xref: /titanic_41/usr/src/uts/i86pc/os/cpuid.c (revision 3fb517f786391b507780c78aabb8d98bfea9efe9)
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) 2009, 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/auxv_386.h>
52 #include <sys/bitmap.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_feature 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_feature = 0;
115 uint_t x86_vendor = X86_VENDOR_IntelClone;
116 uint_t x86_type = X86_TYPE_OTHER;
117 uint_t x86_clflush_size = 0;
118 
119 uint_t pentiumpro_bug4046376;
120 uint_t pentiumpro_bug4064495;
121 
122 uint_t enable486;
123 /*
124  * This is set to platform type Solaris is running on.
125  */
126 static int platform_type = -1;
127 
128 #if !defined(__xpv)
129 /*
130  * Variable to patch if hypervisor platform detection needs to be
131  * disabled (e.g. platform_type will always be HW_NATIVE if this is 0).
132  */
133 int enable_platform_detection = 1;
134 #endif
135 
136 /*
137  * monitor/mwait info.
138  *
139  * size_actual and buf_actual are the real address and size allocated to get
140  * proper mwait_buf alignement.  buf_actual and size_actual should be passed
141  * to kmem_free().  Currently kmem_alloc() and mwait happen to both use
142  * processor cache-line alignment, but this is not guarantied in the furture.
143  */
144 struct mwait_info {
145 	size_t		mon_min;	/* min size to avoid missed wakeups */
146 	size_t		mon_max;	/* size to avoid false wakeups */
147 	size_t		size_actual;	/* size actually allocated */
148 	void		*buf_actual;	/* memory actually allocated */
149 	uint32_t	support;	/* processor support of monitor/mwait */
150 };
151 
152 /*
153  * These constants determine how many of the elements of the
154  * cpuid we cache in the cpuid_info data structure; the
155  * remaining elements are accessible via the cpuid instruction.
156  */
157 
158 #define	NMAX_CPI_STD	6		/* eax = 0 .. 5 */
159 #define	NMAX_CPI_EXTD	0x1c		/* eax = 0x80000000 .. 0x8000001b */
160 
161 /*
162  * Some terminology needs to be explained:
163  *  - Socket: Something that can be plugged into a motherboard.
164  *  - Package: Same as socket
165  *  - Chip: Same as socket. Note that AMD's documentation uses term "chip"
166  *    differently: there, chip is the same as processor node (below)
167  *  - Processor node: Some AMD processors have more than one
168  *    "subprocessor" embedded in a package. These subprocessors (nodes)
169  *    are fully-functional processors themselves with cores, caches,
170  *    memory controllers, PCI configuration spaces. They are connected
171  *    inside the package with Hypertransport links. On single-node
172  *    processors, processor node is equivalent to chip/socket/package.
173  */
174 
175 struct cpuid_info {
176 	uint_t cpi_pass;		/* last pass completed */
177 	/*
178 	 * standard function information
179 	 */
180 	uint_t cpi_maxeax;		/* fn 0: %eax */
181 	char cpi_vendorstr[13];		/* fn 0: %ebx:%ecx:%edx */
182 	uint_t cpi_vendor;		/* enum of cpi_vendorstr */
183 
184 	uint_t cpi_family;		/* fn 1: extended family */
185 	uint_t cpi_model;		/* fn 1: extended model */
186 	uint_t cpi_step;		/* fn 1: stepping */
187 	chipid_t cpi_chipid;		/* fn 1: %ebx:  Intel: chip # */
188 					/*		AMD: package/socket # */
189 	uint_t cpi_brandid;		/* fn 1: %ebx: brand ID */
190 	int cpi_clogid;			/* fn 1: %ebx: thread # */
191 	uint_t cpi_ncpu_per_chip;	/* fn 1: %ebx: logical cpu count */
192 	uint8_t cpi_cacheinfo[16];	/* fn 2: intel-style cache desc */
193 	uint_t cpi_ncache;		/* fn 2: number of elements */
194 	uint_t cpi_ncpu_shr_last_cache;	/* fn 4: %eax: ncpus sharing cache */
195 	id_t cpi_last_lvl_cacheid;	/* fn 4: %eax: derived cache id */
196 	uint_t cpi_std_4_size;		/* fn 4: number of fn 4 elements */
197 	struct cpuid_regs **cpi_std_4;	/* fn 4: %ecx == 0 .. fn4_size */
198 	struct cpuid_regs cpi_std[NMAX_CPI_STD];	/* 0 .. 5 */
199 	/*
200 	 * extended function information
201 	 */
202 	uint_t cpi_xmaxeax;		/* fn 0x80000000: %eax */
203 	char cpi_brandstr[49];		/* fn 0x8000000[234] */
204 	uint8_t cpi_pabits;		/* fn 0x80000006: %eax */
205 	uint8_t	cpi_vabits;		/* fn 0x80000006: %eax */
206 	struct	cpuid_regs cpi_extd[NMAX_CPI_EXTD];	/* 0x800000XX */
207 
208 	id_t cpi_coreid;		/* same coreid => strands share core */
209 	int cpi_pkgcoreid;		/* core number within single package */
210 	uint_t cpi_ncore_per_chip;	/* AMD: fn 0x80000008: %ecx[7-0] */
211 					/* Intel: fn 4: %eax[31-26] */
212 	/*
213 	 * supported feature information
214 	 */
215 	uint32_t cpi_support[5];
216 #define	STD_EDX_FEATURES	0
217 #define	AMD_EDX_FEATURES	1
218 #define	TM_EDX_FEATURES		2
219 #define	STD_ECX_FEATURES	3
220 #define	AMD_ECX_FEATURES	4
221 	/*
222 	 * Synthesized information, where known.
223 	 */
224 	uint32_t cpi_chiprev;		/* See X86_CHIPREV_* in x86_archext.h */
225 	const char *cpi_chiprevstr;	/* May be NULL if chiprev unknown */
226 	uint32_t cpi_socket;		/* Chip package/socket type */
227 
228 	struct mwait_info cpi_mwait;	/* fn 5: monitor/mwait info */
229 	uint32_t cpi_apicid;
230 	uint_t cpi_procnodeid;		/* AMD: nodeID on HT, Intel: chipid */
231 	uint_t cpi_procnodes_per_pkg;	/* AMD: # of nodes in the package */
232 					/* Intel: 1 */
233 };
234 
235 
236 static struct cpuid_info cpuid_info0;
237 
238 /*
239  * These bit fields are defined by the Intel Application Note AP-485
240  * "Intel Processor Identification and the CPUID Instruction"
241  */
242 #define	CPI_FAMILY_XTD(cpi)	BITX((cpi)->cpi_std[1].cp_eax, 27, 20)
243 #define	CPI_MODEL_XTD(cpi)	BITX((cpi)->cpi_std[1].cp_eax, 19, 16)
244 #define	CPI_TYPE(cpi)		BITX((cpi)->cpi_std[1].cp_eax, 13, 12)
245 #define	CPI_FAMILY(cpi)		BITX((cpi)->cpi_std[1].cp_eax, 11, 8)
246 #define	CPI_STEP(cpi)		BITX((cpi)->cpi_std[1].cp_eax, 3, 0)
247 #define	CPI_MODEL(cpi)		BITX((cpi)->cpi_std[1].cp_eax, 7, 4)
248 
249 #define	CPI_FEATURES_EDX(cpi)		((cpi)->cpi_std[1].cp_edx)
250 #define	CPI_FEATURES_ECX(cpi)		((cpi)->cpi_std[1].cp_ecx)
251 #define	CPI_FEATURES_XTD_EDX(cpi)	((cpi)->cpi_extd[1].cp_edx)
252 #define	CPI_FEATURES_XTD_ECX(cpi)	((cpi)->cpi_extd[1].cp_ecx)
253 
254 #define	CPI_BRANDID(cpi)	BITX((cpi)->cpi_std[1].cp_ebx, 7, 0)
255 #define	CPI_CHUNKS(cpi)		BITX((cpi)->cpi_std[1].cp_ebx, 15, 7)
256 #define	CPI_CPU_COUNT(cpi)	BITX((cpi)->cpi_std[1].cp_ebx, 23, 16)
257 #define	CPI_APIC_ID(cpi)	BITX((cpi)->cpi_std[1].cp_ebx, 31, 24)
258 
259 #define	CPI_MAXEAX_MAX		0x100		/* sanity control */
260 #define	CPI_XMAXEAX_MAX		0x80000100
261 #define	CPI_FN4_ECX_MAX		0x20		/* sanity: max fn 4 levels */
262 #define	CPI_FNB_ECX_MAX		0x20		/* sanity: max fn B levels */
263 
264 /*
265  * Function 4 (Deterministic Cache Parameters) macros
266  * Defined by Intel Application Note AP-485
267  */
268 #define	CPI_NUM_CORES(regs)		BITX((regs)->cp_eax, 31, 26)
269 #define	CPI_NTHR_SHR_CACHE(regs)	BITX((regs)->cp_eax, 25, 14)
270 #define	CPI_FULL_ASSOC_CACHE(regs)	BITX((regs)->cp_eax, 9, 9)
271 #define	CPI_SELF_INIT_CACHE(regs)	BITX((regs)->cp_eax, 8, 8)
272 #define	CPI_CACHE_LVL(regs)		BITX((regs)->cp_eax, 7, 5)
273 #define	CPI_CACHE_TYPE(regs)		BITX((regs)->cp_eax, 4, 0)
274 #define	CPI_CPU_LEVEL_TYPE(regs)	BITX((regs)->cp_ecx, 15, 8)
275 
276 #define	CPI_CACHE_WAYS(regs)		BITX((regs)->cp_ebx, 31, 22)
277 #define	CPI_CACHE_PARTS(regs)		BITX((regs)->cp_ebx, 21, 12)
278 #define	CPI_CACHE_COH_LN_SZ(regs)	BITX((regs)->cp_ebx, 11, 0)
279 
280 #define	CPI_CACHE_SETS(regs)		BITX((regs)->cp_ecx, 31, 0)
281 
282 #define	CPI_PREFCH_STRIDE(regs)		BITX((regs)->cp_edx, 9, 0)
283 
284 
285 /*
286  * A couple of shorthand macros to identify "later" P6-family chips
287  * like the Pentium M and Core.  First, the "older" P6-based stuff
288  * (loosely defined as "pre-Pentium-4"):
289  * P6, PII, Mobile PII, PII Xeon, PIII, Mobile PIII, PIII Xeon
290  */
291 
292 #define	IS_LEGACY_P6(cpi) (			\
293 	cpi->cpi_family == 6 && 		\
294 		(cpi->cpi_model == 1 ||		\
295 		cpi->cpi_model == 3 ||		\
296 		cpi->cpi_model == 5 ||		\
297 		cpi->cpi_model == 6 ||		\
298 		cpi->cpi_model == 7 ||		\
299 		cpi->cpi_model == 8 ||		\
300 		cpi->cpi_model == 0xA ||	\
301 		cpi->cpi_model == 0xB)		\
302 )
303 
304 /* A "new F6" is everything with family 6 that's not the above */
305 #define	IS_NEW_F6(cpi) ((cpi->cpi_family == 6) && !IS_LEGACY_P6(cpi))
306 
307 /* Extended family/model support */
308 #define	IS_EXTENDED_MODEL_INTEL(cpi) (cpi->cpi_family == 0x6 || \
309 	cpi->cpi_family >= 0xf)
310 
311 /*
312  * Info for monitor/mwait idle loop.
313  *
314  * See cpuid section of "Intel 64 and IA-32 Architectures Software Developer's
315  * Manual Volume 2A: Instruction Set Reference, A-M" #25366-022US, November
316  * 2006.
317  * See MONITOR/MWAIT section of "AMD64 Architecture Programmer's Manual
318  * Documentation Updates" #33633, Rev 2.05, December 2006.
319  */
320 #define	MWAIT_SUPPORT		(0x00000001)	/* mwait supported */
321 #define	MWAIT_EXTENSIONS	(0x00000002)	/* extenstion supported */
322 #define	MWAIT_ECX_INT_ENABLE	(0x00000004)	/* ecx 1 extension supported */
323 #define	MWAIT_SUPPORTED(cpi)	((cpi)->cpi_std[1].cp_ecx & CPUID_INTC_ECX_MON)
324 #define	MWAIT_INT_ENABLE(cpi)	((cpi)->cpi_std[5].cp_ecx & 0x2)
325 #define	MWAIT_EXTENSION(cpi)	((cpi)->cpi_std[5].cp_ecx & 0x1)
326 #define	MWAIT_SIZE_MIN(cpi)	BITX((cpi)->cpi_std[5].cp_eax, 15, 0)
327 #define	MWAIT_SIZE_MAX(cpi)	BITX((cpi)->cpi_std[5].cp_ebx, 15, 0)
328 /*
329  * Number of sub-cstates for a given c-state.
330  */
331 #define	MWAIT_NUM_SUBC_STATES(cpi, c_state)			\
332 	BITX((cpi)->cpi_std[5].cp_edx, c_state + 3, c_state)
333 
334 /*
335  * Functions we consune from cpuid_subr.c;  don't publish these in a header
336  * file to try and keep people using the expected cpuid_* interfaces.
337  */
338 extern uint32_t _cpuid_skt(uint_t, uint_t, uint_t, uint_t);
339 extern const char *_cpuid_sktstr(uint_t, uint_t, uint_t, uint_t);
340 extern uint32_t _cpuid_chiprev(uint_t, uint_t, uint_t, uint_t);
341 extern const char *_cpuid_chiprevstr(uint_t, uint_t, uint_t, uint_t);
342 extern uint_t _cpuid_vendorstr_to_vendorcode(char *);
343 
344 /*
345  * Apply up various platform-dependent restrictions where the
346  * underlying platform restrictions mean the CPU can be marked
347  * as less capable than its cpuid instruction would imply.
348  */
349 #if defined(__xpv)
350 static void
351 platform_cpuid_mangle(uint_t vendor, uint32_t eax, struct cpuid_regs *cp)
352 {
353 	switch (eax) {
354 	case 1: {
355 		uint32_t mcamask = DOMAIN_IS_INITDOMAIN(xen_info) ?
356 		    0 : CPUID_INTC_EDX_MCA;
357 		cp->cp_edx &=
358 		    ~(mcamask |
359 		    CPUID_INTC_EDX_PSE |
360 		    CPUID_INTC_EDX_VME | CPUID_INTC_EDX_DE |
361 		    CPUID_INTC_EDX_SEP | CPUID_INTC_EDX_MTRR |
362 		    CPUID_INTC_EDX_PGE | CPUID_INTC_EDX_PAT |
363 		    CPUID_AMD_EDX_SYSC | CPUID_INTC_EDX_SEP |
364 		    CPUID_INTC_EDX_PSE36 | CPUID_INTC_EDX_HTT);
365 		break;
366 	}
367 
368 	case 0x80000001:
369 		cp->cp_edx &=
370 		    ~(CPUID_AMD_EDX_PSE |
371 		    CPUID_INTC_EDX_VME | CPUID_INTC_EDX_DE |
372 		    CPUID_AMD_EDX_MTRR | CPUID_AMD_EDX_PGE |
373 		    CPUID_AMD_EDX_PAT | CPUID_AMD_EDX_PSE36 |
374 		    CPUID_AMD_EDX_SYSC | CPUID_INTC_EDX_SEP |
375 		    CPUID_AMD_EDX_TSCP);
376 		cp->cp_ecx &= ~CPUID_AMD_ECX_CMP_LGCY;
377 		break;
378 	default:
379 		break;
380 	}
381 
382 	switch (vendor) {
383 	case X86_VENDOR_Intel:
384 		switch (eax) {
385 		case 4:
386 			/*
387 			 * Zero out the (ncores-per-chip - 1) field
388 			 */
389 			cp->cp_eax &= 0x03fffffff;
390 			break;
391 		default:
392 			break;
393 		}
394 		break;
395 	case X86_VENDOR_AMD:
396 		switch (eax) {
397 
398 		case 0x80000001:
399 			cp->cp_ecx &= ~CPUID_AMD_ECX_CR8D;
400 			break;
401 
402 		case 0x80000008:
403 			/*
404 			 * Zero out the (ncores-per-chip - 1) field
405 			 */
406 			cp->cp_ecx &= 0xffffff00;
407 			break;
408 		default:
409 			break;
410 		}
411 		break;
412 	default:
413 		break;
414 	}
415 }
416 #else
417 #define	platform_cpuid_mangle(vendor, eax, cp)	/* nothing */
418 #endif
419 
420 /*
421  *  Some undocumented ways of patching the results of the cpuid
422  *  instruction to permit running Solaris 10 on future cpus that
423  *  we don't currently support.  Could be set to non-zero values
424  *  via settings in eeprom.
425  */
426 
427 uint32_t cpuid_feature_ecx_include;
428 uint32_t cpuid_feature_ecx_exclude;
429 uint32_t cpuid_feature_edx_include;
430 uint32_t cpuid_feature_edx_exclude;
431 
432 /*
433  * Allocate space for mcpu_cpi in the machcpu structure for all non-boot CPUs.
434  */
435 void
436 cpuid_alloc_space(cpu_t *cpu)
437 {
438 	/*
439 	 * By convention, cpu0 is the boot cpu, which is set up
440 	 * before memory allocation is available.  All other cpus get
441 	 * their cpuid_info struct allocated here.
442 	 */
443 	ASSERT(cpu->cpu_id != 0);
444 	ASSERT(cpu->cpu_m.mcpu_cpi == NULL);
445 	cpu->cpu_m.mcpu_cpi =
446 	    kmem_zalloc(sizeof (*cpu->cpu_m.mcpu_cpi), KM_SLEEP);
447 }
448 
449 void
450 cpuid_free_space(cpu_t *cpu)
451 {
452 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
453 	int i;
454 
455 	ASSERT(cpi != NULL);
456 	ASSERT(cpi != &cpuid_info0);
457 
458 	/*
459 	 * Free up any function 4 related dynamic storage
460 	 */
461 	for (i = 1; i < cpi->cpi_std_4_size; i++)
462 		kmem_free(cpi->cpi_std_4[i], sizeof (struct cpuid_regs));
463 	if (cpi->cpi_std_4_size > 0)
464 		kmem_free(cpi->cpi_std_4,
465 		    cpi->cpi_std_4_size * sizeof (struct cpuid_regs *));
466 
467 	kmem_free(cpi, sizeof (*cpi));
468 	cpu->cpu_m.mcpu_cpi = NULL;
469 }
470 
471 #if !defined(__xpv)
472 
473 static void
474 determine_platform()
475 {
476 	struct cpuid_regs cp;
477 	char *xen_str;
478 	uint32_t xen_signature[4], base;
479 
480 	platform_type = HW_NATIVE;
481 
482 	if (!enable_platform_detection)
483 		return;
484 
485 	/*
486 	 * In a fully virtualized domain, Xen's pseudo-cpuid function
487 	 * returns a string representing the Xen signature in %ebx, %ecx,
488 	 * and %edx. %eax contains the maximum supported cpuid function.
489 	 * We need at least a (base + 2) leaf value to do what we want
490 	 * to do. Try different base values, since the hypervisor might
491 	 * use a different one depending on whether hyper-v emulation
492 	 * is switched on by default or not.
493 	 */
494 	for (base = 0x40000000; base < 0x40010000; base += 0x100) {
495 		cp.cp_eax = base;
496 		(void) __cpuid_insn(&cp);
497 		xen_signature[0] = cp.cp_ebx;
498 		xen_signature[1] = cp.cp_ecx;
499 		xen_signature[2] = cp.cp_edx;
500 		xen_signature[3] = 0;
501 		xen_str = (char *)xen_signature;
502 		if (strcmp("XenVMMXenVMM", xen_str) == 0 &&
503 		    cp.cp_eax >= (base + 2)) {
504 			platform_type = HW_XEN_HVM;
505 			return;
506 		}
507 	}
508 
509 	if (vmware_platform()) /* running under vmware hypervisor? */
510 		platform_type = HW_VMWARE;
511 }
512 
513 int
514 get_hwenv(void)
515 {
516 	if (platform_type == -1)
517 		determine_platform();
518 
519 	return (platform_type);
520 }
521 
522 int
523 is_controldom(void)
524 {
525 	return (0);
526 }
527 
528 #else
529 
530 int
531 get_hwenv(void)
532 {
533 	return (HW_XEN_PV);
534 }
535 
536 int
537 is_controldom(void)
538 {
539 	return (DOMAIN_IS_INITDOMAIN(xen_info));
540 }
541 
542 #endif	/* __xpv */
543 
544 static void
545 cpuid_intel_getids(cpu_t *cpu, uint_t feature)
546 {
547 	uint_t i;
548 	uint_t chipid_shift = 0;
549 	uint_t coreid_shift = 0;
550 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
551 
552 	for (i = 1; i < cpi->cpi_ncpu_per_chip; i <<= 1)
553 		chipid_shift++;
554 
555 	cpi->cpi_chipid = cpi->cpi_apicid >> chipid_shift;
556 	cpi->cpi_clogid = cpi->cpi_apicid & ((1 << chipid_shift) - 1);
557 
558 	if (feature & X86_CMP) {
559 		/*
560 		 * Multi-core (and possibly multi-threaded)
561 		 * processors.
562 		 */
563 		uint_t ncpu_per_core;
564 		if (cpi->cpi_ncore_per_chip == 1)
565 			ncpu_per_core = cpi->cpi_ncpu_per_chip;
566 		else if (cpi->cpi_ncore_per_chip > 1)
567 			ncpu_per_core = cpi->cpi_ncpu_per_chip /
568 			    cpi->cpi_ncore_per_chip;
569 		/*
570 		 * 8bit APIC IDs on dual core Pentiums
571 		 * look like this:
572 		 *
573 		 * +-----------------------+------+------+
574 		 * | Physical Package ID   |  MC  |  HT  |
575 		 * +-----------------------+------+------+
576 		 * <------- chipid -------->
577 		 * <------- coreid --------------->
578 		 *			   <--- clogid -->
579 		 *			   <------>
580 		 *			   pkgcoreid
581 		 *
582 		 * Where the number of bits necessary to
583 		 * represent MC and HT fields together equals
584 		 * to the minimum number of bits necessary to
585 		 * store the value of cpi->cpi_ncpu_per_chip.
586 		 * Of those bits, the MC part uses the number
587 		 * of bits necessary to store the value of
588 		 * cpi->cpi_ncore_per_chip.
589 		 */
590 		for (i = 1; i < ncpu_per_core; i <<= 1)
591 			coreid_shift++;
592 		cpi->cpi_coreid = cpi->cpi_apicid >> coreid_shift;
593 		cpi->cpi_pkgcoreid = cpi->cpi_clogid >> coreid_shift;
594 	} else if (feature & X86_HTT) {
595 		/*
596 		 * Single-core multi-threaded processors.
597 		 */
598 		cpi->cpi_coreid = cpi->cpi_chipid;
599 		cpi->cpi_pkgcoreid = 0;
600 	}
601 	cpi->cpi_procnodeid = cpi->cpi_chipid;
602 }
603 
604 static void
605 cpuid_amd_getids(cpu_t *cpu)
606 {
607 	int i, first_half, coreidsz;
608 	uint32_t nb_caps_reg;
609 	uint_t node2_1;
610 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
611 
612 	/*
613 	 * AMD CMP chips currently have a single thread per core.
614 	 *
615 	 * Since no two cpus share a core we must assign a distinct coreid
616 	 * per cpu, and we do this by using the cpu_id.  This scheme does not,
617 	 * however, guarantee that sibling cores of a chip will have sequential
618 	 * coreids starting at a multiple of the number of cores per chip -
619 	 * that is usually the case, but if the ACPI MADT table is presented
620 	 * in a different order then we need to perform a few more gymnastics
621 	 * for the pkgcoreid.
622 	 *
623 	 * All processors in the system have the same number of enabled
624 	 * cores. Cores within a processor are always numbered sequentially
625 	 * from 0 regardless of how many or which are disabled, and there
626 	 * is no way for operating system to discover the real core id when some
627 	 * are disabled.
628 	 */
629 
630 	cpi->cpi_coreid = cpu->cpu_id;
631 
632 	if (cpi->cpi_xmaxeax >= 0x80000008) {
633 
634 		coreidsz = BITX((cpi)->cpi_extd[8].cp_ecx, 15, 12);
635 
636 		/*
637 		 * In AMD parlance chip is really a node while Solaris
638 		 * sees chip as equivalent to socket/package.
639 		 */
640 		cpi->cpi_ncore_per_chip =
641 		    BITX((cpi)->cpi_extd[8].cp_ecx, 7, 0) + 1;
642 		if (coreidsz == 0) {
643 			/* Use legacy method */
644 			for (i = 1; i < cpi->cpi_ncore_per_chip; i <<= 1)
645 				coreidsz++;
646 			if (coreidsz == 0)
647 				coreidsz = 1;
648 		}
649 	} else {
650 		/* Assume single-core part */
651 		cpi->cpi_ncore_per_chip = 1;
652 	}
653 
654 	cpi->cpi_clogid = cpi->cpi_pkgcoreid =
655 	    cpi->cpi_apicid & ((1<<coreidsz) - 1);
656 	cpi->cpi_ncpu_per_chip = cpi->cpi_ncore_per_chip;
657 
658 	/* Get nodeID */
659 	if (cpi->cpi_family == 0xf) {
660 		cpi->cpi_procnodeid = (cpi->cpi_apicid >> coreidsz) & 7;
661 		cpi->cpi_chipid = cpi->cpi_procnodeid;
662 	} else if (cpi->cpi_family == 0x10) {
663 		/*
664 		 * See if we are a multi-node processor.
665 		 * All processors in the system have the same number of nodes
666 		 */
667 		nb_caps_reg =  pci_getl_func(0, 24, 3, 0xe8);
668 		if ((cpi->cpi_model < 8) || BITX(nb_caps_reg, 29, 29) == 0) {
669 			/* Single-node */
670 			cpi->cpi_procnodeid = BITX(cpi->cpi_apicid, 5,
671 			    coreidsz);
672 			cpi->cpi_chipid = cpi->cpi_procnodeid;
673 		} else {
674 
675 			/*
676 			 * Multi-node revision D (2 nodes per package
677 			 * are supported)
678 			 */
679 			cpi->cpi_procnodes_per_pkg = 2;
680 
681 			first_half = (cpi->cpi_pkgcoreid <=
682 			    (cpi->cpi_ncore_per_chip/2 - 1));
683 
684 			if (cpi->cpi_apicid == cpi->cpi_pkgcoreid) {
685 				/* We are BSP */
686 				cpi->cpi_procnodeid = (first_half ? 0 : 1);
687 				cpi->cpi_chipid = cpi->cpi_procnodeid >> 1;
688 			} else {
689 
690 				/* We are AP */
691 				/* NodeId[2:1] bits to use for reading F3xe8 */
692 				node2_1 = BITX(cpi->cpi_apicid, 5, 4) << 1;
693 
694 				nb_caps_reg =
695 				    pci_getl_func(0, 24 + node2_1, 3, 0xe8);
696 
697 				/*
698 				 * Check IntNodeNum bit (31:30, but bit 31 is
699 				 * always 0 on dual-node processors)
700 				 */
701 				if (BITX(nb_caps_reg, 30, 30) == 0)
702 					cpi->cpi_procnodeid = node2_1 +
703 					    !first_half;
704 				else
705 					cpi->cpi_procnodeid = node2_1 +
706 					    first_half;
707 
708 				cpi->cpi_chipid = cpi->cpi_procnodeid >> 1;
709 			}
710 		}
711 	} else if (cpi->cpi_family >= 0x11) {
712 		cpi->cpi_procnodeid = (cpi->cpi_apicid >> coreidsz) & 7;
713 		cpi->cpi_chipid = cpi->cpi_procnodeid;
714 	} else {
715 		cpi->cpi_procnodeid = 0;
716 		cpi->cpi_chipid = cpi->cpi_procnodeid;
717 	}
718 }
719 
720 uint_t
721 cpuid_pass1(cpu_t *cpu)
722 {
723 	uint32_t mask_ecx, mask_edx;
724 	uint_t feature = X86_CPUID;
725 	struct cpuid_info *cpi;
726 	struct cpuid_regs *cp;
727 	int xcpuid;
728 #if !defined(__xpv)
729 	extern int idle_cpu_prefer_mwait;
730 #endif
731 
732 
733 #if !defined(__xpv)
734 	determine_platform();
735 #endif
736 	/*
737 	 * Space statically allocated for BSP, ensure pointer is set
738 	 */
739 	if (cpu->cpu_id == 0 && cpu->cpu_m.mcpu_cpi == NULL)
740 		cpu->cpu_m.mcpu_cpi = &cpuid_info0;
741 	cpi = cpu->cpu_m.mcpu_cpi;
742 	ASSERT(cpi != NULL);
743 	cp = &cpi->cpi_std[0];
744 	cp->cp_eax = 0;
745 	cpi->cpi_maxeax = __cpuid_insn(cp);
746 	{
747 		uint32_t *iptr = (uint32_t *)cpi->cpi_vendorstr;
748 		*iptr++ = cp->cp_ebx;
749 		*iptr++ = cp->cp_edx;
750 		*iptr++ = cp->cp_ecx;
751 		*(char *)&cpi->cpi_vendorstr[12] = '\0';
752 	}
753 
754 	cpi->cpi_vendor = _cpuid_vendorstr_to_vendorcode(cpi->cpi_vendorstr);
755 	x86_vendor = cpi->cpi_vendor; /* for compatibility */
756 
757 	/*
758 	 * Limit the range in case of weird hardware
759 	 */
760 	if (cpi->cpi_maxeax > CPI_MAXEAX_MAX)
761 		cpi->cpi_maxeax = CPI_MAXEAX_MAX;
762 	if (cpi->cpi_maxeax < 1)
763 		goto pass1_done;
764 
765 	cp = &cpi->cpi_std[1];
766 	cp->cp_eax = 1;
767 	(void) __cpuid_insn(cp);
768 
769 	/*
770 	 * Extract identifying constants for easy access.
771 	 */
772 	cpi->cpi_model = CPI_MODEL(cpi);
773 	cpi->cpi_family = CPI_FAMILY(cpi);
774 
775 	if (cpi->cpi_family == 0xf)
776 		cpi->cpi_family += CPI_FAMILY_XTD(cpi);
777 
778 	/*
779 	 * Beware: AMD uses "extended model" iff base *FAMILY* == 0xf.
780 	 * Intel, and presumably everyone else, uses model == 0xf, as
781 	 * one would expect (max value means possible overflow).  Sigh.
782 	 */
783 
784 	switch (cpi->cpi_vendor) {
785 	case X86_VENDOR_Intel:
786 		if (IS_EXTENDED_MODEL_INTEL(cpi))
787 			cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4;
788 		break;
789 	case X86_VENDOR_AMD:
790 		if (CPI_FAMILY(cpi) == 0xf)
791 			cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4;
792 		break;
793 	default:
794 		if (cpi->cpi_model == 0xf)
795 			cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4;
796 		break;
797 	}
798 
799 	cpi->cpi_step = CPI_STEP(cpi);
800 	cpi->cpi_brandid = CPI_BRANDID(cpi);
801 
802 	/*
803 	 * *default* assumptions:
804 	 * - believe %edx feature word
805 	 * - ignore %ecx feature word
806 	 * - 32-bit virtual and physical addressing
807 	 */
808 	mask_edx = 0xffffffff;
809 	mask_ecx = 0;
810 
811 	cpi->cpi_pabits = cpi->cpi_vabits = 32;
812 
813 	switch (cpi->cpi_vendor) {
814 	case X86_VENDOR_Intel:
815 		if (cpi->cpi_family == 5)
816 			x86_type = X86_TYPE_P5;
817 		else if (IS_LEGACY_P6(cpi)) {
818 			x86_type = X86_TYPE_P6;
819 			pentiumpro_bug4046376 = 1;
820 			pentiumpro_bug4064495 = 1;
821 			/*
822 			 * Clear the SEP bit when it was set erroneously
823 			 */
824 			if (cpi->cpi_model < 3 && cpi->cpi_step < 3)
825 				cp->cp_edx &= ~CPUID_INTC_EDX_SEP;
826 		} else if (IS_NEW_F6(cpi) || cpi->cpi_family == 0xf) {
827 			x86_type = X86_TYPE_P4;
828 			/*
829 			 * We don't currently depend on any of the %ecx
830 			 * features until Prescott, so we'll only check
831 			 * this from P4 onwards.  We might want to revisit
832 			 * that idea later.
833 			 */
834 			mask_ecx = 0xffffffff;
835 		} else if (cpi->cpi_family > 0xf)
836 			mask_ecx = 0xffffffff;
837 		/*
838 		 * We don't support MONITOR/MWAIT if leaf 5 is not available
839 		 * to obtain the monitor linesize.
840 		 */
841 		if (cpi->cpi_maxeax < 5)
842 			mask_ecx &= ~CPUID_INTC_ECX_MON;
843 		break;
844 	case X86_VENDOR_IntelClone:
845 	default:
846 		break;
847 	case X86_VENDOR_AMD:
848 #if defined(OPTERON_ERRATUM_108)
849 		if (cpi->cpi_family == 0xf && cpi->cpi_model == 0xe) {
850 			cp->cp_eax = (0xf0f & cp->cp_eax) | 0xc0;
851 			cpi->cpi_model = 0xc;
852 		} else
853 #endif
854 		if (cpi->cpi_family == 5) {
855 			/*
856 			 * AMD K5 and K6
857 			 *
858 			 * These CPUs have an incomplete implementation
859 			 * of MCA/MCE which we mask away.
860 			 */
861 			mask_edx &= ~(CPUID_INTC_EDX_MCE | CPUID_INTC_EDX_MCA);
862 
863 			/*
864 			 * Model 0 uses the wrong (APIC) bit
865 			 * to indicate PGE.  Fix it here.
866 			 */
867 			if (cpi->cpi_model == 0) {
868 				if (cp->cp_edx & 0x200) {
869 					cp->cp_edx &= ~0x200;
870 					cp->cp_edx |= CPUID_INTC_EDX_PGE;
871 				}
872 			}
873 
874 			/*
875 			 * Early models had problems w/ MMX; disable.
876 			 */
877 			if (cpi->cpi_model < 6)
878 				mask_edx &= ~CPUID_INTC_EDX_MMX;
879 		}
880 
881 		/*
882 		 * For newer families, SSE3 and CX16, at least, are valid;
883 		 * enable all
884 		 */
885 		if (cpi->cpi_family >= 0xf)
886 			mask_ecx = 0xffffffff;
887 		/*
888 		 * We don't support MONITOR/MWAIT if leaf 5 is not available
889 		 * to obtain the monitor linesize.
890 		 */
891 		if (cpi->cpi_maxeax < 5)
892 			mask_ecx &= ~CPUID_INTC_ECX_MON;
893 
894 #if !defined(__xpv)
895 		/*
896 		 * Do not use MONITOR/MWAIT to halt in the idle loop on any AMD
897 		 * processors.  AMD does not intend MWAIT to be used in the cpu
898 		 * idle loop on current and future processors.  10h and future
899 		 * AMD processors use more power in MWAIT than HLT.
900 		 * Pre-family-10h Opterons do not have the MWAIT instruction.
901 		 */
902 		idle_cpu_prefer_mwait = 0;
903 #endif
904 
905 		break;
906 	case X86_VENDOR_TM:
907 		/*
908 		 * workaround the NT workaround in CMS 4.1
909 		 */
910 		if (cpi->cpi_family == 5 && cpi->cpi_model == 4 &&
911 		    (cpi->cpi_step == 2 || cpi->cpi_step == 3))
912 			cp->cp_edx |= CPUID_INTC_EDX_CX8;
913 		break;
914 	case X86_VENDOR_Centaur:
915 		/*
916 		 * workaround the NT workarounds again
917 		 */
918 		if (cpi->cpi_family == 6)
919 			cp->cp_edx |= CPUID_INTC_EDX_CX8;
920 		break;
921 	case X86_VENDOR_Cyrix:
922 		/*
923 		 * We rely heavily on the probing in locore
924 		 * to actually figure out what parts, if any,
925 		 * of the Cyrix cpuid instruction to believe.
926 		 */
927 		switch (x86_type) {
928 		case X86_TYPE_CYRIX_486:
929 			mask_edx = 0;
930 			break;
931 		case X86_TYPE_CYRIX_6x86:
932 			mask_edx = 0;
933 			break;
934 		case X86_TYPE_CYRIX_6x86L:
935 			mask_edx =
936 			    CPUID_INTC_EDX_DE |
937 			    CPUID_INTC_EDX_CX8;
938 			break;
939 		case X86_TYPE_CYRIX_6x86MX:
940 			mask_edx =
941 			    CPUID_INTC_EDX_DE |
942 			    CPUID_INTC_EDX_MSR |
943 			    CPUID_INTC_EDX_CX8 |
944 			    CPUID_INTC_EDX_PGE |
945 			    CPUID_INTC_EDX_CMOV |
946 			    CPUID_INTC_EDX_MMX;
947 			break;
948 		case X86_TYPE_CYRIX_GXm:
949 			mask_edx =
950 			    CPUID_INTC_EDX_MSR |
951 			    CPUID_INTC_EDX_CX8 |
952 			    CPUID_INTC_EDX_CMOV |
953 			    CPUID_INTC_EDX_MMX;
954 			break;
955 		case X86_TYPE_CYRIX_MediaGX:
956 			break;
957 		case X86_TYPE_CYRIX_MII:
958 		case X86_TYPE_VIA_CYRIX_III:
959 			mask_edx =
960 			    CPUID_INTC_EDX_DE |
961 			    CPUID_INTC_EDX_TSC |
962 			    CPUID_INTC_EDX_MSR |
963 			    CPUID_INTC_EDX_CX8 |
964 			    CPUID_INTC_EDX_PGE |
965 			    CPUID_INTC_EDX_CMOV |
966 			    CPUID_INTC_EDX_MMX;
967 			break;
968 		default:
969 			break;
970 		}
971 		break;
972 	}
973 
974 #if defined(__xpv)
975 	/*
976 	 * Do not support MONITOR/MWAIT under a hypervisor
977 	 */
978 	mask_ecx &= ~CPUID_INTC_ECX_MON;
979 #endif	/* __xpv */
980 
981 	/*
982 	 * Now we've figured out the masks that determine
983 	 * which bits we choose to believe, apply the masks
984 	 * to the feature words, then map the kernel's view
985 	 * of these feature words into its feature word.
986 	 */
987 	cp->cp_edx &= mask_edx;
988 	cp->cp_ecx &= mask_ecx;
989 
990 	/*
991 	 * apply any platform restrictions (we don't call this
992 	 * immediately after __cpuid_insn here, because we need the
993 	 * workarounds applied above first)
994 	 */
995 	platform_cpuid_mangle(cpi->cpi_vendor, 1, cp);
996 
997 	/*
998 	 * fold in overrides from the "eeprom" mechanism
999 	 */
1000 	cp->cp_edx |= cpuid_feature_edx_include;
1001 	cp->cp_edx &= ~cpuid_feature_edx_exclude;
1002 
1003 	cp->cp_ecx |= cpuid_feature_ecx_include;
1004 	cp->cp_ecx &= ~cpuid_feature_ecx_exclude;
1005 
1006 	if (cp->cp_edx & CPUID_INTC_EDX_PSE)
1007 		feature |= X86_LARGEPAGE;
1008 	if (cp->cp_edx & CPUID_INTC_EDX_TSC)
1009 		feature |= X86_TSC;
1010 	if (cp->cp_edx & CPUID_INTC_EDX_MSR)
1011 		feature |= X86_MSR;
1012 	if (cp->cp_edx & CPUID_INTC_EDX_MTRR)
1013 		feature |= X86_MTRR;
1014 	if (cp->cp_edx & CPUID_INTC_EDX_PGE)
1015 		feature |= X86_PGE;
1016 	if (cp->cp_edx & CPUID_INTC_EDX_CMOV)
1017 		feature |= X86_CMOV;
1018 	if (cp->cp_edx & CPUID_INTC_EDX_MMX)
1019 		feature |= X86_MMX;
1020 	if ((cp->cp_edx & CPUID_INTC_EDX_MCE) != 0 &&
1021 	    (cp->cp_edx & CPUID_INTC_EDX_MCA) != 0)
1022 		feature |= X86_MCA;
1023 	if (cp->cp_edx & CPUID_INTC_EDX_PAE)
1024 		feature |= X86_PAE;
1025 	if (cp->cp_edx & CPUID_INTC_EDX_CX8)
1026 		feature |= X86_CX8;
1027 	if (cp->cp_ecx & CPUID_INTC_ECX_CX16)
1028 		feature |= X86_CX16;
1029 	if (cp->cp_edx & CPUID_INTC_EDX_PAT)
1030 		feature |= X86_PAT;
1031 	if (cp->cp_edx & CPUID_INTC_EDX_SEP)
1032 		feature |= X86_SEP;
1033 	if (cp->cp_edx & CPUID_INTC_EDX_FXSR) {
1034 		/*
1035 		 * In our implementation, fxsave/fxrstor
1036 		 * are prerequisites before we'll even
1037 		 * try and do SSE things.
1038 		 */
1039 		if (cp->cp_edx & CPUID_INTC_EDX_SSE)
1040 			feature |= X86_SSE;
1041 		if (cp->cp_edx & CPUID_INTC_EDX_SSE2)
1042 			feature |= X86_SSE2;
1043 		if (cp->cp_ecx & CPUID_INTC_ECX_SSE3)
1044 			feature |= X86_SSE3;
1045 		if (cpi->cpi_vendor == X86_VENDOR_Intel) {
1046 			if (cp->cp_ecx & CPUID_INTC_ECX_SSSE3)
1047 				feature |= X86_SSSE3;
1048 			if (cp->cp_ecx & CPUID_INTC_ECX_SSE4_1)
1049 				feature |= X86_SSE4_1;
1050 			if (cp->cp_ecx & CPUID_INTC_ECX_SSE4_2)
1051 				feature |= X86_SSE4_2;
1052 			if (cp->cp_ecx & CPUID_INTC_ECX_AES)
1053 				feature |= X86_AES;
1054 		}
1055 	}
1056 	if (cp->cp_edx & CPUID_INTC_EDX_DE)
1057 		feature |= X86_DE;
1058 #if !defined(__xpv)
1059 	if (cp->cp_ecx & CPUID_INTC_ECX_MON) {
1060 
1061 		/*
1062 		 * We require the CLFLUSH instruction for erratum workaround
1063 		 * to use MONITOR/MWAIT.
1064 		 */
1065 		if (cp->cp_edx & CPUID_INTC_EDX_CLFSH) {
1066 			cpi->cpi_mwait.support |= MWAIT_SUPPORT;
1067 			feature |= X86_MWAIT;
1068 		} else {
1069 			extern int idle_cpu_assert_cflush_monitor;
1070 
1071 			/*
1072 			 * All processors we are aware of which have
1073 			 * MONITOR/MWAIT also have CLFLUSH.
1074 			 */
1075 			if (idle_cpu_assert_cflush_monitor) {
1076 				ASSERT((cp->cp_ecx & CPUID_INTC_ECX_MON) &&
1077 				    (cp->cp_edx & CPUID_INTC_EDX_CLFSH));
1078 			}
1079 		}
1080 	}
1081 #endif	/* __xpv */
1082 
1083 	/*
1084 	 * Only need it first time, rest of the cpus would follow suite.
1085 	 * we only capture this for the bootcpu.
1086 	 */
1087 	if (cp->cp_edx & CPUID_INTC_EDX_CLFSH) {
1088 		feature |= X86_CLFSH;
1089 		x86_clflush_size = (BITX(cp->cp_ebx, 15, 8) * 8);
1090 	}
1091 
1092 	if (feature & X86_PAE)
1093 		cpi->cpi_pabits = 36;
1094 
1095 	/*
1096 	 * Hyperthreading configuration is slightly tricky on Intel
1097 	 * and pure clones, and even trickier on AMD.
1098 	 *
1099 	 * (AMD chose to set the HTT bit on their CMP processors,
1100 	 * even though they're not actually hyperthreaded.  Thus it
1101 	 * takes a bit more work to figure out what's really going
1102 	 * on ... see the handling of the CMP_LGCY bit below)
1103 	 */
1104 	if (cp->cp_edx & CPUID_INTC_EDX_HTT) {
1105 		cpi->cpi_ncpu_per_chip = CPI_CPU_COUNT(cpi);
1106 		if (cpi->cpi_ncpu_per_chip > 1)
1107 			feature |= X86_HTT;
1108 	} else {
1109 		cpi->cpi_ncpu_per_chip = 1;
1110 	}
1111 
1112 	/*
1113 	 * Work on the "extended" feature information, doing
1114 	 * some basic initialization for cpuid_pass2()
1115 	 */
1116 	xcpuid = 0;
1117 	switch (cpi->cpi_vendor) {
1118 	case X86_VENDOR_Intel:
1119 		if (IS_NEW_F6(cpi) || cpi->cpi_family >= 0xf)
1120 			xcpuid++;
1121 		break;
1122 	case X86_VENDOR_AMD:
1123 		if (cpi->cpi_family > 5 ||
1124 		    (cpi->cpi_family == 5 && cpi->cpi_model >= 1))
1125 			xcpuid++;
1126 		break;
1127 	case X86_VENDOR_Cyrix:
1128 		/*
1129 		 * Only these Cyrix CPUs are -known- to support
1130 		 * extended cpuid operations.
1131 		 */
1132 		if (x86_type == X86_TYPE_VIA_CYRIX_III ||
1133 		    x86_type == X86_TYPE_CYRIX_GXm)
1134 			xcpuid++;
1135 		break;
1136 	case X86_VENDOR_Centaur:
1137 	case X86_VENDOR_TM:
1138 	default:
1139 		xcpuid++;
1140 		break;
1141 	}
1142 
1143 	if (xcpuid) {
1144 		cp = &cpi->cpi_extd[0];
1145 		cp->cp_eax = 0x80000000;
1146 		cpi->cpi_xmaxeax = __cpuid_insn(cp);
1147 	}
1148 
1149 	if (cpi->cpi_xmaxeax & 0x80000000) {
1150 
1151 		if (cpi->cpi_xmaxeax > CPI_XMAXEAX_MAX)
1152 			cpi->cpi_xmaxeax = CPI_XMAXEAX_MAX;
1153 
1154 		switch (cpi->cpi_vendor) {
1155 		case X86_VENDOR_Intel:
1156 		case X86_VENDOR_AMD:
1157 			if (cpi->cpi_xmaxeax < 0x80000001)
1158 				break;
1159 			cp = &cpi->cpi_extd[1];
1160 			cp->cp_eax = 0x80000001;
1161 			(void) __cpuid_insn(cp);
1162 
1163 			if (cpi->cpi_vendor == X86_VENDOR_AMD &&
1164 			    cpi->cpi_family == 5 &&
1165 			    cpi->cpi_model == 6 &&
1166 			    cpi->cpi_step == 6) {
1167 				/*
1168 				 * K6 model 6 uses bit 10 to indicate SYSC
1169 				 * Later models use bit 11. Fix it here.
1170 				 */
1171 				if (cp->cp_edx & 0x400) {
1172 					cp->cp_edx &= ~0x400;
1173 					cp->cp_edx |= CPUID_AMD_EDX_SYSC;
1174 				}
1175 			}
1176 
1177 			platform_cpuid_mangle(cpi->cpi_vendor, 0x80000001, cp);
1178 
1179 			/*
1180 			 * Compute the additions to the kernel's feature word.
1181 			 */
1182 			if (cp->cp_edx & CPUID_AMD_EDX_NX)
1183 				feature |= X86_NX;
1184 
1185 			/*
1186 			 * Regardless whether or not we boot 64-bit,
1187 			 * we should have a way to identify whether
1188 			 * the CPU is capable of running 64-bit.
1189 			 */
1190 			if (cp->cp_edx & CPUID_AMD_EDX_LM)
1191 				feature |= X86_64;
1192 
1193 #if defined(__amd64)
1194 			/* 1 GB large page - enable only for 64 bit kernel */
1195 			if (cp->cp_edx & CPUID_AMD_EDX_1GPG)
1196 				feature |= X86_1GPG;
1197 #endif
1198 
1199 			if ((cpi->cpi_vendor == X86_VENDOR_AMD) &&
1200 			    (cpi->cpi_std[1].cp_edx & CPUID_INTC_EDX_FXSR) &&
1201 			    (cp->cp_ecx & CPUID_AMD_ECX_SSE4A))
1202 				feature |= X86_SSE4A;
1203 
1204 			/*
1205 			 * If both the HTT and CMP_LGCY bits are set,
1206 			 * then we're not actually HyperThreaded.  Read
1207 			 * "AMD CPUID Specification" for more details.
1208 			 */
1209 			if (cpi->cpi_vendor == X86_VENDOR_AMD &&
1210 			    (feature & X86_HTT) &&
1211 			    (cp->cp_ecx & CPUID_AMD_ECX_CMP_LGCY)) {
1212 				feature &= ~X86_HTT;
1213 				feature |= X86_CMP;
1214 			}
1215 #if defined(__amd64)
1216 			/*
1217 			 * It's really tricky to support syscall/sysret in
1218 			 * the i386 kernel; we rely on sysenter/sysexit
1219 			 * instead.  In the amd64 kernel, things are -way-
1220 			 * better.
1221 			 */
1222 			if (cp->cp_edx & CPUID_AMD_EDX_SYSC)
1223 				feature |= X86_ASYSC;
1224 
1225 			/*
1226 			 * While we're thinking about system calls, note
1227 			 * that AMD processors don't support sysenter
1228 			 * in long mode at all, so don't try to program them.
1229 			 */
1230 			if (x86_vendor == X86_VENDOR_AMD)
1231 				feature &= ~X86_SEP;
1232 #endif
1233 			if (cp->cp_edx & CPUID_AMD_EDX_TSCP)
1234 				feature |= X86_TSCP;
1235 			break;
1236 		default:
1237 			break;
1238 		}
1239 
1240 		/*
1241 		 * Get CPUID data about processor cores and hyperthreads.
1242 		 */
1243 		switch (cpi->cpi_vendor) {
1244 		case X86_VENDOR_Intel:
1245 			if (cpi->cpi_maxeax >= 4) {
1246 				cp = &cpi->cpi_std[4];
1247 				cp->cp_eax = 4;
1248 				cp->cp_ecx = 0;
1249 				(void) __cpuid_insn(cp);
1250 				platform_cpuid_mangle(cpi->cpi_vendor, 4, cp);
1251 			}
1252 			/*FALLTHROUGH*/
1253 		case X86_VENDOR_AMD:
1254 			if (cpi->cpi_xmaxeax < 0x80000008)
1255 				break;
1256 			cp = &cpi->cpi_extd[8];
1257 			cp->cp_eax = 0x80000008;
1258 			(void) __cpuid_insn(cp);
1259 			platform_cpuid_mangle(cpi->cpi_vendor, 0x80000008, cp);
1260 
1261 			/*
1262 			 * Virtual and physical address limits from
1263 			 * cpuid override previously guessed values.
1264 			 */
1265 			cpi->cpi_pabits = BITX(cp->cp_eax, 7, 0);
1266 			cpi->cpi_vabits = BITX(cp->cp_eax, 15, 8);
1267 			break;
1268 		default:
1269 			break;
1270 		}
1271 
1272 		/*
1273 		 * Derive the number of cores per chip
1274 		 */
1275 		switch (cpi->cpi_vendor) {
1276 		case X86_VENDOR_Intel:
1277 			if (cpi->cpi_maxeax < 4) {
1278 				cpi->cpi_ncore_per_chip = 1;
1279 				break;
1280 			} else {
1281 				cpi->cpi_ncore_per_chip =
1282 				    BITX((cpi)->cpi_std[4].cp_eax, 31, 26) + 1;
1283 			}
1284 			break;
1285 		case X86_VENDOR_AMD:
1286 			if (cpi->cpi_xmaxeax < 0x80000008) {
1287 				cpi->cpi_ncore_per_chip = 1;
1288 				break;
1289 			} else {
1290 				/*
1291 				 * On family 0xf cpuid fn 2 ECX[7:0] "NC" is
1292 				 * 1 less than the number of physical cores on
1293 				 * the chip.  In family 0x10 this value can
1294 				 * be affected by "downcoring" - it reflects
1295 				 * 1 less than the number of cores actually
1296 				 * enabled on this node.
1297 				 */
1298 				cpi->cpi_ncore_per_chip =
1299 				    BITX((cpi)->cpi_extd[8].cp_ecx, 7, 0) + 1;
1300 			}
1301 			break;
1302 		default:
1303 			cpi->cpi_ncore_per_chip = 1;
1304 			break;
1305 		}
1306 
1307 		/*
1308 		 * Get CPUID data about TSC Invariance in Deep C-State.
1309 		 */
1310 		switch (cpi->cpi_vendor) {
1311 		case X86_VENDOR_Intel:
1312 			if (cpi->cpi_maxeax >= 7) {
1313 				cp = &cpi->cpi_extd[7];
1314 				cp->cp_eax = 0x80000007;
1315 				cp->cp_ecx = 0;
1316 				(void) __cpuid_insn(cp);
1317 			}
1318 			break;
1319 		default:
1320 			break;
1321 		}
1322 	} else {
1323 		cpi->cpi_ncore_per_chip = 1;
1324 	}
1325 
1326 	/*
1327 	 * If more than one core, then this processor is CMP.
1328 	 */
1329 	if (cpi->cpi_ncore_per_chip > 1)
1330 		feature |= X86_CMP;
1331 
1332 	/*
1333 	 * If the number of cores is the same as the number
1334 	 * of CPUs, then we cannot have HyperThreading.
1335 	 */
1336 	if (cpi->cpi_ncpu_per_chip == cpi->cpi_ncore_per_chip)
1337 		feature &= ~X86_HTT;
1338 
1339 	cpi->cpi_apicid = CPI_APIC_ID(cpi);
1340 	cpi->cpi_procnodes_per_pkg = 1;
1341 
1342 	if ((feature & (X86_HTT | X86_CMP)) == 0) {
1343 		/*
1344 		 * Single-core single-threaded processors.
1345 		 */
1346 		cpi->cpi_chipid = -1;
1347 		cpi->cpi_clogid = 0;
1348 		cpi->cpi_coreid = cpu->cpu_id;
1349 		cpi->cpi_pkgcoreid = 0;
1350 		if (cpi->cpi_vendor == X86_VENDOR_AMD)
1351 			cpi->cpi_procnodeid = BITX(cpi->cpi_apicid, 3, 0);
1352 		else
1353 			cpi->cpi_procnodeid = cpi->cpi_chipid;
1354 	} else if (cpi->cpi_ncpu_per_chip > 1) {
1355 		if (cpi->cpi_vendor == X86_VENDOR_Intel)
1356 			cpuid_intel_getids(cpu, feature);
1357 		else if (cpi->cpi_vendor == X86_VENDOR_AMD)
1358 			cpuid_amd_getids(cpu);
1359 		else {
1360 			/*
1361 			 * All other processors are currently
1362 			 * assumed to have single cores.
1363 			 */
1364 			cpi->cpi_coreid = cpi->cpi_chipid;
1365 			cpi->cpi_pkgcoreid = 0;
1366 			cpi->cpi_procnodeid = cpi->cpi_chipid;
1367 		}
1368 	}
1369 
1370 	/*
1371 	 * Synthesize chip "revision" and socket type
1372 	 */
1373 	cpi->cpi_chiprev = _cpuid_chiprev(cpi->cpi_vendor, cpi->cpi_family,
1374 	    cpi->cpi_model, cpi->cpi_step);
1375 	cpi->cpi_chiprevstr = _cpuid_chiprevstr(cpi->cpi_vendor,
1376 	    cpi->cpi_family, cpi->cpi_model, cpi->cpi_step);
1377 	cpi->cpi_socket = _cpuid_skt(cpi->cpi_vendor, cpi->cpi_family,
1378 	    cpi->cpi_model, cpi->cpi_step);
1379 
1380 pass1_done:
1381 	cpi->cpi_pass = 1;
1382 	return (feature);
1383 }
1384 
1385 /*
1386  * Make copies of the cpuid table entries we depend on, in
1387  * part for ease of parsing now, in part so that we have only
1388  * one place to correct any of it, in part for ease of
1389  * later export to userland, and in part so we can look at
1390  * this stuff in a crash dump.
1391  */
1392 
1393 /*ARGSUSED*/
1394 void
1395 cpuid_pass2(cpu_t *cpu)
1396 {
1397 	uint_t n, nmax;
1398 	int i;
1399 	struct cpuid_regs *cp;
1400 	uint8_t *dp;
1401 	uint32_t *iptr;
1402 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
1403 
1404 	ASSERT(cpi->cpi_pass == 1);
1405 
1406 	if (cpi->cpi_maxeax < 1)
1407 		goto pass2_done;
1408 
1409 	if ((nmax = cpi->cpi_maxeax + 1) > NMAX_CPI_STD)
1410 		nmax = NMAX_CPI_STD;
1411 	/*
1412 	 * (We already handled n == 0 and n == 1 in pass 1)
1413 	 */
1414 	for (n = 2, cp = &cpi->cpi_std[2]; n < nmax; n++, cp++) {
1415 		cp->cp_eax = n;
1416 
1417 		/*
1418 		 * CPUID function 4 expects %ecx to be initialized
1419 		 * with an index which indicates which cache to return
1420 		 * information about. The OS is expected to call function 4
1421 		 * with %ecx set to 0, 1, 2, ... until it returns with
1422 		 * EAX[4:0] set to 0, which indicates there are no more
1423 		 * caches.
1424 		 *
1425 		 * Here, populate cpi_std[4] with the information returned by
1426 		 * function 4 when %ecx == 0, and do the rest in cpuid_pass3()
1427 		 * when dynamic memory allocation becomes available.
1428 		 *
1429 		 * Note: we need to explicitly initialize %ecx here, since
1430 		 * function 4 may have been previously invoked.
1431 		 */
1432 		if (n == 4)
1433 			cp->cp_ecx = 0;
1434 
1435 		(void) __cpuid_insn(cp);
1436 		platform_cpuid_mangle(cpi->cpi_vendor, n, cp);
1437 		switch (n) {
1438 		case 2:
1439 			/*
1440 			 * "the lower 8 bits of the %eax register
1441 			 * contain a value that identifies the number
1442 			 * of times the cpuid [instruction] has to be
1443 			 * executed to obtain a complete image of the
1444 			 * processor's caching systems."
1445 			 *
1446 			 * How *do* they make this stuff up?
1447 			 */
1448 			cpi->cpi_ncache = sizeof (*cp) *
1449 			    BITX(cp->cp_eax, 7, 0);
1450 			if (cpi->cpi_ncache == 0)
1451 				break;
1452 			cpi->cpi_ncache--;	/* skip count byte */
1453 
1454 			/*
1455 			 * Well, for now, rather than attempt to implement
1456 			 * this slightly dubious algorithm, we just look
1457 			 * at the first 15 ..
1458 			 */
1459 			if (cpi->cpi_ncache > (sizeof (*cp) - 1))
1460 				cpi->cpi_ncache = sizeof (*cp) - 1;
1461 
1462 			dp = cpi->cpi_cacheinfo;
1463 			if (BITX(cp->cp_eax, 31, 31) == 0) {
1464 				uint8_t *p = (void *)&cp->cp_eax;
1465 				for (i = 1; i < 4; i++)
1466 					if (p[i] != 0)
1467 						*dp++ = p[i];
1468 			}
1469 			if (BITX(cp->cp_ebx, 31, 31) == 0) {
1470 				uint8_t *p = (void *)&cp->cp_ebx;
1471 				for (i = 0; i < 4; i++)
1472 					if (p[i] != 0)
1473 						*dp++ = p[i];
1474 			}
1475 			if (BITX(cp->cp_ecx, 31, 31) == 0) {
1476 				uint8_t *p = (void *)&cp->cp_ecx;
1477 				for (i = 0; i < 4; i++)
1478 					if (p[i] != 0)
1479 						*dp++ = p[i];
1480 			}
1481 			if (BITX(cp->cp_edx, 31, 31) == 0) {
1482 				uint8_t *p = (void *)&cp->cp_edx;
1483 				for (i = 0; i < 4; i++)
1484 					if (p[i] != 0)
1485 						*dp++ = p[i];
1486 			}
1487 			break;
1488 
1489 		case 3:	/* Processor serial number, if PSN supported */
1490 			break;
1491 
1492 		case 4:	/* Deterministic cache parameters */
1493 			break;
1494 
1495 		case 5:	/* Monitor/Mwait parameters */
1496 		{
1497 			size_t mwait_size;
1498 
1499 			/*
1500 			 * check cpi_mwait.support which was set in cpuid_pass1
1501 			 */
1502 			if (!(cpi->cpi_mwait.support & MWAIT_SUPPORT))
1503 				break;
1504 
1505 			/*
1506 			 * Protect ourself from insane mwait line size.
1507 			 * Workaround for incomplete hardware emulator(s).
1508 			 */
1509 			mwait_size = (size_t)MWAIT_SIZE_MAX(cpi);
1510 			if (mwait_size < sizeof (uint32_t) ||
1511 			    !ISP2(mwait_size)) {
1512 #if DEBUG
1513 				cmn_err(CE_NOTE, "Cannot handle cpu %d mwait "
1514 				    "size %ld", cpu->cpu_id, (long)mwait_size);
1515 #endif
1516 				break;
1517 			}
1518 
1519 			cpi->cpi_mwait.mon_min = (size_t)MWAIT_SIZE_MIN(cpi);
1520 			cpi->cpi_mwait.mon_max = mwait_size;
1521 			if (MWAIT_EXTENSION(cpi)) {
1522 				cpi->cpi_mwait.support |= MWAIT_EXTENSIONS;
1523 				if (MWAIT_INT_ENABLE(cpi))
1524 					cpi->cpi_mwait.support |=
1525 					    MWAIT_ECX_INT_ENABLE;
1526 			}
1527 			break;
1528 		}
1529 		default:
1530 			break;
1531 		}
1532 	}
1533 
1534 	if (cpi->cpi_maxeax >= 0xB && cpi->cpi_vendor == X86_VENDOR_Intel) {
1535 		struct cpuid_regs regs;
1536 
1537 		cp = &regs;
1538 		cp->cp_eax = 0xB;
1539 		cp->cp_edx = cp->cp_ebx = cp->cp_ecx = 0;
1540 
1541 		(void) __cpuid_insn(cp);
1542 
1543 		/*
1544 		 * Check CPUID.EAX=0BH, ECX=0H:EBX is non-zero, which
1545 		 * indicates that the extended topology enumeration leaf is
1546 		 * available.
1547 		 */
1548 		if (cp->cp_ebx) {
1549 			uint32_t x2apic_id;
1550 			uint_t coreid_shift = 0;
1551 			uint_t ncpu_per_core = 1;
1552 			uint_t chipid_shift = 0;
1553 			uint_t ncpu_per_chip = 1;
1554 			uint_t i;
1555 			uint_t level;
1556 
1557 			for (i = 0; i < CPI_FNB_ECX_MAX; i++) {
1558 				cp->cp_eax = 0xB;
1559 				cp->cp_ecx = i;
1560 
1561 				(void) __cpuid_insn(cp);
1562 				level = CPI_CPU_LEVEL_TYPE(cp);
1563 
1564 				if (level == 1) {
1565 					x2apic_id = cp->cp_edx;
1566 					coreid_shift = BITX(cp->cp_eax, 4, 0);
1567 					ncpu_per_core = BITX(cp->cp_ebx, 15, 0);
1568 				} else if (level == 2) {
1569 					x2apic_id = cp->cp_edx;
1570 					chipid_shift = BITX(cp->cp_eax, 4, 0);
1571 					ncpu_per_chip = BITX(cp->cp_ebx, 15, 0);
1572 				}
1573 			}
1574 
1575 			cpi->cpi_apicid = x2apic_id;
1576 			cpi->cpi_ncpu_per_chip = ncpu_per_chip;
1577 			cpi->cpi_ncore_per_chip = ncpu_per_chip /
1578 			    ncpu_per_core;
1579 			cpi->cpi_chipid = x2apic_id >> chipid_shift;
1580 			cpi->cpi_clogid = x2apic_id & ((1 << chipid_shift) - 1);
1581 			cpi->cpi_coreid = x2apic_id >> coreid_shift;
1582 			cpi->cpi_pkgcoreid = cpi->cpi_clogid >> coreid_shift;
1583 		}
1584 
1585 		/* Make cp NULL so that we don't stumble on others */
1586 		cp = NULL;
1587 	}
1588 
1589 	if ((cpi->cpi_xmaxeax & 0x80000000) == 0)
1590 		goto pass2_done;
1591 
1592 	if ((nmax = cpi->cpi_xmaxeax - 0x80000000 + 1) > NMAX_CPI_EXTD)
1593 		nmax = NMAX_CPI_EXTD;
1594 	/*
1595 	 * Copy the extended properties, fixing them as we go.
1596 	 * (We already handled n == 0 and n == 1 in pass 1)
1597 	 */
1598 	iptr = (void *)cpi->cpi_brandstr;
1599 	for (n = 2, cp = &cpi->cpi_extd[2]; n < nmax; cp++, n++) {
1600 		cp->cp_eax = 0x80000000 + n;
1601 		(void) __cpuid_insn(cp);
1602 		platform_cpuid_mangle(cpi->cpi_vendor, 0x80000000 + n, cp);
1603 		switch (n) {
1604 		case 2:
1605 		case 3:
1606 		case 4:
1607 			/*
1608 			 * Extract the brand string
1609 			 */
1610 			*iptr++ = cp->cp_eax;
1611 			*iptr++ = cp->cp_ebx;
1612 			*iptr++ = cp->cp_ecx;
1613 			*iptr++ = cp->cp_edx;
1614 			break;
1615 		case 5:
1616 			switch (cpi->cpi_vendor) {
1617 			case X86_VENDOR_AMD:
1618 				/*
1619 				 * The Athlon and Duron were the first
1620 				 * parts to report the sizes of the
1621 				 * TLB for large pages. Before then,
1622 				 * we don't trust the data.
1623 				 */
1624 				if (cpi->cpi_family < 6 ||
1625 				    (cpi->cpi_family == 6 &&
1626 				    cpi->cpi_model < 1))
1627 					cp->cp_eax = 0;
1628 				break;
1629 			default:
1630 				break;
1631 			}
1632 			break;
1633 		case 6:
1634 			switch (cpi->cpi_vendor) {
1635 			case X86_VENDOR_AMD:
1636 				/*
1637 				 * The Athlon and Duron were the first
1638 				 * AMD parts with L2 TLB's.
1639 				 * Before then, don't trust the data.
1640 				 */
1641 				if (cpi->cpi_family < 6 ||
1642 				    cpi->cpi_family == 6 &&
1643 				    cpi->cpi_model < 1)
1644 					cp->cp_eax = cp->cp_ebx = 0;
1645 				/*
1646 				 * AMD Duron rev A0 reports L2
1647 				 * cache size incorrectly as 1K
1648 				 * when it is really 64K
1649 				 */
1650 				if (cpi->cpi_family == 6 &&
1651 				    cpi->cpi_model == 3 &&
1652 				    cpi->cpi_step == 0) {
1653 					cp->cp_ecx &= 0xffff;
1654 					cp->cp_ecx |= 0x400000;
1655 				}
1656 				break;
1657 			case X86_VENDOR_Cyrix:	/* VIA C3 */
1658 				/*
1659 				 * VIA C3 processors are a bit messed
1660 				 * up w.r.t. encoding cache sizes in %ecx
1661 				 */
1662 				if (cpi->cpi_family != 6)
1663 					break;
1664 				/*
1665 				 * model 7 and 8 were incorrectly encoded
1666 				 *
1667 				 * xxx is model 8 really broken?
1668 				 */
1669 				if (cpi->cpi_model == 7 ||
1670 				    cpi->cpi_model == 8)
1671 					cp->cp_ecx =
1672 					    BITX(cp->cp_ecx, 31, 24) << 16 |
1673 					    BITX(cp->cp_ecx, 23, 16) << 12 |
1674 					    BITX(cp->cp_ecx, 15, 8) << 8 |
1675 					    BITX(cp->cp_ecx, 7, 0);
1676 				/*
1677 				 * model 9 stepping 1 has wrong associativity
1678 				 */
1679 				if (cpi->cpi_model == 9 && cpi->cpi_step == 1)
1680 					cp->cp_ecx |= 8 << 12;
1681 				break;
1682 			case X86_VENDOR_Intel:
1683 				/*
1684 				 * Extended L2 Cache features function.
1685 				 * First appeared on Prescott.
1686 				 */
1687 			default:
1688 				break;
1689 			}
1690 			break;
1691 		default:
1692 			break;
1693 		}
1694 	}
1695 
1696 pass2_done:
1697 	cpi->cpi_pass = 2;
1698 }
1699 
1700 static const char *
1701 intel_cpubrand(const struct cpuid_info *cpi)
1702 {
1703 	int i;
1704 
1705 	if ((x86_feature & X86_CPUID) == 0 ||
1706 	    cpi->cpi_maxeax < 1 || cpi->cpi_family < 5)
1707 		return ("i486");
1708 
1709 	switch (cpi->cpi_family) {
1710 	case 5:
1711 		return ("Intel Pentium(r)");
1712 	case 6:
1713 		switch (cpi->cpi_model) {
1714 			uint_t celeron, xeon;
1715 			const struct cpuid_regs *cp;
1716 		case 0:
1717 		case 1:
1718 		case 2:
1719 			return ("Intel Pentium(r) Pro");
1720 		case 3:
1721 		case 4:
1722 			return ("Intel Pentium(r) II");
1723 		case 6:
1724 			return ("Intel Celeron(r)");
1725 		case 5:
1726 		case 7:
1727 			celeron = xeon = 0;
1728 			cp = &cpi->cpi_std[2];	/* cache info */
1729 
1730 			for (i = 1; i < 4; i++) {
1731 				uint_t tmp;
1732 
1733 				tmp = (cp->cp_eax >> (8 * i)) & 0xff;
1734 				if (tmp == 0x40)
1735 					celeron++;
1736 				if (tmp >= 0x44 && tmp <= 0x45)
1737 					xeon++;
1738 			}
1739 
1740 			for (i = 0; i < 2; i++) {
1741 				uint_t tmp;
1742 
1743 				tmp = (cp->cp_ebx >> (8 * i)) & 0xff;
1744 				if (tmp == 0x40)
1745 					celeron++;
1746 				else if (tmp >= 0x44 && tmp <= 0x45)
1747 					xeon++;
1748 			}
1749 
1750 			for (i = 0; i < 4; i++) {
1751 				uint_t tmp;
1752 
1753 				tmp = (cp->cp_ecx >> (8 * i)) & 0xff;
1754 				if (tmp == 0x40)
1755 					celeron++;
1756 				else if (tmp >= 0x44 && tmp <= 0x45)
1757 					xeon++;
1758 			}
1759 
1760 			for (i = 0; i < 4; i++) {
1761 				uint_t tmp;
1762 
1763 				tmp = (cp->cp_edx >> (8 * i)) & 0xff;
1764 				if (tmp == 0x40)
1765 					celeron++;
1766 				else if (tmp >= 0x44 && tmp <= 0x45)
1767 					xeon++;
1768 			}
1769 
1770 			if (celeron)
1771 				return ("Intel Celeron(r)");
1772 			if (xeon)
1773 				return (cpi->cpi_model == 5 ?
1774 				    "Intel Pentium(r) II Xeon(tm)" :
1775 				    "Intel Pentium(r) III Xeon(tm)");
1776 			return (cpi->cpi_model == 5 ?
1777 			    "Intel Pentium(r) II or Pentium(r) II Xeon(tm)" :
1778 			    "Intel Pentium(r) III or Pentium(r) III Xeon(tm)");
1779 		default:
1780 			break;
1781 		}
1782 	default:
1783 		break;
1784 	}
1785 
1786 	/* BrandID is present if the field is nonzero */
1787 	if (cpi->cpi_brandid != 0) {
1788 		static const struct {
1789 			uint_t bt_bid;
1790 			const char *bt_str;
1791 		} brand_tbl[] = {
1792 			{ 0x1,	"Intel(r) Celeron(r)" },
1793 			{ 0x2,	"Intel(r) Pentium(r) III" },
1794 			{ 0x3,	"Intel(r) Pentium(r) III Xeon(tm)" },
1795 			{ 0x4,	"Intel(r) Pentium(r) III" },
1796 			{ 0x6,	"Mobile Intel(r) Pentium(r) III" },
1797 			{ 0x7,	"Mobile Intel(r) Celeron(r)" },
1798 			{ 0x8,	"Intel(r) Pentium(r) 4" },
1799 			{ 0x9,	"Intel(r) Pentium(r) 4" },
1800 			{ 0xa,	"Intel(r) Celeron(r)" },
1801 			{ 0xb,	"Intel(r) Xeon(tm)" },
1802 			{ 0xc,	"Intel(r) Xeon(tm) MP" },
1803 			{ 0xe,	"Mobile Intel(r) Pentium(r) 4" },
1804 			{ 0xf,	"Mobile Intel(r) Celeron(r)" },
1805 			{ 0x11, "Mobile Genuine Intel(r)" },
1806 			{ 0x12, "Intel(r) Celeron(r) M" },
1807 			{ 0x13, "Mobile Intel(r) Celeron(r)" },
1808 			{ 0x14, "Intel(r) Celeron(r)" },
1809 			{ 0x15, "Mobile Genuine Intel(r)" },
1810 			{ 0x16,	"Intel(r) Pentium(r) M" },
1811 			{ 0x17, "Mobile Intel(r) Celeron(r)" }
1812 		};
1813 		uint_t btblmax = sizeof (brand_tbl) / sizeof (brand_tbl[0]);
1814 		uint_t sgn;
1815 
1816 		sgn = (cpi->cpi_family << 8) |
1817 		    (cpi->cpi_model << 4) | cpi->cpi_step;
1818 
1819 		for (i = 0; i < btblmax; i++)
1820 			if (brand_tbl[i].bt_bid == cpi->cpi_brandid)
1821 				break;
1822 		if (i < btblmax) {
1823 			if (sgn == 0x6b1 && cpi->cpi_brandid == 3)
1824 				return ("Intel(r) Celeron(r)");
1825 			if (sgn < 0xf13 && cpi->cpi_brandid == 0xb)
1826 				return ("Intel(r) Xeon(tm) MP");
1827 			if (sgn < 0xf13 && cpi->cpi_brandid == 0xe)
1828 				return ("Intel(r) Xeon(tm)");
1829 			return (brand_tbl[i].bt_str);
1830 		}
1831 	}
1832 
1833 	return (NULL);
1834 }
1835 
1836 static const char *
1837 amd_cpubrand(const struct cpuid_info *cpi)
1838 {
1839 	if ((x86_feature & X86_CPUID) == 0 ||
1840 	    cpi->cpi_maxeax < 1 || cpi->cpi_family < 5)
1841 		return ("i486 compatible");
1842 
1843 	switch (cpi->cpi_family) {
1844 	case 5:
1845 		switch (cpi->cpi_model) {
1846 		case 0:
1847 		case 1:
1848 		case 2:
1849 		case 3:
1850 		case 4:
1851 		case 5:
1852 			return ("AMD-K5(r)");
1853 		case 6:
1854 		case 7:
1855 			return ("AMD-K6(r)");
1856 		case 8:
1857 			return ("AMD-K6(r)-2");
1858 		case 9:
1859 			return ("AMD-K6(r)-III");
1860 		default:
1861 			return ("AMD (family 5)");
1862 		}
1863 	case 6:
1864 		switch (cpi->cpi_model) {
1865 		case 1:
1866 			return ("AMD-K7(tm)");
1867 		case 0:
1868 		case 2:
1869 		case 4:
1870 			return ("AMD Athlon(tm)");
1871 		case 3:
1872 		case 7:
1873 			return ("AMD Duron(tm)");
1874 		case 6:
1875 		case 8:
1876 		case 10:
1877 			/*
1878 			 * Use the L2 cache size to distinguish
1879 			 */
1880 			return ((cpi->cpi_extd[6].cp_ecx >> 16) >= 256 ?
1881 			    "AMD Athlon(tm)" : "AMD Duron(tm)");
1882 		default:
1883 			return ("AMD (family 6)");
1884 		}
1885 	default:
1886 		break;
1887 	}
1888 
1889 	if (cpi->cpi_family == 0xf && cpi->cpi_model == 5 &&
1890 	    cpi->cpi_brandid != 0) {
1891 		switch (BITX(cpi->cpi_brandid, 7, 5)) {
1892 		case 3:
1893 			return ("AMD Opteron(tm) UP 1xx");
1894 		case 4:
1895 			return ("AMD Opteron(tm) DP 2xx");
1896 		case 5:
1897 			return ("AMD Opteron(tm) MP 8xx");
1898 		default:
1899 			return ("AMD Opteron(tm)");
1900 		}
1901 	}
1902 
1903 	return (NULL);
1904 }
1905 
1906 static const char *
1907 cyrix_cpubrand(struct cpuid_info *cpi, uint_t type)
1908 {
1909 	if ((x86_feature & X86_CPUID) == 0 ||
1910 	    cpi->cpi_maxeax < 1 || cpi->cpi_family < 5 ||
1911 	    type == X86_TYPE_CYRIX_486)
1912 		return ("i486 compatible");
1913 
1914 	switch (type) {
1915 	case X86_TYPE_CYRIX_6x86:
1916 		return ("Cyrix 6x86");
1917 	case X86_TYPE_CYRIX_6x86L:
1918 		return ("Cyrix 6x86L");
1919 	case X86_TYPE_CYRIX_6x86MX:
1920 		return ("Cyrix 6x86MX");
1921 	case X86_TYPE_CYRIX_GXm:
1922 		return ("Cyrix GXm");
1923 	case X86_TYPE_CYRIX_MediaGX:
1924 		return ("Cyrix MediaGX");
1925 	case X86_TYPE_CYRIX_MII:
1926 		return ("Cyrix M2");
1927 	case X86_TYPE_VIA_CYRIX_III:
1928 		return ("VIA Cyrix M3");
1929 	default:
1930 		/*
1931 		 * Have another wild guess ..
1932 		 */
1933 		if (cpi->cpi_family == 4 && cpi->cpi_model == 9)
1934 			return ("Cyrix 5x86");
1935 		else if (cpi->cpi_family == 5) {
1936 			switch (cpi->cpi_model) {
1937 			case 2:
1938 				return ("Cyrix 6x86");	/* Cyrix M1 */
1939 			case 4:
1940 				return ("Cyrix MediaGX");
1941 			default:
1942 				break;
1943 			}
1944 		} else if (cpi->cpi_family == 6) {
1945 			switch (cpi->cpi_model) {
1946 			case 0:
1947 				return ("Cyrix 6x86MX"); /* Cyrix M2? */
1948 			case 5:
1949 			case 6:
1950 			case 7:
1951 			case 8:
1952 			case 9:
1953 				return ("VIA C3");
1954 			default:
1955 				break;
1956 			}
1957 		}
1958 		break;
1959 	}
1960 	return (NULL);
1961 }
1962 
1963 /*
1964  * This only gets called in the case that the CPU extended
1965  * feature brand string (0x80000002, 0x80000003, 0x80000004)
1966  * aren't available, or contain null bytes for some reason.
1967  */
1968 static void
1969 fabricate_brandstr(struct cpuid_info *cpi)
1970 {
1971 	const char *brand = NULL;
1972 
1973 	switch (cpi->cpi_vendor) {
1974 	case X86_VENDOR_Intel:
1975 		brand = intel_cpubrand(cpi);
1976 		break;
1977 	case X86_VENDOR_AMD:
1978 		brand = amd_cpubrand(cpi);
1979 		break;
1980 	case X86_VENDOR_Cyrix:
1981 		brand = cyrix_cpubrand(cpi, x86_type);
1982 		break;
1983 	case X86_VENDOR_NexGen:
1984 		if (cpi->cpi_family == 5 && cpi->cpi_model == 0)
1985 			brand = "NexGen Nx586";
1986 		break;
1987 	case X86_VENDOR_Centaur:
1988 		if (cpi->cpi_family == 5)
1989 			switch (cpi->cpi_model) {
1990 			case 4:
1991 				brand = "Centaur C6";
1992 				break;
1993 			case 8:
1994 				brand = "Centaur C2";
1995 				break;
1996 			case 9:
1997 				brand = "Centaur C3";
1998 				break;
1999 			default:
2000 				break;
2001 			}
2002 		break;
2003 	case X86_VENDOR_Rise:
2004 		if (cpi->cpi_family == 5 &&
2005 		    (cpi->cpi_model == 0 || cpi->cpi_model == 2))
2006 			brand = "Rise mP6";
2007 		break;
2008 	case X86_VENDOR_SiS:
2009 		if (cpi->cpi_family == 5 && cpi->cpi_model == 0)
2010 			brand = "SiS 55x";
2011 		break;
2012 	case X86_VENDOR_TM:
2013 		if (cpi->cpi_family == 5 && cpi->cpi_model == 4)
2014 			brand = "Transmeta Crusoe TM3x00 or TM5x00";
2015 		break;
2016 	case X86_VENDOR_NSC:
2017 	case X86_VENDOR_UMC:
2018 	default:
2019 		break;
2020 	}
2021 	if (brand) {
2022 		(void) strcpy((char *)cpi->cpi_brandstr, brand);
2023 		return;
2024 	}
2025 
2026 	/*
2027 	 * If all else fails ...
2028 	 */
2029 	(void) snprintf(cpi->cpi_brandstr, sizeof (cpi->cpi_brandstr),
2030 	    "%s %d.%d.%d", cpi->cpi_vendorstr, cpi->cpi_family,
2031 	    cpi->cpi_model, cpi->cpi_step);
2032 }
2033 
2034 /*
2035  * This routine is called just after kernel memory allocation
2036  * becomes available on cpu0, and as part of mp_startup() on
2037  * the other cpus.
2038  *
2039  * Fixup the brand string, and collect any information from cpuid
2040  * that requires dynamicically allocated storage to represent.
2041  */
2042 /*ARGSUSED*/
2043 void
2044 cpuid_pass3(cpu_t *cpu)
2045 {
2046 	int	i, max, shft, level, size;
2047 	struct cpuid_regs regs;
2048 	struct cpuid_regs *cp;
2049 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2050 
2051 	ASSERT(cpi->cpi_pass == 2);
2052 
2053 	/*
2054 	 * Function 4: Deterministic cache parameters
2055 	 *
2056 	 * Take this opportunity to detect the number of threads
2057 	 * sharing the last level cache, and construct a corresponding
2058 	 * cache id. The respective cpuid_info members are initialized
2059 	 * to the default case of "no last level cache sharing".
2060 	 */
2061 	cpi->cpi_ncpu_shr_last_cache = 1;
2062 	cpi->cpi_last_lvl_cacheid = cpu->cpu_id;
2063 
2064 	if (cpi->cpi_maxeax >= 4 && cpi->cpi_vendor == X86_VENDOR_Intel) {
2065 
2066 		/*
2067 		 * Find the # of elements (size) returned by fn 4, and along
2068 		 * the way detect last level cache sharing details.
2069 		 */
2070 		bzero(&regs, sizeof (regs));
2071 		cp = &regs;
2072 		for (i = 0, max = 0; i < CPI_FN4_ECX_MAX; i++) {
2073 			cp->cp_eax = 4;
2074 			cp->cp_ecx = i;
2075 
2076 			(void) __cpuid_insn(cp);
2077 
2078 			if (CPI_CACHE_TYPE(cp) == 0)
2079 				break;
2080 			level = CPI_CACHE_LVL(cp);
2081 			if (level > max) {
2082 				max = level;
2083 				cpi->cpi_ncpu_shr_last_cache =
2084 				    CPI_NTHR_SHR_CACHE(cp) + 1;
2085 			}
2086 		}
2087 		cpi->cpi_std_4_size = size = i;
2088 
2089 		/*
2090 		 * Allocate the cpi_std_4 array. The first element
2091 		 * references the regs for fn 4, %ecx == 0, which
2092 		 * cpuid_pass2() stashed in cpi->cpi_std[4].
2093 		 */
2094 		if (size > 0) {
2095 			cpi->cpi_std_4 =
2096 			    kmem_alloc(size * sizeof (cp), KM_SLEEP);
2097 			cpi->cpi_std_4[0] = &cpi->cpi_std[4];
2098 
2099 			/*
2100 			 * Allocate storage to hold the additional regs
2101 			 * for function 4, %ecx == 1 .. cpi_std_4_size.
2102 			 *
2103 			 * The regs for fn 4, %ecx == 0 has already
2104 			 * been allocated as indicated above.
2105 			 */
2106 			for (i = 1; i < size; i++) {
2107 				cp = cpi->cpi_std_4[i] =
2108 				    kmem_zalloc(sizeof (regs), KM_SLEEP);
2109 				cp->cp_eax = 4;
2110 				cp->cp_ecx = i;
2111 
2112 				(void) __cpuid_insn(cp);
2113 			}
2114 		}
2115 		/*
2116 		 * Determine the number of bits needed to represent
2117 		 * the number of CPUs sharing the last level cache.
2118 		 *
2119 		 * Shift off that number of bits from the APIC id to
2120 		 * derive the cache id.
2121 		 */
2122 		shft = 0;
2123 		for (i = 1; i < cpi->cpi_ncpu_shr_last_cache; i <<= 1)
2124 			shft++;
2125 		cpi->cpi_last_lvl_cacheid = cpi->cpi_apicid >> shft;
2126 	}
2127 
2128 	/*
2129 	 * Now fixup the brand string
2130 	 */
2131 	if ((cpi->cpi_xmaxeax & 0x80000000) == 0) {
2132 		fabricate_brandstr(cpi);
2133 	} else {
2134 
2135 		/*
2136 		 * If we successfully extracted a brand string from the cpuid
2137 		 * instruction, clean it up by removing leading spaces and
2138 		 * similar junk.
2139 		 */
2140 		if (cpi->cpi_brandstr[0]) {
2141 			size_t maxlen = sizeof (cpi->cpi_brandstr);
2142 			char *src, *dst;
2143 
2144 			dst = src = (char *)cpi->cpi_brandstr;
2145 			src[maxlen - 1] = '\0';
2146 			/*
2147 			 * strip leading spaces
2148 			 */
2149 			while (*src == ' ')
2150 				src++;
2151 			/*
2152 			 * Remove any 'Genuine' or "Authentic" prefixes
2153 			 */
2154 			if (strncmp(src, "Genuine ", 8) == 0)
2155 				src += 8;
2156 			if (strncmp(src, "Authentic ", 10) == 0)
2157 				src += 10;
2158 
2159 			/*
2160 			 * Now do an in-place copy.
2161 			 * Map (R) to (r) and (TM) to (tm).
2162 			 * The era of teletypes is long gone, and there's
2163 			 * -really- no need to shout.
2164 			 */
2165 			while (*src != '\0') {
2166 				if (src[0] == '(') {
2167 					if (strncmp(src + 1, "R)", 2) == 0) {
2168 						(void) strncpy(dst, "(r)", 3);
2169 						src += 3;
2170 						dst += 3;
2171 						continue;
2172 					}
2173 					if (strncmp(src + 1, "TM)", 3) == 0) {
2174 						(void) strncpy(dst, "(tm)", 4);
2175 						src += 4;
2176 						dst += 4;
2177 						continue;
2178 					}
2179 				}
2180 				*dst++ = *src++;
2181 			}
2182 			*dst = '\0';
2183 
2184 			/*
2185 			 * Finally, remove any trailing spaces
2186 			 */
2187 			while (--dst > cpi->cpi_brandstr)
2188 				if (*dst == ' ')
2189 					*dst = '\0';
2190 				else
2191 					break;
2192 		} else
2193 			fabricate_brandstr(cpi);
2194 	}
2195 	cpi->cpi_pass = 3;
2196 }
2197 
2198 /*
2199  * This routine is called out of bind_hwcap() much later in the life
2200  * of the kernel (post_startup()).  The job of this routine is to resolve
2201  * the hardware feature support and kernel support for those features into
2202  * what we're actually going to tell applications via the aux vector.
2203  */
2204 uint_t
2205 cpuid_pass4(cpu_t *cpu)
2206 {
2207 	struct cpuid_info *cpi;
2208 	uint_t hwcap_flags = 0;
2209 
2210 	if (cpu == NULL)
2211 		cpu = CPU;
2212 	cpi = cpu->cpu_m.mcpu_cpi;
2213 
2214 	ASSERT(cpi->cpi_pass == 3);
2215 
2216 	if (cpi->cpi_maxeax >= 1) {
2217 		uint32_t *edx = &cpi->cpi_support[STD_EDX_FEATURES];
2218 		uint32_t *ecx = &cpi->cpi_support[STD_ECX_FEATURES];
2219 
2220 		*edx = CPI_FEATURES_EDX(cpi);
2221 		*ecx = CPI_FEATURES_ECX(cpi);
2222 
2223 		/*
2224 		 * [these require explicit kernel support]
2225 		 */
2226 		if ((x86_feature & X86_SEP) == 0)
2227 			*edx &= ~CPUID_INTC_EDX_SEP;
2228 
2229 		if ((x86_feature & X86_SSE) == 0)
2230 			*edx &= ~(CPUID_INTC_EDX_FXSR|CPUID_INTC_EDX_SSE);
2231 		if ((x86_feature & X86_SSE2) == 0)
2232 			*edx &= ~CPUID_INTC_EDX_SSE2;
2233 
2234 		if ((x86_feature & X86_HTT) == 0)
2235 			*edx &= ~CPUID_INTC_EDX_HTT;
2236 
2237 		if ((x86_feature & X86_SSE3) == 0)
2238 			*ecx &= ~CPUID_INTC_ECX_SSE3;
2239 
2240 		if (cpi->cpi_vendor == X86_VENDOR_Intel) {
2241 			if ((x86_feature & X86_SSSE3) == 0)
2242 				*ecx &= ~CPUID_INTC_ECX_SSSE3;
2243 			if ((x86_feature & X86_SSE4_1) == 0)
2244 				*ecx &= ~CPUID_INTC_ECX_SSE4_1;
2245 			if ((x86_feature & X86_SSE4_2) == 0)
2246 				*ecx &= ~CPUID_INTC_ECX_SSE4_2;
2247 			if ((x86_feature & X86_AES) == 0)
2248 				*ecx &= ~CPUID_INTC_ECX_AES;
2249 		}
2250 
2251 		/*
2252 		 * [no explicit support required beyond x87 fp context]
2253 		 */
2254 		if (!fpu_exists)
2255 			*edx &= ~(CPUID_INTC_EDX_FPU | CPUID_INTC_EDX_MMX);
2256 
2257 		/*
2258 		 * Now map the supported feature vector to things that we
2259 		 * think userland will care about.
2260 		 */
2261 		if (*edx & CPUID_INTC_EDX_SEP)
2262 			hwcap_flags |= AV_386_SEP;
2263 		if (*edx & CPUID_INTC_EDX_SSE)
2264 			hwcap_flags |= AV_386_FXSR | AV_386_SSE;
2265 		if (*edx & CPUID_INTC_EDX_SSE2)
2266 			hwcap_flags |= AV_386_SSE2;
2267 		if (*ecx & CPUID_INTC_ECX_SSE3)
2268 			hwcap_flags |= AV_386_SSE3;
2269 		if (cpi->cpi_vendor == X86_VENDOR_Intel) {
2270 			if (*ecx & CPUID_INTC_ECX_SSSE3)
2271 				hwcap_flags |= AV_386_SSSE3;
2272 			if (*ecx & CPUID_INTC_ECX_SSE4_1)
2273 				hwcap_flags |= AV_386_SSE4_1;
2274 			if (*ecx & CPUID_INTC_ECX_SSE4_2)
2275 				hwcap_flags |= AV_386_SSE4_2;
2276 			if (*ecx & CPUID_INTC_ECX_MOVBE)
2277 				hwcap_flags |= AV_386_MOVBE;
2278 			if (*ecx & CPUID_INTC_ECX_AES)
2279 				hwcap_flags |= AV_386_AES;
2280 			if (*ecx & CPUID_INTC_ECX_PCLMULQDQ)
2281 				hwcap_flags |= AV_386_PCLMULQDQ;
2282 		}
2283 		if (*ecx & CPUID_INTC_ECX_POPCNT)
2284 			hwcap_flags |= AV_386_POPCNT;
2285 		if (*edx & CPUID_INTC_EDX_FPU)
2286 			hwcap_flags |= AV_386_FPU;
2287 		if (*edx & CPUID_INTC_EDX_MMX)
2288 			hwcap_flags |= AV_386_MMX;
2289 
2290 		if (*edx & CPUID_INTC_EDX_TSC)
2291 			hwcap_flags |= AV_386_TSC;
2292 		if (*edx & CPUID_INTC_EDX_CX8)
2293 			hwcap_flags |= AV_386_CX8;
2294 		if (*edx & CPUID_INTC_EDX_CMOV)
2295 			hwcap_flags |= AV_386_CMOV;
2296 		if (*ecx & CPUID_INTC_ECX_CX16)
2297 			hwcap_flags |= AV_386_CX16;
2298 	}
2299 
2300 	if (cpi->cpi_xmaxeax < 0x80000001)
2301 		goto pass4_done;
2302 
2303 	switch (cpi->cpi_vendor) {
2304 		struct cpuid_regs cp;
2305 		uint32_t *edx, *ecx;
2306 
2307 	case X86_VENDOR_Intel:
2308 		/*
2309 		 * Seems like Intel duplicated what we necessary
2310 		 * here to make the initial crop of 64-bit OS's work.
2311 		 * Hopefully, those are the only "extended" bits
2312 		 * they'll add.
2313 		 */
2314 		/*FALLTHROUGH*/
2315 
2316 	case X86_VENDOR_AMD:
2317 		edx = &cpi->cpi_support[AMD_EDX_FEATURES];
2318 		ecx = &cpi->cpi_support[AMD_ECX_FEATURES];
2319 
2320 		*edx = CPI_FEATURES_XTD_EDX(cpi);
2321 		*ecx = CPI_FEATURES_XTD_ECX(cpi);
2322 
2323 		/*
2324 		 * [these features require explicit kernel support]
2325 		 */
2326 		switch (cpi->cpi_vendor) {
2327 		case X86_VENDOR_Intel:
2328 			if ((x86_feature & X86_TSCP) == 0)
2329 				*edx &= ~CPUID_AMD_EDX_TSCP;
2330 			break;
2331 
2332 		case X86_VENDOR_AMD:
2333 			if ((x86_feature & X86_TSCP) == 0)
2334 				*edx &= ~CPUID_AMD_EDX_TSCP;
2335 			if ((x86_feature & X86_SSE4A) == 0)
2336 				*ecx &= ~CPUID_AMD_ECX_SSE4A;
2337 			break;
2338 
2339 		default:
2340 			break;
2341 		}
2342 
2343 		/*
2344 		 * [no explicit support required beyond
2345 		 * x87 fp context and exception handlers]
2346 		 */
2347 		if (!fpu_exists)
2348 			*edx &= ~(CPUID_AMD_EDX_MMXamd |
2349 			    CPUID_AMD_EDX_3DNow | CPUID_AMD_EDX_3DNowx);
2350 
2351 		if ((x86_feature & X86_NX) == 0)
2352 			*edx &= ~CPUID_AMD_EDX_NX;
2353 #if !defined(__amd64)
2354 		*edx &= ~CPUID_AMD_EDX_LM;
2355 #endif
2356 		/*
2357 		 * Now map the supported feature vector to
2358 		 * things that we think userland will care about.
2359 		 */
2360 #if defined(__amd64)
2361 		if (*edx & CPUID_AMD_EDX_SYSC)
2362 			hwcap_flags |= AV_386_AMD_SYSC;
2363 #endif
2364 		if (*edx & CPUID_AMD_EDX_MMXamd)
2365 			hwcap_flags |= AV_386_AMD_MMX;
2366 		if (*edx & CPUID_AMD_EDX_3DNow)
2367 			hwcap_flags |= AV_386_AMD_3DNow;
2368 		if (*edx & CPUID_AMD_EDX_3DNowx)
2369 			hwcap_flags |= AV_386_AMD_3DNowx;
2370 
2371 		switch (cpi->cpi_vendor) {
2372 		case X86_VENDOR_AMD:
2373 			if (*edx & CPUID_AMD_EDX_TSCP)
2374 				hwcap_flags |= AV_386_TSCP;
2375 			if (*ecx & CPUID_AMD_ECX_AHF64)
2376 				hwcap_flags |= AV_386_AHF;
2377 			if (*ecx & CPUID_AMD_ECX_SSE4A)
2378 				hwcap_flags |= AV_386_AMD_SSE4A;
2379 			if (*ecx & CPUID_AMD_ECX_LZCNT)
2380 				hwcap_flags |= AV_386_AMD_LZCNT;
2381 			break;
2382 
2383 		case X86_VENDOR_Intel:
2384 			if (*edx & CPUID_AMD_EDX_TSCP)
2385 				hwcap_flags |= AV_386_TSCP;
2386 			/*
2387 			 * Aarrgh.
2388 			 * Intel uses a different bit in the same word.
2389 			 */
2390 			if (*ecx & CPUID_INTC_ECX_AHF64)
2391 				hwcap_flags |= AV_386_AHF;
2392 			break;
2393 
2394 		default:
2395 			break;
2396 		}
2397 		break;
2398 
2399 	case X86_VENDOR_TM:
2400 		cp.cp_eax = 0x80860001;
2401 		(void) __cpuid_insn(&cp);
2402 		cpi->cpi_support[TM_EDX_FEATURES] = cp.cp_edx;
2403 		break;
2404 
2405 	default:
2406 		break;
2407 	}
2408 
2409 pass4_done:
2410 	cpi->cpi_pass = 4;
2411 	return (hwcap_flags);
2412 }
2413 
2414 
2415 /*
2416  * Simulate the cpuid instruction using the data we previously
2417  * captured about this CPU.  We try our best to return the truth
2418  * about the hardware, independently of kernel support.
2419  */
2420 uint32_t
2421 cpuid_insn(cpu_t *cpu, struct cpuid_regs *cp)
2422 {
2423 	struct cpuid_info *cpi;
2424 	struct cpuid_regs *xcp;
2425 
2426 	if (cpu == NULL)
2427 		cpu = CPU;
2428 	cpi = cpu->cpu_m.mcpu_cpi;
2429 
2430 	ASSERT(cpuid_checkpass(cpu, 3));
2431 
2432 	/*
2433 	 * CPUID data is cached in two separate places: cpi_std for standard
2434 	 * CPUID functions, and cpi_extd for extended CPUID functions.
2435 	 */
2436 	if (cp->cp_eax <= cpi->cpi_maxeax && cp->cp_eax < NMAX_CPI_STD)
2437 		xcp = &cpi->cpi_std[cp->cp_eax];
2438 	else if (cp->cp_eax >= 0x80000000 && cp->cp_eax <= cpi->cpi_xmaxeax &&
2439 	    cp->cp_eax < 0x80000000 + NMAX_CPI_EXTD)
2440 		xcp = &cpi->cpi_extd[cp->cp_eax - 0x80000000];
2441 	else
2442 		/*
2443 		 * The caller is asking for data from an input parameter which
2444 		 * the kernel has not cached.  In this case we go fetch from
2445 		 * the hardware and return the data directly to the user.
2446 		 */
2447 		return (__cpuid_insn(cp));
2448 
2449 	cp->cp_eax = xcp->cp_eax;
2450 	cp->cp_ebx = xcp->cp_ebx;
2451 	cp->cp_ecx = xcp->cp_ecx;
2452 	cp->cp_edx = xcp->cp_edx;
2453 	return (cp->cp_eax);
2454 }
2455 
2456 int
2457 cpuid_checkpass(cpu_t *cpu, int pass)
2458 {
2459 	return (cpu != NULL && cpu->cpu_m.mcpu_cpi != NULL &&
2460 	    cpu->cpu_m.mcpu_cpi->cpi_pass >= pass);
2461 }
2462 
2463 int
2464 cpuid_getbrandstr(cpu_t *cpu, char *s, size_t n)
2465 {
2466 	ASSERT(cpuid_checkpass(cpu, 3));
2467 
2468 	return (snprintf(s, n, "%s", cpu->cpu_m.mcpu_cpi->cpi_brandstr));
2469 }
2470 
2471 int
2472 cpuid_is_cmt(cpu_t *cpu)
2473 {
2474 	if (cpu == NULL)
2475 		cpu = CPU;
2476 
2477 	ASSERT(cpuid_checkpass(cpu, 1));
2478 
2479 	return (cpu->cpu_m.mcpu_cpi->cpi_chipid >= 0);
2480 }
2481 
2482 /*
2483  * AMD and Intel both implement the 64-bit variant of the syscall
2484  * instruction (syscallq), so if there's -any- support for syscall,
2485  * cpuid currently says "yes, we support this".
2486  *
2487  * However, Intel decided to -not- implement the 32-bit variant of the
2488  * syscall instruction, so we provide a predicate to allow our caller
2489  * to test that subtlety here.
2490  *
2491  * XXPV	Currently, 32-bit syscall instructions don't work via the hypervisor,
2492  *	even in the case where the hardware would in fact support it.
2493  */
2494 /*ARGSUSED*/
2495 int
2496 cpuid_syscall32_insn(cpu_t *cpu)
2497 {
2498 	ASSERT(cpuid_checkpass((cpu == NULL ? CPU : cpu), 1));
2499 
2500 #if !defined(__xpv)
2501 	if (cpu == NULL)
2502 		cpu = CPU;
2503 
2504 	/*CSTYLED*/
2505 	{
2506 		struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2507 
2508 		if (cpi->cpi_vendor == X86_VENDOR_AMD &&
2509 		    cpi->cpi_xmaxeax >= 0x80000001 &&
2510 		    (CPI_FEATURES_XTD_EDX(cpi) & CPUID_AMD_EDX_SYSC))
2511 			return (1);
2512 	}
2513 #endif
2514 	return (0);
2515 }
2516 
2517 int
2518 cpuid_getidstr(cpu_t *cpu, char *s, size_t n)
2519 {
2520 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2521 
2522 	static const char fmt[] =
2523 	    "x86 (%s %X family %d model %d step %d clock %d MHz)";
2524 	static const char fmt_ht[] =
2525 	    "x86 (chipid 0x%x %s %X family %d model %d step %d clock %d MHz)";
2526 
2527 	ASSERT(cpuid_checkpass(cpu, 1));
2528 
2529 	if (cpuid_is_cmt(cpu))
2530 		return (snprintf(s, n, fmt_ht, cpi->cpi_chipid,
2531 		    cpi->cpi_vendorstr, cpi->cpi_std[1].cp_eax,
2532 		    cpi->cpi_family, cpi->cpi_model,
2533 		    cpi->cpi_step, cpu->cpu_type_info.pi_clock));
2534 	return (snprintf(s, n, fmt,
2535 	    cpi->cpi_vendorstr, cpi->cpi_std[1].cp_eax,
2536 	    cpi->cpi_family, cpi->cpi_model,
2537 	    cpi->cpi_step, cpu->cpu_type_info.pi_clock));
2538 }
2539 
2540 const char *
2541 cpuid_getvendorstr(cpu_t *cpu)
2542 {
2543 	ASSERT(cpuid_checkpass(cpu, 1));
2544 	return ((const char *)cpu->cpu_m.mcpu_cpi->cpi_vendorstr);
2545 }
2546 
2547 uint_t
2548 cpuid_getvendor(cpu_t *cpu)
2549 {
2550 	ASSERT(cpuid_checkpass(cpu, 1));
2551 	return (cpu->cpu_m.mcpu_cpi->cpi_vendor);
2552 }
2553 
2554 uint_t
2555 cpuid_getfamily(cpu_t *cpu)
2556 {
2557 	ASSERT(cpuid_checkpass(cpu, 1));
2558 	return (cpu->cpu_m.mcpu_cpi->cpi_family);
2559 }
2560 
2561 uint_t
2562 cpuid_getmodel(cpu_t *cpu)
2563 {
2564 	ASSERT(cpuid_checkpass(cpu, 1));
2565 	return (cpu->cpu_m.mcpu_cpi->cpi_model);
2566 }
2567 
2568 uint_t
2569 cpuid_get_ncpu_per_chip(cpu_t *cpu)
2570 {
2571 	ASSERT(cpuid_checkpass(cpu, 1));
2572 	return (cpu->cpu_m.mcpu_cpi->cpi_ncpu_per_chip);
2573 }
2574 
2575 uint_t
2576 cpuid_get_ncore_per_chip(cpu_t *cpu)
2577 {
2578 	ASSERT(cpuid_checkpass(cpu, 1));
2579 	return (cpu->cpu_m.mcpu_cpi->cpi_ncore_per_chip);
2580 }
2581 
2582 uint_t
2583 cpuid_get_ncpu_sharing_last_cache(cpu_t *cpu)
2584 {
2585 	ASSERT(cpuid_checkpass(cpu, 2));
2586 	return (cpu->cpu_m.mcpu_cpi->cpi_ncpu_shr_last_cache);
2587 }
2588 
2589 id_t
2590 cpuid_get_last_lvl_cacheid(cpu_t *cpu)
2591 {
2592 	ASSERT(cpuid_checkpass(cpu, 2));
2593 	return (cpu->cpu_m.mcpu_cpi->cpi_last_lvl_cacheid);
2594 }
2595 
2596 uint_t
2597 cpuid_getstep(cpu_t *cpu)
2598 {
2599 	ASSERT(cpuid_checkpass(cpu, 1));
2600 	return (cpu->cpu_m.mcpu_cpi->cpi_step);
2601 }
2602 
2603 uint_t
2604 cpuid_getsig(struct cpu *cpu)
2605 {
2606 	ASSERT(cpuid_checkpass(cpu, 1));
2607 	return (cpu->cpu_m.mcpu_cpi->cpi_std[1].cp_eax);
2608 }
2609 
2610 uint32_t
2611 cpuid_getchiprev(struct cpu *cpu)
2612 {
2613 	ASSERT(cpuid_checkpass(cpu, 1));
2614 	return (cpu->cpu_m.mcpu_cpi->cpi_chiprev);
2615 }
2616 
2617 const char *
2618 cpuid_getchiprevstr(struct cpu *cpu)
2619 {
2620 	ASSERT(cpuid_checkpass(cpu, 1));
2621 	return (cpu->cpu_m.mcpu_cpi->cpi_chiprevstr);
2622 }
2623 
2624 uint32_t
2625 cpuid_getsockettype(struct cpu *cpu)
2626 {
2627 	ASSERT(cpuid_checkpass(cpu, 1));
2628 	return (cpu->cpu_m.mcpu_cpi->cpi_socket);
2629 }
2630 
2631 const char *
2632 cpuid_getsocketstr(cpu_t *cpu)
2633 {
2634 	static const char *socketstr = NULL;
2635 	struct cpuid_info *cpi;
2636 
2637 	ASSERT(cpuid_checkpass(cpu, 1));
2638 	cpi = cpu->cpu_m.mcpu_cpi;
2639 
2640 	/* Assume that socket types are the same across the system */
2641 	if (socketstr == NULL)
2642 		socketstr = _cpuid_sktstr(cpi->cpi_vendor, cpi->cpi_family,
2643 		    cpi->cpi_model, cpi->cpi_step);
2644 
2645 
2646 	return (socketstr);
2647 }
2648 
2649 int
2650 cpuid_get_chipid(cpu_t *cpu)
2651 {
2652 	ASSERT(cpuid_checkpass(cpu, 1));
2653 
2654 	if (cpuid_is_cmt(cpu))
2655 		return (cpu->cpu_m.mcpu_cpi->cpi_chipid);
2656 	return (cpu->cpu_id);
2657 }
2658 
2659 id_t
2660 cpuid_get_coreid(cpu_t *cpu)
2661 {
2662 	ASSERT(cpuid_checkpass(cpu, 1));
2663 	return (cpu->cpu_m.mcpu_cpi->cpi_coreid);
2664 }
2665 
2666 int
2667 cpuid_get_pkgcoreid(cpu_t *cpu)
2668 {
2669 	ASSERT(cpuid_checkpass(cpu, 1));
2670 	return (cpu->cpu_m.mcpu_cpi->cpi_pkgcoreid);
2671 }
2672 
2673 int
2674 cpuid_get_clogid(cpu_t *cpu)
2675 {
2676 	ASSERT(cpuid_checkpass(cpu, 1));
2677 	return (cpu->cpu_m.mcpu_cpi->cpi_clogid);
2678 }
2679 
2680 int
2681 cpuid_get_cacheid(cpu_t *cpu)
2682 {
2683 	ASSERT(cpuid_checkpass(cpu, 1));
2684 	return (cpu->cpu_m.mcpu_cpi->cpi_last_lvl_cacheid);
2685 }
2686 
2687 uint_t
2688 cpuid_get_procnodeid(cpu_t *cpu)
2689 {
2690 	ASSERT(cpuid_checkpass(cpu, 1));
2691 	return (cpu->cpu_m.mcpu_cpi->cpi_procnodeid);
2692 }
2693 
2694 uint_t
2695 cpuid_get_procnodes_per_pkg(cpu_t *cpu)
2696 {
2697 	ASSERT(cpuid_checkpass(cpu, 1));
2698 	return (cpu->cpu_m.mcpu_cpi->cpi_procnodes_per_pkg);
2699 }
2700 
2701 /*ARGSUSED*/
2702 int
2703 cpuid_have_cr8access(cpu_t *cpu)
2704 {
2705 #if defined(__amd64)
2706 	return (1);
2707 #else
2708 	struct cpuid_info *cpi;
2709 
2710 	ASSERT(cpu != NULL);
2711 	cpi = cpu->cpu_m.mcpu_cpi;
2712 	if (cpi->cpi_vendor == X86_VENDOR_AMD && cpi->cpi_maxeax >= 1 &&
2713 	    (CPI_FEATURES_XTD_ECX(cpi) & CPUID_AMD_ECX_CR8D) != 0)
2714 		return (1);
2715 	return (0);
2716 #endif
2717 }
2718 
2719 uint32_t
2720 cpuid_get_apicid(cpu_t *cpu)
2721 {
2722 	ASSERT(cpuid_checkpass(cpu, 1));
2723 	if (cpu->cpu_m.mcpu_cpi->cpi_maxeax < 1) {
2724 		return (UINT32_MAX);
2725 	} else {
2726 		return (cpu->cpu_m.mcpu_cpi->cpi_apicid);
2727 	}
2728 }
2729 
2730 void
2731 cpuid_get_addrsize(cpu_t *cpu, uint_t *pabits, uint_t *vabits)
2732 {
2733 	struct cpuid_info *cpi;
2734 
2735 	if (cpu == NULL)
2736 		cpu = CPU;
2737 	cpi = cpu->cpu_m.mcpu_cpi;
2738 
2739 	ASSERT(cpuid_checkpass(cpu, 1));
2740 
2741 	if (pabits)
2742 		*pabits = cpi->cpi_pabits;
2743 	if (vabits)
2744 		*vabits = cpi->cpi_vabits;
2745 }
2746 
2747 /*
2748  * Returns the number of data TLB entries for a corresponding
2749  * pagesize.  If it can't be computed, or isn't known, the
2750  * routine returns zero.  If you ask about an architecturally
2751  * impossible pagesize, the routine will panic (so that the
2752  * hat implementor knows that things are inconsistent.)
2753  */
2754 uint_t
2755 cpuid_get_dtlb_nent(cpu_t *cpu, size_t pagesize)
2756 {
2757 	struct cpuid_info *cpi;
2758 	uint_t dtlb_nent = 0;
2759 
2760 	if (cpu == NULL)
2761 		cpu = CPU;
2762 	cpi = cpu->cpu_m.mcpu_cpi;
2763 
2764 	ASSERT(cpuid_checkpass(cpu, 1));
2765 
2766 	/*
2767 	 * Check the L2 TLB info
2768 	 */
2769 	if (cpi->cpi_xmaxeax >= 0x80000006) {
2770 		struct cpuid_regs *cp = &cpi->cpi_extd[6];
2771 
2772 		switch (pagesize) {
2773 
2774 		case 4 * 1024:
2775 			/*
2776 			 * All zero in the top 16 bits of the register
2777 			 * indicates a unified TLB. Size is in low 16 bits.
2778 			 */
2779 			if ((cp->cp_ebx & 0xffff0000) == 0)
2780 				dtlb_nent = cp->cp_ebx & 0x0000ffff;
2781 			else
2782 				dtlb_nent = BITX(cp->cp_ebx, 27, 16);
2783 			break;
2784 
2785 		case 2 * 1024 * 1024:
2786 			if ((cp->cp_eax & 0xffff0000) == 0)
2787 				dtlb_nent = cp->cp_eax & 0x0000ffff;
2788 			else
2789 				dtlb_nent = BITX(cp->cp_eax, 27, 16);
2790 			break;
2791 
2792 		default:
2793 			panic("unknown L2 pagesize");
2794 			/*NOTREACHED*/
2795 		}
2796 	}
2797 
2798 	if (dtlb_nent != 0)
2799 		return (dtlb_nent);
2800 
2801 	/*
2802 	 * No L2 TLB support for this size, try L1.
2803 	 */
2804 	if (cpi->cpi_xmaxeax >= 0x80000005) {
2805 		struct cpuid_regs *cp = &cpi->cpi_extd[5];
2806 
2807 		switch (pagesize) {
2808 		case 4 * 1024:
2809 			dtlb_nent = BITX(cp->cp_ebx, 23, 16);
2810 			break;
2811 		case 2 * 1024 * 1024:
2812 			dtlb_nent = BITX(cp->cp_eax, 23, 16);
2813 			break;
2814 		default:
2815 			panic("unknown L1 d-TLB pagesize");
2816 			/*NOTREACHED*/
2817 		}
2818 	}
2819 
2820 	return (dtlb_nent);
2821 }
2822 
2823 /*
2824  * Return 0 if the erratum is not present or not applicable, positive
2825  * if it is, and negative if the status of the erratum is unknown.
2826  *
2827  * See "Revision Guide for AMD Athlon(tm) 64 and AMD Opteron(tm)
2828  * Processors" #25759, Rev 3.57, August 2005
2829  */
2830 int
2831 cpuid_opteron_erratum(cpu_t *cpu, uint_t erratum)
2832 {
2833 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2834 	uint_t eax;
2835 
2836 	/*
2837 	 * Bail out if this CPU isn't an AMD CPU, or if it's
2838 	 * a legacy (32-bit) AMD CPU.
2839 	 */
2840 	if (cpi->cpi_vendor != X86_VENDOR_AMD ||
2841 	    cpi->cpi_family == 4 || cpi->cpi_family == 5 ||
2842 	    cpi->cpi_family == 6)
2843 
2844 		return (0);
2845 
2846 	eax = cpi->cpi_std[1].cp_eax;
2847 
2848 #define	SH_B0(eax)	(eax == 0xf40 || eax == 0xf50)
2849 #define	SH_B3(eax) 	(eax == 0xf51)
2850 #define	B(eax)		(SH_B0(eax) || SH_B3(eax))
2851 
2852 #define	SH_C0(eax)	(eax == 0xf48 || eax == 0xf58)
2853 
2854 #define	SH_CG(eax)	(eax == 0xf4a || eax == 0xf5a || eax == 0xf7a)
2855 #define	DH_CG(eax)	(eax == 0xfc0 || eax == 0xfe0 || eax == 0xff0)
2856 #define	CH_CG(eax)	(eax == 0xf82 || eax == 0xfb2)
2857 #define	CG(eax)		(SH_CG(eax) || DH_CG(eax) || CH_CG(eax))
2858 
2859 #define	SH_D0(eax)	(eax == 0x10f40 || eax == 0x10f50 || eax == 0x10f70)
2860 #define	DH_D0(eax)	(eax == 0x10fc0 || eax == 0x10ff0)
2861 #define	CH_D0(eax)	(eax == 0x10f80 || eax == 0x10fb0)
2862 #define	D0(eax)		(SH_D0(eax) || DH_D0(eax) || CH_D0(eax))
2863 
2864 #define	SH_E0(eax)	(eax == 0x20f50 || eax == 0x20f40 || eax == 0x20f70)
2865 #define	JH_E1(eax)	(eax == 0x20f10)	/* JH8_E0 had 0x20f30 */
2866 #define	DH_E3(eax)	(eax == 0x20fc0 || eax == 0x20ff0)
2867 #define	SH_E4(eax)	(eax == 0x20f51 || eax == 0x20f71)
2868 #define	BH_E4(eax)	(eax == 0x20fb1)
2869 #define	SH_E5(eax)	(eax == 0x20f42)
2870 #define	DH_E6(eax)	(eax == 0x20ff2 || eax == 0x20fc2)
2871 #define	JH_E6(eax)	(eax == 0x20f12 || eax == 0x20f32)
2872 #define	EX(eax)		(SH_E0(eax) || JH_E1(eax) || DH_E3(eax) || \
2873 			    SH_E4(eax) || BH_E4(eax) || SH_E5(eax) || \
2874 			    DH_E6(eax) || JH_E6(eax))
2875 
2876 #define	DR_AX(eax)	(eax == 0x100f00 || eax == 0x100f01 || eax == 0x100f02)
2877 #define	DR_B0(eax)	(eax == 0x100f20)
2878 #define	DR_B1(eax)	(eax == 0x100f21)
2879 #define	DR_BA(eax)	(eax == 0x100f2a)
2880 #define	DR_B2(eax)	(eax == 0x100f22)
2881 #define	DR_B3(eax)	(eax == 0x100f23)
2882 #define	RB_C0(eax)	(eax == 0x100f40)
2883 
2884 	switch (erratum) {
2885 	case 1:
2886 		return (cpi->cpi_family < 0x10);
2887 	case 51:	/* what does the asterisk mean? */
2888 		return (B(eax) || SH_C0(eax) || CG(eax));
2889 	case 52:
2890 		return (B(eax));
2891 	case 57:
2892 		return (cpi->cpi_family <= 0x11);
2893 	case 58:
2894 		return (B(eax));
2895 	case 60:
2896 		return (cpi->cpi_family <= 0x11);
2897 	case 61:
2898 	case 62:
2899 	case 63:
2900 	case 64:
2901 	case 65:
2902 	case 66:
2903 	case 68:
2904 	case 69:
2905 	case 70:
2906 	case 71:
2907 		return (B(eax));
2908 	case 72:
2909 		return (SH_B0(eax));
2910 	case 74:
2911 		return (B(eax));
2912 	case 75:
2913 		return (cpi->cpi_family < 0x10);
2914 	case 76:
2915 		return (B(eax));
2916 	case 77:
2917 		return (cpi->cpi_family <= 0x11);
2918 	case 78:
2919 		return (B(eax) || SH_C0(eax));
2920 	case 79:
2921 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
2922 	case 80:
2923 	case 81:
2924 	case 82:
2925 		return (B(eax));
2926 	case 83:
2927 		return (B(eax) || SH_C0(eax) || CG(eax));
2928 	case 85:
2929 		return (cpi->cpi_family < 0x10);
2930 	case 86:
2931 		return (SH_C0(eax) || CG(eax));
2932 	case 88:
2933 #if !defined(__amd64)
2934 		return (0);
2935 #else
2936 		return (B(eax) || SH_C0(eax));
2937 #endif
2938 	case 89:
2939 		return (cpi->cpi_family < 0x10);
2940 	case 90:
2941 		return (B(eax) || SH_C0(eax) || CG(eax));
2942 	case 91:
2943 	case 92:
2944 		return (B(eax) || SH_C0(eax));
2945 	case 93:
2946 		return (SH_C0(eax));
2947 	case 94:
2948 		return (B(eax) || SH_C0(eax) || CG(eax));
2949 	case 95:
2950 #if !defined(__amd64)
2951 		return (0);
2952 #else
2953 		return (B(eax) || SH_C0(eax));
2954 #endif
2955 	case 96:
2956 		return (B(eax) || SH_C0(eax) || CG(eax));
2957 	case 97:
2958 	case 98:
2959 		return (SH_C0(eax) || CG(eax));
2960 	case 99:
2961 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
2962 	case 100:
2963 		return (B(eax) || SH_C0(eax));
2964 	case 101:
2965 	case 103:
2966 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
2967 	case 104:
2968 		return (SH_C0(eax) || CG(eax) || D0(eax));
2969 	case 105:
2970 	case 106:
2971 	case 107:
2972 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
2973 	case 108:
2974 		return (DH_CG(eax));
2975 	case 109:
2976 		return (SH_C0(eax) || CG(eax) || D0(eax));
2977 	case 110:
2978 		return (D0(eax) || EX(eax));
2979 	case 111:
2980 		return (CG(eax));
2981 	case 112:
2982 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
2983 	case 113:
2984 		return (eax == 0x20fc0);
2985 	case 114:
2986 		return (SH_E0(eax) || JH_E1(eax) || DH_E3(eax));
2987 	case 115:
2988 		return (SH_E0(eax) || JH_E1(eax));
2989 	case 116:
2990 		return (SH_E0(eax) || JH_E1(eax) || DH_E3(eax));
2991 	case 117:
2992 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
2993 	case 118:
2994 		return (SH_E0(eax) || JH_E1(eax) || SH_E4(eax) || BH_E4(eax) ||
2995 		    JH_E6(eax));
2996 	case 121:
2997 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
2998 	case 122:
2999 		return (cpi->cpi_family < 0x10 || cpi->cpi_family == 0x11);
3000 	case 123:
3001 		return (JH_E1(eax) || BH_E4(eax) || JH_E6(eax));
3002 	case 131:
3003 		return (cpi->cpi_family < 0x10);
3004 	case 6336786:
3005 		/*
3006 		 * Test for AdvPowerMgmtInfo.TscPStateInvariant
3007 		 * if this is a K8 family or newer processor
3008 		 */
3009 		if (CPI_FAMILY(cpi) == 0xf) {
3010 			struct cpuid_regs regs;
3011 			regs.cp_eax = 0x80000007;
3012 			(void) __cpuid_insn(&regs);
3013 			return (!(regs.cp_edx & 0x100));
3014 		}
3015 		return (0);
3016 	case 6323525:
3017 		return (((((eax >> 12) & 0xff00) + (eax & 0xf00)) |
3018 		    (((eax >> 4) & 0xf) | ((eax >> 12) & 0xf0))) < 0xf40);
3019 
3020 	case 6671130:
3021 		/*
3022 		 * check for processors (pre-Shanghai) that do not provide
3023 		 * optimal management of 1gb ptes in its tlb.
3024 		 */
3025 		return (cpi->cpi_family == 0x10 && cpi->cpi_model < 4);
3026 
3027 	case 298:
3028 		return (DR_AX(eax) || DR_B0(eax) || DR_B1(eax) || DR_BA(eax) ||
3029 		    DR_B2(eax) || RB_C0(eax));
3030 
3031 	default:
3032 		return (-1);
3033 
3034 	}
3035 }
3036 
3037 /*
3038  * Determine if specified erratum is present via OSVW (OS Visible Workaround).
3039  * Return 1 if erratum is present, 0 if not present and -1 if indeterminate.
3040  */
3041 int
3042 osvw_opteron_erratum(cpu_t *cpu, uint_t erratum)
3043 {
3044 	struct cpuid_info	*cpi;
3045 	uint_t			osvwid;
3046 	static int		osvwfeature = -1;
3047 	uint64_t		osvwlength;
3048 
3049 
3050 	cpi = cpu->cpu_m.mcpu_cpi;
3051 
3052 	/* confirm OSVW supported */
3053 	if (osvwfeature == -1) {
3054 		osvwfeature = cpi->cpi_extd[1].cp_ecx & CPUID_AMD_ECX_OSVW;
3055 	} else {
3056 		/* assert that osvw feature setting is consistent on all cpus */
3057 		ASSERT(osvwfeature ==
3058 		    (cpi->cpi_extd[1].cp_ecx & CPUID_AMD_ECX_OSVW));
3059 	}
3060 	if (!osvwfeature)
3061 		return (-1);
3062 
3063 	osvwlength = rdmsr(MSR_AMD_OSVW_ID_LEN) & OSVW_ID_LEN_MASK;
3064 
3065 	switch (erratum) {
3066 	case 298:	/* osvwid is 0 */
3067 		osvwid = 0;
3068 		if (osvwlength <= (uint64_t)osvwid) {
3069 			/* osvwid 0 is unknown */
3070 			return (-1);
3071 		}
3072 
3073 		/*
3074 		 * Check the OSVW STATUS MSR to determine the state
3075 		 * of the erratum where:
3076 		 *   0 - fixed by HW
3077 		 *   1 - BIOS has applied the workaround when BIOS
3078 		 *   workaround is available. (Or for other errata,
3079 		 *   OS workaround is required.)
3080 		 * For a value of 1, caller will confirm that the
3081 		 * erratum 298 workaround has indeed been applied by BIOS.
3082 		 *
3083 		 * A 1 may be set in cpus that have a HW fix
3084 		 * in a mixed cpu system. Regarding erratum 298:
3085 		 *   In a multiprocessor platform, the workaround above
3086 		 *   should be applied to all processors regardless of
3087 		 *   silicon revision when an affected processor is
3088 		 *   present.
3089 		 */
3090 
3091 		return (rdmsr(MSR_AMD_OSVW_STATUS +
3092 		    (osvwid / OSVW_ID_CNT_PER_MSR)) &
3093 		    (1ULL << (osvwid % OSVW_ID_CNT_PER_MSR)));
3094 
3095 	default:
3096 		return (-1);
3097 	}
3098 }
3099 
3100 static const char assoc_str[] = "associativity";
3101 static const char line_str[] = "line-size";
3102 static const char size_str[] = "size";
3103 
3104 static void
3105 add_cache_prop(dev_info_t *devi, const char *label, const char *type,
3106     uint32_t val)
3107 {
3108 	char buf[128];
3109 
3110 	/*
3111 	 * ndi_prop_update_int() is used because it is desirable for
3112 	 * DDI_PROP_HW_DEF and DDI_PROP_DONTSLEEP to be set.
3113 	 */
3114 	if (snprintf(buf, sizeof (buf), "%s-%s", label, type) < sizeof (buf))
3115 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, devi, buf, val);
3116 }
3117 
3118 /*
3119  * Intel-style cache/tlb description
3120  *
3121  * Standard cpuid level 2 gives a randomly ordered
3122  * selection of tags that index into a table that describes
3123  * cache and tlb properties.
3124  */
3125 
3126 static const char l1_icache_str[] = "l1-icache";
3127 static const char l1_dcache_str[] = "l1-dcache";
3128 static const char l2_cache_str[] = "l2-cache";
3129 static const char l3_cache_str[] = "l3-cache";
3130 static const char itlb4k_str[] = "itlb-4K";
3131 static const char dtlb4k_str[] = "dtlb-4K";
3132 static const char itlb2M_str[] = "itlb-2M";
3133 static const char itlb4M_str[] = "itlb-4M";
3134 static const char dtlb4M_str[] = "dtlb-4M";
3135 static const char dtlb24_str[] = "dtlb0-2M-4M";
3136 static const char itlb424_str[] = "itlb-4K-2M-4M";
3137 static const char itlb24_str[] = "itlb-2M-4M";
3138 static const char dtlb44_str[] = "dtlb-4K-4M";
3139 static const char sl1_dcache_str[] = "sectored-l1-dcache";
3140 static const char sl2_cache_str[] = "sectored-l2-cache";
3141 static const char itrace_str[] = "itrace-cache";
3142 static const char sl3_cache_str[] = "sectored-l3-cache";
3143 static const char sh_l2_tlb4k_str[] = "shared-l2-tlb-4k";
3144 
3145 static const struct cachetab {
3146 	uint8_t 	ct_code;
3147 	uint8_t		ct_assoc;
3148 	uint16_t 	ct_line_size;
3149 	size_t		ct_size;
3150 	const char	*ct_label;
3151 } intel_ctab[] = {
3152 	/*
3153 	 * maintain descending order!
3154 	 *
3155 	 * Codes ignored - Reason
3156 	 * ----------------------
3157 	 * 40H - intel_cpuid_4_cache_info() disambiguates l2/l3 cache
3158 	 * f0H/f1H - Currently we do not interpret prefetch size by design
3159 	 */
3160 	{ 0xe4, 16, 64, 8*1024*1024, l3_cache_str},
3161 	{ 0xe3, 16, 64, 4*1024*1024, l3_cache_str},
3162 	{ 0xe2, 16, 64, 2*1024*1024, l3_cache_str},
3163 	{ 0xde, 12, 64, 6*1024*1024, l3_cache_str},
3164 	{ 0xdd, 12, 64, 3*1024*1024, l3_cache_str},
3165 	{ 0xdc, 12, 64, ((1*1024*1024)+(512*1024)), l3_cache_str},
3166 	{ 0xd8, 8, 64, 4*1024*1024, l3_cache_str},
3167 	{ 0xd7, 8, 64, 2*1024*1024, l3_cache_str},
3168 	{ 0xd6, 8, 64, 1*1024*1024, l3_cache_str},
3169 	{ 0xd2, 4, 64, 2*1024*1024, l3_cache_str},
3170 	{ 0xd1, 4, 64, 1*1024*1024, l3_cache_str},
3171 	{ 0xd0, 4, 64, 512*1024, l3_cache_str},
3172 	{ 0xca, 4, 0, 512, sh_l2_tlb4k_str},
3173 	{ 0xc0, 4, 0, 8, dtlb44_str },
3174 	{ 0xba, 4, 0, 64, dtlb4k_str },
3175 	{ 0xb4, 4, 0, 256, dtlb4k_str },
3176 	{ 0xb3, 4, 0, 128, dtlb4k_str },
3177 	{ 0xb2, 4, 0, 64, itlb4k_str },
3178 	{ 0xb0, 4, 0, 128, itlb4k_str },
3179 	{ 0x87, 8, 64, 1024*1024, l2_cache_str},
3180 	{ 0x86, 4, 64, 512*1024, l2_cache_str},
3181 	{ 0x85, 8, 32, 2*1024*1024, l2_cache_str},
3182 	{ 0x84, 8, 32, 1024*1024, l2_cache_str},
3183 	{ 0x83, 8, 32, 512*1024, l2_cache_str},
3184 	{ 0x82, 8, 32, 256*1024, l2_cache_str},
3185 	{ 0x80, 8, 64, 512*1024, l2_cache_str},
3186 	{ 0x7f, 2, 64, 512*1024, l2_cache_str},
3187 	{ 0x7d, 8, 64, 2*1024*1024, sl2_cache_str},
3188 	{ 0x7c, 8, 64, 1024*1024, sl2_cache_str},
3189 	{ 0x7b, 8, 64, 512*1024, sl2_cache_str},
3190 	{ 0x7a, 8, 64, 256*1024, sl2_cache_str},
3191 	{ 0x79, 8, 64, 128*1024, sl2_cache_str},
3192 	{ 0x78, 8, 64, 1024*1024, l2_cache_str},
3193 	{ 0x73, 8, 0, 64*1024, itrace_str},
3194 	{ 0x72, 8, 0, 32*1024, itrace_str},
3195 	{ 0x71, 8, 0, 16*1024, itrace_str},
3196 	{ 0x70, 8, 0, 12*1024, itrace_str},
3197 	{ 0x68, 4, 64, 32*1024, sl1_dcache_str},
3198 	{ 0x67, 4, 64, 16*1024, sl1_dcache_str},
3199 	{ 0x66, 4, 64, 8*1024, sl1_dcache_str},
3200 	{ 0x60, 8, 64, 16*1024, sl1_dcache_str},
3201 	{ 0x5d, 0, 0, 256, dtlb44_str},
3202 	{ 0x5c, 0, 0, 128, dtlb44_str},
3203 	{ 0x5b, 0, 0, 64, dtlb44_str},
3204 	{ 0x5a, 4, 0, 32, dtlb24_str},
3205 	{ 0x59, 0, 0, 16, dtlb4k_str},
3206 	{ 0x57, 4, 0, 16, dtlb4k_str},
3207 	{ 0x56, 4, 0, 16, dtlb4M_str},
3208 	{ 0x55, 0, 0, 7, itlb24_str},
3209 	{ 0x52, 0, 0, 256, itlb424_str},
3210 	{ 0x51, 0, 0, 128, itlb424_str},
3211 	{ 0x50, 0, 0, 64, itlb424_str},
3212 	{ 0x4f, 0, 0, 32, itlb4k_str},
3213 	{ 0x4e, 24, 64, 6*1024*1024, l2_cache_str},
3214 	{ 0x4d, 16, 64, 16*1024*1024, l3_cache_str},
3215 	{ 0x4c, 12, 64, 12*1024*1024, l3_cache_str},
3216 	{ 0x4b, 16, 64, 8*1024*1024, l3_cache_str},
3217 	{ 0x4a, 12, 64, 6*1024*1024, l3_cache_str},
3218 	{ 0x49, 16, 64, 4*1024*1024, l3_cache_str},
3219 	{ 0x48, 12, 64, 3*1024*1024, l2_cache_str},
3220 	{ 0x47, 8, 64, 8*1024*1024, l3_cache_str},
3221 	{ 0x46, 4, 64, 4*1024*1024, l3_cache_str},
3222 	{ 0x45, 4, 32, 2*1024*1024, l2_cache_str},
3223 	{ 0x44, 4, 32, 1024*1024, l2_cache_str},
3224 	{ 0x43, 4, 32, 512*1024, l2_cache_str},
3225 	{ 0x42, 4, 32, 256*1024, l2_cache_str},
3226 	{ 0x41, 4, 32, 128*1024, l2_cache_str},
3227 	{ 0x3e, 4, 64, 512*1024, sl2_cache_str},
3228 	{ 0x3d, 6, 64, 384*1024, sl2_cache_str},
3229 	{ 0x3c, 4, 64, 256*1024, sl2_cache_str},
3230 	{ 0x3b, 2, 64, 128*1024, sl2_cache_str},
3231 	{ 0x3a, 6, 64, 192*1024, sl2_cache_str},
3232 	{ 0x39, 4, 64, 128*1024, sl2_cache_str},
3233 	{ 0x30, 8, 64, 32*1024, l1_icache_str},
3234 	{ 0x2c, 8, 64, 32*1024, l1_dcache_str},
3235 	{ 0x29, 8, 64, 4096*1024, sl3_cache_str},
3236 	{ 0x25, 8, 64, 2048*1024, sl3_cache_str},
3237 	{ 0x23, 8, 64, 1024*1024, sl3_cache_str},
3238 	{ 0x22, 4, 64, 512*1024, sl3_cache_str},
3239 	{ 0x0e, 6, 64, 24*1024, l1_dcache_str},
3240 	{ 0x0d, 4, 32, 16*1024, l1_dcache_str},
3241 	{ 0x0c, 4, 32, 16*1024, l1_dcache_str},
3242 	{ 0x0b, 4, 0, 4, itlb4M_str},
3243 	{ 0x0a, 2, 32, 8*1024, l1_dcache_str},
3244 	{ 0x08, 4, 32, 16*1024, l1_icache_str},
3245 	{ 0x06, 4, 32, 8*1024, l1_icache_str},
3246 	{ 0x05, 4, 0, 32, dtlb4M_str},
3247 	{ 0x04, 4, 0, 8, dtlb4M_str},
3248 	{ 0x03, 4, 0, 64, dtlb4k_str},
3249 	{ 0x02, 4, 0, 2, itlb4M_str},
3250 	{ 0x01, 4, 0, 32, itlb4k_str},
3251 	{ 0 }
3252 };
3253 
3254 static const struct cachetab cyrix_ctab[] = {
3255 	{ 0x70, 4, 0, 32, "tlb-4K" },
3256 	{ 0x80, 4, 16, 16*1024, "l1-cache" },
3257 	{ 0 }
3258 };
3259 
3260 /*
3261  * Search a cache table for a matching entry
3262  */
3263 static const struct cachetab *
3264 find_cacheent(const struct cachetab *ct, uint_t code)
3265 {
3266 	if (code != 0) {
3267 		for (; ct->ct_code != 0; ct++)
3268 			if (ct->ct_code <= code)
3269 				break;
3270 		if (ct->ct_code == code)
3271 			return (ct);
3272 	}
3273 	return (NULL);
3274 }
3275 
3276 /*
3277  * Populate cachetab entry with L2 or L3 cache-information using
3278  * cpuid function 4. This function is called from intel_walk_cacheinfo()
3279  * when descriptor 0x49 is encountered. It returns 0 if no such cache
3280  * information is found.
3281  */
3282 static int
3283 intel_cpuid_4_cache_info(struct cachetab *ct, struct cpuid_info *cpi)
3284 {
3285 	uint32_t level, i;
3286 	int ret = 0;
3287 
3288 	for (i = 0; i < cpi->cpi_std_4_size; i++) {
3289 		level = CPI_CACHE_LVL(cpi->cpi_std_4[i]);
3290 
3291 		if (level == 2 || level == 3) {
3292 			ct->ct_assoc = CPI_CACHE_WAYS(cpi->cpi_std_4[i]) + 1;
3293 			ct->ct_line_size =
3294 			    CPI_CACHE_COH_LN_SZ(cpi->cpi_std_4[i]) + 1;
3295 			ct->ct_size = ct->ct_assoc *
3296 			    (CPI_CACHE_PARTS(cpi->cpi_std_4[i]) + 1) *
3297 			    ct->ct_line_size *
3298 			    (cpi->cpi_std_4[i]->cp_ecx + 1);
3299 
3300 			if (level == 2) {
3301 				ct->ct_label = l2_cache_str;
3302 			} else if (level == 3) {
3303 				ct->ct_label = l3_cache_str;
3304 			}
3305 			ret = 1;
3306 		}
3307 	}
3308 
3309 	return (ret);
3310 }
3311 
3312 /*
3313  * Walk the cacheinfo descriptor, applying 'func' to every valid element
3314  * The walk is terminated if the walker returns non-zero.
3315  */
3316 static void
3317 intel_walk_cacheinfo(struct cpuid_info *cpi,
3318     void *arg, int (*func)(void *, const struct cachetab *))
3319 {
3320 	const struct cachetab *ct;
3321 	struct cachetab des_49_ct, des_b1_ct;
3322 	uint8_t *dp;
3323 	int i;
3324 
3325 	if ((dp = cpi->cpi_cacheinfo) == NULL)
3326 		return;
3327 	for (i = 0; i < cpi->cpi_ncache; i++, dp++) {
3328 		/*
3329 		 * For overloaded descriptor 0x49 we use cpuid function 4
3330 		 * if supported by the current processor, to create
3331 		 * cache information.
3332 		 * For overloaded descriptor 0xb1 we use X86_PAE flag
3333 		 * to disambiguate the cache information.
3334 		 */
3335 		if (*dp == 0x49 && cpi->cpi_maxeax >= 0x4 &&
3336 		    intel_cpuid_4_cache_info(&des_49_ct, cpi) == 1) {
3337 				ct = &des_49_ct;
3338 		} else if (*dp == 0xb1) {
3339 			des_b1_ct.ct_code = 0xb1;
3340 			des_b1_ct.ct_assoc = 4;
3341 			des_b1_ct.ct_line_size = 0;
3342 			if (x86_feature & X86_PAE) {
3343 				des_b1_ct.ct_size = 8;
3344 				des_b1_ct.ct_label = itlb2M_str;
3345 			} else {
3346 				des_b1_ct.ct_size = 4;
3347 				des_b1_ct.ct_label = itlb4M_str;
3348 			}
3349 			ct = &des_b1_ct;
3350 		} else {
3351 			if ((ct = find_cacheent(intel_ctab, *dp)) == NULL) {
3352 				continue;
3353 			}
3354 		}
3355 
3356 		if (func(arg, ct) != 0) {
3357 			break;
3358 		}
3359 	}
3360 }
3361 
3362 /*
3363  * (Like the Intel one, except for Cyrix CPUs)
3364  */
3365 static void
3366 cyrix_walk_cacheinfo(struct cpuid_info *cpi,
3367     void *arg, int (*func)(void *, const struct cachetab *))
3368 {
3369 	const struct cachetab *ct;
3370 	uint8_t *dp;
3371 	int i;
3372 
3373 	if ((dp = cpi->cpi_cacheinfo) == NULL)
3374 		return;
3375 	for (i = 0; i < cpi->cpi_ncache; i++, dp++) {
3376 		/*
3377 		 * Search Cyrix-specific descriptor table first ..
3378 		 */
3379 		if ((ct = find_cacheent(cyrix_ctab, *dp)) != NULL) {
3380 			if (func(arg, ct) != 0)
3381 				break;
3382 			continue;
3383 		}
3384 		/*
3385 		 * .. else fall back to the Intel one
3386 		 */
3387 		if ((ct = find_cacheent(intel_ctab, *dp)) != NULL) {
3388 			if (func(arg, ct) != 0)
3389 				break;
3390 			continue;
3391 		}
3392 	}
3393 }
3394 
3395 /*
3396  * A cacheinfo walker that adds associativity, line-size, and size properties
3397  * to the devinfo node it is passed as an argument.
3398  */
3399 static int
3400 add_cacheent_props(void *arg, const struct cachetab *ct)
3401 {
3402 	dev_info_t *devi = arg;
3403 
3404 	add_cache_prop(devi, ct->ct_label, assoc_str, ct->ct_assoc);
3405 	if (ct->ct_line_size != 0)
3406 		add_cache_prop(devi, ct->ct_label, line_str,
3407 		    ct->ct_line_size);
3408 	add_cache_prop(devi, ct->ct_label, size_str, ct->ct_size);
3409 	return (0);
3410 }
3411 
3412 
3413 static const char fully_assoc[] = "fully-associative?";
3414 
3415 /*
3416  * AMD style cache/tlb description
3417  *
3418  * Extended functions 5 and 6 directly describe properties of
3419  * tlbs and various cache levels.
3420  */
3421 static void
3422 add_amd_assoc(dev_info_t *devi, const char *label, uint_t assoc)
3423 {
3424 	switch (assoc) {
3425 	case 0:	/* reserved; ignore */
3426 		break;
3427 	default:
3428 		add_cache_prop(devi, label, assoc_str, assoc);
3429 		break;
3430 	case 0xff:
3431 		add_cache_prop(devi, label, fully_assoc, 1);
3432 		break;
3433 	}
3434 }
3435 
3436 static void
3437 add_amd_tlb(dev_info_t *devi, const char *label, uint_t assoc, uint_t size)
3438 {
3439 	if (size == 0)
3440 		return;
3441 	add_cache_prop(devi, label, size_str, size);
3442 	add_amd_assoc(devi, label, assoc);
3443 }
3444 
3445 static void
3446 add_amd_cache(dev_info_t *devi, const char *label,
3447     uint_t size, uint_t assoc, uint_t lines_per_tag, uint_t line_size)
3448 {
3449 	if (size == 0 || line_size == 0)
3450 		return;
3451 	add_amd_assoc(devi, label, assoc);
3452 	/*
3453 	 * Most AMD parts have a sectored cache. Multiple cache lines are
3454 	 * associated with each tag. A sector consists of all cache lines
3455 	 * associated with a tag. For example, the AMD K6-III has a sector
3456 	 * size of 2 cache lines per tag.
3457 	 */
3458 	if (lines_per_tag != 0)
3459 		add_cache_prop(devi, label, "lines-per-tag", lines_per_tag);
3460 	add_cache_prop(devi, label, line_str, line_size);
3461 	add_cache_prop(devi, label, size_str, size * 1024);
3462 }
3463 
3464 static void
3465 add_amd_l2_assoc(dev_info_t *devi, const char *label, uint_t assoc)
3466 {
3467 	switch (assoc) {
3468 	case 0:	/* off */
3469 		break;
3470 	case 1:
3471 	case 2:
3472 	case 4:
3473 		add_cache_prop(devi, label, assoc_str, assoc);
3474 		break;
3475 	case 6:
3476 		add_cache_prop(devi, label, assoc_str, 8);
3477 		break;
3478 	case 8:
3479 		add_cache_prop(devi, label, assoc_str, 16);
3480 		break;
3481 	case 0xf:
3482 		add_cache_prop(devi, label, fully_assoc, 1);
3483 		break;
3484 	default: /* reserved; ignore */
3485 		break;
3486 	}
3487 }
3488 
3489 static void
3490 add_amd_l2_tlb(dev_info_t *devi, const char *label, uint_t assoc, uint_t size)
3491 {
3492 	if (size == 0 || assoc == 0)
3493 		return;
3494 	add_amd_l2_assoc(devi, label, assoc);
3495 	add_cache_prop(devi, label, size_str, size);
3496 }
3497 
3498 static void
3499 add_amd_l2_cache(dev_info_t *devi, const char *label,
3500     uint_t size, uint_t assoc, uint_t lines_per_tag, uint_t line_size)
3501 {
3502 	if (size == 0 || assoc == 0 || line_size == 0)
3503 		return;
3504 	add_amd_l2_assoc(devi, label, assoc);
3505 	if (lines_per_tag != 0)
3506 		add_cache_prop(devi, label, "lines-per-tag", lines_per_tag);
3507 	add_cache_prop(devi, label, line_str, line_size);
3508 	add_cache_prop(devi, label, size_str, size * 1024);
3509 }
3510 
3511 static void
3512 amd_cache_info(struct cpuid_info *cpi, dev_info_t *devi)
3513 {
3514 	struct cpuid_regs *cp;
3515 
3516 	if (cpi->cpi_xmaxeax < 0x80000005)
3517 		return;
3518 	cp = &cpi->cpi_extd[5];
3519 
3520 	/*
3521 	 * 4M/2M L1 TLB configuration
3522 	 *
3523 	 * We report the size for 2M pages because AMD uses two
3524 	 * TLB entries for one 4M page.
3525 	 */
3526 	add_amd_tlb(devi, "dtlb-2M",
3527 	    BITX(cp->cp_eax, 31, 24), BITX(cp->cp_eax, 23, 16));
3528 	add_amd_tlb(devi, "itlb-2M",
3529 	    BITX(cp->cp_eax, 15, 8), BITX(cp->cp_eax, 7, 0));
3530 
3531 	/*
3532 	 * 4K L1 TLB configuration
3533 	 */
3534 
3535 	switch (cpi->cpi_vendor) {
3536 		uint_t nentries;
3537 	case X86_VENDOR_TM:
3538 		if (cpi->cpi_family >= 5) {
3539 			/*
3540 			 * Crusoe processors have 256 TLB entries, but
3541 			 * cpuid data format constrains them to only
3542 			 * reporting 255 of them.
3543 			 */
3544 			if ((nentries = BITX(cp->cp_ebx, 23, 16)) == 255)
3545 				nentries = 256;
3546 			/*
3547 			 * Crusoe processors also have a unified TLB
3548 			 */
3549 			add_amd_tlb(devi, "tlb-4K", BITX(cp->cp_ebx, 31, 24),
3550 			    nentries);
3551 			break;
3552 		}
3553 		/*FALLTHROUGH*/
3554 	default:
3555 		add_amd_tlb(devi, itlb4k_str,
3556 		    BITX(cp->cp_ebx, 31, 24), BITX(cp->cp_ebx, 23, 16));
3557 		add_amd_tlb(devi, dtlb4k_str,
3558 		    BITX(cp->cp_ebx, 15, 8), BITX(cp->cp_ebx, 7, 0));
3559 		break;
3560 	}
3561 
3562 	/*
3563 	 * data L1 cache configuration
3564 	 */
3565 
3566 	add_amd_cache(devi, l1_dcache_str,
3567 	    BITX(cp->cp_ecx, 31, 24), BITX(cp->cp_ecx, 23, 16),
3568 	    BITX(cp->cp_ecx, 15, 8), BITX(cp->cp_ecx, 7, 0));
3569 
3570 	/*
3571 	 * code L1 cache configuration
3572 	 */
3573 
3574 	add_amd_cache(devi, l1_icache_str,
3575 	    BITX(cp->cp_edx, 31, 24), BITX(cp->cp_edx, 23, 16),
3576 	    BITX(cp->cp_edx, 15, 8), BITX(cp->cp_edx, 7, 0));
3577 
3578 	if (cpi->cpi_xmaxeax < 0x80000006)
3579 		return;
3580 	cp = &cpi->cpi_extd[6];
3581 
3582 	/* Check for a unified L2 TLB for large pages */
3583 
3584 	if (BITX(cp->cp_eax, 31, 16) == 0)
3585 		add_amd_l2_tlb(devi, "l2-tlb-2M",
3586 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
3587 	else {
3588 		add_amd_l2_tlb(devi, "l2-dtlb-2M",
3589 		    BITX(cp->cp_eax, 31, 28), BITX(cp->cp_eax, 27, 16));
3590 		add_amd_l2_tlb(devi, "l2-itlb-2M",
3591 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
3592 	}
3593 
3594 	/* Check for a unified L2 TLB for 4K pages */
3595 
3596 	if (BITX(cp->cp_ebx, 31, 16) == 0) {
3597 		add_amd_l2_tlb(devi, "l2-tlb-4K",
3598 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
3599 	} else {
3600 		add_amd_l2_tlb(devi, "l2-dtlb-4K",
3601 		    BITX(cp->cp_eax, 31, 28), BITX(cp->cp_eax, 27, 16));
3602 		add_amd_l2_tlb(devi, "l2-itlb-4K",
3603 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
3604 	}
3605 
3606 	add_amd_l2_cache(devi, l2_cache_str,
3607 	    BITX(cp->cp_ecx, 31, 16), BITX(cp->cp_ecx, 15, 12),
3608 	    BITX(cp->cp_ecx, 11, 8), BITX(cp->cp_ecx, 7, 0));
3609 }
3610 
3611 /*
3612  * There are two basic ways that the x86 world describes it cache
3613  * and tlb architecture - Intel's way and AMD's way.
3614  *
3615  * Return which flavor of cache architecture we should use
3616  */
3617 static int
3618 x86_which_cacheinfo(struct cpuid_info *cpi)
3619 {
3620 	switch (cpi->cpi_vendor) {
3621 	case X86_VENDOR_Intel:
3622 		if (cpi->cpi_maxeax >= 2)
3623 			return (X86_VENDOR_Intel);
3624 		break;
3625 	case X86_VENDOR_AMD:
3626 		/*
3627 		 * The K5 model 1 was the first part from AMD that reported
3628 		 * cache sizes via extended cpuid functions.
3629 		 */
3630 		if (cpi->cpi_family > 5 ||
3631 		    (cpi->cpi_family == 5 && cpi->cpi_model >= 1))
3632 			return (X86_VENDOR_AMD);
3633 		break;
3634 	case X86_VENDOR_TM:
3635 		if (cpi->cpi_family >= 5)
3636 			return (X86_VENDOR_AMD);
3637 		/*FALLTHROUGH*/
3638 	default:
3639 		/*
3640 		 * If they have extended CPU data for 0x80000005
3641 		 * then we assume they have AMD-format cache
3642 		 * information.
3643 		 *
3644 		 * If not, and the vendor happens to be Cyrix,
3645 		 * then try our-Cyrix specific handler.
3646 		 *
3647 		 * If we're not Cyrix, then assume we're using Intel's
3648 		 * table-driven format instead.
3649 		 */
3650 		if (cpi->cpi_xmaxeax >= 0x80000005)
3651 			return (X86_VENDOR_AMD);
3652 		else if (cpi->cpi_vendor == X86_VENDOR_Cyrix)
3653 			return (X86_VENDOR_Cyrix);
3654 		else if (cpi->cpi_maxeax >= 2)
3655 			return (X86_VENDOR_Intel);
3656 		break;
3657 	}
3658 	return (-1);
3659 }
3660 
3661 void
3662 cpuid_set_cpu_properties(void *dip, processorid_t cpu_id,
3663     struct cpuid_info *cpi)
3664 {
3665 	dev_info_t *cpu_devi;
3666 	int create;
3667 
3668 	cpu_devi = (dev_info_t *)dip;
3669 
3670 	/* device_type */
3671 	(void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi,
3672 	    "device_type", "cpu");
3673 
3674 	/* reg */
3675 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3676 	    "reg", cpu_id);
3677 
3678 	/* cpu-mhz, and clock-frequency */
3679 	if (cpu_freq > 0) {
3680 		long long mul;
3681 
3682 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3683 		    "cpu-mhz", cpu_freq);
3684 		if ((mul = cpu_freq * 1000000LL) <= INT_MAX)
3685 			(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3686 			    "clock-frequency", (int)mul);
3687 	}
3688 
3689 	if ((x86_feature & X86_CPUID) == 0) {
3690 		return;
3691 	}
3692 
3693 	/* vendor-id */
3694 	(void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi,
3695 	    "vendor-id", cpi->cpi_vendorstr);
3696 
3697 	if (cpi->cpi_maxeax == 0) {
3698 		return;
3699 	}
3700 
3701 	/*
3702 	 * family, model, and step
3703 	 */
3704 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3705 	    "family", CPI_FAMILY(cpi));
3706 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3707 	    "cpu-model", CPI_MODEL(cpi));
3708 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3709 	    "stepping-id", CPI_STEP(cpi));
3710 
3711 	/* type */
3712 	switch (cpi->cpi_vendor) {
3713 	case X86_VENDOR_Intel:
3714 		create = 1;
3715 		break;
3716 	default:
3717 		create = 0;
3718 		break;
3719 	}
3720 	if (create)
3721 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3722 		    "type", CPI_TYPE(cpi));
3723 
3724 	/* ext-family */
3725 	switch (cpi->cpi_vendor) {
3726 	case X86_VENDOR_Intel:
3727 	case X86_VENDOR_AMD:
3728 		create = cpi->cpi_family >= 0xf;
3729 		break;
3730 	default:
3731 		create = 0;
3732 		break;
3733 	}
3734 	if (create)
3735 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3736 		    "ext-family", CPI_FAMILY_XTD(cpi));
3737 
3738 	/* ext-model */
3739 	switch (cpi->cpi_vendor) {
3740 	case X86_VENDOR_Intel:
3741 		create = IS_EXTENDED_MODEL_INTEL(cpi);
3742 		break;
3743 	case X86_VENDOR_AMD:
3744 		create = CPI_FAMILY(cpi) == 0xf;
3745 		break;
3746 	default:
3747 		create = 0;
3748 		break;
3749 	}
3750 	if (create)
3751 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3752 		    "ext-model", CPI_MODEL_XTD(cpi));
3753 
3754 	/* generation */
3755 	switch (cpi->cpi_vendor) {
3756 	case X86_VENDOR_AMD:
3757 		/*
3758 		 * AMD K5 model 1 was the first part to support this
3759 		 */
3760 		create = cpi->cpi_xmaxeax >= 0x80000001;
3761 		break;
3762 	default:
3763 		create = 0;
3764 		break;
3765 	}
3766 	if (create)
3767 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3768 		    "generation", BITX((cpi)->cpi_extd[1].cp_eax, 11, 8));
3769 
3770 	/* brand-id */
3771 	switch (cpi->cpi_vendor) {
3772 	case X86_VENDOR_Intel:
3773 		/*
3774 		 * brand id first appeared on Pentium III Xeon model 8,
3775 		 * and Celeron model 8 processors and Opteron
3776 		 */
3777 		create = cpi->cpi_family > 6 ||
3778 		    (cpi->cpi_family == 6 && cpi->cpi_model >= 8);
3779 		break;
3780 	case X86_VENDOR_AMD:
3781 		create = cpi->cpi_family >= 0xf;
3782 		break;
3783 	default:
3784 		create = 0;
3785 		break;
3786 	}
3787 	if (create && cpi->cpi_brandid != 0) {
3788 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3789 		    "brand-id", cpi->cpi_brandid);
3790 	}
3791 
3792 	/* chunks, and apic-id */
3793 	switch (cpi->cpi_vendor) {
3794 		/*
3795 		 * first available on Pentium IV and Opteron (K8)
3796 		 */
3797 	case X86_VENDOR_Intel:
3798 		create = IS_NEW_F6(cpi) || cpi->cpi_family >= 0xf;
3799 		break;
3800 	case X86_VENDOR_AMD:
3801 		create = cpi->cpi_family >= 0xf;
3802 		break;
3803 	default:
3804 		create = 0;
3805 		break;
3806 	}
3807 	if (create) {
3808 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3809 		    "chunks", CPI_CHUNKS(cpi));
3810 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3811 		    "apic-id", cpi->cpi_apicid);
3812 		if (cpi->cpi_chipid >= 0) {
3813 			(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3814 			    "chip#", cpi->cpi_chipid);
3815 			(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3816 			    "clog#", cpi->cpi_clogid);
3817 		}
3818 	}
3819 
3820 	/* cpuid-features */
3821 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3822 	    "cpuid-features", CPI_FEATURES_EDX(cpi));
3823 
3824 
3825 	/* cpuid-features-ecx */
3826 	switch (cpi->cpi_vendor) {
3827 	case X86_VENDOR_Intel:
3828 		create = IS_NEW_F6(cpi) || cpi->cpi_family >= 0xf;
3829 		break;
3830 	default:
3831 		create = 0;
3832 		break;
3833 	}
3834 	if (create)
3835 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3836 		    "cpuid-features-ecx", CPI_FEATURES_ECX(cpi));
3837 
3838 	/* ext-cpuid-features */
3839 	switch (cpi->cpi_vendor) {
3840 	case X86_VENDOR_Intel:
3841 	case X86_VENDOR_AMD:
3842 	case X86_VENDOR_Cyrix:
3843 	case X86_VENDOR_TM:
3844 	case X86_VENDOR_Centaur:
3845 		create = cpi->cpi_xmaxeax >= 0x80000001;
3846 		break;
3847 	default:
3848 		create = 0;
3849 		break;
3850 	}
3851 	if (create) {
3852 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3853 		    "ext-cpuid-features", CPI_FEATURES_XTD_EDX(cpi));
3854 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3855 		    "ext-cpuid-features-ecx", CPI_FEATURES_XTD_ECX(cpi));
3856 	}
3857 
3858 	/*
3859 	 * Brand String first appeared in Intel Pentium IV, AMD K5
3860 	 * model 1, and Cyrix GXm.  On earlier models we try and
3861 	 * simulate something similar .. so this string should always
3862 	 * same -something- about the processor, however lame.
3863 	 */
3864 	(void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi,
3865 	    "brand-string", cpi->cpi_brandstr);
3866 
3867 	/*
3868 	 * Finally, cache and tlb information
3869 	 */
3870 	switch (x86_which_cacheinfo(cpi)) {
3871 	case X86_VENDOR_Intel:
3872 		intel_walk_cacheinfo(cpi, cpu_devi, add_cacheent_props);
3873 		break;
3874 	case X86_VENDOR_Cyrix:
3875 		cyrix_walk_cacheinfo(cpi, cpu_devi, add_cacheent_props);
3876 		break;
3877 	case X86_VENDOR_AMD:
3878 		amd_cache_info(cpi, cpu_devi);
3879 		break;
3880 	default:
3881 		break;
3882 	}
3883 }
3884 
3885 struct l2info {
3886 	int *l2i_csz;
3887 	int *l2i_lsz;
3888 	int *l2i_assoc;
3889 	int l2i_ret;
3890 };
3891 
3892 /*
3893  * A cacheinfo walker that fetches the size, line-size and associativity
3894  * of the L2 cache
3895  */
3896 static int
3897 intel_l2cinfo(void *arg, const struct cachetab *ct)
3898 {
3899 	struct l2info *l2i = arg;
3900 	int *ip;
3901 
3902 	if (ct->ct_label != l2_cache_str &&
3903 	    ct->ct_label != sl2_cache_str)
3904 		return (0);	/* not an L2 -- keep walking */
3905 
3906 	if ((ip = l2i->l2i_csz) != NULL)
3907 		*ip = ct->ct_size;
3908 	if ((ip = l2i->l2i_lsz) != NULL)
3909 		*ip = ct->ct_line_size;
3910 	if ((ip = l2i->l2i_assoc) != NULL)
3911 		*ip = ct->ct_assoc;
3912 	l2i->l2i_ret = ct->ct_size;
3913 	return (1);		/* was an L2 -- terminate walk */
3914 }
3915 
3916 /*
3917  * AMD L2/L3 Cache and TLB Associativity Field Definition:
3918  *
3919  *	Unlike the associativity for the L1 cache and tlb where the 8 bit
3920  *	value is the associativity, the associativity for the L2 cache and
3921  *	tlb is encoded in the following table. The 4 bit L2 value serves as
3922  *	an index into the amd_afd[] array to determine the associativity.
3923  *	-1 is undefined. 0 is fully associative.
3924  */
3925 
3926 static int amd_afd[] =
3927 	{-1, 1, 2, -1, 4, -1, 8, -1, 16, -1, 32, 48, 64, 96, 128, 0};
3928 
3929 static void
3930 amd_l2cacheinfo(struct cpuid_info *cpi, struct l2info *l2i)
3931 {
3932 	struct cpuid_regs *cp;
3933 	uint_t size, assoc;
3934 	int i;
3935 	int *ip;
3936 
3937 	if (cpi->cpi_xmaxeax < 0x80000006)
3938 		return;
3939 	cp = &cpi->cpi_extd[6];
3940 
3941 	if ((i = BITX(cp->cp_ecx, 15, 12)) != 0 &&
3942 	    (size = BITX(cp->cp_ecx, 31, 16)) != 0) {
3943 		uint_t cachesz = size * 1024;
3944 		assoc = amd_afd[i];
3945 
3946 		ASSERT(assoc != -1);
3947 
3948 		if ((ip = l2i->l2i_csz) != NULL)
3949 			*ip = cachesz;
3950 		if ((ip = l2i->l2i_lsz) != NULL)
3951 			*ip = BITX(cp->cp_ecx, 7, 0);
3952 		if ((ip = l2i->l2i_assoc) != NULL)
3953 			*ip = assoc;
3954 		l2i->l2i_ret = cachesz;
3955 	}
3956 }
3957 
3958 int
3959 getl2cacheinfo(cpu_t *cpu, int *csz, int *lsz, int *assoc)
3960 {
3961 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
3962 	struct l2info __l2info, *l2i = &__l2info;
3963 
3964 	l2i->l2i_csz = csz;
3965 	l2i->l2i_lsz = lsz;
3966 	l2i->l2i_assoc = assoc;
3967 	l2i->l2i_ret = -1;
3968 
3969 	switch (x86_which_cacheinfo(cpi)) {
3970 	case X86_VENDOR_Intel:
3971 		intel_walk_cacheinfo(cpi, l2i, intel_l2cinfo);
3972 		break;
3973 	case X86_VENDOR_Cyrix:
3974 		cyrix_walk_cacheinfo(cpi, l2i, intel_l2cinfo);
3975 		break;
3976 	case X86_VENDOR_AMD:
3977 		amd_l2cacheinfo(cpi, l2i);
3978 		break;
3979 	default:
3980 		break;
3981 	}
3982 	return (l2i->l2i_ret);
3983 }
3984 
3985 #if !defined(__xpv)
3986 
3987 uint32_t *
3988 cpuid_mwait_alloc(cpu_t *cpu)
3989 {
3990 	uint32_t	*ret;
3991 	size_t		mwait_size;
3992 
3993 	ASSERT(cpuid_checkpass(CPU, 2));
3994 
3995 	mwait_size = CPU->cpu_m.mcpu_cpi->cpi_mwait.mon_max;
3996 	if (mwait_size == 0)
3997 		return (NULL);
3998 
3999 	/*
4000 	 * kmem_alloc() returns cache line size aligned data for mwait_size
4001 	 * allocations.  mwait_size is currently cache line sized.  Neither
4002 	 * of these implementation details are guarantied to be true in the
4003 	 * future.
4004 	 *
4005 	 * First try allocating mwait_size as kmem_alloc() currently returns
4006 	 * correctly aligned memory.  If kmem_alloc() does not return
4007 	 * mwait_size aligned memory, then use mwait_size ROUNDUP.
4008 	 *
4009 	 * Set cpi_mwait.buf_actual and cpi_mwait.size_actual in case we
4010 	 * decide to free this memory.
4011 	 */
4012 	ret = kmem_zalloc(mwait_size, KM_SLEEP);
4013 	if (ret == (uint32_t *)P2ROUNDUP((uintptr_t)ret, mwait_size)) {
4014 		cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual = ret;
4015 		cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual = mwait_size;
4016 		*ret = MWAIT_RUNNING;
4017 		return (ret);
4018 	} else {
4019 		kmem_free(ret, mwait_size);
4020 		ret = kmem_zalloc(mwait_size * 2, KM_SLEEP);
4021 		cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual = ret;
4022 		cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual = mwait_size * 2;
4023 		ret = (uint32_t *)P2ROUNDUP((uintptr_t)ret, mwait_size);
4024 		*ret = MWAIT_RUNNING;
4025 		return (ret);
4026 	}
4027 }
4028 
4029 void
4030 cpuid_mwait_free(cpu_t *cpu)
4031 {
4032 	if (cpu->cpu_m.mcpu_cpi == NULL) {
4033 		return;
4034 	}
4035 
4036 	if (cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual != NULL &&
4037 	    cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual > 0) {
4038 		kmem_free(cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual,
4039 		    cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual);
4040 	}
4041 
4042 	cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual = NULL;
4043 	cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual = 0;
4044 }
4045 
4046 void
4047 patch_tsc_read(int flag)
4048 {
4049 	size_t cnt;
4050 
4051 	switch (flag) {
4052 	case X86_NO_TSC:
4053 		cnt = &_no_rdtsc_end - &_no_rdtsc_start;
4054 		(void) memcpy((void *)tsc_read, (void *)&_no_rdtsc_start, cnt);
4055 		break;
4056 	case X86_HAVE_TSCP:
4057 		cnt = &_tscp_end - &_tscp_start;
4058 		(void) memcpy((void *)tsc_read, (void *)&_tscp_start, cnt);
4059 		break;
4060 	case X86_TSC_MFENCE:
4061 		cnt = &_tsc_mfence_end - &_tsc_mfence_start;
4062 		(void) memcpy((void *)tsc_read,
4063 		    (void *)&_tsc_mfence_start, cnt);
4064 		break;
4065 	case X86_TSC_LFENCE:
4066 		cnt = &_tsc_lfence_end - &_tsc_lfence_start;
4067 		(void) memcpy((void *)tsc_read,
4068 		    (void *)&_tsc_lfence_start, cnt);
4069 		break;
4070 	default:
4071 		break;
4072 	}
4073 }
4074 
4075 int
4076 cpuid_deep_cstates_supported(void)
4077 {
4078 	struct cpuid_info *cpi;
4079 	struct cpuid_regs regs;
4080 
4081 	ASSERT(cpuid_checkpass(CPU, 1));
4082 
4083 	cpi = CPU->cpu_m.mcpu_cpi;
4084 
4085 	if (!(x86_feature & X86_CPUID))
4086 		return (0);
4087 
4088 	switch (cpi->cpi_vendor) {
4089 	case X86_VENDOR_Intel:
4090 		if (cpi->cpi_xmaxeax < 0x80000007)
4091 			return (0);
4092 
4093 		/*
4094 		 * TSC run at a constant rate in all ACPI C-states?
4095 		 */
4096 		regs.cp_eax = 0x80000007;
4097 		(void) __cpuid_insn(&regs);
4098 		return (regs.cp_edx & CPUID_TSC_CSTATE_INVARIANCE);
4099 
4100 	default:
4101 		return (0);
4102 	}
4103 }
4104 
4105 #endif	/* !__xpv */
4106 
4107 void
4108 post_startup_cpu_fixups(void)
4109 {
4110 #ifndef __xpv
4111 	/*
4112 	 * Some AMD processors support C1E state. Entering this state will
4113 	 * cause the local APIC timer to stop, which we can't deal with at
4114 	 * this time.
4115 	 */
4116 	if (cpuid_getvendor(CPU) == X86_VENDOR_AMD) {
4117 		on_trap_data_t otd;
4118 		uint64_t reg;
4119 
4120 		if (!on_trap(&otd, OT_DATA_ACCESS)) {
4121 			reg = rdmsr(MSR_AMD_INT_PENDING_CMP_HALT);
4122 			/* Disable C1E state if it is enabled by BIOS */
4123 			if ((reg >> AMD_ACTONCMPHALT_SHIFT) &
4124 			    AMD_ACTONCMPHALT_MASK) {
4125 				reg &= ~(AMD_ACTONCMPHALT_MASK <<
4126 				    AMD_ACTONCMPHALT_SHIFT);
4127 				wrmsr(MSR_AMD_INT_PENDING_CMP_HALT, reg);
4128 			}
4129 		}
4130 		no_trap();
4131 	}
4132 #endif	/* !__xpv */
4133 }
4134 
4135 /*
4136  * Starting with the Westmere processor the local
4137  * APIC timer will continue running in all C-states,
4138  * including the deepest C-states.
4139  */
4140 int
4141 cpuid_arat_supported(void)
4142 {
4143 	struct cpuid_info *cpi;
4144 	struct cpuid_regs regs;
4145 
4146 	ASSERT(cpuid_checkpass(CPU, 1));
4147 	ASSERT(x86_feature & X86_CPUID);
4148 
4149 	cpi = CPU->cpu_m.mcpu_cpi;
4150 
4151 	switch (cpi->cpi_vendor) {
4152 	case X86_VENDOR_Intel:
4153 		/*
4154 		 * Always-running Local APIC Timer is
4155 		 * indicated by CPUID.6.EAX[2].
4156 		 */
4157 		if (cpi->cpi_maxeax >= 6) {
4158 			regs.cp_eax = 6;
4159 			(void) cpuid_insn(NULL, &regs);
4160 			return (regs.cp_eax & CPUID_CSTATE_ARAT);
4161 		} else {
4162 			return (0);
4163 		}
4164 	default:
4165 		return (0);
4166 	}
4167 }
4168 
4169 /*
4170  * Check support for Intel ENERGY_PERF_BIAS feature
4171  */
4172 int
4173 cpuid_iepb_supported(struct cpu *cp)
4174 {
4175 	struct cpuid_info *cpi = cp->cpu_m.mcpu_cpi;
4176 	struct cpuid_regs regs;
4177 
4178 	ASSERT(cpuid_checkpass(cp, 1));
4179 
4180 	if (!(x86_feature & X86_CPUID) || !(x86_feature & X86_MSR)) {
4181 		return (0);
4182 	}
4183 
4184 	/*
4185 	 * Intel ENERGY_PERF_BIAS MSR is indicated by
4186 	 * capability bit CPUID.6.ECX.3
4187 	 */
4188 	if ((cpi->cpi_vendor != X86_VENDOR_Intel) || (cpi->cpi_maxeax < 6))
4189 		return (0);
4190 
4191 	regs.cp_eax = 0x6;
4192 	(void) cpuid_insn(NULL, &regs);
4193 	return (regs.cp_ecx & CPUID_EPB_SUPPORT);
4194 }
4195 
4196 #if defined(__amd64) && !defined(__xpv)
4197 /*
4198  * Patch in versions of bcopy for high performance Intel Nhm processors
4199  * and later...
4200  */
4201 void
4202 patch_memops(uint_t vendor)
4203 {
4204 	size_t cnt, i;
4205 	caddr_t to, from;
4206 
4207 	if ((vendor == X86_VENDOR_Intel) && ((x86_feature & X86_SSE4_2) != 0)) {
4208 		cnt = &bcopy_patch_end - &bcopy_patch_start;
4209 		to = &bcopy_ck_size;
4210 		from = &bcopy_patch_start;
4211 		for (i = 0; i < cnt; i++) {
4212 			*to++ = *from++;
4213 		}
4214 	}
4215 }
4216 #endif  /* __amd64 && !__xpv */
4217 
4218 /*
4219  * This function finds the number of bits to represent the number of cores per
4220  * chip and the number of strands per core for the Intel platforms.
4221  * It re-uses the x2APIC cpuid code of the cpuid_pass2().
4222  */
4223 void
4224 cpuid_get_ext_topo(uint_t vendor, uint_t *core_nbits, uint_t *strand_nbits)
4225 {
4226 	struct cpuid_regs regs;
4227 	struct cpuid_regs *cp = &regs;
4228 
4229 	if (vendor != X86_VENDOR_Intel) {
4230 		return;
4231 	}
4232 
4233 	/* if the cpuid level is 0xB, extended topo is available. */
4234 	cp->cp_eax = 0;
4235 	if (__cpuid_insn(cp) >= 0xB) {
4236 
4237 		cp->cp_eax = 0xB;
4238 		cp->cp_edx = cp->cp_ebx = cp->cp_ecx = 0;
4239 		(void) __cpuid_insn(cp);
4240 
4241 		/*
4242 		 * Check CPUID.EAX=0BH, ECX=0H:EBX is non-zero, which
4243 		 * indicates that the extended topology enumeration leaf is
4244 		 * available.
4245 		 */
4246 		if (cp->cp_ebx) {
4247 			uint_t coreid_shift = 0;
4248 			uint_t chipid_shift = 0;
4249 			uint_t i;
4250 			uint_t level;
4251 
4252 			for (i = 0; i < CPI_FNB_ECX_MAX; i++) {
4253 				cp->cp_eax = 0xB;
4254 				cp->cp_ecx = i;
4255 
4256 				(void) __cpuid_insn(cp);
4257 				level = CPI_CPU_LEVEL_TYPE(cp);
4258 
4259 				if (level == 1) {
4260 					/*
4261 					 * Thread level processor topology
4262 					 * Number of bits shift right APIC ID
4263 					 * to get the coreid.
4264 					 */
4265 					coreid_shift = BITX(cp->cp_eax, 4, 0);
4266 				} else if (level == 2) {
4267 					/*
4268 					 * Core level processor topology
4269 					 * Number of bits shift right APIC ID
4270 					 * to get the chipid.
4271 					 */
4272 					chipid_shift = BITX(cp->cp_eax, 4, 0);
4273 				}
4274 			}
4275 
4276 			if (coreid_shift > 0 && chipid_shift > coreid_shift) {
4277 				*strand_nbits = coreid_shift;
4278 				*core_nbits = chipid_shift - coreid_shift;
4279 			}
4280 		}
4281 	}
4282 }
4283