xref: /titanic_41/usr/src/uts/i86pc/os/cpuid.c (revision 25ccf31477c2a409aed8d0cf2027cd1073b0fe00)
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 2009 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 /*
27  * Various routines to handle identification
28  * and classification of x86 processors.
29  */
30 
31 #include <sys/types.h>
32 #include <sys/archsystm.h>
33 #include <sys/x86_archext.h>
34 #include <sys/kmem.h>
35 #include <sys/systm.h>
36 #include <sys/cmn_err.h>
37 #include <sys/sunddi.h>
38 #include <sys/sunndi.h>
39 #include <sys/cpuvar.h>
40 #include <sys/processor.h>
41 #include <sys/sysmacros.h>
42 #include <sys/pg.h>
43 #include <sys/fp.h>
44 #include <sys/controlregs.h>
45 #include <sys/auxv_386.h>
46 #include <sys/bitmap.h>
47 #include <sys/memnode.h>
48 
49 #ifdef __xpv
50 #include <sys/hypervisor.h>
51 #endif
52 
53 /*
54  * Pass 0 of cpuid feature analysis happens in locore. It contains special code
55  * to recognize Cyrix processors that are not cpuid-compliant, and to deal with
56  * them accordingly. For most modern processors, feature detection occurs here
57  * in pass 1.
58  *
59  * Pass 1 of cpuid feature analysis happens just at the beginning of mlsetup()
60  * for the boot CPU and does the basic analysis that the early kernel needs.
61  * x86_feature is set based on the return value of cpuid_pass1() of the boot
62  * CPU.
63  *
64  * Pass 1 includes:
65  *
66  *	o Determining vendor/model/family/stepping and setting x86_type and
67  *	  x86_vendor accordingly.
68  *	o Processing the feature flags returned by the cpuid instruction while
69  *	  applying any workarounds or tricks for the specific processor.
70  *	o Mapping the feature flags into Solaris feature bits (X86_*).
71  *	o Processing extended feature flags if supported by the processor,
72  *	  again while applying specific processor knowledge.
73  *	o Determining the CMT characteristics of the system.
74  *
75  * Pass 1 is done on non-boot CPUs during their initialization and the results
76  * are used only as a meager attempt at ensuring that all processors within the
77  * system support the same features.
78  *
79  * Pass 2 of cpuid feature analysis happens just at the beginning
80  * of startup().  It just copies in and corrects the remainder
81  * of the cpuid data we depend on: standard cpuid functions that we didn't
82  * need for pass1 feature analysis, and extended cpuid functions beyond the
83  * simple feature processing done in pass1.
84  *
85  * Pass 3 of cpuid analysis is invoked after basic kernel services; in
86  * particular kernel memory allocation has been made available. It creates a
87  * readable brand string based on the data collected in the first two passes.
88  *
89  * Pass 4 of cpuid analysis is invoked after post_startup() when all
90  * the support infrastructure for various hardware features has been
91  * initialized. It determines which processor features will be reported
92  * to userland via the aux vector.
93  *
94  * All passes are executed on all CPUs, but only the boot CPU determines what
95  * features the kernel will use.
96  *
97  * Much of the worst junk in this file is for the support of processors
98  * that didn't really implement the cpuid instruction properly.
99  *
100  * NOTE: The accessor functions (cpuid_get*) are aware of, and ASSERT upon,
101  * the pass numbers.  Accordingly, changes to the pass code may require changes
102  * to the accessor code.
103  */
104 
105 uint_t x86_feature = 0;
106 uint_t x86_vendor = X86_VENDOR_IntelClone;
107 uint_t x86_type = X86_TYPE_OTHER;
108 uint_t x86_clflush_size = 0;
109 
110 uint_t pentiumpro_bug4046376;
111 uint_t pentiumpro_bug4064495;
112 
113 uint_t enable486;
114 
115 /*
116  * monitor/mwait info.
117  *
118  * size_actual and buf_actual are the real address and size allocated to get
119  * proper mwait_buf alignement.  buf_actual and size_actual should be passed
120  * to kmem_free().  Currently kmem_alloc() and mwait happen to both use
121  * processor cache-line alignment, but this is not guarantied in the furture.
122  */
123 struct mwait_info {
124 	size_t		mon_min;	/* min size to avoid missed wakeups */
125 	size_t		mon_max;	/* size to avoid false wakeups */
126 	size_t		size_actual;	/* size actually allocated */
127 	void		*buf_actual;	/* memory actually allocated */
128 	uint32_t	support;	/* processor support of monitor/mwait */
129 };
130 
131 /*
132  * These constants determine how many of the elements of the
133  * cpuid we cache in the cpuid_info data structure; the
134  * remaining elements are accessible via the cpuid instruction.
135  */
136 
137 #define	NMAX_CPI_STD	6		/* eax = 0 .. 5 */
138 #define	NMAX_CPI_EXTD	9		/* eax = 0x80000000 .. 0x80000008 */
139 
140 struct cpuid_info {
141 	uint_t cpi_pass;		/* last pass completed */
142 	/*
143 	 * standard function information
144 	 */
145 	uint_t cpi_maxeax;		/* fn 0: %eax */
146 	char cpi_vendorstr[13];		/* fn 0: %ebx:%ecx:%edx */
147 	uint_t cpi_vendor;		/* enum of cpi_vendorstr */
148 
149 	uint_t cpi_family;		/* fn 1: extended family */
150 	uint_t cpi_model;		/* fn 1: extended model */
151 	uint_t cpi_step;		/* fn 1: stepping */
152 	chipid_t cpi_chipid;		/* fn 1: %ebx: chip # on ht cpus */
153 	uint_t cpi_brandid;		/* fn 1: %ebx: brand ID */
154 	int cpi_clogid;			/* fn 1: %ebx: thread # */
155 	uint_t cpi_ncpu_per_chip;	/* fn 1: %ebx: logical cpu count */
156 	uint8_t cpi_cacheinfo[16];	/* fn 2: intel-style cache desc */
157 	uint_t cpi_ncache;		/* fn 2: number of elements */
158 	uint_t cpi_ncpu_shr_last_cache;	/* fn 4: %eax: ncpus sharing cache */
159 	id_t cpi_last_lvl_cacheid;	/* fn 4: %eax: derived cache id */
160 	uint_t cpi_std_4_size;		/* fn 4: number of fn 4 elements */
161 	struct cpuid_regs **cpi_std_4;	/* fn 4: %ecx == 0 .. fn4_size */
162 	struct cpuid_regs cpi_std[NMAX_CPI_STD];	/* 0 .. 5 */
163 	/*
164 	 * extended function information
165 	 */
166 	uint_t cpi_xmaxeax;		/* fn 0x80000000: %eax */
167 	char cpi_brandstr[49];		/* fn 0x8000000[234] */
168 	uint8_t cpi_pabits;		/* fn 0x80000006: %eax */
169 	uint8_t cpi_vabits;		/* fn 0x80000006: %eax */
170 	struct cpuid_regs cpi_extd[NMAX_CPI_EXTD]; /* 0x8000000[0-8] */
171 	id_t cpi_coreid;		/* same coreid => strands share core */
172 	int cpi_pkgcoreid;		/* core number within single package */
173 	uint_t cpi_ncore_per_chip;	/* AMD: fn 0x80000008: %ecx[7-0] */
174 					/* Intel: fn 4: %eax[31-26] */
175 	/*
176 	 * supported feature information
177 	 */
178 	uint32_t cpi_support[5];
179 #define	STD_EDX_FEATURES	0
180 #define	AMD_EDX_FEATURES	1
181 #define	TM_EDX_FEATURES		2
182 #define	STD_ECX_FEATURES	3
183 #define	AMD_ECX_FEATURES	4
184 	/*
185 	 * Synthesized information, where known.
186 	 */
187 	uint32_t cpi_chiprev;		/* See X86_CHIPREV_* in x86_archext.h */
188 	const char *cpi_chiprevstr;	/* May be NULL if chiprev unknown */
189 	uint32_t cpi_socket;		/* Chip package/socket type */
190 
191 	struct mwait_info cpi_mwait;	/* fn 5: monitor/mwait info */
192 	uint32_t cpi_apicid;
193 };
194 
195 
196 static struct cpuid_info cpuid_info0;
197 
198 /*
199  * These bit fields are defined by the Intel Application Note AP-485
200  * "Intel Processor Identification and the CPUID Instruction"
201  */
202 #define	CPI_FAMILY_XTD(cpi)	BITX((cpi)->cpi_std[1].cp_eax, 27, 20)
203 #define	CPI_MODEL_XTD(cpi)	BITX((cpi)->cpi_std[1].cp_eax, 19, 16)
204 #define	CPI_TYPE(cpi)		BITX((cpi)->cpi_std[1].cp_eax, 13, 12)
205 #define	CPI_FAMILY(cpi)		BITX((cpi)->cpi_std[1].cp_eax, 11, 8)
206 #define	CPI_STEP(cpi)		BITX((cpi)->cpi_std[1].cp_eax, 3, 0)
207 #define	CPI_MODEL(cpi)		BITX((cpi)->cpi_std[1].cp_eax, 7, 4)
208 
209 #define	CPI_FEATURES_EDX(cpi)		((cpi)->cpi_std[1].cp_edx)
210 #define	CPI_FEATURES_ECX(cpi)		((cpi)->cpi_std[1].cp_ecx)
211 #define	CPI_FEATURES_XTD_EDX(cpi)	((cpi)->cpi_extd[1].cp_edx)
212 #define	CPI_FEATURES_XTD_ECX(cpi)	((cpi)->cpi_extd[1].cp_ecx)
213 
214 #define	CPI_BRANDID(cpi)	BITX((cpi)->cpi_std[1].cp_ebx, 7, 0)
215 #define	CPI_CHUNKS(cpi)		BITX((cpi)->cpi_std[1].cp_ebx, 15, 7)
216 #define	CPI_CPU_COUNT(cpi)	BITX((cpi)->cpi_std[1].cp_ebx, 23, 16)
217 #define	CPI_APIC_ID(cpi)	BITX((cpi)->cpi_std[1].cp_ebx, 31, 24)
218 
219 #define	CPI_MAXEAX_MAX		0x100		/* sanity control */
220 #define	CPI_XMAXEAX_MAX		0x80000100
221 #define	CPI_FN4_ECX_MAX		0x20		/* sanity: max fn 4 levels */
222 #define	CPI_FNB_ECX_MAX		0x20		/* sanity: max fn B levels */
223 
224 /*
225  * Function 4 (Deterministic Cache Parameters) macros
226  * Defined by Intel Application Note AP-485
227  */
228 #define	CPI_NUM_CORES(regs)		BITX((regs)->cp_eax, 31, 26)
229 #define	CPI_NTHR_SHR_CACHE(regs)	BITX((regs)->cp_eax, 25, 14)
230 #define	CPI_FULL_ASSOC_CACHE(regs)	BITX((regs)->cp_eax, 9, 9)
231 #define	CPI_SELF_INIT_CACHE(regs)	BITX((regs)->cp_eax, 8, 8)
232 #define	CPI_CACHE_LVL(regs)		BITX((regs)->cp_eax, 7, 5)
233 #define	CPI_CACHE_TYPE(regs)		BITX((regs)->cp_eax, 4, 0)
234 #define	CPI_CPU_LEVEL_TYPE(regs)	BITX((regs)->cp_ecx, 15, 8)
235 
236 #define	CPI_CACHE_WAYS(regs)		BITX((regs)->cp_ebx, 31, 22)
237 #define	CPI_CACHE_PARTS(regs)		BITX((regs)->cp_ebx, 21, 12)
238 #define	CPI_CACHE_COH_LN_SZ(regs)	BITX((regs)->cp_ebx, 11, 0)
239 
240 #define	CPI_CACHE_SETS(regs)		BITX((regs)->cp_ecx, 31, 0)
241 
242 #define	CPI_PREFCH_STRIDE(regs)		BITX((regs)->cp_edx, 9, 0)
243 
244 
245 /*
246  * A couple of shorthand macros to identify "later" P6-family chips
247  * like the Pentium M and Core.  First, the "older" P6-based stuff
248  * (loosely defined as "pre-Pentium-4"):
249  * P6, PII, Mobile PII, PII Xeon, PIII, Mobile PIII, PIII Xeon
250  */
251 
252 #define	IS_LEGACY_P6(cpi) (			\
253 	cpi->cpi_family == 6 && 		\
254 		(cpi->cpi_model == 1 ||		\
255 		cpi->cpi_model == 3 ||		\
256 		cpi->cpi_model == 5 ||		\
257 		cpi->cpi_model == 6 ||		\
258 		cpi->cpi_model == 7 ||		\
259 		cpi->cpi_model == 8 ||		\
260 		cpi->cpi_model == 0xA ||	\
261 		cpi->cpi_model == 0xB)		\
262 )
263 
264 /* A "new F6" is everything with family 6 that's not the above */
265 #define	IS_NEW_F6(cpi) ((cpi->cpi_family == 6) && !IS_LEGACY_P6(cpi))
266 
267 /* Extended family/model support */
268 #define	IS_EXTENDED_MODEL_INTEL(cpi) (cpi->cpi_family == 0x6 || \
269 	cpi->cpi_family >= 0xf)
270 
271 /*
272  * Info for monitor/mwait idle loop.
273  *
274  * See cpuid section of "Intel 64 and IA-32 Architectures Software Developer's
275  * Manual Volume 2A: Instruction Set Reference, A-M" #25366-022US, November
276  * 2006.
277  * See MONITOR/MWAIT section of "AMD64 Architecture Programmer's Manual
278  * Documentation Updates" #33633, Rev 2.05, December 2006.
279  */
280 #define	MWAIT_SUPPORT		(0x00000001)	/* mwait supported */
281 #define	MWAIT_EXTENSIONS	(0x00000002)	/* extenstion supported */
282 #define	MWAIT_ECX_INT_ENABLE	(0x00000004)	/* ecx 1 extension supported */
283 #define	MWAIT_SUPPORTED(cpi)	((cpi)->cpi_std[1].cp_ecx & CPUID_INTC_ECX_MON)
284 #define	MWAIT_INT_ENABLE(cpi)	((cpi)->cpi_std[5].cp_ecx & 0x2)
285 #define	MWAIT_EXTENSION(cpi)	((cpi)->cpi_std[5].cp_ecx & 0x1)
286 #define	MWAIT_SIZE_MIN(cpi)	BITX((cpi)->cpi_std[5].cp_eax, 15, 0)
287 #define	MWAIT_SIZE_MAX(cpi)	BITX((cpi)->cpi_std[5].cp_ebx, 15, 0)
288 /*
289  * Number of sub-cstates for a given c-state.
290  */
291 #define	MWAIT_NUM_SUBC_STATES(cpi, c_state)			\
292 	BITX((cpi)->cpi_std[5].cp_edx, c_state + 3, c_state)
293 
294 /*
295  * Functions we consune from cpuid_subr.c;  don't publish these in a header
296  * file to try and keep people using the expected cpuid_* interfaces.
297  */
298 extern uint32_t _cpuid_skt(uint_t, uint_t, uint_t, uint_t);
299 extern uint32_t _cpuid_chiprev(uint_t, uint_t, uint_t, uint_t);
300 extern const char *_cpuid_chiprevstr(uint_t, uint_t, uint_t, uint_t);
301 extern uint_t _cpuid_vendorstr_to_vendorcode(char *);
302 
303 /*
304  * Apply up various platform-dependent restrictions where the
305  * underlying platform restrictions mean the CPU can be marked
306  * as less capable than its cpuid instruction would imply.
307  */
308 #if defined(__xpv)
309 static void
310 platform_cpuid_mangle(uint_t vendor, uint32_t eax, struct cpuid_regs *cp)
311 {
312 	switch (eax) {
313 	case 1: {
314 		uint32_t mcamask = DOMAIN_IS_INITDOMAIN(xen_info) ?
315 		    0 : CPUID_INTC_EDX_MCA;
316 		cp->cp_edx &=
317 		    ~(mcamask |
318 		    CPUID_INTC_EDX_PSE |
319 		    CPUID_INTC_EDX_VME | CPUID_INTC_EDX_DE |
320 		    CPUID_INTC_EDX_SEP | CPUID_INTC_EDX_MTRR |
321 		    CPUID_INTC_EDX_PGE | CPUID_INTC_EDX_PAT |
322 		    CPUID_AMD_EDX_SYSC | CPUID_INTC_EDX_SEP |
323 		    CPUID_INTC_EDX_PSE36 | CPUID_INTC_EDX_HTT);
324 		break;
325 	}
326 
327 	case 0x80000001:
328 		cp->cp_edx &=
329 		    ~(CPUID_AMD_EDX_PSE |
330 		    CPUID_INTC_EDX_VME | CPUID_INTC_EDX_DE |
331 		    CPUID_AMD_EDX_MTRR | CPUID_AMD_EDX_PGE |
332 		    CPUID_AMD_EDX_PAT | CPUID_AMD_EDX_PSE36 |
333 		    CPUID_AMD_EDX_SYSC | CPUID_INTC_EDX_SEP |
334 		    CPUID_AMD_EDX_TSCP);
335 		cp->cp_ecx &= ~CPUID_AMD_ECX_CMP_LGCY;
336 		break;
337 	default:
338 		break;
339 	}
340 
341 	switch (vendor) {
342 	case X86_VENDOR_Intel:
343 		switch (eax) {
344 		case 4:
345 			/*
346 			 * Zero out the (ncores-per-chip - 1) field
347 			 */
348 			cp->cp_eax &= 0x03fffffff;
349 			break;
350 		default:
351 			break;
352 		}
353 		break;
354 	case X86_VENDOR_AMD:
355 		switch (eax) {
356 		case 0x80000008:
357 			/*
358 			 * Zero out the (ncores-per-chip - 1) field
359 			 */
360 			cp->cp_ecx &= 0xffffff00;
361 			break;
362 		default:
363 			break;
364 		}
365 		break;
366 	default:
367 		break;
368 	}
369 }
370 #else
371 #define	platform_cpuid_mangle(vendor, eax, cp)	/* nothing */
372 #endif
373 
374 /*
375  *  Some undocumented ways of patching the results of the cpuid
376  *  instruction to permit running Solaris 10 on future cpus that
377  *  we don't currently support.  Could be set to non-zero values
378  *  via settings in eeprom.
379  */
380 
381 uint32_t cpuid_feature_ecx_include;
382 uint32_t cpuid_feature_ecx_exclude;
383 uint32_t cpuid_feature_edx_include;
384 uint32_t cpuid_feature_edx_exclude;
385 
386 void
387 cpuid_alloc_space(cpu_t *cpu)
388 {
389 	/*
390 	 * By convention, cpu0 is the boot cpu, which is set up
391 	 * before memory allocation is available.  All other cpus get
392 	 * their cpuid_info struct allocated here.
393 	 */
394 	ASSERT(cpu->cpu_id != 0);
395 	cpu->cpu_m.mcpu_cpi =
396 	    kmem_zalloc(sizeof (*cpu->cpu_m.mcpu_cpi), KM_SLEEP);
397 }
398 
399 void
400 cpuid_free_space(cpu_t *cpu)
401 {
402 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
403 	int i;
404 
405 	ASSERT(cpu->cpu_id != 0);
406 
407 	/*
408 	 * Free up any function 4 related dynamic storage
409 	 */
410 	for (i = 1; i < cpi->cpi_std_4_size; i++)
411 		kmem_free(cpi->cpi_std_4[i], sizeof (struct cpuid_regs));
412 	if (cpi->cpi_std_4_size > 0)
413 		kmem_free(cpi->cpi_std_4,
414 		    cpi->cpi_std_4_size * sizeof (struct cpuid_regs *));
415 
416 	kmem_free(cpu->cpu_m.mcpu_cpi, sizeof (*cpu->cpu_m.mcpu_cpi));
417 }
418 
419 #if !defined(__xpv)
420 
421 static void
422 check_for_hvm()
423 {
424 	struct cpuid_regs cp;
425 	char *xen_str;
426 	uint32_t xen_signature[4];
427 	extern int xpv_is_hvm;
428 
429 	/*
430 	 * In a fully virtualized domain, Xen's pseudo-cpuid function
431 	 * 0x40000000 returns a string representing the Xen signature in
432 	 * %ebx, %ecx, and %edx.  %eax contains the maximum supported cpuid
433 	 * function.
434 	 */
435 	cp.cp_eax = 0x40000000;
436 	(void) __cpuid_insn(&cp);
437 	xen_signature[0] = cp.cp_ebx;
438 	xen_signature[1] = cp.cp_ecx;
439 	xen_signature[2] = cp.cp_edx;
440 	xen_signature[3] = 0;
441 	xen_str = (char *)xen_signature;
442 	if (strcmp("XenVMMXenVMM", xen_str) == 0 && cp.cp_eax <= 0x40000002)
443 		xpv_is_hvm = 1;
444 }
445 #endif	/* __xpv */
446 
447 uint_t
448 cpuid_pass1(cpu_t *cpu)
449 {
450 	uint32_t mask_ecx, mask_edx;
451 	uint_t feature = X86_CPUID;
452 	struct cpuid_info *cpi;
453 	struct cpuid_regs *cp;
454 	int xcpuid;
455 #if !defined(__xpv)
456 	extern int idle_cpu_prefer_mwait;
457 #endif
458 
459 	/*
460 	 * Space statically allocated for cpu0, ensure pointer is set
461 	 */
462 	if (cpu->cpu_id == 0)
463 		cpu->cpu_m.mcpu_cpi = &cpuid_info0;
464 	cpi = cpu->cpu_m.mcpu_cpi;
465 	ASSERT(cpi != NULL);
466 	cp = &cpi->cpi_std[0];
467 	cp->cp_eax = 0;
468 	cpi->cpi_maxeax = __cpuid_insn(cp);
469 	{
470 		uint32_t *iptr = (uint32_t *)cpi->cpi_vendorstr;
471 		*iptr++ = cp->cp_ebx;
472 		*iptr++ = cp->cp_edx;
473 		*iptr++ = cp->cp_ecx;
474 		*(char *)&cpi->cpi_vendorstr[12] = '\0';
475 	}
476 
477 	cpi->cpi_vendor = _cpuid_vendorstr_to_vendorcode(cpi->cpi_vendorstr);
478 	x86_vendor = cpi->cpi_vendor; /* for compatibility */
479 
480 	/*
481 	 * Limit the range in case of weird hardware
482 	 */
483 	if (cpi->cpi_maxeax > CPI_MAXEAX_MAX)
484 		cpi->cpi_maxeax = CPI_MAXEAX_MAX;
485 	if (cpi->cpi_maxeax < 1)
486 		goto pass1_done;
487 
488 	cp = &cpi->cpi_std[1];
489 	cp->cp_eax = 1;
490 	(void) __cpuid_insn(cp);
491 
492 	/*
493 	 * Extract identifying constants for easy access.
494 	 */
495 	cpi->cpi_model = CPI_MODEL(cpi);
496 	cpi->cpi_family = CPI_FAMILY(cpi);
497 
498 	if (cpi->cpi_family == 0xf)
499 		cpi->cpi_family += CPI_FAMILY_XTD(cpi);
500 
501 	/*
502 	 * Beware: AMD uses "extended model" iff base *FAMILY* == 0xf.
503 	 * Intel, and presumably everyone else, uses model == 0xf, as
504 	 * one would expect (max value means possible overflow).  Sigh.
505 	 */
506 
507 	switch (cpi->cpi_vendor) {
508 	case X86_VENDOR_Intel:
509 		if (IS_EXTENDED_MODEL_INTEL(cpi))
510 			cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4;
511 		break;
512 	case X86_VENDOR_AMD:
513 		if (CPI_FAMILY(cpi) == 0xf)
514 			cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4;
515 		break;
516 	default:
517 		if (cpi->cpi_model == 0xf)
518 			cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4;
519 		break;
520 	}
521 
522 	cpi->cpi_step = CPI_STEP(cpi);
523 	cpi->cpi_brandid = CPI_BRANDID(cpi);
524 
525 	/*
526 	 * *default* assumptions:
527 	 * - believe %edx feature word
528 	 * - ignore %ecx feature word
529 	 * - 32-bit virtual and physical addressing
530 	 */
531 	mask_edx = 0xffffffff;
532 	mask_ecx = 0;
533 
534 	cpi->cpi_pabits = cpi->cpi_vabits = 32;
535 
536 	switch (cpi->cpi_vendor) {
537 	case X86_VENDOR_Intel:
538 		if (cpi->cpi_family == 5)
539 			x86_type = X86_TYPE_P5;
540 		else if (IS_LEGACY_P6(cpi)) {
541 			x86_type = X86_TYPE_P6;
542 			pentiumpro_bug4046376 = 1;
543 			pentiumpro_bug4064495 = 1;
544 			/*
545 			 * Clear the SEP bit when it was set erroneously
546 			 */
547 			if (cpi->cpi_model < 3 && cpi->cpi_step < 3)
548 				cp->cp_edx &= ~CPUID_INTC_EDX_SEP;
549 		} else if (IS_NEW_F6(cpi) || cpi->cpi_family == 0xf) {
550 			x86_type = X86_TYPE_P4;
551 			/*
552 			 * We don't currently depend on any of the %ecx
553 			 * features until Prescott, so we'll only check
554 			 * this from P4 onwards.  We might want to revisit
555 			 * that idea later.
556 			 */
557 			mask_ecx = 0xffffffff;
558 		} else if (cpi->cpi_family > 0xf)
559 			mask_ecx = 0xffffffff;
560 		/*
561 		 * We don't support MONITOR/MWAIT if leaf 5 is not available
562 		 * to obtain the monitor linesize.
563 		 */
564 		if (cpi->cpi_maxeax < 5)
565 			mask_ecx &= ~CPUID_INTC_ECX_MON;
566 		break;
567 	case X86_VENDOR_IntelClone:
568 	default:
569 		break;
570 	case X86_VENDOR_AMD:
571 #if defined(OPTERON_ERRATUM_108)
572 		if (cpi->cpi_family == 0xf && cpi->cpi_model == 0xe) {
573 			cp->cp_eax = (0xf0f & cp->cp_eax) | 0xc0;
574 			cpi->cpi_model = 0xc;
575 		} else
576 #endif
577 		if (cpi->cpi_family == 5) {
578 			/*
579 			 * AMD K5 and K6
580 			 *
581 			 * These CPUs have an incomplete implementation
582 			 * of MCA/MCE which we mask away.
583 			 */
584 			mask_edx &= ~(CPUID_INTC_EDX_MCE | CPUID_INTC_EDX_MCA);
585 
586 			/*
587 			 * Model 0 uses the wrong (APIC) bit
588 			 * to indicate PGE.  Fix it here.
589 			 */
590 			if (cpi->cpi_model == 0) {
591 				if (cp->cp_edx & 0x200) {
592 					cp->cp_edx &= ~0x200;
593 					cp->cp_edx |= CPUID_INTC_EDX_PGE;
594 				}
595 			}
596 
597 			/*
598 			 * Early models had problems w/ MMX; disable.
599 			 */
600 			if (cpi->cpi_model < 6)
601 				mask_edx &= ~CPUID_INTC_EDX_MMX;
602 		}
603 
604 		/*
605 		 * For newer families, SSE3 and CX16, at least, are valid;
606 		 * enable all
607 		 */
608 		if (cpi->cpi_family >= 0xf)
609 			mask_ecx = 0xffffffff;
610 		/*
611 		 * We don't support MONITOR/MWAIT if leaf 5 is not available
612 		 * to obtain the monitor linesize.
613 		 */
614 		if (cpi->cpi_maxeax < 5)
615 			mask_ecx &= ~CPUID_INTC_ECX_MON;
616 
617 #if !defined(__xpv)
618 		/*
619 		 * Do not use MONITOR/MWAIT to halt in the idle loop on any AMD
620 		 * processors.  AMD does not intend MWAIT to be used in the cpu
621 		 * idle loop on current and future processors.  10h and future
622 		 * AMD processors use more power in MWAIT than HLT.
623 		 * Pre-family-10h Opterons do not have the MWAIT instruction.
624 		 */
625 		idle_cpu_prefer_mwait = 0;
626 #endif
627 
628 		break;
629 	case X86_VENDOR_TM:
630 		/*
631 		 * workaround the NT workaround in CMS 4.1
632 		 */
633 		if (cpi->cpi_family == 5 && cpi->cpi_model == 4 &&
634 		    (cpi->cpi_step == 2 || cpi->cpi_step == 3))
635 			cp->cp_edx |= CPUID_INTC_EDX_CX8;
636 		break;
637 	case X86_VENDOR_Centaur:
638 		/*
639 		 * workaround the NT workarounds again
640 		 */
641 		if (cpi->cpi_family == 6)
642 			cp->cp_edx |= CPUID_INTC_EDX_CX8;
643 		break;
644 	case X86_VENDOR_Cyrix:
645 		/*
646 		 * We rely heavily on the probing in locore
647 		 * to actually figure out what parts, if any,
648 		 * of the Cyrix cpuid instruction to believe.
649 		 */
650 		switch (x86_type) {
651 		case X86_TYPE_CYRIX_486:
652 			mask_edx = 0;
653 			break;
654 		case X86_TYPE_CYRIX_6x86:
655 			mask_edx = 0;
656 			break;
657 		case X86_TYPE_CYRIX_6x86L:
658 			mask_edx =
659 			    CPUID_INTC_EDX_DE |
660 			    CPUID_INTC_EDX_CX8;
661 			break;
662 		case X86_TYPE_CYRIX_6x86MX:
663 			mask_edx =
664 			    CPUID_INTC_EDX_DE |
665 			    CPUID_INTC_EDX_MSR |
666 			    CPUID_INTC_EDX_CX8 |
667 			    CPUID_INTC_EDX_PGE |
668 			    CPUID_INTC_EDX_CMOV |
669 			    CPUID_INTC_EDX_MMX;
670 			break;
671 		case X86_TYPE_CYRIX_GXm:
672 			mask_edx =
673 			    CPUID_INTC_EDX_MSR |
674 			    CPUID_INTC_EDX_CX8 |
675 			    CPUID_INTC_EDX_CMOV |
676 			    CPUID_INTC_EDX_MMX;
677 			break;
678 		case X86_TYPE_CYRIX_MediaGX:
679 			break;
680 		case X86_TYPE_CYRIX_MII:
681 		case X86_TYPE_VIA_CYRIX_III:
682 			mask_edx =
683 			    CPUID_INTC_EDX_DE |
684 			    CPUID_INTC_EDX_TSC |
685 			    CPUID_INTC_EDX_MSR |
686 			    CPUID_INTC_EDX_CX8 |
687 			    CPUID_INTC_EDX_PGE |
688 			    CPUID_INTC_EDX_CMOV |
689 			    CPUID_INTC_EDX_MMX;
690 			break;
691 		default:
692 			break;
693 		}
694 		break;
695 	}
696 
697 #if defined(__xpv)
698 	/*
699 	 * Do not support MONITOR/MWAIT under a hypervisor
700 	 */
701 	mask_ecx &= ~CPUID_INTC_ECX_MON;
702 #endif	/* __xpv */
703 
704 	/*
705 	 * Now we've figured out the masks that determine
706 	 * which bits we choose to believe, apply the masks
707 	 * to the feature words, then map the kernel's view
708 	 * of these feature words into its feature word.
709 	 */
710 	cp->cp_edx &= mask_edx;
711 	cp->cp_ecx &= mask_ecx;
712 
713 	/*
714 	 * apply any platform restrictions (we don't call this
715 	 * immediately after __cpuid_insn here, because we need the
716 	 * workarounds applied above first)
717 	 */
718 	platform_cpuid_mangle(cpi->cpi_vendor, 1, cp);
719 
720 	/*
721 	 * fold in overrides from the "eeprom" mechanism
722 	 */
723 	cp->cp_edx |= cpuid_feature_edx_include;
724 	cp->cp_edx &= ~cpuid_feature_edx_exclude;
725 
726 	cp->cp_ecx |= cpuid_feature_ecx_include;
727 	cp->cp_ecx &= ~cpuid_feature_ecx_exclude;
728 
729 	if (cp->cp_edx & CPUID_INTC_EDX_PSE)
730 		feature |= X86_LARGEPAGE;
731 	if (cp->cp_edx & CPUID_INTC_EDX_TSC)
732 		feature |= X86_TSC;
733 	if (cp->cp_edx & CPUID_INTC_EDX_MSR)
734 		feature |= X86_MSR;
735 	if (cp->cp_edx & CPUID_INTC_EDX_MTRR)
736 		feature |= X86_MTRR;
737 	if (cp->cp_edx & CPUID_INTC_EDX_PGE)
738 		feature |= X86_PGE;
739 	if (cp->cp_edx & CPUID_INTC_EDX_CMOV)
740 		feature |= X86_CMOV;
741 	if (cp->cp_edx & CPUID_INTC_EDX_MMX)
742 		feature |= X86_MMX;
743 	if ((cp->cp_edx & CPUID_INTC_EDX_MCE) != 0 &&
744 	    (cp->cp_edx & CPUID_INTC_EDX_MCA) != 0)
745 		feature |= X86_MCA;
746 	if (cp->cp_edx & CPUID_INTC_EDX_PAE)
747 		feature |= X86_PAE;
748 	if (cp->cp_edx & CPUID_INTC_EDX_CX8)
749 		feature |= X86_CX8;
750 	if (cp->cp_ecx & CPUID_INTC_ECX_CX16)
751 		feature |= X86_CX16;
752 	if (cp->cp_edx & CPUID_INTC_EDX_PAT)
753 		feature |= X86_PAT;
754 	if (cp->cp_edx & CPUID_INTC_EDX_SEP)
755 		feature |= X86_SEP;
756 	if (cp->cp_edx & CPUID_INTC_EDX_FXSR) {
757 		/*
758 		 * In our implementation, fxsave/fxrstor
759 		 * are prerequisites before we'll even
760 		 * try and do SSE things.
761 		 */
762 		if (cp->cp_edx & CPUID_INTC_EDX_SSE)
763 			feature |= X86_SSE;
764 		if (cp->cp_edx & CPUID_INTC_EDX_SSE2)
765 			feature |= X86_SSE2;
766 		if (cp->cp_ecx & CPUID_INTC_ECX_SSE3)
767 			feature |= X86_SSE3;
768 		if (cpi->cpi_vendor == X86_VENDOR_Intel) {
769 			if (cp->cp_ecx & CPUID_INTC_ECX_SSSE3)
770 				feature |= X86_SSSE3;
771 			if (cp->cp_ecx & CPUID_INTC_ECX_SSE4_1)
772 				feature |= X86_SSE4_1;
773 			if (cp->cp_ecx & CPUID_INTC_ECX_SSE4_2)
774 				feature |= X86_SSE4_2;
775 		}
776 	}
777 	if (cp->cp_edx & CPUID_INTC_EDX_DE)
778 		feature |= X86_DE;
779 #if !defined(__xpv)
780 	if (cp->cp_ecx & CPUID_INTC_ECX_MON) {
781 
782 		/*
783 		 * We require the CLFLUSH instruction for erratum workaround
784 		 * to use MONITOR/MWAIT.
785 		 */
786 		if (cp->cp_edx & CPUID_INTC_EDX_CLFSH) {
787 			cpi->cpi_mwait.support |= MWAIT_SUPPORT;
788 			feature |= X86_MWAIT;
789 		} else {
790 			extern int idle_cpu_assert_cflush_monitor;
791 
792 			/*
793 			 * All processors we are aware of which have
794 			 * MONITOR/MWAIT also have CLFLUSH.
795 			 */
796 			if (idle_cpu_assert_cflush_monitor) {
797 				ASSERT((cp->cp_ecx & CPUID_INTC_ECX_MON) &&
798 				    (cp->cp_edx & CPUID_INTC_EDX_CLFSH));
799 			}
800 		}
801 	}
802 #endif	/* __xpv */
803 
804 	/*
805 	 * Only need it first time, rest of the cpus would follow suite.
806 	 * we only capture this for the bootcpu.
807 	 */
808 	if (cp->cp_edx & CPUID_INTC_EDX_CLFSH) {
809 		feature |= X86_CLFSH;
810 		x86_clflush_size = (BITX(cp->cp_ebx, 15, 8) * 8);
811 	}
812 
813 	if (feature & X86_PAE)
814 		cpi->cpi_pabits = 36;
815 
816 	/*
817 	 * Hyperthreading configuration is slightly tricky on Intel
818 	 * and pure clones, and even trickier on AMD.
819 	 *
820 	 * (AMD chose to set the HTT bit on their CMP processors,
821 	 * even though they're not actually hyperthreaded.  Thus it
822 	 * takes a bit more work to figure out what's really going
823 	 * on ... see the handling of the CMP_LGCY bit below)
824 	 */
825 	if (cp->cp_edx & CPUID_INTC_EDX_HTT) {
826 		cpi->cpi_ncpu_per_chip = CPI_CPU_COUNT(cpi);
827 		if (cpi->cpi_ncpu_per_chip > 1)
828 			feature |= X86_HTT;
829 	} else {
830 		cpi->cpi_ncpu_per_chip = 1;
831 	}
832 
833 	/*
834 	 * Work on the "extended" feature information, doing
835 	 * some basic initialization for cpuid_pass2()
836 	 */
837 	xcpuid = 0;
838 	switch (cpi->cpi_vendor) {
839 	case X86_VENDOR_Intel:
840 		if (IS_NEW_F6(cpi) || cpi->cpi_family >= 0xf)
841 			xcpuid++;
842 		break;
843 	case X86_VENDOR_AMD:
844 		if (cpi->cpi_family > 5 ||
845 		    (cpi->cpi_family == 5 && cpi->cpi_model >= 1))
846 			xcpuid++;
847 		break;
848 	case X86_VENDOR_Cyrix:
849 		/*
850 		 * Only these Cyrix CPUs are -known- to support
851 		 * extended cpuid operations.
852 		 */
853 		if (x86_type == X86_TYPE_VIA_CYRIX_III ||
854 		    x86_type == X86_TYPE_CYRIX_GXm)
855 			xcpuid++;
856 		break;
857 	case X86_VENDOR_Centaur:
858 	case X86_VENDOR_TM:
859 	default:
860 		xcpuid++;
861 		break;
862 	}
863 
864 	if (xcpuid) {
865 		cp = &cpi->cpi_extd[0];
866 		cp->cp_eax = 0x80000000;
867 		cpi->cpi_xmaxeax = __cpuid_insn(cp);
868 	}
869 
870 	if (cpi->cpi_xmaxeax & 0x80000000) {
871 
872 		if (cpi->cpi_xmaxeax > CPI_XMAXEAX_MAX)
873 			cpi->cpi_xmaxeax = CPI_XMAXEAX_MAX;
874 
875 		switch (cpi->cpi_vendor) {
876 		case X86_VENDOR_Intel:
877 		case X86_VENDOR_AMD:
878 			if (cpi->cpi_xmaxeax < 0x80000001)
879 				break;
880 			cp = &cpi->cpi_extd[1];
881 			cp->cp_eax = 0x80000001;
882 			(void) __cpuid_insn(cp);
883 
884 			if (cpi->cpi_vendor == X86_VENDOR_AMD &&
885 			    cpi->cpi_family == 5 &&
886 			    cpi->cpi_model == 6 &&
887 			    cpi->cpi_step == 6) {
888 				/*
889 				 * K6 model 6 uses bit 10 to indicate SYSC
890 				 * Later models use bit 11. Fix it here.
891 				 */
892 				if (cp->cp_edx & 0x400) {
893 					cp->cp_edx &= ~0x400;
894 					cp->cp_edx |= CPUID_AMD_EDX_SYSC;
895 				}
896 			}
897 
898 			platform_cpuid_mangle(cpi->cpi_vendor, 0x80000001, cp);
899 
900 			/*
901 			 * Compute the additions to the kernel's feature word.
902 			 */
903 			if (cp->cp_edx & CPUID_AMD_EDX_NX)
904 				feature |= X86_NX;
905 
906 			/*
907 			 * Regardless whether or not we boot 64-bit,
908 			 * we should have a way to identify whether
909 			 * the CPU is capable of running 64-bit.
910 			 */
911 			if (cp->cp_edx & CPUID_AMD_EDX_LM)
912 				feature |= X86_64;
913 
914 #if defined(__amd64)
915 			/* 1 GB large page - enable only for 64 bit kernel */
916 			if (cp->cp_edx & CPUID_AMD_EDX_1GPG)
917 				feature |= X86_1GPG;
918 #endif
919 
920 			if ((cpi->cpi_vendor == X86_VENDOR_AMD) &&
921 			    (cpi->cpi_std[1].cp_edx & CPUID_INTC_EDX_FXSR) &&
922 			    (cp->cp_ecx & CPUID_AMD_ECX_SSE4A))
923 				feature |= X86_SSE4A;
924 
925 			/*
926 			 * If both the HTT and CMP_LGCY bits are set,
927 			 * then we're not actually HyperThreaded.  Read
928 			 * "AMD CPUID Specification" for more details.
929 			 */
930 			if (cpi->cpi_vendor == X86_VENDOR_AMD &&
931 			    (feature & X86_HTT) &&
932 			    (cp->cp_ecx & CPUID_AMD_ECX_CMP_LGCY)) {
933 				feature &= ~X86_HTT;
934 				feature |= X86_CMP;
935 			}
936 #if defined(__amd64)
937 			/*
938 			 * It's really tricky to support syscall/sysret in
939 			 * the i386 kernel; we rely on sysenter/sysexit
940 			 * instead.  In the amd64 kernel, things are -way-
941 			 * better.
942 			 */
943 			if (cp->cp_edx & CPUID_AMD_EDX_SYSC)
944 				feature |= X86_ASYSC;
945 
946 			/*
947 			 * While we're thinking about system calls, note
948 			 * that AMD processors don't support sysenter
949 			 * in long mode at all, so don't try to program them.
950 			 */
951 			if (x86_vendor == X86_VENDOR_AMD)
952 				feature &= ~X86_SEP;
953 #endif
954 			if (cp->cp_edx & CPUID_AMD_EDX_TSCP)
955 				feature |= X86_TSCP;
956 			break;
957 		default:
958 			break;
959 		}
960 
961 		/*
962 		 * Get CPUID data about processor cores and hyperthreads.
963 		 */
964 		switch (cpi->cpi_vendor) {
965 		case X86_VENDOR_Intel:
966 			if (cpi->cpi_maxeax >= 4) {
967 				cp = &cpi->cpi_std[4];
968 				cp->cp_eax = 4;
969 				cp->cp_ecx = 0;
970 				(void) __cpuid_insn(cp);
971 				platform_cpuid_mangle(cpi->cpi_vendor, 4, cp);
972 			}
973 			/*FALLTHROUGH*/
974 		case X86_VENDOR_AMD:
975 			if (cpi->cpi_xmaxeax < 0x80000008)
976 				break;
977 			cp = &cpi->cpi_extd[8];
978 			cp->cp_eax = 0x80000008;
979 			(void) __cpuid_insn(cp);
980 			platform_cpuid_mangle(cpi->cpi_vendor, 0x80000008, cp);
981 
982 			/*
983 			 * Virtual and physical address limits from
984 			 * cpuid override previously guessed values.
985 			 */
986 			cpi->cpi_pabits = BITX(cp->cp_eax, 7, 0);
987 			cpi->cpi_vabits = BITX(cp->cp_eax, 15, 8);
988 			break;
989 		default:
990 			break;
991 		}
992 
993 		/*
994 		 * Derive the number of cores per chip
995 		 */
996 		switch (cpi->cpi_vendor) {
997 		case X86_VENDOR_Intel:
998 			if (cpi->cpi_maxeax < 4) {
999 				cpi->cpi_ncore_per_chip = 1;
1000 				break;
1001 			} else {
1002 				cpi->cpi_ncore_per_chip =
1003 				    BITX((cpi)->cpi_std[4].cp_eax, 31, 26) + 1;
1004 			}
1005 			break;
1006 		case X86_VENDOR_AMD:
1007 			if (cpi->cpi_xmaxeax < 0x80000008) {
1008 				cpi->cpi_ncore_per_chip = 1;
1009 				break;
1010 			} else {
1011 				/*
1012 				 * On family 0xf cpuid fn 2 ECX[7:0] "NC" is
1013 				 * 1 less than the number of physical cores on
1014 				 * the chip.  In family 0x10 this value can
1015 				 * be affected by "downcoring" - it reflects
1016 				 * 1 less than the number of cores actually
1017 				 * enabled on this node.
1018 				 */
1019 				cpi->cpi_ncore_per_chip =
1020 				    BITX((cpi)->cpi_extd[8].cp_ecx, 7, 0) + 1;
1021 			}
1022 			break;
1023 		default:
1024 			cpi->cpi_ncore_per_chip = 1;
1025 			break;
1026 		}
1027 
1028 		/*
1029 		 * Get CPUID data about TSC Invariance in Deep C-State.
1030 		 */
1031 		switch (cpi->cpi_vendor) {
1032 		case X86_VENDOR_Intel:
1033 			if (cpi->cpi_maxeax >= 7) {
1034 				cp = &cpi->cpi_extd[7];
1035 				cp->cp_eax = 0x80000007;
1036 				cp->cp_ecx = 0;
1037 				(void) __cpuid_insn(cp);
1038 			}
1039 			break;
1040 		default:
1041 			break;
1042 		}
1043 	} else {
1044 		cpi->cpi_ncore_per_chip = 1;
1045 	}
1046 
1047 	/*
1048 	 * If more than one core, then this processor is CMP.
1049 	 */
1050 	if (cpi->cpi_ncore_per_chip > 1)
1051 		feature |= X86_CMP;
1052 
1053 	/*
1054 	 * If the number of cores is the same as the number
1055 	 * of CPUs, then we cannot have HyperThreading.
1056 	 */
1057 	if (cpi->cpi_ncpu_per_chip == cpi->cpi_ncore_per_chip)
1058 		feature &= ~X86_HTT;
1059 
1060 	if ((feature & (X86_HTT | X86_CMP)) == 0) {
1061 		/*
1062 		 * Single-core single-threaded processors.
1063 		 */
1064 		cpi->cpi_chipid = -1;
1065 		cpi->cpi_clogid = 0;
1066 		cpi->cpi_coreid = cpu->cpu_id;
1067 		cpi->cpi_pkgcoreid = 0;
1068 	} else if (cpi->cpi_ncpu_per_chip > 1) {
1069 		uint_t i;
1070 		uint_t chipid_shift = 0;
1071 		uint_t coreid_shift = 0;
1072 		uint_t apic_id = CPI_APIC_ID(cpi);
1073 
1074 		for (i = 1; i < cpi->cpi_ncpu_per_chip; i <<= 1)
1075 			chipid_shift++;
1076 		cpi->cpi_chipid = apic_id >> chipid_shift;
1077 		cpi->cpi_clogid = apic_id & ((1 << chipid_shift) - 1);
1078 
1079 		if (cpi->cpi_vendor == X86_VENDOR_Intel) {
1080 			if (feature & X86_CMP) {
1081 				/*
1082 				 * Multi-core (and possibly multi-threaded)
1083 				 * processors.
1084 				 */
1085 				uint_t ncpu_per_core;
1086 				if (cpi->cpi_ncore_per_chip == 1)
1087 					ncpu_per_core = cpi->cpi_ncpu_per_chip;
1088 				else if (cpi->cpi_ncore_per_chip > 1)
1089 					ncpu_per_core = cpi->cpi_ncpu_per_chip /
1090 					    cpi->cpi_ncore_per_chip;
1091 				/*
1092 				 * 8bit APIC IDs on dual core Pentiums
1093 				 * look like this:
1094 				 *
1095 				 * +-----------------------+------+------+
1096 				 * | Physical Package ID   |  MC  |  HT  |
1097 				 * +-----------------------+------+------+
1098 				 * <------- chipid -------->
1099 				 * <------- coreid --------------->
1100 				 *			   <--- clogid -->
1101 				 *			   <------>
1102 				 *			   pkgcoreid
1103 				 *
1104 				 * Where the number of bits necessary to
1105 				 * represent MC and HT fields together equals
1106 				 * to the minimum number of bits necessary to
1107 				 * store the value of cpi->cpi_ncpu_per_chip.
1108 				 * Of those bits, the MC part uses the number
1109 				 * of bits necessary to store the value of
1110 				 * cpi->cpi_ncore_per_chip.
1111 				 */
1112 				for (i = 1; i < ncpu_per_core; i <<= 1)
1113 					coreid_shift++;
1114 				cpi->cpi_coreid = apic_id >> coreid_shift;
1115 				cpi->cpi_pkgcoreid = cpi->cpi_clogid >>
1116 				    coreid_shift;
1117 			} else if (feature & X86_HTT) {
1118 				/*
1119 				 * Single-core multi-threaded processors.
1120 				 */
1121 				cpi->cpi_coreid = cpi->cpi_chipid;
1122 				cpi->cpi_pkgcoreid = 0;
1123 			}
1124 		} else if (cpi->cpi_vendor == X86_VENDOR_AMD) {
1125 			/*
1126 			 * AMD CMP chips currently have a single thread per
1127 			 * core, with 2 cores on family 0xf and 2, 3 or 4
1128 			 * cores on family 0x10.
1129 			 *
1130 			 * Since no two cpus share a core we must assign a
1131 			 * distinct coreid per cpu, and we do this by using
1132 			 * the cpu_id.  This scheme does not, however,
1133 			 * guarantee that sibling cores of a chip will have
1134 			 * sequential coreids starting at a multiple of the
1135 			 * number of cores per chip - that is usually the
1136 			 * case, but if the ACPI MADT table is presented
1137 			 * in a different order then we need to perform a
1138 			 * few more gymnastics for the pkgcoreid.
1139 			 *
1140 			 * In family 0xf CMPs there are 2 cores on all nodes
1141 			 * present - no mixing of single and dual core parts.
1142 			 *
1143 			 * In family 0x10 CMPs cpuid fn 2 ECX[15:12]
1144 			 * "ApicIdCoreIdSize[3:0]" tells us how
1145 			 * many least-significant bits in the ApicId
1146 			 * are used to represent the core number
1147 			 * within the node.  Cores are always
1148 			 * numbered sequentially from 0 regardless
1149 			 * of how many or which are disabled, and
1150 			 * there seems to be no way to discover the
1151 			 * real core id when some are disabled.
1152 			 */
1153 			cpi->cpi_coreid = cpu->cpu_id;
1154 
1155 			if (cpi->cpi_family == 0x10 &&
1156 			    cpi->cpi_xmaxeax >= 0x80000008) {
1157 				int coreidsz =
1158 				    BITX((cpi)->cpi_extd[8].cp_ecx, 15, 12);
1159 
1160 				cpi->cpi_pkgcoreid =
1161 				    apic_id & ((1 << coreidsz) - 1);
1162 			} else {
1163 				cpi->cpi_pkgcoreid = cpi->cpi_clogid;
1164 			}
1165 		} else {
1166 			/*
1167 			 * All other processors are currently
1168 			 * assumed to have single cores.
1169 			 */
1170 			cpi->cpi_coreid = cpi->cpi_chipid;
1171 			cpi->cpi_pkgcoreid = 0;
1172 		}
1173 	}
1174 
1175 	cpi->cpi_apicid = CPI_APIC_ID(cpi);
1176 
1177 	/*
1178 	 * Synthesize chip "revision" and socket type
1179 	 */
1180 	cpi->cpi_chiprev = _cpuid_chiprev(cpi->cpi_vendor, cpi->cpi_family,
1181 	    cpi->cpi_model, cpi->cpi_step);
1182 	cpi->cpi_chiprevstr = _cpuid_chiprevstr(cpi->cpi_vendor,
1183 	    cpi->cpi_family, cpi->cpi_model, cpi->cpi_step);
1184 	cpi->cpi_socket = _cpuid_skt(cpi->cpi_vendor, cpi->cpi_family,
1185 	    cpi->cpi_model, cpi->cpi_step);
1186 
1187 pass1_done:
1188 #if !defined(__xpv)
1189 	check_for_hvm();
1190 #endif
1191 	cpi->cpi_pass = 1;
1192 	return (feature);
1193 }
1194 
1195 /*
1196  * Make copies of the cpuid table entries we depend on, in
1197  * part for ease of parsing now, in part so that we have only
1198  * one place to correct any of it, in part for ease of
1199  * later export to userland, and in part so we can look at
1200  * this stuff in a crash dump.
1201  */
1202 
1203 /*ARGSUSED*/
1204 void
1205 cpuid_pass2(cpu_t *cpu)
1206 {
1207 	uint_t n, nmax;
1208 	int i;
1209 	struct cpuid_regs *cp;
1210 	uint8_t *dp;
1211 	uint32_t *iptr;
1212 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
1213 
1214 	ASSERT(cpi->cpi_pass == 1);
1215 
1216 	if (cpi->cpi_maxeax < 1)
1217 		goto pass2_done;
1218 
1219 	if ((nmax = cpi->cpi_maxeax + 1) > NMAX_CPI_STD)
1220 		nmax = NMAX_CPI_STD;
1221 	/*
1222 	 * (We already handled n == 0 and n == 1 in pass 1)
1223 	 */
1224 	for (n = 2, cp = &cpi->cpi_std[2]; n < nmax; n++, cp++) {
1225 		cp->cp_eax = n;
1226 
1227 		/*
1228 		 * CPUID function 4 expects %ecx to be initialized
1229 		 * with an index which indicates which cache to return
1230 		 * information about. The OS is expected to call function 4
1231 		 * with %ecx set to 0, 1, 2, ... until it returns with
1232 		 * EAX[4:0] set to 0, which indicates there are no more
1233 		 * caches.
1234 		 *
1235 		 * Here, populate cpi_std[4] with the information returned by
1236 		 * function 4 when %ecx == 0, and do the rest in cpuid_pass3()
1237 		 * when dynamic memory allocation becomes available.
1238 		 *
1239 		 * Note: we need to explicitly initialize %ecx here, since
1240 		 * function 4 may have been previously invoked.
1241 		 */
1242 		if (n == 4)
1243 			cp->cp_ecx = 0;
1244 
1245 		(void) __cpuid_insn(cp);
1246 		platform_cpuid_mangle(cpi->cpi_vendor, n, cp);
1247 		switch (n) {
1248 		case 2:
1249 			/*
1250 			 * "the lower 8 bits of the %eax register
1251 			 * contain a value that identifies the number
1252 			 * of times the cpuid [instruction] has to be
1253 			 * executed to obtain a complete image of the
1254 			 * processor's caching systems."
1255 			 *
1256 			 * How *do* they make this stuff up?
1257 			 */
1258 			cpi->cpi_ncache = sizeof (*cp) *
1259 			    BITX(cp->cp_eax, 7, 0);
1260 			if (cpi->cpi_ncache == 0)
1261 				break;
1262 			cpi->cpi_ncache--;	/* skip count byte */
1263 
1264 			/*
1265 			 * Well, for now, rather than attempt to implement
1266 			 * this slightly dubious algorithm, we just look
1267 			 * at the first 15 ..
1268 			 */
1269 			if (cpi->cpi_ncache > (sizeof (*cp) - 1))
1270 				cpi->cpi_ncache = sizeof (*cp) - 1;
1271 
1272 			dp = cpi->cpi_cacheinfo;
1273 			if (BITX(cp->cp_eax, 31, 31) == 0) {
1274 				uint8_t *p = (void *)&cp->cp_eax;
1275 				for (i = 1; i < 4; i++)
1276 					if (p[i] != 0)
1277 						*dp++ = p[i];
1278 			}
1279 			if (BITX(cp->cp_ebx, 31, 31) == 0) {
1280 				uint8_t *p = (void *)&cp->cp_ebx;
1281 				for (i = 0; i < 4; i++)
1282 					if (p[i] != 0)
1283 						*dp++ = p[i];
1284 			}
1285 			if (BITX(cp->cp_ecx, 31, 31) == 0) {
1286 				uint8_t *p = (void *)&cp->cp_ecx;
1287 				for (i = 0; i < 4; i++)
1288 					if (p[i] != 0)
1289 						*dp++ = p[i];
1290 			}
1291 			if (BITX(cp->cp_edx, 31, 31) == 0) {
1292 				uint8_t *p = (void *)&cp->cp_edx;
1293 				for (i = 0; i < 4; i++)
1294 					if (p[i] != 0)
1295 						*dp++ = p[i];
1296 			}
1297 			break;
1298 
1299 		case 3:	/* Processor serial number, if PSN supported */
1300 			break;
1301 
1302 		case 4:	/* Deterministic cache parameters */
1303 			break;
1304 
1305 		case 5:	/* Monitor/Mwait parameters */
1306 		{
1307 			size_t mwait_size;
1308 
1309 			/*
1310 			 * check cpi_mwait.support which was set in cpuid_pass1
1311 			 */
1312 			if (!(cpi->cpi_mwait.support & MWAIT_SUPPORT))
1313 				break;
1314 
1315 			/*
1316 			 * Protect ourself from insane mwait line size.
1317 			 * Workaround for incomplete hardware emulator(s).
1318 			 */
1319 			mwait_size = (size_t)MWAIT_SIZE_MAX(cpi);
1320 			if (mwait_size < sizeof (uint32_t) ||
1321 			    !ISP2(mwait_size)) {
1322 #if DEBUG
1323 				cmn_err(CE_NOTE, "Cannot handle cpu %d mwait "
1324 				    "size %ld", cpu->cpu_id, (long)mwait_size);
1325 #endif
1326 				break;
1327 			}
1328 
1329 			cpi->cpi_mwait.mon_min = (size_t)MWAIT_SIZE_MIN(cpi);
1330 			cpi->cpi_mwait.mon_max = mwait_size;
1331 			if (MWAIT_EXTENSION(cpi)) {
1332 				cpi->cpi_mwait.support |= MWAIT_EXTENSIONS;
1333 				if (MWAIT_INT_ENABLE(cpi))
1334 					cpi->cpi_mwait.support |=
1335 					    MWAIT_ECX_INT_ENABLE;
1336 			}
1337 			break;
1338 		}
1339 		default:
1340 			break;
1341 		}
1342 	}
1343 
1344 	if (cpi->cpi_maxeax >= 0xB && cpi->cpi_vendor == X86_VENDOR_Intel) {
1345 		struct cpuid_regs regs;
1346 
1347 		cp = &regs;
1348 		cp->cp_eax = 0xB;
1349 		cp->cp_edx = cp->cp_ebx = cp->cp_ecx = 0;
1350 
1351 		(void) __cpuid_insn(cp);
1352 
1353 		/*
1354 		 * Check CPUID.EAX=0BH, ECX=0H:EBX is non-zero, which
1355 		 * indicates that the extended topology enumeration leaf is
1356 		 * available.
1357 		 */
1358 		if (cp->cp_ebx) {
1359 			uint32_t x2apic_id;
1360 			uint_t coreid_shift = 0;
1361 			uint_t ncpu_per_core = 1;
1362 			uint_t chipid_shift = 0;
1363 			uint_t ncpu_per_chip = 1;
1364 			uint_t i;
1365 			uint_t level;
1366 
1367 			for (i = 0; i < CPI_FNB_ECX_MAX; i++) {
1368 				cp->cp_eax = 0xB;
1369 				cp->cp_ecx = i;
1370 
1371 				(void) __cpuid_insn(cp);
1372 				level = CPI_CPU_LEVEL_TYPE(cp);
1373 
1374 				if (level == 1) {
1375 					x2apic_id = cp->cp_edx;
1376 					coreid_shift = BITX(cp->cp_eax, 4, 0);
1377 					ncpu_per_core = BITX(cp->cp_ebx, 15, 0);
1378 				} else if (level == 2) {
1379 					x2apic_id = cp->cp_edx;
1380 					chipid_shift = BITX(cp->cp_eax, 4, 0);
1381 					ncpu_per_chip = BITX(cp->cp_ebx, 15, 0);
1382 				}
1383 			}
1384 
1385 			cpi->cpi_apicid = x2apic_id;
1386 			cpi->cpi_ncpu_per_chip = ncpu_per_chip;
1387 			cpi->cpi_ncore_per_chip = ncpu_per_chip /
1388 			    ncpu_per_core;
1389 			cpi->cpi_chipid = x2apic_id >> chipid_shift;
1390 			cpi->cpi_clogid = x2apic_id & ((1 << chipid_shift) - 1);
1391 			cpi->cpi_coreid = x2apic_id >> coreid_shift;
1392 			cpi->cpi_pkgcoreid = cpi->cpi_clogid >> coreid_shift;
1393 		}
1394 
1395 		/* Make cp NULL so that we don't stumble on others */
1396 		cp = NULL;
1397 	}
1398 
1399 	if ((cpi->cpi_xmaxeax & 0x80000000) == 0)
1400 		goto pass2_done;
1401 
1402 	if ((nmax = cpi->cpi_xmaxeax - 0x80000000 + 1) > NMAX_CPI_EXTD)
1403 		nmax = NMAX_CPI_EXTD;
1404 	/*
1405 	 * Copy the extended properties, fixing them as we go.
1406 	 * (We already handled n == 0 and n == 1 in pass 1)
1407 	 */
1408 	iptr = (void *)cpi->cpi_brandstr;
1409 	for (n = 2, cp = &cpi->cpi_extd[2]; n < nmax; cp++, n++) {
1410 		cp->cp_eax = 0x80000000 + n;
1411 		(void) __cpuid_insn(cp);
1412 		platform_cpuid_mangle(cpi->cpi_vendor, 0x80000000 + n, cp);
1413 		switch (n) {
1414 		case 2:
1415 		case 3:
1416 		case 4:
1417 			/*
1418 			 * Extract the brand string
1419 			 */
1420 			*iptr++ = cp->cp_eax;
1421 			*iptr++ = cp->cp_ebx;
1422 			*iptr++ = cp->cp_ecx;
1423 			*iptr++ = cp->cp_edx;
1424 			break;
1425 		case 5:
1426 			switch (cpi->cpi_vendor) {
1427 			case X86_VENDOR_AMD:
1428 				/*
1429 				 * The Athlon and Duron were the first
1430 				 * parts to report the sizes of the
1431 				 * TLB for large pages. Before then,
1432 				 * we don't trust the data.
1433 				 */
1434 				if (cpi->cpi_family < 6 ||
1435 				    (cpi->cpi_family == 6 &&
1436 				    cpi->cpi_model < 1))
1437 					cp->cp_eax = 0;
1438 				break;
1439 			default:
1440 				break;
1441 			}
1442 			break;
1443 		case 6:
1444 			switch (cpi->cpi_vendor) {
1445 			case X86_VENDOR_AMD:
1446 				/*
1447 				 * The Athlon and Duron were the first
1448 				 * AMD parts with L2 TLB's.
1449 				 * Before then, don't trust the data.
1450 				 */
1451 				if (cpi->cpi_family < 6 ||
1452 				    cpi->cpi_family == 6 &&
1453 				    cpi->cpi_model < 1)
1454 					cp->cp_eax = cp->cp_ebx = 0;
1455 				/*
1456 				 * AMD Duron rev A0 reports L2
1457 				 * cache size incorrectly as 1K
1458 				 * when it is really 64K
1459 				 */
1460 				if (cpi->cpi_family == 6 &&
1461 				    cpi->cpi_model == 3 &&
1462 				    cpi->cpi_step == 0) {
1463 					cp->cp_ecx &= 0xffff;
1464 					cp->cp_ecx |= 0x400000;
1465 				}
1466 				break;
1467 			case X86_VENDOR_Cyrix:	/* VIA C3 */
1468 				/*
1469 				 * VIA C3 processors are a bit messed
1470 				 * up w.r.t. encoding cache sizes in %ecx
1471 				 */
1472 				if (cpi->cpi_family != 6)
1473 					break;
1474 				/*
1475 				 * model 7 and 8 were incorrectly encoded
1476 				 *
1477 				 * xxx is model 8 really broken?
1478 				 */
1479 				if (cpi->cpi_model == 7 ||
1480 				    cpi->cpi_model == 8)
1481 					cp->cp_ecx =
1482 					    BITX(cp->cp_ecx, 31, 24) << 16 |
1483 					    BITX(cp->cp_ecx, 23, 16) << 12 |
1484 					    BITX(cp->cp_ecx, 15, 8) << 8 |
1485 					    BITX(cp->cp_ecx, 7, 0);
1486 				/*
1487 				 * model 9 stepping 1 has wrong associativity
1488 				 */
1489 				if (cpi->cpi_model == 9 && cpi->cpi_step == 1)
1490 					cp->cp_ecx |= 8 << 12;
1491 				break;
1492 			case X86_VENDOR_Intel:
1493 				/*
1494 				 * Extended L2 Cache features function.
1495 				 * First appeared on Prescott.
1496 				 */
1497 			default:
1498 				break;
1499 			}
1500 			break;
1501 		default:
1502 			break;
1503 		}
1504 	}
1505 
1506 pass2_done:
1507 	cpi->cpi_pass = 2;
1508 }
1509 
1510 static const char *
1511 intel_cpubrand(const struct cpuid_info *cpi)
1512 {
1513 	int i;
1514 
1515 	if ((x86_feature & X86_CPUID) == 0 ||
1516 	    cpi->cpi_maxeax < 1 || cpi->cpi_family < 5)
1517 		return ("i486");
1518 
1519 	switch (cpi->cpi_family) {
1520 	case 5:
1521 		return ("Intel Pentium(r)");
1522 	case 6:
1523 		switch (cpi->cpi_model) {
1524 			uint_t celeron, xeon;
1525 			const struct cpuid_regs *cp;
1526 		case 0:
1527 		case 1:
1528 		case 2:
1529 			return ("Intel Pentium(r) Pro");
1530 		case 3:
1531 		case 4:
1532 			return ("Intel Pentium(r) II");
1533 		case 6:
1534 			return ("Intel Celeron(r)");
1535 		case 5:
1536 		case 7:
1537 			celeron = xeon = 0;
1538 			cp = &cpi->cpi_std[2];	/* cache info */
1539 
1540 			for (i = 1; i < 4; i++) {
1541 				uint_t tmp;
1542 
1543 				tmp = (cp->cp_eax >> (8 * i)) & 0xff;
1544 				if (tmp == 0x40)
1545 					celeron++;
1546 				if (tmp >= 0x44 && tmp <= 0x45)
1547 					xeon++;
1548 			}
1549 
1550 			for (i = 0; i < 2; i++) {
1551 				uint_t tmp;
1552 
1553 				tmp = (cp->cp_ebx >> (8 * i)) & 0xff;
1554 				if (tmp == 0x40)
1555 					celeron++;
1556 				else if (tmp >= 0x44 && tmp <= 0x45)
1557 					xeon++;
1558 			}
1559 
1560 			for (i = 0; i < 4; i++) {
1561 				uint_t tmp;
1562 
1563 				tmp = (cp->cp_ecx >> (8 * i)) & 0xff;
1564 				if (tmp == 0x40)
1565 					celeron++;
1566 				else if (tmp >= 0x44 && tmp <= 0x45)
1567 					xeon++;
1568 			}
1569 
1570 			for (i = 0; i < 4; i++) {
1571 				uint_t tmp;
1572 
1573 				tmp = (cp->cp_edx >> (8 * i)) & 0xff;
1574 				if (tmp == 0x40)
1575 					celeron++;
1576 				else if (tmp >= 0x44 && tmp <= 0x45)
1577 					xeon++;
1578 			}
1579 
1580 			if (celeron)
1581 				return ("Intel Celeron(r)");
1582 			if (xeon)
1583 				return (cpi->cpi_model == 5 ?
1584 				    "Intel Pentium(r) II Xeon(tm)" :
1585 				    "Intel Pentium(r) III Xeon(tm)");
1586 			return (cpi->cpi_model == 5 ?
1587 			    "Intel Pentium(r) II or Pentium(r) II Xeon(tm)" :
1588 			    "Intel Pentium(r) III or Pentium(r) III Xeon(tm)");
1589 		default:
1590 			break;
1591 		}
1592 	default:
1593 		break;
1594 	}
1595 
1596 	/* BrandID is present if the field is nonzero */
1597 	if (cpi->cpi_brandid != 0) {
1598 		static const struct {
1599 			uint_t bt_bid;
1600 			const char *bt_str;
1601 		} brand_tbl[] = {
1602 			{ 0x1,	"Intel(r) Celeron(r)" },
1603 			{ 0x2,	"Intel(r) Pentium(r) III" },
1604 			{ 0x3,	"Intel(r) Pentium(r) III Xeon(tm)" },
1605 			{ 0x4,	"Intel(r) Pentium(r) III" },
1606 			{ 0x6,	"Mobile Intel(r) Pentium(r) III" },
1607 			{ 0x7,	"Mobile Intel(r) Celeron(r)" },
1608 			{ 0x8,	"Intel(r) Pentium(r) 4" },
1609 			{ 0x9,	"Intel(r) Pentium(r) 4" },
1610 			{ 0xa,	"Intel(r) Celeron(r)" },
1611 			{ 0xb,	"Intel(r) Xeon(tm)" },
1612 			{ 0xc,	"Intel(r) Xeon(tm) MP" },
1613 			{ 0xe,	"Mobile Intel(r) Pentium(r) 4" },
1614 			{ 0xf,	"Mobile Intel(r) Celeron(r)" },
1615 			{ 0x11, "Mobile Genuine Intel(r)" },
1616 			{ 0x12, "Intel(r) Celeron(r) M" },
1617 			{ 0x13, "Mobile Intel(r) Celeron(r)" },
1618 			{ 0x14, "Intel(r) Celeron(r)" },
1619 			{ 0x15, "Mobile Genuine Intel(r)" },
1620 			{ 0x16,	"Intel(r) Pentium(r) M" },
1621 			{ 0x17, "Mobile Intel(r) Celeron(r)" }
1622 		};
1623 		uint_t btblmax = sizeof (brand_tbl) / sizeof (brand_tbl[0]);
1624 		uint_t sgn;
1625 
1626 		sgn = (cpi->cpi_family << 8) |
1627 		    (cpi->cpi_model << 4) | cpi->cpi_step;
1628 
1629 		for (i = 0; i < btblmax; i++)
1630 			if (brand_tbl[i].bt_bid == cpi->cpi_brandid)
1631 				break;
1632 		if (i < btblmax) {
1633 			if (sgn == 0x6b1 && cpi->cpi_brandid == 3)
1634 				return ("Intel(r) Celeron(r)");
1635 			if (sgn < 0xf13 && cpi->cpi_brandid == 0xb)
1636 				return ("Intel(r) Xeon(tm) MP");
1637 			if (sgn < 0xf13 && cpi->cpi_brandid == 0xe)
1638 				return ("Intel(r) Xeon(tm)");
1639 			return (brand_tbl[i].bt_str);
1640 		}
1641 	}
1642 
1643 	return (NULL);
1644 }
1645 
1646 static const char *
1647 amd_cpubrand(const struct cpuid_info *cpi)
1648 {
1649 	if ((x86_feature & X86_CPUID) == 0 ||
1650 	    cpi->cpi_maxeax < 1 || cpi->cpi_family < 5)
1651 		return ("i486 compatible");
1652 
1653 	switch (cpi->cpi_family) {
1654 	case 5:
1655 		switch (cpi->cpi_model) {
1656 		case 0:
1657 		case 1:
1658 		case 2:
1659 		case 3:
1660 		case 4:
1661 		case 5:
1662 			return ("AMD-K5(r)");
1663 		case 6:
1664 		case 7:
1665 			return ("AMD-K6(r)");
1666 		case 8:
1667 			return ("AMD-K6(r)-2");
1668 		case 9:
1669 			return ("AMD-K6(r)-III");
1670 		default:
1671 			return ("AMD (family 5)");
1672 		}
1673 	case 6:
1674 		switch (cpi->cpi_model) {
1675 		case 1:
1676 			return ("AMD-K7(tm)");
1677 		case 0:
1678 		case 2:
1679 		case 4:
1680 			return ("AMD Athlon(tm)");
1681 		case 3:
1682 		case 7:
1683 			return ("AMD Duron(tm)");
1684 		case 6:
1685 		case 8:
1686 		case 10:
1687 			/*
1688 			 * Use the L2 cache size to distinguish
1689 			 */
1690 			return ((cpi->cpi_extd[6].cp_ecx >> 16) >= 256 ?
1691 			    "AMD Athlon(tm)" : "AMD Duron(tm)");
1692 		default:
1693 			return ("AMD (family 6)");
1694 		}
1695 	default:
1696 		break;
1697 	}
1698 
1699 	if (cpi->cpi_family == 0xf && cpi->cpi_model == 5 &&
1700 	    cpi->cpi_brandid != 0) {
1701 		switch (BITX(cpi->cpi_brandid, 7, 5)) {
1702 		case 3:
1703 			return ("AMD Opteron(tm) UP 1xx");
1704 		case 4:
1705 			return ("AMD Opteron(tm) DP 2xx");
1706 		case 5:
1707 			return ("AMD Opteron(tm) MP 8xx");
1708 		default:
1709 			return ("AMD Opteron(tm)");
1710 		}
1711 	}
1712 
1713 	return (NULL);
1714 }
1715 
1716 static const char *
1717 cyrix_cpubrand(struct cpuid_info *cpi, uint_t type)
1718 {
1719 	if ((x86_feature & X86_CPUID) == 0 ||
1720 	    cpi->cpi_maxeax < 1 || cpi->cpi_family < 5 ||
1721 	    type == X86_TYPE_CYRIX_486)
1722 		return ("i486 compatible");
1723 
1724 	switch (type) {
1725 	case X86_TYPE_CYRIX_6x86:
1726 		return ("Cyrix 6x86");
1727 	case X86_TYPE_CYRIX_6x86L:
1728 		return ("Cyrix 6x86L");
1729 	case X86_TYPE_CYRIX_6x86MX:
1730 		return ("Cyrix 6x86MX");
1731 	case X86_TYPE_CYRIX_GXm:
1732 		return ("Cyrix GXm");
1733 	case X86_TYPE_CYRIX_MediaGX:
1734 		return ("Cyrix MediaGX");
1735 	case X86_TYPE_CYRIX_MII:
1736 		return ("Cyrix M2");
1737 	case X86_TYPE_VIA_CYRIX_III:
1738 		return ("VIA Cyrix M3");
1739 	default:
1740 		/*
1741 		 * Have another wild guess ..
1742 		 */
1743 		if (cpi->cpi_family == 4 && cpi->cpi_model == 9)
1744 			return ("Cyrix 5x86");
1745 		else if (cpi->cpi_family == 5) {
1746 			switch (cpi->cpi_model) {
1747 			case 2:
1748 				return ("Cyrix 6x86");	/* Cyrix M1 */
1749 			case 4:
1750 				return ("Cyrix MediaGX");
1751 			default:
1752 				break;
1753 			}
1754 		} else if (cpi->cpi_family == 6) {
1755 			switch (cpi->cpi_model) {
1756 			case 0:
1757 				return ("Cyrix 6x86MX"); /* Cyrix M2? */
1758 			case 5:
1759 			case 6:
1760 			case 7:
1761 			case 8:
1762 			case 9:
1763 				return ("VIA C3");
1764 			default:
1765 				break;
1766 			}
1767 		}
1768 		break;
1769 	}
1770 	return (NULL);
1771 }
1772 
1773 /*
1774  * This only gets called in the case that the CPU extended
1775  * feature brand string (0x80000002, 0x80000003, 0x80000004)
1776  * aren't available, or contain null bytes for some reason.
1777  */
1778 static void
1779 fabricate_brandstr(struct cpuid_info *cpi)
1780 {
1781 	const char *brand = NULL;
1782 
1783 	switch (cpi->cpi_vendor) {
1784 	case X86_VENDOR_Intel:
1785 		brand = intel_cpubrand(cpi);
1786 		break;
1787 	case X86_VENDOR_AMD:
1788 		brand = amd_cpubrand(cpi);
1789 		break;
1790 	case X86_VENDOR_Cyrix:
1791 		brand = cyrix_cpubrand(cpi, x86_type);
1792 		break;
1793 	case X86_VENDOR_NexGen:
1794 		if (cpi->cpi_family == 5 && cpi->cpi_model == 0)
1795 			brand = "NexGen Nx586";
1796 		break;
1797 	case X86_VENDOR_Centaur:
1798 		if (cpi->cpi_family == 5)
1799 			switch (cpi->cpi_model) {
1800 			case 4:
1801 				brand = "Centaur C6";
1802 				break;
1803 			case 8:
1804 				brand = "Centaur C2";
1805 				break;
1806 			case 9:
1807 				brand = "Centaur C3";
1808 				break;
1809 			default:
1810 				break;
1811 			}
1812 		break;
1813 	case X86_VENDOR_Rise:
1814 		if (cpi->cpi_family == 5 &&
1815 		    (cpi->cpi_model == 0 || cpi->cpi_model == 2))
1816 			brand = "Rise mP6";
1817 		break;
1818 	case X86_VENDOR_SiS:
1819 		if (cpi->cpi_family == 5 && cpi->cpi_model == 0)
1820 			brand = "SiS 55x";
1821 		break;
1822 	case X86_VENDOR_TM:
1823 		if (cpi->cpi_family == 5 && cpi->cpi_model == 4)
1824 			brand = "Transmeta Crusoe TM3x00 or TM5x00";
1825 		break;
1826 	case X86_VENDOR_NSC:
1827 	case X86_VENDOR_UMC:
1828 	default:
1829 		break;
1830 	}
1831 	if (brand) {
1832 		(void) strcpy((char *)cpi->cpi_brandstr, brand);
1833 		return;
1834 	}
1835 
1836 	/*
1837 	 * If all else fails ...
1838 	 */
1839 	(void) snprintf(cpi->cpi_brandstr, sizeof (cpi->cpi_brandstr),
1840 	    "%s %d.%d.%d", cpi->cpi_vendorstr, cpi->cpi_family,
1841 	    cpi->cpi_model, cpi->cpi_step);
1842 }
1843 
1844 /*
1845  * This routine is called just after kernel memory allocation
1846  * becomes available on cpu0, and as part of mp_startup() on
1847  * the other cpus.
1848  *
1849  * Fixup the brand string, and collect any information from cpuid
1850  * that requires dynamicically allocated storage to represent.
1851  */
1852 /*ARGSUSED*/
1853 void
1854 cpuid_pass3(cpu_t *cpu)
1855 {
1856 	int	i, max, shft, level, size;
1857 	struct cpuid_regs regs;
1858 	struct cpuid_regs *cp;
1859 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
1860 
1861 	ASSERT(cpi->cpi_pass == 2);
1862 
1863 	/*
1864 	 * Function 4: Deterministic cache parameters
1865 	 *
1866 	 * Take this opportunity to detect the number of threads
1867 	 * sharing the last level cache, and construct a corresponding
1868 	 * cache id. The respective cpuid_info members are initialized
1869 	 * to the default case of "no last level cache sharing".
1870 	 */
1871 	cpi->cpi_ncpu_shr_last_cache = 1;
1872 	cpi->cpi_last_lvl_cacheid = cpu->cpu_id;
1873 
1874 	if (cpi->cpi_maxeax >= 4 && cpi->cpi_vendor == X86_VENDOR_Intel) {
1875 
1876 		/*
1877 		 * Find the # of elements (size) returned by fn 4, and along
1878 		 * the way detect last level cache sharing details.
1879 		 */
1880 		bzero(&regs, sizeof (regs));
1881 		cp = &regs;
1882 		for (i = 0, max = 0; i < CPI_FN4_ECX_MAX; i++) {
1883 			cp->cp_eax = 4;
1884 			cp->cp_ecx = i;
1885 
1886 			(void) __cpuid_insn(cp);
1887 
1888 			if (CPI_CACHE_TYPE(cp) == 0)
1889 				break;
1890 			level = CPI_CACHE_LVL(cp);
1891 			if (level > max) {
1892 				max = level;
1893 				cpi->cpi_ncpu_shr_last_cache =
1894 				    CPI_NTHR_SHR_CACHE(cp) + 1;
1895 			}
1896 		}
1897 		cpi->cpi_std_4_size = size = i;
1898 
1899 		/*
1900 		 * Allocate the cpi_std_4 array. The first element
1901 		 * references the regs for fn 4, %ecx == 0, which
1902 		 * cpuid_pass2() stashed in cpi->cpi_std[4].
1903 		 */
1904 		if (size > 0) {
1905 			cpi->cpi_std_4 =
1906 			    kmem_alloc(size * sizeof (cp), KM_SLEEP);
1907 			cpi->cpi_std_4[0] = &cpi->cpi_std[4];
1908 
1909 			/*
1910 			 * Allocate storage to hold the additional regs
1911 			 * for function 4, %ecx == 1 .. cpi_std_4_size.
1912 			 *
1913 			 * The regs for fn 4, %ecx == 0 has already
1914 			 * been allocated as indicated above.
1915 			 */
1916 			for (i = 1; i < size; i++) {
1917 				cp = cpi->cpi_std_4[i] =
1918 				    kmem_zalloc(sizeof (regs), KM_SLEEP);
1919 				cp->cp_eax = 4;
1920 				cp->cp_ecx = i;
1921 
1922 				(void) __cpuid_insn(cp);
1923 			}
1924 		}
1925 		/*
1926 		 * Determine the number of bits needed to represent
1927 		 * the number of CPUs sharing the last level cache.
1928 		 *
1929 		 * Shift off that number of bits from the APIC id to
1930 		 * derive the cache id.
1931 		 */
1932 		shft = 0;
1933 		for (i = 1; i < cpi->cpi_ncpu_shr_last_cache; i <<= 1)
1934 			shft++;
1935 		cpi->cpi_last_lvl_cacheid = cpi->cpi_apicid >> shft;
1936 	}
1937 
1938 	/*
1939 	 * Now fixup the brand string
1940 	 */
1941 	if ((cpi->cpi_xmaxeax & 0x80000000) == 0) {
1942 		fabricate_brandstr(cpi);
1943 	} else {
1944 
1945 		/*
1946 		 * If we successfully extracted a brand string from the cpuid
1947 		 * instruction, clean it up by removing leading spaces and
1948 		 * similar junk.
1949 		 */
1950 		if (cpi->cpi_brandstr[0]) {
1951 			size_t maxlen = sizeof (cpi->cpi_brandstr);
1952 			char *src, *dst;
1953 
1954 			dst = src = (char *)cpi->cpi_brandstr;
1955 			src[maxlen - 1] = '\0';
1956 			/*
1957 			 * strip leading spaces
1958 			 */
1959 			while (*src == ' ')
1960 				src++;
1961 			/*
1962 			 * Remove any 'Genuine' or "Authentic" prefixes
1963 			 */
1964 			if (strncmp(src, "Genuine ", 8) == 0)
1965 				src += 8;
1966 			if (strncmp(src, "Authentic ", 10) == 0)
1967 				src += 10;
1968 
1969 			/*
1970 			 * Now do an in-place copy.
1971 			 * Map (R) to (r) and (TM) to (tm).
1972 			 * The era of teletypes is long gone, and there's
1973 			 * -really- no need to shout.
1974 			 */
1975 			while (*src != '\0') {
1976 				if (src[0] == '(') {
1977 					if (strncmp(src + 1, "R)", 2) == 0) {
1978 						(void) strncpy(dst, "(r)", 3);
1979 						src += 3;
1980 						dst += 3;
1981 						continue;
1982 					}
1983 					if (strncmp(src + 1, "TM)", 3) == 0) {
1984 						(void) strncpy(dst, "(tm)", 4);
1985 						src += 4;
1986 						dst += 4;
1987 						continue;
1988 					}
1989 				}
1990 				*dst++ = *src++;
1991 			}
1992 			*dst = '\0';
1993 
1994 			/*
1995 			 * Finally, remove any trailing spaces
1996 			 */
1997 			while (--dst > cpi->cpi_brandstr)
1998 				if (*dst == ' ')
1999 					*dst = '\0';
2000 				else
2001 					break;
2002 		} else
2003 			fabricate_brandstr(cpi);
2004 	}
2005 	cpi->cpi_pass = 3;
2006 }
2007 
2008 /*
2009  * This routine is called out of bind_hwcap() much later in the life
2010  * of the kernel (post_startup()).  The job of this routine is to resolve
2011  * the hardware feature support and kernel support for those features into
2012  * what we're actually going to tell applications via the aux vector.
2013  */
2014 uint_t
2015 cpuid_pass4(cpu_t *cpu)
2016 {
2017 	struct cpuid_info *cpi;
2018 	uint_t hwcap_flags = 0;
2019 
2020 	if (cpu == NULL)
2021 		cpu = CPU;
2022 	cpi = cpu->cpu_m.mcpu_cpi;
2023 
2024 	ASSERT(cpi->cpi_pass == 3);
2025 
2026 	if (cpi->cpi_maxeax >= 1) {
2027 		uint32_t *edx = &cpi->cpi_support[STD_EDX_FEATURES];
2028 		uint32_t *ecx = &cpi->cpi_support[STD_ECX_FEATURES];
2029 
2030 		*edx = CPI_FEATURES_EDX(cpi);
2031 		*ecx = CPI_FEATURES_ECX(cpi);
2032 
2033 		/*
2034 		 * [these require explicit kernel support]
2035 		 */
2036 		if ((x86_feature & X86_SEP) == 0)
2037 			*edx &= ~CPUID_INTC_EDX_SEP;
2038 
2039 		if ((x86_feature & X86_SSE) == 0)
2040 			*edx &= ~(CPUID_INTC_EDX_FXSR|CPUID_INTC_EDX_SSE);
2041 		if ((x86_feature & X86_SSE2) == 0)
2042 			*edx &= ~CPUID_INTC_EDX_SSE2;
2043 
2044 		if ((x86_feature & X86_HTT) == 0)
2045 			*edx &= ~CPUID_INTC_EDX_HTT;
2046 
2047 		if ((x86_feature & X86_SSE3) == 0)
2048 			*ecx &= ~CPUID_INTC_ECX_SSE3;
2049 
2050 		if (cpi->cpi_vendor == X86_VENDOR_Intel) {
2051 			if ((x86_feature & X86_SSSE3) == 0)
2052 				*ecx &= ~CPUID_INTC_ECX_SSSE3;
2053 			if ((x86_feature & X86_SSE4_1) == 0)
2054 				*ecx &= ~CPUID_INTC_ECX_SSE4_1;
2055 			if ((x86_feature & X86_SSE4_2) == 0)
2056 				*ecx &= ~CPUID_INTC_ECX_SSE4_2;
2057 		}
2058 
2059 		/*
2060 		 * [no explicit support required beyond x87 fp context]
2061 		 */
2062 		if (!fpu_exists)
2063 			*edx &= ~(CPUID_INTC_EDX_FPU | CPUID_INTC_EDX_MMX);
2064 
2065 		/*
2066 		 * Now map the supported feature vector to things that we
2067 		 * think userland will care about.
2068 		 */
2069 		if (*edx & CPUID_INTC_EDX_SEP)
2070 			hwcap_flags |= AV_386_SEP;
2071 		if (*edx & CPUID_INTC_EDX_SSE)
2072 			hwcap_flags |= AV_386_FXSR | AV_386_SSE;
2073 		if (*edx & CPUID_INTC_EDX_SSE2)
2074 			hwcap_flags |= AV_386_SSE2;
2075 		if (*ecx & CPUID_INTC_ECX_SSE3)
2076 			hwcap_flags |= AV_386_SSE3;
2077 		if (cpi->cpi_vendor == X86_VENDOR_Intel) {
2078 			if (*ecx & CPUID_INTC_ECX_SSSE3)
2079 				hwcap_flags |= AV_386_SSSE3;
2080 			if (*ecx & CPUID_INTC_ECX_SSE4_1)
2081 				hwcap_flags |= AV_386_SSE4_1;
2082 			if (*ecx & CPUID_INTC_ECX_SSE4_2)
2083 				hwcap_flags |= AV_386_SSE4_2;
2084 			if (*ecx & CPUID_INTC_ECX_MOVBE)
2085 				hwcap_flags |= AV_386_MOVBE;
2086 		}
2087 		if (*ecx & CPUID_INTC_ECX_POPCNT)
2088 			hwcap_flags |= AV_386_POPCNT;
2089 		if (*edx & CPUID_INTC_EDX_FPU)
2090 			hwcap_flags |= AV_386_FPU;
2091 		if (*edx & CPUID_INTC_EDX_MMX)
2092 			hwcap_flags |= AV_386_MMX;
2093 
2094 		if (*edx & CPUID_INTC_EDX_TSC)
2095 			hwcap_flags |= AV_386_TSC;
2096 		if (*edx & CPUID_INTC_EDX_CX8)
2097 			hwcap_flags |= AV_386_CX8;
2098 		if (*edx & CPUID_INTC_EDX_CMOV)
2099 			hwcap_flags |= AV_386_CMOV;
2100 		if (*ecx & CPUID_INTC_ECX_MON)
2101 			hwcap_flags |= AV_386_MON;
2102 		if (*ecx & CPUID_INTC_ECX_CX16)
2103 			hwcap_flags |= AV_386_CX16;
2104 	}
2105 
2106 	if (x86_feature & X86_HTT)
2107 		hwcap_flags |= AV_386_PAUSE;
2108 
2109 	if (cpi->cpi_xmaxeax < 0x80000001)
2110 		goto pass4_done;
2111 
2112 	switch (cpi->cpi_vendor) {
2113 		struct cpuid_regs cp;
2114 		uint32_t *edx, *ecx;
2115 
2116 	case X86_VENDOR_Intel:
2117 		/*
2118 		 * Seems like Intel duplicated what we necessary
2119 		 * here to make the initial crop of 64-bit OS's work.
2120 		 * Hopefully, those are the only "extended" bits
2121 		 * they'll add.
2122 		 */
2123 		/*FALLTHROUGH*/
2124 
2125 	case X86_VENDOR_AMD:
2126 		edx = &cpi->cpi_support[AMD_EDX_FEATURES];
2127 		ecx = &cpi->cpi_support[AMD_ECX_FEATURES];
2128 
2129 		*edx = CPI_FEATURES_XTD_EDX(cpi);
2130 		*ecx = CPI_FEATURES_XTD_ECX(cpi);
2131 
2132 		/*
2133 		 * [these features require explicit kernel support]
2134 		 */
2135 		switch (cpi->cpi_vendor) {
2136 		case X86_VENDOR_Intel:
2137 			if ((x86_feature & X86_TSCP) == 0)
2138 				*edx &= ~CPUID_AMD_EDX_TSCP;
2139 			break;
2140 
2141 		case X86_VENDOR_AMD:
2142 			if ((x86_feature & X86_TSCP) == 0)
2143 				*edx &= ~CPUID_AMD_EDX_TSCP;
2144 			if ((x86_feature & X86_SSE4A) == 0)
2145 				*ecx &= ~CPUID_AMD_ECX_SSE4A;
2146 			break;
2147 
2148 		default:
2149 			break;
2150 		}
2151 
2152 		/*
2153 		 * [no explicit support required beyond
2154 		 * x87 fp context and exception handlers]
2155 		 */
2156 		if (!fpu_exists)
2157 			*edx &= ~(CPUID_AMD_EDX_MMXamd |
2158 			    CPUID_AMD_EDX_3DNow | CPUID_AMD_EDX_3DNowx);
2159 
2160 		if ((x86_feature & X86_NX) == 0)
2161 			*edx &= ~CPUID_AMD_EDX_NX;
2162 #if !defined(__amd64)
2163 		*edx &= ~CPUID_AMD_EDX_LM;
2164 #endif
2165 		/*
2166 		 * Now map the supported feature vector to
2167 		 * things that we think userland will care about.
2168 		 */
2169 #if defined(__amd64)
2170 		if (*edx & CPUID_AMD_EDX_SYSC)
2171 			hwcap_flags |= AV_386_AMD_SYSC;
2172 #endif
2173 		if (*edx & CPUID_AMD_EDX_MMXamd)
2174 			hwcap_flags |= AV_386_AMD_MMX;
2175 		if (*edx & CPUID_AMD_EDX_3DNow)
2176 			hwcap_flags |= AV_386_AMD_3DNow;
2177 		if (*edx & CPUID_AMD_EDX_3DNowx)
2178 			hwcap_flags |= AV_386_AMD_3DNowx;
2179 
2180 		switch (cpi->cpi_vendor) {
2181 		case X86_VENDOR_AMD:
2182 			if (*edx & CPUID_AMD_EDX_TSCP)
2183 				hwcap_flags |= AV_386_TSCP;
2184 			if (*ecx & CPUID_AMD_ECX_AHF64)
2185 				hwcap_flags |= AV_386_AHF;
2186 			if (*ecx & CPUID_AMD_ECX_SSE4A)
2187 				hwcap_flags |= AV_386_AMD_SSE4A;
2188 			if (*ecx & CPUID_AMD_ECX_LZCNT)
2189 				hwcap_flags |= AV_386_AMD_LZCNT;
2190 			break;
2191 
2192 		case X86_VENDOR_Intel:
2193 			if (*edx & CPUID_AMD_EDX_TSCP)
2194 				hwcap_flags |= AV_386_TSCP;
2195 			/*
2196 			 * Aarrgh.
2197 			 * Intel uses a different bit in the same word.
2198 			 */
2199 			if (*ecx & CPUID_INTC_ECX_AHF64)
2200 				hwcap_flags |= AV_386_AHF;
2201 			break;
2202 
2203 		default:
2204 			break;
2205 		}
2206 		break;
2207 
2208 	case X86_VENDOR_TM:
2209 		cp.cp_eax = 0x80860001;
2210 		(void) __cpuid_insn(&cp);
2211 		cpi->cpi_support[TM_EDX_FEATURES] = cp.cp_edx;
2212 		break;
2213 
2214 	default:
2215 		break;
2216 	}
2217 
2218 pass4_done:
2219 	cpi->cpi_pass = 4;
2220 	return (hwcap_flags);
2221 }
2222 
2223 
2224 /*
2225  * Simulate the cpuid instruction using the data we previously
2226  * captured about this CPU.  We try our best to return the truth
2227  * about the hardware, independently of kernel support.
2228  */
2229 uint32_t
2230 cpuid_insn(cpu_t *cpu, struct cpuid_regs *cp)
2231 {
2232 	struct cpuid_info *cpi;
2233 	struct cpuid_regs *xcp;
2234 
2235 	if (cpu == NULL)
2236 		cpu = CPU;
2237 	cpi = cpu->cpu_m.mcpu_cpi;
2238 
2239 	ASSERT(cpuid_checkpass(cpu, 3));
2240 
2241 	/*
2242 	 * CPUID data is cached in two separate places: cpi_std for standard
2243 	 * CPUID functions, and cpi_extd for extended CPUID functions.
2244 	 */
2245 	if (cp->cp_eax <= cpi->cpi_maxeax && cp->cp_eax < NMAX_CPI_STD)
2246 		xcp = &cpi->cpi_std[cp->cp_eax];
2247 	else if (cp->cp_eax >= 0x80000000 && cp->cp_eax <= cpi->cpi_xmaxeax &&
2248 	    cp->cp_eax < 0x80000000 + NMAX_CPI_EXTD)
2249 		xcp = &cpi->cpi_extd[cp->cp_eax - 0x80000000];
2250 	else
2251 		/*
2252 		 * The caller is asking for data from an input parameter which
2253 		 * the kernel has not cached.  In this case we go fetch from
2254 		 * the hardware and return the data directly to the user.
2255 		 */
2256 		return (__cpuid_insn(cp));
2257 
2258 	cp->cp_eax = xcp->cp_eax;
2259 	cp->cp_ebx = xcp->cp_ebx;
2260 	cp->cp_ecx = xcp->cp_ecx;
2261 	cp->cp_edx = xcp->cp_edx;
2262 	return (cp->cp_eax);
2263 }
2264 
2265 int
2266 cpuid_checkpass(cpu_t *cpu, int pass)
2267 {
2268 	return (cpu != NULL && cpu->cpu_m.mcpu_cpi != NULL &&
2269 	    cpu->cpu_m.mcpu_cpi->cpi_pass >= pass);
2270 }
2271 
2272 int
2273 cpuid_getbrandstr(cpu_t *cpu, char *s, size_t n)
2274 {
2275 	ASSERT(cpuid_checkpass(cpu, 3));
2276 
2277 	return (snprintf(s, n, "%s", cpu->cpu_m.mcpu_cpi->cpi_brandstr));
2278 }
2279 
2280 int
2281 cpuid_is_cmt(cpu_t *cpu)
2282 {
2283 	if (cpu == NULL)
2284 		cpu = CPU;
2285 
2286 	ASSERT(cpuid_checkpass(cpu, 1));
2287 
2288 	return (cpu->cpu_m.mcpu_cpi->cpi_chipid >= 0);
2289 }
2290 
2291 /*
2292  * AMD and Intel both implement the 64-bit variant of the syscall
2293  * instruction (syscallq), so if there's -any- support for syscall,
2294  * cpuid currently says "yes, we support this".
2295  *
2296  * However, Intel decided to -not- implement the 32-bit variant of the
2297  * syscall instruction, so we provide a predicate to allow our caller
2298  * to test that subtlety here.
2299  *
2300  * XXPV	Currently, 32-bit syscall instructions don't work via the hypervisor,
2301  *	even in the case where the hardware would in fact support it.
2302  */
2303 /*ARGSUSED*/
2304 int
2305 cpuid_syscall32_insn(cpu_t *cpu)
2306 {
2307 	ASSERT(cpuid_checkpass((cpu == NULL ? CPU : cpu), 1));
2308 
2309 #if !defined(__xpv)
2310 	if (cpu == NULL)
2311 		cpu = CPU;
2312 
2313 	/*CSTYLED*/
2314 	{
2315 		struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2316 
2317 		if (cpi->cpi_vendor == X86_VENDOR_AMD &&
2318 		    cpi->cpi_xmaxeax >= 0x80000001 &&
2319 		    (CPI_FEATURES_XTD_EDX(cpi) & CPUID_AMD_EDX_SYSC))
2320 			return (1);
2321 	}
2322 #endif
2323 	return (0);
2324 }
2325 
2326 int
2327 cpuid_getidstr(cpu_t *cpu, char *s, size_t n)
2328 {
2329 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2330 
2331 	static const char fmt[] =
2332 	    "x86 (%s %X family %d model %d step %d clock %d MHz)";
2333 	static const char fmt_ht[] =
2334 	    "x86 (chipid 0x%x %s %X family %d model %d step %d clock %d MHz)";
2335 
2336 	ASSERT(cpuid_checkpass(cpu, 1));
2337 
2338 	if (cpuid_is_cmt(cpu))
2339 		return (snprintf(s, n, fmt_ht, cpi->cpi_chipid,
2340 		    cpi->cpi_vendorstr, cpi->cpi_std[1].cp_eax,
2341 		    cpi->cpi_family, cpi->cpi_model,
2342 		    cpi->cpi_step, cpu->cpu_type_info.pi_clock));
2343 	return (snprintf(s, n, fmt,
2344 	    cpi->cpi_vendorstr, cpi->cpi_std[1].cp_eax,
2345 	    cpi->cpi_family, cpi->cpi_model,
2346 	    cpi->cpi_step, cpu->cpu_type_info.pi_clock));
2347 }
2348 
2349 const char *
2350 cpuid_getvendorstr(cpu_t *cpu)
2351 {
2352 	ASSERT(cpuid_checkpass(cpu, 1));
2353 	return ((const char *)cpu->cpu_m.mcpu_cpi->cpi_vendorstr);
2354 }
2355 
2356 uint_t
2357 cpuid_getvendor(cpu_t *cpu)
2358 {
2359 	ASSERT(cpuid_checkpass(cpu, 1));
2360 	return (cpu->cpu_m.mcpu_cpi->cpi_vendor);
2361 }
2362 
2363 uint_t
2364 cpuid_getfamily(cpu_t *cpu)
2365 {
2366 	ASSERT(cpuid_checkpass(cpu, 1));
2367 	return (cpu->cpu_m.mcpu_cpi->cpi_family);
2368 }
2369 
2370 uint_t
2371 cpuid_getmodel(cpu_t *cpu)
2372 {
2373 	ASSERT(cpuid_checkpass(cpu, 1));
2374 	return (cpu->cpu_m.mcpu_cpi->cpi_model);
2375 }
2376 
2377 uint_t
2378 cpuid_get_ncpu_per_chip(cpu_t *cpu)
2379 {
2380 	ASSERT(cpuid_checkpass(cpu, 1));
2381 	return (cpu->cpu_m.mcpu_cpi->cpi_ncpu_per_chip);
2382 }
2383 
2384 uint_t
2385 cpuid_get_ncore_per_chip(cpu_t *cpu)
2386 {
2387 	ASSERT(cpuid_checkpass(cpu, 1));
2388 	return (cpu->cpu_m.mcpu_cpi->cpi_ncore_per_chip);
2389 }
2390 
2391 uint_t
2392 cpuid_get_ncpu_sharing_last_cache(cpu_t *cpu)
2393 {
2394 	ASSERT(cpuid_checkpass(cpu, 2));
2395 	return (cpu->cpu_m.mcpu_cpi->cpi_ncpu_shr_last_cache);
2396 }
2397 
2398 id_t
2399 cpuid_get_last_lvl_cacheid(cpu_t *cpu)
2400 {
2401 	ASSERT(cpuid_checkpass(cpu, 2));
2402 	return (cpu->cpu_m.mcpu_cpi->cpi_last_lvl_cacheid);
2403 }
2404 
2405 uint_t
2406 cpuid_getstep(cpu_t *cpu)
2407 {
2408 	ASSERT(cpuid_checkpass(cpu, 1));
2409 	return (cpu->cpu_m.mcpu_cpi->cpi_step);
2410 }
2411 
2412 uint_t
2413 cpuid_getsig(struct cpu *cpu)
2414 {
2415 	ASSERT(cpuid_checkpass(cpu, 1));
2416 	return (cpu->cpu_m.mcpu_cpi->cpi_std[1].cp_eax);
2417 }
2418 
2419 uint32_t
2420 cpuid_getchiprev(struct cpu *cpu)
2421 {
2422 	ASSERT(cpuid_checkpass(cpu, 1));
2423 	return (cpu->cpu_m.mcpu_cpi->cpi_chiprev);
2424 }
2425 
2426 const char *
2427 cpuid_getchiprevstr(struct cpu *cpu)
2428 {
2429 	ASSERT(cpuid_checkpass(cpu, 1));
2430 	return (cpu->cpu_m.mcpu_cpi->cpi_chiprevstr);
2431 }
2432 
2433 uint32_t
2434 cpuid_getsockettype(struct cpu *cpu)
2435 {
2436 	ASSERT(cpuid_checkpass(cpu, 1));
2437 	return (cpu->cpu_m.mcpu_cpi->cpi_socket);
2438 }
2439 
2440 int
2441 cpuid_get_chipid(cpu_t *cpu)
2442 {
2443 	ASSERT(cpuid_checkpass(cpu, 1));
2444 
2445 	if (cpuid_is_cmt(cpu))
2446 		return (cpu->cpu_m.mcpu_cpi->cpi_chipid);
2447 	return (cpu->cpu_id);
2448 }
2449 
2450 id_t
2451 cpuid_get_coreid(cpu_t *cpu)
2452 {
2453 	ASSERT(cpuid_checkpass(cpu, 1));
2454 	return (cpu->cpu_m.mcpu_cpi->cpi_coreid);
2455 }
2456 
2457 int
2458 cpuid_get_pkgcoreid(cpu_t *cpu)
2459 {
2460 	ASSERT(cpuid_checkpass(cpu, 1));
2461 	return (cpu->cpu_m.mcpu_cpi->cpi_pkgcoreid);
2462 }
2463 
2464 int
2465 cpuid_get_clogid(cpu_t *cpu)
2466 {
2467 	ASSERT(cpuid_checkpass(cpu, 1));
2468 	return (cpu->cpu_m.mcpu_cpi->cpi_clogid);
2469 }
2470 
2471 void
2472 cpuid_get_addrsize(cpu_t *cpu, uint_t *pabits, uint_t *vabits)
2473 {
2474 	struct cpuid_info *cpi;
2475 
2476 	if (cpu == NULL)
2477 		cpu = CPU;
2478 	cpi = cpu->cpu_m.mcpu_cpi;
2479 
2480 	ASSERT(cpuid_checkpass(cpu, 1));
2481 
2482 	if (pabits)
2483 		*pabits = cpi->cpi_pabits;
2484 	if (vabits)
2485 		*vabits = cpi->cpi_vabits;
2486 }
2487 
2488 /*
2489  * Returns the number of data TLB entries for a corresponding
2490  * pagesize.  If it can't be computed, or isn't known, the
2491  * routine returns zero.  If you ask about an architecturally
2492  * impossible pagesize, the routine will panic (so that the
2493  * hat implementor knows that things are inconsistent.)
2494  */
2495 uint_t
2496 cpuid_get_dtlb_nent(cpu_t *cpu, size_t pagesize)
2497 {
2498 	struct cpuid_info *cpi;
2499 	uint_t dtlb_nent = 0;
2500 
2501 	if (cpu == NULL)
2502 		cpu = CPU;
2503 	cpi = cpu->cpu_m.mcpu_cpi;
2504 
2505 	ASSERT(cpuid_checkpass(cpu, 1));
2506 
2507 	/*
2508 	 * Check the L2 TLB info
2509 	 */
2510 	if (cpi->cpi_xmaxeax >= 0x80000006) {
2511 		struct cpuid_regs *cp = &cpi->cpi_extd[6];
2512 
2513 		switch (pagesize) {
2514 
2515 		case 4 * 1024:
2516 			/*
2517 			 * All zero in the top 16 bits of the register
2518 			 * indicates a unified TLB. Size is in low 16 bits.
2519 			 */
2520 			if ((cp->cp_ebx & 0xffff0000) == 0)
2521 				dtlb_nent = cp->cp_ebx & 0x0000ffff;
2522 			else
2523 				dtlb_nent = BITX(cp->cp_ebx, 27, 16);
2524 			break;
2525 
2526 		case 2 * 1024 * 1024:
2527 			if ((cp->cp_eax & 0xffff0000) == 0)
2528 				dtlb_nent = cp->cp_eax & 0x0000ffff;
2529 			else
2530 				dtlb_nent = BITX(cp->cp_eax, 27, 16);
2531 			break;
2532 
2533 		default:
2534 			panic("unknown L2 pagesize");
2535 			/*NOTREACHED*/
2536 		}
2537 	}
2538 
2539 	if (dtlb_nent != 0)
2540 		return (dtlb_nent);
2541 
2542 	/*
2543 	 * No L2 TLB support for this size, try L1.
2544 	 */
2545 	if (cpi->cpi_xmaxeax >= 0x80000005) {
2546 		struct cpuid_regs *cp = &cpi->cpi_extd[5];
2547 
2548 		switch (pagesize) {
2549 		case 4 * 1024:
2550 			dtlb_nent = BITX(cp->cp_ebx, 23, 16);
2551 			break;
2552 		case 2 * 1024 * 1024:
2553 			dtlb_nent = BITX(cp->cp_eax, 23, 16);
2554 			break;
2555 		default:
2556 			panic("unknown L1 d-TLB pagesize");
2557 			/*NOTREACHED*/
2558 		}
2559 	}
2560 
2561 	return (dtlb_nent);
2562 }
2563 
2564 /*
2565  * Return 0 if the erratum is not present or not applicable, positive
2566  * if it is, and negative if the status of the erratum is unknown.
2567  *
2568  * See "Revision Guide for AMD Athlon(tm) 64 and AMD Opteron(tm)
2569  * Processors" #25759, Rev 3.57, August 2005
2570  */
2571 int
2572 cpuid_opteron_erratum(cpu_t *cpu, uint_t erratum)
2573 {
2574 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2575 	uint_t eax;
2576 
2577 	/*
2578 	 * Bail out if this CPU isn't an AMD CPU, or if it's
2579 	 * a legacy (32-bit) AMD CPU.
2580 	 */
2581 	if (cpi->cpi_vendor != X86_VENDOR_AMD ||
2582 	    cpi->cpi_family == 4 || cpi->cpi_family == 5 ||
2583 	    cpi->cpi_family == 6)
2584 
2585 		return (0);
2586 
2587 	eax = cpi->cpi_std[1].cp_eax;
2588 
2589 #define	SH_B0(eax)	(eax == 0xf40 || eax == 0xf50)
2590 #define	SH_B3(eax) 	(eax == 0xf51)
2591 #define	B(eax)		(SH_B0(eax) || SH_B3(eax))
2592 
2593 #define	SH_C0(eax)	(eax == 0xf48 || eax == 0xf58)
2594 
2595 #define	SH_CG(eax)	(eax == 0xf4a || eax == 0xf5a || eax == 0xf7a)
2596 #define	DH_CG(eax)	(eax == 0xfc0 || eax == 0xfe0 || eax == 0xff0)
2597 #define	CH_CG(eax)	(eax == 0xf82 || eax == 0xfb2)
2598 #define	CG(eax)		(SH_CG(eax) || DH_CG(eax) || CH_CG(eax))
2599 
2600 #define	SH_D0(eax)	(eax == 0x10f40 || eax == 0x10f50 || eax == 0x10f70)
2601 #define	DH_D0(eax)	(eax == 0x10fc0 || eax == 0x10ff0)
2602 #define	CH_D0(eax)	(eax == 0x10f80 || eax == 0x10fb0)
2603 #define	D0(eax)		(SH_D0(eax) || DH_D0(eax) || CH_D0(eax))
2604 
2605 #define	SH_E0(eax)	(eax == 0x20f50 || eax == 0x20f40 || eax == 0x20f70)
2606 #define	JH_E1(eax)	(eax == 0x20f10)	/* JH8_E0 had 0x20f30 */
2607 #define	DH_E3(eax)	(eax == 0x20fc0 || eax == 0x20ff0)
2608 #define	SH_E4(eax)	(eax == 0x20f51 || eax == 0x20f71)
2609 #define	BH_E4(eax)	(eax == 0x20fb1)
2610 #define	SH_E5(eax)	(eax == 0x20f42)
2611 #define	DH_E6(eax)	(eax == 0x20ff2 || eax == 0x20fc2)
2612 #define	JH_E6(eax)	(eax == 0x20f12 || eax == 0x20f32)
2613 #define	EX(eax)		(SH_E0(eax) || JH_E1(eax) || DH_E3(eax) || \
2614 			    SH_E4(eax) || BH_E4(eax) || SH_E5(eax) || \
2615 			    DH_E6(eax) || JH_E6(eax))
2616 
2617 #define	DR_AX(eax)	(eax == 0x100f00 || eax == 0x100f01 || eax == 0x100f02)
2618 #define	DR_B0(eax)	(eax == 0x100f20)
2619 #define	DR_B1(eax)	(eax == 0x100f21)
2620 #define	DR_BA(eax)	(eax == 0x100f2a)
2621 #define	DR_B2(eax)	(eax == 0x100f22)
2622 #define	DR_B3(eax)	(eax == 0x100f23)
2623 #define	RB_C0(eax)	(eax == 0x100f40)
2624 
2625 	switch (erratum) {
2626 	case 1:
2627 		return (cpi->cpi_family < 0x10);
2628 	case 51:	/* what does the asterisk mean? */
2629 		return (B(eax) || SH_C0(eax) || CG(eax));
2630 	case 52:
2631 		return (B(eax));
2632 	case 57:
2633 		return (cpi->cpi_family <= 0x11);
2634 	case 58:
2635 		return (B(eax));
2636 	case 60:
2637 		return (cpi->cpi_family <= 0x11);
2638 	case 61:
2639 	case 62:
2640 	case 63:
2641 	case 64:
2642 	case 65:
2643 	case 66:
2644 	case 68:
2645 	case 69:
2646 	case 70:
2647 	case 71:
2648 		return (B(eax));
2649 	case 72:
2650 		return (SH_B0(eax));
2651 	case 74:
2652 		return (B(eax));
2653 	case 75:
2654 		return (cpi->cpi_family < 0x10);
2655 	case 76:
2656 		return (B(eax));
2657 	case 77:
2658 		return (cpi->cpi_family <= 0x11);
2659 	case 78:
2660 		return (B(eax) || SH_C0(eax));
2661 	case 79:
2662 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
2663 	case 80:
2664 	case 81:
2665 	case 82:
2666 		return (B(eax));
2667 	case 83:
2668 		return (B(eax) || SH_C0(eax) || CG(eax));
2669 	case 85:
2670 		return (cpi->cpi_family < 0x10);
2671 	case 86:
2672 		return (SH_C0(eax) || CG(eax));
2673 	case 88:
2674 #if !defined(__amd64)
2675 		return (0);
2676 #else
2677 		return (B(eax) || SH_C0(eax));
2678 #endif
2679 	case 89:
2680 		return (cpi->cpi_family < 0x10);
2681 	case 90:
2682 		return (B(eax) || SH_C0(eax) || CG(eax));
2683 	case 91:
2684 	case 92:
2685 		return (B(eax) || SH_C0(eax));
2686 	case 93:
2687 		return (SH_C0(eax));
2688 	case 94:
2689 		return (B(eax) || SH_C0(eax) || CG(eax));
2690 	case 95:
2691 #if !defined(__amd64)
2692 		return (0);
2693 #else
2694 		return (B(eax) || SH_C0(eax));
2695 #endif
2696 	case 96:
2697 		return (B(eax) || SH_C0(eax) || CG(eax));
2698 	case 97:
2699 	case 98:
2700 		return (SH_C0(eax) || CG(eax));
2701 	case 99:
2702 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
2703 	case 100:
2704 		return (B(eax) || SH_C0(eax));
2705 	case 101:
2706 	case 103:
2707 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
2708 	case 104:
2709 		return (SH_C0(eax) || CG(eax) || D0(eax));
2710 	case 105:
2711 	case 106:
2712 	case 107:
2713 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
2714 	case 108:
2715 		return (DH_CG(eax));
2716 	case 109:
2717 		return (SH_C0(eax) || CG(eax) || D0(eax));
2718 	case 110:
2719 		return (D0(eax) || EX(eax));
2720 	case 111:
2721 		return (CG(eax));
2722 	case 112:
2723 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
2724 	case 113:
2725 		return (eax == 0x20fc0);
2726 	case 114:
2727 		return (SH_E0(eax) || JH_E1(eax) || DH_E3(eax));
2728 	case 115:
2729 		return (SH_E0(eax) || JH_E1(eax));
2730 	case 116:
2731 		return (SH_E0(eax) || JH_E1(eax) || DH_E3(eax));
2732 	case 117:
2733 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
2734 	case 118:
2735 		return (SH_E0(eax) || JH_E1(eax) || SH_E4(eax) || BH_E4(eax) ||
2736 		    JH_E6(eax));
2737 	case 121:
2738 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
2739 	case 122:
2740 		return (cpi->cpi_family < 0x10 || cpi->cpi_family == 0x11);
2741 	case 123:
2742 		return (JH_E1(eax) || BH_E4(eax) || JH_E6(eax));
2743 	case 131:
2744 		return (cpi->cpi_family < 0x10);
2745 	case 6336786:
2746 		/*
2747 		 * Test for AdvPowerMgmtInfo.TscPStateInvariant
2748 		 * if this is a K8 family or newer processor
2749 		 */
2750 		if (CPI_FAMILY(cpi) == 0xf) {
2751 			struct cpuid_regs regs;
2752 			regs.cp_eax = 0x80000007;
2753 			(void) __cpuid_insn(&regs);
2754 			return (!(regs.cp_edx & 0x100));
2755 		}
2756 		return (0);
2757 	case 6323525:
2758 		return (((((eax >> 12) & 0xff00) + (eax & 0xf00)) |
2759 		    (((eax >> 4) & 0xf) | ((eax >> 12) & 0xf0))) < 0xf40);
2760 
2761 	case 6671130:
2762 		/*
2763 		 * check for processors (pre-Shanghai) that do not provide
2764 		 * optimal management of 1gb ptes in its tlb.
2765 		 */
2766 		return (cpi->cpi_family == 0x10 && cpi->cpi_model < 4);
2767 
2768 	case 298:
2769 		return (DR_AX(eax) || DR_B0(eax) || DR_B1(eax) || DR_BA(eax) ||
2770 		    DR_B2(eax) || RB_C0(eax));
2771 
2772 	default:
2773 		return (-1);
2774 
2775 	}
2776 }
2777 
2778 /*
2779  * Determine if specified erratum is present via OSVW (OS Visible Workaround).
2780  * Return 1 if erratum is present, 0 if not present and -1 if indeterminate.
2781  */
2782 int
2783 osvw_opteron_erratum(cpu_t *cpu, uint_t erratum)
2784 {
2785 	struct cpuid_info	*cpi;
2786 	uint_t			osvwid;
2787 	static int		osvwfeature = -1;
2788 	uint64_t		osvwlength;
2789 
2790 
2791 	cpi = cpu->cpu_m.mcpu_cpi;
2792 
2793 	/* confirm OSVW supported */
2794 	if (osvwfeature == -1) {
2795 		osvwfeature = cpi->cpi_extd[1].cp_ecx & CPUID_AMD_ECX_OSVW;
2796 	} else {
2797 		/* assert that osvw feature setting is consistent on all cpus */
2798 		ASSERT(osvwfeature ==
2799 		    (cpi->cpi_extd[1].cp_ecx & CPUID_AMD_ECX_OSVW));
2800 	}
2801 	if (!osvwfeature)
2802 		return (-1);
2803 
2804 	osvwlength = rdmsr(MSR_AMD_OSVW_ID_LEN) & OSVW_ID_LEN_MASK;
2805 
2806 	switch (erratum) {
2807 	case 298:	/* osvwid is 0 */
2808 		osvwid = 0;
2809 		if (osvwlength <= (uint64_t)osvwid) {
2810 			/* osvwid 0 is unknown */
2811 			return (-1);
2812 		}
2813 
2814 		/*
2815 		 * Check the OSVW STATUS MSR to determine the state
2816 		 * of the erratum where:
2817 		 *   0 - fixed by HW
2818 		 *   1 - BIOS has applied the workaround when BIOS
2819 		 *   workaround is available. (Or for other errata,
2820 		 *   OS workaround is required.)
2821 		 * For a value of 1, caller will confirm that the
2822 		 * erratum 298 workaround has indeed been applied by BIOS.
2823 		 *
2824 		 * A 1 may be set in cpus that have a HW fix
2825 		 * in a mixed cpu system. Regarding erratum 298:
2826 		 *   In a multiprocessor platform, the workaround above
2827 		 *   should be applied to all processors regardless of
2828 		 *   silicon revision when an affected processor is
2829 		 *   present.
2830 		 */
2831 
2832 		return (rdmsr(MSR_AMD_OSVW_STATUS +
2833 		    (osvwid / OSVW_ID_CNT_PER_MSR)) &
2834 		    (1ULL << (osvwid % OSVW_ID_CNT_PER_MSR)));
2835 
2836 	default:
2837 		return (-1);
2838 	}
2839 }
2840 
2841 static const char assoc_str[] = "associativity";
2842 static const char line_str[] = "line-size";
2843 static const char size_str[] = "size";
2844 
2845 static void
2846 add_cache_prop(dev_info_t *devi, const char *label, const char *type,
2847     uint32_t val)
2848 {
2849 	char buf[128];
2850 
2851 	/*
2852 	 * ndi_prop_update_int() is used because it is desirable for
2853 	 * DDI_PROP_HW_DEF and DDI_PROP_DONTSLEEP to be set.
2854 	 */
2855 	if (snprintf(buf, sizeof (buf), "%s-%s", label, type) < sizeof (buf))
2856 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, devi, buf, val);
2857 }
2858 
2859 /*
2860  * Intel-style cache/tlb description
2861  *
2862  * Standard cpuid level 2 gives a randomly ordered
2863  * selection of tags that index into a table that describes
2864  * cache and tlb properties.
2865  */
2866 
2867 static const char l1_icache_str[] = "l1-icache";
2868 static const char l1_dcache_str[] = "l1-dcache";
2869 static const char l2_cache_str[] = "l2-cache";
2870 static const char l3_cache_str[] = "l3-cache";
2871 static const char itlb4k_str[] = "itlb-4K";
2872 static const char dtlb4k_str[] = "dtlb-4K";
2873 static const char itlb2M_str[] = "itlb-2M";
2874 static const char itlb4M_str[] = "itlb-4M";
2875 static const char dtlb4M_str[] = "dtlb-4M";
2876 static const char dtlb24_str[] = "dtlb0-2M-4M";
2877 static const char itlb424_str[] = "itlb-4K-2M-4M";
2878 static const char itlb24_str[] = "itlb-2M-4M";
2879 static const char dtlb44_str[] = "dtlb-4K-4M";
2880 static const char sl1_dcache_str[] = "sectored-l1-dcache";
2881 static const char sl2_cache_str[] = "sectored-l2-cache";
2882 static const char itrace_str[] = "itrace-cache";
2883 static const char sl3_cache_str[] = "sectored-l3-cache";
2884 static const char sh_l2_tlb4k_str[] = "shared-l2-tlb-4k";
2885 
2886 static const struct cachetab {
2887 	uint8_t 	ct_code;
2888 	uint8_t		ct_assoc;
2889 	uint16_t 	ct_line_size;
2890 	size_t		ct_size;
2891 	const char	*ct_label;
2892 } intel_ctab[] = {
2893 	/*
2894 	 * maintain descending order!
2895 	 *
2896 	 * Codes ignored - Reason
2897 	 * ----------------------
2898 	 * 40H - intel_cpuid_4_cache_info() disambiguates l2/l3 cache
2899 	 * f0H/f1H - Currently we do not interpret prefetch size by design
2900 	 */
2901 	{ 0xe4, 16, 64, 8*1024*1024, l3_cache_str},
2902 	{ 0xe3, 16, 64, 4*1024*1024, l3_cache_str},
2903 	{ 0xe2, 16, 64, 2*1024*1024, l3_cache_str},
2904 	{ 0xde, 12, 64, 6*1024*1024, l3_cache_str},
2905 	{ 0xdd, 12, 64, 3*1024*1024, l3_cache_str},
2906 	{ 0xdc, 12, 64, ((1*1024*1024)+(512*1024)), l3_cache_str},
2907 	{ 0xd8, 8, 64, 4*1024*1024, l3_cache_str},
2908 	{ 0xd7, 8, 64, 2*1024*1024, l3_cache_str},
2909 	{ 0xd6, 8, 64, 1*1024*1024, l3_cache_str},
2910 	{ 0xd2, 4, 64, 2*1024*1024, l3_cache_str},
2911 	{ 0xd1, 4, 64, 1*1024*1024, l3_cache_str},
2912 	{ 0xd0, 4, 64, 512*1024, l3_cache_str},
2913 	{ 0xca, 4, 0, 512, sh_l2_tlb4k_str},
2914 	{ 0xc0, 4, 0, 8, dtlb44_str },
2915 	{ 0xba, 4, 0, 64, dtlb4k_str },
2916 	{ 0xb4, 4, 0, 256, dtlb4k_str },
2917 	{ 0xb3, 4, 0, 128, dtlb4k_str },
2918 	{ 0xb2, 4, 0, 64, itlb4k_str },
2919 	{ 0xb0, 4, 0, 128, itlb4k_str },
2920 	{ 0x87, 8, 64, 1024*1024, l2_cache_str},
2921 	{ 0x86, 4, 64, 512*1024, l2_cache_str},
2922 	{ 0x85, 8, 32, 2*1024*1024, l2_cache_str},
2923 	{ 0x84, 8, 32, 1024*1024, l2_cache_str},
2924 	{ 0x83, 8, 32, 512*1024, l2_cache_str},
2925 	{ 0x82, 8, 32, 256*1024, l2_cache_str},
2926 	{ 0x80, 8, 64, 512*1024, l2_cache_str},
2927 	{ 0x7f, 2, 64, 512*1024, l2_cache_str},
2928 	{ 0x7d, 8, 64, 2*1024*1024, sl2_cache_str},
2929 	{ 0x7c, 8, 64, 1024*1024, sl2_cache_str},
2930 	{ 0x7b, 8, 64, 512*1024, sl2_cache_str},
2931 	{ 0x7a, 8, 64, 256*1024, sl2_cache_str},
2932 	{ 0x79, 8, 64, 128*1024, sl2_cache_str},
2933 	{ 0x78, 8, 64, 1024*1024, l2_cache_str},
2934 	{ 0x73, 8, 0, 64*1024, itrace_str},
2935 	{ 0x72, 8, 0, 32*1024, itrace_str},
2936 	{ 0x71, 8, 0, 16*1024, itrace_str},
2937 	{ 0x70, 8, 0, 12*1024, itrace_str},
2938 	{ 0x68, 4, 64, 32*1024, sl1_dcache_str},
2939 	{ 0x67, 4, 64, 16*1024, sl1_dcache_str},
2940 	{ 0x66, 4, 64, 8*1024, sl1_dcache_str},
2941 	{ 0x60, 8, 64, 16*1024, sl1_dcache_str},
2942 	{ 0x5d, 0, 0, 256, dtlb44_str},
2943 	{ 0x5c, 0, 0, 128, dtlb44_str},
2944 	{ 0x5b, 0, 0, 64, dtlb44_str},
2945 	{ 0x5a, 4, 0, 32, dtlb24_str},
2946 	{ 0x59, 0, 0, 16, dtlb4k_str},
2947 	{ 0x57, 4, 0, 16, dtlb4k_str},
2948 	{ 0x56, 4, 0, 16, dtlb4M_str},
2949 	{ 0x55, 0, 0, 7, itlb24_str},
2950 	{ 0x52, 0, 0, 256, itlb424_str},
2951 	{ 0x51, 0, 0, 128, itlb424_str},
2952 	{ 0x50, 0, 0, 64, itlb424_str},
2953 	{ 0x4f, 0, 0, 32, itlb4k_str},
2954 	{ 0x4e, 24, 64, 6*1024*1024, l2_cache_str},
2955 	{ 0x4d, 16, 64, 16*1024*1024, l3_cache_str},
2956 	{ 0x4c, 12, 64, 12*1024*1024, l3_cache_str},
2957 	{ 0x4b, 16, 64, 8*1024*1024, l3_cache_str},
2958 	{ 0x4a, 12, 64, 6*1024*1024, l3_cache_str},
2959 	{ 0x49, 16, 64, 4*1024*1024, l3_cache_str},
2960 	{ 0x48, 12, 64, 3*1024*1024, l2_cache_str},
2961 	{ 0x47, 8, 64, 8*1024*1024, l3_cache_str},
2962 	{ 0x46, 4, 64, 4*1024*1024, l3_cache_str},
2963 	{ 0x45, 4, 32, 2*1024*1024, l2_cache_str},
2964 	{ 0x44, 4, 32, 1024*1024, l2_cache_str},
2965 	{ 0x43, 4, 32, 512*1024, l2_cache_str},
2966 	{ 0x42, 4, 32, 256*1024, l2_cache_str},
2967 	{ 0x41, 4, 32, 128*1024, l2_cache_str},
2968 	{ 0x3e, 4, 64, 512*1024, sl2_cache_str},
2969 	{ 0x3d, 6, 64, 384*1024, sl2_cache_str},
2970 	{ 0x3c, 4, 64, 256*1024, sl2_cache_str},
2971 	{ 0x3b, 2, 64, 128*1024, sl2_cache_str},
2972 	{ 0x3a, 6, 64, 192*1024, sl2_cache_str},
2973 	{ 0x39, 4, 64, 128*1024, sl2_cache_str},
2974 	{ 0x30, 8, 64, 32*1024, l1_icache_str},
2975 	{ 0x2c, 8, 64, 32*1024, l1_dcache_str},
2976 	{ 0x29, 8, 64, 4096*1024, sl3_cache_str},
2977 	{ 0x25, 8, 64, 2048*1024, sl3_cache_str},
2978 	{ 0x23, 8, 64, 1024*1024, sl3_cache_str},
2979 	{ 0x22, 4, 64, 512*1024, sl3_cache_str},
2980 	{ 0x0e, 6, 64, 24*1024, l1_dcache_str},
2981 	{ 0x0d, 4, 32, 16*1024, l1_dcache_str},
2982 	{ 0x0c, 4, 32, 16*1024, l1_dcache_str},
2983 	{ 0x0b, 4, 0, 4, itlb4M_str},
2984 	{ 0x0a, 2, 32, 8*1024, l1_dcache_str},
2985 	{ 0x08, 4, 32, 16*1024, l1_icache_str},
2986 	{ 0x06, 4, 32, 8*1024, l1_icache_str},
2987 	{ 0x05, 4, 0, 32, dtlb4M_str},
2988 	{ 0x04, 4, 0, 8, dtlb4M_str},
2989 	{ 0x03, 4, 0, 64, dtlb4k_str},
2990 	{ 0x02, 4, 0, 2, itlb4M_str},
2991 	{ 0x01, 4, 0, 32, itlb4k_str},
2992 	{ 0 }
2993 };
2994 
2995 static const struct cachetab cyrix_ctab[] = {
2996 	{ 0x70, 4, 0, 32, "tlb-4K" },
2997 	{ 0x80, 4, 16, 16*1024, "l1-cache" },
2998 	{ 0 }
2999 };
3000 
3001 /*
3002  * Search a cache table for a matching entry
3003  */
3004 static const struct cachetab *
3005 find_cacheent(const struct cachetab *ct, uint_t code)
3006 {
3007 	if (code != 0) {
3008 		for (; ct->ct_code != 0; ct++)
3009 			if (ct->ct_code <= code)
3010 				break;
3011 		if (ct->ct_code == code)
3012 			return (ct);
3013 	}
3014 	return (NULL);
3015 }
3016 
3017 /*
3018  * Populate cachetab entry with L2 or L3 cache-information using
3019  * cpuid function 4. This function is called from intel_walk_cacheinfo()
3020  * when descriptor 0x49 is encountered. It returns 0 if no such cache
3021  * information is found.
3022  */
3023 static int
3024 intel_cpuid_4_cache_info(struct cachetab *ct, struct cpuid_info *cpi)
3025 {
3026 	uint32_t level, i;
3027 	int ret = 0;
3028 
3029 	for (i = 0; i < cpi->cpi_std_4_size; i++) {
3030 		level = CPI_CACHE_LVL(cpi->cpi_std_4[i]);
3031 
3032 		if (level == 2 || level == 3) {
3033 			ct->ct_assoc = CPI_CACHE_WAYS(cpi->cpi_std_4[i]) + 1;
3034 			ct->ct_line_size =
3035 			    CPI_CACHE_COH_LN_SZ(cpi->cpi_std_4[i]) + 1;
3036 			ct->ct_size = ct->ct_assoc *
3037 			    (CPI_CACHE_PARTS(cpi->cpi_std_4[i]) + 1) *
3038 			    ct->ct_line_size *
3039 			    (cpi->cpi_std_4[i]->cp_ecx + 1);
3040 
3041 			if (level == 2) {
3042 				ct->ct_label = l2_cache_str;
3043 			} else if (level == 3) {
3044 				ct->ct_label = l3_cache_str;
3045 			}
3046 			ret = 1;
3047 		}
3048 	}
3049 
3050 	return (ret);
3051 }
3052 
3053 /*
3054  * Walk the cacheinfo descriptor, applying 'func' to every valid element
3055  * The walk is terminated if the walker returns non-zero.
3056  */
3057 static void
3058 intel_walk_cacheinfo(struct cpuid_info *cpi,
3059     void *arg, int (*func)(void *, const struct cachetab *))
3060 {
3061 	const struct cachetab *ct;
3062 	struct cachetab des_49_ct, des_b1_ct;
3063 	uint8_t *dp;
3064 	int i;
3065 
3066 	if ((dp = cpi->cpi_cacheinfo) == NULL)
3067 		return;
3068 	for (i = 0; i < cpi->cpi_ncache; i++, dp++) {
3069 		/*
3070 		 * For overloaded descriptor 0x49 we use cpuid function 4
3071 		 * if supported by the current processor, to create
3072 		 * cache information.
3073 		 * For overloaded descriptor 0xb1 we use X86_PAE flag
3074 		 * to disambiguate the cache information.
3075 		 */
3076 		if (*dp == 0x49 && cpi->cpi_maxeax >= 0x4 &&
3077 		    intel_cpuid_4_cache_info(&des_49_ct, cpi) == 1) {
3078 				ct = &des_49_ct;
3079 		} else if (*dp == 0xb1) {
3080 			des_b1_ct.ct_code = 0xb1;
3081 			des_b1_ct.ct_assoc = 4;
3082 			des_b1_ct.ct_line_size = 0;
3083 			if (x86_feature & X86_PAE) {
3084 				des_b1_ct.ct_size = 8;
3085 				des_b1_ct.ct_label = itlb2M_str;
3086 			} else {
3087 				des_b1_ct.ct_size = 4;
3088 				des_b1_ct.ct_label = itlb4M_str;
3089 			}
3090 			ct = &des_b1_ct;
3091 		} else {
3092 			if ((ct = find_cacheent(intel_ctab, *dp)) == NULL) {
3093 				continue;
3094 			}
3095 		}
3096 
3097 		if (func(arg, ct) != 0) {
3098 			break;
3099 		}
3100 	}
3101 }
3102 
3103 /*
3104  * (Like the Intel one, except for Cyrix CPUs)
3105  */
3106 static void
3107 cyrix_walk_cacheinfo(struct cpuid_info *cpi,
3108     void *arg, int (*func)(void *, const struct cachetab *))
3109 {
3110 	const struct cachetab *ct;
3111 	uint8_t *dp;
3112 	int i;
3113 
3114 	if ((dp = cpi->cpi_cacheinfo) == NULL)
3115 		return;
3116 	for (i = 0; i < cpi->cpi_ncache; i++, dp++) {
3117 		/*
3118 		 * Search Cyrix-specific descriptor table first ..
3119 		 */
3120 		if ((ct = find_cacheent(cyrix_ctab, *dp)) != NULL) {
3121 			if (func(arg, ct) != 0)
3122 				break;
3123 			continue;
3124 		}
3125 		/*
3126 		 * .. else fall back to the Intel one
3127 		 */
3128 		if ((ct = find_cacheent(intel_ctab, *dp)) != NULL) {
3129 			if (func(arg, ct) != 0)
3130 				break;
3131 			continue;
3132 		}
3133 	}
3134 }
3135 
3136 /*
3137  * A cacheinfo walker that adds associativity, line-size, and size properties
3138  * to the devinfo node it is passed as an argument.
3139  */
3140 static int
3141 add_cacheent_props(void *arg, const struct cachetab *ct)
3142 {
3143 	dev_info_t *devi = arg;
3144 
3145 	add_cache_prop(devi, ct->ct_label, assoc_str, ct->ct_assoc);
3146 	if (ct->ct_line_size != 0)
3147 		add_cache_prop(devi, ct->ct_label, line_str,
3148 		    ct->ct_line_size);
3149 	add_cache_prop(devi, ct->ct_label, size_str, ct->ct_size);
3150 	return (0);
3151 }
3152 
3153 
3154 static const char fully_assoc[] = "fully-associative?";
3155 
3156 /*
3157  * AMD style cache/tlb description
3158  *
3159  * Extended functions 5 and 6 directly describe properties of
3160  * tlbs and various cache levels.
3161  */
3162 static void
3163 add_amd_assoc(dev_info_t *devi, const char *label, uint_t assoc)
3164 {
3165 	switch (assoc) {
3166 	case 0:	/* reserved; ignore */
3167 		break;
3168 	default:
3169 		add_cache_prop(devi, label, assoc_str, assoc);
3170 		break;
3171 	case 0xff:
3172 		add_cache_prop(devi, label, fully_assoc, 1);
3173 		break;
3174 	}
3175 }
3176 
3177 static void
3178 add_amd_tlb(dev_info_t *devi, const char *label, uint_t assoc, uint_t size)
3179 {
3180 	if (size == 0)
3181 		return;
3182 	add_cache_prop(devi, label, size_str, size);
3183 	add_amd_assoc(devi, label, assoc);
3184 }
3185 
3186 static void
3187 add_amd_cache(dev_info_t *devi, const char *label,
3188     uint_t size, uint_t assoc, uint_t lines_per_tag, uint_t line_size)
3189 {
3190 	if (size == 0 || line_size == 0)
3191 		return;
3192 	add_amd_assoc(devi, label, assoc);
3193 	/*
3194 	 * Most AMD parts have a sectored cache. Multiple cache lines are
3195 	 * associated with each tag. A sector consists of all cache lines
3196 	 * associated with a tag. For example, the AMD K6-III has a sector
3197 	 * size of 2 cache lines per tag.
3198 	 */
3199 	if (lines_per_tag != 0)
3200 		add_cache_prop(devi, label, "lines-per-tag", lines_per_tag);
3201 	add_cache_prop(devi, label, line_str, line_size);
3202 	add_cache_prop(devi, label, size_str, size * 1024);
3203 }
3204 
3205 static void
3206 add_amd_l2_assoc(dev_info_t *devi, const char *label, uint_t assoc)
3207 {
3208 	switch (assoc) {
3209 	case 0:	/* off */
3210 		break;
3211 	case 1:
3212 	case 2:
3213 	case 4:
3214 		add_cache_prop(devi, label, assoc_str, assoc);
3215 		break;
3216 	case 6:
3217 		add_cache_prop(devi, label, assoc_str, 8);
3218 		break;
3219 	case 8:
3220 		add_cache_prop(devi, label, assoc_str, 16);
3221 		break;
3222 	case 0xf:
3223 		add_cache_prop(devi, label, fully_assoc, 1);
3224 		break;
3225 	default: /* reserved; ignore */
3226 		break;
3227 	}
3228 }
3229 
3230 static void
3231 add_amd_l2_tlb(dev_info_t *devi, const char *label, uint_t assoc, uint_t size)
3232 {
3233 	if (size == 0 || assoc == 0)
3234 		return;
3235 	add_amd_l2_assoc(devi, label, assoc);
3236 	add_cache_prop(devi, label, size_str, size);
3237 }
3238 
3239 static void
3240 add_amd_l2_cache(dev_info_t *devi, const char *label,
3241     uint_t size, uint_t assoc, uint_t lines_per_tag, uint_t line_size)
3242 {
3243 	if (size == 0 || assoc == 0 || line_size == 0)
3244 		return;
3245 	add_amd_l2_assoc(devi, label, assoc);
3246 	if (lines_per_tag != 0)
3247 		add_cache_prop(devi, label, "lines-per-tag", lines_per_tag);
3248 	add_cache_prop(devi, label, line_str, line_size);
3249 	add_cache_prop(devi, label, size_str, size * 1024);
3250 }
3251 
3252 static void
3253 amd_cache_info(struct cpuid_info *cpi, dev_info_t *devi)
3254 {
3255 	struct cpuid_regs *cp;
3256 
3257 	if (cpi->cpi_xmaxeax < 0x80000005)
3258 		return;
3259 	cp = &cpi->cpi_extd[5];
3260 
3261 	/*
3262 	 * 4M/2M L1 TLB configuration
3263 	 *
3264 	 * We report the size for 2M pages because AMD uses two
3265 	 * TLB entries for one 4M page.
3266 	 */
3267 	add_amd_tlb(devi, "dtlb-2M",
3268 	    BITX(cp->cp_eax, 31, 24), BITX(cp->cp_eax, 23, 16));
3269 	add_amd_tlb(devi, "itlb-2M",
3270 	    BITX(cp->cp_eax, 15, 8), BITX(cp->cp_eax, 7, 0));
3271 
3272 	/*
3273 	 * 4K L1 TLB configuration
3274 	 */
3275 
3276 	switch (cpi->cpi_vendor) {
3277 		uint_t nentries;
3278 	case X86_VENDOR_TM:
3279 		if (cpi->cpi_family >= 5) {
3280 			/*
3281 			 * Crusoe processors have 256 TLB entries, but
3282 			 * cpuid data format constrains them to only
3283 			 * reporting 255 of them.
3284 			 */
3285 			if ((nentries = BITX(cp->cp_ebx, 23, 16)) == 255)
3286 				nentries = 256;
3287 			/*
3288 			 * Crusoe processors also have a unified TLB
3289 			 */
3290 			add_amd_tlb(devi, "tlb-4K", BITX(cp->cp_ebx, 31, 24),
3291 			    nentries);
3292 			break;
3293 		}
3294 		/*FALLTHROUGH*/
3295 	default:
3296 		add_amd_tlb(devi, itlb4k_str,
3297 		    BITX(cp->cp_ebx, 31, 24), BITX(cp->cp_ebx, 23, 16));
3298 		add_amd_tlb(devi, dtlb4k_str,
3299 		    BITX(cp->cp_ebx, 15, 8), BITX(cp->cp_ebx, 7, 0));
3300 		break;
3301 	}
3302 
3303 	/*
3304 	 * data L1 cache configuration
3305 	 */
3306 
3307 	add_amd_cache(devi, l1_dcache_str,
3308 	    BITX(cp->cp_ecx, 31, 24), BITX(cp->cp_ecx, 23, 16),
3309 	    BITX(cp->cp_ecx, 15, 8), BITX(cp->cp_ecx, 7, 0));
3310 
3311 	/*
3312 	 * code L1 cache configuration
3313 	 */
3314 
3315 	add_amd_cache(devi, l1_icache_str,
3316 	    BITX(cp->cp_edx, 31, 24), BITX(cp->cp_edx, 23, 16),
3317 	    BITX(cp->cp_edx, 15, 8), BITX(cp->cp_edx, 7, 0));
3318 
3319 	if (cpi->cpi_xmaxeax < 0x80000006)
3320 		return;
3321 	cp = &cpi->cpi_extd[6];
3322 
3323 	/* Check for a unified L2 TLB for large pages */
3324 
3325 	if (BITX(cp->cp_eax, 31, 16) == 0)
3326 		add_amd_l2_tlb(devi, "l2-tlb-2M",
3327 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
3328 	else {
3329 		add_amd_l2_tlb(devi, "l2-dtlb-2M",
3330 		    BITX(cp->cp_eax, 31, 28), BITX(cp->cp_eax, 27, 16));
3331 		add_amd_l2_tlb(devi, "l2-itlb-2M",
3332 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
3333 	}
3334 
3335 	/* Check for a unified L2 TLB for 4K pages */
3336 
3337 	if (BITX(cp->cp_ebx, 31, 16) == 0) {
3338 		add_amd_l2_tlb(devi, "l2-tlb-4K",
3339 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
3340 	} else {
3341 		add_amd_l2_tlb(devi, "l2-dtlb-4K",
3342 		    BITX(cp->cp_eax, 31, 28), BITX(cp->cp_eax, 27, 16));
3343 		add_amd_l2_tlb(devi, "l2-itlb-4K",
3344 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
3345 	}
3346 
3347 	add_amd_l2_cache(devi, l2_cache_str,
3348 	    BITX(cp->cp_ecx, 31, 16), BITX(cp->cp_ecx, 15, 12),
3349 	    BITX(cp->cp_ecx, 11, 8), BITX(cp->cp_ecx, 7, 0));
3350 }
3351 
3352 /*
3353  * There are two basic ways that the x86 world describes it cache
3354  * and tlb architecture - Intel's way and AMD's way.
3355  *
3356  * Return which flavor of cache architecture we should use
3357  */
3358 static int
3359 x86_which_cacheinfo(struct cpuid_info *cpi)
3360 {
3361 	switch (cpi->cpi_vendor) {
3362 	case X86_VENDOR_Intel:
3363 		if (cpi->cpi_maxeax >= 2)
3364 			return (X86_VENDOR_Intel);
3365 		break;
3366 	case X86_VENDOR_AMD:
3367 		/*
3368 		 * The K5 model 1 was the first part from AMD that reported
3369 		 * cache sizes via extended cpuid functions.
3370 		 */
3371 		if (cpi->cpi_family > 5 ||
3372 		    (cpi->cpi_family == 5 && cpi->cpi_model >= 1))
3373 			return (X86_VENDOR_AMD);
3374 		break;
3375 	case X86_VENDOR_TM:
3376 		if (cpi->cpi_family >= 5)
3377 			return (X86_VENDOR_AMD);
3378 		/*FALLTHROUGH*/
3379 	default:
3380 		/*
3381 		 * If they have extended CPU data for 0x80000005
3382 		 * then we assume they have AMD-format cache
3383 		 * information.
3384 		 *
3385 		 * If not, and the vendor happens to be Cyrix,
3386 		 * then try our-Cyrix specific handler.
3387 		 *
3388 		 * If we're not Cyrix, then assume we're using Intel's
3389 		 * table-driven format instead.
3390 		 */
3391 		if (cpi->cpi_xmaxeax >= 0x80000005)
3392 			return (X86_VENDOR_AMD);
3393 		else if (cpi->cpi_vendor == X86_VENDOR_Cyrix)
3394 			return (X86_VENDOR_Cyrix);
3395 		else if (cpi->cpi_maxeax >= 2)
3396 			return (X86_VENDOR_Intel);
3397 		break;
3398 	}
3399 	return (-1);
3400 }
3401 
3402 /*
3403  * create a node for the given cpu under the prom root node.
3404  * Also, create a cpu node in the device tree.
3405  */
3406 static dev_info_t *cpu_nex_devi = NULL;
3407 static kmutex_t cpu_node_lock;
3408 
3409 /*
3410  * Called from post_startup() and mp_startup()
3411  */
3412 void
3413 add_cpunode2devtree(processorid_t cpu_id, struct cpuid_info *cpi)
3414 {
3415 	dev_info_t *cpu_devi;
3416 	int create;
3417 
3418 	mutex_enter(&cpu_node_lock);
3419 
3420 	/*
3421 	 * create a nexus node for all cpus identified as 'cpu_id' under
3422 	 * the root node.
3423 	 */
3424 	if (cpu_nex_devi == NULL) {
3425 		if (ndi_devi_alloc(ddi_root_node(), "cpus",
3426 		    (pnode_t)DEVI_SID_NODEID, &cpu_nex_devi) != NDI_SUCCESS) {
3427 			mutex_exit(&cpu_node_lock);
3428 			return;
3429 		}
3430 		(void) ndi_devi_online(cpu_nex_devi, 0);
3431 	}
3432 
3433 	/*
3434 	 * create a child node for cpu identified as 'cpu_id'
3435 	 */
3436 	cpu_devi = ddi_add_child(cpu_nex_devi, "cpu", DEVI_SID_NODEID,
3437 	    cpu_id);
3438 	if (cpu_devi == NULL) {
3439 		mutex_exit(&cpu_node_lock);
3440 		return;
3441 	}
3442 
3443 	/* device_type */
3444 
3445 	(void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi,
3446 	    "device_type", "cpu");
3447 
3448 	/* reg */
3449 
3450 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3451 	    "reg", cpu_id);
3452 
3453 	/* cpu-mhz, and clock-frequency */
3454 
3455 	if (cpu_freq > 0) {
3456 		long long mul;
3457 
3458 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3459 		    "cpu-mhz", cpu_freq);
3460 
3461 		if ((mul = cpu_freq * 1000000LL) <= INT_MAX)
3462 			(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3463 			    "clock-frequency", (int)mul);
3464 	}
3465 
3466 	(void) ndi_devi_online(cpu_devi, 0);
3467 
3468 	if ((x86_feature & X86_CPUID) == 0) {
3469 		mutex_exit(&cpu_node_lock);
3470 		return;
3471 	}
3472 
3473 	/* vendor-id */
3474 
3475 	(void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi,
3476 	    "vendor-id", cpi->cpi_vendorstr);
3477 
3478 	if (cpi->cpi_maxeax == 0) {
3479 		mutex_exit(&cpu_node_lock);
3480 		return;
3481 	}
3482 
3483 	/*
3484 	 * family, model, and step
3485 	 */
3486 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3487 	    "family", CPI_FAMILY(cpi));
3488 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3489 	    "cpu-model", CPI_MODEL(cpi));
3490 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3491 	    "stepping-id", CPI_STEP(cpi));
3492 
3493 	/* type */
3494 
3495 	switch (cpi->cpi_vendor) {
3496 	case X86_VENDOR_Intel:
3497 		create = 1;
3498 		break;
3499 	default:
3500 		create = 0;
3501 		break;
3502 	}
3503 	if (create)
3504 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3505 		    "type", CPI_TYPE(cpi));
3506 
3507 	/* ext-family */
3508 
3509 	switch (cpi->cpi_vendor) {
3510 	case X86_VENDOR_Intel:
3511 	case X86_VENDOR_AMD:
3512 		create = cpi->cpi_family >= 0xf;
3513 		break;
3514 	default:
3515 		create = 0;
3516 		break;
3517 	}
3518 	if (create)
3519 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3520 		    "ext-family", CPI_FAMILY_XTD(cpi));
3521 
3522 	/* ext-model */
3523 
3524 	switch (cpi->cpi_vendor) {
3525 	case X86_VENDOR_Intel:
3526 		create = IS_EXTENDED_MODEL_INTEL(cpi);
3527 		break;
3528 	case X86_VENDOR_AMD:
3529 		create = CPI_FAMILY(cpi) == 0xf;
3530 		break;
3531 	default:
3532 		create = 0;
3533 		break;
3534 	}
3535 	if (create)
3536 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3537 		    "ext-model", CPI_MODEL_XTD(cpi));
3538 
3539 	/* generation */
3540 
3541 	switch (cpi->cpi_vendor) {
3542 	case X86_VENDOR_AMD:
3543 		/*
3544 		 * AMD K5 model 1 was the first part to support this
3545 		 */
3546 		create = cpi->cpi_xmaxeax >= 0x80000001;
3547 		break;
3548 	default:
3549 		create = 0;
3550 		break;
3551 	}
3552 	if (create)
3553 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3554 		    "generation", BITX((cpi)->cpi_extd[1].cp_eax, 11, 8));
3555 
3556 	/* brand-id */
3557 
3558 	switch (cpi->cpi_vendor) {
3559 	case X86_VENDOR_Intel:
3560 		/*
3561 		 * brand id first appeared on Pentium III Xeon model 8,
3562 		 * and Celeron model 8 processors and Opteron
3563 		 */
3564 		create = cpi->cpi_family > 6 ||
3565 		    (cpi->cpi_family == 6 && cpi->cpi_model >= 8);
3566 		break;
3567 	case X86_VENDOR_AMD:
3568 		create = cpi->cpi_family >= 0xf;
3569 		break;
3570 	default:
3571 		create = 0;
3572 		break;
3573 	}
3574 	if (create && cpi->cpi_brandid != 0) {
3575 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3576 		    "brand-id", cpi->cpi_brandid);
3577 	}
3578 
3579 	/* chunks, and apic-id */
3580 
3581 	switch (cpi->cpi_vendor) {
3582 		/*
3583 		 * first available on Pentium IV and Opteron (K8)
3584 		 */
3585 	case X86_VENDOR_Intel:
3586 		create = IS_NEW_F6(cpi) || cpi->cpi_family >= 0xf;
3587 		break;
3588 	case X86_VENDOR_AMD:
3589 		create = cpi->cpi_family >= 0xf;
3590 		break;
3591 	default:
3592 		create = 0;
3593 		break;
3594 	}
3595 	if (create) {
3596 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3597 		    "chunks", CPI_CHUNKS(cpi));
3598 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3599 		    "apic-id", cpi->cpi_apicid);
3600 		if (cpi->cpi_chipid >= 0) {
3601 			(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3602 			    "chip#", cpi->cpi_chipid);
3603 			(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3604 			    "clog#", cpi->cpi_clogid);
3605 		}
3606 	}
3607 
3608 	/* cpuid-features */
3609 
3610 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3611 	    "cpuid-features", CPI_FEATURES_EDX(cpi));
3612 
3613 
3614 	/* cpuid-features-ecx */
3615 
3616 	switch (cpi->cpi_vendor) {
3617 	case X86_VENDOR_Intel:
3618 		create = IS_NEW_F6(cpi) || cpi->cpi_family >= 0xf;
3619 		break;
3620 	default:
3621 		create = 0;
3622 		break;
3623 	}
3624 	if (create)
3625 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3626 		    "cpuid-features-ecx", CPI_FEATURES_ECX(cpi));
3627 
3628 	/* ext-cpuid-features */
3629 
3630 	switch (cpi->cpi_vendor) {
3631 	case X86_VENDOR_Intel:
3632 	case X86_VENDOR_AMD:
3633 	case X86_VENDOR_Cyrix:
3634 	case X86_VENDOR_TM:
3635 	case X86_VENDOR_Centaur:
3636 		create = cpi->cpi_xmaxeax >= 0x80000001;
3637 		break;
3638 	default:
3639 		create = 0;
3640 		break;
3641 	}
3642 	if (create) {
3643 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3644 		    "ext-cpuid-features", CPI_FEATURES_XTD_EDX(cpi));
3645 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3646 		    "ext-cpuid-features-ecx", CPI_FEATURES_XTD_ECX(cpi));
3647 	}
3648 
3649 	/*
3650 	 * Brand String first appeared in Intel Pentium IV, AMD K5
3651 	 * model 1, and Cyrix GXm.  On earlier models we try and
3652 	 * simulate something similar .. so this string should always
3653 	 * same -something- about the processor, however lame.
3654 	 */
3655 	(void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi,
3656 	    "brand-string", cpi->cpi_brandstr);
3657 
3658 	/*
3659 	 * Finally, cache and tlb information
3660 	 */
3661 	switch (x86_which_cacheinfo(cpi)) {
3662 	case X86_VENDOR_Intel:
3663 		intel_walk_cacheinfo(cpi, cpu_devi, add_cacheent_props);
3664 		break;
3665 	case X86_VENDOR_Cyrix:
3666 		cyrix_walk_cacheinfo(cpi, cpu_devi, add_cacheent_props);
3667 		break;
3668 	case X86_VENDOR_AMD:
3669 		amd_cache_info(cpi, cpu_devi);
3670 		break;
3671 	default:
3672 		break;
3673 	}
3674 
3675 	mutex_exit(&cpu_node_lock);
3676 }
3677 
3678 struct l2info {
3679 	int *l2i_csz;
3680 	int *l2i_lsz;
3681 	int *l2i_assoc;
3682 	int l2i_ret;
3683 };
3684 
3685 /*
3686  * A cacheinfo walker that fetches the size, line-size and associativity
3687  * of the L2 cache
3688  */
3689 static int
3690 intel_l2cinfo(void *arg, const struct cachetab *ct)
3691 {
3692 	struct l2info *l2i = arg;
3693 	int *ip;
3694 
3695 	if (ct->ct_label != l2_cache_str &&
3696 	    ct->ct_label != sl2_cache_str)
3697 		return (0);	/* not an L2 -- keep walking */
3698 
3699 	if ((ip = l2i->l2i_csz) != NULL)
3700 		*ip = ct->ct_size;
3701 	if ((ip = l2i->l2i_lsz) != NULL)
3702 		*ip = ct->ct_line_size;
3703 	if ((ip = l2i->l2i_assoc) != NULL)
3704 		*ip = ct->ct_assoc;
3705 	l2i->l2i_ret = ct->ct_size;
3706 	return (1);		/* was an L2 -- terminate walk */
3707 }
3708 
3709 /*
3710  * AMD L2/L3 Cache and TLB Associativity Field Definition:
3711  *
3712  *	Unlike the associativity for the L1 cache and tlb where the 8 bit
3713  *	value is the associativity, the associativity for the L2 cache and
3714  *	tlb is encoded in the following table. The 4 bit L2 value serves as
3715  *	an index into the amd_afd[] array to determine the associativity.
3716  *	-1 is undefined. 0 is fully associative.
3717  */
3718 
3719 static int amd_afd[] =
3720 	{-1, 1, 2, -1, 4, -1, 8, -1, 16, -1, 32, 48, 64, 96, 128, 0};
3721 
3722 static void
3723 amd_l2cacheinfo(struct cpuid_info *cpi, struct l2info *l2i)
3724 {
3725 	struct cpuid_regs *cp;
3726 	uint_t size, assoc;
3727 	int i;
3728 	int *ip;
3729 
3730 	if (cpi->cpi_xmaxeax < 0x80000006)
3731 		return;
3732 	cp = &cpi->cpi_extd[6];
3733 
3734 	if ((i = BITX(cp->cp_ecx, 15, 12)) != 0 &&
3735 	    (size = BITX(cp->cp_ecx, 31, 16)) != 0) {
3736 		uint_t cachesz = size * 1024;
3737 		assoc = amd_afd[i];
3738 
3739 		ASSERT(assoc != -1);
3740 
3741 		if ((ip = l2i->l2i_csz) != NULL)
3742 			*ip = cachesz;
3743 		if ((ip = l2i->l2i_lsz) != NULL)
3744 			*ip = BITX(cp->cp_ecx, 7, 0);
3745 		if ((ip = l2i->l2i_assoc) != NULL)
3746 			*ip = assoc;
3747 		l2i->l2i_ret = cachesz;
3748 	}
3749 }
3750 
3751 int
3752 getl2cacheinfo(cpu_t *cpu, int *csz, int *lsz, int *assoc)
3753 {
3754 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
3755 	struct l2info __l2info, *l2i = &__l2info;
3756 
3757 	l2i->l2i_csz = csz;
3758 	l2i->l2i_lsz = lsz;
3759 	l2i->l2i_assoc = assoc;
3760 	l2i->l2i_ret = -1;
3761 
3762 	switch (x86_which_cacheinfo(cpi)) {
3763 	case X86_VENDOR_Intel:
3764 		intel_walk_cacheinfo(cpi, l2i, intel_l2cinfo);
3765 		break;
3766 	case X86_VENDOR_Cyrix:
3767 		cyrix_walk_cacheinfo(cpi, l2i, intel_l2cinfo);
3768 		break;
3769 	case X86_VENDOR_AMD:
3770 		amd_l2cacheinfo(cpi, l2i);
3771 		break;
3772 	default:
3773 		break;
3774 	}
3775 	return (l2i->l2i_ret);
3776 }
3777 
3778 #if !defined(__xpv)
3779 
3780 uint32_t *
3781 cpuid_mwait_alloc(cpu_t *cpu)
3782 {
3783 	uint32_t	*ret;
3784 	size_t		mwait_size;
3785 
3786 	ASSERT(cpuid_checkpass(cpu, 2));
3787 
3788 	mwait_size = cpu->cpu_m.mcpu_cpi->cpi_mwait.mon_max;
3789 	if (mwait_size == 0)
3790 		return (NULL);
3791 
3792 	/*
3793 	 * kmem_alloc() returns cache line size aligned data for mwait_size
3794 	 * allocations.  mwait_size is currently cache line sized.  Neither
3795 	 * of these implementation details are guarantied to be true in the
3796 	 * future.
3797 	 *
3798 	 * First try allocating mwait_size as kmem_alloc() currently returns
3799 	 * correctly aligned memory.  If kmem_alloc() does not return
3800 	 * mwait_size aligned memory, then use mwait_size ROUNDUP.
3801 	 *
3802 	 * Set cpi_mwait.buf_actual and cpi_mwait.size_actual in case we
3803 	 * decide to free this memory.
3804 	 */
3805 	ret = kmem_zalloc(mwait_size, KM_SLEEP);
3806 	if (ret == (uint32_t *)P2ROUNDUP((uintptr_t)ret, mwait_size)) {
3807 		cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual = ret;
3808 		cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual = mwait_size;
3809 		*ret = MWAIT_RUNNING;
3810 		return (ret);
3811 	} else {
3812 		kmem_free(ret, mwait_size);
3813 		ret = kmem_zalloc(mwait_size * 2, KM_SLEEP);
3814 		cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual = ret;
3815 		cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual = mwait_size * 2;
3816 		ret = (uint32_t *)P2ROUNDUP((uintptr_t)ret, mwait_size);
3817 		*ret = MWAIT_RUNNING;
3818 		return (ret);
3819 	}
3820 }
3821 
3822 void
3823 cpuid_mwait_free(cpu_t *cpu)
3824 {
3825 	ASSERT(cpuid_checkpass(cpu, 2));
3826 
3827 	if (cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual != NULL &&
3828 	    cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual > 0) {
3829 		kmem_free(cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual,
3830 		    cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual);
3831 	}
3832 
3833 	cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual = NULL;
3834 	cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual = 0;
3835 }
3836 
3837 void
3838 patch_tsc_read(int flag)
3839 {
3840 	size_t cnt;
3841 
3842 	switch (flag) {
3843 	case X86_NO_TSC:
3844 		cnt = &_no_rdtsc_end - &_no_rdtsc_start;
3845 		(void) memcpy((void *)tsc_read, (void *)&_no_rdtsc_start, cnt);
3846 		break;
3847 	case X86_HAVE_TSCP:
3848 		cnt = &_tscp_end - &_tscp_start;
3849 		(void) memcpy((void *)tsc_read, (void *)&_tscp_start, cnt);
3850 		break;
3851 	case X86_TSC_MFENCE:
3852 		cnt = &_tsc_mfence_end - &_tsc_mfence_start;
3853 		(void) memcpy((void *)tsc_read,
3854 		    (void *)&_tsc_mfence_start, cnt);
3855 		break;
3856 	case X86_TSC_LFENCE:
3857 		cnt = &_tsc_lfence_end - &_tsc_lfence_start;
3858 		(void) memcpy((void *)tsc_read,
3859 		    (void *)&_tsc_lfence_start, cnt);
3860 		break;
3861 	default:
3862 		break;
3863 	}
3864 }
3865 
3866 int
3867 cpuid_deep_cstates_supported(void)
3868 {
3869 	struct cpuid_info *cpi;
3870 	struct cpuid_regs regs;
3871 
3872 	ASSERT(cpuid_checkpass(CPU, 1));
3873 
3874 	cpi = CPU->cpu_m.mcpu_cpi;
3875 
3876 	if (!(x86_feature & X86_CPUID))
3877 		return (0);
3878 
3879 	switch (cpi->cpi_vendor) {
3880 	case X86_VENDOR_Intel:
3881 		if (cpi->cpi_xmaxeax < 0x80000007)
3882 			return (0);
3883 
3884 		/*
3885 		 * TSC run at a constant rate in all ACPI C-states?
3886 		 */
3887 		regs.cp_eax = 0x80000007;
3888 		(void) __cpuid_insn(&regs);
3889 		return (regs.cp_edx & CPUID_TSC_CSTATE_INVARIANCE);
3890 
3891 	default:
3892 		return (0);
3893 	}
3894 }
3895 
3896 #if defined(__amd64) && !defined(__xpv)
3897 /*
3898  * Patch in versions of bcopy for high performance Intel Nhm processors
3899  * and later...
3900  */
3901 void
3902 patch_memops(uint_t vendor)
3903 {
3904 	size_t cnt, i;
3905 	caddr_t to, from;
3906 
3907 	if ((vendor == X86_VENDOR_Intel) && ((x86_feature & X86_SSE4_2) != 0)) {
3908 		cnt = &bcopy_patch_end - &bcopy_patch_start;
3909 		to = &bcopy_ck_size;
3910 		from = &bcopy_patch_start;
3911 		for (i = 0; i < cnt; i++) {
3912 			*to++ = *from++;
3913 		}
3914 	}
3915 }
3916 #endif  /* __amd64 && !__xpv */
3917 
3918 #endif	/* !__xpv */
3919