return;
2303 	add_amd_l2_assoc(devi, label, assoc);
2304 	add_cache_prop(devi, label, size_str, size);
2305 }
2306 
2307 static void
2308 add_amd_l2_cache(dev_info_t *devi, const char *label,
2309     uint_t size, uint_t assoc, uint_t lines_per_tag, uint_t line_size)
2310 {
2311 	if (size == 0 || assoc == 0 || line_size == 0)
2312 		return;
2313 	add_amd_l2_assoc(devi, label, assoc);
2314 	if (lines_per_tag != 0)
2315 		add_cache_prop(devi, label, "lines-per-tag", lines_per_tag);
2316 	add_cache_prop(devi, label, line_str, line_size);
2317 	add_cache_prop(devi, label, size_str, size * 1024);
2318 }
2319 
2320 static void
2321 amd_cache_info(struct cpuid_info *cpi, dev_info_t *devi)
2322 {
2323 	struct cpuid_regs *cp;
2324 
2325 	if (cpi->cpi_xmaxeax < 0x80000005)
2326 		return;
2327 	cp = &cpi->cpi_extd[5];
2328 
2329 	/*
2330 	 * 4M/2M L1 TLB configuration
2331 	 *
2332 	 * We report the size for 2M pages because AMD uses two
2333 	 * TLB entries for one 4M page.
2334 	 */
2335 	add_amd_tlb(devi, "dtlb-2M",
2336 	    BITX(cp->cp_eax, 31, 24), BITX(cp->cp_eax, 23, 16));
2337 	add_amd_tlb(devi, "itlb-2M",
2338 	    BITX(cp->cp_eax, 15, 8), BITX(cp->cp_eax, 7, 0));
2339 
2340 	/*
2341 	 * 4K L1 TLB configuration
2342 	 */
2343 
2344 	switch (cpi->cpi_vendor) {
2345 		uint_t nentries;
2346 	case X86_VENDOR_TM:
2347 		if (cpi->cpi_family >= 5) {
2348 			/*
2349 			 * Crusoe processors have 256 TLB entries, but
2350 			 * cpuid data format constrains them to only
2351 			 * reporting 255 of them.
2352 			 */
2353 			if ((nentries = BITX(cp->cp_ebx, 23, 16)) == 255)
2354 				nentries = 256;
2355 			/*
2356 			 * Crusoe processors also have a unified TLB
2357 			 */
2358 			add_amd_tlb(devi, "tlb-4K", BITX(cp->cp_ebx, 31, 24),
2359 			    nentries);
2360 			break;
2361 		}
2362 		/*FALLTHROUGH*/
2363 	default:
2364 		add_amd_tlb(devi, itlb4k_str,
2365 		    BITX(cp->cp_ebx, 31, 24), BITX(cp->cp_ebx, 23, 16));
2366 		add_amd_tlb(devi, dtlb4k_str,
2367 		    BITX(cp->cp_ebx, 15, 8), BITX(cp->cp_ebx, 7, 0));
2368 		break;
2369 	}
2370 
2371 	/*
2372 	 * data L1 cache configuration
2373 	 */
2374 
2375 	add_amd_cache(devi, l1_dcache_str,
2376 	    BITX(cp->cp_ecx, 31, 24), BITX(cp->cp_ecx, 23, 16),
2377 	    BITX(cp->cp_ecx, 15, 8), BITX(cp->cp_ecx, 7, 0));
2378 
2379 	/*
2380 	 * code L1 cache configuration
2381 	 */
2382 
2383 	add_amd_cache(devi, l1_icache_str,
2384 	    BITX(cp->cp_edx, 31, 24), BITX(cp->cp_edx, 23, 16),
2385 	    BITX(cp->cp_edx, 15, 8), BITX(cp->cp_edx, 7, 0));
2386 
2387 	if (cpi->cpi_xmaxeax < 0x80000006)
2388 		return;
2389 	cp = &cpi->cpi_extd[6];
2390 
2391 	/* Check for a unified L2 TLB for large pages */
2392 
2393 	if (BITX(cp->cp_eax, 31, 16) == 0)
2394 		add_amd_l2_tlb(devi, "l2-tlb-2M",
2395 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
2396 	else {
2397 		add_amd_l2_tlb(devi, "l2-dtlb-2M",
2398 		    BITX(cp->cp_eax, 31, 28), BITX(cp->cp_eax, 27, 16));
2399 		add_amd_l2_tlb(devi, "l2-itlb-2M",
2400 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
2401 	}
2402 
2403 	/* Check for a unified L2 TLB for 4K pages */
2404 
2405 	if (BITX(cp->cp_ebx, 31, 16) == 0) {
2406 		add_amd_l2_tlb(devi, "l2-tlb-4K",
2407 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
2408 	} else {
2409 		add_amd_l2_tlb(devi, "l2-dtlb-4K",
2410 		    BITX(cp->cp_eax, 31, 28), BITX(cp->cp_eax, 27, 16));
2411 		add_amd_l2_tlb(devi, "l2-itlb-4K",
2412 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
2413 	}
2414 
2415 	add_amd_l2_cache(devi, l2_cache_str,
2416 	    BITX(cp->cp_ecx, 31, 16), BITX(cp->cp_ecx, 15, 12),
2417 	    BITX(cp->cp_ecx, 11, 8), BITX(cp->cp_ecx, 7, 0));
2418 }
2419 
2420 /*
2421  * There are two basic ways that the x86 world describes it cache
2422  * and tlb architecture - Intel's way and AMD's way.
2423  *
2424  * Return which flavor of cache architecture we should use
2425  */
2426 static int
2427 x86_which_cacheinfo(struct cpuid_info *cpi)
2428 {
2429 	switch (cpi->cpi_vendor) {
2430 	case X86_VENDOR_Intel:
2431 		if (cpi->cpi_maxeax >= 2)
2432 			return (X86_VENDOR_Intel);
2433 		break;
2434 	case X86_VENDOR_AMD:
2435 		/*
2436 		 * The K5 model 1 was the first part from AMD that reported
2437 		 * cache sizes via extended cpuid functions.
2438 		 */
2439 		if (cpi->cpi_family > 5 ||
2440 		    (cpi->cpi_family == 5 && cpi->cpi_model >= 1))
2441 			return (X86_VENDOR_AMD);
2442 		break;
2443 	case X86_VENDOR_TM:
2444 		if (cpi->cpi_family >= 5)
2445 			return (X86_VENDOR_AMD);
2446 		/*FALLTHROUGH*/
2447 	default:
2448 		/*
2449 		 * If they have extended CPU data for 0x80000005
2450 		 * then we assume they have AMD-format cache
2451 		 * information.
2452 		 *
2453 		 * If not, and the vendor happens to be Cyrix,
2454 		 * then try our-Cyrix specific handler.
2455 		 *
2456 		 * If we're not Cyrix, then assume we're using Intel's
2457 		 * table-driven format instead.
2458 		 */
2459 		if (cpi->cpi_xmaxeax >= 0x80000005)
2460 			return (X86_VENDOR_AMD);
2461 		else if (cpi->cpi_vendor == X86_VENDOR_Cyrix)
2462 			return (X86_VENDOR_Cyrix);
2463 		else if (cpi->cpi_maxeax >= 2)
2464 			return (X86_VENDOR_Intel);
2465 		break;
2466 	}
2467 	return (-1);
2468 }
2469 
2470 /*
2471  * create a node for the given cpu under the prom root node.
2472  * Also, create a cpu node in the device tree.
2473  */
2474 static dev_info_t *cpu_nex_devi = NULL;
2475 static kmutex_t cpu_node_lock;
2476 
2477 /*
2478  * Called from post_startup() and mp_startup()
2479  */
2480 void
2481 add_cpunode2devtree(processorid_t cpu_id, struct cpuid_info *cpi)
2482 {
2483 	dev_info_t *cpu_devi;
2484 	int create;
2485 
2486 	mutex_enter(&cpu_node_lock);
2487 
2488 	/*
2489 	 * create a nexus node for all cpus identified as 'cpu_id' under
2490 	 * the root node.
2491 	 */
2492 	if (cpu_nex_devi == NULL) {
2493 		if (ndi_devi_alloc(ddi_root_node(), "cpus",
2494 		    (pnode_t)DEVI_SID_NODEID, &cpu_nex_devi) != NDI_SUCCESS) {
2495 			mutex_exit(&cpu_node_lock);
2496 			return;
2497 		}
2498 		(void) ndi_devi_online(cpu_nex_devi, 0);
2499 	}
2500 
2501 	/*
2502 	 * create a child node for cpu identified as 'cpu_id'
2503 	 */
2504 	cpu_devi = ddi_add_child(cpu_nex_devi, "cpu", DEVI_SID_NODEID,
2505 		cpu_id);
2506 	if (cpu_devi == NULL) {
2507 		mutex_exit(&cpu_node_lock);
2508 		return;
2509 	}
2510 
2511 	/* device_type */
2512 
2513 	(void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi,
2514 	    "device_type", "cpu");
2515 
2516 	/* reg */
2517 
2518 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2519 	    "reg", cpu_id);
2520 
2521 	/* cpu-mhz, and clock-frequency */
2522 
2523 	if (cpu_freq > 0) {
2524 		long long mul;
2525 
2526 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2527 		    "cpu-mhz", cpu_freq);
2528 
2529 		if ((mul = cpu_freq * 1000000LL) <= INT_MAX)
2530 			(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2531 			    "clock-frequency", (int)mul);
2532 	}
2533 
2534 	(void) ndi_devi_online(cpu_devi, 0);
2535 
2536 	if ((x86_feature & X86_CPUID) == 0) {
2537 		mutex_exit(&cpu_node_lock);
2538 		return;
2539 	}
2540 
2541 	/* vendor-id */
2542 
2543 	(void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi,
2544 		"vendor-id", cpi->cpi_vendorstr);
2545 
2546 	if (cpi->cpi_maxeax == 0) {
2547 		mutex_exit(&cpu_node_lock);
2548 		return;
2549 	}
2550 
2551 	/*
2552 	 * family, model, and step
2553 	 */
2554 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2555 		"family", CPI_FAMILY(cpi));
2556 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2557 		"cpu-model", CPI_MODEL(cpi));
2558 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2559 		"stepping-id", CPI_STEP(cpi));
2560 
2561 	/* type */
2562 
2563 	switch (cpi->cpi_vendor) {
2564 	case X86_VENDOR_Intel:
2565 		create = 1;
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 			"type", CPI_TYPE(cpi));
2574 
2575 	/* ext-family */
2576 
2577 	switch (cpi->cpi_vendor) {
2578 	case X86_VENDOR_Intel:
2579 	case X86_VENDOR_AMD:
2580 		create = cpi->cpi_family >= 0xf;
2581 		break;
2582 	default:
2583 		create = 0;
2584 		break;
2585 	}
2586 	if (create)
2587 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2588 		    "ext-family", CPI_FAMILY_XTD(cpi));
2589 
2590 	/* ext-model */
2591 
2592 	switch (cpi->cpi_vendor) {
2593 	case X86_VENDOR_Intel:
2594 	case X86_VENDOR_AMD:
2595 		create = CPI_FAMILY(cpi) == 0xf;
2596 		break;
2597 	default:
2598 		create = 0;
2599 		break;
2600 	}
2601 	if (create)
2602 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2603 			"ext-model", CPI_MODEL_XTD(cpi));
2604 
2605 	/* generation */
2606 
2607 	switch (cpi->cpi_vendor) {
2608 	case X86_VENDOR_AMD:
2609 		/*
2610 		 * AMD K5 model 1 was the first part to support this
2611 		 */
2612 		create = cpi->cpi_xmaxeax >= 0x80000001;
2613 		break;
2614 	default:
2615 		create = 0;
2616 		break;
2617 	}
2618 	if (create)
2619 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2620 		    "generation", BITX((cpi)->cpi_extd[1].cp_eax, 11, 8));
2621 
2622 	/* brand-id */
2623 
2624 	switch (cpi->cpi_vendor) {
2625 	case X86_VENDOR_Intel:
2626 		/*
2627 		 * brand id first appeared on Pentium III Xeon model 8,
2628 		 * and Celeron model 8 processors and Opteron
2629 		 */
2630 		create = cpi->cpi_family > 6 ||
2631 		    (cpi->cpi_family == 6 && cpi->cpi_model >= 8);
2632 		break;
2633 	case X86_VENDOR_AMD:
2634 		create = cpi->cpi_family >= 0xf;
2635 		break;
2636 	default:
2637 		create = 0;
2638 		break;
2639 	}
2640 	if (create && cpi->cpi_brandid != 0) {
2641 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2642 		    "brand-id", cpi->cpi_brandid);
2643 	}
2644 
2645 	/* chunks, and apic-id */
2646 
2647 	switch (cpi->cpi_vendor) {
2648 		/*
2649 		 * first available on Pentium IV and Opteron (K8)
2650 		 */
2651 	case X86_VENDOR_Intel:
2652 		create = IS_NEW_F6(cpi) || cpi->cpi_family >= 0xf;
2653 		break;
2654 	case X86_VENDOR_AMD:
2655 		create = cpi->cpi_family >= 0xf;
2656 		break;
2657 	default:
2658 		create = 0;
2659 		break;
2660 	}
2661 	if (create) {
2662 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2663 			"chunks", CPI_CHUNKS(cpi));
2664 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2665 			"apic-id", CPI_APIC_ID(cpi));
2666 		if (cpi->cpi_chipid >= 0) {
2667 			(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2668 			    "chip#", cpi->cpi_chipid);
2669 			(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2670 			    "clog#", cpi->cpi_clogid);
2671 		}
2672 	}
2673 
2674 	/* cpuid-features */
2675 
2676 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2677 	    "cpuid-features", CPI_FEATURES_EDX(cpi));
2678 
2679 
2680 	/* cpuid-features-ecx */
2681 
2682 	switch (cpi->cpi_vendor) {
2683 	case X86_VENDOR_Intel:
2684 		create = IS_NEW_F6(cpi) || cpi->cpi_family >= 0xf;
2685 		break;
2686 	default:
2687 		create = 0;
2688 		break;
2689 	}
2690 	if (create)
2691 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2692 		    "cpuid-features-ecx", CPI_FEATURES_ECX(cpi));
2693 
2694 	/* ext-cpuid-features */
2695 
2696 	switch (cpi->cpi_vendor) {
2697 	case X86_VENDOR_Intel:
2698 	case X86_VENDOR_AMD:
2699 	case X86_VENDOR_Cyrix:
2700 	case X86_VENDOR_TM:
2701 	case X86_VENDOR_Centaur:
2702 		create = cpi->cpi_xmaxeax >= 0x80000001;
2703 		break;
2704 	default:
2705 		create = 0;
2706 		break;
2707 	}
2708 	if (create) {
2709 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2710 			"ext-cpuid-features", CPI_FEATURES_XTD_EDX(cpi));
2711 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
2712 			"ext-cpuid-features-ecx", CPI_FEATURES_XTD_ECX(cpi));
2713 	}
2714 
2715 	/*
2716 	 * Brand String first appeared in Intel Pentium IV, AMD K5
2717 	 * model 1, and Cyrix GXm.  On earlier models we try and
2718 	 * simulate something similar .. so this string should always
2719 	 * same -something- about the processor, however lame.
2720 	 */
2721 	(void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi,
2722 	    "brand-string", cpi->cpi_brandstr);
2723 
2724 	/*
2725 	 * Finally, cache and tlb information
2726 	 */
2727 	switch (x86_which_cacheinfo(cpi)) {
2728 	case X86_VENDOR_Intel:
2729 		intel_walk_cacheinfo(cpi, cpu_devi, add_cacheent_props);
2730 		break;
2731 	case X86_VENDOR_Cyrix:
2732 		cyrix_walk_cacheinfo(cpi, cpu_devi, add_cacheent_props);
2733 		break;
2734 	case X86_VENDOR_AMD:
2735 		amd_cache_info(cpi, cpu_devi);
2736 		break;
2737 	default:
2738 		break;
2739 	}
2740 
2741 	mutex_exit(&cpu_node_lock);
2742 }
2743 
2744 struct l2info {
2745 	int *l2i_csz;
2746 	int *l2i_lsz;
2747 	int *l2i_assoc;
2748 	int l2i_ret;
2749 };
2750 
2751 /*
2752  * A cacheinfo walker that fetches the size, line-size and associativity
2753  * of the L2 cache
2754  */
2755 static int
2756 intel_l2cinfo(void *arg, const struct cachetab *ct)
2757 {
2758 	struct l2info *l2i = arg;
2759 	int *ip;
2760 
2761 	if (ct->ct_label != l2_cache_str &&
2762 	    ct->ct_label != sl2_cache_str)
2763 		return (0);	/* not an L2 -- keep walking */
2764 
2765 	if ((ip = l2i->l2i_csz) != NULL)
2766 		*ip = ct->ct_size;
2767 	if ((ip = l2i->l2i_lsz) != NULL)
2768 		*ip = ct->ct_line_size;
2769 	if ((ip = l2i->l2i_assoc) != NULL)
2770 		*ip = ct->ct_assoc;
2771 	l2i->l2i_ret = ct->ct_size;
2772 	return (1);		/* was an L2 -- terminate walk */
2773 }
2774 
2775 static void
2776 amd_l2cacheinfo(struct cpuid_info *cpi, struct l2info *l2i)
2777 {
2778 	struct cpuid_regs *cp;
2779 	uint_t size, assoc;
2780 	int *ip;
2781 
2782 	if (cpi->cpi_xmaxeax < 0x80000006)
2783 		return;
2784 	cp = &cpi->cpi_extd[6];
2785 
2786 	if ((assoc = BITX(cp->cp_ecx, 15, 12)) != 0 &&
2787 	    (size = BITX(cp->cp_ecx, 31, 16)) != 0) {
2788 		uint_t cachesz = size * 1024;
2789 
2790 
2791 		if ((ip = l2i->l2i_csz) != NULL)
2792 			*ip = cachesz;
2793 		if ((ip = l2i->l2i_lsz) != NULL)
2794 			*ip = BITX(cp->cp_ecx, 7, 0);
2795 		if ((ip = l2i->l2i_assoc) != NULL)
2796 			*ip = assoc;
2797 		l2i->l2i_ret = cachesz;
2798 	}
2799 }
2800 
2801 int
2802 getl2cacheinfo(cpu_t *cpu, int *csz, int *lsz, int *assoc)
2803 {
2804 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2805 	struct l2info __l2info, *l2i = &__l2info;
2806 
2807 	l2i->l2i_csz = csz;
2808 	l2i->l2i_lsz = lsz;
2809 	l2i->l2i_assoc = assoc;
2810 	l2i->l2i_ret = -1;
2811 
2812 	switch (x86_which_cacheinfo(cpi)) {
2813 	case X86_VENDOR_Intel:
2814 		intel_walk_cacheinfo(cpi, l2i, intel_l2cinfo);
2815 		break;
2816 	case X86_VENDOR_Cyrix:
2817 		cyrix_walk_cacheinfo(cpi, l2i, intel_l2cinfo);
2818 		break;
2819 	case X86_VENDOR_AMD:
2820 		amd_l2cacheinfo(cpi, l2i);
2821 		break;
2822 	default:
2823 		break;
2824 	}
2825 	return (l2i->l2i_ret);
2826 }
2827