xref: /titanic_51/usr/src/uts/i86pc/os/cpuid.c (revision 72b703890acc1682901e7ab4df40758e3c4399d8)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
23  */
24 /*
25  * Copyright (c) 2009, Intel Corporation.
26  * All rights reserved.
27  */
28 /*
29  * Portions Copyright 2009 Advanced Micro Devices, Inc.
30  */
31 
32 /*
33  * Various routines to handle identification
34  * and classification of x86 processors.
35  */
36 
37 #include <sys/types.h>
38 #include <sys/archsystm.h>
39 #include <sys/x86_archext.h>
40 #include <sys/kmem.h>
41 #include <sys/systm.h>
42 #include <sys/cmn_err.h>
43 #include <sys/sunddi.h>
44 #include <sys/sunndi.h>
45 #include <sys/cpuvar.h>
46 #include <sys/processor.h>
47 #include <sys/sysmacros.h>
48 #include <sys/pg.h>
49 #include <sys/fp.h>
50 #include <sys/controlregs.h>
51 #include <sys/auxv_386.h>
52 #include <sys/bitmap.h>
53 #include <sys/memnode.h>
54 #include <sys/pci_cfgspace.h>
55 
56 #ifdef __xpv
57 #include <sys/hypervisor.h>
58 #else
59 #include <sys/ontrap.h>
60 #endif
61 
62 /*
63  * Pass 0 of cpuid feature analysis happens in locore. It contains special code
64  * to recognize Cyrix processors that are not cpuid-compliant, and to deal with
65  * them accordingly. For most modern processors, feature detection occurs here
66  * in pass 1.
67  *
68  * Pass 1 of cpuid feature analysis happens just at the beginning of mlsetup()
69  * for the boot CPU and does the basic analysis that the early kernel needs.
70  * x86_feature is set based on the return value of cpuid_pass1() of the boot
71  * CPU.
72  *
73  * Pass 1 includes:
74  *
75  *	o Determining vendor/model/family/stepping and setting x86_type and
76  *	  x86_vendor accordingly.
77  *	o Processing the feature flags returned by the cpuid instruction while
78  *	  applying any workarounds or tricks for the specific processor.
79  *	o Mapping the feature flags into Solaris feature bits (X86_*).
80  *	o Processing extended feature flags if supported by the processor,
81  *	  again while applying specific processor knowledge.
82  *	o Determining the CMT characteristics of the system.
83  *
84  * Pass 1 is done on non-boot CPUs during their initialization and the results
85  * are used only as a meager attempt at ensuring that all processors within the
86  * system support the same features.
87  *
88  * Pass 2 of cpuid feature analysis happens just at the beginning
89  * of startup().  It just copies in and corrects the remainder
90  * of the cpuid data we depend on: standard cpuid functions that we didn't
91  * need for pass1 feature analysis, and extended cpuid functions beyond the
92  * simple feature processing done in pass1.
93  *
94  * Pass 3 of cpuid analysis is invoked after basic kernel services; in
95  * particular kernel memory allocation has been made available. It creates a
96  * readable brand string based on the data collected in the first two passes.
97  *
98  * Pass 4 of cpuid analysis is invoked after post_startup() when all
99  * the support infrastructure for various hardware features has been
100  * initialized. It determines which processor features will be reported
101  * to userland via the aux vector.
102  *
103  * All passes are executed on all CPUs, but only the boot CPU determines what
104  * features the kernel will use.
105  *
106  * Much of the worst junk in this file is for the support of processors
107  * that didn't really implement the cpuid instruction properly.
108  *
109  * NOTE: The accessor functions (cpuid_get*) are aware of, and ASSERT upon,
110  * the pass numbers.  Accordingly, changes to the pass code may require changes
111  * to the accessor code.
112  */
113 
114 uint_t x86_feature = 0;
115 uint_t x86_vendor = X86_VENDOR_IntelClone;
116 uint_t x86_type = X86_TYPE_OTHER;
117 uint_t x86_clflush_size = 0;
118 
119 uint_t pentiumpro_bug4046376;
120 uint_t pentiumpro_bug4064495;
121 
122 uint_t enable486;
123 /*
124  * This is set to platform type Solaris is running on.
125  */
126 static int platform_type = -1;
127 
128 #if !defined(__xpv)
129 /*
130  * Variable to patch if hypervisor platform detection needs to be
131  * disabled (e.g. platform_type will always be HW_NATIVE if this is 0).
132  */
133 int enable_platform_detection = 1;
134 #endif
135 
136 /*
137  * monitor/mwait info.
138  *
139  * size_actual and buf_actual are the real address and size allocated to get
140  * proper mwait_buf alignement.  buf_actual and size_actual should be passed
141  * to kmem_free().  Currently kmem_alloc() and mwait happen to both use
142  * processor cache-line alignment, but this is not guarantied in the furture.
143  */
144 struct mwait_info {
145 	size_t		mon_min;	/* min size to avoid missed wakeups */
146 	size_t		mon_max;	/* size to avoid false wakeups */
147 	size_t		size_actual;	/* size actually allocated */
148 	void		*buf_actual;	/* memory actually allocated */
149 	uint32_t	support;	/* processor support of monitor/mwait */
150 };
151 
152 /*
153  * These constants determine how many of the elements of the
154  * cpuid we cache in the cpuid_info data structure; the
155  * remaining elements are accessible via the cpuid instruction.
156  */
157 
158 #define	NMAX_CPI_STD	6		/* eax = 0 .. 5 */
159 #define	NMAX_CPI_EXTD	0x1c		/* eax = 0x80000000 .. 0x8000001b */
160 
161 /*
162  * Some terminology needs to be explained:
163  *  - Socket: Something that can be plugged into a motherboard.
164  *  - Package: Same as socket
165  *  - Chip: Same as socket. Note that AMD's documentation uses term "chip"
166  *    differently: there, chip is the same as processor node (below)
167  *  - Processor node: Some AMD processors have more than one
168  *    "subprocessor" embedded in a package. These subprocessors (nodes)
169  *    are fully-functional processors themselves with cores, caches,
170  *    memory controllers, PCI configuration spaces. They are connected
171  *    inside the package with Hypertransport links. On single-node
172  *    processors, processor node is equivalent to chip/socket/package.
173  */
174 
175 struct cpuid_info {
176 	uint_t cpi_pass;		/* last pass completed */
177 	/*
178 	 * standard function information
179 	 */
180 	uint_t cpi_maxeax;		/* fn 0: %eax */
181 	char cpi_vendorstr[13];		/* fn 0: %ebx:%ecx:%edx */
182 	uint_t cpi_vendor;		/* enum of cpi_vendorstr */
183 
184 	uint_t cpi_family;		/* fn 1: extended family */
185 	uint_t cpi_model;		/* fn 1: extended model */
186 	uint_t cpi_step;		/* fn 1: stepping */
187 	chipid_t cpi_chipid;		/* fn 1: %ebx:  Intel: chip # */
188 					/*		AMD: package/socket # */
189 	uint_t cpi_brandid;		/* fn 1: %ebx: brand ID */
190 	int cpi_clogid;			/* fn 1: %ebx: thread # */
191 	uint_t cpi_ncpu_per_chip;	/* fn 1: %ebx: logical cpu count */
192 	uint8_t cpi_cacheinfo[16];	/* fn 2: intel-style cache desc */
193 	uint_t cpi_ncache;		/* fn 2: number of elements */
194 	uint_t cpi_ncpu_shr_last_cache;	/* fn 4: %eax: ncpus sharing cache */
195 	id_t cpi_last_lvl_cacheid;	/* fn 4: %eax: derived cache id */
196 	uint_t cpi_std_4_size;		/* fn 4: number of fn 4 elements */
197 	struct cpuid_regs **cpi_std_4;	/* fn 4: %ecx == 0 .. fn4_size */
198 	struct cpuid_regs cpi_std[NMAX_CPI_STD];	/* 0 .. 5 */
199 	/*
200 	 * extended function information
201 	 */
202 	uint_t cpi_xmaxeax;		/* fn 0x80000000: %eax */
203 	char cpi_brandstr[49];		/* fn 0x8000000[234] */
204 	uint8_t cpi_pabits;		/* fn 0x80000006: %eax */
205 	uint8_t	cpi_vabits;		/* fn 0x80000006: %eax */
206 	struct	cpuid_regs cpi_extd[NMAX_CPI_EXTD];	/* 0x800000XX */
207 
208 	id_t cpi_coreid;		/* same coreid => strands share core */
209 	int cpi_pkgcoreid;		/* core number within single package */
210 	uint_t cpi_ncore_per_chip;	/* AMD: fn 0x80000008: %ecx[7-0] */
211 					/* Intel: fn 4: %eax[31-26] */
212 	/*
213 	 * supported feature information
214 	 */
215 	uint32_t cpi_support[5];
216 #define	STD_EDX_FEATURES	0
217 #define	AMD_EDX_FEATURES	1
218 #define	TM_EDX_FEATURES		2
219 #define	STD_ECX_FEATURES	3
220 #define	AMD_ECX_FEATURES	4
221 	/*
222 	 * Synthesized information, where known.
223 	 */
224 	uint32_t cpi_chiprev;		/* See X86_CHIPREV_* in x86_archext.h */
225 	const char *cpi_chiprevstr;	/* May be NULL if chiprev unknown */
226 	uint32_t cpi_socket;		/* Chip package/socket type */
227 
228 	struct mwait_info cpi_mwait;	/* fn 5: monitor/mwait info */
229 	uint32_t cpi_apicid;
230 	uint_t cpi_procnodeid;		/* AMD: nodeID on HT, Intel: chipid */
231 	uint_t cpi_procnodes_per_pkg;	/* AMD: # of nodes in the package */
232 					/* Intel: 1 */
233 };
234 
235 
236 static struct cpuid_info cpuid_info0;
237 
238 /*
239  * These bit fields are defined by the Intel Application Note AP-485
240  * "Intel Processor Identification and the CPUID Instruction"
241  */
242 #define	CPI_FAMILY_XTD(cpi)	BITX((cpi)->cpi_std[1].cp_eax, 27, 20)
243 #define	CPI_MODEL_XTD(cpi)	BITX((cpi)->cpi_std[1].cp_eax, 19, 16)
244 #define	CPI_TYPE(cpi)		BITX((cpi)->cpi_std[1].cp_eax, 13, 12)
245 #define	CPI_FAMILY(cpi)		BITX((cpi)->cpi_std[1].cp_eax, 11, 8)
246 #define	CPI_STEP(cpi)		BITX((cpi)->cpi_std[1].cp_eax, 3, 0)
247 #define	CPI_MODEL(cpi)		BITX((cpi)->cpi_std[1].cp_eax, 7, 4)
248 
249 #define	CPI_FEATURES_EDX(cpi)		((cpi)->cpi_std[1].cp_edx)
250 #define	CPI_FEATURES_ECX(cpi)		((cpi)->cpi_std[1].cp_ecx)
251 #define	CPI_FEATURES_XTD_EDX(cpi)	((cpi)->cpi_extd[1].cp_edx)
252 #define	CPI_FEATURES_XTD_ECX(cpi)	((cpi)->cpi_extd[1].cp_ecx)
253 
254 #define	CPI_BRANDID(cpi)	BITX((cpi)->cpi_std[1].cp_ebx, 7, 0)
255 #define	CPI_CHUNKS(cpi)		BITX((cpi)->cpi_std[1].cp_ebx, 15, 7)
256 #define	CPI_CPU_COUNT(cpi)	BITX((cpi)->cpi_std[1].cp_ebx, 23, 16)
257 #define	CPI_APIC_ID(cpi)	BITX((cpi)->cpi_std[1].cp_ebx, 31, 24)
258 
259 #define	CPI_MAXEAX_MAX		0x100		/* sanity control */
260 #define	CPI_XMAXEAX_MAX		0x80000100
261 #define	CPI_FN4_ECX_MAX		0x20		/* sanity: max fn 4 levels */
262 #define	CPI_FNB_ECX_MAX		0x20		/* sanity: max fn B levels */
263 
264 /*
265  * Function 4 (Deterministic Cache Parameters) macros
266  * Defined by Intel Application Note AP-485
267  */
268 #define	CPI_NUM_CORES(regs)		BITX((regs)->cp_eax, 31, 26)
269 #define	CPI_NTHR_SHR_CACHE(regs)	BITX((regs)->cp_eax, 25, 14)
270 #define	CPI_FULL_ASSOC_CACHE(regs)	BITX((regs)->cp_eax, 9, 9)
271 #define	CPI_SELF_INIT_CACHE(regs)	BITX((regs)->cp_eax, 8, 8)
272 #define	CPI_CACHE_LVL(regs)		BITX((regs)->cp_eax, 7, 5)
273 #define	CPI_CACHE_TYPE(regs)		BITX((regs)->cp_eax, 4, 0)
274 #define	CPI_CPU_LEVEL_TYPE(regs)	BITX((regs)->cp_ecx, 15, 8)
275 
276 #define	CPI_CACHE_WAYS(regs)		BITX((regs)->cp_ebx, 31, 22)
277 #define	CPI_CACHE_PARTS(regs)		BITX((regs)->cp_ebx, 21, 12)
278 #define	CPI_CACHE_COH_LN_SZ(regs)	BITX((regs)->cp_ebx, 11, 0)
279 
280 #define	CPI_CACHE_SETS(regs)		BITX((regs)->cp_ecx, 31, 0)
281 
282 #define	CPI_PREFCH_STRIDE(regs)		BITX((regs)->cp_edx, 9, 0)
283 
284 
285 /*
286  * A couple of shorthand macros to identify "later" P6-family chips
287  * like the Pentium M and Core.  First, the "older" P6-based stuff
288  * (loosely defined as "pre-Pentium-4"):
289  * P6, PII, Mobile PII, PII Xeon, PIII, Mobile PIII, PIII Xeon
290  */
291 
292 #define	IS_LEGACY_P6(cpi) (			\
293 	cpi->cpi_family == 6 && 		\
294 		(cpi->cpi_model == 1 ||		\
295 		cpi->cpi_model == 3 ||		\
296 		cpi->cpi_model == 5 ||		\
297 		cpi->cpi_model == 6 ||		\
298 		cpi->cpi_model == 7 ||		\
299 		cpi->cpi_model == 8 ||		\
300 		cpi->cpi_model == 0xA ||	\
301 		cpi->cpi_model == 0xB)		\
302 )
303 
304 /* A "new F6" is everything with family 6 that's not the above */
305 #define	IS_NEW_F6(cpi) ((cpi->cpi_family == 6) && !IS_LEGACY_P6(cpi))
306 
307 /* Extended family/model support */
308 #define	IS_EXTENDED_MODEL_INTEL(cpi) (cpi->cpi_family == 0x6 || \
309 	cpi->cpi_family >= 0xf)
310 
311 /*
312  * Info for monitor/mwait idle loop.
313  *
314  * See cpuid section of "Intel 64 and IA-32 Architectures Software Developer's
315  * Manual Volume 2A: Instruction Set Reference, A-M" #25366-022US, November
316  * 2006.
317  * See MONITOR/MWAIT section of "AMD64 Architecture Programmer's Manual
318  * Documentation Updates" #33633, Rev 2.05, December 2006.
319  */
320 #define	MWAIT_SUPPORT		(0x00000001)	/* mwait supported */
321 #define	MWAIT_EXTENSIONS	(0x00000002)	/* extenstion supported */
322 #define	MWAIT_ECX_INT_ENABLE	(0x00000004)	/* ecx 1 extension supported */
323 #define	MWAIT_SUPPORTED(cpi)	((cpi)->cpi_std[1].cp_ecx & CPUID_INTC_ECX_MON)
324 #define	MWAIT_INT_ENABLE(cpi)	((cpi)->cpi_std[5].cp_ecx & 0x2)
325 #define	MWAIT_EXTENSION(cpi)	((cpi)->cpi_std[5].cp_ecx & 0x1)
326 #define	MWAIT_SIZE_MIN(cpi)	BITX((cpi)->cpi_std[5].cp_eax, 15, 0)
327 #define	MWAIT_SIZE_MAX(cpi)	BITX((cpi)->cpi_std[5].cp_ebx, 15, 0)
328 /*
329  * Number of sub-cstates for a given c-state.
330  */
331 #define	MWAIT_NUM_SUBC_STATES(cpi, c_state)			\
332 	BITX((cpi)->cpi_std[5].cp_edx, c_state + 3, c_state)
333 
334 /*
335  * Functions we consune from cpuid_subr.c;  don't publish these in a header
336  * file to try and keep people using the expected cpuid_* interfaces.
337  */
338 extern uint32_t _cpuid_skt(uint_t, uint_t, uint_t, uint_t);
339 extern const char *_cpuid_sktstr(uint_t, uint_t, uint_t, uint_t);
340 extern uint32_t _cpuid_chiprev(uint_t, uint_t, uint_t, uint_t);
341 extern const char *_cpuid_chiprevstr(uint_t, uint_t, uint_t, uint_t);
342 extern uint_t _cpuid_vendorstr_to_vendorcode(char *);
343 
344 /*
345  * Apply up various platform-dependent restrictions where the
346  * underlying platform restrictions mean the CPU can be marked
347  * as less capable than its cpuid instruction would imply.
348  */
349 #if defined(__xpv)
350 static void
351 platform_cpuid_mangle(uint_t vendor, uint32_t eax, struct cpuid_regs *cp)
352 {
353 	switch (eax) {
354 	case 1: {
355 		uint32_t mcamask = DOMAIN_IS_INITDOMAIN(xen_info) ?
356 		    0 : CPUID_INTC_EDX_MCA;
357 		cp->cp_edx &=
358 		    ~(mcamask |
359 		    CPUID_INTC_EDX_PSE |
360 		    CPUID_INTC_EDX_VME | CPUID_INTC_EDX_DE |
361 		    CPUID_INTC_EDX_SEP | CPUID_INTC_EDX_MTRR |
362 		    CPUID_INTC_EDX_PGE | CPUID_INTC_EDX_PAT |
363 		    CPUID_AMD_EDX_SYSC | CPUID_INTC_EDX_SEP |
364 		    CPUID_INTC_EDX_PSE36 | CPUID_INTC_EDX_HTT);
365 		break;
366 	}
367 
368 	case 0x80000001:
369 		cp->cp_edx &=
370 		    ~(CPUID_AMD_EDX_PSE |
371 		    CPUID_INTC_EDX_VME | CPUID_INTC_EDX_DE |
372 		    CPUID_AMD_EDX_MTRR | CPUID_AMD_EDX_PGE |
373 		    CPUID_AMD_EDX_PAT | CPUID_AMD_EDX_PSE36 |
374 		    CPUID_AMD_EDX_SYSC | CPUID_INTC_EDX_SEP |
375 		    CPUID_AMD_EDX_TSCP);
376 		cp->cp_ecx &= ~CPUID_AMD_ECX_CMP_LGCY;
377 		break;
378 	default:
379 		break;
380 	}
381 
382 	switch (vendor) {
383 	case X86_VENDOR_Intel:
384 		switch (eax) {
385 		case 4:
386 			/*
387 			 * Zero out the (ncores-per-chip - 1) field
388 			 */
389 			cp->cp_eax &= 0x03fffffff;
390 			break;
391 		default:
392 			break;
393 		}
394 		break;
395 	case X86_VENDOR_AMD:
396 		switch (eax) {
397 
398 		case 0x80000001:
399 			cp->cp_ecx &= ~CPUID_AMD_ECX_CR8D;
400 			break;
401 
402 		case 0x80000008:
403 			/*
404 			 * Zero out the (ncores-per-chip - 1) field
405 			 */
406 			cp->cp_ecx &= 0xffffff00;
407 			break;
408 		default:
409 			break;
410 		}
411 		break;
412 	default:
413 		break;
414 	}
415 }
416 #else
417 #define	platform_cpuid_mangle(vendor, eax, cp)	/* nothing */
418 #endif
419 
420 /*
421  *  Some undocumented ways of patching the results of the cpuid
422  *  instruction to permit running Solaris 10 on future cpus that
423  *  we don't currently support.  Could be set to non-zero values
424  *  via settings in eeprom.
425  */
426 
427 uint32_t cpuid_feature_ecx_include;
428 uint32_t cpuid_feature_ecx_exclude;
429 uint32_t cpuid_feature_edx_include;
430 uint32_t cpuid_feature_edx_exclude;
431 
432 /*
433  * Allocate space for mcpu_cpi in the machcpu structure for all non-boot CPUs.
434  */
435 void
436 cpuid_alloc_space(cpu_t *cpu)
437 {
438 	/*
439 	 * By convention, cpu0 is the boot cpu, which is set up
440 	 * before memory allocation is available.  All other cpus get
441 	 * their cpuid_info struct allocated here.
442 	 */
443 	ASSERT(cpu->cpu_id != 0);
444 	ASSERT(cpu->cpu_m.mcpu_cpi == NULL);
445 	cpu->cpu_m.mcpu_cpi =
446 	    kmem_zalloc(sizeof (*cpu->cpu_m.mcpu_cpi), KM_SLEEP);
447 }
448 
449 void
450 cpuid_free_space(cpu_t *cpu)
451 {
452 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
453 	int i;
454 
455 	ASSERT(cpi != NULL);
456 	ASSERT(cpi != &cpuid_info0);
457 
458 	/*
459 	 * Free up any function 4 related dynamic storage
460 	 */
461 	for (i = 1; i < cpi->cpi_std_4_size; i++)
462 		kmem_free(cpi->cpi_std_4[i], sizeof (struct cpuid_regs));
463 	if (cpi->cpi_std_4_size > 0)
464 		kmem_free(cpi->cpi_std_4,
465 		    cpi->cpi_std_4_size * sizeof (struct cpuid_regs *));
466 
467 	kmem_free(cpi, sizeof (*cpi));
468 	cpu->cpu_m.mcpu_cpi = NULL;
469 }
470 
471 #if !defined(__xpv)
472 
473 static void
474 determine_platform()
475 {
476 	struct cpuid_regs cp;
477 	char *xen_str;
478 	uint32_t xen_signature[4], base;
479 
480 	platform_type = HW_NATIVE;
481 
482 	if (!enable_platform_detection)
483 		return;
484 
485 	/*
486 	 * In a fully virtualized domain, Xen's pseudo-cpuid function
487 	 * returns a string representing the Xen signature in %ebx, %ecx,
488 	 * and %edx. %eax contains the maximum supported cpuid function.
489 	 * We need at least a (base + 2) leaf value to do what we want
490 	 * to do. Try different base values, since the hypervisor might
491 	 * use a different one depending on whether hyper-v emulation
492 	 * is switched on by default or not.
493 	 */
494 	for (base = 0x40000000; base < 0x40010000; base += 0x100) {
495 		cp.cp_eax = base;
496 		(void) __cpuid_insn(&cp);
497 		xen_signature[0] = cp.cp_ebx;
498 		xen_signature[1] = cp.cp_ecx;
499 		xen_signature[2] = cp.cp_edx;
500 		xen_signature[3] = 0;
501 		xen_str = (char *)xen_signature;
502 		if (strcmp("XenVMMXenVMM", xen_str) == 0 &&
503 		    cp.cp_eax >= (base + 2)) {
504 			platform_type = HW_XEN_HVM;
505 			return;
506 		}
507 	}
508 
509 	if (vmware_platform()) /* running under vmware hypervisor? */
510 		platform_type = HW_VMWARE;
511 }
512 
513 int
514 get_hwenv(void)
515 {
516 	if (platform_type == -1)
517 		determine_platform();
518 
519 	return (platform_type);
520 }
521 
522 int
523 is_controldom(void)
524 {
525 	return (0);
526 }
527 
528 #else
529 
530 int
531 get_hwenv(void)
532 {
533 	return (HW_XEN_PV);
534 }
535 
536 int
537 is_controldom(void)
538 {
539 	return (DOMAIN_IS_INITDOMAIN(xen_info));
540 }
541 
542 #endif	/* __xpv */
543 
544 static void
545 cpuid_intel_getids(cpu_t *cpu, uint_t feature)
546 {
547 	uint_t i;
548 	uint_t chipid_shift = 0;
549 	uint_t coreid_shift = 0;
550 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
551 
552 	for (i = 1; i < cpi->cpi_ncpu_per_chip; i <<= 1)
553 		chipid_shift++;
554 
555 	cpi->cpi_chipid = cpi->cpi_apicid >> chipid_shift;
556 	cpi->cpi_clogid = cpi->cpi_apicid & ((1 << chipid_shift) - 1);
557 
558 	if (feature & X86_CMP) {
559 		/*
560 		 * Multi-core (and possibly multi-threaded)
561 		 * processors.
562 		 */
563 		uint_t ncpu_per_core;
564 		if (cpi->cpi_ncore_per_chip == 1)
565 			ncpu_per_core = cpi->cpi_ncpu_per_chip;
566 		else if (cpi->cpi_ncore_per_chip > 1)
567 			ncpu_per_core = cpi->cpi_ncpu_per_chip /
568 			    cpi->cpi_ncore_per_chip;
569 		/*
570 		 * 8bit APIC IDs on dual core Pentiums
571 		 * look like this:
572 		 *
573 		 * +-----------------------+------+------+
574 		 * | Physical Package ID   |  MC  |  HT  |
575 		 * +-----------------------+------+------+
576 		 * <------- chipid -------->
577 		 * <------- coreid --------------->
578 		 *			   <--- clogid -->
579 		 *			   <------>
580 		 *			   pkgcoreid
581 		 *
582 		 * Where the number of bits necessary to
583 		 * represent MC and HT fields together equals
584 		 * to the minimum number of bits necessary to
585 		 * store the value of cpi->cpi_ncpu_per_chip.
586 		 * Of those bits, the MC part uses the number
587 		 * of bits necessary to store the value of
588 		 * cpi->cpi_ncore_per_chip.
589 		 */
590 		for (i = 1; i < ncpu_per_core; i <<= 1)
591 			coreid_shift++;
592 		cpi->cpi_coreid = cpi->cpi_apicid >> coreid_shift;
593 		cpi->cpi_pkgcoreid = cpi->cpi_clogid >> coreid_shift;
594 	} else if (feature & X86_HTT) {
595 		/*
596 		 * Single-core multi-threaded processors.
597 		 */
598 		cpi->cpi_coreid = cpi->cpi_chipid;
599 		cpi->cpi_pkgcoreid = 0;
600 	}
601 	cpi->cpi_procnodeid = cpi->cpi_chipid;
602 }
603 
604 static void
605 cpuid_amd_getids(cpu_t *cpu)
606 {
607 	int i, first_half, coreidsz;
608 	uint32_t nb_caps_reg;
609 	uint_t node2_1;
610 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
611 
612 	/*
613 	 * AMD CMP chips currently have a single thread per core.
614 	 *
615 	 * Since no two cpus share a core we must assign a distinct coreid
616 	 * per cpu, and we do this by using the cpu_id.  This scheme does not,
617 	 * however, guarantee that sibling cores of a chip will have sequential
618 	 * coreids starting at a multiple of the number of cores per chip -
619 	 * that is usually the case, but if the ACPI MADT table is presented
620 	 * in a different order then we need to perform a few more gymnastics
621 	 * for the pkgcoreid.
622 	 *
623 	 * All processors in the system have the same number of enabled
624 	 * cores. Cores within a processor are always numbered sequentially
625 	 * from 0 regardless of how many or which are disabled, and there
626 	 * is no way for operating system to discover the real core id when some
627 	 * are disabled.
628 	 */
629 
630 	cpi->cpi_coreid = cpu->cpu_id;
631 
632 	if (cpi->cpi_xmaxeax >= 0x80000008) {
633 
634 		coreidsz = BITX((cpi)->cpi_extd[8].cp_ecx, 15, 12);
635 
636 		/*
637 		 * In AMD parlance chip is really a node while Solaris
638 		 * sees chip as equivalent to socket/package.
639 		 */
640 		cpi->cpi_ncore_per_chip =
641 		    BITX((cpi)->cpi_extd[8].cp_ecx, 7, 0) + 1;
642 		if (coreidsz == 0) {
643 			/* Use legacy method */
644 			for (i = 1; i < cpi->cpi_ncore_per_chip; i <<= 1)
645 				coreidsz++;
646 			if (coreidsz == 0)
647 				coreidsz = 1;
648 		}
649 	} else {
650 		/* Assume single-core part */
651 		cpi->cpi_ncore_per_chip = 1;
652 		coreidsz = 1;
653 	}
654 
655 	cpi->cpi_clogid = cpi->cpi_pkgcoreid =
656 	    cpi->cpi_apicid & ((1<<coreidsz) - 1);
657 	cpi->cpi_ncpu_per_chip = cpi->cpi_ncore_per_chip;
658 
659 	/* Get nodeID */
660 	if (cpi->cpi_family == 0xf) {
661 		cpi->cpi_procnodeid = (cpi->cpi_apicid >> coreidsz) & 7;
662 		cpi->cpi_chipid = cpi->cpi_procnodeid;
663 	} else if (cpi->cpi_family == 0x10) {
664 		/*
665 		 * See if we are a multi-node processor.
666 		 * All processors in the system have the same number of nodes
667 		 */
668 		nb_caps_reg =  pci_getl_func(0, 24, 3, 0xe8);
669 		if ((cpi->cpi_model < 8) || BITX(nb_caps_reg, 29, 29) == 0) {
670 			/* Single-node */
671 			cpi->cpi_procnodeid = BITX(cpi->cpi_apicid, 5,
672 			    coreidsz);
673 			cpi->cpi_chipid = cpi->cpi_procnodeid;
674 		} else {
675 
676 			/*
677 			 * Multi-node revision D (2 nodes per package
678 			 * are supported)
679 			 */
680 			cpi->cpi_procnodes_per_pkg = 2;
681 
682 			first_half = (cpi->cpi_pkgcoreid <=
683 			    (cpi->cpi_ncore_per_chip/2 - 1));
684 
685 			if (cpi->cpi_apicid == cpi->cpi_pkgcoreid) {
686 				/* We are BSP */
687 				cpi->cpi_procnodeid = (first_half ? 0 : 1);
688 				cpi->cpi_chipid = cpi->cpi_procnodeid >> 1;
689 			} else {
690 
691 				/* We are AP */
692 				/* NodeId[2:1] bits to use for reading F3xe8 */
693 				node2_1 = BITX(cpi->cpi_apicid, 5, 4) << 1;
694 
695 				nb_caps_reg =
696 				    pci_getl_func(0, 24 + node2_1, 3, 0xe8);
697 
698 				/*
699 				 * Check IntNodeNum bit (31:30, but bit 31 is
700 				 * always 0 on dual-node processors)
701 				 */
702 				if (BITX(nb_caps_reg, 30, 30) == 0)
703 					cpi->cpi_procnodeid = node2_1 +
704 					    !first_half;
705 				else
706 					cpi->cpi_procnodeid = node2_1 +
707 					    first_half;
708 
709 				cpi->cpi_chipid = cpi->cpi_procnodeid >> 1;
710 			}
711 		}
712 	} else if (cpi->cpi_family >= 0x11) {
713 		cpi->cpi_procnodeid = (cpi->cpi_apicid >> coreidsz) & 7;
714 		cpi->cpi_chipid = cpi->cpi_procnodeid;
715 	} else {
716 		cpi->cpi_procnodeid = 0;
717 		cpi->cpi_chipid = cpi->cpi_procnodeid;
718 	}
719 }
720 
721 uint_t
722 cpuid_pass1(cpu_t *cpu)
723 {
724 	uint32_t mask_ecx, mask_edx;
725 	uint_t feature = X86_CPUID;
726 	struct cpuid_info *cpi;
727 	struct cpuid_regs *cp;
728 	int xcpuid;
729 #if !defined(__xpv)
730 	extern int idle_cpu_prefer_mwait;
731 #endif
732 
733 
734 #if !defined(__xpv)
735 	determine_platform();
736 #endif
737 	/*
738 	 * Space statically allocated for BSP, ensure pointer is set
739 	 */
740 	if (cpu->cpu_id == 0 && cpu->cpu_m.mcpu_cpi == NULL)
741 		cpu->cpu_m.mcpu_cpi = &cpuid_info0;
742 	cpi = cpu->cpu_m.mcpu_cpi;
743 	ASSERT(cpi != NULL);
744 	cp = &cpi->cpi_std[0];
745 	cp->cp_eax = 0;
746 	cpi->cpi_maxeax = __cpuid_insn(cp);
747 	{
748 		uint32_t *iptr = (uint32_t *)cpi->cpi_vendorstr;
749 		*iptr++ = cp->cp_ebx;
750 		*iptr++ = cp->cp_edx;
751 		*iptr++ = cp->cp_ecx;
752 		*(char *)&cpi->cpi_vendorstr[12] = '\0';
753 	}
754 
755 	cpi->cpi_vendor = _cpuid_vendorstr_to_vendorcode(cpi->cpi_vendorstr);
756 	x86_vendor = cpi->cpi_vendor; /* for compatibility */
757 
758 	/*
759 	 * Limit the range in case of weird hardware
760 	 */
761 	if (cpi->cpi_maxeax > CPI_MAXEAX_MAX)
762 		cpi->cpi_maxeax = CPI_MAXEAX_MAX;
763 	if (cpi->cpi_maxeax < 1)
764 		goto pass1_done;
765 
766 	cp = &cpi->cpi_std[1];
767 	cp->cp_eax = 1;
768 	(void) __cpuid_insn(cp);
769 
770 	/*
771 	 * Extract identifying constants for easy access.
772 	 */
773 	cpi->cpi_model = CPI_MODEL(cpi);
774 	cpi->cpi_family = CPI_FAMILY(cpi);
775 
776 	if (cpi->cpi_family == 0xf)
777 		cpi->cpi_family += CPI_FAMILY_XTD(cpi);
778 
779 	/*
780 	 * Beware: AMD uses "extended model" iff base *FAMILY* == 0xf.
781 	 * Intel, and presumably everyone else, uses model == 0xf, as
782 	 * one would expect (max value means possible overflow).  Sigh.
783 	 */
784 
785 	switch (cpi->cpi_vendor) {
786 	case X86_VENDOR_Intel:
787 		if (IS_EXTENDED_MODEL_INTEL(cpi))
788 			cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4;
789 		break;
790 	case X86_VENDOR_AMD:
791 		if (CPI_FAMILY(cpi) == 0xf)
792 			cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4;
793 		break;
794 	default:
795 		if (cpi->cpi_model == 0xf)
796 			cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4;
797 		break;
798 	}
799 
800 	cpi->cpi_step = CPI_STEP(cpi);
801 	cpi->cpi_brandid = CPI_BRANDID(cpi);
802 
803 	/*
804 	 * *default* assumptions:
805 	 * - believe %edx feature word
806 	 * - ignore %ecx feature word
807 	 * - 32-bit virtual and physical addressing
808 	 */
809 	mask_edx = 0xffffffff;
810 	mask_ecx = 0;
811 
812 	cpi->cpi_pabits = cpi->cpi_vabits = 32;
813 
814 	switch (cpi->cpi_vendor) {
815 	case X86_VENDOR_Intel:
816 		if (cpi->cpi_family == 5)
817 			x86_type = X86_TYPE_P5;
818 		else if (IS_LEGACY_P6(cpi)) {
819 			x86_type = X86_TYPE_P6;
820 			pentiumpro_bug4046376 = 1;
821 			pentiumpro_bug4064495 = 1;
822 			/*
823 			 * Clear the SEP bit when it was set erroneously
824 			 */
825 			if (cpi->cpi_model < 3 && cpi->cpi_step < 3)
826 				cp->cp_edx &= ~CPUID_INTC_EDX_SEP;
827 		} else if (IS_NEW_F6(cpi) || cpi->cpi_family == 0xf) {
828 			x86_type = X86_TYPE_P4;
829 			/*
830 			 * We don't currently depend on any of the %ecx
831 			 * features until Prescott, so we'll only check
832 			 * this from P4 onwards.  We might want to revisit
833 			 * that idea later.
834 			 */
835 			mask_ecx = 0xffffffff;
836 		} else if (cpi->cpi_family > 0xf)
837 			mask_ecx = 0xffffffff;
838 		/*
839 		 * We don't support MONITOR/MWAIT if leaf 5 is not available
840 		 * to obtain the monitor linesize.
841 		 */
842 		if (cpi->cpi_maxeax < 5)
843 			mask_ecx &= ~CPUID_INTC_ECX_MON;
844 		break;
845 	case X86_VENDOR_IntelClone:
846 	default:
847 		break;
848 	case X86_VENDOR_AMD:
849 #if defined(OPTERON_ERRATUM_108)
850 		if (cpi->cpi_family == 0xf && cpi->cpi_model == 0xe) {
851 			cp->cp_eax = (0xf0f & cp->cp_eax) | 0xc0;
852 			cpi->cpi_model = 0xc;
853 		} else
854 #endif
855 		if (cpi->cpi_family == 5) {
856 			/*
857 			 * AMD K5 and K6
858 			 *
859 			 * These CPUs have an incomplete implementation
860 			 * of MCA/MCE which we mask away.
861 			 */
862 			mask_edx &= ~(CPUID_INTC_EDX_MCE | CPUID_INTC_EDX_MCA);
863 
864 			/*
865 			 * Model 0 uses the wrong (APIC) bit
866 			 * to indicate PGE.  Fix it here.
867 			 */
868 			if (cpi->cpi_model == 0) {
869 				if (cp->cp_edx & 0x200) {
870 					cp->cp_edx &= ~0x200;
871 					cp->cp_edx |= CPUID_INTC_EDX_PGE;
872 				}
873 			}
874 
875 			/*
876 			 * Early models had problems w/ MMX; disable.
877 			 */
878 			if (cpi->cpi_model < 6)
879 				mask_edx &= ~CPUID_INTC_EDX_MMX;
880 		}
881 
882 		/*
883 		 * For newer families, SSE3 and CX16, at least, are valid;
884 		 * enable all
885 		 */
886 		if (cpi->cpi_family >= 0xf)
887 			mask_ecx = 0xffffffff;
888 		/*
889 		 * We don't support MONITOR/MWAIT if leaf 5 is not available
890 		 * to obtain the monitor linesize.
891 		 */
892 		if (cpi->cpi_maxeax < 5)
893 			mask_ecx &= ~CPUID_INTC_ECX_MON;
894 
895 #if !defined(__xpv)
896 		/*
897 		 * Do not use MONITOR/MWAIT to halt in the idle loop on any AMD
898 		 * processors.  AMD does not intend MWAIT to be used in the cpu
899 		 * idle loop on current and future processors.  10h and future
900 		 * AMD processors use more power in MWAIT than HLT.
901 		 * Pre-family-10h Opterons do not have the MWAIT instruction.
902 		 */
903 		idle_cpu_prefer_mwait = 0;
904 #endif
905 
906 		break;
907 	case X86_VENDOR_TM:
908 		/*
909 		 * workaround the NT workaround in CMS 4.1
910 		 */
911 		if (cpi->cpi_family == 5 && cpi->cpi_model == 4 &&
912 		    (cpi->cpi_step == 2 || cpi->cpi_step == 3))
913 			cp->cp_edx |= CPUID_INTC_EDX_CX8;
914 		break;
915 	case X86_VENDOR_Centaur:
916 		/*
917 		 * workaround the NT workarounds again
918 		 */
919 		if (cpi->cpi_family == 6)
920 			cp->cp_edx |= CPUID_INTC_EDX_CX8;
921 		break;
922 	case X86_VENDOR_Cyrix:
923 		/*
924 		 * We rely heavily on the probing in locore
925 		 * to actually figure out what parts, if any,
926 		 * of the Cyrix cpuid instruction to believe.
927 		 */
928 		switch (x86_type) {
929 		case X86_TYPE_CYRIX_486:
930 			mask_edx = 0;
931 			break;
932 		case X86_TYPE_CYRIX_6x86:
933 			mask_edx = 0;
934 			break;
935 		case X86_TYPE_CYRIX_6x86L:
936 			mask_edx =
937 			    CPUID_INTC_EDX_DE |
938 			    CPUID_INTC_EDX_CX8;
939 			break;
940 		case X86_TYPE_CYRIX_6x86MX:
941 			mask_edx =
942 			    CPUID_INTC_EDX_DE |
943 			    CPUID_INTC_EDX_MSR |
944 			    CPUID_INTC_EDX_CX8 |
945 			    CPUID_INTC_EDX_PGE |
946 			    CPUID_INTC_EDX_CMOV |
947 			    CPUID_INTC_EDX_MMX;
948 			break;
949 		case X86_TYPE_CYRIX_GXm:
950 			mask_edx =
951 			    CPUID_INTC_EDX_MSR |
952 			    CPUID_INTC_EDX_CX8 |
953 			    CPUID_INTC_EDX_CMOV |
954 			    CPUID_INTC_EDX_MMX;
955 			break;
956 		case X86_TYPE_CYRIX_MediaGX:
957 			break;
958 		case X86_TYPE_CYRIX_MII:
959 		case X86_TYPE_VIA_CYRIX_III:
960 			mask_edx =
961 			    CPUID_INTC_EDX_DE |
962 			    CPUID_INTC_EDX_TSC |
963 			    CPUID_INTC_EDX_MSR |
964 			    CPUID_INTC_EDX_CX8 |
965 			    CPUID_INTC_EDX_PGE |
966 			    CPUID_INTC_EDX_CMOV |
967 			    CPUID_INTC_EDX_MMX;
968 			break;
969 		default:
970 			break;
971 		}
972 		break;
973 	}
974 
975 #if defined(__xpv)
976 	/*
977 	 * Do not support MONITOR/MWAIT under a hypervisor
978 	 */
979 	mask_ecx &= ~CPUID_INTC_ECX_MON;
980 #endif	/* __xpv */
981 
982 	/*
983 	 * Now we've figured out the masks that determine
984 	 * which bits we choose to believe, apply the masks
985 	 * to the feature words, then map the kernel's view
986 	 * of these feature words into its feature word.
987 	 */
988 	cp->cp_edx &= mask_edx;
989 	cp->cp_ecx &= mask_ecx;
990 
991 	/*
992 	 * apply any platform restrictions (we don't call this
993 	 * immediately after __cpuid_insn here, because we need the
994 	 * workarounds applied above first)
995 	 */
996 	platform_cpuid_mangle(cpi->cpi_vendor, 1, cp);
997 
998 	/*
999 	 * fold in overrides from the "eeprom" mechanism
1000 	 */
1001 	cp->cp_edx |= cpuid_feature_edx_include;
1002 	cp->cp_edx &= ~cpuid_feature_edx_exclude;
1003 
1004 	cp->cp_ecx |= cpuid_feature_ecx_include;
1005 	cp->cp_ecx &= ~cpuid_feature_ecx_exclude;
1006 
1007 	if (cp->cp_edx & CPUID_INTC_EDX_PSE)
1008 		feature |= X86_LARGEPAGE;
1009 	if (cp->cp_edx & CPUID_INTC_EDX_TSC)
1010 		feature |= X86_TSC;
1011 	if (cp->cp_edx & CPUID_INTC_EDX_MSR)
1012 		feature |= X86_MSR;
1013 	if (cp->cp_edx & CPUID_INTC_EDX_MTRR)
1014 		feature |= X86_MTRR;
1015 	if (cp->cp_edx & CPUID_INTC_EDX_PGE)
1016 		feature |= X86_PGE;
1017 	if (cp->cp_edx & CPUID_INTC_EDX_CMOV)
1018 		feature |= X86_CMOV;
1019 	if (cp->cp_edx & CPUID_INTC_EDX_MMX)
1020 		feature |= X86_MMX;
1021 	if ((cp->cp_edx & CPUID_INTC_EDX_MCE) != 0 &&
1022 	    (cp->cp_edx & CPUID_INTC_EDX_MCA) != 0)
1023 		feature |= X86_MCA;
1024 	if (cp->cp_edx & CPUID_INTC_EDX_PAE)
1025 		feature |= X86_PAE;
1026 	if (cp->cp_edx & CPUID_INTC_EDX_CX8)
1027 		feature |= X86_CX8;
1028 	if (cp->cp_ecx & CPUID_INTC_ECX_CX16)
1029 		feature |= X86_CX16;
1030 	if (cp->cp_edx & CPUID_INTC_EDX_PAT)
1031 		feature |= X86_PAT;
1032 	if (cp->cp_edx & CPUID_INTC_EDX_SEP)
1033 		feature |= X86_SEP;
1034 	if (cp->cp_edx & CPUID_INTC_EDX_FXSR) {
1035 		/*
1036 		 * In our implementation, fxsave/fxrstor
1037 		 * are prerequisites before we'll even
1038 		 * try and do SSE things.
1039 		 */
1040 		if (cp->cp_edx & CPUID_INTC_EDX_SSE)
1041 			feature |= X86_SSE;
1042 		if (cp->cp_edx & CPUID_INTC_EDX_SSE2)
1043 			feature |= X86_SSE2;
1044 		if (cp->cp_ecx & CPUID_INTC_ECX_SSE3)
1045 			feature |= X86_SSE3;
1046 		if (cpi->cpi_vendor == X86_VENDOR_Intel) {
1047 			if (cp->cp_ecx & CPUID_INTC_ECX_SSSE3)
1048 				feature |= X86_SSSE3;
1049 			if (cp->cp_ecx & CPUID_INTC_ECX_SSE4_1)
1050 				feature |= X86_SSE4_1;
1051 			if (cp->cp_ecx & CPUID_INTC_ECX_SSE4_2)
1052 				feature |= X86_SSE4_2;
1053 			if (cp->cp_ecx & CPUID_INTC_ECX_AES)
1054 				feature |= X86_AES;
1055 		}
1056 	}
1057 	if (cp->cp_edx & CPUID_INTC_EDX_DE)
1058 		feature |= X86_DE;
1059 #if !defined(__xpv)
1060 	if (cp->cp_ecx & CPUID_INTC_ECX_MON) {
1061 
1062 		/*
1063 		 * We require the CLFLUSH instruction for erratum workaround
1064 		 * to use MONITOR/MWAIT.
1065 		 */
1066 		if (cp->cp_edx & CPUID_INTC_EDX_CLFSH) {
1067 			cpi->cpi_mwait.support |= MWAIT_SUPPORT;
1068 			feature |= X86_MWAIT;
1069 		} else {
1070 			extern int idle_cpu_assert_cflush_monitor;
1071 
1072 			/*
1073 			 * All processors we are aware of which have
1074 			 * MONITOR/MWAIT also have CLFLUSH.
1075 			 */
1076 			if (idle_cpu_assert_cflush_monitor) {
1077 				ASSERT((cp->cp_ecx & CPUID_INTC_ECX_MON) &&
1078 				    (cp->cp_edx & CPUID_INTC_EDX_CLFSH));
1079 			}
1080 		}
1081 	}
1082 #endif	/* __xpv */
1083 
1084 	/*
1085 	 * Only need it first time, rest of the cpus would follow suite.
1086 	 * we only capture this for the bootcpu.
1087 	 */
1088 	if (cp->cp_edx & CPUID_INTC_EDX_CLFSH) {
1089 		feature |= X86_CLFSH;
1090 		x86_clflush_size = (BITX(cp->cp_ebx, 15, 8) * 8);
1091 	}
1092 
1093 	if (feature & X86_PAE)
1094 		cpi->cpi_pabits = 36;
1095 
1096 	/*
1097 	 * Hyperthreading configuration is slightly tricky on Intel
1098 	 * and pure clones, and even trickier on AMD.
1099 	 *
1100 	 * (AMD chose to set the HTT bit on their CMP processors,
1101 	 * even though they're not actually hyperthreaded.  Thus it
1102 	 * takes a bit more work to figure out what's really going
1103 	 * on ... see the handling of the CMP_LGCY bit below)
1104 	 */
1105 	if (cp->cp_edx & CPUID_INTC_EDX_HTT) {
1106 		cpi->cpi_ncpu_per_chip = CPI_CPU_COUNT(cpi);
1107 		if (cpi->cpi_ncpu_per_chip > 1)
1108 			feature |= X86_HTT;
1109 	} else {
1110 		cpi->cpi_ncpu_per_chip = 1;
1111 	}
1112 
1113 	/*
1114 	 * Work on the "extended" feature information, doing
1115 	 * some basic initialization for cpuid_pass2()
1116 	 */
1117 	xcpuid = 0;
1118 	switch (cpi->cpi_vendor) {
1119 	case X86_VENDOR_Intel:
1120 		if (IS_NEW_F6(cpi) || cpi->cpi_family >= 0xf)
1121 			xcpuid++;
1122 		break;
1123 	case X86_VENDOR_AMD:
1124 		if (cpi->cpi_family > 5 ||
1125 		    (cpi->cpi_family == 5 && cpi->cpi_model >= 1))
1126 			xcpuid++;
1127 		break;
1128 	case X86_VENDOR_Cyrix:
1129 		/*
1130 		 * Only these Cyrix CPUs are -known- to support
1131 		 * extended cpuid operations.
1132 		 */
1133 		if (x86_type == X86_TYPE_VIA_CYRIX_III ||
1134 		    x86_type == X86_TYPE_CYRIX_GXm)
1135 			xcpuid++;
1136 		break;
1137 	case X86_VENDOR_Centaur:
1138 	case X86_VENDOR_TM:
1139 	default:
1140 		xcpuid++;
1141 		break;
1142 	}
1143 
1144 	if (xcpuid) {
1145 		cp = &cpi->cpi_extd[0];
1146 		cp->cp_eax = 0x80000000;
1147 		cpi->cpi_xmaxeax = __cpuid_insn(cp);
1148 	}
1149 
1150 	if (cpi->cpi_xmaxeax & 0x80000000) {
1151 
1152 		if (cpi->cpi_xmaxeax > CPI_XMAXEAX_MAX)
1153 			cpi->cpi_xmaxeax = CPI_XMAXEAX_MAX;
1154 
1155 		switch (cpi->cpi_vendor) {
1156 		case X86_VENDOR_Intel:
1157 		case X86_VENDOR_AMD:
1158 			if (cpi->cpi_xmaxeax < 0x80000001)
1159 				break;
1160 			cp = &cpi->cpi_extd[1];
1161 			cp->cp_eax = 0x80000001;
1162 			(void) __cpuid_insn(cp);
1163 
1164 			if (cpi->cpi_vendor == X86_VENDOR_AMD &&
1165 			    cpi->cpi_family == 5 &&
1166 			    cpi->cpi_model == 6 &&
1167 			    cpi->cpi_step == 6) {
1168 				/*
1169 				 * K6 model 6 uses bit 10 to indicate SYSC
1170 				 * Later models use bit 11. Fix it here.
1171 				 */
1172 				if (cp->cp_edx & 0x400) {
1173 					cp->cp_edx &= ~0x400;
1174 					cp->cp_edx |= CPUID_AMD_EDX_SYSC;
1175 				}
1176 			}
1177 
1178 			platform_cpuid_mangle(cpi->cpi_vendor, 0x80000001, cp);
1179 
1180 			/*
1181 			 * Compute the additions to the kernel's feature word.
1182 			 */
1183 			if (cp->cp_edx & CPUID_AMD_EDX_NX)
1184 				feature |= X86_NX;
1185 
1186 			/*
1187 			 * Regardless whether or not we boot 64-bit,
1188 			 * we should have a way to identify whether
1189 			 * the CPU is capable of running 64-bit.
1190 			 */
1191 			if (cp->cp_edx & CPUID_AMD_EDX_LM)
1192 				feature |= X86_64;
1193 
1194 #if defined(__amd64)
1195 			/* 1 GB large page - enable only for 64 bit kernel */
1196 			if (cp->cp_edx & CPUID_AMD_EDX_1GPG)
1197 				feature |= X86_1GPG;
1198 #endif
1199 
1200 			if ((cpi->cpi_vendor == X86_VENDOR_AMD) &&
1201 			    (cpi->cpi_std[1].cp_edx & CPUID_INTC_EDX_FXSR) &&
1202 			    (cp->cp_ecx & CPUID_AMD_ECX_SSE4A))
1203 				feature |= X86_SSE4A;
1204 
1205 			/*
1206 			 * If both the HTT and CMP_LGCY bits are set,
1207 			 * then we're not actually HyperThreaded.  Read
1208 			 * "AMD CPUID Specification" for more details.
1209 			 */
1210 			if (cpi->cpi_vendor == X86_VENDOR_AMD &&
1211 			    (feature & X86_HTT) &&
1212 			    (cp->cp_ecx & CPUID_AMD_ECX_CMP_LGCY)) {
1213 				feature &= ~X86_HTT;
1214 				feature |= X86_CMP;
1215 			}
1216 #if defined(__amd64)
1217 			/*
1218 			 * It's really tricky to support syscall/sysret in
1219 			 * the i386 kernel; we rely on sysenter/sysexit
1220 			 * instead.  In the amd64 kernel, things are -way-
1221 			 * better.
1222 			 */
1223 			if (cp->cp_edx & CPUID_AMD_EDX_SYSC)
1224 				feature |= X86_ASYSC;
1225 
1226 			/*
1227 			 * While we're thinking about system calls, note
1228 			 * that AMD processors don't support sysenter
1229 			 * in long mode at all, so don't try to program them.
1230 			 */
1231 			if (x86_vendor == X86_VENDOR_AMD)
1232 				feature &= ~X86_SEP;
1233 #endif
1234 			if (cp->cp_edx & CPUID_AMD_EDX_TSCP)
1235 				feature |= X86_TSCP;
1236 			break;
1237 		default:
1238 			break;
1239 		}
1240 
1241 		/*
1242 		 * Get CPUID data about processor cores and hyperthreads.
1243 		 */
1244 		switch (cpi->cpi_vendor) {
1245 		case X86_VENDOR_Intel:
1246 			if (cpi->cpi_maxeax >= 4) {
1247 				cp = &cpi->cpi_std[4];
1248 				cp->cp_eax = 4;
1249 				cp->cp_ecx = 0;
1250 				(void) __cpuid_insn(cp);
1251 				platform_cpuid_mangle(cpi->cpi_vendor, 4, cp);
1252 			}
1253 			/*FALLTHROUGH*/
1254 		case X86_VENDOR_AMD:
1255 			if (cpi->cpi_xmaxeax < 0x80000008)
1256 				break;
1257 			cp = &cpi->cpi_extd[8];
1258 			cp->cp_eax = 0x80000008;
1259 			(void) __cpuid_insn(cp);
1260 			platform_cpuid_mangle(cpi->cpi_vendor, 0x80000008, cp);
1261 
1262 			/*
1263 			 * Virtual and physical address limits from
1264 			 * cpuid override previously guessed values.
1265 			 */
1266 			cpi->cpi_pabits = BITX(cp->cp_eax, 7, 0);
1267 			cpi->cpi_vabits = BITX(cp->cp_eax, 15, 8);
1268 			break;
1269 		default:
1270 			break;
1271 		}
1272 
1273 		/*
1274 		 * Derive the number of cores per chip
1275 		 */
1276 		switch (cpi->cpi_vendor) {
1277 		case X86_VENDOR_Intel:
1278 			if (cpi->cpi_maxeax < 4) {
1279 				cpi->cpi_ncore_per_chip = 1;
1280 				break;
1281 			} else {
1282 				cpi->cpi_ncore_per_chip =
1283 				    BITX((cpi)->cpi_std[4].cp_eax, 31, 26) + 1;
1284 			}
1285 			break;
1286 		case X86_VENDOR_AMD:
1287 			if (cpi->cpi_xmaxeax < 0x80000008) {
1288 				cpi->cpi_ncore_per_chip = 1;
1289 				break;
1290 			} else {
1291 				/*
1292 				 * On family 0xf cpuid fn 2 ECX[7:0] "NC" is
1293 				 * 1 less than the number of physical cores on
1294 				 * the chip.  In family 0x10 this value can
1295 				 * be affected by "downcoring" - it reflects
1296 				 * 1 less than the number of cores actually
1297 				 * enabled on this node.
1298 				 */
1299 				cpi->cpi_ncore_per_chip =
1300 				    BITX((cpi)->cpi_extd[8].cp_ecx, 7, 0) + 1;
1301 			}
1302 			break;
1303 		default:
1304 			cpi->cpi_ncore_per_chip = 1;
1305 			break;
1306 		}
1307 
1308 		/*
1309 		 * Get CPUID data about TSC Invariance in Deep C-State.
1310 		 */
1311 		switch (cpi->cpi_vendor) {
1312 		case X86_VENDOR_Intel:
1313 			if (cpi->cpi_maxeax >= 7) {
1314 				cp = &cpi->cpi_extd[7];
1315 				cp->cp_eax = 0x80000007;
1316 				cp->cp_ecx = 0;
1317 				(void) __cpuid_insn(cp);
1318 			}
1319 			break;
1320 		default:
1321 			break;
1322 		}
1323 	} else {
1324 		cpi->cpi_ncore_per_chip = 1;
1325 	}
1326 
1327 	/*
1328 	 * If more than one core, then this processor is CMP.
1329 	 */
1330 	if (cpi->cpi_ncore_per_chip > 1)
1331 		feature |= X86_CMP;
1332 
1333 	/*
1334 	 * If the number of cores is the same as the number
1335 	 * of CPUs, then we cannot have HyperThreading.
1336 	 */
1337 	if (cpi->cpi_ncpu_per_chip == cpi->cpi_ncore_per_chip)
1338 		feature &= ~X86_HTT;
1339 
1340 	cpi->cpi_apicid = CPI_APIC_ID(cpi);
1341 	cpi->cpi_procnodes_per_pkg = 1;
1342 
1343 	if ((feature & (X86_HTT | X86_CMP)) == 0) {
1344 		/*
1345 		 * Single-core single-threaded processors.
1346 		 */
1347 		cpi->cpi_chipid = -1;
1348 		cpi->cpi_clogid = 0;
1349 		cpi->cpi_coreid = cpu->cpu_id;
1350 		cpi->cpi_pkgcoreid = 0;
1351 		if (cpi->cpi_vendor == X86_VENDOR_AMD)
1352 			cpi->cpi_procnodeid = BITX(cpi->cpi_apicid, 3, 0);
1353 		else
1354 			cpi->cpi_procnodeid = cpi->cpi_chipid;
1355 	} else if (cpi->cpi_ncpu_per_chip > 1) {
1356 		if (cpi->cpi_vendor == X86_VENDOR_Intel)
1357 			cpuid_intel_getids(cpu, feature);
1358 		else if (cpi->cpi_vendor == X86_VENDOR_AMD)
1359 			cpuid_amd_getids(cpu);
1360 		else {
1361 			/*
1362 			 * All other processors are currently
1363 			 * assumed to have single cores.
1364 			 */
1365 			cpi->cpi_coreid = cpi->cpi_chipid;
1366 			cpi->cpi_pkgcoreid = 0;
1367 			cpi->cpi_procnodeid = cpi->cpi_chipid;
1368 		}
1369 	}
1370 
1371 	/*
1372 	 * Synthesize chip "revision" and socket type
1373 	 */
1374 	cpi->cpi_chiprev = _cpuid_chiprev(cpi->cpi_vendor, cpi->cpi_family,
1375 	    cpi->cpi_model, cpi->cpi_step);
1376 	cpi->cpi_chiprevstr = _cpuid_chiprevstr(cpi->cpi_vendor,
1377 	    cpi->cpi_family, cpi->cpi_model, cpi->cpi_step);
1378 	cpi->cpi_socket = _cpuid_skt(cpi->cpi_vendor, cpi->cpi_family,
1379 	    cpi->cpi_model, cpi->cpi_step);
1380 
1381 pass1_done:
1382 	cpi->cpi_pass = 1;
1383 	return (feature);
1384 }
1385 
1386 /*
1387  * Make copies of the cpuid table entries we depend on, in
1388  * part for ease of parsing now, in part so that we have only
1389  * one place to correct any of it, in part for ease of
1390  * later export to userland, and in part so we can look at
1391  * this stuff in a crash dump.
1392  */
1393 
1394 /*ARGSUSED*/
1395 void
1396 cpuid_pass2(cpu_t *cpu)
1397 {
1398 	uint_t n, nmax;
1399 	int i;
1400 	struct cpuid_regs *cp;
1401 	uint8_t *dp;
1402 	uint32_t *iptr;
1403 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
1404 
1405 	ASSERT(cpi->cpi_pass == 1);
1406 
1407 	if (cpi->cpi_maxeax < 1)
1408 		goto pass2_done;
1409 
1410 	if ((nmax = cpi->cpi_maxeax + 1) > NMAX_CPI_STD)
1411 		nmax = NMAX_CPI_STD;
1412 	/*
1413 	 * (We already handled n == 0 and n == 1 in pass 1)
1414 	 */
1415 	for (n = 2, cp = &cpi->cpi_std[2]; n < nmax; n++, cp++) {
1416 		cp->cp_eax = n;
1417 
1418 		/*
1419 		 * CPUID function 4 expects %ecx to be initialized
1420 		 * with an index which indicates which cache to return
1421 		 * information about. The OS is expected to call function 4
1422 		 * with %ecx set to 0, 1, 2, ... until it returns with
1423 		 * EAX[4:0] set to 0, which indicates there are no more
1424 		 * caches.
1425 		 *
1426 		 * Here, populate cpi_std[4] with the information returned by
1427 		 * function 4 when %ecx == 0, and do the rest in cpuid_pass3()
1428 		 * when dynamic memory allocation becomes available.
1429 		 *
1430 		 * Note: we need to explicitly initialize %ecx here, since
1431 		 * function 4 may have been previously invoked.
1432 		 */
1433 		if (n == 4)
1434 			cp->cp_ecx = 0;
1435 
1436 		(void) __cpuid_insn(cp);
1437 		platform_cpuid_mangle(cpi->cpi_vendor, n, cp);
1438 		switch (n) {
1439 		case 2:
1440 			/*
1441 			 * "the lower 8 bits of the %eax register
1442 			 * contain a value that identifies the number
1443 			 * of times the cpuid [instruction] has to be
1444 			 * executed to obtain a complete image of the
1445 			 * processor's caching systems."
1446 			 *
1447 			 * How *do* they make this stuff up?
1448 			 */
1449 			cpi->cpi_ncache = sizeof (*cp) *
1450 			    BITX(cp->cp_eax, 7, 0);
1451 			if (cpi->cpi_ncache == 0)
1452 				break;
1453 			cpi->cpi_ncache--;	/* skip count byte */
1454 
1455 			/*
1456 			 * Well, for now, rather than attempt to implement
1457 			 * this slightly dubious algorithm, we just look
1458 			 * at the first 15 ..
1459 			 */
1460 			if (cpi->cpi_ncache > (sizeof (*cp) - 1))
1461 				cpi->cpi_ncache = sizeof (*cp) - 1;
1462 
1463 			dp = cpi->cpi_cacheinfo;
1464 			if (BITX(cp->cp_eax, 31, 31) == 0) {
1465 				uint8_t *p = (void *)&cp->cp_eax;
1466 				for (i = 1; i < 4; i++)
1467 					if (p[i] != 0)
1468 						*dp++ = p[i];
1469 			}
1470 			if (BITX(cp->cp_ebx, 31, 31) == 0) {
1471 				uint8_t *p = (void *)&cp->cp_ebx;
1472 				for (i = 0; i < 4; i++)
1473 					if (p[i] != 0)
1474 						*dp++ = p[i];
1475 			}
1476 			if (BITX(cp->cp_ecx, 31, 31) == 0) {
1477 				uint8_t *p = (void *)&cp->cp_ecx;
1478 				for (i = 0; i < 4; i++)
1479 					if (p[i] != 0)
1480 						*dp++ = p[i];
1481 			}
1482 			if (BITX(cp->cp_edx, 31, 31) == 0) {
1483 				uint8_t *p = (void *)&cp->cp_edx;
1484 				for (i = 0; i < 4; i++)
1485 					if (p[i] != 0)
1486 						*dp++ = p[i];
1487 			}
1488 			break;
1489 
1490 		case 3:	/* Processor serial number, if PSN supported */
1491 			break;
1492 
1493 		case 4:	/* Deterministic cache parameters */
1494 			break;
1495 
1496 		case 5:	/* Monitor/Mwait parameters */
1497 		{
1498 			size_t mwait_size;
1499 
1500 			/*
1501 			 * check cpi_mwait.support which was set in cpuid_pass1
1502 			 */
1503 			if (!(cpi->cpi_mwait.support & MWAIT_SUPPORT))
1504 				break;
1505 
1506 			/*
1507 			 * Protect ourself from insane mwait line size.
1508 			 * Workaround for incomplete hardware emulator(s).
1509 			 */
1510 			mwait_size = (size_t)MWAIT_SIZE_MAX(cpi);
1511 			if (mwait_size < sizeof (uint32_t) ||
1512 			    !ISP2(mwait_size)) {
1513 #if DEBUG
1514 				cmn_err(CE_NOTE, "Cannot handle cpu %d mwait "
1515 				    "size %ld", cpu->cpu_id, (long)mwait_size);
1516 #endif
1517 				break;
1518 			}
1519 
1520 			cpi->cpi_mwait.mon_min = (size_t)MWAIT_SIZE_MIN(cpi);
1521 			cpi->cpi_mwait.mon_max = mwait_size;
1522 			if (MWAIT_EXTENSION(cpi)) {
1523 				cpi->cpi_mwait.support |= MWAIT_EXTENSIONS;
1524 				if (MWAIT_INT_ENABLE(cpi))
1525 					cpi->cpi_mwait.support |=
1526 					    MWAIT_ECX_INT_ENABLE;
1527 			}
1528 			break;
1529 		}
1530 		default:
1531 			break;
1532 		}
1533 	}
1534 
1535 	if (cpi->cpi_maxeax >= 0xB && cpi->cpi_vendor == X86_VENDOR_Intel) {
1536 		struct cpuid_regs regs;
1537 
1538 		cp = &regs;
1539 		cp->cp_eax = 0xB;
1540 		cp->cp_edx = cp->cp_ebx = cp->cp_ecx = 0;
1541 
1542 		(void) __cpuid_insn(cp);
1543 
1544 		/*
1545 		 * Check CPUID.EAX=0BH, ECX=0H:EBX is non-zero, which
1546 		 * indicates that the extended topology enumeration leaf is
1547 		 * available.
1548 		 */
1549 		if (cp->cp_ebx) {
1550 			uint32_t x2apic_id;
1551 			uint_t coreid_shift = 0;
1552 			uint_t ncpu_per_core = 1;
1553 			uint_t chipid_shift = 0;
1554 			uint_t ncpu_per_chip = 1;
1555 			uint_t i;
1556 			uint_t level;
1557 
1558 			for (i = 0; i < CPI_FNB_ECX_MAX; i++) {
1559 				cp->cp_eax = 0xB;
1560 				cp->cp_ecx = i;
1561 
1562 				(void) __cpuid_insn(cp);
1563 				level = CPI_CPU_LEVEL_TYPE(cp);
1564 
1565 				if (level == 1) {
1566 					x2apic_id = cp->cp_edx;
1567 					coreid_shift = BITX(cp->cp_eax, 4, 0);
1568 					ncpu_per_core = BITX(cp->cp_ebx, 15, 0);
1569 				} else if (level == 2) {
1570 					x2apic_id = cp->cp_edx;
1571 					chipid_shift = BITX(cp->cp_eax, 4, 0);
1572 					ncpu_per_chip = BITX(cp->cp_ebx, 15, 0);
1573 				}
1574 			}
1575 
1576 			cpi->cpi_apicid = x2apic_id;
1577 			cpi->cpi_ncpu_per_chip = ncpu_per_chip;
1578 			cpi->cpi_ncore_per_chip = ncpu_per_chip /
1579 			    ncpu_per_core;
1580 			cpi->cpi_chipid = x2apic_id >> chipid_shift;
1581 			cpi->cpi_clogid = x2apic_id & ((1 << chipid_shift) - 1);
1582 			cpi->cpi_coreid = x2apic_id >> coreid_shift;
1583 			cpi->cpi_pkgcoreid = cpi->cpi_clogid >> coreid_shift;
1584 		}
1585 
1586 		/* Make cp NULL so that we don't stumble on others */
1587 		cp = NULL;
1588 	}
1589 
1590 	if ((cpi->cpi_xmaxeax & 0x80000000) == 0)
1591 		goto pass2_done;
1592 
1593 	if ((nmax = cpi->cpi_xmaxeax - 0x80000000 + 1) > NMAX_CPI_EXTD)
1594 		nmax = NMAX_CPI_EXTD;
1595 	/*
1596 	 * Copy the extended properties, fixing them as we go.
1597 	 * (We already handled n == 0 and n == 1 in pass 1)
1598 	 */
1599 	iptr = (void *)cpi->cpi_brandstr;
1600 	for (n = 2, cp = &cpi->cpi_extd[2]; n < nmax; cp++, n++) {
1601 		cp->cp_eax = 0x80000000 + n;
1602 		(void) __cpuid_insn(cp);
1603 		platform_cpuid_mangle(cpi->cpi_vendor, 0x80000000 + n, cp);
1604 		switch (n) {
1605 		case 2:
1606 		case 3:
1607 		case 4:
1608 			/*
1609 			 * Extract the brand string
1610 			 */
1611 			*iptr++ = cp->cp_eax;
1612 			*iptr++ = cp->cp_ebx;
1613 			*iptr++ = cp->cp_ecx;
1614 			*iptr++ = cp->cp_edx;
1615 			break;
1616 		case 5:
1617 			switch (cpi->cpi_vendor) {
1618 			case X86_VENDOR_AMD:
1619 				/*
1620 				 * The Athlon and Duron were the first
1621 				 * parts to report the sizes of the
1622 				 * TLB for large pages. Before then,
1623 				 * we don't trust the data.
1624 				 */
1625 				if (cpi->cpi_family < 6 ||
1626 				    (cpi->cpi_family == 6 &&
1627 				    cpi->cpi_model < 1))
1628 					cp->cp_eax = 0;
1629 				break;
1630 			default:
1631 				break;
1632 			}
1633 			break;
1634 		case 6:
1635 			switch (cpi->cpi_vendor) {
1636 			case X86_VENDOR_AMD:
1637 				/*
1638 				 * The Athlon and Duron were the first
1639 				 * AMD parts with L2 TLB's.
1640 				 * Before then, don't trust the data.
1641 				 */
1642 				if (cpi->cpi_family < 6 ||
1643 				    cpi->cpi_family == 6 &&
1644 				    cpi->cpi_model < 1)
1645 					cp->cp_eax = cp->cp_ebx = 0;
1646 				/*
1647 				 * AMD Duron rev A0 reports L2
1648 				 * cache size incorrectly as 1K
1649 				 * when it is really 64K
1650 				 */
1651 				if (cpi->cpi_family == 6 &&
1652 				    cpi->cpi_model == 3 &&
1653 				    cpi->cpi_step == 0) {
1654 					cp->cp_ecx &= 0xffff;
1655 					cp->cp_ecx |= 0x400000;
1656 				}
1657 				break;
1658 			case X86_VENDOR_Cyrix:	/* VIA C3 */
1659 				/*
1660 				 * VIA C3 processors are a bit messed
1661 				 * up w.r.t. encoding cache sizes in %ecx
1662 				 */
1663 				if (cpi->cpi_family != 6)
1664 					break;
1665 				/*
1666 				 * model 7 and 8 were incorrectly encoded
1667 				 *
1668 				 * xxx is model 8 really broken?
1669 				 */
1670 				if (cpi->cpi_model == 7 ||
1671 				    cpi->cpi_model == 8)
1672 					cp->cp_ecx =
1673 					    BITX(cp->cp_ecx, 31, 24) << 16 |
1674 					    BITX(cp->cp_ecx, 23, 16) << 12 |
1675 					    BITX(cp->cp_ecx, 15, 8) << 8 |
1676 					    BITX(cp->cp_ecx, 7, 0);
1677 				/*
1678 				 * model 9 stepping 1 has wrong associativity
1679 				 */
1680 				if (cpi->cpi_model == 9 && cpi->cpi_step == 1)
1681 					cp->cp_ecx |= 8 << 12;
1682 				break;
1683 			case X86_VENDOR_Intel:
1684 				/*
1685 				 * Extended L2 Cache features function.
1686 				 * First appeared on Prescott.
1687 				 */
1688 			default:
1689 				break;
1690 			}
1691 			break;
1692 		default:
1693 			break;
1694 		}
1695 	}
1696 
1697 pass2_done:
1698 	cpi->cpi_pass = 2;
1699 }
1700 
1701 static const char *
1702 intel_cpubrand(const struct cpuid_info *cpi)
1703 {
1704 	int i;
1705 
1706 	if ((x86_feature & X86_CPUID) == 0 ||
1707 	    cpi->cpi_maxeax < 1 || cpi->cpi_family < 5)
1708 		return ("i486");
1709 
1710 	switch (cpi->cpi_family) {
1711 	case 5:
1712 		return ("Intel Pentium(r)");
1713 	case 6:
1714 		switch (cpi->cpi_model) {
1715 			uint_t celeron, xeon;
1716 			const struct cpuid_regs *cp;
1717 		case 0:
1718 		case 1:
1719 		case 2:
1720 			return ("Intel Pentium(r) Pro");
1721 		case 3:
1722 		case 4:
1723 			return ("Intel Pentium(r) II");
1724 		case 6:
1725 			return ("Intel Celeron(r)");
1726 		case 5:
1727 		case 7:
1728 			celeron = xeon = 0;
1729 			cp = &cpi->cpi_std[2];	/* cache info */
1730 
1731 			for (i = 1; i < 4; i++) {
1732 				uint_t tmp;
1733 
1734 				tmp = (cp->cp_eax >> (8 * i)) & 0xff;
1735 				if (tmp == 0x40)
1736 					celeron++;
1737 				if (tmp >= 0x44 && tmp <= 0x45)
1738 					xeon++;
1739 			}
1740 
1741 			for (i = 0; i < 2; i++) {
1742 				uint_t tmp;
1743 
1744 				tmp = (cp->cp_ebx >> (8 * i)) & 0xff;
1745 				if (tmp == 0x40)
1746 					celeron++;
1747 				else if (tmp >= 0x44 && tmp <= 0x45)
1748 					xeon++;
1749 			}
1750 
1751 			for (i = 0; i < 4; i++) {
1752 				uint_t tmp;
1753 
1754 				tmp = (cp->cp_ecx >> (8 * i)) & 0xff;
1755 				if (tmp == 0x40)
1756 					celeron++;
1757 				else if (tmp >= 0x44 && tmp <= 0x45)
1758 					xeon++;
1759 			}
1760 
1761 			for (i = 0; i < 4; i++) {
1762 				uint_t tmp;
1763 
1764 				tmp = (cp->cp_edx >> (8 * i)) & 0xff;
1765 				if (tmp == 0x40)
1766 					celeron++;
1767 				else if (tmp >= 0x44 && tmp <= 0x45)
1768 					xeon++;
1769 			}
1770 
1771 			if (celeron)
1772 				return ("Intel Celeron(r)");
1773 			if (xeon)
1774 				return (cpi->cpi_model == 5 ?
1775 				    "Intel Pentium(r) II Xeon(tm)" :
1776 				    "Intel Pentium(r) III Xeon(tm)");
1777 			return (cpi->cpi_model == 5 ?
1778 			    "Intel Pentium(r) II or Pentium(r) II Xeon(tm)" :
1779 			    "Intel Pentium(r) III or Pentium(r) III Xeon(tm)");
1780 		default:
1781 			break;
1782 		}
1783 	default:
1784 		break;
1785 	}
1786 
1787 	/* BrandID is present if the field is nonzero */
1788 	if (cpi->cpi_brandid != 0) {
1789 		static const struct {
1790 			uint_t bt_bid;
1791 			const char *bt_str;
1792 		} brand_tbl[] = {
1793 			{ 0x1,	"Intel(r) Celeron(r)" },
1794 			{ 0x2,	"Intel(r) Pentium(r) III" },
1795 			{ 0x3,	"Intel(r) Pentium(r) III Xeon(tm)" },
1796 			{ 0x4,	"Intel(r) Pentium(r) III" },
1797 			{ 0x6,	"Mobile Intel(r) Pentium(r) III" },
1798 			{ 0x7,	"Mobile Intel(r) Celeron(r)" },
1799 			{ 0x8,	"Intel(r) Pentium(r) 4" },
1800 			{ 0x9,	"Intel(r) Pentium(r) 4" },
1801 			{ 0xa,	"Intel(r) Celeron(r)" },
1802 			{ 0xb,	"Intel(r) Xeon(tm)" },
1803 			{ 0xc,	"Intel(r) Xeon(tm) MP" },
1804 			{ 0xe,	"Mobile Intel(r) Pentium(r) 4" },
1805 			{ 0xf,	"Mobile Intel(r) Celeron(r)" },
1806 			{ 0x11, "Mobile Genuine Intel(r)" },
1807 			{ 0x12, "Intel(r) Celeron(r) M" },
1808 			{ 0x13, "Mobile Intel(r) Celeron(r)" },
1809 			{ 0x14, "Intel(r) Celeron(r)" },
1810 			{ 0x15, "Mobile Genuine Intel(r)" },
1811 			{ 0x16,	"Intel(r) Pentium(r) M" },
1812 			{ 0x17, "Mobile Intel(r) Celeron(r)" }
1813 		};
1814 		uint_t btblmax = sizeof (brand_tbl) / sizeof (brand_tbl[0]);
1815 		uint_t sgn;
1816 
1817 		sgn = (cpi->cpi_family << 8) |
1818 		    (cpi->cpi_model << 4) | cpi->cpi_step;
1819 
1820 		for (i = 0; i < btblmax; i++)
1821 			if (brand_tbl[i].bt_bid == cpi->cpi_brandid)
1822 				break;
1823 		if (i < btblmax) {
1824 			if (sgn == 0x6b1 && cpi->cpi_brandid == 3)
1825 				return ("Intel(r) Celeron(r)");
1826 			if (sgn < 0xf13 && cpi->cpi_brandid == 0xb)
1827 				return ("Intel(r) Xeon(tm) MP");
1828 			if (sgn < 0xf13 && cpi->cpi_brandid == 0xe)
1829 				return ("Intel(r) Xeon(tm)");
1830 			return (brand_tbl[i].bt_str);
1831 		}
1832 	}
1833 
1834 	return (NULL);
1835 }
1836 
1837 static const char *
1838 amd_cpubrand(const struct cpuid_info *cpi)
1839 {
1840 	if ((x86_feature & X86_CPUID) == 0 ||
1841 	    cpi->cpi_maxeax < 1 || cpi->cpi_family < 5)
1842 		return ("i486 compatible");
1843 
1844 	switch (cpi->cpi_family) {
1845 	case 5:
1846 		switch (cpi->cpi_model) {
1847 		case 0:
1848 		case 1:
1849 		case 2:
1850 		case 3:
1851 		case 4:
1852 		case 5:
1853 			return ("AMD-K5(r)");
1854 		case 6:
1855 		case 7:
1856 			return ("AMD-K6(r)");
1857 		case 8:
1858 			return ("AMD-K6(r)-2");
1859 		case 9:
1860 			return ("AMD-K6(r)-III");
1861 		default:
1862 			return ("AMD (family 5)");
1863 		}
1864 	case 6:
1865 		switch (cpi->cpi_model) {
1866 		case 1:
1867 			return ("AMD-K7(tm)");
1868 		case 0:
1869 		case 2:
1870 		case 4:
1871 			return ("AMD Athlon(tm)");
1872 		case 3:
1873 		case 7:
1874 			return ("AMD Duron(tm)");
1875 		case 6:
1876 		case 8:
1877 		case 10:
1878 			/*
1879 			 * Use the L2 cache size to distinguish
1880 			 */
1881 			return ((cpi->cpi_extd[6].cp_ecx >> 16) >= 256 ?
1882 			    "AMD Athlon(tm)" : "AMD Duron(tm)");
1883 		default:
1884 			return ("AMD (family 6)");
1885 		}
1886 	default:
1887 		break;
1888 	}
1889 
1890 	if (cpi->cpi_family == 0xf && cpi->cpi_model == 5 &&
1891 	    cpi->cpi_brandid != 0) {
1892 		switch (BITX(cpi->cpi_brandid, 7, 5)) {
1893 		case 3:
1894 			return ("AMD Opteron(tm) UP 1xx");
1895 		case 4:
1896 			return ("AMD Opteron(tm) DP 2xx");
1897 		case 5:
1898 			return ("AMD Opteron(tm) MP 8xx");
1899 		default:
1900 			return ("AMD Opteron(tm)");
1901 		}
1902 	}
1903 
1904 	return (NULL);
1905 }
1906 
1907 static const char *
1908 cyrix_cpubrand(struct cpuid_info *cpi, uint_t type)
1909 {
1910 	if ((x86_feature & X86_CPUID) == 0 ||
1911 	    cpi->cpi_maxeax < 1 || cpi->cpi_family < 5 ||
1912 	    type == X86_TYPE_CYRIX_486)
1913 		return ("i486 compatible");
1914 
1915 	switch (type) {
1916 	case X86_TYPE_CYRIX_6x86:
1917 		return ("Cyrix 6x86");
1918 	case X86_TYPE_CYRIX_6x86L:
1919 		return ("Cyrix 6x86L");
1920 	case X86_TYPE_CYRIX_6x86MX:
1921 		return ("Cyrix 6x86MX");
1922 	case X86_TYPE_CYRIX_GXm:
1923 		return ("Cyrix GXm");
1924 	case X86_TYPE_CYRIX_MediaGX:
1925 		return ("Cyrix MediaGX");
1926 	case X86_TYPE_CYRIX_MII:
1927 		return ("Cyrix M2");
1928 	case X86_TYPE_VIA_CYRIX_III:
1929 		return ("VIA Cyrix M3");
1930 	default:
1931 		/*
1932 		 * Have another wild guess ..
1933 		 */
1934 		if (cpi->cpi_family == 4 && cpi->cpi_model == 9)
1935 			return ("Cyrix 5x86");
1936 		else if (cpi->cpi_family == 5) {
1937 			switch (cpi->cpi_model) {
1938 			case 2:
1939 				return ("Cyrix 6x86");	/* Cyrix M1 */
1940 			case 4:
1941 				return ("Cyrix MediaGX");
1942 			default:
1943 				break;
1944 			}
1945 		} else if (cpi->cpi_family == 6) {
1946 			switch (cpi->cpi_model) {
1947 			case 0:
1948 				return ("Cyrix 6x86MX"); /* Cyrix M2? */
1949 			case 5:
1950 			case 6:
1951 			case 7:
1952 			case 8:
1953 			case 9:
1954 				return ("VIA C3");
1955 			default:
1956 				break;
1957 			}
1958 		}
1959 		break;
1960 	}
1961 	return (NULL);
1962 }
1963 
1964 /*
1965  * This only gets called in the case that the CPU extended
1966  * feature brand string (0x80000002, 0x80000003, 0x80000004)
1967  * aren't available, or contain null bytes for some reason.
1968  */
1969 static void
1970 fabricate_brandstr(struct cpuid_info *cpi)
1971 {
1972 	const char *brand = NULL;
1973 
1974 	switch (cpi->cpi_vendor) {
1975 	case X86_VENDOR_Intel:
1976 		brand = intel_cpubrand(cpi);
1977 		break;
1978 	case X86_VENDOR_AMD:
1979 		brand = amd_cpubrand(cpi);
1980 		break;
1981 	case X86_VENDOR_Cyrix:
1982 		brand = cyrix_cpubrand(cpi, x86_type);
1983 		break;
1984 	case X86_VENDOR_NexGen:
1985 		if (cpi->cpi_family == 5 && cpi->cpi_model == 0)
1986 			brand = "NexGen Nx586";
1987 		break;
1988 	case X86_VENDOR_Centaur:
1989 		if (cpi->cpi_family == 5)
1990 			switch (cpi->cpi_model) {
1991 			case 4:
1992 				brand = "Centaur C6";
1993 				break;
1994 			case 8:
1995 				brand = "Centaur C2";
1996 				break;
1997 			case 9:
1998 				brand = "Centaur C3";
1999 				break;
2000 			default:
2001 				break;
2002 			}
2003 		break;
2004 	case X86_VENDOR_Rise:
2005 		if (cpi->cpi_family == 5 &&
2006 		    (cpi->cpi_model == 0 || cpi->cpi_model == 2))
2007 			brand = "Rise mP6";
2008 		break;
2009 	case X86_VENDOR_SiS:
2010 		if (cpi->cpi_family == 5 && cpi->cpi_model == 0)
2011 			brand = "SiS 55x";
2012 		break;
2013 	case X86_VENDOR_TM:
2014 		if (cpi->cpi_family == 5 && cpi->cpi_model == 4)
2015 			brand = "Transmeta Crusoe TM3x00 or TM5x00";
2016 		break;
2017 	case X86_VENDOR_NSC:
2018 	case X86_VENDOR_UMC:
2019 	default:
2020 		break;
2021 	}
2022 	if (brand) {
2023 		(void) strcpy((char *)cpi->cpi_brandstr, brand);
2024 		return;
2025 	}
2026 
2027 	/*
2028 	 * If all else fails ...
2029 	 */
2030 	(void) snprintf(cpi->cpi_brandstr, sizeof (cpi->cpi_brandstr),
2031 	    "%s %d.%d.%d", cpi->cpi_vendorstr, cpi->cpi_family,
2032 	    cpi->cpi_model, cpi->cpi_step);
2033 }
2034 
2035 /*
2036  * This routine is called just after kernel memory allocation
2037  * becomes available on cpu0, and as part of mp_startup() on
2038  * the other cpus.
2039  *
2040  * Fixup the brand string, and collect any information from cpuid
2041  * that requires dynamicically allocated storage to represent.
2042  */
2043 /*ARGSUSED*/
2044 void
2045 cpuid_pass3(cpu_t *cpu)
2046 {
2047 	int	i, max, shft, level, size;
2048 	struct cpuid_regs regs;
2049 	struct cpuid_regs *cp;
2050 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2051 
2052 	ASSERT(cpi->cpi_pass == 2);
2053 
2054 	/*
2055 	 * Function 4: Deterministic cache parameters
2056 	 *
2057 	 * Take this opportunity to detect the number of threads
2058 	 * sharing the last level cache, and construct a corresponding
2059 	 * cache id. The respective cpuid_info members are initialized
2060 	 * to the default case of "no last level cache sharing".
2061 	 */
2062 	cpi->cpi_ncpu_shr_last_cache = 1;
2063 	cpi->cpi_last_lvl_cacheid = cpu->cpu_id;
2064 
2065 	if (cpi->cpi_maxeax >= 4 && cpi->cpi_vendor == X86_VENDOR_Intel) {
2066 
2067 		/*
2068 		 * Find the # of elements (size) returned by fn 4, and along
2069 		 * the way detect last level cache sharing details.
2070 		 */
2071 		bzero(&regs, sizeof (regs));
2072 		cp = &regs;
2073 		for (i = 0, max = 0; i < CPI_FN4_ECX_MAX; i++) {
2074 			cp->cp_eax = 4;
2075 			cp->cp_ecx = i;
2076 
2077 			(void) __cpuid_insn(cp);
2078 
2079 			if (CPI_CACHE_TYPE(cp) == 0)
2080 				break;
2081 			level = CPI_CACHE_LVL(cp);
2082 			if (level > max) {
2083 				max = level;
2084 				cpi->cpi_ncpu_shr_last_cache =
2085 				    CPI_NTHR_SHR_CACHE(cp) + 1;
2086 			}
2087 		}
2088 		cpi->cpi_std_4_size = size = i;
2089 
2090 		/*
2091 		 * Allocate the cpi_std_4 array. The first element
2092 		 * references the regs for fn 4, %ecx == 0, which
2093 		 * cpuid_pass2() stashed in cpi->cpi_std[4].
2094 		 */
2095 		if (size > 0) {
2096 			cpi->cpi_std_4 =
2097 			    kmem_alloc(size * sizeof (cp), KM_SLEEP);
2098 			cpi->cpi_std_4[0] = &cpi->cpi_std[4];
2099 
2100 			/*
2101 			 * Allocate storage to hold the additional regs
2102 			 * for function 4, %ecx == 1 .. cpi_std_4_size.
2103 			 *
2104 			 * The regs for fn 4, %ecx == 0 has already
2105 			 * been allocated as indicated above.
2106 			 */
2107 			for (i = 1; i < size; i++) {
2108 				cp = cpi->cpi_std_4[i] =
2109 				    kmem_zalloc(sizeof (regs), KM_SLEEP);
2110 				cp->cp_eax = 4;
2111 				cp->cp_ecx = i;
2112 
2113 				(void) __cpuid_insn(cp);
2114 			}
2115 		}
2116 		/*
2117 		 * Determine the number of bits needed to represent
2118 		 * the number of CPUs sharing the last level cache.
2119 		 *
2120 		 * Shift off that number of bits from the APIC id to
2121 		 * derive the cache id.
2122 		 */
2123 		shft = 0;
2124 		for (i = 1; i < cpi->cpi_ncpu_shr_last_cache; i <<= 1)
2125 			shft++;
2126 		cpi->cpi_last_lvl_cacheid = cpi->cpi_apicid >> shft;
2127 	}
2128 
2129 	/*
2130 	 * Now fixup the brand string
2131 	 */
2132 	if ((cpi->cpi_xmaxeax & 0x80000000) == 0) {
2133 		fabricate_brandstr(cpi);
2134 	} else {
2135 
2136 		/*
2137 		 * If we successfully extracted a brand string from the cpuid
2138 		 * instruction, clean it up by removing leading spaces and
2139 		 * similar junk.
2140 		 */
2141 		if (cpi->cpi_brandstr[0]) {
2142 			size_t maxlen = sizeof (cpi->cpi_brandstr);
2143 			char *src, *dst;
2144 
2145 			dst = src = (char *)cpi->cpi_brandstr;
2146 			src[maxlen - 1] = '\0';
2147 			/*
2148 			 * strip leading spaces
2149 			 */
2150 			while (*src == ' ')
2151 				src++;
2152 			/*
2153 			 * Remove any 'Genuine' or "Authentic" prefixes
2154 			 */
2155 			if (strncmp(src, "Genuine ", 8) == 0)
2156 				src += 8;
2157 			if (strncmp(src, "Authentic ", 10) == 0)
2158 				src += 10;
2159 
2160 			/*
2161 			 * Now do an in-place copy.
2162 			 * Map (R) to (r) and (TM) to (tm).
2163 			 * The era of teletypes is long gone, and there's
2164 			 * -really- no need to shout.
2165 			 */
2166 			while (*src != '\0') {
2167 				if (src[0] == '(') {
2168 					if (strncmp(src + 1, "R)", 2) == 0) {
2169 						(void) strncpy(dst, "(r)", 3);
2170 						src += 3;
2171 						dst += 3;
2172 						continue;
2173 					}
2174 					if (strncmp(src + 1, "TM)", 3) == 0) {
2175 						(void) strncpy(dst, "(tm)", 4);
2176 						src += 4;
2177 						dst += 4;
2178 						continue;
2179 					}
2180 				}
2181 				*dst++ = *src++;
2182 			}
2183 			*dst = '\0';
2184 
2185 			/*
2186 			 * Finally, remove any trailing spaces
2187 			 */
2188 			while (--dst > cpi->cpi_brandstr)
2189 				if (*dst == ' ')
2190 					*dst = '\0';
2191 				else
2192 					break;
2193 		} else
2194 			fabricate_brandstr(cpi);
2195 	}
2196 	cpi->cpi_pass = 3;
2197 }
2198 
2199 /*
2200  * This routine is called out of bind_hwcap() much later in the life
2201  * of the kernel (post_startup()).  The job of this routine is to resolve
2202  * the hardware feature support and kernel support for those features into
2203  * what we're actually going to tell applications via the aux vector.
2204  */
2205 uint_t
2206 cpuid_pass4(cpu_t *cpu)
2207 {
2208 	struct cpuid_info *cpi;
2209 	uint_t hwcap_flags = 0;
2210 
2211 	if (cpu == NULL)
2212 		cpu = CPU;
2213 	cpi = cpu->cpu_m.mcpu_cpi;
2214 
2215 	ASSERT(cpi->cpi_pass == 3);
2216 
2217 	if (cpi->cpi_maxeax >= 1) {
2218 		uint32_t *edx = &cpi->cpi_support[STD_EDX_FEATURES];
2219 		uint32_t *ecx = &cpi->cpi_support[STD_ECX_FEATURES];
2220 
2221 		*edx = CPI_FEATURES_EDX(cpi);
2222 		*ecx = CPI_FEATURES_ECX(cpi);
2223 
2224 		/*
2225 		 * [these require explicit kernel support]
2226 		 */
2227 		if ((x86_feature & X86_SEP) == 0)
2228 			*edx &= ~CPUID_INTC_EDX_SEP;
2229 
2230 		if ((x86_feature & X86_SSE) == 0)
2231 			*edx &= ~(CPUID_INTC_EDX_FXSR|CPUID_INTC_EDX_SSE);
2232 		if ((x86_feature & X86_SSE2) == 0)
2233 			*edx &= ~CPUID_INTC_EDX_SSE2;
2234 
2235 		if ((x86_feature & X86_HTT) == 0)
2236 			*edx &= ~CPUID_INTC_EDX_HTT;
2237 
2238 		if ((x86_feature & X86_SSE3) == 0)
2239 			*ecx &= ~CPUID_INTC_ECX_SSE3;
2240 
2241 		if (cpi->cpi_vendor == X86_VENDOR_Intel) {
2242 			if ((x86_feature & X86_SSSE3) == 0)
2243 				*ecx &= ~CPUID_INTC_ECX_SSSE3;
2244 			if ((x86_feature & X86_SSE4_1) == 0)
2245 				*ecx &= ~CPUID_INTC_ECX_SSE4_1;
2246 			if ((x86_feature & X86_SSE4_2) == 0)
2247 				*ecx &= ~CPUID_INTC_ECX_SSE4_2;
2248 			if ((x86_feature & X86_AES) == 0)
2249 				*ecx &= ~CPUID_INTC_ECX_AES;
2250 		}
2251 
2252 		/*
2253 		 * [no explicit support required beyond x87 fp context]
2254 		 */
2255 		if (!fpu_exists)
2256 			*edx &= ~(CPUID_INTC_EDX_FPU | CPUID_INTC_EDX_MMX);
2257 
2258 		/*
2259 		 * Now map the supported feature vector to things that we
2260 		 * think userland will care about.
2261 		 */
2262 		if (*edx & CPUID_INTC_EDX_SEP)
2263 			hwcap_flags |= AV_386_SEP;
2264 		if (*edx & CPUID_INTC_EDX_SSE)
2265 			hwcap_flags |= AV_386_FXSR | AV_386_SSE;
2266 		if (*edx & CPUID_INTC_EDX_SSE2)
2267 			hwcap_flags |= AV_386_SSE2;
2268 		if (*ecx & CPUID_INTC_ECX_SSE3)
2269 			hwcap_flags |= AV_386_SSE3;
2270 		if (cpi->cpi_vendor == X86_VENDOR_Intel) {
2271 			if (*ecx & CPUID_INTC_ECX_SSSE3)
2272 				hwcap_flags |= AV_386_SSSE3;
2273 			if (*ecx & CPUID_INTC_ECX_SSE4_1)
2274 				hwcap_flags |= AV_386_SSE4_1;
2275 			if (*ecx & CPUID_INTC_ECX_SSE4_2)
2276 				hwcap_flags |= AV_386_SSE4_2;
2277 			if (*ecx & CPUID_INTC_ECX_MOVBE)
2278 				hwcap_flags |= AV_386_MOVBE;
2279 			if (*ecx & CPUID_INTC_ECX_AES)
2280 				hwcap_flags |= AV_386_AES;
2281 			if (*ecx & CPUID_INTC_ECX_PCLMULQDQ)
2282 				hwcap_flags |= AV_386_PCLMULQDQ;
2283 		}
2284 		if (*ecx & CPUID_INTC_ECX_POPCNT)
2285 			hwcap_flags |= AV_386_POPCNT;
2286 		if (*edx & CPUID_INTC_EDX_FPU)
2287 			hwcap_flags |= AV_386_FPU;
2288 		if (*edx & CPUID_INTC_EDX_MMX)
2289 			hwcap_flags |= AV_386_MMX;
2290 
2291 		if (*edx & CPUID_INTC_EDX_TSC)
2292 			hwcap_flags |= AV_386_TSC;
2293 		if (*edx & CPUID_INTC_EDX_CX8)
2294 			hwcap_flags |= AV_386_CX8;
2295 		if (*edx & CPUID_INTC_EDX_CMOV)
2296 			hwcap_flags |= AV_386_CMOV;
2297 		if (*ecx & CPUID_INTC_ECX_CX16)
2298 			hwcap_flags |= AV_386_CX16;
2299 	}
2300 
2301 	if (cpi->cpi_xmaxeax < 0x80000001)
2302 		goto pass4_done;
2303 
2304 	switch (cpi->cpi_vendor) {
2305 		struct cpuid_regs cp;
2306 		uint32_t *edx, *ecx;
2307 
2308 	case X86_VENDOR_Intel:
2309 		/*
2310 		 * Seems like Intel duplicated what we necessary
2311 		 * here to make the initial crop of 64-bit OS's work.
2312 		 * Hopefully, those are the only "extended" bits
2313 		 * they'll add.
2314 		 */
2315 		/*FALLTHROUGH*/
2316 
2317 	case X86_VENDOR_AMD:
2318 		edx = &cpi->cpi_support[AMD_EDX_FEATURES];
2319 		ecx = &cpi->cpi_support[AMD_ECX_FEATURES];
2320 
2321 		*edx = CPI_FEATURES_XTD_EDX(cpi);
2322 		*ecx = CPI_FEATURES_XTD_ECX(cpi);
2323 
2324 		/*
2325 		 * [these features require explicit kernel support]
2326 		 */
2327 		switch (cpi->cpi_vendor) {
2328 		case X86_VENDOR_Intel:
2329 			if ((x86_feature & X86_TSCP) == 0)
2330 				*edx &= ~CPUID_AMD_EDX_TSCP;
2331 			break;
2332 
2333 		case X86_VENDOR_AMD:
2334 			if ((x86_feature & X86_TSCP) == 0)
2335 				*edx &= ~CPUID_AMD_EDX_TSCP;
2336 			if ((x86_feature & X86_SSE4A) == 0)
2337 				*ecx &= ~CPUID_AMD_ECX_SSE4A;
2338 			break;
2339 
2340 		default:
2341 			break;
2342 		}
2343 
2344 		/*
2345 		 * [no explicit support required beyond
2346 		 * x87 fp context and exception handlers]
2347 		 */
2348 		if (!fpu_exists)
2349 			*edx &= ~(CPUID_AMD_EDX_MMXamd |
2350 			    CPUID_AMD_EDX_3DNow | CPUID_AMD_EDX_3DNowx);
2351 
2352 		if ((x86_feature & X86_NX) == 0)
2353 			*edx &= ~CPUID_AMD_EDX_NX;
2354 #if !defined(__amd64)
2355 		*edx &= ~CPUID_AMD_EDX_LM;
2356 #endif
2357 		/*
2358 		 * Now map the supported feature vector to
2359 		 * things that we think userland will care about.
2360 		 */
2361 #if defined(__amd64)
2362 		if (*edx & CPUID_AMD_EDX_SYSC)
2363 			hwcap_flags |= AV_386_AMD_SYSC;
2364 #endif
2365 		if (*edx & CPUID_AMD_EDX_MMXamd)
2366 			hwcap_flags |= AV_386_AMD_MMX;
2367 		if (*edx & CPUID_AMD_EDX_3DNow)
2368 			hwcap_flags |= AV_386_AMD_3DNow;
2369 		if (*edx & CPUID_AMD_EDX_3DNowx)
2370 			hwcap_flags |= AV_386_AMD_3DNowx;
2371 
2372 		switch (cpi->cpi_vendor) {
2373 		case X86_VENDOR_AMD:
2374 			if (*edx & CPUID_AMD_EDX_TSCP)
2375 				hwcap_flags |= AV_386_TSCP;
2376 			if (*ecx & CPUID_AMD_ECX_AHF64)
2377 				hwcap_flags |= AV_386_AHF;
2378 			if (*ecx & CPUID_AMD_ECX_SSE4A)
2379 				hwcap_flags |= AV_386_AMD_SSE4A;
2380 			if (*ecx & CPUID_AMD_ECX_LZCNT)
2381 				hwcap_flags |= AV_386_AMD_LZCNT;
2382 			break;
2383 
2384 		case X86_VENDOR_Intel:
2385 			if (*edx & CPUID_AMD_EDX_TSCP)
2386 				hwcap_flags |= AV_386_TSCP;
2387 			/*
2388 			 * Aarrgh.
2389 			 * Intel uses a different bit in the same word.
2390 			 */
2391 			if (*ecx & CPUID_INTC_ECX_AHF64)
2392 				hwcap_flags |= AV_386_AHF;
2393 			break;
2394 
2395 		default:
2396 			break;
2397 		}
2398 		break;
2399 
2400 	case X86_VENDOR_TM:
2401 		cp.cp_eax = 0x80860001;
2402 		(void) __cpuid_insn(&cp);
2403 		cpi->cpi_support[TM_EDX_FEATURES] = cp.cp_edx;
2404 		break;
2405 
2406 	default:
2407 		break;
2408 	}
2409 
2410 pass4_done:
2411 	cpi->cpi_pass = 4;
2412 	return (hwcap_flags);
2413 }
2414 
2415 
2416 /*
2417  * Simulate the cpuid instruction using the data we previously
2418  * captured about this CPU.  We try our best to return the truth
2419  * about the hardware, independently of kernel support.
2420  */
2421 uint32_t
2422 cpuid_insn(cpu_t *cpu, struct cpuid_regs *cp)
2423 {
2424 	struct cpuid_info *cpi;
2425 	struct cpuid_regs *xcp;
2426 
2427 	if (cpu == NULL)
2428 		cpu = CPU;
2429 	cpi = cpu->cpu_m.mcpu_cpi;
2430 
2431 	ASSERT(cpuid_checkpass(cpu, 3));
2432 
2433 	/*
2434 	 * CPUID data is cached in two separate places: cpi_std for standard
2435 	 * CPUID functions, and cpi_extd for extended CPUID functions.
2436 	 */
2437 	if (cp->cp_eax <= cpi->cpi_maxeax && cp->cp_eax < NMAX_CPI_STD)
2438 		xcp = &cpi->cpi_std[cp->cp_eax];
2439 	else if (cp->cp_eax >= 0x80000000 && cp->cp_eax <= cpi->cpi_xmaxeax &&
2440 	    cp->cp_eax < 0x80000000 + NMAX_CPI_EXTD)
2441 		xcp = &cpi->cpi_extd[cp->cp_eax - 0x80000000];
2442 	else
2443 		/*
2444 		 * The caller is asking for data from an input parameter which
2445 		 * the kernel has not cached.  In this case we go fetch from
2446 		 * the hardware and return the data directly to the user.
2447 		 */
2448 		return (__cpuid_insn(cp));
2449 
2450 	cp->cp_eax = xcp->cp_eax;
2451 	cp->cp_ebx = xcp->cp_ebx;
2452 	cp->cp_ecx = xcp->cp_ecx;
2453 	cp->cp_edx = xcp->cp_edx;
2454 	return (cp->cp_eax);
2455 }
2456 
2457 int
2458 cpuid_checkpass(cpu_t *cpu, int pass)
2459 {
2460 	return (cpu != NULL && cpu->cpu_m.mcpu_cpi != NULL &&
2461 	    cpu->cpu_m.mcpu_cpi->cpi_pass >= pass);
2462 }
2463 
2464 int
2465 cpuid_getbrandstr(cpu_t *cpu, char *s, size_t n)
2466 {
2467 	ASSERT(cpuid_checkpass(cpu, 3));
2468 
2469 	return (snprintf(s, n, "%s", cpu->cpu_m.mcpu_cpi->cpi_brandstr));
2470 }
2471 
2472 int
2473 cpuid_is_cmt(cpu_t *cpu)
2474 {
2475 	if (cpu == NULL)
2476 		cpu = CPU;
2477 
2478 	ASSERT(cpuid_checkpass(cpu, 1));
2479 
2480 	return (cpu->cpu_m.mcpu_cpi->cpi_chipid >= 0);
2481 }
2482 
2483 /*
2484  * AMD and Intel both implement the 64-bit variant of the syscall
2485  * instruction (syscallq), so if there's -any- support for syscall,
2486  * cpuid currently says "yes, we support this".
2487  *
2488  * However, Intel decided to -not- implement the 32-bit variant of the
2489  * syscall instruction, so we provide a predicate to allow our caller
2490  * to test that subtlety here.
2491  *
2492  * XXPV	Currently, 32-bit syscall instructions don't work via the hypervisor,
2493  *	even in the case where the hardware would in fact support it.
2494  */
2495 /*ARGSUSED*/
2496 int
2497 cpuid_syscall32_insn(cpu_t *cpu)
2498 {
2499 	ASSERT(cpuid_checkpass((cpu == NULL ? CPU : cpu), 1));
2500 
2501 #if !defined(__xpv)
2502 	if (cpu == NULL)
2503 		cpu = CPU;
2504 
2505 	/*CSTYLED*/
2506 	{
2507 		struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2508 
2509 		if (cpi->cpi_vendor == X86_VENDOR_AMD &&
2510 		    cpi->cpi_xmaxeax >= 0x80000001 &&
2511 		    (CPI_FEATURES_XTD_EDX(cpi) & CPUID_AMD_EDX_SYSC))
2512 			return (1);
2513 	}
2514 #endif
2515 	return (0);
2516 }
2517 
2518 int
2519 cpuid_getidstr(cpu_t *cpu, char *s, size_t n)
2520 {
2521 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2522 
2523 	static const char fmt[] =
2524 	    "x86 (%s %X family %d model %d step %d clock %d MHz)";
2525 	static const char fmt_ht[] =
2526 	    "x86 (chipid 0x%x %s %X family %d model %d step %d clock %d MHz)";
2527 
2528 	ASSERT(cpuid_checkpass(cpu, 1));
2529 
2530 	if (cpuid_is_cmt(cpu))
2531 		return (snprintf(s, n, fmt_ht, cpi->cpi_chipid,
2532 		    cpi->cpi_vendorstr, cpi->cpi_std[1].cp_eax,
2533 		    cpi->cpi_family, cpi->cpi_model,
2534 		    cpi->cpi_step, cpu->cpu_type_info.pi_clock));
2535 	return (snprintf(s, n, fmt,
2536 	    cpi->cpi_vendorstr, cpi->cpi_std[1].cp_eax,
2537 	    cpi->cpi_family, cpi->cpi_model,
2538 	    cpi->cpi_step, cpu->cpu_type_info.pi_clock));
2539 }
2540 
2541 const char *
2542 cpuid_getvendorstr(cpu_t *cpu)
2543 {
2544 	ASSERT(cpuid_checkpass(cpu, 1));
2545 	return ((const char *)cpu->cpu_m.mcpu_cpi->cpi_vendorstr);
2546 }
2547 
2548 uint_t
2549 cpuid_getvendor(cpu_t *cpu)
2550 {
2551 	ASSERT(cpuid_checkpass(cpu, 1));
2552 	return (cpu->cpu_m.mcpu_cpi->cpi_vendor);
2553 }
2554 
2555 uint_t
2556 cpuid_getfamily(cpu_t *cpu)
2557 {
2558 	ASSERT(cpuid_checkpass(cpu, 1));
2559 	return (cpu->cpu_m.mcpu_cpi->cpi_family);
2560 }
2561 
2562 uint_t
2563 cpuid_getmodel(cpu_t *cpu)
2564 {
2565 	ASSERT(cpuid_checkpass(cpu, 1));
2566 	return (cpu->cpu_m.mcpu_cpi->cpi_model);
2567 }
2568 
2569 uint_t
2570 cpuid_get_ncpu_per_chip(cpu_t *cpu)
2571 {
2572 	ASSERT(cpuid_checkpass(cpu, 1));
2573 	return (cpu->cpu_m.mcpu_cpi->cpi_ncpu_per_chip);
2574 }
2575 
2576 uint_t
2577 cpuid_get_ncore_per_chip(cpu_t *cpu)
2578 {
2579 	ASSERT(cpuid_checkpass(cpu, 1));
2580 	return (cpu->cpu_m.mcpu_cpi->cpi_ncore_per_chip);
2581 }
2582 
2583 uint_t
2584 cpuid_get_ncpu_sharing_last_cache(cpu_t *cpu)
2585 {
2586 	ASSERT(cpuid_checkpass(cpu, 2));
2587 	return (cpu->cpu_m.mcpu_cpi->cpi_ncpu_shr_last_cache);
2588 }
2589 
2590 id_t
2591 cpuid_get_last_lvl_cacheid(cpu_t *cpu)
2592 {
2593 	ASSERT(cpuid_checkpass(cpu, 2));
2594 	return (cpu->cpu_m.mcpu_cpi->cpi_last_lvl_cacheid);
2595 }
2596 
2597 uint_t
2598 cpuid_getstep(cpu_t *cpu)
2599 {
2600 	ASSERT(cpuid_checkpass(cpu, 1));
2601 	return (cpu->cpu_m.mcpu_cpi->cpi_step);
2602 }
2603 
2604 uint_t
2605 cpuid_getsig(struct cpu *cpu)
2606 {
2607 	ASSERT(cpuid_checkpass(cpu, 1));
2608 	return (cpu->cpu_m.mcpu_cpi->cpi_std[1].cp_eax);
2609 }
2610 
2611 uint32_t
2612 cpuid_getchiprev(struct cpu *cpu)
2613 {
2614 	ASSERT(cpuid_checkpass(cpu, 1));
2615 	return (cpu->cpu_m.mcpu_cpi->cpi_chiprev);
2616 }
2617 
2618 const char *
2619 cpuid_getchiprevstr(struct cpu *cpu)
2620 {
2621 	ASSERT(cpuid_checkpass(cpu, 1));
2622 	return (cpu->cpu_m.mcpu_cpi->cpi_chiprevstr);
2623 }
2624 
2625 uint32_t
2626 cpuid_getsockettype(struct cpu *cpu)
2627 {
2628 	ASSERT(cpuid_checkpass(cpu, 1));
2629 	return (cpu->cpu_m.mcpu_cpi->cpi_socket);
2630 }
2631 
2632 const char *
2633 cpuid_getsocketstr(cpu_t *cpu)
2634 {
2635 	static const char *socketstr = NULL;
2636 	struct cpuid_info *cpi;
2637 
2638 	ASSERT(cpuid_checkpass(cpu, 1));
2639 	cpi = cpu->cpu_m.mcpu_cpi;
2640 
2641 	/* Assume that socket types are the same across the system */
2642 	if (socketstr == NULL)
2643 		socketstr = _cpuid_sktstr(cpi->cpi_vendor, cpi->cpi_family,
2644 		    cpi->cpi_model, cpi->cpi_step);
2645 
2646 
2647 	return (socketstr);
2648 }
2649 
2650 int
2651 cpuid_get_chipid(cpu_t *cpu)
2652 {
2653 	ASSERT(cpuid_checkpass(cpu, 1));
2654 
2655 	if (cpuid_is_cmt(cpu))
2656 		return (cpu->cpu_m.mcpu_cpi->cpi_chipid);
2657 	return (cpu->cpu_id);
2658 }
2659 
2660 id_t
2661 cpuid_get_coreid(cpu_t *cpu)
2662 {
2663 	ASSERT(cpuid_checkpass(cpu, 1));
2664 	return (cpu->cpu_m.mcpu_cpi->cpi_coreid);
2665 }
2666 
2667 int
2668 cpuid_get_pkgcoreid(cpu_t *cpu)
2669 {
2670 	ASSERT(cpuid_checkpass(cpu, 1));
2671 	return (cpu->cpu_m.mcpu_cpi->cpi_pkgcoreid);
2672 }
2673 
2674 int
2675 cpuid_get_clogid(cpu_t *cpu)
2676 {
2677 	ASSERT(cpuid_checkpass(cpu, 1));
2678 	return (cpu->cpu_m.mcpu_cpi->cpi_clogid);
2679 }
2680 
2681 int
2682 cpuid_get_cacheid(cpu_t *cpu)
2683 {
2684 	ASSERT(cpuid_checkpass(cpu, 1));
2685 	return (cpu->cpu_m.mcpu_cpi->cpi_last_lvl_cacheid);
2686 }
2687 
2688 uint_t
2689 cpuid_get_procnodeid(cpu_t *cpu)
2690 {
2691 	ASSERT(cpuid_checkpass(cpu, 1));
2692 	return (cpu->cpu_m.mcpu_cpi->cpi_procnodeid);
2693 }
2694 
2695 uint_t
2696 cpuid_get_procnodes_per_pkg(cpu_t *cpu)
2697 {
2698 	ASSERT(cpuid_checkpass(cpu, 1));
2699 	return (cpu->cpu_m.mcpu_cpi->cpi_procnodes_per_pkg);
2700 }
2701 
2702 /*ARGSUSED*/
2703 int
2704 cpuid_have_cr8access(cpu_t *cpu)
2705 {
2706 #if defined(__amd64)
2707 	return (1);
2708 #else
2709 	struct cpuid_info *cpi;
2710 
2711 	ASSERT(cpu != NULL);
2712 	cpi = cpu->cpu_m.mcpu_cpi;
2713 	if (cpi->cpi_vendor == X86_VENDOR_AMD && cpi->cpi_maxeax >= 1 &&
2714 	    (CPI_FEATURES_XTD_ECX(cpi) & CPUID_AMD_ECX_CR8D) != 0)
2715 		return (1);
2716 	return (0);
2717 #endif
2718 }
2719 
2720 uint32_t
2721 cpuid_get_apicid(cpu_t *cpu)
2722 {
2723 	ASSERT(cpuid_checkpass(cpu, 1));
2724 	if (cpu->cpu_m.mcpu_cpi->cpi_maxeax < 1) {
2725 		return (UINT32_MAX);
2726 	} else {
2727 		return (cpu->cpu_m.mcpu_cpi->cpi_apicid);
2728 	}
2729 }
2730 
2731 void
2732 cpuid_get_addrsize(cpu_t *cpu, uint_t *pabits, uint_t *vabits)
2733 {
2734 	struct cpuid_info *cpi;
2735 
2736 	if (cpu == NULL)
2737 		cpu = CPU;
2738 	cpi = cpu->cpu_m.mcpu_cpi;
2739 
2740 	ASSERT(cpuid_checkpass(cpu, 1));
2741 
2742 	if (pabits)
2743 		*pabits = cpi->cpi_pabits;
2744 	if (vabits)
2745 		*vabits = cpi->cpi_vabits;
2746 }
2747 
2748 /*
2749  * Returns the number of data TLB entries for a corresponding
2750  * pagesize.  If it can't be computed, or isn't known, the
2751  * routine returns zero.  If you ask about an architecturally
2752  * impossible pagesize, the routine will panic (so that the
2753  * hat implementor knows that things are inconsistent.)
2754  */
2755 uint_t
2756 cpuid_get_dtlb_nent(cpu_t *cpu, size_t pagesize)
2757 {
2758 	struct cpuid_info *cpi;
2759 	uint_t dtlb_nent = 0;
2760 
2761 	if (cpu == NULL)
2762 		cpu = CPU;
2763 	cpi = cpu->cpu_m.mcpu_cpi;
2764 
2765 	ASSERT(cpuid_checkpass(cpu, 1));
2766 
2767 	/*
2768 	 * Check the L2 TLB info
2769 	 */
2770 	if (cpi->cpi_xmaxeax >= 0x80000006) {
2771 		struct cpuid_regs *cp = &cpi->cpi_extd[6];
2772 
2773 		switch (pagesize) {
2774 
2775 		case 4 * 1024:
2776 			/*
2777 			 * All zero in the top 16 bits of the register
2778 			 * indicates a unified TLB. Size is in low 16 bits.
2779 			 */
2780 			if ((cp->cp_ebx & 0xffff0000) == 0)
2781 				dtlb_nent = cp->cp_ebx & 0x0000ffff;
2782 			else
2783 				dtlb_nent = BITX(cp->cp_ebx, 27, 16);
2784 			break;
2785 
2786 		case 2 * 1024 * 1024:
2787 			if ((cp->cp_eax & 0xffff0000) == 0)
2788 				dtlb_nent = cp->cp_eax & 0x0000ffff;
2789 			else
2790 				dtlb_nent = BITX(cp->cp_eax, 27, 16);
2791 			break;
2792 
2793 		default:
2794 			panic("unknown L2 pagesize");
2795 			/*NOTREACHED*/
2796 		}
2797 	}
2798 
2799 	if (dtlb_nent != 0)
2800 		return (dtlb_nent);
2801 
2802 	/*
2803 	 * No L2 TLB support for this size, try L1.
2804 	 */
2805 	if (cpi->cpi_xmaxeax >= 0x80000005) {
2806 		struct cpuid_regs *cp = &cpi->cpi_extd[5];
2807 
2808 		switch (pagesize) {
2809 		case 4 * 1024:
2810 			dtlb_nent = BITX(cp->cp_ebx, 23, 16);
2811 			break;
2812 		case 2 * 1024 * 1024:
2813 			dtlb_nent = BITX(cp->cp_eax, 23, 16);
2814 			break;
2815 		default:
2816 			panic("unknown L1 d-TLB pagesize");
2817 			/*NOTREACHED*/
2818 		}
2819 	}
2820 
2821 	return (dtlb_nent);
2822 }
2823 
2824 /*
2825  * Return 0 if the erratum is not present or not applicable, positive
2826  * if it is, and negative if the status of the erratum is unknown.
2827  *
2828  * See "Revision Guide for AMD Athlon(tm) 64 and AMD Opteron(tm)
2829  * Processors" #25759, Rev 3.57, August 2005
2830  */
2831 int
2832 cpuid_opteron_erratum(cpu_t *cpu, uint_t erratum)
2833 {
2834 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2835 	uint_t eax;
2836 
2837 	/*
2838 	 * Bail out if this CPU isn't an AMD CPU, or if it's
2839 	 * a legacy (32-bit) AMD CPU.
2840 	 */
2841 	if (cpi->cpi_vendor != X86_VENDOR_AMD ||
2842 	    cpi->cpi_family == 4 || cpi->cpi_family == 5 ||
2843 	    cpi->cpi_family == 6)
2844 
2845 		return (0);
2846 
2847 	eax = cpi->cpi_std[1].cp_eax;
2848 
2849 #define	SH_B0(eax)	(eax == 0xf40 || eax == 0xf50)
2850 #define	SH_B3(eax) 	(eax == 0xf51)
2851 #define	B(eax)		(SH_B0(eax) || SH_B3(eax))
2852 
2853 #define	SH_C0(eax)	(eax == 0xf48 || eax == 0xf58)
2854 
2855 #define	SH_CG(eax)	(eax == 0xf4a || eax == 0xf5a || eax == 0xf7a)
2856 #define	DH_CG(eax)	(eax == 0xfc0 || eax == 0xfe0 || eax == 0xff0)
2857 #define	CH_CG(eax)	(eax == 0xf82 || eax == 0xfb2)
2858 #define	CG(eax)		(SH_CG(eax) || DH_CG(eax) || CH_CG(eax))
2859 
2860 #define	SH_D0(eax)	(eax == 0x10f40 || eax == 0x10f50 || eax == 0x10f70)
2861 #define	DH_D0(eax)	(eax == 0x10fc0 || eax == 0x10ff0)
2862 #define	CH_D0(eax)	(eax == 0x10f80 || eax == 0x10fb0)
2863 #define	D0(eax)		(SH_D0(eax) || DH_D0(eax) || CH_D0(eax))
2864 
2865 #define	SH_E0(eax)	(eax == 0x20f50 || eax == 0x20f40 || eax == 0x20f70)
2866 #define	JH_E1(eax)	(eax == 0x20f10)	/* JH8_E0 had 0x20f30 */
2867 #define	DH_E3(eax)	(eax == 0x20fc0 || eax == 0x20ff0)
2868 #define	SH_E4(eax)	(eax == 0x20f51 || eax == 0x20f71)
2869 #define	BH_E4(eax)	(eax == 0x20fb1)
2870 #define	SH_E5(eax)	(eax == 0x20f42)
2871 #define	DH_E6(eax)	(eax == 0x20ff2 || eax == 0x20fc2)
2872 #define	JH_E6(eax)	(eax == 0x20f12 || eax == 0x20f32)
2873 #define	EX(eax)		(SH_E0(eax) || JH_E1(eax) || DH_E3(eax) || \
2874 			    SH_E4(eax) || BH_E4(eax) || SH_E5(eax) || \
2875 			    DH_E6(eax) || JH_E6(eax))
2876 
2877 #define	DR_AX(eax)	(eax == 0x100f00 || eax == 0x100f01 || eax == 0x100f02)
2878 #define	DR_B0(eax)	(eax == 0x100f20)
2879 #define	DR_B1(eax)	(eax == 0x100f21)
2880 #define	DR_BA(eax)	(eax == 0x100f2a)
2881 #define	DR_B2(eax)	(eax == 0x100f22)
2882 #define	DR_B3(eax)	(eax == 0x100f23)
2883 #define	RB_C0(eax)	(eax == 0x100f40)
2884 
2885 	switch (erratum) {
2886 	case 1:
2887 		return (cpi->cpi_family < 0x10);
2888 	case 51:	/* what does the asterisk mean? */
2889 		return (B(eax) || SH_C0(eax) || CG(eax));
2890 	case 52:
2891 		return (B(eax));
2892 	case 57:
2893 		return (cpi->cpi_family <= 0x11);
2894 	case 58:
2895 		return (B(eax));
2896 	case 60:
2897 		return (cpi->cpi_family <= 0x11);
2898 	case 61:
2899 	case 62:
2900 	case 63:
2901 	case 64:
2902 	case 65:
2903 	case 66:
2904 	case 68:
2905 	case 69:
2906 	case 70:
2907 	case 71:
2908 		return (B(eax));
2909 	case 72:
2910 		return (SH_B0(eax));
2911 	case 74:
2912 		return (B(eax));
2913 	case 75:
2914 		return (cpi->cpi_family < 0x10);
2915 	case 76:
2916 		return (B(eax));
2917 	case 77:
2918 		return (cpi->cpi_family <= 0x11);
2919 	case 78:
2920 		return (B(eax) || SH_C0(eax));
2921 	case 79:
2922 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
2923 	case 80:
2924 	case 81:
2925 	case 82:
2926 		return (B(eax));
2927 	case 83:
2928 		return (B(eax) || SH_C0(eax) || CG(eax));
2929 	case 85:
2930 		return (cpi->cpi_family < 0x10);
2931 	case 86:
2932 		return (SH_C0(eax) || CG(eax));
2933 	case 88:
2934 #if !defined(__amd64)
2935 		return (0);
2936 #else
2937 		return (B(eax) || SH_C0(eax));
2938 #endif
2939 	case 89:
2940 		return (cpi->cpi_family < 0x10);
2941 	case 90:
2942 		return (B(eax) || SH_C0(eax) || CG(eax));
2943 	case 91:
2944 	case 92:
2945 		return (B(eax) || SH_C0(eax));
2946 	case 93:
2947 		return (SH_C0(eax));
2948 	case 94:
2949 		return (B(eax) || SH_C0(eax) || CG(eax));
2950 	case 95:
2951 #if !defined(__amd64)
2952 		return (0);
2953 #else
2954 		return (B(eax) || SH_C0(eax));
2955 #endif
2956 	case 96:
2957 		return (B(eax) || SH_C0(eax) || CG(eax));
2958 	case 97:
2959 	case 98:
2960 		return (SH_C0(eax) || CG(eax));
2961 	case 99:
2962 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
2963 	case 100:
2964 		return (B(eax) || SH_C0(eax));
2965 	case 101:
2966 	case 103:
2967 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
2968 	case 104:
2969 		return (SH_C0(eax) || CG(eax) || D0(eax));
2970 	case 105:
2971 	case 106:
2972 	case 107:
2973 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
2974 	case 108:
2975 		return (DH_CG(eax));
2976 	case 109:
2977 		return (SH_C0(eax) || CG(eax) || D0(eax));
2978 	case 110:
2979 		return (D0(eax) || EX(eax));
2980 	case 111:
2981 		return (CG(eax));
2982 	case 112:
2983 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
2984 	case 113:
2985 		return (eax == 0x20fc0);
2986 	case 114:
2987 		return (SH_E0(eax) || JH_E1(eax) || DH_E3(eax));
2988 	case 115:
2989 		return (SH_E0(eax) || JH_E1(eax));
2990 	case 116:
2991 		return (SH_E0(eax) || JH_E1(eax) || DH_E3(eax));
2992 	case 117:
2993 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
2994 	case 118:
2995 		return (SH_E0(eax) || JH_E1(eax) || SH_E4(eax) || BH_E4(eax) ||
2996 		    JH_E6(eax));
2997 	case 121:
2998 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
2999 	case 122:
3000 		return (cpi->cpi_family < 0x10 || cpi->cpi_family == 0x11);
3001 	case 123:
3002 		return (JH_E1(eax) || BH_E4(eax) || JH_E6(eax));
3003 	case 131:
3004 		return (cpi->cpi_family < 0x10);
3005 	case 6336786:
3006 		/*
3007 		 * Test for AdvPowerMgmtInfo.TscPStateInvariant
3008 		 * if this is a K8 family or newer processor
3009 		 */
3010 		if (CPI_FAMILY(cpi) == 0xf) {
3011 			struct cpuid_regs regs;
3012 			regs.cp_eax = 0x80000007;
3013 			(void) __cpuid_insn(&regs);
3014 			return (!(regs.cp_edx & 0x100));
3015 		}
3016 		return (0);
3017 	case 6323525:
3018 		return (((((eax >> 12) & 0xff00) + (eax & 0xf00)) |
3019 		    (((eax >> 4) & 0xf) | ((eax >> 12) & 0xf0))) < 0xf40);
3020 
3021 	case 6671130:
3022 		/*
3023 		 * check for processors (pre-Shanghai) that do not provide
3024 		 * optimal management of 1gb ptes in its tlb.
3025 		 */
3026 		return (cpi->cpi_family == 0x10 && cpi->cpi_model < 4);
3027 
3028 	case 298:
3029 		return (DR_AX(eax) || DR_B0(eax) || DR_B1(eax) || DR_BA(eax) ||
3030 		    DR_B2(eax) || RB_C0(eax));
3031 
3032 	default:
3033 		return (-1);
3034 
3035 	}
3036 }
3037 
3038 /*
3039  * Determine if specified erratum is present via OSVW (OS Visible Workaround).
3040  * Return 1 if erratum is present, 0 if not present and -1 if indeterminate.
3041  */
3042 int
3043 osvw_opteron_erratum(cpu_t *cpu, uint_t erratum)
3044 {
3045 	struct cpuid_info	*cpi;
3046 	uint_t			osvwid;
3047 	static int		osvwfeature = -1;
3048 	uint64_t		osvwlength;
3049 
3050 
3051 	cpi = cpu->cpu_m.mcpu_cpi;
3052 
3053 	/* confirm OSVW supported */
3054 	if (osvwfeature == -1) {
3055 		osvwfeature = cpi->cpi_extd[1].cp_ecx & CPUID_AMD_ECX_OSVW;
3056 	} else {
3057 		/* assert that osvw feature setting is consistent on all cpus */
3058 		ASSERT(osvwfeature ==
3059 		    (cpi->cpi_extd[1].cp_ecx & CPUID_AMD_ECX_OSVW));
3060 	}
3061 	if (!osvwfeature)
3062 		return (-1);
3063 
3064 	osvwlength = rdmsr(MSR_AMD_OSVW_ID_LEN) & OSVW_ID_LEN_MASK;
3065 
3066 	switch (erratum) {
3067 	case 298:	/* osvwid is 0 */
3068 		osvwid = 0;
3069 		if (osvwlength <= (uint64_t)osvwid) {
3070 			/* osvwid 0 is unknown */
3071 			return (-1);
3072 		}
3073 
3074 		/*
3075 		 * Check the OSVW STATUS MSR to determine the state
3076 		 * of the erratum where:
3077 		 *   0 - fixed by HW
3078 		 *   1 - BIOS has applied the workaround when BIOS
3079 		 *   workaround is available. (Or for other errata,
3080 		 *   OS workaround is required.)
3081 		 * For a value of 1, caller will confirm that the
3082 		 * erratum 298 workaround has indeed been applied by BIOS.
3083 		 *
3084 		 * A 1 may be set in cpus that have a HW fix
3085 		 * in a mixed cpu system. Regarding erratum 298:
3086 		 *   In a multiprocessor platform, the workaround above
3087 		 *   should be applied to all processors regardless of
3088 		 *   silicon revision when an affected processor is
3089 		 *   present.
3090 		 */
3091 
3092 		return (rdmsr(MSR_AMD_OSVW_STATUS +
3093 		    (osvwid / OSVW_ID_CNT_PER_MSR)) &
3094 		    (1ULL << (osvwid % OSVW_ID_CNT_PER_MSR)));
3095 
3096 	default:
3097 		return (-1);
3098 	}
3099 }
3100 
3101 static const char assoc_str[] = "associativity";
3102 static const char line_str[] = "line-size";
3103 static const char size_str[] = "size";
3104 
3105 static void
3106 add_cache_prop(dev_info_t *devi, const char *label, const char *type,
3107     uint32_t val)
3108 {
3109 	char buf[128];
3110 
3111 	/*
3112 	 * ndi_prop_update_int() is used because it is desirable for
3113 	 * DDI_PROP_HW_DEF and DDI_PROP_DONTSLEEP to be set.
3114 	 */
3115 	if (snprintf(buf, sizeof (buf), "%s-%s", label, type) < sizeof (buf))
3116 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, devi, buf, val);
3117 }
3118 
3119 /*
3120  * Intel-style cache/tlb description
3121  *
3122  * Standard cpuid level 2 gives a randomly ordered
3123  * selection of tags that index into a table that describes
3124  * cache and tlb properties.
3125  */
3126 
3127 static const char l1_icache_str[] = "l1-icache";
3128 static const char l1_dcache_str[] = "l1-dcache";
3129 static const char l2_cache_str[] = "l2-cache";
3130 static const char l3_cache_str[] = "l3-cache";
3131 static const char itlb4k_str[] = "itlb-4K";
3132 static const char dtlb4k_str[] = "dtlb-4K";
3133 static const char itlb2M_str[] = "itlb-2M";
3134 static const char itlb4M_str[] = "itlb-4M";
3135 static const char dtlb4M_str[] = "dtlb-4M";
3136 static const char dtlb24_str[] = "dtlb0-2M-4M";
3137 static const char itlb424_str[] = "itlb-4K-2M-4M";
3138 static const char itlb24_str[] = "itlb-2M-4M";
3139 static const char dtlb44_str[] = "dtlb-4K-4M";
3140 static const char sl1_dcache_str[] = "sectored-l1-dcache";
3141 static const char sl2_cache_str[] = "sectored-l2-cache";
3142 static const char itrace_str[] = "itrace-cache";
3143 static const char sl3_cache_str[] = "sectored-l3-cache";
3144 static const char sh_l2_tlb4k_str[] = "shared-l2-tlb-4k";
3145 
3146 static const struct cachetab {
3147 	uint8_t 	ct_code;
3148 	uint8_t		ct_assoc;
3149 	uint16_t 	ct_line_size;
3150 	size_t		ct_size;
3151 	const char	*ct_label;
3152 } intel_ctab[] = {
3153 	/*
3154 	 * maintain descending order!
3155 	 *
3156 	 * Codes ignored - Reason
3157 	 * ----------------------
3158 	 * 40H - intel_cpuid_4_cache_info() disambiguates l2/l3 cache
3159 	 * f0H/f1H - Currently we do not interpret prefetch size by design
3160 	 */
3161 	{ 0xe4, 16, 64, 8*1024*1024, l3_cache_str},
3162 	{ 0xe3, 16, 64, 4*1024*1024, l3_cache_str},
3163 	{ 0xe2, 16, 64, 2*1024*1024, l3_cache_str},
3164 	{ 0xde, 12, 64, 6*1024*1024, l3_cache_str},
3165 	{ 0xdd, 12, 64, 3*1024*1024, l3_cache_str},
3166 	{ 0xdc, 12, 64, ((1*1024*1024)+(512*1024)), l3_cache_str},
3167 	{ 0xd8, 8, 64, 4*1024*1024, l3_cache_str},
3168 	{ 0xd7, 8, 64, 2*1024*1024, l3_cache_str},
3169 	{ 0xd6, 8, 64, 1*1024*1024, l3_cache_str},
3170 	{ 0xd2, 4, 64, 2*1024*1024, l3_cache_str},
3171 	{ 0xd1, 4, 64, 1*1024*1024, l3_cache_str},
3172 	{ 0xd0, 4, 64, 512*1024, l3_cache_str},
3173 	{ 0xca, 4, 0, 512, sh_l2_tlb4k_str},
3174 	{ 0xc0, 4, 0, 8, dtlb44_str },
3175 	{ 0xba, 4, 0, 64, dtlb4k_str },
3176 	{ 0xb4, 4, 0, 256, dtlb4k_str },
3177 	{ 0xb3, 4, 0, 128, dtlb4k_str },
3178 	{ 0xb2, 4, 0, 64, itlb4k_str },
3179 	{ 0xb0, 4, 0, 128, itlb4k_str },
3180 	{ 0x87, 8, 64, 1024*1024, l2_cache_str},
3181 	{ 0x86, 4, 64, 512*1024, l2_cache_str},
3182 	{ 0x85, 8, 32, 2*1024*1024, l2_cache_str},
3183 	{ 0x84, 8, 32, 1024*1024, l2_cache_str},
3184 	{ 0x83, 8, 32, 512*1024, l2_cache_str},
3185 	{ 0x82, 8, 32, 256*1024, l2_cache_str},
3186 	{ 0x80, 8, 64, 512*1024, l2_cache_str},
3187 	{ 0x7f, 2, 64, 512*1024, l2_cache_str},
3188 	{ 0x7d, 8, 64, 2*1024*1024, sl2_cache_str},
3189 	{ 0x7c, 8, 64, 1024*1024, sl2_cache_str},
3190 	{ 0x7b, 8, 64, 512*1024, sl2_cache_str},
3191 	{ 0x7a, 8, 64, 256*1024, sl2_cache_str},
3192 	{ 0x79, 8, 64, 128*1024, sl2_cache_str},
3193 	{ 0x78, 8, 64, 1024*1024, l2_cache_str},
3194 	{ 0x73, 8, 0, 64*1024, itrace_str},
3195 	{ 0x72, 8, 0, 32*1024, itrace_str},
3196 	{ 0x71, 8, 0, 16*1024, itrace_str},
3197 	{ 0x70, 8, 0, 12*1024, itrace_str},
3198 	{ 0x68, 4, 64, 32*1024, sl1_dcache_str},
3199 	{ 0x67, 4, 64, 16*1024, sl1_dcache_str},
3200 	{ 0x66, 4, 64, 8*1024, sl1_dcache_str},
3201 	{ 0x60, 8, 64, 16*1024, sl1_dcache_str},
3202 	{ 0x5d, 0, 0, 256, dtlb44_str},
3203 	{ 0x5c, 0, 0, 128, dtlb44_str},
3204 	{ 0x5b, 0, 0, 64, dtlb44_str},
3205 	{ 0x5a, 4, 0, 32, dtlb24_str},
3206 	{ 0x59, 0, 0, 16, dtlb4k_str},
3207 	{ 0x57, 4, 0, 16, dtlb4k_str},
3208 	{ 0x56, 4, 0, 16, dtlb4M_str},
3209 	{ 0x55, 0, 0, 7, itlb24_str},
3210 	{ 0x52, 0, 0, 256, itlb424_str},
3211 	{ 0x51, 0, 0, 128, itlb424_str},
3212 	{ 0x50, 0, 0, 64, itlb424_str},
3213 	{ 0x4f, 0, 0, 32, itlb4k_str},
3214 	{ 0x4e, 24, 64, 6*1024*1024, l2_cache_str},
3215 	{ 0x4d, 16, 64, 16*1024*1024, l3_cache_str},
3216 	{ 0x4c, 12, 64, 12*1024*1024, l3_cache_str},
3217 	{ 0x4b, 16, 64, 8*1024*1024, l3_cache_str},
3218 	{ 0x4a, 12, 64, 6*1024*1024, l3_cache_str},
3219 	{ 0x49, 16, 64, 4*1024*1024, l3_cache_str},
3220 	{ 0x48, 12, 64, 3*1024*1024, l2_cache_str},
3221 	{ 0x47, 8, 64, 8*1024*1024, l3_cache_str},
3222 	{ 0x46, 4, 64, 4*1024*1024, l3_cache_str},
3223 	{ 0x45, 4, 32, 2*1024*1024, l2_cache_str},
3224 	{ 0x44, 4, 32, 1024*1024, l2_cache_str},
3225 	{ 0x43, 4, 32, 512*1024, l2_cache_str},
3226 	{ 0x42, 4, 32, 256*1024, l2_cache_str},
3227 	{ 0x41, 4, 32, 128*1024, l2_cache_str},
3228 	{ 0x3e, 4, 64, 512*1024, sl2_cache_str},
3229 	{ 0x3d, 6, 64, 384*1024, sl2_cache_str},
3230 	{ 0x3c, 4, 64, 256*1024, sl2_cache_str},
3231 	{ 0x3b, 2, 64, 128*1024, sl2_cache_str},
3232 	{ 0x3a, 6, 64, 192*1024, sl2_cache_str},
3233 	{ 0x39, 4, 64, 128*1024, sl2_cache_str},
3234 	{ 0x30, 8, 64, 32*1024, l1_icache_str},
3235 	{ 0x2c, 8, 64, 32*1024, l1_dcache_str},
3236 	{ 0x29, 8, 64, 4096*1024, sl3_cache_str},
3237 	{ 0x25, 8, 64, 2048*1024, sl3_cache_str},
3238 	{ 0x23, 8, 64, 1024*1024, sl3_cache_str},
3239 	{ 0x22, 4, 64, 512*1024, sl3_cache_str},
3240 	{ 0x0e, 6, 64, 24*1024, l1_dcache_str},
3241 	{ 0x0d, 4, 32, 16*1024, l1_dcache_str},
3242 	{ 0x0c, 4, 32, 16*1024, l1_dcache_str},
3243 	{ 0x0b, 4, 0, 4, itlb4M_str},
3244 	{ 0x0a, 2, 32, 8*1024, l1_dcache_str},
3245 	{ 0x08, 4, 32, 16*1024, l1_icache_str},
3246 	{ 0x06, 4, 32, 8*1024, l1_icache_str},
3247 	{ 0x05, 4, 0, 32, dtlb4M_str},
3248 	{ 0x04, 4, 0, 8, dtlb4M_str},
3249 	{ 0x03, 4, 0, 64, dtlb4k_str},
3250 	{ 0x02, 4, 0, 2, itlb4M_str},
3251 	{ 0x01, 4, 0, 32, itlb4k_str},
3252 	{ 0 }
3253 };
3254 
3255 static const struct cachetab cyrix_ctab[] = {
3256 	{ 0x70, 4, 0, 32, "tlb-4K" },
3257 	{ 0x80, 4, 16, 16*1024, "l1-cache" },
3258 	{ 0 }
3259 };
3260 
3261 /*
3262  * Search a cache table for a matching entry
3263  */
3264 static const struct cachetab *
3265 find_cacheent(const struct cachetab *ct, uint_t code)
3266 {
3267 	if (code != 0) {
3268 		for (; ct->ct_code != 0; ct++)
3269 			if (ct->ct_code <= code)
3270 				break;
3271 		if (ct->ct_code == code)
3272 			return (ct);
3273 	}
3274 	return (NULL);
3275 }
3276 
3277 /*
3278  * Populate cachetab entry with L2 or L3 cache-information using
3279  * cpuid function 4. This function is called from intel_walk_cacheinfo()
3280  * when descriptor 0x49 is encountered. It returns 0 if no such cache
3281  * information is found.
3282  */
3283 static int
3284 intel_cpuid_4_cache_info(struct cachetab *ct, struct cpuid_info *cpi)
3285 {
3286 	uint32_t level, i;
3287 	int ret = 0;
3288 
3289 	for (i = 0; i < cpi->cpi_std_4_size; i++) {
3290 		level = CPI_CACHE_LVL(cpi->cpi_std_4[i]);
3291 
3292 		if (level == 2 || level == 3) {
3293 			ct->ct_assoc = CPI_CACHE_WAYS(cpi->cpi_std_4[i]) + 1;
3294 			ct->ct_line_size =
3295 			    CPI_CACHE_COH_LN_SZ(cpi->cpi_std_4[i]) + 1;
3296 			ct->ct_size = ct->ct_assoc *
3297 			    (CPI_CACHE_PARTS(cpi->cpi_std_4[i]) + 1) *
3298 			    ct->ct_line_size *
3299 			    (cpi->cpi_std_4[i]->cp_ecx + 1);
3300 
3301 			if (level == 2) {
3302 				ct->ct_label = l2_cache_str;
3303 			} else if (level == 3) {
3304 				ct->ct_label = l3_cache_str;
3305 			}
3306 			ret = 1;
3307 		}
3308 	}
3309 
3310 	return (ret);
3311 }
3312 
3313 /*
3314  * Walk the cacheinfo descriptor, applying 'func' to every valid element
3315  * The walk is terminated if the walker returns non-zero.
3316  */
3317 static void
3318 intel_walk_cacheinfo(struct cpuid_info *cpi,
3319     void *arg, int (*func)(void *, const struct cachetab *))
3320 {
3321 	const struct cachetab *ct;
3322 	struct cachetab des_49_ct, des_b1_ct;
3323 	uint8_t *dp;
3324 	int i;
3325 
3326 	if ((dp = cpi->cpi_cacheinfo) == NULL)
3327 		return;
3328 	for (i = 0; i < cpi->cpi_ncache; i++, dp++) {
3329 		/*
3330 		 * For overloaded descriptor 0x49 we use cpuid function 4
3331 		 * if supported by the current processor, to create
3332 		 * cache information.
3333 		 * For overloaded descriptor 0xb1 we use X86_PAE flag
3334 		 * to disambiguate the cache information.
3335 		 */
3336 		if (*dp == 0x49 && cpi->cpi_maxeax >= 0x4 &&
3337 		    intel_cpuid_4_cache_info(&des_49_ct, cpi) == 1) {
3338 				ct = &des_49_ct;
3339 		} else if (*dp == 0xb1) {
3340 			des_b1_ct.ct_code = 0xb1;
3341 			des_b1_ct.ct_assoc = 4;
3342 			des_b1_ct.ct_line_size = 0;
3343 			if (x86_feature & X86_PAE) {
3344 				des_b1_ct.ct_size = 8;
3345 				des_b1_ct.ct_label = itlb2M_str;
3346 			} else {
3347 				des_b1_ct.ct_size = 4;
3348 				des_b1_ct.ct_label = itlb4M_str;
3349 			}
3350 			ct = &des_b1_ct;
3351 		} else {
3352 			if ((ct = find_cacheent(intel_ctab, *dp)) == NULL) {
3353 				continue;
3354 			}
3355 		}
3356 
3357 		if (func(arg, ct) != 0) {
3358 			break;
3359 		}
3360 	}
3361 }
3362 
3363 /*
3364  * (Like the Intel one, except for Cyrix CPUs)
3365  */
3366 static void
3367 cyrix_walk_cacheinfo(struct cpuid_info *cpi,
3368     void *arg, int (*func)(void *, const struct cachetab *))
3369 {
3370 	const struct cachetab *ct;
3371 	uint8_t *dp;
3372 	int i;
3373 
3374 	if ((dp = cpi->cpi_cacheinfo) == NULL)
3375 		return;
3376 	for (i = 0; i < cpi->cpi_ncache; i++, dp++) {
3377 		/*
3378 		 * Search Cyrix-specific descriptor table first ..
3379 		 */
3380 		if ((ct = find_cacheent(cyrix_ctab, *dp)) != NULL) {
3381 			if (func(arg, ct) != 0)
3382 				break;
3383 			continue;
3384 		}
3385 		/*
3386 		 * .. else fall back to the Intel one
3387 		 */
3388 		if ((ct = find_cacheent(intel_ctab, *dp)) != NULL) {
3389 			if (func(arg, ct) != 0)
3390 				break;
3391 			continue;
3392 		}
3393 	}
3394 }
3395 
3396 /*
3397  * A cacheinfo walker that adds associativity, line-size, and size properties
3398  * to the devinfo node it is passed as an argument.
3399  */
3400 static int
3401 add_cacheent_props(void *arg, const struct cachetab *ct)
3402 {
3403 	dev_info_t *devi = arg;
3404 
3405 	add_cache_prop(devi, ct->ct_label, assoc_str, ct->ct_assoc);
3406 	if (ct->ct_line_size != 0)
3407 		add_cache_prop(devi, ct->ct_label, line_str,
3408 		    ct->ct_line_size);
3409 	add_cache_prop(devi, ct->ct_label, size_str, ct->ct_size);
3410 	return (0);
3411 }
3412 
3413 
3414 static const char fully_assoc[] = "fully-associative?";
3415 
3416 /*
3417  * AMD style cache/tlb description
3418  *
3419  * Extended functions 5 and 6 directly describe properties of
3420  * tlbs and various cache levels.
3421  */
3422 static void
3423 add_amd_assoc(dev_info_t *devi, const char *label, uint_t assoc)
3424 {
3425 	switch (assoc) {
3426 	case 0:	/* reserved; ignore */
3427 		break;
3428 	default:
3429 		add_cache_prop(devi, label, assoc_str, assoc);
3430 		break;
3431 	case 0xff:
3432 		add_cache_prop(devi, label, fully_assoc, 1);
3433 		break;
3434 	}
3435 }
3436 
3437 static void
3438 add_amd_tlb(dev_info_t *devi, const char *label, uint_t assoc, uint_t size)
3439 {
3440 	if (size == 0)
3441 		return;
3442 	add_cache_prop(devi, label, size_str, size);
3443 	add_amd_assoc(devi, label, assoc);
3444 }
3445 
3446 static void
3447 add_amd_cache(dev_info_t *devi, const char *label,
3448     uint_t size, uint_t assoc, uint_t lines_per_tag, uint_t line_size)
3449 {
3450 	if (size == 0 || line_size == 0)
3451 		return;
3452 	add_amd_assoc(devi, label, assoc);
3453 	/*
3454 	 * Most AMD parts have a sectored cache. Multiple cache lines are
3455 	 * associated with each tag. A sector consists of all cache lines
3456 	 * associated with a tag. For example, the AMD K6-III has a sector
3457 	 * size of 2 cache lines per tag.
3458 	 */
3459 	if (lines_per_tag != 0)
3460 		add_cache_prop(devi, label, "lines-per-tag", lines_per_tag);
3461 	add_cache_prop(devi, label, line_str, line_size);
3462 	add_cache_prop(devi, label, size_str, size * 1024);
3463 }
3464 
3465 static void
3466 add_amd_l2_assoc(dev_info_t *devi, const char *label, uint_t assoc)
3467 {
3468 	switch (assoc) {
3469 	case 0:	/* off */
3470 		break;
3471 	case 1:
3472 	case 2:
3473 	case 4:
3474 		add_cache_prop(devi, label, assoc_str, assoc);
3475 		break;
3476 	case 6:
3477 		add_cache_prop(devi, label, assoc_str, 8);
3478 		break;
3479 	case 8:
3480 		add_cache_prop(devi, label, assoc_str, 16);
3481 		break;
3482 	case 0xf:
3483 		add_cache_prop(devi, label, fully_assoc, 1);
3484 		break;
3485 	default: /* reserved; ignore */
3486 		break;
3487 	}
3488 }
3489 
3490 static void
3491 add_amd_l2_tlb(dev_info_t *devi, const char *label, uint_t assoc, uint_t size)
3492 {
3493 	if (size == 0 || assoc == 0)
3494 		return;
3495 	add_amd_l2_assoc(devi, label, assoc);
3496 	add_cache_prop(devi, label, size_str, size);
3497 }
3498 
3499 static void
3500 add_amd_l2_cache(dev_info_t *devi, const char *label,
3501     uint_t size, uint_t assoc, uint_t lines_per_tag, uint_t line_size)
3502 {
3503 	if (size == 0 || assoc == 0 || line_size == 0)
3504 		return;
3505 	add_amd_l2_assoc(devi, label, assoc);
3506 	if (lines_per_tag != 0)
3507 		add_cache_prop(devi, label, "lines-per-tag", lines_per_tag);
3508 	add_cache_prop(devi, label, line_str, line_size);
3509 	add_cache_prop(devi, label, size_str, size * 1024);
3510 }
3511 
3512 static void
3513 amd_cache_info(struct cpuid_info *cpi, dev_info_t *devi)
3514 {
3515 	struct cpuid_regs *cp;
3516 
3517 	if (cpi->cpi_xmaxeax < 0x80000005)
3518 		return;
3519 	cp = &cpi->cpi_extd[5];
3520 
3521 	/*
3522 	 * 4M/2M L1 TLB configuration
3523 	 *
3524 	 * We report the size for 2M pages because AMD uses two
3525 	 * TLB entries for one 4M page.
3526 	 */
3527 	add_amd_tlb(devi, "dtlb-2M",
3528 	    BITX(cp->cp_eax, 31, 24), BITX(cp->cp_eax, 23, 16));
3529 	add_amd_tlb(devi, "itlb-2M",
3530 	    BITX(cp->cp_eax, 15, 8), BITX(cp->cp_eax, 7, 0));
3531 
3532 	/*
3533 	 * 4K L1 TLB configuration
3534 	 */
3535 
3536 	switch (cpi->cpi_vendor) {
3537 		uint_t nentries;
3538 	case X86_VENDOR_TM:
3539 		if (cpi->cpi_family >= 5) {
3540 			/*
3541 			 * Crusoe processors have 256 TLB entries, but
3542 			 * cpuid data format constrains them to only
3543 			 * reporting 255 of them.
3544 			 */
3545 			if ((nentries = BITX(cp->cp_ebx, 23, 16)) == 255)
3546 				nentries = 256;
3547 			/*
3548 			 * Crusoe processors also have a unified TLB
3549 			 */
3550 			add_amd_tlb(devi, "tlb-4K", BITX(cp->cp_ebx, 31, 24),
3551 			    nentries);
3552 			break;
3553 		}
3554 		/*FALLTHROUGH*/
3555 	default:
3556 		add_amd_tlb(devi, itlb4k_str,
3557 		    BITX(cp->cp_ebx, 31, 24), BITX(cp->cp_ebx, 23, 16));
3558 		add_amd_tlb(devi, dtlb4k_str,
3559 		    BITX(cp->cp_ebx, 15, 8), BITX(cp->cp_ebx, 7, 0));
3560 		break;
3561 	}
3562 
3563 	/*
3564 	 * data L1 cache configuration
3565 	 */
3566 
3567 	add_amd_cache(devi, l1_dcache_str,
3568 	    BITX(cp->cp_ecx, 31, 24), BITX(cp->cp_ecx, 23, 16),
3569 	    BITX(cp->cp_ecx, 15, 8), BITX(cp->cp_ecx, 7, 0));
3570 
3571 	/*
3572 	 * code L1 cache configuration
3573 	 */
3574 
3575 	add_amd_cache(devi, l1_icache_str,
3576 	    BITX(cp->cp_edx, 31, 24), BITX(cp->cp_edx, 23, 16),
3577 	    BITX(cp->cp_edx, 15, 8), BITX(cp->cp_edx, 7, 0));
3578 
3579 	if (cpi->cpi_xmaxeax < 0x80000006)
3580 		return;
3581 	cp = &cpi->cpi_extd[6];
3582 
3583 	/* Check for a unified L2 TLB for large pages */
3584 
3585 	if (BITX(cp->cp_eax, 31, 16) == 0)
3586 		add_amd_l2_tlb(devi, "l2-tlb-2M",
3587 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
3588 	else {
3589 		add_amd_l2_tlb(devi, "l2-dtlb-2M",
3590 		    BITX(cp->cp_eax, 31, 28), BITX(cp->cp_eax, 27, 16));
3591 		add_amd_l2_tlb(devi, "l2-itlb-2M",
3592 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
3593 	}
3594 
3595 	/* Check for a unified L2 TLB for 4K pages */
3596 
3597 	if (BITX(cp->cp_ebx, 31, 16) == 0) {
3598 		add_amd_l2_tlb(devi, "l2-tlb-4K",
3599 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
3600 	} else {
3601 		add_amd_l2_tlb(devi, "l2-dtlb-4K",
3602 		    BITX(cp->cp_eax, 31, 28), BITX(cp->cp_eax, 27, 16));
3603 		add_amd_l2_tlb(devi, "l2-itlb-4K",
3604 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
3605 	}
3606 
3607 	add_amd_l2_cache(devi, l2_cache_str,
3608 	    BITX(cp->cp_ecx, 31, 16), BITX(cp->cp_ecx, 15, 12),
3609 	    BITX(cp->cp_ecx, 11, 8), BITX(cp->cp_ecx, 7, 0));
3610 }
3611 
3612 /*
3613  * There are two basic ways that the x86 world describes it cache
3614  * and tlb architecture - Intel's way and AMD's way.
3615  *
3616  * Return which flavor of cache architecture we should use
3617  */
3618 static int
3619 x86_which_cacheinfo(struct cpuid_info *cpi)
3620 {
3621 	switch (cpi->cpi_vendor) {
3622 	case X86_VENDOR_Intel:
3623 		if (cpi->cpi_maxeax >= 2)
3624 			return (X86_VENDOR_Intel);
3625 		break;
3626 	case X86_VENDOR_AMD:
3627 		/*
3628 		 * The K5 model 1 was the first part from AMD that reported
3629 		 * cache sizes via extended cpuid functions.
3630 		 */
3631 		if (cpi->cpi_family > 5 ||
3632 		    (cpi->cpi_family == 5 && cpi->cpi_model >= 1))
3633 			return (X86_VENDOR_AMD);
3634 		break;
3635 	case X86_VENDOR_TM:
3636 		if (cpi->cpi_family >= 5)
3637 			return (X86_VENDOR_AMD);
3638 		/*FALLTHROUGH*/
3639 	default:
3640 		/*
3641 		 * If they have extended CPU data for 0x80000005
3642 		 * then we assume they have AMD-format cache
3643 		 * information.
3644 		 *
3645 		 * If not, and the vendor happens to be Cyrix,
3646 		 * then try our-Cyrix specific handler.
3647 		 *
3648 		 * If we're not Cyrix, then assume we're using Intel's
3649 		 * table-driven format instead.
3650 		 */
3651 		if (cpi->cpi_xmaxeax >= 0x80000005)
3652 			return (X86_VENDOR_AMD);
3653 		else if (cpi->cpi_vendor == X86_VENDOR_Cyrix)
3654 			return (X86_VENDOR_Cyrix);
3655 		else if (cpi->cpi_maxeax >= 2)
3656 			return (X86_VENDOR_Intel);
3657 		break;
3658 	}
3659 	return (-1);
3660 }
3661 
3662 void
3663 cpuid_set_cpu_properties(void *dip, processorid_t cpu_id,
3664     struct cpuid_info *cpi)
3665 {
3666 	dev_info_t *cpu_devi;
3667 	int create;
3668 
3669 	cpu_devi = (dev_info_t *)dip;
3670 
3671 	/* device_type */
3672 	(void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi,
3673 	    "device_type", "cpu");
3674 
3675 	/* reg */
3676 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3677 	    "reg", cpu_id);
3678 
3679 	/* cpu-mhz, and clock-frequency */
3680 	if (cpu_freq > 0) {
3681 		long long mul;
3682 
3683 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3684 		    "cpu-mhz", cpu_freq);
3685 		if ((mul = cpu_freq * 1000000LL) <= INT_MAX)
3686 			(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3687 			    "clock-frequency", (int)mul);
3688 	}
3689 
3690 	if ((x86_feature & X86_CPUID) == 0) {
3691 		return;
3692 	}
3693 
3694 	/* vendor-id */
3695 	(void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi,
3696 	    "vendor-id", cpi->cpi_vendorstr);
3697 
3698 	if (cpi->cpi_maxeax == 0) {
3699 		return;
3700 	}
3701 
3702 	/*
3703 	 * family, model, and step
3704 	 */
3705 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3706 	    "family", CPI_FAMILY(cpi));
3707 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3708 	    "cpu-model", CPI_MODEL(cpi));
3709 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3710 	    "stepping-id", CPI_STEP(cpi));
3711 
3712 	/* type */
3713 	switch (cpi->cpi_vendor) {
3714 	case X86_VENDOR_Intel:
3715 		create = 1;
3716 		break;
3717 	default:
3718 		create = 0;
3719 		break;
3720 	}
3721 	if (create)
3722 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3723 		    "type", CPI_TYPE(cpi));
3724 
3725 	/* ext-family */
3726 	switch (cpi->cpi_vendor) {
3727 	case X86_VENDOR_Intel:
3728 	case X86_VENDOR_AMD:
3729 		create = cpi->cpi_family >= 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-family", CPI_FAMILY_XTD(cpi));
3738 
3739 	/* ext-model */
3740 	switch (cpi->cpi_vendor) {
3741 	case X86_VENDOR_Intel:
3742 		create = IS_EXTENDED_MODEL_INTEL(cpi);
3743 		break;
3744 	case X86_VENDOR_AMD:
3745 		create = CPI_FAMILY(cpi) == 0xf;
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 		    "ext-model", CPI_MODEL_XTD(cpi));
3754 
3755 	/* generation */
3756 	switch (cpi->cpi_vendor) {
3757 	case X86_VENDOR_AMD:
3758 		/*
3759 		 * AMD K5 model 1 was the first part to support this
3760 		 */
3761 		create = cpi->cpi_xmaxeax >= 0x80000001;
3762 		break;
3763 	default:
3764 		create = 0;
3765 		break;
3766 	}
3767 	if (create)
3768 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3769 		    "generation", BITX((cpi)->cpi_extd[1].cp_eax, 11, 8));
3770 
3771 	/* brand-id */
3772 	switch (cpi->cpi_vendor) {
3773 	case X86_VENDOR_Intel:
3774 		/*
3775 		 * brand id first appeared on Pentium III Xeon model 8,
3776 		 * and Celeron model 8 processors and Opteron
3777 		 */
3778 		create = cpi->cpi_family > 6 ||
3779 		    (cpi->cpi_family == 6 && cpi->cpi_model >= 8);
3780 		break;
3781 	case X86_VENDOR_AMD:
3782 		create = cpi->cpi_family >= 0xf;
3783 		break;
3784 	default:
3785 		create = 0;
3786 		break;
3787 	}
3788 	if (create && cpi->cpi_brandid != 0) {
3789 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3790 		    "brand-id", cpi->cpi_brandid);
3791 	}
3792 
3793 	/* chunks, and apic-id */
3794 	switch (cpi->cpi_vendor) {
3795 		/*
3796 		 * first available on Pentium IV and Opteron (K8)
3797 		 */
3798 	case X86_VENDOR_Intel:
3799 		create = IS_NEW_F6(cpi) || cpi->cpi_family >= 0xf;
3800 		break;
3801 	case X86_VENDOR_AMD:
3802 		create = cpi->cpi_family >= 0xf;
3803 		break;
3804 	default:
3805 		create = 0;
3806 		break;
3807 	}
3808 	if (create) {
3809 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3810 		    "chunks", CPI_CHUNKS(cpi));
3811 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3812 		    "apic-id", cpi->cpi_apicid);
3813 		if (cpi->cpi_chipid >= 0) {
3814 			(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3815 			    "chip#", cpi->cpi_chipid);
3816 			(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3817 			    "clog#", cpi->cpi_clogid);
3818 		}
3819 	}
3820 
3821 	/* cpuid-features */
3822 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3823 	    "cpuid-features", CPI_FEATURES_EDX(cpi));
3824 
3825 
3826 	/* cpuid-features-ecx */
3827 	switch (cpi->cpi_vendor) {
3828 	case X86_VENDOR_Intel:
3829 		create = IS_NEW_F6(cpi) || cpi->cpi_family >= 0xf;
3830 		break;
3831 	default:
3832 		create = 0;
3833 		break;
3834 	}
3835 	if (create)
3836 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3837 		    "cpuid-features-ecx", CPI_FEATURES_ECX(cpi));
3838 
3839 	/* ext-cpuid-features */
3840 	switch (cpi->cpi_vendor) {
3841 	case X86_VENDOR_Intel:
3842 	case X86_VENDOR_AMD:
3843 	case X86_VENDOR_Cyrix:
3844 	case X86_VENDOR_TM:
3845 	case X86_VENDOR_Centaur:
3846 		create = cpi->cpi_xmaxeax >= 0x80000001;
3847 		break;
3848 	default:
3849 		create = 0;
3850 		break;
3851 	}
3852 	if (create) {
3853 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3854 		    "ext-cpuid-features", CPI_FEATURES_XTD_EDX(cpi));
3855 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
3856 		    "ext-cpuid-features-ecx", CPI_FEATURES_XTD_ECX(cpi));
3857 	}
3858 
3859 	/*
3860 	 * Brand String first appeared in Intel Pentium IV, AMD K5
3861 	 * model 1, and Cyrix GXm.  On earlier models we try and
3862 	 * simulate something similar .. so this string should always
3863 	 * same -something- about the processor, however lame.
3864 	 */
3865 	(void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi,
3866 	    "brand-string", cpi->cpi_brandstr);
3867 
3868 	/*
3869 	 * Finally, cache and tlb information
3870 	 */
3871 	switch (x86_which_cacheinfo(cpi)) {
3872 	case X86_VENDOR_Intel:
3873 		intel_walk_cacheinfo(cpi, cpu_devi, add_cacheent_props);
3874 		break;
3875 	case X86_VENDOR_Cyrix:
3876 		cyrix_walk_cacheinfo(cpi, cpu_devi, add_cacheent_props);
3877 		break;
3878 	case X86_VENDOR_AMD:
3879 		amd_cache_info(cpi, cpu_devi);
3880 		break;
3881 	default:
3882 		break;
3883 	}
3884 }
3885 
3886 struct l2info {
3887 	int *l2i_csz;
3888 	int *l2i_lsz;
3889 	int *l2i_assoc;
3890 	int l2i_ret;
3891 };
3892 
3893 /*
3894  * A cacheinfo walker that fetches the size, line-size and associativity
3895  * of the L2 cache
3896  */
3897 static int
3898 intel_l2cinfo(void *arg, const struct cachetab *ct)
3899 {
3900 	struct l2info *l2i = arg;
3901 	int *ip;
3902 
3903 	if (ct->ct_label != l2_cache_str &&
3904 	    ct->ct_label != sl2_cache_str)
3905 		return (0);	/* not an L2 -- keep walking */
3906 
3907 	if ((ip = l2i->l2i_csz) != NULL)
3908 		*ip = ct->ct_size;
3909 	if ((ip = l2i->l2i_lsz) != NULL)
3910 		*ip = ct->ct_line_size;
3911 	if ((ip = l2i->l2i_assoc) != NULL)
3912 		*ip = ct->ct_assoc;
3913 	l2i->l2i_ret = ct->ct_size;
3914 	return (1);		/* was an L2 -- terminate walk */
3915 }
3916 
3917 /*
3918  * AMD L2/L3 Cache and TLB Associativity Field Definition:
3919  *
3920  *	Unlike the associativity for the L1 cache and tlb where the 8 bit
3921  *	value is the associativity, the associativity for the L2 cache and
3922  *	tlb is encoded in the following table. The 4 bit L2 value serves as
3923  *	an index into the amd_afd[] array to determine the associativity.
3924  *	-1 is undefined. 0 is fully associative.
3925  */
3926 
3927 static int amd_afd[] =
3928 	{-1, 1, 2, -1, 4, -1, 8, -1, 16, -1, 32, 48, 64, 96, 128, 0};
3929 
3930 static void
3931 amd_l2cacheinfo(struct cpuid_info *cpi, struct l2info *l2i)
3932 {
3933 	struct cpuid_regs *cp;
3934 	uint_t size, assoc;
3935 	int i;
3936 	int *ip;
3937 
3938 	if (cpi->cpi_xmaxeax < 0x80000006)
3939 		return;
3940 	cp = &cpi->cpi_extd[6];
3941 
3942 	if ((i = BITX(cp->cp_ecx, 15, 12)) != 0 &&
3943 	    (size = BITX(cp->cp_ecx, 31, 16)) != 0) {
3944 		uint_t cachesz = size * 1024;
3945 		assoc = amd_afd[i];
3946 
3947 		ASSERT(assoc != -1);
3948 
3949 		if ((ip = l2i->l2i_csz) != NULL)
3950 			*ip = cachesz;
3951 		if ((ip = l2i->l2i_lsz) != NULL)
3952 			*ip = BITX(cp->cp_ecx, 7, 0);
3953 		if ((ip = l2i->l2i_assoc) != NULL)
3954 			*ip = assoc;
3955 		l2i->l2i_ret = cachesz;
3956 	}
3957 }
3958 
3959 int
3960 getl2cacheinfo(cpu_t *cpu, int *csz, int *lsz, int *assoc)
3961 {
3962 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
3963 	struct l2info __l2info, *l2i = &__l2info;
3964 
3965 	l2i->l2i_csz = csz;
3966 	l2i->l2i_lsz = lsz;
3967 	l2i->l2i_assoc = assoc;
3968 	l2i->l2i_ret = -1;
3969 
3970 	switch (x86_which_cacheinfo(cpi)) {
3971 	case X86_VENDOR_Intel:
3972 		intel_walk_cacheinfo(cpi, l2i, intel_l2cinfo);
3973 		break;
3974 	case X86_VENDOR_Cyrix:
3975 		cyrix_walk_cacheinfo(cpi, l2i, intel_l2cinfo);
3976 		break;
3977 	case X86_VENDOR_AMD:
3978 		amd_l2cacheinfo(cpi, l2i);
3979 		break;
3980 	default:
3981 		break;
3982 	}
3983 	return (l2i->l2i_ret);
3984 }
3985 
3986 #if !defined(__xpv)
3987 
3988 uint32_t *
3989 cpuid_mwait_alloc(cpu_t *cpu)
3990 {
3991 	uint32_t	*ret;
3992 	size_t		mwait_size;
3993 
3994 	ASSERT(cpuid_checkpass(CPU, 2));
3995 
3996 	mwait_size = CPU->cpu_m.mcpu_cpi->cpi_mwait.mon_max;
3997 	if (mwait_size == 0)
3998 		return (NULL);
3999 
4000 	/*
4001 	 * kmem_alloc() returns cache line size aligned data for mwait_size
4002 	 * allocations.  mwait_size is currently cache line sized.  Neither
4003 	 * of these implementation details are guarantied to be true in the
4004 	 * future.
4005 	 *
4006 	 * First try allocating mwait_size as kmem_alloc() currently returns
4007 	 * correctly aligned memory.  If kmem_alloc() does not return
4008 	 * mwait_size aligned memory, then use mwait_size ROUNDUP.
4009 	 *
4010 	 * Set cpi_mwait.buf_actual and cpi_mwait.size_actual in case we
4011 	 * decide to free this memory.
4012 	 */
4013 	ret = kmem_zalloc(mwait_size, KM_SLEEP);
4014 	if (ret == (uint32_t *)P2ROUNDUP((uintptr_t)ret, mwait_size)) {
4015 		cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual = ret;
4016 		cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual = mwait_size;
4017 		*ret = MWAIT_RUNNING;
4018 		return (ret);
4019 	} else {
4020 		kmem_free(ret, mwait_size);
4021 		ret = kmem_zalloc(mwait_size * 2, KM_SLEEP);
4022 		cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual = ret;
4023 		cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual = mwait_size * 2;
4024 		ret = (uint32_t *)P2ROUNDUP((uintptr_t)ret, mwait_size);
4025 		*ret = MWAIT_RUNNING;
4026 		return (ret);
4027 	}
4028 }
4029 
4030 void
4031 cpuid_mwait_free(cpu_t *cpu)
4032 {
4033 	if (cpu->cpu_m.mcpu_cpi == NULL) {
4034 		return;
4035 	}
4036 
4037 	if (cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual != NULL &&
4038 	    cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual > 0) {
4039 		kmem_free(cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual,
4040 		    cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual);
4041 	}
4042 
4043 	cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual = NULL;
4044 	cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual = 0;
4045 }
4046 
4047 void
4048 patch_tsc_read(int flag)
4049 {
4050 	size_t cnt;
4051 
4052 	switch (flag) {
4053 	case X86_NO_TSC:
4054 		cnt = &_no_rdtsc_end - &_no_rdtsc_start;
4055 		(void) memcpy((void *)tsc_read, (void *)&_no_rdtsc_start, cnt);
4056 		break;
4057 	case X86_HAVE_TSCP:
4058 		cnt = &_tscp_end - &_tscp_start;
4059 		(void) memcpy((void *)tsc_read, (void *)&_tscp_start, cnt);
4060 		break;
4061 	case X86_TSC_MFENCE:
4062 		cnt = &_tsc_mfence_end - &_tsc_mfence_start;
4063 		(void) memcpy((void *)tsc_read,
4064 		    (void *)&_tsc_mfence_start, cnt);
4065 		break;
4066 	case X86_TSC_LFENCE:
4067 		cnt = &_tsc_lfence_end - &_tsc_lfence_start;
4068 		(void) memcpy((void *)tsc_read,
4069 		    (void *)&_tsc_lfence_start, cnt);
4070 		break;
4071 	default:
4072 		break;
4073 	}
4074 }
4075 
4076 int
4077 cpuid_deep_cstates_supported(void)
4078 {
4079 	struct cpuid_info *cpi;
4080 	struct cpuid_regs regs;
4081 
4082 	ASSERT(cpuid_checkpass(CPU, 1));
4083 
4084 	cpi = CPU->cpu_m.mcpu_cpi;
4085 
4086 	if (!(x86_feature & X86_CPUID))
4087 		return (0);
4088 
4089 	switch (cpi->cpi_vendor) {
4090 	case X86_VENDOR_Intel:
4091 		if (cpi->cpi_xmaxeax < 0x80000007)
4092 			return (0);
4093 
4094 		/*
4095 		 * TSC run at a constant rate in all ACPI C-states?
4096 		 */
4097 		regs.cp_eax = 0x80000007;
4098 		(void) __cpuid_insn(&regs);
4099 		return (regs.cp_edx & CPUID_TSC_CSTATE_INVARIANCE);
4100 
4101 	default:
4102 		return (0);
4103 	}
4104 }
4105 
4106 #endif	/* !__xpv */
4107 
4108 void
4109 post_startup_cpu_fixups(void)
4110 {
4111 #ifndef __xpv
4112 	/*
4113 	 * Some AMD processors support C1E state. Entering this state will
4114 	 * cause the local APIC timer to stop, which we can't deal with at
4115 	 * this time.
4116 	 */
4117 	if (cpuid_getvendor(CPU) == X86_VENDOR_AMD) {
4118 		on_trap_data_t otd;
4119 		uint64_t reg;
4120 
4121 		if (!on_trap(&otd, OT_DATA_ACCESS)) {
4122 			reg = rdmsr(MSR_AMD_INT_PENDING_CMP_HALT);
4123 			/* Disable C1E state if it is enabled by BIOS */
4124 			if ((reg >> AMD_ACTONCMPHALT_SHIFT) &
4125 			    AMD_ACTONCMPHALT_MASK) {
4126 				reg &= ~(AMD_ACTONCMPHALT_MASK <<
4127 				    AMD_ACTONCMPHALT_SHIFT);
4128 				wrmsr(MSR_AMD_INT_PENDING_CMP_HALT, reg);
4129 			}
4130 		}
4131 		no_trap();
4132 	}
4133 #endif	/* !__xpv */
4134 }
4135 
4136 /*
4137  * Starting with the Westmere processor the local
4138  * APIC timer will continue running in all C-states,
4139  * including the deepest C-states.
4140  */
4141 int
4142 cpuid_arat_supported(void)
4143 {
4144 	struct cpuid_info *cpi;
4145 	struct cpuid_regs regs;
4146 
4147 	ASSERT(cpuid_checkpass(CPU, 1));
4148 	ASSERT(x86_feature & X86_CPUID);
4149 
4150 	cpi = CPU->cpu_m.mcpu_cpi;
4151 
4152 	switch (cpi->cpi_vendor) {
4153 	case X86_VENDOR_Intel:
4154 		/*
4155 		 * Always-running Local APIC Timer is
4156 		 * indicated by CPUID.6.EAX[2].
4157 		 */
4158 		if (cpi->cpi_maxeax >= 6) {
4159 			regs.cp_eax = 6;
4160 			(void) cpuid_insn(NULL, &regs);
4161 			return (regs.cp_eax & CPUID_CSTATE_ARAT);
4162 		} else {
4163 			return (0);
4164 		}
4165 	default:
4166 		return (0);
4167 	}
4168 }
4169 
4170 /*
4171  * Check support for Intel ENERGY_PERF_BIAS feature
4172  */
4173 int
4174 cpuid_iepb_supported(struct cpu *cp)
4175 {
4176 	struct cpuid_info *cpi = cp->cpu_m.mcpu_cpi;
4177 	struct cpuid_regs regs;
4178 
4179 	ASSERT(cpuid_checkpass(cp, 1));
4180 
4181 	if (!(x86_feature & X86_CPUID) || !(x86_feature & X86_MSR)) {
4182 		return (0);
4183 	}
4184 
4185 	/*
4186 	 * Intel ENERGY_PERF_BIAS MSR is indicated by
4187 	 * capability bit CPUID.6.ECX.3
4188 	 */
4189 	if ((cpi->cpi_vendor != X86_VENDOR_Intel) || (cpi->cpi_maxeax < 6))
4190 		return (0);
4191 
4192 	regs.cp_eax = 0x6;
4193 	(void) cpuid_insn(NULL, &regs);
4194 	return (regs.cp_ecx & CPUID_EPB_SUPPORT);
4195 }
4196 
4197 #if defined(__amd64) && !defined(__xpv)
4198 /*
4199  * Patch in versions of bcopy for high performance Intel Nhm processors
4200  * and later...
4201  */
4202 void
4203 patch_memops(uint_t vendor)
4204 {
4205 	size_t cnt, i;
4206 	caddr_t to, from;
4207 
4208 	if ((vendor == X86_VENDOR_Intel) && ((x86_feature & X86_SSE4_2) != 0)) {
4209 		cnt = &bcopy_patch_end - &bcopy_patch_start;
4210 		to = &bcopy_ck_size;
4211 		from = &bcopy_patch_start;
4212 		for (i = 0; i < cnt; i++) {
4213 			*to++ = *from++;
4214 		}
4215 	}
4216 }
4217 #endif  /* __amd64 && !__xpv */
4218 
4219 /*
4220  * This function finds the number of bits to represent the number of cores per
4221  * chip and the number of strands per core for the Intel platforms.
4222  * It re-uses the x2APIC cpuid code of the cpuid_pass2().
4223  */
4224 void
4225 cpuid_get_ext_topo(uint_t vendor, uint_t *core_nbits, uint_t *strand_nbits)
4226 {
4227 	struct cpuid_regs regs;
4228 	struct cpuid_regs *cp = &regs;
4229 
4230 	if (vendor != X86_VENDOR_Intel) {
4231 		return;
4232 	}
4233 
4234 	/* if the cpuid level is 0xB, extended topo is available. */
4235 	cp->cp_eax = 0;
4236 	if (__cpuid_insn(cp) >= 0xB) {
4237 
4238 		cp->cp_eax = 0xB;
4239 		cp->cp_edx = cp->cp_ebx = cp->cp_ecx = 0;
4240 		(void) __cpuid_insn(cp);
4241 
4242 		/*
4243 		 * Check CPUID.EAX=0BH, ECX=0H:EBX is non-zero, which
4244 		 * indicates that the extended topology enumeration leaf is
4245 		 * available.
4246 		 */
4247 		if (cp->cp_ebx) {
4248 			uint_t coreid_shift = 0;
4249 			uint_t chipid_shift = 0;
4250 			uint_t i;
4251 			uint_t level;
4252 
4253 			for (i = 0; i < CPI_FNB_ECX_MAX; i++) {
4254 				cp->cp_eax = 0xB;
4255 				cp->cp_ecx = i;
4256 
4257 				(void) __cpuid_insn(cp);
4258 				level = CPI_CPU_LEVEL_TYPE(cp);
4259 
4260 				if (level == 1) {
4261 					/*
4262 					 * Thread level processor topology
4263 					 * Number of bits shift right APIC ID
4264 					 * to get the coreid.
4265 					 */
4266 					coreid_shift = BITX(cp->cp_eax, 4, 0);
4267 				} else if (level == 2) {
4268 					/*
4269 					 * Core level processor topology
4270 					 * Number of bits shift right APIC ID
4271 					 * to get the chipid.
4272 					 */
4273 					chipid_shift = BITX(cp->cp_eax, 4, 0);
4274 				}
4275 			}
4276 
4277 			if (coreid_shift > 0 && chipid_shift > coreid_shift) {
4278 				*strand_nbits = coreid_shift;
4279 				*core_nbits = chipid_shift - coreid_shift;
4280 			}
4281 		}
4282 	}
4283 }
4284