xref: /titanic_52/usr/src/uts/i86pc/os/cpuid.c (revision e98fafb9956429b59c817d4fbd27720c73879203)
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/pg.h>
44 #include <sys/fp.h>
45 #include <sys/controlregs.h>
46 #include <sys/auxv_386.h>
47 #include <sys/bitmap.h>
48 #include <sys/memnode.h>
49 
50 /*
51  * Pass 0 of cpuid feature analysis happens in locore. It contains special code
52  * to recognize Cyrix processors that are not cpuid-compliant, and to deal with
53  * them accordingly. For most modern processors, feature detection occurs here
54  * in pass 1.
55  *
56  * Pass 1 of cpuid feature analysis happens just at the beginning of mlsetup()
57  * for the boot CPU and does the basic analysis that the early kernel needs.
58  * x86_feature is set based on the return value of cpuid_pass1() of the boot
59  * CPU.
60  *
61  * Pass 1 includes:
62  *
63  *	o Determining vendor/model/family/stepping and setting x86_type and
64  *	  x86_vendor accordingly.
65  *	o Processing the feature flags returned by the cpuid instruction while
66  *	  applying any workarounds or tricks for the specific processor.
67  *	o Mapping the feature flags into Solaris feature bits (X86_*).
68  *	o Processing extended feature flags if supported by the processor,
69  *	  again while applying specific processor knowledge.
70  *	o Determining the CMT characteristics of the system.
71  *
72  * Pass 1 is done on non-boot CPUs during their initialization and the results
73  * are used only as a meager attempt at ensuring that all processors within the
74  * system support the same features.
75  *
76  * Pass 2 of cpuid feature analysis happens just at the beginning
77  * of startup().  It just copies in and corrects the remainder
78  * of the cpuid data we depend on: standard cpuid functions that we didn't
79  * need for pass1 feature analysis, and extended cpuid functions beyond the
80  * simple feature processing done in pass1.
81  *
82  * Pass 3 of cpuid analysis is invoked after basic kernel services; in
83  * particular kernel memory allocation has been made available. It creates a
84  * readable brand string based on the data collected in the first two passes.
85  *
86  * Pass 4 of cpuid analysis is invoked after post_startup() when all
87  * the support infrastructure for various hardware features has been
88  * initialized. It determines which processor features will be reported
89  * to userland via the aux vector.
90  *
91  * All passes are executed on all CPUs, but only the boot CPU determines what
92  * features the kernel will use.
93  *
94  * Much of the worst junk in this file is for the support of processors
95  * that didn't really implement the cpuid instruction properly.
96  *
97  * NOTE: The accessor functions (cpuid_get*) are aware of, and ASSERT upon,
98  * the pass numbers.  Accordingly, changes to the pass code may require changes
99  * to the accessor code.
100  */
101 
102 uint_t x86_feature = 0;
103 uint_t x86_vendor = X86_VENDOR_IntelClone;
104 uint_t x86_type = X86_TYPE_OTHER;
105 
106 uint_t pentiumpro_bug4046376;
107 uint_t pentiumpro_bug4064495;
108 
109 uint_t enable486;
110 
111 /*
112  * This set of strings are for processors rumored to support the cpuid
113  * instruction, and is used by locore.s to figure out how to set x86_vendor
114  */
115 const char CyrixInstead[] = "CyrixInstead";
116 
117 /*
118  * monitor/mwait info.
119  */
120 struct mwait_info {
121 	size_t		mon_min;	/* min size to avoid missed wakeups */
122 	size_t		mon_max;	/* size to avoid false wakeups */
123 	uint32_t	support;	/* processor support of monitor/mwait */
124 };
125 
126 /*
127  * These constants determine how many of the elements of the
128  * cpuid we cache in the cpuid_info data structure; the
129  * remaining elements are accessible via the cpuid instruction.
130  */
131 
132 #define	NMAX_CPI_STD	6		/* eax = 0 .. 5 */
133 #define	NMAX_CPI_EXTD	9		/* eax = 0x80000000 .. 0x80000008 */
134 
135 struct cpuid_info {
136 	uint_t cpi_pass;		/* last pass completed */
137 	/*
138 	 * standard function information
139 	 */
140 	uint_t cpi_maxeax;		/* fn 0: %eax */
141 	char cpi_vendorstr[13];		/* fn 0: %ebx:%ecx:%edx */
142 	uint_t cpi_vendor;		/* enum of cpi_vendorstr */
143 
144 	uint_t cpi_family;		/* fn 1: extended family */
145 	uint_t cpi_model;		/* fn 1: extended model */
146 	uint_t cpi_step;		/* fn 1: stepping */
147 	chipid_t cpi_chipid;		/* fn 1: %ebx: chip # on ht cpus */
148 	uint_t cpi_brandid;		/* fn 1: %ebx: brand ID */
149 	int cpi_clogid;			/* fn 1: %ebx: thread # */
150 	uint_t cpi_ncpu_per_chip;	/* fn 1: %ebx: logical cpu count */
151 	uint8_t cpi_cacheinfo[16];	/* fn 2: intel-style cache desc */
152 	uint_t cpi_ncache;		/* fn 2: number of elements */
153 	uint_t cpi_ncpu_shr_last_cache;	/* fn 4: %eax: ncpus sharing cache */
154 	id_t cpi_last_lvl_cacheid;	/* fn 4: %eax: derived cache id */
155 	uint_t cpi_std_4_size;		/* fn 4: number of fn 4 elements */
156 	struct cpuid_regs **cpi_std_4;	/* fn 4: %ecx == 0 .. fn4_size */
157 	struct cpuid_regs cpi_std[NMAX_CPI_STD];	/* 0 .. 5 */
158 	/*
159 	 * extended function information
160 	 */
161 	uint_t cpi_xmaxeax;		/* fn 0x80000000: %eax */
162 	char cpi_brandstr[49];		/* fn 0x8000000[234] */
163 	uint8_t cpi_pabits;		/* fn 0x80000006: %eax */
164 	uint8_t cpi_vabits;		/* fn 0x80000006: %eax */
165 	struct cpuid_regs cpi_extd[NMAX_CPI_EXTD]; /* 0x8000000[0-8] */
166 	id_t cpi_coreid;
167 	uint_t cpi_ncore_per_chip;	/* AMD: fn 0x80000008: %ecx[7-0] */
168 					/* Intel: fn 4: %eax[31-26] */
169 	/*
170 	 * supported feature information
171 	 */
172 	uint32_t cpi_support[5];
173 #define	STD_EDX_FEATURES	0
174 #define	AMD_EDX_FEATURES	1
175 #define	TM_EDX_FEATURES		2
176 #define	STD_ECX_FEATURES	3
177 #define	AMD_ECX_FEATURES	4
178 	/*
179 	 * Synthesized information, where known.
180 	 */
181 	uint32_t cpi_chiprev;		/* See X86_CHIPREV_* in x86_archext.h */
182 	const char *cpi_chiprevstr;	/* May be NULL if chiprev unknown */
183 	uint32_t cpi_socket;		/* Chip package/socket type */
184 
185 	struct mwait_info cpi_mwait;	/* fn 5: monitor/mwait info */
186 };
187 
188 
189 static struct cpuid_info cpuid_info0;
190 
191 /*
192  * These bit fields are defined by the Intel Application Note AP-485
193  * "Intel Processor Identification and the CPUID Instruction"
194  */
195 #define	CPI_FAMILY_XTD(cpi)	BITX((cpi)->cpi_std[1].cp_eax, 27, 20)
196 #define	CPI_MODEL_XTD(cpi)	BITX((cpi)->cpi_std[1].cp_eax, 19, 16)
197 #define	CPI_TYPE(cpi)		BITX((cpi)->cpi_std[1].cp_eax, 13, 12)
198 #define	CPI_FAMILY(cpi)		BITX((cpi)->cpi_std[1].cp_eax, 11, 8)
199 #define	CPI_STEP(cpi)		BITX((cpi)->cpi_std[1].cp_eax, 3, 0)
200 #define	CPI_MODEL(cpi)		BITX((cpi)->cpi_std[1].cp_eax, 7, 4)
201 
202 #define	CPI_FEATURES_EDX(cpi)		((cpi)->cpi_std[1].cp_edx)
203 #define	CPI_FEATURES_ECX(cpi)		((cpi)->cpi_std[1].cp_ecx)
204 #define	CPI_FEATURES_XTD_EDX(cpi)	((cpi)->cpi_extd[1].cp_edx)
205 #define	CPI_FEATURES_XTD_ECX(cpi)	((cpi)->cpi_extd[1].cp_ecx)
206 
207 #define	CPI_BRANDID(cpi)	BITX((cpi)->cpi_std[1].cp_ebx, 7, 0)
208 #define	CPI_CHUNKS(cpi)		BITX((cpi)->cpi_std[1].cp_ebx, 15, 7)
209 #define	CPI_CPU_COUNT(cpi)	BITX((cpi)->cpi_std[1].cp_ebx, 23, 16)
210 #define	CPI_APIC_ID(cpi)	BITX((cpi)->cpi_std[1].cp_ebx, 31, 24)
211 
212 #define	CPI_MAXEAX_MAX		0x100		/* sanity control */
213 #define	CPI_XMAXEAX_MAX		0x80000100
214 #define	CPI_FN4_ECX_MAX		0x20		/* sanity: max fn 4 levels */
215 
216 /*
217  * Function 4 (Deterministic Cache Parameters) macros
218  * Defined by Intel Application Note AP-485
219  */
220 #define	CPI_NUM_CORES(regs)		BITX((regs)->cp_eax, 31, 26)
221 #define	CPI_NTHR_SHR_CACHE(regs)	BITX((regs)->cp_eax, 25, 14)
222 #define	CPI_FULL_ASSOC_CACHE(regs)	BITX((regs)->cp_eax, 9, 9)
223 #define	CPI_SELF_INIT_CACHE(regs)	BITX((regs)->cp_eax, 8, 8)
224 #define	CPI_CACHE_LVL(regs)		BITX((regs)->cp_eax, 7, 5)
225 #define	CPI_CACHE_TYPE(regs)		BITX((regs)->cp_eax, 4, 0)
226 
227 #define	CPI_CACHE_WAYS(regs)		BITX((regs)->cp_ebx, 31, 22)
228 #define	CPI_CACHE_PARTS(regs)		BITX((regs)->cp_ebx, 21, 12)
229 #define	CPI_CACHE_COH_LN_SZ(regs)	BITX((regs)->cp_ebx, 11, 0)
230 
231 #define	CPI_CACHE_SETS(regs)		BITX((regs)->cp_ecx, 31, 0)
232 
233 #define	CPI_PREFCH_STRIDE(regs)		BITX((regs)->cp_edx, 9, 0)
234 
235 
236 /*
237  * A couple of shorthand macros to identify "later" P6-family chips
238  * like the Pentium M and Core.  First, the "older" P6-based stuff
239  * (loosely defined as "pre-Pentium-4"):
240  * P6, PII, Mobile PII, PII Xeon, PIII, Mobile PIII, PIII Xeon
241  */
242 
243 #define	IS_LEGACY_P6(cpi) (			\
244 	cpi->cpi_family == 6 && 		\
245 		(cpi->cpi_model == 1 ||		\
246 		cpi->cpi_model == 3 ||		\
247 		cpi->cpi_model == 5 ||		\
248 		cpi->cpi_model == 6 ||		\
249 		cpi->cpi_model == 7 ||		\
250 		cpi->cpi_model == 8 ||		\
251 		cpi->cpi_model == 0xA ||	\
252 		cpi->cpi_model == 0xB)		\
253 )
254 
255 /* A "new F6" is everything with family 6 that's not the above */
256 #define	IS_NEW_F6(cpi) ((cpi->cpi_family == 6) && !IS_LEGACY_P6(cpi))
257 
258 /* Extended family/model support */
259 #define	IS_EXTENDED_MODEL_INTEL(cpi) (cpi->cpi_family == 0x6 || \
260 	cpi->cpi_family >= 0xf)
261 
262 /*
263  * AMD family 0xf socket types.
264  * First index is 0 for revs B thru E, 1 for F and G.
265  * Second index by (model & 0x3)
266  */
267 static uint32_t amd_skts[2][4] = {
268 	{
269 		X86_SOCKET_754,		/* 0b00 */
270 		X86_SOCKET_940,		/* 0b01 */
271 		X86_SOCKET_754,		/* 0b10 */
272 		X86_SOCKET_939		/* 0b11 */
273 	},
274 	{
275 		X86_SOCKET_S1g1,	/* 0b00 */
276 		X86_SOCKET_F1207,	/* 0b01 */
277 		X86_SOCKET_UNKNOWN,	/* 0b10 */
278 		X86_SOCKET_AM2		/* 0b11 */
279 	}
280 };
281 
282 /*
283  * Table for mapping AMD Family 0xf model/stepping combination to
284  * chip "revision" and socket type.  Only rm_family 0xf is used at the
285  * moment, but AMD family 0x10 will extend the exsiting revision names
286  * so will likely also use this table.
287  *
288  * The first member of this array that matches a given family, extended model
289  * plus model range, and stepping range will be considered a match.
290  */
291 static const struct amd_rev_mapent {
292 	uint_t rm_family;
293 	uint_t rm_modello;
294 	uint_t rm_modelhi;
295 	uint_t rm_steplo;
296 	uint_t rm_stephi;
297 	uint32_t rm_chiprev;
298 	const char *rm_chiprevstr;
299 	int rm_sktidx;
300 } amd_revmap[] = {
301 	/*
302 	 * Rev B includes model 0x4 stepping 0 and model 0x5 stepping 0 and 1.
303 	 */
304 	{ 0xf, 0x04, 0x04, 0x0, 0x0, X86_CHIPREV_AMD_F_REV_B, "B", 0 },
305 	{ 0xf, 0x05, 0x05, 0x0, 0x1, X86_CHIPREV_AMD_F_REV_B, "B", 0 },
306 	/*
307 	 * Rev C0 includes model 0x4 stepping 8 and model 0x5 stepping 8
308 	 */
309 	{ 0xf, 0x04, 0x05, 0x8, 0x8, X86_CHIPREV_AMD_F_REV_C0, "C0", 0 },
310 	/*
311 	 * Rev CG is the rest of extended model 0x0 - i.e., everything
312 	 * but the rev B and C0 combinations covered above.
313 	 */
314 	{ 0xf, 0x00, 0x0f, 0x0, 0xf, X86_CHIPREV_AMD_F_REV_CG, "CG", 0 },
315 	/*
316 	 * Rev D has extended model 0x1.
317 	 */
318 	{ 0xf, 0x10, 0x1f, 0x0, 0xf, X86_CHIPREV_AMD_F_REV_D, "D", 0 },
319 	/*
320 	 * Rev E has extended model 0x2.
321 	 * Extended model 0x3 is unused but available to grow into.
322 	 */
323 	{ 0xf, 0x20, 0x3f, 0x0, 0xf, X86_CHIPREV_AMD_F_REV_E, "E", 0 },
324 	/*
325 	 * Rev F has extended models 0x4 and 0x5.
326 	 */
327 	{ 0xf, 0x40, 0x5f, 0x0, 0xf, X86_CHIPREV_AMD_F_REV_F, "F", 1 },
328 	/*
329 	 * Rev G has extended model 0x6.
330 	 */
331 	{ 0xf, 0x60, 0x6f, 0x0, 0xf, X86_CHIPREV_AMD_F_REV_G, "G", 1 },
332 };
333 
334 /*
335  * Info for monitor/mwait idle loop.
336  *
337  * See cpuid section of "Intel 64 and IA-32 Architectures Software Developer's
338  * Manual Volume 2A: Instruction Set Reference, A-M" #25366-022US, November
339  * 2006.
340  * See MONITOR/MWAIT section of "AMD64 Architecture Programmer's Manual
341  * Documentation Updates" #33633, Rev 2.05, December 2006.
342  */
343 #define	MWAIT_SUPPORT		(0x00000001)	/* mwait supported */
344 #define	MWAIT_EXTENSIONS	(0x00000002)	/* extenstion supported */
345 #define	MWAIT_ECX_INT_ENABLE	(0x00000004)	/* ecx 1 extension supported */
346 #define	MWAIT_SUPPORTED(cpi)	((cpi)->cpi_std[1].cp_ecx & CPUID_INTC_ECX_MON)
347 #define	MWAIT_INT_ENABLE(cpi)	((cpi)->cpi_std[5].cp_ecx & 0x2)
348 #define	MWAIT_EXTENSION(cpi)	((cpi)->cpi_std[5].cp_ecx & 0x1)
349 #define	MWAIT_SIZE_MIN(cpi)	BITX((cpi)->cpi_std[5].cp_eax, 15, 0)
350 #define	MWAIT_SIZE_MAX(cpi)	BITX((cpi)->cpi_std[5].cp_ebx, 15, 0)
351 /*
352  * Number of sub-cstates for a given c-state.
353  */
354 #define	MWAIT_NUM_SUBC_STATES(cpi, c_state)			\
355 	BITX((cpi)->cpi_std[5].cp_edx, c_state + 3, c_state)
356 
357 static void intel_cpuid_4_cache_info(void *, struct cpuid_info *);
358 
359 static void
360 synth_amd_info(struct cpuid_info *cpi)
361 {
362 	const struct amd_rev_mapent *rmp;
363 	uint_t family, model, step;
364 	int i;
365 
366 	/*
367 	 * Currently only AMD family 0xf uses these fields.
368 	 */
369 	if (cpi->cpi_family != 0xf)
370 		return;
371 
372 	family = cpi->cpi_family;
373 	model = cpi->cpi_model;
374 	step = cpi->cpi_step;
375 
376 	for (i = 0, rmp = amd_revmap; i < sizeof (amd_revmap) / sizeof (*rmp);
377 	    i++, rmp++) {
378 		if (family == rmp->rm_family &&
379 		    model >= rmp->rm_modello && model <= rmp->rm_modelhi &&
380 		    step >= rmp->rm_steplo && step <= rmp->rm_stephi) {
381 			cpi->cpi_chiprev = rmp->rm_chiprev;
382 			cpi->cpi_chiprevstr = rmp->rm_chiprevstr;
383 			cpi->cpi_socket = amd_skts[rmp->rm_sktidx][model & 0x3];
384 			return;
385 		}
386 	}
387 }
388 
389 static void
390 synth_info(struct cpuid_info *cpi)
391 {
392 	cpi->cpi_chiprev = X86_CHIPREV_UNKNOWN;
393 	cpi->cpi_chiprevstr = "Unknown";
394 	cpi->cpi_socket = X86_SOCKET_UNKNOWN;
395 
396 	switch (cpi->cpi_vendor) {
397 	case X86_VENDOR_AMD:
398 		synth_amd_info(cpi);
399 		break;
400 
401 	default:
402 		break;
403 
404 	}
405 }
406 
407 /*
408  * Apply up various platform-dependent restrictions where the
409  * underlying platform restrictions mean the CPU can be marked
410  * as less capable than its cpuid instruction would imply.
411  */
412 
413 #define	platform_cpuid_mangle(vendor, eax, cp)	/* nothing */
414 
415 /*
416  *  Some undocumented ways of patching the results of the cpuid
417  *  instruction to permit running Solaris 10 on future cpus that
418  *  we don't currently support.  Could be set to non-zero values
419  *  via settings in eeprom.
420  */
421 
422 uint32_t cpuid_feature_ecx_include;
423 uint32_t cpuid_feature_ecx_exclude;
424 uint32_t cpuid_feature_edx_include;
425 uint32_t cpuid_feature_edx_exclude;
426 
427 void
428 cpuid_alloc_space(cpu_t *cpu)
429 {
430 	/*
431 	 * By convention, cpu0 is the boot cpu, which is set up
432 	 * before memory allocation is available.  All other cpus get
433 	 * their cpuid_info struct allocated here.
434 	 */
435 	ASSERT(cpu->cpu_id != 0);
436 	cpu->cpu_m.mcpu_cpi =
437 	    kmem_zalloc(sizeof (*cpu->cpu_m.mcpu_cpi), KM_SLEEP);
438 }
439 
440 void
441 cpuid_free_space(cpu_t *cpu)
442 {
443 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
444 	int i;
445 
446 	ASSERT(cpu->cpu_id != 0);
447 
448 	/*
449 	 * Free up any function 4 related dynamic storage
450 	 */
451 	for (i = 1; i < cpi->cpi_std_4_size; i++)
452 		kmem_free(cpi->cpi_std_4[i], sizeof (struct cpuid_regs));
453 	if (cpi->cpi_std_4_size > 0)
454 		kmem_free(cpi->cpi_std_4,
455 		    cpi->cpi_std_4_size * sizeof (struct cpuid_regs *));
456 
457 	kmem_free(cpu->cpu_m.mcpu_cpi, sizeof (*cpu->cpu_m.mcpu_cpi));
458 }
459 
460 uint_t
461 cpuid_pass1(cpu_t *cpu)
462 {
463 	uint32_t mask_ecx, mask_edx;
464 	uint_t feature = X86_CPUID;
465 	struct cpuid_info *cpi;
466 	struct cpuid_regs *cp;
467 	int xcpuid;
468 
469 
470 	/*
471 	 * Space statically allocated for cpu0, ensure pointer is set
472 	 */
473 	if (cpu->cpu_id == 0)
474 		cpu->cpu_m.mcpu_cpi = &cpuid_info0;
475 	cpi = cpu->cpu_m.mcpu_cpi;
476 	ASSERT(cpi != NULL);
477 	cp = &cpi->cpi_std[0];
478 	cp->cp_eax = 0;
479 	cpi->cpi_maxeax = __cpuid_insn(cp);
480 	{
481 		uint32_t *iptr = (uint32_t *)cpi->cpi_vendorstr;
482 		*iptr++ = cp->cp_ebx;
483 		*iptr++ = cp->cp_edx;
484 		*iptr++ = cp->cp_ecx;
485 		*(char *)&cpi->cpi_vendorstr[12] = '\0';
486 	}
487 
488 	/*
489 	 * Map the vendor string to a type code
490 	 */
491 	if (strcmp(cpi->cpi_vendorstr, "GenuineIntel") == 0)
492 		cpi->cpi_vendor = X86_VENDOR_Intel;
493 	else if (strcmp(cpi->cpi_vendorstr, "AuthenticAMD") == 0)
494 		cpi->cpi_vendor = X86_VENDOR_AMD;
495 	else if (strcmp(cpi->cpi_vendorstr, "GenuineTMx86") == 0)
496 		cpi->cpi_vendor = X86_VENDOR_TM;
497 	else if (strcmp(cpi->cpi_vendorstr, CyrixInstead) == 0)
498 		/*
499 		 * CyrixInstead is a variable used by the Cyrix detection code
500 		 * in locore.
501 		 */
502 		cpi->cpi_vendor = X86_VENDOR_Cyrix;
503 	else if (strcmp(cpi->cpi_vendorstr, "UMC UMC UMC ") == 0)
504 		cpi->cpi_vendor = X86_VENDOR_UMC;
505 	else if (strcmp(cpi->cpi_vendorstr, "NexGenDriven") == 0)
506 		cpi->cpi_vendor = X86_VENDOR_NexGen;
507 	else if (strcmp(cpi->cpi_vendorstr, "CentaurHauls") == 0)
508 		cpi->cpi_vendor = X86_VENDOR_Centaur;
509 	else if (strcmp(cpi->cpi_vendorstr, "RiseRiseRise") == 0)
510 		cpi->cpi_vendor = X86_VENDOR_Rise;
511 	else if (strcmp(cpi->cpi_vendorstr, "SiS SiS SiS ") == 0)
512 		cpi->cpi_vendor = X86_VENDOR_SiS;
513 	else if (strcmp(cpi->cpi_vendorstr, "Geode by NSC") == 0)
514 		cpi->cpi_vendor = X86_VENDOR_NSC;
515 	else
516 		cpi->cpi_vendor = X86_VENDOR_IntelClone;
517 
518 	x86_vendor = cpi->cpi_vendor; /* for compatibility */
519 
520 	/*
521 	 * Limit the range in case of weird hardware
522 	 */
523 	if (cpi->cpi_maxeax > CPI_MAXEAX_MAX)
524 		cpi->cpi_maxeax = CPI_MAXEAX_MAX;
525 	if (cpi->cpi_maxeax < 1)
526 		goto pass1_done;
527 
528 	cp = &cpi->cpi_std[1];
529 	cp->cp_eax = 1;
530 	(void) __cpuid_insn(cp);
531 
532 	/*
533 	 * Extract identifying constants for easy access.
534 	 */
535 	cpi->cpi_model = CPI_MODEL(cpi);
536 	cpi->cpi_family = CPI_FAMILY(cpi);
537 
538 	if (cpi->cpi_family == 0xf)
539 		cpi->cpi_family += CPI_FAMILY_XTD(cpi);
540 
541 	/*
542 	 * Beware: AMD uses "extended model" iff base *FAMILY* == 0xf.
543 	 * Intel, and presumably everyone else, uses model == 0xf, as
544 	 * one would expect (max value means possible overflow).  Sigh.
545 	 */
546 
547 	switch (cpi->cpi_vendor) {
548 	case X86_VENDOR_Intel:
549 		if (IS_EXTENDED_MODEL_INTEL(cpi))
550 			cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4;
551 		break;
552 	case X86_VENDOR_AMD:
553 		if (CPI_FAMILY(cpi) == 0xf)
554 			cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4;
555 		break;
556 	default:
557 		if (cpi->cpi_model == 0xf)
558 			cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4;
559 		break;
560 	}
561 
562 	cpi->cpi_step = CPI_STEP(cpi);
563 	cpi->cpi_brandid = CPI_BRANDID(cpi);
564 
565 	/*
566 	 * *default* assumptions:
567 	 * - believe %edx feature word
568 	 * - ignore %ecx feature word
569 	 * - 32-bit virtual and physical addressing
570 	 */
571 	mask_edx = 0xffffffff;
572 	mask_ecx = 0;
573 
574 	cpi->cpi_pabits = cpi->cpi_vabits = 32;
575 
576 	switch (cpi->cpi_vendor) {
577 	case X86_VENDOR_Intel:
578 		if (cpi->cpi_family == 5)
579 			x86_type = X86_TYPE_P5;
580 		else if (IS_LEGACY_P6(cpi)) {
581 			x86_type = X86_TYPE_P6;
582 			pentiumpro_bug4046376 = 1;
583 			pentiumpro_bug4064495 = 1;
584 			/*
585 			 * Clear the SEP bit when it was set erroneously
586 			 */
587 			if (cpi->cpi_model < 3 && cpi->cpi_step < 3)
588 				cp->cp_edx &= ~CPUID_INTC_EDX_SEP;
589 		} else if (IS_NEW_F6(cpi) || cpi->cpi_family == 0xf) {
590 			x86_type = X86_TYPE_P4;
591 			/*
592 			 * We don't currently depend on any of the %ecx
593 			 * features until Prescott, so we'll only check
594 			 * this from P4 onwards.  We might want to revisit
595 			 * that idea later.
596 			 */
597 			mask_ecx = 0xffffffff;
598 		} else if (cpi->cpi_family > 0xf)
599 			mask_ecx = 0xffffffff;
600 		/*
601 		 * We don't support MONITOR/MWAIT if leaf 5 is not available
602 		 * to obtain the monitor linesize.
603 		 */
604 		if (cpi->cpi_maxeax < 5)
605 			mask_ecx &= ~CPUID_INTC_ECX_MON;
606 		break;
607 	case X86_VENDOR_IntelClone:
608 	default:
609 		break;
610 	case X86_VENDOR_AMD:
611 #if defined(OPTERON_ERRATUM_108)
612 		if (cpi->cpi_family == 0xf && cpi->cpi_model == 0xe) {
613 			cp->cp_eax = (0xf0f & cp->cp_eax) | 0xc0;
614 			cpi->cpi_model = 0xc;
615 		} else
616 #endif
617 		if (cpi->cpi_family == 5) {
618 			/*
619 			 * AMD K5 and K6
620 			 *
621 			 * These CPUs have an incomplete implementation
622 			 * of MCA/MCE which we mask away.
623 			 */
624 			mask_edx &= ~(CPUID_INTC_EDX_MCE | CPUID_INTC_EDX_MCA);
625 
626 			/*
627 			 * Model 0 uses the wrong (APIC) bit
628 			 * to indicate PGE.  Fix it here.
629 			 */
630 			if (cpi->cpi_model == 0) {
631 				if (cp->cp_edx & 0x200) {
632 					cp->cp_edx &= ~0x200;
633 					cp->cp_edx |= CPUID_INTC_EDX_PGE;
634 				}
635 			}
636 
637 			/*
638 			 * Early models had problems w/ MMX; disable.
639 			 */
640 			if (cpi->cpi_model < 6)
641 				mask_edx &= ~CPUID_INTC_EDX_MMX;
642 		}
643 
644 		/*
645 		 * For newer families, SSE3 and CX16, at least, are valid;
646 		 * enable all
647 		 */
648 		if (cpi->cpi_family >= 0xf)
649 			mask_ecx = 0xffffffff;
650 		/*
651 		 * We don't support MONITOR/MWAIT if leaf 5 is not available
652 		 * to obtain the monitor linesize.
653 		 */
654 		if (cpi->cpi_maxeax < 5)
655 			mask_ecx &= ~CPUID_INTC_ECX_MON;
656 		break;
657 	case X86_VENDOR_TM:
658 		/*
659 		 * workaround the NT workaround in CMS 4.1
660 		 */
661 		if (cpi->cpi_family == 5 && cpi->cpi_model == 4 &&
662 		    (cpi->cpi_step == 2 || cpi->cpi_step == 3))
663 			cp->cp_edx |= CPUID_INTC_EDX_CX8;
664 		break;
665 	case X86_VENDOR_Centaur:
666 		/*
667 		 * workaround the NT workarounds again
668 		 */
669 		if (cpi->cpi_family == 6)
670 			cp->cp_edx |= CPUID_INTC_EDX_CX8;
671 		break;
672 	case X86_VENDOR_Cyrix:
673 		/*
674 		 * We rely heavily on the probing in locore
675 		 * to actually figure out what parts, if any,
676 		 * of the Cyrix cpuid instruction to believe.
677 		 */
678 		switch (x86_type) {
679 		case X86_TYPE_CYRIX_486:
680 			mask_edx = 0;
681 			break;
682 		case X86_TYPE_CYRIX_6x86:
683 			mask_edx = 0;
684 			break;
685 		case X86_TYPE_CYRIX_6x86L:
686 			mask_edx =
687 			    CPUID_INTC_EDX_DE |
688 			    CPUID_INTC_EDX_CX8;
689 			break;
690 		case X86_TYPE_CYRIX_6x86MX:
691 			mask_edx =
692 			    CPUID_INTC_EDX_DE |
693 			    CPUID_INTC_EDX_MSR |
694 			    CPUID_INTC_EDX_CX8 |
695 			    CPUID_INTC_EDX_PGE |
696 			    CPUID_INTC_EDX_CMOV |
697 			    CPUID_INTC_EDX_MMX;
698 			break;
699 		case X86_TYPE_CYRIX_GXm:
700 			mask_edx =
701 			    CPUID_INTC_EDX_MSR |
702 			    CPUID_INTC_EDX_CX8 |
703 			    CPUID_INTC_EDX_CMOV |
704 			    CPUID_INTC_EDX_MMX;
705 			break;
706 		case X86_TYPE_CYRIX_MediaGX:
707 			break;
708 		case X86_TYPE_CYRIX_MII:
709 		case X86_TYPE_VIA_CYRIX_III:
710 			mask_edx =
711 			    CPUID_INTC_EDX_DE |
712 			    CPUID_INTC_EDX_TSC |
713 			    CPUID_INTC_EDX_MSR |
714 			    CPUID_INTC_EDX_CX8 |
715 			    CPUID_INTC_EDX_PGE |
716 			    CPUID_INTC_EDX_CMOV |
717 			    CPUID_INTC_EDX_MMX;
718 			break;
719 		default:
720 			break;
721 		}
722 		break;
723 	}
724 
725 	/*
726 	 * Now we've figured out the masks that determine
727 	 * which bits we choose to believe, apply the masks
728 	 * to the feature words, then map the kernel's view
729 	 * of these feature words into its feature word.
730 	 */
731 	cp->cp_edx &= mask_edx;
732 	cp->cp_ecx &= mask_ecx;
733 
734 	/*
735 	 * apply any platform restrictions (we don't call this
736 	 * immediately after __cpuid_insn here, because we need the
737 	 * workarounds applied above first)
738 	 */
739 	platform_cpuid_mangle(cpi->cpi_vendor, 1, cp);
740 
741 	/*
742 	 * fold in overrides from the "eeprom" mechanism
743 	 */
744 	cp->cp_edx |= cpuid_feature_edx_include;
745 	cp->cp_edx &= ~cpuid_feature_edx_exclude;
746 
747 	cp->cp_ecx |= cpuid_feature_ecx_include;
748 	cp->cp_ecx &= ~cpuid_feature_ecx_exclude;
749 
750 	if (cp->cp_edx & CPUID_INTC_EDX_PSE)
751 		feature |= X86_LARGEPAGE;
752 	if (cp->cp_edx & CPUID_INTC_EDX_TSC)
753 		feature |= X86_TSC;
754 	if (cp->cp_edx & CPUID_INTC_EDX_MSR)
755 		feature |= X86_MSR;
756 	if (cp->cp_edx & CPUID_INTC_EDX_MTRR)
757 		feature |= X86_MTRR;
758 	if (cp->cp_edx & CPUID_INTC_EDX_PGE)
759 		feature |= X86_PGE;
760 	if (cp->cp_edx & CPUID_INTC_EDX_CMOV)
761 		feature |= X86_CMOV;
762 	if (cp->cp_edx & CPUID_INTC_EDX_MMX)
763 		feature |= X86_MMX;
764 	if ((cp->cp_edx & CPUID_INTC_EDX_MCE) != 0 &&
765 	    (cp->cp_edx & CPUID_INTC_EDX_MCA) != 0)
766 		feature |= X86_MCA;
767 	if (cp->cp_edx & CPUID_INTC_EDX_PAE)
768 		feature |= X86_PAE;
769 	if (cp->cp_edx & CPUID_INTC_EDX_CX8)
770 		feature |= X86_CX8;
771 	if (cp->cp_ecx & CPUID_INTC_ECX_CX16)
772 		feature |= X86_CX16;
773 	if (cp->cp_edx & CPUID_INTC_EDX_PAT)
774 		feature |= X86_PAT;
775 	if (cp->cp_edx & CPUID_INTC_EDX_SEP)
776 		feature |= X86_SEP;
777 	if (cp->cp_edx & CPUID_INTC_EDX_FXSR) {
778 		/*
779 		 * In our implementation, fxsave/fxrstor
780 		 * are prerequisites before we'll even
781 		 * try and do SSE things.
782 		 */
783 		if (cp->cp_edx & CPUID_INTC_EDX_SSE)
784 			feature |= X86_SSE;
785 		if (cp->cp_edx & CPUID_INTC_EDX_SSE2)
786 			feature |= X86_SSE2;
787 		if (cp->cp_ecx & CPUID_INTC_ECX_SSE3)
788 			feature |= X86_SSE3;
789 	}
790 	if (cp->cp_edx & CPUID_INTC_EDX_DE)
791 		feature |= X86_DE;
792 	if (cp->cp_ecx & CPUID_INTC_ECX_MON) {
793 		cpi->cpi_mwait.support |= MWAIT_SUPPORT;
794 		feature |= X86_MWAIT;
795 	}
796 
797 	if (feature & X86_PAE)
798 		cpi->cpi_pabits = 36;
799 
800 	/*
801 	 * Hyperthreading configuration is slightly tricky on Intel
802 	 * and pure clones, and even trickier on AMD.
803 	 *
804 	 * (AMD chose to set the HTT bit on their CMP processors,
805 	 * even though they're not actually hyperthreaded.  Thus it
806 	 * takes a bit more work to figure out what's really going
807 	 * on ... see the handling of the CMP_LGCY bit below)
808 	 */
809 	if (cp->cp_edx & CPUID_INTC_EDX_HTT) {
810 		cpi->cpi_ncpu_per_chip = CPI_CPU_COUNT(cpi);
811 		if (cpi->cpi_ncpu_per_chip > 1)
812 			feature |= X86_HTT;
813 	} else {
814 		cpi->cpi_ncpu_per_chip = 1;
815 	}
816 
817 	/*
818 	 * Work on the "extended" feature information, doing
819 	 * some basic initialization for cpuid_pass2()
820 	 */
821 	xcpuid = 0;
822 	switch (cpi->cpi_vendor) {
823 	case X86_VENDOR_Intel:
824 		if (IS_NEW_F6(cpi) || cpi->cpi_family >= 0xf)
825 			xcpuid++;
826 		break;
827 	case X86_VENDOR_AMD:
828 		if (cpi->cpi_family > 5 ||
829 		    (cpi->cpi_family == 5 && cpi->cpi_model >= 1))
830 			xcpuid++;
831 		break;
832 	case X86_VENDOR_Cyrix:
833 		/*
834 		 * Only these Cyrix CPUs are -known- to support
835 		 * extended cpuid operations.
836 		 */
837 		if (x86_type == X86_TYPE_VIA_CYRIX_III ||
838 		    x86_type == X86_TYPE_CYRIX_GXm)
839 			xcpuid++;
840 		break;
841 	case X86_VENDOR_Centaur:
842 	case X86_VENDOR_TM:
843 	default:
844 		xcpuid++;
845 		break;
846 	}
847 
848 	if (xcpuid) {
849 		cp = &cpi->cpi_extd[0];
850 		cp->cp_eax = 0x80000000;
851 		cpi->cpi_xmaxeax = __cpuid_insn(cp);
852 	}
853 
854 	if (cpi->cpi_xmaxeax & 0x80000000) {
855 
856 		if (cpi->cpi_xmaxeax > CPI_XMAXEAX_MAX)
857 			cpi->cpi_xmaxeax = CPI_XMAXEAX_MAX;
858 
859 		switch (cpi->cpi_vendor) {
860 		case X86_VENDOR_Intel:
861 		case X86_VENDOR_AMD:
862 			if (cpi->cpi_xmaxeax < 0x80000001)
863 				break;
864 			cp = &cpi->cpi_extd[1];
865 			cp->cp_eax = 0x80000001;
866 			(void) __cpuid_insn(cp);
867 
868 			if (cpi->cpi_vendor == X86_VENDOR_AMD &&
869 			    cpi->cpi_family == 5 &&
870 			    cpi->cpi_model == 6 &&
871 			    cpi->cpi_step == 6) {
872 				/*
873 				 * K6 model 6 uses bit 10 to indicate SYSC
874 				 * Later models use bit 11. Fix it here.
875 				 */
876 				if (cp->cp_edx & 0x400) {
877 					cp->cp_edx &= ~0x400;
878 					cp->cp_edx |= CPUID_AMD_EDX_SYSC;
879 				}
880 			}
881 
882 			platform_cpuid_mangle(cpi->cpi_vendor, 0x80000001, cp);
883 
884 			/*
885 			 * Compute the additions to the kernel's feature word.
886 			 */
887 			if (cp->cp_edx & CPUID_AMD_EDX_NX)
888 				feature |= X86_NX;
889 
890 			if ((cpi->cpi_vendor == X86_VENDOR_AMD) &&
891 			    (cpi->cpi_std[1].cp_edx & CPUID_INTC_EDX_FXSR) &&
892 			    (cp->cp_ecx & CPUID_AMD_ECX_SSE4A))
893 				feature |= X86_SSE4A;
894 
895 			/*
896 			 * If both the HTT and CMP_LGCY bits are set,
897 			 * then we're not actually HyperThreaded.  Read
898 			 * "AMD CPUID Specification" for more details.
899 			 */
900 			if (cpi->cpi_vendor == X86_VENDOR_AMD &&
901 			    (feature & X86_HTT) &&
902 			    (cp->cp_ecx & CPUID_AMD_ECX_CMP_LGCY)) {
903 				feature &= ~X86_HTT;
904 				feature |= X86_CMP;
905 			}
906 #if defined(__amd64)
907 			/*
908 			 * It's really tricky to support syscall/sysret in
909 			 * the i386 kernel; we rely on sysenter/sysexit
910 			 * instead.  In the amd64 kernel, things are -way-
911 			 * better.
912 			 */
913 			if (cp->cp_edx & CPUID_AMD_EDX_SYSC)
914 				feature |= X86_ASYSC;
915 
916 			/*
917 			 * While we're thinking about system calls, note
918 			 * that AMD processors don't support sysenter
919 			 * in long mode at all, so don't try to program them.
920 			 */
921 			if (x86_vendor == X86_VENDOR_AMD)
922 				feature &= ~X86_SEP;
923 #endif
924 			if (cp->cp_edx & CPUID_AMD_EDX_TSCP)
925 				feature |= X86_TSCP;
926 			break;
927 		default:
928 			break;
929 		}
930 
931 		/*
932 		 * Get CPUID data about processor cores and hyperthreads.
933 		 */
934 		switch (cpi->cpi_vendor) {
935 		case X86_VENDOR_Intel:
936 			if (cpi->cpi_maxeax >= 4) {
937 				cp = &cpi->cpi_std[4];
938 				cp->cp_eax = 4;
939 				cp->cp_ecx = 0;
940 				(void) __cpuid_insn(cp);
941 				platform_cpuid_mangle(cpi->cpi_vendor, 4, cp);
942 			}
943 			/*FALLTHROUGH*/
944 		case X86_VENDOR_AMD:
945 			if (cpi->cpi_xmaxeax < 0x80000008)
946 				break;
947 			cp = &cpi->cpi_extd[8];
948 			cp->cp_eax = 0x80000008;
949 			(void) __cpuid_insn(cp);
950 			platform_cpuid_mangle(cpi->cpi_vendor, 0x80000008, cp);
951 
952 			/*
953 			 * Virtual and physical address limits from
954 			 * cpuid override previously guessed values.
955 			 */
956 			cpi->cpi_pabits = BITX(cp->cp_eax, 7, 0);
957 			cpi->cpi_vabits = BITX(cp->cp_eax, 15, 8);
958 			break;
959 		default:
960 			break;
961 		}
962 
963 		/*
964 		 * Derive the number of cores per chip
965 		 */
966 		switch (cpi->cpi_vendor) {
967 		case X86_VENDOR_Intel:
968 			if (cpi->cpi_maxeax < 4) {
969 				cpi->cpi_ncore_per_chip = 1;
970 				break;
971 			} else {
972 				cpi->cpi_ncore_per_chip =
973 				    BITX((cpi)->cpi_std[4].cp_eax, 31, 26) + 1;
974 			}
975 			break;
976 		case X86_VENDOR_AMD:
977 			if (cpi->cpi_xmaxeax < 0x80000008) {
978 				cpi->cpi_ncore_per_chip = 1;
979 				break;
980 			} else {
981 				cpi->cpi_ncore_per_chip =
982 				    BITX((cpi)->cpi_extd[8].cp_ecx, 7, 0) + 1;
983 			}
984 			break;
985 		default:
986 			cpi->cpi_ncore_per_chip = 1;
987 			break;
988 		}
989 	}
990 
991 	/*
992 	 * If more than one core, then this processor is CMP.
993 	 */
994 	if (cpi->cpi_ncore_per_chip > 1)
995 		feature |= X86_CMP;
996 
997 	/*
998 	 * If the number of cores is the same as the number
999 	 * of CPUs, then we cannot have HyperThreading.
1000 	 */
1001 	if (cpi->cpi_ncpu_per_chip == cpi->cpi_ncore_per_chip)
1002 		feature &= ~X86_HTT;
1003 
1004 	if ((feature & (X86_HTT | X86_CMP)) == 0) {
1005 		/*
1006 		 * Single-core single-threaded processors.
1007 		 */
1008 		cpi->cpi_chipid = -1;
1009 		cpi->cpi_clogid = 0;
1010 		cpi->cpi_coreid = cpu->cpu_id;
1011 	} else if (cpi->cpi_ncpu_per_chip > 1) {
1012 		uint_t i;
1013 		uint_t chipid_shift = 0;
1014 		uint_t coreid_shift = 0;
1015 		uint_t apic_id = CPI_APIC_ID(cpi);
1016 
1017 		for (i = 1; i < cpi->cpi_ncpu_per_chip; i <<= 1)
1018 			chipid_shift++;
1019 		cpi->cpi_chipid = apic_id >> chipid_shift;
1020 		cpi->cpi_clogid = apic_id & ((1 << chipid_shift) - 1);
1021 
1022 		if (cpi->cpi_vendor == X86_VENDOR_Intel) {
1023 			if (feature & X86_CMP) {
1024 				/*
1025 				 * Multi-core (and possibly multi-threaded)
1026 				 * processors.
1027 				 */
1028 				uint_t ncpu_per_core;
1029 				if (cpi->cpi_ncore_per_chip == 1)
1030 					ncpu_per_core = cpi->cpi_ncpu_per_chip;
1031 				else if (cpi->cpi_ncore_per_chip > 1)
1032 					ncpu_per_core = cpi->cpi_ncpu_per_chip /
1033 					    cpi->cpi_ncore_per_chip;
1034 				/*
1035 				 * 8bit APIC IDs on dual core Pentiums
1036 				 * look like this:
1037 				 *
1038 				 * +-----------------------+------+------+
1039 				 * | Physical Package ID   |  MC  |  HT  |
1040 				 * +-----------------------+------+------+
1041 				 * <------- chipid -------->
1042 				 * <------- coreid --------------->
1043 				 *			   <--- clogid -->
1044 				 *
1045 				 * Where the number of bits necessary to
1046 				 * represent MC and HT fields together equals
1047 				 * to the minimum number of bits necessary to
1048 				 * store the value of cpi->cpi_ncpu_per_chip.
1049 				 * Of those bits, the MC part uses the number
1050 				 * of bits necessary to store the value of
1051 				 * cpi->cpi_ncore_per_chip.
1052 				 */
1053 				for (i = 1; i < ncpu_per_core; i <<= 1)
1054 					coreid_shift++;
1055 				cpi->cpi_coreid = apic_id >> coreid_shift;
1056 			} else if (feature & X86_HTT) {
1057 				/*
1058 				 * Single-core multi-threaded processors.
1059 				 */
1060 				cpi->cpi_coreid = cpi->cpi_chipid;
1061 			}
1062 		} else if (cpi->cpi_vendor == X86_VENDOR_AMD) {
1063 			/*
1064 			 * AMD currently only has dual-core processors with
1065 			 * single-threaded cores.  If they ever release
1066 			 * multi-threaded processors, then this code
1067 			 * will have to be updated.
1068 			 */
1069 			cpi->cpi_coreid = cpu->cpu_id;
1070 		} else {
1071 			/*
1072 			 * All other processors are currently
1073 			 * assumed to have single cores.
1074 			 */
1075 			cpi->cpi_coreid = cpi->cpi_chipid;
1076 		}
1077 	}
1078 
1079 	/*
1080 	 * Synthesize chip "revision" and socket type
1081 	 */
1082 	synth_info(cpi);
1083 
1084 pass1_done:
1085 	cpi->cpi_pass = 1;
1086 	return (feature);
1087 }
1088 
1089 /*
1090  * Make copies of the cpuid table entries we depend on, in
1091  * part for ease of parsing now, in part so that we have only
1092  * one place to correct any of it, in part for ease of
1093  * later export to userland, and in part so we can look at
1094  * this stuff in a crash dump.
1095  */
1096 
1097 /*ARGSUSED*/
1098 void
1099 cpuid_pass2(cpu_t *cpu)
1100 {
1101 	uint_t n, nmax;
1102 	int i;
1103 	struct cpuid_regs *cp;
1104 	uint8_t *dp;
1105 	uint32_t *iptr;
1106 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
1107 
1108 	ASSERT(cpi->cpi_pass == 1);
1109 
1110 	if (cpi->cpi_maxeax < 1)
1111 		goto pass2_done;
1112 
1113 	if ((nmax = cpi->cpi_maxeax + 1) > NMAX_CPI_STD)
1114 		nmax = NMAX_CPI_STD;
1115 	/*
1116 	 * (We already handled n == 0 and n == 1 in pass 1)
1117 	 */
1118 	for (n = 2, cp = &cpi->cpi_std[2]; n < nmax; n++, cp++) {
1119 		cp->cp_eax = n;
1120 
1121 		/*
1122 		 * CPUID function 4 expects %ecx to be initialized
1123 		 * with an index which indicates which cache to return
1124 		 * information about. The OS is expected to call function 4
1125 		 * with %ecx set to 0, 1, 2, ... until it returns with
1126 		 * EAX[4:0] set to 0, which indicates there are no more
1127 		 * caches.
1128 		 *
1129 		 * Here, populate cpi_std[4] with the information returned by
1130 		 * function 4 when %ecx == 0, and do the rest in cpuid_pass3()
1131 		 * when dynamic memory allocation becomes available.
1132 		 *
1133 		 * Note: we need to explicitly initialize %ecx here, since
1134 		 * function 4 may have been previously invoked.
1135 		 */
1136 		if (n == 4)
1137 			cp->cp_ecx = 0;
1138 
1139 		(void) __cpuid_insn(cp);
1140 		platform_cpuid_mangle(cpi->cpi_vendor, n, cp);
1141 		switch (n) {
1142 		case 2:
1143 			/*
1144 			 * "the lower 8 bits of the %eax register
1145 			 * contain a value that identifies the number
1146 			 * of times the cpuid [instruction] has to be
1147 			 * executed to obtain a complete image of the
1148 			 * processor's caching systems."
1149 			 *
1150 			 * How *do* they make this stuff up?
1151 			 */
1152 			cpi->cpi_ncache = sizeof (*cp) *
1153 			    BITX(cp->cp_eax, 7, 0);
1154 			if (cpi->cpi_ncache == 0)
1155 				break;
1156 			cpi->cpi_ncache--;	/* skip count byte */
1157 
1158 			/*
1159 			 * Well, for now, rather than attempt to implement
1160 			 * this slightly dubious algorithm, we just look
1161 			 * at the first 15 ..
1162 			 */
1163 			if (cpi->cpi_ncache > (sizeof (*cp) - 1))
1164 				cpi->cpi_ncache = sizeof (*cp) - 1;
1165 
1166 			dp = cpi->cpi_cacheinfo;
1167 			if (BITX(cp->cp_eax, 31, 31) == 0) {
1168 				uint8_t *p = (void *)&cp->cp_eax;
1169 				for (i = 1; i < 3; i++)
1170 					if (p[i] != 0)
1171 						*dp++ = p[i];
1172 			}
1173 			if (BITX(cp->cp_ebx, 31, 31) == 0) {
1174 				uint8_t *p = (void *)&cp->cp_ebx;
1175 				for (i = 0; i < 4; i++)
1176 					if (p[i] != 0)
1177 						*dp++ = p[i];
1178 			}
1179 			if (BITX(cp->cp_ecx, 31, 31) == 0) {
1180 				uint8_t *p = (void *)&cp->cp_ecx;
1181 				for (i = 0; i < 4; i++)
1182 					if (p[i] != 0)
1183 						*dp++ = p[i];
1184 			}
1185 			if (BITX(cp->cp_edx, 31, 31) == 0) {
1186 				uint8_t *p = (void *)&cp->cp_edx;
1187 				for (i = 0; i < 4; i++)
1188 					if (p[i] != 0)
1189 						*dp++ = p[i];
1190 			}
1191 			break;
1192 
1193 		case 3:	/* Processor serial number, if PSN supported */
1194 			break;
1195 
1196 		case 4:	/* Deterministic cache parameters */
1197 			break;
1198 
1199 		case 5:	/* Monitor/Mwait parameters */
1200 
1201 			/*
1202 			 * check cpi_mwait.support which was set in cpuid_pass1
1203 			 */
1204 			if (!(cpi->cpi_mwait.support & MWAIT_SUPPORT))
1205 				break;
1206 
1207 			cpi->cpi_mwait.mon_min = (size_t)MWAIT_SIZE_MIN(cpi);
1208 			cpi->cpi_mwait.mon_max = (size_t)MWAIT_SIZE_MAX(cpi);
1209 			if (MWAIT_EXTENSION(cpi)) {
1210 				cpi->cpi_mwait.support |= MWAIT_EXTENSIONS;
1211 				if (MWAIT_INT_ENABLE(cpi))
1212 					cpi->cpi_mwait.support |=
1213 					    MWAIT_ECX_INT_ENABLE;
1214 			}
1215 			break;
1216 		default:
1217 			break;
1218 		}
1219 	}
1220 
1221 	if ((cpi->cpi_xmaxeax & 0x80000000) == 0)
1222 		goto pass2_done;
1223 
1224 	if ((nmax = cpi->cpi_xmaxeax - 0x80000000 + 1) > NMAX_CPI_EXTD)
1225 		nmax = NMAX_CPI_EXTD;
1226 	/*
1227 	 * Copy the extended properties, fixing them as we go.
1228 	 * (We already handled n == 0 and n == 1 in pass 1)
1229 	 */
1230 	iptr = (void *)cpi->cpi_brandstr;
1231 	for (n = 2, cp = &cpi->cpi_extd[2]; n < nmax; cp++, n++) {
1232 		cp->cp_eax = 0x80000000 + n;
1233 		(void) __cpuid_insn(cp);
1234 		platform_cpuid_mangle(cpi->cpi_vendor, 0x80000000 + n, cp);
1235 		switch (n) {
1236 		case 2:
1237 		case 3:
1238 		case 4:
1239 			/*
1240 			 * Extract the brand string
1241 			 */
1242 			*iptr++ = cp->cp_eax;
1243 			*iptr++ = cp->cp_ebx;
1244 			*iptr++ = cp->cp_ecx;
1245 			*iptr++ = cp->cp_edx;
1246 			break;
1247 		case 5:
1248 			switch (cpi->cpi_vendor) {
1249 			case X86_VENDOR_AMD:
1250 				/*
1251 				 * The Athlon and Duron were the first
1252 				 * parts to report the sizes of the
1253 				 * TLB for large pages. Before then,
1254 				 * we don't trust the data.
1255 				 */
1256 				if (cpi->cpi_family < 6 ||
1257 				    (cpi->cpi_family == 6 &&
1258 				    cpi->cpi_model < 1))
1259 					cp->cp_eax = 0;
1260 				break;
1261 			default:
1262 				break;
1263 			}
1264 			break;
1265 		case 6:
1266 			switch (cpi->cpi_vendor) {
1267 			case X86_VENDOR_AMD:
1268 				/*
1269 				 * The Athlon and Duron were the first
1270 				 * AMD parts with L2 TLB's.
1271 				 * Before then, don't trust the data.
1272 				 */
1273 				if (cpi->cpi_family < 6 ||
1274 				    cpi->cpi_family == 6 &&
1275 				    cpi->cpi_model < 1)
1276 					cp->cp_eax = cp->cp_ebx = 0;
1277 				/*
1278 				 * AMD Duron rev A0 reports L2
1279 				 * cache size incorrectly as 1K
1280 				 * when it is really 64K
1281 				 */
1282 				if (cpi->cpi_family == 6 &&
1283 				    cpi->cpi_model == 3 &&
1284 				    cpi->cpi_step == 0) {
1285 					cp->cp_ecx &= 0xffff;
1286 					cp->cp_ecx |= 0x400000;
1287 				}
1288 				break;
1289 			case X86_VENDOR_Cyrix:	/* VIA C3 */
1290 				/*
1291 				 * VIA C3 processors are a bit messed
1292 				 * up w.r.t. encoding cache sizes in %ecx
1293 				 */
1294 				if (cpi->cpi_family != 6)
1295 					break;
1296 				/*
1297 				 * model 7 and 8 were incorrectly encoded
1298 				 *
1299 				 * xxx is model 8 really broken?
1300 				 */
1301 				if (cpi->cpi_model == 7 ||
1302 				    cpi->cpi_model == 8)
1303 					cp->cp_ecx =
1304 					    BITX(cp->cp_ecx, 31, 24) << 16 |
1305 					    BITX(cp->cp_ecx, 23, 16) << 12 |
1306 					    BITX(cp->cp_ecx, 15, 8) << 8 |
1307 					    BITX(cp->cp_ecx, 7, 0);
1308 				/*
1309 				 * model 9 stepping 1 has wrong associativity
1310 				 */
1311 				if (cpi->cpi_model == 9 && cpi->cpi_step == 1)
1312 					cp->cp_ecx |= 8 << 12;
1313 				break;
1314 			case X86_VENDOR_Intel:
1315 				/*
1316 				 * Extended L2 Cache features function.
1317 				 * First appeared on Prescott.
1318 				 */
1319 			default:
1320 				break;
1321 			}
1322 			break;
1323 		default:
1324 			break;
1325 		}
1326 	}
1327 
1328 pass2_done:
1329 	cpi->cpi_pass = 2;
1330 }
1331 
1332 static const char *
1333 intel_cpubrand(const struct cpuid_info *cpi)
1334 {
1335 	int i;
1336 
1337 	if ((x86_feature & X86_CPUID) == 0 ||
1338 	    cpi->cpi_maxeax < 1 || cpi->cpi_family < 5)
1339 		return ("i486");
1340 
1341 	switch (cpi->cpi_family) {
1342 	case 5:
1343 		return ("Intel Pentium(r)");
1344 	case 6:
1345 		switch (cpi->cpi_model) {
1346 			uint_t celeron, xeon;
1347 			const struct cpuid_regs *cp;
1348 		case 0:
1349 		case 1:
1350 		case 2:
1351 			return ("Intel Pentium(r) Pro");
1352 		case 3:
1353 		case 4:
1354 			return ("Intel Pentium(r) II");
1355 		case 6:
1356 			return ("Intel Celeron(r)");
1357 		case 5:
1358 		case 7:
1359 			celeron = xeon = 0;
1360 			cp = &cpi->cpi_std[2];	/* cache info */
1361 
1362 			for (i = 1; i < 3; i++) {
1363 				uint_t tmp;
1364 
1365 				tmp = (cp->cp_eax >> (8 * i)) & 0xff;
1366 				if (tmp == 0x40)
1367 					celeron++;
1368 				if (tmp >= 0x44 && tmp <= 0x45)
1369 					xeon++;
1370 			}
1371 
1372 			for (i = 0; i < 2; i++) {
1373 				uint_t tmp;
1374 
1375 				tmp = (cp->cp_ebx >> (8 * i)) & 0xff;
1376 				if (tmp == 0x40)
1377 					celeron++;
1378 				else if (tmp >= 0x44 && tmp <= 0x45)
1379 					xeon++;
1380 			}
1381 
1382 			for (i = 0; i < 4; i++) {
1383 				uint_t tmp;
1384 
1385 				tmp = (cp->cp_ecx >> (8 * i)) & 0xff;
1386 				if (tmp == 0x40)
1387 					celeron++;
1388 				else if (tmp >= 0x44 && tmp <= 0x45)
1389 					xeon++;
1390 			}
1391 
1392 			for (i = 0; i < 4; i++) {
1393 				uint_t tmp;
1394 
1395 				tmp = (cp->cp_edx >> (8 * i)) & 0xff;
1396 				if (tmp == 0x40)
1397 					celeron++;
1398 				else if (tmp >= 0x44 && tmp <= 0x45)
1399 					xeon++;
1400 			}
1401 
1402 			if (celeron)
1403 				return ("Intel Celeron(r)");
1404 			if (xeon)
1405 				return (cpi->cpi_model == 5 ?
1406 				    "Intel Pentium(r) II Xeon(tm)" :
1407 				    "Intel Pentium(r) III Xeon(tm)");
1408 			return (cpi->cpi_model == 5 ?
1409 			    "Intel Pentium(r) II or Pentium(r) II Xeon(tm)" :
1410 			    "Intel Pentium(r) III or Pentium(r) III Xeon(tm)");
1411 		default:
1412 			break;
1413 		}
1414 	default:
1415 		break;
1416 	}
1417 
1418 	/* BrandID is present if the field is nonzero */
1419 	if (cpi->cpi_brandid != 0) {
1420 		static const struct {
1421 			uint_t bt_bid;
1422 			const char *bt_str;
1423 		} brand_tbl[] = {
1424 			{ 0x1,	"Intel(r) Celeron(r)" },
1425 			{ 0x2,	"Intel(r) Pentium(r) III" },
1426 			{ 0x3,	"Intel(r) Pentium(r) III Xeon(tm)" },
1427 			{ 0x4,	"Intel(r) Pentium(r) III" },
1428 			{ 0x6,	"Mobile Intel(r) Pentium(r) III" },
1429 			{ 0x7,	"Mobile Intel(r) Celeron(r)" },
1430 			{ 0x8,	"Intel(r) Pentium(r) 4" },
1431 			{ 0x9,	"Intel(r) Pentium(r) 4" },
1432 			{ 0xa,	"Intel(r) Celeron(r)" },
1433 			{ 0xb,	"Intel(r) Xeon(tm)" },
1434 			{ 0xc,	"Intel(r) Xeon(tm) MP" },
1435 			{ 0xe,	"Mobile Intel(r) Pentium(r) 4" },
1436 			{ 0xf,	"Mobile Intel(r) Celeron(r)" },
1437 			{ 0x11, "Mobile Genuine Intel(r)" },
1438 			{ 0x12, "Intel(r) Celeron(r) M" },
1439 			{ 0x13, "Mobile Intel(r) Celeron(r)" },
1440 			{ 0x14, "Intel(r) Celeron(r)" },
1441 			{ 0x15, "Mobile Genuine Intel(r)" },
1442 			{ 0x16,	"Intel(r) Pentium(r) M" },
1443 			{ 0x17, "Mobile Intel(r) Celeron(r)" }
1444 		};
1445 		uint_t btblmax = sizeof (brand_tbl) / sizeof (brand_tbl[0]);
1446 		uint_t sgn;
1447 
1448 		sgn = (cpi->cpi_family << 8) |
1449 		    (cpi->cpi_model << 4) | cpi->cpi_step;
1450 
1451 		for (i = 0; i < btblmax; i++)
1452 			if (brand_tbl[i].bt_bid == cpi->cpi_brandid)
1453 				break;
1454 		if (i < btblmax) {
1455 			if (sgn == 0x6b1 && cpi->cpi_brandid == 3)
1456 				return ("Intel(r) Celeron(r)");
1457 			if (sgn < 0xf13 && cpi->cpi_brandid == 0xb)
1458 				return ("Intel(r) Xeon(tm) MP");
1459 			if (sgn < 0xf13 && cpi->cpi_brandid == 0xe)
1460 				return ("Intel(r) Xeon(tm)");
1461 			return (brand_tbl[i].bt_str);
1462 		}
1463 	}
1464 
1465 	return (NULL);
1466 }
1467 
1468 static const char *
1469 amd_cpubrand(const struct cpuid_info *cpi)
1470 {
1471 	if ((x86_feature & X86_CPUID) == 0 ||
1472 	    cpi->cpi_maxeax < 1 || cpi->cpi_family < 5)
1473 		return ("i486 compatible");
1474 
1475 	switch (cpi->cpi_family) {
1476 	case 5:
1477 		switch (cpi->cpi_model) {
1478 		case 0:
1479 		case 1:
1480 		case 2:
1481 		case 3:
1482 		case 4:
1483 		case 5:
1484 			return ("AMD-K5(r)");
1485 		case 6:
1486 		case 7:
1487 			return ("AMD-K6(r)");
1488 		case 8:
1489 			return ("AMD-K6(r)-2");
1490 		case 9:
1491 			return ("AMD-K6(r)-III");
1492 		default:
1493 			return ("AMD (family 5)");
1494 		}
1495 	case 6:
1496 		switch (cpi->cpi_model) {
1497 		case 1:
1498 			return ("AMD-K7(tm)");
1499 		case 0:
1500 		case 2:
1501 		case 4:
1502 			return ("AMD Athlon(tm)");
1503 		case 3:
1504 		case 7:
1505 			return ("AMD Duron(tm)");
1506 		case 6:
1507 		case 8:
1508 		case 10:
1509 			/*
1510 			 * Use the L2 cache size to distinguish
1511 			 */
1512 			return ((cpi->cpi_extd[6].cp_ecx >> 16) >= 256 ?
1513 			    "AMD Athlon(tm)" : "AMD Duron(tm)");
1514 		default:
1515 			return ("AMD (family 6)");
1516 		}
1517 	default:
1518 		break;
1519 	}
1520 
1521 	if (cpi->cpi_family == 0xf && cpi->cpi_model == 5 &&
1522 	    cpi->cpi_brandid != 0) {
1523 		switch (BITX(cpi->cpi_brandid, 7, 5)) {
1524 		case 3:
1525 			return ("AMD Opteron(tm) UP 1xx");
1526 		case 4:
1527 			return ("AMD Opteron(tm) DP 2xx");
1528 		case 5:
1529 			return ("AMD Opteron(tm) MP 8xx");
1530 		default:
1531 			return ("AMD Opteron(tm)");
1532 		}
1533 	}
1534 
1535 	return (NULL);
1536 }
1537 
1538 static const char *
1539 cyrix_cpubrand(struct cpuid_info *cpi, uint_t type)
1540 {
1541 	if ((x86_feature & X86_CPUID) == 0 ||
1542 	    cpi->cpi_maxeax < 1 || cpi->cpi_family < 5 ||
1543 	    type == X86_TYPE_CYRIX_486)
1544 		return ("i486 compatible");
1545 
1546 	switch (type) {
1547 	case X86_TYPE_CYRIX_6x86:
1548 		return ("Cyrix 6x86");
1549 	case X86_TYPE_CYRIX_6x86L:
1550 		return ("Cyrix 6x86L");
1551 	case X86_TYPE_CYRIX_6x86MX:
1552 		return ("Cyrix 6x86MX");
1553 	case X86_TYPE_CYRIX_GXm:
1554 		return ("Cyrix GXm");
1555 	case X86_TYPE_CYRIX_MediaGX:
1556 		return ("Cyrix MediaGX");
1557 	case X86_TYPE_CYRIX_MII:
1558 		return ("Cyrix M2");
1559 	case X86_TYPE_VIA_CYRIX_III:
1560 		return ("VIA Cyrix M3");
1561 	default:
1562 		/*
1563 		 * Have another wild guess ..
1564 		 */
1565 		if (cpi->cpi_family == 4 && cpi->cpi_model == 9)
1566 			return ("Cyrix 5x86");
1567 		else if (cpi->cpi_family == 5) {
1568 			switch (cpi->cpi_model) {
1569 			case 2:
1570 				return ("Cyrix 6x86");	/* Cyrix M1 */
1571 			case 4:
1572 				return ("Cyrix MediaGX");
1573 			default:
1574 				break;
1575 			}
1576 		} else if (cpi->cpi_family == 6) {
1577 			switch (cpi->cpi_model) {
1578 			case 0:
1579 				return ("Cyrix 6x86MX"); /* Cyrix M2? */
1580 			case 5:
1581 			case 6:
1582 			case 7:
1583 			case 8:
1584 			case 9:
1585 				return ("VIA C3");
1586 			default:
1587 				break;
1588 			}
1589 		}
1590 		break;
1591 	}
1592 	return (NULL);
1593 }
1594 
1595 /*
1596  * This only gets called in the case that the CPU extended
1597  * feature brand string (0x80000002, 0x80000003, 0x80000004)
1598  * aren't available, or contain null bytes for some reason.
1599  */
1600 static void
1601 fabricate_brandstr(struct cpuid_info *cpi)
1602 {
1603 	const char *brand = NULL;
1604 
1605 	switch (cpi->cpi_vendor) {
1606 	case X86_VENDOR_Intel:
1607 		brand = intel_cpubrand(cpi);
1608 		break;
1609 	case X86_VENDOR_AMD:
1610 		brand = amd_cpubrand(cpi);
1611 		break;
1612 	case X86_VENDOR_Cyrix:
1613 		brand = cyrix_cpubrand(cpi, x86_type);
1614 		break;
1615 	case X86_VENDOR_NexGen:
1616 		if (cpi->cpi_family == 5 && cpi->cpi_model == 0)
1617 			brand = "NexGen Nx586";
1618 		break;
1619 	case X86_VENDOR_Centaur:
1620 		if (cpi->cpi_family == 5)
1621 			switch (cpi->cpi_model) {
1622 			case 4:
1623 				brand = "Centaur C6";
1624 				break;
1625 			case 8:
1626 				brand = "Centaur C2";
1627 				break;
1628 			case 9:
1629 				brand = "Centaur C3";
1630 				break;
1631 			default:
1632 				break;
1633 			}
1634 		break;
1635 	case X86_VENDOR_Rise:
1636 		if (cpi->cpi_family == 5 &&
1637 		    (cpi->cpi_model == 0 || cpi->cpi_model == 2))
1638 			brand = "Rise mP6";
1639 		break;
1640 	case X86_VENDOR_SiS:
1641 		if (cpi->cpi_family == 5 && cpi->cpi_model == 0)
1642 			brand = "SiS 55x";
1643 		break;
1644 	case X86_VENDOR_TM:
1645 		if (cpi->cpi_family == 5 && cpi->cpi_model == 4)
1646 			brand = "Transmeta Crusoe TM3x00 or TM5x00";
1647 		break;
1648 	case X86_VENDOR_NSC:
1649 	case X86_VENDOR_UMC:
1650 	default:
1651 		break;
1652 	}
1653 	if (brand) {
1654 		(void) strcpy((char *)cpi->cpi_brandstr, brand);
1655 		return;
1656 	}
1657 
1658 	/*
1659 	 * If all else fails ...
1660 	 */
1661 	(void) snprintf(cpi->cpi_brandstr, sizeof (cpi->cpi_brandstr),
1662 	    "%s %d.%d.%d", cpi->cpi_vendorstr, cpi->cpi_family,
1663 	    cpi->cpi_model, cpi->cpi_step);
1664 }
1665 
1666 /*
1667  * This routine is called just after kernel memory allocation
1668  * becomes available on cpu0, and as part of mp_startup() on
1669  * the other cpus.
1670  *
1671  * Fixup the brand string, and collect any information from cpuid
1672  * that requires dynamicically allocated storage to represent.
1673  */
1674 /*ARGSUSED*/
1675 void
1676 cpuid_pass3(cpu_t *cpu)
1677 {
1678 	int	i, max, shft, level, size;
1679 	struct cpuid_regs regs;
1680 	struct cpuid_regs *cp;
1681 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
1682 
1683 	ASSERT(cpi->cpi_pass == 2);
1684 
1685 	/*
1686 	 * Function 4: Deterministic cache parameters
1687 	 *
1688 	 * Take this opportunity to detect the number of threads
1689 	 * sharing the last level cache, and construct a corresponding
1690 	 * cache id. The respective cpuid_info members are initialized
1691 	 * to the default case of "no last level cache sharing".
1692 	 */
1693 	cpi->cpi_ncpu_shr_last_cache = 1;
1694 	cpi->cpi_last_lvl_cacheid = cpu->cpu_id;
1695 
1696 	if (cpi->cpi_maxeax >= 4 && cpi->cpi_vendor == X86_VENDOR_Intel) {
1697 
1698 		/*
1699 		 * Find the # of elements (size) returned by fn 4, and along
1700 		 * the way detect last level cache sharing details.
1701 		 */
1702 		bzero(&regs, sizeof (regs));
1703 		cp = &regs;
1704 		for (i = 0, max = 0; i < CPI_FN4_ECX_MAX; i++) {
1705 			cp->cp_eax = 4;
1706 			cp->cp_ecx = i;
1707 
1708 			(void) __cpuid_insn(cp);
1709 
1710 			if (CPI_CACHE_TYPE(cp) == 0)
1711 				break;
1712 			level = CPI_CACHE_LVL(cp);
1713 			if (level > max) {
1714 				max = level;
1715 				cpi->cpi_ncpu_shr_last_cache =
1716 				    CPI_NTHR_SHR_CACHE(cp) + 1;
1717 			}
1718 		}
1719 		cpi->cpi_std_4_size = size = i;
1720 
1721 		/*
1722 		 * Allocate the cpi_std_4 array. The first element
1723 		 * references the regs for fn 4, %ecx == 0, which
1724 		 * cpuid_pass2() stashed in cpi->cpi_std[4].
1725 		 */
1726 		if (size > 0) {
1727 			cpi->cpi_std_4 =
1728 			    kmem_alloc(size * sizeof (cp), KM_SLEEP);
1729 			cpi->cpi_std_4[0] = &cpi->cpi_std[4];
1730 
1731 			/*
1732 			 * Allocate storage to hold the additional regs
1733 			 * for function 4, %ecx == 1 .. cpi_std_4_size.
1734 			 *
1735 			 * The regs for fn 4, %ecx == 0 has already
1736 			 * been allocated as indicated above.
1737 			 */
1738 			for (i = 1; i < size; i++) {
1739 				cp = cpi->cpi_std_4[i] =
1740 				    kmem_zalloc(sizeof (regs), KM_SLEEP);
1741 				cp->cp_eax = 4;
1742 				cp->cp_ecx = i;
1743 
1744 				(void) __cpuid_insn(cp);
1745 			}
1746 		}
1747 		/*
1748 		 * Determine the number of bits needed to represent
1749 		 * the number of CPUs sharing the last level cache.
1750 		 *
1751 		 * Shift off that number of bits from the APIC id to
1752 		 * derive the cache id.
1753 		 */
1754 		shft = 0;
1755 		for (i = 1; i < cpi->cpi_ncpu_shr_last_cache; i <<= 1)
1756 			shft++;
1757 		cpi->cpi_last_lvl_cacheid = CPI_APIC_ID(cpi) >> shft;
1758 	}
1759 
1760 	/*
1761 	 * Now fixup the brand string
1762 	 */
1763 	if ((cpi->cpi_xmaxeax & 0x80000000) == 0) {
1764 		fabricate_brandstr(cpi);
1765 	} else {
1766 
1767 		/*
1768 		 * If we successfully extracted a brand string from the cpuid
1769 		 * instruction, clean it up by removing leading spaces and
1770 		 * similar junk.
1771 		 */
1772 		if (cpi->cpi_brandstr[0]) {
1773 			size_t maxlen = sizeof (cpi->cpi_brandstr);
1774 			char *src, *dst;
1775 
1776 			dst = src = (char *)cpi->cpi_brandstr;
1777 			src[maxlen - 1] = '\0';
1778 			/*
1779 			 * strip leading spaces
1780 			 */
1781 			while (*src == ' ')
1782 				src++;
1783 			/*
1784 			 * Remove any 'Genuine' or "Authentic" prefixes
1785 			 */
1786 			if (strncmp(src, "Genuine ", 8) == 0)
1787 				src += 8;
1788 			if (strncmp(src, "Authentic ", 10) == 0)
1789 				src += 10;
1790 
1791 			/*
1792 			 * Now do an in-place copy.
1793 			 * Map (R) to (r) and (TM) to (tm).
1794 			 * The era of teletypes is long gone, and there's
1795 			 * -really- no need to shout.
1796 			 */
1797 			while (*src != '\0') {
1798 				if (src[0] == '(') {
1799 					if (strncmp(src + 1, "R)", 2) == 0) {
1800 						(void) strncpy(dst, "(r)", 3);
1801 						src += 3;
1802 						dst += 3;
1803 						continue;
1804 					}
1805 					if (strncmp(src + 1, "TM)", 3) == 0) {
1806 						(void) strncpy(dst, "(tm)", 4);
1807 						src += 4;
1808 						dst += 4;
1809 						continue;
1810 					}
1811 				}
1812 				*dst++ = *src++;
1813 			}
1814 			*dst = '\0';
1815 
1816 			/*
1817 			 * Finally, remove any trailing spaces
1818 			 */
1819 			while (--dst > cpi->cpi_brandstr)
1820 				if (*dst == ' ')
1821 					*dst = '\0';
1822 				else
1823 					break;
1824 		} else
1825 			fabricate_brandstr(cpi);
1826 	}
1827 	cpi->cpi_pass = 3;
1828 }
1829 
1830 /*
1831  * This routine is called out of bind_hwcap() much later in the life
1832  * of the kernel (post_startup()).  The job of this routine is to resolve
1833  * the hardware feature support and kernel support for those features into
1834  * what we're actually going to tell applications via the aux vector.
1835  */
1836 uint_t
1837 cpuid_pass4(cpu_t *cpu)
1838 {
1839 	struct cpuid_info *cpi;
1840 	uint_t hwcap_flags = 0;
1841 
1842 	if (cpu == NULL)
1843 		cpu = CPU;
1844 	cpi = cpu->cpu_m.mcpu_cpi;
1845 
1846 	ASSERT(cpi->cpi_pass == 3);
1847 
1848 	if (cpi->cpi_maxeax >= 1) {
1849 		uint32_t *edx = &cpi->cpi_support[STD_EDX_FEATURES];
1850 		uint32_t *ecx = &cpi->cpi_support[STD_ECX_FEATURES];
1851 
1852 		*edx = CPI_FEATURES_EDX(cpi);
1853 		*ecx = CPI_FEATURES_ECX(cpi);
1854 
1855 		/*
1856 		 * [these require explicit kernel support]
1857 		 */
1858 		if ((x86_feature & X86_SEP) == 0)
1859 			*edx &= ~CPUID_INTC_EDX_SEP;
1860 
1861 		if ((x86_feature & X86_SSE) == 0)
1862 			*edx &= ~(CPUID_INTC_EDX_FXSR|CPUID_INTC_EDX_SSE);
1863 		if ((x86_feature & X86_SSE2) == 0)
1864 			*edx &= ~CPUID_INTC_EDX_SSE2;
1865 
1866 		if ((x86_feature & X86_HTT) == 0)
1867 			*edx &= ~CPUID_INTC_EDX_HTT;
1868 
1869 		if ((x86_feature & X86_SSE3) == 0)
1870 			*ecx &= ~CPUID_INTC_ECX_SSE3;
1871 
1872 		/*
1873 		 * [no explicit support required beyond x87 fp context]
1874 		 */
1875 		if (!fpu_exists)
1876 			*edx &= ~(CPUID_INTC_EDX_FPU | CPUID_INTC_EDX_MMX);
1877 
1878 		/*
1879 		 * Now map the supported feature vector to things that we
1880 		 * think userland will care about.
1881 		 */
1882 		if (*edx & CPUID_INTC_EDX_SEP)
1883 			hwcap_flags |= AV_386_SEP;
1884 		if (*edx & CPUID_INTC_EDX_SSE)
1885 			hwcap_flags |= AV_386_FXSR | AV_386_SSE;
1886 		if (*edx & CPUID_INTC_EDX_SSE2)
1887 			hwcap_flags |= AV_386_SSE2;
1888 		if (*ecx & CPUID_INTC_ECX_SSE3)
1889 			hwcap_flags |= AV_386_SSE3;
1890 		if (*ecx & CPUID_INTC_ECX_POPCNT)
1891 			hwcap_flags |= AV_386_POPCNT;
1892 		if (*edx & CPUID_INTC_EDX_FPU)
1893 			hwcap_flags |= AV_386_FPU;
1894 		if (*edx & CPUID_INTC_EDX_MMX)
1895 			hwcap_flags |= AV_386_MMX;
1896 
1897 		if (*edx & CPUID_INTC_EDX_TSC)
1898 			hwcap_flags |= AV_386_TSC;
1899 		if (*edx & CPUID_INTC_EDX_CX8)
1900 			hwcap_flags |= AV_386_CX8;
1901 		if (*edx & CPUID_INTC_EDX_CMOV)
1902 			hwcap_flags |= AV_386_CMOV;
1903 		if (*ecx & CPUID_INTC_ECX_MON)
1904 			hwcap_flags |= AV_386_MON;
1905 		if (*ecx & CPUID_INTC_ECX_CX16)
1906 			hwcap_flags |= AV_386_CX16;
1907 	}
1908 
1909 	if (x86_feature & X86_HTT)
1910 		hwcap_flags |= AV_386_PAUSE;
1911 
1912 	if (cpi->cpi_xmaxeax < 0x80000001)
1913 		goto pass4_done;
1914 
1915 	switch (cpi->cpi_vendor) {
1916 		struct cpuid_regs cp;
1917 		uint32_t *edx, *ecx;
1918 
1919 	case X86_VENDOR_Intel:
1920 		/*
1921 		 * Seems like Intel duplicated what we necessary
1922 		 * here to make the initial crop of 64-bit OS's work.
1923 		 * Hopefully, those are the only "extended" bits
1924 		 * they'll add.
1925 		 */
1926 		/*FALLTHROUGH*/
1927 
1928 	case X86_VENDOR_AMD:
1929 		edx = &cpi->cpi_support[AMD_EDX_FEATURES];
1930 		ecx = &cpi->cpi_support[AMD_ECX_FEATURES];
1931 
1932 		*edx = CPI_FEATURES_XTD_EDX(cpi);
1933 		*ecx = CPI_FEATURES_XTD_ECX(cpi);
1934 
1935 		/*
1936 		 * [these features require explicit kernel support]
1937 		 */
1938 		switch (cpi->cpi_vendor) {
1939 		case X86_VENDOR_Intel:
1940 			break;
1941 
1942 		case X86_VENDOR_AMD:
1943 			if ((x86_feature & X86_TSCP) == 0)
1944 				*edx &= ~CPUID_AMD_EDX_TSCP;
1945 			if ((x86_feature & X86_SSE4A) == 0)
1946 				*ecx &= ~CPUID_AMD_ECX_SSE4A;
1947 			break;
1948 
1949 		default:
1950 			break;
1951 		}
1952 
1953 		/*
1954 		 * [no explicit support required beyond
1955 		 * x87 fp context and exception handlers]
1956 		 */
1957 		if (!fpu_exists)
1958 			*edx &= ~(CPUID_AMD_EDX_MMXamd |
1959 			    CPUID_AMD_EDX_3DNow | CPUID_AMD_EDX_3DNowx);
1960 
1961 		if ((x86_feature & X86_NX) == 0)
1962 			*edx &= ~CPUID_AMD_EDX_NX;
1963 #if !defined(__amd64)
1964 		*edx &= ~CPUID_AMD_EDX_LM;
1965 #endif
1966 		/*
1967 		 * Now map the supported feature vector to
1968 		 * things that we think userland will care about.
1969 		 */
1970 #if defined(__amd64)
1971 		if (*edx & CPUID_AMD_EDX_SYSC)
1972 			hwcap_flags |= AV_386_AMD_SYSC;
1973 #endif
1974 		if (*edx & CPUID_AMD_EDX_MMXamd)
1975 			hwcap_flags |= AV_386_AMD_MMX;
1976 		if (*edx & CPUID_AMD_EDX_3DNow)
1977 			hwcap_flags |= AV_386_AMD_3DNow;
1978 		if (*edx & CPUID_AMD_EDX_3DNowx)
1979 			hwcap_flags |= AV_386_AMD_3DNowx;
1980 
1981 		switch (cpi->cpi_vendor) {
1982 		case X86_VENDOR_AMD:
1983 			if (*edx & CPUID_AMD_EDX_TSCP)
1984 				hwcap_flags |= AV_386_TSCP;
1985 			if (*ecx & CPUID_AMD_ECX_AHF64)
1986 				hwcap_flags |= AV_386_AHF;
1987 			if (*ecx & CPUID_AMD_ECX_SSE4A)
1988 				hwcap_flags |= AV_386_AMD_SSE4A;
1989 			if (*ecx & CPUID_AMD_ECX_LZCNT)
1990 				hwcap_flags |= AV_386_AMD_LZCNT;
1991 			break;
1992 
1993 		case X86_VENDOR_Intel:
1994 			/*
1995 			 * Aarrgh.
1996 			 * Intel uses a different bit in the same word.
1997 			 */
1998 			if (*ecx & CPUID_INTC_ECX_AHF64)
1999 				hwcap_flags |= AV_386_AHF;
2000 			break;
2001 
2002 		default:
2003 			break;
2004 		}
2005 		break;
2006 
2007 	case X86_VENDOR_TM:
2008 		cp.cp_eax = 0x80860001;
2009 		(void) __cpuid_insn(&cp);
2010 		cpi->cpi_support[TM_EDX_FEATURES] = cp.cp_edx;
2011 		break;
2012 
2013 	default:
2014 		break;
2015 	}
2016 
2017 pass4_done:
2018 	cpi->cpi_pass = 4;
2019 	return (hwcap_flags);
2020 }
2021 
2022 
2023 /*
2024  * Simulate the cpuid instruction using the data we previously
2025  * captured about this CPU.  We try our best to return the truth
2026  * about the hardware, independently of kernel support.
2027  */
2028 uint32_t
2029 cpuid_insn(cpu_t *cpu, struct cpuid_regs *cp)
2030 {
2031 	struct cpuid_info *cpi;
2032 	struct cpuid_regs *xcp;
2033 
2034 	if (cpu == NULL)
2035 		cpu = CPU;
2036 	cpi = cpu->cpu_m.mcpu_cpi;
2037 
2038 	ASSERT(cpuid_checkpass(cpu, 3));
2039 
2040 	/*
2041 	 * CPUID data is cached in two separate places: cpi_std for standard
2042 	 * CPUID functions, and cpi_extd for extended CPUID functions.
2043 	 */
2044 	if (cp->cp_eax <= cpi->cpi_maxeax && cp->cp_eax < NMAX_CPI_STD)
2045 		xcp = &cpi->cpi_std[cp->cp_eax];
2046 	else if (cp->cp_eax >= 0x80000000 && cp->cp_eax <= cpi->cpi_xmaxeax &&
2047 	    cp->cp_eax < 0x80000000 + NMAX_CPI_EXTD)
2048 		xcp = &cpi->cpi_extd[cp->cp_eax - 0x80000000];
2049 	else
2050 		/*
2051 		 * The caller is asking for data from an input parameter which
2052 		 * the kernel has not cached.  In this case we go fetch from
2053 		 * the hardware and return the data directly to the user.
2054 		 */
2055 		return (__cpuid_insn(cp));
2056 
2057 	cp->cp_eax = xcp->cp_eax;
2058 	cp->cp_ebx = xcp->cp_ebx;
2059 	cp->cp_ecx = xcp->cp_ecx;
2060 	cp->cp_edx = xcp->cp_edx;
2061 	return (cp->cp_eax);
2062 }
2063 
2064 int
2065 cpuid_checkpass(cpu_t *cpu, int pass)
2066 {
2067 	return (cpu != NULL && cpu->cpu_m.mcpu_cpi != NULL &&
2068 	    cpu->cpu_m.mcpu_cpi->cpi_pass >= pass);
2069 }
2070 
2071 int
2072 cpuid_getbrandstr(cpu_t *cpu, char *s, size_t n)
2073 {
2074 	ASSERT(cpuid_checkpass(cpu, 3));
2075 
2076 	return (snprintf(s, n, "%s", cpu->cpu_m.mcpu_cpi->cpi_brandstr));
2077 }
2078 
2079 int
2080 cpuid_is_cmt(cpu_t *cpu)
2081 {
2082 	if (cpu == NULL)
2083 		cpu = CPU;
2084 
2085 	ASSERT(cpuid_checkpass(cpu, 1));
2086 
2087 	return (cpu->cpu_m.mcpu_cpi->cpi_chipid >= 0);
2088 }
2089 
2090 /*
2091  * AMD and Intel both implement the 64-bit variant of the syscall
2092  * instruction (syscallq), so if there's -any- support for syscall,
2093  * cpuid currently says "yes, we support this".
2094  *
2095  * However, Intel decided to -not- implement the 32-bit variant of the
2096  * syscall instruction, so we provide a predicate to allow our caller
2097  * to test that subtlety here.
2098  */
2099 /*ARGSUSED*/
2100 int
2101 cpuid_syscall32_insn(cpu_t *cpu)
2102 {
2103 	ASSERT(cpuid_checkpass((cpu == NULL ? CPU : cpu), 1));
2104 
2105 	if (cpu == NULL)
2106 		cpu = CPU;
2107 
2108 	/*CSTYLED*/
2109 	{
2110 		struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2111 
2112 		if (cpi->cpi_vendor == X86_VENDOR_AMD &&
2113 		    cpi->cpi_xmaxeax >= 0x80000001 &&
2114 		    (CPI_FEATURES_XTD_EDX(cpi) & CPUID_AMD_EDX_SYSC))
2115 			return (1);
2116 	}
2117 	return (0);
2118 }
2119 
2120 int
2121 cpuid_getidstr(cpu_t *cpu, char *s, size_t n)
2122 {
2123 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2124 
2125 	static const char fmt[] =
2126 	    "x86 (%s %X family %d model %d step %d clock %d MHz)";
2127 	static const char fmt_ht[] =
2128 	    "x86 (chipid 0x%x %s %X family %d model %d step %d clock %d MHz)";
2129 
2130 	ASSERT(cpuid_checkpass(cpu, 1));
2131 
2132 	if (cpuid_is_cmt(cpu))
2133 		return (snprintf(s, n, fmt_ht, cpi->cpi_chipid,
2134 		    cpi->cpi_vendorstr, cpi->cpi_std[1].cp_eax,
2135 		    cpi->cpi_family, cpi->cpi_model,
2136 		    cpi->cpi_step, cpu->cpu_type_info.pi_clock));
2137 	return (snprintf(s, n, fmt,
2138 	    cpi->cpi_vendorstr, cpi->cpi_std[1].cp_eax,
2139 	    cpi->cpi_family, cpi->cpi_model,
2140 	    cpi->cpi_step, cpu->cpu_type_info.pi_clock));
2141 }
2142 
2143 const char *
2144 cpuid_getvendorstr(cpu_t *cpu)
2145 {
2146 	ASSERT(cpuid_checkpass(cpu, 1));
2147 	return ((const char *)cpu->cpu_m.mcpu_cpi->cpi_vendorstr);
2148 }
2149 
2150 uint_t
2151 cpuid_getvendor(cpu_t *cpu)
2152 {
2153 	ASSERT(cpuid_checkpass(cpu, 1));
2154 	return (cpu->cpu_m.mcpu_cpi->cpi_vendor);
2155 }
2156 
2157 uint_t
2158 cpuid_getfamily(cpu_t *cpu)
2159 {
2160 	ASSERT(cpuid_checkpass(cpu, 1));
2161 	return (cpu->cpu_m.mcpu_cpi->cpi_family);
2162 }
2163 
2164 uint_t
2165 cpuid_getmodel(cpu_t *cpu)
2166 {
2167 	ASSERT(cpuid_checkpass(cpu, 1));
2168 	return (cpu->cpu_m.mcpu_cpi->cpi_model);
2169 }
2170 
2171 uint_t
2172 cpuid_get_ncpu_per_chip(cpu_t *cpu)
2173 {
2174 	ASSERT(cpuid_checkpass(cpu, 1));
2175 	return (cpu->cpu_m.mcpu_cpi->cpi_ncpu_per_chip);
2176 }
2177 
2178 uint_t
2179 cpuid_get_ncore_per_chip(cpu_t *cpu)
2180 {
2181 	ASSERT(cpuid_checkpass(cpu, 1));
2182 	return (cpu->cpu_m.mcpu_cpi->cpi_ncore_per_chip);
2183 }
2184 
2185 uint_t
2186 cpuid_get_ncpu_sharing_last_cache(cpu_t *cpu)
2187 {
2188 	ASSERT(cpuid_checkpass(cpu, 2));
2189 	return (cpu->cpu_m.mcpu_cpi->cpi_ncpu_shr_last_cache);
2190 }
2191 
2192 id_t
2193 cpuid_get_last_lvl_cacheid(cpu_t *cpu)
2194 {
2195 	ASSERT(cpuid_checkpass(cpu, 2));
2196 	return (cpu->cpu_m.mcpu_cpi->cpi_last_lvl_cacheid);
2197 }
2198 
2199 uint_t
2200 cpuid_getstep(cpu_t *cpu)
2201 {
2202 	ASSERT(cpuid_checkpass(cpu, 1));
2203 	return (cpu->cpu_m.mcpu_cpi->cpi_step);
2204 }
2205 
2206 uint_t
2207 cpuid_getsig(struct cpu *cpu)
2208 {
2209 	ASSERT(cpuid_checkpass(cpu, 1));
2210 	return (cpu->cpu_m.mcpu_cpi->cpi_std[1].cp_eax);
2211 }
2212 
2213 uint32_t
2214 cpuid_getchiprev(struct cpu *cpu)
2215 {
2216 	ASSERT(cpuid_checkpass(cpu, 1));
2217 	return (cpu->cpu_m.mcpu_cpi->cpi_chiprev);
2218 }
2219 
2220 const char *
2221 cpuid_getchiprevstr(struct cpu *cpu)
2222 {
2223 	ASSERT(cpuid_checkpass(cpu, 1));
2224 	return (cpu->cpu_m.mcpu_cpi->cpi_chiprevstr);
2225 }
2226 
2227 uint32_t
2228 cpuid_getsockettype(struct cpu *cpu)
2229 {
2230 	ASSERT(cpuid_checkpass(cpu, 1));
2231 	return (cpu->cpu_m.mcpu_cpi->cpi_socket);
2232 }
2233 
2234 int
2235 cpuid_get_chipid(cpu_t *cpu)
2236 {
2237 	ASSERT(cpuid_checkpass(cpu, 1));
2238 
2239 	if (cpuid_is_cmt(cpu))
2240 		return (cpu->cpu_m.mcpu_cpi->cpi_chipid);
2241 	return (cpu->cpu_id);
2242 }
2243 
2244 id_t
2245 cpuid_get_coreid(cpu_t *cpu)
2246 {
2247 	ASSERT(cpuid_checkpass(cpu, 1));
2248 	return (cpu->cpu_m.mcpu_cpi->cpi_coreid);
2249 }
2250 
2251 int
2252 cpuid_get_clogid(cpu_t *cpu)
2253 {
2254 	ASSERT(cpuid_checkpass(cpu, 1));
2255 	return (cpu->cpu_m.mcpu_cpi->cpi_clogid);
2256 }
2257 
2258 void
2259 cpuid_get_addrsize(cpu_t *cpu, uint_t *pabits, uint_t *vabits)
2260 {
2261 	struct cpuid_info *cpi;
2262 
2263 	if (cpu == NULL)
2264 		cpu = CPU;
2265 	cpi = cpu->cpu_m.mcpu_cpi;
2266 
2267 	ASSERT(cpuid_checkpass(cpu, 1));
2268 
2269 	if (pabits)
2270 		*pabits = cpi->cpi_pabits;
2271 	if (vabits)
2272 		*vabits = cpi->cpi_vabits;
2273 }
2274 
2275 /*
2276  * Returns the number of data TLB entries for a corresponding
2277  * pagesize.  If it can't be computed, or isn't known, the
2278  * routine returns zero.  If you ask about an architecturally
2279  * impossible pagesize, the routine will panic (so that the
2280  * hat implementor knows that things are inconsistent.)
2281  */
2282 uint_t
2283 cpuid_get_dtlb_nent(cpu_t *cpu, size_t pagesize)
2284 {
2285 	struct cpuid_info *cpi;
2286 	uint_t dtlb_nent = 0;
2287 
2288 	if (cpu == NULL)
2289 		cpu = CPU;
2290 	cpi = cpu->cpu_m.mcpu_cpi;
2291 
2292 	ASSERT(cpuid_checkpass(cpu, 1));
2293 
2294 	/*
2295 	 * Check the L2 TLB info
2296 	 */
2297 	if (cpi->cpi_xmaxeax >= 0x80000006) {
2298 		struct cpuid_regs *cp = &cpi->cpi_extd[6];
2299 
2300 		switch (pagesize) {
2301 
2302 		case 4 * 1024:
2303 			/*
2304 			 * All zero in the top 16 bits of the register
2305 			 * indicates a unified TLB. Size is in low 16 bits.
2306 			 */
2307 			if ((cp->cp_ebx & 0xffff0000) == 0)
2308 				dtlb_nent = cp->cp_ebx & 0x0000ffff;
2309 			else
2310 				dtlb_nent = BITX(cp->cp_ebx, 27, 16);
2311 			break;
2312 
2313 		case 2 * 1024 * 1024:
2314 			if ((cp->cp_eax & 0xffff0000) == 0)
2315 				dtlb_nent = cp->cp_eax & 0x0000ffff;
2316 			else
2317 				dtlb_nent = BITX(cp->cp_eax, 27, 16);
2318 			break;
2319 
2320 		default:
2321 			panic("unknown L2 pagesize");
2322 			/*NOTREACHED*/
2323 		}
2324 	}
2325 
2326 	if (dtlb_nent != 0)
2327 		return (dtlb_nent);
2328 
2329 	/*
2330 	 * No L2 TLB support for this size, try L1.
2331 	 */
2332 	if (cpi->cpi_xmaxeax >= 0x80000005) {
2333 		struct cpuid_regs *cp = &cpi->cpi_extd[5];
2334 
2335 		switch (pagesize) {
2336 		case 4 * 1024:
2337 			dtlb_nent = BITX(cp->cp_ebx, 23, 16);
2338 			break;
2339 		case 2 * 1024 * 1024:
2340 			dtlb_nent = BITX(cp->cp_eax, 23, 16);
2341 			break;
2342 		default:
2343 			panic("unknown L1 d-TLB pagesize");
2344 			/*NOTREACHED*/
2345 		}
2346 	}
2347 
2348 	return (dtlb_nent);
2349 }
2350 
2351 /*
2352  * Return 0 if the erratum is not present or not applicable, positive
2353  * if it is, and negative if the status of the erratum is unknown.
2354  *
2355  * See "Revision Guide for AMD Athlon(tm) 64 and AMD Opteron(tm)
2356  * Processors" #25759, Rev 3.57, August 2005
2357  */
2358 int
2359 cpuid_opteron_erratum(cpu_t *cpu, uint_t erratum)
2360 {
2361 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2362 	uint_t eax;
2363 
2364 	/*
2365 	 * Bail out if this CPU isn't an AMD CPU, or if it's
2366 	 * a legacy (32-bit) AMD CPU.
2367 	 */
2368 	if (cpi->cpi_vendor != X86_VENDOR_AMD ||
2369 	    cpi->cpi_family == 4 || cpi->cpi_family == 5 ||
2370 	    cpi->cpi_family == 6)
2371 
2372 		return (0);
2373 
2374 	eax = cpi->cpi_std[1].cp_eax;
2375 
2376 #define	SH_B0(eax)	(eax == 0xf40 || eax == 0xf50)
2377 #define	SH_B3(eax) 	(eax == 0xf51)
2378 #define	B(eax)		(SH_B0(eax) || SH_B3(eax))
2379 
2380 #define	SH_C0(eax)	(eax == 0xf48 || eax == 0xf58)
2381 
2382 #define	SH_CG(eax)	(eax == 0xf4a || eax == 0xf5a || eax == 0xf7a)
2383 #define	DH_CG(eax)	(eax == 0xfc0 || eax == 0xfe0 || eax == 0xff0)
2384 #define	CH_CG(eax)	(eax == 0xf82 || eax == 0xfb2)
2385 #define	CG(eax)		(SH_CG(eax) || DH_CG(eax) || CH_CG(eax))
2386 
2387 #define	SH_D0(eax)	(eax == 0x10f40 || eax == 0x10f50 || eax == 0x10f70)
2388 #define	DH_D0(eax)	(eax == 0x10fc0 || eax == 0x10ff0)
2389 #define	CH_D0(eax)	(eax == 0x10f80 || eax == 0x10fb0)
2390 #define	D0(eax)		(SH_D0(eax) || DH_D0(eax) || CH_D0(eax))
2391 
2392 #define	SH_E0(eax)	(eax == 0x20f50 || eax == 0x20f40 || eax == 0x20f70)
2393 #define	JH_E1(eax)	(eax == 0x20f10)	/* JH8_E0 had 0x20f30 */
2394 #define	DH_E3(eax)	(eax == 0x20fc0 || eax == 0x20ff0)
2395 #define	SH_E4(eax)	(eax == 0x20f51 || eax == 0x20f71)
2396 #define	BH_E4(eax)	(eax == 0x20fb1)
2397 #define	SH_E5(eax)	(eax == 0x20f42)
2398 #define	DH_E6(eax)	(eax == 0x20ff2 || eax == 0x20fc2)
2399 #define	JH_E6(eax)	(eax == 0x20f12 || eax == 0x20f32)
2400 #define	EX(eax)		(SH_E0(eax) || JH_E1(eax) || DH_E3(eax) || \
2401 			    SH_E4(eax) || BH_E4(eax) || SH_E5(eax) || \
2402 			    DH_E6(eax) || JH_E6(eax))
2403 
2404 	switch (erratum) {
2405 	case 1:
2406 		return (cpi->cpi_family < 0x10);
2407 	case 51:	/* what does the asterisk mean? */
2408 		return (B(eax) || SH_C0(eax) || CG(eax));
2409 	case 52:
2410 		return (B(eax));
2411 	case 57:
2412 		return (cpi->cpi_family <= 0x10);
2413 	case 58:
2414 		return (B(eax));
2415 	case 60:
2416 		return (cpi->cpi_family <= 0x10);
2417 	case 61:
2418 	case 62:
2419 	case 63:
2420 	case 64:
2421 	case 65:
2422 	case 66:
2423 	case 68:
2424 	case 69:
2425 	case 70:
2426 	case 71:
2427 		return (B(eax));
2428 	case 72:
2429 		return (SH_B0(eax));
2430 	case 74:
2431 		return (B(eax));
2432 	case 75:
2433 		return (cpi->cpi_family < 0x10);
2434 	case 76:
2435 		return (B(eax));
2436 	case 77:
2437 		return (cpi->cpi_family <= 0x10);
2438 	case 78:
2439 		return (B(eax) || SH_C0(eax));
2440 	case 79:
2441 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
2442 	case 80:
2443 	case 81:
2444 	case 82:
2445 		return (B(eax));
2446 	case 83:
2447 		return (B(eax) || SH_C0(eax) || CG(eax));
2448 	case 85:
2449 		return (cpi->cpi_family < 0x10);
2450 	case 86:
2451 		return (SH_C0(eax) || CG(eax));
2452 	case 88:
2453 #if !defined(__amd64)
2454 		return (0);
2455 #else
2456 		return (B(eax) || SH_C0(eax));
2457 #endif
2458 	case 89:
2459 		return (cpi->cpi_family < 0x10);
2460 	case 90:
2461 		return (B(eax) || SH_C0(eax) || CG(eax));
2462 	case 91:
2463 	case 92:
2464 		return (B(eax) || SH_C0(eax));
2465 	case 93:
2466 		return (SH_C0(eax));
2467 	case 94:
2468 		return (B(eax) || SH_C0(eax) || CG(eax));
2469 	case 95:
2470 #if !defined(__amd64)
2471 		return (0);
2472 #else
2473 		return (B(eax) || SH_C0(eax));
2474 #endif
2475 	case 96:
2476 		return (B(eax) || SH_C0(eax) || CG(eax));
2477 	case 97:
2478 	case 98:
2479 		return (SH_C0(eax) || CG(eax));
2480 	case 99:
2481 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
2482 	case 100:
2483 		return (B(eax) || SH_C0(eax));
2484 	case 101:
2485 	case 103:
2486 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
2487 	case 104:
2488 		return (SH_C0(eax) || CG(eax) || D0(eax));
2489 	case 105:
2490 	case 106:
2491 	case 107:
2492 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
2493 	case 108:
2494 		return (DH_CG(eax));
2495 	case 109:
2496 		return (SH_C0(eax) || CG(eax) || D0(eax));
2497 	case 110:
2498 		return (D0(eax) || EX(eax));
2499 	case 111:
2500 		return (CG(eax));
2501 	case 112:
2502 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
2503 	case 113:
2504 		return (eax == 0x20fc0);
2505 	case 114:
2506 		return (SH_E0(eax) || JH_E1(eax) || DH_E3(eax));
2507 	case 115:
2508 		return (SH_E0(eax) || JH_E1(eax));
2509 	case 116:
2510 		return (SH_E0(eax) || JH_E1(eax) || DH_E3(eax));
2511 	case 117:
2512 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
2513 	case 118:
2514 		return (SH_E0(eax) || JH_E1(eax) || SH_E4(eax) || BH_E4(eax) ||
2515 		    JH_E6(eax));
2516 	case 121:
2517 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
2518 	case 122:
2519 		return (cpi->cpi_family < 0x10);
2520 	case 123:
2521 		return (JH_E1(eax) || BH_E4(eax) || JH_E6(eax));
2522 	case 131:
2523 		return (cpi->cpi_family < 0x10);
2524 	case 6336786:
2525 		/*
2526 		 * Test for AdvPowerMgmtInfo.TscPStateInvariant
2527 		 * if this is a K8 family or newer processor
2528 		 */
2529 		if (CPI_FAMILY(cpi) == 0xf) {
2530 			struct cpuid_regs regs;
2531 			regs.cp_eax = 0x80000007;
2532 			(void) __cpuid_insn(&regs);
2533 			return (!(regs.cp_edx & 0x100));
2534 		}
2535 		return (0);
2536 	case 6323525:
2537 		return (((((eax >> 12) & 0xff00) + (eax & 0xf00)) |
2538 		    (((eax >> 4) & 0xf) | ((eax >> 12) & 0xf0))) < 0xf40);
2539 
2540 	default:
2541 		return (-1);
2542 	}
2543 }
2544 
2545 static const char assoc_str[] = "associativity";
2546 static const char line_str[] = "line-size";
2547 static const char size_str[] = "size";
2548 
2549 static void
2550 add_cache_prop(dev_info_t *devi, const char *label, const char *type,
2551     uint32_t val)
2552 {
2553 	char buf[128];
2554 
2555 	/*
2556 	 * ndi_prop_update_int() is used because it is desirable for
2557 	 * DDI_PROP_HW_DEF and DDI_PROP_DONTSLEEP to be set.
2558 	 */
2559 	if (snprintf(buf, sizeof (buf), "%s-%s", label, type) < sizeof (buf))
2560 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, devi, buf, val);
2561 }
2562 
2563 /*
2564  * Intel-style cache/tlb description
2565  *
2566  * Standard cpuid level 2 gives a randomly ordered
2567  * selection of tags that index into a table that describes
2568  * cache and tlb properties.
2569  */
2570 
2571 static const char l1_icache_str[] = "l1-icache";
2572 static const char l1_dcache_str[] = "l1-dcache";
2573 static const char l2_cache_str[] = "l2-cache";
2574 static const char l3_cache_str[] = "l3-cache";
2575 static const char itlb4k_str[] = "itlb-4K";
2576 static const char dtlb4k_str[] = "dtlb-4K";
2577 static const char itlb4M_str[] = "itlb-4M";
2578 static const char dtlb4M_str[] = "dtlb-4M";
2579 static const char itlb424_str[] = "itlb-4K-2M-4M";
2580 static const char dtlb44_str[] = "dtlb-4K-4M";
2581 static const char sl1_dcache_str[] = "sectored-l1-dcache";
2582 static const char sl2_cache_str[] = "sectored-l2-cache";
2583 static const char itrace_str[] = "itrace-cache";
2584 static const char sl3_cache_str[] = "sectored-l3-cache";
2585 
2586 static const struct cachetab {
2587 	uint8_t 	ct_code;
2588 	uint8_t		ct_assoc;
2589 	uint16_t 	ct_line_size;
2590 	size_t		ct_size;
2591 	const char	*ct_label;
2592 } intel_ctab[] = {
2593 	/* maintain descending order! */
2594 	{ 0xb4, 4, 0, 256, dtlb4k_str },
2595 	{ 0xb3, 4, 0, 128, dtlb4k_str },
2596 	{ 0xb0, 4, 0, 128, itlb4k_str },
2597 	{ 0x87, 8, 64, 1024*1024, l2_cache_str},
2598 	{ 0x86, 4, 64, 512*1024, l2_cache_str},
2599 	{ 0x85, 8, 32, 2*1024*1024, l2_cache_str},
2600 	{ 0x84, 8, 32, 1024*1024, l2_cache_str},
2601 	{ 0x83, 8, 32, 512*1024, l2_cache_str},
2602 	{ 0x82, 8, 32, 256*1024, l2_cache_str},
2603 	{ 0x7f, 2, 64, 512*1024, l2_cache_str},
2604 	{ 0x7d, 8, 64, 2*1024*1024, sl2_cache_str},
2605 	{ 0x7c, 8, 64, 1024*1024, sl2_cache_str},
2606 	{ 0x7b, 8, 64, 512*1024, sl2_cache_str},
2607 	{ 0x7a, 8, 64, 256*1024, sl2_cache_str},
2608 	{ 0x79, 8, 64, 128*1024, sl2_cache_str},
2609 	{ 0x78, 8, 64, 1024*1024, l2_cache_str},
2610 	{ 0x73, 8, 0, 64*1024, itrace_str},
2611 	{ 0x72, 8, 0, 32*1024, itrace_str},
2612 	{ 0x71, 8, 0, 16*1024, itrace_str},
2613 	{ 0x70, 8, 0, 12*1024, itrace_str},
2614 	{ 0x68, 4, 64, 32*1024, sl1_dcache_str},
2615 	{ 0x67, 4, 64, 16*1024, sl1_dcache_str},
2616 	{ 0x66, 4, 64, 8*1024, sl1_dcache_str},
2617 	{ 0x60, 8, 64, 16*1024, sl1_dcache_str},
2618 	{ 0x5d, 0, 0, 256, dtlb44_str},
2619 	{ 0x5c, 0, 0, 128, dtlb44_str},
2620 	{ 0x5b, 0, 0, 64, dtlb44_str},
2621 	{ 0x52, 0, 0, 256, itlb424_str},
2622 	{ 0x51, 0, 0, 128, itlb424_str},
2623 	{ 0x50, 0, 0, 64, itlb424_str},
2624 	{ 0x4d, 16, 64, 16*1024*1024, l3_cache_str},
2625 	{ 0x4c, 12, 64, 12*1024*1024, l3_cache_str},
2626 	{ 0x4b, 16, 64, 8*1024*1024, l3_cache_str},
2627 	{ 0x4a, 12, 64, 6*1024*1024, l3_cache_str},
2628 	{ 0x49, 16, 64, 4*1024*1024, l3_cache_str},
2629 	{ 0x47, 8, 64, 8*1024*1024, l3_cache_str},
2630 	{ 0x46, 4, 64, 4*1024*1024, l3_cache_str},
2631 	{ 0x45, 4, 32, 2*1024*1024, l2_cache_str},
2632 	{ 0x44, 4, 32, 1024*1024, l2_cache_str},
2633 	{ 0x43, 4, 32, 512*1024, l2_cache_str},
2634 	{ 0x42, 4, 32, 256*1024, l2_cache_str},
2635 	{ 0x41, 4, 32, 128*1024, l2_cache_str},
2636 	{ 0x3e, 4, 64, 512*1024, sl2_cache_str},
2637 	{ 0x3d, 6, 64, 384*1024, sl2_cache_str},
2638 	{ 0x3c, 4, 64, 256*1024, sl2_cache_str},
2639 	{ 0x3b, 2, 64, 128*1024, sl2_cache_str},
2640 	{ 0x3a, 6, 64, 192*1024, sl2_cache_str},
2641 	{ 0x39, 4, 64, 128*1024, sl2_cache_str},
2642 	{ 0x30, 8, 64, 32*1024, l1_icache_str},
2643 	{ 0x2c, 8, 64, 32*1024, l1_dcache_str},
2644 	{ 0x29, 8, 64, 4096*1024, sl3_cache_str},
2645 	{ 0x25, 8, 64, 2048*1024, sl3_cache_str},
2646 	{ 0x23, 8, 64, 1024*1024, sl3_cache_str},
2647 	{ 0x22, 4, 64, 512*1024, sl3_cache_str},
2648 	{ 0x0c, 4, 32, 16*1024, l1_dcache_str},
2649 	{ 0x0b, 4, 0, 4, itlb4M_str},
2650 	{ 0x0a, 2, 32, 8*1024, l1_dcache_str},
2651 	{ 0x08, 4, 32, 16*1024, l1_icache_str},
2652 	{ 0x06, 4, 32, 8*1024, l1_icache_str},
2653 	{ 0x04, 4, 0, 8, dtlb4M_str},
2654 	{ 0x03, 4, 0, 64, dtlb4k_str},
2655 	{ 0x02, 4, 0, 2, itlb4M_str},
2656 	{ 0x01, 4, 0, 32, itlb4k_str},
2657 	{ 0 }
2658 };
2659 
2660 static const struct cachetab cyrix_ctab[] = {
2661 	{ 0x70, 4, 0, 32, "tlb-4K" },
2662 	{ 0x80, 4, 16, 16*1024, "l1-cache" },
2663 	{ 0 }
2664 };
2665 
2666 /*
2667  * Search a cache table for a matching entry
2668  */
2669 static const struct cachetab *
2670 find_cacheent(const struct cachetab *ct, uint_t code)
2671 {
2672 	if (code != 0) {
2673 		for (; ct->ct_code != 0; ct++)
2674 			if (ct->ct_code <= code)
2675 				break;
2676 		if (ct->ct_code == code)
2677 			return (ct);
2678 	}
2679 	return (NULL);
2680 }
2681 
2682 /*
2683  * Walk the cacheinfo descriptor, applying 'func' to every valid element
2684  * The walk is terminated if the walker returns non-zero.
2685  */
2686 static void
2687 intel_walk_cacheinfo(struct cpuid_info *cpi,
2688     void *arg, int (*func)(void *, const struct cachetab *))
2689 {
2690 	const struct cachetab *ct;
2691 	uint8_t *dp;
2692 	int i;
2693 
2694 	if ((dp = cpi->cpi_cacheinfo) == NULL)
2695 		return;
2696 	for (i = 0; i < cpi->cpi_ncache; i++, dp++) {
2697 		/*
2698 		 * For overloaded descriptor 0x49 we use cpuid function 4
2699 		 * if supported by the current processor, to update
2700 		 * cache information.
2701 		 */
2702 		if (*dp == 0x49 && cpi->cpi_maxeax >= 0x4) {
2703 			intel_cpuid_4_cache_info(arg, cpi);
2704 			continue;
2705 		}
2706 
2707 		if ((ct = find_cacheent(intel_ctab, *dp)) != NULL) {
2708 			if (func(arg, ct) != 0)
2709 				break;
2710 		}
2711 	}
2712 }
2713 
2714 /*
2715  * (Like the Intel one, except for Cyrix CPUs)
2716  */
2717 static void
2718 cyrix_walk_cacheinfo(struct cpuid_info *cpi,
2719     void *arg, int (*func)(void *, const struct cachetab *))
2720 {
2721 	const struct cachetab *ct;
2722 	uint8_t *dp;
2723 	int i;
2724 
2725 	if ((dp = cpi->cpi_cacheinfo) == NULL)
2726 		return;
2727 	for (i = 0; i < cpi->cpi_ncache; i++, dp++) {
2728 		/*
2729 		 * Search Cyrix-specific descriptor table first ..
2730 		 */
2731 		if ((ct = find_cacheent(cyrix_ctab, *dp)) != NULL) {
2732 			if (func(arg, ct) != 0)
2733 				break;
2734 			continue;
2735 		}
2736 		/*
2737 		 * .. else fall back to the Intel one
2738 		 */
2739 		if ((ct = find_cacheent(intel_ctab, *dp)) != NULL) {
2740 			if (func(arg, ct) != 0)
2741 				break;
2742 			continue;
2743 		}
2744 	}
2745 }
2746 
2747 /*
2748  * A cacheinfo walker that adds associativity, line-size, and size properties
2749  * to the devinfo node it is passed as an argument.
2750  */
2751 static int
2752 add_cacheent_props(void *arg, const struct cachetab *ct)
2753 {
2754 	dev_info_t *devi = arg;
2755 
2756 	add_cache_prop(devi, ct->ct_label, assoc_str, ct->ct_assoc);
2757 	if (ct->ct_line_size != 0)
2758 		add_cache_prop(devi, ct->ct_label, line_str,
2759 		    ct->ct_line_size);
2760 	add_cache_prop(devi, ct->ct_label, size_str, ct->ct_size);
2761 	return (0);
2762 }
2763 
2764 /*
2765  * Add L2 or L3 cache-information using cpuid function 4. This
2766  * function is called from intel_walk_cacheinfo() when descriptor
2767  * 0x49 is encountered.
2768  */
2769 static void
2770 intel_cpuid_4_cache_info(void *arg, struct cpuid_info *cpi)
2771 {
2772 	uint32_t level, i;
2773 
2774 	struct cachetab ct;
2775 
2776 	for (i = 0; i < cpi->cpi_std_4_size; i++) {
2777 		level = CPI_CACHE_LVL(cpi->cpi_std_4[i]);
2778 
2779 		if (level == 2 || level == 3) {
2780 			ct.ct_assoc = CPI_CACHE_WAYS(cpi->cpi_std_4[i]) + 1;
2781 			ct.ct_line_size =
2782 			    CPI_CACHE_COH_LN_SZ(cpi->cpi_std_4[i]) + 1;
2783 			ct.ct_size = ct.ct_assoc *
2784 			    (CPI_CACHE_PARTS(cpi->cpi_std_4[i]) + 1) *
2785 			    ct.ct_line_size *
2786 			    (cpi->cpi_std_4[i]->cp_ecx + 1);
2787 
2788 			if (level == 2) {
2789 				ct.ct_label = l2_cache_str;
2790 			} else if (level == 3) {
2791 				ct.ct_label = l3_cache_str;
2792 			}
2793 
2794 			(void) add_cacheent_props(arg,
2795 			    (const struct cachetab *) (&ct));
2796 		}
2797 	}
2798 }
2799 
2800 static const char fully_assoc[] = "fully-associative?";
2801 
2802 /*
2803  * AMD style cache/tlb description
2804  *
2805  * Extended functions 5 and 6 directly describe properties of
2806  * tlbs and various cache levels.
2807  */
2808 static void
2809 add_amd_assoc(dev_info_t *devi, const char *label, uint_t assoc)
2810 {
2811 	switch (assoc) {
2812 	case 0:	/* reserved; ignore */
2813 		break;
2814 	default:
2815 		add_cache_prop(devi, label, assoc_str, assoc);
2816 		break;
2817 	case 0xff:
2818 		add_cache_prop(devi, label, fully_assoc, 1);
2819 		break;
2820 	}
2821 }
2822 
2823 static void
2824 add_amd_tlb(dev_info_t *devi, const char *label, uint_t assoc, uint_t size)
2825 {
2826 	if (size == 0)
2827 		return;
2828 	add_cache_prop(devi, label, size_str, size);
2829 	add_amd_assoc(devi, label, assoc);
2830 }
2831 
2832 static void
2833 add_amd_cache(dev_info_t *devi, const char *label,
2834     uint_t size, uint_t assoc, uint_t lines_per_tag, uint_t line_size)
2835 {
2836 	if (size == 0 || line_size == 0)
2837 		return;
2838 	add_amd_assoc(devi, label, assoc);
2839 	/*
2840 	 * Most AMD parts have a sectored cache. Multiple cache lines are
2841 	 * associated with each tag. A sector consists of all cache lines
2842 	 * associated with a tag. For example, the AMD K6-III has a sector
2843 	 * size of 2 cache lines per tag.
2844 	 */
2845 	if (lines_per_tag != 0)
2846 		add_cache_prop(devi, label, "lines-per-tag", lines_per_tag);
2847 	add_cache_prop(devi, label, line_str, line_size);
2848 	add_cache_prop(devi, label, size_str, size * 1024);
2849 }
2850 
2851 static void
2852 add_amd_l2_assoc(dev_info_t *devi, const char *label, uint_t assoc)
2853 {
2854 	switch (assoc) {
2855 	case 0:	/* off */
2856 		break;
2857 	case 1:
2858 	case 2:
2859 	case 4:
2860 		add_cache_prop(devi, label, assoc_str, assoc);
2861 		break;
2862 	case 6:
2863 		add_cache_prop(devi, label, assoc_str, 8);
2864 		break;
2865 	case 8:
2866 		add_cache_prop(devi, label, assoc_str, 16);
2867 		break;
2868 	case 0xf:
2869 		add_cache_prop(devi, label, fully_assoc, 1);
2870 		break;
2871 	default: /* reserved; ignore */
2872 		break;
2873 	}
2874 }
2875 
2876 static void
2877 add_amd_l2_tlb(dev_info_t *devi, const char *label, uint_t assoc, uint_t size)
2878 {
2879 	if (size == 0 || assoc == 0)
2880 		return;
2881 	add_amd_l2_assoc(devi, label, assoc);
2882 	add_cache_prop(devi, label, size_str, size);
2883 }
2884 
2885 static void
2886 add_amd_l2_cache(dev_info_t *devi, const char *label,
2887     uint_t size, uint_t assoc, uint_t lines_per_tag, uint_t line_size)
2888 {
2889 	if (size == 0 || assoc == 0 || line_size == 0)
2890 		return;
2891 	add_amd_l2_assoc(devi, label, assoc);
2892 	if (lines_per_tag != 0)
2893 		add_cache_prop(devi, label, "lines-per-tag", lines_per_tag);
2894 	add_cache_prop(devi, label, line_str, line_size);
2895 	add_cache_prop(devi, label, size_str, size * 1024);
2896 }
2897 
2898 static void
2899 amd_cache_info(struct cpuid_info *cpi, dev_info_t *devi)
2900 {
2901 	struct cpuid_regs *cp;
2902 
2903 	if (cpi->cpi_xmaxeax < 0x80000005)
2904 		return;
2905 	cp = &cpi->cpi_extd[5];
2906 
2907 	/*
2908 	 * 4M/2M L1 TLB configuration
2909 	 *
2910 	 * We report the size for 2M pages because AMD uses two
2911 	 * TLB entries for one 4M page.
2912 	 */
2913 	add_amd_tlb(devi, "dtlb-2M",
2914 	    BITX(cp->cp_eax, 31, 24), BITX(cp->cp_eax, 23, 16));
2915 	add_amd_tlb(devi, "itlb-2M",
2916 	    BITX(cp->cp_eax, 15, 8), BITX(cp->cp_eax, 7, 0));
2917 
2918 	/*
2919 	 * 4K L1 TLB configuration
2920 	 */
2921 
2922 	switch (cpi->cpi_vendor) {
2923 		uint_t nentries;
2924 	case X86_VENDOR_TM:
2925 		if (cpi->cpi_family >= 5) {
2926 			/*
2927 			 * Crusoe processors have 256 TLB entries, but
2928 			 * cpuid data format constrains them to only
2929 			 * reporting 255 of them.
2930 			 */
2931 			if ((nentries = BITX(cp->cp_ebx, 23, 16)) == 255)
2932 				nentries = 256;
2933 			/*
2934 			 * Crusoe processors also have a unified TLB
2935 			 */
2936 			add_amd_tlb(devi, "tlb-4K", BITX(cp->cp_ebx, 31, 24),
2937 			    nentries);
2938 			break;
2939 		}
2940 		/*FALLTHROUGH*/
2941 	default:
2942 		add_amd_tlb(devi, itlb4k_str,
2943 		    BITX(cp->cp_ebx, 31, 24), BITX(cp->cp_ebx, 23, 16));
2944 		add_amd_tlb(devi, dtlb4k_str,
2945 		    BITX(cp->cp_ebx, 15, 8), BITX(cp->cp_ebx, 7, 0));
2946 		break;
2947 	}
2948 
2949 	/*
2950 	 * data L1 cache configuration
2951 	 */
2952 
2953 	add_amd_cache(devi, l1_dcache_str,
2954 	    BITX(cp->cp_ecx, 31, 24), BITX(cp->cp_ecx, 23, 16),
2955 	    BITX(cp->cp_ecx, 15, 8), BITX(cp->cp_ecx, 7, 0));
2956 
2957 	/*
2958 	 * code L1 cache configuration
2959 	 */
2960 
2961 	add_amd_cache(devi, l1_icache_str,
2962 	    BITX(cp->cp_edx, 31, 24), BITX(cp->cp_edx, 23, 16),
2963 	    BITX(cp->cp_edx, 15, 8), BITX(cp->cp_edx, 7, 0));
2964 
2965 	if (cpi->cpi_xmaxeax < 0x80000006)
2966 		return;
2967 	cp = &cpi->cpi_extd[6];
2968 
2969 	/* Check for a unified L2 TLB for large pages */
2970 
2971 	if (BITX(cp->cp_eax, 31, 16) == 0)
2972 		add_amd_l2_tlb(devi, "l2-tlb-2M",
2973 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
2974 	else {
2975 		add_amd_l2_tlb(devi, "l2-dtlb-2M",
2976 		    BITX(cp->cp_eax, 31, 28), BITX(cp->cp_eax, 27, 16));
2977 		add_amd_l2_tlb(devi, "l2-itlb-2M",
2978 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
2979 	}
2980 
2981 	/* Check for a unified L2 TLB for 4K pages */
2982 
2983 	if (BITX(cp->cp_ebx, 31, 16) == 0) {
2984 		add_amd_l2_tlb(devi, "l2-tlb-4K",
2985 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
2986 	} else {
2987 		add_amd_l2_tlb(devi, "l2-dtlb-4K",
2988 		    BITX(cp->cp_eax, 31, 28), BITX(cp->cp_eax, 27, 16));
2989 		add_amd_l2_tlb(devi, "l2-itlb-4K",
2990 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
2991 	}
2992 
2993 	add_amd_l2_cache(devi, l2_cache_str,
2994 	    BITX(cp->cp_ecx, 31, 16), BITX(cp->cp_ecx, 15, 12),
2995 	    BITX(cp->cp_ecx, 11, 8), BITX(cp->cp_ecx, 7, 0));
2996 }
2997 
2998 /*
2999  * There are two basic ways that the x86 world describes it cache
3000  * and tlb architecture - Intel's way and AMD's way.
3001  *
3002  * Return which flavor of cache architecture we should use
3003  */
3004 static int
3005 x86_which_cacheinfo(struct cpuid_info *cpi)
3006 {
3007 	switch (cpi->cpi_vendor) {
3008 	case X86_VENDOR_Intel:
3009 		if (cpi->cpi_maxeax >= 2)
3010 			return (X86_VENDOR_Intel);
3011 		break;
3012 	case X86_VENDOR_AMD:
3013 		/*
3014 		 * The K5 model 1 was the first part from AMD that reported
3015 		 * cache sizes via extended cpuid functions.
3016 		 */
3017 		if (cpi->cpi_family > 5 ||
3018 		    (cpi->cpi_family == 5 && cpi->cpi_model >= 1))
3019 			return (X86_VENDOR_AMD);
3020 		break;
3021 	case X86_VENDOR_TM:
3022 		if (cpi->cpi_family >= 5)
3023 			return (X86_VENDOR_AMD);
3024 		/*FALLTHROUGH*/
3025 	default:
3026 		/*
3027 		 * If they have extended CPU data for 0x80000005
3028 		 * then we assume they have AMD-format cache
3029 		 * information.
3030 		 *
3031 		 * If not, and the vendor happens to be Cyrix,
3032 		 * then try our-Cyrix specific handler.
3033 		 *
3034 		 * If we're not Cyrix, then assume we're using Intel's
3035 		 * table-driven format instead.
3036 		 */
3037 		if (cpi->cpi_xmaxeax >= 0x80000005)
3038 			return (X86_VENDOR_AMD);
3039 		else if (cpi->cpi_vendor == X86_VENDOR_Cyrix)
3040 			return (X86_VENDOR_Cyrix);
3041 		else if (cpi->cpi_maxeax >= 2)
3042 			return (X86_VENDOR_Intel);
3043 		break;
3044 	}
3045 	return (-1);
3046 }
3047 
3048 /*
3049  * create a node for the given cpu under the prom root node.
3050  * Also, create a cpu node in the device tree.
3051  */
3052 static dev_info_t *cpu_nex_devi = NULL;
3053 static kmutex_t cpu_node_lock;
3054 
3055 /*
3056  * Called from post_startup() and mp_startup()
3057  */
3058 void
3059 add_cpunode2devtree(processorid_t cpu_id, struct cpuid_info *cpi)
3060 {
3061 	dev_info_t *cpu_devi;
3062 	int create;
3063 
3064 	mutex_enter(&cpu_node_lock);
3065 
3066 	/*
3067 	 * create a nexus node for all cpus identified as 'cpu_id' under
3068 	 * the root node.
3069 	 */
3070 	if (cpu_nex_devi == NULL) {
3071 		if (ndi_devi_alloc(ddi_root_node(), "cpus",
3072 		    (pnode_t)DEVI_SID_NODEID, &cpu_nex_devi) != NDI_SUCCESS) {
3073 			mutex_exit(&cpu_node_lock);
3074 			return;
3075 		}
3076 		(void) ndi_devi_online(cpu_nex_devi, 0);
3077 	}
3078 
3079 	/*
3080 	 * create a child node for cpu identified as 'cpu_id'
3081 	 */
3082 	cpu_devi = ddi_add_child(cpu_nex_devi, "cpu", DEVI_SID_NODEID,
3083 	    cpu_id);
3084 	if (cpu_devi == NULL) {
3085 		mutex_exit(&cpu_node_lock);
3086 		return;
3087 	}
3088 
3089 	/* device_type */
3090 
3091 	(void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi,
3092 	    "device_type", "cpu");
3093 
3094 	/* reg */
3095 
3096 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3097 	    "reg", cpu_id);
3098 
3099 	/* cpu-mhz, and clock-frequency */
3100 
3101 	if (cpu_freq > 0) {
3102 		long long mul;
3103 
3104 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3105 		    "cpu-mhz", cpu_freq);
3106 
3107 		if ((mul = cpu_freq * 1000000LL) <= INT_MAX)
3108 			(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3109 			    "clock-frequency", (int)mul);
3110 	}
3111 
3112 	(void) ndi_devi_online(cpu_devi, 0);
3113 
3114 	if ((x86_feature & X86_CPUID) == 0) {
3115 		mutex_exit(&cpu_node_lock);
3116 		return;
3117 	}
3118 
3119 	/* vendor-id */
3120 
3121 	(void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi,
3122 	    "vendor-id", cpi->cpi_vendorstr);
3123 
3124 	if (cpi->cpi_maxeax == 0) {
3125 		mutex_exit(&cpu_node_lock);
3126 		return;
3127 	}
3128 
3129 	/*
3130 	 * family, model, and step
3131 	 */
3132 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3133 	    "family", CPI_FAMILY(cpi));
3134 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3135 	    "cpu-model", CPI_MODEL(cpi));
3136 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3137 	    "stepping-id", CPI_STEP(cpi));
3138 
3139 	/* type */
3140 
3141 	switch (cpi->cpi_vendor) {
3142 	case X86_VENDOR_Intel:
3143 		create = 1;
3144 		break;
3145 	default:
3146 		create = 0;
3147 		break;
3148 	}
3149 	if (create)
3150 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3151 		    "type", CPI_TYPE(cpi));
3152 
3153 	/* ext-family */
3154 
3155 	switch (cpi->cpi_vendor) {
3156 	case X86_VENDOR_Intel:
3157 	case X86_VENDOR_AMD:
3158 		create = cpi->cpi_family >= 0xf;
3159 		break;
3160 	default:
3161 		create = 0;
3162 		break;
3163 	}
3164 	if (create)
3165 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3166 		    "ext-family", CPI_FAMILY_XTD(cpi));
3167 
3168 	/* ext-model */
3169 
3170 	switch (cpi->cpi_vendor) {
3171 	case X86_VENDOR_Intel:
3172 		create = CPI_MODEL(cpi) == 0xf;
3173 		break;
3174 	case X86_VENDOR_AMD:
3175 		create = CPI_FAMILY(cpi) == 0xf;
3176 		break;
3177 	default:
3178 		create = 0;
3179 		break;
3180 	}
3181 	if (create)
3182 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3183 		    "ext-model", CPI_MODEL_XTD(cpi));
3184 
3185 	/* generation */
3186 
3187 	switch (cpi->cpi_vendor) {
3188 	case X86_VENDOR_AMD:
3189 		/*
3190 		 * AMD K5 model 1 was the first part to support this
3191 		 */
3192 		create = cpi->cpi_xmaxeax >= 0x80000001;
3193 		break;
3194 	default:
3195 		create = 0;
3196 		break;
3197 	}
3198 	if (create)
3199 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3200 		    "generation", BITX((cpi)->cpi_extd[1].cp_eax, 11, 8));
3201 
3202 	/* brand-id */
3203 
3204 	switch (cpi->cpi_vendor) {
3205 	case X86_VENDOR_Intel:
3206 		/*
3207 		 * brand id first appeared on Pentium III Xeon model 8,
3208 		 * and Celeron model 8 processors and Opteron
3209 		 */
3210 		create = cpi->cpi_family > 6 ||
3211 		    (cpi->cpi_family == 6 && cpi->cpi_model >= 8);
3212 		break;
3213 	case X86_VENDOR_AMD:
3214 		create = cpi->cpi_family >= 0xf;
3215 		break;
3216 	default:
3217 		create = 0;
3218 		break;
3219 	}
3220 	if (create && cpi->cpi_brandid != 0) {
3221 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3222 		    "brand-id", cpi->cpi_brandid);
3223 	}
3224 
3225 	/* chunks, and apic-id */
3226 
3227 	switch (cpi->cpi_vendor) {
3228 		/*
3229 		 * first available on Pentium IV and Opteron (K8)
3230 		 */
3231 	case X86_VENDOR_Intel:
3232 		create = IS_NEW_F6(cpi) || cpi->cpi_family >= 0xf;
3233 		break;
3234 	case X86_VENDOR_AMD:
3235 		create = cpi->cpi_family >= 0xf;
3236 		break;
3237 	default:
3238 		create = 0;
3239 		break;
3240 	}
3241 	if (create) {
3242 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3243 		    "chunks", CPI_CHUNKS(cpi));
3244 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3245 		    "apic-id", CPI_APIC_ID(cpi));
3246 		if (cpi->cpi_chipid >= 0) {
3247 			(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3248 			    "chip#", cpi->cpi_chipid);
3249 			(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3250 			    "clog#", cpi->cpi_clogid);
3251 		}
3252 	}
3253 
3254 	/* cpuid-features */
3255 
3256 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3257 	    "cpuid-features", CPI_FEATURES_EDX(cpi));
3258 
3259 
3260 	/* cpuid-features-ecx */
3261 
3262 	switch (cpi->cpi_vendor) {
3263 	case X86_VENDOR_Intel:
3264 		create = IS_NEW_F6(cpi) || cpi->cpi_family >= 0xf;
3265 		break;
3266 	default:
3267 		create = 0;
3268 		break;
3269 	}
3270 	if (create)
3271 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3272 		    "cpuid-features-ecx", CPI_FEATURES_ECX(cpi));
3273 
3274 	/* ext-cpuid-features */
3275 
3276 	switch (cpi->cpi_vendor) {
3277 	case X86_VENDOR_Intel:
3278 	case X86_VENDOR_AMD:
3279 	case X86_VENDOR_Cyrix:
3280 	case X86_VENDOR_TM:
3281 	case X86_VENDOR_Centaur:
3282 		create = cpi->cpi_xmaxeax >= 0x80000001;
3283 		break;
3284 	default:
3285 		create = 0;
3286 		break;
3287 	}
3288 	if (create) {
3289 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3290 		    "ext-cpuid-features", CPI_FEATURES_XTD_EDX(cpi));
3291 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3292 		    "ext-cpuid-features-ecx", CPI_FEATURES_XTD_ECX(cpi));
3293 	}
3294 
3295 	/*
3296 	 * Brand String first appeared in Intel Pentium IV, AMD K5
3297 	 * model 1, and Cyrix GXm.  On earlier models we try and
3298 	 * simulate something similar .. so this string should always
3299 	 * same -something- about the processor, however lame.
3300 	 */
3301 	(void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi,
3302 	    "brand-string", cpi->cpi_brandstr);
3303 
3304 	/*
3305 	 * Finally, cache and tlb information
3306 	 */
3307 	switch (x86_which_cacheinfo(cpi)) {
3308 	case X86_VENDOR_Intel:
3309 		intel_walk_cacheinfo(cpi, cpu_devi, add_cacheent_props);
3310 		break;
3311 	case X86_VENDOR_Cyrix:
3312 		cyrix_walk_cacheinfo(cpi, cpu_devi, add_cacheent_props);
3313 		break;
3314 	case X86_VENDOR_AMD:
3315 		amd_cache_info(cpi, cpu_devi);
3316 		break;
3317 	default:
3318 		break;
3319 	}
3320 
3321 	mutex_exit(&cpu_node_lock);
3322 }
3323 
3324 struct l2info {
3325 	int *l2i_csz;
3326 	int *l2i_lsz;
3327 	int *l2i_assoc;
3328 	int l2i_ret;
3329 };
3330 
3331 /*
3332  * A cacheinfo walker that fetches the size, line-size and associativity
3333  * of the L2 cache
3334  */
3335 static int
3336 intel_l2cinfo(void *arg, const struct cachetab *ct)
3337 {
3338 	struct l2info *l2i = arg;
3339 	int *ip;
3340 
3341 	if (ct->ct_label != l2_cache_str &&
3342 	    ct->ct_label != sl2_cache_str)
3343 		return (0);	/* not an L2 -- keep walking */
3344 
3345 	if ((ip = l2i->l2i_csz) != NULL)
3346 		*ip = ct->ct_size;
3347 	if ((ip = l2i->l2i_lsz) != NULL)
3348 		*ip = ct->ct_line_size;
3349 	if ((ip = l2i->l2i_assoc) != NULL)
3350 		*ip = ct->ct_assoc;
3351 	l2i->l2i_ret = ct->ct_size;
3352 	return (1);		/* was an L2 -- terminate walk */
3353 }
3354 
3355 static void
3356 amd_l2cacheinfo(struct cpuid_info *cpi, struct l2info *l2i)
3357 {
3358 	struct cpuid_regs *cp;
3359 	uint_t size, assoc;
3360 	int *ip;
3361 
3362 	if (cpi->cpi_xmaxeax < 0x80000006)
3363 		return;
3364 	cp = &cpi->cpi_extd[6];
3365 
3366 	if ((assoc = BITX(cp->cp_ecx, 15, 12)) != 0 &&
3367 	    (size = BITX(cp->cp_ecx, 31, 16)) != 0) {
3368 		uint_t cachesz = size * 1024;
3369 
3370 
3371 		if ((ip = l2i->l2i_csz) != NULL)
3372 			*ip = cachesz;
3373 		if ((ip = l2i->l2i_lsz) != NULL)
3374 			*ip = BITX(cp->cp_ecx, 7, 0);
3375 		if ((ip = l2i->l2i_assoc) != NULL)
3376 			*ip = assoc;
3377 		l2i->l2i_ret = cachesz;
3378 	}
3379 }
3380 
3381 int
3382 getl2cacheinfo(cpu_t *cpu, int *csz, int *lsz, int *assoc)
3383 {
3384 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
3385 	struct l2info __l2info, *l2i = &__l2info;
3386 
3387 	l2i->l2i_csz = csz;
3388 	l2i->l2i_lsz = lsz;
3389 	l2i->l2i_assoc = assoc;
3390 	l2i->l2i_ret = -1;
3391 
3392 	switch (x86_which_cacheinfo(cpi)) {
3393 	case X86_VENDOR_Intel:
3394 		intel_walk_cacheinfo(cpi, l2i, intel_l2cinfo);
3395 		break;
3396 	case X86_VENDOR_Cyrix:
3397 		cyrix_walk_cacheinfo(cpi, l2i, intel_l2cinfo);
3398 		break;
3399 	case X86_VENDOR_AMD:
3400 		amd_l2cacheinfo(cpi, l2i);
3401 		break;
3402 	default:
3403 		break;
3404 	}
3405 	return (l2i->l2i_ret);
3406 }
3407 
3408 size_t
3409 cpuid_get_mwait_size(cpu_t *cpu)
3410 {
3411 	ASSERT(cpuid_checkpass(cpu, 2));
3412 	return (cpu->cpu_m.mcpu_cpi->cpi_mwait.mon_max);
3413 }
3414