xref: /titanic_51/usr/src/uts/i86pc/os/cpuid.c (revision cee1ddad9783a59a09fe42ce3e5f4d3328e5a111)
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 2007 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/sysmacros.h>
44 #include <sys/pg.h>
45 #include <sys/fp.h>
46 #include <sys/controlregs.h>
47 #include <sys/auxv_386.h>
48 #include <sys/bitmap.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 uint_t pentiumpro_bug4046376;
108 uint_t pentiumpro_bug4064495;
109 
110 uint_t enable486;
111 
112 /*
113  * This set of strings are for processors rumored to support the cpuid
114  * instruction, and is used by locore.s to figure out how to set x86_vendor
115  */
116 const char CyrixInstead[] = "CyrixInstead";
117 
118 /*
119  * monitor/mwait info.
120  *
121  * size_actual and buf_actual are the real address and size allocated to get
122  * proper mwait_buf alignement.  buf_actual and size_actual should be passed
123  * to kmem_free().  Currently kmem_alloc() and mwait happen to both use
124  * processor cache-line alignment, but this is not guarantied in the furture.
125  */
126 struct mwait_info {
127 	size_t		mon_min;	/* min size to avoid missed wakeups */
128 	size_t		mon_max;	/* size to avoid false wakeups */
129 	size_t		size_actual;	/* size actually allocated */
130 	void		*buf_actual;	/* memory actually allocated */
131 	uint32_t	support;	/* processor support of monitor/mwait */
132 };
133 
134 /*
135  * These constants determine how many of the elements of the
136  * cpuid we cache in the cpuid_info data structure; the
137  * remaining elements are accessible via the cpuid instruction.
138  */
139 
140 #define	NMAX_CPI_STD	6		/* eax = 0 .. 5 */
141 #define	NMAX_CPI_EXTD	9		/* eax = 0x80000000 .. 0x80000008 */
142 
143 struct cpuid_info {
144 	uint_t cpi_pass;		/* last pass completed */
145 	/*
146 	 * standard function information
147 	 */
148 	uint_t cpi_maxeax;		/* fn 0: %eax */
149 	char cpi_vendorstr[13];		/* fn 0: %ebx:%ecx:%edx */
150 	uint_t cpi_vendor;		/* enum of cpi_vendorstr */
151 
152 	uint_t cpi_family;		/* fn 1: extended family */
153 	uint_t cpi_model;		/* fn 1: extended model */
154 	uint_t cpi_step;		/* fn 1: stepping */
155 	chipid_t cpi_chipid;		/* fn 1: %ebx: chip # on ht cpus */
156 	uint_t cpi_brandid;		/* fn 1: %ebx: brand ID */
157 	int cpi_clogid;			/* fn 1: %ebx: thread # */
158 	uint_t cpi_ncpu_per_chip;	/* fn 1: %ebx: logical cpu count */
159 	uint8_t cpi_cacheinfo[16];	/* fn 2: intel-style cache desc */
160 	uint_t cpi_ncache;		/* fn 2: number of elements */
161 	uint_t cpi_ncpu_shr_last_cache;	/* fn 4: %eax: ncpus sharing cache */
162 	id_t cpi_last_lvl_cacheid;	/* fn 4: %eax: derived cache id */
163 	uint_t cpi_std_4_size;		/* fn 4: number of fn 4 elements */
164 	struct cpuid_regs **cpi_std_4;	/* fn 4: %ecx == 0 .. fn4_size */
165 	struct cpuid_regs cpi_std[NMAX_CPI_STD];	/* 0 .. 5 */
166 	/*
167 	 * extended function information
168 	 */
169 	uint_t cpi_xmaxeax;		/* fn 0x80000000: %eax */
170 	char cpi_brandstr[49];		/* fn 0x8000000[234] */
171 	uint8_t cpi_pabits;		/* fn 0x80000006: %eax */
172 	uint8_t cpi_vabits;		/* fn 0x80000006: %eax */
173 	struct cpuid_regs cpi_extd[NMAX_CPI_EXTD]; /* 0x8000000[0-8] */
174 	id_t cpi_coreid;
175 	uint_t cpi_ncore_per_chip;	/* AMD: fn 0x80000008: %ecx[7-0] */
176 					/* Intel: fn 4: %eax[31-26] */
177 	/*
178 	 * supported feature information
179 	 */
180 	uint32_t cpi_support[5];
181 #define	STD_EDX_FEATURES	0
182 #define	AMD_EDX_FEATURES	1
183 #define	TM_EDX_FEATURES		2
184 #define	STD_ECX_FEATURES	3
185 #define	AMD_ECX_FEATURES	4
186 	/*
187 	 * Synthesized information, where known.
188 	 */
189 	uint32_t cpi_chiprev;		/* See X86_CHIPREV_* in x86_archext.h */
190 	const char *cpi_chiprevstr;	/* May be NULL if chiprev unknown */
191 	uint32_t cpi_socket;		/* Chip package/socket type */
192 
193 	struct mwait_info cpi_mwait;	/* fn 5: monitor/mwait info */
194 };
195 
196 
197 static struct cpuid_info cpuid_info0;
198 
199 /*
200  * These bit fields are defined by the Intel Application Note AP-485
201  * "Intel Processor Identification and the CPUID Instruction"
202  */
203 #define	CPI_FAMILY_XTD(cpi)	BITX((cpi)->cpi_std[1].cp_eax, 27, 20)
204 #define	CPI_MODEL_XTD(cpi)	BITX((cpi)->cpi_std[1].cp_eax, 19, 16)
205 #define	CPI_TYPE(cpi)		BITX((cpi)->cpi_std[1].cp_eax, 13, 12)
206 #define	CPI_FAMILY(cpi)		BITX((cpi)->cpi_std[1].cp_eax, 11, 8)
207 #define	CPI_STEP(cpi)		BITX((cpi)->cpi_std[1].cp_eax, 3, 0)
208 #define	CPI_MODEL(cpi)		BITX((cpi)->cpi_std[1].cp_eax, 7, 4)
209 
210 #define	CPI_FEATURES_EDX(cpi)		((cpi)->cpi_std[1].cp_edx)
211 #define	CPI_FEATURES_ECX(cpi)		((cpi)->cpi_std[1].cp_ecx)
212 #define	CPI_FEATURES_XTD_EDX(cpi)	((cpi)->cpi_extd[1].cp_edx)
213 #define	CPI_FEATURES_XTD_ECX(cpi)	((cpi)->cpi_extd[1].cp_ecx)
214 
215 #define	CPI_BRANDID(cpi)	BITX((cpi)->cpi_std[1].cp_ebx, 7, 0)
216 #define	CPI_CHUNKS(cpi)		BITX((cpi)->cpi_std[1].cp_ebx, 15, 7)
217 #define	CPI_CPU_COUNT(cpi)	BITX((cpi)->cpi_std[1].cp_ebx, 23, 16)
218 #define	CPI_APIC_ID(cpi)	BITX((cpi)->cpi_std[1].cp_ebx, 31, 24)
219 
220 #define	CPI_MAXEAX_MAX		0x100		/* sanity control */
221 #define	CPI_XMAXEAX_MAX		0x80000100
222 #define	CPI_FN4_ECX_MAX		0x20		/* sanity: max fn 4 levels */
223 
224 /*
225  * Function 4 (Deterministic Cache Parameters) macros
226  * Defined by Intel Application Note AP-485
227  */
228 #define	CPI_NUM_CORES(regs)		BITX((regs)->cp_eax, 31, 26)
229 #define	CPI_NTHR_SHR_CACHE(regs)	BITX((regs)->cp_eax, 25, 14)
230 #define	CPI_FULL_ASSOC_CACHE(regs)	BITX((regs)->cp_eax, 9, 9)
231 #define	CPI_SELF_INIT_CACHE(regs)	BITX((regs)->cp_eax, 8, 8)
232 #define	CPI_CACHE_LVL(regs)		BITX((regs)->cp_eax, 7, 5)
233 #define	CPI_CACHE_TYPE(regs)		BITX((regs)->cp_eax, 4, 0)
234 
235 #define	CPI_CACHE_WAYS(regs)		BITX((regs)->cp_ebx, 31, 22)
236 #define	CPI_CACHE_PARTS(regs)		BITX((regs)->cp_ebx, 21, 12)
237 #define	CPI_CACHE_COH_LN_SZ(regs)	BITX((regs)->cp_ebx, 11, 0)
238 
239 #define	CPI_CACHE_SETS(regs)		BITX((regs)->cp_ecx, 31, 0)
240 
241 #define	CPI_PREFCH_STRIDE(regs)		BITX((regs)->cp_edx, 9, 0)
242 
243 
244 /*
245  * A couple of shorthand macros to identify "later" P6-family chips
246  * like the Pentium M and Core.  First, the "older" P6-based stuff
247  * (loosely defined as "pre-Pentium-4"):
248  * P6, PII, Mobile PII, PII Xeon, PIII, Mobile PIII, PIII Xeon
249  */
250 
251 #define	IS_LEGACY_P6(cpi) (			\
252 	cpi->cpi_family == 6 && 		\
253 		(cpi->cpi_model == 1 ||		\
254 		cpi->cpi_model == 3 ||		\
255 		cpi->cpi_model == 5 ||		\
256 		cpi->cpi_model == 6 ||		\
257 		cpi->cpi_model == 7 ||		\
258 		cpi->cpi_model == 8 ||		\
259 		cpi->cpi_model == 0xA ||	\
260 		cpi->cpi_model == 0xB)		\
261 )
262 
263 /* A "new F6" is everything with family 6 that's not the above */
264 #define	IS_NEW_F6(cpi) ((cpi->cpi_family == 6) && !IS_LEGACY_P6(cpi))
265 
266 /* Extended family/model support */
267 #define	IS_EXTENDED_MODEL_INTEL(cpi) (cpi->cpi_family == 0x6 || \
268 	cpi->cpi_family >= 0xf)
269 
270 /*
271  * AMD family 0xf and family 0x10 socket types.
272  * First index :
273  *		0 for family 0xf, revs B thru E
274  *		1 for family 0xf, revs F and G
275  *		2 for family 0x10, rev B
276  * Second index by (model & 0x3)
277  */
278 static uint32_t amd_skts[3][4] = {
279 	/*
280 	 * Family 0xf revisions B through E
281 	 */
282 #define	A_SKTS_0			0
283 	{
284 		X86_SOCKET_754,		/* 0b00 */
285 		X86_SOCKET_940,		/* 0b01 */
286 		X86_SOCKET_754,		/* 0b10 */
287 		X86_SOCKET_939		/* 0b11 */
288 	},
289 	/*
290 	 * Family 0xf revisions F and G
291 	 */
292 #define	A_SKTS_1			1
293 	{
294 		X86_SOCKET_S1g1,	/* 0b00 */
295 		X86_SOCKET_F1207,	/* 0b01 */
296 		X86_SOCKET_UNKNOWN,	/* 0b10 */
297 		X86_SOCKET_AM2		/* 0b11 */
298 	},
299 	/*
300 	 * Family 0x10 revisions A and B
301 	 * It is not clear whether, as new sockets release, that
302 	 * model & 0x3 will id socket for this family
303 	 */
304 #define	A_SKTS_2			2
305 	{
306 		X86_SOCKET_F1207,	/* 0b00 */
307 		X86_SOCKET_F1207,	/* 0b01 */
308 		X86_SOCKET_F1207,	/* 0b10 */
309 		X86_SOCKET_F1207,	/* 0b11 */
310 	}
311 };
312 
313 /*
314  * Table for mapping AMD Family 0xf and AMD Family 0x10 model/stepping
315  * combination to chip "revision" and socket type.
316  *
317  * The first member of this array that matches a given family, extended model
318  * plus model range, and stepping range will be considered a match.
319  */
320 static const struct amd_rev_mapent {
321 	uint_t rm_family;
322 	uint_t rm_modello;
323 	uint_t rm_modelhi;
324 	uint_t rm_steplo;
325 	uint_t rm_stephi;
326 	uint32_t rm_chiprev;
327 	const char *rm_chiprevstr;
328 	int rm_sktidx;
329 } amd_revmap[] = {
330 	/*
331 	 * =============== AuthenticAMD Family 0xf ===============
332 	 */
333 
334 	/*
335 	 * Rev B includes model 0x4 stepping 0 and model 0x5 stepping 0 and 1.
336 	 */
337 	{ 0xf, 0x04, 0x04, 0x0, 0x0, X86_CHIPREV_AMD_F_REV_B, "B", A_SKTS_0 },
338 	{ 0xf, 0x05, 0x05, 0x0, 0x1, X86_CHIPREV_AMD_F_REV_B, "B", A_SKTS_0 },
339 	/*
340 	 * Rev C0 includes model 0x4 stepping 8 and model 0x5 stepping 8
341 	 */
342 	{ 0xf, 0x04, 0x05, 0x8, 0x8, X86_CHIPREV_AMD_F_REV_C0, "C0", A_SKTS_0 },
343 	/*
344 	 * Rev CG is the rest of extended model 0x0 - i.e., everything
345 	 * but the rev B and C0 combinations covered above.
346 	 */
347 	{ 0xf, 0x00, 0x0f, 0x0, 0xf, X86_CHIPREV_AMD_F_REV_CG, "CG", A_SKTS_0 },
348 	/*
349 	 * Rev D has extended model 0x1.
350 	 */
351 	{ 0xf, 0x10, 0x1f, 0x0, 0xf, X86_CHIPREV_AMD_F_REV_D, "D", A_SKTS_0 },
352 	/*
353 	 * Rev E has extended model 0x2.
354 	 * Extended model 0x3 is unused but available to grow into.
355 	 */
356 	{ 0xf, 0x20, 0x3f, 0x0, 0xf, X86_CHIPREV_AMD_F_REV_E, "E", A_SKTS_0 },
357 	/*
358 	 * Rev F has extended models 0x4 and 0x5.
359 	 */
360 	{ 0xf, 0x40, 0x5f, 0x0, 0xf, X86_CHIPREV_AMD_F_REV_F, "F", A_SKTS_1 },
361 	/*
362 	 * Rev G has extended model 0x6.
363 	 */
364 	{ 0xf, 0x60, 0x6f, 0x0, 0xf, X86_CHIPREV_AMD_F_REV_G, "G", A_SKTS_1 },
365 
366 	/*
367 	 * =============== AuthenticAMD Family 0x10 ===============
368 	 */
369 
370 	/*
371 	 * Rev A has model 0 and stepping 0/1/2 for DR-{A0,A1,A2}.
372 	 * Give all of model 0 stepping range to rev A.
373 	 */
374 	{ 0x10, 0x00, 0x00, 0x0, 0x2, X86_CHIPREV_AMD_10_REV_A, "A", A_SKTS_2 },
375 
376 	/*
377 	 * Rev B has model 2 and steppings 0/1/0xa/2 for DR-{B0,B1,BA,B2}.
378 	 * Give all of model 2 stepping range to rev B.
379 	 */
380 	{ 0x10, 0x02, 0x02, 0x0, 0xf, X86_CHIPREV_AMD_10_REV_B, "B", A_SKTS_2 },
381 };
382 
383 /*
384  * Info for monitor/mwait idle loop.
385  *
386  * See cpuid section of "Intel 64 and IA-32 Architectures Software Developer's
387  * Manual Volume 2A: Instruction Set Reference, A-M" #25366-022US, November
388  * 2006.
389  * See MONITOR/MWAIT section of "AMD64 Architecture Programmer's Manual
390  * Documentation Updates" #33633, Rev 2.05, December 2006.
391  */
392 #define	MWAIT_SUPPORT		(0x00000001)	/* mwait supported */
393 #define	MWAIT_EXTENSIONS	(0x00000002)	/* extenstion supported */
394 #define	MWAIT_ECX_INT_ENABLE	(0x00000004)	/* ecx 1 extension supported */
395 #define	MWAIT_SUPPORTED(cpi)	((cpi)->cpi_std[1].cp_ecx & CPUID_INTC_ECX_MON)
396 #define	MWAIT_INT_ENABLE(cpi)	((cpi)->cpi_std[5].cp_ecx & 0x2)
397 #define	MWAIT_EXTENSION(cpi)	((cpi)->cpi_std[5].cp_ecx & 0x1)
398 #define	MWAIT_SIZE_MIN(cpi)	BITX((cpi)->cpi_std[5].cp_eax, 15, 0)
399 #define	MWAIT_SIZE_MAX(cpi)	BITX((cpi)->cpi_std[5].cp_ebx, 15, 0)
400 /*
401  * Number of sub-cstates for a given c-state.
402  */
403 #define	MWAIT_NUM_SUBC_STATES(cpi, c_state)			\
404 	BITX((cpi)->cpi_std[5].cp_edx, c_state + 3, c_state)
405 
406 static void intel_cpuid_4_cache_info(void *, struct cpuid_info *);
407 
408 static void
409 synth_amd_info(struct cpuid_info *cpi)
410 {
411 	const struct amd_rev_mapent *rmp;
412 	uint_t family, model, step;
413 	int i;
414 
415 	/*
416 	 * Currently only AMD family 0xf and family 0x10 use these fields.
417 	 */
418 	if (cpi->cpi_family != 0xf && cpi->cpi_family != 0x10)
419 		return;
420 
421 	family = cpi->cpi_family;
422 	model = cpi->cpi_model;
423 	step = cpi->cpi_step;
424 
425 	for (i = 0, rmp = amd_revmap; i < sizeof (amd_revmap) / sizeof (*rmp);
426 	    i++, rmp++) {
427 		if (family == rmp->rm_family &&
428 		    model >= rmp->rm_modello && model <= rmp->rm_modelhi &&
429 		    step >= rmp->rm_steplo && step <= rmp->rm_stephi) {
430 			cpi->cpi_chiprev = rmp->rm_chiprev;
431 			cpi->cpi_chiprevstr = rmp->rm_chiprevstr;
432 			cpi->cpi_socket = amd_skts[rmp->rm_sktidx][model & 0x3];
433 			return;
434 		}
435 	}
436 }
437 
438 static void
439 synth_info(struct cpuid_info *cpi)
440 {
441 	cpi->cpi_chiprev = X86_CHIPREV_UNKNOWN;
442 	cpi->cpi_chiprevstr = "Unknown";
443 	cpi->cpi_socket = X86_SOCKET_UNKNOWN;
444 
445 	switch (cpi->cpi_vendor) {
446 	case X86_VENDOR_AMD:
447 		synth_amd_info(cpi);
448 		break;
449 
450 	default:
451 		break;
452 
453 	}
454 }
455 
456 /*
457  * Apply up various platform-dependent restrictions where the
458  * underlying platform restrictions mean the CPU can be marked
459  * as less capable than its cpuid instruction would imply.
460  */
461 #if defined(__xpv)
462 static void
463 platform_cpuid_mangle(uint_t vendor, uint32_t eax, struct cpuid_regs *cp)
464 {
465 	switch (eax) {
466 	case 1:
467 		cp->cp_edx &=
468 		    ~(CPUID_INTC_EDX_PSE |
469 		    CPUID_INTC_EDX_VME | CPUID_INTC_EDX_DE |
470 		    CPUID_INTC_EDX_MCA |	/* XXPV true on dom0? */
471 		    CPUID_INTC_EDX_SEP | CPUID_INTC_EDX_MTRR |
472 		    CPUID_INTC_EDX_PGE | CPUID_INTC_EDX_PAT |
473 		    CPUID_AMD_EDX_SYSC | CPUID_INTC_EDX_SEP |
474 		    CPUID_INTC_EDX_PSE36 | CPUID_INTC_EDX_HTT);
475 		break;
476 
477 	case 0x80000001:
478 		cp->cp_edx &=
479 		    ~(CPUID_AMD_EDX_PSE |
480 		    CPUID_INTC_EDX_VME | CPUID_INTC_EDX_DE |
481 		    CPUID_AMD_EDX_MTRR | CPUID_AMD_EDX_PGE |
482 		    CPUID_AMD_EDX_PAT | CPUID_AMD_EDX_PSE36 |
483 		    CPUID_AMD_EDX_SYSC | CPUID_INTC_EDX_SEP |
484 		    CPUID_AMD_EDX_TSCP);
485 		cp->cp_ecx &= ~CPUID_AMD_ECX_CMP_LGCY;
486 		break;
487 	default:
488 		break;
489 	}
490 
491 	switch (vendor) {
492 	case X86_VENDOR_Intel:
493 		switch (eax) {
494 		case 4:
495 			/*
496 			 * Zero out the (ncores-per-chip - 1) field
497 			 */
498 			cp->cp_eax &= 0x03fffffff;
499 			break;
500 		default:
501 			break;
502 		}
503 		break;
504 	case X86_VENDOR_AMD:
505 		switch (eax) {
506 		case 0x80000008:
507 			/*
508 			 * Zero out the (ncores-per-chip - 1) field
509 			 */
510 			cp->cp_ecx &= 0xffffff00;
511 			break;
512 		default:
513 			break;
514 		}
515 		break;
516 	default:
517 		break;
518 	}
519 }
520 #else
521 #define	platform_cpuid_mangle(vendor, eax, cp)	/* nothing */
522 #endif
523 
524 /*
525  *  Some undocumented ways of patching the results of the cpuid
526  *  instruction to permit running Solaris 10 on future cpus that
527  *  we don't currently support.  Could be set to non-zero values
528  *  via settings in eeprom.
529  */
530 
531 uint32_t cpuid_feature_ecx_include;
532 uint32_t cpuid_feature_ecx_exclude;
533 uint32_t cpuid_feature_edx_include;
534 uint32_t cpuid_feature_edx_exclude;
535 
536 void
537 cpuid_alloc_space(cpu_t *cpu)
538 {
539 	/*
540 	 * By convention, cpu0 is the boot cpu, which is set up
541 	 * before memory allocation is available.  All other cpus get
542 	 * their cpuid_info struct allocated here.
543 	 */
544 	ASSERT(cpu->cpu_id != 0);
545 	cpu->cpu_m.mcpu_cpi =
546 	    kmem_zalloc(sizeof (*cpu->cpu_m.mcpu_cpi), KM_SLEEP);
547 }
548 
549 void
550 cpuid_free_space(cpu_t *cpu)
551 {
552 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
553 	int i;
554 
555 	ASSERT(cpu->cpu_id != 0);
556 
557 	/*
558 	 * Free up any function 4 related dynamic storage
559 	 */
560 	for (i = 1; i < cpi->cpi_std_4_size; i++)
561 		kmem_free(cpi->cpi_std_4[i], sizeof (struct cpuid_regs));
562 	if (cpi->cpi_std_4_size > 0)
563 		kmem_free(cpi->cpi_std_4,
564 		    cpi->cpi_std_4_size * sizeof (struct cpuid_regs *));
565 
566 	kmem_free(cpu->cpu_m.mcpu_cpi, sizeof (*cpu->cpu_m.mcpu_cpi));
567 }
568 
569 uint_t
570 cpuid_pass1(cpu_t *cpu)
571 {
572 	uint32_t mask_ecx, mask_edx;
573 	uint_t feature = X86_CPUID;
574 	struct cpuid_info *cpi;
575 	struct cpuid_regs *cp;
576 	int xcpuid;
577 #if !defined(__xpv)
578 	extern int idle_cpu_prefer_mwait;
579 #endif
580 
581 	/*
582 	 * Space statically allocated for cpu0, ensure pointer is set
583 	 */
584 	if (cpu->cpu_id == 0)
585 		cpu->cpu_m.mcpu_cpi = &cpuid_info0;
586 	cpi = cpu->cpu_m.mcpu_cpi;
587 	ASSERT(cpi != NULL);
588 	cp = &cpi->cpi_std[0];
589 	cp->cp_eax = 0;
590 	cpi->cpi_maxeax = __cpuid_insn(cp);
591 	{
592 		uint32_t *iptr = (uint32_t *)cpi->cpi_vendorstr;
593 		*iptr++ = cp->cp_ebx;
594 		*iptr++ = cp->cp_edx;
595 		*iptr++ = cp->cp_ecx;
596 		*(char *)&cpi->cpi_vendorstr[12] = '\0';
597 	}
598 
599 	/*
600 	 * Map the vendor string to a type code
601 	 */
602 	if (strcmp(cpi->cpi_vendorstr, "GenuineIntel") == 0)
603 		cpi->cpi_vendor = X86_VENDOR_Intel;
604 	else if (strcmp(cpi->cpi_vendorstr, "AuthenticAMD") == 0)
605 		cpi->cpi_vendor = X86_VENDOR_AMD;
606 	else if (strcmp(cpi->cpi_vendorstr, "GenuineTMx86") == 0)
607 		cpi->cpi_vendor = X86_VENDOR_TM;
608 	else if (strcmp(cpi->cpi_vendorstr, CyrixInstead) == 0)
609 		/*
610 		 * CyrixInstead is a variable used by the Cyrix detection code
611 		 * in locore.
612 		 */
613 		cpi->cpi_vendor = X86_VENDOR_Cyrix;
614 	else if (strcmp(cpi->cpi_vendorstr, "UMC UMC UMC ") == 0)
615 		cpi->cpi_vendor = X86_VENDOR_UMC;
616 	else if (strcmp(cpi->cpi_vendorstr, "NexGenDriven") == 0)
617 		cpi->cpi_vendor = X86_VENDOR_NexGen;
618 	else if (strcmp(cpi->cpi_vendorstr, "CentaurHauls") == 0)
619 		cpi->cpi_vendor = X86_VENDOR_Centaur;
620 	else if (strcmp(cpi->cpi_vendorstr, "RiseRiseRise") == 0)
621 		cpi->cpi_vendor = X86_VENDOR_Rise;
622 	else if (strcmp(cpi->cpi_vendorstr, "SiS SiS SiS ") == 0)
623 		cpi->cpi_vendor = X86_VENDOR_SiS;
624 	else if (strcmp(cpi->cpi_vendorstr, "Geode by NSC") == 0)
625 		cpi->cpi_vendor = X86_VENDOR_NSC;
626 	else
627 		cpi->cpi_vendor = X86_VENDOR_IntelClone;
628 
629 	x86_vendor = cpi->cpi_vendor; /* for compatibility */
630 
631 	/*
632 	 * Limit the range in case of weird hardware
633 	 */
634 	if (cpi->cpi_maxeax > CPI_MAXEAX_MAX)
635 		cpi->cpi_maxeax = CPI_MAXEAX_MAX;
636 	if (cpi->cpi_maxeax < 1)
637 		goto pass1_done;
638 
639 	cp = &cpi->cpi_std[1];
640 	cp->cp_eax = 1;
641 	(void) __cpuid_insn(cp);
642 
643 	/*
644 	 * Extract identifying constants for easy access.
645 	 */
646 	cpi->cpi_model = CPI_MODEL(cpi);
647 	cpi->cpi_family = CPI_FAMILY(cpi);
648 
649 	if (cpi->cpi_family == 0xf)
650 		cpi->cpi_family += CPI_FAMILY_XTD(cpi);
651 
652 	/*
653 	 * Beware: AMD uses "extended model" iff base *FAMILY* == 0xf.
654 	 * Intel, and presumably everyone else, uses model == 0xf, as
655 	 * one would expect (max value means possible overflow).  Sigh.
656 	 */
657 
658 	switch (cpi->cpi_vendor) {
659 	case X86_VENDOR_Intel:
660 		if (IS_EXTENDED_MODEL_INTEL(cpi))
661 			cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4;
662 		break;
663 	case X86_VENDOR_AMD:
664 		if (CPI_FAMILY(cpi) == 0xf)
665 			cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4;
666 		break;
667 	default:
668 		if (cpi->cpi_model == 0xf)
669 			cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4;
670 		break;
671 	}
672 
673 	cpi->cpi_step = CPI_STEP(cpi);
674 	cpi->cpi_brandid = CPI_BRANDID(cpi);
675 
676 	/*
677 	 * *default* assumptions:
678 	 * - believe %edx feature word
679 	 * - ignore %ecx feature word
680 	 * - 32-bit virtual and physical addressing
681 	 */
682 	mask_edx = 0xffffffff;
683 	mask_ecx = 0;
684 
685 	cpi->cpi_pabits = cpi->cpi_vabits = 32;
686 
687 	switch (cpi->cpi_vendor) {
688 	case X86_VENDOR_Intel:
689 		if (cpi->cpi_family == 5)
690 			x86_type = X86_TYPE_P5;
691 		else if (IS_LEGACY_P6(cpi)) {
692 			x86_type = X86_TYPE_P6;
693 			pentiumpro_bug4046376 = 1;
694 			pentiumpro_bug4064495 = 1;
695 			/*
696 			 * Clear the SEP bit when it was set erroneously
697 			 */
698 			if (cpi->cpi_model < 3 && cpi->cpi_step < 3)
699 				cp->cp_edx &= ~CPUID_INTC_EDX_SEP;
700 		} else if (IS_NEW_F6(cpi) || cpi->cpi_family == 0xf) {
701 			x86_type = X86_TYPE_P4;
702 			/*
703 			 * We don't currently depend on any of the %ecx
704 			 * features until Prescott, so we'll only check
705 			 * this from P4 onwards.  We might want to revisit
706 			 * that idea later.
707 			 */
708 			mask_ecx = 0xffffffff;
709 		} else if (cpi->cpi_family > 0xf)
710 			mask_ecx = 0xffffffff;
711 		/*
712 		 * We don't support MONITOR/MWAIT if leaf 5 is not available
713 		 * to obtain the monitor linesize.
714 		 */
715 		if (cpi->cpi_maxeax < 5)
716 			mask_ecx &= ~CPUID_INTC_ECX_MON;
717 		break;
718 	case X86_VENDOR_IntelClone:
719 	default:
720 		break;
721 	case X86_VENDOR_AMD:
722 #if defined(OPTERON_ERRATUM_108)
723 		if (cpi->cpi_family == 0xf && cpi->cpi_model == 0xe) {
724 			cp->cp_eax = (0xf0f & cp->cp_eax) | 0xc0;
725 			cpi->cpi_model = 0xc;
726 		} else
727 #endif
728 		if (cpi->cpi_family == 5) {
729 			/*
730 			 * AMD K5 and K6
731 			 *
732 			 * These CPUs have an incomplete implementation
733 			 * of MCA/MCE which we mask away.
734 			 */
735 			mask_edx &= ~(CPUID_INTC_EDX_MCE | CPUID_INTC_EDX_MCA);
736 
737 			/*
738 			 * Model 0 uses the wrong (APIC) bit
739 			 * to indicate PGE.  Fix it here.
740 			 */
741 			if (cpi->cpi_model == 0) {
742 				if (cp->cp_edx & 0x200) {
743 					cp->cp_edx &= ~0x200;
744 					cp->cp_edx |= CPUID_INTC_EDX_PGE;
745 				}
746 			}
747 
748 			/*
749 			 * Early models had problems w/ MMX; disable.
750 			 */
751 			if (cpi->cpi_model < 6)
752 				mask_edx &= ~CPUID_INTC_EDX_MMX;
753 		}
754 
755 		/*
756 		 * For newer families, SSE3 and CX16, at least, are valid;
757 		 * enable all
758 		 */
759 		if (cpi->cpi_family >= 0xf)
760 			mask_ecx = 0xffffffff;
761 		/*
762 		 * We don't support MONITOR/MWAIT if leaf 5 is not available
763 		 * to obtain the monitor linesize.
764 		 */
765 		if (cpi->cpi_maxeax < 5)
766 			mask_ecx &= ~CPUID_INTC_ECX_MON;
767 
768 #if !defined(__xpv)
769 		/*
770 		 * Do not use MONITOR/MWAIT to halt in the idle loop on any AMD
771 		 * processors.  AMD does not intend MWAIT to be used in the cpu
772 		 * idle loop on current and future processors.  10h and future
773 		 * AMD processors use more power in MWAIT than HLT.
774 		 * Pre-family-10h Opterons do not have the MWAIT instruction.
775 		 */
776 		idle_cpu_prefer_mwait = 0;
777 #endif
778 
779 		break;
780 	case X86_VENDOR_TM:
781 		/*
782 		 * workaround the NT workaround in CMS 4.1
783 		 */
784 		if (cpi->cpi_family == 5 && cpi->cpi_model == 4 &&
785 		    (cpi->cpi_step == 2 || cpi->cpi_step == 3))
786 			cp->cp_edx |= CPUID_INTC_EDX_CX8;
787 		break;
788 	case X86_VENDOR_Centaur:
789 		/*
790 		 * workaround the NT workarounds again
791 		 */
792 		if (cpi->cpi_family == 6)
793 			cp->cp_edx |= CPUID_INTC_EDX_CX8;
794 		break;
795 	case X86_VENDOR_Cyrix:
796 		/*
797 		 * We rely heavily on the probing in locore
798 		 * to actually figure out what parts, if any,
799 		 * of the Cyrix cpuid instruction to believe.
800 		 */
801 		switch (x86_type) {
802 		case X86_TYPE_CYRIX_486:
803 			mask_edx = 0;
804 			break;
805 		case X86_TYPE_CYRIX_6x86:
806 			mask_edx = 0;
807 			break;
808 		case X86_TYPE_CYRIX_6x86L:
809 			mask_edx =
810 			    CPUID_INTC_EDX_DE |
811 			    CPUID_INTC_EDX_CX8;
812 			break;
813 		case X86_TYPE_CYRIX_6x86MX:
814 			mask_edx =
815 			    CPUID_INTC_EDX_DE |
816 			    CPUID_INTC_EDX_MSR |
817 			    CPUID_INTC_EDX_CX8 |
818 			    CPUID_INTC_EDX_PGE |
819 			    CPUID_INTC_EDX_CMOV |
820 			    CPUID_INTC_EDX_MMX;
821 			break;
822 		case X86_TYPE_CYRIX_GXm:
823 			mask_edx =
824 			    CPUID_INTC_EDX_MSR |
825 			    CPUID_INTC_EDX_CX8 |
826 			    CPUID_INTC_EDX_CMOV |
827 			    CPUID_INTC_EDX_MMX;
828 			break;
829 		case X86_TYPE_CYRIX_MediaGX:
830 			break;
831 		case X86_TYPE_CYRIX_MII:
832 		case X86_TYPE_VIA_CYRIX_III:
833 			mask_edx =
834 			    CPUID_INTC_EDX_DE |
835 			    CPUID_INTC_EDX_TSC |
836 			    CPUID_INTC_EDX_MSR |
837 			    CPUID_INTC_EDX_CX8 |
838 			    CPUID_INTC_EDX_PGE |
839 			    CPUID_INTC_EDX_CMOV |
840 			    CPUID_INTC_EDX_MMX;
841 			break;
842 		default:
843 			break;
844 		}
845 		break;
846 	}
847 
848 #if defined(__xpv)
849 	/*
850 	 * Do not support MONITOR/MWAIT under a hypervisor
851 	 */
852 	mask_ecx &= ~CPUID_INTC_ECX_MON;
853 #endif	/* __xpv */
854 
855 	/*
856 	 * Now we've figured out the masks that determine
857 	 * which bits we choose to believe, apply the masks
858 	 * to the feature words, then map the kernel's view
859 	 * of these feature words into its feature word.
860 	 */
861 	cp->cp_edx &= mask_edx;
862 	cp->cp_ecx &= mask_ecx;
863 
864 	/*
865 	 * apply any platform restrictions (we don't call this
866 	 * immediately after __cpuid_insn here, because we need the
867 	 * workarounds applied above first)
868 	 */
869 	platform_cpuid_mangle(cpi->cpi_vendor, 1, cp);
870 
871 	/*
872 	 * fold in overrides from the "eeprom" mechanism
873 	 */
874 	cp->cp_edx |= cpuid_feature_edx_include;
875 	cp->cp_edx &= ~cpuid_feature_edx_exclude;
876 
877 	cp->cp_ecx |= cpuid_feature_ecx_include;
878 	cp->cp_ecx &= ~cpuid_feature_ecx_exclude;
879 
880 	if (cp->cp_edx & CPUID_INTC_EDX_PSE)
881 		feature |= X86_LARGEPAGE;
882 	if (cp->cp_edx & CPUID_INTC_EDX_TSC)
883 		feature |= X86_TSC;
884 	if (cp->cp_edx & CPUID_INTC_EDX_MSR)
885 		feature |= X86_MSR;
886 	if (cp->cp_edx & CPUID_INTC_EDX_MTRR)
887 		feature |= X86_MTRR;
888 	if (cp->cp_edx & CPUID_INTC_EDX_PGE)
889 		feature |= X86_PGE;
890 	if (cp->cp_edx & CPUID_INTC_EDX_CMOV)
891 		feature |= X86_CMOV;
892 	if (cp->cp_edx & CPUID_INTC_EDX_MMX)
893 		feature |= X86_MMX;
894 	if ((cp->cp_edx & CPUID_INTC_EDX_MCE) != 0 &&
895 	    (cp->cp_edx & CPUID_INTC_EDX_MCA) != 0)
896 		feature |= X86_MCA;
897 	if (cp->cp_edx & CPUID_INTC_EDX_PAE)
898 		feature |= X86_PAE;
899 	if (cp->cp_edx & CPUID_INTC_EDX_CX8)
900 		feature |= X86_CX8;
901 	if (cp->cp_ecx & CPUID_INTC_ECX_CX16)
902 		feature |= X86_CX16;
903 	if (cp->cp_edx & CPUID_INTC_EDX_PAT)
904 		feature |= X86_PAT;
905 	if (cp->cp_edx & CPUID_INTC_EDX_SEP)
906 		feature |= X86_SEP;
907 	if (cp->cp_edx & CPUID_INTC_EDX_FXSR) {
908 		/*
909 		 * In our implementation, fxsave/fxrstor
910 		 * are prerequisites before we'll even
911 		 * try and do SSE things.
912 		 */
913 		if (cp->cp_edx & CPUID_INTC_EDX_SSE)
914 			feature |= X86_SSE;
915 		if (cp->cp_edx & CPUID_INTC_EDX_SSE2)
916 			feature |= X86_SSE2;
917 		if (cp->cp_ecx & CPUID_INTC_ECX_SSE3)
918 			feature |= X86_SSE3;
919 		if (cpi->cpi_vendor == X86_VENDOR_Intel) {
920 			if (cp->cp_ecx & CPUID_INTC_ECX_SSSE3)
921 				feature |= X86_SSSE3;
922 			if (cp->cp_ecx & CPUID_INTC_ECX_SSE4_1)
923 				feature |= X86_SSE4_1;
924 			if (cp->cp_ecx & CPUID_INTC_ECX_SSE4_2)
925 				feature |= X86_SSE4_2;
926 		}
927 	}
928 	if (cp->cp_edx & CPUID_INTC_EDX_DE)
929 		feature |= X86_DE;
930 	if (cp->cp_ecx & CPUID_INTC_ECX_MON) {
931 		cpi->cpi_mwait.support |= MWAIT_SUPPORT;
932 		feature |= X86_MWAIT;
933 	}
934 
935 	if (feature & X86_PAE)
936 		cpi->cpi_pabits = 36;
937 
938 	/*
939 	 * Hyperthreading configuration is slightly tricky on Intel
940 	 * and pure clones, and even trickier on AMD.
941 	 *
942 	 * (AMD chose to set the HTT bit on their CMP processors,
943 	 * even though they're not actually hyperthreaded.  Thus it
944 	 * takes a bit more work to figure out what's really going
945 	 * on ... see the handling of the CMP_LGCY bit below)
946 	 */
947 	if (cp->cp_edx & CPUID_INTC_EDX_HTT) {
948 		cpi->cpi_ncpu_per_chip = CPI_CPU_COUNT(cpi);
949 		if (cpi->cpi_ncpu_per_chip > 1)
950 			feature |= X86_HTT;
951 	} else {
952 		cpi->cpi_ncpu_per_chip = 1;
953 	}
954 
955 	/*
956 	 * Work on the "extended" feature information, doing
957 	 * some basic initialization for cpuid_pass2()
958 	 */
959 	xcpuid = 0;
960 	switch (cpi->cpi_vendor) {
961 	case X86_VENDOR_Intel:
962 		if (IS_NEW_F6(cpi) || cpi->cpi_family >= 0xf)
963 			xcpuid++;
964 		break;
965 	case X86_VENDOR_AMD:
966 		if (cpi->cpi_family > 5 ||
967 		    (cpi->cpi_family == 5 && cpi->cpi_model >= 1))
968 			xcpuid++;
969 		break;
970 	case X86_VENDOR_Cyrix:
971 		/*
972 		 * Only these Cyrix CPUs are -known- to support
973 		 * extended cpuid operations.
974 		 */
975 		if (x86_type == X86_TYPE_VIA_CYRIX_III ||
976 		    x86_type == X86_TYPE_CYRIX_GXm)
977 			xcpuid++;
978 		break;
979 	case X86_VENDOR_Centaur:
980 	case X86_VENDOR_TM:
981 	default:
982 		xcpuid++;
983 		break;
984 	}
985 
986 	if (xcpuid) {
987 		cp = &cpi->cpi_extd[0];
988 		cp->cp_eax = 0x80000000;
989 		cpi->cpi_xmaxeax = __cpuid_insn(cp);
990 	}
991 
992 	if (cpi->cpi_xmaxeax & 0x80000000) {
993 
994 		if (cpi->cpi_xmaxeax > CPI_XMAXEAX_MAX)
995 			cpi->cpi_xmaxeax = CPI_XMAXEAX_MAX;
996 
997 		switch (cpi->cpi_vendor) {
998 		case X86_VENDOR_Intel:
999 		case X86_VENDOR_AMD:
1000 			if (cpi->cpi_xmaxeax < 0x80000001)
1001 				break;
1002 			cp = &cpi->cpi_extd[1];
1003 			cp->cp_eax = 0x80000001;
1004 			(void) __cpuid_insn(cp);
1005 
1006 			if (cpi->cpi_vendor == X86_VENDOR_AMD &&
1007 			    cpi->cpi_family == 5 &&
1008 			    cpi->cpi_model == 6 &&
1009 			    cpi->cpi_step == 6) {
1010 				/*
1011 				 * K6 model 6 uses bit 10 to indicate SYSC
1012 				 * Later models use bit 11. Fix it here.
1013 				 */
1014 				if (cp->cp_edx & 0x400) {
1015 					cp->cp_edx &= ~0x400;
1016 					cp->cp_edx |= CPUID_AMD_EDX_SYSC;
1017 				}
1018 			}
1019 
1020 			platform_cpuid_mangle(cpi->cpi_vendor, 0x80000001, cp);
1021 
1022 			/*
1023 			 * Compute the additions to the kernel's feature word.
1024 			 */
1025 			if (cp->cp_edx & CPUID_AMD_EDX_NX)
1026 				feature |= X86_NX;
1027 
1028 #if defined(__amd64)
1029 			/* 1 GB large page - enable only for 64 bit kernel */
1030 			if (cp->cp_edx & CPUID_AMD_EDX_1GPG)
1031 				feature |= X86_1GPG;
1032 #endif
1033 
1034 			if ((cpi->cpi_vendor == X86_VENDOR_AMD) &&
1035 			    (cpi->cpi_std[1].cp_edx & CPUID_INTC_EDX_FXSR) &&
1036 			    (cp->cp_ecx & CPUID_AMD_ECX_SSE4A))
1037 				feature |= X86_SSE4A;
1038 
1039 			/*
1040 			 * If both the HTT and CMP_LGCY bits are set,
1041 			 * then we're not actually HyperThreaded.  Read
1042 			 * "AMD CPUID Specification" for more details.
1043 			 */
1044 			if (cpi->cpi_vendor == X86_VENDOR_AMD &&
1045 			    (feature & X86_HTT) &&
1046 			    (cp->cp_ecx & CPUID_AMD_ECX_CMP_LGCY)) {
1047 				feature &= ~X86_HTT;
1048 				feature |= X86_CMP;
1049 			}
1050 #if defined(__amd64)
1051 			/*
1052 			 * It's really tricky to support syscall/sysret in
1053 			 * the i386 kernel; we rely on sysenter/sysexit
1054 			 * instead.  In the amd64 kernel, things are -way-
1055 			 * better.
1056 			 */
1057 			if (cp->cp_edx & CPUID_AMD_EDX_SYSC)
1058 				feature |= X86_ASYSC;
1059 
1060 			/*
1061 			 * While we're thinking about system calls, note
1062 			 * that AMD processors don't support sysenter
1063 			 * in long mode at all, so don't try to program them.
1064 			 */
1065 			if (x86_vendor == X86_VENDOR_AMD)
1066 				feature &= ~X86_SEP;
1067 #endif
1068 			if (x86_vendor == X86_VENDOR_AMD &&
1069 			    cp->cp_edx & CPUID_AMD_EDX_TSCP)
1070 				feature |= X86_TSCP;
1071 			break;
1072 		default:
1073 			break;
1074 		}
1075 
1076 		/*
1077 		 * Get CPUID data about processor cores and hyperthreads.
1078 		 */
1079 		switch (cpi->cpi_vendor) {
1080 		case X86_VENDOR_Intel:
1081 			if (cpi->cpi_maxeax >= 4) {
1082 				cp = &cpi->cpi_std[4];
1083 				cp->cp_eax = 4;
1084 				cp->cp_ecx = 0;
1085 				(void) __cpuid_insn(cp);
1086 				platform_cpuid_mangle(cpi->cpi_vendor, 4, cp);
1087 			}
1088 			/*FALLTHROUGH*/
1089 		case X86_VENDOR_AMD:
1090 			if (cpi->cpi_xmaxeax < 0x80000008)
1091 				break;
1092 			cp = &cpi->cpi_extd[8];
1093 			cp->cp_eax = 0x80000008;
1094 			(void) __cpuid_insn(cp);
1095 			platform_cpuid_mangle(cpi->cpi_vendor, 0x80000008, cp);
1096 
1097 			/*
1098 			 * Virtual and physical address limits from
1099 			 * cpuid override previously guessed values.
1100 			 */
1101 			cpi->cpi_pabits = BITX(cp->cp_eax, 7, 0);
1102 			cpi->cpi_vabits = BITX(cp->cp_eax, 15, 8);
1103 			break;
1104 		default:
1105 			break;
1106 		}
1107 
1108 		/*
1109 		 * Derive the number of cores per chip
1110 		 */
1111 		switch (cpi->cpi_vendor) {
1112 		case X86_VENDOR_Intel:
1113 			if (cpi->cpi_maxeax < 4) {
1114 				cpi->cpi_ncore_per_chip = 1;
1115 				break;
1116 			} else {
1117 				cpi->cpi_ncore_per_chip =
1118 				    BITX((cpi)->cpi_std[4].cp_eax, 31, 26) + 1;
1119 			}
1120 			break;
1121 		case X86_VENDOR_AMD:
1122 			if (cpi->cpi_xmaxeax < 0x80000008) {
1123 				cpi->cpi_ncore_per_chip = 1;
1124 				break;
1125 			} else {
1126 				cpi->cpi_ncore_per_chip =
1127 				    BITX((cpi)->cpi_extd[8].cp_ecx, 7, 0) + 1;
1128 			}
1129 			break;
1130 		default:
1131 			cpi->cpi_ncore_per_chip = 1;
1132 			break;
1133 		}
1134 	} else {
1135 		cpi->cpi_ncore_per_chip = 1;
1136 	}
1137 
1138 	/*
1139 	 * If more than one core, then this processor is CMP.
1140 	 */
1141 	if (cpi->cpi_ncore_per_chip > 1)
1142 		feature |= X86_CMP;
1143 
1144 	/*
1145 	 * If the number of cores is the same as the number
1146 	 * of CPUs, then we cannot have HyperThreading.
1147 	 */
1148 	if (cpi->cpi_ncpu_per_chip == cpi->cpi_ncore_per_chip)
1149 		feature &= ~X86_HTT;
1150 
1151 	if ((feature & (X86_HTT | X86_CMP)) == 0) {
1152 		/*
1153 		 * Single-core single-threaded processors.
1154 		 */
1155 		cpi->cpi_chipid = -1;
1156 		cpi->cpi_clogid = 0;
1157 		cpi->cpi_coreid = cpu->cpu_id;
1158 	} else if (cpi->cpi_ncpu_per_chip > 1) {
1159 		uint_t i;
1160 		uint_t chipid_shift = 0;
1161 		uint_t coreid_shift = 0;
1162 		uint_t apic_id = CPI_APIC_ID(cpi);
1163 
1164 		for (i = 1; i < cpi->cpi_ncpu_per_chip; i <<= 1)
1165 			chipid_shift++;
1166 		cpi->cpi_chipid = apic_id >> chipid_shift;
1167 		cpi->cpi_clogid = apic_id & ((1 << chipid_shift) - 1);
1168 
1169 		if (cpi->cpi_vendor == X86_VENDOR_Intel) {
1170 			if (feature & X86_CMP) {
1171 				/*
1172 				 * Multi-core (and possibly multi-threaded)
1173 				 * processors.
1174 				 */
1175 				uint_t ncpu_per_core;
1176 				if (cpi->cpi_ncore_per_chip == 1)
1177 					ncpu_per_core = cpi->cpi_ncpu_per_chip;
1178 				else if (cpi->cpi_ncore_per_chip > 1)
1179 					ncpu_per_core = cpi->cpi_ncpu_per_chip /
1180 					    cpi->cpi_ncore_per_chip;
1181 				/*
1182 				 * 8bit APIC IDs on dual core Pentiums
1183 				 * look like this:
1184 				 *
1185 				 * +-----------------------+------+------+
1186 				 * | Physical Package ID   |  MC  |  HT  |
1187 				 * +-----------------------+------+------+
1188 				 * <------- chipid -------->
1189 				 * <------- coreid --------------->
1190 				 *			   <--- clogid -->
1191 				 *
1192 				 * Where the number of bits necessary to
1193 				 * represent MC and HT fields together equals
1194 				 * to the minimum number of bits necessary to
1195 				 * store the value of cpi->cpi_ncpu_per_chip.
1196 				 * Of those bits, the MC part uses the number
1197 				 * of bits necessary to store the value of
1198 				 * cpi->cpi_ncore_per_chip.
1199 				 */
1200 				for (i = 1; i < ncpu_per_core; i <<= 1)
1201 					coreid_shift++;
1202 				cpi->cpi_coreid = apic_id >> coreid_shift;
1203 			} else if (feature & X86_HTT) {
1204 				/*
1205 				 * Single-core multi-threaded processors.
1206 				 */
1207 				cpi->cpi_coreid = cpi->cpi_chipid;
1208 			}
1209 		} else if (cpi->cpi_vendor == X86_VENDOR_AMD) {
1210 			/*
1211 			 * AMD currently only has dual-core processors with
1212 			 * single-threaded cores.  If they ever release
1213 			 * multi-threaded processors, then this code
1214 			 * will have to be updated.
1215 			 */
1216 			cpi->cpi_coreid = cpu->cpu_id;
1217 		} else {
1218 			/*
1219 			 * All other processors are currently
1220 			 * assumed to have single cores.
1221 			 */
1222 			cpi->cpi_coreid = cpi->cpi_chipid;
1223 		}
1224 	}
1225 
1226 	/*
1227 	 * Synthesize chip "revision" and socket type
1228 	 */
1229 	synth_info(cpi);
1230 
1231 pass1_done:
1232 	cpi->cpi_pass = 1;
1233 	return (feature);
1234 }
1235 
1236 /*
1237  * Make copies of the cpuid table entries we depend on, in
1238  * part for ease of parsing now, in part so that we have only
1239  * one place to correct any of it, in part for ease of
1240  * later export to userland, and in part so we can look at
1241  * this stuff in a crash dump.
1242  */
1243 
1244 /*ARGSUSED*/
1245 void
1246 cpuid_pass2(cpu_t *cpu)
1247 {
1248 	uint_t n, nmax;
1249 	int i;
1250 	struct cpuid_regs *cp;
1251 	uint8_t *dp;
1252 	uint32_t *iptr;
1253 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
1254 
1255 	ASSERT(cpi->cpi_pass == 1);
1256 
1257 	if (cpi->cpi_maxeax < 1)
1258 		goto pass2_done;
1259 
1260 	if ((nmax = cpi->cpi_maxeax + 1) > NMAX_CPI_STD)
1261 		nmax = NMAX_CPI_STD;
1262 	/*
1263 	 * (We already handled n == 0 and n == 1 in pass 1)
1264 	 */
1265 	for (n = 2, cp = &cpi->cpi_std[2]; n < nmax; n++, cp++) {
1266 		cp->cp_eax = n;
1267 
1268 		/*
1269 		 * CPUID function 4 expects %ecx to be initialized
1270 		 * with an index which indicates which cache to return
1271 		 * information about. The OS is expected to call function 4
1272 		 * with %ecx set to 0, 1, 2, ... until it returns with
1273 		 * EAX[4:0] set to 0, which indicates there are no more
1274 		 * caches.
1275 		 *
1276 		 * Here, populate cpi_std[4] with the information returned by
1277 		 * function 4 when %ecx == 0, and do the rest in cpuid_pass3()
1278 		 * when dynamic memory allocation becomes available.
1279 		 *
1280 		 * Note: we need to explicitly initialize %ecx here, since
1281 		 * function 4 may have been previously invoked.
1282 		 */
1283 		if (n == 4)
1284 			cp->cp_ecx = 0;
1285 
1286 		(void) __cpuid_insn(cp);
1287 		platform_cpuid_mangle(cpi->cpi_vendor, n, cp);
1288 		switch (n) {
1289 		case 2:
1290 			/*
1291 			 * "the lower 8 bits of the %eax register
1292 			 * contain a value that identifies the number
1293 			 * of times the cpuid [instruction] has to be
1294 			 * executed to obtain a complete image of the
1295 			 * processor's caching systems."
1296 			 *
1297 			 * How *do* they make this stuff up?
1298 			 */
1299 			cpi->cpi_ncache = sizeof (*cp) *
1300 			    BITX(cp->cp_eax, 7, 0);
1301 			if (cpi->cpi_ncache == 0)
1302 				break;
1303 			cpi->cpi_ncache--;	/* skip count byte */
1304 
1305 			/*
1306 			 * Well, for now, rather than attempt to implement
1307 			 * this slightly dubious algorithm, we just look
1308 			 * at the first 15 ..
1309 			 */
1310 			if (cpi->cpi_ncache > (sizeof (*cp) - 1))
1311 				cpi->cpi_ncache = sizeof (*cp) - 1;
1312 
1313 			dp = cpi->cpi_cacheinfo;
1314 			if (BITX(cp->cp_eax, 31, 31) == 0) {
1315 				uint8_t *p = (void *)&cp->cp_eax;
1316 				for (i = 1; i < 3; i++)
1317 					if (p[i] != 0)
1318 						*dp++ = p[i];
1319 			}
1320 			if (BITX(cp->cp_ebx, 31, 31) == 0) {
1321 				uint8_t *p = (void *)&cp->cp_ebx;
1322 				for (i = 0; i < 4; i++)
1323 					if (p[i] != 0)
1324 						*dp++ = p[i];
1325 			}
1326 			if (BITX(cp->cp_ecx, 31, 31) == 0) {
1327 				uint8_t *p = (void *)&cp->cp_ecx;
1328 				for (i = 0; i < 4; i++)
1329 					if (p[i] != 0)
1330 						*dp++ = p[i];
1331 			}
1332 			if (BITX(cp->cp_edx, 31, 31) == 0) {
1333 				uint8_t *p = (void *)&cp->cp_edx;
1334 				for (i = 0; i < 4; i++)
1335 					if (p[i] != 0)
1336 						*dp++ = p[i];
1337 			}
1338 			break;
1339 
1340 		case 3:	/* Processor serial number, if PSN supported */
1341 			break;
1342 
1343 		case 4:	/* Deterministic cache parameters */
1344 			break;
1345 
1346 		case 5:	/* Monitor/Mwait parameters */
1347 		{
1348 			size_t mwait_size;
1349 
1350 			/*
1351 			 * check cpi_mwait.support which was set in cpuid_pass1
1352 			 */
1353 			if (!(cpi->cpi_mwait.support & MWAIT_SUPPORT))
1354 				break;
1355 
1356 			/*
1357 			 * Protect ourself from insane mwait line size.
1358 			 * Workaround for incomplete hardware emulator(s).
1359 			 */
1360 			mwait_size = (size_t)MWAIT_SIZE_MAX(cpi);
1361 			if (mwait_size < sizeof (uint32_t) ||
1362 			    !ISP2(mwait_size)) {
1363 #if DEBUG
1364 				cmn_err(CE_NOTE, "Cannot handle cpu %d mwait "
1365 				    "size %ld",
1366 				    cpu->cpu_id, (long)mwait_size);
1367 #endif
1368 				break;
1369 			}
1370 
1371 			cpi->cpi_mwait.mon_min = (size_t)MWAIT_SIZE_MIN(cpi);
1372 			cpi->cpi_mwait.mon_max = mwait_size;
1373 			if (MWAIT_EXTENSION(cpi)) {
1374 				cpi->cpi_mwait.support |= MWAIT_EXTENSIONS;
1375 				if (MWAIT_INT_ENABLE(cpi))
1376 					cpi->cpi_mwait.support |=
1377 					    MWAIT_ECX_INT_ENABLE;
1378 			}
1379 			break;
1380 		}
1381 		default:
1382 			break;
1383 		}
1384 	}
1385 
1386 	if ((cpi->cpi_xmaxeax & 0x80000000) == 0)
1387 		goto pass2_done;
1388 
1389 	if ((nmax = cpi->cpi_xmaxeax - 0x80000000 + 1) > NMAX_CPI_EXTD)
1390 		nmax = NMAX_CPI_EXTD;
1391 	/*
1392 	 * Copy the extended properties, fixing them as we go.
1393 	 * (We already handled n == 0 and n == 1 in pass 1)
1394 	 */
1395 	iptr = (void *)cpi->cpi_brandstr;
1396 	for (n = 2, cp = &cpi->cpi_extd[2]; n < nmax; cp++, n++) {
1397 		cp->cp_eax = 0x80000000 + n;
1398 		(void) __cpuid_insn(cp);
1399 		platform_cpuid_mangle(cpi->cpi_vendor, 0x80000000 + n, cp);
1400 		switch (n) {
1401 		case 2:
1402 		case 3:
1403 		case 4:
1404 			/*
1405 			 * Extract the brand string
1406 			 */
1407 			*iptr++ = cp->cp_eax;
1408 			*iptr++ = cp->cp_ebx;
1409 			*iptr++ = cp->cp_ecx;
1410 			*iptr++ = cp->cp_edx;
1411 			break;
1412 		case 5:
1413 			switch (cpi->cpi_vendor) {
1414 			case X86_VENDOR_AMD:
1415 				/*
1416 				 * The Athlon and Duron were the first
1417 				 * parts to report the sizes of the
1418 				 * TLB for large pages. Before then,
1419 				 * we don't trust the data.
1420 				 */
1421 				if (cpi->cpi_family < 6 ||
1422 				    (cpi->cpi_family == 6 &&
1423 				    cpi->cpi_model < 1))
1424 					cp->cp_eax = 0;
1425 				break;
1426 			default:
1427 				break;
1428 			}
1429 			break;
1430 		case 6:
1431 			switch (cpi->cpi_vendor) {
1432 			case X86_VENDOR_AMD:
1433 				/*
1434 				 * The Athlon and Duron were the first
1435 				 * AMD parts with L2 TLB's.
1436 				 * Before then, don't trust the data.
1437 				 */
1438 				if (cpi->cpi_family < 6 ||
1439 				    cpi->cpi_family == 6 &&
1440 				    cpi->cpi_model < 1)
1441 					cp->cp_eax = cp->cp_ebx = 0;
1442 				/*
1443 				 * AMD Duron rev A0 reports L2
1444 				 * cache size incorrectly as 1K
1445 				 * when it is really 64K
1446 				 */
1447 				if (cpi->cpi_family == 6 &&
1448 				    cpi->cpi_model == 3 &&
1449 				    cpi->cpi_step == 0) {
1450 					cp->cp_ecx &= 0xffff;
1451 					cp->cp_ecx |= 0x400000;
1452 				}
1453 				break;
1454 			case X86_VENDOR_Cyrix:	/* VIA C3 */
1455 				/*
1456 				 * VIA C3 processors are a bit messed
1457 				 * up w.r.t. encoding cache sizes in %ecx
1458 				 */
1459 				if (cpi->cpi_family != 6)
1460 					break;
1461 				/*
1462 				 * model 7 and 8 were incorrectly encoded
1463 				 *
1464 				 * xxx is model 8 really broken?
1465 				 */
1466 				if (cpi->cpi_model == 7 ||
1467 				    cpi->cpi_model == 8)
1468 					cp->cp_ecx =
1469 					    BITX(cp->cp_ecx, 31, 24) << 16 |
1470 					    BITX(cp->cp_ecx, 23, 16) << 12 |
1471 					    BITX(cp->cp_ecx, 15, 8) << 8 |
1472 					    BITX(cp->cp_ecx, 7, 0);
1473 				/*
1474 				 * model 9 stepping 1 has wrong associativity
1475 				 */
1476 				if (cpi->cpi_model == 9 && cpi->cpi_step == 1)
1477 					cp->cp_ecx |= 8 << 12;
1478 				break;
1479 			case X86_VENDOR_Intel:
1480 				/*
1481 				 * Extended L2 Cache features function.
1482 				 * First appeared on Prescott.
1483 				 */
1484 			default:
1485 				break;
1486 			}
1487 			break;
1488 		default:
1489 			break;
1490 		}
1491 	}
1492 
1493 pass2_done:
1494 	cpi->cpi_pass = 2;
1495 }
1496 
1497 static const char *
1498 intel_cpubrand(const struct cpuid_info *cpi)
1499 {
1500 	int i;
1501 
1502 	if ((x86_feature & X86_CPUID) == 0 ||
1503 	    cpi->cpi_maxeax < 1 || cpi->cpi_family < 5)
1504 		return ("i486");
1505 
1506 	switch (cpi->cpi_family) {
1507 	case 5:
1508 		return ("Intel Pentium(r)");
1509 	case 6:
1510 		switch (cpi->cpi_model) {
1511 			uint_t celeron, xeon;
1512 			const struct cpuid_regs *cp;
1513 		case 0:
1514 		case 1:
1515 		case 2:
1516 			return ("Intel Pentium(r) Pro");
1517 		case 3:
1518 		case 4:
1519 			return ("Intel Pentium(r) II");
1520 		case 6:
1521 			return ("Intel Celeron(r)");
1522 		case 5:
1523 		case 7:
1524 			celeron = xeon = 0;
1525 			cp = &cpi->cpi_std[2];	/* cache info */
1526 
1527 			for (i = 1; i < 3; i++) {
1528 				uint_t tmp;
1529 
1530 				tmp = (cp->cp_eax >> (8 * i)) & 0xff;
1531 				if (tmp == 0x40)
1532 					celeron++;
1533 				if (tmp >= 0x44 && tmp <= 0x45)
1534 					xeon++;
1535 			}
1536 
1537 			for (i = 0; i < 2; i++) {
1538 				uint_t tmp;
1539 
1540 				tmp = (cp->cp_ebx >> (8 * i)) & 0xff;
1541 				if (tmp == 0x40)
1542 					celeron++;
1543 				else if (tmp >= 0x44 && tmp <= 0x45)
1544 					xeon++;
1545 			}
1546 
1547 			for (i = 0; i < 4; i++) {
1548 				uint_t tmp;
1549 
1550 				tmp = (cp->cp_ecx >> (8 * i)) & 0xff;
1551 				if (tmp == 0x40)
1552 					celeron++;
1553 				else if (tmp >= 0x44 && tmp <= 0x45)
1554 					xeon++;
1555 			}
1556 
1557 			for (i = 0; i < 4; i++) {
1558 				uint_t tmp;
1559 
1560 				tmp = (cp->cp_edx >> (8 * i)) & 0xff;
1561 				if (tmp == 0x40)
1562 					celeron++;
1563 				else if (tmp >= 0x44 && tmp <= 0x45)
1564 					xeon++;
1565 			}
1566 
1567 			if (celeron)
1568 				return ("Intel Celeron(r)");
1569 			if (xeon)
1570 				return (cpi->cpi_model == 5 ?
1571 				    "Intel Pentium(r) II Xeon(tm)" :
1572 				    "Intel Pentium(r) III Xeon(tm)");
1573 			return (cpi->cpi_model == 5 ?
1574 			    "Intel Pentium(r) II or Pentium(r) II Xeon(tm)" :
1575 			    "Intel Pentium(r) III or Pentium(r) III Xeon(tm)");
1576 		default:
1577 			break;
1578 		}
1579 	default:
1580 		break;
1581 	}
1582 
1583 	/* BrandID is present if the field is nonzero */
1584 	if (cpi->cpi_brandid != 0) {
1585 		static const struct {
1586 			uint_t bt_bid;
1587 			const char *bt_str;
1588 		} brand_tbl[] = {
1589 			{ 0x1,	"Intel(r) Celeron(r)" },
1590 			{ 0x2,	"Intel(r) Pentium(r) III" },
1591 			{ 0x3,	"Intel(r) Pentium(r) III Xeon(tm)" },
1592 			{ 0x4,	"Intel(r) Pentium(r) III" },
1593 			{ 0x6,	"Mobile Intel(r) Pentium(r) III" },
1594 			{ 0x7,	"Mobile Intel(r) Celeron(r)" },
1595 			{ 0x8,	"Intel(r) Pentium(r) 4" },
1596 			{ 0x9,	"Intel(r) Pentium(r) 4" },
1597 			{ 0xa,	"Intel(r) Celeron(r)" },
1598 			{ 0xb,	"Intel(r) Xeon(tm)" },
1599 			{ 0xc,	"Intel(r) Xeon(tm) MP" },
1600 			{ 0xe,	"Mobile Intel(r) Pentium(r) 4" },
1601 			{ 0xf,	"Mobile Intel(r) Celeron(r)" },
1602 			{ 0x11, "Mobile Genuine Intel(r)" },
1603 			{ 0x12, "Intel(r) Celeron(r) M" },
1604 			{ 0x13, "Mobile Intel(r) Celeron(r)" },
1605 			{ 0x14, "Intel(r) Celeron(r)" },
1606 			{ 0x15, "Mobile Genuine Intel(r)" },
1607 			{ 0x16,	"Intel(r) Pentium(r) M" },
1608 			{ 0x17, "Mobile Intel(r) Celeron(r)" }
1609 		};
1610 		uint_t btblmax = sizeof (brand_tbl) / sizeof (brand_tbl[0]);
1611 		uint_t sgn;
1612 
1613 		sgn = (cpi->cpi_family << 8) |
1614 		    (cpi->cpi_model << 4) | cpi->cpi_step;
1615 
1616 		for (i = 0; i < btblmax; i++)
1617 			if (brand_tbl[i].bt_bid == cpi->cpi_brandid)
1618 				break;
1619 		if (i < btblmax) {
1620 			if (sgn == 0x6b1 && cpi->cpi_brandid == 3)
1621 				return ("Intel(r) Celeron(r)");
1622 			if (sgn < 0xf13 && cpi->cpi_brandid == 0xb)
1623 				return ("Intel(r) Xeon(tm) MP");
1624 			if (sgn < 0xf13 && cpi->cpi_brandid == 0xe)
1625 				return ("Intel(r) Xeon(tm)");
1626 			return (brand_tbl[i].bt_str);
1627 		}
1628 	}
1629 
1630 	return (NULL);
1631 }
1632 
1633 static const char *
1634 amd_cpubrand(const struct cpuid_info *cpi)
1635 {
1636 	if ((x86_feature & X86_CPUID) == 0 ||
1637 	    cpi->cpi_maxeax < 1 || cpi->cpi_family < 5)
1638 		return ("i486 compatible");
1639 
1640 	switch (cpi->cpi_family) {
1641 	case 5:
1642 		switch (cpi->cpi_model) {
1643 		case 0:
1644 		case 1:
1645 		case 2:
1646 		case 3:
1647 		case 4:
1648 		case 5:
1649 			return ("AMD-K5(r)");
1650 		case 6:
1651 		case 7:
1652 			return ("AMD-K6(r)");
1653 		case 8:
1654 			return ("AMD-K6(r)-2");
1655 		case 9:
1656 			return ("AMD-K6(r)-III");
1657 		default:
1658 			return ("AMD (family 5)");
1659 		}
1660 	case 6:
1661 		switch (cpi->cpi_model) {
1662 		case 1:
1663 			return ("AMD-K7(tm)");
1664 		case 0:
1665 		case 2:
1666 		case 4:
1667 			return ("AMD Athlon(tm)");
1668 		case 3:
1669 		case 7:
1670 			return ("AMD Duron(tm)");
1671 		case 6:
1672 		case 8:
1673 		case 10:
1674 			/*
1675 			 * Use the L2 cache size to distinguish
1676 			 */
1677 			return ((cpi->cpi_extd[6].cp_ecx >> 16) >= 256 ?
1678 			    "AMD Athlon(tm)" : "AMD Duron(tm)");
1679 		default:
1680 			return ("AMD (family 6)");
1681 		}
1682 	default:
1683 		break;
1684 	}
1685 
1686 	if (cpi->cpi_family == 0xf && cpi->cpi_model == 5 &&
1687 	    cpi->cpi_brandid != 0) {
1688 		switch (BITX(cpi->cpi_brandid, 7, 5)) {
1689 		case 3:
1690 			return ("AMD Opteron(tm) UP 1xx");
1691 		case 4:
1692 			return ("AMD Opteron(tm) DP 2xx");
1693 		case 5:
1694 			return ("AMD Opteron(tm) MP 8xx");
1695 		default:
1696 			return ("AMD Opteron(tm)");
1697 		}
1698 	}
1699 
1700 	return (NULL);
1701 }
1702 
1703 static const char *
1704 cyrix_cpubrand(struct cpuid_info *cpi, uint_t type)
1705 {
1706 	if ((x86_feature & X86_CPUID) == 0 ||
1707 	    cpi->cpi_maxeax < 1 || cpi->cpi_family < 5 ||
1708 	    type == X86_TYPE_CYRIX_486)
1709 		return ("i486 compatible");
1710 
1711 	switch (type) {
1712 	case X86_TYPE_CYRIX_6x86:
1713 		return ("Cyrix 6x86");
1714 	case X86_TYPE_CYRIX_6x86L:
1715 		return ("Cyrix 6x86L");
1716 	case X86_TYPE_CYRIX_6x86MX:
1717 		return ("Cyrix 6x86MX");
1718 	case X86_TYPE_CYRIX_GXm:
1719 		return ("Cyrix GXm");
1720 	case X86_TYPE_CYRIX_MediaGX:
1721 		return ("Cyrix MediaGX");
1722 	case X86_TYPE_CYRIX_MII:
1723 		return ("Cyrix M2");
1724 	case X86_TYPE_VIA_CYRIX_III:
1725 		return ("VIA Cyrix M3");
1726 	default:
1727 		/*
1728 		 * Have another wild guess ..
1729 		 */
1730 		if (cpi->cpi_family == 4 && cpi->cpi_model == 9)
1731 			return ("Cyrix 5x86");
1732 		else if (cpi->cpi_family == 5) {
1733 			switch (cpi->cpi_model) {
1734 			case 2:
1735 				return ("Cyrix 6x86");	/* Cyrix M1 */
1736 			case 4:
1737 				return ("Cyrix MediaGX");
1738 			default:
1739 				break;
1740 			}
1741 		} else if (cpi->cpi_family == 6) {
1742 			switch (cpi->cpi_model) {
1743 			case 0:
1744 				return ("Cyrix 6x86MX"); /* Cyrix M2? */
1745 			case 5:
1746 			case 6:
1747 			case 7:
1748 			case 8:
1749 			case 9:
1750 				return ("VIA C3");
1751 			default:
1752 				break;
1753 			}
1754 		}
1755 		break;
1756 	}
1757 	return (NULL);
1758 }
1759 
1760 /*
1761  * This only gets called in the case that the CPU extended
1762  * feature brand string (0x80000002, 0x80000003, 0x80000004)
1763  * aren't available, or contain null bytes for some reason.
1764  */
1765 static void
1766 fabricate_brandstr(struct cpuid_info *cpi)
1767 {
1768 	const char *brand = NULL;
1769 
1770 	switch (cpi->cpi_vendor) {
1771 	case X86_VENDOR_Intel:
1772 		brand = intel_cpubrand(cpi);
1773 		break;
1774 	case X86_VENDOR_AMD:
1775 		brand = amd_cpubrand(cpi);
1776 		break;
1777 	case X86_VENDOR_Cyrix:
1778 		brand = cyrix_cpubrand(cpi, x86_type);
1779 		break;
1780 	case X86_VENDOR_NexGen:
1781 		if (cpi->cpi_family == 5 && cpi->cpi_model == 0)
1782 			brand = "NexGen Nx586";
1783 		break;
1784 	case X86_VENDOR_Centaur:
1785 		if (cpi->cpi_family == 5)
1786 			switch (cpi->cpi_model) {
1787 			case 4:
1788 				brand = "Centaur C6";
1789 				break;
1790 			case 8:
1791 				brand = "Centaur C2";
1792 				break;
1793 			case 9:
1794 				brand = "Centaur C3";
1795 				break;
1796 			default:
1797 				break;
1798 			}
1799 		break;
1800 	case X86_VENDOR_Rise:
1801 		if (cpi->cpi_family == 5 &&
1802 		    (cpi->cpi_model == 0 || cpi->cpi_model == 2))
1803 			brand = "Rise mP6";
1804 		break;
1805 	case X86_VENDOR_SiS:
1806 		if (cpi->cpi_family == 5 && cpi->cpi_model == 0)
1807 			brand = "SiS 55x";
1808 		break;
1809 	case X86_VENDOR_TM:
1810 		if (cpi->cpi_family == 5 && cpi->cpi_model == 4)
1811 			brand = "Transmeta Crusoe TM3x00 or TM5x00";
1812 		break;
1813 	case X86_VENDOR_NSC:
1814 	case X86_VENDOR_UMC:
1815 	default:
1816 		break;
1817 	}
1818 	if (brand) {
1819 		(void) strcpy((char *)cpi->cpi_brandstr, brand);
1820 		return;
1821 	}
1822 
1823 	/*
1824 	 * If all else fails ...
1825 	 */
1826 	(void) snprintf(cpi->cpi_brandstr, sizeof (cpi->cpi_brandstr),
1827 	    "%s %d.%d.%d", cpi->cpi_vendorstr, cpi->cpi_family,
1828 	    cpi->cpi_model, cpi->cpi_step);
1829 }
1830 
1831 /*
1832  * This routine is called just after kernel memory allocation
1833  * becomes available on cpu0, and as part of mp_startup() on
1834  * the other cpus.
1835  *
1836  * Fixup the brand string, and collect any information from cpuid
1837  * that requires dynamicically allocated storage to represent.
1838  */
1839 /*ARGSUSED*/
1840 void
1841 cpuid_pass3(cpu_t *cpu)
1842 {
1843 	int	i, max, shft, level, size;
1844 	struct cpuid_regs regs;
1845 	struct cpuid_regs *cp;
1846 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
1847 
1848 	ASSERT(cpi->cpi_pass == 2);
1849 
1850 	/*
1851 	 * Function 4: Deterministic cache parameters
1852 	 *
1853 	 * Take this opportunity to detect the number of threads
1854 	 * sharing the last level cache, and construct a corresponding
1855 	 * cache id. The respective cpuid_info members are initialized
1856 	 * to the default case of "no last level cache sharing".
1857 	 */
1858 	cpi->cpi_ncpu_shr_last_cache = 1;
1859 	cpi->cpi_last_lvl_cacheid = cpu->cpu_id;
1860 
1861 	if (cpi->cpi_maxeax >= 4 && cpi->cpi_vendor == X86_VENDOR_Intel) {
1862 
1863 		/*
1864 		 * Find the # of elements (size) returned by fn 4, and along
1865 		 * the way detect last level cache sharing details.
1866 		 */
1867 		bzero(&regs, sizeof (regs));
1868 		cp = &regs;
1869 		for (i = 0, max = 0; i < CPI_FN4_ECX_MAX; i++) {
1870 			cp->cp_eax = 4;
1871 			cp->cp_ecx = i;
1872 
1873 			(void) __cpuid_insn(cp);
1874 
1875 			if (CPI_CACHE_TYPE(cp) == 0)
1876 				break;
1877 			level = CPI_CACHE_LVL(cp);
1878 			if (level > max) {
1879 				max = level;
1880 				cpi->cpi_ncpu_shr_last_cache =
1881 				    CPI_NTHR_SHR_CACHE(cp) + 1;
1882 			}
1883 		}
1884 		cpi->cpi_std_4_size = size = i;
1885 
1886 		/*
1887 		 * Allocate the cpi_std_4 array. The first element
1888 		 * references the regs for fn 4, %ecx == 0, which
1889 		 * cpuid_pass2() stashed in cpi->cpi_std[4].
1890 		 */
1891 		if (size > 0) {
1892 			cpi->cpi_std_4 =
1893 			    kmem_alloc(size * sizeof (cp), KM_SLEEP);
1894 			cpi->cpi_std_4[0] = &cpi->cpi_std[4];
1895 
1896 			/*
1897 			 * Allocate storage to hold the additional regs
1898 			 * for function 4, %ecx == 1 .. cpi_std_4_size.
1899 			 *
1900 			 * The regs for fn 4, %ecx == 0 has already
1901 			 * been allocated as indicated above.
1902 			 */
1903 			for (i = 1; i < size; i++) {
1904 				cp = cpi->cpi_std_4[i] =
1905 				    kmem_zalloc(sizeof (regs), KM_SLEEP);
1906 				cp->cp_eax = 4;
1907 				cp->cp_ecx = i;
1908 
1909 				(void) __cpuid_insn(cp);
1910 			}
1911 		}
1912 		/*
1913 		 * Determine the number of bits needed to represent
1914 		 * the number of CPUs sharing the last level cache.
1915 		 *
1916 		 * Shift off that number of bits from the APIC id to
1917 		 * derive the cache id.
1918 		 */
1919 		shft = 0;
1920 		for (i = 1; i < cpi->cpi_ncpu_shr_last_cache; i <<= 1)
1921 			shft++;
1922 		cpi->cpi_last_lvl_cacheid = CPI_APIC_ID(cpi) >> shft;
1923 	}
1924 
1925 	/*
1926 	 * Now fixup the brand string
1927 	 */
1928 	if ((cpi->cpi_xmaxeax & 0x80000000) == 0) {
1929 		fabricate_brandstr(cpi);
1930 	} else {
1931 
1932 		/*
1933 		 * If we successfully extracted a brand string from the cpuid
1934 		 * instruction, clean it up by removing leading spaces and
1935 		 * similar junk.
1936 		 */
1937 		if (cpi->cpi_brandstr[0]) {
1938 			size_t maxlen = sizeof (cpi->cpi_brandstr);
1939 			char *src, *dst;
1940 
1941 			dst = src = (char *)cpi->cpi_brandstr;
1942 			src[maxlen - 1] = '\0';
1943 			/*
1944 			 * strip leading spaces
1945 			 */
1946 			while (*src == ' ')
1947 				src++;
1948 			/*
1949 			 * Remove any 'Genuine' or "Authentic" prefixes
1950 			 */
1951 			if (strncmp(src, "Genuine ", 8) == 0)
1952 				src += 8;
1953 			if (strncmp(src, "Authentic ", 10) == 0)
1954 				src += 10;
1955 
1956 			/*
1957 			 * Now do an in-place copy.
1958 			 * Map (R) to (r) and (TM) to (tm).
1959 			 * The era of teletypes is long gone, and there's
1960 			 * -really- no need to shout.
1961 			 */
1962 			while (*src != '\0') {
1963 				if (src[0] == '(') {
1964 					if (strncmp(src + 1, "R)", 2) == 0) {
1965 						(void) strncpy(dst, "(r)", 3);
1966 						src += 3;
1967 						dst += 3;
1968 						continue;
1969 					}
1970 					if (strncmp(src + 1, "TM)", 3) == 0) {
1971 						(void) strncpy(dst, "(tm)", 4);
1972 						src += 4;
1973 						dst += 4;
1974 						continue;
1975 					}
1976 				}
1977 				*dst++ = *src++;
1978 			}
1979 			*dst = '\0';
1980 
1981 			/*
1982 			 * Finally, remove any trailing spaces
1983 			 */
1984 			while (--dst > cpi->cpi_brandstr)
1985 				if (*dst == ' ')
1986 					*dst = '\0';
1987 				else
1988 					break;
1989 		} else
1990 			fabricate_brandstr(cpi);
1991 	}
1992 	cpi->cpi_pass = 3;
1993 }
1994 
1995 /*
1996  * This routine is called out of bind_hwcap() much later in the life
1997  * of the kernel (post_startup()).  The job of this routine is to resolve
1998  * the hardware feature support and kernel support for those features into
1999  * what we're actually going to tell applications via the aux vector.
2000  */
2001 uint_t
2002 cpuid_pass4(cpu_t *cpu)
2003 {
2004 	struct cpuid_info *cpi;
2005 	uint_t hwcap_flags = 0;
2006 
2007 	if (cpu == NULL)
2008 		cpu = CPU;
2009 	cpi = cpu->cpu_m.mcpu_cpi;
2010 
2011 	ASSERT(cpi->cpi_pass == 3);
2012 
2013 	if (cpi->cpi_maxeax >= 1) {
2014 		uint32_t *edx = &cpi->cpi_support[STD_EDX_FEATURES];
2015 		uint32_t *ecx = &cpi->cpi_support[STD_ECX_FEATURES];
2016 
2017 		*edx = CPI_FEATURES_EDX(cpi);
2018 		*ecx = CPI_FEATURES_ECX(cpi);
2019 
2020 		/*
2021 		 * [these require explicit kernel support]
2022 		 */
2023 		if ((x86_feature & X86_SEP) == 0)
2024 			*edx &= ~CPUID_INTC_EDX_SEP;
2025 
2026 		if ((x86_feature & X86_SSE) == 0)
2027 			*edx &= ~(CPUID_INTC_EDX_FXSR|CPUID_INTC_EDX_SSE);
2028 		if ((x86_feature & X86_SSE2) == 0)
2029 			*edx &= ~CPUID_INTC_EDX_SSE2;
2030 
2031 		if ((x86_feature & X86_HTT) == 0)
2032 			*edx &= ~CPUID_INTC_EDX_HTT;
2033 
2034 		if ((x86_feature & X86_SSE3) == 0)
2035 			*ecx &= ~CPUID_INTC_ECX_SSE3;
2036 
2037 		if (cpi->cpi_vendor == X86_VENDOR_Intel) {
2038 			if ((x86_feature & X86_SSSE3) == 0)
2039 				*ecx &= ~CPUID_INTC_ECX_SSSE3;
2040 			if ((x86_feature & X86_SSE4_1) == 0)
2041 				*ecx &= ~CPUID_INTC_ECX_SSE4_1;
2042 			if ((x86_feature & X86_SSE4_2) == 0)
2043 				*ecx &= ~CPUID_INTC_ECX_SSE4_2;
2044 		}
2045 
2046 		/*
2047 		 * [no explicit support required beyond x87 fp context]
2048 		 */
2049 		if (!fpu_exists)
2050 			*edx &= ~(CPUID_INTC_EDX_FPU | CPUID_INTC_EDX_MMX);
2051 
2052 		/*
2053 		 * Now map the supported feature vector to things that we
2054 		 * think userland will care about.
2055 		 */
2056 		if (*edx & CPUID_INTC_EDX_SEP)
2057 			hwcap_flags |= AV_386_SEP;
2058 		if (*edx & CPUID_INTC_EDX_SSE)
2059 			hwcap_flags |= AV_386_FXSR | AV_386_SSE;
2060 		if (*edx & CPUID_INTC_EDX_SSE2)
2061 			hwcap_flags |= AV_386_SSE2;
2062 		if (*ecx & CPUID_INTC_ECX_SSE3)
2063 			hwcap_flags |= AV_386_SSE3;
2064 		if (cpi->cpi_vendor == X86_VENDOR_Intel) {
2065 			if (*ecx & CPUID_INTC_ECX_SSSE3)
2066 				hwcap_flags |= AV_386_SSSE3;
2067 			if (*ecx & CPUID_INTC_ECX_SSE4_1)
2068 				hwcap_flags |= AV_386_SSE4_1;
2069 			if (*ecx & CPUID_INTC_ECX_SSE4_2)
2070 				hwcap_flags |= AV_386_SSE4_2;
2071 		}
2072 		if (*ecx & CPUID_INTC_ECX_POPCNT)
2073 			hwcap_flags |= AV_386_POPCNT;
2074 		if (*edx & CPUID_INTC_EDX_FPU)
2075 			hwcap_flags |= AV_386_FPU;
2076 		if (*edx & CPUID_INTC_EDX_MMX)
2077 			hwcap_flags |= AV_386_MMX;
2078 
2079 		if (*edx & CPUID_INTC_EDX_TSC)
2080 			hwcap_flags |= AV_386_TSC;
2081 		if (*edx & CPUID_INTC_EDX_CX8)
2082 			hwcap_flags |= AV_386_CX8;
2083 		if (*edx & CPUID_INTC_EDX_CMOV)
2084 			hwcap_flags |= AV_386_CMOV;
2085 		if (*ecx & CPUID_INTC_ECX_MON)
2086 			hwcap_flags |= AV_386_MON;
2087 		if (*ecx & CPUID_INTC_ECX_CX16)
2088 			hwcap_flags |= AV_386_CX16;
2089 	}
2090 
2091 	if (x86_feature & X86_HTT)
2092 		hwcap_flags |= AV_386_PAUSE;
2093 
2094 	if (cpi->cpi_xmaxeax < 0x80000001)
2095 		goto pass4_done;
2096 
2097 	switch (cpi->cpi_vendor) {
2098 		struct cpuid_regs cp;
2099 		uint32_t *edx, *ecx;
2100 
2101 	case X86_VENDOR_Intel:
2102 		/*
2103 		 * Seems like Intel duplicated what we necessary
2104 		 * here to make the initial crop of 64-bit OS's work.
2105 		 * Hopefully, those are the only "extended" bits
2106 		 * they'll add.
2107 		 */
2108 		/*FALLTHROUGH*/
2109 
2110 	case X86_VENDOR_AMD:
2111 		edx = &cpi->cpi_support[AMD_EDX_FEATURES];
2112 		ecx = &cpi->cpi_support[AMD_ECX_FEATURES];
2113 
2114 		*edx = CPI_FEATURES_XTD_EDX(cpi);
2115 		*ecx = CPI_FEATURES_XTD_ECX(cpi);
2116 
2117 		/*
2118 		 * [these features require explicit kernel support]
2119 		 */
2120 		switch (cpi->cpi_vendor) {
2121 		case X86_VENDOR_Intel:
2122 			break;
2123 
2124 		case X86_VENDOR_AMD:
2125 			if ((x86_feature & X86_TSCP) == 0)
2126 				*edx &= ~CPUID_AMD_EDX_TSCP;
2127 			if ((x86_feature & X86_SSE4A) == 0)
2128 				*ecx &= ~CPUID_AMD_ECX_SSE4A;
2129 			break;
2130 
2131 		default:
2132 			break;
2133 		}
2134 
2135 		/*
2136 		 * [no explicit support required beyond
2137 		 * x87 fp context and exception handlers]
2138 		 */
2139 		if (!fpu_exists)
2140 			*edx &= ~(CPUID_AMD_EDX_MMXamd |
2141 			    CPUID_AMD_EDX_3DNow | CPUID_AMD_EDX_3DNowx);
2142 
2143 		if ((x86_feature & X86_NX) == 0)
2144 			*edx &= ~CPUID_AMD_EDX_NX;
2145 #if !defined(__amd64)
2146 		*edx &= ~CPUID_AMD_EDX_LM;
2147 #endif
2148 		/*
2149 		 * Now map the supported feature vector to
2150 		 * things that we think userland will care about.
2151 		 */
2152 #if defined(__amd64)
2153 		if (*edx & CPUID_AMD_EDX_SYSC)
2154 			hwcap_flags |= AV_386_AMD_SYSC;
2155 #endif
2156 		if (*edx & CPUID_AMD_EDX_MMXamd)
2157 			hwcap_flags |= AV_386_AMD_MMX;
2158 		if (*edx & CPUID_AMD_EDX_3DNow)
2159 			hwcap_flags |= AV_386_AMD_3DNow;
2160 		if (*edx & CPUID_AMD_EDX_3DNowx)
2161 			hwcap_flags |= AV_386_AMD_3DNowx;
2162 
2163 		switch (cpi->cpi_vendor) {
2164 		case X86_VENDOR_AMD:
2165 			if (*edx & CPUID_AMD_EDX_TSCP)
2166 				hwcap_flags |= AV_386_TSCP;
2167 			if (*ecx & CPUID_AMD_ECX_AHF64)
2168 				hwcap_flags |= AV_386_AHF;
2169 			if (*ecx & CPUID_AMD_ECX_SSE4A)
2170 				hwcap_flags |= AV_386_AMD_SSE4A;
2171 			if (*ecx & CPUID_AMD_ECX_LZCNT)
2172 				hwcap_flags |= AV_386_AMD_LZCNT;
2173 			break;
2174 
2175 		case X86_VENDOR_Intel:
2176 			/*
2177 			 * Aarrgh.
2178 			 * Intel uses a different bit in the same word.
2179 			 */
2180 			if (*ecx & CPUID_INTC_ECX_AHF64)
2181 				hwcap_flags |= AV_386_AHF;
2182 			break;
2183 
2184 		default:
2185 			break;
2186 		}
2187 		break;
2188 
2189 	case X86_VENDOR_TM:
2190 		cp.cp_eax = 0x80860001;
2191 		(void) __cpuid_insn(&cp);
2192 		cpi->cpi_support[TM_EDX_FEATURES] = cp.cp_edx;
2193 		break;
2194 
2195 	default:
2196 		break;
2197 	}
2198 
2199 pass4_done:
2200 	cpi->cpi_pass = 4;
2201 	return (hwcap_flags);
2202 }
2203 
2204 
2205 /*
2206  * Simulate the cpuid instruction using the data we previously
2207  * captured about this CPU.  We try our best to return the truth
2208  * about the hardware, independently of kernel support.
2209  */
2210 uint32_t
2211 cpuid_insn(cpu_t *cpu, struct cpuid_regs *cp)
2212 {
2213 	struct cpuid_info *cpi;
2214 	struct cpuid_regs *xcp;
2215 
2216 	if (cpu == NULL)
2217 		cpu = CPU;
2218 	cpi = cpu->cpu_m.mcpu_cpi;
2219 
2220 	ASSERT(cpuid_checkpass(cpu, 3));
2221 
2222 	/*
2223 	 * CPUID data is cached in two separate places: cpi_std for standard
2224 	 * CPUID functions, and cpi_extd for extended CPUID functions.
2225 	 */
2226 	if (cp->cp_eax <= cpi->cpi_maxeax && cp->cp_eax < NMAX_CPI_STD)
2227 		xcp = &cpi->cpi_std[cp->cp_eax];
2228 	else if (cp->cp_eax >= 0x80000000 && cp->cp_eax <= cpi->cpi_xmaxeax &&
2229 	    cp->cp_eax < 0x80000000 + NMAX_CPI_EXTD)
2230 		xcp = &cpi->cpi_extd[cp->cp_eax - 0x80000000];
2231 	else
2232 		/*
2233 		 * The caller is asking for data from an input parameter which
2234 		 * the kernel has not cached.  In this case we go fetch from
2235 		 * the hardware and return the data directly to the user.
2236 		 */
2237 		return (__cpuid_insn(cp));
2238 
2239 	cp->cp_eax = xcp->cp_eax;
2240 	cp->cp_ebx = xcp->cp_ebx;
2241 	cp->cp_ecx = xcp->cp_ecx;
2242 	cp->cp_edx = xcp->cp_edx;
2243 	return (cp->cp_eax);
2244 }
2245 
2246 int
2247 cpuid_checkpass(cpu_t *cpu, int pass)
2248 {
2249 	return (cpu != NULL && cpu->cpu_m.mcpu_cpi != NULL &&
2250 	    cpu->cpu_m.mcpu_cpi->cpi_pass >= pass);
2251 }
2252 
2253 int
2254 cpuid_getbrandstr(cpu_t *cpu, char *s, size_t n)
2255 {
2256 	ASSERT(cpuid_checkpass(cpu, 3));
2257 
2258 	return (snprintf(s, n, "%s", cpu->cpu_m.mcpu_cpi->cpi_brandstr));
2259 }
2260 
2261 int
2262 cpuid_is_cmt(cpu_t *cpu)
2263 {
2264 	if (cpu == NULL)
2265 		cpu = CPU;
2266 
2267 	ASSERT(cpuid_checkpass(cpu, 1));
2268 
2269 	return (cpu->cpu_m.mcpu_cpi->cpi_chipid >= 0);
2270 }
2271 
2272 /*
2273  * AMD and Intel both implement the 64-bit variant of the syscall
2274  * instruction (syscallq), so if there's -any- support for syscall,
2275  * cpuid currently says "yes, we support this".
2276  *
2277  * However, Intel decided to -not- implement the 32-bit variant of the
2278  * syscall instruction, so we provide a predicate to allow our caller
2279  * to test that subtlety here.
2280  *
2281  * XXPV	Currently, 32-bit syscall instructions don't work via the hypervisor,
2282  *	even in the case where the hardware would in fact support it.
2283  */
2284 /*ARGSUSED*/
2285 int
2286 cpuid_syscall32_insn(cpu_t *cpu)
2287 {
2288 	ASSERT(cpuid_checkpass((cpu == NULL ? CPU : cpu), 1));
2289 
2290 #if !defined(__xpv)
2291 	if (cpu == NULL)
2292 		cpu = CPU;
2293 
2294 	/*CSTYLED*/
2295 	{
2296 		struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2297 
2298 		if (cpi->cpi_vendor == X86_VENDOR_AMD &&
2299 		    cpi->cpi_xmaxeax >= 0x80000001 &&
2300 		    (CPI_FEATURES_XTD_EDX(cpi) & CPUID_AMD_EDX_SYSC))
2301 			return (1);
2302 	}
2303 #endif
2304 	return (0);
2305 }
2306 
2307 int
2308 cpuid_getidstr(cpu_t *cpu, char *s, size_t n)
2309 {
2310 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2311 
2312 	static const char fmt[] =
2313 	    "x86 (%s %X family %d model %d step %d clock %d MHz)";
2314 	static const char fmt_ht[] =
2315 	    "x86 (chipid 0x%x %s %X family %d model %d step %d clock %d MHz)";
2316 
2317 	ASSERT(cpuid_checkpass(cpu, 1));
2318 
2319 	if (cpuid_is_cmt(cpu))
2320 		return (snprintf(s, n, fmt_ht, cpi->cpi_chipid,
2321 		    cpi->cpi_vendorstr, cpi->cpi_std[1].cp_eax,
2322 		    cpi->cpi_family, cpi->cpi_model,
2323 		    cpi->cpi_step, cpu->cpu_type_info.pi_clock));
2324 	return (snprintf(s, n, fmt,
2325 	    cpi->cpi_vendorstr, cpi->cpi_std[1].cp_eax,
2326 	    cpi->cpi_family, cpi->cpi_model,
2327 	    cpi->cpi_step, cpu->cpu_type_info.pi_clock));
2328 }
2329 
2330 const char *
2331 cpuid_getvendorstr(cpu_t *cpu)
2332 {
2333 	ASSERT(cpuid_checkpass(cpu, 1));
2334 	return ((const char *)cpu->cpu_m.mcpu_cpi->cpi_vendorstr);
2335 }
2336 
2337 uint_t
2338 cpuid_getvendor(cpu_t *cpu)
2339 {
2340 	ASSERT(cpuid_checkpass(cpu, 1));
2341 	return (cpu->cpu_m.mcpu_cpi->cpi_vendor);
2342 }
2343 
2344 uint_t
2345 cpuid_getfamily(cpu_t *cpu)
2346 {
2347 	ASSERT(cpuid_checkpass(cpu, 1));
2348 	return (cpu->cpu_m.mcpu_cpi->cpi_family);
2349 }
2350 
2351 uint_t
2352 cpuid_getmodel(cpu_t *cpu)
2353 {
2354 	ASSERT(cpuid_checkpass(cpu, 1));
2355 	return (cpu->cpu_m.mcpu_cpi->cpi_model);
2356 }
2357 
2358 uint_t
2359 cpuid_get_ncpu_per_chip(cpu_t *cpu)
2360 {
2361 	ASSERT(cpuid_checkpass(cpu, 1));
2362 	return (cpu->cpu_m.mcpu_cpi->cpi_ncpu_per_chip);
2363 }
2364 
2365 uint_t
2366 cpuid_get_ncore_per_chip(cpu_t *cpu)
2367 {
2368 	ASSERT(cpuid_checkpass(cpu, 1));
2369 	return (cpu->cpu_m.mcpu_cpi->cpi_ncore_per_chip);
2370 }
2371 
2372 uint_t
2373 cpuid_get_ncpu_sharing_last_cache(cpu_t *cpu)
2374 {
2375 	ASSERT(cpuid_checkpass(cpu, 2));
2376 	return (cpu->cpu_m.mcpu_cpi->cpi_ncpu_shr_last_cache);
2377 }
2378 
2379 id_t
2380 cpuid_get_last_lvl_cacheid(cpu_t *cpu)
2381 {
2382 	ASSERT(cpuid_checkpass(cpu, 2));
2383 	return (cpu->cpu_m.mcpu_cpi->cpi_last_lvl_cacheid);
2384 }
2385 
2386 uint_t
2387 cpuid_getstep(cpu_t *cpu)
2388 {
2389 	ASSERT(cpuid_checkpass(cpu, 1));
2390 	return (cpu->cpu_m.mcpu_cpi->cpi_step);
2391 }
2392 
2393 uint_t
2394 cpuid_getsig(struct cpu *cpu)
2395 {
2396 	ASSERT(cpuid_checkpass(cpu, 1));
2397 	return (cpu->cpu_m.mcpu_cpi->cpi_std[1].cp_eax);
2398 }
2399 
2400 uint32_t
2401 cpuid_getchiprev(struct cpu *cpu)
2402 {
2403 	ASSERT(cpuid_checkpass(cpu, 1));
2404 	return (cpu->cpu_m.mcpu_cpi->cpi_chiprev);
2405 }
2406 
2407 const char *
2408 cpuid_getchiprevstr(struct cpu *cpu)
2409 {
2410 	ASSERT(cpuid_checkpass(cpu, 1));
2411 	return (cpu->cpu_m.mcpu_cpi->cpi_chiprevstr);
2412 }
2413 
2414 uint32_t
2415 cpuid_getsockettype(struct cpu *cpu)
2416 {
2417 	ASSERT(cpuid_checkpass(cpu, 1));
2418 	return (cpu->cpu_m.mcpu_cpi->cpi_socket);
2419 }
2420 
2421 int
2422 cpuid_get_chipid(cpu_t *cpu)
2423 {
2424 	ASSERT(cpuid_checkpass(cpu, 1));
2425 
2426 	if (cpuid_is_cmt(cpu))
2427 		return (cpu->cpu_m.mcpu_cpi->cpi_chipid);
2428 	return (cpu->cpu_id);
2429 }
2430 
2431 id_t
2432 cpuid_get_coreid(cpu_t *cpu)
2433 {
2434 	ASSERT(cpuid_checkpass(cpu, 1));
2435 	return (cpu->cpu_m.mcpu_cpi->cpi_coreid);
2436 }
2437 
2438 int
2439 cpuid_get_clogid(cpu_t *cpu)
2440 {
2441 	ASSERT(cpuid_checkpass(cpu, 1));
2442 	return (cpu->cpu_m.mcpu_cpi->cpi_clogid);
2443 }
2444 
2445 void
2446 cpuid_get_addrsize(cpu_t *cpu, uint_t *pabits, uint_t *vabits)
2447 {
2448 	struct cpuid_info *cpi;
2449 
2450 	if (cpu == NULL)
2451 		cpu = CPU;
2452 	cpi = cpu->cpu_m.mcpu_cpi;
2453 
2454 	ASSERT(cpuid_checkpass(cpu, 1));
2455 
2456 	if (pabits)
2457 		*pabits = cpi->cpi_pabits;
2458 	if (vabits)
2459 		*vabits = cpi->cpi_vabits;
2460 }
2461 
2462 /*
2463  * Returns the number of data TLB entries for a corresponding
2464  * pagesize.  If it can't be computed, or isn't known, the
2465  * routine returns zero.  If you ask about an architecturally
2466  * impossible pagesize, the routine will panic (so that the
2467  * hat implementor knows that things are inconsistent.)
2468  */
2469 uint_t
2470 cpuid_get_dtlb_nent(cpu_t *cpu, size_t pagesize)
2471 {
2472 	struct cpuid_info *cpi;
2473 	uint_t dtlb_nent = 0;
2474 
2475 	if (cpu == NULL)
2476 		cpu = CPU;
2477 	cpi = cpu->cpu_m.mcpu_cpi;
2478 
2479 	ASSERT(cpuid_checkpass(cpu, 1));
2480 
2481 	/*
2482 	 * Check the L2 TLB info
2483 	 */
2484 	if (cpi->cpi_xmaxeax >= 0x80000006) {
2485 		struct cpuid_regs *cp = &cpi->cpi_extd[6];
2486 
2487 		switch (pagesize) {
2488 
2489 		case 4 * 1024:
2490 			/*
2491 			 * All zero in the top 16 bits of the register
2492 			 * indicates a unified TLB. Size is in low 16 bits.
2493 			 */
2494 			if ((cp->cp_ebx & 0xffff0000) == 0)
2495 				dtlb_nent = cp->cp_ebx & 0x0000ffff;
2496 			else
2497 				dtlb_nent = BITX(cp->cp_ebx, 27, 16);
2498 			break;
2499 
2500 		case 2 * 1024 * 1024:
2501 			if ((cp->cp_eax & 0xffff0000) == 0)
2502 				dtlb_nent = cp->cp_eax & 0x0000ffff;
2503 			else
2504 				dtlb_nent = BITX(cp->cp_eax, 27, 16);
2505 			break;
2506 
2507 		default:
2508 			panic("unknown L2 pagesize");
2509 			/*NOTREACHED*/
2510 		}
2511 	}
2512 
2513 	if (dtlb_nent != 0)
2514 		return (dtlb_nent);
2515 
2516 	/*
2517 	 * No L2 TLB support for this size, try L1.
2518 	 */
2519 	if (cpi->cpi_xmaxeax >= 0x80000005) {
2520 		struct cpuid_regs *cp = &cpi->cpi_extd[5];
2521 
2522 		switch (pagesize) {
2523 		case 4 * 1024:
2524 			dtlb_nent = BITX(cp->cp_ebx, 23, 16);
2525 			break;
2526 		case 2 * 1024 * 1024:
2527 			dtlb_nent = BITX(cp->cp_eax, 23, 16);
2528 			break;
2529 		default:
2530 			panic("unknown L1 d-TLB pagesize");
2531 			/*NOTREACHED*/
2532 		}
2533 	}
2534 
2535 	return (dtlb_nent);
2536 }
2537 
2538 /*
2539  * Return 0 if the erratum is not present or not applicable, positive
2540  * if it is, and negative if the status of the erratum is unknown.
2541  *
2542  * See "Revision Guide for AMD Athlon(tm) 64 and AMD Opteron(tm)
2543  * Processors" #25759, Rev 3.57, August 2005
2544  */
2545 int
2546 cpuid_opteron_erratum(cpu_t *cpu, uint_t erratum)
2547 {
2548 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2549 	uint_t eax;
2550 
2551 	/*
2552 	 * Bail out if this CPU isn't an AMD CPU, or if it's
2553 	 * a legacy (32-bit) AMD CPU.
2554 	 */
2555 	if (cpi->cpi_vendor != X86_VENDOR_AMD ||
2556 	    cpi->cpi_family == 4 || cpi->cpi_family == 5 ||
2557 	    cpi->cpi_family == 6)
2558 
2559 		return (0);
2560 
2561 	eax = cpi->cpi_std[1].cp_eax;
2562 
2563 #define	SH_B0(eax)	(eax == 0xf40 || eax == 0xf50)
2564 #define	SH_B3(eax) 	(eax == 0xf51)
2565 #define	B(eax)		(SH_B0(eax) || SH_B3(eax))
2566 
2567 #define	SH_C0(eax)	(eax == 0xf48 || eax == 0xf58)
2568 
2569 #define	SH_CG(eax)	(eax == 0xf4a || eax == 0xf5a || eax == 0xf7a)
2570 #define	DH_CG(eax)	(eax == 0xfc0 || eax == 0xfe0 || eax == 0xff0)
2571 #define	CH_CG(eax)	(eax == 0xf82 || eax == 0xfb2)
2572 #define	CG(eax)		(SH_CG(eax) || DH_CG(eax) || CH_CG(eax))
2573 
2574 #define	SH_D0(eax)	(eax == 0x10f40 || eax == 0x10f50 || eax == 0x10f70)
2575 #define	DH_D0(eax)	(eax == 0x10fc0 || eax == 0x10ff0)
2576 #define	CH_D0(eax)	(eax == 0x10f80 || eax == 0x10fb0)
2577 #define	D0(eax)		(SH_D0(eax) || DH_D0(eax) || CH_D0(eax))
2578 
2579 #define	SH_E0(eax)	(eax == 0x20f50 || eax == 0x20f40 || eax == 0x20f70)
2580 #define	JH_E1(eax)	(eax == 0x20f10)	/* JH8_E0 had 0x20f30 */
2581 #define	DH_E3(eax)	(eax == 0x20fc0 || eax == 0x20ff0)
2582 #define	SH_E4(eax)	(eax == 0x20f51 || eax == 0x20f71)
2583 #define	BH_E4(eax)	(eax == 0x20fb1)
2584 #define	SH_E5(eax)	(eax == 0x20f42)
2585 #define	DH_E6(eax)	(eax == 0x20ff2 || eax == 0x20fc2)
2586 #define	JH_E6(eax)	(eax == 0x20f12 || eax == 0x20f32)
2587 #define	EX(eax)		(SH_E0(eax) || JH_E1(eax) || DH_E3(eax) || \
2588 			    SH_E4(eax) || BH_E4(eax) || SH_E5(eax) || \
2589 			    DH_E6(eax) || JH_E6(eax))
2590 
2591 	switch (erratum) {
2592 	case 1:
2593 		return (cpi->cpi_family < 0x10);
2594 	case 51:	/* what does the asterisk mean? */
2595 		return (B(eax) || SH_C0(eax) || CG(eax));
2596 	case 52:
2597 		return (B(eax));
2598 	case 57:
2599 		return (cpi->cpi_family <= 0x10);
2600 	case 58:
2601 		return (B(eax));
2602 	case 60:
2603 		return (cpi->cpi_family <= 0x10);
2604 	case 61:
2605 	case 62:
2606 	case 63:
2607 	case 64:
2608 	case 65:
2609 	case 66:
2610 	case 68:
2611 	case 69:
2612 	case 70:
2613 	case 71:
2614 		return (B(eax));
2615 	case 72:
2616 		return (SH_B0(eax));
2617 	case 74:
2618 		return (B(eax));
2619 	case 75:
2620 		return (cpi->cpi_family < 0x10);
2621 	case 76:
2622 		return (B(eax));
2623 	case 77:
2624 		return (cpi->cpi_family <= 0x10);
2625 	case 78:
2626 		return (B(eax) || SH_C0(eax));
2627 	case 79:
2628 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
2629 	case 80:
2630 	case 81:
2631 	case 82:
2632 		return (B(eax));
2633 	case 83:
2634 		return (B(eax) || SH_C0(eax) || CG(eax));
2635 	case 85:
2636 		return (cpi->cpi_family < 0x10);
2637 	case 86:
2638 		return (SH_C0(eax) || CG(eax));
2639 	case 88:
2640 #if !defined(__amd64)
2641 		return (0);
2642 #else
2643 		return (B(eax) || SH_C0(eax));
2644 #endif
2645 	case 89:
2646 		return (cpi->cpi_family < 0x10);
2647 	case 90:
2648 		return (B(eax) || SH_C0(eax) || CG(eax));
2649 	case 91:
2650 	case 92:
2651 		return (B(eax) || SH_C0(eax));
2652 	case 93:
2653 		return (SH_C0(eax));
2654 	case 94:
2655 		return (B(eax) || SH_C0(eax) || CG(eax));
2656 	case 95:
2657 #if !defined(__amd64)
2658 		return (0);
2659 #else
2660 		return (B(eax) || SH_C0(eax));
2661 #endif
2662 	case 96:
2663 		return (B(eax) || SH_C0(eax) || CG(eax));
2664 	case 97:
2665 	case 98:
2666 		return (SH_C0(eax) || CG(eax));
2667 	case 99:
2668 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
2669 	case 100:
2670 		return (B(eax) || SH_C0(eax));
2671 	case 101:
2672 	case 103:
2673 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
2674 	case 104:
2675 		return (SH_C0(eax) || CG(eax) || D0(eax));
2676 	case 105:
2677 	case 106:
2678 	case 107:
2679 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
2680 	case 108:
2681 		return (DH_CG(eax));
2682 	case 109:
2683 		return (SH_C0(eax) || CG(eax) || D0(eax));
2684 	case 110:
2685 		return (D0(eax) || EX(eax));
2686 	case 111:
2687 		return (CG(eax));
2688 	case 112:
2689 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
2690 	case 113:
2691 		return (eax == 0x20fc0);
2692 	case 114:
2693 		return (SH_E0(eax) || JH_E1(eax) || DH_E3(eax));
2694 	case 115:
2695 		return (SH_E0(eax) || JH_E1(eax));
2696 	case 116:
2697 		return (SH_E0(eax) || JH_E1(eax) || DH_E3(eax));
2698 	case 117:
2699 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
2700 	case 118:
2701 		return (SH_E0(eax) || JH_E1(eax) || SH_E4(eax) || BH_E4(eax) ||
2702 		    JH_E6(eax));
2703 	case 121:
2704 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
2705 	case 122:
2706 		return (cpi->cpi_family < 0x10);
2707 	case 123:
2708 		return (JH_E1(eax) || BH_E4(eax) || JH_E6(eax));
2709 	case 131:
2710 		return (cpi->cpi_family < 0x10);
2711 	case 6336786:
2712 		/*
2713 		 * Test for AdvPowerMgmtInfo.TscPStateInvariant
2714 		 * if this is a K8 family or newer processor
2715 		 */
2716 		if (CPI_FAMILY(cpi) == 0xf) {
2717 			struct cpuid_regs regs;
2718 			regs.cp_eax = 0x80000007;
2719 			(void) __cpuid_insn(&regs);
2720 			return (!(regs.cp_edx & 0x100));
2721 		}
2722 		return (0);
2723 	case 6323525:
2724 		return (((((eax >> 12) & 0xff00) + (eax & 0xf00)) |
2725 		    (((eax >> 4) & 0xf) | ((eax >> 12) & 0xf0))) < 0xf40);
2726 
2727 	default:
2728 		return (-1);
2729 	}
2730 }
2731 
2732 static const char assoc_str[] = "associativity";
2733 static const char line_str[] = "line-size";
2734 static const char size_str[] = "size";
2735 
2736 static void
2737 add_cache_prop(dev_info_t *devi, const char *label, const char *type,
2738     uint32_t val)
2739 {
2740 	char buf[128];
2741 
2742 	/*
2743 	 * ndi_prop_update_int() is used because it is desirable for
2744 	 * DDI_PROP_HW_DEF and DDI_PROP_DONTSLEEP to be set.
2745 	 */
2746 	if (snprintf(buf, sizeof (buf), "%s-%s", label, type) < sizeof (buf))
2747 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, devi, buf, val);
2748 }
2749 
2750 /*
2751  * Intel-style cache/tlb description
2752  *
2753  * Standard cpuid level 2 gives a randomly ordered
2754  * selection of tags that index into a table that describes
2755  * cache and tlb properties.
2756  */
2757 
2758 static const char l1_icache_str[] = "l1-icache";
2759 static const char l1_dcache_str[] = "l1-dcache";
2760 static const char l2_cache_str[] = "l2-cache";
2761 static const char l3_cache_str[] = "l3-cache";
2762 static const char itlb4k_str[] = "itlb-4K";
2763 static const char dtlb4k_str[] = "dtlb-4K";
2764 static const char itlb4M_str[] = "itlb-4M";
2765 static const char dtlb4M_str[] = "dtlb-4M";
2766 static const char itlb424_str[] = "itlb-4K-2M-4M";
2767 static const char dtlb44_str[] = "dtlb-4K-4M";
2768 static const char sl1_dcache_str[] = "sectored-l1-dcache";
2769 static const char sl2_cache_str[] = "sectored-l2-cache";
2770 static const char itrace_str[] = "itrace-cache";
2771 static const char sl3_cache_str[] = "sectored-l3-cache";
2772 
2773 static const struct cachetab {
2774 	uint8_t 	ct_code;
2775 	uint8_t		ct_assoc;
2776 	uint16_t 	ct_line_size;
2777 	size_t		ct_size;
2778 	const char	*ct_label;
2779 } intel_ctab[] = {
2780 	/* maintain descending order! */
2781 	{ 0xb4, 4, 0, 256, dtlb4k_str },
2782 	{ 0xb3, 4, 0, 128, dtlb4k_str },
2783 	{ 0xb0, 4, 0, 128, itlb4k_str },
2784 	{ 0x87, 8, 64, 1024*1024, l2_cache_str},
2785 	{ 0x86, 4, 64, 512*1024, l2_cache_str},
2786 	{ 0x85, 8, 32, 2*1024*1024, l2_cache_str},
2787 	{ 0x84, 8, 32, 1024*1024, l2_cache_str},
2788 	{ 0x83, 8, 32, 512*1024, l2_cache_str},
2789 	{ 0x82, 8, 32, 256*1024, l2_cache_str},
2790 	{ 0x7f, 2, 64, 512*1024, l2_cache_str},
2791 	{ 0x7d, 8, 64, 2*1024*1024, sl2_cache_str},
2792 	{ 0x7c, 8, 64, 1024*1024, sl2_cache_str},
2793 	{ 0x7b, 8, 64, 512*1024, sl2_cache_str},
2794 	{ 0x7a, 8, 64, 256*1024, sl2_cache_str},
2795 	{ 0x79, 8, 64, 128*1024, sl2_cache_str},
2796 	{ 0x78, 8, 64, 1024*1024, l2_cache_str},
2797 	{ 0x73, 8, 0, 64*1024, itrace_str},
2798 	{ 0x72, 8, 0, 32*1024, itrace_str},
2799 	{ 0x71, 8, 0, 16*1024, itrace_str},
2800 	{ 0x70, 8, 0, 12*1024, itrace_str},
2801 	{ 0x68, 4, 64, 32*1024, sl1_dcache_str},
2802 	{ 0x67, 4, 64, 16*1024, sl1_dcache_str},
2803 	{ 0x66, 4, 64, 8*1024, sl1_dcache_str},
2804 	{ 0x60, 8, 64, 16*1024, sl1_dcache_str},
2805 	{ 0x5d, 0, 0, 256, dtlb44_str},
2806 	{ 0x5c, 0, 0, 128, dtlb44_str},
2807 	{ 0x5b, 0, 0, 64, dtlb44_str},
2808 	{ 0x52, 0, 0, 256, itlb424_str},
2809 	{ 0x51, 0, 0, 128, itlb424_str},
2810 	{ 0x50, 0, 0, 64, itlb424_str},
2811 	{ 0x4d, 16, 64, 16*1024*1024, l3_cache_str},
2812 	{ 0x4c, 12, 64, 12*1024*1024, l3_cache_str},
2813 	{ 0x4b, 16, 64, 8*1024*1024, l3_cache_str},
2814 	{ 0x4a, 12, 64, 6*1024*1024, l3_cache_str},
2815 	{ 0x49, 16, 64, 4*1024*1024, l3_cache_str},
2816 	{ 0x47, 8, 64, 8*1024*1024, l3_cache_str},
2817 	{ 0x46, 4, 64, 4*1024*1024, l3_cache_str},
2818 	{ 0x45, 4, 32, 2*1024*1024, l2_cache_str},
2819 	{ 0x44, 4, 32, 1024*1024, l2_cache_str},
2820 	{ 0x43, 4, 32, 512*1024, l2_cache_str},
2821 	{ 0x42, 4, 32, 256*1024, l2_cache_str},
2822 	{ 0x41, 4, 32, 128*1024, l2_cache_str},
2823 	{ 0x3e, 4, 64, 512*1024, sl2_cache_str},
2824 	{ 0x3d, 6, 64, 384*1024, sl2_cache_str},
2825 	{ 0x3c, 4, 64, 256*1024, sl2_cache_str},
2826 	{ 0x3b, 2, 64, 128*1024, sl2_cache_str},
2827 	{ 0x3a, 6, 64, 192*1024, sl2_cache_str},
2828 	{ 0x39, 4, 64, 128*1024, sl2_cache_str},
2829 	{ 0x30, 8, 64, 32*1024, l1_icache_str},
2830 	{ 0x2c, 8, 64, 32*1024, l1_dcache_str},
2831 	{ 0x29, 8, 64, 4096*1024, sl3_cache_str},
2832 	{ 0x25, 8, 64, 2048*1024, sl3_cache_str},
2833 	{ 0x23, 8, 64, 1024*1024, sl3_cache_str},
2834 	{ 0x22, 4, 64, 512*1024, sl3_cache_str},
2835 	{ 0x0c, 4, 32, 16*1024, l1_dcache_str},
2836 	{ 0x0b, 4, 0, 4, itlb4M_str},
2837 	{ 0x0a, 2, 32, 8*1024, l1_dcache_str},
2838 	{ 0x08, 4, 32, 16*1024, l1_icache_str},
2839 	{ 0x06, 4, 32, 8*1024, l1_icache_str},
2840 	{ 0x04, 4, 0, 8, dtlb4M_str},
2841 	{ 0x03, 4, 0, 64, dtlb4k_str},
2842 	{ 0x02, 4, 0, 2, itlb4M_str},
2843 	{ 0x01, 4, 0, 32, itlb4k_str},
2844 	{ 0 }
2845 };
2846 
2847 static const struct cachetab cyrix_ctab[] = {
2848 	{ 0x70, 4, 0, 32, "tlb-4K" },
2849 	{ 0x80, 4, 16, 16*1024, "l1-cache" },
2850 	{ 0 }
2851 };
2852 
2853 /*
2854  * Search a cache table for a matching entry
2855  */
2856 static const struct cachetab *
2857 find_cacheent(const struct cachetab *ct, uint_t code)
2858 {
2859 	if (code != 0) {
2860 		for (; ct->ct_code != 0; ct++)
2861 			if (ct->ct_code <= code)
2862 				break;
2863 		if (ct->ct_code == code)
2864 			return (ct);
2865 	}
2866 	return (NULL);
2867 }
2868 
2869 /*
2870  * Walk the cacheinfo descriptor, applying 'func' to every valid element
2871  * The walk is terminated if the walker returns non-zero.
2872  */
2873 static void
2874 intel_walk_cacheinfo(struct cpuid_info *cpi,
2875     void *arg, int (*func)(void *, const struct cachetab *))
2876 {
2877 	const struct cachetab *ct;
2878 	uint8_t *dp;
2879 	int i;
2880 
2881 	if ((dp = cpi->cpi_cacheinfo) == NULL)
2882 		return;
2883 	for (i = 0; i < cpi->cpi_ncache; i++, dp++) {
2884 		/*
2885 		 * For overloaded descriptor 0x49 we use cpuid function 4
2886 		 * if supported by the current processor, to update
2887 		 * cache information.
2888 		 */
2889 		if (*dp == 0x49 && cpi->cpi_maxeax >= 0x4) {
2890 			intel_cpuid_4_cache_info(arg, cpi);
2891 			continue;
2892 		}
2893 
2894 		if ((ct = find_cacheent(intel_ctab, *dp)) != NULL) {
2895 			if (func(arg, ct) != 0)
2896 				break;
2897 		}
2898 	}
2899 }
2900 
2901 /*
2902  * (Like the Intel one, except for Cyrix CPUs)
2903  */
2904 static void
2905 cyrix_walk_cacheinfo(struct cpuid_info *cpi,
2906     void *arg, int (*func)(void *, const struct cachetab *))
2907 {
2908 	const struct cachetab *ct;
2909 	uint8_t *dp;
2910 	int i;
2911 
2912 	if ((dp = cpi->cpi_cacheinfo) == NULL)
2913 		return;
2914 	for (i = 0; i < cpi->cpi_ncache; i++, dp++) {
2915 		/*
2916 		 * Search Cyrix-specific descriptor table first ..
2917 		 */
2918 		if ((ct = find_cacheent(cyrix_ctab, *dp)) != NULL) {
2919 			if (func(arg, ct) != 0)
2920 				break;
2921 			continue;
2922 		}
2923 		/*
2924 		 * .. else fall back to the Intel one
2925 		 */
2926 		if ((ct = find_cacheent(intel_ctab, *dp)) != NULL) {
2927 			if (func(arg, ct) != 0)
2928 				break;
2929 			continue;
2930 		}
2931 	}
2932 }
2933 
2934 /*
2935  * A cacheinfo walker that adds associativity, line-size, and size properties
2936  * to the devinfo node it is passed as an argument.
2937  */
2938 static int
2939 add_cacheent_props(void *arg, const struct cachetab *ct)
2940 {
2941 	dev_info_t *devi = arg;
2942 
2943 	add_cache_prop(devi, ct->ct_label, assoc_str, ct->ct_assoc);
2944 	if (ct->ct_line_size != 0)
2945 		add_cache_prop(devi, ct->ct_label, line_str,
2946 		    ct->ct_line_size);
2947 	add_cache_prop(devi, ct->ct_label, size_str, ct->ct_size);
2948 	return (0);
2949 }
2950 
2951 /*
2952  * Add L2 or L3 cache-information using cpuid function 4. This
2953  * function is called from intel_walk_cacheinfo() when descriptor
2954  * 0x49 is encountered.
2955  */
2956 static void
2957 intel_cpuid_4_cache_info(void *arg, struct cpuid_info *cpi)
2958 {
2959 	uint32_t level, i;
2960 
2961 	struct cachetab ct;
2962 
2963 	for (i = 0; i < cpi->cpi_std_4_size; i++) {
2964 		level = CPI_CACHE_LVL(cpi->cpi_std_4[i]);
2965 
2966 		if (level == 2 || level == 3) {
2967 			ct.ct_assoc = CPI_CACHE_WAYS(cpi->cpi_std_4[i]) + 1;
2968 			ct.ct_line_size =
2969 			    CPI_CACHE_COH_LN_SZ(cpi->cpi_std_4[i]) + 1;
2970 			ct.ct_size = ct.ct_assoc *
2971 			    (CPI_CACHE_PARTS(cpi->cpi_std_4[i]) + 1) *
2972 			    ct.ct_line_size *
2973 			    (cpi->cpi_std_4[i]->cp_ecx + 1);
2974 
2975 			if (level == 2) {
2976 				ct.ct_label = l2_cache_str;
2977 			} else if (level == 3) {
2978 				ct.ct_label = l3_cache_str;
2979 			}
2980 
2981 			(void) add_cacheent_props(arg,
2982 			    (const struct cachetab *) (&ct));
2983 		}
2984 	}
2985 }
2986 
2987 static const char fully_assoc[] = "fully-associative?";
2988 
2989 /*
2990  * AMD style cache/tlb description
2991  *
2992  * Extended functions 5 and 6 directly describe properties of
2993  * tlbs and various cache levels.
2994  */
2995 static void
2996 add_amd_assoc(dev_info_t *devi, const char *label, uint_t assoc)
2997 {
2998 	switch (assoc) {
2999 	case 0:	/* reserved; ignore */
3000 		break;
3001 	default:
3002 		add_cache_prop(devi, label, assoc_str, assoc);
3003 		break;
3004 	case 0xff:
3005 		add_cache_prop(devi, label, fully_assoc, 1);
3006 		break;
3007 	}
3008 }
3009 
3010 static void
3011 add_amd_tlb(dev_info_t *devi, const char *label, uint_t assoc, uint_t size)
3012 {
3013 	if (size == 0)
3014 		return;
3015 	add_cache_prop(devi, label, size_str, size);
3016 	add_amd_assoc(devi, label, assoc);
3017 }
3018 
3019 static void
3020 add_amd_cache(dev_info_t *devi, const char *label,
3021     uint_t size, uint_t assoc, uint_t lines_per_tag, uint_t line_size)
3022 {
3023 	if (size == 0 || line_size == 0)
3024 		return;
3025 	add_amd_assoc(devi, label, assoc);
3026 	/*
3027 	 * Most AMD parts have a sectored cache. Multiple cache lines are
3028 	 * associated with each tag. A sector consists of all cache lines
3029 	 * associated with a tag. For example, the AMD K6-III has a sector
3030 	 * size of 2 cache lines per tag.
3031 	 */
3032 	if (lines_per_tag != 0)
3033 		add_cache_prop(devi, label, "lines-per-tag", lines_per_tag);
3034 	add_cache_prop(devi, label, line_str, line_size);
3035 	add_cache_prop(devi, label, size_str, size * 1024);
3036 }
3037 
3038 static void
3039 add_amd_l2_assoc(dev_info_t *devi, const char *label, uint_t assoc)
3040 {
3041 	switch (assoc) {
3042 	case 0:	/* off */
3043 		break;
3044 	case 1:
3045 	case 2:
3046 	case 4:
3047 		add_cache_prop(devi, label, assoc_str, assoc);
3048 		break;
3049 	case 6:
3050 		add_cache_prop(devi, label, assoc_str, 8);
3051 		break;
3052 	case 8:
3053 		add_cache_prop(devi, label, assoc_str, 16);
3054 		break;
3055 	case 0xf:
3056 		add_cache_prop(devi, label, fully_assoc, 1);
3057 		break;
3058 	default: /* reserved; ignore */
3059 		break;
3060 	}
3061 }
3062 
3063 static void
3064 add_amd_l2_tlb(dev_info_t *devi, const char *label, uint_t assoc, uint_t size)
3065 {
3066 	if (size == 0 || assoc == 0)
3067 		return;
3068 	add_amd_l2_assoc(devi, label, assoc);
3069 	add_cache_prop(devi, label, size_str, size);
3070 }
3071 
3072 static void
3073 add_amd_l2_cache(dev_info_t *devi, const char *label,
3074     uint_t size, uint_t assoc, uint_t lines_per_tag, uint_t line_size)
3075 {
3076 	if (size == 0 || assoc == 0 || line_size == 0)
3077 		return;
3078 	add_amd_l2_assoc(devi, label, assoc);
3079 	if (lines_per_tag != 0)
3080 		add_cache_prop(devi, label, "lines-per-tag", lines_per_tag);
3081 	add_cache_prop(devi, label, line_str, line_size);
3082 	add_cache_prop(devi, label, size_str, size * 1024);
3083 }
3084 
3085 static void
3086 amd_cache_info(struct cpuid_info *cpi, dev_info_t *devi)
3087 {
3088 	struct cpuid_regs *cp;
3089 
3090 	if (cpi->cpi_xmaxeax < 0x80000005)
3091 		return;
3092 	cp = &cpi->cpi_extd[5];
3093 
3094 	/*
3095 	 * 4M/2M L1 TLB configuration
3096 	 *
3097 	 * We report the size for 2M pages because AMD uses two
3098 	 * TLB entries for one 4M page.
3099 	 */
3100 	add_amd_tlb(devi, "dtlb-2M",
3101 	    BITX(cp->cp_eax, 31, 24), BITX(cp->cp_eax, 23, 16));
3102 	add_amd_tlb(devi, "itlb-2M",
3103 	    BITX(cp->cp_eax, 15, 8), BITX(cp->cp_eax, 7, 0));
3104 
3105 	/*
3106 	 * 4K L1 TLB configuration
3107 	 */
3108 
3109 	switch (cpi->cpi_vendor) {
3110 		uint_t nentries;
3111 	case X86_VENDOR_TM:
3112 		if (cpi->cpi_family >= 5) {
3113 			/*
3114 			 * Crusoe processors have 256 TLB entries, but
3115 			 * cpuid data format constrains them to only
3116 			 * reporting 255 of them.
3117 			 */
3118 			if ((nentries = BITX(cp->cp_ebx, 23, 16)) == 255)
3119 				nentries = 256;
3120 			/*
3121 			 * Crusoe processors also have a unified TLB
3122 			 */
3123 			add_amd_tlb(devi, "tlb-4K", BITX(cp->cp_ebx, 31, 24),
3124 			    nentries);
3125 			break;
3126 		}
3127 		/*FALLTHROUGH*/
3128 	default:
3129 		add_amd_tlb(devi, itlb4k_str,
3130 		    BITX(cp->cp_ebx, 31, 24), BITX(cp->cp_ebx, 23, 16));
3131 		add_amd_tlb(devi, dtlb4k_str,
3132 		    BITX(cp->cp_ebx, 15, 8), BITX(cp->cp_ebx, 7, 0));
3133 		break;
3134 	}
3135 
3136 	/*
3137 	 * data L1 cache configuration
3138 	 */
3139 
3140 	add_amd_cache(devi, l1_dcache_str,
3141 	    BITX(cp->cp_ecx, 31, 24), BITX(cp->cp_ecx, 23, 16),
3142 	    BITX(cp->cp_ecx, 15, 8), BITX(cp->cp_ecx, 7, 0));
3143 
3144 	/*
3145 	 * code L1 cache configuration
3146 	 */
3147 
3148 	add_amd_cache(devi, l1_icache_str,
3149 	    BITX(cp->cp_edx, 31, 24), BITX(cp->cp_edx, 23, 16),
3150 	    BITX(cp->cp_edx, 15, 8), BITX(cp->cp_edx, 7, 0));
3151 
3152 	if (cpi->cpi_xmaxeax < 0x80000006)
3153 		return;
3154 	cp = &cpi->cpi_extd[6];
3155 
3156 	/* Check for a unified L2 TLB for large pages */
3157 
3158 	if (BITX(cp->cp_eax, 31, 16) == 0)
3159 		add_amd_l2_tlb(devi, "l2-tlb-2M",
3160 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
3161 	else {
3162 		add_amd_l2_tlb(devi, "l2-dtlb-2M",
3163 		    BITX(cp->cp_eax, 31, 28), BITX(cp->cp_eax, 27, 16));
3164 		add_amd_l2_tlb(devi, "l2-itlb-2M",
3165 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
3166 	}
3167 
3168 	/* Check for a unified L2 TLB for 4K pages */
3169 
3170 	if (BITX(cp->cp_ebx, 31, 16) == 0) {
3171 		add_amd_l2_tlb(devi, "l2-tlb-4K",
3172 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
3173 	} else {
3174 		add_amd_l2_tlb(devi, "l2-dtlb-4K",
3175 		    BITX(cp->cp_eax, 31, 28), BITX(cp->cp_eax, 27, 16));
3176 		add_amd_l2_tlb(devi, "l2-itlb-4K",
3177 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
3178 	}
3179 
3180 	add_amd_l2_cache(devi, l2_cache_str,
3181 	    BITX(cp->cp_ecx, 31, 16), BITX(cp->cp_ecx, 15, 12),
3182 	    BITX(cp->cp_ecx, 11, 8), BITX(cp->cp_ecx, 7, 0));
3183 }
3184 
3185 /*
3186  * There are two basic ways that the x86 world describes it cache
3187  * and tlb architecture - Intel's way and AMD's way.
3188  *
3189  * Return which flavor of cache architecture we should use
3190  */
3191 static int
3192 x86_which_cacheinfo(struct cpuid_info *cpi)
3193 {
3194 	switch (cpi->cpi_vendor) {
3195 	case X86_VENDOR_Intel:
3196 		if (cpi->cpi_maxeax >= 2)
3197 			return (X86_VENDOR_Intel);
3198 		break;
3199 	case X86_VENDOR_AMD:
3200 		/*
3201 		 * The K5 model 1 was the first part from AMD that reported
3202 		 * cache sizes via extended cpuid functions.
3203 		 */
3204 		if (cpi->cpi_family > 5 ||
3205 		    (cpi->cpi_family == 5 && cpi->cpi_model >= 1))
3206 			return (X86_VENDOR_AMD);
3207 		break;
3208 	case X86_VENDOR_TM:
3209 		if (cpi->cpi_family >= 5)
3210 			return (X86_VENDOR_AMD);
3211 		/*FALLTHROUGH*/
3212 	default:
3213 		/*
3214 		 * If they have extended CPU data for 0x80000005
3215 		 * then we assume they have AMD-format cache
3216 		 * information.
3217 		 *
3218 		 * If not, and the vendor happens to be Cyrix,
3219 		 * then try our-Cyrix specific handler.
3220 		 *
3221 		 * If we're not Cyrix, then assume we're using Intel's
3222 		 * table-driven format instead.
3223 		 */
3224 		if (cpi->cpi_xmaxeax >= 0x80000005)
3225 			return (X86_VENDOR_AMD);
3226 		else if (cpi->cpi_vendor == X86_VENDOR_Cyrix)
3227 			return (X86_VENDOR_Cyrix);
3228 		else if (cpi->cpi_maxeax >= 2)
3229 			return (X86_VENDOR_Intel);
3230 		break;
3231 	}
3232 	return (-1);
3233 }
3234 
3235 /*
3236  * create a node for the given cpu under the prom root node.
3237  * Also, create a cpu node in the device tree.
3238  */
3239 static dev_info_t *cpu_nex_devi = NULL;
3240 static kmutex_t cpu_node_lock;
3241 
3242 /*
3243  * Called from post_startup() and mp_startup()
3244  */
3245 void
3246 add_cpunode2devtree(processorid_t cpu_id, struct cpuid_info *cpi)
3247 {
3248 	dev_info_t *cpu_devi;
3249 	int create;
3250 
3251 	mutex_enter(&cpu_node_lock);
3252 
3253 	/*
3254 	 * create a nexus node for all cpus identified as 'cpu_id' under
3255 	 * the root node.
3256 	 */
3257 	if (cpu_nex_devi == NULL) {
3258 		if (ndi_devi_alloc(ddi_root_node(), "cpus",
3259 		    (pnode_t)DEVI_SID_NODEID, &cpu_nex_devi) != NDI_SUCCESS) {
3260 			mutex_exit(&cpu_node_lock);
3261 			return;
3262 		}
3263 		(void) ndi_devi_online(cpu_nex_devi, 0);
3264 	}
3265 
3266 	/*
3267 	 * create a child node for cpu identified as 'cpu_id'
3268 	 */
3269 	cpu_devi = ddi_add_child(cpu_nex_devi, "cpu", DEVI_SID_NODEID,
3270 	    cpu_id);
3271 	if (cpu_devi == NULL) {
3272 		mutex_exit(&cpu_node_lock);
3273 		return;
3274 	}
3275 
3276 	/* device_type */
3277 
3278 	(void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi,
3279 	    "device_type", "cpu");
3280 
3281 	/* reg */
3282 
3283 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3284 	    "reg", cpu_id);
3285 
3286 	/* cpu-mhz, and clock-frequency */
3287 
3288 	if (cpu_freq > 0) {
3289 		long long mul;
3290 
3291 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3292 		    "cpu-mhz", cpu_freq);
3293 
3294 		if ((mul = cpu_freq * 1000000LL) <= INT_MAX)
3295 			(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3296 			    "clock-frequency", (int)mul);
3297 	}
3298 
3299 	(void) ndi_devi_online(cpu_devi, 0);
3300 
3301 	if ((x86_feature & X86_CPUID) == 0) {
3302 		mutex_exit(&cpu_node_lock);
3303 		return;
3304 	}
3305 
3306 	/* vendor-id */
3307 
3308 	(void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi,
3309 	    "vendor-id", cpi->cpi_vendorstr);
3310 
3311 	if (cpi->cpi_maxeax == 0) {
3312 		mutex_exit(&cpu_node_lock);
3313 		return;
3314 	}
3315 
3316 	/*
3317 	 * family, model, and step
3318 	 */
3319 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3320 	    "family", CPI_FAMILY(cpi));
3321 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3322 	    "cpu-model", CPI_MODEL(cpi));
3323 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3324 	    "stepping-id", CPI_STEP(cpi));
3325 
3326 	/* type */
3327 
3328 	switch (cpi->cpi_vendor) {
3329 	case X86_VENDOR_Intel:
3330 		create = 1;
3331 		break;
3332 	default:
3333 		create = 0;
3334 		break;
3335 	}
3336 	if (create)
3337 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3338 		    "type", CPI_TYPE(cpi));
3339 
3340 	/* ext-family */
3341 
3342 	switch (cpi->cpi_vendor) {
3343 	case X86_VENDOR_Intel:
3344 	case X86_VENDOR_AMD:
3345 		create = cpi->cpi_family >= 0xf;
3346 		break;
3347 	default:
3348 		create = 0;
3349 		break;
3350 	}
3351 	if (create)
3352 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3353 		    "ext-family", CPI_FAMILY_XTD(cpi));
3354 
3355 	/* ext-model */
3356 
3357 	switch (cpi->cpi_vendor) {
3358 	case X86_VENDOR_Intel:
3359 		create = CPI_MODEL(cpi) == 0xf;
3360 		break;
3361 	case X86_VENDOR_AMD:
3362 		create = CPI_FAMILY(cpi) == 0xf;
3363 		break;
3364 	default:
3365 		create = 0;
3366 		break;
3367 	}
3368 	if (create)
3369 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3370 		    "ext-model", CPI_MODEL_XTD(cpi));
3371 
3372 	/* generation */
3373 
3374 	switch (cpi->cpi_vendor) {
3375 	case X86_VENDOR_AMD:
3376 		/*
3377 		 * AMD K5 model 1 was the first part to support this
3378 		 */
3379 		create = cpi->cpi_xmaxeax >= 0x80000001;
3380 		break;
3381 	default:
3382 		create = 0;
3383 		break;
3384 	}
3385 	if (create)
3386 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3387 		    "generation", BITX((cpi)->cpi_extd[1].cp_eax, 11, 8));
3388 
3389 	/* brand-id */
3390 
3391 	switch (cpi->cpi_vendor) {
3392 	case X86_VENDOR_Intel:
3393 		/*
3394 		 * brand id first appeared on Pentium III Xeon model 8,
3395 		 * and Celeron model 8 processors and Opteron
3396 		 */
3397 		create = cpi->cpi_family > 6 ||
3398 		    (cpi->cpi_family == 6 && cpi->cpi_model >= 8);
3399 		break;
3400 	case X86_VENDOR_AMD:
3401 		create = cpi->cpi_family >= 0xf;
3402 		break;
3403 	default:
3404 		create = 0;
3405 		break;
3406 	}
3407 	if (create && cpi->cpi_brandid != 0) {
3408 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3409 		    "brand-id", cpi->cpi_brandid);
3410 	}
3411 
3412 	/* chunks, and apic-id */
3413 
3414 	switch (cpi->cpi_vendor) {
3415 		/*
3416 		 * first available on Pentium IV and Opteron (K8)
3417 		 */
3418 	case X86_VENDOR_Intel:
3419 		create = IS_NEW_F6(cpi) || cpi->cpi_family >= 0xf;
3420 		break;
3421 	case X86_VENDOR_AMD:
3422 		create = cpi->cpi_family >= 0xf;
3423 		break;
3424 	default:
3425 		create = 0;
3426 		break;
3427 	}
3428 	if (create) {
3429 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3430 		    "chunks", CPI_CHUNKS(cpi));
3431 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3432 		    "apic-id", CPI_APIC_ID(cpi));
3433 		if (cpi->cpi_chipid >= 0) {
3434 			(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3435 			    "chip#", cpi->cpi_chipid);
3436 			(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3437 			    "clog#", cpi->cpi_clogid);
3438 		}
3439 	}
3440 
3441 	/* cpuid-features */
3442 
3443 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3444 	    "cpuid-features", CPI_FEATURES_EDX(cpi));
3445 
3446 
3447 	/* cpuid-features-ecx */
3448 
3449 	switch (cpi->cpi_vendor) {
3450 	case X86_VENDOR_Intel:
3451 		create = IS_NEW_F6(cpi) || cpi->cpi_family >= 0xf;
3452 		break;
3453 	default:
3454 		create = 0;
3455 		break;
3456 	}
3457 	if (create)
3458 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3459 		    "cpuid-features-ecx", CPI_FEATURES_ECX(cpi));
3460 
3461 	/* ext-cpuid-features */
3462 
3463 	switch (cpi->cpi_vendor) {
3464 	case X86_VENDOR_Intel:
3465 	case X86_VENDOR_AMD:
3466 	case X86_VENDOR_Cyrix:
3467 	case X86_VENDOR_TM:
3468 	case X86_VENDOR_Centaur:
3469 		create = cpi->cpi_xmaxeax >= 0x80000001;
3470 		break;
3471 	default:
3472 		create = 0;
3473 		break;
3474 	}
3475 	if (create) {
3476 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3477 		    "ext-cpuid-features", CPI_FEATURES_XTD_EDX(cpi));
3478 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3479 		    "ext-cpuid-features-ecx", CPI_FEATURES_XTD_ECX(cpi));
3480 	}
3481 
3482 	/*
3483 	 * Brand String first appeared in Intel Pentium IV, AMD K5
3484 	 * model 1, and Cyrix GXm.  On earlier models we try and
3485 	 * simulate something similar .. so this string should always
3486 	 * same -something- about the processor, however lame.
3487 	 */
3488 	(void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi,
3489 	    "brand-string", cpi->cpi_brandstr);
3490 
3491 	/*
3492 	 * Finally, cache and tlb information
3493 	 */
3494 	switch (x86_which_cacheinfo(cpi)) {
3495 	case X86_VENDOR_Intel:
3496 		intel_walk_cacheinfo(cpi, cpu_devi, add_cacheent_props);
3497 		break;
3498 	case X86_VENDOR_Cyrix:
3499 		cyrix_walk_cacheinfo(cpi, cpu_devi, add_cacheent_props);
3500 		break;
3501 	case X86_VENDOR_AMD:
3502 		amd_cache_info(cpi, cpu_devi);
3503 		break;
3504 	default:
3505 		break;
3506 	}
3507 
3508 	mutex_exit(&cpu_node_lock);
3509 }
3510 
3511 struct l2info {
3512 	int *l2i_csz;
3513 	int *l2i_lsz;
3514 	int *l2i_assoc;
3515 	int l2i_ret;
3516 };
3517 
3518 /*
3519  * A cacheinfo walker that fetches the size, line-size and associativity
3520  * of the L2 cache
3521  */
3522 static int
3523 intel_l2cinfo(void *arg, const struct cachetab *ct)
3524 {
3525 	struct l2info *l2i = arg;
3526 	int *ip;
3527 
3528 	if (ct->ct_label != l2_cache_str &&
3529 	    ct->ct_label != sl2_cache_str)
3530 		return (0);	/* not an L2 -- keep walking */
3531 
3532 	if ((ip = l2i->l2i_csz) != NULL)
3533 		*ip = ct->ct_size;
3534 	if ((ip = l2i->l2i_lsz) != NULL)
3535 		*ip = ct->ct_line_size;
3536 	if ((ip = l2i->l2i_assoc) != NULL)
3537 		*ip = ct->ct_assoc;
3538 	l2i->l2i_ret = ct->ct_size;
3539 	return (1);		/* was an L2 -- terminate walk */
3540 }
3541 
3542 /*
3543  * AMD L2/L3 Cache and TLB Associativity Field Definition:
3544  *
3545  *	Unlike the associativity for the L1 cache and tlb where the 8 bit
3546  *	value is the associativity, the associativity for the L2 cache and
3547  *	tlb is encoded in the following table. The 4 bit L2 value serves as
3548  *	an index into the amd_afd[] array to determine the associativity.
3549  *	-1 is undefined. 0 is fully associative.
3550  */
3551 
3552 static int amd_afd[] =
3553 	{-1, 1, 2, -1, 4, -1, 8, -1, 16, -1, 32, 48, 64, 96, 128, 0};
3554 
3555 static void
3556 amd_l2cacheinfo(struct cpuid_info *cpi, struct l2info *l2i)
3557 {
3558 	struct cpuid_regs *cp;
3559 	uint_t size, assoc;
3560 	int i;
3561 	int *ip;
3562 
3563 	if (cpi->cpi_xmaxeax < 0x80000006)
3564 		return;
3565 	cp = &cpi->cpi_extd[6];
3566 
3567 	if ((i = BITX(cp->cp_ecx, 15, 12)) != 0 &&
3568 	    (size = BITX(cp->cp_ecx, 31, 16)) != 0) {
3569 		uint_t cachesz = size * 1024;
3570 		assoc = amd_afd[i];
3571 
3572 		ASSERT(assoc != -1);
3573 
3574 		if ((ip = l2i->l2i_csz) != NULL)
3575 			*ip = cachesz;
3576 		if ((ip = l2i->l2i_lsz) != NULL)
3577 			*ip = BITX(cp->cp_ecx, 7, 0);
3578 		if ((ip = l2i->l2i_assoc) != NULL)
3579 			*ip = assoc;
3580 		l2i->l2i_ret = cachesz;
3581 	}
3582 }
3583 
3584 int
3585 getl2cacheinfo(cpu_t *cpu, int *csz, int *lsz, int *assoc)
3586 {
3587 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
3588 	struct l2info __l2info, *l2i = &__l2info;
3589 
3590 	l2i->l2i_csz = csz;
3591 	l2i->l2i_lsz = lsz;
3592 	l2i->l2i_assoc = assoc;
3593 	l2i->l2i_ret = -1;
3594 
3595 	switch (x86_which_cacheinfo(cpi)) {
3596 	case X86_VENDOR_Intel:
3597 		intel_walk_cacheinfo(cpi, l2i, intel_l2cinfo);
3598 		break;
3599 	case X86_VENDOR_Cyrix:
3600 		cyrix_walk_cacheinfo(cpi, l2i, intel_l2cinfo);
3601 		break;
3602 	case X86_VENDOR_AMD:
3603 		amd_l2cacheinfo(cpi, l2i);
3604 		break;
3605 	default:
3606 		break;
3607 	}
3608 	return (l2i->l2i_ret);
3609 }
3610 
3611 #if !defined(__xpv)
3612 
3613 uint32_t *
3614 cpuid_mwait_alloc(cpu_t *cpu)
3615 {
3616 	uint32_t	*ret;
3617 	size_t		mwait_size;
3618 
3619 	ASSERT(cpuid_checkpass(cpu, 2));
3620 
3621 	mwait_size = cpu->cpu_m.mcpu_cpi->cpi_mwait.mon_max;
3622 	if (mwait_size == 0)
3623 		return (NULL);
3624 
3625 	/*
3626 	 * kmem_alloc() returns cache line size aligned data for mwait_size
3627 	 * allocations.  mwait_size is currently cache line sized.  Neither
3628 	 * of these implementation details are guarantied to be true in the
3629 	 * future.
3630 	 *
3631 	 * First try allocating mwait_size as kmem_alloc() currently returns
3632 	 * correctly aligned memory.  If kmem_alloc() does not return
3633 	 * mwait_size aligned memory, then use mwait_size ROUNDUP.
3634 	 *
3635 	 * Set cpi_mwait.buf_actual and cpi_mwait.size_actual in case we
3636 	 * decide to free this memory.
3637 	 */
3638 	ret = kmem_zalloc(mwait_size, KM_SLEEP);
3639 	if (ret == (uint32_t *)P2ROUNDUP((uintptr_t)ret, mwait_size)) {
3640 		cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual = ret;
3641 		cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual = mwait_size;
3642 		*ret = MWAIT_RUNNING;
3643 		return (ret);
3644 	} else {
3645 		kmem_free(ret, mwait_size);
3646 		ret = kmem_zalloc(mwait_size * 2, KM_SLEEP);
3647 		cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual = ret;
3648 		cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual = mwait_size * 2;
3649 		ret = (uint32_t *)P2ROUNDUP((uintptr_t)ret, mwait_size);
3650 		*ret = MWAIT_RUNNING;
3651 		return (ret);
3652 	}
3653 }
3654 
3655 void
3656 cpuid_mwait_free(cpu_t *cpu)
3657 {
3658 	ASSERT(cpuid_checkpass(cpu, 2));
3659 
3660 	if (cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual != NULL &&
3661 	    cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual > 0) {
3662 		kmem_free(cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual,
3663 		    cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual);
3664 	}
3665 
3666 	cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual = NULL;
3667 	cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual = 0;
3668 }
3669 
3670 void
3671 patch_tsc_read(int flag)
3672 {
3673 	size_t cnt;
3674 
3675 	switch (flag) {
3676 	case X86_NO_TSC:
3677 		cnt = &_no_rdtsc_end - &_no_rdtsc_start;
3678 		(void) memcpy((void *)tsc_read, (void *)&_no_rdtsc_start, cnt);
3679 		break;
3680 	case X86_HAVE_TSCP:
3681 		cnt = &_tscp_end - &_tscp_start;
3682 		(void) memcpy((void *)tsc_read, (void *)&_tscp_start, cnt);
3683 		break;
3684 	case X86_TSC_MFENCE:
3685 		cnt = &_tsc_mfence_end - &_tsc_mfence_start;
3686 		(void) memcpy((void *)tsc_read,
3687 		    (void *)&_tsc_mfence_start, cnt);
3688 		break;
3689 	default:
3690 		break;
3691 	}
3692 }
3693 
3694 #endif	/* !__xpv */
3695