xref: /titanic_51/usr/src/uts/i86pc/os/cpuid.c (revision 6bbe05905a1c10a2703f95fb4912eb14b87f6670)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 /*
26  * Copyright (c) 2009, Intel Corporation.
27  * All rights reserved.
28  */
29 /*
30  * Portions Copyright 2009 Advanced Micro Devices, Inc.
31  */
32 
33 /*
34  * Various routines to handle identification
35  * and classification of x86 processors.
36  */
37 
38 #include <sys/types.h>
39 #include <sys/archsystm.h>
40 #include <sys/x86_archext.h>
41 #include <sys/kmem.h>
42 #include <sys/systm.h>
43 #include <sys/cmn_err.h>
44 #include <sys/sunddi.h>
45 #include <sys/sunndi.h>
46 #include <sys/cpuvar.h>
47 #include <sys/processor.h>
48 #include <sys/sysmacros.h>
49 #include <sys/pg.h>
50 #include <sys/fp.h>
51 #include <sys/controlregs.h>
52 #include <sys/auxv_386.h>
53 #include <sys/bitmap.h>
54 #include <sys/memnode.h>
55 #include <sys/pci_cfgspace.h>
56 
57 #ifdef __xpv
58 #include <sys/hypervisor.h>
59 #else
60 #include <sys/ontrap.h>
61 #endif
62 
63 /*
64  * Pass 0 of cpuid feature analysis happens in locore. It contains special code
65  * to recognize Cyrix processors that are not cpuid-compliant, and to deal with
66  * them accordingly. For most modern processors, feature detection occurs here
67  * in pass 1.
68  *
69  * Pass 1 of cpuid feature analysis happens just at the beginning of mlsetup()
70  * for the boot CPU and does the basic analysis that the early kernel needs.
71  * x86_feature is set based on the return value of cpuid_pass1() of the boot
72  * CPU.
73  *
74  * Pass 1 includes:
75  *
76  *	o Determining vendor/model/family/stepping and setting x86_type and
77  *	  x86_vendor accordingly.
78  *	o Processing the feature flags returned by the cpuid instruction while
79  *	  applying any workarounds or tricks for the specific processor.
80  *	o Mapping the feature flags into Solaris feature bits (X86_*).
81  *	o Processing extended feature flags if supported by the processor,
82  *	  again while applying specific processor knowledge.
83  *	o Determining the CMT characteristics of the system.
84  *
85  * Pass 1 is done on non-boot CPUs during their initialization and the results
86  * are used only as a meager attempt at ensuring that all processors within the
87  * system support the same features.
88  *
89  * Pass 2 of cpuid feature analysis happens just at the beginning
90  * of startup().  It just copies in and corrects the remainder
91  * of the cpuid data we depend on: standard cpuid functions that we didn't
92  * need for pass1 feature analysis, and extended cpuid functions beyond the
93  * simple feature processing done in pass1.
94  *
95  * Pass 3 of cpuid analysis is invoked after basic kernel services; in
96  * particular kernel memory allocation has been made available. It creates a
97  * readable brand string based on the data collected in the first two passes.
98  *
99  * Pass 4 of cpuid analysis is invoked after post_startup() when all
100  * the support infrastructure for various hardware features has been
101  * initialized. It determines which processor features will be reported
102  * to userland via the aux vector.
103  *
104  * All passes are executed on all CPUs, but only the boot CPU determines what
105  * features the kernel will use.
106  *
107  * Much of the worst junk in this file is for the support of processors
108  * that didn't really implement the cpuid instruction properly.
109  *
110  * NOTE: The accessor functions (cpuid_get*) are aware of, and ASSERT upon,
111  * the pass numbers.  Accordingly, changes to the pass code may require changes
112  * to the accessor code.
113  */
114 
115 uint_t x86_feature = 0;
116 uint_t x86_vendor = X86_VENDOR_IntelClone;
117 uint_t x86_type = X86_TYPE_OTHER;
118 uint_t x86_clflush_size = 0;
119 
120 uint_t pentiumpro_bug4046376;
121 uint_t pentiumpro_bug4064495;
122 
123 uint_t enable486;
124 /*
125  * This is set to platform type Solaris is running on.
126  */
127 static int platform_type = -1;
128 
129 #if !defined(__xpv)
130 /*
131  * Variable to patch if hypervisor platform detection needs to be
132  * disabled (e.g. platform_type will always be HW_NATIVE if this is 0).
133  */
134 int enable_platform_detection = 1;
135 #endif
136 
137 /*
138  * monitor/mwait info.
139  *
140  * size_actual and buf_actual are the real address and size allocated to get
141  * proper mwait_buf alignement.  buf_actual and size_actual should be passed
142  * to kmem_free().  Currently kmem_alloc() and mwait happen to both use
143  * processor cache-line alignment, but this is not guarantied in the furture.
144  */
145 struct mwait_info {
146 	size_t		mon_min;	/* min size to avoid missed wakeups */
147 	size_t		mon_max;	/* size to avoid false wakeups */
148 	size_t		size_actual;	/* size actually allocated */
149 	void		*buf_actual;	/* memory actually allocated */
150 	uint32_t	support;	/* processor support of monitor/mwait */
151 };
152 
153 /*
154  * These constants determine how many of the elements of the
155  * cpuid we cache in the cpuid_info data structure; the
156  * remaining elements are accessible via the cpuid instruction.
157  */
158 
159 #define	NMAX_CPI_STD	6		/* eax = 0 .. 5 */
160 #define	NMAX_CPI_EXTD	0x1c		/* eax = 0x80000000 .. 0x8000001b */
161 
162 /*
163  * Some terminology needs to be explained:
164  *  - Socket: Something that can be plugged into a motherboard.
165  *  - Package: Same as socket
166  *  - Chip: Same as socket. Note that AMD's documentation uses term "chip"
167  *    differently: there, chip is the same as processor node (below)
168  *  - Processor node: Some AMD processors have more than one
169  *    "subprocessor" embedded in a package. These subprocessors (nodes)
170  *    are fully-functional processors themselves with cores, caches,
171  *    memory controllers, PCI configuration spaces. They are connected
172  *    inside the package with Hypertransport links. On single-node
173  *    processors, processor node is equivalent to chip/socket/package.
174  */
175 
176 struct cpuid_info {
177 	uint_t cpi_pass;		/* last pass completed */
178 	/*
179 	 * standard function information
180 	 */
181 	uint_t cpi_maxeax;		/* fn 0: %eax */
182 	char cpi_vendorstr[13];		/* fn 0: %ebx:%ecx:%edx */
183 	uint_t cpi_vendor;		/* enum of cpi_vendorstr */
184 
185 	uint_t cpi_family;		/* fn 1: extended family */
186 	uint_t cpi_model;		/* fn 1: extended model */
187 	uint_t cpi_step;		/* fn 1: stepping */
188 	chipid_t cpi_chipid;		/* fn 1: %ebx:  Intel: chip # */
189 					/*		AMD: package/socket # */
190 	uint_t cpi_brandid;		/* fn 1: %ebx: brand ID */
191 	int cpi_clogid;			/* fn 1: %ebx: thread # */
192 	uint_t cpi_ncpu_per_chip;	/* fn 1: %ebx: logical cpu count */
193 	uint8_t cpi_cacheinfo[16];	/* fn 2: intel-style cache desc */
194 	uint_t cpi_ncache;		/* fn 2: number of elements */
195 	uint_t cpi_ncpu_shr_last_cache;	/* fn 4: %eax: ncpus sharing cache */
196 	id_t cpi_last_lvl_cacheid;	/* fn 4: %eax: derived cache id */
197 	uint_t cpi_std_4_size;		/* fn 4: number of fn 4 elements */
198 	struct cpuid_regs **cpi_std_4;	/* fn 4: %ecx == 0 .. fn4_size */
199 	struct cpuid_regs cpi_std[NMAX_CPI_STD];	/* 0 .. 5 */
200 	/*
201 	 * extended function information
202 	 */
203 	uint_t cpi_xmaxeax;		/* fn 0x80000000: %eax */
204 	char cpi_brandstr[49];		/* fn 0x8000000[234] */
205 	uint8_t cpi_pabits;		/* fn 0x80000006: %eax */
206 	uint8_t	cpi_vabits;		/* fn 0x80000006: %eax */
207 	struct	cpuid_regs cpi_extd[NMAX_CPI_EXTD];	/* 0x800000XX */
208 
209 	id_t cpi_coreid;		/* same coreid => strands share core */
210 	int cpi_pkgcoreid;		/* core number within single package */
211 	uint_t cpi_ncore_per_chip;	/* AMD: fn 0x80000008: %ecx[7-0] */
212 					/* Intel: fn 4: %eax[31-26] */
213 	/*
214 	 * supported feature information
215 	 */
216 	uint32_t cpi_support[5];
217 #define	STD_EDX_FEATURES	0
218 #define	AMD_EDX_FEATURES	1
219 #define	TM_EDX_FEATURES		2
220 #define	STD_ECX_FEATURES	3
221 #define	AMD_ECX_FEATURES	4
222 	/*
223 	 * Synthesized information, where known.
224 	 */
225 	uint32_t cpi_chiprev;		/* See X86_CHIPREV_* in x86_archext.h */
226 	const char *cpi_chiprevstr;	/* May be NULL if chiprev unknown */
227 	uint32_t cpi_socket;		/* Chip package/socket type */
228 
229 	struct mwait_info cpi_mwait;	/* fn 5: monitor/mwait info */
230 	uint32_t cpi_apicid;
231 	uint_t cpi_procnodeid;		/* AMD: nodeID on HT, Intel: chipid */
232 	uint_t cpi_procnodes_per_pkg;	/* AMD: # of nodes in the package */
233 					/* Intel: 1 */
234 };
235 
236 
237 static struct cpuid_info cpuid_info0;
238 
239 /*
240  * These bit fields are defined by the Intel Application Note AP-485
241  * "Intel Processor Identification and the CPUID Instruction"
242  */
243 #define	CPI_FAMILY_XTD(cpi)	BITX((cpi)->cpi_std[1].cp_eax, 27, 20)
244 #define	CPI_MODEL_XTD(cpi)	BITX((cpi)->cpi_std[1].cp_eax, 19, 16)
245 #define	CPI_TYPE(cpi)		BITX((cpi)->cpi_std[1].cp_eax, 13, 12)
246 #define	CPI_FAMILY(cpi)		BITX((cpi)->cpi_std[1].cp_eax, 11, 8)
247 #define	CPI_STEP(cpi)		BITX((cpi)->cpi_std[1].cp_eax, 3, 0)
248 #define	CPI_MODEL(cpi)		BITX((cpi)->cpi_std[1].cp_eax, 7, 4)
249 
250 #define	CPI_FEATURES_EDX(cpi)		((cpi)->cpi_std[1].cp_edx)
251 #define	CPI_FEATURES_ECX(cpi)		((cpi)->cpi_std[1].cp_ecx)
252 #define	CPI_FEATURES_XTD_EDX(cpi)	((cpi)->cpi_extd[1].cp_edx)
253 #define	CPI_FEATURES_XTD_ECX(cpi)	((cpi)->cpi_extd[1].cp_ecx)
254 
255 #define	CPI_BRANDID(cpi)	BITX((cpi)->cpi_std[1].cp_ebx, 7, 0)
256 #define	CPI_CHUNKS(cpi)		BITX((cpi)->cpi_std[1].cp_ebx, 15, 7)
257 #define	CPI_CPU_COUNT(cpi)	BITX((cpi)->cpi_std[1].cp_ebx, 23, 16)
258 #define	CPI_APIC_ID(cpi)	BITX((cpi)->cpi_std[1].cp_ebx, 31, 24)
259 
260 #define	CPI_MAXEAX_MAX		0x100		/* sanity control */
261 #define	CPI_XMAXEAX_MAX		0x80000100
262 #define	CPI_FN4_ECX_MAX		0x20		/* sanity: max fn 4 levels */
263 #define	CPI_FNB_ECX_MAX		0x20		/* sanity: max fn B levels */
264 
265 /*
266  * Function 4 (Deterministic Cache Parameters) macros
267  * Defined by Intel Application Note AP-485
268  */
269 #define	CPI_NUM_CORES(regs)		BITX((regs)->cp_eax, 31, 26)
270 #define	CPI_NTHR_SHR_CACHE(regs)	BITX((regs)->cp_eax, 25, 14)
271 #define	CPI_FULL_ASSOC_CACHE(regs)	BITX((regs)->cp_eax, 9, 9)
272 #define	CPI_SELF_INIT_CACHE(regs)	BITX((regs)->cp_eax, 8, 8)
273 #define	CPI_CACHE_LVL(regs)		BITX((regs)->cp_eax, 7, 5)
274 #define	CPI_CACHE_TYPE(regs)		BITX((regs)->cp_eax, 4, 0)
275 #define	CPI_CPU_LEVEL_TYPE(regs)	BITX((regs)->cp_ecx, 15, 8)
276 
277 #define	CPI_CACHE_WAYS(regs)		BITX((regs)->cp_ebx, 31, 22)
278 #define	CPI_CACHE_PARTS(regs)		BITX((regs)->cp_ebx, 21, 12)
279 #define	CPI_CACHE_COH_LN_SZ(regs)	BITX((regs)->cp_ebx, 11, 0)
280 
281 #define	CPI_CACHE_SETS(regs)		BITX((regs)->cp_ecx, 31, 0)
282 
283 #define	CPI_PREFCH_STRIDE(regs)		BITX((regs)->cp_edx, 9, 0)
284 
285 
286 /*
287  * A couple of shorthand macros to identify "later" P6-family chips
288  * like the Pentium M and Core.  First, the "older" P6-based stuff
289  * (loosely defined as "pre-Pentium-4"):
290  * P6, PII, Mobile PII, PII Xeon, PIII, Mobile PIII, PIII Xeon
291  */
292 
293 #define	IS_LEGACY_P6(cpi) (			\
294 	cpi->cpi_family == 6 && 		\
295 		(cpi->cpi_model == 1 ||		\
296 		cpi->cpi_model == 3 ||		\
297 		cpi->cpi_model == 5 ||		\
298 		cpi->cpi_model == 6 ||		\
299 		cpi->cpi_model == 7 ||		\
300 		cpi->cpi_model == 8 ||		\
301 		cpi->cpi_model == 0xA ||	\
302 		cpi->cpi_model == 0xB)		\
303 )
304 
305 /* A "new F6" is everything with family 6 that's not the above */
306 #define	IS_NEW_F6(cpi) ((cpi->cpi_family == 6) && !IS_LEGACY_P6(cpi))
307 
308 /* Extended family/model support */
309 #define	IS_EXTENDED_MODEL_INTEL(cpi) (cpi->cpi_family == 0x6 || \
310 	cpi->cpi_family >= 0xf)
311 
312 /*
313  * Info for monitor/mwait idle loop.
314  *
315  * See cpuid section of "Intel 64 and IA-32 Architectures Software Developer's
316  * Manual Volume 2A: Instruction Set Reference, A-M" #25366-022US, November
317  * 2006.
318  * See MONITOR/MWAIT section of "AMD64 Architecture Programmer's Manual
319  * Documentation Updates" #33633, Rev 2.05, December 2006.
320  */
321 #define	MWAIT_SUPPORT		(0x00000001)	/* mwait supported */
322 #define	MWAIT_EXTENSIONS	(0x00000002)	/* extenstion supported */
323 #define	MWAIT_ECX_INT_ENABLE	(0x00000004)	/* ecx 1 extension supported */
324 #define	MWAIT_SUPPORTED(cpi)	((cpi)->cpi_std[1].cp_ecx & CPUID_INTC_ECX_MON)
325 #define	MWAIT_INT_ENABLE(cpi)	((cpi)->cpi_std[5].cp_ecx & 0x2)
326 #define	MWAIT_EXTENSION(cpi)	((cpi)->cpi_std[5].cp_ecx & 0x1)
327 #define	MWAIT_SIZE_MIN(cpi)	BITX((cpi)->cpi_std[5].cp_eax, 15, 0)
328 #define	MWAIT_SIZE_MAX(cpi)	BITX((cpi)->cpi_std[5].cp_ebx, 15, 0)
329 /*
330  * Number of sub-cstates for a given c-state.
331  */
332 #define	MWAIT_NUM_SUBC_STATES(cpi, c_state)			\
333 	BITX((cpi)->cpi_std[5].cp_edx, c_state + 3, c_state)
334 
335 /*
336  * Functions we consune from cpuid_subr.c;  don't publish these in a header
337  * file to try and keep people using the expected cpuid_* interfaces.
338  */
339 extern uint32_t _cpuid_skt(uint_t, uint_t, uint_t, uint_t);
340 extern const char *_cpuid_sktstr(uint_t, uint_t, uint_t, uint_t);
341 extern uint32_t _cpuid_chiprev(uint_t, uint_t, uint_t, uint_t);
342 extern const char *_cpuid_chiprevstr(uint_t, uint_t, uint_t, uint_t);
343 extern uint_t _cpuid_vendorstr_to_vendorcode(char *);
344 
345 /*
346  * Apply up various platform-dependent restrictions where the
347  * underlying platform restrictions mean the CPU can be marked
348  * as less capable than its cpuid instruction would imply.
349  */
350 #if defined(__xpv)
351 static void
352 platform_cpuid_mangle(uint_t vendor, uint32_t eax, struct cpuid_regs *cp)
353 {
354 	switch (eax) {
355 	case 1: {
356 		uint32_t mcamask = DOMAIN_IS_INITDOMAIN(xen_info) ?
357 		    0 : CPUID_INTC_EDX_MCA;
358 		cp->cp_edx &=
359 		    ~(mcamask |
360 		    CPUID_INTC_EDX_PSE |
361 		    CPUID_INTC_EDX_VME | CPUID_INTC_EDX_DE |
362 		    CPUID_INTC_EDX_SEP | CPUID_INTC_EDX_MTRR |
363 		    CPUID_INTC_EDX_PGE | CPUID_INTC_EDX_PAT |
364 		    CPUID_AMD_EDX_SYSC | CPUID_INTC_EDX_SEP |
365 		    CPUID_INTC_EDX_PSE36 | CPUID_INTC_EDX_HTT);
366 		break;
367 	}
368 
369 	case 0x80000001:
370 		cp->cp_edx &=
371 		    ~(CPUID_AMD_EDX_PSE |
372 		    CPUID_INTC_EDX_VME | CPUID_INTC_EDX_DE |
373 		    CPUID_AMD_EDX_MTRR | CPUID_AMD_EDX_PGE |
374 		    CPUID_AMD_EDX_PAT | CPUID_AMD_EDX_PSE36 |
375 		    CPUID_AMD_EDX_SYSC | CPUID_INTC_EDX_SEP |
376 		    CPUID_AMD_EDX_TSCP);
377 		cp->cp_ecx &= ~CPUID_AMD_ECX_CMP_LGCY;
378 		break;
379 	default:
380 		break;
381 	}
382 
383 	switch (vendor) {
384 	case X86_VENDOR_Intel:
385 		switch (eax) {
386 		case 4:
387 			/*
388 			 * Zero out the (ncores-per-chip - 1) field
389 			 */
390 			cp->cp_eax &= 0x03fffffff;
391 			break;
392 		default:
393 			break;
394 		}
395 		break;
396 	case X86_VENDOR_AMD:
397 		switch (eax) {
398 
399 		case 0x80000001:
400 			cp->cp_ecx &= ~CPUID_AMD_ECX_CR8D;
401 			break;
402 
403 		case 0x80000008:
404 			/*
405 			 * Zero out the (ncores-per-chip - 1) field
406 			 */
407 			cp->cp_ecx &= 0xffffff00;
408 			break;
409 		default:
410 			break;
411 		}
412 		break;
413 	default:
414 		break;
415 	}
416 }
417 #else
418 #define	platform_cpuid_mangle(vendor, eax, cp)	/* nothing */
419 #endif
420 
421 /*
422  *  Some undocumented ways of patching the results of the cpuid
423  *  instruction to permit running Solaris 10 on future cpus that
424  *  we don't currently support.  Could be set to non-zero values
425  *  via settings in eeprom.
426  */
427 
428 uint32_t cpuid_feature_ecx_include;
429 uint32_t cpuid_feature_ecx_exclude;
430 uint32_t cpuid_feature_edx_include;
431 uint32_t cpuid_feature_edx_exclude;
432 
433 void
434 cpuid_alloc_space(cpu_t *cpu)
435 {
436 	/*
437 	 * By convention, cpu0 is the boot cpu, which is set up
438 	 * before memory allocation is available.  All other cpus get
439 	 * their cpuid_info struct allocated here.
440 	 */
441 	ASSERT(cpu->cpu_id != 0);
442 	cpu->cpu_m.mcpu_cpi =
443 	    kmem_zalloc(sizeof (*cpu->cpu_m.mcpu_cpi), KM_SLEEP);
444 }
445 
446 void
447 cpuid_free_space(cpu_t *cpu)
448 {
449 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
450 	int i;
451 
452 	ASSERT(cpu->cpu_id != 0);
453 
454 	/*
455 	 * Free up any function 4 related dynamic storage
456 	 */
457 	for (i = 1; i < cpi->cpi_std_4_size; i++)
458 		kmem_free(cpi->cpi_std_4[i], sizeof (struct cpuid_regs));
459 	if (cpi->cpi_std_4_size > 0)
460 		kmem_free(cpi->cpi_std_4,
461 		    cpi->cpi_std_4_size * sizeof (struct cpuid_regs *));
462 
463 	kmem_free(cpu->cpu_m.mcpu_cpi, sizeof (*cpu->cpu_m.mcpu_cpi));
464 }
465 
466 #if !defined(__xpv)
467 
468 static void
469 determine_platform()
470 {
471 	struct cpuid_regs cp;
472 	char *xen_str;
473 	uint32_t xen_signature[4];
474 
475 	platform_type = HW_NATIVE;
476 
477 	if (!enable_platform_detection)
478 		return;
479 
480 	/*
481 	 * In a fully virtualized domain, Xen's pseudo-cpuid function
482 	 * 0x40000000 returns a string representing the Xen signature in
483 	 * %ebx, %ecx, and %edx.  %eax contains the maximum supported cpuid
484 	 * function.
485 	 */
486 	cp.cp_eax = 0x40000000;
487 	(void) __cpuid_insn(&cp);
488 	xen_signature[0] = cp.cp_ebx;
489 	xen_signature[1] = cp.cp_ecx;
490 	xen_signature[2] = cp.cp_edx;
491 	xen_signature[3] = 0;
492 	xen_str = (char *)xen_signature;
493 	if (strcmp("XenVMMXenVMM", xen_str) == 0 && cp.cp_eax <= 0x40000002) {
494 		platform_type = HW_XEN_HVM;
495 	} else if (vmware_platform()) { /* running under vmware hypervisor? */
496 		platform_type = HW_VMWARE;
497 	}
498 }
499 
500 int
501 get_hwenv(void)
502 {
503 	if (platform_type == -1)
504 		determine_platform();
505 
506 	return (platform_type);
507 }
508 
509 int
510 is_controldom(void)
511 {
512 	return (0);
513 }
514 
515 #else
516 
517 int
518 get_hwenv(void)
519 {
520 	return (HW_XEN_PV);
521 }
522 
523 int
524 is_controldom(void)
525 {
526 	return (DOMAIN_IS_INITDOMAIN(xen_info));
527 }
528 
529 #endif	/* __xpv */
530 
531 static void
532 cpuid_intel_getids(cpu_t *cpu, uint_t feature)
533 {
534 	uint_t i;
535 	uint_t chipid_shift = 0;
536 	uint_t coreid_shift = 0;
537 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
538 
539 	for (i = 1; i < cpi->cpi_ncpu_per_chip; i <<= 1)
540 		chipid_shift++;
541 
542 	cpi->cpi_chipid = cpi->cpi_apicid >> chipid_shift;
543 	cpi->cpi_clogid = cpi->cpi_apicid & ((1 << chipid_shift) - 1);
544 
545 	if (feature & X86_CMP) {
546 		/*
547 		 * Multi-core (and possibly multi-threaded)
548 		 * processors.
549 		 */
550 		uint_t ncpu_per_core;
551 		if (cpi->cpi_ncore_per_chip == 1)
552 			ncpu_per_core = cpi->cpi_ncpu_per_chip;
553 		else if (cpi->cpi_ncore_per_chip > 1)
554 			ncpu_per_core = cpi->cpi_ncpu_per_chip /
555 			    cpi->cpi_ncore_per_chip;
556 		/*
557 		 * 8bit APIC IDs on dual core Pentiums
558 		 * look like this:
559 		 *
560 		 * +-----------------------+------+------+
561 		 * | Physical Package ID   |  MC  |  HT  |
562 		 * +-----------------------+------+------+
563 		 * <------- chipid -------->
564 		 * <------- coreid --------------->
565 		 *			   <--- clogid -->
566 		 *			   <------>
567 		 *			   pkgcoreid
568 		 *
569 		 * Where the number of bits necessary to
570 		 * represent MC and HT fields together equals
571 		 * to the minimum number of bits necessary to
572 		 * store the value of cpi->cpi_ncpu_per_chip.
573 		 * Of those bits, the MC part uses the number
574 		 * of bits necessary to store the value of
575 		 * cpi->cpi_ncore_per_chip.
576 		 */
577 		for (i = 1; i < ncpu_per_core; i <<= 1)
578 			coreid_shift++;
579 		cpi->cpi_coreid = cpi->cpi_apicid >> coreid_shift;
580 		cpi->cpi_pkgcoreid = cpi->cpi_clogid >> coreid_shift;
581 	} else if (feature & X86_HTT) {
582 		/*
583 		 * Single-core multi-threaded processors.
584 		 */
585 		cpi->cpi_coreid = cpi->cpi_chipid;
586 		cpi->cpi_pkgcoreid = 0;
587 	}
588 	cpi->cpi_procnodeid = cpi->cpi_chipid;
589 }
590 
591 static void
592 cpuid_amd_getids(cpu_t *cpu)
593 {
594 	int i, first_half, coreidsz;
595 	uint32_t nb_caps_reg;
596 	uint_t node2_1;
597 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
598 
599 	/*
600 	 * AMD CMP chips currently have a single thread per core.
601 	 *
602 	 * Since no two cpus share a core we must assign a distinct coreid
603 	 * per cpu, and we do this by using the cpu_id.  This scheme does not,
604 	 * however, guarantee that sibling cores of a chip will have sequential
605 	 * coreids starting at a multiple of the number of cores per chip -
606 	 * that is usually the case, but if the ACPI MADT table is presented
607 	 * in a different order then we need to perform a few more gymnastics
608 	 * for the pkgcoreid.
609 	 *
610 	 * All processors in the system have the same number of enabled
611 	 * cores. Cores within a processor are always numbered sequentially
612 	 * from 0 regardless of how many or which are disabled, and there
613 	 * is no way for operating system to discover the real core id when some
614 	 * are disabled.
615 	 */
616 
617 	cpi->cpi_coreid = cpu->cpu_id;
618 
619 	if (cpi->cpi_xmaxeax >= 0x80000008) {
620 
621 		coreidsz = BITX((cpi)->cpi_extd[8].cp_ecx, 15, 12);
622 
623 		/*
624 		 * In AMD parlance chip is really a node while Solaris
625 		 * sees chip as equivalent to socket/package.
626 		 */
627 		cpi->cpi_ncore_per_chip =
628 		    BITX((cpi)->cpi_extd[8].cp_ecx, 7, 0) + 1;
629 		if (coreidsz == 0) {
630 			/* Use legacy method */
631 			for (i = 1; i < cpi->cpi_ncore_per_chip; i <<= 1)
632 				coreidsz++;
633 			if (coreidsz == 0)
634 				coreidsz = 1;
635 		}
636 	} else {
637 		/* Assume single-core part */
638 		cpi->cpi_ncore_per_chip = 1;
639 	}
640 
641 	cpi->cpi_clogid = cpi->cpi_pkgcoreid =
642 	    cpi->cpi_apicid & ((1<<coreidsz) - 1);
643 	cpi->cpi_ncpu_per_chip = cpi->cpi_ncore_per_chip;
644 
645 	/* Get nodeID */
646 	if (cpi->cpi_family == 0xf) {
647 		cpi->cpi_procnodeid = (cpi->cpi_apicid >> coreidsz) & 7;
648 		cpi->cpi_chipid = cpi->cpi_procnodeid;
649 	} else if (cpi->cpi_family == 0x10) {
650 		/*
651 		 * See if we are a multi-node processor.
652 		 * All processors in the system have the same number of nodes
653 		 */
654 		nb_caps_reg =  pci_getl_func(0, 24, 3, 0xe8);
655 		if ((cpi->cpi_model < 8) || BITX(nb_caps_reg, 29, 29) == 0) {
656 			/* Single-node */
657 			cpi->cpi_procnodeid = BITX(cpi->cpi_apicid, 5,
658 			    coreidsz);
659 			cpi->cpi_chipid = cpi->cpi_procnodeid;
660 		} else {
661 
662 			/*
663 			 * Multi-node revision D (2 nodes per package
664 			 * are supported)
665 			 */
666 			cpi->cpi_procnodes_per_pkg = 2;
667 
668 			first_half = (cpi->cpi_pkgcoreid <=
669 			    (cpi->cpi_ncore_per_chip/2 - 1));
670 
671 			if (cpi->cpi_apicid == cpi->cpi_pkgcoreid) {
672 				/* We are BSP */
673 				cpi->cpi_procnodeid = (first_half ? 0 : 1);
674 				cpi->cpi_chipid = cpi->cpi_procnodeid >> 1;
675 			} else {
676 
677 				/* We are AP */
678 				/* NodeId[2:1] bits to use for reading F3xe8 */
679 				node2_1 = BITX(cpi->cpi_apicid, 5, 4) << 1;
680 
681 				nb_caps_reg =
682 				    pci_getl_func(0, 24 + node2_1, 3, 0xe8);
683 
684 				/*
685 				 * Check IntNodeNum bit (31:30, but bit 31 is
686 				 * always 0 on dual-node processors)
687 				 */
688 				if (BITX(nb_caps_reg, 30, 30) == 0)
689 					cpi->cpi_procnodeid = node2_1 +
690 					    !first_half;
691 				else
692 					cpi->cpi_procnodeid = node2_1 +
693 					    first_half;
694 
695 				cpi->cpi_chipid = cpi->cpi_procnodeid >> 1;
696 			}
697 		}
698 	} else if (cpi->cpi_family >= 0x11) {
699 		cpi->cpi_procnodeid = (cpi->cpi_apicid >> coreidsz) & 7;
700 		cpi->cpi_chipid = cpi->cpi_procnodeid;
701 	} else {
702 		cpi->cpi_procnodeid = 0;
703 		cpi->cpi_chipid = cpi->cpi_procnodeid;
704 	}
705 }
706 
707 uint_t
708 cpuid_pass1(cpu_t *cpu)
709 {
710 	uint32_t mask_ecx, mask_edx;
711 	uint_t feature = X86_CPUID;
712 	struct cpuid_info *cpi;
713 	struct cpuid_regs *cp;
714 	int xcpuid;
715 #if !defined(__xpv)
716 	extern int idle_cpu_prefer_mwait;
717 #endif
718 
719 
720 #if !defined(__xpv)
721 	determine_platform();
722 #endif
723 	/*
724 	 * Space statically allocated for cpu0, ensure pointer is set
725 	 */
726 	if (cpu->cpu_id == 0)
727 		cpu->cpu_m.mcpu_cpi = &cpuid_info0;
728 	cpi = cpu->cpu_m.mcpu_cpi;
729 	ASSERT(cpi != NULL);
730 	cp = &cpi->cpi_std[0];
731 	cp->cp_eax = 0;
732 	cpi->cpi_maxeax = __cpuid_insn(cp);
733 	{
734 		uint32_t *iptr = (uint32_t *)cpi->cpi_vendorstr;
735 		*iptr++ = cp->cp_ebx;
736 		*iptr++ = cp->cp_edx;
737 		*iptr++ = cp->cp_ecx;
738 		*(char *)&cpi->cpi_vendorstr[12] = '\0';
739 	}
740 
741 	cpi->cpi_vendor = _cpuid_vendorstr_to_vendorcode(cpi->cpi_vendorstr);
742 	x86_vendor = cpi->cpi_vendor; /* for compatibility */
743 
744 	/*
745 	 * Limit the range in case of weird hardware
746 	 */
747 	if (cpi->cpi_maxeax > CPI_MAXEAX_MAX)
748 		cpi->cpi_maxeax = CPI_MAXEAX_MAX;
749 	if (cpi->cpi_maxeax < 1)
750 		goto pass1_done;
751 
752 	cp = &cpi->cpi_std[1];
753 	cp->cp_eax = 1;
754 	(void) __cpuid_insn(cp);
755 
756 	/*
757 	 * Extract identifying constants for easy access.
758 	 */
759 	cpi->cpi_model = CPI_MODEL(cpi);
760 	cpi->cpi_family = CPI_FAMILY(cpi);
761 
762 	if (cpi->cpi_family == 0xf)
763 		cpi->cpi_family += CPI_FAMILY_XTD(cpi);
764 
765 	/*
766 	 * Beware: AMD uses "extended model" iff base *FAMILY* == 0xf.
767 	 * Intel, and presumably everyone else, uses model == 0xf, as
768 	 * one would expect (max value means possible overflow).  Sigh.
769 	 */
770 
771 	switch (cpi->cpi_vendor) {
772 	case X86_VENDOR_Intel:
773 		if (IS_EXTENDED_MODEL_INTEL(cpi))
774 			cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4;
775 		break;
776 	case X86_VENDOR_AMD:
777 		if (CPI_FAMILY(cpi) == 0xf)
778 			cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4;
779 		break;
780 	default:
781 		if (cpi->cpi_model == 0xf)
782 			cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4;
783 		break;
784 	}
785 
786 	cpi->cpi_step = CPI_STEP(cpi);
787 	cpi->cpi_brandid = CPI_BRANDID(cpi);
788 
789 	/*
790 	 * *default* assumptions:
791 	 * - believe %edx feature word
792 	 * - ignore %ecx feature word
793 	 * - 32-bit virtual and physical addressing
794 	 */
795 	mask_edx = 0xffffffff;
796 	mask_ecx = 0;
797 
798 	cpi->cpi_pabits = cpi->cpi_vabits = 32;
799 
800 	switch (cpi->cpi_vendor) {
801 	case X86_VENDOR_Intel:
802 		if (cpi->cpi_family == 5)
803 			x86_type = X86_TYPE_P5;
804 		else if (IS_LEGACY_P6(cpi)) {
805 			x86_type = X86_TYPE_P6;
806 			pentiumpro_bug4046376 = 1;
807 			pentiumpro_bug4064495 = 1;
808 			/*
809 			 * Clear the SEP bit when it was set erroneously
810 			 */
811 			if (cpi->cpi_model < 3 && cpi->cpi_step < 3)
812 				cp->cp_edx &= ~CPUID_INTC_EDX_SEP;
813 		} else if (IS_NEW_F6(cpi) || cpi->cpi_family == 0xf) {
814 			x86_type = X86_TYPE_P4;
815 			/*
816 			 * We don't currently depend on any of the %ecx
817 			 * features until Prescott, so we'll only check
818 			 * this from P4 onwards.  We might want to revisit
819 			 * that idea later.
820 			 */
821 			mask_ecx = 0xffffffff;
822 		} else if (cpi->cpi_family > 0xf)
823 			mask_ecx = 0xffffffff;
824 		/*
825 		 * We don't support MONITOR/MWAIT if leaf 5 is not available
826 		 * to obtain the monitor linesize.
827 		 */
828 		if (cpi->cpi_maxeax < 5)
829 			mask_ecx &= ~CPUID_INTC_ECX_MON;
830 		break;
831 	case X86_VENDOR_IntelClone:
832 	default:
833 		break;
834 	case X86_VENDOR_AMD:
835 #if defined(OPTERON_ERRATUM_108)
836 		if (cpi->cpi_family == 0xf && cpi->cpi_model == 0xe) {
837 			cp->cp_eax = (0xf0f & cp->cp_eax) | 0xc0;
838 			cpi->cpi_model = 0xc;
839 		} else
840 #endif
841 		if (cpi->cpi_family == 5) {
842 			/*
843 			 * AMD K5 and K6
844 			 *
845 			 * These CPUs have an incomplete implementation
846 			 * of MCA/MCE which we mask away.
847 			 */
848 			mask_edx &= ~(CPUID_INTC_EDX_MCE | CPUID_INTC_EDX_MCA);
849 
850 			/*
851 			 * Model 0 uses the wrong (APIC) bit
852 			 * to indicate PGE.  Fix it here.
853 			 */
854 			if (cpi->cpi_model == 0) {
855 				if (cp->cp_edx & 0x200) {
856 					cp->cp_edx &= ~0x200;
857 					cp->cp_edx |= CPUID_INTC_EDX_PGE;
858 				}
859 			}
860 
861 			/*
862 			 * Early models had problems w/ MMX; disable.
863 			 */
864 			if (cpi->cpi_model < 6)
865 				mask_edx &= ~CPUID_INTC_EDX_MMX;
866 		}
867 
868 		/*
869 		 * For newer families, SSE3 and CX16, at least, are valid;
870 		 * enable all
871 		 */
872 		if (cpi->cpi_family >= 0xf)
873 			mask_ecx = 0xffffffff;
874 		/*
875 		 * We don't support MONITOR/MWAIT if leaf 5 is not available
876 		 * to obtain the monitor linesize.
877 		 */
878 		if (cpi->cpi_maxeax < 5)
879 			mask_ecx &= ~CPUID_INTC_ECX_MON;
880 
881 #if !defined(__xpv)
882 		/*
883 		 * Do not use MONITOR/MWAIT to halt in the idle loop on any AMD
884 		 * processors.  AMD does not intend MWAIT to be used in the cpu
885 		 * idle loop on current and future processors.  10h and future
886 		 * AMD processors use more power in MWAIT than HLT.
887 		 * Pre-family-10h Opterons do not have the MWAIT instruction.
888 		 */
889 		idle_cpu_prefer_mwait = 0;
890 #endif
891 
892 		break;
893 	case X86_VENDOR_TM:
894 		/*
895 		 * workaround the NT workaround in CMS 4.1
896 		 */
897 		if (cpi->cpi_family == 5 && cpi->cpi_model == 4 &&
898 		    (cpi->cpi_step == 2 || cpi->cpi_step == 3))
899 			cp->cp_edx |= CPUID_INTC_EDX_CX8;
900 		break;
901 	case X86_VENDOR_Centaur:
902 		/*
903 		 * workaround the NT workarounds again
904 		 */
905 		if (cpi->cpi_family == 6)
906 			cp->cp_edx |= CPUID_INTC_EDX_CX8;
907 		break;
908 	case X86_VENDOR_Cyrix:
909 		/*
910 		 * We rely heavily on the probing in locore
911 		 * to actually figure out what parts, if any,
912 		 * of the Cyrix cpuid instruction to believe.
913 		 */
914 		switch (x86_type) {
915 		case X86_TYPE_CYRIX_486:
916 			mask_edx = 0;
917 			break;
918 		case X86_TYPE_CYRIX_6x86:
919 			mask_edx = 0;
920 			break;
921 		case X86_TYPE_CYRIX_6x86L:
922 			mask_edx =
923 			    CPUID_INTC_EDX_DE |
924 			    CPUID_INTC_EDX_CX8;
925 			break;
926 		case X86_TYPE_CYRIX_6x86MX:
927 			mask_edx =
928 			    CPUID_INTC_EDX_DE |
929 			    CPUID_INTC_EDX_MSR |
930 			    CPUID_INTC_EDX_CX8 |
931 			    CPUID_INTC_EDX_PGE |
932 			    CPUID_INTC_EDX_CMOV |
933 			    CPUID_INTC_EDX_MMX;
934 			break;
935 		case X86_TYPE_CYRIX_GXm:
936 			mask_edx =
937 			    CPUID_INTC_EDX_MSR |
938 			    CPUID_INTC_EDX_CX8 |
939 			    CPUID_INTC_EDX_CMOV |
940 			    CPUID_INTC_EDX_MMX;
941 			break;
942 		case X86_TYPE_CYRIX_MediaGX:
943 			break;
944 		case X86_TYPE_CYRIX_MII:
945 		case X86_TYPE_VIA_CYRIX_III:
946 			mask_edx =
947 			    CPUID_INTC_EDX_DE |
948 			    CPUID_INTC_EDX_TSC |
949 			    CPUID_INTC_EDX_MSR |
950 			    CPUID_INTC_EDX_CX8 |
951 			    CPUID_INTC_EDX_PGE |
952 			    CPUID_INTC_EDX_CMOV |
953 			    CPUID_INTC_EDX_MMX;
954 			break;
955 		default:
956 			break;
957 		}
958 		break;
959 	}
960 
961 #if defined(__xpv)
962 	/*
963 	 * Do not support MONITOR/MWAIT under a hypervisor
964 	 */
965 	mask_ecx &= ~CPUID_INTC_ECX_MON;
966 #endif	/* __xpv */
967 
968 	/*
969 	 * Now we've figured out the masks that determine
970 	 * which bits we choose to believe, apply the masks
971 	 * to the feature words, then map the kernel's view
972 	 * of these feature words into its feature word.
973 	 */
974 	cp->cp_edx &= mask_edx;
975 	cp->cp_ecx &= mask_ecx;
976 
977 	/*
978 	 * apply any platform restrictions (we don't call this
979 	 * immediately after __cpuid_insn here, because we need the
980 	 * workarounds applied above first)
981 	 */
982 	platform_cpuid_mangle(cpi->cpi_vendor, 1, cp);
983 
984 	/*
985 	 * fold in overrides from the "eeprom" mechanism
986 	 */
987 	cp->cp_edx |= cpuid_feature_edx_include;
988 	cp->cp_edx &= ~cpuid_feature_edx_exclude;
989 
990 	cp->cp_ecx |= cpuid_feature_ecx_include;
991 	cp->cp_ecx &= ~cpuid_feature_ecx_exclude;
992 
993 	if (cp->cp_edx & CPUID_INTC_EDX_PSE)
994 		feature |= X86_LARGEPAGE;
995 	if (cp->cp_edx & CPUID_INTC_EDX_TSC)
996 		feature |= X86_TSC;
997 	if (cp->cp_edx & CPUID_INTC_EDX_MSR)
998 		feature |= X86_MSR;
999 	if (cp->cp_edx & CPUID_INTC_EDX_MTRR)
1000 		feature |= X86_MTRR;
1001 	if (cp->cp_edx & CPUID_INTC_EDX_PGE)
1002 		feature |= X86_PGE;
1003 	if (cp->cp_edx & CPUID_INTC_EDX_CMOV)
1004 		feature |= X86_CMOV;
1005 	if (cp->cp_edx & CPUID_INTC_EDX_MMX)
1006 		feature |= X86_MMX;
1007 	if ((cp->cp_edx & CPUID_INTC_EDX_MCE) != 0 &&
1008 	    (cp->cp_edx & CPUID_INTC_EDX_MCA) != 0)
1009 		feature |= X86_MCA;
1010 	if (cp->cp_edx & CPUID_INTC_EDX_PAE)
1011 		feature |= X86_PAE;
1012 	if (cp->cp_edx & CPUID_INTC_EDX_CX8)
1013 		feature |= X86_CX8;
1014 	if (cp->cp_ecx & CPUID_INTC_ECX_CX16)
1015 		feature |= X86_CX16;
1016 	if (cp->cp_edx & CPUID_INTC_EDX_PAT)
1017 		feature |= X86_PAT;
1018 	if (cp->cp_edx & CPUID_INTC_EDX_SEP)
1019 		feature |= X86_SEP;
1020 	if (cp->cp_edx & CPUID_INTC_EDX_FXSR) {
1021 		/*
1022 		 * In our implementation, fxsave/fxrstor
1023 		 * are prerequisites before we'll even
1024 		 * try and do SSE things.
1025 		 */
1026 		if (cp->cp_edx & CPUID_INTC_EDX_SSE)
1027 			feature |= X86_SSE;
1028 		if (cp->cp_edx & CPUID_INTC_EDX_SSE2)
1029 			feature |= X86_SSE2;
1030 		if (cp->cp_ecx & CPUID_INTC_ECX_SSE3)
1031 			feature |= X86_SSE3;
1032 		if (cpi->cpi_vendor == X86_VENDOR_Intel) {
1033 			if (cp->cp_ecx & CPUID_INTC_ECX_SSSE3)
1034 				feature |= X86_SSSE3;
1035 			if (cp->cp_ecx & CPUID_INTC_ECX_SSE4_1)
1036 				feature |= X86_SSE4_1;
1037 			if (cp->cp_ecx & CPUID_INTC_ECX_SSE4_2)
1038 				feature |= X86_SSE4_2;
1039 			if (cp->cp_ecx & CPUID_INTC_ECX_AES)
1040 				feature |= X86_AES;
1041 		}
1042 	}
1043 	if (cp->cp_edx & CPUID_INTC_EDX_DE)
1044 		feature |= X86_DE;
1045 #if !defined(__xpv)
1046 	if (cp->cp_ecx & CPUID_INTC_ECX_MON) {
1047 
1048 		/*
1049 		 * We require the CLFLUSH instruction for erratum workaround
1050 		 * to use MONITOR/MWAIT.
1051 		 */
1052 		if (cp->cp_edx & CPUID_INTC_EDX_CLFSH) {
1053 			cpi->cpi_mwait.support |= MWAIT_SUPPORT;
1054 			feature |= X86_MWAIT;
1055 		} else {
1056 			extern int idle_cpu_assert_cflush_monitor;
1057 
1058 			/*
1059 			 * All processors we are aware of which have
1060 			 * MONITOR/MWAIT also have CLFLUSH.
1061 			 */
1062 			if (idle_cpu_assert_cflush_monitor) {
1063 				ASSERT((cp->cp_ecx & CPUID_INTC_ECX_MON) &&
1064 				    (cp->cp_edx & CPUID_INTC_EDX_CLFSH));
1065 			}
1066 		}
1067 	}
1068 #endif	/* __xpv */
1069 
1070 	/*
1071 	 * Only need it first time, rest of the cpus would follow suite.
1072 	 * we only capture this for the bootcpu.
1073 	 */
1074 	if (cp->cp_edx & CPUID_INTC_EDX_CLFSH) {
1075 		feature |= X86_CLFSH;
1076 		x86_clflush_size = (BITX(cp->cp_ebx, 15, 8) * 8);
1077 	}
1078 
1079 	if (feature & X86_PAE)
1080 		cpi->cpi_pabits = 36;
1081 
1082 	/*
1083 	 * Hyperthreading configuration is slightly tricky on Intel
1084 	 * and pure clones, and even trickier on AMD.
1085 	 *
1086 	 * (AMD chose to set the HTT bit on their CMP processors,
1087 	 * even though they're not actually hyperthreaded.  Thus it
1088 	 * takes a bit more work to figure out what's really going
1089 	 * on ... see the handling of the CMP_LGCY bit below)
1090 	 */
1091 	if (cp->cp_edx & CPUID_INTC_EDX_HTT) {
1092 		cpi->cpi_ncpu_per_chip = CPI_CPU_COUNT(cpi);
1093 		if (cpi->cpi_ncpu_per_chip > 1)
1094 			feature |= X86_HTT;
1095 	} else {
1096 		cpi->cpi_ncpu_per_chip = 1;
1097 	}
1098 
1099 	/*
1100 	 * Work on the "extended" feature information, doing
1101 	 * some basic initialization for cpuid_pass2()
1102 	 */
1103 	xcpuid = 0;
1104 	switch (cpi->cpi_vendor) {
1105 	case X86_VENDOR_Intel:
1106 		if (IS_NEW_F6(cpi) || cpi->cpi_family >= 0xf)
1107 			xcpuid++;
1108 		break;
1109 	case X86_VENDOR_AMD:
1110 		if (cpi->cpi_family > 5 ||
1111 		    (cpi->cpi_family == 5 && cpi->cpi_model >= 1))
1112 			xcpuid++;
1113 		break;
1114 	case X86_VENDOR_Cyrix:
1115 		/*
1116 		 * Only these Cyrix CPUs are -known- to support
1117 		 * extended cpuid operations.
1118 		 */
1119 		if (x86_type == X86_TYPE_VIA_CYRIX_III ||
1120 		    x86_type == X86_TYPE_CYRIX_GXm)
1121 			xcpuid++;
1122 		break;
1123 	case X86_VENDOR_Centaur:
1124 	case X86_VENDOR_TM:
1125 	default:
1126 		xcpuid++;
1127 		break;
1128 	}
1129 
1130 	if (xcpuid) {
1131 		cp = &cpi->cpi_extd[0];
1132 		cp->cp_eax = 0x80000000;
1133 		cpi->cpi_xmaxeax = __cpuid_insn(cp);
1134 	}
1135 
1136 	if (cpi->cpi_xmaxeax & 0x80000000) {
1137 
1138 		if (cpi->cpi_xmaxeax > CPI_XMAXEAX_MAX)
1139 			cpi->cpi_xmaxeax = CPI_XMAXEAX_MAX;
1140 
1141 		switch (cpi->cpi_vendor) {
1142 		case X86_VENDOR_Intel:
1143 		case X86_VENDOR_AMD:
1144 			if (cpi->cpi_xmaxeax < 0x80000001)
1145 				break;
1146 			cp = &cpi->cpi_extd[1];
1147 			cp->cp_eax = 0x80000001;
1148 			(void) __cpuid_insn(cp);
1149 
1150 			if (cpi->cpi_vendor == X86_VENDOR_AMD &&
1151 			    cpi->cpi_family == 5 &&
1152 			    cpi->cpi_model == 6 &&
1153 			    cpi->cpi_step == 6) {
1154 				/*
1155 				 * K6 model 6 uses bit 10 to indicate SYSC
1156 				 * Later models use bit 11. Fix it here.
1157 				 */
1158 				if (cp->cp_edx & 0x400) {
1159 					cp->cp_edx &= ~0x400;
1160 					cp->cp_edx |= CPUID_AMD_EDX_SYSC;
1161 				}
1162 			}
1163 
1164 			platform_cpuid_mangle(cpi->cpi_vendor, 0x80000001, cp);
1165 
1166 			/*
1167 			 * Compute the additions to the kernel's feature word.
1168 			 */
1169 			if (cp->cp_edx & CPUID_AMD_EDX_NX)
1170 				feature |= X86_NX;
1171 
1172 			/*
1173 			 * Regardless whether or not we boot 64-bit,
1174 			 * we should have a way to identify whether
1175 			 * the CPU is capable of running 64-bit.
1176 			 */
1177 			if (cp->cp_edx & CPUID_AMD_EDX_LM)
1178 				feature |= X86_64;
1179 
1180 #if defined(__amd64)
1181 			/* 1 GB large page - enable only for 64 bit kernel */
1182 			if (cp->cp_edx & CPUID_AMD_EDX_1GPG)
1183 				feature |= X86_1GPG;
1184 #endif
1185 
1186 			if ((cpi->cpi_vendor == X86_VENDOR_AMD) &&
1187 			    (cpi->cpi_std[1].cp_edx & CPUID_INTC_EDX_FXSR) &&
1188 			    (cp->cp_ecx & CPUID_AMD_ECX_SSE4A))
1189 				feature |= X86_SSE4A;
1190 
1191 			/*
1192 			 * If both the HTT and CMP_LGCY bits are set,
1193 			 * then we're not actually HyperThreaded.  Read
1194 			 * "AMD CPUID Specification" for more details.
1195 			 */
1196 			if (cpi->cpi_vendor == X86_VENDOR_AMD &&
1197 			    (feature & X86_HTT) &&
1198 			    (cp->cp_ecx & CPUID_AMD_ECX_CMP_LGCY)) {
1199 				feature &= ~X86_HTT;
1200 				feature |= X86_CMP;
1201 			}
1202 #if defined(__amd64)
1203 			/*
1204 			 * It's really tricky to support syscall/sysret in
1205 			 * the i386 kernel; we rely on sysenter/sysexit
1206 			 * instead.  In the amd64 kernel, things are -way-
1207 			 * better.
1208 			 */
1209 			if (cp->cp_edx & CPUID_AMD_EDX_SYSC)
1210 				feature |= X86_ASYSC;
1211 
1212 			/*
1213 			 * While we're thinking about system calls, note
1214 			 * that AMD processors don't support sysenter
1215 			 * in long mode at all, so don't try to program them.
1216 			 */
1217 			if (x86_vendor == X86_VENDOR_AMD)
1218 				feature &= ~X86_SEP;
1219 #endif
1220 			if (cp->cp_edx & CPUID_AMD_EDX_TSCP)
1221 				feature |= X86_TSCP;
1222 			break;
1223 		default:
1224 			break;
1225 		}
1226 
1227 		/*
1228 		 * Get CPUID data about processor cores and hyperthreads.
1229 		 */
1230 		switch (cpi->cpi_vendor) {
1231 		case X86_VENDOR_Intel:
1232 			if (cpi->cpi_maxeax >= 4) {
1233 				cp = &cpi->cpi_std[4];
1234 				cp->cp_eax = 4;
1235 				cp->cp_ecx = 0;
1236 				(void) __cpuid_insn(cp);
1237 				platform_cpuid_mangle(cpi->cpi_vendor, 4, cp);
1238 			}
1239 			/*FALLTHROUGH*/
1240 		case X86_VENDOR_AMD:
1241 			if (cpi->cpi_xmaxeax < 0x80000008)
1242 				break;
1243 			cp = &cpi->cpi_extd[8];
1244 			cp->cp_eax = 0x80000008;
1245 			(void) __cpuid_insn(cp);
1246 			platform_cpuid_mangle(cpi->cpi_vendor, 0x80000008, cp);
1247 
1248 			/*
1249 			 * Virtual and physical address limits from
1250 			 * cpuid override previously guessed values.
1251 			 */
1252 			cpi->cpi_pabits = BITX(cp->cp_eax, 7, 0);
1253 			cpi->cpi_vabits = BITX(cp->cp_eax, 15, 8);
1254 			break;
1255 		default:
1256 			break;
1257 		}
1258 
1259 		/*
1260 		 * Derive the number of cores per chip
1261 		 */
1262 		switch (cpi->cpi_vendor) {
1263 		case X86_VENDOR_Intel:
1264 			if (cpi->cpi_maxeax < 4) {
1265 				cpi->cpi_ncore_per_chip = 1;
1266 				break;
1267 			} else {
1268 				cpi->cpi_ncore_per_chip =
1269 				    BITX((cpi)->cpi_std[4].cp_eax, 31, 26) + 1;
1270 			}
1271 			break;
1272 		case X86_VENDOR_AMD:
1273 			if (cpi->cpi_xmaxeax < 0x80000008) {
1274 				cpi->cpi_ncore_per_chip = 1;
1275 				break;
1276 			} else {
1277 				/*
1278 				 * On family 0xf cpuid fn 2 ECX[7:0] "NC" is
1279 				 * 1 less than the number of physical cores on
1280 				 * the chip.  In family 0x10 this value can
1281 				 * be affected by "downcoring" - it reflects
1282 				 * 1 less than the number of cores actually
1283 				 * enabled on this node.
1284 				 */
1285 				cpi->cpi_ncore_per_chip =
1286 				    BITX((cpi)->cpi_extd[8].cp_ecx, 7, 0) + 1;
1287 			}
1288 			break;
1289 		default:
1290 			cpi->cpi_ncore_per_chip = 1;
1291 			break;
1292 		}
1293 
1294 		/*
1295 		 * Get CPUID data about TSC Invariance in Deep C-State.
1296 		 */
1297 		switch (cpi->cpi_vendor) {
1298 		case X86_VENDOR_Intel:
1299 			if (cpi->cpi_maxeax >= 7) {
1300 				cp = &cpi->cpi_extd[7];
1301 				cp->cp_eax = 0x80000007;
1302 				cp->cp_ecx = 0;
1303 				(void) __cpuid_insn(cp);
1304 			}
1305 			break;
1306 		default:
1307 			break;
1308 		}
1309 	} else {
1310 		cpi->cpi_ncore_per_chip = 1;
1311 	}
1312 
1313 	/*
1314 	 * If more than one core, then this processor is CMP.
1315 	 */
1316 	if (cpi->cpi_ncore_per_chip > 1)
1317 		feature |= X86_CMP;
1318 
1319 	/*
1320 	 * If the number of cores is the same as the number
1321 	 * of CPUs, then we cannot have HyperThreading.
1322 	 */
1323 	if (cpi->cpi_ncpu_per_chip == cpi->cpi_ncore_per_chip)
1324 		feature &= ~X86_HTT;
1325 
1326 	cpi->cpi_apicid = CPI_APIC_ID(cpi);
1327 	cpi->cpi_procnodes_per_pkg = 1;
1328 
1329 	if ((feature & (X86_HTT | X86_CMP)) == 0) {
1330 		/*
1331 		 * Single-core single-threaded processors.
1332 		 */
1333 		cpi->cpi_chipid = -1;
1334 		cpi->cpi_clogid = 0;
1335 		cpi->cpi_coreid = cpu->cpu_id;
1336 		cpi->cpi_pkgcoreid = 0;
1337 		if (cpi->cpi_vendor == X86_VENDOR_AMD)
1338 			cpi->cpi_procnodeid = BITX(cpi->cpi_apicid, 3, 0);
1339 		else
1340 			cpi->cpi_procnodeid = cpi->cpi_chipid;
1341 	} else if (cpi->cpi_ncpu_per_chip > 1) {
1342 		if (cpi->cpi_vendor == X86_VENDOR_Intel)
1343 			cpuid_intel_getids(cpu, feature);
1344 		else if (cpi->cpi_vendor == X86_VENDOR_AMD)
1345 			cpuid_amd_getids(cpu);
1346 		else {
1347 			/*
1348 			 * All other processors are currently
1349 			 * assumed to have single cores.
1350 			 */
1351 			cpi->cpi_coreid = cpi->cpi_chipid;
1352 			cpi->cpi_pkgcoreid = 0;
1353 			cpi->cpi_procnodeid = cpi->cpi_chipid;
1354 		}
1355 	}
1356 
1357 	/*
1358 	 * Synthesize chip "revision" and socket type
1359 	 */
1360 	cpi->cpi_chiprev = _cpuid_chiprev(cpi->cpi_vendor, cpi->cpi_family,
1361 	    cpi->cpi_model, cpi->cpi_step);
1362 	cpi->cpi_chiprevstr = _cpuid_chiprevstr(cpi->cpi_vendor,
1363 	    cpi->cpi_family, cpi->cpi_model, cpi->cpi_step);
1364 	cpi->cpi_socket = _cpuid_skt(cpi->cpi_vendor, cpi->cpi_family,
1365 	    cpi->cpi_model, cpi->cpi_step);
1366 
1367 pass1_done:
1368 	cpi->cpi_pass = 1;
1369 	return (feature);
1370 }
1371 
1372 /*
1373  * Make copies of the cpuid table entries we depend on, in
1374  * part for ease of parsing now, in part so that we have only
1375  * one place to correct any of it, in part for ease of
1376  * later export to userland, and in part so we can look at
1377  * this stuff in a crash dump.
1378  */
1379 
1380 /*ARGSUSED*/
1381 void
1382 cpuid_pass2(cpu_t *cpu)
1383 {
1384 	uint_t n, nmax;
1385 	int i;
1386 	struct cpuid_regs *cp;
1387 	uint8_t *dp;
1388 	uint32_t *iptr;
1389 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
1390 
1391 	ASSERT(cpi->cpi_pass == 1);
1392 
1393 	if (cpi->cpi_maxeax < 1)
1394 		goto pass2_done;
1395 
1396 	if ((nmax = cpi->cpi_maxeax + 1) > NMAX_CPI_STD)
1397 		nmax = NMAX_CPI_STD;
1398 	/*
1399 	 * (We already handled n == 0 and n == 1 in pass 1)
1400 	 */
1401 	for (n = 2, cp = &cpi->cpi_std[2]; n < nmax; n++, cp++) {
1402 		cp->cp_eax = n;
1403 
1404 		/*
1405 		 * CPUID function 4 expects %ecx to be initialized
1406 		 * with an index which indicates which cache to return
1407 		 * information about. The OS is expected to call function 4
1408 		 * with %ecx set to 0, 1, 2, ... until it returns with
1409 		 * EAX[4:0] set to 0, which indicates there are no more
1410 		 * caches.
1411 		 *
1412 		 * Here, populate cpi_std[4] with the information returned by
1413 		 * function 4 when %ecx == 0, and do the rest in cpuid_pass3()
1414 		 * when dynamic memory allocation becomes available.
1415 		 *
1416 		 * Note: we need to explicitly initialize %ecx here, since
1417 		 * function 4 may have been previously invoked.
1418 		 */
1419 		if (n == 4)
1420 			cp->cp_ecx = 0;
1421 
1422 		(void) __cpuid_insn(cp);
1423 		platform_cpuid_mangle(cpi->cpi_vendor, n, cp);
1424 		switch (n) {
1425 		case 2:
1426 			/*
1427 			 * "the lower 8 bits of the %eax register
1428 			 * contain a value that identifies the number
1429 			 * of times the cpuid [instruction] has to be
1430 			 * executed to obtain a complete image of the
1431 			 * processor's caching systems."
1432 			 *
1433 			 * How *do* they make this stuff up?
1434 			 */
1435 			cpi->cpi_ncache = sizeof (*cp) *
1436 			    BITX(cp->cp_eax, 7, 0);
1437 			if (cpi->cpi_ncache == 0)
1438 				break;
1439 			cpi->cpi_ncache--;	/* skip count byte */
1440 
1441 			/*
1442 			 * Well, for now, rather than attempt to implement
1443 			 * this slightly dubious algorithm, we just look
1444 			 * at the first 15 ..
1445 			 */
1446 			if (cpi->cpi_ncache > (sizeof (*cp) - 1))
1447 				cpi->cpi_ncache = sizeof (*cp) - 1;
1448 
1449 			dp = cpi->cpi_cacheinfo;
1450 			if (BITX(cp->cp_eax, 31, 31) == 0) {
1451 				uint8_t *p = (void *)&cp->cp_eax;
1452 				for (i = 1; i < 4; i++)
1453 					if (p[i] != 0)
1454 						*dp++ = p[i];
1455 			}
1456 			if (BITX(cp->cp_ebx, 31, 31) == 0) {
1457 				uint8_t *p = (void *)&cp->cp_ebx;
1458 				for (i = 0; i < 4; i++)
1459 					if (p[i] != 0)
1460 						*dp++ = p[i];
1461 			}
1462 			if (BITX(cp->cp_ecx, 31, 31) == 0) {
1463 				uint8_t *p = (void *)&cp->cp_ecx;
1464 				for (i = 0; i < 4; i++)
1465 					if (p[i] != 0)
1466 						*dp++ = p[i];
1467 			}
1468 			if (BITX(cp->cp_edx, 31, 31) == 0) {
1469 				uint8_t *p = (void *)&cp->cp_edx;
1470 				for (i = 0; i < 4; i++)
1471 					if (p[i] != 0)
1472 						*dp++ = p[i];
1473 			}
1474 			break;
1475 
1476 		case 3:	/* Processor serial number, if PSN supported */
1477 			break;
1478 
1479 		case 4:	/* Deterministic cache parameters */
1480 			break;
1481 
1482 		case 5:	/* Monitor/Mwait parameters */
1483 		{
1484 			size_t mwait_size;
1485 
1486 			/*
1487 			 * check cpi_mwait.support which was set in cpuid_pass1
1488 			 */
1489 			if (!(cpi->cpi_mwait.support & MWAIT_SUPPORT))
1490 				break;
1491 
1492 			/*
1493 			 * Protect ourself from insane mwait line size.
1494 			 * Workaround for incomplete hardware emulator(s).
1495 			 */
1496 			mwait_size = (size_t)MWAIT_SIZE_MAX(cpi);
1497 			if (mwait_size < sizeof (uint32_t) ||
1498 			    !ISP2(mwait_size)) {
1499 #if DEBUG
1500 				cmn_err(CE_NOTE, "Cannot handle cpu %d mwait "
1501 				    "size %ld", cpu->cpu_id, (long)mwait_size);
1502 #endif
1503 				break;
1504 			}
1505 
1506 			cpi->cpi_mwait.mon_min = (size_t)MWAIT_SIZE_MIN(cpi);
1507 			cpi->cpi_mwait.mon_max = mwait_size;
1508 			if (MWAIT_EXTENSION(cpi)) {
1509 				cpi->cpi_mwait.support |= MWAIT_EXTENSIONS;
1510 				if (MWAIT_INT_ENABLE(cpi))
1511 					cpi->cpi_mwait.support |=
1512 					    MWAIT_ECX_INT_ENABLE;
1513 			}
1514 			break;
1515 		}
1516 		default:
1517 			break;
1518 		}
1519 	}
1520 
1521 	if (cpi->cpi_maxeax >= 0xB && cpi->cpi_vendor == X86_VENDOR_Intel) {
1522 		struct cpuid_regs regs;
1523 
1524 		cp = &regs;
1525 		cp->cp_eax = 0xB;
1526 		cp->cp_edx = cp->cp_ebx = cp->cp_ecx = 0;
1527 
1528 		(void) __cpuid_insn(cp);
1529 
1530 		/*
1531 		 * Check CPUID.EAX=0BH, ECX=0H:EBX is non-zero, which
1532 		 * indicates that the extended topology enumeration leaf is
1533 		 * available.
1534 		 */
1535 		if (cp->cp_ebx) {
1536 			uint32_t x2apic_id;
1537 			uint_t coreid_shift = 0;
1538 			uint_t ncpu_per_core = 1;
1539 			uint_t chipid_shift = 0;
1540 			uint_t ncpu_per_chip = 1;
1541 			uint_t i;
1542 			uint_t level;
1543 
1544 			for (i = 0; i < CPI_FNB_ECX_MAX; i++) {
1545 				cp->cp_eax = 0xB;
1546 				cp->cp_ecx = i;
1547 
1548 				(void) __cpuid_insn(cp);
1549 				level = CPI_CPU_LEVEL_TYPE(cp);
1550 
1551 				if (level == 1) {
1552 					x2apic_id = cp->cp_edx;
1553 					coreid_shift = BITX(cp->cp_eax, 4, 0);
1554 					ncpu_per_core = BITX(cp->cp_ebx, 15, 0);
1555 				} else if (level == 2) {
1556 					x2apic_id = cp->cp_edx;
1557 					chipid_shift = BITX(cp->cp_eax, 4, 0);
1558 					ncpu_per_chip = BITX(cp->cp_ebx, 15, 0);
1559 				}
1560 			}
1561 
1562 			cpi->cpi_apicid = x2apic_id;
1563 			cpi->cpi_ncpu_per_chip = ncpu_per_chip;
1564 			cpi->cpi_ncore_per_chip = ncpu_per_chip /
1565 			    ncpu_per_core;
1566 			cpi->cpi_chipid = x2apic_id >> chipid_shift;
1567 			cpi->cpi_clogid = x2apic_id & ((1 << chipid_shift) - 1);
1568 			cpi->cpi_coreid = x2apic_id >> coreid_shift;
1569 			cpi->cpi_pkgcoreid = cpi->cpi_clogid >> coreid_shift;
1570 		}
1571 
1572 		/* Make cp NULL so that we don't stumble on others */
1573 		cp = NULL;
1574 	}
1575 
1576 	if ((cpi->cpi_xmaxeax & 0x80000000) == 0)
1577 		goto pass2_done;
1578 
1579 	if ((nmax = cpi->cpi_xmaxeax - 0x80000000 + 1) > NMAX_CPI_EXTD)
1580 		nmax = NMAX_CPI_EXTD;
1581 	/*
1582 	 * Copy the extended properties, fixing them as we go.
1583 	 * (We already handled n == 0 and n == 1 in pass 1)
1584 	 */
1585 	iptr = (void *)cpi->cpi_brandstr;
1586 	for (n = 2, cp = &cpi->cpi_extd[2]; n < nmax; cp++, n++) {
1587 		cp->cp_eax = 0x80000000 + n;
1588 		(void) __cpuid_insn(cp);
1589 		platform_cpuid_mangle(cpi->cpi_vendor, 0x80000000 + n, cp);
1590 		switch (n) {
1591 		case 2:
1592 		case 3:
1593 		case 4:
1594 			/*
1595 			 * Extract the brand string
1596 			 */
1597 			*iptr++ = cp->cp_eax;
1598 			*iptr++ = cp->cp_ebx;
1599 			*iptr++ = cp->cp_ecx;
1600 			*iptr++ = cp->cp_edx;
1601 			break;
1602 		case 5:
1603 			switch (cpi->cpi_vendor) {
1604 			case X86_VENDOR_AMD:
1605 				/*
1606 				 * The Athlon and Duron were the first
1607 				 * parts to report the sizes of the
1608 				 * TLB for large pages. Before then,
1609 				 * we don't trust the data.
1610 				 */
1611 				if (cpi->cpi_family < 6 ||
1612 				    (cpi->cpi_family == 6 &&
1613 				    cpi->cpi_model < 1))
1614 					cp->cp_eax = 0;
1615 				break;
1616 			default:
1617 				break;
1618 			}
1619 			break;
1620 		case 6:
1621 			switch (cpi->cpi_vendor) {
1622 			case X86_VENDOR_AMD:
1623 				/*
1624 				 * The Athlon and Duron were the first
1625 				 * AMD parts with L2 TLB's.
1626 				 * Before then, don't trust the data.
1627 				 */
1628 				if (cpi->cpi_family < 6 ||
1629 				    cpi->cpi_family == 6 &&
1630 				    cpi->cpi_model < 1)
1631 					cp->cp_eax = cp->cp_ebx = 0;
1632 				/*
1633 				 * AMD Duron rev A0 reports L2
1634 				 * cache size incorrectly as 1K
1635 				 * when it is really 64K
1636 				 */
1637 				if (cpi->cpi_family == 6 &&
1638 				    cpi->cpi_model == 3 &&
1639 				    cpi->cpi_step == 0) {
1640 					cp->cp_ecx &= 0xffff;
1641 					cp->cp_ecx |= 0x400000;
1642 				}
1643 				break;
1644 			case X86_VENDOR_Cyrix:	/* VIA C3 */
1645 				/*
1646 				 * VIA C3 processors are a bit messed
1647 				 * up w.r.t. encoding cache sizes in %ecx
1648 				 */
1649 				if (cpi->cpi_family != 6)
1650 					break;
1651 				/*
1652 				 * model 7 and 8 were incorrectly encoded
1653 				 *
1654 				 * xxx is model 8 really broken?
1655 				 */
1656 				if (cpi->cpi_model == 7 ||
1657 				    cpi->cpi_model == 8)
1658 					cp->cp_ecx =
1659 					    BITX(cp->cp_ecx, 31, 24) << 16 |
1660 					    BITX(cp->cp_ecx, 23, 16) << 12 |
1661 					    BITX(cp->cp_ecx, 15, 8) << 8 |
1662 					    BITX(cp->cp_ecx, 7, 0);
1663 				/*
1664 				 * model 9 stepping 1 has wrong associativity
1665 				 */
1666 				if (cpi->cpi_model == 9 && cpi->cpi_step == 1)
1667 					cp->cp_ecx |= 8 << 12;
1668 				break;
1669 			case X86_VENDOR_Intel:
1670 				/*
1671 				 * Extended L2 Cache features function.
1672 				 * First appeared on Prescott.
1673 				 */
1674 			default:
1675 				break;
1676 			}
1677 			break;
1678 		default:
1679 			break;
1680 		}
1681 	}
1682 
1683 pass2_done:
1684 	cpi->cpi_pass = 2;
1685 }
1686 
1687 static const char *
1688 intel_cpubrand(const struct cpuid_info *cpi)
1689 {
1690 	int i;
1691 
1692 	if ((x86_feature & X86_CPUID) == 0 ||
1693 	    cpi->cpi_maxeax < 1 || cpi->cpi_family < 5)
1694 		return ("i486");
1695 
1696 	switch (cpi->cpi_family) {
1697 	case 5:
1698 		return ("Intel Pentium(r)");
1699 	case 6:
1700 		switch (cpi->cpi_model) {
1701 			uint_t celeron, xeon;
1702 			const struct cpuid_regs *cp;
1703 		case 0:
1704 		case 1:
1705 		case 2:
1706 			return ("Intel Pentium(r) Pro");
1707 		case 3:
1708 		case 4:
1709 			return ("Intel Pentium(r) II");
1710 		case 6:
1711 			return ("Intel Celeron(r)");
1712 		case 5:
1713 		case 7:
1714 			celeron = xeon = 0;
1715 			cp = &cpi->cpi_std[2];	/* cache info */
1716 
1717 			for (i = 1; i < 4; i++) {
1718 				uint_t tmp;
1719 
1720 				tmp = (cp->cp_eax >> (8 * i)) & 0xff;
1721 				if (tmp == 0x40)
1722 					celeron++;
1723 				if (tmp >= 0x44 && tmp <= 0x45)
1724 					xeon++;
1725 			}
1726 
1727 			for (i = 0; i < 2; i++) {
1728 				uint_t tmp;
1729 
1730 				tmp = (cp->cp_ebx >> (8 * i)) & 0xff;
1731 				if (tmp == 0x40)
1732 					celeron++;
1733 				else if (tmp >= 0x44 && tmp <= 0x45)
1734 					xeon++;
1735 			}
1736 
1737 			for (i = 0; i < 4; i++) {
1738 				uint_t tmp;
1739 
1740 				tmp = (cp->cp_ecx >> (8 * i)) & 0xff;
1741 				if (tmp == 0x40)
1742 					celeron++;
1743 				else if (tmp >= 0x44 && tmp <= 0x45)
1744 					xeon++;
1745 			}
1746 
1747 			for (i = 0; i < 4; i++) {
1748 				uint_t tmp;
1749 
1750 				tmp = (cp->cp_edx >> (8 * i)) & 0xff;
1751 				if (tmp == 0x40)
1752 					celeron++;
1753 				else if (tmp >= 0x44 && tmp <= 0x45)
1754 					xeon++;
1755 			}
1756 
1757 			if (celeron)
1758 				return ("Intel Celeron(r)");
1759 			if (xeon)
1760 				return (cpi->cpi_model == 5 ?
1761 				    "Intel Pentium(r) II Xeon(tm)" :
1762 				    "Intel Pentium(r) III Xeon(tm)");
1763 			return (cpi->cpi_model == 5 ?
1764 			    "Intel Pentium(r) II or Pentium(r) II Xeon(tm)" :
1765 			    "Intel Pentium(r) III or Pentium(r) III Xeon(tm)");
1766 		default:
1767 			break;
1768 		}
1769 	default:
1770 		break;
1771 	}
1772 
1773 	/* BrandID is present if the field is nonzero */
1774 	if (cpi->cpi_brandid != 0) {
1775 		static const struct {
1776 			uint_t bt_bid;
1777 			const char *bt_str;
1778 		} brand_tbl[] = {
1779 			{ 0x1,	"Intel(r) Celeron(r)" },
1780 			{ 0x2,	"Intel(r) Pentium(r) III" },
1781 			{ 0x3,	"Intel(r) Pentium(r) III Xeon(tm)" },
1782 			{ 0x4,	"Intel(r) Pentium(r) III" },
1783 			{ 0x6,	"Mobile Intel(r) Pentium(r) III" },
1784 			{ 0x7,	"Mobile Intel(r) Celeron(r)" },
1785 			{ 0x8,	"Intel(r) Pentium(r) 4" },
1786 			{ 0x9,	"Intel(r) Pentium(r) 4" },
1787 			{ 0xa,	"Intel(r) Celeron(r)" },
1788 			{ 0xb,	"Intel(r) Xeon(tm)" },
1789 			{ 0xc,	"Intel(r) Xeon(tm) MP" },
1790 			{ 0xe,	"Mobile Intel(r) Pentium(r) 4" },
1791 			{ 0xf,	"Mobile Intel(r) Celeron(r)" },
1792 			{ 0x11, "Mobile Genuine Intel(r)" },
1793 			{ 0x12, "Intel(r) Celeron(r) M" },
1794 			{ 0x13, "Mobile Intel(r) Celeron(r)" },
1795 			{ 0x14, "Intel(r) Celeron(r)" },
1796 			{ 0x15, "Mobile Genuine Intel(r)" },
1797 			{ 0x16,	"Intel(r) Pentium(r) M" },
1798 			{ 0x17, "Mobile Intel(r) Celeron(r)" }
1799 		};
1800 		uint_t btblmax = sizeof (brand_tbl) / sizeof (brand_tbl[0]);
1801 		uint_t sgn;
1802 
1803 		sgn = (cpi->cpi_family << 8) |
1804 		    (cpi->cpi_model << 4) | cpi->cpi_step;
1805 
1806 		for (i = 0; i < btblmax; i++)
1807 			if (brand_tbl[i].bt_bid == cpi->cpi_brandid)
1808 				break;
1809 		if (i < btblmax) {
1810 			if (sgn == 0x6b1 && cpi->cpi_brandid == 3)
1811 				return ("Intel(r) Celeron(r)");
1812 			if (sgn < 0xf13 && cpi->cpi_brandid == 0xb)
1813 				return ("Intel(r) Xeon(tm) MP");
1814 			if (sgn < 0xf13 && cpi->cpi_brandid == 0xe)
1815 				return ("Intel(r) Xeon(tm)");
1816 			return (brand_tbl[i].bt_str);
1817 		}
1818 	}
1819 
1820 	return (NULL);
1821 }
1822 
1823 static const char *
1824 amd_cpubrand(const struct cpuid_info *cpi)
1825 {
1826 	if ((x86_feature & X86_CPUID) == 0 ||
1827 	    cpi->cpi_maxeax < 1 || cpi->cpi_family < 5)
1828 		return ("i486 compatible");
1829 
1830 	switch (cpi->cpi_family) {
1831 	case 5:
1832 		switch (cpi->cpi_model) {
1833 		case 0:
1834 		case 1:
1835 		case 2:
1836 		case 3:
1837 		case 4:
1838 		case 5:
1839 			return ("AMD-K5(r)");
1840 		case 6:
1841 		case 7:
1842 			return ("AMD-K6(r)");
1843 		case 8:
1844 			return ("AMD-K6(r)-2");
1845 		case 9:
1846 			return ("AMD-K6(r)-III");
1847 		default:
1848 			return ("AMD (family 5)");
1849 		}
1850 	case 6:
1851 		switch (cpi->cpi_model) {
1852 		case 1:
1853 			return ("AMD-K7(tm)");
1854 		case 0:
1855 		case 2:
1856 		case 4:
1857 			return ("AMD Athlon(tm)");
1858 		case 3:
1859 		case 7:
1860 			return ("AMD Duron(tm)");
1861 		case 6:
1862 		case 8:
1863 		case 10:
1864 			/*
1865 			 * Use the L2 cache size to distinguish
1866 			 */
1867 			return ((cpi->cpi_extd[6].cp_ecx >> 16) >= 256 ?
1868 			    "AMD Athlon(tm)" : "AMD Duron(tm)");
1869 		default:
1870 			return ("AMD (family 6)");
1871 		}
1872 	default:
1873 		break;
1874 	}
1875 
1876 	if (cpi->cpi_family == 0xf && cpi->cpi_model == 5 &&
1877 	    cpi->cpi_brandid != 0) {
1878 		switch (BITX(cpi->cpi_brandid, 7, 5)) {
1879 		case 3:
1880 			return ("AMD Opteron(tm) UP 1xx");
1881 		case 4:
1882 			return ("AMD Opteron(tm) DP 2xx");
1883 		case 5:
1884 			return ("AMD Opteron(tm) MP 8xx");
1885 		default:
1886 			return ("AMD Opteron(tm)");
1887 		}
1888 	}
1889 
1890 	return (NULL);
1891 }
1892 
1893 static const char *
1894 cyrix_cpubrand(struct cpuid_info *cpi, uint_t type)
1895 {
1896 	if ((x86_feature & X86_CPUID) == 0 ||
1897 	    cpi->cpi_maxeax < 1 || cpi->cpi_family < 5 ||
1898 	    type == X86_TYPE_CYRIX_486)
1899 		return ("i486 compatible");
1900 
1901 	switch (type) {
1902 	case X86_TYPE_CYRIX_6x86:
1903 		return ("Cyrix 6x86");
1904 	case X86_TYPE_CYRIX_6x86L:
1905 		return ("Cyrix 6x86L");
1906 	case X86_TYPE_CYRIX_6x86MX:
1907 		return ("Cyrix 6x86MX");
1908 	case X86_TYPE_CYRIX_GXm:
1909 		return ("Cyrix GXm");
1910 	case X86_TYPE_CYRIX_MediaGX:
1911 		return ("Cyrix MediaGX");
1912 	case X86_TYPE_CYRIX_MII:
1913 		return ("Cyrix M2");
1914 	case X86_TYPE_VIA_CYRIX_III:
1915 		return ("VIA Cyrix M3");
1916 	default:
1917 		/*
1918 		 * Have another wild guess ..
1919 		 */
1920 		if (cpi->cpi_family == 4 && cpi->cpi_model == 9)
1921 			return ("Cyrix 5x86");
1922 		else if (cpi->cpi_family == 5) {
1923 			switch (cpi->cpi_model) {
1924 			case 2:
1925 				return ("Cyrix 6x86");	/* Cyrix M1 */
1926 			case 4:
1927 				return ("Cyrix MediaGX");
1928 			default:
1929 				break;
1930 			}
1931 		} else if (cpi->cpi_family == 6) {
1932 			switch (cpi->cpi_model) {
1933 			case 0:
1934 				return ("Cyrix 6x86MX"); /* Cyrix M2? */
1935 			case 5:
1936 			case 6:
1937 			case 7:
1938 			case 8:
1939 			case 9:
1940 				return ("VIA C3");
1941 			default:
1942 				break;
1943 			}
1944 		}
1945 		break;
1946 	}
1947 	return (NULL);
1948 }
1949 
1950 /*
1951  * This only gets called in the case that the CPU extended
1952  * feature brand string (0x80000002, 0x80000003, 0x80000004)
1953  * aren't available, or contain null bytes for some reason.
1954  */
1955 static void
1956 fabricate_brandstr(struct cpuid_info *cpi)
1957 {
1958 	const char *brand = NULL;
1959 
1960 	switch (cpi->cpi_vendor) {
1961 	case X86_VENDOR_Intel:
1962 		brand = intel_cpubrand(cpi);
1963 		break;
1964 	case X86_VENDOR_AMD:
1965 		brand = amd_cpubrand(cpi);
1966 		break;
1967 	case X86_VENDOR_Cyrix:
1968 		brand = cyrix_cpubrand(cpi, x86_type);
1969 		break;
1970 	case X86_VENDOR_NexGen:
1971 		if (cpi->cpi_family == 5 && cpi->cpi_model == 0)
1972 			brand = "NexGen Nx586";
1973 		break;
1974 	case X86_VENDOR_Centaur:
1975 		if (cpi->cpi_family == 5)
1976 			switch (cpi->cpi_model) {
1977 			case 4:
1978 				brand = "Centaur C6";
1979 				break;
1980 			case 8:
1981 				brand = "Centaur C2";
1982 				break;
1983 			case 9:
1984 				brand = "Centaur C3";
1985 				break;
1986 			default:
1987 				break;
1988 			}
1989 		break;
1990 	case X86_VENDOR_Rise:
1991 		if (cpi->cpi_family == 5 &&
1992 		    (cpi->cpi_model == 0 || cpi->cpi_model == 2))
1993 			brand = "Rise mP6";
1994 		break;
1995 	case X86_VENDOR_SiS:
1996 		if (cpi->cpi_family == 5 && cpi->cpi_model == 0)
1997 			brand = "SiS 55x";
1998 		break;
1999 	case X86_VENDOR_TM:
2000 		if (cpi->cpi_family == 5 && cpi->cpi_model == 4)
2001 			brand = "Transmeta Crusoe TM3x00 or TM5x00";
2002 		break;
2003 	case X86_VENDOR_NSC:
2004 	case X86_VENDOR_UMC:
2005 	default:
2006 		break;
2007 	}
2008 	if (brand) {
2009 		(void) strcpy((char *)cpi->cpi_brandstr, brand);
2010 		return;
2011 	}
2012 
2013 	/*
2014 	 * If all else fails ...
2015 	 */
2016 	(void) snprintf(cpi->cpi_brandstr, sizeof (cpi->cpi_brandstr),
2017 	    "%s %d.%d.%d", cpi->cpi_vendorstr, cpi->cpi_family,
2018 	    cpi->cpi_model, cpi->cpi_step);
2019 }
2020 
2021 /*
2022  * This routine is called just after kernel memory allocation
2023  * becomes available on cpu0, and as part of mp_startup() on
2024  * the other cpus.
2025  *
2026  * Fixup the brand string, and collect any information from cpuid
2027  * that requires dynamicically allocated storage to represent.
2028  */
2029 /*ARGSUSED*/
2030 void
2031 cpuid_pass3(cpu_t *cpu)
2032 {
2033 	int	i, max, shft, level, size;
2034 	struct cpuid_regs regs;
2035 	struct cpuid_regs *cp;
2036 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2037 
2038 	ASSERT(cpi->cpi_pass == 2);
2039 
2040 	/*
2041 	 * Function 4: Deterministic cache parameters
2042 	 *
2043 	 * Take this opportunity to detect the number of threads
2044 	 * sharing the last level cache, and construct a corresponding
2045 	 * cache id. The respective cpuid_info members are initialized
2046 	 * to the default case of "no last level cache sharing".
2047 	 */
2048 	cpi->cpi_ncpu_shr_last_cache = 1;
2049 	cpi->cpi_last_lvl_cacheid = cpu->cpu_id;
2050 
2051 	if (cpi->cpi_maxeax >= 4 && cpi->cpi_vendor == X86_VENDOR_Intel) {
2052 
2053 		/*
2054 		 * Find the # of elements (size) returned by fn 4, and along
2055 		 * the way detect last level cache sharing details.
2056 		 */
2057 		bzero(&regs, sizeof (regs));
2058 		cp = &regs;
2059 		for (i = 0, max = 0; i < CPI_FN4_ECX_MAX; i++) {
2060 			cp->cp_eax = 4;
2061 			cp->cp_ecx = i;
2062 
2063 			(void) __cpuid_insn(cp);
2064 
2065 			if (CPI_CACHE_TYPE(cp) == 0)
2066 				break;
2067 			level = CPI_CACHE_LVL(cp);
2068 			if (level > max) {
2069 				max = level;
2070 				cpi->cpi_ncpu_shr_last_cache =
2071 				    CPI_NTHR_SHR_CACHE(cp) + 1;
2072 			}
2073 		}
2074 		cpi->cpi_std_4_size = size = i;
2075 
2076 		/*
2077 		 * Allocate the cpi_std_4 array. The first element
2078 		 * references the regs for fn 4, %ecx == 0, which
2079 		 * cpuid_pass2() stashed in cpi->cpi_std[4].
2080 		 */
2081 		if (size > 0) {
2082 			cpi->cpi_std_4 =
2083 			    kmem_alloc(size * sizeof (cp), KM_SLEEP);
2084 			cpi->cpi_std_4[0] = &cpi->cpi_std[4];
2085 
2086 			/*
2087 			 * Allocate storage to hold the additional regs
2088 			 * for function 4, %ecx == 1 .. cpi_std_4_size.
2089 			 *
2090 			 * The regs for fn 4, %ecx == 0 has already
2091 			 * been allocated as indicated above.
2092 			 */
2093 			for (i = 1; i < size; i++) {
2094 				cp = cpi->cpi_std_4[i] =
2095 				    kmem_zalloc(sizeof (regs), KM_SLEEP);
2096 				cp->cp_eax = 4;
2097 				cp->cp_ecx = i;
2098 
2099 				(void) __cpuid_insn(cp);
2100 			}
2101 		}
2102 		/*
2103 		 * Determine the number of bits needed to represent
2104 		 * the number of CPUs sharing the last level cache.
2105 		 *
2106 		 * Shift off that number of bits from the APIC id to
2107 		 * derive the cache id.
2108 		 */
2109 		shft = 0;
2110 		for (i = 1; i < cpi->cpi_ncpu_shr_last_cache; i <<= 1)
2111 			shft++;
2112 		cpi->cpi_last_lvl_cacheid = cpi->cpi_apicid >> shft;
2113 	}
2114 
2115 	/*
2116 	 * Now fixup the brand string
2117 	 */
2118 	if ((cpi->cpi_xmaxeax & 0x80000000) == 0) {
2119 		fabricate_brandstr(cpi);
2120 	} else {
2121 
2122 		/*
2123 		 * If we successfully extracted a brand string from the cpuid
2124 		 * instruction, clean it up by removing leading spaces and
2125 		 * similar junk.
2126 		 */
2127 		if (cpi->cpi_brandstr[0]) {
2128 			size_t maxlen = sizeof (cpi->cpi_brandstr);
2129 			char *src, *dst;
2130 
2131 			dst = src = (char *)cpi->cpi_brandstr;
2132 			src[maxlen - 1] = '\0';
2133 			/*
2134 			 * strip leading spaces
2135 			 */
2136 			while (*src == ' ')
2137 				src++;
2138 			/*
2139 			 * Remove any 'Genuine' or "Authentic" prefixes
2140 			 */
2141 			if (strncmp(src, "Genuine ", 8) == 0)
2142 				src += 8;
2143 			if (strncmp(src, "Authentic ", 10) == 0)
2144 				src += 10;
2145 
2146 			/*
2147 			 * Now do an in-place copy.
2148 			 * Map (R) to (r) and (TM) to (tm).
2149 			 * The era of teletypes is long gone, and there's
2150 			 * -really- no need to shout.
2151 			 */
2152 			while (*src != '\0') {
2153 				if (src[0] == '(') {
2154 					if (strncmp(src + 1, "R)", 2) == 0) {
2155 						(void) strncpy(dst, "(r)", 3);
2156 						src += 3;
2157 						dst += 3;
2158 						continue;
2159 					}
2160 					if (strncmp(src + 1, "TM)", 3) == 0) {
2161 						(void) strncpy(dst, "(tm)", 4);
2162 						src += 4;
2163 						dst += 4;
2164 						continue;
2165 					}
2166 				}
2167 				*dst++ = *src++;
2168 			}
2169 			*dst = '\0';
2170 
2171 			/*
2172 			 * Finally, remove any trailing spaces
2173 			 */
2174 			while (--dst > cpi->cpi_brandstr)
2175 				if (*dst == ' ')
2176 					*dst = '\0';
2177 				else
2178 					break;
2179 		} else
2180 			fabricate_brandstr(cpi);
2181 	}
2182 	cpi->cpi_pass = 3;
2183 }
2184 
2185 /*
2186  * This routine is called out of bind_hwcap() much later in the life
2187  * of the kernel (post_startup()).  The job of this routine is to resolve
2188  * the hardware feature support and kernel support for those features into
2189  * what we're actually going to tell applications via the aux vector.
2190  */
2191 uint_t
2192 cpuid_pass4(cpu_t *cpu)
2193 {
2194 	struct cpuid_info *cpi;
2195 	uint_t hwcap_flags = 0;
2196 
2197 	if (cpu == NULL)
2198 		cpu = CPU;
2199 	cpi = cpu->cpu_m.mcpu_cpi;
2200 
2201 	ASSERT(cpi->cpi_pass == 3);
2202 
2203 	if (cpi->cpi_maxeax >= 1) {
2204 		uint32_t *edx = &cpi->cpi_support[STD_EDX_FEATURES];
2205 		uint32_t *ecx = &cpi->cpi_support[STD_ECX_FEATURES];
2206 
2207 		*edx = CPI_FEATURES_EDX(cpi);
2208 		*ecx = CPI_FEATURES_ECX(cpi);
2209 
2210 		/*
2211 		 * [these require explicit kernel support]
2212 		 */
2213 		if ((x86_feature & X86_SEP) == 0)
2214 			*edx &= ~CPUID_INTC_EDX_SEP;
2215 
2216 		if ((x86_feature & X86_SSE) == 0)
2217 			*edx &= ~(CPUID_INTC_EDX_FXSR|CPUID_INTC_EDX_SSE);
2218 		if ((x86_feature & X86_SSE2) == 0)
2219 			*edx &= ~CPUID_INTC_EDX_SSE2;
2220 
2221 		if ((x86_feature & X86_HTT) == 0)
2222 			*edx &= ~CPUID_INTC_EDX_HTT;
2223 
2224 		if ((x86_feature & X86_SSE3) == 0)
2225 			*ecx &= ~CPUID_INTC_ECX_SSE3;
2226 
2227 		if (cpi->cpi_vendor == X86_VENDOR_Intel) {
2228 			if ((x86_feature & X86_SSSE3) == 0)
2229 				*ecx &= ~CPUID_INTC_ECX_SSSE3;
2230 			if ((x86_feature & X86_SSE4_1) == 0)
2231 				*ecx &= ~CPUID_INTC_ECX_SSE4_1;
2232 			if ((x86_feature & X86_SSE4_2) == 0)
2233 				*ecx &= ~CPUID_INTC_ECX_SSE4_2;
2234 			if ((x86_feature & X86_AES) == 0)
2235 				*ecx &= ~CPUID_INTC_ECX_AES;
2236 		}
2237 
2238 		/*
2239 		 * [no explicit support required beyond x87 fp context]
2240 		 */
2241 		if (!fpu_exists)
2242 			*edx &= ~(CPUID_INTC_EDX_FPU | CPUID_INTC_EDX_MMX);
2243 
2244 		/*
2245 		 * Now map the supported feature vector to things that we
2246 		 * think userland will care about.
2247 		 */
2248 		if (*edx & CPUID_INTC_EDX_SEP)
2249 			hwcap_flags |= AV_386_SEP;
2250 		if (*edx & CPUID_INTC_EDX_SSE)
2251 			hwcap_flags |= AV_386_FXSR | AV_386_SSE;
2252 		if (*edx & CPUID_INTC_EDX_SSE2)
2253 			hwcap_flags |= AV_386_SSE2;
2254 		if (*ecx & CPUID_INTC_ECX_SSE3)
2255 			hwcap_flags |= AV_386_SSE3;
2256 		if (cpi->cpi_vendor == X86_VENDOR_Intel) {
2257 			if (*ecx & CPUID_INTC_ECX_SSSE3)
2258 				hwcap_flags |= AV_386_SSSE3;
2259 			if (*ecx & CPUID_INTC_ECX_SSE4_1)
2260 				hwcap_flags |= AV_386_SSE4_1;
2261 			if (*ecx & CPUID_INTC_ECX_SSE4_2)
2262 				hwcap_flags |= AV_386_SSE4_2;
2263 			if (*ecx & CPUID_INTC_ECX_MOVBE)
2264 				hwcap_flags |= AV_386_MOVBE;
2265 			if (*ecx & CPUID_INTC_ECX_AES)
2266 				hwcap_flags |= AV_386_AES;
2267 			if (*ecx & CPUID_INTC_ECX_PCLMULQDQ)
2268 				hwcap_flags |= AV_386_PCLMULQDQ;
2269 		}
2270 		if (*ecx & CPUID_INTC_ECX_POPCNT)
2271 			hwcap_flags |= AV_386_POPCNT;
2272 		if (*edx & CPUID_INTC_EDX_FPU)
2273 			hwcap_flags |= AV_386_FPU;
2274 		if (*edx & CPUID_INTC_EDX_MMX)
2275 			hwcap_flags |= AV_386_MMX;
2276 
2277 		if (*edx & CPUID_INTC_EDX_TSC)
2278 			hwcap_flags |= AV_386_TSC;
2279 		if (*edx & CPUID_INTC_EDX_CX8)
2280 			hwcap_flags |= AV_386_CX8;
2281 		if (*edx & CPUID_INTC_EDX_CMOV)
2282 			hwcap_flags |= AV_386_CMOV;
2283 		if (*ecx & CPUID_INTC_ECX_MON)
2284 			hwcap_flags |= AV_386_MON;
2285 		if (*ecx & CPUID_INTC_ECX_CX16)
2286 			hwcap_flags |= AV_386_CX16;
2287 	}
2288 
2289 	if (x86_feature & X86_HTT)
2290 		hwcap_flags |= AV_386_PAUSE;
2291 
2292 	if (cpi->cpi_xmaxeax < 0x80000001)
2293 		goto pass4_done;
2294 
2295 	switch (cpi->cpi_vendor) {
2296 		struct cpuid_regs cp;
2297 		uint32_t *edx, *ecx;
2298 
2299 	case X86_VENDOR_Intel:
2300 		/*
2301 		 * Seems like Intel duplicated what we necessary
2302 		 * here to make the initial crop of 64-bit OS's work.
2303 		 * Hopefully, those are the only "extended" bits
2304 		 * they'll add.
2305 		 */
2306 		/*FALLTHROUGH*/
2307 
2308 	case X86_VENDOR_AMD:
2309 		edx = &cpi->cpi_support[AMD_EDX_FEATURES];
2310 		ecx = &cpi->cpi_support[AMD_ECX_FEATURES];
2311 
2312 		*edx = CPI_FEATURES_XTD_EDX(cpi);
2313 		*ecx = CPI_FEATURES_XTD_ECX(cpi);
2314 
2315 		/*
2316 		 * [these features require explicit kernel support]
2317 		 */
2318 		switch (cpi->cpi_vendor) {
2319 		case X86_VENDOR_Intel:
2320 			if ((x86_feature & X86_TSCP) == 0)
2321 				*edx &= ~CPUID_AMD_EDX_TSCP;
2322 			break;
2323 
2324 		case X86_VENDOR_AMD:
2325 			if ((x86_feature & X86_TSCP) == 0)
2326 				*edx &= ~CPUID_AMD_EDX_TSCP;
2327 			if ((x86_feature & X86_SSE4A) == 0)
2328 				*ecx &= ~CPUID_AMD_ECX_SSE4A;
2329 			break;
2330 
2331 		default:
2332 			break;
2333 		}
2334 
2335 		/*
2336 		 * [no explicit support required beyond
2337 		 * x87 fp context and exception handlers]
2338 		 */
2339 		if (!fpu_exists)
2340 			*edx &= ~(CPUID_AMD_EDX_MMXamd |
2341 			    CPUID_AMD_EDX_3DNow | CPUID_AMD_EDX_3DNowx);
2342 
2343 		if ((x86_feature & X86_NX) == 0)
2344 			*edx &= ~CPUID_AMD_EDX_NX;
2345 #if !defined(__amd64)
2346 		*edx &= ~CPUID_AMD_EDX_LM;
2347 #endif
2348 		/*
2349 		 * Now map the supported feature vector to
2350 		 * things that we think userland will care about.
2351 		 */
2352 #if defined(__amd64)
2353 		if (*edx & CPUID_AMD_EDX_SYSC)
2354 			hwcap_flags |= AV_386_AMD_SYSC;
2355 #endif
2356 		if (*edx & CPUID_AMD_EDX_MMXamd)
2357 			hwcap_flags |= AV_386_AMD_MMX;
2358 		if (*edx & CPUID_AMD_EDX_3DNow)
2359 			hwcap_flags |= AV_386_AMD_3DNow;
2360 		if (*edx & CPUID_AMD_EDX_3DNowx)
2361 			hwcap_flags |= AV_386_AMD_3DNowx;
2362 
2363 		switch (cpi->cpi_vendor) {
2364 		case X86_VENDOR_AMD:
2365 			if (*edx & CPUID_AMD_EDX_TSCP)
2366 				hwcap_flags |= AV_386_TSCP;
2367 			if (*ecx & CPUID_AMD_ECX_AHF64)
2368 				hwcap_flags |= AV_386_AHF;
2369 			if (*ecx & CPUID_AMD_ECX_SSE4A)
2370 				hwcap_flags |= AV_386_AMD_SSE4A;
2371 			if (*ecx & CPUID_AMD_ECX_LZCNT)
2372 				hwcap_flags |= AV_386_AMD_LZCNT;
2373 			break;
2374 
2375 		case X86_VENDOR_Intel:
2376 			if (*edx & CPUID_AMD_EDX_TSCP)
2377 				hwcap_flags |= AV_386_TSCP;
2378 			/*
2379 			 * Aarrgh.
2380 			 * Intel uses a different bit in the same word.
2381 			 */
2382 			if (*ecx & CPUID_INTC_ECX_AHF64)
2383 				hwcap_flags |= AV_386_AHF;
2384 			break;
2385 
2386 		default:
2387 			break;
2388 		}
2389 		break;
2390 
2391 	case X86_VENDOR_TM:
2392 		cp.cp_eax = 0x80860001;
2393 		(void) __cpuid_insn(&cp);
2394 		cpi->cpi_support[TM_EDX_FEATURES] = cp.cp_edx;
2395 		break;
2396 
2397 	default:
2398 		break;
2399 	}
2400 
2401 pass4_done:
2402 	cpi->cpi_pass = 4;
2403 	return (hwcap_flags);
2404 }
2405 
2406 
2407 /*
2408  * Simulate the cpuid instruction using the data we previously
2409  * captured about this CPU.  We try our best to return the truth
2410  * about the hardware, independently of kernel support.
2411  */
2412 uint32_t
2413 cpuid_insn(cpu_t *cpu, struct cpuid_regs *cp)
2414 {
2415 	struct cpuid_info *cpi;
2416 	struct cpuid_regs *xcp;
2417 
2418 	if (cpu == NULL)
2419 		cpu = CPU;
2420 	cpi = cpu->cpu_m.mcpu_cpi;
2421 
2422 	ASSERT(cpuid_checkpass(cpu, 3));
2423 
2424 	/*
2425 	 * CPUID data is cached in two separate places: cpi_std for standard
2426 	 * CPUID functions, and cpi_extd for extended CPUID functions.
2427 	 */
2428 	if (cp->cp_eax <= cpi->cpi_maxeax && cp->cp_eax < NMAX_CPI_STD)
2429 		xcp = &cpi->cpi_std[cp->cp_eax];
2430 	else if (cp->cp_eax >= 0x80000000 && cp->cp_eax <= cpi->cpi_xmaxeax &&
2431 	    cp->cp_eax < 0x80000000 + NMAX_CPI_EXTD)
2432 		xcp = &cpi->cpi_extd[cp->cp_eax - 0x80000000];
2433 	else
2434 		/*
2435 		 * The caller is asking for data from an input parameter which
2436 		 * the kernel has not cached.  In this case we go fetch from
2437 		 * the hardware and return the data directly to the user.
2438 		 */
2439 		return (__cpuid_insn(cp));
2440 
2441 	cp->cp_eax = xcp->cp_eax;
2442 	cp->cp_ebx = xcp->cp_ebx;
2443 	cp->cp_ecx = xcp->cp_ecx;
2444 	cp->cp_edx = xcp->cp_edx;
2445 	return (cp->cp_eax);
2446 }
2447 
2448 int
2449 cpuid_checkpass(cpu_t *cpu, int pass)
2450 {
2451 	return (cpu != NULL && cpu->cpu_m.mcpu_cpi != NULL &&
2452 	    cpu->cpu_m.mcpu_cpi->cpi_pass >= pass);
2453 }
2454 
2455 int
2456 cpuid_getbrandstr(cpu_t *cpu, char *s, size_t n)
2457 {
2458 	ASSERT(cpuid_checkpass(cpu, 3));
2459 
2460 	return (snprintf(s, n, "%s", cpu->cpu_m.mcpu_cpi->cpi_brandstr));
2461 }
2462 
2463 int
2464 cpuid_is_cmt(cpu_t *cpu)
2465 {
2466 	if (cpu == NULL)
2467 		cpu = CPU;
2468 
2469 	ASSERT(cpuid_checkpass(cpu, 1));
2470 
2471 	return (cpu->cpu_m.mcpu_cpi->cpi_chipid >= 0);
2472 }
2473 
2474 /*
2475  * AMD and Intel both implement the 64-bit variant of the syscall
2476  * instruction (syscallq), so if there's -any- support for syscall,
2477  * cpuid currently says "yes, we support this".
2478  *
2479  * However, Intel decided to -not- implement the 32-bit variant of the
2480  * syscall instruction, so we provide a predicate to allow our caller
2481  * to test that subtlety here.
2482  *
2483  * XXPV	Currently, 32-bit syscall instructions don't work via the hypervisor,
2484  *	even in the case where the hardware would in fact support it.
2485  */
2486 /*ARGSUSED*/
2487 int
2488 cpuid_syscall32_insn(cpu_t *cpu)
2489 {
2490 	ASSERT(cpuid_checkpass((cpu == NULL ? CPU : cpu), 1));
2491 
2492 #if !defined(__xpv)
2493 	if (cpu == NULL)
2494 		cpu = CPU;
2495 
2496 	/*CSTYLED*/
2497 	{
2498 		struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2499 
2500 		if (cpi->cpi_vendor == X86_VENDOR_AMD &&
2501 		    cpi->cpi_xmaxeax >= 0x80000001 &&
2502 		    (CPI_FEATURES_XTD_EDX(cpi) & CPUID_AMD_EDX_SYSC))
2503 			return (1);
2504 	}
2505 #endif
2506 	return (0);
2507 }
2508 
2509 int
2510 cpuid_getidstr(cpu_t *cpu, char *s, size_t n)
2511 {
2512 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2513 
2514 	static const char fmt[] =
2515 	    "x86 (%s %X family %d model %d step %d clock %d MHz)";
2516 	static const char fmt_ht[] =
2517 	    "x86 (chipid 0x%x %s %X family %d model %d step %d clock %d MHz)";
2518 
2519 	ASSERT(cpuid_checkpass(cpu, 1));
2520 
2521 	if (cpuid_is_cmt(cpu))
2522 		return (snprintf(s, n, fmt_ht, cpi->cpi_chipid,
2523 		    cpi->cpi_vendorstr, cpi->cpi_std[1].cp_eax,
2524 		    cpi->cpi_family, cpi->cpi_model,
2525 		    cpi->cpi_step, cpu->cpu_type_info.pi_clock));
2526 	return (snprintf(s, n, fmt,
2527 	    cpi->cpi_vendorstr, cpi->cpi_std[1].cp_eax,
2528 	    cpi->cpi_family, cpi->cpi_model,
2529 	    cpi->cpi_step, cpu->cpu_type_info.pi_clock));
2530 }
2531 
2532 const char *
2533 cpuid_getvendorstr(cpu_t *cpu)
2534 {
2535 	ASSERT(cpuid_checkpass(cpu, 1));
2536 	return ((const char *)cpu->cpu_m.mcpu_cpi->cpi_vendorstr);
2537 }
2538 
2539 uint_t
2540 cpuid_getvendor(cpu_t *cpu)
2541 {
2542 	ASSERT(cpuid_checkpass(cpu, 1));
2543 	return (cpu->cpu_m.mcpu_cpi->cpi_vendor);
2544 }
2545 
2546 uint_t
2547 cpuid_getfamily(cpu_t *cpu)
2548 {
2549 	ASSERT(cpuid_checkpass(cpu, 1));
2550 	return (cpu->cpu_m.mcpu_cpi->cpi_family);
2551 }
2552 
2553 uint_t
2554 cpuid_getmodel(cpu_t *cpu)
2555 {
2556 	ASSERT(cpuid_checkpass(cpu, 1));
2557 	return (cpu->cpu_m.mcpu_cpi->cpi_model);
2558 }
2559 
2560 uint_t
2561 cpuid_get_ncpu_per_chip(cpu_t *cpu)
2562 {
2563 	ASSERT(cpuid_checkpass(cpu, 1));
2564 	return (cpu->cpu_m.mcpu_cpi->cpi_ncpu_per_chip);
2565 }
2566 
2567 uint_t
2568 cpuid_get_ncore_per_chip(cpu_t *cpu)
2569 {
2570 	ASSERT(cpuid_checkpass(cpu, 1));
2571 	return (cpu->cpu_m.mcpu_cpi->cpi_ncore_per_chip);
2572 }
2573 
2574 uint_t
2575 cpuid_get_ncpu_sharing_last_cache(cpu_t *cpu)
2576 {
2577 	ASSERT(cpuid_checkpass(cpu, 2));
2578 	return (cpu->cpu_m.mcpu_cpi->cpi_ncpu_shr_last_cache);
2579 }
2580 
2581 id_t
2582 cpuid_get_last_lvl_cacheid(cpu_t *cpu)
2583 {
2584 	ASSERT(cpuid_checkpass(cpu, 2));
2585 	return (cpu->cpu_m.mcpu_cpi->cpi_last_lvl_cacheid);
2586 }
2587 
2588 uint_t
2589 cpuid_getstep(cpu_t *cpu)
2590 {
2591 	ASSERT(cpuid_checkpass(cpu, 1));
2592 	return (cpu->cpu_m.mcpu_cpi->cpi_step);
2593 }
2594 
2595 uint_t
2596 cpuid_getsig(struct cpu *cpu)
2597 {
2598 	ASSERT(cpuid_checkpass(cpu, 1));
2599 	return (cpu->cpu_m.mcpu_cpi->cpi_std[1].cp_eax);
2600 }
2601 
2602 uint32_t
2603 cpuid_getchiprev(struct cpu *cpu)
2604 {
2605 	ASSERT(cpuid_checkpass(cpu, 1));
2606 	return (cpu->cpu_m.mcpu_cpi->cpi_chiprev);
2607 }
2608 
2609 const char *
2610 cpuid_getchiprevstr(struct cpu *cpu)
2611 {
2612 	ASSERT(cpuid_checkpass(cpu, 1));
2613 	return (cpu->cpu_m.mcpu_cpi->cpi_chiprevstr);
2614 }
2615 
2616 uint32_t
2617 cpuid_getsockettype(struct cpu *cpu)
2618 {
2619 	ASSERT(cpuid_checkpass(cpu, 1));
2620 	return (cpu->cpu_m.mcpu_cpi->cpi_socket);
2621 }
2622 
2623 const char *
2624 cpuid_getsocketstr(cpu_t *cpu)
2625 {
2626 	static const char *socketstr = NULL;
2627 	struct cpuid_info *cpi;
2628 
2629 	ASSERT(cpuid_checkpass(cpu, 1));
2630 	cpi = cpu->cpu_m.mcpu_cpi;
2631 
2632 	/* Assume that socket types are the same across the system */
2633 	if (socketstr == NULL)
2634 		socketstr = _cpuid_sktstr(cpi->cpi_vendor, cpi->cpi_family,
2635 		    cpi->cpi_model, cpi->cpi_step);
2636 
2637 
2638 	return (socketstr);
2639 }
2640 
2641 int
2642 cpuid_get_chipid(cpu_t *cpu)
2643 {
2644 	ASSERT(cpuid_checkpass(cpu, 1));
2645 
2646 	if (cpuid_is_cmt(cpu))
2647 		return (cpu->cpu_m.mcpu_cpi->cpi_chipid);
2648 	return (cpu->cpu_id);
2649 }
2650 
2651 id_t
2652 cpuid_get_coreid(cpu_t *cpu)
2653 {
2654 	ASSERT(cpuid_checkpass(cpu, 1));
2655 	return (cpu->cpu_m.mcpu_cpi->cpi_coreid);
2656 }
2657 
2658 int
2659 cpuid_get_pkgcoreid(cpu_t *cpu)
2660 {
2661 	ASSERT(cpuid_checkpass(cpu, 1));
2662 	return (cpu->cpu_m.mcpu_cpi->cpi_pkgcoreid);
2663 }
2664 
2665 int
2666 cpuid_get_clogid(cpu_t *cpu)
2667 {
2668 	ASSERT(cpuid_checkpass(cpu, 1));
2669 	return (cpu->cpu_m.mcpu_cpi->cpi_clogid);
2670 }
2671 
2672 uint_t
2673 cpuid_get_procnodeid(cpu_t *cpu)
2674 {
2675 	ASSERT(cpuid_checkpass(cpu, 1));
2676 	return (cpu->cpu_m.mcpu_cpi->cpi_procnodeid);
2677 }
2678 
2679 uint_t
2680 cpuid_get_procnodes_per_pkg(cpu_t *cpu)
2681 {
2682 	ASSERT(cpuid_checkpass(cpu, 1));
2683 	return (cpu->cpu_m.mcpu_cpi->cpi_procnodes_per_pkg);
2684 }
2685 
2686 /*ARGSUSED*/
2687 int
2688 cpuid_have_cr8access(cpu_t *cpu)
2689 {
2690 #if defined(__amd64)
2691 	return (1);
2692 #else
2693 	struct cpuid_info *cpi;
2694 
2695 	ASSERT(cpu != NULL);
2696 	cpi = cpu->cpu_m.mcpu_cpi;
2697 	if (cpi->cpi_vendor == X86_VENDOR_AMD && cpi->cpi_maxeax >= 1 &&
2698 	    (CPI_FEATURES_XTD_ECX(cpi) & CPUID_AMD_ECX_CR8D) != 0)
2699 		return (1);
2700 	return (0);
2701 #endif
2702 }
2703 
2704 uint32_t
2705 cpuid_get_apicid(cpu_t *cpu)
2706 {
2707 	ASSERT(cpuid_checkpass(cpu, 1));
2708 	if (cpu->cpu_m.mcpu_cpi->cpi_maxeax < 1) {
2709 		return (UINT32_MAX);
2710 	} else {
2711 		return (cpu->cpu_m.mcpu_cpi->cpi_apicid);
2712 	}
2713 }
2714 
2715 void
2716 cpuid_get_addrsize(cpu_t *cpu, uint_t *pabits, uint_t *vabits)
2717 {
2718 	struct cpuid_info *cpi;
2719 
2720 	if (cpu == NULL)
2721 		cpu = CPU;
2722 	cpi = cpu->cpu_m.mcpu_cpi;
2723 
2724 	ASSERT(cpuid_checkpass(cpu, 1));
2725 
2726 	if (pabits)
2727 		*pabits = cpi->cpi_pabits;
2728 	if (vabits)
2729 		*vabits = cpi->cpi_vabits;
2730 }
2731 
2732 /*
2733  * Returns the number of data TLB entries for a corresponding
2734  * pagesize.  If it can't be computed, or isn't known, the
2735  * routine returns zero.  If you ask about an architecturally
2736  * impossible pagesize, the routine will panic (so that the
2737  * hat implementor knows that things are inconsistent.)
2738  */
2739 uint_t
2740 cpuid_get_dtlb_nent(cpu_t *cpu, size_t pagesize)
2741 {
2742 	struct cpuid_info *cpi;
2743 	uint_t dtlb_nent = 0;
2744 
2745 	if (cpu == NULL)
2746 		cpu = CPU;
2747 	cpi = cpu->cpu_m.mcpu_cpi;
2748 
2749 	ASSERT(cpuid_checkpass(cpu, 1));
2750 
2751 	/*
2752 	 * Check the L2 TLB info
2753 	 */
2754 	if (cpi->cpi_xmaxeax >= 0x80000006) {
2755 		struct cpuid_regs *cp = &cpi->cpi_extd[6];
2756 
2757 		switch (pagesize) {
2758 
2759 		case 4 * 1024:
2760 			/*
2761 			 * All zero in the top 16 bits of the register
2762 			 * indicates a unified TLB. Size is in low 16 bits.
2763 			 */
2764 			if ((cp->cp_ebx & 0xffff0000) == 0)
2765 				dtlb_nent = cp->cp_ebx & 0x0000ffff;
2766 			else
2767 				dtlb_nent = BITX(cp->cp_ebx, 27, 16);
2768 			break;
2769 
2770 		case 2 * 1024 * 1024:
2771 			if ((cp->cp_eax & 0xffff0000) == 0)
2772 				dtlb_nent = cp->cp_eax & 0x0000ffff;
2773 			else
2774 				dtlb_nent = BITX(cp->cp_eax, 27, 16);
2775 			break;
2776 
2777 		default:
2778 			panic("unknown L2 pagesize");
2779 			/*NOTREACHED*/
2780 		}
2781 	}
2782 
2783 	if (dtlb_nent != 0)
2784 		return (dtlb_nent);
2785 
2786 	/*
2787 	 * No L2 TLB support for this size, try L1.
2788 	 */
2789 	if (cpi->cpi_xmaxeax >= 0x80000005) {
2790 		struct cpuid_regs *cp = &cpi->cpi_extd[5];
2791 
2792 		switch (pagesize) {
2793 		case 4 * 1024:
2794 			dtlb_nent = BITX(cp->cp_ebx, 23, 16);
2795 			break;
2796 		case 2 * 1024 * 1024:
2797 			dtlb_nent = BITX(cp->cp_eax, 23, 16);
2798 			break;
2799 		default:
2800 			panic("unknown L1 d-TLB pagesize");
2801 			/*NOTREACHED*/
2802 		}
2803 	}
2804 
2805 	return (dtlb_nent);
2806 }
2807 
2808 /*
2809  * Return 0 if the erratum is not present or not applicable, positive
2810  * if it is, and negative if the status of the erratum is unknown.
2811  *
2812  * See "Revision Guide for AMD Athlon(tm) 64 and AMD Opteron(tm)
2813  * Processors" #25759, Rev 3.57, August 2005
2814  */
2815 int
2816 cpuid_opteron_erratum(cpu_t *cpu, uint_t erratum)
2817 {
2818 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2819 	uint_t eax;
2820 
2821 	/*
2822 	 * Bail out if this CPU isn't an AMD CPU, or if it's
2823 	 * a legacy (32-bit) AMD CPU.
2824 	 */
2825 	if (cpi->cpi_vendor != X86_VENDOR_AMD ||
2826 	    cpi->cpi_family == 4 || cpi->cpi_family == 5 ||
2827 	    cpi->cpi_family == 6)
2828 
2829 		return (0);
2830 
2831 	eax = cpi->cpi_std[1].cp_eax;
2832 
2833 #define	SH_B0(eax)	(eax == 0xf40 || eax == 0xf50)
2834 #define	SH_B3(eax) 	(eax == 0xf51)
2835 #define	B(eax)		(SH_B0(eax) || SH_B3(eax))
2836 
2837 #define	SH_C0(eax)	(eax == 0xf48 || eax == 0xf58)
2838 
2839 #define	SH_CG(eax)	(eax == 0xf4a || eax == 0xf5a || eax == 0xf7a)
2840 #define	DH_CG(eax)	(eax == 0xfc0 || eax == 0xfe0 || eax == 0xff0)
2841 #define	CH_CG(eax)	(eax == 0xf82 || eax == 0xfb2)
2842 #define	CG(eax)		(SH_CG(eax) || DH_CG(eax) || CH_CG(eax))
2843 
2844 #define	SH_D0(eax)	(eax == 0x10f40 || eax == 0x10f50 || eax == 0x10f70)
2845 #define	DH_D0(eax)	(eax == 0x10fc0 || eax == 0x10ff0)
2846 #define	CH_D0(eax)	(eax == 0x10f80 || eax == 0x10fb0)
2847 #define	D0(eax)		(SH_D0(eax) || DH_D0(eax) || CH_D0(eax))
2848 
2849 #define	SH_E0(eax)	(eax == 0x20f50 || eax == 0x20f40 || eax == 0x20f70)
2850 #define	JH_E1(eax)	(eax == 0x20f10)	/* JH8_E0 had 0x20f30 */
2851 #define	DH_E3(eax)	(eax == 0x20fc0 || eax == 0x20ff0)
2852 #define	SH_E4(eax)	(eax == 0x20f51 || eax == 0x20f71)
2853 #define	BH_E4(eax)	(eax == 0x20fb1)
2854 #define	SH_E5(eax)	(eax == 0x20f42)
2855 #define	DH_E6(eax)	(eax == 0x20ff2 || eax == 0x20fc2)
2856 #define	JH_E6(eax)	(eax == 0x20f12 || eax == 0x20f32)
2857 #define	EX(eax)		(SH_E0(eax) || JH_E1(eax) || DH_E3(eax) || \
2858 			    SH_E4(eax) || BH_E4(eax) || SH_E5(eax) || \
2859 			    DH_E6(eax) || JH_E6(eax))
2860 
2861 #define	DR_AX(eax)	(eax == 0x100f00 || eax == 0x100f01 || eax == 0x100f02)
2862 #define	DR_B0(eax)	(eax == 0x100f20)
2863 #define	DR_B1(eax)	(eax == 0x100f21)
2864 #define	DR_BA(eax)	(eax == 0x100f2a)
2865 #define	DR_B2(eax)	(eax == 0x100f22)
2866 #define	DR_B3(eax)	(eax == 0x100f23)
2867 #define	RB_C0(eax)	(eax == 0x100f40)
2868 
2869 	switch (erratum) {
2870 	case 1:
2871 		return (cpi->cpi_family < 0x10);
2872 	case 51:	/* what does the asterisk mean? */
2873 		return (B(eax) || SH_C0(eax) || CG(eax));
2874 	case 52:
2875 		return (B(eax));
2876 	case 57:
2877 		return (cpi->cpi_family <= 0x11);
2878 	case 58:
2879 		return (B(eax));
2880 	case 60:
2881 		return (cpi->cpi_family <= 0x11);
2882 	case 61:
2883 	case 62:
2884 	case 63:
2885 	case 64:
2886 	case 65:
2887 	case 66:
2888 	case 68:
2889 	case 69:
2890 	case 70:
2891 	case 71:
2892 		return (B(eax));
2893 	case 72:
2894 		return (SH_B0(eax));
2895 	case 74:
2896 		return (B(eax));
2897 	case 75:
2898 		return (cpi->cpi_family < 0x10);
2899 	case 76:
2900 		return (B(eax));
2901 	case 77:
2902 		return (cpi->cpi_family <= 0x11);
2903 	case 78:
2904 		return (B(eax) || SH_C0(eax));
2905 	case 79:
2906 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
2907 	case 80:
2908 	case 81:
2909 	case 82:
2910 		return (B(eax));
2911 	case 83:
2912 		return (B(eax) || SH_C0(eax) || CG(eax));
2913 	case 85:
2914 		return (cpi->cpi_family < 0x10);
2915 	case 86:
2916 		return (SH_C0(eax) || CG(eax));
2917 	case 88:
2918 #if !defined(__amd64)
2919 		return (0);
2920 #else
2921 		return (B(eax) || SH_C0(eax));
2922 #endif
2923 	case 89:
2924 		return (cpi->cpi_family < 0x10);
2925 	case 90:
2926 		return (B(eax) || SH_C0(eax) || CG(eax));
2927 	case 91:
2928 	case 92:
2929 		return (B(eax) || SH_C0(eax));
2930 	case 93:
2931 		return (SH_C0(eax));
2932 	case 94:
2933 		return (B(eax) || SH_C0(eax) || CG(eax));
2934 	case 95:
2935 #if !defined(__amd64)
2936 		return (0);
2937 #else
2938 		return (B(eax) || SH_C0(eax));
2939 #endif
2940 	case 96:
2941 		return (B(eax) || SH_C0(eax) || CG(eax));
2942 	case 97:
2943 	case 98:
2944 		return (SH_C0(eax) || CG(eax));
2945 	case 99:
2946 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
2947 	case 100:
2948 		return (B(eax) || SH_C0(eax));
2949 	case 101:
2950 	case 103:
2951 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
2952 	case 104:
2953 		return (SH_C0(eax) || CG(eax) || D0(eax));
2954 	case 105:
2955 	case 106:
2956 	case 107:
2957 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
2958 	case 108:
2959 		return (DH_CG(eax));
2960 	case 109:
2961 		return (SH_C0(eax) || CG(eax) || D0(eax));
2962 	case 110:
2963 		return (D0(eax) || EX(eax));
2964 	case 111:
2965 		return (CG(eax));
2966 	case 112:
2967 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
2968 	case 113:
2969 		return (eax == 0x20fc0);
2970 	case 114:
2971 		return (SH_E0(eax) || JH_E1(eax) || DH_E3(eax));
2972 	case 115:
2973 		return (SH_E0(eax) || JH_E1(eax));
2974 	case 116:
2975 		return (SH_E0(eax) || JH_E1(eax) || DH_E3(eax));
2976 	case 117:
2977 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
2978 	case 118:
2979 		return (SH_E0(eax) || JH_E1(eax) || SH_E4(eax) || BH_E4(eax) ||
2980 		    JH_E6(eax));
2981 	case 121:
2982 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
2983 	case 122:
2984 		return (cpi->cpi_family < 0x10 || cpi->cpi_family == 0x11);
2985 	case 123:
2986 		return (JH_E1(eax) || BH_E4(eax) || JH_E6(eax));
2987 	case 131:
2988 		return (cpi->cpi_family < 0x10);
2989 	case 6336786:
2990 		/*
2991 		 * Test for AdvPowerMgmtInfo.TscPStateInvariant
2992 		 * if this is a K8 family or newer processor
2993 		 */
2994 		if (CPI_FAMILY(cpi) == 0xf) {
2995 			struct cpuid_regs regs;
2996 			regs.cp_eax = 0x80000007;
2997 			(void) __cpuid_insn(&regs);
2998 			return (!(regs.cp_edx & 0x100));
2999 		}
3000 		return (0);
3001 	case 6323525:
3002 		return (((((eax >> 12) & 0xff00) + (eax & 0xf00)) |
3003 		    (((eax >> 4) & 0xf) | ((eax >> 12) & 0xf0))) < 0xf40);
3004 
3005 	case 6671130:
3006 		/*
3007 		 * check for processors (pre-Shanghai) that do not provide
3008 		 * optimal management of 1gb ptes in its tlb.
3009 		 */
3010 		return (cpi->cpi_family == 0x10 && cpi->cpi_model < 4);
3011 
3012 	case 298:
3013 		return (DR_AX(eax) || DR_B0(eax) || DR_B1(eax) || DR_BA(eax) ||
3014 		    DR_B2(eax) || RB_C0(eax));
3015 
3016 	default:
3017 		return (-1);
3018 
3019 	}
3020 }
3021 
3022 /*
3023  * Determine if specified erratum is present via OSVW (OS Visible Workaround).
3024  * Return 1 if erratum is present, 0 if not present and -1 if indeterminate.
3025  */
3026 int
3027 osvw_opteron_erratum(cpu_t *cpu, uint_t erratum)
3028 {
3029 	struct cpuid_info	*cpi;
3030 	uint_t			osvwid;
3031 	static int		osvwfeature = -1;
3032 	uint64_t		osvwlength;
3033 
3034 
3035 	cpi = cpu->cpu_m.mcpu_cpi;
3036 
3037 	/* confirm OSVW supported */
3038 	if (osvwfeature == -1) {
3039 		osvwfeature = cpi->cpi_extd[1].cp_ecx & CPUID_AMD_ECX_OSVW;
3040 	} else {
3041 		/* assert that osvw feature setting is consistent on all cpus */
3042 		ASSERT(osvwfeature ==
3043 		    (cpi->cpi_extd[1].cp_ecx & CPUID_AMD_ECX_OSVW));
3044 	}
3045 	if (!osvwfeature)
3046 		return (-1);
3047 
3048 	osvwlength = rdmsr(MSR_AMD_OSVW_ID_LEN) & OSVW_ID_LEN_MASK;
3049 
3050 	switch (erratum) {
3051 	case 298:	/* osvwid is 0 */
3052 		osvwid = 0;
3053 		if (osvwlength <= (uint64_t)osvwid) {
3054 			/* osvwid 0 is unknown */
3055 			return (-1);
3056 		}
3057 
3058 		/*
3059 		 * Check the OSVW STATUS MSR to determine the state
3060 		 * of the erratum where:
3061 		 *   0 - fixed by HW
3062 		 *   1 - BIOS has applied the workaround when BIOS
3063 		 *   workaround is available. (Or for other errata,
3064 		 *   OS workaround is required.)
3065 		 * For a value of 1, caller will confirm that the
3066 		 * erratum 298 workaround has indeed been applied by BIOS.
3067 		 *
3068 		 * A 1 may be set in cpus that have a HW fix
3069 		 * in a mixed cpu system. Regarding erratum 298:
3070 		 *   In a multiprocessor platform, the workaround above
3071 		 *   should be applied to all processors regardless of
3072 		 *   silicon revision when an affected processor is
3073 		 *   present.
3074 		 */
3075 
3076 		return (rdmsr(MSR_AMD_OSVW_STATUS +
3077 		    (osvwid / OSVW_ID_CNT_PER_MSR)) &
3078 		    (1ULL << (osvwid % OSVW_ID_CNT_PER_MSR)));
3079 
3080 	default:
3081 		return (-1);
3082 	}
3083 }
3084 
3085 static const char assoc_str[] = "associativity";
3086 static const char line_str[] = "line-size";
3087 static const char size_str[] = "size";
3088 
3089 static void
3090 add_cache_prop(dev_info_t *devi, const char *label, const char *type,
3091     uint32_t val)
3092 {
3093 	char buf[128];
3094 
3095 	/*
3096 	 * ndi_prop_update_int() is used because it is desirable for
3097 	 * DDI_PROP_HW_DEF and DDI_PROP_DONTSLEEP to be set.
3098 	 */
3099 	if (snprintf(buf, sizeof (buf), "%s-%s", label, type) < sizeof (buf))
3100 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, devi, buf, val);
3101 }
3102 
3103 /*
3104  * Intel-style cache/tlb description
3105  *
3106  * Standard cpuid level 2 gives a randomly ordered
3107  * selection of tags that index into a table that describes
3108  * cache and tlb properties.
3109  */
3110 
3111 static const char l1_icache_str[] = "l1-icache";
3112 static const char l1_dcache_str[] = "l1-dcache";
3113 static const char l2_cache_str[] = "l2-cache";
3114 static const char l3_cache_str[] = "l3-cache";
3115 static const char itlb4k_str[] = "itlb-4K";
3116 static const char dtlb4k_str[] = "dtlb-4K";
3117 static const char itlb2M_str[] = "itlb-2M";
3118 static const char itlb4M_str[] = "itlb-4M";
3119 static const char dtlb4M_str[] = "dtlb-4M";
3120 static const char dtlb24_str[] = "dtlb0-2M-4M";
3121 static const char itlb424_str[] = "itlb-4K-2M-4M";
3122 static const char itlb24_str[] = "itlb-2M-4M";
3123 static const char dtlb44_str[] = "dtlb-4K-4M";
3124 static const char sl1_dcache_str[] = "sectored-l1-dcache";
3125 static const char sl2_cache_str[] = "sectored-l2-cache";
3126 static const char itrace_str[] = "itrace-cache";
3127 static const char sl3_cache_str[] = "sectored-l3-cache";
3128 static const char sh_l2_tlb4k_str[] = "shared-l2-tlb-4k";
3129 
3130 static const struct cachetab {
3131 	uint8_t 	ct_code;
3132 	uint8_t		ct_assoc;
3133 	uint16_t 	ct_line_size;
3134 	size_t		ct_size;
3135 	const char	*ct_label;
3136 } intel_ctab[] = {
3137 	/*
3138 	 * maintain descending order!
3139 	 *
3140 	 * Codes ignored - Reason
3141 	 * ----------------------
3142 	 * 40H - intel_cpuid_4_cache_info() disambiguates l2/l3 cache
3143 	 * f0H/f1H - Currently we do not interpret prefetch size by design
3144 	 */
3145 	{ 0xe4, 16, 64, 8*1024*1024, l3_cache_str},
3146 	{ 0xe3, 16, 64, 4*1024*1024, l3_cache_str},
3147 	{ 0xe2, 16, 64, 2*1024*1024, l3_cache_str},
3148 	{ 0xde, 12, 64, 6*1024*1024, l3_cache_str},
3149 	{ 0xdd, 12, 64, 3*1024*1024, l3_cache_str},
3150 	{ 0xdc, 12, 64, ((1*1024*1024)+(512*1024)), l3_cache_str},
3151 	{ 0xd8, 8, 64, 4*1024*1024, l3_cache_str},
3152 	{ 0xd7, 8, 64, 2*1024*1024, l3_cache_str},
3153 	{ 0xd6, 8, 64, 1*1024*1024, l3_cache_str},
3154 	{ 0xd2, 4, 64, 2*1024*1024, l3_cache_str},
3155 	{ 0xd1, 4, 64, 1*1024*1024, l3_cache_str},
3156 	{ 0xd0, 4, 64, 512*1024, l3_cache_str},
3157 	{ 0xca, 4, 0, 512, sh_l2_tlb4k_str},
3158 	{ 0xc0, 4, 0, 8, dtlb44_str },
3159 	{ 0xba, 4, 0, 64, dtlb4k_str },
3160 	{ 0xb4, 4, 0, 256, dtlb4k_str },
3161 	{ 0xb3, 4, 0, 128, dtlb4k_str },
3162 	{ 0xb2, 4, 0, 64, itlb4k_str },
3163 	{ 0xb0, 4, 0, 128, itlb4k_str },
3164 	{ 0x87, 8, 64, 1024*1024, l2_cache_str},
3165 	{ 0x86, 4, 64, 512*1024, l2_cache_str},
3166 	{ 0x85, 8, 32, 2*1024*1024, l2_cache_str},
3167 	{ 0x84, 8, 32, 1024*1024, l2_cache_str},
3168 	{ 0x83, 8, 32, 512*1024, l2_cache_str},
3169 	{ 0x82, 8, 32, 256*1024, l2_cache_str},
3170 	{ 0x80, 8, 64, 512*1024, l2_cache_str},
3171 	{ 0x7f, 2, 64, 512*1024, l2_cache_str},
3172 	{ 0x7d, 8, 64, 2*1024*1024, sl2_cache_str},
3173 	{ 0x7c, 8, 64, 1024*1024, sl2_cache_str},
3174 	{ 0x7b, 8, 64, 512*1024, sl2_cache_str},
3175 	{ 0x7a, 8, 64, 256*1024, sl2_cache_str},
3176 	{ 0x79, 8, 64, 128*1024, sl2_cache_str},
3177 	{ 0x78, 8, 64, 1024*1024, l2_cache_str},
3178 	{ 0x73, 8, 0, 64*1024, itrace_str},
3179 	{ 0x72, 8, 0, 32*1024, itrace_str},
3180 	{ 0x71, 8, 0, 16*1024, itrace_str},
3181 	{ 0x70, 8, 0, 12*1024, itrace_str},
3182 	{ 0x68, 4, 64, 32*1024, sl1_dcache_str},
3183 	{ 0x67, 4, 64, 16*1024, sl1_dcache_str},
3184 	{ 0x66, 4, 64, 8*1024, sl1_dcache_str},
3185 	{ 0x60, 8, 64, 16*1024, sl1_dcache_str},
3186 	{ 0x5d, 0, 0, 256, dtlb44_str},
3187 	{ 0x5c, 0, 0, 128, dtlb44_str},
3188 	{ 0x5b, 0, 0, 64, dtlb44_str},
3189 	{ 0x5a, 4, 0, 32, dtlb24_str},
3190 	{ 0x59, 0, 0, 16, dtlb4k_str},
3191 	{ 0x57, 4, 0, 16, dtlb4k_str},
3192 	{ 0x56, 4, 0, 16, dtlb4M_str},
3193 	{ 0x55, 0, 0, 7, itlb24_str},
3194 	{ 0x52, 0, 0, 256, itlb424_str},
3195 	{ 0x51, 0, 0, 128, itlb424_str},
3196 	{ 0x50, 0, 0, 64, itlb424_str},
3197 	{ 0x4f, 0, 0, 32, itlb4k_str},
3198 	{ 0x4e, 24, 64, 6*1024*1024, l2_cache_str},
3199 	{ 0x4d, 16, 64, 16*1024*1024, l3_cache_str},
3200 	{ 0x4c, 12, 64, 12*1024*1024, l3_cache_str},
3201 	{ 0x4b, 16, 64, 8*1024*1024, l3_cache_str},
3202 	{ 0x4a, 12, 64, 6*1024*1024, l3_cache_str},
3203 	{ 0x49, 16, 64, 4*1024*1024, l3_cache_str},
3204 	{ 0x48, 12, 64, 3*1024*1024, l2_cache_str},
3205 	{ 0x47, 8, 64, 8*1024*1024, l3_cache_str},
3206 	{ 0x46, 4, 64, 4*1024*1024, l3_cache_str},
3207 	{ 0x45, 4, 32, 2*1024*1024, l2_cache_str},
3208 	{ 0x44, 4, 32, 1024*1024, l2_cache_str},
3209 	{ 0x43, 4, 32, 512*1024, l2_cache_str},
3210 	{ 0x42, 4, 32, 256*1024, l2_cache_str},
3211 	{ 0x41, 4, 32, 128*1024, l2_cache_str},
3212 	{ 0x3e, 4, 64, 512*1024, sl2_cache_str},
3213 	{ 0x3d, 6, 64, 384*1024, sl2_cache_str},
3214 	{ 0x3c, 4, 64, 256*1024, sl2_cache_str},
3215 	{ 0x3b, 2, 64, 128*1024, sl2_cache_str},
3216 	{ 0x3a, 6, 64, 192*1024, sl2_cache_str},
3217 	{ 0x39, 4, 64, 128*1024, sl2_cache_str},
3218 	{ 0x30, 8, 64, 32*1024, l1_icache_str},
3219 	{ 0x2c, 8, 64, 32*1024, l1_dcache_str},
3220 	{ 0x29, 8, 64, 4096*1024, sl3_cache_str},
3221 	{ 0x25, 8, 64, 2048*1024, sl3_cache_str},
3222 	{ 0x23, 8, 64, 1024*1024, sl3_cache_str},
3223 	{ 0x22, 4, 64, 512*1024, sl3_cache_str},
3224 	{ 0x0e, 6, 64, 24*1024, l1_dcache_str},
3225 	{ 0x0d, 4, 32, 16*1024, l1_dcache_str},
3226 	{ 0x0c, 4, 32, 16*1024, l1_dcache_str},
3227 	{ 0x0b, 4, 0, 4, itlb4M_str},
3228 	{ 0x0a, 2, 32, 8*1024, l1_dcache_str},
3229 	{ 0x08, 4, 32, 16*1024, l1_icache_str},
3230 	{ 0x06, 4, 32, 8*1024, l1_icache_str},
3231 	{ 0x05, 4, 0, 32, dtlb4M_str},
3232 	{ 0x04, 4, 0, 8, dtlb4M_str},
3233 	{ 0x03, 4, 0, 64, dtlb4k_str},
3234 	{ 0x02, 4, 0, 2, itlb4M_str},
3235 	{ 0x01, 4, 0, 32, itlb4k_str},
3236 	{ 0 }
3237 };
3238 
3239 static const struct cachetab cyrix_ctab[] = {
3240 	{ 0x70, 4, 0, 32, "tlb-4K" },
3241 	{ 0x80, 4, 16, 16*1024, "l1-cache" },
3242 	{ 0 }
3243 };
3244 
3245 /*
3246  * Search a cache table for a matching entry
3247  */
3248 static const struct cachetab *
3249 find_cacheent(const struct cachetab *ct, uint_t code)
3250 {
3251 	if (code != 0) {
3252 		for (; ct->ct_code != 0; ct++)
3253 			if (ct->ct_code <= code)
3254 				break;
3255 		if (ct->ct_code == code)
3256 			return (ct);
3257 	}
3258 	return (NULL);
3259 }
3260 
3261 /*
3262  * Populate cachetab entry with L2 or L3 cache-information using
3263  * cpuid function 4. This function is called from intel_walk_cacheinfo()
3264  * when descriptor 0x49 is encountered. It returns 0 if no such cache
3265  * information is found.
3266  */
3267 static int
3268 intel_cpuid_4_cache_info(struct cachetab *ct, struct cpuid_info *cpi)
3269 {
3270 	uint32_t level, i;
3271 	int ret = 0;
3272 
3273 	for (i = 0; i < cpi->cpi_std_4_size; i++) {
3274 		level = CPI_CACHE_LVL(cpi->cpi_std_4[i]);
3275 
3276 		if (level == 2 || level == 3) {
3277 			ct->ct_assoc = CPI_CACHE_WAYS(cpi->cpi_std_4[i]) + 1;
3278 			ct->ct_line_size =
3279 			    CPI_CACHE_COH_LN_SZ(cpi->cpi_std_4[i]) + 1;
3280 			ct->ct_size = ct->ct_assoc *
3281 			    (CPI_CACHE_PARTS(cpi->cpi_std_4[i]) + 1) *
3282 			    ct->ct_line_size *
3283 			    (cpi->cpi_std_4[i]->cp_ecx + 1);
3284 
3285 			if (level == 2) {
3286 				ct->ct_label = l2_cache_str;
3287 			} else if (level == 3) {
3288 				ct->ct_label = l3_cache_str;
3289 			}
3290 			ret = 1;
3291 		}
3292 	}
3293 
3294 	return (ret);
3295 }
3296 
3297 /*
3298  * Walk the cacheinfo descriptor, applying 'func' to every valid element
3299  * The walk is terminated if the walker returns non-zero.
3300  */
3301 static void
3302 intel_walk_cacheinfo(struct cpuid_info *cpi,
3303     void *arg, int (*func)(void *, const struct cachetab *))
3304 {
3305 	const struct cachetab *ct;
3306 	struct cachetab des_49_ct, des_b1_ct;
3307 	uint8_t *dp;
3308 	int i;
3309 
3310 	if ((dp = cpi->cpi_cacheinfo) == NULL)
3311 		return;
3312 	for (i = 0; i < cpi->cpi_ncache; i++, dp++) {
3313 		/*
3314 		 * For overloaded descriptor 0x49 we use cpuid function 4
3315 		 * if supported by the current processor, to create
3316 		 * cache information.
3317 		 * For overloaded descriptor 0xb1 we use X86_PAE flag
3318 		 * to disambiguate the cache information.
3319 		 */
3320 		if (*dp == 0x49 && cpi->cpi_maxeax >= 0x4 &&
3321 		    intel_cpuid_4_cache_info(&des_49_ct, cpi) == 1) {
3322 				ct = &des_49_ct;
3323 		} else if (*dp == 0xb1) {
3324 			des_b1_ct.ct_code = 0xb1;
3325 			des_b1_ct.ct_assoc = 4;
3326 			des_b1_ct.ct_line_size = 0;
3327 			if (x86_feature & X86_PAE) {
3328 				des_b1_ct.ct_size = 8;
3329 				des_b1_ct.ct_label = itlb2M_str;
3330 			} else {
3331 				des_b1_ct.ct_size = 4;
3332 				des_b1_ct.ct_label = itlb4M_str;
3333 			}
3334 			ct = &des_b1_ct;
3335 		} else {
3336 			if ((ct = find_cacheent(intel_ctab, *dp)) == NULL) {
3337 				continue;
3338 			}
3339 		}
3340 
3341 		if (func(arg, ct) != 0) {
3342 			break;
3343 		}
3344 	}
3345 }
3346 
3347 /*
3348  * (Like the Intel one, except for Cyrix CPUs)
3349  */
3350 static void
3351 cyrix_walk_cacheinfo(struct cpuid_info *cpi,
3352     void *arg, int (*func)(void *, const struct cachetab *))
3353 {
3354 	const struct cachetab *ct;
3355 	uint8_t *dp;
3356 	int i;
3357 
3358 	if ((dp = cpi->cpi_cacheinfo) == NULL)
3359 		return;
3360 	for (i = 0; i < cpi->cpi_ncache; i++, dp++) {
3361 		/*
3362 		 * Search Cyrix-specific descriptor table first ..
3363 		 */
3364 		if ((ct = find_cacheent(cyrix_ctab, *dp)) != NULL) {
3365 			if (func(arg, ct) != 0)
3366 				break;
3367 			continue;
3368 		}
3369 		/*
3370 		 * .. else fall back to the Intel one
3371 		 */
3372 		if ((ct = find_cacheent(intel_ctab, *dp)) != NULL) {
3373 			if (func(arg, ct) != 0)
3374 				break;
3375 			continue;
3376 		}
3377 	}
3378 }
3379 
3380 /*
3381  * A cacheinfo walker that adds associativity, line-size, and size properties
3382  * to the devinfo node it is passed as an argument.
3383  */
3384 static int
3385 add_cacheent_props(void *arg, const struct cachetab *ct)
3386 {
3387 	dev_info_t *devi = arg;
3388 
3389 	add_cache_prop(devi, ct->ct_label, assoc_str, ct->ct_assoc);
3390 	if (ct->ct_line_size != 0)
3391 		add_cache_prop(devi, ct->ct_label, line_str,
3392 		    ct->ct_line_size);
3393 	add_cache_prop(devi, ct->ct_label, size_str, ct->ct_size);
3394 	return (0);
3395 }
3396 
3397 
3398 static const char fully_assoc[] = "fully-associative?";
3399 
3400 /*
3401  * AMD style cache/tlb description
3402  *
3403  * Extended functions 5 and 6 directly describe properties of
3404  * tlbs and various cache levels.
3405  */
3406 static void
3407 add_amd_assoc(dev_info_t *devi, const char *label, uint_t assoc)
3408 {
3409 	switch (assoc) {
3410 	case 0:	/* reserved; ignore */
3411 		break;
3412 	default:
3413 		add_cache_prop(devi, label, assoc_str, assoc);
3414 		break;
3415 	case 0xff:
3416 		add_cache_prop(devi, label, fully_assoc, 1);
3417 		break;
3418 	}
3419 }
3420 
3421 static void
3422 add_amd_tlb(dev_info_t *devi, const char *label, uint_t assoc, uint_t size)
3423 {
3424 	if (size == 0)
3425 		return;
3426 	add_cache_prop(devi, label, size_str, size);
3427 	add_amd_assoc(devi, label, assoc);
3428 }
3429 
3430 static void
3431 add_amd_cache(dev_info_t *devi, const char *label,
3432     uint_t size, uint_t assoc, uint_t lines_per_tag, uint_t line_size)
3433 {
3434 	if (size == 0 || line_size == 0)
3435 		return;
3436 	add_amd_assoc(devi, label, assoc);
3437 	/*
3438 	 * Most AMD parts have a sectored cache. Multiple cache lines are
3439 	 * associated with each tag. A sector consists of all cache lines
3440 	 * associated with a tag. For example, the AMD K6-III has a sector
3441 	 * size of 2 cache lines per tag.
3442 	 */
3443 	if (lines_per_tag != 0)
3444 		add_cache_prop(devi, label, "lines-per-tag", lines_per_tag);
3445 	add_cache_prop(devi, label, line_str, line_size);
3446 	add_cache_prop(devi, label, size_str, size * 1024);
3447 }
3448 
3449 static void
3450 add_amd_l2_assoc(dev_info_t *devi, const char *label, uint_t assoc)
3451 {
3452 	switch (assoc) {
3453 	case 0:	/* off */
3454 		break;
3455 	case 1:
3456 	case 2:
3457 	case 4:
3458 		add_cache_prop(devi, label, assoc_str, assoc);
3459 		break;
3460 	case 6:
3461 		add_cache_prop(devi, label, assoc_str, 8);
3462 		break;
3463 	case 8:
3464 		add_cache_prop(devi, label, assoc_str, 16);
3465 		break;
3466 	case 0xf:
3467 		add_cache_prop(devi, label, fully_assoc, 1);
3468 		break;
3469 	default: /* reserved; ignore */
3470 		break;
3471 	}
3472 }
3473 
3474 static void
3475 add_amd_l2_tlb(dev_info_t *devi, const char *label, uint_t assoc, uint_t size)
3476 {
3477 	if (size == 0 || assoc == 0)
3478 		return;
3479 	add_amd_l2_assoc(devi, label, assoc);
3480 	add_cache_prop(devi, label, size_str, size);
3481 }
3482 
3483 static void
3484 add_amd_l2_cache(dev_info_t *devi, const char *label,
3485     uint_t size, uint_t assoc, uint_t lines_per_tag, uint_t line_size)
3486 {
3487 	if (size == 0 || assoc == 0 || line_size == 0)
3488 		return;
3489 	add_amd_l2_assoc(devi, label, assoc);
3490 	if (lines_per_tag != 0)
3491 		add_cache_prop(devi, label, "lines-per-tag", lines_per_tag);
3492 	add_cache_prop(devi, label, line_str, line_size);
3493 	add_cache_prop(devi, label, size_str, size * 1024);
3494 }
3495 
3496 static void
3497 amd_cache_info(struct cpuid_info *cpi, dev_info_t *devi)
3498 {
3499 	struct cpuid_regs *cp;
3500 
3501 	if (cpi->cpi_xmaxeax < 0x80000005)
3502 		return;
3503 	cp = &cpi->cpi_extd[5];
3504 
3505 	/*
3506 	 * 4M/2M L1 TLB configuration
3507 	 *
3508 	 * We report the size for 2M pages because AMD uses two
3509 	 * TLB entries for one 4M page.
3510 	 */
3511 	add_amd_tlb(devi, "dtlb-2M",
3512 	    BITX(cp->cp_eax, 31, 24), BITX(cp->cp_eax, 23, 16));
3513 	add_amd_tlb(devi, "itlb-2M",
3514 	    BITX(cp->cp_eax, 15, 8), BITX(cp->cp_eax, 7, 0));
3515 
3516 	/*
3517 	 * 4K L1 TLB configuration
3518 	 */
3519 
3520 	switch (cpi->cpi_vendor) {
3521 		uint_t nentries;
3522 	case X86_VENDOR_TM:
3523 		if (cpi->cpi_family >= 5) {
3524 			/*
3525 			 * Crusoe processors have 256 TLB entries, but
3526 			 * cpuid data format constrains them to only
3527 			 * reporting 255 of them.
3528 			 */
3529 			if ((nentries = BITX(cp->cp_ebx, 23, 16)) == 255)
3530 				nentries = 256;
3531 			/*
3532 			 * Crusoe processors also have a unified TLB
3533 			 */
3534 			add_amd_tlb(devi, "tlb-4K", BITX(cp->cp_ebx, 31, 24),
3535 			    nentries);
3536 			break;
3537 		}
3538 		/*FALLTHROUGH*/
3539 	default:
3540 		add_amd_tlb(devi, itlb4k_str,
3541 		    BITX(cp->cp_ebx, 31, 24), BITX(cp->cp_ebx, 23, 16));
3542 		add_amd_tlb(devi, dtlb4k_str,
3543 		    BITX(cp->cp_ebx, 15, 8), BITX(cp->cp_ebx, 7, 0));
3544 		break;
3545 	}
3546 
3547 	/*
3548 	 * data L1 cache configuration
3549 	 */
3550 
3551 	add_amd_cache(devi, l1_dcache_str,
3552 	    BITX(cp->cp_ecx, 31, 24), BITX(cp->cp_ecx, 23, 16),
3553 	    BITX(cp->cp_ecx, 15, 8), BITX(cp->cp_ecx, 7, 0));
3554 
3555 	/*
3556 	 * code L1 cache configuration
3557 	 */
3558 
3559 	add_amd_cache(devi, l1_icache_str,
3560 	    BITX(cp->cp_edx, 31, 24), BITX(cp->cp_edx, 23, 16),
3561 	    BITX(cp->cp_edx, 15, 8), BITX(cp->cp_edx, 7, 0));
3562 
3563 	if (cpi->cpi_xmaxeax < 0x80000006)
3564 		return;
3565 	cp = &cpi->cpi_extd[6];
3566 
3567 	/* Check for a unified L2 TLB for large pages */
3568 
3569 	if (BITX(cp->cp_eax, 31, 16) == 0)
3570 		add_amd_l2_tlb(devi, "l2-tlb-2M",
3571 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
3572 	else {
3573 		add_amd_l2_tlb(devi, "l2-dtlb-2M",
3574 		    BITX(cp->cp_eax, 31, 28), BITX(cp->cp_eax, 27, 16));
3575 		add_amd_l2_tlb(devi, "l2-itlb-2M",
3576 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
3577 	}
3578 
3579 	/* Check for a unified L2 TLB for 4K pages */
3580 
3581 	if (BITX(cp->cp_ebx, 31, 16) == 0) {
3582 		add_amd_l2_tlb(devi, "l2-tlb-4K",
3583 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
3584 	} else {
3585 		add_amd_l2_tlb(devi, "l2-dtlb-4K",
3586 		    BITX(cp->cp_eax, 31, 28), BITX(cp->cp_eax, 27, 16));
3587 		add_amd_l2_tlb(devi, "l2-itlb-4K",
3588 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
3589 	}
3590 
3591 	add_amd_l2_cache(devi, l2_cache_str,
3592 	    BITX(cp->cp_ecx, 31, 16), BITX(cp->cp_ecx, 15, 12),
3593 	    BITX(cp->cp_ecx, 11, 8), BITX(cp->cp_ecx, 7, 0));
3594 }
3595 
3596 /*
3597  * There are two basic ways that the x86 world describes it cache
3598  * and tlb architecture - Intel's way and AMD's way.
3599  *
3600  * Return which flavor of cache architecture we should use
3601  */
3602 static int
3603 x86_which_cacheinfo(struct cpuid_info *cpi)
3604 {
3605 	switch (cpi->cpi_vendor) {
3606 	case X86_VENDOR_Intel:
3607 		if (cpi->cpi_maxeax >= 2)
3608 			return (X86_VENDOR_Intel);
3609 		break;
3610 	case X86_VENDOR_AMD:
3611 		/*
3612 		 * The K5 model 1 was the first part from AMD that reported
3613 		 * cache sizes via extended cpuid functions.
3614 		 */
3615 		if (cpi->cpi_family > 5 ||
3616 		    (cpi->cpi_family == 5 && cpi->cpi_model >= 1))
3617 			return (X86_VENDOR_AMD);
3618 		break;
3619 	case X86_VENDOR_TM:
3620 		if (cpi->cpi_family >= 5)
3621 			return (X86_VENDOR_AMD);
3622 		/*FALLTHROUGH*/
3623 	default:
3624 		/*
3625 		 * If they have extended CPU data for 0x80000005
3626 		 * then we assume they have AMD-format cache
3627 		 * information.
3628 		 *
3629 		 * If not, and the vendor happens to be Cyrix,
3630 		 * then try our-Cyrix specific handler.
3631 		 *
3632 		 * If we're not Cyrix, then assume we're using Intel's
3633 		 * table-driven format instead.
3634 		 */
3635 		if (cpi->cpi_xmaxeax >= 0x80000005)
3636 			return (X86_VENDOR_AMD);
3637 		else if (cpi->cpi_vendor == X86_VENDOR_Cyrix)
3638 			return (X86_VENDOR_Cyrix);
3639 		else if (cpi->cpi_maxeax >= 2)
3640 			return (X86_VENDOR_Intel);
3641 		break;
3642 	}
3643 	return (-1);
3644 }
3645 
3646 void
3647 cpuid_set_cpu_properties(void *dip, processorid_t cpu_id,
3648     struct cpuid_info *cpi)
3649 {
3650 	dev_info_t *cpu_devi;
3651 	int create;
3652 
3653 	cpu_devi = (dev_info_t *)dip;
3654 
3655 	/* device_type */
3656 	(void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi,
3657 	    "device_type", "cpu");
3658 
3659 	/* reg */
3660 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3661 	    "reg", cpu_id);
3662 
3663 	/* cpu-mhz, and clock-frequency */
3664 	if (cpu_freq > 0) {
3665 		long long mul;
3666 
3667 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3668 		    "cpu-mhz", cpu_freq);
3669 		if ((mul = cpu_freq * 1000000LL) <= INT_MAX)
3670 			(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3671 			    "clock-frequency", (int)mul);
3672 	}
3673 
3674 	if ((x86_feature & X86_CPUID) == 0) {
3675 		return;
3676 	}
3677 
3678 	/* vendor-id */
3679 	(void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi,
3680 	    "vendor-id", cpi->cpi_vendorstr);
3681 
3682 	if (cpi->cpi_maxeax == 0) {
3683 		return;
3684 	}
3685 
3686 	/*
3687 	 * family, model, and step
3688 	 */
3689 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3690 	    "family", CPI_FAMILY(cpi));
3691 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3692 	    "cpu-model", CPI_MODEL(cpi));
3693 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3694 	    "stepping-id", CPI_STEP(cpi));
3695 
3696 	/* type */
3697 	switch (cpi->cpi_vendor) {
3698 	case X86_VENDOR_Intel:
3699 		create = 1;
3700 		break;
3701 	default:
3702 		create = 0;
3703 		break;
3704 	}
3705 	if (create)
3706 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3707 		    "type", CPI_TYPE(cpi));
3708 
3709 	/* ext-family */
3710 	switch (cpi->cpi_vendor) {
3711 	case X86_VENDOR_Intel:
3712 	case X86_VENDOR_AMD:
3713 		create = cpi->cpi_family >= 0xf;
3714 		break;
3715 	default:
3716 		create = 0;
3717 		break;
3718 	}
3719 	if (create)
3720 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3721 		    "ext-family", CPI_FAMILY_XTD(cpi));
3722 
3723 	/* ext-model */
3724 	switch (cpi->cpi_vendor) {
3725 	case X86_VENDOR_Intel:
3726 		create = IS_EXTENDED_MODEL_INTEL(cpi);
3727 		break;
3728 	case X86_VENDOR_AMD:
3729 		create = CPI_FAMILY(cpi) == 0xf;
3730 		break;
3731 	default:
3732 		create = 0;
3733 		break;
3734 	}
3735 	if (create)
3736 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3737 		    "ext-model", CPI_MODEL_XTD(cpi));
3738 
3739 	/* generation */
3740 	switch (cpi->cpi_vendor) {
3741 	case X86_VENDOR_AMD:
3742 		/*
3743 		 * AMD K5 model 1 was the first part to support this
3744 		 */
3745 		create = cpi->cpi_xmaxeax >= 0x80000001;
3746 		break;
3747 	default:
3748 		create = 0;
3749 		break;
3750 	}
3751 	if (create)
3752 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3753 		    "generation", BITX((cpi)->cpi_extd[1].cp_eax, 11, 8));
3754 
3755 	/* brand-id */
3756 	switch (cpi->cpi_vendor) {
3757 	case X86_VENDOR_Intel:
3758 		/*
3759 		 * brand id first appeared on Pentium III Xeon model 8,
3760 		 * and Celeron model 8 processors and Opteron
3761 		 */
3762 		create = cpi->cpi_family > 6 ||
3763 		    (cpi->cpi_family == 6 && cpi->cpi_model >= 8);
3764 		break;
3765 	case X86_VENDOR_AMD:
3766 		create = cpi->cpi_family >= 0xf;
3767 		break;
3768 	default:
3769 		create = 0;
3770 		break;
3771 	}
3772 	if (create && cpi->cpi_brandid != 0) {
3773 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3774 		    "brand-id", cpi->cpi_brandid);
3775 	}
3776 
3777 	/* chunks, and apic-id */
3778 	switch (cpi->cpi_vendor) {
3779 		/*
3780 		 * first available on Pentium IV and Opteron (K8)
3781 		 */
3782 	case X86_VENDOR_Intel:
3783 		create = IS_NEW_F6(cpi) || cpi->cpi_family >= 0xf;
3784 		break;
3785 	case X86_VENDOR_AMD:
3786 		create = cpi->cpi_family >= 0xf;
3787 		break;
3788 	default:
3789 		create = 0;
3790 		break;
3791 	}
3792 	if (create) {
3793 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3794 		    "chunks", CPI_CHUNKS(cpi));
3795 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3796 		    "apic-id", cpi->cpi_apicid);
3797 		if (cpi->cpi_chipid >= 0) {
3798 			(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3799 			    "chip#", cpi->cpi_chipid);
3800 			(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3801 			    "clog#", cpi->cpi_clogid);
3802 		}
3803 	}
3804 
3805 	/* cpuid-features */
3806 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3807 	    "cpuid-features", CPI_FEATURES_EDX(cpi));
3808 
3809 
3810 	/* cpuid-features-ecx */
3811 	switch (cpi->cpi_vendor) {
3812 	case X86_VENDOR_Intel:
3813 		create = IS_NEW_F6(cpi) || cpi->cpi_family >= 0xf;
3814 		break;
3815 	default:
3816 		create = 0;
3817 		break;
3818 	}
3819 	if (create)
3820 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3821 		    "cpuid-features-ecx", CPI_FEATURES_ECX(cpi));
3822 
3823 	/* ext-cpuid-features */
3824 	switch (cpi->cpi_vendor) {
3825 	case X86_VENDOR_Intel:
3826 	case X86_VENDOR_AMD:
3827 	case X86_VENDOR_Cyrix:
3828 	case X86_VENDOR_TM:
3829 	case X86_VENDOR_Centaur:
3830 		create = cpi->cpi_xmaxeax >= 0x80000001;
3831 		break;
3832 	default:
3833 		create = 0;
3834 		break;
3835 	}
3836 	if (create) {
3837 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3838 		    "ext-cpuid-features", CPI_FEATURES_XTD_EDX(cpi));
3839 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3840 		    "ext-cpuid-features-ecx", CPI_FEATURES_XTD_ECX(cpi));
3841 	}
3842 
3843 	/*
3844 	 * Brand String first appeared in Intel Pentium IV, AMD K5
3845 	 * model 1, and Cyrix GXm.  On earlier models we try and
3846 	 * simulate something similar .. so this string should always
3847 	 * same -something- about the processor, however lame.
3848 	 */
3849 	(void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi,
3850 	    "brand-string", cpi->cpi_brandstr);
3851 
3852 	/*
3853 	 * Finally, cache and tlb information
3854 	 */
3855 	switch (x86_which_cacheinfo(cpi)) {
3856 	case X86_VENDOR_Intel:
3857 		intel_walk_cacheinfo(cpi, cpu_devi, add_cacheent_props);
3858 		break;
3859 	case X86_VENDOR_Cyrix:
3860 		cyrix_walk_cacheinfo(cpi, cpu_devi, add_cacheent_props);
3861 		break;
3862 	case X86_VENDOR_AMD:
3863 		amd_cache_info(cpi, cpu_devi);
3864 		break;
3865 	default:
3866 		break;
3867 	}
3868 }
3869 
3870 struct l2info {
3871 	int *l2i_csz;
3872 	int *l2i_lsz;
3873 	int *l2i_assoc;
3874 	int l2i_ret;
3875 };
3876 
3877 /*
3878  * A cacheinfo walker that fetches the size, line-size and associativity
3879  * of the L2 cache
3880  */
3881 static int
3882 intel_l2cinfo(void *arg, const struct cachetab *ct)
3883 {
3884 	struct l2info *l2i = arg;
3885 	int *ip;
3886 
3887 	if (ct->ct_label != l2_cache_str &&
3888 	    ct->ct_label != sl2_cache_str)
3889 		return (0);	/* not an L2 -- keep walking */
3890 
3891 	if ((ip = l2i->l2i_csz) != NULL)
3892 		*ip = ct->ct_size;
3893 	if ((ip = l2i->l2i_lsz) != NULL)
3894 		*ip = ct->ct_line_size;
3895 	if ((ip = l2i->l2i_assoc) != NULL)
3896 		*ip = ct->ct_assoc;
3897 	l2i->l2i_ret = ct->ct_size;
3898 	return (1);		/* was an L2 -- terminate walk */
3899 }
3900 
3901 /*
3902  * AMD L2/L3 Cache and TLB Associativity Field Definition:
3903  *
3904  *	Unlike the associativity for the L1 cache and tlb where the 8 bit
3905  *	value is the associativity, the associativity for the L2 cache and
3906  *	tlb is encoded in the following table. The 4 bit L2 value serves as
3907  *	an index into the amd_afd[] array to determine the associativity.
3908  *	-1 is undefined. 0 is fully associative.
3909  */
3910 
3911 static int amd_afd[] =
3912 	{-1, 1, 2, -1, 4, -1, 8, -1, 16, -1, 32, 48, 64, 96, 128, 0};
3913 
3914 static void
3915 amd_l2cacheinfo(struct cpuid_info *cpi, struct l2info *l2i)
3916 {
3917 	struct cpuid_regs *cp;
3918 	uint_t size, assoc;
3919 	int i;
3920 	int *ip;
3921 
3922 	if (cpi->cpi_xmaxeax < 0x80000006)
3923 		return;
3924 	cp = &cpi->cpi_extd[6];
3925 
3926 	if ((i = BITX(cp->cp_ecx, 15, 12)) != 0 &&
3927 	    (size = BITX(cp->cp_ecx, 31, 16)) != 0) {
3928 		uint_t cachesz = size * 1024;
3929 		assoc = amd_afd[i];
3930 
3931 		ASSERT(assoc != -1);
3932 
3933 		if ((ip = l2i->l2i_csz) != NULL)
3934 			*ip = cachesz;
3935 		if ((ip = l2i->l2i_lsz) != NULL)
3936 			*ip = BITX(cp->cp_ecx, 7, 0);
3937 		if ((ip = l2i->l2i_assoc) != NULL)
3938 			*ip = assoc;
3939 		l2i->l2i_ret = cachesz;
3940 	}
3941 }
3942 
3943 int
3944 getl2cacheinfo(cpu_t *cpu, int *csz, int *lsz, int *assoc)
3945 {
3946 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
3947 	struct l2info __l2info, *l2i = &__l2info;
3948 
3949 	l2i->l2i_csz = csz;
3950 	l2i->l2i_lsz = lsz;
3951 	l2i->l2i_assoc = assoc;
3952 	l2i->l2i_ret = -1;
3953 
3954 	switch (x86_which_cacheinfo(cpi)) {
3955 	case X86_VENDOR_Intel:
3956 		intel_walk_cacheinfo(cpi, l2i, intel_l2cinfo);
3957 		break;
3958 	case X86_VENDOR_Cyrix:
3959 		cyrix_walk_cacheinfo(cpi, l2i, intel_l2cinfo);
3960 		break;
3961 	case X86_VENDOR_AMD:
3962 		amd_l2cacheinfo(cpi, l2i);
3963 		break;
3964 	default:
3965 		break;
3966 	}
3967 	return (l2i->l2i_ret);
3968 }
3969 
3970 #if !defined(__xpv)
3971 
3972 uint32_t *
3973 cpuid_mwait_alloc(cpu_t *cpu)
3974 {
3975 	uint32_t	*ret;
3976 	size_t		mwait_size;
3977 
3978 	ASSERT(cpuid_checkpass(cpu, 2));
3979 
3980 	mwait_size = cpu->cpu_m.mcpu_cpi->cpi_mwait.mon_max;
3981 	if (mwait_size == 0)
3982 		return (NULL);
3983 
3984 	/*
3985 	 * kmem_alloc() returns cache line size aligned data for mwait_size
3986 	 * allocations.  mwait_size is currently cache line sized.  Neither
3987 	 * of these implementation details are guarantied to be true in the
3988 	 * future.
3989 	 *
3990 	 * First try allocating mwait_size as kmem_alloc() currently returns
3991 	 * correctly aligned memory.  If kmem_alloc() does not return
3992 	 * mwait_size aligned memory, then use mwait_size ROUNDUP.
3993 	 *
3994 	 * Set cpi_mwait.buf_actual and cpi_mwait.size_actual in case we
3995 	 * decide to free this memory.
3996 	 */
3997 	ret = kmem_zalloc(mwait_size, KM_SLEEP);
3998 	if (ret == (uint32_t *)P2ROUNDUP((uintptr_t)ret, mwait_size)) {
3999 		cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual = ret;
4000 		cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual = mwait_size;
4001 		*ret = MWAIT_RUNNING;
4002 		return (ret);
4003 	} else {
4004 		kmem_free(ret, mwait_size);
4005 		ret = kmem_zalloc(mwait_size * 2, KM_SLEEP);
4006 		cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual = ret;
4007 		cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual = mwait_size * 2;
4008 		ret = (uint32_t *)P2ROUNDUP((uintptr_t)ret, mwait_size);
4009 		*ret = MWAIT_RUNNING;
4010 		return (ret);
4011 	}
4012 }
4013 
4014 void
4015 cpuid_mwait_free(cpu_t *cpu)
4016 {
4017 	ASSERT(cpuid_checkpass(cpu, 2));
4018 
4019 	if (cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual != NULL &&
4020 	    cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual > 0) {
4021 		kmem_free(cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual,
4022 		    cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual);
4023 	}
4024 
4025 	cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual = NULL;
4026 	cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual = 0;
4027 }
4028 
4029 void
4030 patch_tsc_read(int flag)
4031 {
4032 	size_t cnt;
4033 
4034 	switch (flag) {
4035 	case X86_NO_TSC:
4036 		cnt = &_no_rdtsc_end - &_no_rdtsc_start;
4037 		(void) memcpy((void *)tsc_read, (void *)&_no_rdtsc_start, cnt);
4038 		break;
4039 	case X86_HAVE_TSCP:
4040 		cnt = &_tscp_end - &_tscp_start;
4041 		(void) memcpy((void *)tsc_read, (void *)&_tscp_start, cnt);
4042 		break;
4043 	case X86_TSC_MFENCE:
4044 		cnt = &_tsc_mfence_end - &_tsc_mfence_start;
4045 		(void) memcpy((void *)tsc_read,
4046 		    (void *)&_tsc_mfence_start, cnt);
4047 		break;
4048 	case X86_TSC_LFENCE:
4049 		cnt = &_tsc_lfence_end - &_tsc_lfence_start;
4050 		(void) memcpy((void *)tsc_read,
4051 		    (void *)&_tsc_lfence_start, cnt);
4052 		break;
4053 	default:
4054 		break;
4055 	}
4056 }
4057 
4058 int
4059 cpuid_deep_cstates_supported(void)
4060 {
4061 	struct cpuid_info *cpi;
4062 	struct cpuid_regs regs;
4063 
4064 	ASSERT(cpuid_checkpass(CPU, 1));
4065 
4066 	cpi = CPU->cpu_m.mcpu_cpi;
4067 
4068 	if (!(x86_feature & X86_CPUID))
4069 		return (0);
4070 
4071 	switch (cpi->cpi_vendor) {
4072 	case X86_VENDOR_Intel:
4073 		if (cpi->cpi_xmaxeax < 0x80000007)
4074 			return (0);
4075 
4076 		/*
4077 		 * TSC run at a constant rate in all ACPI C-states?
4078 		 */
4079 		regs.cp_eax = 0x80000007;
4080 		(void) __cpuid_insn(&regs);
4081 		return (regs.cp_edx & CPUID_TSC_CSTATE_INVARIANCE);
4082 
4083 	default:
4084 		return (0);
4085 	}
4086 }
4087 
4088 #endif	/* !__xpv */
4089 
4090 void
4091 post_startup_cpu_fixups(void)
4092 {
4093 #ifndef __xpv
4094 	/*
4095 	 * Some AMD processors support C1E state. Entering this state will
4096 	 * cause the local APIC timer to stop, which we can't deal with at
4097 	 * this time.
4098 	 */
4099 	if (cpuid_getvendor(CPU) == X86_VENDOR_AMD) {
4100 		on_trap_data_t otd;
4101 		uint64_t reg;
4102 
4103 		if (!on_trap(&otd, OT_DATA_ACCESS)) {
4104 			reg = rdmsr(MSR_AMD_INT_PENDING_CMP_HALT);
4105 			/* Disable C1E state if it is enabled by BIOS */
4106 			if ((reg >> AMD_ACTONCMPHALT_SHIFT) &
4107 			    AMD_ACTONCMPHALT_MASK) {
4108 				reg &= ~(AMD_ACTONCMPHALT_MASK <<
4109 				    AMD_ACTONCMPHALT_SHIFT);
4110 				wrmsr(MSR_AMD_INT_PENDING_CMP_HALT, reg);
4111 			}
4112 		}
4113 		no_trap();
4114 	}
4115 #endif	/* !__xpv */
4116 }
4117 
4118 /*
4119  * Starting with the Westmere processor the local
4120  * APIC timer will continue running in all C-states,
4121  * including the deepest C-states.
4122  */
4123 int
4124 cpuid_arat_supported(void)
4125 {
4126 	struct cpuid_info *cpi;
4127 	struct cpuid_regs regs;
4128 
4129 	ASSERT(cpuid_checkpass(CPU, 1));
4130 	ASSERT(x86_feature & X86_CPUID);
4131 
4132 	cpi = CPU->cpu_m.mcpu_cpi;
4133 
4134 	switch (cpi->cpi_vendor) {
4135 	case X86_VENDOR_Intel:
4136 		/*
4137 		 * Always-running Local APIC Timer is
4138 		 * indicated by CPUID.6.EAX[2].
4139 		 */
4140 		if (cpi->cpi_maxeax >= 6) {
4141 			regs.cp_eax = 6;
4142 			(void) cpuid_insn(NULL, &regs);
4143 			return (regs.cp_eax & CPUID_CSTATE_ARAT);
4144 		} else {
4145 			return (0);
4146 		}
4147 	default:
4148 		return (0);
4149 	}
4150 }
4151 
4152 /*
4153  * Check support for Intel ENERGY_PERF_BIAS feature
4154  */
4155 int
4156 cpuid_iepb_supported(struct cpu *cp)
4157 {
4158 	struct cpuid_info *cpi = cp->cpu_m.mcpu_cpi;
4159 	struct cpuid_regs regs;
4160 
4161 	ASSERT(cpuid_checkpass(cp, 1));
4162 
4163 	if (!(x86_feature & X86_CPUID) || !(x86_feature & X86_MSR)) {
4164 		return (0);
4165 	}
4166 
4167 	/*
4168 	 * Intel ENERGY_PERF_BIAS MSR is indicated by
4169 	 * capability bit CPUID.6.ECX.3
4170 	 */
4171 	if ((cpi->cpi_vendor != X86_VENDOR_Intel) || (cpi->cpi_maxeax < 6))
4172 		return (0);
4173 
4174 	regs.cp_eax = 0x6;
4175 	(void) cpuid_insn(NULL, &regs);
4176 	return (regs.cp_ecx & CPUID_EPB_SUPPORT);
4177 }
4178 
4179 #if defined(__amd64) && !defined(__xpv)
4180 /*
4181  * Patch in versions of bcopy for high performance Intel Nhm processors
4182  * and later...
4183  */
4184 void
4185 patch_memops(uint_t vendor)
4186 {
4187 	size_t cnt, i;
4188 	caddr_t to, from;
4189 
4190 	if ((vendor == X86_VENDOR_Intel) && ((x86_feature & X86_SSE4_2) != 0)) {
4191 		cnt = &bcopy_patch_end - &bcopy_patch_start;
4192 		to = &bcopy_ck_size;
4193 		from = &bcopy_patch_start;
4194 		for (i = 0; i < cnt; i++) {
4195 			*to++ = *from++;
4196 		}
4197 	}
4198 }
4199 #endif  /* __amd64 && !__xpv */
4200