xref: /titanic_41/usr/src/uts/i86pc/os/cpuid.c (revision 49d3bc91e27cd871b950d56c01398fa2f2e12ab4)
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 (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
23  */
24 /*
25  * Copyright (c) 2010, Intel Corporation.
26  * All rights reserved.
27  */
28 /*
29  * Portions Copyright 2009 Advanced Micro Devices, Inc.
30  */
31 
32 /*
33  * Various routines to handle identification
34  * and classification of x86 processors.
35  */
36 
37 #include <sys/types.h>
38 #include <sys/archsystm.h>
39 #include <sys/x86_archext.h>
40 #include <sys/kmem.h>
41 #include <sys/systm.h>
42 #include <sys/cmn_err.h>
43 #include <sys/sunddi.h>
44 #include <sys/sunndi.h>
45 #include <sys/cpuvar.h>
46 #include <sys/processor.h>
47 #include <sys/sysmacros.h>
48 #include <sys/pg.h>
49 #include <sys/fp.h>
50 #include <sys/controlregs.h>
51 #include <sys/bitmap.h>
52 #include <sys/auxv_386.h>
53 #include <sys/memnode.h>
54 #include <sys/pci_cfgspace.h>
55 
56 #ifdef __xpv
57 #include <sys/hypervisor.h>
58 #else
59 #include <sys/ontrap.h>
60 #endif
61 
62 /*
63  * Pass 0 of cpuid feature analysis happens in locore. It contains special code
64  * to recognize Cyrix processors that are not cpuid-compliant, and to deal with
65  * them accordingly. For most modern processors, feature detection occurs here
66  * in pass 1.
67  *
68  * Pass 1 of cpuid feature analysis happens just at the beginning of mlsetup()
69  * for the boot CPU and does the basic analysis that the early kernel needs.
70  * x86_featureset is set based on the return value of cpuid_pass1() of the boot
71  * CPU.
72  *
73  * Pass 1 includes:
74  *
75  *	o Determining vendor/model/family/stepping and setting x86_type and
76  *	  x86_vendor accordingly.
77  *	o Processing the feature flags returned by the cpuid instruction while
78  *	  applying any workarounds or tricks for the specific processor.
79  *	o Mapping the feature flags into Solaris feature bits (X86_*).
80  *	o Processing extended feature flags if supported by the processor,
81  *	  again while applying specific processor knowledge.
82  *	o Determining the CMT characteristics of the system.
83  *
84  * Pass 1 is done on non-boot CPUs during their initialization and the results
85  * are used only as a meager attempt at ensuring that all processors within the
86  * system support the same features.
87  *
88  * Pass 2 of cpuid feature analysis happens just at the beginning
89  * of startup().  It just copies in and corrects the remainder
90  * of the cpuid data we depend on: standard cpuid functions that we didn't
91  * need for pass1 feature analysis, and extended cpuid functions beyond the
92  * simple feature processing done in pass1.
93  *
94  * Pass 3 of cpuid analysis is invoked after basic kernel services; in
95  * particular kernel memory allocation has been made available. It creates a
96  * readable brand string based on the data collected in the first two passes.
97  *
98  * Pass 4 of cpuid analysis is invoked after post_startup() when all
99  * the support infrastructure for various hardware features has been
100  * initialized. It determines which processor features will be reported
101  * to userland via the aux vector.
102  *
103  * All passes are executed on all CPUs, but only the boot CPU determines what
104  * features the kernel will use.
105  *
106  * Much of the worst junk in this file is for the support of processors
107  * that didn't really implement the cpuid instruction properly.
108  *
109  * NOTE: The accessor functions (cpuid_get*) are aware of, and ASSERT upon,
110  * the pass numbers.  Accordingly, changes to the pass code may require changes
111  * to the accessor code.
112  */
113 
114 uint_t x86_vendor = X86_VENDOR_IntelClone;
115 uint_t x86_type = X86_TYPE_OTHER;
116 uint_t x86_clflush_size = 0;
117 
118 uint_t pentiumpro_bug4046376;
119 uint_t pentiumpro_bug4064495;
120 
121 uchar_t x86_featureset[BT_SIZEOFMAP(NUM_X86_FEATURES)];
122 
123 static char *x86_feature_names[NUM_X86_FEATURES] = {
124 	"lgpg",
125 	"tsc",
126 	"msr",
127 	"mtrr",
128 	"pge",
129 	"de",
130 	"cmov",
131 	"mmx",
132 	"mca",
133 	"pae",
134 	"cv8",
135 	"pat",
136 	"sep",
137 	"sse",
138 	"sse2",
139 	"htt",
140 	"asysc",
141 	"nx",
142 	"sse3",
143 	"cx16",
144 	"cmp",
145 	"tscp",
146 	"mwait",
147 	"sse4a",
148 	"cpuid",
149 	"ssse3",
150 	"sse4_1",
151 	"sse4_2",
152 	"1gpg",
153 	"clfsh",
154 	"64",
155 	"aes",
156 	"pclmulqdq",
157 	"xsave",
158 	"avx" };
159 
160 boolean_t
161 is_x86_feature(void *featureset, uint_t feature)
162 {
163 	ASSERT(feature < NUM_X86_FEATURES);
164 	return (BT_TEST((ulong_t *)featureset, feature));
165 }
166 
167 void
168 add_x86_feature(void *featureset, uint_t feature)
169 {
170 	ASSERT(feature < NUM_X86_FEATURES);
171 	BT_SET((ulong_t *)featureset, feature);
172 }
173 
174 void
175 remove_x86_feature(void *featureset, uint_t feature)
176 {
177 	ASSERT(feature < NUM_X86_FEATURES);
178 	BT_CLEAR((ulong_t *)featureset, feature);
179 }
180 
181 boolean_t
182 compare_x86_featureset(void *setA, void *setB)
183 {
184 	/*
185 	 * We assume that the unused bits of the bitmap are always zero.
186 	 */
187 	if (memcmp(setA, setB, BT_SIZEOFMAP(NUM_X86_FEATURES)) == 0) {
188 		return (B_TRUE);
189 	} else {
190 		return (B_FALSE);
191 	}
192 }
193 
194 void
195 print_x86_featureset(void *featureset)
196 {
197 	uint_t i;
198 
199 	for (i = 0; i < NUM_X86_FEATURES; i++) {
200 		if (is_x86_feature(featureset, i)) {
201 			cmn_err(CE_CONT, "?x86_feature: %s\n",
202 			    x86_feature_names[i]);
203 		}
204 	}
205 }
206 
207 uint_t enable486;
208 
209 static size_t xsave_state_size = 0;
210 uint64_t xsave_bv_all = (XFEATURE_LEGACY_FP | XFEATURE_SSE);
211 boolean_t xsave_force_disable = B_FALSE;
212 
213 /*
214  * This is set to platform type Solaris is running on.
215  */
216 static int platform_type = -1;
217 
218 #if !defined(__xpv)
219 /*
220  * Variable to patch if hypervisor platform detection needs to be
221  * disabled (e.g. platform_type will always be HW_NATIVE if this is 0).
222  */
223 int enable_platform_detection = 1;
224 #endif
225 
226 /*
227  * monitor/mwait info.
228  *
229  * size_actual and buf_actual are the real address and size allocated to get
230  * proper mwait_buf alignement.  buf_actual and size_actual should be passed
231  * to kmem_free().  Currently kmem_alloc() and mwait happen to both use
232  * processor cache-line alignment, but this is not guarantied in the furture.
233  */
234 struct mwait_info {
235 	size_t		mon_min;	/* min size to avoid missed wakeups */
236 	size_t		mon_max;	/* size to avoid false wakeups */
237 	size_t		size_actual;	/* size actually allocated */
238 	void		*buf_actual;	/* memory actually allocated */
239 	uint32_t	support;	/* processor support of monitor/mwait */
240 };
241 
242 /*
243  * xsave/xrestor info.
244  *
245  * This structure contains HW feature bits and size of the xsave save area.
246  * Note: the kernel will use the maximum size required for all hardware
247  * features. It is not optimize for potential memory savings if features at
248  * the end of the save area are not enabled.
249  */
250 struct xsave_info {
251 	uint32_t	xsav_hw_features_low;   /* Supported HW features */
252 	uint32_t	xsav_hw_features_high;  /* Supported HW features */
253 	size_t		xsav_max_size;  /* max size save area for HW features */
254 	size_t		ymm_size;	/* AVX: size of ymm save area */
255 	size_t		ymm_offset;	/* AVX: offset for ymm save area */
256 };
257 
258 
259 /*
260  * These constants determine how many of the elements of the
261  * cpuid we cache in the cpuid_info data structure; the
262  * remaining elements are accessible via the cpuid instruction.
263  */
264 
265 #define	NMAX_CPI_STD	6		/* eax = 0 .. 5 */
266 #define	NMAX_CPI_EXTD	0x1c		/* eax = 0x80000000 .. 0x8000001b */
267 
268 /*
269  * Some terminology needs to be explained:
270  *  - Socket: Something that can be plugged into a motherboard.
271  *  - Package: Same as socket
272  *  - Chip: Same as socket. Note that AMD's documentation uses term "chip"
273  *    differently: there, chip is the same as processor node (below)
274  *  - Processor node: Some AMD processors have more than one
275  *    "subprocessor" embedded in a package. These subprocessors (nodes)
276  *    are fully-functional processors themselves with cores, caches,
277  *    memory controllers, PCI configuration spaces. They are connected
278  *    inside the package with Hypertransport links. On single-node
279  *    processors, processor node is equivalent to chip/socket/package.
280  */
281 
282 struct cpuid_info {
283 	uint_t cpi_pass;		/* last pass completed */
284 	/*
285 	 * standard function information
286 	 */
287 	uint_t cpi_maxeax;		/* fn 0: %eax */
288 	char cpi_vendorstr[13];		/* fn 0: %ebx:%ecx:%edx */
289 	uint_t cpi_vendor;		/* enum of cpi_vendorstr */
290 
291 	uint_t cpi_family;		/* fn 1: extended family */
292 	uint_t cpi_model;		/* fn 1: extended model */
293 	uint_t cpi_step;		/* fn 1: stepping */
294 	chipid_t cpi_chipid;		/* fn 1: %ebx:  Intel: chip # */
295 					/*		AMD: package/socket # */
296 	uint_t cpi_brandid;		/* fn 1: %ebx: brand ID */
297 	int cpi_clogid;			/* fn 1: %ebx: thread # */
298 	uint_t cpi_ncpu_per_chip;	/* fn 1: %ebx: logical cpu count */
299 	uint8_t cpi_cacheinfo[16];	/* fn 2: intel-style cache desc */
300 	uint_t cpi_ncache;		/* fn 2: number of elements */
301 	uint_t cpi_ncpu_shr_last_cache;	/* fn 4: %eax: ncpus sharing cache */
302 	id_t cpi_last_lvl_cacheid;	/* fn 4: %eax: derived cache id */
303 	uint_t cpi_std_4_size;		/* fn 4: number of fn 4 elements */
304 	struct cpuid_regs **cpi_std_4;	/* fn 4: %ecx == 0 .. fn4_size */
305 	struct cpuid_regs cpi_std[NMAX_CPI_STD];	/* 0 .. 5 */
306 	/*
307 	 * extended function information
308 	 */
309 	uint_t cpi_xmaxeax;		/* fn 0x80000000: %eax */
310 	char cpi_brandstr[49];		/* fn 0x8000000[234] */
311 	uint8_t cpi_pabits;		/* fn 0x80000006: %eax */
312 	uint8_t	cpi_vabits;		/* fn 0x80000006: %eax */
313 	struct	cpuid_regs cpi_extd[NMAX_CPI_EXTD];	/* 0x800000XX */
314 
315 	id_t cpi_coreid;		/* same coreid => strands share core */
316 	int cpi_pkgcoreid;		/* core number within single package */
317 	uint_t cpi_ncore_per_chip;	/* AMD: fn 0x80000008: %ecx[7-0] */
318 					/* Intel: fn 4: %eax[31-26] */
319 	/*
320 	 * supported feature information
321 	 */
322 	uint32_t cpi_support[5];
323 #define	STD_EDX_FEATURES	0
324 #define	AMD_EDX_FEATURES	1
325 #define	TM_EDX_FEATURES		2
326 #define	STD_ECX_FEATURES	3
327 #define	AMD_ECX_FEATURES	4
328 	/*
329 	 * Synthesized information, where known.
330 	 */
331 	uint32_t cpi_chiprev;		/* See X86_CHIPREV_* in x86_archext.h */
332 	const char *cpi_chiprevstr;	/* May be NULL if chiprev unknown */
333 	uint32_t cpi_socket;		/* Chip package/socket type */
334 
335 	struct mwait_info cpi_mwait;	/* fn 5: monitor/mwait info */
336 	uint32_t cpi_apicid;
337 	uint_t cpi_procnodeid;		/* AMD: nodeID on HT, Intel: chipid */
338 	uint_t cpi_procnodes_per_pkg;	/* AMD: # of nodes in the package */
339 					/* Intel: 1 */
340 
341 	struct xsave_info cpi_xsave;	/* fn D: xsave/xrestor info */
342 };
343 
344 
345 static struct cpuid_info cpuid_info0;
346 
347 /*
348  * These bit fields are defined by the Intel Application Note AP-485
349  * "Intel Processor Identification and the CPUID Instruction"
350  */
351 #define	CPI_FAMILY_XTD(cpi)	BITX((cpi)->cpi_std[1].cp_eax, 27, 20)
352 #define	CPI_MODEL_XTD(cpi)	BITX((cpi)->cpi_std[1].cp_eax, 19, 16)
353 #define	CPI_TYPE(cpi)		BITX((cpi)->cpi_std[1].cp_eax, 13, 12)
354 #define	CPI_FAMILY(cpi)		BITX((cpi)->cpi_std[1].cp_eax, 11, 8)
355 #define	CPI_STEP(cpi)		BITX((cpi)->cpi_std[1].cp_eax, 3, 0)
356 #define	CPI_MODEL(cpi)		BITX((cpi)->cpi_std[1].cp_eax, 7, 4)
357 
358 #define	CPI_FEATURES_EDX(cpi)		((cpi)->cpi_std[1].cp_edx)
359 #define	CPI_FEATURES_ECX(cpi)		((cpi)->cpi_std[1].cp_ecx)
360 #define	CPI_FEATURES_XTD_EDX(cpi)	((cpi)->cpi_extd[1].cp_edx)
361 #define	CPI_FEATURES_XTD_ECX(cpi)	((cpi)->cpi_extd[1].cp_ecx)
362 
363 #define	CPI_BRANDID(cpi)	BITX((cpi)->cpi_std[1].cp_ebx, 7, 0)
364 #define	CPI_CHUNKS(cpi)		BITX((cpi)->cpi_std[1].cp_ebx, 15, 7)
365 #define	CPI_CPU_COUNT(cpi)	BITX((cpi)->cpi_std[1].cp_ebx, 23, 16)
366 #define	CPI_APIC_ID(cpi)	BITX((cpi)->cpi_std[1].cp_ebx, 31, 24)
367 
368 #define	CPI_MAXEAX_MAX		0x100		/* sanity control */
369 #define	CPI_XMAXEAX_MAX		0x80000100
370 #define	CPI_FN4_ECX_MAX		0x20		/* sanity: max fn 4 levels */
371 #define	CPI_FNB_ECX_MAX		0x20		/* sanity: max fn B levels */
372 
373 /*
374  * Function 4 (Deterministic Cache Parameters) macros
375  * Defined by Intel Application Note AP-485
376  */
377 #define	CPI_NUM_CORES(regs)		BITX((regs)->cp_eax, 31, 26)
378 #define	CPI_NTHR_SHR_CACHE(regs)	BITX((regs)->cp_eax, 25, 14)
379 #define	CPI_FULL_ASSOC_CACHE(regs)	BITX((regs)->cp_eax, 9, 9)
380 #define	CPI_SELF_INIT_CACHE(regs)	BITX((regs)->cp_eax, 8, 8)
381 #define	CPI_CACHE_LVL(regs)		BITX((regs)->cp_eax, 7, 5)
382 #define	CPI_CACHE_TYPE(regs)		BITX((regs)->cp_eax, 4, 0)
383 #define	CPI_CPU_LEVEL_TYPE(regs)	BITX((regs)->cp_ecx, 15, 8)
384 
385 #define	CPI_CACHE_WAYS(regs)		BITX((regs)->cp_ebx, 31, 22)
386 #define	CPI_CACHE_PARTS(regs)		BITX((regs)->cp_ebx, 21, 12)
387 #define	CPI_CACHE_COH_LN_SZ(regs)	BITX((regs)->cp_ebx, 11, 0)
388 
389 #define	CPI_CACHE_SETS(regs)		BITX((regs)->cp_ecx, 31, 0)
390 
391 #define	CPI_PREFCH_STRIDE(regs)		BITX((regs)->cp_edx, 9, 0)
392 
393 
394 /*
395  * A couple of shorthand macros to identify "later" P6-family chips
396  * like the Pentium M and Core.  First, the "older" P6-based stuff
397  * (loosely defined as "pre-Pentium-4"):
398  * P6, PII, Mobile PII, PII Xeon, PIII, Mobile PIII, PIII Xeon
399  */
400 
401 #define	IS_LEGACY_P6(cpi) (			\
402 	cpi->cpi_family == 6 && 		\
403 		(cpi->cpi_model == 1 ||		\
404 		cpi->cpi_model == 3 ||		\
405 		cpi->cpi_model == 5 ||		\
406 		cpi->cpi_model == 6 ||		\
407 		cpi->cpi_model == 7 ||		\
408 		cpi->cpi_model == 8 ||		\
409 		cpi->cpi_model == 0xA ||	\
410 		cpi->cpi_model == 0xB)		\
411 )
412 
413 /* A "new F6" is everything with family 6 that's not the above */
414 #define	IS_NEW_F6(cpi) ((cpi->cpi_family == 6) && !IS_LEGACY_P6(cpi))
415 
416 /* Extended family/model support */
417 #define	IS_EXTENDED_MODEL_INTEL(cpi) (cpi->cpi_family == 0x6 || \
418 	cpi->cpi_family >= 0xf)
419 
420 /*
421  * Info for monitor/mwait idle loop.
422  *
423  * See cpuid section of "Intel 64 and IA-32 Architectures Software Developer's
424  * Manual Volume 2A: Instruction Set Reference, A-M" #25366-022US, November
425  * 2006.
426  * See MONITOR/MWAIT section of "AMD64 Architecture Programmer's Manual
427  * Documentation Updates" #33633, Rev 2.05, December 2006.
428  */
429 #define	MWAIT_SUPPORT		(0x00000001)	/* mwait supported */
430 #define	MWAIT_EXTENSIONS	(0x00000002)	/* extenstion supported */
431 #define	MWAIT_ECX_INT_ENABLE	(0x00000004)	/* ecx 1 extension supported */
432 #define	MWAIT_SUPPORTED(cpi)	((cpi)->cpi_std[1].cp_ecx & CPUID_INTC_ECX_MON)
433 #define	MWAIT_INT_ENABLE(cpi)	((cpi)->cpi_std[5].cp_ecx & 0x2)
434 #define	MWAIT_EXTENSION(cpi)	((cpi)->cpi_std[5].cp_ecx & 0x1)
435 #define	MWAIT_SIZE_MIN(cpi)	BITX((cpi)->cpi_std[5].cp_eax, 15, 0)
436 #define	MWAIT_SIZE_MAX(cpi)	BITX((cpi)->cpi_std[5].cp_ebx, 15, 0)
437 /*
438  * Number of sub-cstates for a given c-state.
439  */
440 #define	MWAIT_NUM_SUBC_STATES(cpi, c_state)			\
441 	BITX((cpi)->cpi_std[5].cp_edx, c_state + 3, c_state)
442 
443 /*
444  * XSAVE leaf 0xD enumeration
445  */
446 #define	CPUID_LEAFD_2_YMM_OFFSET	576
447 #define	CPUID_LEAFD_2_YMM_SIZE		256
448 
449 /*
450  * Functions we consune from cpuid_subr.c;  don't publish these in a header
451  * file to try and keep people using the expected cpuid_* interfaces.
452  */
453 extern uint32_t _cpuid_skt(uint_t, uint_t, uint_t, uint_t);
454 extern const char *_cpuid_sktstr(uint_t, uint_t, uint_t, uint_t);
455 extern uint32_t _cpuid_chiprev(uint_t, uint_t, uint_t, uint_t);
456 extern const char *_cpuid_chiprevstr(uint_t, uint_t, uint_t, uint_t);
457 extern uint_t _cpuid_vendorstr_to_vendorcode(char *);
458 
459 /*
460  * Apply up various platform-dependent restrictions where the
461  * underlying platform restrictions mean the CPU can be marked
462  * as less capable than its cpuid instruction would imply.
463  */
464 #if defined(__xpv)
465 static void
466 platform_cpuid_mangle(uint_t vendor, uint32_t eax, struct cpuid_regs *cp)
467 {
468 	switch (eax) {
469 	case 1: {
470 		uint32_t mcamask = DOMAIN_IS_INITDOMAIN(xen_info) ?
471 		    0 : CPUID_INTC_EDX_MCA;
472 		cp->cp_edx &=
473 		    ~(mcamask |
474 		    CPUID_INTC_EDX_PSE |
475 		    CPUID_INTC_EDX_VME | CPUID_INTC_EDX_DE |
476 		    CPUID_INTC_EDX_SEP | CPUID_INTC_EDX_MTRR |
477 		    CPUID_INTC_EDX_PGE | CPUID_INTC_EDX_PAT |
478 		    CPUID_AMD_EDX_SYSC | CPUID_INTC_EDX_SEP |
479 		    CPUID_INTC_EDX_PSE36 | CPUID_INTC_EDX_HTT);
480 		break;
481 	}
482 
483 	case 0x80000001:
484 		cp->cp_edx &=
485 		    ~(CPUID_AMD_EDX_PSE |
486 		    CPUID_INTC_EDX_VME | CPUID_INTC_EDX_DE |
487 		    CPUID_AMD_EDX_MTRR | CPUID_AMD_EDX_PGE |
488 		    CPUID_AMD_EDX_PAT | CPUID_AMD_EDX_PSE36 |
489 		    CPUID_AMD_EDX_SYSC | CPUID_INTC_EDX_SEP |
490 		    CPUID_AMD_EDX_TSCP);
491 		cp->cp_ecx &= ~CPUID_AMD_ECX_CMP_LGCY;
492 		break;
493 	default:
494 		break;
495 	}
496 
497 	switch (vendor) {
498 	case X86_VENDOR_Intel:
499 		switch (eax) {
500 		case 4:
501 			/*
502 			 * Zero out the (ncores-per-chip - 1) field
503 			 */
504 			cp->cp_eax &= 0x03fffffff;
505 			break;
506 		default:
507 			break;
508 		}
509 		break;
510 	case X86_VENDOR_AMD:
511 		switch (eax) {
512 
513 		case 0x80000001:
514 			cp->cp_ecx &= ~CPUID_AMD_ECX_CR8D;
515 			break;
516 
517 		case 0x80000008:
518 			/*
519 			 * Zero out the (ncores-per-chip - 1) field
520 			 */
521 			cp->cp_ecx &= 0xffffff00;
522 			break;
523 		default:
524 			break;
525 		}
526 		break;
527 	default:
528 		break;
529 	}
530 }
531 #else
532 #define	platform_cpuid_mangle(vendor, eax, cp)	/* nothing */
533 #endif
534 
535 /*
536  *  Some undocumented ways of patching the results of the cpuid
537  *  instruction to permit running Solaris 10 on future cpus that
538  *  we don't currently support.  Could be set to non-zero values
539  *  via settings in eeprom.
540  */
541 
542 uint32_t cpuid_feature_ecx_include;
543 uint32_t cpuid_feature_ecx_exclude;
544 uint32_t cpuid_feature_edx_include;
545 uint32_t cpuid_feature_edx_exclude;
546 
547 /*
548  * Allocate space for mcpu_cpi in the machcpu structure for all non-boot CPUs.
549  */
550 void
551 cpuid_alloc_space(cpu_t *cpu)
552 {
553 	/*
554 	 * By convention, cpu0 is the boot cpu, which is set up
555 	 * before memory allocation is available.  All other cpus get
556 	 * their cpuid_info struct allocated here.
557 	 */
558 	ASSERT(cpu->cpu_id != 0);
559 	ASSERT(cpu->cpu_m.mcpu_cpi == NULL);
560 	cpu->cpu_m.mcpu_cpi =
561 	    kmem_zalloc(sizeof (*cpu->cpu_m.mcpu_cpi), KM_SLEEP);
562 }
563 
564 void
565 cpuid_free_space(cpu_t *cpu)
566 {
567 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
568 	int i;
569 
570 	ASSERT(cpi != NULL);
571 	ASSERT(cpi != &cpuid_info0);
572 
573 	/*
574 	 * Free up any function 4 related dynamic storage
575 	 */
576 	for (i = 1; i < cpi->cpi_std_4_size; i++)
577 		kmem_free(cpi->cpi_std_4[i], sizeof (struct cpuid_regs));
578 	if (cpi->cpi_std_4_size > 0)
579 		kmem_free(cpi->cpi_std_4,
580 		    cpi->cpi_std_4_size * sizeof (struct cpuid_regs *));
581 
582 	kmem_free(cpi, sizeof (*cpi));
583 	cpu->cpu_m.mcpu_cpi = NULL;
584 }
585 
586 #if !defined(__xpv)
587 
588 static void
589 determine_platform()
590 {
591 	struct cpuid_regs cp;
592 	char *xen_str;
593 	uint32_t xen_signature[4], base;
594 
595 	platform_type = HW_NATIVE;
596 
597 	if (!enable_platform_detection)
598 		return;
599 
600 	/*
601 	 * In a fully virtualized domain, Xen's pseudo-cpuid function
602 	 * returns a string representing the Xen signature in %ebx, %ecx,
603 	 * and %edx. %eax contains the maximum supported cpuid function.
604 	 * We need at least a (base + 2) leaf value to do what we want
605 	 * to do. Try different base values, since the hypervisor might
606 	 * use a different one depending on whether hyper-v emulation
607 	 * is switched on by default or not.
608 	 */
609 	for (base = 0x40000000; base < 0x40010000; base += 0x100) {
610 		cp.cp_eax = base;
611 		(void) __cpuid_insn(&cp);
612 		xen_signature[0] = cp.cp_ebx;
613 		xen_signature[1] = cp.cp_ecx;
614 		xen_signature[2] = cp.cp_edx;
615 		xen_signature[3] = 0;
616 		xen_str = (char *)xen_signature;
617 		if (strcmp("XenVMMXenVMM", xen_str) == 0 &&
618 		    cp.cp_eax >= (base + 2)) {
619 			platform_type = HW_XEN_HVM;
620 			return;
621 		}
622 	}
623 
624 	if (vmware_platform()) /* running under vmware hypervisor? */
625 		platform_type = HW_VMWARE;
626 }
627 
628 int
629 get_hwenv(void)
630 {
631 	if (platform_type == -1)
632 		determine_platform();
633 
634 	return (platform_type);
635 }
636 
637 int
638 is_controldom(void)
639 {
640 	return (0);
641 }
642 
643 #else
644 
645 int
646 get_hwenv(void)
647 {
648 	return (HW_XEN_PV);
649 }
650 
651 int
652 is_controldom(void)
653 {
654 	return (DOMAIN_IS_INITDOMAIN(xen_info));
655 }
656 
657 #endif	/* __xpv */
658 
659 static void
660 cpuid_intel_getids(cpu_t *cpu, void *feature)
661 {
662 	uint_t i;
663 	uint_t chipid_shift = 0;
664 	uint_t coreid_shift = 0;
665 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
666 
667 	for (i = 1; i < cpi->cpi_ncpu_per_chip; i <<= 1)
668 		chipid_shift++;
669 
670 	cpi->cpi_chipid = cpi->cpi_apicid >> chipid_shift;
671 	cpi->cpi_clogid = cpi->cpi_apicid & ((1 << chipid_shift) - 1);
672 
673 	if (is_x86_feature(feature, X86FSET_CMP)) {
674 		/*
675 		 * Multi-core (and possibly multi-threaded)
676 		 * processors.
677 		 */
678 		uint_t ncpu_per_core;
679 		if (cpi->cpi_ncore_per_chip == 1)
680 			ncpu_per_core = cpi->cpi_ncpu_per_chip;
681 		else if (cpi->cpi_ncore_per_chip > 1)
682 			ncpu_per_core = cpi->cpi_ncpu_per_chip /
683 			    cpi->cpi_ncore_per_chip;
684 		/*
685 		 * 8bit APIC IDs on dual core Pentiums
686 		 * look like this:
687 		 *
688 		 * +-----------------------+------+------+
689 		 * | Physical Package ID   |  MC  |  HT  |
690 		 * +-----------------------+------+------+
691 		 * <------- chipid -------->
692 		 * <------- coreid --------------->
693 		 *			   <--- clogid -->
694 		 *			   <------>
695 		 *			   pkgcoreid
696 		 *
697 		 * Where the number of bits necessary to
698 		 * represent MC and HT fields together equals
699 		 * to the minimum number of bits necessary to
700 		 * store the value of cpi->cpi_ncpu_per_chip.
701 		 * Of those bits, the MC part uses the number
702 		 * of bits necessary to store the value of
703 		 * cpi->cpi_ncore_per_chip.
704 		 */
705 		for (i = 1; i < ncpu_per_core; i <<= 1)
706 			coreid_shift++;
707 		cpi->cpi_coreid = cpi->cpi_apicid >> coreid_shift;
708 		cpi->cpi_pkgcoreid = cpi->cpi_clogid >> coreid_shift;
709 	} else if (is_x86_feature(feature, X86FSET_HTT)) {
710 		/*
711 		 * Single-core multi-threaded processors.
712 		 */
713 		cpi->cpi_coreid = cpi->cpi_chipid;
714 		cpi->cpi_pkgcoreid = 0;
715 	}
716 	cpi->cpi_procnodeid = cpi->cpi_chipid;
717 }
718 
719 static void
720 cpuid_amd_getids(cpu_t *cpu)
721 {
722 	int i, first_half, coreidsz;
723 	uint32_t nb_caps_reg;
724 	uint_t node2_1;
725 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
726 
727 	/*
728 	 * AMD CMP chips currently have a single thread per core.
729 	 *
730 	 * Since no two cpus share a core we must assign a distinct coreid
731 	 * per cpu, and we do this by using the cpu_id.  This scheme does not,
732 	 * however, guarantee that sibling cores of a chip will have sequential
733 	 * coreids starting at a multiple of the number of cores per chip -
734 	 * that is usually the case, but if the ACPI MADT table is presented
735 	 * in a different order then we need to perform a few more gymnastics
736 	 * for the pkgcoreid.
737 	 *
738 	 * All processors in the system have the same number of enabled
739 	 * cores. Cores within a processor are always numbered sequentially
740 	 * from 0 regardless of how many or which are disabled, and there
741 	 * is no way for operating system to discover the real core id when some
742 	 * are disabled.
743 	 */
744 
745 	cpi->cpi_coreid = cpu->cpu_id;
746 
747 	if (cpi->cpi_xmaxeax >= 0x80000008) {
748 
749 		coreidsz = BITX((cpi)->cpi_extd[8].cp_ecx, 15, 12);
750 
751 		/*
752 		 * In AMD parlance chip is really a node while Solaris
753 		 * sees chip as equivalent to socket/package.
754 		 */
755 		cpi->cpi_ncore_per_chip =
756 		    BITX((cpi)->cpi_extd[8].cp_ecx, 7, 0) + 1;
757 		if (coreidsz == 0) {
758 			/* Use legacy method */
759 			for (i = 1; i < cpi->cpi_ncore_per_chip; i <<= 1)
760 				coreidsz++;
761 			if (coreidsz == 0)
762 				coreidsz = 1;
763 		}
764 	} else {
765 		/* Assume single-core part */
766 		cpi->cpi_ncore_per_chip = 1;
767 		coreidsz = 1;
768 	}
769 
770 	cpi->cpi_clogid = cpi->cpi_pkgcoreid =
771 	    cpi->cpi_apicid & ((1<<coreidsz) - 1);
772 	cpi->cpi_ncpu_per_chip = cpi->cpi_ncore_per_chip;
773 
774 	/* Get nodeID */
775 	if (cpi->cpi_family == 0xf) {
776 		cpi->cpi_procnodeid = (cpi->cpi_apicid >> coreidsz) & 7;
777 		cpi->cpi_chipid = cpi->cpi_procnodeid;
778 	} else if (cpi->cpi_family == 0x10) {
779 		/*
780 		 * See if we are a multi-node processor.
781 		 * All processors in the system have the same number of nodes
782 		 */
783 		nb_caps_reg =  pci_getl_func(0, 24, 3, 0xe8);
784 		if ((cpi->cpi_model < 8) || BITX(nb_caps_reg, 29, 29) == 0) {
785 			/* Single-node */
786 			cpi->cpi_procnodeid = BITX(cpi->cpi_apicid, 5,
787 			    coreidsz);
788 			cpi->cpi_chipid = cpi->cpi_procnodeid;
789 		} else {
790 
791 			/*
792 			 * Multi-node revision D (2 nodes per package
793 			 * are supported)
794 			 */
795 			cpi->cpi_procnodes_per_pkg = 2;
796 
797 			first_half = (cpi->cpi_pkgcoreid <=
798 			    (cpi->cpi_ncore_per_chip/2 - 1));
799 
800 			if (cpi->cpi_apicid == cpi->cpi_pkgcoreid) {
801 				/* We are BSP */
802 				cpi->cpi_procnodeid = (first_half ? 0 : 1);
803 				cpi->cpi_chipid = cpi->cpi_procnodeid >> 1;
804 			} else {
805 
806 				/* We are AP */
807 				/* NodeId[2:1] bits to use for reading F3xe8 */
808 				node2_1 = BITX(cpi->cpi_apicid, 5, 4) << 1;
809 
810 				nb_caps_reg =
811 				    pci_getl_func(0, 24 + node2_1, 3, 0xe8);
812 
813 				/*
814 				 * Check IntNodeNum bit (31:30, but bit 31 is
815 				 * always 0 on dual-node processors)
816 				 */
817 				if (BITX(nb_caps_reg, 30, 30) == 0)
818 					cpi->cpi_procnodeid = node2_1 +
819 					    !first_half;
820 				else
821 					cpi->cpi_procnodeid = node2_1 +
822 					    first_half;
823 
824 				cpi->cpi_chipid = cpi->cpi_procnodeid >> 1;
825 			}
826 		}
827 	} else if (cpi->cpi_family >= 0x11) {
828 		cpi->cpi_procnodeid = (cpi->cpi_apicid >> coreidsz) & 7;
829 		cpi->cpi_chipid = cpi->cpi_procnodeid;
830 	} else {
831 		cpi->cpi_procnodeid = 0;
832 		cpi->cpi_chipid = cpi->cpi_procnodeid;
833 	}
834 }
835 
836 /*
837  * Setup XFeature_Enabled_Mask register. Required by xsave feature.
838  */
839 void
840 setup_xfem(void)
841 {
842 	uint64_t flags = XFEATURE_LEGACY_FP;
843 
844 	ASSERT(is_x86_feature(x86_featureset, X86FSET_XSAVE));
845 
846 	if (is_x86_feature(x86_featureset, X86FSET_SSE))
847 		flags |= XFEATURE_SSE;
848 
849 	if (is_x86_feature(x86_featureset, X86FSET_AVX))
850 		flags |= XFEATURE_AVX;
851 
852 	set_xcr(XFEATURE_ENABLED_MASK, flags);
853 
854 	xsave_bv_all = flags;
855 }
856 
857 void
858 cpuid_pass1(cpu_t *cpu, uchar_t *featureset)
859 {
860 	uint32_t mask_ecx, mask_edx;
861 	struct cpuid_info *cpi;
862 	struct cpuid_regs *cp;
863 	int xcpuid;
864 #if !defined(__xpv)
865 	extern int idle_cpu_prefer_mwait;
866 #endif
867 
868 #if !defined(__xpv)
869 	determine_platform();
870 #endif
871 	/*
872 	 * Space statically allocated for BSP, ensure pointer is set
873 	 */
874 	if (cpu->cpu_id == 0) {
875 		if (cpu->cpu_m.mcpu_cpi == NULL)
876 			cpu->cpu_m.mcpu_cpi = &cpuid_info0;
877 	}
878 
879 	add_x86_feature(featureset, X86FSET_CPUID);
880 
881 	cpi = cpu->cpu_m.mcpu_cpi;
882 	ASSERT(cpi != NULL);
883 	cp = &cpi->cpi_std[0];
884 	cp->cp_eax = 0;
885 	cpi->cpi_maxeax = __cpuid_insn(cp);
886 	{
887 		uint32_t *iptr = (uint32_t *)cpi->cpi_vendorstr;
888 		*iptr++ = cp->cp_ebx;
889 		*iptr++ = cp->cp_edx;
890 		*iptr++ = cp->cp_ecx;
891 		*(char *)&cpi->cpi_vendorstr[12] = '\0';
892 	}
893 
894 	cpi->cpi_vendor = _cpuid_vendorstr_to_vendorcode(cpi->cpi_vendorstr);
895 	x86_vendor = cpi->cpi_vendor; /* for compatibility */
896 
897 	/*
898 	 * Limit the range in case of weird hardware
899 	 */
900 	if (cpi->cpi_maxeax > CPI_MAXEAX_MAX)
901 		cpi->cpi_maxeax = CPI_MAXEAX_MAX;
902 	if (cpi->cpi_maxeax < 1)
903 		goto pass1_done;
904 
905 	cp = &cpi->cpi_std[1];
906 	cp->cp_eax = 1;
907 	(void) __cpuid_insn(cp);
908 
909 	/*
910 	 * Extract identifying constants for easy access.
911 	 */
912 	cpi->cpi_model = CPI_MODEL(cpi);
913 	cpi->cpi_family = CPI_FAMILY(cpi);
914 
915 	if (cpi->cpi_family == 0xf)
916 		cpi->cpi_family += CPI_FAMILY_XTD(cpi);
917 
918 	/*
919 	 * Beware: AMD uses "extended model" iff base *FAMILY* == 0xf.
920 	 * Intel, and presumably everyone else, uses model == 0xf, as
921 	 * one would expect (max value means possible overflow).  Sigh.
922 	 */
923 
924 	switch (cpi->cpi_vendor) {
925 	case X86_VENDOR_Intel:
926 		if (IS_EXTENDED_MODEL_INTEL(cpi))
927 			cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4;
928 		break;
929 	case X86_VENDOR_AMD:
930 		if (CPI_FAMILY(cpi) == 0xf)
931 			cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4;
932 		break;
933 	default:
934 		if (cpi->cpi_model == 0xf)
935 			cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4;
936 		break;
937 	}
938 
939 	cpi->cpi_step = CPI_STEP(cpi);
940 	cpi->cpi_brandid = CPI_BRANDID(cpi);
941 
942 	/*
943 	 * *default* assumptions:
944 	 * - believe %edx feature word
945 	 * - ignore %ecx feature word
946 	 * - 32-bit virtual and physical addressing
947 	 */
948 	mask_edx = 0xffffffff;
949 	mask_ecx = 0;
950 
951 	cpi->cpi_pabits = cpi->cpi_vabits = 32;
952 
953 	switch (cpi->cpi_vendor) {
954 	case X86_VENDOR_Intel:
955 		if (cpi->cpi_family == 5)
956 			x86_type = X86_TYPE_P5;
957 		else if (IS_LEGACY_P6(cpi)) {
958 			x86_type = X86_TYPE_P6;
959 			pentiumpro_bug4046376 = 1;
960 			pentiumpro_bug4064495 = 1;
961 			/*
962 			 * Clear the SEP bit when it was set erroneously
963 			 */
964 			if (cpi->cpi_model < 3 && cpi->cpi_step < 3)
965 				cp->cp_edx &= ~CPUID_INTC_EDX_SEP;
966 		} else if (IS_NEW_F6(cpi) || cpi->cpi_family == 0xf) {
967 			x86_type = X86_TYPE_P4;
968 			/*
969 			 * We don't currently depend on any of the %ecx
970 			 * features until Prescott, so we'll only check
971 			 * this from P4 onwards.  We might want to revisit
972 			 * that idea later.
973 			 */
974 			mask_ecx = 0xffffffff;
975 		} else if (cpi->cpi_family > 0xf)
976 			mask_ecx = 0xffffffff;
977 		/*
978 		 * We don't support MONITOR/MWAIT if leaf 5 is not available
979 		 * to obtain the monitor linesize.
980 		 */
981 		if (cpi->cpi_maxeax < 5)
982 			mask_ecx &= ~CPUID_INTC_ECX_MON;
983 		break;
984 	case X86_VENDOR_IntelClone:
985 	default:
986 		break;
987 	case X86_VENDOR_AMD:
988 #if defined(OPTERON_ERRATUM_108)
989 		if (cpi->cpi_family == 0xf && cpi->cpi_model == 0xe) {
990 			cp->cp_eax = (0xf0f & cp->cp_eax) | 0xc0;
991 			cpi->cpi_model = 0xc;
992 		} else
993 #endif
994 		if (cpi->cpi_family == 5) {
995 			/*
996 			 * AMD K5 and K6
997 			 *
998 			 * These CPUs have an incomplete implementation
999 			 * of MCA/MCE which we mask away.
1000 			 */
1001 			mask_edx &= ~(CPUID_INTC_EDX_MCE | CPUID_INTC_EDX_MCA);
1002 
1003 			/*
1004 			 * Model 0 uses the wrong (APIC) bit
1005 			 * to indicate PGE.  Fix it here.
1006 			 */
1007 			if (cpi->cpi_model == 0) {
1008 				if (cp->cp_edx & 0x200) {
1009 					cp->cp_edx &= ~0x200;
1010 					cp->cp_edx |= CPUID_INTC_EDX_PGE;
1011 				}
1012 			}
1013 
1014 			/*
1015 			 * Early models had problems w/ MMX; disable.
1016 			 */
1017 			if (cpi->cpi_model < 6)
1018 				mask_edx &= ~CPUID_INTC_EDX_MMX;
1019 		}
1020 
1021 		/*
1022 		 * For newer families, SSE3 and CX16, at least, are valid;
1023 		 * enable all
1024 		 */
1025 		if (cpi->cpi_family >= 0xf)
1026 			mask_ecx = 0xffffffff;
1027 		/*
1028 		 * We don't support MONITOR/MWAIT if leaf 5 is not available
1029 		 * to obtain the monitor linesize.
1030 		 */
1031 		if (cpi->cpi_maxeax < 5)
1032 			mask_ecx &= ~CPUID_INTC_ECX_MON;
1033 
1034 #if !defined(__xpv)
1035 		/*
1036 		 * Do not use MONITOR/MWAIT to halt in the idle loop on any AMD
1037 		 * processors.  AMD does not intend MWAIT to be used in the cpu
1038 		 * idle loop on current and future processors.  10h and future
1039 		 * AMD processors use more power in MWAIT than HLT.
1040 		 * Pre-family-10h Opterons do not have the MWAIT instruction.
1041 		 */
1042 		idle_cpu_prefer_mwait = 0;
1043 #endif
1044 
1045 		break;
1046 	case X86_VENDOR_TM:
1047 		/*
1048 		 * workaround the NT workaround in CMS 4.1
1049 		 */
1050 		if (cpi->cpi_family == 5 && cpi->cpi_model == 4 &&
1051 		    (cpi->cpi_step == 2 || cpi->cpi_step == 3))
1052 			cp->cp_edx |= CPUID_INTC_EDX_CX8;
1053 		break;
1054 	case X86_VENDOR_Centaur:
1055 		/*
1056 		 * workaround the NT workarounds again
1057 		 */
1058 		if (cpi->cpi_family == 6)
1059 			cp->cp_edx |= CPUID_INTC_EDX_CX8;
1060 		break;
1061 	case X86_VENDOR_Cyrix:
1062 		/*
1063 		 * We rely heavily on the probing in locore
1064 		 * to actually figure out what parts, if any,
1065 		 * of the Cyrix cpuid instruction to believe.
1066 		 */
1067 		switch (x86_type) {
1068 		case X86_TYPE_CYRIX_486:
1069 			mask_edx = 0;
1070 			break;
1071 		case X86_TYPE_CYRIX_6x86:
1072 			mask_edx = 0;
1073 			break;
1074 		case X86_TYPE_CYRIX_6x86L:
1075 			mask_edx =
1076 			    CPUID_INTC_EDX_DE |
1077 			    CPUID_INTC_EDX_CX8;
1078 			break;
1079 		case X86_TYPE_CYRIX_6x86MX:
1080 			mask_edx =
1081 			    CPUID_INTC_EDX_DE |
1082 			    CPUID_INTC_EDX_MSR |
1083 			    CPUID_INTC_EDX_CX8 |
1084 			    CPUID_INTC_EDX_PGE |
1085 			    CPUID_INTC_EDX_CMOV |
1086 			    CPUID_INTC_EDX_MMX;
1087 			break;
1088 		case X86_TYPE_CYRIX_GXm:
1089 			mask_edx =
1090 			    CPUID_INTC_EDX_MSR |
1091 			    CPUID_INTC_EDX_CX8 |
1092 			    CPUID_INTC_EDX_CMOV |
1093 			    CPUID_INTC_EDX_MMX;
1094 			break;
1095 		case X86_TYPE_CYRIX_MediaGX:
1096 			break;
1097 		case X86_TYPE_CYRIX_MII:
1098 		case X86_TYPE_VIA_CYRIX_III:
1099 			mask_edx =
1100 			    CPUID_INTC_EDX_DE |
1101 			    CPUID_INTC_EDX_TSC |
1102 			    CPUID_INTC_EDX_MSR |
1103 			    CPUID_INTC_EDX_CX8 |
1104 			    CPUID_INTC_EDX_PGE |
1105 			    CPUID_INTC_EDX_CMOV |
1106 			    CPUID_INTC_EDX_MMX;
1107 			break;
1108 		default:
1109 			break;
1110 		}
1111 		break;
1112 	}
1113 
1114 #if defined(__xpv)
1115 	/*
1116 	 * Do not support MONITOR/MWAIT under a hypervisor
1117 	 */
1118 	mask_ecx &= ~CPUID_INTC_ECX_MON;
1119 	/*
1120 	 * Do not support XSAVE under a hypervisor for now
1121 	 */
1122 	xsave_force_disable = B_TRUE;
1123 
1124 #endif	/* __xpv */
1125 
1126 	if (xsave_force_disable) {
1127 		mask_ecx &= ~CPUID_INTC_ECX_XSAVE;
1128 		mask_ecx &= ~CPUID_INTC_ECX_AVX;
1129 	}
1130 
1131 	/*
1132 	 * Now we've figured out the masks that determine
1133 	 * which bits we choose to believe, apply the masks
1134 	 * to the feature words, then map the kernel's view
1135 	 * of these feature words into its feature word.
1136 	 */
1137 	cp->cp_edx &= mask_edx;
1138 	cp->cp_ecx &= mask_ecx;
1139 
1140 	/*
1141 	 * apply any platform restrictions (we don't call this
1142 	 * immediately after __cpuid_insn here, because we need the
1143 	 * workarounds applied above first)
1144 	 */
1145 	platform_cpuid_mangle(cpi->cpi_vendor, 1, cp);
1146 
1147 	/*
1148 	 * fold in overrides from the "eeprom" mechanism
1149 	 */
1150 	cp->cp_edx |= cpuid_feature_edx_include;
1151 	cp->cp_edx &= ~cpuid_feature_edx_exclude;
1152 
1153 	cp->cp_ecx |= cpuid_feature_ecx_include;
1154 	cp->cp_ecx &= ~cpuid_feature_ecx_exclude;
1155 
1156 	if (cp->cp_edx & CPUID_INTC_EDX_PSE) {
1157 		add_x86_feature(featureset, X86FSET_LARGEPAGE);
1158 	}
1159 	if (cp->cp_edx & CPUID_INTC_EDX_TSC) {
1160 		add_x86_feature(featureset, X86FSET_TSC);
1161 	}
1162 	if (cp->cp_edx & CPUID_INTC_EDX_MSR) {
1163 		add_x86_feature(featureset, X86FSET_MSR);
1164 	}
1165 	if (cp->cp_edx & CPUID_INTC_EDX_MTRR) {
1166 		add_x86_feature(featureset, X86FSET_MTRR);
1167 	}
1168 	if (cp->cp_edx & CPUID_INTC_EDX_PGE) {
1169 		add_x86_feature(featureset, X86FSET_PGE);
1170 	}
1171 	if (cp->cp_edx & CPUID_INTC_EDX_CMOV) {
1172 		add_x86_feature(featureset, X86FSET_CMOV);
1173 	}
1174 	if (cp->cp_edx & CPUID_INTC_EDX_MMX) {
1175 		add_x86_feature(featureset, X86FSET_MMX);
1176 	}
1177 	if ((cp->cp_edx & CPUID_INTC_EDX_MCE) != 0 &&
1178 	    (cp->cp_edx & CPUID_INTC_EDX_MCA) != 0) {
1179 		add_x86_feature(featureset, X86FSET_MCA);
1180 	}
1181 	if (cp->cp_edx & CPUID_INTC_EDX_PAE) {
1182 		add_x86_feature(featureset, X86FSET_PAE);
1183 	}
1184 	if (cp->cp_edx & CPUID_INTC_EDX_CX8) {
1185 		add_x86_feature(featureset, X86FSET_CX8);
1186 	}
1187 	if (cp->cp_ecx & CPUID_INTC_ECX_CX16) {
1188 		add_x86_feature(featureset, X86FSET_CX16);
1189 	}
1190 	if (cp->cp_edx & CPUID_INTC_EDX_PAT) {
1191 		add_x86_feature(featureset, X86FSET_PAT);
1192 	}
1193 	if (cp->cp_edx & CPUID_INTC_EDX_SEP) {
1194 		add_x86_feature(featureset, X86FSET_SEP);
1195 	}
1196 	if (cp->cp_edx & CPUID_INTC_EDX_FXSR) {
1197 		/*
1198 		 * In our implementation, fxsave/fxrstor
1199 		 * are prerequisites before we'll even
1200 		 * try and do SSE things.
1201 		 */
1202 		if (cp->cp_edx & CPUID_INTC_EDX_SSE) {
1203 			add_x86_feature(featureset, X86FSET_SSE);
1204 		}
1205 		if (cp->cp_edx & CPUID_INTC_EDX_SSE2) {
1206 			add_x86_feature(featureset, X86FSET_SSE2);
1207 		}
1208 		if (cp->cp_ecx & CPUID_INTC_ECX_SSE3) {
1209 			add_x86_feature(featureset, X86FSET_SSE3);
1210 		}
1211 		if (cpi->cpi_vendor == X86_VENDOR_Intel) {
1212 			if (cp->cp_ecx & CPUID_INTC_ECX_SSSE3) {
1213 				add_x86_feature(featureset, X86FSET_SSSE3);
1214 			}
1215 			if (cp->cp_ecx & CPUID_INTC_ECX_SSE4_1) {
1216 				add_x86_feature(featureset, X86FSET_SSE4_1);
1217 			}
1218 			if (cp->cp_ecx & CPUID_INTC_ECX_SSE4_2) {
1219 				add_x86_feature(featureset, X86FSET_SSE4_2);
1220 			}
1221 			if (cp->cp_ecx & CPUID_INTC_ECX_AES) {
1222 				add_x86_feature(featureset, X86FSET_AES);
1223 			}
1224 			if (cp->cp_ecx & CPUID_INTC_ECX_PCLMULQDQ) {
1225 				add_x86_feature(featureset, X86FSET_PCLMULQDQ);
1226 			}
1227 
1228 			if (cp->cp_ecx & CPUID_INTC_ECX_XSAVE) {
1229 				add_x86_feature(featureset, X86FSET_XSAVE);
1230 				/* We only test AVX when there is XSAVE */
1231 				if (cp->cp_ecx & CPUID_INTC_ECX_AVX) {
1232 					add_x86_feature(featureset,
1233 					    X86FSET_AVX);
1234 				}
1235 			}
1236 		}
1237 	}
1238 	if (cp->cp_edx & CPUID_INTC_EDX_DE) {
1239 		add_x86_feature(featureset, X86FSET_DE);
1240 	}
1241 #if !defined(__xpv)
1242 	if (cp->cp_ecx & CPUID_INTC_ECX_MON) {
1243 
1244 		/*
1245 		 * We require the CLFLUSH instruction for erratum workaround
1246 		 * to use MONITOR/MWAIT.
1247 		 */
1248 		if (cp->cp_edx & CPUID_INTC_EDX_CLFSH) {
1249 			cpi->cpi_mwait.support |= MWAIT_SUPPORT;
1250 			add_x86_feature(featureset, X86FSET_MWAIT);
1251 		} else {
1252 			extern int idle_cpu_assert_cflush_monitor;
1253 
1254 			/*
1255 			 * All processors we are aware of which have
1256 			 * MONITOR/MWAIT also have CLFLUSH.
1257 			 */
1258 			if (idle_cpu_assert_cflush_monitor) {
1259 				ASSERT((cp->cp_ecx & CPUID_INTC_ECX_MON) &&
1260 				    (cp->cp_edx & CPUID_INTC_EDX_CLFSH));
1261 			}
1262 		}
1263 	}
1264 #endif	/* __xpv */
1265 
1266 	/*
1267 	 * Only need it first time, rest of the cpus would follow suite.
1268 	 * we only capture this for the bootcpu.
1269 	 */
1270 	if (cp->cp_edx & CPUID_INTC_EDX_CLFSH) {
1271 		add_x86_feature(featureset, X86FSET_CLFSH);
1272 		x86_clflush_size = (BITX(cp->cp_ebx, 15, 8) * 8);
1273 	}
1274 	if (is_x86_feature(featureset, X86FSET_PAE))
1275 		cpi->cpi_pabits = 36;
1276 
1277 	/*
1278 	 * Hyperthreading configuration is slightly tricky on Intel
1279 	 * and pure clones, and even trickier on AMD.
1280 	 *
1281 	 * (AMD chose to set the HTT bit on their CMP processors,
1282 	 * even though they're not actually hyperthreaded.  Thus it
1283 	 * takes a bit more work to figure out what's really going
1284 	 * on ... see the handling of the CMP_LGCY bit below)
1285 	 */
1286 	if (cp->cp_edx & CPUID_INTC_EDX_HTT) {
1287 		cpi->cpi_ncpu_per_chip = CPI_CPU_COUNT(cpi);
1288 		if (cpi->cpi_ncpu_per_chip > 1)
1289 			add_x86_feature(featureset, X86FSET_HTT);
1290 	} else {
1291 		cpi->cpi_ncpu_per_chip = 1;
1292 	}
1293 
1294 	/*
1295 	 * Work on the "extended" feature information, doing
1296 	 * some basic initialization for cpuid_pass2()
1297 	 */
1298 	xcpuid = 0;
1299 	switch (cpi->cpi_vendor) {
1300 	case X86_VENDOR_Intel:
1301 		if (IS_NEW_F6(cpi) || cpi->cpi_family >= 0xf)
1302 			xcpuid++;
1303 		break;
1304 	case X86_VENDOR_AMD:
1305 		if (cpi->cpi_family > 5 ||
1306 		    (cpi->cpi_family == 5 && cpi->cpi_model >= 1))
1307 			xcpuid++;
1308 		break;
1309 	case X86_VENDOR_Cyrix:
1310 		/*
1311 		 * Only these Cyrix CPUs are -known- to support
1312 		 * extended cpuid operations.
1313 		 */
1314 		if (x86_type == X86_TYPE_VIA_CYRIX_III ||
1315 		    x86_type == X86_TYPE_CYRIX_GXm)
1316 			xcpuid++;
1317 		break;
1318 	case X86_VENDOR_Centaur:
1319 	case X86_VENDOR_TM:
1320 	default:
1321 		xcpuid++;
1322 		break;
1323 	}
1324 
1325 	if (xcpuid) {
1326 		cp = &cpi->cpi_extd[0];
1327 		cp->cp_eax = 0x80000000;
1328 		cpi->cpi_xmaxeax = __cpuid_insn(cp);
1329 	}
1330 
1331 	if (cpi->cpi_xmaxeax & 0x80000000) {
1332 
1333 		if (cpi->cpi_xmaxeax > CPI_XMAXEAX_MAX)
1334 			cpi->cpi_xmaxeax = CPI_XMAXEAX_MAX;
1335 
1336 		switch (cpi->cpi_vendor) {
1337 		case X86_VENDOR_Intel:
1338 		case X86_VENDOR_AMD:
1339 			if (cpi->cpi_xmaxeax < 0x80000001)
1340 				break;
1341 			cp = &cpi->cpi_extd[1];
1342 			cp->cp_eax = 0x80000001;
1343 			(void) __cpuid_insn(cp);
1344 
1345 			if (cpi->cpi_vendor == X86_VENDOR_AMD &&
1346 			    cpi->cpi_family == 5 &&
1347 			    cpi->cpi_model == 6 &&
1348 			    cpi->cpi_step == 6) {
1349 				/*
1350 				 * K6 model 6 uses bit 10 to indicate SYSC
1351 				 * Later models use bit 11. Fix it here.
1352 				 */
1353 				if (cp->cp_edx & 0x400) {
1354 					cp->cp_edx &= ~0x400;
1355 					cp->cp_edx |= CPUID_AMD_EDX_SYSC;
1356 				}
1357 			}
1358 
1359 			platform_cpuid_mangle(cpi->cpi_vendor, 0x80000001, cp);
1360 
1361 			/*
1362 			 * Compute the additions to the kernel's feature word.
1363 			 */
1364 			if (cp->cp_edx & CPUID_AMD_EDX_NX) {
1365 				add_x86_feature(featureset, X86FSET_NX);
1366 			}
1367 
1368 			/*
1369 			 * Regardless whether or not we boot 64-bit,
1370 			 * we should have a way to identify whether
1371 			 * the CPU is capable of running 64-bit.
1372 			 */
1373 			if (cp->cp_edx & CPUID_AMD_EDX_LM) {
1374 				add_x86_feature(featureset, X86FSET_64);
1375 			}
1376 
1377 #if defined(__amd64)
1378 			/* 1 GB large page - enable only for 64 bit kernel */
1379 			if (cp->cp_edx & CPUID_AMD_EDX_1GPG) {
1380 				add_x86_feature(featureset, X86FSET_1GPG);
1381 			}
1382 #endif
1383 
1384 			if ((cpi->cpi_vendor == X86_VENDOR_AMD) &&
1385 			    (cpi->cpi_std[1].cp_edx & CPUID_INTC_EDX_FXSR) &&
1386 			    (cp->cp_ecx & CPUID_AMD_ECX_SSE4A)) {
1387 				add_x86_feature(featureset, X86FSET_SSE4A);
1388 			}
1389 
1390 			/*
1391 			 * If both the HTT and CMP_LGCY bits are set,
1392 			 * then we're not actually HyperThreaded.  Read
1393 			 * "AMD CPUID Specification" for more details.
1394 			 */
1395 			if (cpi->cpi_vendor == X86_VENDOR_AMD &&
1396 			    is_x86_feature(featureset, X86FSET_HTT) &&
1397 			    (cp->cp_ecx & CPUID_AMD_ECX_CMP_LGCY)) {
1398 				remove_x86_feature(featureset, X86FSET_HTT);
1399 				add_x86_feature(featureset, X86FSET_CMP);
1400 			}
1401 #if defined(__amd64)
1402 			/*
1403 			 * It's really tricky to support syscall/sysret in
1404 			 * the i386 kernel; we rely on sysenter/sysexit
1405 			 * instead.  In the amd64 kernel, things are -way-
1406 			 * better.
1407 			 */
1408 			if (cp->cp_edx & CPUID_AMD_EDX_SYSC) {
1409 				add_x86_feature(featureset, X86FSET_ASYSC);
1410 			}
1411 
1412 			/*
1413 			 * While we're thinking about system calls, note
1414 			 * that AMD processors don't support sysenter
1415 			 * in long mode at all, so don't try to program them.
1416 			 */
1417 			if (x86_vendor == X86_VENDOR_AMD) {
1418 				remove_x86_feature(featureset, X86FSET_SEP);
1419 			}
1420 #endif
1421 			if (cp->cp_edx & CPUID_AMD_EDX_TSCP) {
1422 				add_x86_feature(featureset, X86FSET_TSCP);
1423 			}
1424 			break;
1425 		default:
1426 			break;
1427 		}
1428 
1429 		/*
1430 		 * Get CPUID data about processor cores and hyperthreads.
1431 		 */
1432 		switch (cpi->cpi_vendor) {
1433 		case X86_VENDOR_Intel:
1434 			if (cpi->cpi_maxeax >= 4) {
1435 				cp = &cpi->cpi_std[4];
1436 				cp->cp_eax = 4;
1437 				cp->cp_ecx = 0;
1438 				(void) __cpuid_insn(cp);
1439 				platform_cpuid_mangle(cpi->cpi_vendor, 4, cp);
1440 			}
1441 			/*FALLTHROUGH*/
1442 		case X86_VENDOR_AMD:
1443 			if (cpi->cpi_xmaxeax < 0x80000008)
1444 				break;
1445 			cp = &cpi->cpi_extd[8];
1446 			cp->cp_eax = 0x80000008;
1447 			(void) __cpuid_insn(cp);
1448 			platform_cpuid_mangle(cpi->cpi_vendor, 0x80000008, cp);
1449 
1450 			/*
1451 			 * Virtual and physical address limits from
1452 			 * cpuid override previously guessed values.
1453 			 */
1454 			cpi->cpi_pabits = BITX(cp->cp_eax, 7, 0);
1455 			cpi->cpi_vabits = BITX(cp->cp_eax, 15, 8);
1456 			break;
1457 		default:
1458 			break;
1459 		}
1460 
1461 		/*
1462 		 * Derive the number of cores per chip
1463 		 */
1464 		switch (cpi->cpi_vendor) {
1465 		case X86_VENDOR_Intel:
1466 			if (cpi->cpi_maxeax < 4) {
1467 				cpi->cpi_ncore_per_chip = 1;
1468 				break;
1469 			} else {
1470 				cpi->cpi_ncore_per_chip =
1471 				    BITX((cpi)->cpi_std[4].cp_eax, 31, 26) + 1;
1472 			}
1473 			break;
1474 		case X86_VENDOR_AMD:
1475 			if (cpi->cpi_xmaxeax < 0x80000008) {
1476 				cpi->cpi_ncore_per_chip = 1;
1477 				break;
1478 			} else {
1479 				/*
1480 				 * On family 0xf cpuid fn 2 ECX[7:0] "NC" is
1481 				 * 1 less than the number of physical cores on
1482 				 * the chip.  In family 0x10 this value can
1483 				 * be affected by "downcoring" - it reflects
1484 				 * 1 less than the number of cores actually
1485 				 * enabled on this node.
1486 				 */
1487 				cpi->cpi_ncore_per_chip =
1488 				    BITX((cpi)->cpi_extd[8].cp_ecx, 7, 0) + 1;
1489 			}
1490 			break;
1491 		default:
1492 			cpi->cpi_ncore_per_chip = 1;
1493 			break;
1494 		}
1495 
1496 		/*
1497 		 * Get CPUID data about TSC Invariance in Deep C-State.
1498 		 */
1499 		switch (cpi->cpi_vendor) {
1500 		case X86_VENDOR_Intel:
1501 			if (cpi->cpi_maxeax >= 7) {
1502 				cp = &cpi->cpi_extd[7];
1503 				cp->cp_eax = 0x80000007;
1504 				cp->cp_ecx = 0;
1505 				(void) __cpuid_insn(cp);
1506 			}
1507 			break;
1508 		default:
1509 			break;
1510 		}
1511 	} else {
1512 		cpi->cpi_ncore_per_chip = 1;
1513 	}
1514 
1515 	/*
1516 	 * If more than one core, then this processor is CMP.
1517 	 */
1518 	if (cpi->cpi_ncore_per_chip > 1) {
1519 		add_x86_feature(featureset, X86FSET_CMP);
1520 	}
1521 
1522 	/*
1523 	 * If the number of cores is the same as the number
1524 	 * of CPUs, then we cannot have HyperThreading.
1525 	 */
1526 	if (cpi->cpi_ncpu_per_chip == cpi->cpi_ncore_per_chip) {
1527 		remove_x86_feature(featureset, X86FSET_HTT);
1528 	}
1529 
1530 	cpi->cpi_apicid = CPI_APIC_ID(cpi);
1531 	cpi->cpi_procnodes_per_pkg = 1;
1532 	if (is_x86_feature(featureset, X86FSET_HTT) == B_FALSE &&
1533 	    is_x86_feature(featureset, X86FSET_CMP) == B_FALSE) {
1534 		/*
1535 		 * Single-core single-threaded processors.
1536 		 */
1537 		cpi->cpi_chipid = -1;
1538 		cpi->cpi_clogid = 0;
1539 		cpi->cpi_coreid = cpu->cpu_id;
1540 		cpi->cpi_pkgcoreid = 0;
1541 		if (cpi->cpi_vendor == X86_VENDOR_AMD)
1542 			cpi->cpi_procnodeid = BITX(cpi->cpi_apicid, 3, 0);
1543 		else
1544 			cpi->cpi_procnodeid = cpi->cpi_chipid;
1545 	} else if (cpi->cpi_ncpu_per_chip > 1) {
1546 		if (cpi->cpi_vendor == X86_VENDOR_Intel)
1547 			cpuid_intel_getids(cpu, featureset);
1548 		else if (cpi->cpi_vendor == X86_VENDOR_AMD)
1549 			cpuid_amd_getids(cpu);
1550 		else {
1551 			/*
1552 			 * All other processors are currently
1553 			 * assumed to have single cores.
1554 			 */
1555 			cpi->cpi_coreid = cpi->cpi_chipid;
1556 			cpi->cpi_pkgcoreid = 0;
1557 			cpi->cpi_procnodeid = cpi->cpi_chipid;
1558 		}
1559 	}
1560 
1561 	/*
1562 	 * Synthesize chip "revision" and socket type
1563 	 */
1564 	cpi->cpi_chiprev = _cpuid_chiprev(cpi->cpi_vendor, cpi->cpi_family,
1565 	    cpi->cpi_model, cpi->cpi_step);
1566 	cpi->cpi_chiprevstr = _cpuid_chiprevstr(cpi->cpi_vendor,
1567 	    cpi->cpi_family, cpi->cpi_model, cpi->cpi_step);
1568 	cpi->cpi_socket = _cpuid_skt(cpi->cpi_vendor, cpi->cpi_family,
1569 	    cpi->cpi_model, cpi->cpi_step);
1570 
1571 pass1_done:
1572 	cpi->cpi_pass = 1;
1573 }
1574 
1575 /*
1576  * Make copies of the cpuid table entries we depend on, in
1577  * part for ease of parsing now, in part so that we have only
1578  * one place to correct any of it, in part for ease of
1579  * later export to userland, and in part so we can look at
1580  * this stuff in a crash dump.
1581  */
1582 
1583 /*ARGSUSED*/
1584 void
1585 cpuid_pass2(cpu_t *cpu)
1586 {
1587 	uint_t n, nmax;
1588 	int i;
1589 	struct cpuid_regs *cp;
1590 	uint8_t *dp;
1591 	uint32_t *iptr;
1592 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
1593 
1594 	ASSERT(cpi->cpi_pass == 1);
1595 
1596 	if (cpi->cpi_maxeax < 1)
1597 		goto pass2_done;
1598 
1599 	if ((nmax = cpi->cpi_maxeax + 1) > NMAX_CPI_STD)
1600 		nmax = NMAX_CPI_STD;
1601 	/*
1602 	 * (We already handled n == 0 and n == 1 in pass 1)
1603 	 */
1604 	for (n = 2, cp = &cpi->cpi_std[2]; n < nmax; n++, cp++) {
1605 		cp->cp_eax = n;
1606 
1607 		/*
1608 		 * CPUID function 4 expects %ecx to be initialized
1609 		 * with an index which indicates which cache to return
1610 		 * information about. The OS is expected to call function 4
1611 		 * with %ecx set to 0, 1, 2, ... until it returns with
1612 		 * EAX[4:0] set to 0, which indicates there are no more
1613 		 * caches.
1614 		 *
1615 		 * Here, populate cpi_std[4] with the information returned by
1616 		 * function 4 when %ecx == 0, and do the rest in cpuid_pass3()
1617 		 * when dynamic memory allocation becomes available.
1618 		 *
1619 		 * Note: we need to explicitly initialize %ecx here, since
1620 		 * function 4 may have been previously invoked.
1621 		 */
1622 		if (n == 4)
1623 			cp->cp_ecx = 0;
1624 
1625 		(void) __cpuid_insn(cp);
1626 		platform_cpuid_mangle(cpi->cpi_vendor, n, cp);
1627 		switch (n) {
1628 		case 2:
1629 			/*
1630 			 * "the lower 8 bits of the %eax register
1631 			 * contain a value that identifies the number
1632 			 * of times the cpuid [instruction] has to be
1633 			 * executed to obtain a complete image of the
1634 			 * processor's caching systems."
1635 			 *
1636 			 * How *do* they make this stuff up?
1637 			 */
1638 			cpi->cpi_ncache = sizeof (*cp) *
1639 			    BITX(cp->cp_eax, 7, 0);
1640 			if (cpi->cpi_ncache == 0)
1641 				break;
1642 			cpi->cpi_ncache--;	/* skip count byte */
1643 
1644 			/*
1645 			 * Well, for now, rather than attempt to implement
1646 			 * this slightly dubious algorithm, we just look
1647 			 * at the first 15 ..
1648 			 */
1649 			if (cpi->cpi_ncache > (sizeof (*cp) - 1))
1650 				cpi->cpi_ncache = sizeof (*cp) - 1;
1651 
1652 			dp = cpi->cpi_cacheinfo;
1653 			if (BITX(cp->cp_eax, 31, 31) == 0) {
1654 				uint8_t *p = (void *)&cp->cp_eax;
1655 				for (i = 1; i < 4; i++)
1656 					if (p[i] != 0)
1657 						*dp++ = p[i];
1658 			}
1659 			if (BITX(cp->cp_ebx, 31, 31) == 0) {
1660 				uint8_t *p = (void *)&cp->cp_ebx;
1661 				for (i = 0; i < 4; i++)
1662 					if (p[i] != 0)
1663 						*dp++ = p[i];
1664 			}
1665 			if (BITX(cp->cp_ecx, 31, 31) == 0) {
1666 				uint8_t *p = (void *)&cp->cp_ecx;
1667 				for (i = 0; i < 4; i++)
1668 					if (p[i] != 0)
1669 						*dp++ = p[i];
1670 			}
1671 			if (BITX(cp->cp_edx, 31, 31) == 0) {
1672 				uint8_t *p = (void *)&cp->cp_edx;
1673 				for (i = 0; i < 4; i++)
1674 					if (p[i] != 0)
1675 						*dp++ = p[i];
1676 			}
1677 			break;
1678 
1679 		case 3:	/* Processor serial number, if PSN supported */
1680 			break;
1681 
1682 		case 4:	/* Deterministic cache parameters */
1683 			break;
1684 
1685 		case 5:	/* Monitor/Mwait parameters */
1686 		{
1687 			size_t mwait_size;
1688 
1689 			/*
1690 			 * check cpi_mwait.support which was set in cpuid_pass1
1691 			 */
1692 			if (!(cpi->cpi_mwait.support & MWAIT_SUPPORT))
1693 				break;
1694 
1695 			/*
1696 			 * Protect ourself from insane mwait line size.
1697 			 * Workaround for incomplete hardware emulator(s).
1698 			 */
1699 			mwait_size = (size_t)MWAIT_SIZE_MAX(cpi);
1700 			if (mwait_size < sizeof (uint32_t) ||
1701 			    !ISP2(mwait_size)) {
1702 #if DEBUG
1703 				cmn_err(CE_NOTE, "Cannot handle cpu %d mwait "
1704 				    "size %ld", cpu->cpu_id, (long)mwait_size);
1705 #endif
1706 				break;
1707 			}
1708 
1709 			cpi->cpi_mwait.mon_min = (size_t)MWAIT_SIZE_MIN(cpi);
1710 			cpi->cpi_mwait.mon_max = mwait_size;
1711 			if (MWAIT_EXTENSION(cpi)) {
1712 				cpi->cpi_mwait.support |= MWAIT_EXTENSIONS;
1713 				if (MWAIT_INT_ENABLE(cpi))
1714 					cpi->cpi_mwait.support |=
1715 					    MWAIT_ECX_INT_ENABLE;
1716 			}
1717 			break;
1718 		}
1719 		default:
1720 			break;
1721 		}
1722 	}
1723 
1724 	if (cpi->cpi_maxeax >= 0xB && cpi->cpi_vendor == X86_VENDOR_Intel) {
1725 		struct cpuid_regs regs;
1726 
1727 		cp = &regs;
1728 		cp->cp_eax = 0xB;
1729 		cp->cp_edx = cp->cp_ebx = cp->cp_ecx = 0;
1730 
1731 		(void) __cpuid_insn(cp);
1732 
1733 		/*
1734 		 * Check CPUID.EAX=0BH, ECX=0H:EBX is non-zero, which
1735 		 * indicates that the extended topology enumeration leaf is
1736 		 * available.
1737 		 */
1738 		if (cp->cp_ebx) {
1739 			uint32_t x2apic_id;
1740 			uint_t coreid_shift = 0;
1741 			uint_t ncpu_per_core = 1;
1742 			uint_t chipid_shift = 0;
1743 			uint_t ncpu_per_chip = 1;
1744 			uint_t i;
1745 			uint_t level;
1746 
1747 			for (i = 0; i < CPI_FNB_ECX_MAX; i++) {
1748 				cp->cp_eax = 0xB;
1749 				cp->cp_ecx = i;
1750 
1751 				(void) __cpuid_insn(cp);
1752 				level = CPI_CPU_LEVEL_TYPE(cp);
1753 
1754 				if (level == 1) {
1755 					x2apic_id = cp->cp_edx;
1756 					coreid_shift = BITX(cp->cp_eax, 4, 0);
1757 					ncpu_per_core = BITX(cp->cp_ebx, 15, 0);
1758 				} else if (level == 2) {
1759 					x2apic_id = cp->cp_edx;
1760 					chipid_shift = BITX(cp->cp_eax, 4, 0);
1761 					ncpu_per_chip = BITX(cp->cp_ebx, 15, 0);
1762 				}
1763 			}
1764 
1765 			cpi->cpi_apicid = x2apic_id;
1766 			cpi->cpi_ncpu_per_chip = ncpu_per_chip;
1767 			cpi->cpi_ncore_per_chip = ncpu_per_chip /
1768 			    ncpu_per_core;
1769 			cpi->cpi_chipid = x2apic_id >> chipid_shift;
1770 			cpi->cpi_clogid = x2apic_id & ((1 << chipid_shift) - 1);
1771 			cpi->cpi_coreid = x2apic_id >> coreid_shift;
1772 			cpi->cpi_pkgcoreid = cpi->cpi_clogid >> coreid_shift;
1773 		}
1774 
1775 		/* Make cp NULL so that we don't stumble on others */
1776 		cp = NULL;
1777 	}
1778 
1779 	/*
1780 	 * XSAVE enumeration
1781 	 */
1782 	if (cpi->cpi_maxeax >= 0xD && cpi->cpi_vendor == X86_VENDOR_Intel) {
1783 		struct cpuid_regs regs;
1784 		boolean_t cpuid_d_valid = B_TRUE;
1785 
1786 		cp = &regs;
1787 		cp->cp_eax = 0xD;
1788 		cp->cp_edx = cp->cp_ebx = cp->cp_ecx = 0;
1789 
1790 		(void) __cpuid_insn(cp);
1791 
1792 		/*
1793 		 * Sanity checks for debug
1794 		 */
1795 		if ((cp->cp_eax & XFEATURE_LEGACY_FP) == 0 ||
1796 		    (cp->cp_eax & XFEATURE_SSE) == 0) {
1797 			cpuid_d_valid = B_FALSE;
1798 		}
1799 
1800 		cpi->cpi_xsave.xsav_hw_features_low = cp->cp_eax;
1801 		cpi->cpi_xsave.xsav_hw_features_high = cp->cp_edx;
1802 		cpi->cpi_xsave.xsav_max_size = cp->cp_ecx;
1803 
1804 		/*
1805 		 * If the hw supports AVX, get the size and offset in the save
1806 		 * area for the ymm state.
1807 		 */
1808 		if (cpi->cpi_xsave.xsav_hw_features_low & XFEATURE_AVX) {
1809 			cp->cp_eax = 0xD;
1810 			cp->cp_ecx = 2;
1811 			cp->cp_edx = cp->cp_ebx = 0;
1812 
1813 			(void) __cpuid_insn(cp);
1814 
1815 			if (cp->cp_ebx != CPUID_LEAFD_2_YMM_OFFSET ||
1816 			    cp->cp_eax != CPUID_LEAFD_2_YMM_SIZE) {
1817 				cpuid_d_valid = B_FALSE;
1818 			}
1819 
1820 			cpi->cpi_xsave.ymm_size = cp->cp_eax;
1821 			cpi->cpi_xsave.ymm_offset = cp->cp_ebx;
1822 		}
1823 
1824 		if (is_x86_feature(x86_featureset, X86FSET_XSAVE)) {
1825 			xsave_state_size = 0;
1826 		} else if (cpuid_d_valid) {
1827 			xsave_state_size = cpi->cpi_xsave.xsav_max_size;
1828 		} else {
1829 			/* Broken CPUID 0xD, probably in HVM */
1830 			cmn_err(CE_WARN, "cpu%d: CPUID.0xD returns invalid "
1831 			    "value: hw_low = %d, hw_high = %d, xsave_size = %d"
1832 			    ", ymm_size = %d, ymm_offset = %d\n",
1833 			    cpu->cpu_id, cpi->cpi_xsave.xsav_hw_features_low,
1834 			    cpi->cpi_xsave.xsav_hw_features_high,
1835 			    (int)cpi->cpi_xsave.xsav_max_size,
1836 			    (int)cpi->cpi_xsave.ymm_size,
1837 			    (int)cpi->cpi_xsave.ymm_offset);
1838 
1839 			if (xsave_state_size != 0) {
1840 				/*
1841 				 * This must be a non-boot CPU. We cannot
1842 				 * continue, because boot cpu has already
1843 				 * enabled XSAVE.
1844 				 */
1845 				ASSERT(cpu->cpu_id != 0);
1846 				cmn_err(CE_PANIC, "cpu%d: we have already "
1847 				    "enabled XSAVE on boot cpu, cannot "
1848 				    "continue.", cpu->cpu_id);
1849 			} else {
1850 				/*
1851 				 * Must be from boot CPU, OK to disable XSAVE.
1852 				 */
1853 				ASSERT(cpu->cpu_id == 0);
1854 				remove_x86_feature(x86_featureset,
1855 				    X86FSET_XSAVE);
1856 				remove_x86_feature(x86_featureset, X86FSET_AVX);
1857 				CPI_FEATURES_ECX(cpi) &= ~CPUID_INTC_ECX_XSAVE;
1858 				CPI_FEATURES_ECX(cpi) &= ~CPUID_INTC_ECX_AVX;
1859 				xsave_force_disable = B_TRUE;
1860 			}
1861 		}
1862 	}
1863 
1864 
1865 	if ((cpi->cpi_xmaxeax & 0x80000000) == 0)
1866 		goto pass2_done;
1867 
1868 	if ((nmax = cpi->cpi_xmaxeax - 0x80000000 + 1) > NMAX_CPI_EXTD)
1869 		nmax = NMAX_CPI_EXTD;
1870 	/*
1871 	 * Copy the extended properties, fixing them as we go.
1872 	 * (We already handled n == 0 and n == 1 in pass 1)
1873 	 */
1874 	iptr = (void *)cpi->cpi_brandstr;
1875 	for (n = 2, cp = &cpi->cpi_extd[2]; n < nmax; cp++, n++) {
1876 		cp->cp_eax = 0x80000000 + n;
1877 		(void) __cpuid_insn(cp);
1878 		platform_cpuid_mangle(cpi->cpi_vendor, 0x80000000 + n, cp);
1879 		switch (n) {
1880 		case 2:
1881 		case 3:
1882 		case 4:
1883 			/*
1884 			 * Extract the brand string
1885 			 */
1886 			*iptr++ = cp->cp_eax;
1887 			*iptr++ = cp->cp_ebx;
1888 			*iptr++ = cp->cp_ecx;
1889 			*iptr++ = cp->cp_edx;
1890 			break;
1891 		case 5:
1892 			switch (cpi->cpi_vendor) {
1893 			case X86_VENDOR_AMD:
1894 				/*
1895 				 * The Athlon and Duron were the first
1896 				 * parts to report the sizes of the
1897 				 * TLB for large pages. Before then,
1898 				 * we don't trust the data.
1899 				 */
1900 				if (cpi->cpi_family < 6 ||
1901 				    (cpi->cpi_family == 6 &&
1902 				    cpi->cpi_model < 1))
1903 					cp->cp_eax = 0;
1904 				break;
1905 			default:
1906 				break;
1907 			}
1908 			break;
1909 		case 6:
1910 			switch (cpi->cpi_vendor) {
1911 			case X86_VENDOR_AMD:
1912 				/*
1913 				 * The Athlon and Duron were the first
1914 				 * AMD parts with L2 TLB's.
1915 				 * Before then, don't trust the data.
1916 				 */
1917 				if (cpi->cpi_family < 6 ||
1918 				    cpi->cpi_family == 6 &&
1919 				    cpi->cpi_model < 1)
1920 					cp->cp_eax = cp->cp_ebx = 0;
1921 				/*
1922 				 * AMD Duron rev A0 reports L2
1923 				 * cache size incorrectly as 1K
1924 				 * when it is really 64K
1925 				 */
1926 				if (cpi->cpi_family == 6 &&
1927 				    cpi->cpi_model == 3 &&
1928 				    cpi->cpi_step == 0) {
1929 					cp->cp_ecx &= 0xffff;
1930 					cp->cp_ecx |= 0x400000;
1931 				}
1932 				break;
1933 			case X86_VENDOR_Cyrix:	/* VIA C3 */
1934 				/*
1935 				 * VIA C3 processors are a bit messed
1936 				 * up w.r.t. encoding cache sizes in %ecx
1937 				 */
1938 				if (cpi->cpi_family != 6)
1939 					break;
1940 				/*
1941 				 * model 7 and 8 were incorrectly encoded
1942 				 *
1943 				 * xxx is model 8 really broken?
1944 				 */
1945 				if (cpi->cpi_model == 7 ||
1946 				    cpi->cpi_model == 8)
1947 					cp->cp_ecx =
1948 					    BITX(cp->cp_ecx, 31, 24) << 16 |
1949 					    BITX(cp->cp_ecx, 23, 16) << 12 |
1950 					    BITX(cp->cp_ecx, 15, 8) << 8 |
1951 					    BITX(cp->cp_ecx, 7, 0);
1952 				/*
1953 				 * model 9 stepping 1 has wrong associativity
1954 				 */
1955 				if (cpi->cpi_model == 9 && cpi->cpi_step == 1)
1956 					cp->cp_ecx |= 8 << 12;
1957 				break;
1958 			case X86_VENDOR_Intel:
1959 				/*
1960 				 * Extended L2 Cache features function.
1961 				 * First appeared on Prescott.
1962 				 */
1963 			default:
1964 				break;
1965 			}
1966 			break;
1967 		default:
1968 			break;
1969 		}
1970 	}
1971 
1972 pass2_done:
1973 	cpi->cpi_pass = 2;
1974 }
1975 
1976 static const char *
1977 intel_cpubrand(const struct cpuid_info *cpi)
1978 {
1979 	int i;
1980 
1981 	if (!is_x86_feature(x86_featureset, X86FSET_CPUID) ||
1982 	    cpi->cpi_maxeax < 1 || cpi->cpi_family < 5)
1983 		return ("i486");
1984 
1985 	switch (cpi->cpi_family) {
1986 	case 5:
1987 		return ("Intel Pentium(r)");
1988 	case 6:
1989 		switch (cpi->cpi_model) {
1990 			uint_t celeron, xeon;
1991 			const struct cpuid_regs *cp;
1992 		case 0:
1993 		case 1:
1994 		case 2:
1995 			return ("Intel Pentium(r) Pro");
1996 		case 3:
1997 		case 4:
1998 			return ("Intel Pentium(r) II");
1999 		case 6:
2000 			return ("Intel Celeron(r)");
2001 		case 5:
2002 		case 7:
2003 			celeron = xeon = 0;
2004 			cp = &cpi->cpi_std[2];	/* cache info */
2005 
2006 			for (i = 1; i < 4; i++) {
2007 				uint_t tmp;
2008 
2009 				tmp = (cp->cp_eax >> (8 * i)) & 0xff;
2010 				if (tmp == 0x40)
2011 					celeron++;
2012 				if (tmp >= 0x44 && tmp <= 0x45)
2013 					xeon++;
2014 			}
2015 
2016 			for (i = 0; i < 2; i++) {
2017 				uint_t tmp;
2018 
2019 				tmp = (cp->cp_ebx >> (8 * i)) & 0xff;
2020 				if (tmp == 0x40)
2021 					celeron++;
2022 				else if (tmp >= 0x44 && tmp <= 0x45)
2023 					xeon++;
2024 			}
2025 
2026 			for (i = 0; i < 4; i++) {
2027 				uint_t tmp;
2028 
2029 				tmp = (cp->cp_ecx >> (8 * i)) & 0xff;
2030 				if (tmp == 0x40)
2031 					celeron++;
2032 				else if (tmp >= 0x44 && tmp <= 0x45)
2033 					xeon++;
2034 			}
2035 
2036 			for (i = 0; i < 4; i++) {
2037 				uint_t tmp;
2038 
2039 				tmp = (cp->cp_edx >> (8 * i)) & 0xff;
2040 				if (tmp == 0x40)
2041 					celeron++;
2042 				else if (tmp >= 0x44 && tmp <= 0x45)
2043 					xeon++;
2044 			}
2045 
2046 			if (celeron)
2047 				return ("Intel Celeron(r)");
2048 			if (xeon)
2049 				return (cpi->cpi_model == 5 ?
2050 				    "Intel Pentium(r) II Xeon(tm)" :
2051 				    "Intel Pentium(r) III Xeon(tm)");
2052 			return (cpi->cpi_model == 5 ?
2053 			    "Intel Pentium(r) II or Pentium(r) II Xeon(tm)" :
2054 			    "Intel Pentium(r) III or Pentium(r) III Xeon(tm)");
2055 		default:
2056 			break;
2057 		}
2058 	default:
2059 		break;
2060 	}
2061 
2062 	/* BrandID is present if the field is nonzero */
2063 	if (cpi->cpi_brandid != 0) {
2064 		static const struct {
2065 			uint_t bt_bid;
2066 			const char *bt_str;
2067 		} brand_tbl[] = {
2068 			{ 0x1,	"Intel(r) Celeron(r)" },
2069 			{ 0x2,	"Intel(r) Pentium(r) III" },
2070 			{ 0x3,	"Intel(r) Pentium(r) III Xeon(tm)" },
2071 			{ 0x4,	"Intel(r) Pentium(r) III" },
2072 			{ 0x6,	"Mobile Intel(r) Pentium(r) III" },
2073 			{ 0x7,	"Mobile Intel(r) Celeron(r)" },
2074 			{ 0x8,	"Intel(r) Pentium(r) 4" },
2075 			{ 0x9,	"Intel(r) Pentium(r) 4" },
2076 			{ 0xa,	"Intel(r) Celeron(r)" },
2077 			{ 0xb,	"Intel(r) Xeon(tm)" },
2078 			{ 0xc,	"Intel(r) Xeon(tm) MP" },
2079 			{ 0xe,	"Mobile Intel(r) Pentium(r) 4" },
2080 			{ 0xf,	"Mobile Intel(r) Celeron(r)" },
2081 			{ 0x11, "Mobile Genuine Intel(r)" },
2082 			{ 0x12, "Intel(r) Celeron(r) M" },
2083 			{ 0x13, "Mobile Intel(r) Celeron(r)" },
2084 			{ 0x14, "Intel(r) Celeron(r)" },
2085 			{ 0x15, "Mobile Genuine Intel(r)" },
2086 			{ 0x16,	"Intel(r) Pentium(r) M" },
2087 			{ 0x17, "Mobile Intel(r) Celeron(r)" }
2088 		};
2089 		uint_t btblmax = sizeof (brand_tbl) / sizeof (brand_tbl[0]);
2090 		uint_t sgn;
2091 
2092 		sgn = (cpi->cpi_family << 8) |
2093 		    (cpi->cpi_model << 4) | cpi->cpi_step;
2094 
2095 		for (i = 0; i < btblmax; i++)
2096 			if (brand_tbl[i].bt_bid == cpi->cpi_brandid)
2097 				break;
2098 		if (i < btblmax) {
2099 			if (sgn == 0x6b1 && cpi->cpi_brandid == 3)
2100 				return ("Intel(r) Celeron(r)");
2101 			if (sgn < 0xf13 && cpi->cpi_brandid == 0xb)
2102 				return ("Intel(r) Xeon(tm) MP");
2103 			if (sgn < 0xf13 && cpi->cpi_brandid == 0xe)
2104 				return ("Intel(r) Xeon(tm)");
2105 			return (brand_tbl[i].bt_str);
2106 		}
2107 	}
2108 
2109 	return (NULL);
2110 }
2111 
2112 static const char *
2113 amd_cpubrand(const struct cpuid_info *cpi)
2114 {
2115 	if (!is_x86_feature(x86_featureset, X86FSET_CPUID) ||
2116 	    cpi->cpi_maxeax < 1 || cpi->cpi_family < 5)
2117 		return ("i486 compatible");
2118 
2119 	switch (cpi->cpi_family) {
2120 	case 5:
2121 		switch (cpi->cpi_model) {
2122 		case 0:
2123 		case 1:
2124 		case 2:
2125 		case 3:
2126 		case 4:
2127 		case 5:
2128 			return ("AMD-K5(r)");
2129 		case 6:
2130 		case 7:
2131 			return ("AMD-K6(r)");
2132 		case 8:
2133 			return ("AMD-K6(r)-2");
2134 		case 9:
2135 			return ("AMD-K6(r)-III");
2136 		default:
2137 			return ("AMD (family 5)");
2138 		}
2139 	case 6:
2140 		switch (cpi->cpi_model) {
2141 		case 1:
2142 			return ("AMD-K7(tm)");
2143 		case 0:
2144 		case 2:
2145 		case 4:
2146 			return ("AMD Athlon(tm)");
2147 		case 3:
2148 		case 7:
2149 			return ("AMD Duron(tm)");
2150 		case 6:
2151 		case 8:
2152 		case 10:
2153 			/*
2154 			 * Use the L2 cache size to distinguish
2155 			 */
2156 			return ((cpi->cpi_extd[6].cp_ecx >> 16) >= 256 ?
2157 			    "AMD Athlon(tm)" : "AMD Duron(tm)");
2158 		default:
2159 			return ("AMD (family 6)");
2160 		}
2161 	default:
2162 		break;
2163 	}
2164 
2165 	if (cpi->cpi_family == 0xf && cpi->cpi_model == 5 &&
2166 	    cpi->cpi_brandid != 0) {
2167 		switch (BITX(cpi->cpi_brandid, 7, 5)) {
2168 		case 3:
2169 			return ("AMD Opteron(tm) UP 1xx");
2170 		case 4:
2171 			return ("AMD Opteron(tm) DP 2xx");
2172 		case 5:
2173 			return ("AMD Opteron(tm) MP 8xx");
2174 		default:
2175 			return ("AMD Opteron(tm)");
2176 		}
2177 	}
2178 
2179 	return (NULL);
2180 }
2181 
2182 static const char *
2183 cyrix_cpubrand(struct cpuid_info *cpi, uint_t type)
2184 {
2185 	if (!is_x86_feature(x86_featureset, X86FSET_CPUID) ||
2186 	    cpi->cpi_maxeax < 1 || cpi->cpi_family < 5 ||
2187 	    type == X86_TYPE_CYRIX_486)
2188 		return ("i486 compatible");
2189 
2190 	switch (type) {
2191 	case X86_TYPE_CYRIX_6x86:
2192 		return ("Cyrix 6x86");
2193 	case X86_TYPE_CYRIX_6x86L:
2194 		return ("Cyrix 6x86L");
2195 	case X86_TYPE_CYRIX_6x86MX:
2196 		return ("Cyrix 6x86MX");
2197 	case X86_TYPE_CYRIX_GXm:
2198 		return ("Cyrix GXm");
2199 	case X86_TYPE_CYRIX_MediaGX:
2200 		return ("Cyrix MediaGX");
2201 	case X86_TYPE_CYRIX_MII:
2202 		return ("Cyrix M2");
2203 	case X86_TYPE_VIA_CYRIX_III:
2204 		return ("VIA Cyrix M3");
2205 	default:
2206 		/*
2207 		 * Have another wild guess ..
2208 		 */
2209 		if (cpi->cpi_family == 4 && cpi->cpi_model == 9)
2210 			return ("Cyrix 5x86");
2211 		else if (cpi->cpi_family == 5) {
2212 			switch (cpi->cpi_model) {
2213 			case 2:
2214 				return ("Cyrix 6x86");	/* Cyrix M1 */
2215 			case 4:
2216 				return ("Cyrix MediaGX");
2217 			default:
2218 				break;
2219 			}
2220 		} else if (cpi->cpi_family == 6) {
2221 			switch (cpi->cpi_model) {
2222 			case 0:
2223 				return ("Cyrix 6x86MX"); /* Cyrix M2? */
2224 			case 5:
2225 			case 6:
2226 			case 7:
2227 			case 8:
2228 			case 9:
2229 				return ("VIA C3");
2230 			default:
2231 				break;
2232 			}
2233 		}
2234 		break;
2235 	}
2236 	return (NULL);
2237 }
2238 
2239 /*
2240  * This only gets called in the case that the CPU extended
2241  * feature brand string (0x80000002, 0x80000003, 0x80000004)
2242  * aren't available, or contain null bytes for some reason.
2243  */
2244 static void
2245 fabricate_brandstr(struct cpuid_info *cpi)
2246 {
2247 	const char *brand = NULL;
2248 
2249 	switch (cpi->cpi_vendor) {
2250 	case X86_VENDOR_Intel:
2251 		brand = intel_cpubrand(cpi);
2252 		break;
2253 	case X86_VENDOR_AMD:
2254 		brand = amd_cpubrand(cpi);
2255 		break;
2256 	case X86_VENDOR_Cyrix:
2257 		brand = cyrix_cpubrand(cpi, x86_type);
2258 		break;
2259 	case X86_VENDOR_NexGen:
2260 		if (cpi->cpi_family == 5 && cpi->cpi_model == 0)
2261 			brand = "NexGen Nx586";
2262 		break;
2263 	case X86_VENDOR_Centaur:
2264 		if (cpi->cpi_family == 5)
2265 			switch (cpi->cpi_model) {
2266 			case 4:
2267 				brand = "Centaur C6";
2268 				break;
2269 			case 8:
2270 				brand = "Centaur C2";
2271 				break;
2272 			case 9:
2273 				brand = "Centaur C3";
2274 				break;
2275 			default:
2276 				break;
2277 			}
2278 		break;
2279 	case X86_VENDOR_Rise:
2280 		if (cpi->cpi_family == 5 &&
2281 		    (cpi->cpi_model == 0 || cpi->cpi_model == 2))
2282 			brand = "Rise mP6";
2283 		break;
2284 	case X86_VENDOR_SiS:
2285 		if (cpi->cpi_family == 5 && cpi->cpi_model == 0)
2286 			brand = "SiS 55x";
2287 		break;
2288 	case X86_VENDOR_TM:
2289 		if (cpi->cpi_family == 5 && cpi->cpi_model == 4)
2290 			brand = "Transmeta Crusoe TM3x00 or TM5x00";
2291 		break;
2292 	case X86_VENDOR_NSC:
2293 	case X86_VENDOR_UMC:
2294 	default:
2295 		break;
2296 	}
2297 	if (brand) {
2298 		(void) strcpy((char *)cpi->cpi_brandstr, brand);
2299 		return;
2300 	}
2301 
2302 	/*
2303 	 * If all else fails ...
2304 	 */
2305 	(void) snprintf(cpi->cpi_brandstr, sizeof (cpi->cpi_brandstr),
2306 	    "%s %d.%d.%d", cpi->cpi_vendorstr, cpi->cpi_family,
2307 	    cpi->cpi_model, cpi->cpi_step);
2308 }
2309 
2310 /*
2311  * This routine is called just after kernel memory allocation
2312  * becomes available on cpu0, and as part of mp_startup() on
2313  * the other cpus.
2314  *
2315  * Fixup the brand string, and collect any information from cpuid
2316  * that requires dynamicically allocated storage to represent.
2317  */
2318 /*ARGSUSED*/
2319 void
2320 cpuid_pass3(cpu_t *cpu)
2321 {
2322 	int	i, max, shft, level, size;
2323 	struct cpuid_regs regs;
2324 	struct cpuid_regs *cp;
2325 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2326 
2327 	ASSERT(cpi->cpi_pass == 2);
2328 
2329 	/*
2330 	 * Function 4: Deterministic cache parameters
2331 	 *
2332 	 * Take this opportunity to detect the number of threads
2333 	 * sharing the last level cache, and construct a corresponding
2334 	 * cache id. The respective cpuid_info members are initialized
2335 	 * to the default case of "no last level cache sharing".
2336 	 */
2337 	cpi->cpi_ncpu_shr_last_cache = 1;
2338 	cpi->cpi_last_lvl_cacheid = cpu->cpu_id;
2339 
2340 	if (cpi->cpi_maxeax >= 4 && cpi->cpi_vendor == X86_VENDOR_Intel) {
2341 
2342 		/*
2343 		 * Find the # of elements (size) returned by fn 4, and along
2344 		 * the way detect last level cache sharing details.
2345 		 */
2346 		bzero(&regs, sizeof (regs));
2347 		cp = &regs;
2348 		for (i = 0, max = 0; i < CPI_FN4_ECX_MAX; i++) {
2349 			cp->cp_eax = 4;
2350 			cp->cp_ecx = i;
2351 
2352 			(void) __cpuid_insn(cp);
2353 
2354 			if (CPI_CACHE_TYPE(cp) == 0)
2355 				break;
2356 			level = CPI_CACHE_LVL(cp);
2357 			if (level > max) {
2358 				max = level;
2359 				cpi->cpi_ncpu_shr_last_cache =
2360 				    CPI_NTHR_SHR_CACHE(cp) + 1;
2361 			}
2362 		}
2363 		cpi->cpi_std_4_size = size = i;
2364 
2365 		/*
2366 		 * Allocate the cpi_std_4 array. The first element
2367 		 * references the regs for fn 4, %ecx == 0, which
2368 		 * cpuid_pass2() stashed in cpi->cpi_std[4].
2369 		 */
2370 		if (size > 0) {
2371 			cpi->cpi_std_4 =
2372 			    kmem_alloc(size * sizeof (cp), KM_SLEEP);
2373 			cpi->cpi_std_4[0] = &cpi->cpi_std[4];
2374 
2375 			/*
2376 			 * Allocate storage to hold the additional regs
2377 			 * for function 4, %ecx == 1 .. cpi_std_4_size.
2378 			 *
2379 			 * The regs for fn 4, %ecx == 0 has already
2380 			 * been allocated as indicated above.
2381 			 */
2382 			for (i = 1; i < size; i++) {
2383 				cp = cpi->cpi_std_4[i] =
2384 				    kmem_zalloc(sizeof (regs), KM_SLEEP);
2385 				cp->cp_eax = 4;
2386 				cp->cp_ecx = i;
2387 
2388 				(void) __cpuid_insn(cp);
2389 			}
2390 		}
2391 		/*
2392 		 * Determine the number of bits needed to represent
2393 		 * the number of CPUs sharing the last level cache.
2394 		 *
2395 		 * Shift off that number of bits from the APIC id to
2396 		 * derive the cache id.
2397 		 */
2398 		shft = 0;
2399 		for (i = 1; i < cpi->cpi_ncpu_shr_last_cache; i <<= 1)
2400 			shft++;
2401 		cpi->cpi_last_lvl_cacheid = cpi->cpi_apicid >> shft;
2402 	}
2403 
2404 	/*
2405 	 * Now fixup the brand string
2406 	 */
2407 	if ((cpi->cpi_xmaxeax & 0x80000000) == 0) {
2408 		fabricate_brandstr(cpi);
2409 	} else {
2410 
2411 		/*
2412 		 * If we successfully extracted a brand string from the cpuid
2413 		 * instruction, clean it up by removing leading spaces and
2414 		 * similar junk.
2415 		 */
2416 		if (cpi->cpi_brandstr[0]) {
2417 			size_t maxlen = sizeof (cpi->cpi_brandstr);
2418 			char *src, *dst;
2419 
2420 			dst = src = (char *)cpi->cpi_brandstr;
2421 			src[maxlen - 1] = '\0';
2422 			/*
2423 			 * strip leading spaces
2424 			 */
2425 			while (*src == ' ')
2426 				src++;
2427 			/*
2428 			 * Remove any 'Genuine' or "Authentic" prefixes
2429 			 */
2430 			if (strncmp(src, "Genuine ", 8) == 0)
2431 				src += 8;
2432 			if (strncmp(src, "Authentic ", 10) == 0)
2433 				src += 10;
2434 
2435 			/*
2436 			 * Now do an in-place copy.
2437 			 * Map (R) to (r) and (TM) to (tm).
2438 			 * The era of teletypes is long gone, and there's
2439 			 * -really- no need to shout.
2440 			 */
2441 			while (*src != '\0') {
2442 				if (src[0] == '(') {
2443 					if (strncmp(src + 1, "R)", 2) == 0) {
2444 						(void) strncpy(dst, "(r)", 3);
2445 						src += 3;
2446 						dst += 3;
2447 						continue;
2448 					}
2449 					if (strncmp(src + 1, "TM)", 3) == 0) {
2450 						(void) strncpy(dst, "(tm)", 4);
2451 						src += 4;
2452 						dst += 4;
2453 						continue;
2454 					}
2455 				}
2456 				*dst++ = *src++;
2457 			}
2458 			*dst = '\0';
2459 
2460 			/*
2461 			 * Finally, remove any trailing spaces
2462 			 */
2463 			while (--dst > cpi->cpi_brandstr)
2464 				if (*dst == ' ')
2465 					*dst = '\0';
2466 				else
2467 					break;
2468 		} else
2469 			fabricate_brandstr(cpi);
2470 	}
2471 	cpi->cpi_pass = 3;
2472 }
2473 
2474 /*
2475  * This routine is called out of bind_hwcap() much later in the life
2476  * of the kernel (post_startup()).  The job of this routine is to resolve
2477  * the hardware feature support and kernel support for those features into
2478  * what we're actually going to tell applications via the aux vector.
2479  */
2480 uint_t
2481 cpuid_pass4(cpu_t *cpu)
2482 {
2483 	struct cpuid_info *cpi;
2484 	uint_t hwcap_flags = 0;
2485 
2486 	if (cpu == NULL)
2487 		cpu = CPU;
2488 	cpi = cpu->cpu_m.mcpu_cpi;
2489 
2490 	ASSERT(cpi->cpi_pass == 3);
2491 
2492 	if (cpi->cpi_maxeax >= 1) {
2493 		uint32_t *edx = &cpi->cpi_support[STD_EDX_FEATURES];
2494 		uint32_t *ecx = &cpi->cpi_support[STD_ECX_FEATURES];
2495 
2496 		*edx = CPI_FEATURES_EDX(cpi);
2497 		*ecx = CPI_FEATURES_ECX(cpi);
2498 
2499 		/*
2500 		 * [these require explicit kernel support]
2501 		 */
2502 		if (!is_x86_feature(x86_featureset, X86FSET_SEP))
2503 			*edx &= ~CPUID_INTC_EDX_SEP;
2504 
2505 		if (!is_x86_feature(x86_featureset, X86FSET_SSE))
2506 			*edx &= ~(CPUID_INTC_EDX_FXSR|CPUID_INTC_EDX_SSE);
2507 		if (!is_x86_feature(x86_featureset, X86FSET_SSE2))
2508 			*edx &= ~CPUID_INTC_EDX_SSE2;
2509 
2510 		if (!is_x86_feature(x86_featureset, X86FSET_HTT))
2511 			*edx &= ~CPUID_INTC_EDX_HTT;
2512 
2513 		if (!is_x86_feature(x86_featureset, X86FSET_SSE3))
2514 			*ecx &= ~CPUID_INTC_ECX_SSE3;
2515 
2516 		if (cpi->cpi_vendor == X86_VENDOR_Intel) {
2517 			if (!is_x86_feature(x86_featureset, X86FSET_SSSE3))
2518 				*ecx &= ~CPUID_INTC_ECX_SSSE3;
2519 			if (!is_x86_feature(x86_featureset, X86FSET_SSE4_1))
2520 				*ecx &= ~CPUID_INTC_ECX_SSE4_1;
2521 			if (!is_x86_feature(x86_featureset, X86FSET_SSE4_2))
2522 				*ecx &= ~CPUID_INTC_ECX_SSE4_2;
2523 			if (!is_x86_feature(x86_featureset, X86FSET_AES))
2524 				*ecx &= ~CPUID_INTC_ECX_AES;
2525 			if (!is_x86_feature(x86_featureset, X86FSET_PCLMULQDQ))
2526 				*ecx &= ~CPUID_INTC_ECX_PCLMULQDQ;
2527 			if (!is_x86_feature(x86_featureset, X86FSET_XSAVE))
2528 				*ecx &= ~(CPUID_INTC_ECX_XSAVE |
2529 				    CPUID_INTC_ECX_OSXSAVE);
2530 			if (!is_x86_feature(x86_featureset, X86FSET_AVX))
2531 				*ecx &= ~CPUID_INTC_ECX_AVX;
2532 		}
2533 
2534 		/*
2535 		 * [no explicit support required beyond x87 fp context]
2536 		 */
2537 		if (!fpu_exists)
2538 			*edx &= ~(CPUID_INTC_EDX_FPU | CPUID_INTC_EDX_MMX);
2539 
2540 		/*
2541 		 * Now map the supported feature vector to things that we
2542 		 * think userland will care about.
2543 		 */
2544 		if (*edx & CPUID_INTC_EDX_SEP)
2545 			hwcap_flags |= AV_386_SEP;
2546 		if (*edx & CPUID_INTC_EDX_SSE)
2547 			hwcap_flags |= AV_386_FXSR | AV_386_SSE;
2548 		if (*edx & CPUID_INTC_EDX_SSE2)
2549 			hwcap_flags |= AV_386_SSE2;
2550 		if (*ecx & CPUID_INTC_ECX_SSE3)
2551 			hwcap_flags |= AV_386_SSE3;
2552 		if (cpi->cpi_vendor == X86_VENDOR_Intel) {
2553 			if (*ecx & CPUID_INTC_ECX_SSSE3)
2554 				hwcap_flags |= AV_386_SSSE3;
2555 			if (*ecx & CPUID_INTC_ECX_SSE4_1)
2556 				hwcap_flags |= AV_386_SSE4_1;
2557 			if (*ecx & CPUID_INTC_ECX_SSE4_2)
2558 				hwcap_flags |= AV_386_SSE4_2;
2559 			if (*ecx & CPUID_INTC_ECX_MOVBE)
2560 				hwcap_flags |= AV_386_MOVBE;
2561 			if (*ecx & CPUID_INTC_ECX_AES)
2562 				hwcap_flags |= AV_386_AES;
2563 			if (*ecx & CPUID_INTC_ECX_PCLMULQDQ)
2564 				hwcap_flags |= AV_386_PCLMULQDQ;
2565 			if ((*ecx & CPUID_INTC_ECX_XSAVE) &&
2566 			    (*ecx & CPUID_INTC_ECX_OSXSAVE))
2567 				hwcap_flags |= AV_386_XSAVE;
2568 		}
2569 		if (*ecx & CPUID_INTC_ECX_POPCNT)
2570 			hwcap_flags |= AV_386_POPCNT;
2571 		if (*edx & CPUID_INTC_EDX_FPU)
2572 			hwcap_flags |= AV_386_FPU;
2573 		if (*edx & CPUID_INTC_EDX_MMX)
2574 			hwcap_flags |= AV_386_MMX;
2575 
2576 		if (*edx & CPUID_INTC_EDX_TSC)
2577 			hwcap_flags |= AV_386_TSC;
2578 		if (*edx & CPUID_INTC_EDX_CX8)
2579 			hwcap_flags |= AV_386_CX8;
2580 		if (*edx & CPUID_INTC_EDX_CMOV)
2581 			hwcap_flags |= AV_386_CMOV;
2582 		if (*ecx & CPUID_INTC_ECX_CX16)
2583 			hwcap_flags |= AV_386_CX16;
2584 	}
2585 
2586 	if (cpi->cpi_xmaxeax < 0x80000001)
2587 		goto pass4_done;
2588 
2589 	switch (cpi->cpi_vendor) {
2590 		struct cpuid_regs cp;
2591 		uint32_t *edx, *ecx;
2592 
2593 	case X86_VENDOR_Intel:
2594 		/*
2595 		 * Seems like Intel duplicated what we necessary
2596 		 * here to make the initial crop of 64-bit OS's work.
2597 		 * Hopefully, those are the only "extended" bits
2598 		 * they'll add.
2599 		 */
2600 		/*FALLTHROUGH*/
2601 
2602 	case X86_VENDOR_AMD:
2603 		edx = &cpi->cpi_support[AMD_EDX_FEATURES];
2604 		ecx = &cpi->cpi_support[AMD_ECX_FEATURES];
2605 
2606 		*edx = CPI_FEATURES_XTD_EDX(cpi);
2607 		*ecx = CPI_FEATURES_XTD_ECX(cpi);
2608 
2609 		/*
2610 		 * [these features require explicit kernel support]
2611 		 */
2612 		switch (cpi->cpi_vendor) {
2613 		case X86_VENDOR_Intel:
2614 			if (!is_x86_feature(x86_featureset, X86FSET_TSCP))
2615 				*edx &= ~CPUID_AMD_EDX_TSCP;
2616 			break;
2617 
2618 		case X86_VENDOR_AMD:
2619 			if (!is_x86_feature(x86_featureset, X86FSET_TSCP))
2620 				*edx &= ~CPUID_AMD_EDX_TSCP;
2621 			if (!is_x86_feature(x86_featureset, X86FSET_SSE4A))
2622 				*ecx &= ~CPUID_AMD_ECX_SSE4A;
2623 			break;
2624 
2625 		default:
2626 			break;
2627 		}
2628 
2629 		/*
2630 		 * [no explicit support required beyond
2631 		 * x87 fp context and exception handlers]
2632 		 */
2633 		if (!fpu_exists)
2634 			*edx &= ~(CPUID_AMD_EDX_MMXamd |
2635 			    CPUID_AMD_EDX_3DNow | CPUID_AMD_EDX_3DNowx);
2636 
2637 		if (!is_x86_feature(x86_featureset, X86FSET_NX))
2638 			*edx &= ~CPUID_AMD_EDX_NX;
2639 #if !defined(__amd64)
2640 		*edx &= ~CPUID_AMD_EDX_LM;
2641 #endif
2642 		/*
2643 		 * Now map the supported feature vector to
2644 		 * things that we think userland will care about.
2645 		 */
2646 #if defined(__amd64)
2647 		if (*edx & CPUID_AMD_EDX_SYSC)
2648 			hwcap_flags |= AV_386_AMD_SYSC;
2649 #endif
2650 		if (*edx & CPUID_AMD_EDX_MMXamd)
2651 			hwcap_flags |= AV_386_AMD_MMX;
2652 		if (*edx & CPUID_AMD_EDX_3DNow)
2653 			hwcap_flags |= AV_386_AMD_3DNow;
2654 		if (*edx & CPUID_AMD_EDX_3DNowx)
2655 			hwcap_flags |= AV_386_AMD_3DNowx;
2656 
2657 		switch (cpi->cpi_vendor) {
2658 		case X86_VENDOR_AMD:
2659 			if (*edx & CPUID_AMD_EDX_TSCP)
2660 				hwcap_flags |= AV_386_TSCP;
2661 			if (*ecx & CPUID_AMD_ECX_AHF64)
2662 				hwcap_flags |= AV_386_AHF;
2663 			if (*ecx & CPUID_AMD_ECX_SSE4A)
2664 				hwcap_flags |= AV_386_AMD_SSE4A;
2665 			if (*ecx & CPUID_AMD_ECX_LZCNT)
2666 				hwcap_flags |= AV_386_AMD_LZCNT;
2667 			break;
2668 
2669 		case X86_VENDOR_Intel:
2670 			if (*edx & CPUID_AMD_EDX_TSCP)
2671 				hwcap_flags |= AV_386_TSCP;
2672 			/*
2673 			 * Aarrgh.
2674 			 * Intel uses a different bit in the same word.
2675 			 */
2676 			if (*ecx & CPUID_INTC_ECX_AHF64)
2677 				hwcap_flags |= AV_386_AHF;
2678 			break;
2679 
2680 		default:
2681 			break;
2682 		}
2683 		break;
2684 
2685 	case X86_VENDOR_TM:
2686 		cp.cp_eax = 0x80860001;
2687 		(void) __cpuid_insn(&cp);
2688 		cpi->cpi_support[TM_EDX_FEATURES] = cp.cp_edx;
2689 		break;
2690 
2691 	default:
2692 		break;
2693 	}
2694 
2695 pass4_done:
2696 	cpi->cpi_pass = 4;
2697 	return (hwcap_flags);
2698 }
2699 
2700 
2701 /*
2702  * Simulate the cpuid instruction using the data we previously
2703  * captured about this CPU.  We try our best to return the truth
2704  * about the hardware, independently of kernel support.
2705  */
2706 uint32_t
2707 cpuid_insn(cpu_t *cpu, struct cpuid_regs *cp)
2708 {
2709 	struct cpuid_info *cpi;
2710 	struct cpuid_regs *xcp;
2711 
2712 	if (cpu == NULL)
2713 		cpu = CPU;
2714 	cpi = cpu->cpu_m.mcpu_cpi;
2715 
2716 	ASSERT(cpuid_checkpass(cpu, 3));
2717 
2718 	/*
2719 	 * CPUID data is cached in two separate places: cpi_std for standard
2720 	 * CPUID functions, and cpi_extd for extended CPUID functions.
2721 	 */
2722 	if (cp->cp_eax <= cpi->cpi_maxeax && cp->cp_eax < NMAX_CPI_STD)
2723 		xcp = &cpi->cpi_std[cp->cp_eax];
2724 	else if (cp->cp_eax >= 0x80000000 && cp->cp_eax <= cpi->cpi_xmaxeax &&
2725 	    cp->cp_eax < 0x80000000 + NMAX_CPI_EXTD)
2726 		xcp = &cpi->cpi_extd[cp->cp_eax - 0x80000000];
2727 	else
2728 		/*
2729 		 * The caller is asking for data from an input parameter which
2730 		 * the kernel has not cached.  In this case we go fetch from
2731 		 * the hardware and return the data directly to the user.
2732 		 */
2733 		return (__cpuid_insn(cp));
2734 
2735 	cp->cp_eax = xcp->cp_eax;
2736 	cp->cp_ebx = xcp->cp_ebx;
2737 	cp->cp_ecx = xcp->cp_ecx;
2738 	cp->cp_edx = xcp->cp_edx;
2739 	return (cp->cp_eax);
2740 }
2741 
2742 int
2743 cpuid_checkpass(cpu_t *cpu, int pass)
2744 {
2745 	return (cpu != NULL && cpu->cpu_m.mcpu_cpi != NULL &&
2746 	    cpu->cpu_m.mcpu_cpi->cpi_pass >= pass);
2747 }
2748 
2749 int
2750 cpuid_getbrandstr(cpu_t *cpu, char *s, size_t n)
2751 {
2752 	ASSERT(cpuid_checkpass(cpu, 3));
2753 
2754 	return (snprintf(s, n, "%s", cpu->cpu_m.mcpu_cpi->cpi_brandstr));
2755 }
2756 
2757 int
2758 cpuid_is_cmt(cpu_t *cpu)
2759 {
2760 	if (cpu == NULL)
2761 		cpu = CPU;
2762 
2763 	ASSERT(cpuid_checkpass(cpu, 1));
2764 
2765 	return (cpu->cpu_m.mcpu_cpi->cpi_chipid >= 0);
2766 }
2767 
2768 /*
2769  * AMD and Intel both implement the 64-bit variant of the syscall
2770  * instruction (syscallq), so if there's -any- support for syscall,
2771  * cpuid currently says "yes, we support this".
2772  *
2773  * However, Intel decided to -not- implement the 32-bit variant of the
2774  * syscall instruction, so we provide a predicate to allow our caller
2775  * to test that subtlety here.
2776  *
2777  * XXPV	Currently, 32-bit syscall instructions don't work via the hypervisor,
2778  *	even in the case where the hardware would in fact support it.
2779  */
2780 /*ARGSUSED*/
2781 int
2782 cpuid_syscall32_insn(cpu_t *cpu)
2783 {
2784 	ASSERT(cpuid_checkpass((cpu == NULL ? CPU : cpu), 1));
2785 
2786 #if !defined(__xpv)
2787 	if (cpu == NULL)
2788 		cpu = CPU;
2789 
2790 	/*CSTYLED*/
2791 	{
2792 		struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2793 
2794 		if (cpi->cpi_vendor == X86_VENDOR_AMD &&
2795 		    cpi->cpi_xmaxeax >= 0x80000001 &&
2796 		    (CPI_FEATURES_XTD_EDX(cpi) & CPUID_AMD_EDX_SYSC))
2797 			return (1);
2798 	}
2799 #endif
2800 	return (0);
2801 }
2802 
2803 int
2804 cpuid_getidstr(cpu_t *cpu, char *s, size_t n)
2805 {
2806 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2807 
2808 	static const char fmt[] =
2809 	    "x86 (%s %X family %d model %d step %d clock %d MHz)";
2810 	static const char fmt_ht[] =
2811 	    "x86 (chipid 0x%x %s %X family %d model %d step %d clock %d MHz)";
2812 
2813 	ASSERT(cpuid_checkpass(cpu, 1));
2814 
2815 	if (cpuid_is_cmt(cpu))
2816 		return (snprintf(s, n, fmt_ht, cpi->cpi_chipid,
2817 		    cpi->cpi_vendorstr, cpi->cpi_std[1].cp_eax,
2818 		    cpi->cpi_family, cpi->cpi_model,
2819 		    cpi->cpi_step, cpu->cpu_type_info.pi_clock));
2820 	return (snprintf(s, n, fmt,
2821 	    cpi->cpi_vendorstr, cpi->cpi_std[1].cp_eax,
2822 	    cpi->cpi_family, cpi->cpi_model,
2823 	    cpi->cpi_step, cpu->cpu_type_info.pi_clock));
2824 }
2825 
2826 const char *
2827 cpuid_getvendorstr(cpu_t *cpu)
2828 {
2829 	ASSERT(cpuid_checkpass(cpu, 1));
2830 	return ((const char *)cpu->cpu_m.mcpu_cpi->cpi_vendorstr);
2831 }
2832 
2833 uint_t
2834 cpuid_getvendor(cpu_t *cpu)
2835 {
2836 	ASSERT(cpuid_checkpass(cpu, 1));
2837 	return (cpu->cpu_m.mcpu_cpi->cpi_vendor);
2838 }
2839 
2840 uint_t
2841 cpuid_getfamily(cpu_t *cpu)
2842 {
2843 	ASSERT(cpuid_checkpass(cpu, 1));
2844 	return (cpu->cpu_m.mcpu_cpi->cpi_family);
2845 }
2846 
2847 uint_t
2848 cpuid_getmodel(cpu_t *cpu)
2849 {
2850 	ASSERT(cpuid_checkpass(cpu, 1));
2851 	return (cpu->cpu_m.mcpu_cpi->cpi_model);
2852 }
2853 
2854 uint_t
2855 cpuid_get_ncpu_per_chip(cpu_t *cpu)
2856 {
2857 	ASSERT(cpuid_checkpass(cpu, 1));
2858 	return (cpu->cpu_m.mcpu_cpi->cpi_ncpu_per_chip);
2859 }
2860 
2861 uint_t
2862 cpuid_get_ncore_per_chip(cpu_t *cpu)
2863 {
2864 	ASSERT(cpuid_checkpass(cpu, 1));
2865 	return (cpu->cpu_m.mcpu_cpi->cpi_ncore_per_chip);
2866 }
2867 
2868 uint_t
2869 cpuid_get_ncpu_sharing_last_cache(cpu_t *cpu)
2870 {
2871 	ASSERT(cpuid_checkpass(cpu, 2));
2872 	return (cpu->cpu_m.mcpu_cpi->cpi_ncpu_shr_last_cache);
2873 }
2874 
2875 id_t
2876 cpuid_get_last_lvl_cacheid(cpu_t *cpu)
2877 {
2878 	ASSERT(cpuid_checkpass(cpu, 2));
2879 	return (cpu->cpu_m.mcpu_cpi->cpi_last_lvl_cacheid);
2880 }
2881 
2882 uint_t
2883 cpuid_getstep(cpu_t *cpu)
2884 {
2885 	ASSERT(cpuid_checkpass(cpu, 1));
2886 	return (cpu->cpu_m.mcpu_cpi->cpi_step);
2887 }
2888 
2889 uint_t
2890 cpuid_getsig(struct cpu *cpu)
2891 {
2892 	ASSERT(cpuid_checkpass(cpu, 1));
2893 	return (cpu->cpu_m.mcpu_cpi->cpi_std[1].cp_eax);
2894 }
2895 
2896 uint32_t
2897 cpuid_getchiprev(struct cpu *cpu)
2898 {
2899 	ASSERT(cpuid_checkpass(cpu, 1));
2900 	return (cpu->cpu_m.mcpu_cpi->cpi_chiprev);
2901 }
2902 
2903 const char *
2904 cpuid_getchiprevstr(struct cpu *cpu)
2905 {
2906 	ASSERT(cpuid_checkpass(cpu, 1));
2907 	return (cpu->cpu_m.mcpu_cpi->cpi_chiprevstr);
2908 }
2909 
2910 uint32_t
2911 cpuid_getsockettype(struct cpu *cpu)
2912 {
2913 	ASSERT(cpuid_checkpass(cpu, 1));
2914 	return (cpu->cpu_m.mcpu_cpi->cpi_socket);
2915 }
2916 
2917 const char *
2918 cpuid_getsocketstr(cpu_t *cpu)
2919 {
2920 	static const char *socketstr = NULL;
2921 	struct cpuid_info *cpi;
2922 
2923 	ASSERT(cpuid_checkpass(cpu, 1));
2924 	cpi = cpu->cpu_m.mcpu_cpi;
2925 
2926 	/* Assume that socket types are the same across the system */
2927 	if (socketstr == NULL)
2928 		socketstr = _cpuid_sktstr(cpi->cpi_vendor, cpi->cpi_family,
2929 		    cpi->cpi_model, cpi->cpi_step);
2930 
2931 
2932 	return (socketstr);
2933 }
2934 
2935 int
2936 cpuid_get_chipid(cpu_t *cpu)
2937 {
2938 	ASSERT(cpuid_checkpass(cpu, 1));
2939 
2940 	if (cpuid_is_cmt(cpu))
2941 		return (cpu->cpu_m.mcpu_cpi->cpi_chipid);
2942 	return (cpu->cpu_id);
2943 }
2944 
2945 id_t
2946 cpuid_get_coreid(cpu_t *cpu)
2947 {
2948 	ASSERT(cpuid_checkpass(cpu, 1));
2949 	return (cpu->cpu_m.mcpu_cpi->cpi_coreid);
2950 }
2951 
2952 int
2953 cpuid_get_pkgcoreid(cpu_t *cpu)
2954 {
2955 	ASSERT(cpuid_checkpass(cpu, 1));
2956 	return (cpu->cpu_m.mcpu_cpi->cpi_pkgcoreid);
2957 }
2958 
2959 int
2960 cpuid_get_clogid(cpu_t *cpu)
2961 {
2962 	ASSERT(cpuid_checkpass(cpu, 1));
2963 	return (cpu->cpu_m.mcpu_cpi->cpi_clogid);
2964 }
2965 
2966 int
2967 cpuid_get_cacheid(cpu_t *cpu)
2968 {
2969 	ASSERT(cpuid_checkpass(cpu, 1));
2970 	return (cpu->cpu_m.mcpu_cpi->cpi_last_lvl_cacheid);
2971 }
2972 
2973 uint_t
2974 cpuid_get_procnodeid(cpu_t *cpu)
2975 {
2976 	ASSERT(cpuid_checkpass(cpu, 1));
2977 	return (cpu->cpu_m.mcpu_cpi->cpi_procnodeid);
2978 }
2979 
2980 uint_t
2981 cpuid_get_procnodes_per_pkg(cpu_t *cpu)
2982 {
2983 	ASSERT(cpuid_checkpass(cpu, 1));
2984 	return (cpu->cpu_m.mcpu_cpi->cpi_procnodes_per_pkg);
2985 }
2986 
2987 /*ARGSUSED*/
2988 int
2989 cpuid_have_cr8access(cpu_t *cpu)
2990 {
2991 #if defined(__amd64)
2992 	return (1);
2993 #else
2994 	struct cpuid_info *cpi;
2995 
2996 	ASSERT(cpu != NULL);
2997 	cpi = cpu->cpu_m.mcpu_cpi;
2998 	if (cpi->cpi_vendor == X86_VENDOR_AMD && cpi->cpi_maxeax >= 1 &&
2999 	    (CPI_FEATURES_XTD_ECX(cpi) & CPUID_AMD_ECX_CR8D) != 0)
3000 		return (1);
3001 	return (0);
3002 #endif
3003 }
3004 
3005 uint32_t
3006 cpuid_get_apicid(cpu_t *cpu)
3007 {
3008 	ASSERT(cpuid_checkpass(cpu, 1));
3009 	if (cpu->cpu_m.mcpu_cpi->cpi_maxeax < 1) {
3010 		return (UINT32_MAX);
3011 	} else {
3012 		return (cpu->cpu_m.mcpu_cpi->cpi_apicid);
3013 	}
3014 }
3015 
3016 void
3017 cpuid_get_addrsize(cpu_t *cpu, uint_t *pabits, uint_t *vabits)
3018 {
3019 	struct cpuid_info *cpi;
3020 
3021 	if (cpu == NULL)
3022 		cpu = CPU;
3023 	cpi = cpu->cpu_m.mcpu_cpi;
3024 
3025 	ASSERT(cpuid_checkpass(cpu, 1));
3026 
3027 	if (pabits)
3028 		*pabits = cpi->cpi_pabits;
3029 	if (vabits)
3030 		*vabits = cpi->cpi_vabits;
3031 }
3032 
3033 /*
3034  * Returns the number of data TLB entries for a corresponding
3035  * pagesize.  If it can't be computed, or isn't known, the
3036  * routine returns zero.  If you ask about an architecturally
3037  * impossible pagesize, the routine will panic (so that the
3038  * hat implementor knows that things are inconsistent.)
3039  */
3040 uint_t
3041 cpuid_get_dtlb_nent(cpu_t *cpu, size_t pagesize)
3042 {
3043 	struct cpuid_info *cpi;
3044 	uint_t dtlb_nent = 0;
3045 
3046 	if (cpu == NULL)
3047 		cpu = CPU;
3048 	cpi = cpu->cpu_m.mcpu_cpi;
3049 
3050 	ASSERT(cpuid_checkpass(cpu, 1));
3051 
3052 	/*
3053 	 * Check the L2 TLB info
3054 	 */
3055 	if (cpi->cpi_xmaxeax >= 0x80000006) {
3056 		struct cpuid_regs *cp = &cpi->cpi_extd[6];
3057 
3058 		switch (pagesize) {
3059 
3060 		case 4 * 1024:
3061 			/*
3062 			 * All zero in the top 16 bits of the register
3063 			 * indicates a unified TLB. Size is in low 16 bits.
3064 			 */
3065 			if ((cp->cp_ebx & 0xffff0000) == 0)
3066 				dtlb_nent = cp->cp_ebx & 0x0000ffff;
3067 			else
3068 				dtlb_nent = BITX(cp->cp_ebx, 27, 16);
3069 			break;
3070 
3071 		case 2 * 1024 * 1024:
3072 			if ((cp->cp_eax & 0xffff0000) == 0)
3073 				dtlb_nent = cp->cp_eax & 0x0000ffff;
3074 			else
3075 				dtlb_nent = BITX(cp->cp_eax, 27, 16);
3076 			break;
3077 
3078 		default:
3079 			panic("unknown L2 pagesize");
3080 			/*NOTREACHED*/
3081 		}
3082 	}
3083 
3084 	if (dtlb_nent != 0)
3085 		return (dtlb_nent);
3086 
3087 	/*
3088 	 * No L2 TLB support for this size, try L1.
3089 	 */
3090 	if (cpi->cpi_xmaxeax >= 0x80000005) {
3091 		struct cpuid_regs *cp = &cpi->cpi_extd[5];
3092 
3093 		switch (pagesize) {
3094 		case 4 * 1024:
3095 			dtlb_nent = BITX(cp->cp_ebx, 23, 16);
3096 			break;
3097 		case 2 * 1024 * 1024:
3098 			dtlb_nent = BITX(cp->cp_eax, 23, 16);
3099 			break;
3100 		default:
3101 			panic("unknown L1 d-TLB pagesize");
3102 			/*NOTREACHED*/
3103 		}
3104 	}
3105 
3106 	return (dtlb_nent);
3107 }
3108 
3109 /*
3110  * Return 0 if the erratum is not present or not applicable, positive
3111  * if it is, and negative if the status of the erratum is unknown.
3112  *
3113  * See "Revision Guide for AMD Athlon(tm) 64 and AMD Opteron(tm)
3114  * Processors" #25759, Rev 3.57, August 2005
3115  */
3116 int
3117 cpuid_opteron_erratum(cpu_t *cpu, uint_t erratum)
3118 {
3119 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
3120 	uint_t eax;
3121 
3122 	/*
3123 	 * Bail out if this CPU isn't an AMD CPU, or if it's
3124 	 * a legacy (32-bit) AMD CPU.
3125 	 */
3126 	if (cpi->cpi_vendor != X86_VENDOR_AMD ||
3127 	    cpi->cpi_family == 4 || cpi->cpi_family == 5 ||
3128 	    cpi->cpi_family == 6)
3129 
3130 		return (0);
3131 
3132 	eax = cpi->cpi_std[1].cp_eax;
3133 
3134 #define	SH_B0(eax)	(eax == 0xf40 || eax == 0xf50)
3135 #define	SH_B3(eax) 	(eax == 0xf51)
3136 #define	B(eax)		(SH_B0(eax) || SH_B3(eax))
3137 
3138 #define	SH_C0(eax)	(eax == 0xf48 || eax == 0xf58)
3139 
3140 #define	SH_CG(eax)	(eax == 0xf4a || eax == 0xf5a || eax == 0xf7a)
3141 #define	DH_CG(eax)	(eax == 0xfc0 || eax == 0xfe0 || eax == 0xff0)
3142 #define	CH_CG(eax)	(eax == 0xf82 || eax == 0xfb2)
3143 #define	CG(eax)		(SH_CG(eax) || DH_CG(eax) || CH_CG(eax))
3144 
3145 #define	SH_D0(eax)	(eax == 0x10f40 || eax == 0x10f50 || eax == 0x10f70)
3146 #define	DH_D0(eax)	(eax == 0x10fc0 || eax == 0x10ff0)
3147 #define	CH_D0(eax)	(eax == 0x10f80 || eax == 0x10fb0)
3148 #define	D0(eax)		(SH_D0(eax) || DH_D0(eax) || CH_D0(eax))
3149 
3150 #define	SH_E0(eax)	(eax == 0x20f50 || eax == 0x20f40 || eax == 0x20f70)
3151 #define	JH_E1(eax)	(eax == 0x20f10)	/* JH8_E0 had 0x20f30 */
3152 #define	DH_E3(eax)	(eax == 0x20fc0 || eax == 0x20ff0)
3153 #define	SH_E4(eax)	(eax == 0x20f51 || eax == 0x20f71)
3154 #define	BH_E4(eax)	(eax == 0x20fb1)
3155 #define	SH_E5(eax)	(eax == 0x20f42)
3156 #define	DH_E6(eax)	(eax == 0x20ff2 || eax == 0x20fc2)
3157 #define	JH_E6(eax)	(eax == 0x20f12 || eax == 0x20f32)
3158 #define	EX(eax)		(SH_E0(eax) || JH_E1(eax) || DH_E3(eax) || \
3159 			    SH_E4(eax) || BH_E4(eax) || SH_E5(eax) || \
3160 			    DH_E6(eax) || JH_E6(eax))
3161 
3162 #define	DR_AX(eax)	(eax == 0x100f00 || eax == 0x100f01 || eax == 0x100f02)
3163 #define	DR_B0(eax)	(eax == 0x100f20)
3164 #define	DR_B1(eax)	(eax == 0x100f21)
3165 #define	DR_BA(eax)	(eax == 0x100f2a)
3166 #define	DR_B2(eax)	(eax == 0x100f22)
3167 #define	DR_B3(eax)	(eax == 0x100f23)
3168 #define	RB_C0(eax)	(eax == 0x100f40)
3169 
3170 	switch (erratum) {
3171 	case 1:
3172 		return (cpi->cpi_family < 0x10);
3173 	case 51:	/* what does the asterisk mean? */
3174 		return (B(eax) || SH_C0(eax) || CG(eax));
3175 	case 52:
3176 		return (B(eax));
3177 	case 57:
3178 		return (cpi->cpi_family <= 0x11);
3179 	case 58:
3180 		return (B(eax));
3181 	case 60:
3182 		return (cpi->cpi_family <= 0x11);
3183 	case 61:
3184 	case 62:
3185 	case 63:
3186 	case 64:
3187 	case 65:
3188 	case 66:
3189 	case 68:
3190 	case 69:
3191 	case 70:
3192 	case 71:
3193 		return (B(eax));
3194 	case 72:
3195 		return (SH_B0(eax));
3196 	case 74:
3197 		return (B(eax));
3198 	case 75:
3199 		return (cpi->cpi_family < 0x10);
3200 	case 76:
3201 		return (B(eax));
3202 	case 77:
3203 		return (cpi->cpi_family <= 0x11);
3204 	case 78:
3205 		return (B(eax) || SH_C0(eax));
3206 	case 79:
3207 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
3208 	case 80:
3209 	case 81:
3210 	case 82:
3211 		return (B(eax));
3212 	case 83:
3213 		return (B(eax) || SH_C0(eax) || CG(eax));
3214 	case 85:
3215 		return (cpi->cpi_family < 0x10);
3216 	case 86:
3217 		return (SH_C0(eax) || CG(eax));
3218 	case 88:
3219 #if !defined(__amd64)
3220 		return (0);
3221 #else
3222 		return (B(eax) || SH_C0(eax));
3223 #endif
3224 	case 89:
3225 		return (cpi->cpi_family < 0x10);
3226 	case 90:
3227 		return (B(eax) || SH_C0(eax) || CG(eax));
3228 	case 91:
3229 	case 92:
3230 		return (B(eax) || SH_C0(eax));
3231 	case 93:
3232 		return (SH_C0(eax));
3233 	case 94:
3234 		return (B(eax) || SH_C0(eax) || CG(eax));
3235 	case 95:
3236 #if !defined(__amd64)
3237 		return (0);
3238 #else
3239 		return (B(eax) || SH_C0(eax));
3240 #endif
3241 	case 96:
3242 		return (B(eax) || SH_C0(eax) || CG(eax));
3243 	case 97:
3244 	case 98:
3245 		return (SH_C0(eax) || CG(eax));
3246 	case 99:
3247 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
3248 	case 100:
3249 		return (B(eax) || SH_C0(eax));
3250 	case 101:
3251 	case 103:
3252 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
3253 	case 104:
3254 		return (SH_C0(eax) || CG(eax) || D0(eax));
3255 	case 105:
3256 	case 106:
3257 	case 107:
3258 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
3259 	case 108:
3260 		return (DH_CG(eax));
3261 	case 109:
3262 		return (SH_C0(eax) || CG(eax) || D0(eax));
3263 	case 110:
3264 		return (D0(eax) || EX(eax));
3265 	case 111:
3266 		return (CG(eax));
3267 	case 112:
3268 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
3269 	case 113:
3270 		return (eax == 0x20fc0);
3271 	case 114:
3272 		return (SH_E0(eax) || JH_E1(eax) || DH_E3(eax));
3273 	case 115:
3274 		return (SH_E0(eax) || JH_E1(eax));
3275 	case 116:
3276 		return (SH_E0(eax) || JH_E1(eax) || DH_E3(eax));
3277 	case 117:
3278 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
3279 	case 118:
3280 		return (SH_E0(eax) || JH_E1(eax) || SH_E4(eax) || BH_E4(eax) ||
3281 		    JH_E6(eax));
3282 	case 121:
3283 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
3284 	case 122:
3285 		return (cpi->cpi_family < 0x10 || cpi->cpi_family == 0x11);
3286 	case 123:
3287 		return (JH_E1(eax) || BH_E4(eax) || JH_E6(eax));
3288 	case 131:
3289 		return (cpi->cpi_family < 0x10);
3290 	case 6336786:
3291 		/*
3292 		 * Test for AdvPowerMgmtInfo.TscPStateInvariant
3293 		 * if this is a K8 family or newer processor
3294 		 */
3295 		if (CPI_FAMILY(cpi) == 0xf) {
3296 			struct cpuid_regs regs;
3297 			regs.cp_eax = 0x80000007;
3298 			(void) __cpuid_insn(&regs);
3299 			return (!(regs.cp_edx & 0x100));
3300 		}
3301 		return (0);
3302 	case 6323525:
3303 		return (((((eax >> 12) & 0xff00) + (eax & 0xf00)) |
3304 		    (((eax >> 4) & 0xf) | ((eax >> 12) & 0xf0))) < 0xf40);
3305 
3306 	case 6671130:
3307 		/*
3308 		 * check for processors (pre-Shanghai) that do not provide
3309 		 * optimal management of 1gb ptes in its tlb.
3310 		 */
3311 		return (cpi->cpi_family == 0x10 && cpi->cpi_model < 4);
3312 
3313 	case 298:
3314 		return (DR_AX(eax) || DR_B0(eax) || DR_B1(eax) || DR_BA(eax) ||
3315 		    DR_B2(eax) || RB_C0(eax));
3316 
3317 	default:
3318 		return (-1);
3319 
3320 	}
3321 }
3322 
3323 /*
3324  * Determine if specified erratum is present via OSVW (OS Visible Workaround).
3325  * Return 1 if erratum is present, 0 if not present and -1 if indeterminate.
3326  */
3327 int
3328 osvw_opteron_erratum(cpu_t *cpu, uint_t erratum)
3329 {
3330 	struct cpuid_info	*cpi;
3331 	uint_t			osvwid;
3332 	static int		osvwfeature = -1;
3333 	uint64_t		osvwlength;
3334 
3335 
3336 	cpi = cpu->cpu_m.mcpu_cpi;
3337 
3338 	/* confirm OSVW supported */
3339 	if (osvwfeature == -1) {
3340 		osvwfeature = cpi->cpi_extd[1].cp_ecx & CPUID_AMD_ECX_OSVW;
3341 	} else {
3342 		/* assert that osvw feature setting is consistent on all cpus */
3343 		ASSERT(osvwfeature ==
3344 		    (cpi->cpi_extd[1].cp_ecx & CPUID_AMD_ECX_OSVW));
3345 	}
3346 	if (!osvwfeature)
3347 		return (-1);
3348 
3349 	osvwlength = rdmsr(MSR_AMD_OSVW_ID_LEN) & OSVW_ID_LEN_MASK;
3350 
3351 	switch (erratum) {
3352 	case 298:	/* osvwid is 0 */
3353 		osvwid = 0;
3354 		if (osvwlength <= (uint64_t)osvwid) {
3355 			/* osvwid 0 is unknown */
3356 			return (-1);
3357 		}
3358 
3359 		/*
3360 		 * Check the OSVW STATUS MSR to determine the state
3361 		 * of the erratum where:
3362 		 *   0 - fixed by HW
3363 		 *   1 - BIOS has applied the workaround when BIOS
3364 		 *   workaround is available. (Or for other errata,
3365 		 *   OS workaround is required.)
3366 		 * For a value of 1, caller will confirm that the
3367 		 * erratum 298 workaround has indeed been applied by BIOS.
3368 		 *
3369 		 * A 1 may be set in cpus that have a HW fix
3370 		 * in a mixed cpu system. Regarding erratum 298:
3371 		 *   In a multiprocessor platform, the workaround above
3372 		 *   should be applied to all processors regardless of
3373 		 *   silicon revision when an affected processor is
3374 		 *   present.
3375 		 */
3376 
3377 		return (rdmsr(MSR_AMD_OSVW_STATUS +
3378 		    (osvwid / OSVW_ID_CNT_PER_MSR)) &
3379 		    (1ULL << (osvwid % OSVW_ID_CNT_PER_MSR)));
3380 
3381 	default:
3382 		return (-1);
3383 	}
3384 }
3385 
3386 static const char assoc_str[] = "associativity";
3387 static const char line_str[] = "line-size";
3388 static const char size_str[] = "size";
3389 
3390 static void
3391 add_cache_prop(dev_info_t *devi, const char *label, const char *type,
3392     uint32_t val)
3393 {
3394 	char buf[128];
3395 
3396 	/*
3397 	 * ndi_prop_update_int() is used because it is desirable for
3398 	 * DDI_PROP_HW_DEF and DDI_PROP_DONTSLEEP to be set.
3399 	 */
3400 	if (snprintf(buf, sizeof (buf), "%s-%s", label, type) < sizeof (buf))
3401 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, devi, buf, val);
3402 }
3403 
3404 /*
3405  * Intel-style cache/tlb description
3406  *
3407  * Standard cpuid level 2 gives a randomly ordered
3408  * selection of tags that index into a table that describes
3409  * cache and tlb properties.
3410  */
3411 
3412 static const char l1_icache_str[] = "l1-icache";
3413 static const char l1_dcache_str[] = "l1-dcache";
3414 static const char l2_cache_str[] = "l2-cache";
3415 static const char l3_cache_str[] = "l3-cache";
3416 static const char itlb4k_str[] = "itlb-4K";
3417 static const char dtlb4k_str[] = "dtlb-4K";
3418 static const char itlb2M_str[] = "itlb-2M";
3419 static const char itlb4M_str[] = "itlb-4M";
3420 static const char dtlb4M_str[] = "dtlb-4M";
3421 static const char dtlb24_str[] = "dtlb0-2M-4M";
3422 static const char itlb424_str[] = "itlb-4K-2M-4M";
3423 static const char itlb24_str[] = "itlb-2M-4M";
3424 static const char dtlb44_str[] = "dtlb-4K-4M";
3425 static const char sl1_dcache_str[] = "sectored-l1-dcache";
3426 static const char sl2_cache_str[] = "sectored-l2-cache";
3427 static const char itrace_str[] = "itrace-cache";
3428 static const char sl3_cache_str[] = "sectored-l3-cache";
3429 static const char sh_l2_tlb4k_str[] = "shared-l2-tlb-4k";
3430 
3431 static const struct cachetab {
3432 	uint8_t 	ct_code;
3433 	uint8_t		ct_assoc;
3434 	uint16_t 	ct_line_size;
3435 	size_t		ct_size;
3436 	const char	*ct_label;
3437 } intel_ctab[] = {
3438 	/*
3439 	 * maintain descending order!
3440 	 *
3441 	 * Codes ignored - Reason
3442 	 * ----------------------
3443 	 * 40H - intel_cpuid_4_cache_info() disambiguates l2/l3 cache
3444 	 * f0H/f1H - Currently we do not interpret prefetch size by design
3445 	 */
3446 	{ 0xe4, 16, 64, 8*1024*1024, l3_cache_str},
3447 	{ 0xe3, 16, 64, 4*1024*1024, l3_cache_str},
3448 	{ 0xe2, 16, 64, 2*1024*1024, l3_cache_str},
3449 	{ 0xde, 12, 64, 6*1024*1024, l3_cache_str},
3450 	{ 0xdd, 12, 64, 3*1024*1024, l3_cache_str},
3451 	{ 0xdc, 12, 64, ((1*1024*1024)+(512*1024)), l3_cache_str},
3452 	{ 0xd8, 8, 64, 4*1024*1024, l3_cache_str},
3453 	{ 0xd7, 8, 64, 2*1024*1024, l3_cache_str},
3454 	{ 0xd6, 8, 64, 1*1024*1024, l3_cache_str},
3455 	{ 0xd2, 4, 64, 2*1024*1024, l3_cache_str},
3456 	{ 0xd1, 4, 64, 1*1024*1024, l3_cache_str},
3457 	{ 0xd0, 4, 64, 512*1024, l3_cache_str},
3458 	{ 0xca, 4, 0, 512, sh_l2_tlb4k_str},
3459 	{ 0xc0, 4, 0, 8, dtlb44_str },
3460 	{ 0xba, 4, 0, 64, dtlb4k_str },
3461 	{ 0xb4, 4, 0, 256, dtlb4k_str },
3462 	{ 0xb3, 4, 0, 128, dtlb4k_str },
3463 	{ 0xb2, 4, 0, 64, itlb4k_str },
3464 	{ 0xb0, 4, 0, 128, itlb4k_str },
3465 	{ 0x87, 8, 64, 1024*1024, l2_cache_str},
3466 	{ 0x86, 4, 64, 512*1024, l2_cache_str},
3467 	{ 0x85, 8, 32, 2*1024*1024, l2_cache_str},
3468 	{ 0x84, 8, 32, 1024*1024, l2_cache_str},
3469 	{ 0x83, 8, 32, 512*1024, l2_cache_str},
3470 	{ 0x82, 8, 32, 256*1024, l2_cache_str},
3471 	{ 0x80, 8, 64, 512*1024, l2_cache_str},
3472 	{ 0x7f, 2, 64, 512*1024, l2_cache_str},
3473 	{ 0x7d, 8, 64, 2*1024*1024, sl2_cache_str},
3474 	{ 0x7c, 8, 64, 1024*1024, sl2_cache_str},
3475 	{ 0x7b, 8, 64, 512*1024, sl2_cache_str},
3476 	{ 0x7a, 8, 64, 256*1024, sl2_cache_str},
3477 	{ 0x79, 8, 64, 128*1024, sl2_cache_str},
3478 	{ 0x78, 8, 64, 1024*1024, l2_cache_str},
3479 	{ 0x73, 8, 0, 64*1024, itrace_str},
3480 	{ 0x72, 8, 0, 32*1024, itrace_str},
3481 	{ 0x71, 8, 0, 16*1024, itrace_str},
3482 	{ 0x70, 8, 0, 12*1024, itrace_str},
3483 	{ 0x68, 4, 64, 32*1024, sl1_dcache_str},
3484 	{ 0x67, 4, 64, 16*1024, sl1_dcache_str},
3485 	{ 0x66, 4, 64, 8*1024, sl1_dcache_str},
3486 	{ 0x60, 8, 64, 16*1024, sl1_dcache_str},
3487 	{ 0x5d, 0, 0, 256, dtlb44_str},
3488 	{ 0x5c, 0, 0, 128, dtlb44_str},
3489 	{ 0x5b, 0, 0, 64, dtlb44_str},
3490 	{ 0x5a, 4, 0, 32, dtlb24_str},
3491 	{ 0x59, 0, 0, 16, dtlb4k_str},
3492 	{ 0x57, 4, 0, 16, dtlb4k_str},
3493 	{ 0x56, 4, 0, 16, dtlb4M_str},
3494 	{ 0x55, 0, 0, 7, itlb24_str},
3495 	{ 0x52, 0, 0, 256, itlb424_str},
3496 	{ 0x51, 0, 0, 128, itlb424_str},
3497 	{ 0x50, 0, 0, 64, itlb424_str},
3498 	{ 0x4f, 0, 0, 32, itlb4k_str},
3499 	{ 0x4e, 24, 64, 6*1024*1024, l2_cache_str},
3500 	{ 0x4d, 16, 64, 16*1024*1024, l3_cache_str},
3501 	{ 0x4c, 12, 64, 12*1024*1024, l3_cache_str},
3502 	{ 0x4b, 16, 64, 8*1024*1024, l3_cache_str},
3503 	{ 0x4a, 12, 64, 6*1024*1024, l3_cache_str},
3504 	{ 0x49, 16, 64, 4*1024*1024, l3_cache_str},
3505 	{ 0x48, 12, 64, 3*1024*1024, l2_cache_str},
3506 	{ 0x47, 8, 64, 8*1024*1024, l3_cache_str},
3507 	{ 0x46, 4, 64, 4*1024*1024, l3_cache_str},
3508 	{ 0x45, 4, 32, 2*1024*1024, l2_cache_str},
3509 	{ 0x44, 4, 32, 1024*1024, l2_cache_str},
3510 	{ 0x43, 4, 32, 512*1024, l2_cache_str},
3511 	{ 0x42, 4, 32, 256*1024, l2_cache_str},
3512 	{ 0x41, 4, 32, 128*1024, l2_cache_str},
3513 	{ 0x3e, 4, 64, 512*1024, sl2_cache_str},
3514 	{ 0x3d, 6, 64, 384*1024, sl2_cache_str},
3515 	{ 0x3c, 4, 64, 256*1024, sl2_cache_str},
3516 	{ 0x3b, 2, 64, 128*1024, sl2_cache_str},
3517 	{ 0x3a, 6, 64, 192*1024, sl2_cache_str},
3518 	{ 0x39, 4, 64, 128*1024, sl2_cache_str},
3519 	{ 0x30, 8, 64, 32*1024, l1_icache_str},
3520 	{ 0x2c, 8, 64, 32*1024, l1_dcache_str},
3521 	{ 0x29, 8, 64, 4096*1024, sl3_cache_str},
3522 	{ 0x25, 8, 64, 2048*1024, sl3_cache_str},
3523 	{ 0x23, 8, 64, 1024*1024, sl3_cache_str},
3524 	{ 0x22, 4, 64, 512*1024, sl3_cache_str},
3525 	{ 0x0e, 6, 64, 24*1024, l1_dcache_str},
3526 	{ 0x0d, 4, 32, 16*1024, l1_dcache_str},
3527 	{ 0x0c, 4, 32, 16*1024, l1_dcache_str},
3528 	{ 0x0b, 4, 0, 4, itlb4M_str},
3529 	{ 0x0a, 2, 32, 8*1024, l1_dcache_str},
3530 	{ 0x08, 4, 32, 16*1024, l1_icache_str},
3531 	{ 0x06, 4, 32, 8*1024, l1_icache_str},
3532 	{ 0x05, 4, 0, 32, dtlb4M_str},
3533 	{ 0x04, 4, 0, 8, dtlb4M_str},
3534 	{ 0x03, 4, 0, 64, dtlb4k_str},
3535 	{ 0x02, 4, 0, 2, itlb4M_str},
3536 	{ 0x01, 4, 0, 32, itlb4k_str},
3537 	{ 0 }
3538 };
3539 
3540 static const struct cachetab cyrix_ctab[] = {
3541 	{ 0x70, 4, 0, 32, "tlb-4K" },
3542 	{ 0x80, 4, 16, 16*1024, "l1-cache" },
3543 	{ 0 }
3544 };
3545 
3546 /*
3547  * Search a cache table for a matching entry
3548  */
3549 static const struct cachetab *
3550 find_cacheent(const struct cachetab *ct, uint_t code)
3551 {
3552 	if (code != 0) {
3553 		for (; ct->ct_code != 0; ct++)
3554 			if (ct->ct_code <= code)
3555 				break;
3556 		if (ct->ct_code == code)
3557 			return (ct);
3558 	}
3559 	return (NULL);
3560 }
3561 
3562 /*
3563  * Populate cachetab entry with L2 or L3 cache-information using
3564  * cpuid function 4. This function is called from intel_walk_cacheinfo()
3565  * when descriptor 0x49 is encountered. It returns 0 if no such cache
3566  * information is found.
3567  */
3568 static int
3569 intel_cpuid_4_cache_info(struct cachetab *ct, struct cpuid_info *cpi)
3570 {
3571 	uint32_t level, i;
3572 	int ret = 0;
3573 
3574 	for (i = 0; i < cpi->cpi_std_4_size; i++) {
3575 		level = CPI_CACHE_LVL(cpi->cpi_std_4[i]);
3576 
3577 		if (level == 2 || level == 3) {
3578 			ct->ct_assoc = CPI_CACHE_WAYS(cpi->cpi_std_4[i]) + 1;
3579 			ct->ct_line_size =
3580 			    CPI_CACHE_COH_LN_SZ(cpi->cpi_std_4[i]) + 1;
3581 			ct->ct_size = ct->ct_assoc *
3582 			    (CPI_CACHE_PARTS(cpi->cpi_std_4[i]) + 1) *
3583 			    ct->ct_line_size *
3584 			    (cpi->cpi_std_4[i]->cp_ecx + 1);
3585 
3586 			if (level == 2) {
3587 				ct->ct_label = l2_cache_str;
3588 			} else if (level == 3) {
3589 				ct->ct_label = l3_cache_str;
3590 			}
3591 			ret = 1;
3592 		}
3593 	}
3594 
3595 	return (ret);
3596 }
3597 
3598 /*
3599  * Walk the cacheinfo descriptor, applying 'func' to every valid element
3600  * The walk is terminated if the walker returns non-zero.
3601  */
3602 static void
3603 intel_walk_cacheinfo(struct cpuid_info *cpi,
3604     void *arg, int (*func)(void *, const struct cachetab *))
3605 {
3606 	const struct cachetab *ct;
3607 	struct cachetab des_49_ct, des_b1_ct;
3608 	uint8_t *dp;
3609 	int i;
3610 
3611 	if ((dp = cpi->cpi_cacheinfo) == NULL)
3612 		return;
3613 	for (i = 0; i < cpi->cpi_ncache; i++, dp++) {
3614 		/*
3615 		 * For overloaded descriptor 0x49 we use cpuid function 4
3616 		 * if supported by the current processor, to create
3617 		 * cache information.
3618 		 * For overloaded descriptor 0xb1 we use X86_PAE flag
3619 		 * to disambiguate the cache information.
3620 		 */
3621 		if (*dp == 0x49 && cpi->cpi_maxeax >= 0x4 &&
3622 		    intel_cpuid_4_cache_info(&des_49_ct, cpi) == 1) {
3623 				ct = &des_49_ct;
3624 		} else if (*dp == 0xb1) {
3625 			des_b1_ct.ct_code = 0xb1;
3626 			des_b1_ct.ct_assoc = 4;
3627 			des_b1_ct.ct_line_size = 0;
3628 			if (is_x86_feature(x86_featureset, X86FSET_PAE)) {
3629 				des_b1_ct.ct_size = 8;
3630 				des_b1_ct.ct_label = itlb2M_str;
3631 			} else {
3632 				des_b1_ct.ct_size = 4;
3633 				des_b1_ct.ct_label = itlb4M_str;
3634 			}
3635 			ct = &des_b1_ct;
3636 		} else {
3637 			if ((ct = find_cacheent(intel_ctab, *dp)) == NULL) {
3638 				continue;
3639 			}
3640 		}
3641 
3642 		if (func(arg, ct) != 0) {
3643 			break;
3644 		}
3645 	}
3646 }
3647 
3648 /*
3649  * (Like the Intel one, except for Cyrix CPUs)
3650  */
3651 static void
3652 cyrix_walk_cacheinfo(struct cpuid_info *cpi,
3653     void *arg, int (*func)(void *, const struct cachetab *))
3654 {
3655 	const struct cachetab *ct;
3656 	uint8_t *dp;
3657 	int i;
3658 
3659 	if ((dp = cpi->cpi_cacheinfo) == NULL)
3660 		return;
3661 	for (i = 0; i < cpi->cpi_ncache; i++, dp++) {
3662 		/*
3663 		 * Search Cyrix-specific descriptor table first ..
3664 		 */
3665 		if ((ct = find_cacheent(cyrix_ctab, *dp)) != NULL) {
3666 			if (func(arg, ct) != 0)
3667 				break;
3668 			continue;
3669 		}
3670 		/*
3671 		 * .. else fall back to the Intel one
3672 		 */
3673 		if ((ct = find_cacheent(intel_ctab, *dp)) != NULL) {
3674 			if (func(arg, ct) != 0)
3675 				break;
3676 			continue;
3677 		}
3678 	}
3679 }
3680 
3681 /*
3682  * A cacheinfo walker that adds associativity, line-size, and size properties
3683  * to the devinfo node it is passed as an argument.
3684  */
3685 static int
3686 add_cacheent_props(void *arg, const struct cachetab *ct)
3687 {
3688 	dev_info_t *devi = arg;
3689 
3690 	add_cache_prop(devi, ct->ct_label, assoc_str, ct->ct_assoc);
3691 	if (ct->ct_line_size != 0)
3692 		add_cache_prop(devi, ct->ct_label, line_str,
3693 		    ct->ct_line_size);
3694 	add_cache_prop(devi, ct->ct_label, size_str, ct->ct_size);
3695 	return (0);
3696 }
3697 
3698 
3699 static const char fully_assoc[] = "fully-associative?";
3700 
3701 /*
3702  * AMD style cache/tlb description
3703  *
3704  * Extended functions 5 and 6 directly describe properties of
3705  * tlbs and various cache levels.
3706  */
3707 static void
3708 add_amd_assoc(dev_info_t *devi, const char *label, uint_t assoc)
3709 {
3710 	switch (assoc) {
3711 	case 0:	/* reserved; ignore */
3712 		break;
3713 	default:
3714 		add_cache_prop(devi, label, assoc_str, assoc);
3715 		break;
3716 	case 0xff:
3717 		add_cache_prop(devi, label, fully_assoc, 1);
3718 		break;
3719 	}
3720 }
3721 
3722 static void
3723 add_amd_tlb(dev_info_t *devi, const char *label, uint_t assoc, uint_t size)
3724 {
3725 	if (size == 0)
3726 		return;
3727 	add_cache_prop(devi, label, size_str, size);
3728 	add_amd_assoc(devi, label, assoc);
3729 }
3730 
3731 static void
3732 add_amd_cache(dev_info_t *devi, const char *label,
3733     uint_t size, uint_t assoc, uint_t lines_per_tag, uint_t line_size)
3734 {
3735 	if (size == 0 || line_size == 0)
3736 		return;
3737 	add_amd_assoc(devi, label, assoc);
3738 	/*
3739 	 * Most AMD parts have a sectored cache. Multiple cache lines are
3740 	 * associated with each tag. A sector consists of all cache lines
3741 	 * associated with a tag. For example, the AMD K6-III has a sector
3742 	 * size of 2 cache lines per tag.
3743 	 */
3744 	if (lines_per_tag != 0)
3745 		add_cache_prop(devi, label, "lines-per-tag", lines_per_tag);
3746 	add_cache_prop(devi, label, line_str, line_size);
3747 	add_cache_prop(devi, label, size_str, size * 1024);
3748 }
3749 
3750 static void
3751 add_amd_l2_assoc(dev_info_t *devi, const char *label, uint_t assoc)
3752 {
3753 	switch (assoc) {
3754 	case 0:	/* off */
3755 		break;
3756 	case 1:
3757 	case 2:
3758 	case 4:
3759 		add_cache_prop(devi, label, assoc_str, assoc);
3760 		break;
3761 	case 6:
3762 		add_cache_prop(devi, label, assoc_str, 8);
3763 		break;
3764 	case 8:
3765 		add_cache_prop(devi, label, assoc_str, 16);
3766 		break;
3767 	case 0xf:
3768 		add_cache_prop(devi, label, fully_assoc, 1);
3769 		break;
3770 	default: /* reserved; ignore */
3771 		break;
3772 	}
3773 }
3774 
3775 static void
3776 add_amd_l2_tlb(dev_info_t *devi, const char *label, uint_t assoc, uint_t size)
3777 {
3778 	if (size == 0 || assoc == 0)
3779 		return;
3780 	add_amd_l2_assoc(devi, label, assoc);
3781 	add_cache_prop(devi, label, size_str, size);
3782 }
3783 
3784 static void
3785 add_amd_l2_cache(dev_info_t *devi, const char *label,
3786     uint_t size, uint_t assoc, uint_t lines_per_tag, uint_t line_size)
3787 {
3788 	if (size == 0 || assoc == 0 || line_size == 0)
3789 		return;
3790 	add_amd_l2_assoc(devi, label, assoc);
3791 	if (lines_per_tag != 0)
3792 		add_cache_prop(devi, label, "lines-per-tag", lines_per_tag);
3793 	add_cache_prop(devi, label, line_str, line_size);
3794 	add_cache_prop(devi, label, size_str, size * 1024);
3795 }
3796 
3797 static void
3798 amd_cache_info(struct cpuid_info *cpi, dev_info_t *devi)
3799 {
3800 	struct cpuid_regs *cp;
3801 
3802 	if (cpi->cpi_xmaxeax < 0x80000005)
3803 		return;
3804 	cp = &cpi->cpi_extd[5];
3805 
3806 	/*
3807 	 * 4M/2M L1 TLB configuration
3808 	 *
3809 	 * We report the size for 2M pages because AMD uses two
3810 	 * TLB entries for one 4M page.
3811 	 */
3812 	add_amd_tlb(devi, "dtlb-2M",
3813 	    BITX(cp->cp_eax, 31, 24), BITX(cp->cp_eax, 23, 16));
3814 	add_amd_tlb(devi, "itlb-2M",
3815 	    BITX(cp->cp_eax, 15, 8), BITX(cp->cp_eax, 7, 0));
3816 
3817 	/*
3818 	 * 4K L1 TLB configuration
3819 	 */
3820 
3821 	switch (cpi->cpi_vendor) {
3822 		uint_t nentries;
3823 	case X86_VENDOR_TM:
3824 		if (cpi->cpi_family >= 5) {
3825 			/*
3826 			 * Crusoe processors have 256 TLB entries, but
3827 			 * cpuid data format constrains them to only
3828 			 * reporting 255 of them.
3829 			 */
3830 			if ((nentries = BITX(cp->cp_ebx, 23, 16)) == 255)
3831 				nentries = 256;
3832 			/*
3833 			 * Crusoe processors also have a unified TLB
3834 			 */
3835 			add_amd_tlb(devi, "tlb-4K", BITX(cp->cp_ebx, 31, 24),
3836 			    nentries);
3837 			break;
3838 		}
3839 		/*FALLTHROUGH*/
3840 	default:
3841 		add_amd_tlb(devi, itlb4k_str,
3842 		    BITX(cp->cp_ebx, 31, 24), BITX(cp->cp_ebx, 23, 16));
3843 		add_amd_tlb(devi, dtlb4k_str,
3844 		    BITX(cp->cp_ebx, 15, 8), BITX(cp->cp_ebx, 7, 0));
3845 		break;
3846 	}
3847 
3848 	/*
3849 	 * data L1 cache configuration
3850 	 */
3851 
3852 	add_amd_cache(devi, l1_dcache_str,
3853 	    BITX(cp->cp_ecx, 31, 24), BITX(cp->cp_ecx, 23, 16),
3854 	    BITX(cp->cp_ecx, 15, 8), BITX(cp->cp_ecx, 7, 0));
3855 
3856 	/*
3857 	 * code L1 cache configuration
3858 	 */
3859 
3860 	add_amd_cache(devi, l1_icache_str,
3861 	    BITX(cp->cp_edx, 31, 24), BITX(cp->cp_edx, 23, 16),
3862 	    BITX(cp->cp_edx, 15, 8), BITX(cp->cp_edx, 7, 0));
3863 
3864 	if (cpi->cpi_xmaxeax < 0x80000006)
3865 		return;
3866 	cp = &cpi->cpi_extd[6];
3867 
3868 	/* Check for a unified L2 TLB for large pages */
3869 
3870 	if (BITX(cp->cp_eax, 31, 16) == 0)
3871 		add_amd_l2_tlb(devi, "l2-tlb-2M",
3872 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
3873 	else {
3874 		add_amd_l2_tlb(devi, "l2-dtlb-2M",
3875 		    BITX(cp->cp_eax, 31, 28), BITX(cp->cp_eax, 27, 16));
3876 		add_amd_l2_tlb(devi, "l2-itlb-2M",
3877 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
3878 	}
3879 
3880 	/* Check for a unified L2 TLB for 4K pages */
3881 
3882 	if (BITX(cp->cp_ebx, 31, 16) == 0) {
3883 		add_amd_l2_tlb(devi, "l2-tlb-4K",
3884 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
3885 	} else {
3886 		add_amd_l2_tlb(devi, "l2-dtlb-4K",
3887 		    BITX(cp->cp_eax, 31, 28), BITX(cp->cp_eax, 27, 16));
3888 		add_amd_l2_tlb(devi, "l2-itlb-4K",
3889 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
3890 	}
3891 
3892 	add_amd_l2_cache(devi, l2_cache_str,
3893 	    BITX(cp->cp_ecx, 31, 16), BITX(cp->cp_ecx, 15, 12),
3894 	    BITX(cp->cp_ecx, 11, 8), BITX(cp->cp_ecx, 7, 0));
3895 }
3896 
3897 /*
3898  * There are two basic ways that the x86 world describes it cache
3899  * and tlb architecture - Intel's way and AMD's way.
3900  *
3901  * Return which flavor of cache architecture we should use
3902  */
3903 static int
3904 x86_which_cacheinfo(struct cpuid_info *cpi)
3905 {
3906 	switch (cpi->cpi_vendor) {
3907 	case X86_VENDOR_Intel:
3908 		if (cpi->cpi_maxeax >= 2)
3909 			return (X86_VENDOR_Intel);
3910 		break;
3911 	case X86_VENDOR_AMD:
3912 		/*
3913 		 * The K5 model 1 was the first part from AMD that reported
3914 		 * cache sizes via extended cpuid functions.
3915 		 */
3916 		if (cpi->cpi_family > 5 ||
3917 		    (cpi->cpi_family == 5 && cpi->cpi_model >= 1))
3918 			return (X86_VENDOR_AMD);
3919 		break;
3920 	case X86_VENDOR_TM:
3921 		if (cpi->cpi_family >= 5)
3922 			return (X86_VENDOR_AMD);
3923 		/*FALLTHROUGH*/
3924 	default:
3925 		/*
3926 		 * If they have extended CPU data for 0x80000005
3927 		 * then we assume they have AMD-format cache
3928 		 * information.
3929 		 *
3930 		 * If not, and the vendor happens to be Cyrix,
3931 		 * then try our-Cyrix specific handler.
3932 		 *
3933 		 * If we're not Cyrix, then assume we're using Intel's
3934 		 * table-driven format instead.
3935 		 */
3936 		if (cpi->cpi_xmaxeax >= 0x80000005)
3937 			return (X86_VENDOR_AMD);
3938 		else if (cpi->cpi_vendor == X86_VENDOR_Cyrix)
3939 			return (X86_VENDOR_Cyrix);
3940 		else if (cpi->cpi_maxeax >= 2)
3941 			return (X86_VENDOR_Intel);
3942 		break;
3943 	}
3944 	return (-1);
3945 }
3946 
3947 void
3948 cpuid_set_cpu_properties(void *dip, processorid_t cpu_id,
3949     struct cpuid_info *cpi)
3950 {
3951 	dev_info_t *cpu_devi;
3952 	int create;
3953 
3954 	cpu_devi = (dev_info_t *)dip;
3955 
3956 	/* device_type */
3957 	(void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi,
3958 	    "device_type", "cpu");
3959 
3960 	/* reg */
3961 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3962 	    "reg", cpu_id);
3963 
3964 	/* cpu-mhz, and clock-frequency */
3965 	if (cpu_freq > 0) {
3966 		long long mul;
3967 
3968 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3969 		    "cpu-mhz", cpu_freq);
3970 		if ((mul = cpu_freq * 1000000LL) <= INT_MAX)
3971 			(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3972 			    "clock-frequency", (int)mul);
3973 	}
3974 
3975 	if (!is_x86_feature(x86_featureset, X86FSET_CPUID)) {
3976 		return;
3977 	}
3978 
3979 	/* vendor-id */
3980 	(void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi,
3981 	    "vendor-id", cpi->cpi_vendorstr);
3982 
3983 	if (cpi->cpi_maxeax == 0) {
3984 		return;
3985 	}
3986 
3987 	/*
3988 	 * family, model, and step
3989 	 */
3990 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3991 	    "family", CPI_FAMILY(cpi));
3992 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3993 	    "cpu-model", CPI_MODEL(cpi));
3994 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3995 	    "stepping-id", CPI_STEP(cpi));
3996 
3997 	/* type */
3998 	switch (cpi->cpi_vendor) {
3999 	case X86_VENDOR_Intel:
4000 		create = 1;
4001 		break;
4002 	default:
4003 		create = 0;
4004 		break;
4005 	}
4006 	if (create)
4007 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
4008 		    "type", CPI_TYPE(cpi));
4009 
4010 	/* ext-family */
4011 	switch (cpi->cpi_vendor) {
4012 	case X86_VENDOR_Intel:
4013 	case X86_VENDOR_AMD:
4014 		create = cpi->cpi_family >= 0xf;
4015 		break;
4016 	default:
4017 		create = 0;
4018 		break;
4019 	}
4020 	if (create)
4021 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
4022 		    "ext-family", CPI_FAMILY_XTD(cpi));
4023 
4024 	/* ext-model */
4025 	switch (cpi->cpi_vendor) {
4026 	case X86_VENDOR_Intel:
4027 		create = IS_EXTENDED_MODEL_INTEL(cpi);
4028 		break;
4029 	case X86_VENDOR_AMD:
4030 		create = CPI_FAMILY(cpi) == 0xf;
4031 		break;
4032 	default:
4033 		create = 0;
4034 		break;
4035 	}
4036 	if (create)
4037 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
4038 		    "ext-model", CPI_MODEL_XTD(cpi));
4039 
4040 	/* generation */
4041 	switch (cpi->cpi_vendor) {
4042 	case X86_VENDOR_AMD:
4043 		/*
4044 		 * AMD K5 model 1 was the first part to support this
4045 		 */
4046 		create = cpi->cpi_xmaxeax >= 0x80000001;
4047 		break;
4048 	default:
4049 		create = 0;
4050 		break;
4051 	}
4052 	if (create)
4053 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
4054 		    "generation", BITX((cpi)->cpi_extd[1].cp_eax, 11, 8));
4055 
4056 	/* brand-id */
4057 	switch (cpi->cpi_vendor) {
4058 	case X86_VENDOR_Intel:
4059 		/*
4060 		 * brand id first appeared on Pentium III Xeon model 8,
4061 		 * and Celeron model 8 processors and Opteron
4062 		 */
4063 		create = cpi->cpi_family > 6 ||
4064 		    (cpi->cpi_family == 6 && cpi->cpi_model >= 8);
4065 		break;
4066 	case X86_VENDOR_AMD:
4067 		create = cpi->cpi_family >= 0xf;
4068 		break;
4069 	default:
4070 		create = 0;
4071 		break;
4072 	}
4073 	if (create && cpi->cpi_brandid != 0) {
4074 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
4075 		    "brand-id", cpi->cpi_brandid);
4076 	}
4077 
4078 	/* chunks, and apic-id */
4079 	switch (cpi->cpi_vendor) {
4080 		/*
4081 		 * first available on Pentium IV and Opteron (K8)
4082 		 */
4083 	case X86_VENDOR_Intel:
4084 		create = IS_NEW_F6(cpi) || cpi->cpi_family >= 0xf;
4085 		break;
4086 	case X86_VENDOR_AMD:
4087 		create = cpi->cpi_family >= 0xf;
4088 		break;
4089 	default:
4090 		create = 0;
4091 		break;
4092 	}
4093 	if (create) {
4094 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
4095 		    "chunks", CPI_CHUNKS(cpi));
4096 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
4097 		    "apic-id", cpi->cpi_apicid);
4098 		if (cpi->cpi_chipid >= 0) {
4099 			(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
4100 			    "chip#", cpi->cpi_chipid);
4101 			(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
4102 			    "clog#", cpi->cpi_clogid);
4103 		}
4104 	}
4105 
4106 	/* cpuid-features */
4107 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
4108 	    "cpuid-features", CPI_FEATURES_EDX(cpi));
4109 
4110 
4111 	/* cpuid-features-ecx */
4112 	switch (cpi->cpi_vendor) {
4113 	case X86_VENDOR_Intel:
4114 		create = IS_NEW_F6(cpi) || cpi->cpi_family >= 0xf;
4115 		break;
4116 	default:
4117 		create = 0;
4118 		break;
4119 	}
4120 	if (create)
4121 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
4122 		    "cpuid-features-ecx", CPI_FEATURES_ECX(cpi));
4123 
4124 	/* ext-cpuid-features */
4125 	switch (cpi->cpi_vendor) {
4126 	case X86_VENDOR_Intel:
4127 	case X86_VENDOR_AMD:
4128 	case X86_VENDOR_Cyrix:
4129 	case X86_VENDOR_TM:
4130 	case X86_VENDOR_Centaur:
4131 		create = cpi->cpi_xmaxeax >= 0x80000001;
4132 		break;
4133 	default:
4134 		create = 0;
4135 		break;
4136 	}
4137 	if (create) {
4138 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
4139 		    "ext-cpuid-features", CPI_FEATURES_XTD_EDX(cpi));
4140 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
4141 		    "ext-cpuid-features-ecx", CPI_FEATURES_XTD_ECX(cpi));
4142 	}
4143 
4144 	/*
4145 	 * Brand String first appeared in Intel Pentium IV, AMD K5
4146 	 * model 1, and Cyrix GXm.  On earlier models we try and
4147 	 * simulate something similar .. so this string should always
4148 	 * same -something- about the processor, however lame.
4149 	 */
4150 	(void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi,
4151 	    "brand-string", cpi->cpi_brandstr);
4152 
4153 	/*
4154 	 * Finally, cache and tlb information
4155 	 */
4156 	switch (x86_which_cacheinfo(cpi)) {
4157 	case X86_VENDOR_Intel:
4158 		intel_walk_cacheinfo(cpi, cpu_devi, add_cacheent_props);
4159 		break;
4160 	case X86_VENDOR_Cyrix:
4161 		cyrix_walk_cacheinfo(cpi, cpu_devi, add_cacheent_props);
4162 		break;
4163 	case X86_VENDOR_AMD:
4164 		amd_cache_info(cpi, cpu_devi);
4165 		break;
4166 	default:
4167 		break;
4168 	}
4169 }
4170 
4171 struct l2info {
4172 	int *l2i_csz;
4173 	int *l2i_lsz;
4174 	int *l2i_assoc;
4175 	int l2i_ret;
4176 };
4177 
4178 /*
4179  * A cacheinfo walker that fetches the size, line-size and associativity
4180  * of the L2 cache
4181  */
4182 static int
4183 intel_l2cinfo(void *arg, const struct cachetab *ct)
4184 {
4185 	struct l2info *l2i = arg;
4186 	int *ip;
4187 
4188 	if (ct->ct_label != l2_cache_str &&
4189 	    ct->ct_label != sl2_cache_str)
4190 		return (0);	/* not an L2 -- keep walking */
4191 
4192 	if ((ip = l2i->l2i_csz) != NULL)
4193 		*ip = ct->ct_size;
4194 	if ((ip = l2i->l2i_lsz) != NULL)
4195 		*ip = ct->ct_line_size;
4196 	if ((ip = l2i->l2i_assoc) != NULL)
4197 		*ip = ct->ct_assoc;
4198 	l2i->l2i_ret = ct->ct_size;
4199 	return (1);		/* was an L2 -- terminate walk */
4200 }
4201 
4202 /*
4203  * AMD L2/L3 Cache and TLB Associativity Field Definition:
4204  *
4205  *	Unlike the associativity for the L1 cache and tlb where the 8 bit
4206  *	value is the associativity, the associativity for the L2 cache and
4207  *	tlb is encoded in the following table. The 4 bit L2 value serves as
4208  *	an index into the amd_afd[] array to determine the associativity.
4209  *	-1 is undefined. 0 is fully associative.
4210  */
4211 
4212 static int amd_afd[] =
4213 	{-1, 1, 2, -1, 4, -1, 8, -1, 16, -1, 32, 48, 64, 96, 128, 0};
4214 
4215 static void
4216 amd_l2cacheinfo(struct cpuid_info *cpi, struct l2info *l2i)
4217 {
4218 	struct cpuid_regs *cp;
4219 	uint_t size, assoc;
4220 	int i;
4221 	int *ip;
4222 
4223 	if (cpi->cpi_xmaxeax < 0x80000006)
4224 		return;
4225 	cp = &cpi->cpi_extd[6];
4226 
4227 	if ((i = BITX(cp->cp_ecx, 15, 12)) != 0 &&
4228 	    (size = BITX(cp->cp_ecx, 31, 16)) != 0) {
4229 		uint_t cachesz = size * 1024;
4230 		assoc = amd_afd[i];
4231 
4232 		ASSERT(assoc != -1);
4233 
4234 		if ((ip = l2i->l2i_csz) != NULL)
4235 			*ip = cachesz;
4236 		if ((ip = l2i->l2i_lsz) != NULL)
4237 			*ip = BITX(cp->cp_ecx, 7, 0);
4238 		if ((ip = l2i->l2i_assoc) != NULL)
4239 			*ip = assoc;
4240 		l2i->l2i_ret = cachesz;
4241 	}
4242 }
4243 
4244 int
4245 getl2cacheinfo(cpu_t *cpu, int *csz, int *lsz, int *assoc)
4246 {
4247 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
4248 	struct l2info __l2info, *l2i = &__l2info;
4249 
4250 	l2i->l2i_csz = csz;
4251 	l2i->l2i_lsz = lsz;
4252 	l2i->l2i_assoc = assoc;
4253 	l2i->l2i_ret = -1;
4254 
4255 	switch (x86_which_cacheinfo(cpi)) {
4256 	case X86_VENDOR_Intel:
4257 		intel_walk_cacheinfo(cpi, l2i, intel_l2cinfo);
4258 		break;
4259 	case X86_VENDOR_Cyrix:
4260 		cyrix_walk_cacheinfo(cpi, l2i, intel_l2cinfo);
4261 		break;
4262 	case X86_VENDOR_AMD:
4263 		amd_l2cacheinfo(cpi, l2i);
4264 		break;
4265 	default:
4266 		break;
4267 	}
4268 	return (l2i->l2i_ret);
4269 }
4270 
4271 #if !defined(__xpv)
4272 
4273 uint32_t *
4274 cpuid_mwait_alloc(cpu_t *cpu)
4275 {
4276 	uint32_t	*ret;
4277 	size_t		mwait_size;
4278 
4279 	ASSERT(cpuid_checkpass(CPU, 2));
4280 
4281 	mwait_size = CPU->cpu_m.mcpu_cpi->cpi_mwait.mon_max;
4282 	if (mwait_size == 0)
4283 		return (NULL);
4284 
4285 	/*
4286 	 * kmem_alloc() returns cache line size aligned data for mwait_size
4287 	 * allocations.  mwait_size is currently cache line sized.  Neither
4288 	 * of these implementation details are guarantied to be true in the
4289 	 * future.
4290 	 *
4291 	 * First try allocating mwait_size as kmem_alloc() currently returns
4292 	 * correctly aligned memory.  If kmem_alloc() does not return
4293 	 * mwait_size aligned memory, then use mwait_size ROUNDUP.
4294 	 *
4295 	 * Set cpi_mwait.buf_actual and cpi_mwait.size_actual in case we
4296 	 * decide to free this memory.
4297 	 */
4298 	ret = kmem_zalloc(mwait_size, KM_SLEEP);
4299 	if (ret == (uint32_t *)P2ROUNDUP((uintptr_t)ret, mwait_size)) {
4300 		cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual = ret;
4301 		cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual = mwait_size;
4302 		*ret = MWAIT_RUNNING;
4303 		return (ret);
4304 	} else {
4305 		kmem_free(ret, mwait_size);
4306 		ret = kmem_zalloc(mwait_size * 2, KM_SLEEP);
4307 		cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual = ret;
4308 		cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual = mwait_size * 2;
4309 		ret = (uint32_t *)P2ROUNDUP((uintptr_t)ret, mwait_size);
4310 		*ret = MWAIT_RUNNING;
4311 		return (ret);
4312 	}
4313 }
4314 
4315 void
4316 cpuid_mwait_free(cpu_t *cpu)
4317 {
4318 	if (cpu->cpu_m.mcpu_cpi == NULL) {
4319 		return;
4320 	}
4321 
4322 	if (cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual != NULL &&
4323 	    cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual > 0) {
4324 		kmem_free(cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual,
4325 		    cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual);
4326 	}
4327 
4328 	cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual = NULL;
4329 	cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual = 0;
4330 }
4331 
4332 void
4333 patch_tsc_read(int flag)
4334 {
4335 	size_t cnt;
4336 
4337 	switch (flag) {
4338 	case X86_NO_TSC:
4339 		cnt = &_no_rdtsc_end - &_no_rdtsc_start;
4340 		(void) memcpy((void *)tsc_read, (void *)&_no_rdtsc_start, cnt);
4341 		break;
4342 	case X86_HAVE_TSCP:
4343 		cnt = &_tscp_end - &_tscp_start;
4344 		(void) memcpy((void *)tsc_read, (void *)&_tscp_start, cnt);
4345 		break;
4346 	case X86_TSC_MFENCE:
4347 		cnt = &_tsc_mfence_end - &_tsc_mfence_start;
4348 		(void) memcpy((void *)tsc_read,
4349 		    (void *)&_tsc_mfence_start, cnt);
4350 		break;
4351 	case X86_TSC_LFENCE:
4352 		cnt = &_tsc_lfence_end - &_tsc_lfence_start;
4353 		(void) memcpy((void *)tsc_read,
4354 		    (void *)&_tsc_lfence_start, cnt);
4355 		break;
4356 	default:
4357 		break;
4358 	}
4359 }
4360 
4361 int
4362 cpuid_deep_cstates_supported(void)
4363 {
4364 	struct cpuid_info *cpi;
4365 	struct cpuid_regs regs;
4366 
4367 	ASSERT(cpuid_checkpass(CPU, 1));
4368 
4369 	cpi = CPU->cpu_m.mcpu_cpi;
4370 
4371 	if (!is_x86_feature(x86_featureset, X86FSET_CPUID))
4372 		return (0);
4373 
4374 	switch (cpi->cpi_vendor) {
4375 	case X86_VENDOR_Intel:
4376 		if (cpi->cpi_xmaxeax < 0x80000007)
4377 			return (0);
4378 
4379 		/*
4380 		 * TSC run at a constant rate in all ACPI C-states?
4381 		 */
4382 		regs.cp_eax = 0x80000007;
4383 		(void) __cpuid_insn(&regs);
4384 		return (regs.cp_edx & CPUID_TSC_CSTATE_INVARIANCE);
4385 
4386 	default:
4387 		return (0);
4388 	}
4389 }
4390 
4391 #endif	/* !__xpv */
4392 
4393 void
4394 post_startup_cpu_fixups(void)
4395 {
4396 #ifndef __xpv
4397 	/*
4398 	 * Some AMD processors support C1E state. Entering this state will
4399 	 * cause the local APIC timer to stop, which we can't deal with at
4400 	 * this time.
4401 	 */
4402 	if (cpuid_getvendor(CPU) == X86_VENDOR_AMD) {
4403 		on_trap_data_t otd;
4404 		uint64_t reg;
4405 
4406 		if (!on_trap(&otd, OT_DATA_ACCESS)) {
4407 			reg = rdmsr(MSR_AMD_INT_PENDING_CMP_HALT);
4408 			/* Disable C1E state if it is enabled by BIOS */
4409 			if ((reg >> AMD_ACTONCMPHALT_SHIFT) &
4410 			    AMD_ACTONCMPHALT_MASK) {
4411 				reg &= ~(AMD_ACTONCMPHALT_MASK <<
4412 				    AMD_ACTONCMPHALT_SHIFT);
4413 				wrmsr(MSR_AMD_INT_PENDING_CMP_HALT, reg);
4414 			}
4415 		}
4416 		no_trap();
4417 	}
4418 #endif	/* !__xpv */
4419 }
4420 
4421 /*
4422  * Setup necessary registers to enable XSAVE feature on this processor.
4423  * This function needs to be called early enough, so that no xsave/xrstor
4424  * ops will execute on the processor before the MSRs are properly set up.
4425  *
4426  * Current implementation has the following assumption:
4427  * - cpuid_pass1() is done, so that X86 features are known.
4428  * - fpu_probe() is done, so that fp_save_mech is chosen.
4429  */
4430 void
4431 xsave_setup_msr(cpu_t *cpu)
4432 {
4433 	ASSERT(fp_save_mech == FP_XSAVE);
4434 	ASSERT(is_x86_feature(x86_featureset, X86FSET_XSAVE));
4435 
4436 	/* Enable OSXSAVE in CR4. */
4437 	setcr4(getcr4() | CR4_OSXSAVE);
4438 	/*
4439 	 * Update SW copy of ECX, so that /dev/cpu/self/cpuid will report
4440 	 * correct value.
4441 	 */
4442 	cpu->cpu_m.mcpu_cpi->cpi_std[1].cp_ecx |= CPUID_INTC_ECX_OSXSAVE;
4443 	setup_xfem();
4444 }
4445 
4446 /*
4447  * Starting with the Westmere processor the local
4448  * APIC timer will continue running in all C-states,
4449  * including the deepest C-states.
4450  */
4451 int
4452 cpuid_arat_supported(void)
4453 {
4454 	struct cpuid_info *cpi;
4455 	struct cpuid_regs regs;
4456 
4457 	ASSERT(cpuid_checkpass(CPU, 1));
4458 	ASSERT(is_x86_feature(x86_featureset, X86FSET_CPUID));
4459 
4460 	cpi = CPU->cpu_m.mcpu_cpi;
4461 
4462 	switch (cpi->cpi_vendor) {
4463 	case X86_VENDOR_Intel:
4464 		/*
4465 		 * Always-running Local APIC Timer is
4466 		 * indicated by CPUID.6.EAX[2].
4467 		 */
4468 		if (cpi->cpi_maxeax >= 6) {
4469 			regs.cp_eax = 6;
4470 			(void) cpuid_insn(NULL, &regs);
4471 			return (regs.cp_eax & CPUID_CSTATE_ARAT);
4472 		} else {
4473 			return (0);
4474 		}
4475 	default:
4476 		return (0);
4477 	}
4478 }
4479 
4480 /*
4481  * Check support for Intel ENERGY_PERF_BIAS feature
4482  */
4483 int
4484 cpuid_iepb_supported(struct cpu *cp)
4485 {
4486 	struct cpuid_info *cpi = cp->cpu_m.mcpu_cpi;
4487 	struct cpuid_regs regs;
4488 
4489 	ASSERT(cpuid_checkpass(cp, 1));
4490 
4491 	if (!(is_x86_feature(x86_featureset, X86FSET_CPUID)) ||
4492 	    !(is_x86_feature(x86_featureset, X86FSET_MSR))) {
4493 		return (0);
4494 	}
4495 
4496 	/*
4497 	 * Intel ENERGY_PERF_BIAS MSR is indicated by
4498 	 * capability bit CPUID.6.ECX.3
4499 	 */
4500 	if ((cpi->cpi_vendor != X86_VENDOR_Intel) || (cpi->cpi_maxeax < 6))
4501 		return (0);
4502 
4503 	regs.cp_eax = 0x6;
4504 	(void) cpuid_insn(NULL, &regs);
4505 	return (regs.cp_ecx & CPUID_EPB_SUPPORT);
4506 }
4507 
4508 /*
4509  * Check support for TSC deadline timer
4510  *
4511  * TSC deadline timer provides a superior software programming
4512  * model over local APIC timer that eliminates "time drifts".
4513  * Instead of specifying a relative time, software specifies an
4514  * absolute time as the target at which the processor should
4515  * generate a timer event.
4516  */
4517 int
4518 cpuid_deadline_tsc_supported(void)
4519 {
4520 	struct cpuid_info *cpi = CPU->cpu_m.mcpu_cpi;
4521 	struct cpuid_regs regs;
4522 
4523 	ASSERT(cpuid_checkpass(CPU, 1));
4524 	ASSERT(is_x86_feature(x86_featureset, X86FSET_CPUID));
4525 
4526 	switch (cpi->cpi_vendor) {
4527 	case X86_VENDOR_Intel:
4528 		if (cpi->cpi_maxeax >= 1) {
4529 			regs.cp_eax = 1;
4530 			(void) cpuid_insn(NULL, &regs);
4531 			return (regs.cp_ecx & CPUID_DEADLINE_TSC);
4532 		} else {
4533 			return (0);
4534 		}
4535 	default:
4536 		return (0);
4537 	}
4538 }
4539 
4540 #if defined(__amd64) && !defined(__xpv)
4541 /*
4542  * Patch in versions of bcopy for high performance Intel Nhm processors
4543  * and later...
4544  */
4545 void
4546 patch_memops(uint_t vendor)
4547 {
4548 	size_t cnt, i;
4549 	caddr_t to, from;
4550 
4551 	if ((vendor == X86_VENDOR_Intel) &&
4552 	    is_x86_feature(x86_featureset, X86FSET_SSE4_2)) {
4553 		cnt = &bcopy_patch_end - &bcopy_patch_start;
4554 		to = &bcopy_ck_size;
4555 		from = &bcopy_patch_start;
4556 		for (i = 0; i < cnt; i++) {
4557 			*to++ = *from++;
4558 		}
4559 	}
4560 }
4561 #endif  /* __amd64 && !__xpv */
4562 
4563 /*
4564  * This function finds the number of bits to represent the number of cores per
4565  * chip and the number of strands per core for the Intel platforms.
4566  * It re-uses the x2APIC cpuid code of the cpuid_pass2().
4567  */
4568 void
4569 cpuid_get_ext_topo(uint_t vendor, uint_t *core_nbits, uint_t *strand_nbits)
4570 {
4571 	struct cpuid_regs regs;
4572 	struct cpuid_regs *cp = &regs;
4573 
4574 	if (vendor != X86_VENDOR_Intel) {
4575 		return;
4576 	}
4577 
4578 	/* if the cpuid level is 0xB, extended topo is available. */
4579 	cp->cp_eax = 0;
4580 	if (__cpuid_insn(cp) >= 0xB) {
4581 
4582 		cp->cp_eax = 0xB;
4583 		cp->cp_edx = cp->cp_ebx = cp->cp_ecx = 0;
4584 		(void) __cpuid_insn(cp);
4585 
4586 		/*
4587 		 * Check CPUID.EAX=0BH, ECX=0H:EBX is non-zero, which
4588 		 * indicates that the extended topology enumeration leaf is
4589 		 * available.
4590 		 */
4591 		if (cp->cp_ebx) {
4592 			uint_t coreid_shift = 0;
4593 			uint_t chipid_shift = 0;
4594 			uint_t i;
4595 			uint_t level;
4596 
4597 			for (i = 0; i < CPI_FNB_ECX_MAX; i++) {
4598 				cp->cp_eax = 0xB;
4599 				cp->cp_ecx = i;
4600 
4601 				(void) __cpuid_insn(cp);
4602 				level = CPI_CPU_LEVEL_TYPE(cp);
4603 
4604 				if (level == 1) {
4605 					/*
4606 					 * Thread level processor topology
4607 					 * Number of bits shift right APIC ID
4608 					 * to get the coreid.
4609 					 */
4610 					coreid_shift = BITX(cp->cp_eax, 4, 0);
4611 				} else if (level == 2) {
4612 					/*
4613 					 * Core level processor topology
4614 					 * Number of bits shift right APIC ID
4615 					 * to get the chipid.
4616 					 */
4617 					chipid_shift = BITX(cp->cp_eax, 4, 0);
4618 				}
4619 			}
4620 
4621 			if (coreid_shift > 0 && chipid_shift > coreid_shift) {
4622 				*strand_nbits = coreid_shift;
4623 				*core_nbits = chipid_shift - coreid_shift;
4624 			}
4625 		}
4626 	}
4627 }
4628