xref: /illumos-gate/usr/src/uts/intel/os/cpuid.c (revision b1529121add3ff25fc7b58196363a5439c0f6b67)
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  * Copyright (c) 2011, 2016 by Delphix. All rights reserved.
24  * Copyright 2013 Nexenta Systems, Inc. All rights reserved.
25  * Copyright 2014 Josef "Jeff" Sipek <jeffpc@josefsipek.net>
26  * Copyright 2020 Joyent, Inc.
27  * Copyright 2023 Oxide Computer Company
28  * Copyright 2024 MNX Cloud, Inc.
29  */
30 /*
31  * Copyright (c) 2010, Intel Corporation.
32  * All rights reserved.
33  */
34 /*
35  * Portions Copyright 2009 Advanced Micro Devices, Inc.
36  */
37 
38 /*
39  * CPU Identification logic
40  *
41  * The purpose of this file and its companion, cpuid_subr.c, is to help deal
42  * with the identification of CPUs, their features, and their topologies. More
43  * specifically, this file helps drive the following:
44  *
45  * 1. Enumeration of features of the processor which are used by the kernel to
46  *    determine what features to enable or disable. These may be instruction set
47  *    enhancements or features that we use.
48  *
49  * 2. Enumeration of instruction set architecture (ISA) additions that userland
50  *    will be told about through the auxiliary vector.
51  *
52  * 3. Understanding the physical topology of the CPU such as the number of
53  *    caches, how many cores it has, whether or not it supports symmetric
54  *    multi-processing (SMT), etc.
55  *
56  * ------------------------
57  * CPUID History and Basics
58  * ------------------------
59  *
60  * The cpuid instruction was added by Intel roughly around the time that the
61  * original Pentium was introduced. The purpose of cpuid was to tell in a
62  * programmatic fashion information about the CPU that previously was guessed
63  * at. For example, an important part of cpuid is that we can know what
64  * extensions to the ISA exist. If you use an invalid opcode you would get a
65  * #UD, so this method allows a program (whether a user program or the kernel)
66  * to determine what exists without crashing or getting a SIGILL. Of course,
67  * this was also during the era of the clones and the AMD Am5x86. The vendor
68  * name shows up first in cpuid for a reason.
69  *
70  * cpuid information is broken down into ranges called a 'leaf'. Each leaf puts
71  * unique values into the registers %eax, %ebx, %ecx, and %edx and each leaf has
72  * its own meaning. The different leaves are broken down into different regions:
73  *
74  *	[ 0, 7fffffff ]			This region is called the 'basic'
75  *					region. This region is generally defined
76  *					by Intel, though some of the original
77  *					portions have different meanings based
78  *					on the manufacturer. These days, Intel
79  *					adds most new features to this region.
80  *					AMD adds non-Intel compatible
81  *					information in the third, extended
82  *					region. Intel uses this for everything
83  *					including ISA extensions, CPU
84  *					features, cache information, topology,
85  *					and more.
86  *
87  *					There is a hole carved out of this
88  *					region which is reserved for
89  *					hypervisors.
90  *
91  *	[ 40000000, 4fffffff ]		This region, which is found in the
92  *					middle of the previous region, is
93  *					explicitly promised to never be used by
94  *					CPUs. Instead, it is used by hypervisors
95  *					to communicate information about
96  *					themselves to the operating system. The
97  *					values and details are unique for each
98  *					hypervisor.
99  *
100  *	[ 80000000, ffffffff ]		This region is called the 'extended'
101  *					region. Some of the low leaves mirror
102  *					parts of the basic leaves. This region
103  *					has generally been used by AMD for
104  *					various extensions. For example, AMD-
105  *					specific information about caches,
106  *					features, and topology are found in this
107  *					region.
108  *
109  * To specify a range, you place the desired leaf into %eax, zero %ebx, %ecx,
110  * and %edx, and then issue the cpuid instruction. At the first leaf in each of
111  * the ranges, one of the primary things returned is the maximum valid leaf in
112  * that range. This allows for discovery of what range of CPUID is valid.
113  *
114  * The CPUs have potentially surprising behavior when using an invalid leaf or
115  * unimplemented leaf. If the requested leaf is within the valid basic or
116  * extended range, but is unimplemented, then %eax, %ebx, %ecx, and %edx will be
117  * set to zero. However, if you specify a leaf that is outside of a valid range,
118  * then instead it will be filled with the last valid _basic_ leaf. For example,
119  * if the maximum basic value is on leaf 0x3, then issuing a cpuid for leaf 4 or
120  * an invalid extended leaf will return the information for leaf 3.
121  *
122  * Some leaves are broken down into sub-leaves. This means that the value
123  * depends on both the leaf asked for in %eax and a secondary register. For
124  * example, Intel uses the value in %ecx on leaf 7 to indicate a sub-leaf to get
125  * additional information. Or when getting topology information in leaf 0xb, the
126  * initial value in %ecx changes which level of the topology that you are
127  * getting information about.
128  *
129  * cpuid values are always kept to 32 bits regardless of whether or not the
130  * program is in 64-bit mode. When executing in 64-bit mode, the upper
131  * 32 bits of the register are always set to zero so that way the values are the
132  * same regardless of execution mode.
133  *
134  * ----------------------
135  * Identifying Processors
136  * ----------------------
137  *
138  * We can identify a processor in two steps. The first step looks at cpuid leaf
139  * 0. Leaf 0 contains the processor's vendor information. This is done by
140  * putting a 12 character string in %ebx, %ecx, and %edx. On AMD, it is
141  * 'AuthenticAMD' and on Intel it is 'GenuineIntel'.
142  *
143  * From there, a processor is identified by a combination of three different
144  * values:
145  *
146  *  1. Family
147  *  2. Model
148  *  3. Stepping
149  *
150  * Each vendor uses the family and model to uniquely identify a processor. The
151  * way that family and model are changed depends on the vendor. For example,
152  * Intel has been using family 0x6 for almost all of their processor since the
153  * Pentium Pro/Pentium II era, often called the P6. The model is used to
154  * identify the exact processor. Different models are often used for the client
155  * (consumer) and server parts. Even though each processor often has major
156  * architectural differences, they still are considered the same family by
157  * Intel.
158  *
159  * On the other hand, each major AMD architecture generally has its own family.
160  * For example, the K8 is family 0x10, Bulldozer 0x15, and Zen 0x17. Within it
161  * the model number is used to help identify specific processors.  As AMD's
162  * product lines have expanded, they have started putting a mixed bag of
163  * processors into the same family, with each processor under a single
164  * identifying banner (e.g., Milan, Cezanne) using a range of model numbers.  We
165  * refer to each such collection as a processor family, distinct from cpuid
166  * family.  Importantly, each processor family has a BIOS and Kernel Developer's
167  * Guide (BKDG, older parts) or Processor Programming Reference (PPR) that
168  * defines the processor family's non-architectural features.  In general, we'll
169  * use "family" here to mean the family number reported by the cpuid instruction
170  * and distinguish the processor family from it where appropriate.
171  *
172  * The stepping is used to refer to a revision of a specific microprocessor. The
173  * term comes from equipment used to produce masks that are used to create
174  * integrated circuits.
175  *
176  * The information is present in leaf 1, %eax. In technical documentation you
177  * will see the terms extended model and extended family. The original family,
178  * model, and stepping fields were each 4 bits wide. If the values in either
179  * are 0xf, then one is to consult the extended model and extended family, which
180  * take previously reserved bits and allow for a larger number of models and add
181  * 0xf to them.
182  *
183  * When we process this information, we store the full family, model, and
184  * stepping in the struct cpuid_info members cpi_family, cpi_model, and
185  * cpi_step, respectively. Whenever you are performing comparisons with the
186  * family, model, and stepping, you should use these members and not the raw
187  * values from cpuid. If you must use the raw values from cpuid directly, you
188  * must make sure that you add the extended model and family to the base model
189  * and family.
190  *
191  * In general, we do not use information about the family, model, and stepping
192  * to determine whether or not a feature is present; that is generally driven by
193  * specific leaves. However, when something we care about on the processor is
194  * not considered 'architectural' meaning that it is specific to a set of
195  * processors and not promised in the architecture model to be consistent from
196  * generation to generation, then we will fall back on this information. The
197  * most common cases where this comes up is when we have to workaround errata in
198  * the processor, are dealing with processor-specific features such as CPU
199  * performance counters, or we want to provide additional information for things
200  * such as fault management.
201  *
202  * While processors also do have a brand string, which is the name that people
203  * are familiar with when buying the processor, they are not meant for
204  * programmatic consumption. That is what the family, model, and stepping are
205  * for.
206  *
207  * We use the x86_chiprev_t to encode a combination of vendor, processor family,
208  * and stepping(s) that refer to a single or very closely related set of silicon
209  * implementations; while there are sometimes more specific ways to learn of the
210  * presence or absence of a particular erratum or workaround, one may generally
211  * assume that all processors of the same chiprev have the same errata and we
212  * have chosen to represent them this way precisely because that is how AMD
213  * groups them in their revision guides (errata documentation).  The processor
214  * family (x86_processor_family_t) may be extracted from the chiprev if that
215  * level of detail is not needed.  Processor families are considered unordered
216  * but revisions within a family may be compared for either an exact match or at
217  * least as recent as a reference revision.  See the chiprev_xxx() functions
218  * below.
219  *
220  * Similarly, each processor family implements a particular microarchitecture,
221  * which itself may have multiple revisions.  In general, non-architectural
222  * features are specific to a processor family, but some may exist across
223  * families containing cores that implement the same microarchitectural revision
224  * (and, such cores share common bugs, too).  We provide utility routines
225  * analogous to those for extracting and comparing chiprevs for
226  * microarchitectures as well; see the uarch_xxx() functions.
227  *
228  * Both chiprevs and uarchrevs are defined in x86_archext.h and both are at
229  * present used and available only for AMD and AMD-like processors.
230  *
231  * ------------
232  * CPUID Passes
233  * ------------
234  *
235  * As part of performing feature detection, we break this into several different
236  * passes. There used to be a pass 0 that was done from assembly in locore.s to
237  * support processors that have a missing or broken cpuid instruction (notably
238  * certain Cyrix processors) but those were all 32-bit processors which are no
239  * longer supported. Passes are no longer numbered explicitly to make it easier
240  * to break them up or move them around as needed; however, they still have a
241  * well-defined execution ordering enforced by the definition of cpuid_pass_t in
242  * x86_archext.h. The external interface to execute a cpuid pass or determine
243  * whether a pass has been completed consists of cpuid_execpass() and
244  * cpuid_checkpass() respectively.  The passes now, in that execution order,
245  * are as follows:
246  *
247  *	PRELUDE		This pass does not have any dependencies on system
248  *			setup; in particular, unlike all subsequent passes it is
249  *			guaranteed not to require PCI config space access.  It
250  *			sets the flag indicating that the processor we are
251  *			running on supports the cpuid instruction, which all
252  *			64-bit processors do.  This would also be the place to
253  *			add any other basic state that is required later on and
254  *			can be learned without dependencies.
255  *
256  *	IDENT		Determine which vendor manufactured the CPU, the family,
257  *			model, and stepping information, and compute basic
258  *			identifying tags from those values.  This is done first
259  *			so that machine-dependent code can control the features
260  *			the cpuid instruction will report during subsequent
261  *			passes if needed, and so that any intervening
262  *			machine-dependent code that needs basic identity will
263  *			have it available.  This includes synthesised
264  *			identifiers such as chiprev and uarchrev as well as the
265  *			values obtained directly from cpuid.  Prior to executing
266  *			this pass, machine-depedent boot code is responsible for
267  *			ensuring that the PCI configuration space access
268  *			functions have been set up and, if necessary, that
269  *			determine_platform() has been called.
270  *
271  *	BASIC		This is the primary pass and is responsible for doing a
272  *			large number of different things:
273  *
274  *			1. Gathering a large number of feature flags to
275  *			determine which features the CPU support and which
276  *			indicate things that we need to do other work in the OS
277  *			to enable. Features detected this way are added to the
278  *			x86_featureset which can be queried to
279  *			determine what we should do. This includes processing
280  *			all of the basic and extended CPU features that we care
281  *			about.
282  *
283  *			2. Determining the CPU's topology. This includes
284  *			information about how many cores and threads are present
285  *			in the package. It also is responsible for figuring out
286  *			which logical CPUs are potentially part of the same core
287  *			and what other resources they might share. For more
288  *			information see the 'Topology' section.
289  *
290  *			3. Determining the set of CPU security-specific features
291  *			that we need to worry about and determine the
292  *			appropriate set of workarounds.
293  *
294  *			Pass 1 on the boot CPU occurs before KMDB is started.
295  *
296  *	EXTENDED	The second pass is done after startup(). Here, we check
297  *			other miscellaneous features. Most of this is gathering
298  *			additional basic and extended features that we'll use in
299  *			later passes or for debugging support.
300  *
301  *	DYNAMIC		The third pass occurs after the kernel memory allocator
302  *			has been fully initialized. This gathers information
303  *			where we might need dynamic memory available for our
304  *			uses. This includes several varying width leaves that
305  *			have cache information and the processor's brand string.
306  *
307  *	RESOLVE		The fourth and final normal pass is performed after the
308  *			kernel has brought most everything online. This is
309  *			invoked from post_startup(). In this pass, we go through
310  *			the set of features that we have enabled and turn that
311  *			into the hardware auxiliary vector features that
312  *			userland receives. This is used by userland, primarily
313  *			by the run-time link-editor (RTLD), though userland
314  *			software could also refer to it directly.
315  *
316  * The function that performs a pass is currently assumed to be infallible, and
317  * all existing implementation are.  This simplifies callers by allowing
318  * cpuid_execpass() to return void. Similarly, implementers do not need to check
319  * for a NULL CPU argument; the current CPU's cpu_t is substituted if necessary.
320  * Both of these assumptions can be relaxed if needed by future developments.
321  * Tracking of completed states is handled by cpuid_execpass(). It is programmer
322  * error to attempt to execute a pass before all previous passes have been
323  * completed on the specified CPU, or to request cpuid information before the
324  * pass that captures it has been executed.  These conditions can be tested
325  * using cpuid_checkpass().
326  *
327  * The Microcode Pass
328  *
329  * After a microcode update, we do a selective rescan of the cpuid leaves to
330  * determine what features have changed. Microcode updates can provide more
331  * details about security related features to deal with issues like Spectre and
332  * L1TF. On occasion, vendors have violated their contract and removed bits.
333  * However, we don't try to detect that because that puts us in a situation that
334  * we really can't deal with. As such, the only thing we rescan are security
335  * related features today. See cpuid_pass_ucode().  This pass may be run in a
336  * different sequence on APs and therefore is not part of the sequential order;
337  * It is invoked directly instead of by cpuid_execpass() and its completion
338  * status cannot be checked by cpuid_checkpass().  This could be integrated with
339  * a more complex dependency mechanism if warranted by future developments.
340  *
341  * All of the passes are run on all CPUs. However, for the most part we only
342  * care about what the boot CPU says about this information and use the other
343  * CPUs as a rough guide to sanity check that we have the same feature set.
344  *
345  * We do not support running multiple logical CPUs with disjoint, let alone
346  * different, feature sets.
347  *
348  * ------------------
349  * Processor Topology
350  * ------------------
351  *
352  * One of the important things that we need to do is to understand the topology
353  * of the underlying processor. When we say topology in this case, we're trying
354  * to understand the relationship between the logical CPUs that the operating
355  * system sees and the underlying physical layout. Different logical CPUs may
356  * share different resources which can have important consequences for the
357  * performance of the system. For example, they may share caches, execution
358  * units, and more.
359  *
360  * The topology of the processor changes from generation to generation and
361  * vendor to vendor.  Along with that, different vendors use different
362  * terminology, and the operating system itself uses occasionally overlapping
363  * terminology. It's important to understand what this topology looks like so
364  * one can understand the different things that we try to calculate and
365  * determine.
366  *
367  * To get started, let's talk about a little bit of terminology that we've used
368  * so far, is used throughout this file, and is fairly generic across multiple
369  * vendors:
370  *
371  * CPU
372  *	A central processing unit (CPU) refers to a logical and/or virtual
373  *	entity that the operating system can execute instructions on. The
374  *	underlying resources for this CPU may be shared between multiple
375  *	entities; however, to the operating system it is a discrete unit.
376  *
377  * PROCESSOR and PACKAGE
378  *
379  *	Generally, when we use the term 'processor' on its own, we are referring
380  *	to the physical entity that one buys and plugs into a board. However,
381  *	because processor has been overloaded and one might see it used to mean
382  *	multiple different levels, we will instead use the term 'package' for
383  *	the rest of this file. The term package comes from the electrical
384  *	engineering side and refers to the physical entity that encloses the
385  *	electronics inside. Strictly speaking the package can contain more than
386  *	just the CPU, for example, on many processors it may also have what's
387  *	called an 'integrated graphical processing unit (GPU)'. Because the
388  *	package can encapsulate multiple units, it is the largest physical unit
389  *	that we refer to.
390  *
391  * SOCKET
392  *
393  *	A socket refers to unit on a system board (generally the motherboard)
394  *	that can receive a package. A single package, or processor, is plugged
395  *	into a single socket. A system may have multiple sockets. Often times,
396  *	the term socket is used interchangeably with package and refers to the
397  *	electrical component that has plugged in, and not the receptacle itself.
398  *
399  * CORE
400  *
401  *	A core refers to the physical instantiation of a CPU, generally, with a
402  *	full set of hardware resources available to it. A package may contain
403  *	multiple cores inside of it or it may just have a single one. A
404  *	processor with more than one core is often referred to as 'multi-core'.
405  *	In illumos, we will use the feature X86FSET_CMP to refer to a system
406  *	that has 'multi-core' processors.
407  *
408  *	A core may expose a single logical CPU to the operating system, or it
409  *	may expose multiple CPUs, which we call threads, defined below.
410  *
411  *	Some resources may still be shared by cores in the same package. For
412  *	example, many processors will share the level 3 cache between cores.
413  *	Some AMD generations share hardware resources between cores. For more
414  *	information on that see the section 'AMD Topology'.
415  *
416  * THREAD and STRAND
417  *
418  *	In this file, generally a thread refers to a hardware resources and not
419  *	the operating system's logical abstraction. A thread is always exposed
420  *	as an independent logical CPU to the operating system. A thread belongs
421  *	to a specific core. A core may have more than one thread. When that is
422  *	the case, the threads that are part of the same core are often referred
423  *	to as 'siblings'.
424  *
425  *	When multiple threads exist, this is generally referred to as
426  *	simultaneous multi-threading (SMT). When Intel introduced this in their
427  *	processors they called it hyper-threading (HT). When multiple threads
428  *	are active in a core, they split the resources of the core. For example,
429  *	two threads may share the same set of hardware execution units.
430  *
431  *	The operating system often uses the term 'strand' to refer to a thread.
432  *	This helps disambiguate it from the software concept.
433  *
434  * CHIP
435  *
436  *	Unfortunately, the term 'chip' is dramatically overloaded. At its most
437  *	base meaning, it is used to refer to a single integrated circuit, which
438  *	may or may not be the only thing in the package. In illumos, when you
439  *	see the term 'chip' it is almost always referring to the same thing as
440  *	the 'package'. However, many vendors may use chip to refer to one of
441  *	many integrated circuits that have been placed in the package. As an
442  *	example, see the subsequent definition.
443  *
444  *	To try and keep things consistent, we will only use chip when referring
445  *	to the entire integrated circuit package, with the exception of the
446  *	definition of multi-chip module (because it is in the name) and use the
447  *	term 'die' when we want the more general, potential sub-component
448  *	definition.
449  *
450  * DIE
451  *
452  *	A die refers to an integrated circuit. Inside of the package there may
453  *	be a single die or multiple dies. This is sometimes called a 'chip' in
454  *	vendor's parlance, but in this file, we use the term die to refer to a
455  *	subcomponent.
456  *
457  * MULTI-CHIP MODULE
458  *
459  *	A multi-chip module (MCM) refers to putting multiple distinct chips that
460  *	are connected together in the same package. When a multi-chip design is
461  *	used, generally each chip is manufactured independently and then joined
462  *	together in the package. For example, on AMD's Zen microarchitecture
463  *	(family 0x17), the package contains several dies (the second meaning of
464  *	chip from above) that are connected together.
465  *
466  * CACHE
467  *
468  *	A cache is a part of the processor that maintains copies of recently
469  *	accessed memory. Caches are split into levels and then into types.
470  *	Commonly there are one to three levels, called level one, two, and
471  *	three. The lower the level, the smaller it is, the closer it is to the
472  *	execution units of the CPU, and the faster it is to access. The layout
473  *	and design of the cache come in many different flavors, consult other
474  *	resources for a discussion of those.
475  *
476  *	Caches are generally split into two types, the instruction and data
477  *	cache. The caches contain what their names suggest, the instruction
478  *	cache has executable program text, while the data cache has all other
479  *	memory that the processor accesses. As of this writing, data is kept
480  *	coherent between all of the caches on x86, so if one modifies program
481  *	text before it is executed, that will be in the data cache, and the
482  *	instruction cache will be synchronized with that change when the
483  *	processor actually executes those instructions. This coherency also
484  *	covers the fact that data could show up in multiple caches.
485  *
486  *	Generally, the lowest level caches are specific to a core. However, the
487  *	last layer cache is shared between some number of cores. The number of
488  *	CPUs sharing this last level cache is important. This has implications
489  *	for the choices that the scheduler makes, as accessing memory that might
490  *	be in a remote cache after thread migration can be quite expensive.
491  *
492  *	Sometimes, the word cache is abbreviated with a '$', because in US
493  *	English the word cache is pronounced the same as cash. So L1D$ refers to
494  *	the L1 data cache, and L2$ would be the L2 cache. This will not be used
495  *	in the rest of this theory statement for clarity.
496  *
497  * MEMORY CONTROLLER
498  *
499  *	The memory controller is a component that provides access to DRAM. Each
500  *	memory controller can access a set number of DRAM channels. Each channel
501  *	can have a number of DIMMs (sticks of memory) associated with it. A
502  *	given package may have more than one memory controller. The association
503  *	of the memory controller to a group of cores is important as it is
504  *	cheaper to access memory on the controller that you are associated with.
505  *
506  * NUMA
507  *
508  *	NUMA or non-uniform memory access, describes a way that systems are
509  *	built. On x86, any processor core can address all of the memory in the
510  *	system. However, When using multiple sockets or possibly within a
511  *	multi-chip module, some of that memory is physically closer and some of
512  *	it is further. Memory that is further away is more expensive to access.
513  *	Consider the following image of multiple sockets with memory:
514  *
515  *	+--------+                                                +--------+
516  *	| DIMM A |         +----------+      +----------+         | DIMM D |
517  *	+--------+-+       |          |      |          |       +-+------+-+
518  *	  | DIMM B |=======| Socket 0 |======| Socket 1 |=======| DIMM E |
519  *	  +--------+-+     |          |      |          |     +-+------+-+
520  *	    | DIMM C |     +----------+      +----------+     | DIMM F |
521  *	    +--------+                                        +--------+
522  *
523  *	In this example, Socket 0 is closer to DIMMs A-C while Socket 1 is
524  *	closer to DIMMs D-F. This means that it is cheaper for socket 0 to
525  *	access DIMMs A-C and more expensive to access D-F as it has to go
526  *	through Socket 1 to get there. The inverse is true for Socket 1. DIMMs
527  *	D-F are cheaper than A-C. While the socket form is the most common, when
528  *	using multi-chip modules, this can also sometimes occur. For another
529  *	example of this that's more involved, see the AMD topology section.
530  *
531  *
532  * Intel Topology
533  * --------------
534  *
535  * Most Intel processors since Nehalem, (as of this writing the current gen
536  * is Skylake / Cannon Lake) follow a fairly similar pattern. The CPU portion of
537  * the package is a single monolithic die. MCMs currently aren't used. Most
538  * parts have three levels of caches, with the L3 cache being shared between
539  * all of the cores on the package. The L1/L2 cache is generally specific to
540  * an individual core. The following image shows at a simplified level what
541  * this looks like. The memory controller is commonly part of something called
542  * the 'Uncore', that used to be separate physical chips that were not a part of
543  * the package, but are now part of the same chip.
544  *
545  *  +-----------------------------------------------------------------------+
546  *  | Package                                                               |
547  *  |  +-------------------+  +-------------------+  +-------------------+  |
548  *  |  | Core              |  | Core              |  | Core              |  |
549  *  |  |  +--------+ +---+ |  |  +--------+ +---+ |  |  +--------+ +---+ |  |
550  *  |  |  | Thread | | L | |  |  | Thread | | L | |  |  | Thread | | L | |  |
551  *  |  |  +--------+ | 1 | |  |  +--------+ | 1 | |  |  +--------+ | 1 | |  |
552  *  |  |  +--------+ |   | |  |  +--------+ |   | |  |  +--------+ |   | |  |
553  *  |  |  | Thread | |   | |  |  | Thread | |   | |  |  | Thread | |   | |  |
554  *  |  |  +--------+ +---+ |  |  +--------+ +---+ |  |  +--------+ +---+ |  |
555  *  |  |  +--------------+ |  |  +--------------+ |  |  +--------------+ |  |
556  *  |  |  | L2 Cache     | |  |  | L2 Cache     | |  |  | L2 Cache     | |  |
557  *  |  |  +--------------+ |  |  +--------------+ |  |  +--------------+ |  |
558  *  |  +-------------------+  +-------------------+  +-------------------+  |
559  *  | +-------------------------------------------------------------------+ |
560  *  | |                         Shared L3 Cache                           | |
561  *  | +-------------------------------------------------------------------+ |
562  *  | +-------------------------------------------------------------------+ |
563  *  | |                        Memory Controller                          | |
564  *  | +-------------------------------------------------------------------+ |
565  *  +-----------------------------------------------------------------------+
566  *
567  * A side effect of this current architecture is that what we care about from a
568  * scheduling and topology perspective, is simplified. In general we care about
569  * understanding which logical CPUs are part of the same core and socket.
570  *
571  * To determine the relationship between threads and cores, Intel initially used
572  * the identifier in the advanced programmable interrupt controller (APIC). They
573  * also added cpuid leaf 4 to give additional information about the number of
574  * threads and CPUs in the processor. With the addition of x2apic (which
575  * increased the number of addressable logical CPUs from 8-bits to 32-bits), an
576  * additional cpuid topology leaf 0xB was added.
577  *
578  * AMD Topology
579  * ------------
580  *
581  * When discussing AMD topology, we want to break this into three distinct
582  * generations of topology. There's the basic topology that has been used in
583  * family 0xf+ (Opteron, Athlon64), there's the topology that was introduced
584  * with family 0x15 (Bulldozer), and there's the topology that was introduced
585  * with family 0x17 (Zen), evolved more dramatically in Zen 2 (still family
586  * 0x17), and tweaked slightly in Zen 3 (family 19h). AMD also has some
587  * additional terminology that's worth talking about.
588  *
589  * Until the introduction of family 0x17 (Zen), AMD did not implement something
590  * that they considered SMT. Whether or not the AMD processors have SMT
591  * influences many things including scheduling and reliability, availability,
592  * and serviceability (RAS) features.
593  *
594  * NODE
595  *
596  *	AMD uses the term node to refer to a die that contains a number of cores
597  *	and I/O resources. Depending on the processor family and model, more
598  *	than one node can be present in the package. When there is more than one
599  *	node this indicates a multi-chip module. Usually each node has its own
600  *	access to memory and I/O devices. This is important and generally
601  *	different from the corresponding Intel Nehalem-Skylake+ processors. As a
602  *	result, we track this relationship in the operating system.
603  *
604  *	In processors with an L3 cache, the L3 cache is generally shared across
605  *	the entire node, though the way this is carved up varies from generation
606  *	to generation.
607  *
608  * BULLDOZER
609  *
610  *	Starting with the Bulldozer family (0x15) and continuing until the
611  *	introduction of the Zen microarchitecture, AMD introduced the idea of a
612  *	compute unit. In a compute unit, two traditional cores share a number of
613  *	hardware resources. Critically, they share the FPU, L1 instruction
614  *	cache, and the L2 cache. Several compute units were then combined inside
615  *	of a single node.  Because the integer execution units, L1 data cache,
616  *	and some other resources were not shared between the cores, AMD never
617  *	considered this to be SMT.
618  *
619  * ZEN
620  *
621  *	The Zen family (0x17) uses a multi-chip module (MCM) design, the module
622  *	is called Zeppelin. These modules are similar to the idea of nodes used
623  *	previously. Each of these nodes has two DRAM channels which all of the
624  *	cores in the node can access uniformly. These nodes are linked together
625  *	in the package, creating a NUMA environment.
626  *
627  *	The Zeppelin die itself contains two different 'core complexes'. Each
628  *	core complex consists of four cores which each have two threads, for a
629  *	total of 8 logical CPUs per complex. Unlike other generations,
630  *	where all the logical CPUs in a given node share the L3 cache, here each
631  *	core complex has its own shared L3 cache.
632  *
633  *	A further thing that we need to consider is that in some configurations,
634  *	particularly with the Threadripper line of processors, not every die
635  *	actually has its memory controllers wired up to actual memory channels.
636  *	This means that some cores have memory attached to them and others
637  *	don't.
638  *
639  *	To put Zen in perspective, consider the following images:
640  *
641  *      +--------------------------------------------------------+
642  *      | Core Complex                                           |
643  *      | +-------------------+    +-------------------+  +---+  |
644  *      | | Core       +----+ |    | Core       +----+ |  |   |  |
645  *      | | +--------+ | L2 | |    | +--------+ | L2 | |  |   |  |
646  *      | | | Thread | +----+ |    | | Thread | +----+ |  |   |  |
647  *      | | +--------+-+ +--+ |    | +--------+-+ +--+ |  | L |  |
648  *      | |   | Thread | |L1| |    |   | Thread | |L1| |  | 3 |  |
649  *      | |   +--------+ +--+ |    |   +--------+ +--+ |  |   |  |
650  *      | +-------------------+    +-------------------+  | C |  |
651  *      | +-------------------+    +-------------------+  | a |  |
652  *      | | Core       +----+ |    | Core       +----+ |  | c |  |
653  *      | | +--------+ | L2 | |    | +--------+ | L2 | |  | h |  |
654  *      | | | Thread | +----+ |    | | Thread | +----+ |  | e |  |
655  *      | | +--------+-+ +--+ |    | +--------+-+ +--+ |  |   |  |
656  *      | |   | Thread | |L1| |    |   | Thread | |L1| |  |   |  |
657  *      | |   +--------+ +--+ |    |   +--------+ +--+ |  |   |  |
658  *      | +-------------------+    +-------------------+  +---+  |
659  *      |                                                        |
660  *	+--------------------------------------------------------+
661  *
662  *  This first image represents a single Zen core complex that consists of four
663  *  cores.
664  *
665  *
666  *	+--------------------------------------------------------+
667  *	| Zeppelin Die                                           |
668  *	|  +--------------------------------------------------+  |
669  *	|  |         I/O Units (PCIe, SATA, USB, etc.)        |  |
670  *	|  +--------------------------------------------------+  |
671  *      |                           HH                           |
672  *	|          +-----------+    HH    +-----------+          |
673  *	|          |           |    HH    |           |          |
674  *	|          |    Core   |==========|    Core   |          |
675  *	|          |  Complex  |==========|  Complex  |          |
676  *	|          |           |    HH    |           |          |
677  *	|          +-----------+    HH    +-----------+          |
678  *      |                           HH                           |
679  *	|  +--------------------------------------------------+  |
680  *	|  |                Memory Controller                 |  |
681  *	|  +--------------------------------------------------+  |
682  *      |                                                        |
683  *	+--------------------------------------------------------+
684  *
685  *  This image represents a single Zeppelin Die. Note how both cores are
686  *  connected to the same memory controller and I/O units. While each core
687  *  complex has its own L3 cache as seen in the first image, they both have
688  *  uniform access to memory.
689  *
690  *
691  *                      PP                     PP
692  *                      PP                     PP
693  *           +----------PP---------------------PP---------+
694  *           |          PP                     PP         |
695  *           |    +-----------+          +-----------+    |
696  *           |    |           |          |           |    |
697  *       MMMMMMMMM|  Zeppelin |==========|  Zeppelin |MMMMMMMMM
698  *       MMMMMMMMM|    Die    |==========|    Die    |MMMMMMMMM
699  *           |    |           |          |           |    |
700  *           |    +-----------+ooo    ...+-----------+    |
701  *           |          HH      ooo  ...       HH         |
702  *           |          HH        oo..         HH         |
703  *           |          HH        ..oo         HH         |
704  *           |          HH      ...  ooo       HH         |
705  *           |    +-----------+...    ooo+-----------+    |
706  *           |    |           |          |           |    |
707  *       MMMMMMMMM|  Zeppelin |==========|  Zeppelin |MMMMMMMMM
708  *       MMMMMMMMM|    Die    |==========|    Die    |MMMMMMMMM
709  *           |    |           |          |           |    |
710  *           |    +-----------+          +-----------+    |
711  *           |          PP                     PP         |
712  *           +----------PP---------------------PP---------+
713  *                      PP                     PP
714  *                      PP                     PP
715  *
716  *  This image represents a single Zen package. In this example, it has four
717  *  Zeppelin dies, though some configurations only have a single one. In this
718  *  example, each die is directly connected to the next. Also, each die is
719  *  represented as being connected to memory by the 'M' character and connected
720  *  to PCIe devices and other I/O, by the 'P' character. Because each Zeppelin
721  *  die is made up of two core complexes, we have multiple different NUMA
722  *  domains that we care about for these systems.
723  *
724  * ZEN 2
725  *
726  *	Zen 2 changes things in a dramatic way from Zen 1. Whereas in Zen 1
727  *	each Zeppelin Die had its own I/O die, that has been moved out of the
728  *	core complex in Zen 2. The actual core complex looks pretty similar, but
729  *	now the die actually looks much simpler:
730  *
731  *      +--------------------------------------------------------+
732  *      | Zen 2 Core Complex Die    HH                           |
733  *      |                           HH                           |
734  *      |          +-----------+    HH    +-----------+          |
735  *      |          |           |    HH    |           |          |
736  *      |          |    Core   |==========|    Core   |          |
737  *      |          |  Complex  |==========|  Complex  |          |
738  *      |          |           |    HH    |           |          |
739  *      |          +-----------+    HH    +-----------+          |
740  *      |                           HH                           |
741  *      |                           HH                           |
742  *      +--------------------------------------------------------+
743  *
744  *	From here, when we add the central I/O die, this changes things a bit.
745  *	Each die is connected to the I/O die, rather than trying to interconnect
746  *	them directly. The following image takes the same Zen 1 image that we
747  *	had earlier and shows what it looks like with the I/O die instead:
748  *
749  *                                 PP    PP
750  *                                 PP    PP
751  *           +---------------------PP----PP---------------------+
752  *           |                     PP    PP                     |
753  *           |  +-----------+      PP    PP      +-----------+  |
754  *           |  |           |      PP    PP      |           |  |
755  *           |  |   Zen 2   |    +-PP----PP-+    |   Zen 2   |  |
756  *           |  |    Die   _|    | PP    PP |    |_   Die    |  |
757  *           |  |         |o|oooo|          |oooo|o|         |  |
758  *           |  +-----------+    |          |    +-----------+  |
759  *           |                   |   I/O    |                   |
760  *       MMMMMMMMMMMMMMMMMMMMMMMMMM  Die   MMMMMMMMMMMMMMMMMMMMMMMMMM
761  *       MMMMMMMMMMMMMMMMMMMMMMMMMM        MMMMMMMMMMMMMMMMMMMMMMMMMM
762  *           |                   |          |                   |
763  *       MMMMMMMMMMMMMMMMMMMMMMMMMM        MMMMMMMMMMMMMMMMMMMMMMMMMM
764  *       MMMMMMMMMMMMMMMMMMMMMMMMMM        MMMMMMMMMMMMMMMMMMMMMMMMMM
765  *           |                   |          |                   |
766  *           |  +-----------+    |          |    +-----------+  |
767  *           |  |         |o|oooo| PP    PP |oooo|o|         |  |
768  *           |  |   Zen 2  -|    +-PP----PP-+    |-  Zen 2   |  |
769  *           |  |    Die    |      PP    PP      |    Die    |  |
770  *           |  |           |      PP    PP      |           |  |
771  *           |  +-----------+      PP    PP      +-----------+  |
772  *           |                     PP    PP                     |
773  *           +---------------------PP----PP---------------------+
774  *                                 PP    PP
775  *                                 PP    PP
776  *
777  *	The above has four core complex dies installed, though the Zen 2 EPYC
778  *	and ThreadRipper parts allow for up to eight, while the Ryzen parts
779  *	generally only have one to two. The more notable difference here is how
780  *	everything communicates. Note that memory and PCIe come out of the
781  *	central die. This changes the way that one die accesses a resource. It
782  *	basically always has to go to the I/O die, where as in Zen 1 it may have
783  *	satisfied it locally. In general, this ends up being a better strategy
784  *	for most things, though it is possible to still treat everything in four
785  *	distinct NUMA domains with each Zen 2 die slightly closer to some memory
786  *	and PCIe than otherwise. This also impacts the 'amdzen' nexus driver as
787  *	now there is only one 'node' present.
788  *
789  * ZEN 3
790  *
791  *	From an architectural perspective, Zen 3 is a much smaller change from
792  *	Zen 2 than Zen 2 was from Zen 1, though it makes up for most of that in
793  *	its microarchitectural changes. The biggest thing for us is how the die
794  *	changes. In Zen 1 and Zen 2, each core complex still had its own L3
795  *	cache. However, in Zen 3, the L3 is now shared between the entire core
796  *	complex die and is no longer partitioned between each core complex. This
797  *	means that all cores on the die can share the same L3 cache. Otherwise,
798  *	the general layout of the overall package with various core complexes
799  *	and an I/O die stays the same. Here's what the Core Complex Die looks
800  *	like in a bit more detail:
801  *
802  *               +-------------------------------------------------+
803  *               | Zen 3 Core Complex Die                          |
804  *               | +-------------------+    +-------------------+  |
805  *               | | Core       +----+ |    | Core       +----+ |  |
806  *               | | +--------+ | L2 | |    | +--------+ | L2 | |  |
807  *               | | | Thread | +----+ |    | | Thread | +----+ |  |
808  *               | | +--------+-+ +--+ |    | +--------+-+ +--+ |  |
809  *               | |   | Thread | |L1| |    |   | Thread | |L1| |  |
810  *               | |   +--------+ +--+ |    |   +--------+ +--+ |  |
811  *               | +-------------------+    +-------------------+  |
812  *               | +-------------------+    +-------------------+  |
813  *               | | Core       +----+ |    | Core       +----+ |  |
814  *               | | +--------+ | L2 | |    | +--------+ | L2 | |  |
815  *               | | | Thread | +----+ |    | | Thread | +----+ |  |
816  *               | | +--------+-+ +--+ |    | +--------+-+ +--+ |  |
817  *               | |   | Thread | |L1| |    |   | Thread | |L1| |  |
818  *               | |   +--------+ +--+ |    |   +--------+ +--+ |  |
819  *               | +-------------------+    +-------------------+  |
820  *               |                                                 |
821  *               | +--------------------------------------------+  |
822  *               | |                 L3 Cache                   |  |
823  *               | +--------------------------------------------+  |
824  *               |                                                 |
825  *               | +-------------------+    +-------------------+  |
826  *               | | Core       +----+ |    | Core       +----+ |  |
827  *               | | +--------+ | L2 | |    | +--------+ | L2 | |  |
828  *               | | | Thread | +----+ |    | | Thread | +----+ |  |
829  *               | | +--------+-+ +--+ |    | +--------+-+ +--+ |  |
830  *               | |   | Thread | |L1| |    |   | Thread | |L1| |  |
831  *               | |   +--------+ +--+ |    |   +--------+ +--+ |  |
832  *               | +-------------------+    +-------------------+  |
833  *               | +-------------------+    +-------------------+  |
834  *               | | Core       +----+ |    | Core       +----+ |  |
835  *               | | +--------+ | L2 | |    | +--------+ | L2 | |  |
836  *               | | | Thread | +----+ |    | | Thread | +----+ |  |
837  *               | | +--------+-+ +--+ |    | +--------+-+ +--+ |  |
838  *               | |   | Thread | |L1| |    |   | Thread | |L1| |  |
839  *               | |   +--------+ +--+ |    |   +--------+ +--+ |  |
840  *               | +-------------------+    +-------------------+  |
841  *               +-------------------------------------------------+
842  *
843  *	While it is not pictured, there are connections from the die to the
844  *	broader data fabric and additional functional blocks to support that
845  *	communication and coherency.
846  *
847  * CPUID LEAVES
848  *
849  * There are a few different CPUID leaves that we can use to try and understand
850  * the actual state of the world. As part of the introduction of family 0xf, AMD
851  * added CPUID leaf 0x80000008. This leaf tells us the number of logical
852  * processors that are in the system. Because families before Zen didn't have
853  * SMT, this was always the number of cores that were in the system. However, it
854  * should always be thought of as the number of logical threads to be consistent
855  * between generations. In addition we also get the size of the APIC ID that is
856  * used to represent the number of logical processors. This is important for
857  * deriving topology information.
858  *
859  * In the Bulldozer family, AMD added leaf 0x8000001E. The information varies a
860  * bit between Bulldozer and later families, but it is quite useful in
861  * determining the topology information. Because this information has changed
862  * across family generations, it's worth calling out what these mean
863  * explicitly. The registers have the following meanings:
864  *
865  *	%eax	The APIC ID. The entire register is defined to have a 32-bit
866  *		APIC ID, even though on systems without x2apic support, it will
867  *		be limited to 8 bits.
868  *
869  *	%ebx	On Bulldozer-era systems this contains information about the
870  *		number of cores that are in a compute unit (cores that share
871  *		resources). It also contains a per-package compute unit ID that
872  *		identifies which compute unit the logical CPU is a part of.
873  *
874  *		On Zen-era systems this instead contains the number of threads
875  *		per core and the ID of the core that the logical CPU is a part
876  *		of. Note, this ID is unique only to the package, it is not
877  *		globally unique across the entire system.
878  *
879  *	%ecx	This contains the number of nodes that exist in the package. It
880  *		also contains an ID that identifies which node the logical CPU
881  *		is a part of.
882  *
883  * Finally, we also use cpuid leaf 0x8000001D to determine information about the
884  * cache layout to determine which logical CPUs are sharing which caches.
885  *
886  * illumos Topology
887  * ----------------
888  *
889  * Based on the above we synthesize the information into several different
890  * variables that we store in the 'struct cpuid_info'. We'll go into the details
891  * of what each member is supposed to represent and their uniqueness. In
892  * general, there are two levels of uniqueness that we care about. We care about
893  * an ID that is globally unique. That means that it will be unique across all
894  * entities in the system. For example, the default logical CPU ID is globally
895  * unique. On the other hand, there is some information that we only care about
896  * being unique within the context of a single package / socket. Here are the
897  * variables that we keep track of and their meaning.
898  *
899  * Several of the values that are asking for an identifier, with the exception
900  * of cpi_apicid, are allowed to be synthetic.
901  *
902  *
903  * cpi_apicid
904  *
905  *	This is the value of the CPU's APIC id. This should be the full 32-bit
906  *	ID if the CPU is using the x2apic. Otherwise, it should be the 8-bit
907  *	APIC ID. This value is globally unique between all logical CPUs across
908  *	all packages. This is usually required by the APIC.
909  *
910  * cpi_chipid
911  *
912  *	This value indicates the ID of the package that the logical CPU is a
913  *	part of. This value is allowed to be synthetic. It is usually derived by
914  *	taking the CPU's APIC ID and determining how many bits are used to
915  *	represent CPU cores in the package. All logical CPUs that are part of
916  *	the same package must have the same value.
917  *
918  * cpi_coreid
919  *
920  *	This represents the ID of a CPU core. Two logical CPUs should only have
921  *	the same cpi_coreid value if they are part of the same core. These
922  *	values may be synthetic. On systems that support SMT, this value is
923  *	usually derived from the APIC ID, otherwise it is often synthetic and
924  *	just set to the value of the cpu_id in the cpu_t.
925  *
926  * cpi_pkgcoreid
927  *
928  *	This is similar to the cpi_coreid in that logical CPUs that are part of
929  *	the same core should have the same ID. The main difference is that these
930  *	values are only required to be unique to a given socket.
931  *
932  * cpi_clogid
933  *
934  *	This represents the logical ID of a logical CPU. This value should be
935  *	unique within a given socket for each logical CPU. This is allowed to be
936  *	synthetic, though it is usually based off of the CPU's apic ID. The
937  *	broader system expects that logical CPUs that have are part of the same
938  *	core have contiguous numbers. For example, if there were two threads per
939  *	core, then the core IDs divided by two should be the same and the first
940  *	modulus two should be zero and the second one. For example, IDs 4 and 5
941  *	indicate two logical CPUs that are part of the same core. But IDs 5 and
942  *	6 represent two logical CPUs that are part of different cores.
943  *
944  *	While it is common for the cpi_coreid and the cpi_clogid to be derived
945  *	from the same source, strictly speaking, they don't have to be and the
946  *	two values should be considered logically independent. One should not
947  *	try to compare a logical CPU's cpi_coreid and cpi_clogid to determine
948  *	some kind of relationship. While this is tempting, we've seen cases on
949  *	AMD family 0xf where the system's cpu id is not related to its APIC ID.
950  *
951  * cpi_ncpu_per_chip
952  *
953  *	This value indicates the total number of logical CPUs that exist in the
954  *	physical package. Critically, this is not the number of logical CPUs
955  *	that exist for just the single core.
956  *
957  *	This value should be the same for all logical CPUs in the same package.
958  *
959  * cpi_ncore_per_chip
960  *
961  *	This value indicates the total number of physical CPU cores that exist
962  *	in the package. The system compares this value with cpi_ncpu_per_chip to
963  *	determine if simultaneous multi-threading (SMT) is enabled. When
964  *	cpi_ncpu_per_chip equals cpi_ncore_per_chip, then there is no SMT and
965  *	the X86FSET_HTT feature is not set. If this value is greater than one,
966  *	than we consider the processor to have the feature X86FSET_CMP, to
967  *	indicate that there is support for more than one core.
968  *
969  *	This value should be the same for all logical CPUs in the same package.
970  *
971  * cpi_procnodes_per_pkg
972  *
973  *	This value indicates the number of 'nodes' that exist in the package.
974  *	When processors are actually a multi-chip module, this represents the
975  *	number of such modules that exist in the package. Currently, on Intel
976  *	based systems this member is always set to 1.
977  *
978  *	This value should be the same for all logical CPUs in the same package.
979  *
980  * cpi_procnodeid
981  *
982  *	This value indicates the ID of the node that the logical CPU is a part
983  *	of. All logical CPUs that are in the same node must have the same value
984  *	here. This value must be unique across all of the packages in the
985  *	system.  On Intel based systems, this is currently set to the value in
986  *	cpi_chipid because there is only one node.
987  *
988  * cpi_cores_per_compunit
989  *
990  *	This value indicates the number of cores that are part of a compute
991  *	unit. See the AMD topology section for this. This member only has real
992  *	meaning currently for AMD Bulldozer family processors. For all other
993  *	processors, this should currently be set to 1.
994  *
995  * cpi_compunitid
996  *
997  *	This indicates the compute unit that the logical CPU belongs to. For
998  *	processors without AMD Bulldozer-style compute units this should be set
999  *	to the value of cpi_coreid.
1000  *
1001  * cpi_ncpu_shr_last_cache
1002  *
1003  *	This indicates the number of logical CPUs that are sharing the same last
1004  *	level cache. This value should be the same for all CPUs that are sharing
1005  *	that cache. The last cache refers to the cache that is closest to memory
1006  *	and furthest away from the CPU.
1007  *
1008  * cpi_last_lvl_cacheid
1009  *
1010  *	This indicates the ID of the last cache that the logical CPU uses. This
1011  *	cache is often shared between multiple logical CPUs and is the cache
1012  *	that is closest to memory and furthest away from the CPU. This value
1013  *	should be the same for a group of logical CPUs only if they actually
1014  *	share the same last level cache. IDs should not overlap between
1015  *	packages.
1016  *
1017  * cpi_ncore_bits
1018  *
1019  *	This indicates the number of bits that are required to represent all of
1020  *	the cores in the system. As cores are derived based on their APIC IDs,
1021  *	we aren't guaranteed a run of APIC IDs starting from zero. It's OK for
1022  *	this value to be larger than the actual number of IDs that are present
1023  *	in the system. This is used to size tables by the CMI framework. It is
1024  *	only filled in for Intel and AMD CPUs.
1025  *
1026  * cpi_nthread_bits
1027  *
1028  *	This indicates the number of bits required to represent all of the IDs
1029  *	that cover the logical CPUs that exist on a given core. It's OK for this
1030  *	value to be larger than the actual number of IDs that are present in the
1031  *	system.  This is used to size tables by the CMI framework. It is
1032  *	only filled in for Intel and AMD CPUs.
1033  *
1034  * -----------
1035  * Hypervisors
1036  * -----------
1037  *
1038  * If trying to manage the differences between vendors wasn't bad enough, it can
1039  * get worse thanks to our friend hardware virtualization. Hypervisors are given
1040  * the ability to interpose on all cpuid instructions and change them to suit
1041  * their purposes. In general, this is necessary as the hypervisor wants to be
1042  * able to present a more uniform set of features or not necessarily give the
1043  * guest operating system kernel knowledge of all features so it can be
1044  * more easily migrated between systems.
1045  *
1046  * When it comes to trying to determine topology information, this can be a
1047  * double edged sword. When a hypervisor doesn't actually implement a cpuid
1048  * leaf, it'll often return all zeros. Because of that, you'll often see various
1049  * checks scattered about fields being non-zero before we assume we can use
1050  * them.
1051  *
1052  * When it comes to topology information, the hypervisor is often incentivized
1053  * to lie to you about topology. This is because it doesn't always actually
1054  * guarantee that topology at all. The topology path we take in the system
1055  * depends on how the CPU advertises itself. If it advertises itself as an Intel
1056  * or AMD CPU, then we basically do our normal path. However, when they don't
1057  * use an actual vendor, then that usually turns into multiple one-core CPUs
1058  * that we enumerate that are often on different sockets. The actual behavior
1059  * depends greatly on what the hypervisor actually exposes to us.
1060  *
1061  * --------------------
1062  * Exposing Information
1063  * --------------------
1064  *
1065  * We expose CPUID information in three different forms in the system.
1066  *
1067  * The first is through the x86_featureset variable. This is used in conjunction
1068  * with the is_x86_feature() function. This is queried by x86-specific functions
1069  * to determine which features are or aren't present in the system and to make
1070  * decisions based upon them. For example, users of this include everything from
1071  * parts of the system dedicated to reliability, availability, and
1072  * serviceability (RAS), to making decisions about how to handle security
1073  * mitigations, to various x86-specific drivers. General purpose or
1074  * architecture independent drivers should never be calling this function.
1075  *
1076  * The second means is through the auxiliary vector. The auxiliary vector is a
1077  * series of tagged data that the kernel passes down to a user program when it
1078  * begins executing. This information is used to indicate to programs what
1079  * instruction set extensions are present. For example, information about the
1080  * CPU supporting the machine check architecture (MCA) wouldn't be passed down
1081  * since user programs cannot make use of it. However, things like the AVX
1082  * instruction sets are. Programs use this information to make run-time
1083  * decisions about what features they should use. As an example, the run-time
1084  * link-editor (rtld) can relocate different functions depending on the hardware
1085  * support available.
1086  *
1087  * The final form is through a series of accessor functions that all have the
1088  * form cpuid_get*. This is used by a number of different subsystems in the
1089  * kernel to determine more detailed information about what we're running on,
1090  * topology information, etc. Some of these subsystems include processor groups
1091  * (uts/common/os/pg.c.), CPU Module Interface (uts/i86pc/os/cmi.c), ACPI,
1092  * microcode, and performance monitoring. These functions all ASSERT that the
1093  * CPU they're being called on has reached a certain cpuid pass. If the passes
1094  * are rearranged, then this needs to be adjusted.
1095  *
1096  * -----------------------------------------------
1097  * Speculative Execution CPU Side Channel Security
1098  * -----------------------------------------------
1099  *
1100  * With the advent of the Spectre and Meltdown attacks which exploit speculative
1101  * execution in the CPU to create side channels there have been a number of
1102  * different attacks and corresponding issues that the operating system needs to
1103  * mitigate against. The following list is some of the common, but not
1104  * exhaustive, set of issues that we know about and have done some or need to do
1105  * more work in the system to mitigate against:
1106  *
1107  *   - Spectre v1
1108  *   - swapgs (Spectre v1 variant)
1109  *   - Spectre v2
1110  *     - Branch History Injection (BHI).
1111  *   - Meltdown (Spectre v3)
1112  *   - Rogue Register Read (Spectre v3a)
1113  *   - Speculative Store Bypass (Spectre v4)
1114  *   - ret2spec, SpectreRSB
1115  *   - L1 Terminal Fault (L1TF)
1116  *   - Microarchitectural Data Sampling (MDS)
1117  *   - Register File Data Sampling (RFDS)
1118  *
1119  * Each of these requires different sets of mitigations and has different attack
1120  * surfaces. For the most part, this discussion is about protecting the kernel
1121  * from non-kernel executing environments such as user processes and hardware
1122  * virtual machines. Unfortunately, there are a number of user vs. user
1123  * scenarios that exist with these. The rest of this section will describe the
1124  * overall approach that the system has taken to address these as well as their
1125  * shortcomings. Unfortunately, not all of the above have been handled today.
1126  *
1127  * SPECTRE v2, ret2spec, SpectreRSB
1128  *
1129  * The second variant of the spectre attack focuses on performing branch target
1130  * injection. This generally impacts indirect call instructions in the system.
1131  * There are four different ways to mitigate this issue that are commonly
1132  * described today:
1133  *
1134  *  1. Using Indirect Branch Restricted Speculation (IBRS).
1135  *  2. Using Retpolines and RSB Stuffing
1136  *  3. Using Enhanced Indirect Branch Restricted Speculation (eIBRS)
1137  *  4. Using Automated Indirect Branch Restricted Speculation (AIBRS)
1138  *
1139  * IBRS uses a feature added to microcode to restrict speculation, among other
1140  * things. This form of mitigation has not been used as it has been generally
1141  * seen as too expensive and requires reactivation upon various transitions in
1142  * the system.
1143  *
1144  * As a less impactful alternative to IBRS, retpolines were developed by
1145  * Google. These basically require one to replace indirect calls with a specific
1146  * trampoline that will cause speculation to fail and break the attack.
1147  * Retpolines require compiler support. We always build with retpolines in the
1148  * external thunk mode. This means that a traditional indirect call is replaced
1149  * with a call to one of the __x86_indirect_thunk_<reg> functions. A side effect
1150  * of this is that all indirect function calls are performed through a register.
1151  *
1152  * We have to use a common external location of the thunk and not inline it into
1153  * the callsite so that way we can have a single place to patch these functions.
1154  * As it turns out, we currently have two different forms of retpolines that
1155  * exist in the system:
1156  *
1157  *  1. A full retpoline
1158  *  2. A no-op version
1159  *
1160  * The first one is used in the general case. Historically, there was an
1161  * AMD-specific optimized retopoline variant that was based around using a
1162  * serializing lfence instruction; however, in March 2022 it was announced that
1163  * this was actually still vulnerable to Spectre v2 and therefore we no longer
1164  * use it and it is no longer available in the system.
1165  *
1166  * The third form described above is the most curious. It turns out that the way
1167  * that retpolines are implemented is that they rely on how speculation is
1168  * performed on a 'ret' instruction. Intel has continued to optimize this
1169  * process (which is partly why we need to have return stack buffer stuffing,
1170  * but more on that in a bit) and in processors starting with Cascade Lake
1171  * on the server side, it's dangerous to rely on retpolines. Instead, a new
1172  * mechanism has been introduced called Enhanced IBRS (eIBRS).
1173  *
1174  * Unlike IBRS, eIBRS is designed to be enabled once at boot and left on each
1175  * physical core. However, if this is the case, we don't want to use retpolines
1176  * any more. Therefore if eIBRS is present, we end up turning each retpoline
1177  * function (called a thunk) into a jmp instruction. This means that we're still
1178  * paying the cost of an extra jump to the external thunk, but it gives us
1179  * flexibility and the ability to have a single kernel image that works across a
1180  * wide variety of systems and hardware features.
1181  *
1182  * Unfortunately, this alone is insufficient. First, Skylake systems have
1183  * additional speculation for the Return Stack Buffer (RSB) which is used to
1184  * return from call instructions which retpolines take advantage of. However,
1185  * this problem is not just limited to Skylake and is actually more pernicious.
1186  * The SpectreRSB paper introduces several more problems that can arise with
1187  * dealing with this. The RSB can be poisoned just like the indirect branch
1188  * predictor. This means that one needs to clear the RSB when transitioning
1189  * between two different privilege domains. Some examples include:
1190  *
1191  *  - Switching between two different user processes
1192  *  - Going between user land and the kernel
1193  *  - Returning to the kernel from a hardware virtual machine
1194  *
1195  * Mitigating this involves combining a couple of different things. The first is
1196  * SMEP (supervisor mode execution protection) which was introduced in Ivy
1197  * Bridge. When an RSB entry refers to a user address and we're executing in the
1198  * kernel, speculation through it will be stopped when SMEP is enabled. This
1199  * protects against a number of the different cases that we would normally be
1200  * worried about such as when we enter the kernel from user land.
1201  *
1202  * To prevent against additional manipulation of the RSB from other contexts
1203  * such as a non-root VMX context attacking the kernel we first look to
1204  * enhanced IBRS. When eIBRS is present and enabled, then there should be
1205  * nothing else that we need to do to protect the kernel at this time.
1206  *
1207  * Unfortunately, not all eIBRS implementations are sufficient to guard
1208  * against RSB manipulations, so we still need to manually overwrite the
1209  * contents of the return stack buffer unless the hardware specifies we are
1210  * covered. We do this through the x86_rsb_stuff() function.  Currently this
1211  * is employed on context switch and vmx_exit. The x86_rsb_stuff() function is
1212  * disabled only when mitigations in general are, or if we have hardware
1213  * indicating no need for post-barrier RSB protections, either in one place
1214  * (old hardware), or on both (newer hardware).
1215  *
1216  * If SMEP is not present, then we would have to stuff the RSB every time we
1217  * transitioned from user mode to the kernel, which isn't very practical right
1218  * now.
1219  *
1220  * To fully protect user to user and vmx to vmx attacks from these classes of
1221  * issues, we would also need to allow them to opt into performing an Indirect
1222  * Branch Prediction Barrier (IBPB) on switch. This is not currently wired up.
1223  *
1224  * The fourth form of mitigation here is specific to AMD and is called Automated
1225  * IBRS (AIBRS). This is similar in spirit to eIBRS; however rather than set the
1226  * IBRS bit in MSR_IA32_SPEC_CTRL (0x48) we instead set a bit in the EFER
1227  * (extended feature enable register) MSR. This bit basically says that IBRS
1228  * acts as though it is always active when executing at CPL0 and when executing
1229  * in the 'host' context when SEV-SNP is enabled.
1230  *
1231  * When this is active, AMD states that the RSB is cleared on VMEXIT and
1232  * therefore it is unnecessary. While this handles RSB stuffing attacks from SVM
1233  * to the kernel, we must still consider the remaining cases that exist, just
1234  * like above. While traditionally AMD employed a 32 entry RSB allowing the
1235  * traditional technique to work, this is not true on all CPUs. While a write to
1236  * IBRS would clear the RSB if the processor supports more than 32 entries (but
1237  * not otherwise), AMD states that as long as at leat a single 4 KiB unmapped
1238  * guard page is present between user and kernel address spaces and SMEP is
1239  * enabled, then there is no need to clear the RSB at all.
1240  *
1241  * By default, the system will enable RSB stuffing and the required variant of
1242  * retpolines and store that information in the x86_spectrev2_mitigation value.
1243  * This will be evaluated after a microcode update as well, though it is
1244  * expected that microcode updates will not take away features. This may mean
1245  * that a late loaded microcode may not end up in the optimal configuration
1246  * (though this should be rare).
1247  *
1248  * Currently we do not build kmdb with retpolines or perform any additional side
1249  * channel security mitigations for it. One complication with kmdb is that it
1250  * requires its own retpoline thunks and it would need to adjust itself based on
1251  * what the kernel does. The threat model of kmdb is more limited and therefore
1252  * it may make more sense to investigate using prediction barriers as the whole
1253  * system is only executing a single instruction at a time while in kmdb.
1254  *
1255  * Branch History Injection (BHI)
1256  *
1257  * BHI is a specific form of SPECTREv2 where an attacker may manipulate branch
1258  * history before transitioning from user to supervisor mode (or from VMX
1259  * non-root/guest to root mode). The attacker can then exploit certain
1260  * compiler-generated code-sequences ("gadgets") to disclose information from
1261  * other contexts or domains.  Recent (late-2023/early-2024) research in
1262  * object code analysis discovered many more potential gadgets than what was
1263  * initially reported (which previously was confined to Linux use of
1264  * unprivileged eBPF).
1265  *
1266  * The BHI threat doesn't exist in processsors that predate eIBRS, or in AMD
1267  * ones. Some eIBRS processors have the ability to disable branch history in
1268  * certain (but not all) cases using an MSR write. eIBRS processors that don't
1269  * have the ability to disable must use a software sequence to scrub the
1270  * branch history buffer.
1271  *
1272  * BHI_DIS_S (the aforementioned MSR) prevents ring 0 from ring 3 (VMX guest
1273  * or VMX root). It does not protect different user processes from each other,
1274  * or ring 3 VMX guest from ring 3 VMX root or vice versa.
1275  *
1276  * The BHI clearing sequence prevents user exploiting kernel gadgets, and user
1277  * A's use of user B's gadgets.
1278  *
1279  * SMEP and eIBRS are a continuing defense-in-depth measure protecting the
1280  * kernel.
1281  *
1282  * SPECTRE v1, v4
1283  *
1284  * The v1 and v4 variants of spectre are not currently mitigated in the
1285  * system and require other classes of changes to occur in the code.
1286  *
1287  * SPECTRE v1 (SWAPGS VARIANT)
1288  *
1289  * The class of Spectre v1 vulnerabilities aren't all about bounds checks, but
1290  * can generally affect any branch-dependent code. The swapgs issue is one
1291  * variant of this. If we are coming in from userspace, we can have code like
1292  * this:
1293  *
1294  *	cmpw	$KCS_SEL, REGOFF_CS(%rsp)
1295  *	je	1f
1296  *	movq	$0, REGOFF_SAVFP(%rsp)
1297  *	swapgs
1298  *	1:
1299  *	movq	%gs:CPU_THREAD, %rax
1300  *
1301  * If an attacker can cause a mis-speculation of the branch here, we could skip
1302  * the needed swapgs, and use the /user/ %gsbase as the base of the %gs-based
1303  * load. If subsequent code can act as the usual Spectre cache gadget, this
1304  * would potentially allow KPTI bypass. To fix this, we need an lfence prior to
1305  * any use of the %gs override.
1306  *
1307  * The other case is also an issue: if we're coming into a trap from kernel
1308  * space, we could mis-speculate and swapgs the user %gsbase back in prior to
1309  * using it. AMD systems are not vulnerable to this version, as a swapgs is
1310  * serializing with respect to subsequent uses. But as AMD /does/ need the other
1311  * case, and the fix is the same in both cases (an lfence at the branch target
1312  * 1: in this example), we'll just do it unconditionally.
1313  *
1314  * Note that we don't enable user-space "wrgsbase" via CR4_FSGSBASE, making it
1315  * harder for user-space to actually set a useful %gsbase value: although it's
1316  * not clear, it might still be feasible via lwp_setprivate(), though, so we
1317  * mitigate anyway.
1318  *
1319  * MELTDOWN
1320  *
1321  * Meltdown, or spectre v3, allowed a user process to read any data in their
1322  * address space regardless of whether or not the page tables in question
1323  * allowed the user to have the ability to read them. The solution to meltdown
1324  * is kernel page table isolation. In this world, there are two page tables that
1325  * are used for a process, one in user land and one in the kernel. To implement
1326  * this we use per-CPU page tables and switch between the user and kernel
1327  * variants when entering and exiting the kernel.  For more information about
1328  * this process and how the trampolines work, please see the big theory
1329  * statements and additional comments in:
1330  *
1331  *  - uts/i86pc/ml/kpti_trampolines.s
1332  *  - uts/i86pc/vm/hat_i86.c
1333  *
1334  * While Meltdown only impacted Intel systems and there are also Intel systems
1335  * that have Meltdown fixed (called Rogue Data Cache Load), we always have
1336  * kernel page table isolation enabled. While this may at first seem weird, an
1337  * important thing to remember is that you can't speculatively read an address
1338  * if it's never in your page table at all. Having user processes without kernel
1339  * pages present provides us with an important layer of defense in the kernel
1340  * against any other side channel attacks that exist and have yet to be
1341  * discovered. As such, kernel page table isolation (KPTI) is always enabled by
1342  * default, no matter the x86 system.
1343  *
1344  * L1 TERMINAL FAULT
1345  *
1346  * L1 Terminal Fault (L1TF) takes advantage of an issue in how speculative
1347  * execution uses page table entries. Effectively, it is two different problems.
1348  * The first is that it ignores the not present bit in the page table entries
1349  * when performing speculative execution. This means that something can
1350  * speculatively read the listed physical address if it's present in the L1
1351  * cache under certain conditions (see Intel's documentation for the full set of
1352  * conditions). Secondly, this can be used to bypass hardware virtualization
1353  * extended page tables (EPT) that are part of Intel's hardware virtual machine
1354  * instructions.
1355  *
1356  * For the non-hardware virtualized case, this is relatively easy to deal with.
1357  * We must make sure that all unmapped pages have an address of zero. This means
1358  * that they could read the first 4k of physical memory; however, we never use
1359  * that first page in the operating system and always skip putting it in our
1360  * memory map, even if firmware tells us we can use it in our memory map. While
1361  * other systems try to put extra metadata in the address and reserved bits,
1362  * which led to this being problematic in those cases, we do not.
1363  *
1364  * For hardware virtual machines things are more complicated. Because they can
1365  * construct their own page tables, it isn't hard for them to perform this
1366  * attack against any physical address. The one wrinkle is that this physical
1367  * address must be in the L1 data cache. Thus Intel added an MSR that we can use
1368  * to flush the L1 data cache. We wrap this up in the function
1369  * spec_uarch_flush(). This function is also used in the mitigation of
1370  * microarchitectural data sampling (MDS) discussed later on. Kernel based
1371  * hypervisors such as KVM or bhyve are responsible for performing this before
1372  * entering the guest.
1373  *
1374  * Because this attack takes place in the L1 cache, there's another wrinkle
1375  * here. The L1 cache is shared between all logical CPUs in a core in most Intel
1376  * designs. This means that when a thread enters a hardware virtualized context
1377  * and flushes the L1 data cache, the other thread on the processor may then go
1378  * ahead and put new data in it that can be potentially attacked. While one
1379  * solution is to disable SMT on the system, another option that is available is
1380  * to use a feature for hardware virtualization called 'SMT exclusion'. This
1381  * goes through and makes sure that if a HVM is being scheduled on one thread,
1382  * then the thing on the other thread is from the same hardware virtual machine.
1383  * If an interrupt comes in or the guest exits to the broader system, then the
1384  * other SMT thread will be kicked out.
1385  *
1386  * L1TF can be fully mitigated by hardware. If the RDCL_NO feature is set in the
1387  * architecture capabilities MSR (MSR_IA32_ARCH_CAPABILITIES), then we will not
1388  * perform L1TF related mitigations.
1389  *
1390  * MICROARCHITECTURAL DATA SAMPLING
1391  *
1392  * Microarchitectural data sampling (MDS) is a combination of four discrete
1393  * vulnerabilities that are similar issues affecting various parts of the CPU's
1394  * microarchitectural implementation around load, store, and fill buffers.
1395  * Specifically it is made up of the following subcomponents:
1396  *
1397  *  1. Microarchitectural Store Buffer Data Sampling (MSBDS)
1398  *  2. Microarchitectural Fill Buffer Data Sampling (MFBDS)
1399  *  3. Microarchitectural Load Port Data Sampling (MLPDS)
1400  *  4. Microarchitectural Data Sampling Uncacheable Memory (MDSUM)
1401  *
1402  * To begin addressing these, Intel has introduced another feature in microcode
1403  * called MD_CLEAR. This changes the verw instruction to operate in a different
1404  * way. This allows us to execute the verw instruction in a particular way to
1405  * flush the state of the affected parts. The L1TF L1D flush mechanism is also
1406  * updated when this microcode is present to flush this state.
1407  *
1408  * Primarily we need to flush this state whenever we transition from the kernel
1409  * to a less privileged context such as user mode or an HVM guest. MSBDS is a
1410  * little bit different. Here the structures are statically sized when a logical
1411  * CPU is in use and resized when it goes to sleep. Therefore, we also need to
1412  * flush the microarchitectural state before the CPU goes idles by calling hlt,
1413  * mwait, or another ACPI method. To perform these flushes, we call
1414  * x86_md_clear() at all of these transition points.
1415  *
1416  * If hardware enumerates RDCL_NO, indicating that it is not vulnerable to L1TF,
1417  * then we change the spec_uarch_flush() function to point to x86_md_clear(). If
1418  * MDS_NO has been set, then this is fully mitigated and x86_md_clear() becomes
1419  * a no-op.
1420  *
1421  * Unfortunately, with this issue hyperthreading rears its ugly head. In
1422  * particular, everything we've discussed above is only valid for a single
1423  * thread executing on a core. In the case where you have hyper-threading
1424  * present, this attack can be performed between threads. The theoretical fix
1425  * for this is to ensure that both threads are always in the same security
1426  * domain. This means that they are executing in the same ring and mutually
1427  * trust each other. Practically speaking, this would mean that a system call
1428  * would have to issue an inter-processor interrupt (IPI) to the other thread.
1429  * Rather than implement this, we recommend that one disables hyper-threading
1430  * through the use of psradm -aS.
1431  *
1432  * TSX ASYNCHRONOUS ABORT
1433  *
1434  * TSX Asynchronous Abort (TAA) is another side-channel vulnerability that
1435  * behaves like MDS, but leverages Intel's transactional instructions as another
1436  * vector. Effectively, when a transaction hits one of these cases (unmapped
1437  * page, various cache snoop activity, etc.) then the same data can be exposed
1438  * as in the case of MDS. This means that you can attack your twin.
1439  *
1440  * Intel has described that there are two different ways that we can mitigate
1441  * this problem on affected processors:
1442  *
1443  *   1) We can use the same techniques used to deal with MDS. Flushing the
1444  *      microarchitectural buffers and disabling hyperthreading will mitigate
1445  *      this in the same way.
1446  *
1447  *   2) Using microcode to disable TSX.
1448  *
1449  * Now, most processors that are subject to MDS (as in they don't have MDS_NO in
1450  * the IA32_ARCH_CAPABILITIES MSR) will not receive microcode to disable TSX.
1451  * That's OK as we're already doing all such mitigations. On the other hand,
1452  * processors with MDS_NO are all supposed to receive microcode updates that
1453  * enumerate support for disabling TSX. In general, we'd rather use this method
1454  * when available as it doesn't require disabling hyperthreading to be
1455  * effective. Currently we basically are relying on microcode for processors
1456  * that enumerate MDS_NO.
1457  *
1458  * Another MDS-variant in a few select Intel Atom CPUs is Register File Data
1459  * Sampling: RFDS. This allows an attacker to sample values that were in any
1460  * of integer, floating point, or vector registers. This was discovered by
1461  * Intel during internal validation work.  The existence of the RFDS_NO
1462  * capability, or the LACK of a RFDS_CLEAR capability, means we do not have to
1463  * act. Intel has said some CPU models immune to RFDS MAY NOT enumerate
1464  * RFDS_NO. If RFDS_NO is not set, but RFDS_CLEAR is, we must set x86_md_clear,
1465  * and make sure it's using VERW. Unlike MDS, RFDS can't be helped by the
1466  * MSR that L1D uses.
1467  *
1468  * The microcode features are enumerated as part of the IA32_ARCH_CAPABILITIES.
1469  * When bit 7 (IA32_ARCH_CAP_TSX_CTRL) is present, then we are given two
1470  * different powers. The first allows us to cause all transactions to
1471  * immediately abort. The second gives us a means of disabling TSX completely,
1472  * which includes removing it from cpuid. If we have support for this in
1473  * microcode during the first cpuid pass, then we'll disable TSX completely such
1474  * that user land never has a chance to observe the bit. However, if we are late
1475  * loading the microcode, then we must use the functionality to cause
1476  * transactions to automatically abort. This is necessary for user land's sake.
1477  * Once a program sees a cpuid bit, it must not be taken away.
1478  *
1479  * We track whether or not we should do this based on what cpuid pass we're in.
1480  * Whenever we hit cpuid_scan_security() on the boot CPU and we're still on pass
1481  * 1 of the cpuid logic, then we can completely turn off TSX. Notably this
1482  * should happen twice. Once in the normal cpuid_pass_basic() code and then a
1483  * second time after we do the initial microcode update.  As a result we need to
1484  * be careful in cpuid_apply_tsx() to only use the MSR if we've loaded a
1485  * suitable microcode on the current CPU (which happens prior to
1486  * cpuid_pass_ucode()).
1487  *
1488  * If TAA has been fixed, then it will be enumerated in IA32_ARCH_CAPABILITIES
1489  * as TAA_NO. In such a case, we will still disable TSX: it's proven to be an
1490  * unfortunate feature in a number of ways, and taking the opportunity to
1491  * finally be able to turn it off is likely to be of benefit in the future.
1492  *
1493  * SUMMARY
1494  *
1495  * The following table attempts to summarize the mitigations for various issues
1496  * and what's done in various places:
1497  *
1498  *  - Spectre v1: Not currently mitigated
1499  *  - swapgs: lfences after swapgs paths
1500  *  - Spectre v2: Retpolines/RSB Stuffing or eIBRS/AIBRS if HW support
1501  *  - Meltdown: Kernel Page Table Isolation
1502  *  - Spectre v3a: Updated CPU microcode
1503  *  - Spectre v4: Not currently mitigated
1504  *  - SpectreRSB: SMEP and RSB Stuffing
1505  *  - L1TF: spec_uarch_flush, SMT exclusion, requires microcode
1506  *  - MDS: x86_md_clear, requires microcode, disabling SMT
1507  *  - TAA: x86_md_clear and disabling SMT OR microcode and disabling TSX
1508  *  - RFDS: microcode with x86_md_clear if RFDS_CLEAR set and RFDS_NO not.
1509  *  - BHI: software sequence, and use of BHI_DIS_S if microcode has it.
1510  *
1511  * The following table indicates the x86 feature set bits that indicate that a
1512  * given problem has been solved or a notable feature is present:
1513  *
1514  *  - RDCL_NO: Meltdown, L1TF, MSBDS subset of MDS
1515  *  - MDS_NO: All forms of MDS
1516  *  - TAA_NO: TAA
1517  *  - RFDS_NO: RFDS
1518  *  - BHI_NO: BHI
1519  */
1520 
1521 #include <sys/types.h>
1522 #include <sys/archsystm.h>
1523 #include <sys/x86_archext.h>
1524 #include <sys/kmem.h>
1525 #include <sys/systm.h>
1526 #include <sys/cmn_err.h>
1527 #include <sys/sunddi.h>
1528 #include <sys/sunndi.h>
1529 #include <sys/cpuvar.h>
1530 #include <sys/processor.h>
1531 #include <sys/sysmacros.h>
1532 #include <sys/pg.h>
1533 #include <sys/fp.h>
1534 #include <sys/controlregs.h>
1535 #include <sys/bitmap.h>
1536 #include <sys/auxv_386.h>
1537 #include <sys/memnode.h>
1538 #include <sys/pci_cfgspace.h>
1539 #include <sys/comm_page.h>
1540 #include <sys/mach_mmu.h>
1541 #include <sys/ucode.h>
1542 #include <sys/tsc.h>
1543 #include <sys/kobj.h>
1544 #include <sys/asm_misc.h>
1545 #include <sys/bitmap.h>
1546 
1547 #ifdef __xpv
1548 #include <sys/hypervisor.h>
1549 #else
1550 #include <sys/ontrap.h>
1551 #endif
1552 
1553 uint_t x86_vendor = X86_VENDOR_IntelClone;
1554 uint_t x86_type = X86_TYPE_OTHER;
1555 uint_t x86_clflush_size = 0;
1556 
1557 #if defined(__xpv)
1558 int x86_use_pcid = 0;
1559 int x86_use_invpcid = 0;
1560 #else
1561 int x86_use_pcid = -1;
1562 int x86_use_invpcid = -1;
1563 #endif
1564 
1565 typedef enum {
1566 	X86_SPECTREV2_RETPOLINE,
1567 	X86_SPECTREV2_ENHANCED_IBRS,
1568 	X86_SPECTREV2_AUTO_IBRS,
1569 	X86_SPECTREV2_DISABLED
1570 } x86_spectrev2_mitigation_t;
1571 
1572 uint_t x86_disable_spectrev2 = 0;
1573 static x86_spectrev2_mitigation_t x86_spectrev2_mitigation =
1574     X86_SPECTREV2_RETPOLINE;
1575 
1576 /*
1577  * The mitigation status for TAA:
1578  * X86_TAA_NOTHING -- no mitigation available for TAA side-channels
1579  * X86_TAA_DISABLED -- mitigation disabled via x86_disable_taa
1580  * X86_TAA_MD_CLEAR -- MDS mitigation also suffices for TAA
1581  * X86_TAA_TSX_FORCE_ABORT -- transactions are forced to abort
1582  * X86_TAA_TSX_DISABLE -- force abort transactions and hide from CPUID
1583  * X86_TAA_HW_MITIGATED -- TSX potentially active but H/W not TAA-vulnerable
1584  */
1585 typedef enum {
1586 	X86_TAA_NOTHING,
1587 	X86_TAA_DISABLED,
1588 	X86_TAA_MD_CLEAR,
1589 	X86_TAA_TSX_FORCE_ABORT,
1590 	X86_TAA_TSX_DISABLE,
1591 	X86_TAA_HW_MITIGATED
1592 } x86_taa_mitigation_t;
1593 
1594 uint_t x86_disable_taa = 0;
1595 static x86_taa_mitigation_t x86_taa_mitigation = X86_TAA_NOTHING;
1596 
1597 uint_t pentiumpro_bug4046376;
1598 
1599 uchar_t x86_featureset[BT_SIZEOFMAP(NUM_X86_FEATURES)];
1600 
1601 static char *x86_feature_names[NUM_X86_FEATURES] = {
1602 	"lgpg",
1603 	"tsc",
1604 	"msr",
1605 	"mtrr",
1606 	"pge",
1607 	"de",
1608 	"cmov",
1609 	"mmx",
1610 	"mca",
1611 	"pae",
1612 	"cv8",
1613 	"pat",
1614 	"sep",
1615 	"sse",
1616 	"sse2",
1617 	"htt",
1618 	"asysc",
1619 	"nx",
1620 	"sse3",
1621 	"cx16",
1622 	"cmp",
1623 	"tscp",
1624 	"mwait",
1625 	"sse4a",
1626 	"cpuid",
1627 	"ssse3",
1628 	"sse4_1",
1629 	"sse4_2",
1630 	"1gpg",
1631 	"clfsh",
1632 	"64",
1633 	"aes",
1634 	"pclmulqdq",
1635 	"xsave",
1636 	"avx",
1637 	"vmx",
1638 	"svm",
1639 	"topoext",
1640 	"f16c",
1641 	"rdrand",
1642 	"x2apic",
1643 	"avx2",
1644 	"bmi1",
1645 	"bmi2",
1646 	"fma",
1647 	"smep",
1648 	"smap",
1649 	"adx",
1650 	"rdseed",
1651 	"mpx",
1652 	"avx512f",
1653 	"avx512dq",
1654 	"avx512pf",
1655 	"avx512er",
1656 	"avx512cd",
1657 	"avx512bw",
1658 	"avx512vl",
1659 	"avx512fma",
1660 	"avx512vbmi",
1661 	"avx512_vpopcntdq",
1662 	"avx512_4vnniw",
1663 	"avx512_4fmaps",
1664 	"xsaveopt",
1665 	"xsavec",
1666 	"xsaves",
1667 	"sha",
1668 	"umip",
1669 	"pku",
1670 	"ospke",
1671 	"pcid",
1672 	"invpcid",
1673 	"ibrs",
1674 	"ibpb",
1675 	"stibp",
1676 	"ssbd",
1677 	"ssbd_virt",
1678 	"rdcl_no",
1679 	"ibrs_all",
1680 	"rsba",
1681 	"ssb_no",
1682 	"stibp_all",
1683 	"flush_cmd",
1684 	"l1d_vmentry_no",
1685 	"fsgsbase",
1686 	"clflushopt",
1687 	"clwb",
1688 	"monitorx",
1689 	"clzero",
1690 	"xop",
1691 	"fma4",
1692 	"tbm",
1693 	"avx512_vnni",
1694 	"amd_pcec",
1695 	"md_clear",
1696 	"mds_no",
1697 	"core_thermal",
1698 	"pkg_thermal",
1699 	"tsx_ctrl",
1700 	"taa_no",
1701 	"ppin",
1702 	"vaes",
1703 	"vpclmulqdq",
1704 	"lfence_serializing",
1705 	"gfni",
1706 	"avx512_vp2intersect",
1707 	"avx512_bitalg",
1708 	"avx512_vbmi2",
1709 	"avx512_bf16",
1710 	"auto_ibrs",
1711 	"rfds_no",
1712 	"rfds_clear",
1713 	"pbrsb_no",
1714 	"bhi_no",
1715 	"bhi_clear"
1716 };
1717 
1718 boolean_t
1719 is_x86_feature(void *featureset, uint_t feature)
1720 {
1721 	ASSERT(feature < NUM_X86_FEATURES);
1722 	return (BT_TEST((ulong_t *)featureset, feature));
1723 }
1724 
1725 void
1726 add_x86_feature(void *featureset, uint_t feature)
1727 {
1728 	ASSERT(feature < NUM_X86_FEATURES);
1729 	BT_SET((ulong_t *)featureset, feature);
1730 }
1731 
1732 void
1733 remove_x86_feature(void *featureset, uint_t feature)
1734 {
1735 	ASSERT(feature < NUM_X86_FEATURES);
1736 	BT_CLEAR((ulong_t *)featureset, feature);
1737 }
1738 
1739 boolean_t
1740 compare_x86_featureset(void *setA, void *setB)
1741 {
1742 	/*
1743 	 * We assume that the unused bits of the bitmap are always zero.
1744 	 */
1745 	if (memcmp(setA, setB, BT_SIZEOFMAP(NUM_X86_FEATURES)) == 0) {
1746 		return (B_TRUE);
1747 	} else {
1748 		return (B_FALSE);
1749 	}
1750 }
1751 
1752 void
1753 print_x86_featureset(void *featureset)
1754 {
1755 	uint_t i;
1756 
1757 	for (i = 0; i < NUM_X86_FEATURES; i++) {
1758 		if (is_x86_feature(featureset, i)) {
1759 			cmn_err(CE_CONT, "?x86_feature: %s\n",
1760 			    x86_feature_names[i]);
1761 		}
1762 	}
1763 }
1764 
1765 /* Note: This is the maximum size for the CPU, not the size of the structure. */
1766 static size_t xsave_state_size = 0;
1767 uint64_t xsave_bv_all = (XFEATURE_LEGACY_FP | XFEATURE_SSE);
1768 boolean_t xsave_force_disable = B_FALSE;
1769 extern int disable_smap;
1770 
1771 /*
1772  * This is set to platform type we are running on.
1773  */
1774 static int platform_type = -1;
1775 
1776 #if !defined(__xpv)
1777 /*
1778  * Variable to patch if hypervisor platform detection needs to be
1779  * disabled (e.g. platform_type will always be HW_NATIVE if this is 0).
1780  */
1781 int enable_platform_detection = 1;
1782 #endif
1783 
1784 /*
1785  * monitor/mwait info.
1786  *
1787  * size_actual and buf_actual are the real address and size allocated to get
1788  * proper mwait_buf alignement.  buf_actual and size_actual should be passed
1789  * to kmem_free().  Currently kmem_alloc() and mwait happen to both use
1790  * processor cache-line alignment, but this is not guarantied in the furture.
1791  */
1792 struct mwait_info {
1793 	size_t		mon_min;	/* min size to avoid missed wakeups */
1794 	size_t		mon_max;	/* size to avoid false wakeups */
1795 	size_t		size_actual;	/* size actually allocated */
1796 	void		*buf_actual;	/* memory actually allocated */
1797 	uint32_t	support;	/* processor support of monitor/mwait */
1798 };
1799 
1800 /*
1801  * xsave/xrestor info.
1802  *
1803  * This structure contains HW feature bits and the size of the xsave save area.
1804  * Note: the kernel declares a fixed size (AVX_XSAVE_SIZE) structure
1805  * (xsave_state) to describe the xsave layout. However, at runtime the
1806  * per-lwp xsave area is dynamically allocated based on xsav_max_size. The
1807  * xsave_state structure simply represents the legacy layout of the beginning
1808  * of the xsave area.
1809  */
1810 struct xsave_info {
1811 	uint32_t	xsav_hw_features_low;   /* Supported HW features */
1812 	uint32_t	xsav_hw_features_high;  /* Supported HW features */
1813 	size_t		xsav_max_size;  /* max size save area for HW features */
1814 	size_t		ymm_size;	/* AVX: size of ymm save area */
1815 	size_t		ymm_offset;	/* AVX: offset for ymm save area */
1816 	size_t		bndregs_size;	/* MPX: size of bndregs save area */
1817 	size_t		bndregs_offset;	/* MPX: offset for bndregs save area */
1818 	size_t		bndcsr_size;	/* MPX: size of bndcsr save area */
1819 	size_t		bndcsr_offset;	/* MPX: offset for bndcsr save area */
1820 	size_t		opmask_size;	/* AVX512: size of opmask save */
1821 	size_t		opmask_offset;	/* AVX512: offset for opmask save */
1822 	size_t		zmmlo_size;	/* AVX512: size of zmm 256 save */
1823 	size_t		zmmlo_offset;	/* AVX512: offset for zmm 256 save */
1824 	size_t		zmmhi_size;	/* AVX512: size of zmm hi reg save */
1825 	size_t		zmmhi_offset;	/* AVX512: offset for zmm hi reg save */
1826 };
1827 
1828 
1829 /*
1830  * These constants determine how many of the elements of the
1831  * cpuid we cache in the cpuid_info data structure; the
1832  * remaining elements are accessible via the cpuid instruction.
1833  */
1834 
1835 #define	NMAX_CPI_STD	8		/* eax = 0 .. 7 */
1836 #define	NMAX_CPI_EXTD	0x22		/* eax = 0x80000000 .. 0x80000021 */
1837 #define	NMAX_CPI_TOPO	0x10		/* Sanity check on leaf 8X26, 1F */
1838 
1839 /*
1840  * See the big theory statement for a more detailed explanation of what some of
1841  * these members mean.
1842  */
1843 struct cpuid_info {
1844 	uint_t cpi_pass;		/* last pass completed */
1845 	/*
1846 	 * standard function information
1847 	 */
1848 	uint_t cpi_maxeax;		/* fn 0: %eax */
1849 	char cpi_vendorstr[13];		/* fn 0: %ebx:%ecx:%edx */
1850 	uint_t cpi_vendor;		/* enum of cpi_vendorstr */
1851 
1852 	uint_t cpi_family;		/* fn 1: extended family */
1853 	uint_t cpi_model;		/* fn 1: extended model */
1854 	uint_t cpi_step;		/* fn 1: stepping */
1855 	chipid_t cpi_chipid;		/* fn 1: %ebx:  Intel: chip # */
1856 					/*		AMD: package/socket # */
1857 	uint_t cpi_brandid;		/* fn 1: %ebx: brand ID */
1858 	int cpi_clogid;			/* fn 1: %ebx: thread # */
1859 	uint_t cpi_ncpu_per_chip;	/* fn 1: %ebx: logical cpu count */
1860 	uint8_t cpi_cacheinfo[16];	/* fn 2: intel-style cache desc */
1861 	uint_t cpi_ncache;		/* fn 2: number of elements */
1862 	uint_t cpi_ncpu_shr_last_cache;	/* fn 4: %eax: ncpus sharing cache */
1863 	id_t cpi_last_lvl_cacheid;	/* fn 4: %eax: derived cache id */
1864 	uint_t cpi_cache_leaf_size;	/* Number of cache elements */
1865 					/* Intel fn: 4, AMD fn: 8000001d */
1866 	struct cpuid_regs **cpi_cache_leaves;	/* Actual leaves from above */
1867 	struct cpuid_regs cpi_std[NMAX_CPI_STD];	/* 0 .. 7 */
1868 	struct cpuid_regs cpi_sub7[2];	/* Leaf 7, sub-leaves 1-2 */
1869 	/*
1870 	 * extended function information
1871 	 */
1872 	uint_t cpi_xmaxeax;		/* fn 0x80000000: %eax */
1873 	char cpi_brandstr[49];		/* fn 0x8000000[234] */
1874 	uint8_t cpi_pabits;		/* fn 0x80000006: %eax */
1875 	uint8_t	cpi_vabits;		/* fn 0x80000006: %eax */
1876 	uint8_t cpi_fp_amd_save;	/* AMD: FP error pointer save rqd. */
1877 	struct	cpuid_regs cpi_extd[NMAX_CPI_EXTD];	/* 0x800000XX */
1878 
1879 	id_t cpi_coreid;		/* same coreid => strands share core */
1880 	int cpi_pkgcoreid;		/* core number within single package */
1881 	uint_t cpi_ncore_per_chip;	/* AMD: fn 0x80000008: %ecx[7-0] */
1882 					/* Intel: fn 4: %eax[31-26] */
1883 
1884 	/*
1885 	 * These values represent the number of bits that are required to store
1886 	 * information about the number of cores and threads.
1887 	 */
1888 	uint_t cpi_ncore_bits;
1889 	uint_t cpi_nthread_bits;
1890 	/*
1891 	 * supported feature information
1892 	 */
1893 	uint32_t cpi_support[6];
1894 #define	STD_EDX_FEATURES	0
1895 #define	AMD_EDX_FEATURES	1
1896 #define	TM_EDX_FEATURES		2
1897 #define	STD_ECX_FEATURES	3
1898 #define	AMD_ECX_FEATURES	4
1899 #define	STD_EBX_FEATURES	5
1900 	/*
1901 	 * Synthesized information, where known.
1902 	 */
1903 	x86_chiprev_t cpi_chiprev;	/* See X86_CHIPREV_* in x86_archext.h */
1904 	const char *cpi_chiprevstr;	/* May be NULL if chiprev unknown */
1905 	uint32_t cpi_socket;		/* Chip package/socket type */
1906 	x86_uarchrev_t cpi_uarchrev;	/* Microarchitecture and revision */
1907 
1908 	struct mwait_info cpi_mwait;	/* fn 5: monitor/mwait info */
1909 	uint32_t cpi_apicid;
1910 	uint_t cpi_procnodeid;		/* AMD: nodeID on HT, Intel: chipid */
1911 	uint_t cpi_procnodes_per_pkg;	/* AMD: # of nodes in the package */
1912 					/* Intel: 1 */
1913 	uint_t cpi_compunitid;		/* AMD: ComputeUnit ID, Intel: coreid */
1914 	uint_t cpi_cores_per_compunit;	/* AMD: # of cores in the ComputeUnit */
1915 
1916 	struct xsave_info cpi_xsave;	/* fn D: xsave/xrestor info */
1917 
1918 	/*
1919 	 * AMD and Intel extended topology information. Leaf 8X26 (AMD) and
1920 	 * eventually leaf 0x1F (Intel).
1921 	 */
1922 	uint_t cpi_topo_nleaves;
1923 	struct cpuid_regs cpi_topo[NMAX_CPI_TOPO];
1924 };
1925 
1926 
1927 static struct cpuid_info cpuid_info0;
1928 
1929 /*
1930  * These bit fields are defined by the Intel Application Note AP-485
1931  * "Intel Processor Identification and the CPUID Instruction"
1932  */
1933 #define	CPI_FAMILY_XTD(cpi)	BITX((cpi)->cpi_std[1].cp_eax, 27, 20)
1934 #define	CPI_MODEL_XTD(cpi)	BITX((cpi)->cpi_std[1].cp_eax, 19, 16)
1935 #define	CPI_TYPE(cpi)		BITX((cpi)->cpi_std[1].cp_eax, 13, 12)
1936 #define	CPI_FAMILY(cpi)		BITX((cpi)->cpi_std[1].cp_eax, 11, 8)
1937 #define	CPI_STEP(cpi)		BITX((cpi)->cpi_std[1].cp_eax, 3, 0)
1938 #define	CPI_MODEL(cpi)		BITX((cpi)->cpi_std[1].cp_eax, 7, 4)
1939 
1940 #define	CPI_FEATURES_EDX(cpi)		((cpi)->cpi_std[1].cp_edx)
1941 #define	CPI_FEATURES_ECX(cpi)		((cpi)->cpi_std[1].cp_ecx)
1942 #define	CPI_FEATURES_XTD_EDX(cpi)	((cpi)->cpi_extd[1].cp_edx)
1943 #define	CPI_FEATURES_XTD_ECX(cpi)	((cpi)->cpi_extd[1].cp_ecx)
1944 #define	CPI_FEATURES_7_0_EBX(cpi)	((cpi)->cpi_std[7].cp_ebx)
1945 #define	CPI_FEATURES_7_0_ECX(cpi)	((cpi)->cpi_std[7].cp_ecx)
1946 #define	CPI_FEATURES_7_0_EDX(cpi)	((cpi)->cpi_std[7].cp_edx)
1947 #define	CPI_FEATURES_7_1_EAX(cpi)	((cpi)->cpi_sub7[0].cp_eax)
1948 #define	CPI_FEATURES_7_2_EDX(cpi)	((cpi)->cpi_sub7[1].cp_edx)
1949 
1950 #define	CPI_BRANDID(cpi)	BITX((cpi)->cpi_std[1].cp_ebx, 7, 0)
1951 #define	CPI_CHUNKS(cpi)		BITX((cpi)->cpi_std[1].cp_ebx, 15, 7)
1952 #define	CPI_CPU_COUNT(cpi)	BITX((cpi)->cpi_std[1].cp_ebx, 23, 16)
1953 #define	CPI_APIC_ID(cpi)	BITX((cpi)->cpi_std[1].cp_ebx, 31, 24)
1954 
1955 #define	CPI_MAXEAX_MAX		0x100		/* sanity control */
1956 #define	CPI_XMAXEAX_MAX		0x80000100
1957 #define	CPI_FN4_ECX_MAX		0x20		/* sanity: max fn 4 levels */
1958 #define	CPI_FNB_ECX_MAX		0x20		/* sanity: max fn B levels */
1959 
1960 /*
1961  * Function 4 (Deterministic Cache Parameters) macros
1962  * Defined by Intel Application Note AP-485
1963  */
1964 #define	CPI_NUM_CORES(regs)		BITX((regs)->cp_eax, 31, 26)
1965 #define	CPI_NTHR_SHR_CACHE(regs)	BITX((regs)->cp_eax, 25, 14)
1966 #define	CPI_FULL_ASSOC_CACHE(regs)	BITX((regs)->cp_eax, 9, 9)
1967 #define	CPI_SELF_INIT_CACHE(regs)	BITX((regs)->cp_eax, 8, 8)
1968 #define	CPI_CACHE_LVL(regs)		BITX((regs)->cp_eax, 7, 5)
1969 #define	CPI_CACHE_TYPE(regs)		BITX((regs)->cp_eax, 4, 0)
1970 #define	CPI_CACHE_TYPE_DONE	0
1971 #define	CPI_CACHE_TYPE_DATA	1
1972 #define	CPI_CACHE_TYPE_INSTR	2
1973 #define	CPI_CACHE_TYPE_UNIFIED	3
1974 #define	CPI_CPU_LEVEL_TYPE(regs)	BITX((regs)->cp_ecx, 15, 8)
1975 
1976 #define	CPI_CACHE_WAYS(regs)		BITX((regs)->cp_ebx, 31, 22)
1977 #define	CPI_CACHE_PARTS(regs)		BITX((regs)->cp_ebx, 21, 12)
1978 #define	CPI_CACHE_COH_LN_SZ(regs)	BITX((regs)->cp_ebx, 11, 0)
1979 
1980 #define	CPI_CACHE_SETS(regs)		BITX((regs)->cp_ecx, 31, 0)
1981 
1982 #define	CPI_PREFCH_STRIDE(regs)		BITX((regs)->cp_edx, 9, 0)
1983 
1984 
1985 /*
1986  * A couple of shorthand macros to identify "later" P6-family chips
1987  * like the Pentium M and Core.  First, the "older" P6-based stuff
1988  * (loosely defined as "pre-Pentium-4"):
1989  * P6, PII, Mobile PII, PII Xeon, PIII, Mobile PIII, PIII Xeon
1990  */
1991 #define	IS_LEGACY_P6(cpi) (			\
1992 	cpi->cpi_family == 6 &&			\
1993 		(cpi->cpi_model == 1 ||		\
1994 		cpi->cpi_model == 3 ||		\
1995 		cpi->cpi_model == 5 ||		\
1996 		cpi->cpi_model == 6 ||		\
1997 		cpi->cpi_model == 7 ||		\
1998 		cpi->cpi_model == 8 ||		\
1999 		cpi->cpi_model == 0xA ||	\
2000 		cpi->cpi_model == 0xB)		\
2001 )
2002 
2003 /* A "new F6" is everything with family 6 that's not the above */
2004 #define	IS_NEW_F6(cpi) ((cpi->cpi_family == 6) && !IS_LEGACY_P6(cpi))
2005 
2006 /* Extended family/model support */
2007 #define	IS_EXTENDED_MODEL_INTEL(cpi) (cpi->cpi_family == 0x6 || \
2008 	cpi->cpi_family >= 0xf)
2009 
2010 /*
2011  * Info for monitor/mwait idle loop.
2012  *
2013  * See cpuid section of "Intel 64 and IA-32 Architectures Software Developer's
2014  * Manual Volume 2A: Instruction Set Reference, A-M" #25366-022US, November
2015  * 2006.
2016  * See MONITOR/MWAIT section of "AMD64 Architecture Programmer's Manual
2017  * Documentation Updates" #33633, Rev 2.05, December 2006.
2018  */
2019 #define	MWAIT_SUPPORT		(0x00000001)	/* mwait supported */
2020 #define	MWAIT_EXTENSIONS	(0x00000002)	/* extenstion supported */
2021 #define	MWAIT_ECX_INT_ENABLE	(0x00000004)	/* ecx 1 extension supported */
2022 #define	MWAIT_SUPPORTED(cpi)	((cpi)->cpi_std[1].cp_ecx & CPUID_INTC_ECX_MON)
2023 #define	MWAIT_INT_ENABLE(cpi)	((cpi)->cpi_std[5].cp_ecx & 0x2)
2024 #define	MWAIT_EXTENSION(cpi)	((cpi)->cpi_std[5].cp_ecx & 0x1)
2025 #define	MWAIT_SIZE_MIN(cpi)	BITX((cpi)->cpi_std[5].cp_eax, 15, 0)
2026 #define	MWAIT_SIZE_MAX(cpi)	BITX((cpi)->cpi_std[5].cp_ebx, 15, 0)
2027 /*
2028  * Number of sub-cstates for a given c-state.
2029  */
2030 #define	MWAIT_NUM_SUBC_STATES(cpi, c_state)			\
2031 	BITX((cpi)->cpi_std[5].cp_edx, c_state + 3, c_state)
2032 
2033 /*
2034  * XSAVE leaf 0xD enumeration
2035  */
2036 #define	CPUID_LEAFD_2_YMM_OFFSET	576
2037 #define	CPUID_LEAFD_2_YMM_SIZE		256
2038 
2039 /*
2040  * Common extended leaf names to cut down on typos.
2041  */
2042 #define	CPUID_LEAF_EXT_0		0x80000000
2043 #define	CPUID_LEAF_EXT_8		0x80000008
2044 #define	CPUID_LEAF_EXT_1d		0x8000001d
2045 #define	CPUID_LEAF_EXT_1e		0x8000001e
2046 #define	CPUID_LEAF_EXT_21		0x80000021
2047 #define	CPUID_LEAF_EXT_26		0x80000026
2048 
2049 /*
2050  * Functions we consume from cpuid_subr.c;  don't publish these in a header
2051  * file to try and keep people using the expected cpuid_* interfaces.
2052  */
2053 extern uint32_t _cpuid_skt(uint_t, uint_t, uint_t, uint_t);
2054 extern const char *_cpuid_sktstr(uint_t, uint_t, uint_t, uint_t);
2055 extern x86_chiprev_t _cpuid_chiprev(uint_t, uint_t, uint_t, uint_t);
2056 extern const char *_cpuid_chiprevstr(uint_t, uint_t, uint_t, uint_t);
2057 extern x86_uarchrev_t _cpuid_uarchrev(uint_t, uint_t, uint_t, uint_t);
2058 extern uint_t _cpuid_vendorstr_to_vendorcode(char *);
2059 
2060 /*
2061  * Apply up various platform-dependent restrictions where the
2062  * underlying platform restrictions mean the CPU can be marked
2063  * as less capable than its cpuid instruction would imply.
2064  */
2065 #if defined(__xpv)
2066 static void
2067 platform_cpuid_mangle(uint_t vendor, uint32_t eax, struct cpuid_regs *cp)
2068 {
2069 	switch (eax) {
2070 	case 1: {
2071 		uint32_t mcamask = DOMAIN_IS_INITDOMAIN(xen_info) ?
2072 		    0 : CPUID_INTC_EDX_MCA;
2073 		cp->cp_edx &=
2074 		    ~(mcamask |
2075 		    CPUID_INTC_EDX_PSE |
2076 		    CPUID_INTC_EDX_VME | CPUID_INTC_EDX_DE |
2077 		    CPUID_INTC_EDX_SEP | CPUID_INTC_EDX_MTRR |
2078 		    CPUID_INTC_EDX_PGE | CPUID_INTC_EDX_PAT |
2079 		    CPUID_AMD_EDX_SYSC | CPUID_INTC_EDX_SEP |
2080 		    CPUID_INTC_EDX_PSE36 | CPUID_INTC_EDX_HTT);
2081 		break;
2082 	}
2083 
2084 	case 0x80000001:
2085 		cp->cp_edx &=
2086 		    ~(CPUID_AMD_EDX_PSE |
2087 		    CPUID_INTC_EDX_VME | CPUID_INTC_EDX_DE |
2088 		    CPUID_AMD_EDX_MTRR | CPUID_AMD_EDX_PGE |
2089 		    CPUID_AMD_EDX_PAT | CPUID_AMD_EDX_PSE36 |
2090 		    CPUID_AMD_EDX_SYSC | CPUID_INTC_EDX_SEP |
2091 		    CPUID_AMD_EDX_TSCP);
2092 		cp->cp_ecx &= ~CPUID_AMD_ECX_CMP_LGCY;
2093 		break;
2094 	default:
2095 		break;
2096 	}
2097 
2098 	switch (vendor) {
2099 	case X86_VENDOR_Intel:
2100 		switch (eax) {
2101 		case 4:
2102 			/*
2103 			 * Zero out the (ncores-per-chip - 1) field
2104 			 */
2105 			cp->cp_eax &= 0x03fffffff;
2106 			break;
2107 		default:
2108 			break;
2109 		}
2110 		break;
2111 	case X86_VENDOR_AMD:
2112 	case X86_VENDOR_HYGON:
2113 		switch (eax) {
2114 
2115 		case 0x80000001:
2116 			cp->cp_ecx &= ~CPUID_AMD_ECX_CR8D;
2117 			break;
2118 
2119 		case CPUID_LEAF_EXT_8:
2120 			/*
2121 			 * Zero out the (ncores-per-chip - 1) field
2122 			 */
2123 			cp->cp_ecx &= 0xffffff00;
2124 			break;
2125 		default:
2126 			break;
2127 		}
2128 		break;
2129 	default:
2130 		break;
2131 	}
2132 }
2133 #else
2134 #define	platform_cpuid_mangle(vendor, eax, cp)	/* nothing */
2135 #endif
2136 
2137 /*
2138  *  Some undocumented ways of patching the results of the cpuid
2139  *  instruction to permit running Solaris 10 on future cpus that
2140  *  we don't currently support.  Could be set to non-zero values
2141  *  via settings in eeprom.
2142  */
2143 
2144 uint32_t cpuid_feature_ecx_include;
2145 uint32_t cpuid_feature_ecx_exclude;
2146 uint32_t cpuid_feature_edx_include;
2147 uint32_t cpuid_feature_edx_exclude;
2148 
2149 /*
2150  * Allocate space for mcpu_cpi in the machcpu structure for all non-boot CPUs.
2151  */
2152 void
2153 cpuid_alloc_space(cpu_t *cpu)
2154 {
2155 	/*
2156 	 * By convention, cpu0 is the boot cpu, which is set up
2157 	 * before memory allocation is available.  All other cpus get
2158 	 * their cpuid_info struct allocated here.
2159 	 */
2160 	ASSERT(cpu->cpu_id != 0);
2161 	ASSERT(cpu->cpu_m.mcpu_cpi == NULL);
2162 	cpu->cpu_m.mcpu_cpi =
2163 	    kmem_zalloc(sizeof (*cpu->cpu_m.mcpu_cpi), KM_SLEEP);
2164 }
2165 
2166 void
2167 cpuid_free_space(cpu_t *cpu)
2168 {
2169 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2170 	int i;
2171 
2172 	ASSERT(cpi != NULL);
2173 	ASSERT(cpi != &cpuid_info0);
2174 
2175 	/*
2176 	 * Free up any cache leaf related dynamic storage. The first entry was
2177 	 * cached from the standard cpuid storage, so we should not free it.
2178 	 */
2179 	for (i = 1; i < cpi->cpi_cache_leaf_size; i++)
2180 		kmem_free(cpi->cpi_cache_leaves[i], sizeof (struct cpuid_regs));
2181 	if (cpi->cpi_cache_leaf_size > 0)
2182 		kmem_free(cpi->cpi_cache_leaves,
2183 		    cpi->cpi_cache_leaf_size * sizeof (struct cpuid_regs *));
2184 
2185 	kmem_free(cpi, sizeof (*cpi));
2186 	cpu->cpu_m.mcpu_cpi = NULL;
2187 }
2188 
2189 #if !defined(__xpv)
2190 /*
2191  * Determine the type of the underlying platform. This is used to customize
2192  * initialization of various subsystems (e.g. TSC). determine_platform() must
2193  * only ever be called once to prevent two processors from seeing different
2194  * values of platform_type. Must be called before cpuid_pass_ident(), the
2195  * earliest consumer to execute; the identification pass will call
2196  * synth_amd_info() to compute the chiprev, which in turn calls get_hwenv().
2197  */
2198 void
2199 determine_platform(void)
2200 {
2201 	struct cpuid_regs cp;
2202 	uint32_t base;
2203 	uint32_t regs[4];
2204 	char *hvstr = (char *)regs;
2205 
2206 	ASSERT(platform_type == -1);
2207 
2208 	platform_type = HW_NATIVE;
2209 
2210 	if (!enable_platform_detection)
2211 		return;
2212 
2213 	/*
2214 	 * If Hypervisor CPUID bit is set, try to determine hypervisor
2215 	 * vendor signature, and set platform type accordingly.
2216 	 *
2217 	 * References:
2218 	 * http://lkml.org/lkml/2008/10/1/246
2219 	 * http://kb.vmware.com/kb/1009458
2220 	 */
2221 	cp.cp_eax = 0x1;
2222 	(void) __cpuid_insn(&cp);
2223 	if ((cp.cp_ecx & CPUID_INTC_ECX_HV) != 0) {
2224 		cp.cp_eax = 0x40000000;
2225 		(void) __cpuid_insn(&cp);
2226 		regs[0] = cp.cp_ebx;
2227 		regs[1] = cp.cp_ecx;
2228 		regs[2] = cp.cp_edx;
2229 		regs[3] = 0;
2230 		if (strcmp(hvstr, HVSIG_XEN_HVM) == 0) {
2231 			platform_type = HW_XEN_HVM;
2232 			return;
2233 		}
2234 		if (strcmp(hvstr, HVSIG_VMWARE) == 0) {
2235 			platform_type = HW_VMWARE;
2236 			return;
2237 		}
2238 		if (strcmp(hvstr, HVSIG_KVM) == 0) {
2239 			platform_type = HW_KVM;
2240 			return;
2241 		}
2242 		if (strcmp(hvstr, HVSIG_BHYVE) == 0) {
2243 			platform_type = HW_BHYVE;
2244 			return;
2245 		}
2246 		if (strcmp(hvstr, HVSIG_MICROSOFT) == 0) {
2247 			platform_type = HW_MICROSOFT;
2248 			return;
2249 		}
2250 		if (strcmp(hvstr, HVSIG_QEMU_TCG) == 0) {
2251 			platform_type = HW_QEMU_TCG;
2252 			return;
2253 		}
2254 		if (strcmp(hvstr, HVSIG_VIRTUALBOX) == 0) {
2255 			platform_type = HW_VIRTUALBOX;
2256 			return;
2257 		}
2258 		if (strcmp(hvstr, HVSIG_ACRN) == 0) {
2259 			platform_type = HW_ACRN;
2260 			return;
2261 		}
2262 	} else {
2263 		/*
2264 		 * Check older VMware hardware versions. VMware hypervisor is
2265 		 * detected by performing an IN operation to VMware hypervisor
2266 		 * port and checking that value returned in %ebx is VMware
2267 		 * hypervisor magic value.
2268 		 *
2269 		 * References: http://kb.vmware.com/kb/1009458
2270 		 */
2271 		vmware_port(VMWARE_HVCMD_GETVERSION, regs);
2272 		if (regs[1] == VMWARE_HVMAGIC) {
2273 			platform_type = HW_VMWARE;
2274 			return;
2275 		}
2276 	}
2277 
2278 	/*
2279 	 * Check Xen hypervisor. In a fully virtualized domain,
2280 	 * Xen's pseudo-cpuid function returns a string representing the
2281 	 * Xen signature in %ebx, %ecx, and %edx. %eax contains the maximum
2282 	 * supported cpuid function. We need at least a (base + 2) leaf value
2283 	 * to do what we want to do. Try different base values, since the
2284 	 * hypervisor might use a different one depending on whether Hyper-V
2285 	 * emulation is switched on by default or not.
2286 	 */
2287 	for (base = 0x40000000; base < 0x40010000; base += 0x100) {
2288 		cp.cp_eax = base;
2289 		(void) __cpuid_insn(&cp);
2290 		regs[0] = cp.cp_ebx;
2291 		regs[1] = cp.cp_ecx;
2292 		regs[2] = cp.cp_edx;
2293 		regs[3] = 0;
2294 		if (strcmp(hvstr, HVSIG_XEN_HVM) == 0 &&
2295 		    cp.cp_eax >= (base + 2)) {
2296 			platform_type &= ~HW_NATIVE;
2297 			platform_type |= HW_XEN_HVM;
2298 			return;
2299 		}
2300 	}
2301 }
2302 
2303 int
2304 get_hwenv(void)
2305 {
2306 	ASSERT(platform_type != -1);
2307 	return (platform_type);
2308 }
2309 
2310 int
2311 is_controldom(void)
2312 {
2313 	return (0);
2314 }
2315 
2316 #else
2317 
2318 int
2319 get_hwenv(void)
2320 {
2321 	return (HW_XEN_PV);
2322 }
2323 
2324 int
2325 is_controldom(void)
2326 {
2327 	return (DOMAIN_IS_INITDOMAIN(xen_info));
2328 }
2329 
2330 #endif	/* __xpv */
2331 
2332 /*
2333  * Gather the extended topology information. This should be the same for both
2334  * AMD leaf 8X26 and Intel leaf 0x1F (though the data interpretation varies).
2335  */
2336 static void
2337 cpuid_gather_ext_topo_leaf(struct cpuid_info *cpi, uint32_t leaf)
2338 {
2339 	uint_t i;
2340 
2341 	for (i = 0; i < ARRAY_SIZE(cpi->cpi_topo); i++) {
2342 		struct cpuid_regs *regs = &cpi->cpi_topo[i];
2343 
2344 		bzero(regs, sizeof (struct cpuid_regs));
2345 		regs->cp_eax = leaf;
2346 		regs->cp_ecx = i;
2347 
2348 		(void) __cpuid_insn(regs);
2349 		if (CPUID_AMD_8X26_ECX_TYPE(regs->cp_ecx) ==
2350 		    CPUID_AMD_8X26_TYPE_DONE) {
2351 			break;
2352 		}
2353 	}
2354 
2355 	cpi->cpi_topo_nleaves = i;
2356 }
2357 
2358 /*
2359  * Make sure that we have gathered all of the CPUID leaves that we might need to
2360  * determine topology. We assume that the standard leaf 1 has already been done
2361  * and that xmaxeax has already been calculated.
2362  */
2363 static void
2364 cpuid_gather_amd_topology_leaves(cpu_t *cpu)
2365 {
2366 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2367 
2368 	if (cpi->cpi_xmaxeax >= CPUID_LEAF_EXT_8) {
2369 		struct cpuid_regs *cp;
2370 
2371 		cp = &cpi->cpi_extd[8];
2372 		cp->cp_eax = CPUID_LEAF_EXT_8;
2373 		(void) __cpuid_insn(cp);
2374 		platform_cpuid_mangle(cpi->cpi_vendor, CPUID_LEAF_EXT_8, cp);
2375 	}
2376 
2377 	if (is_x86_feature(x86_featureset, X86FSET_TOPOEXT) &&
2378 	    cpi->cpi_xmaxeax >= CPUID_LEAF_EXT_1e) {
2379 		struct cpuid_regs *cp;
2380 
2381 		cp = &cpi->cpi_extd[0x1e];
2382 		cp->cp_eax = CPUID_LEAF_EXT_1e;
2383 		(void) __cpuid_insn(cp);
2384 	}
2385 
2386 	if (cpi->cpi_xmaxeax >= CPUID_LEAF_EXT_26) {
2387 		cpuid_gather_ext_topo_leaf(cpi, CPUID_LEAF_EXT_26);
2388 	}
2389 }
2390 
2391 /*
2392  * Get the APIC ID for this processor. If Leaf B is present and valid, we prefer
2393  * it to everything else. If not, and we're on an AMD system where 8000001e is
2394  * valid, then we use that. Othewrise, we fall back to the default value for the
2395  * APIC ID in leaf 1.
2396  */
2397 static uint32_t
2398 cpuid_gather_apicid(struct cpuid_info *cpi)
2399 {
2400 	/*
2401 	 * Leaf B changes based on the arguments to it. Because we don't cache
2402 	 * it, we need to gather it again.
2403 	 */
2404 	if (cpi->cpi_maxeax >= 0xB) {
2405 		struct cpuid_regs regs;
2406 		struct cpuid_regs *cp;
2407 
2408 		cp = &regs;
2409 		cp->cp_eax = 0xB;
2410 		cp->cp_edx = cp->cp_ebx = cp->cp_ecx = 0;
2411 		(void) __cpuid_insn(cp);
2412 
2413 		if (cp->cp_ebx != 0) {
2414 			return (cp->cp_edx);
2415 		}
2416 	}
2417 
2418 	if ((cpi->cpi_vendor == X86_VENDOR_AMD ||
2419 	    cpi->cpi_vendor == X86_VENDOR_HYGON) &&
2420 	    is_x86_feature(x86_featureset, X86FSET_TOPOEXT) &&
2421 	    cpi->cpi_xmaxeax >= CPUID_LEAF_EXT_1e) {
2422 		return (cpi->cpi_extd[0x1e].cp_eax);
2423 	}
2424 
2425 	return (CPI_APIC_ID(cpi));
2426 }
2427 
2428 /*
2429  * For AMD processors, attempt to calculate the number of chips and cores that
2430  * exist. The way that we do this varies based on the generation, because the
2431  * generations themselves have changed dramatically.
2432  *
2433  * If cpuid leaf 0x80000008 exists, that generally tells us the number of cores.
2434  * However, with the advent of family 17h (Zen) it actually tells us the number
2435  * of threads, so we need to look at leaf 0x8000001e if available to determine
2436  * its value. Otherwise, for all prior families, the number of enabled cores is
2437  * the same as threads.
2438  *
2439  * If we do not have leaf 0x80000008, then we assume that this processor does
2440  * not have anything. AMD's older CPUID specification says there's no reason to
2441  * fall back to leaf 1.
2442  *
2443  * In some virtualization cases we will not have leaf 8000001e or it will be
2444  * zero. When that happens we assume the number of threads is one.
2445  */
2446 static void
2447 cpuid_amd_ncores(struct cpuid_info *cpi, uint_t *ncpus, uint_t *ncores)
2448 {
2449 	uint_t nthreads, nthread_per_core;
2450 
2451 	nthreads = nthread_per_core = 1;
2452 
2453 	if (cpi->cpi_xmaxeax >= CPUID_LEAF_EXT_8) {
2454 		nthreads = BITX(cpi->cpi_extd[8].cp_ecx, 7, 0) + 1;
2455 	} else if ((cpi->cpi_std[1].cp_edx & CPUID_INTC_EDX_HTT) != 0) {
2456 		nthreads = CPI_CPU_COUNT(cpi);
2457 	}
2458 
2459 	/*
2460 	 * For us to have threads, and know about it, we have to be at least at
2461 	 * family 17h and have the cpuid bit that says we have extended
2462 	 * topology.
2463 	 */
2464 	if (cpi->cpi_family >= 0x17 &&
2465 	    is_x86_feature(x86_featureset, X86FSET_TOPOEXT) &&
2466 	    cpi->cpi_xmaxeax >= CPUID_LEAF_EXT_1e) {
2467 		nthread_per_core = BITX(cpi->cpi_extd[0x1e].cp_ebx, 15, 8) + 1;
2468 	}
2469 
2470 	*ncpus = nthreads;
2471 	*ncores = nthreads / nthread_per_core;
2472 }
2473 
2474 /*
2475  * Seed the initial values for the cores and threads for an Intel based
2476  * processor. These values will be overwritten if we detect that the processor
2477  * supports CPUID leaf 0xb.
2478  */
2479 static void
2480 cpuid_intel_ncores(struct cpuid_info *cpi, uint_t *ncpus, uint_t *ncores)
2481 {
2482 	/*
2483 	 * Only seed the number of physical cores from the first level leaf 4
2484 	 * information. The number of threads there indicate how many share the
2485 	 * L1 cache, which may or may not have anything to do with the number of
2486 	 * logical CPUs per core.
2487 	 */
2488 	if (cpi->cpi_maxeax >= 4) {
2489 		*ncores = BITX(cpi->cpi_std[4].cp_eax, 31, 26) + 1;
2490 	} else {
2491 		*ncores = 1;
2492 	}
2493 
2494 	if ((cpi->cpi_std[1].cp_edx & CPUID_INTC_EDX_HTT) != 0) {
2495 		*ncpus = CPI_CPU_COUNT(cpi);
2496 	} else {
2497 		*ncpus = *ncores;
2498 	}
2499 }
2500 
2501 static boolean_t
2502 cpuid_leafB_getids(cpu_t *cpu)
2503 {
2504 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2505 	struct cpuid_regs regs;
2506 	struct cpuid_regs *cp;
2507 
2508 	if (cpi->cpi_maxeax < 0xB)
2509 		return (B_FALSE);
2510 
2511 	cp = &regs;
2512 	cp->cp_eax = 0xB;
2513 	cp->cp_edx = cp->cp_ebx = cp->cp_ecx = 0;
2514 
2515 	(void) __cpuid_insn(cp);
2516 
2517 	/*
2518 	 * Check CPUID.EAX=0BH, ECX=0H:EBX is non-zero, which
2519 	 * indicates that the extended topology enumeration leaf is
2520 	 * available.
2521 	 */
2522 	if (cp->cp_ebx != 0) {
2523 		uint32_t x2apic_id = 0;
2524 		uint_t coreid_shift = 0;
2525 		uint_t ncpu_per_core = 1;
2526 		uint_t chipid_shift = 0;
2527 		uint_t ncpu_per_chip = 1;
2528 		uint_t i;
2529 		uint_t level;
2530 
2531 		for (i = 0; i < CPI_FNB_ECX_MAX; i++) {
2532 			cp->cp_eax = 0xB;
2533 			cp->cp_ecx = i;
2534 
2535 			(void) __cpuid_insn(cp);
2536 			level = CPI_CPU_LEVEL_TYPE(cp);
2537 
2538 			if (level == 1) {
2539 				x2apic_id = cp->cp_edx;
2540 				coreid_shift = BITX(cp->cp_eax, 4, 0);
2541 				ncpu_per_core = BITX(cp->cp_ebx, 15, 0);
2542 			} else if (level == 2) {
2543 				x2apic_id = cp->cp_edx;
2544 				chipid_shift = BITX(cp->cp_eax, 4, 0);
2545 				ncpu_per_chip = BITX(cp->cp_ebx, 15, 0);
2546 			}
2547 		}
2548 
2549 		/*
2550 		 * cpi_apicid is taken care of in cpuid_gather_apicid.
2551 		 */
2552 		cpi->cpi_ncpu_per_chip = ncpu_per_chip;
2553 		cpi->cpi_ncore_per_chip = ncpu_per_chip /
2554 		    ncpu_per_core;
2555 		cpi->cpi_chipid = x2apic_id >> chipid_shift;
2556 		cpi->cpi_clogid = x2apic_id & ((1 << chipid_shift) - 1);
2557 		cpi->cpi_coreid = x2apic_id >> coreid_shift;
2558 		cpi->cpi_pkgcoreid = cpi->cpi_clogid >> coreid_shift;
2559 		cpi->cpi_procnodeid = cpi->cpi_chipid;
2560 		cpi->cpi_compunitid = cpi->cpi_coreid;
2561 
2562 		if (coreid_shift > 0 && chipid_shift > coreid_shift) {
2563 			cpi->cpi_nthread_bits = coreid_shift;
2564 			cpi->cpi_ncore_bits = chipid_shift - coreid_shift;
2565 		}
2566 
2567 		return (B_TRUE);
2568 	} else {
2569 		return (B_FALSE);
2570 	}
2571 }
2572 
2573 static void
2574 cpuid_intel_getids(cpu_t *cpu, void *feature)
2575 {
2576 	uint_t i;
2577 	uint_t chipid_shift = 0;
2578 	uint_t coreid_shift = 0;
2579 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2580 
2581 	/*
2582 	 * There are no compute units or processor nodes currently on Intel.
2583 	 * Always set these to one.
2584 	 */
2585 	cpi->cpi_procnodes_per_pkg = 1;
2586 	cpi->cpi_cores_per_compunit = 1;
2587 
2588 	/*
2589 	 * If cpuid Leaf B is present, use that to try and get this information.
2590 	 * It will be the most accurate for Intel CPUs.
2591 	 */
2592 	if (cpuid_leafB_getids(cpu))
2593 		return;
2594 
2595 	/*
2596 	 * In this case, we have the leaf 1 and leaf 4 values for ncpu_per_chip
2597 	 * and ncore_per_chip. These represent the largest power of two values
2598 	 * that we need to cover all of the IDs in the system. Therefore, we use
2599 	 * those values to seed the number of bits needed to cover information
2600 	 * in the case when leaf B is not available. These values will probably
2601 	 * be larger than required, but that's OK.
2602 	 */
2603 	cpi->cpi_nthread_bits = ddi_fls(cpi->cpi_ncpu_per_chip);
2604 	cpi->cpi_ncore_bits = ddi_fls(cpi->cpi_ncore_per_chip);
2605 
2606 	for (i = 1; i < cpi->cpi_ncpu_per_chip; i <<= 1)
2607 		chipid_shift++;
2608 
2609 	cpi->cpi_chipid = cpi->cpi_apicid >> chipid_shift;
2610 	cpi->cpi_clogid = cpi->cpi_apicid & ((1 << chipid_shift) - 1);
2611 
2612 	if (is_x86_feature(feature, X86FSET_CMP)) {
2613 		/*
2614 		 * Multi-core (and possibly multi-threaded)
2615 		 * processors.
2616 		 */
2617 		uint_t ncpu_per_core = 0;
2618 
2619 		if (cpi->cpi_ncore_per_chip == 1)
2620 			ncpu_per_core = cpi->cpi_ncpu_per_chip;
2621 		else if (cpi->cpi_ncore_per_chip > 1)
2622 			ncpu_per_core = cpi->cpi_ncpu_per_chip /
2623 			    cpi->cpi_ncore_per_chip;
2624 		/*
2625 		 * 8bit APIC IDs on dual core Pentiums
2626 		 * look like this:
2627 		 *
2628 		 * +-----------------------+------+------+
2629 		 * | Physical Package ID   |  MC  |  HT  |
2630 		 * +-----------------------+------+------+
2631 		 * <------- chipid -------->
2632 		 * <------- coreid --------------->
2633 		 *			   <--- clogid -->
2634 		 *			   <------>
2635 		 *			   pkgcoreid
2636 		 *
2637 		 * Where the number of bits necessary to
2638 		 * represent MC and HT fields together equals
2639 		 * to the minimum number of bits necessary to
2640 		 * store the value of cpi->cpi_ncpu_per_chip.
2641 		 * Of those bits, the MC part uses the number
2642 		 * of bits necessary to store the value of
2643 		 * cpi->cpi_ncore_per_chip.
2644 		 */
2645 		for (i = 1; i < ncpu_per_core; i <<= 1)
2646 			coreid_shift++;
2647 		cpi->cpi_coreid = cpi->cpi_apicid >> coreid_shift;
2648 		cpi->cpi_pkgcoreid = cpi->cpi_clogid >> coreid_shift;
2649 	} else if (is_x86_feature(feature, X86FSET_HTT)) {
2650 		/*
2651 		 * Single-core multi-threaded processors.
2652 		 */
2653 		cpi->cpi_coreid = cpi->cpi_chipid;
2654 		cpi->cpi_pkgcoreid = 0;
2655 	} else {
2656 		/*
2657 		 * Single-core single-thread processors.
2658 		 */
2659 		cpi->cpi_coreid = cpu->cpu_id;
2660 		cpi->cpi_pkgcoreid = 0;
2661 	}
2662 	cpi->cpi_procnodeid = cpi->cpi_chipid;
2663 	cpi->cpi_compunitid = cpi->cpi_coreid;
2664 }
2665 
2666 /*
2667  * Historically, AMD has had CMP chips with only a single thread per core.
2668  * However, starting in family 17h (Zen), this has changed and they now have
2669  * multiple threads. Our internal core id needs to be a unique value.
2670  *
2671  * To determine the core id of an AMD system, if we're from a family before 17h,
2672  * then we just use the cpu id, as that gives us a good value that will be
2673  * unique for each core. If instead, we're on family 17h or later, then we need
2674  * to do something more complicated. CPUID leaf 0x8000001e can tell us
2675  * how many threads are in the system. Based on that, we'll shift the APIC ID.
2676  * We can't use the normal core id in that leaf as it's only unique within the
2677  * socket, which is perfect for cpi_pkgcoreid, but not us.
2678  */
2679 static id_t
2680 cpuid_amd_get_coreid(cpu_t *cpu)
2681 {
2682 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2683 
2684 	if (cpi->cpi_family >= 0x17 &&
2685 	    is_x86_feature(x86_featureset, X86FSET_TOPOEXT) &&
2686 	    cpi->cpi_xmaxeax >= CPUID_LEAF_EXT_1e) {
2687 		uint_t nthreads = BITX(cpi->cpi_extd[0x1e].cp_ebx, 15, 8) + 1;
2688 		if (nthreads > 1) {
2689 			VERIFY3U(nthreads, ==, 2);
2690 			return (cpi->cpi_apicid >> 1);
2691 		}
2692 	}
2693 
2694 	return (cpu->cpu_id);
2695 }
2696 
2697 /*
2698  * IDs on AMD is a more challenging task. This is notable because of the
2699  * following two facts:
2700  *
2701  *  1. Before family 0x17 (Zen), there was no support for SMT and there was
2702  *     also no way to get an actual unique core id from the system. As such, we
2703  *     synthesize this case by using cpu->cpu_id.  This scheme does not,
2704  *     however, guarantee that sibling cores of a chip will have sequential
2705  *     coreids starting at a multiple of the number of cores per chip - that is
2706  *     usually the case, but if the APIC IDs have been set up in a different
2707  *     order then we need to perform a few more gymnastics for the pkgcoreid.
2708  *
2709  *  2. In families 0x15 and 16x (Bulldozer and co.) the cores came in groups
2710  *     called compute units. These compute units share the L1I cache, L2 cache,
2711  *     and the FPU. To deal with this, a new topology leaf was added in
2712  *     0x8000001e. However, parts of this leaf have different meanings
2713  *     once we get to family 0x17.
2714  */
2715 
2716 static void
2717 cpuid_amd_getids(cpu_t *cpu, uchar_t *features)
2718 {
2719 	int i, first_half, coreidsz;
2720 	uint32_t nb_caps_reg;
2721 	uint_t node2_1;
2722 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2723 	struct cpuid_regs *cp;
2724 
2725 	/*
2726 	 * Calculate the core id (this comes from hardware in family 0x17 if it
2727 	 * hasn't been stripped by virtualization). We always set the compute
2728 	 * unit id to the same value. Also, initialize the default number of
2729 	 * cores per compute unit and nodes per package. This will be
2730 	 * overwritten when we know information about a particular family.
2731 	 */
2732 	cpi->cpi_coreid = cpuid_amd_get_coreid(cpu);
2733 	cpi->cpi_compunitid = cpi->cpi_coreid;
2734 	cpi->cpi_cores_per_compunit = 1;
2735 	cpi->cpi_procnodes_per_pkg = 1;
2736 
2737 	/*
2738 	 * To construct the logical ID, we need to determine how many APIC IDs
2739 	 * are dedicated to the cores and threads. This is provided for us in
2740 	 * 0x80000008. However, if it's not present (say due to virtualization),
2741 	 * then we assume it's one. This should be present on all 64-bit AMD
2742 	 * processors.  It was added in family 0xf (Hammer).
2743 	 */
2744 	if (cpi->cpi_xmaxeax >= CPUID_LEAF_EXT_8) {
2745 		coreidsz = BITX((cpi)->cpi_extd[8].cp_ecx, 15, 12);
2746 
2747 		/*
2748 		 * In AMD parlance chip is really a node while illumos
2749 		 * uses chip as equivalent to socket/package.
2750 		 */
2751 		if (coreidsz == 0) {
2752 			/* Use legacy method */
2753 			for (i = 1; i < cpi->cpi_ncore_per_chip; i <<= 1)
2754 				coreidsz++;
2755 			if (coreidsz == 0)
2756 				coreidsz = 1;
2757 		}
2758 	} else {
2759 		/* Assume single-core part */
2760 		coreidsz = 1;
2761 	}
2762 	cpi->cpi_clogid = cpi->cpi_apicid & ((1 << coreidsz) - 1);
2763 
2764 	/*
2765 	 * The package core ID varies depending on the family. While it may be
2766 	 * tempting to use the CPUID_LEAF_EXT_1e %ebx core id, unfortunately,
2767 	 * this value is the core id in the given node. For non-virtualized
2768 	 * family 17h, we need to take the logical core id and shift off the
2769 	 * threads like we do when getting the core id.  Otherwise, we can use
2770 	 * the clogid as is. When family 17h is virtualized, the clogid should
2771 	 * be sufficient as if we don't have valid data in the leaf, then we
2772 	 * won't think we have SMT, in which case the cpi_clogid should be
2773 	 * sufficient.
2774 	 */
2775 	if (cpi->cpi_family >= 0x17 &&
2776 	    is_x86_feature(x86_featureset, X86FSET_TOPOEXT) &&
2777 	    cpi->cpi_xmaxeax >= CPUID_LEAF_EXT_1e &&
2778 	    cpi->cpi_extd[0x1e].cp_ebx != 0) {
2779 		uint_t nthreads = BITX(cpi->cpi_extd[0x1e].cp_ebx, 15, 8) + 1;
2780 		if (nthreads > 1) {
2781 			VERIFY3U(nthreads, ==, 2);
2782 			cpi->cpi_pkgcoreid = cpi->cpi_clogid >> 1;
2783 		} else {
2784 			cpi->cpi_pkgcoreid = cpi->cpi_clogid;
2785 		}
2786 	} else {
2787 		cpi->cpi_pkgcoreid = cpi->cpi_clogid;
2788 	}
2789 
2790 	/*
2791 	 * Obtain the node ID and compute unit IDs. If we're on family 0x15
2792 	 * (bulldozer) or newer, then we can derive all of this from leaf
2793 	 * CPUID_LEAF_EXT_1e. Otherwise, the method varies by family.
2794 	 */
2795 	if (is_x86_feature(x86_featureset, X86FSET_TOPOEXT) &&
2796 	    cpi->cpi_xmaxeax >= CPUID_LEAF_EXT_1e) {
2797 		cp = &cpi->cpi_extd[0x1e];
2798 
2799 		cpi->cpi_procnodes_per_pkg = BITX(cp->cp_ecx, 10, 8) + 1;
2800 		cpi->cpi_procnodeid = BITX(cp->cp_ecx, 7, 0);
2801 
2802 		/*
2803 		 * For Bulldozer-era CPUs, recalculate the compute unit
2804 		 * information.
2805 		 */
2806 		if (cpi->cpi_family >= 0x15 && cpi->cpi_family < 0x17) {
2807 			cpi->cpi_cores_per_compunit =
2808 			    BITX(cp->cp_ebx, 15, 8) + 1;
2809 			cpi->cpi_compunitid = BITX(cp->cp_ebx, 7, 0) +
2810 			    (cpi->cpi_ncore_per_chip /
2811 			    cpi->cpi_cores_per_compunit) *
2812 			    (cpi->cpi_procnodeid /
2813 			    cpi->cpi_procnodes_per_pkg);
2814 		}
2815 	} else if (cpi->cpi_family == 0xf || cpi->cpi_family >= 0x11) {
2816 		cpi->cpi_procnodeid = (cpi->cpi_apicid >> coreidsz) & 7;
2817 	} else if (cpi->cpi_family == 0x10) {
2818 		/*
2819 		 * See if we are a multi-node processor.
2820 		 * All processors in the system have the same number of nodes
2821 		 */
2822 		nb_caps_reg =  pci_getl_func(0, 24, 3, 0xe8);
2823 		if ((cpi->cpi_model < 8) || BITX(nb_caps_reg, 29, 29) == 0) {
2824 			/* Single-node */
2825 			cpi->cpi_procnodeid = BITX(cpi->cpi_apicid, 5,
2826 			    coreidsz);
2827 		} else {
2828 
2829 			/*
2830 			 * Multi-node revision D (2 nodes per package
2831 			 * are supported)
2832 			 */
2833 			cpi->cpi_procnodes_per_pkg = 2;
2834 
2835 			first_half = (cpi->cpi_pkgcoreid <=
2836 			    (cpi->cpi_ncore_per_chip/2 - 1));
2837 
2838 			if (cpi->cpi_apicid == cpi->cpi_pkgcoreid) {
2839 				/* We are BSP */
2840 				cpi->cpi_procnodeid = (first_half ? 0 : 1);
2841 			} else {
2842 
2843 				/* We are AP */
2844 				/* NodeId[2:1] bits to use for reading F3xe8 */
2845 				node2_1 = BITX(cpi->cpi_apicid, 5, 4) << 1;
2846 
2847 				nb_caps_reg =
2848 				    pci_getl_func(0, 24 + node2_1, 3, 0xe8);
2849 
2850 				/*
2851 				 * Check IntNodeNum bit (31:30, but bit 31 is
2852 				 * always 0 on dual-node processors)
2853 				 */
2854 				if (BITX(nb_caps_reg, 30, 30) == 0)
2855 					cpi->cpi_procnodeid = node2_1 +
2856 					    !first_half;
2857 				else
2858 					cpi->cpi_procnodeid = node2_1 +
2859 					    first_half;
2860 			}
2861 		}
2862 	} else {
2863 		cpi->cpi_procnodeid = 0;
2864 	}
2865 
2866 	cpi->cpi_chipid =
2867 	    cpi->cpi_procnodeid / cpi->cpi_procnodes_per_pkg;
2868 
2869 	cpi->cpi_ncore_bits = coreidsz;
2870 	cpi->cpi_nthread_bits = ddi_fls(cpi->cpi_ncpu_per_chip /
2871 	    cpi->cpi_ncore_per_chip);
2872 }
2873 
2874 static void
2875 spec_uarch_flush_noop(void)
2876 {
2877 }
2878 
2879 /*
2880  * When microcode is present that mitigates MDS, this wrmsr will also flush the
2881  * MDS-related micro-architectural state that would normally happen by calling
2882  * x86_md_clear().
2883  */
2884 static void
2885 spec_uarch_flush_msr(void)
2886 {
2887 	wrmsr(MSR_IA32_FLUSH_CMD, IA32_FLUSH_CMD_L1D);
2888 }
2889 
2890 /*
2891  * This function points to a function that will flush certain
2892  * micro-architectural state on the processor. This flush is used to mitigate
2893  * three different classes of Intel CPU vulnerabilities: L1TF, MDS, and RFDS.
2894  * This function can point to one of three functions:
2895  *
2896  * - A noop which is done because we either are vulnerable, but do not have
2897  *   microcode available to help deal with a fix, or because we aren't
2898  *   vulnerable.
2899  *
2900  * - spec_uarch_flush_msr which will issue an L1D flush and if microcode to
2901  *   mitigate MDS is present, also perform the equivalent of the MDS flush;
2902  *   however, it only flushes the MDS related micro-architectural state on the
2903  *   current hyperthread, it does not do anything for the twin.
2904  *
2905  * - x86_md_clear which will flush the MDS related state. This is done when we
2906  *   have a processor that is vulnerable to MDS, but is not vulnerable to L1TF
2907  *   (RDCL_NO is set); or if the CPU is vulnerable to RFDS and indicates VERW
2908  *   can clear it (RFDS_CLEAR is set).
2909  */
2910 void (*spec_uarch_flush)(void) = spec_uarch_flush_noop;
2911 
2912 static void
2913 cpuid_update_md_clear(cpu_t *cpu, uchar_t *featureset)
2914 {
2915 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2916 
2917 	/* Non-Intel doesn't concern us here. */
2918 	if (cpi->cpi_vendor != X86_VENDOR_Intel)
2919 		return;
2920 
2921 	/*
2922 	 * While RDCL_NO indicates that one of the MDS vulnerabilities (MSBDS)
2923 	 * has been fixed in hardware, it doesn't cover everything related to
2924 	 * MDS. Therefore we can only rely on MDS_NO to determine that we don't
2925 	 * need to mitigate this.
2926 	 *
2927 	 * We must ALSO check the case of RFDS_NO and if RFDS_CLEAR is set,
2928 	 * because of the small cases of RFDS.
2929 	 */
2930 
2931 	if ((!is_x86_feature(featureset, X86FSET_MDS_NO) &&
2932 	    is_x86_feature(featureset, X86FSET_MD_CLEAR)) ||
2933 	    (!is_x86_feature(featureset, X86FSET_RFDS_NO) &&
2934 	    is_x86_feature(featureset, X86FSET_RFDS_CLEAR))) {
2935 		const uint8_t nop = NOP_INSTR;
2936 		uint8_t *md = (uint8_t *)x86_md_clear;
2937 
2938 		*md = nop;
2939 	}
2940 
2941 	membar_producer();
2942 }
2943 
2944 static void
2945 cpuid_update_l1d_flush(cpu_t *cpu, uchar_t *featureset)
2946 {
2947 	boolean_t need_l1d, need_mds, need_rfds;
2948 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2949 
2950 	/*
2951 	 * If we're not on Intel or we've mitigated all of RDCL, MDS, and RFDS
2952 	 * in hardware, then there's nothing left for us to do for enabling
2953 	 * the flush. We can also go ahead and say that SMT exclusion is
2954 	 * unnecessary.
2955 	 */
2956 	if (cpi->cpi_vendor != X86_VENDOR_Intel ||
2957 	    (is_x86_feature(featureset, X86FSET_RDCL_NO) &&
2958 	    is_x86_feature(featureset, X86FSET_MDS_NO) &&
2959 	    is_x86_feature(featureset, X86FSET_RFDS_NO))) {
2960 		extern int smt_exclusion;
2961 		smt_exclusion = 0;
2962 		spec_uarch_flush = spec_uarch_flush_noop;
2963 		membar_producer();
2964 		return;
2965 	}
2966 
2967 	/*
2968 	 * The locations where we need to perform an L1D flush are required both
2969 	 * for mitigating L1TF and MDS. When verw support is present in
2970 	 * microcode, then the L1D flush will take care of doing that as well.
2971 	 * However, if we have a system where RDCL_NO is present, but we don't
2972 	 * have MDS_NO, then we need to do a verw (x86_md_clear) and not a full
2973 	 * L1D flush.
2974 	 */
2975 	if (!is_x86_feature(featureset, X86FSET_RDCL_NO) &&
2976 	    is_x86_feature(featureset, X86FSET_FLUSH_CMD) &&
2977 	    !is_x86_feature(featureset, X86FSET_L1D_VM_NO)) {
2978 		need_l1d = B_TRUE;
2979 	} else {
2980 		need_l1d = B_FALSE;
2981 	}
2982 
2983 	if (!is_x86_feature(featureset, X86FSET_MDS_NO) &&
2984 	    is_x86_feature(featureset, X86FSET_MD_CLEAR)) {
2985 		need_mds = B_TRUE;
2986 	} else {
2987 		need_mds = B_FALSE;
2988 	}
2989 
2990 	if (!is_x86_feature(featureset, X86FSET_RFDS_NO) &&
2991 	    is_x86_feature(featureset, X86FSET_RFDS_CLEAR)) {
2992 		need_rfds = B_TRUE;
2993 	} else {
2994 		need_rfds = B_FALSE;
2995 	}
2996 
2997 	if (need_l1d) {
2998 		/*
2999 		 * As of Feb, 2024, no CPU needs L1D *and* RFDS mitigation
3000 		 * together. If the following VERIFY trips, we need to add
3001 		 * further fixes here.
3002 		 */
3003 		VERIFY(!need_rfds);
3004 		spec_uarch_flush = spec_uarch_flush_msr;
3005 	} else if (need_mds || need_rfds) {
3006 		spec_uarch_flush = x86_md_clear;
3007 	} else {
3008 		/*
3009 		 * We have no hardware mitigations available to us.
3010 		 */
3011 		spec_uarch_flush = spec_uarch_flush_noop;
3012 	}
3013 	membar_producer();
3014 }
3015 
3016 /*
3017  * Branch History Injection (BHI) mitigations.
3018  *
3019  * Intel has provided a software sequence that will scrub the BHB. Like RSB
3020  * (below) we can scribble a return at the beginning to avoid if if the CPU
3021  * is modern enough. We can also scribble a return if the CPU is old enough
3022  * to not have an RSB (pre-eIBRS).
3023  */
3024 typedef enum {
3025 	X86_BHI_TOO_OLD_OR_DISABLED,	/* Pre-eIBRS or disabled */
3026 	X86_BHI_NEW_ENOUGH,		/* AMD, or Intel with BHI_NO set */
3027 	X86_BHI_DIS_S,			/* BHI_NO == 0, but BHI_DIS_S avail. */
3028 	/* NOTE: BHI_DIS_S above will still need the software sequence. */
3029 	X86_BHI_SOFTWARE_SEQUENCE,	/* Use software sequence */
3030 } x86_native_bhi_mitigation_t;
3031 
3032 x86_native_bhi_mitigation_t x86_bhi_mitigation = X86_BHI_SOFTWARE_SEQUENCE;
3033 
3034 static void
3035 cpuid_enable_bhi_dis_s(void)
3036 {
3037 	uint64_t val;
3038 
3039 	val = rdmsr(MSR_IA32_SPEC_CTRL);
3040 	val |= IA32_SPEC_CTRL_BHI_DIS_S;
3041 	wrmsr(MSR_IA32_SPEC_CTRL, val);
3042 }
3043 
3044 /*
3045  * This function scribbles RET into the first instruction of x86_bhb_clear()
3046  * if SPECTREV2 mitigations are disabled, the CPU is too old, the CPU is new
3047  * enough to fix (which includes non-Intel CPUs), or the CPU has an explicit
3048  * disable-Branch-History control.
3049  */
3050 static x86_native_bhi_mitigation_t
3051 cpuid_learn_and_patch_bhi(x86_spectrev2_mitigation_t v2mit, cpu_t *cpu,
3052     uchar_t *featureset)
3053 {
3054 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
3055 	const uint8_t ret = RET_INSTR;
3056 	uint8_t *bhb_clear = (uint8_t *)x86_bhb_clear;
3057 
3058 	ASSERT0(cpu->cpu_id);
3059 
3060 	/* First check for explicitly disabled... */
3061 	if (v2mit == X86_SPECTREV2_DISABLED) {
3062 		*bhb_clear = ret;
3063 		return (X86_BHI_TOO_OLD_OR_DISABLED);
3064 	}
3065 
3066 	/*
3067 	 * Then check for BHI_NO, which means the CPU doesn't have this bug,
3068 	 * or if it's non-Intel, in which case this mitigation mechanism
3069 	 * doesn't apply.
3070 	 */
3071 	if (cpi->cpi_vendor != X86_VENDOR_Intel ||
3072 	    is_x86_feature(featureset, X86FSET_BHI_NO)) {
3073 		*bhb_clear = ret;
3074 		return (X86_BHI_NEW_ENOUGH);
3075 	}
3076 
3077 	/*
3078 	 * Now check for the BHI_CTRL MSR, and then set it if available.
3079 	 * We will still need to use the software sequence, however.
3080 	 */
3081 	if (is_x86_feature(featureset, X86FSET_BHI_CTRL)) {
3082 		cpuid_enable_bhi_dis_s();
3083 		return (X86_BHI_DIS_S);
3084 	}
3085 
3086 	/*
3087 	 * Finally, check if we are too old to bother with RSB:
3088 	 */
3089 	if (v2mit == X86_SPECTREV2_RETPOLINE) {
3090 		*bhb_clear = ret;
3091 		return (X86_BHI_TOO_OLD_OR_DISABLED);
3092 	}
3093 
3094 	ASSERT(*bhb_clear != ret);
3095 	return (X86_BHI_SOFTWARE_SEQUENCE);
3096 }
3097 
3098 /*
3099  * We default to enabling Return Stack Buffer (RSB) mitigations.
3100  *
3101  * We used to skip RSB mitigations with Intel eIBRS, but developments around
3102  * post-barrier RSB (PBRSB) guessing suggests we should enable Intel RSB
3103  * mitigations always unless explicitly bypassed, or unless hardware indicates
3104  * the bug has been fixed.
3105  *
3106  * The current decisions for using, or ignoring, a RSB software stuffing
3107  * sequence are expressed by the following table:
3108  *
3109  * +-------+------------+-----------------+--------+
3110  * | eIBRS |  PBRSB_NO  |  context switch | vmexit |
3111  * +-------+------------+-----------------+--------+
3112  * |   Yes |     No     |  stuff          | stuff  |
3113  * |   Yes |     Yes    |  ignore         | ignore |
3114  * |   No  |     No     |  stuff          | ignore |
3115  * +-------+------------+-----------------+--------+
3116  *
3117  * Note that if an Intel CPU has no eIBRS, it will never enumerate PBRSB_NO,
3118  * because machines with no eIBRS do not have a problem with PBRSB overflow.
3119  * See the Intel document cited below for details.
3120  *
3121  * Also note that AMD AUTO_IBRS has no PBRSB problem, so it is not included in
3122  * the table above, and that there is no situation where vmexit stuffing is
3123  * needed, but context-switch stuffing isn't.
3124  */
3125 
3126 /* BEGIN CSTYLED */
3127 /*
3128  * https://www.intel.com/content/www/us/en/developer/articles/technical/software-security-guidance/advisory-guidance/post-barrier-return-stack-buffer-predictions.html
3129  */
3130 /* END CSTYLED */
3131 
3132 /*
3133  * AMD indicates that when Automatic IBRS is enabled we do not need to implement
3134  * return stack buffer clearing for VMEXIT as it takes care of it. The manual
3135  * also states that as long as SMEP and we maintain at least one page between
3136  * the kernel and user space (we have much more of a red zone), then we do not
3137  * need to clear the RSB. We constrain this to only when Automatic IRBS is
3138  * present.
3139  */
3140 static void
3141 cpuid_patch_rsb(x86_spectrev2_mitigation_t mit, bool intel_pbrsb_no)
3142 {
3143 	const uint8_t ret = RET_INSTR;
3144 	uint8_t *stuff = (uint8_t *)x86_rsb_stuff;
3145 	uint8_t *vmx_stuff = (uint8_t *)x86_rsb_stuff_vmexit;
3146 
3147 	switch (mit) {
3148 	case X86_SPECTREV2_AUTO_IBRS:
3149 	case X86_SPECTREV2_DISABLED:
3150 		/* Don't bother with any RSB stuffing! */
3151 		*stuff = ret;
3152 		*vmx_stuff = ret;
3153 		break;
3154 	case X86_SPECTREV2_RETPOLINE:
3155 		/*
3156 		 * The Intel document on Post-Barrier RSB says that processors
3157 		 * without eIBRS do not have PBRSB problems upon VMEXIT.
3158 		 */
3159 		VERIFY(!intel_pbrsb_no);
3160 		VERIFY3U(*stuff, !=, ret);
3161 		*vmx_stuff = ret;
3162 		break;
3163 	default:
3164 		/*
3165 		 * eIBRS is all that's left.  If CPU claims PBRSB is fixed,
3166 		 * don't use the RSB mitigation in either case.  Otherwise
3167 		 * both vmexit and context-switching require the software
3168 		 * mitigation.
3169 		 */
3170 		if (intel_pbrsb_no) {
3171 			/* CPU claims PBRSB problems are fixed. */
3172 			*stuff = ret;
3173 			*vmx_stuff = ret;
3174 		}
3175 		VERIFY3U(*stuff, ==, *vmx_stuff);
3176 		break;
3177 	}
3178 }
3179 
3180 static void
3181 cpuid_patch_retpolines(x86_spectrev2_mitigation_t mit)
3182 {
3183 	const char *thunks[] = { "_rax", "_rbx", "_rcx", "_rdx", "_rdi",
3184 	    "_rsi", "_rbp", "_r8", "_r9", "_r10", "_r11", "_r12", "_r13",
3185 	    "_r14", "_r15" };
3186 	const uint_t nthunks = ARRAY_SIZE(thunks);
3187 	const char *type;
3188 	uint_t i;
3189 
3190 	if (mit == x86_spectrev2_mitigation)
3191 		return;
3192 
3193 	switch (mit) {
3194 	case X86_SPECTREV2_RETPOLINE:
3195 		type = "gen";
3196 		break;
3197 	case X86_SPECTREV2_AUTO_IBRS:
3198 	case X86_SPECTREV2_ENHANCED_IBRS:
3199 	case X86_SPECTREV2_DISABLED:
3200 		type = "jmp";
3201 		break;
3202 	default:
3203 		panic("asked to update retpoline state with unknown state!");
3204 	}
3205 
3206 	for (i = 0; i < nthunks; i++) {
3207 		uintptr_t source, dest;
3208 		int ssize, dsize;
3209 		char sourcebuf[64], destbuf[64];
3210 
3211 		(void) snprintf(destbuf, sizeof (destbuf),
3212 		    "__x86_indirect_thunk%s", thunks[i]);
3213 		(void) snprintf(sourcebuf, sizeof (sourcebuf),
3214 		    "__x86_indirect_thunk_%s%s", type, thunks[i]);
3215 
3216 		source = kobj_getelfsym(sourcebuf, NULL, &ssize);
3217 		dest = kobj_getelfsym(destbuf, NULL, &dsize);
3218 		VERIFY3U(source, !=, 0);
3219 		VERIFY3U(dest, !=, 0);
3220 		VERIFY3S(dsize, >=, ssize);
3221 		bcopy((void *)source, (void *)dest, ssize);
3222 	}
3223 }
3224 
3225 static void
3226 cpuid_enable_enhanced_ibrs(void)
3227 {
3228 	uint64_t val;
3229 
3230 	val = rdmsr(MSR_IA32_SPEC_CTRL);
3231 	val |= IA32_SPEC_CTRL_IBRS;
3232 	wrmsr(MSR_IA32_SPEC_CTRL, val);
3233 }
3234 
3235 static void
3236 cpuid_enable_auto_ibrs(void)
3237 {
3238 	uint64_t val;
3239 
3240 	val = rdmsr(MSR_AMD_EFER);
3241 	val |= AMD_EFER_AIBRSE;
3242 	wrmsr(MSR_AMD_EFER, val);
3243 }
3244 
3245 /*
3246  * Determine how we should mitigate TAA or if we need to. Regardless of TAA, if
3247  * we can disable TSX, we do so.
3248  *
3249  * This determination is done only on the boot CPU, potentially after loading
3250  * updated microcode.
3251  */
3252 static void
3253 cpuid_update_tsx(cpu_t *cpu, uchar_t *featureset)
3254 {
3255 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
3256 
3257 	VERIFY(cpu->cpu_id == 0);
3258 
3259 	if (cpi->cpi_vendor != X86_VENDOR_Intel) {
3260 		x86_taa_mitigation = X86_TAA_HW_MITIGATED;
3261 		return;
3262 	}
3263 
3264 	if (x86_disable_taa) {
3265 		x86_taa_mitigation = X86_TAA_DISABLED;
3266 		return;
3267 	}
3268 
3269 	/*
3270 	 * If we do not have the ability to disable TSX, then our only
3271 	 * mitigation options are in hardware (TAA_NO), or by using our existing
3272 	 * MDS mitigation as described above.  The latter relies upon us having
3273 	 * configured MDS mitigations correctly! This includes disabling SMT if
3274 	 * we want to cross-CPU-thread protection.
3275 	 */
3276 	if (!is_x86_feature(featureset, X86FSET_TSX_CTRL)) {
3277 		/*
3278 		 * It's not clear whether any parts will enumerate TAA_NO
3279 		 * *without* TSX_CTRL, but let's mark it as such if we see this.
3280 		 */
3281 		if (is_x86_feature(featureset, X86FSET_TAA_NO)) {
3282 			x86_taa_mitigation = X86_TAA_HW_MITIGATED;
3283 			return;
3284 		}
3285 
3286 		if (is_x86_feature(featureset, X86FSET_MD_CLEAR) &&
3287 		    !is_x86_feature(featureset, X86FSET_MDS_NO)) {
3288 			x86_taa_mitigation = X86_TAA_MD_CLEAR;
3289 		} else {
3290 			x86_taa_mitigation = X86_TAA_NOTHING;
3291 		}
3292 		return;
3293 	}
3294 
3295 	/*
3296 	 * We have TSX_CTRL, but we can only fully disable TSX if we're early
3297 	 * enough in boot.
3298 	 *
3299 	 * Otherwise, we'll fall back to causing transactions to abort as our
3300 	 * mitigation. TSX-using code will always take the fallback path.
3301 	 */
3302 	if (cpi->cpi_pass < 4) {
3303 		x86_taa_mitigation = X86_TAA_TSX_DISABLE;
3304 	} else {
3305 		x86_taa_mitigation = X86_TAA_TSX_FORCE_ABORT;
3306 	}
3307 }
3308 
3309 /*
3310  * As mentioned, we should only touch the MSR when we've got a suitable
3311  * microcode loaded on this CPU.
3312  */
3313 static void
3314 cpuid_apply_tsx(x86_taa_mitigation_t taa, uchar_t *featureset)
3315 {
3316 	uint64_t val;
3317 
3318 	switch (taa) {
3319 	case X86_TAA_TSX_DISABLE:
3320 		if (!is_x86_feature(featureset, X86FSET_TSX_CTRL))
3321 			return;
3322 		val = rdmsr(MSR_IA32_TSX_CTRL);
3323 		val |= IA32_TSX_CTRL_CPUID_CLEAR | IA32_TSX_CTRL_RTM_DISABLE;
3324 		wrmsr(MSR_IA32_TSX_CTRL, val);
3325 		break;
3326 	case X86_TAA_TSX_FORCE_ABORT:
3327 		if (!is_x86_feature(featureset, X86FSET_TSX_CTRL))
3328 			return;
3329 		val = rdmsr(MSR_IA32_TSX_CTRL);
3330 		val |= IA32_TSX_CTRL_RTM_DISABLE;
3331 		wrmsr(MSR_IA32_TSX_CTRL, val);
3332 		break;
3333 	case X86_TAA_HW_MITIGATED:
3334 	case X86_TAA_MD_CLEAR:
3335 	case X86_TAA_DISABLED:
3336 	case X86_TAA_NOTHING:
3337 		break;
3338 	}
3339 }
3340 
3341 static void
3342 cpuid_scan_security(cpu_t *cpu, uchar_t *featureset)
3343 {
3344 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
3345 	x86_spectrev2_mitigation_t v2mit;
3346 
3347 	if ((cpi->cpi_vendor == X86_VENDOR_AMD ||
3348 	    cpi->cpi_vendor == X86_VENDOR_HYGON) &&
3349 	    cpi->cpi_xmaxeax >= CPUID_LEAF_EXT_8) {
3350 		if (cpi->cpi_extd[8].cp_ebx & CPUID_AMD_EBX_IBPB)
3351 			add_x86_feature(featureset, X86FSET_IBPB);
3352 		if (cpi->cpi_extd[8].cp_ebx & CPUID_AMD_EBX_IBRS)
3353 			add_x86_feature(featureset, X86FSET_IBRS);
3354 		if (cpi->cpi_extd[8].cp_ebx & CPUID_AMD_EBX_STIBP)
3355 			add_x86_feature(featureset, X86FSET_STIBP);
3356 		if (cpi->cpi_extd[8].cp_ebx & CPUID_AMD_EBX_STIBP_ALL)
3357 			add_x86_feature(featureset, X86FSET_STIBP_ALL);
3358 		if (cpi->cpi_extd[8].cp_ebx & CPUID_AMD_EBX_SSBD)
3359 			add_x86_feature(featureset, X86FSET_SSBD);
3360 		if (cpi->cpi_extd[8].cp_ebx & CPUID_AMD_EBX_VIRT_SSBD)
3361 			add_x86_feature(featureset, X86FSET_SSBD_VIRT);
3362 		if (cpi->cpi_extd[8].cp_ebx & CPUID_AMD_EBX_SSB_NO)
3363 			add_x86_feature(featureset, X86FSET_SSB_NO);
3364 
3365 		/*
3366 		 * Rather than Enhanced IBRS, AMD has a different feature that
3367 		 * is a bit in EFER that can be enabled and will basically do
3368 		 * the right thing while executing in the kernel.
3369 		 */
3370 		if (cpi->cpi_vendor == X86_VENDOR_AMD &&
3371 		    (cpi->cpi_extd[8].cp_ebx & CPUID_AMD_EBX_PREFER_IBRS) &&
3372 		    cpi->cpi_xmaxeax >= CPUID_LEAF_EXT_21 &&
3373 		    (cpi->cpi_extd[0x21].cp_eax & CPUID_AMD_8X21_EAX_AIBRS)) {
3374 			add_x86_feature(featureset, X86FSET_AUTO_IBRS);
3375 		}
3376 
3377 	} else if (cpi->cpi_vendor == X86_VENDOR_Intel &&
3378 	    cpi->cpi_maxeax >= 7) {
3379 		struct cpuid_regs *ecp;
3380 		ecp = &cpi->cpi_std[7];
3381 
3382 		if (ecp->cp_edx & CPUID_INTC_EDX_7_0_MD_CLEAR) {
3383 			add_x86_feature(featureset, X86FSET_MD_CLEAR);
3384 		}
3385 
3386 		if (ecp->cp_edx & CPUID_INTC_EDX_7_0_SPEC_CTRL) {
3387 			add_x86_feature(featureset, X86FSET_IBRS);
3388 			add_x86_feature(featureset, X86FSET_IBPB);
3389 		}
3390 
3391 		if (ecp->cp_edx & CPUID_INTC_EDX_7_0_STIBP) {
3392 			add_x86_feature(featureset, X86FSET_STIBP);
3393 		}
3394 
3395 		/*
3396 		 * Some prediction controls are enumerated by subleaf 2 of
3397 		 * leaf 7.
3398 		 */
3399 		if (CPI_FEATURES_7_2_EDX(cpi) & CPUID_INTC_EDX_7_2_BHI_CTRL) {
3400 			add_x86_feature(featureset, X86FSET_BHI_CTRL);
3401 		}
3402 
3403 		/*
3404 		 * Don't read the arch caps MSR on xpv where we lack the
3405 		 * on_trap().
3406 		 */
3407 #ifndef __xpv
3408 		if (ecp->cp_edx & CPUID_INTC_EDX_7_0_ARCH_CAPS) {
3409 			on_trap_data_t otd;
3410 
3411 			/*
3412 			 * Be paranoid and assume we'll get a #GP.
3413 			 */
3414 			if (!on_trap(&otd, OT_DATA_ACCESS)) {
3415 				uint64_t reg;
3416 
3417 				reg = rdmsr(MSR_IA32_ARCH_CAPABILITIES);
3418 				if (reg & IA32_ARCH_CAP_RDCL_NO) {
3419 					add_x86_feature(featureset,
3420 					    X86FSET_RDCL_NO);
3421 				}
3422 				if (reg & IA32_ARCH_CAP_IBRS_ALL) {
3423 					add_x86_feature(featureset,
3424 					    X86FSET_IBRS_ALL);
3425 				}
3426 				if (reg & IA32_ARCH_CAP_RSBA) {
3427 					add_x86_feature(featureset,
3428 					    X86FSET_RSBA);
3429 				}
3430 				if (reg & IA32_ARCH_CAP_SKIP_L1DFL_VMENTRY) {
3431 					add_x86_feature(featureset,
3432 					    X86FSET_L1D_VM_NO);
3433 				}
3434 				if (reg & IA32_ARCH_CAP_SSB_NO) {
3435 					add_x86_feature(featureset,
3436 					    X86FSET_SSB_NO);
3437 				}
3438 				if (reg & IA32_ARCH_CAP_MDS_NO) {
3439 					add_x86_feature(featureset,
3440 					    X86FSET_MDS_NO);
3441 				}
3442 				if (reg & IA32_ARCH_CAP_TSX_CTRL) {
3443 					add_x86_feature(featureset,
3444 					    X86FSET_TSX_CTRL);
3445 				}
3446 				if (reg & IA32_ARCH_CAP_TAA_NO) {
3447 					add_x86_feature(featureset,
3448 					    X86FSET_TAA_NO);
3449 				}
3450 				if (reg & IA32_ARCH_CAP_RFDS_NO) {
3451 					add_x86_feature(featureset,
3452 					    X86FSET_RFDS_NO);
3453 				}
3454 				if (reg & IA32_ARCH_CAP_RFDS_CLEAR) {
3455 					add_x86_feature(featureset,
3456 					    X86FSET_RFDS_CLEAR);
3457 				}
3458 				if (reg & IA32_ARCH_CAP_PBRSB_NO) {
3459 					add_x86_feature(featureset,
3460 					    X86FSET_PBRSB_NO);
3461 				}
3462 				if (reg & IA32_ARCH_CAP_BHI_NO) {
3463 					add_x86_feature(featureset,
3464 					    X86FSET_BHI_NO);
3465 				}
3466 			}
3467 			no_trap();
3468 		}
3469 #endif	/* !__xpv */
3470 
3471 		if (ecp->cp_edx & CPUID_INTC_EDX_7_0_SSBD)
3472 			add_x86_feature(featureset, X86FSET_SSBD);
3473 
3474 		if (ecp->cp_edx & CPUID_INTC_EDX_7_0_FLUSH_CMD)
3475 			add_x86_feature(featureset, X86FSET_FLUSH_CMD);
3476 	}
3477 
3478 	/*
3479 	 * Take care of certain mitigations on the non-boot CPU. The boot CPU
3480 	 * will have already run this function and determined what we need to
3481 	 * do. This gives us a hook for per-HW thread mitigations such as
3482 	 * enhanced IBRS, or disabling TSX.
3483 	 */
3484 	if (cpu->cpu_id != 0) {
3485 		switch (x86_spectrev2_mitigation) {
3486 		case X86_SPECTREV2_ENHANCED_IBRS:
3487 			cpuid_enable_enhanced_ibrs();
3488 			break;
3489 		case X86_SPECTREV2_AUTO_IBRS:
3490 			cpuid_enable_auto_ibrs();
3491 			break;
3492 		default:
3493 			break;
3494 		}
3495 
3496 		/* If we're committed to BHI_DIS_S, set it for this core. */
3497 		if (x86_bhi_mitigation == X86_BHI_DIS_S)
3498 			cpuid_enable_bhi_dis_s();
3499 
3500 		cpuid_apply_tsx(x86_taa_mitigation, featureset);
3501 		return;
3502 	}
3503 
3504 	/*
3505 	 * Go through and initialize various security mechanisms that we should
3506 	 * only do on a single CPU. This includes Spectre V2, L1TF, MDS, and
3507 	 * TAA.
3508 	 */
3509 
3510 	/*
3511 	 * By default we've come in with retpolines enabled. Check whether we
3512 	 * should disable them or enable enhanced or automatic IBRS.
3513 	 *
3514 	 * Note, we do not allow the use of AMD optimized retpolines as it was
3515 	 * disclosed by AMD in March 2022 that they were still
3516 	 * vulnerable. Prior to that point, we used them.
3517 	 */
3518 	if (x86_disable_spectrev2 != 0) {
3519 		v2mit = X86_SPECTREV2_DISABLED;
3520 	} else if (is_x86_feature(featureset, X86FSET_AUTO_IBRS)) {
3521 		cpuid_enable_auto_ibrs();
3522 		v2mit = X86_SPECTREV2_AUTO_IBRS;
3523 	} else if (is_x86_feature(featureset, X86FSET_IBRS_ALL)) {
3524 		cpuid_enable_enhanced_ibrs();
3525 		v2mit = X86_SPECTREV2_ENHANCED_IBRS;
3526 	} else {
3527 		v2mit = X86_SPECTREV2_RETPOLINE;
3528 	}
3529 
3530 	cpuid_patch_retpolines(v2mit);
3531 	cpuid_patch_rsb(v2mit, is_x86_feature(featureset, X86FSET_PBRSB_NO));
3532 	x86_bhi_mitigation = cpuid_learn_and_patch_bhi(v2mit, cpu, featureset);
3533 	x86_spectrev2_mitigation = v2mit;
3534 	membar_producer();
3535 
3536 	/*
3537 	 * We need to determine what changes are required for mitigating L1TF
3538 	 * and MDS. If the CPU suffers from either of them, then SMT exclusion
3539 	 * is required.
3540 	 *
3541 	 * If any of these are present, then we need to flush u-arch state at
3542 	 * various points. For MDS, we need to do so whenever we change to a
3543 	 * lesser privilege level or we are halting the CPU. For L1TF we need to
3544 	 * flush the L1D cache at VM entry. When we have microcode that handles
3545 	 * MDS, the L1D flush also clears the other u-arch state that the
3546 	 * md_clear does.
3547 	 */
3548 
3549 	/*
3550 	 * Update whether or not we need to be taking explicit action against
3551 	 * MDS or RFDS.
3552 	 */
3553 	cpuid_update_md_clear(cpu, featureset);
3554 
3555 	/*
3556 	 * Determine whether SMT exclusion is required and whether or not we
3557 	 * need to perform an l1d flush.
3558 	 */
3559 	cpuid_update_l1d_flush(cpu, featureset);
3560 
3561 	/*
3562 	 * Determine what our mitigation strategy should be for TAA and then
3563 	 * also apply TAA mitigations.
3564 	 */
3565 	cpuid_update_tsx(cpu, featureset);
3566 	cpuid_apply_tsx(x86_taa_mitigation, featureset);
3567 }
3568 
3569 /*
3570  * Setup XFeature_Enabled_Mask register. Required by xsave feature.
3571  */
3572 void
3573 setup_xfem(void)
3574 {
3575 	uint64_t flags = XFEATURE_LEGACY_FP;
3576 
3577 	ASSERT(is_x86_feature(x86_featureset, X86FSET_XSAVE));
3578 
3579 	if (is_x86_feature(x86_featureset, X86FSET_SSE))
3580 		flags |= XFEATURE_SSE;
3581 
3582 	if (is_x86_feature(x86_featureset, X86FSET_AVX))
3583 		flags |= XFEATURE_AVX;
3584 
3585 	if (is_x86_feature(x86_featureset, X86FSET_AVX512F))
3586 		flags |= XFEATURE_AVX512;
3587 
3588 	set_xcr(XFEATURE_ENABLED_MASK, flags);
3589 
3590 	xsave_bv_all = flags;
3591 }
3592 
3593 static void
3594 cpuid_basic_topology(cpu_t *cpu, uchar_t *featureset)
3595 {
3596 	struct cpuid_info *cpi;
3597 
3598 	cpi = cpu->cpu_m.mcpu_cpi;
3599 
3600 	if (cpi->cpi_vendor == X86_VENDOR_AMD ||
3601 	    cpi->cpi_vendor == X86_VENDOR_HYGON) {
3602 		cpuid_gather_amd_topology_leaves(cpu);
3603 	}
3604 
3605 	cpi->cpi_apicid = cpuid_gather_apicid(cpi);
3606 
3607 	/*
3608 	 * Before we can calculate the IDs that we should assign to this
3609 	 * processor, we need to understand how many cores and threads it has.
3610 	 */
3611 	switch (cpi->cpi_vendor) {
3612 	case X86_VENDOR_Intel:
3613 		cpuid_intel_ncores(cpi, &cpi->cpi_ncpu_per_chip,
3614 		    &cpi->cpi_ncore_per_chip);
3615 		break;
3616 	case X86_VENDOR_AMD:
3617 	case X86_VENDOR_HYGON:
3618 		cpuid_amd_ncores(cpi, &cpi->cpi_ncpu_per_chip,
3619 		    &cpi->cpi_ncore_per_chip);
3620 		break;
3621 	default:
3622 		/*
3623 		 * If we have some other x86 compatible chip, it's not clear how
3624 		 * they would behave. The most common case is virtualization
3625 		 * today, though there are also 64-bit VIA chips. Assume that
3626 		 * all we can get is the basic Leaf 1 HTT information.
3627 		 */
3628 		if ((cpi->cpi_std[1].cp_edx & CPUID_INTC_EDX_HTT) != 0) {
3629 			cpi->cpi_ncore_per_chip = 1;
3630 			cpi->cpi_ncpu_per_chip = CPI_CPU_COUNT(cpi);
3631 		}
3632 		break;
3633 	}
3634 
3635 	/*
3636 	 * Based on the calculated number of threads and cores, potentially
3637 	 * assign the HTT and CMT features.
3638 	 */
3639 	if (cpi->cpi_ncore_per_chip > 1) {
3640 		add_x86_feature(featureset, X86FSET_CMP);
3641 	}
3642 
3643 	if (cpi->cpi_ncpu_per_chip > 1 &&
3644 	    cpi->cpi_ncpu_per_chip != cpi->cpi_ncore_per_chip) {
3645 		add_x86_feature(featureset, X86FSET_HTT);
3646 	}
3647 
3648 	/*
3649 	 * Now that has been set up, we need to go through and calculate all of
3650 	 * the rest of the parameters that exist. If we think the CPU doesn't
3651 	 * have either SMT (HTT) or CMP, then we basically go through and fake
3652 	 * up information in some way. The most likely case for this is
3653 	 * virtualization where we have a lot of partial topology information.
3654 	 */
3655 	if (!is_x86_feature(featureset, X86FSET_HTT) &&
3656 	    !is_x86_feature(featureset, X86FSET_CMP)) {
3657 		/*
3658 		 * This is a single core, single-threaded processor.
3659 		 */
3660 		cpi->cpi_procnodes_per_pkg = 1;
3661 		cpi->cpi_cores_per_compunit = 1;
3662 		cpi->cpi_compunitid = 0;
3663 		cpi->cpi_chipid = -1;
3664 		cpi->cpi_clogid = 0;
3665 		cpi->cpi_coreid = cpu->cpu_id;
3666 		cpi->cpi_pkgcoreid = 0;
3667 		if (cpi->cpi_vendor == X86_VENDOR_AMD ||
3668 		    cpi->cpi_vendor == X86_VENDOR_HYGON) {
3669 			cpi->cpi_procnodeid = BITX(cpi->cpi_apicid, 3, 0);
3670 		} else {
3671 			cpi->cpi_procnodeid = cpi->cpi_chipid;
3672 		}
3673 	} else {
3674 		switch (cpi->cpi_vendor) {
3675 		case X86_VENDOR_Intel:
3676 			cpuid_intel_getids(cpu, featureset);
3677 			break;
3678 		case X86_VENDOR_AMD:
3679 		case X86_VENDOR_HYGON:
3680 			cpuid_amd_getids(cpu, featureset);
3681 			break;
3682 		default:
3683 			/*
3684 			 * In this case, it's hard to say what we should do.
3685 			 * We're going to model them to the OS as single core
3686 			 * threads. We don't have a good identifier for them, so
3687 			 * we're just going to use the cpu id all on a single
3688 			 * chip.
3689 			 *
3690 			 * This case has historically been different from the
3691 			 * case above where we don't have HTT or CMP. While they
3692 			 * could be combined, we've opted to keep it separate to
3693 			 * minimize the risk of topology changes in weird cases.
3694 			 */
3695 			cpi->cpi_procnodes_per_pkg = 1;
3696 			cpi->cpi_cores_per_compunit = 1;
3697 			cpi->cpi_chipid = 0;
3698 			cpi->cpi_coreid = cpu->cpu_id;
3699 			cpi->cpi_clogid = cpu->cpu_id;
3700 			cpi->cpi_pkgcoreid = cpu->cpu_id;
3701 			cpi->cpi_procnodeid = cpi->cpi_chipid;
3702 			cpi->cpi_compunitid = cpi->cpi_coreid;
3703 			break;
3704 		}
3705 	}
3706 }
3707 
3708 /*
3709  * Gather relevant CPU features from leaf 6 which covers thermal information. We
3710  * always gather leaf 6 if it's supported; however, we only look for features on
3711  * Intel systems as AMD does not currently define any of the features we look
3712  * for below.
3713  */
3714 static void
3715 cpuid_basic_thermal(cpu_t *cpu, uchar_t *featureset)
3716 {
3717 	struct cpuid_regs *cp;
3718 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
3719 
3720 	if (cpi->cpi_maxeax < 6) {
3721 		return;
3722 	}
3723 
3724 	cp = &cpi->cpi_std[6];
3725 	cp->cp_eax = 6;
3726 	cp->cp_ebx = cp->cp_ecx = cp->cp_edx = 0;
3727 	(void) __cpuid_insn(cp);
3728 	platform_cpuid_mangle(cpi->cpi_vendor, 6, cp);
3729 
3730 	if (cpi->cpi_vendor != X86_VENDOR_Intel) {
3731 		return;
3732 	}
3733 
3734 	if ((cp->cp_eax & CPUID_INTC_EAX_DTS) != 0) {
3735 		add_x86_feature(featureset, X86FSET_CORE_THERMAL);
3736 	}
3737 
3738 	if ((cp->cp_eax & CPUID_INTC_EAX_PTM) != 0) {
3739 		add_x86_feature(featureset, X86FSET_PKG_THERMAL);
3740 	}
3741 }
3742 
3743 /*
3744  * This is used when we discover that we have AVX support in cpuid. This
3745  * proceeds to scan for the rest of the AVX derived features.
3746  */
3747 static void
3748 cpuid_basic_avx(cpu_t *cpu, uchar_t *featureset)
3749 {
3750 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
3751 
3752 	/*
3753 	 * If we don't have AVX, don't bother with most of this.
3754 	 */
3755 	if ((cpi->cpi_std[1].cp_ecx & CPUID_INTC_ECX_AVX) == 0)
3756 		return;
3757 
3758 	add_x86_feature(featureset, X86FSET_AVX);
3759 
3760 	/*
3761 	 * Intel says we can't check these without also
3762 	 * checking AVX.
3763 	 */
3764 	if (cpi->cpi_std[1].cp_ecx & CPUID_INTC_ECX_F16C)
3765 		add_x86_feature(featureset, X86FSET_F16C);
3766 
3767 	if (cpi->cpi_std[1].cp_ecx & CPUID_INTC_ECX_FMA)
3768 		add_x86_feature(featureset, X86FSET_FMA);
3769 
3770 	if (cpi->cpi_std[7].cp_ebx & CPUID_INTC_EBX_7_0_BMI1)
3771 		add_x86_feature(featureset, X86FSET_BMI1);
3772 
3773 	if (cpi->cpi_std[7].cp_ebx & CPUID_INTC_EBX_7_0_BMI2)
3774 		add_x86_feature(featureset, X86FSET_BMI2);
3775 
3776 	if (cpi->cpi_std[7].cp_ebx & CPUID_INTC_EBX_7_0_AVX2)
3777 		add_x86_feature(featureset, X86FSET_AVX2);
3778 
3779 	if (cpi->cpi_std[7].cp_ecx & CPUID_INTC_ECX_7_0_VAES)
3780 		add_x86_feature(featureset, X86FSET_VAES);
3781 
3782 	if (cpi->cpi_std[7].cp_ecx & CPUID_INTC_ECX_7_0_VPCLMULQDQ)
3783 		add_x86_feature(featureset, X86FSET_VPCLMULQDQ);
3784 
3785 	/*
3786 	 * The rest of the AVX features require AVX512. Do not check them unless
3787 	 * it is present.
3788 	 */
3789 	if ((cpi->cpi_std[7].cp_ebx & CPUID_INTC_EBX_7_0_AVX512F) == 0)
3790 		return;
3791 	add_x86_feature(featureset, X86FSET_AVX512F);
3792 
3793 	if (cpi->cpi_std[7].cp_ebx & CPUID_INTC_EBX_7_0_AVX512DQ)
3794 		add_x86_feature(featureset, X86FSET_AVX512DQ);
3795 
3796 	if (cpi->cpi_std[7].cp_ebx & CPUID_INTC_EBX_7_0_AVX512IFMA)
3797 		add_x86_feature(featureset, X86FSET_AVX512FMA);
3798 
3799 	if (cpi->cpi_std[7].cp_ebx & CPUID_INTC_EBX_7_0_AVX512PF)
3800 		add_x86_feature(featureset, X86FSET_AVX512PF);
3801 
3802 	if (cpi->cpi_std[7].cp_ebx & CPUID_INTC_EBX_7_0_AVX512ER)
3803 		add_x86_feature(featureset, X86FSET_AVX512ER);
3804 
3805 	if (cpi->cpi_std[7].cp_ebx & CPUID_INTC_EBX_7_0_AVX512CD)
3806 		add_x86_feature(featureset, X86FSET_AVX512CD);
3807 
3808 	if (cpi->cpi_std[7].cp_ebx & CPUID_INTC_EBX_7_0_AVX512BW)
3809 		add_x86_feature(featureset, X86FSET_AVX512BW);
3810 
3811 	if (cpi->cpi_std[7].cp_ebx & CPUID_INTC_EBX_7_0_AVX512VL)
3812 		add_x86_feature(featureset, X86FSET_AVX512VL);
3813 
3814 	if (cpi->cpi_std[7].cp_ecx & CPUID_INTC_ECX_7_0_AVX512VBMI)
3815 		add_x86_feature(featureset, X86FSET_AVX512VBMI);
3816 
3817 	if (cpi->cpi_std[7].cp_ecx & CPUID_INTC_ECX_7_0_AVX512VBMI2)
3818 		add_x86_feature(featureset, X86FSET_AVX512_VBMI2);
3819 
3820 	if (cpi->cpi_std[7].cp_ecx & CPUID_INTC_ECX_7_0_AVX512VNNI)
3821 		add_x86_feature(featureset, X86FSET_AVX512VNNI);
3822 
3823 	if (cpi->cpi_std[7].cp_ecx & CPUID_INTC_ECX_7_0_AVX512BITALG)
3824 		add_x86_feature(featureset, X86FSET_AVX512_BITALG);
3825 
3826 	if (cpi->cpi_std[7].cp_ecx & CPUID_INTC_ECX_7_0_AVX512VPOPCDQ)
3827 		add_x86_feature(featureset, X86FSET_AVX512VPOPCDQ);
3828 
3829 	if (cpi->cpi_std[7].cp_edx & CPUID_INTC_EDX_7_0_AVX5124NNIW)
3830 		add_x86_feature(featureset, X86FSET_AVX512NNIW);
3831 
3832 	if (cpi->cpi_std[7].cp_edx & CPUID_INTC_EDX_7_0_AVX5124FMAPS)
3833 		add_x86_feature(featureset, X86FSET_AVX512FMAPS);
3834 
3835 	/*
3836 	 * More features here are in Leaf 7, subleaf 1. Don't bother checking if
3837 	 * we don't need to.
3838 	 */
3839 	if (cpi->cpi_std[7].cp_eax < 1)
3840 		return;
3841 
3842 	if (cpi->cpi_sub7[0].cp_eax & CPUID_INTC_EAX_7_1_AVX512_BF16)
3843 		add_x86_feature(featureset, X86FSET_AVX512_BF16);
3844 }
3845 
3846 /*
3847  * PPIN is the protected processor inventory number. On AMD this is an actual
3848  * feature bit. However, on Intel systems we need to read the platform
3849  * information MSR if we're on a specific model.
3850  */
3851 #if !defined(__xpv)
3852 static void
3853 cpuid_basic_ppin(cpu_t *cpu, uchar_t *featureset)
3854 {
3855 	on_trap_data_t otd;
3856 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
3857 
3858 	switch (cpi->cpi_vendor) {
3859 	case X86_VENDOR_AMD:
3860 		/*
3861 		 * This leaf will have already been gathered in the topology
3862 		 * functions.
3863 		 */
3864 		if (cpi->cpi_xmaxeax >= CPUID_LEAF_EXT_8) {
3865 			if (cpi->cpi_extd[8].cp_ebx & CPUID_AMD_EBX_PPIN) {
3866 				add_x86_feature(featureset, X86FSET_PPIN);
3867 			}
3868 		}
3869 		break;
3870 	case X86_VENDOR_Intel:
3871 		if (cpi->cpi_family != 6)
3872 			break;
3873 		switch (cpi->cpi_model) {
3874 		case INTC_MODEL_IVYBRIDGE_XEON:
3875 		case INTC_MODEL_HASWELL_XEON:
3876 		case INTC_MODEL_BROADWELL_XEON:
3877 		case INTC_MODEL_BROADWELL_XEON_D:
3878 		case INTC_MODEL_SKYLAKE_XEON:
3879 		case INTC_MODEL_ICELAKE_XEON:
3880 			if (!on_trap(&otd, OT_DATA_ACCESS)) {
3881 				uint64_t value;
3882 
3883 				value = rdmsr(MSR_PLATFORM_INFO);
3884 				if ((value & MSR_PLATFORM_INFO_PPIN) != 0) {
3885 					add_x86_feature(featureset,
3886 					    X86FSET_PPIN);
3887 				}
3888 			}
3889 			no_trap();
3890 			break;
3891 		default:
3892 			break;
3893 		}
3894 		break;
3895 	default:
3896 		break;
3897 	}
3898 }
3899 #endif	/* ! __xpv */
3900 
3901 static void
3902 cpuid_pass_prelude(cpu_t *cpu, void *arg)
3903 {
3904 	uchar_t *featureset = (uchar_t *)arg;
3905 
3906 	/*
3907 	 * We don't run on any processor that doesn't have cpuid, and could not
3908 	 * possibly have arrived here.
3909 	 */
3910 	add_x86_feature(featureset, X86FSET_CPUID);
3911 }
3912 
3913 static void
3914 cpuid_pass_ident(cpu_t *cpu, void *arg __unused)
3915 {
3916 	struct cpuid_info *cpi;
3917 	struct cpuid_regs *cp;
3918 
3919 	/*
3920 	 * We require that virtual/native detection be complete and that PCI
3921 	 * config space access has been set up; at present there is no reliable
3922 	 * way to determine the latter.
3923 	 */
3924 #if !defined(__xpv)
3925 	ASSERT3S(platform_type, !=, -1);
3926 #endif	/* !__xpv */
3927 
3928 	cpi = cpu->cpu_m.mcpu_cpi;
3929 	ASSERT(cpi != NULL);
3930 
3931 	cp = &cpi->cpi_std[0];
3932 	cp->cp_eax = 0;
3933 	cpi->cpi_maxeax = __cpuid_insn(cp);
3934 	{
3935 		uint32_t *iptr = (uint32_t *)cpi->cpi_vendorstr;
3936 		*iptr++ = cp->cp_ebx;
3937 		*iptr++ = cp->cp_edx;
3938 		*iptr++ = cp->cp_ecx;
3939 		*(char *)&cpi->cpi_vendorstr[12] = '\0';
3940 	}
3941 
3942 	cpi->cpi_vendor = _cpuid_vendorstr_to_vendorcode(cpi->cpi_vendorstr);
3943 	x86_vendor = cpi->cpi_vendor; /* for compatibility */
3944 
3945 	/*
3946 	 * Limit the range in case of weird hardware
3947 	 */
3948 	if (cpi->cpi_maxeax > CPI_MAXEAX_MAX)
3949 		cpi->cpi_maxeax = CPI_MAXEAX_MAX;
3950 	if (cpi->cpi_maxeax < 1)
3951 		return;
3952 
3953 	cp = &cpi->cpi_std[1];
3954 	cp->cp_eax = 1;
3955 	(void) __cpuid_insn(cp);
3956 
3957 	/*
3958 	 * Extract identifying constants for easy access.
3959 	 */
3960 	cpi->cpi_model = CPI_MODEL(cpi);
3961 	cpi->cpi_family = CPI_FAMILY(cpi);
3962 
3963 	if (cpi->cpi_family == 0xf)
3964 		cpi->cpi_family += CPI_FAMILY_XTD(cpi);
3965 
3966 	/*
3967 	 * Beware: AMD uses "extended model" iff base *FAMILY* == 0xf.
3968 	 * Intel, and presumably everyone else, uses model == 0xf, as
3969 	 * one would expect (max value means possible overflow).  Sigh.
3970 	 */
3971 
3972 	switch (cpi->cpi_vendor) {
3973 	case X86_VENDOR_Intel:
3974 		if (IS_EXTENDED_MODEL_INTEL(cpi))
3975 			cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4;
3976 		break;
3977 	case X86_VENDOR_AMD:
3978 		if (CPI_FAMILY(cpi) == 0xf)
3979 			cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4;
3980 		break;
3981 	case X86_VENDOR_HYGON:
3982 		cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4;
3983 		break;
3984 	default:
3985 		if (cpi->cpi_model == 0xf)
3986 			cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4;
3987 		break;
3988 	}
3989 
3990 	cpi->cpi_step = CPI_STEP(cpi);
3991 	cpi->cpi_brandid = CPI_BRANDID(cpi);
3992 
3993 	/*
3994 	 * Synthesize chip "revision" and socket type
3995 	 */
3996 	cpi->cpi_chiprev = _cpuid_chiprev(cpi->cpi_vendor, cpi->cpi_family,
3997 	    cpi->cpi_model, cpi->cpi_step);
3998 	cpi->cpi_chiprevstr = _cpuid_chiprevstr(cpi->cpi_vendor,
3999 	    cpi->cpi_family, cpi->cpi_model, cpi->cpi_step);
4000 	cpi->cpi_socket = _cpuid_skt(cpi->cpi_vendor, cpi->cpi_family,
4001 	    cpi->cpi_model, cpi->cpi_step);
4002 	cpi->cpi_uarchrev = _cpuid_uarchrev(cpi->cpi_vendor, cpi->cpi_family,
4003 	    cpi->cpi_model, cpi->cpi_step);
4004 }
4005 
4006 static void
4007 cpuid_pass_basic(cpu_t *cpu, void *arg)
4008 {
4009 	uchar_t *featureset = (uchar_t *)arg;
4010 	uint32_t mask_ecx, mask_edx;
4011 	struct cpuid_info *cpi;
4012 	struct cpuid_regs *cp;
4013 	int xcpuid;
4014 #if !defined(__xpv)
4015 	extern int idle_cpu_prefer_mwait;
4016 #endif
4017 
4018 	cpi = cpu->cpu_m.mcpu_cpi;
4019 	ASSERT(cpi != NULL);
4020 
4021 	if (cpi->cpi_maxeax < 1)
4022 		return;
4023 
4024 	/*
4025 	 * This was filled during the identification pass.
4026 	 */
4027 	cp = &cpi->cpi_std[1];
4028 
4029 	/*
4030 	 * *default* assumptions:
4031 	 * - believe %edx feature word
4032 	 * - ignore %ecx feature word
4033 	 * - 32-bit virtual and physical addressing
4034 	 */
4035 	mask_edx = 0xffffffff;
4036 	mask_ecx = 0;
4037 
4038 	cpi->cpi_pabits = cpi->cpi_vabits = 32;
4039 
4040 	switch (cpi->cpi_vendor) {
4041 	case X86_VENDOR_Intel:
4042 		if (cpi->cpi_family == 5)
4043 			x86_type = X86_TYPE_P5;
4044 		else if (IS_LEGACY_P6(cpi)) {
4045 			x86_type = X86_TYPE_P6;
4046 			pentiumpro_bug4046376 = 1;
4047 			/*
4048 			 * Clear the SEP bit when it was set erroneously
4049 			 */
4050 			if (cpi->cpi_model < 3 && cpi->cpi_step < 3)
4051 				cp->cp_edx &= ~CPUID_INTC_EDX_SEP;
4052 		} else if (IS_NEW_F6(cpi) || cpi->cpi_family == 0xf) {
4053 			x86_type = X86_TYPE_P4;
4054 			/*
4055 			 * We don't currently depend on any of the %ecx
4056 			 * features until Prescott, so we'll only check
4057 			 * this from P4 onwards.  We might want to revisit
4058 			 * that idea later.
4059 			 */
4060 			mask_ecx = 0xffffffff;
4061 		} else if (cpi->cpi_family > 0xf)
4062 			mask_ecx = 0xffffffff;
4063 		/*
4064 		 * We don't support MONITOR/MWAIT if leaf 5 is not available
4065 		 * to obtain the monitor linesize.
4066 		 */
4067 		if (cpi->cpi_maxeax < 5)
4068 			mask_ecx &= ~CPUID_INTC_ECX_MON;
4069 		break;
4070 	case X86_VENDOR_IntelClone:
4071 	default:
4072 		break;
4073 	case X86_VENDOR_AMD:
4074 #if defined(OPTERON_ERRATUM_108)
4075 		if (cpi->cpi_family == 0xf && cpi->cpi_model == 0xe) {
4076 			cp->cp_eax = (0xf0f & cp->cp_eax) | 0xc0;
4077 			cpi->cpi_model = 0xc;
4078 		} else
4079 #endif
4080 		if (cpi->cpi_family == 5) {
4081 			/*
4082 			 * AMD K5 and K6
4083 			 *
4084 			 * These CPUs have an incomplete implementation
4085 			 * of MCA/MCE which we mask away.
4086 			 */
4087 			mask_edx &= ~(CPUID_INTC_EDX_MCE | CPUID_INTC_EDX_MCA);
4088 
4089 			/*
4090 			 * Model 0 uses the wrong (APIC) bit
4091 			 * to indicate PGE.  Fix it here.
4092 			 */
4093 			if (cpi->cpi_model == 0) {
4094 				if (cp->cp_edx & 0x200) {
4095 					cp->cp_edx &= ~0x200;
4096 					cp->cp_edx |= CPUID_INTC_EDX_PGE;
4097 				}
4098 			}
4099 
4100 			/*
4101 			 * Early models had problems w/ MMX; disable.
4102 			 */
4103 			if (cpi->cpi_model < 6)
4104 				mask_edx &= ~CPUID_INTC_EDX_MMX;
4105 		}
4106 
4107 		/*
4108 		 * For newer families, SSE3 and CX16, at least, are valid;
4109 		 * enable all
4110 		 */
4111 		if (cpi->cpi_family >= 0xf)
4112 			mask_ecx = 0xffffffff;
4113 		/*
4114 		 * We don't support MONITOR/MWAIT if leaf 5 is not available
4115 		 * to obtain the monitor linesize.
4116 		 */
4117 		if (cpi->cpi_maxeax < 5)
4118 			mask_ecx &= ~CPUID_INTC_ECX_MON;
4119 
4120 #if !defined(__xpv)
4121 		/*
4122 		 * AMD has not historically used MWAIT in the CPU's idle loop.
4123 		 * Pre-family-10h Opterons do not have the MWAIT instruction. We
4124 		 * know for certain that in at least family 17h, per AMD, mwait
4125 		 * is preferred. Families in-between are less certain.
4126 		 */
4127 		if (cpi->cpi_family < 0x17) {
4128 			idle_cpu_prefer_mwait = 0;
4129 		}
4130 #endif
4131 
4132 		break;
4133 	case X86_VENDOR_HYGON:
4134 		/* Enable all for Hygon Dhyana CPU */
4135 		mask_ecx = 0xffffffff;
4136 		break;
4137 	case X86_VENDOR_TM:
4138 		/*
4139 		 * workaround the NT workaround in CMS 4.1
4140 		 */
4141 		if (cpi->cpi_family == 5 && cpi->cpi_model == 4 &&
4142 		    (cpi->cpi_step == 2 || cpi->cpi_step == 3))
4143 			cp->cp_edx |= CPUID_INTC_EDX_CX8;
4144 		break;
4145 	case X86_VENDOR_Centaur:
4146 		/*
4147 		 * workaround the NT workarounds again
4148 		 */
4149 		if (cpi->cpi_family == 6)
4150 			cp->cp_edx |= CPUID_INTC_EDX_CX8;
4151 		break;
4152 	case X86_VENDOR_Cyrix:
4153 		/*
4154 		 * We rely heavily on the probing in locore
4155 		 * to actually figure out what parts, if any,
4156 		 * of the Cyrix cpuid instruction to believe.
4157 		 */
4158 		switch (x86_type) {
4159 		case X86_TYPE_CYRIX_486:
4160 			mask_edx = 0;
4161 			break;
4162 		case X86_TYPE_CYRIX_6x86:
4163 			mask_edx = 0;
4164 			break;
4165 		case X86_TYPE_CYRIX_6x86L:
4166 			mask_edx =
4167 			    CPUID_INTC_EDX_DE |
4168 			    CPUID_INTC_EDX_CX8;
4169 			break;
4170 		case X86_TYPE_CYRIX_6x86MX:
4171 			mask_edx =
4172 			    CPUID_INTC_EDX_DE |
4173 			    CPUID_INTC_EDX_MSR |
4174 			    CPUID_INTC_EDX_CX8 |
4175 			    CPUID_INTC_EDX_PGE |
4176 			    CPUID_INTC_EDX_CMOV |
4177 			    CPUID_INTC_EDX_MMX;
4178 			break;
4179 		case X86_TYPE_CYRIX_GXm:
4180 			mask_edx =
4181 			    CPUID_INTC_EDX_MSR |
4182 			    CPUID_INTC_EDX_CX8 |
4183 			    CPUID_INTC_EDX_CMOV |
4184 			    CPUID_INTC_EDX_MMX;
4185 			break;
4186 		case X86_TYPE_CYRIX_MediaGX:
4187 			break;
4188 		case X86_TYPE_CYRIX_MII:
4189 		case X86_TYPE_VIA_CYRIX_III:
4190 			mask_edx =
4191 			    CPUID_INTC_EDX_DE |
4192 			    CPUID_INTC_EDX_TSC |
4193 			    CPUID_INTC_EDX_MSR |
4194 			    CPUID_INTC_EDX_CX8 |
4195 			    CPUID_INTC_EDX_PGE |
4196 			    CPUID_INTC_EDX_CMOV |
4197 			    CPUID_INTC_EDX_MMX;
4198 			break;
4199 		default:
4200 			break;
4201 		}
4202 		break;
4203 	}
4204 
4205 #if defined(__xpv)
4206 	/*
4207 	 * Do not support MONITOR/MWAIT under a hypervisor
4208 	 */
4209 	mask_ecx &= ~CPUID_INTC_ECX_MON;
4210 	/*
4211 	 * Do not support XSAVE under a hypervisor for now
4212 	 */
4213 	xsave_force_disable = B_TRUE;
4214 
4215 #endif	/* __xpv */
4216 
4217 	if (xsave_force_disable) {
4218 		mask_ecx &= ~CPUID_INTC_ECX_XSAVE;
4219 		mask_ecx &= ~CPUID_INTC_ECX_AVX;
4220 		mask_ecx &= ~CPUID_INTC_ECX_F16C;
4221 		mask_ecx &= ~CPUID_INTC_ECX_FMA;
4222 	}
4223 
4224 	/*
4225 	 * Now we've figured out the masks that determine
4226 	 * which bits we choose to believe, apply the masks
4227 	 * to the feature words, then map the kernel's view
4228 	 * of these feature words into its feature word.
4229 	 */
4230 	cp->cp_edx &= mask_edx;
4231 	cp->cp_ecx &= mask_ecx;
4232 
4233 	/*
4234 	 * apply any platform restrictions (we don't call this
4235 	 * immediately after __cpuid_insn here, because we need the
4236 	 * workarounds applied above first)
4237 	 */
4238 	platform_cpuid_mangle(cpi->cpi_vendor, 1, cp);
4239 
4240 	/*
4241 	 * In addition to ecx and edx, Intel and AMD are storing a bunch of
4242 	 * instruction set extensions in leaf 7's ebx, ecx, and edx. Note, leaf
4243 	 * 7 has sub-leaves determined by ecx.
4244 	 */
4245 	if (cpi->cpi_maxeax >= 7) {
4246 		struct cpuid_regs *ecp;
4247 		ecp = &cpi->cpi_std[7];
4248 		ecp->cp_eax = 7;
4249 		ecp->cp_ecx = 0;
4250 		(void) __cpuid_insn(ecp);
4251 
4252 		/*
4253 		 * If XSAVE has been disabled, just ignore all of the
4254 		 * extended-save-area dependent flags here. By removing most of
4255 		 * the leaf 7, sub-leaf 0 flags, that will ensure tha we don't
4256 		 * end up looking at additional xsave dependent leaves right
4257 		 * now.
4258 		 */
4259 		if (xsave_force_disable) {
4260 			ecp->cp_ebx &= ~CPUID_INTC_EBX_7_0_BMI1;
4261 			ecp->cp_ebx &= ~CPUID_INTC_EBX_7_0_BMI2;
4262 			ecp->cp_ebx &= ~CPUID_INTC_EBX_7_0_AVX2;
4263 			ecp->cp_ebx &= ~CPUID_INTC_EBX_7_0_MPX;
4264 			ecp->cp_ebx &= ~CPUID_INTC_EBX_7_0_ALL_AVX512;
4265 			ecp->cp_ecx &= ~CPUID_INTC_ECX_7_0_ALL_AVX512;
4266 			ecp->cp_edx &= ~CPUID_INTC_EDX_7_0_ALL_AVX512;
4267 			ecp->cp_ecx &= ~CPUID_INTC_ECX_7_0_VAES;
4268 			ecp->cp_ecx &= ~CPUID_INTC_ECX_7_0_VPCLMULQDQ;
4269 			ecp->cp_ecx &= ~CPUID_INTC_ECX_7_0_GFNI;
4270 		}
4271 
4272 		if (ecp->cp_ebx & CPUID_INTC_EBX_7_0_SMEP)
4273 			add_x86_feature(featureset, X86FSET_SMEP);
4274 
4275 		/*
4276 		 * We check disable_smap here in addition to in startup_smap()
4277 		 * to ensure CPUs that aren't the boot CPU don't accidentally
4278 		 * include it in the feature set and thus generate a mismatched
4279 		 * x86 feature set across CPUs.
4280 		 */
4281 		if (ecp->cp_ebx & CPUID_INTC_EBX_7_0_SMAP &&
4282 		    disable_smap == 0)
4283 			add_x86_feature(featureset, X86FSET_SMAP);
4284 
4285 		if (ecp->cp_ebx & CPUID_INTC_EBX_7_0_RDSEED)
4286 			add_x86_feature(featureset, X86FSET_RDSEED);
4287 
4288 		if (ecp->cp_ebx & CPUID_INTC_EBX_7_0_ADX)
4289 			add_x86_feature(featureset, X86FSET_ADX);
4290 
4291 		if (ecp->cp_ebx & CPUID_INTC_EBX_7_0_FSGSBASE)
4292 			add_x86_feature(featureset, X86FSET_FSGSBASE);
4293 
4294 		if (ecp->cp_ebx & CPUID_INTC_EBX_7_0_CLFLUSHOPT)
4295 			add_x86_feature(featureset, X86FSET_CLFLUSHOPT);
4296 
4297 		if (ecp->cp_ebx & CPUID_INTC_EBX_7_0_INVPCID)
4298 			add_x86_feature(featureset, X86FSET_INVPCID);
4299 
4300 		if (ecp->cp_ecx & CPUID_INTC_ECX_7_0_UMIP)
4301 			add_x86_feature(featureset, X86FSET_UMIP);
4302 		if (ecp->cp_ecx & CPUID_INTC_ECX_7_0_PKU)
4303 			add_x86_feature(featureset, X86FSET_PKU);
4304 		if (ecp->cp_ecx & CPUID_INTC_ECX_7_0_OSPKE)
4305 			add_x86_feature(featureset, X86FSET_OSPKE);
4306 		if (ecp->cp_ecx & CPUID_INTC_ECX_7_0_GFNI)
4307 			add_x86_feature(featureset, X86FSET_GFNI);
4308 
4309 		if (ecp->cp_ebx & CPUID_INTC_EBX_7_0_CLWB)
4310 			add_x86_feature(featureset, X86FSET_CLWB);
4311 
4312 		if (cpi->cpi_vendor == X86_VENDOR_Intel) {
4313 			if (ecp->cp_ebx & CPUID_INTC_EBX_7_0_MPX)
4314 				add_x86_feature(featureset, X86FSET_MPX);
4315 		}
4316 
4317 		/*
4318 		 * If we have subleaf 1 or 2 available, grab and store
4319 		 * that. This is used for more AVX and related features.
4320 		 */
4321 		if (ecp->cp_eax >= 1) {
4322 			struct cpuid_regs *c71;
4323 			c71 = &cpi->cpi_sub7[0];
4324 			c71->cp_eax = 7;
4325 			c71->cp_ecx = 1;
4326 			(void) __cpuid_insn(c71);
4327 		}
4328 
4329 		/* Subleaf 2 has certain security indicators in it. */
4330 		if (ecp->cp_eax >= 2) {
4331 			struct cpuid_regs *c72;
4332 			c72 = &cpi->cpi_sub7[1];
4333 			c72->cp_eax = 7;
4334 			c72->cp_ecx = 2;
4335 			(void) __cpuid_insn(c72);
4336 		}
4337 	}
4338 
4339 	/*
4340 	 * fold in overrides from the "eeprom" mechanism
4341 	 */
4342 	cp->cp_edx |= cpuid_feature_edx_include;
4343 	cp->cp_edx &= ~cpuid_feature_edx_exclude;
4344 
4345 	cp->cp_ecx |= cpuid_feature_ecx_include;
4346 	cp->cp_ecx &= ~cpuid_feature_ecx_exclude;
4347 
4348 	if (cp->cp_edx & CPUID_INTC_EDX_PSE) {
4349 		add_x86_feature(featureset, X86FSET_LARGEPAGE);
4350 	}
4351 	if (cp->cp_edx & CPUID_INTC_EDX_TSC) {
4352 		add_x86_feature(featureset, X86FSET_TSC);
4353 	}
4354 	if (cp->cp_edx & CPUID_INTC_EDX_MSR) {
4355 		add_x86_feature(featureset, X86FSET_MSR);
4356 	}
4357 	if (cp->cp_edx & CPUID_INTC_EDX_MTRR) {
4358 		add_x86_feature(featureset, X86FSET_MTRR);
4359 	}
4360 	if (cp->cp_edx & CPUID_INTC_EDX_PGE) {
4361 		add_x86_feature(featureset, X86FSET_PGE);
4362 	}
4363 	if (cp->cp_edx & CPUID_INTC_EDX_CMOV) {
4364 		add_x86_feature(featureset, X86FSET_CMOV);
4365 	}
4366 	if (cp->cp_edx & CPUID_INTC_EDX_MMX) {
4367 		add_x86_feature(featureset, X86FSET_MMX);
4368 	}
4369 	if ((cp->cp_edx & CPUID_INTC_EDX_MCE) != 0 &&
4370 	    (cp->cp_edx & CPUID_INTC_EDX_MCA) != 0) {
4371 		add_x86_feature(featureset, X86FSET_MCA);
4372 	}
4373 	if (cp->cp_edx & CPUID_INTC_EDX_PAE) {
4374 		add_x86_feature(featureset, X86FSET_PAE);
4375 	}
4376 	if (cp->cp_edx & CPUID_INTC_EDX_CX8) {
4377 		add_x86_feature(featureset, X86FSET_CX8);
4378 	}
4379 	if (cp->cp_ecx & CPUID_INTC_ECX_CX16) {
4380 		add_x86_feature(featureset, X86FSET_CX16);
4381 	}
4382 	if (cp->cp_edx & CPUID_INTC_EDX_PAT) {
4383 		add_x86_feature(featureset, X86FSET_PAT);
4384 	}
4385 	if (cp->cp_edx & CPUID_INTC_EDX_SEP) {
4386 		add_x86_feature(featureset, X86FSET_SEP);
4387 	}
4388 	if (cp->cp_edx & CPUID_INTC_EDX_FXSR) {
4389 		/*
4390 		 * In our implementation, fxsave/fxrstor
4391 		 * are prerequisites before we'll even
4392 		 * try and do SSE things.
4393 		 */
4394 		if (cp->cp_edx & CPUID_INTC_EDX_SSE) {
4395 			add_x86_feature(featureset, X86FSET_SSE);
4396 		}
4397 		if (cp->cp_edx & CPUID_INTC_EDX_SSE2) {
4398 			add_x86_feature(featureset, X86FSET_SSE2);
4399 		}
4400 		if (cp->cp_ecx & CPUID_INTC_ECX_SSE3) {
4401 			add_x86_feature(featureset, X86FSET_SSE3);
4402 		}
4403 		if (cp->cp_ecx & CPUID_INTC_ECX_SSSE3) {
4404 			add_x86_feature(featureset, X86FSET_SSSE3);
4405 		}
4406 		if (cp->cp_ecx & CPUID_INTC_ECX_SSE4_1) {
4407 			add_x86_feature(featureset, X86FSET_SSE4_1);
4408 		}
4409 		if (cp->cp_ecx & CPUID_INTC_ECX_SSE4_2) {
4410 			add_x86_feature(featureset, X86FSET_SSE4_2);
4411 		}
4412 		if (cp->cp_ecx & CPUID_INTC_ECX_AES) {
4413 			add_x86_feature(featureset, X86FSET_AES);
4414 		}
4415 		if (cp->cp_ecx & CPUID_INTC_ECX_PCLMULQDQ) {
4416 			add_x86_feature(featureset, X86FSET_PCLMULQDQ);
4417 		}
4418 
4419 		if (cpi->cpi_std[7].cp_ebx & CPUID_INTC_EBX_7_0_SHA)
4420 			add_x86_feature(featureset, X86FSET_SHA);
4421 
4422 		if (cp->cp_ecx & CPUID_INTC_ECX_XSAVE) {
4423 			add_x86_feature(featureset, X86FSET_XSAVE);
4424 
4425 			/* We only test AVX & AVX512 when there is XSAVE */
4426 			cpuid_basic_avx(cpu, featureset);
4427 		}
4428 	}
4429 
4430 	if (cp->cp_ecx & CPUID_INTC_ECX_PCID) {
4431 		add_x86_feature(featureset, X86FSET_PCID);
4432 	}
4433 
4434 	if (cp->cp_ecx & CPUID_INTC_ECX_X2APIC) {
4435 		add_x86_feature(featureset, X86FSET_X2APIC);
4436 	}
4437 	if (cp->cp_edx & CPUID_INTC_EDX_DE) {
4438 		add_x86_feature(featureset, X86FSET_DE);
4439 	}
4440 #if !defined(__xpv)
4441 	if (cp->cp_ecx & CPUID_INTC_ECX_MON) {
4442 
4443 		/*
4444 		 * We require the CLFLUSH instruction for erratum workaround
4445 		 * to use MONITOR/MWAIT.
4446 		 */
4447 		if (cp->cp_edx & CPUID_INTC_EDX_CLFSH) {
4448 			cpi->cpi_mwait.support |= MWAIT_SUPPORT;
4449 			add_x86_feature(featureset, X86FSET_MWAIT);
4450 		} else {
4451 			extern int idle_cpu_assert_cflush_monitor;
4452 
4453 			/*
4454 			 * All processors we are aware of which have
4455 			 * MONITOR/MWAIT also have CLFLUSH.
4456 			 */
4457 			if (idle_cpu_assert_cflush_monitor) {
4458 				ASSERT((cp->cp_ecx & CPUID_INTC_ECX_MON) &&
4459 				    (cp->cp_edx & CPUID_INTC_EDX_CLFSH));
4460 			}
4461 		}
4462 	}
4463 #endif	/* __xpv */
4464 
4465 	if (cp->cp_ecx & CPUID_INTC_ECX_VMX) {
4466 		add_x86_feature(featureset, X86FSET_VMX);
4467 	}
4468 
4469 	if (cp->cp_ecx & CPUID_INTC_ECX_RDRAND)
4470 		add_x86_feature(featureset, X86FSET_RDRAND);
4471 
4472 	/*
4473 	 * Only need it first time, rest of the cpus would follow suit.
4474 	 * we only capture this for the bootcpu.
4475 	 */
4476 	if (cp->cp_edx & CPUID_INTC_EDX_CLFSH) {
4477 		add_x86_feature(featureset, X86FSET_CLFSH);
4478 		x86_clflush_size = (BITX(cp->cp_ebx, 15, 8) * 8);
4479 	}
4480 	if (is_x86_feature(featureset, X86FSET_PAE))
4481 		cpi->cpi_pabits = 36;
4482 
4483 	if (cpi->cpi_maxeax >= 0xD && !xsave_force_disable) {
4484 		struct cpuid_regs r, *ecp;
4485 
4486 		ecp = &r;
4487 		ecp->cp_eax = 0xD;
4488 		ecp->cp_ecx = 1;
4489 		ecp->cp_edx = ecp->cp_ebx = 0;
4490 		(void) __cpuid_insn(ecp);
4491 
4492 		if (ecp->cp_eax & CPUID_INTC_EAX_D_1_XSAVEOPT)
4493 			add_x86_feature(featureset, X86FSET_XSAVEOPT);
4494 		if (ecp->cp_eax & CPUID_INTC_EAX_D_1_XSAVEC)
4495 			add_x86_feature(featureset, X86FSET_XSAVEC);
4496 		if (ecp->cp_eax & CPUID_INTC_EAX_D_1_XSAVES)
4497 			add_x86_feature(featureset, X86FSET_XSAVES);
4498 
4499 		/*
4500 		 * Zen 2 family processors suffer from erratum 1386 that causes
4501 		 * xsaves to not function correctly in some circumstances. There
4502 		 * are no supervisor states in Zen 2 and earlier. Practically
4503 		 * speaking this has no impact for us as we currently do not
4504 		 * leverage compressed xsave formats. To safeguard against
4505 		 * issues in the future where we may opt to using it, we remove
4506 		 * it from the feature set now. While Matisse has a microcode
4507 		 * update available with a fix, not all Zen 2 CPUs do so it's
4508 		 * simpler for the moment to unconditionally remove it.
4509 		 */
4510 		if (cpi->cpi_vendor == X86_VENDOR_AMD &&
4511 		    uarchrev_uarch(cpi->cpi_uarchrev) <= X86_UARCH_AMD_ZEN2) {
4512 			remove_x86_feature(featureset, X86FSET_XSAVES);
4513 		}
4514 	}
4515 
4516 	/*
4517 	 * Work on the "extended" feature information, doing
4518 	 * some basic initialization to be used in the extended pass.
4519 	 */
4520 	xcpuid = 0;
4521 	switch (cpi->cpi_vendor) {
4522 	case X86_VENDOR_Intel:
4523 		/*
4524 		 * On KVM we know we will have proper support for extended
4525 		 * cpuid.
4526 		 */
4527 		if (IS_NEW_F6(cpi) || cpi->cpi_family >= 0xf ||
4528 		    (get_hwenv() == HW_KVM && cpi->cpi_family == 6 &&
4529 		    (cpi->cpi_model == 6 || cpi->cpi_model == 2)))
4530 			xcpuid++;
4531 		break;
4532 	case X86_VENDOR_AMD:
4533 		if (cpi->cpi_family > 5 ||
4534 		    (cpi->cpi_family == 5 && cpi->cpi_model >= 1))
4535 			xcpuid++;
4536 		break;
4537 	case X86_VENDOR_Cyrix:
4538 		/*
4539 		 * Only these Cyrix CPUs are -known- to support
4540 		 * extended cpuid operations.
4541 		 */
4542 		if (x86_type == X86_TYPE_VIA_CYRIX_III ||
4543 		    x86_type == X86_TYPE_CYRIX_GXm)
4544 			xcpuid++;
4545 		break;
4546 	case X86_VENDOR_HYGON:
4547 	case X86_VENDOR_Centaur:
4548 	case X86_VENDOR_TM:
4549 	default:
4550 		xcpuid++;
4551 		break;
4552 	}
4553 
4554 	if (xcpuid) {
4555 		cp = &cpi->cpi_extd[0];
4556 		cp->cp_eax = CPUID_LEAF_EXT_0;
4557 		cpi->cpi_xmaxeax = __cpuid_insn(cp);
4558 	}
4559 
4560 	if (cpi->cpi_xmaxeax & CPUID_LEAF_EXT_0) {
4561 
4562 		if (cpi->cpi_xmaxeax > CPI_XMAXEAX_MAX)
4563 			cpi->cpi_xmaxeax = CPI_XMAXEAX_MAX;
4564 
4565 		switch (cpi->cpi_vendor) {
4566 		case X86_VENDOR_Intel:
4567 		case X86_VENDOR_AMD:
4568 		case X86_VENDOR_HYGON:
4569 			if (cpi->cpi_xmaxeax < 0x80000001)
4570 				break;
4571 			cp = &cpi->cpi_extd[1];
4572 			cp->cp_eax = 0x80000001;
4573 			(void) __cpuid_insn(cp);
4574 
4575 			if (cpi->cpi_vendor == X86_VENDOR_AMD &&
4576 			    cpi->cpi_family == 5 &&
4577 			    cpi->cpi_model == 6 &&
4578 			    cpi->cpi_step == 6) {
4579 				/*
4580 				 * K6 model 6 uses bit 10 to indicate SYSC
4581 				 * Later models use bit 11. Fix it here.
4582 				 */
4583 				if (cp->cp_edx & 0x400) {
4584 					cp->cp_edx &= ~0x400;
4585 					cp->cp_edx |= CPUID_AMD_EDX_SYSC;
4586 				}
4587 			}
4588 
4589 			platform_cpuid_mangle(cpi->cpi_vendor, 0x80000001, cp);
4590 
4591 			/*
4592 			 * Compute the additions to the kernel's feature word.
4593 			 */
4594 			if (cp->cp_edx & CPUID_AMD_EDX_NX) {
4595 				add_x86_feature(featureset, X86FSET_NX);
4596 			}
4597 
4598 			/*
4599 			 * Regardless whether or not we boot 64-bit,
4600 			 * we should have a way to identify whether
4601 			 * the CPU is capable of running 64-bit.
4602 			 */
4603 			if (cp->cp_edx & CPUID_AMD_EDX_LM) {
4604 				add_x86_feature(featureset, X86FSET_64);
4605 			}
4606 
4607 			/* 1 GB large page - enable only for 64 bit kernel */
4608 			if (cp->cp_edx & CPUID_AMD_EDX_1GPG) {
4609 				add_x86_feature(featureset, X86FSET_1GPG);
4610 			}
4611 
4612 			if ((cpi->cpi_vendor == X86_VENDOR_AMD ||
4613 			    cpi->cpi_vendor == X86_VENDOR_HYGON) &&
4614 			    (cpi->cpi_std[1].cp_edx & CPUID_INTC_EDX_FXSR) &&
4615 			    (cp->cp_ecx & CPUID_AMD_ECX_SSE4A)) {
4616 				add_x86_feature(featureset, X86FSET_SSE4A);
4617 			}
4618 
4619 			/*
4620 			 * It's really tricky to support syscall/sysret in
4621 			 * the i386 kernel; we rely on sysenter/sysexit
4622 			 * instead.  In the amd64 kernel, things are -way-
4623 			 * better.
4624 			 */
4625 			if (cp->cp_edx & CPUID_AMD_EDX_SYSC) {
4626 				add_x86_feature(featureset, X86FSET_ASYSC);
4627 			}
4628 
4629 			/*
4630 			 * While we're thinking about system calls, note
4631 			 * that AMD processors don't support sysenter
4632 			 * in long mode at all, so don't try to program them.
4633 			 */
4634 			if (x86_vendor == X86_VENDOR_AMD ||
4635 			    x86_vendor == X86_VENDOR_HYGON) {
4636 				remove_x86_feature(featureset, X86FSET_SEP);
4637 			}
4638 
4639 			if (cp->cp_edx & CPUID_AMD_EDX_TSCP) {
4640 				add_x86_feature(featureset, X86FSET_TSCP);
4641 			}
4642 
4643 			if (cp->cp_ecx & CPUID_AMD_ECX_SVM) {
4644 				add_x86_feature(featureset, X86FSET_SVM);
4645 			}
4646 
4647 			if (cp->cp_ecx & CPUID_AMD_ECX_TOPOEXT) {
4648 				add_x86_feature(featureset, X86FSET_TOPOEXT);
4649 			}
4650 
4651 			if (cp->cp_ecx & CPUID_AMD_ECX_PCEC) {
4652 				add_x86_feature(featureset, X86FSET_AMD_PCEC);
4653 			}
4654 
4655 			if (cp->cp_ecx & CPUID_AMD_ECX_XOP) {
4656 				add_x86_feature(featureset, X86FSET_XOP);
4657 			}
4658 
4659 			if (cp->cp_ecx & CPUID_AMD_ECX_FMA4) {
4660 				add_x86_feature(featureset, X86FSET_FMA4);
4661 			}
4662 
4663 			if (cp->cp_ecx & CPUID_AMD_ECX_TBM) {
4664 				add_x86_feature(featureset, X86FSET_TBM);
4665 			}
4666 
4667 			if (cp->cp_ecx & CPUID_AMD_ECX_MONITORX) {
4668 				add_x86_feature(featureset, X86FSET_MONITORX);
4669 			}
4670 			break;
4671 		default:
4672 			break;
4673 		}
4674 
4675 		/*
4676 		 * Get CPUID data about processor cores and hyperthreads.
4677 		 */
4678 		switch (cpi->cpi_vendor) {
4679 		case X86_VENDOR_Intel:
4680 			if (cpi->cpi_maxeax >= 4) {
4681 				cp = &cpi->cpi_std[4];
4682 				cp->cp_eax = 4;
4683 				cp->cp_ecx = 0;
4684 				(void) __cpuid_insn(cp);
4685 				platform_cpuid_mangle(cpi->cpi_vendor, 4, cp);
4686 			}
4687 			/*FALLTHROUGH*/
4688 		case X86_VENDOR_AMD:
4689 		case X86_VENDOR_HYGON:
4690 			if (cpi->cpi_xmaxeax < CPUID_LEAF_EXT_8)
4691 				break;
4692 			cp = &cpi->cpi_extd[8];
4693 			cp->cp_eax = CPUID_LEAF_EXT_8;
4694 			(void) __cpuid_insn(cp);
4695 			platform_cpuid_mangle(cpi->cpi_vendor, CPUID_LEAF_EXT_8,
4696 			    cp);
4697 
4698 			/*
4699 			 * AMD uses ebx for some extended functions.
4700 			 */
4701 			if (cpi->cpi_vendor == X86_VENDOR_AMD ||
4702 			    cpi->cpi_vendor == X86_VENDOR_HYGON) {
4703 				/*
4704 				 * While we're here, check for the AMD "Error
4705 				 * Pointer Zero/Restore" feature. This can be
4706 				 * used to setup the FP save handlers
4707 				 * appropriately.
4708 				 */
4709 				if (cp->cp_ebx & CPUID_AMD_EBX_ERR_PTR_ZERO) {
4710 					cpi->cpi_fp_amd_save = 0;
4711 				} else {
4712 					cpi->cpi_fp_amd_save = 1;
4713 				}
4714 
4715 				if (cp->cp_ebx & CPUID_AMD_EBX_CLZERO) {
4716 					add_x86_feature(featureset,
4717 					    X86FSET_CLZERO);
4718 				}
4719 			}
4720 
4721 			/*
4722 			 * Virtual and physical address limits from
4723 			 * cpuid override previously guessed values.
4724 			 */
4725 			cpi->cpi_pabits = BITX(cp->cp_eax, 7, 0);
4726 			cpi->cpi_vabits = BITX(cp->cp_eax, 15, 8);
4727 			break;
4728 		default:
4729 			break;
4730 		}
4731 
4732 		/*
4733 		 * Get CPUID data about TSC Invariance in Deep C-State.
4734 		 */
4735 		switch (cpi->cpi_vendor) {
4736 		case X86_VENDOR_Intel:
4737 		case X86_VENDOR_AMD:
4738 		case X86_VENDOR_HYGON:
4739 			if (cpi->cpi_maxeax >= 7) {
4740 				cp = &cpi->cpi_extd[7];
4741 				cp->cp_eax = 0x80000007;
4742 				cp->cp_ecx = 0;
4743 				(void) __cpuid_insn(cp);
4744 			}
4745 			break;
4746 		default:
4747 			break;
4748 		}
4749 	}
4750 
4751 	/*
4752 	 * cpuid_basic_ppin assumes that cpuid_basic_topology has already been
4753 	 * run and thus gathered some of its dependent leaves.
4754 	 */
4755 	cpuid_basic_topology(cpu, featureset);
4756 	cpuid_basic_thermal(cpu, featureset);
4757 #if !defined(__xpv)
4758 	cpuid_basic_ppin(cpu, featureset);
4759 #endif
4760 
4761 	if (cpi->cpi_vendor == X86_VENDOR_AMD ||
4762 	    cpi->cpi_vendor == X86_VENDOR_HYGON) {
4763 		if (cpi->cpi_xmaxeax >= CPUID_LEAF_EXT_8 &&
4764 		    cpi->cpi_extd[8].cp_ebx & CPUID_AMD_EBX_ERR_PTR_ZERO) {
4765 			/* Special handling for AMD FP not necessary. */
4766 			cpi->cpi_fp_amd_save = 0;
4767 		} else {
4768 			cpi->cpi_fp_amd_save = 1;
4769 		}
4770 	}
4771 
4772 	/*
4773 	 * Check (and potentially set) if lfence is serializing.
4774 	 * This is useful for accurate rdtsc measurements and AMD retpolines.
4775 	 */
4776 	if ((cpi->cpi_vendor == X86_VENDOR_AMD ||
4777 	    cpi->cpi_vendor == X86_VENDOR_HYGON) &&
4778 	    is_x86_feature(featureset, X86FSET_SSE2)) {
4779 		/*
4780 		 * The AMD white paper Software Techniques For Managing
4781 		 * Speculation on AMD Processors details circumstances for when
4782 		 * lfence instructions are serializing.
4783 		 *
4784 		 * On family 0xf and 0x11, it is inherently so.  On family 0x10
4785 		 * and later (excluding 0x11), a bit in the DE_CFG MSR
4786 		 * determines the lfence behavior.  Per that whitepaper, AMD has
4787 		 * committed to supporting that MSR on all later CPUs.
4788 		 */
4789 		if (cpi->cpi_family == 0xf || cpi->cpi_family == 0x11) {
4790 			add_x86_feature(featureset, X86FSET_LFENCE_SER);
4791 		} else if (cpi->cpi_family >= 0x10) {
4792 #if !defined(__xpv)
4793 			uint64_t val;
4794 
4795 			/*
4796 			 * Be careful when attempting to enable the bit, and
4797 			 * verify that it was actually set in case we are
4798 			 * running in a hypervisor which is less than faithful
4799 			 * about its emulation of this feature.
4800 			 */
4801 			on_trap_data_t otd;
4802 			if (!on_trap(&otd, OT_DATA_ACCESS)) {
4803 				val = rdmsr(MSR_AMD_DE_CFG);
4804 				val |= AMD_DE_CFG_LFENCE_DISPATCH;
4805 				wrmsr(MSR_AMD_DE_CFG, val);
4806 				val = rdmsr(MSR_AMD_DE_CFG);
4807 			} else {
4808 				val = 0;
4809 			}
4810 			no_trap();
4811 
4812 			if ((val & AMD_DE_CFG_LFENCE_DISPATCH) != 0) {
4813 				add_x86_feature(featureset, X86FSET_LFENCE_SER);
4814 			}
4815 #endif
4816 		}
4817 	} else if (cpi->cpi_vendor == X86_VENDOR_Intel &&
4818 	    is_x86_feature(featureset, X86FSET_SSE2)) {
4819 		/*
4820 		 * Documentation and other OSes indicate that lfence is always
4821 		 * serializing on Intel CPUs.
4822 		 */
4823 		add_x86_feature(featureset, X86FSET_LFENCE_SER);
4824 	}
4825 
4826 
4827 	/*
4828 	 * Check the processor leaves that are used for security features. Grab
4829 	 * any additional processor-specific leaves that we may not have yet.
4830 	 */
4831 	switch (cpi->cpi_vendor) {
4832 	case X86_VENDOR_AMD:
4833 	case X86_VENDOR_HYGON:
4834 		if (cpi->cpi_xmaxeax >= CPUID_LEAF_EXT_21) {
4835 			cp = &cpi->cpi_extd[0x21];
4836 			cp->cp_eax = CPUID_LEAF_EXT_21;
4837 			cp->cp_ecx = 0;
4838 			(void) __cpuid_insn(cp);
4839 		}
4840 		break;
4841 	default:
4842 		break;
4843 	}
4844 
4845 	cpuid_scan_security(cpu, featureset);
4846 }
4847 
4848 /*
4849  * Make copies of the cpuid table entries we depend on, in
4850  * part for ease of parsing now, in part so that we have only
4851  * one place to correct any of it, in part for ease of
4852  * later export to userland, and in part so we can look at
4853  * this stuff in a crash dump.
4854  */
4855 
4856 static void
4857 cpuid_pass_extended(cpu_t *cpu, void *_arg __unused)
4858 {
4859 	uint_t n, nmax;
4860 	int i;
4861 	struct cpuid_regs *cp;
4862 	uint8_t *dp;
4863 	uint32_t *iptr;
4864 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
4865 
4866 	if (cpi->cpi_maxeax < 1)
4867 		return;
4868 
4869 	if ((nmax = cpi->cpi_maxeax + 1) > NMAX_CPI_STD)
4870 		nmax = NMAX_CPI_STD;
4871 	/*
4872 	 * (We already handled n == 0 and n == 1 in the basic pass)
4873 	 */
4874 	for (n = 2, cp = &cpi->cpi_std[2]; n < nmax; n++, cp++) {
4875 		/*
4876 		 * leaves 6 and 7 were handled in the basic pass
4877 		 */
4878 		if (n == 6 || n == 7)
4879 			continue;
4880 
4881 		cp->cp_eax = n;
4882 
4883 		/*
4884 		 * CPUID function 4 expects %ecx to be initialized
4885 		 * with an index which indicates which cache to return
4886 		 * information about. The OS is expected to call function 4
4887 		 * with %ecx set to 0, 1, 2, ... until it returns with
4888 		 * EAX[4:0] set to 0, which indicates there are no more
4889 		 * caches.
4890 		 *
4891 		 * Here, populate cpi_std[4] with the information returned by
4892 		 * function 4 when %ecx == 0, and do the rest in a later pass
4893 		 * when dynamic memory allocation becomes available.
4894 		 *
4895 		 * Note: we need to explicitly initialize %ecx here, since
4896 		 * function 4 may have been previously invoked.
4897 		 */
4898 		if (n == 4)
4899 			cp->cp_ecx = 0;
4900 
4901 		(void) __cpuid_insn(cp);
4902 		platform_cpuid_mangle(cpi->cpi_vendor, n, cp);
4903 		switch (n) {
4904 		case 2:
4905 			/*
4906 			 * "the lower 8 bits of the %eax register
4907 			 * contain a value that identifies the number
4908 			 * of times the cpuid [instruction] has to be
4909 			 * executed to obtain a complete image of the
4910 			 * processor's caching systems."
4911 			 *
4912 			 * How *do* they make this stuff up?
4913 			 */
4914 			cpi->cpi_ncache = sizeof (*cp) *
4915 			    BITX(cp->cp_eax, 7, 0);
4916 			if (cpi->cpi_ncache == 0)
4917 				break;
4918 			cpi->cpi_ncache--;	/* skip count byte */
4919 
4920 			/*
4921 			 * Well, for now, rather than attempt to implement
4922 			 * this slightly dubious algorithm, we just look
4923 			 * at the first 15 ..
4924 			 */
4925 			if (cpi->cpi_ncache > (sizeof (*cp) - 1))
4926 				cpi->cpi_ncache = sizeof (*cp) - 1;
4927 
4928 			dp = cpi->cpi_cacheinfo;
4929 			if (BITX(cp->cp_eax, 31, 31) == 0) {
4930 				uint8_t *p = (void *)&cp->cp_eax;
4931 				for (i = 1; i < 4; i++)
4932 					if (p[i] != 0)
4933 						*dp++ = p[i];
4934 			}
4935 			if (BITX(cp->cp_ebx, 31, 31) == 0) {
4936 				uint8_t *p = (void *)&cp->cp_ebx;
4937 				for (i = 0; i < 4; i++)
4938 					if (p[i] != 0)
4939 						*dp++ = p[i];
4940 			}
4941 			if (BITX(cp->cp_ecx, 31, 31) == 0) {
4942 				uint8_t *p = (void *)&cp->cp_ecx;
4943 				for (i = 0; i < 4; i++)
4944 					if (p[i] != 0)
4945 						*dp++ = p[i];
4946 			}
4947 			if (BITX(cp->cp_edx, 31, 31) == 0) {
4948 				uint8_t *p = (void *)&cp->cp_edx;
4949 				for (i = 0; i < 4; i++)
4950 					if (p[i] != 0)
4951 						*dp++ = p[i];
4952 			}
4953 			break;
4954 
4955 		case 3:	/* Processor serial number, if PSN supported */
4956 			break;
4957 
4958 		case 4:	/* Deterministic cache parameters */
4959 			break;
4960 
4961 		case 5:	/* Monitor/Mwait parameters */
4962 		{
4963 			size_t mwait_size;
4964 
4965 			/*
4966 			 * check cpi_mwait.support which was set in
4967 			 * cpuid_pass_basic()
4968 			 */
4969 			if (!(cpi->cpi_mwait.support & MWAIT_SUPPORT))
4970 				break;
4971 
4972 			/*
4973 			 * Protect ourself from insane mwait line size.
4974 			 * Workaround for incomplete hardware emulator(s).
4975 			 */
4976 			mwait_size = (size_t)MWAIT_SIZE_MAX(cpi);
4977 			if (mwait_size < sizeof (uint32_t) ||
4978 			    !ISP2(mwait_size)) {
4979 #if DEBUG
4980 				cmn_err(CE_NOTE, "Cannot handle cpu %d mwait "
4981 				    "size %ld", cpu->cpu_id, (long)mwait_size);
4982 #endif
4983 				break;
4984 			}
4985 
4986 			cpi->cpi_mwait.mon_min = (size_t)MWAIT_SIZE_MIN(cpi);
4987 			cpi->cpi_mwait.mon_max = mwait_size;
4988 			if (MWAIT_EXTENSION(cpi)) {
4989 				cpi->cpi_mwait.support |= MWAIT_EXTENSIONS;
4990 				if (MWAIT_INT_ENABLE(cpi))
4991 					cpi->cpi_mwait.support |=
4992 					    MWAIT_ECX_INT_ENABLE;
4993 			}
4994 			break;
4995 		}
4996 		default:
4997 			break;
4998 		}
4999 	}
5000 
5001 	/*
5002 	 * XSAVE enumeration
5003 	 */
5004 	if (cpi->cpi_maxeax >= 0xD) {
5005 		struct cpuid_regs regs;
5006 		boolean_t cpuid_d_valid = B_TRUE;
5007 
5008 		cp = &regs;
5009 		cp->cp_eax = 0xD;
5010 		cp->cp_edx = cp->cp_ebx = cp->cp_ecx = 0;
5011 
5012 		(void) __cpuid_insn(cp);
5013 
5014 		/*
5015 		 * Sanity checks for debug
5016 		 */
5017 		if ((cp->cp_eax & XFEATURE_LEGACY_FP) == 0 ||
5018 		    (cp->cp_eax & XFEATURE_SSE) == 0) {
5019 			cpuid_d_valid = B_FALSE;
5020 		}
5021 
5022 		cpi->cpi_xsave.xsav_hw_features_low = cp->cp_eax;
5023 		cpi->cpi_xsave.xsav_hw_features_high = cp->cp_edx;
5024 		cpi->cpi_xsave.xsav_max_size = cp->cp_ecx;
5025 
5026 		/*
5027 		 * If the hw supports AVX, get the size and offset in the save
5028 		 * area for the ymm state.
5029 		 */
5030 		if (cpi->cpi_xsave.xsav_hw_features_low & XFEATURE_AVX) {
5031 			cp->cp_eax = 0xD;
5032 			cp->cp_ecx = 2;
5033 			cp->cp_edx = cp->cp_ebx = 0;
5034 
5035 			(void) __cpuid_insn(cp);
5036 
5037 			if (cp->cp_ebx != CPUID_LEAFD_2_YMM_OFFSET ||
5038 			    cp->cp_eax != CPUID_LEAFD_2_YMM_SIZE) {
5039 				cpuid_d_valid = B_FALSE;
5040 			}
5041 
5042 			cpi->cpi_xsave.ymm_size = cp->cp_eax;
5043 			cpi->cpi_xsave.ymm_offset = cp->cp_ebx;
5044 		}
5045 
5046 		/*
5047 		 * If the hw supports MPX, get the size and offset in the
5048 		 * save area for BNDREGS and BNDCSR.
5049 		 */
5050 		if (cpi->cpi_xsave.xsav_hw_features_low & XFEATURE_MPX) {
5051 			cp->cp_eax = 0xD;
5052 			cp->cp_ecx = 3;
5053 			cp->cp_edx = cp->cp_ebx = 0;
5054 
5055 			(void) __cpuid_insn(cp);
5056 
5057 			cpi->cpi_xsave.bndregs_size = cp->cp_eax;
5058 			cpi->cpi_xsave.bndregs_offset = cp->cp_ebx;
5059 
5060 			cp->cp_eax = 0xD;
5061 			cp->cp_ecx = 4;
5062 			cp->cp_edx = cp->cp_ebx = 0;
5063 
5064 			(void) __cpuid_insn(cp);
5065 
5066 			cpi->cpi_xsave.bndcsr_size = cp->cp_eax;
5067 			cpi->cpi_xsave.bndcsr_offset = cp->cp_ebx;
5068 		}
5069 
5070 		/*
5071 		 * If the hw supports AVX512, get the size and offset in the
5072 		 * save area for the opmask registers and zmm state.
5073 		 */
5074 		if (cpi->cpi_xsave.xsav_hw_features_low & XFEATURE_AVX512) {
5075 			cp->cp_eax = 0xD;
5076 			cp->cp_ecx = 5;
5077 			cp->cp_edx = cp->cp_ebx = 0;
5078 
5079 			(void) __cpuid_insn(cp);
5080 
5081 			cpi->cpi_xsave.opmask_size = cp->cp_eax;
5082 			cpi->cpi_xsave.opmask_offset = cp->cp_ebx;
5083 
5084 			cp->cp_eax = 0xD;
5085 			cp->cp_ecx = 6;
5086 			cp->cp_edx = cp->cp_ebx = 0;
5087 
5088 			(void) __cpuid_insn(cp);
5089 
5090 			cpi->cpi_xsave.zmmlo_size = cp->cp_eax;
5091 			cpi->cpi_xsave.zmmlo_offset = cp->cp_ebx;
5092 
5093 			cp->cp_eax = 0xD;
5094 			cp->cp_ecx = 7;
5095 			cp->cp_edx = cp->cp_ebx = 0;
5096 
5097 			(void) __cpuid_insn(cp);
5098 
5099 			cpi->cpi_xsave.zmmhi_size = cp->cp_eax;
5100 			cpi->cpi_xsave.zmmhi_offset = cp->cp_ebx;
5101 		}
5102 
5103 		if (is_x86_feature(x86_featureset, X86FSET_XSAVE)) {
5104 			xsave_state_size = 0;
5105 		} else if (cpuid_d_valid) {
5106 			xsave_state_size = cpi->cpi_xsave.xsav_max_size;
5107 		} else {
5108 			/* Broken CPUID 0xD, probably in HVM */
5109 			cmn_err(CE_WARN, "cpu%d: CPUID.0xD returns invalid "
5110 			    "value: hw_low = %d, hw_high = %d, xsave_size = %d"
5111 			    ", ymm_size = %d, ymm_offset = %d\n",
5112 			    cpu->cpu_id, cpi->cpi_xsave.xsav_hw_features_low,
5113 			    cpi->cpi_xsave.xsav_hw_features_high,
5114 			    (int)cpi->cpi_xsave.xsav_max_size,
5115 			    (int)cpi->cpi_xsave.ymm_size,
5116 			    (int)cpi->cpi_xsave.ymm_offset);
5117 
5118 			if (xsave_state_size != 0) {
5119 				/*
5120 				 * This must be a non-boot CPU. We cannot
5121 				 * continue, because boot cpu has already
5122 				 * enabled XSAVE.
5123 				 */
5124 				ASSERT(cpu->cpu_id != 0);
5125 				cmn_err(CE_PANIC, "cpu%d: we have already "
5126 				    "enabled XSAVE on boot cpu, cannot "
5127 				    "continue.", cpu->cpu_id);
5128 			} else {
5129 				/*
5130 				 * If we reached here on the boot CPU, it's also
5131 				 * almost certain that we'll reach here on the
5132 				 * non-boot CPUs. When we're here on a boot CPU
5133 				 * we should disable the feature, on a non-boot
5134 				 * CPU we need to confirm that we have.
5135 				 */
5136 				if (cpu->cpu_id == 0) {
5137 					remove_x86_feature(x86_featureset,
5138 					    X86FSET_XSAVE);
5139 					remove_x86_feature(x86_featureset,
5140 					    X86FSET_AVX);
5141 					remove_x86_feature(x86_featureset,
5142 					    X86FSET_F16C);
5143 					remove_x86_feature(x86_featureset,
5144 					    X86FSET_BMI1);
5145 					remove_x86_feature(x86_featureset,
5146 					    X86FSET_BMI2);
5147 					remove_x86_feature(x86_featureset,
5148 					    X86FSET_FMA);
5149 					remove_x86_feature(x86_featureset,
5150 					    X86FSET_AVX2);
5151 					remove_x86_feature(x86_featureset,
5152 					    X86FSET_MPX);
5153 					remove_x86_feature(x86_featureset,
5154 					    X86FSET_AVX512F);
5155 					remove_x86_feature(x86_featureset,
5156 					    X86FSET_AVX512DQ);
5157 					remove_x86_feature(x86_featureset,
5158 					    X86FSET_AVX512PF);
5159 					remove_x86_feature(x86_featureset,
5160 					    X86FSET_AVX512ER);
5161 					remove_x86_feature(x86_featureset,
5162 					    X86FSET_AVX512CD);
5163 					remove_x86_feature(x86_featureset,
5164 					    X86FSET_AVX512BW);
5165 					remove_x86_feature(x86_featureset,
5166 					    X86FSET_AVX512VL);
5167 					remove_x86_feature(x86_featureset,
5168 					    X86FSET_AVX512FMA);
5169 					remove_x86_feature(x86_featureset,
5170 					    X86FSET_AVX512VBMI);
5171 					remove_x86_feature(x86_featureset,
5172 					    X86FSET_AVX512VNNI);
5173 					remove_x86_feature(x86_featureset,
5174 					    X86FSET_AVX512VPOPCDQ);
5175 					remove_x86_feature(x86_featureset,
5176 					    X86FSET_AVX512NNIW);
5177 					remove_x86_feature(x86_featureset,
5178 					    X86FSET_AVX512FMAPS);
5179 					remove_x86_feature(x86_featureset,
5180 					    X86FSET_VAES);
5181 					remove_x86_feature(x86_featureset,
5182 					    X86FSET_VPCLMULQDQ);
5183 					remove_x86_feature(x86_featureset,
5184 					    X86FSET_GFNI);
5185 					remove_x86_feature(x86_featureset,
5186 					    X86FSET_AVX512_VP2INT);
5187 					remove_x86_feature(x86_featureset,
5188 					    X86FSET_AVX512_BITALG);
5189 					remove_x86_feature(x86_featureset,
5190 					    X86FSET_AVX512_VBMI2);
5191 					remove_x86_feature(x86_featureset,
5192 					    X86FSET_AVX512_BF16);
5193 
5194 					xsave_force_disable = B_TRUE;
5195 				} else {
5196 					VERIFY(is_x86_feature(x86_featureset,
5197 					    X86FSET_XSAVE) == B_FALSE);
5198 				}
5199 			}
5200 		}
5201 	}
5202 
5203 
5204 	if ((cpi->cpi_xmaxeax & CPUID_LEAF_EXT_0) == 0)
5205 		return;
5206 
5207 	if ((nmax = cpi->cpi_xmaxeax - CPUID_LEAF_EXT_0 + 1) > NMAX_CPI_EXTD)
5208 		nmax = NMAX_CPI_EXTD;
5209 	/*
5210 	 * Copy the extended properties, fixing them as we go. While we start at
5211 	 * 2 because we've already handled a few cases in the basic pass, the
5212 	 * rest we let ourselves just grab again (e.g. 0x8, 0x21).
5213 	 */
5214 	iptr = (void *)cpi->cpi_brandstr;
5215 	for (n = 2, cp = &cpi->cpi_extd[2]; n < nmax; cp++, n++) {
5216 		cp->cp_eax = CPUID_LEAF_EXT_0 + n;
5217 		(void) __cpuid_insn(cp);
5218 		platform_cpuid_mangle(cpi->cpi_vendor, CPUID_LEAF_EXT_0 + n,
5219 		    cp);
5220 		switch (n) {
5221 		case 2:
5222 		case 3:
5223 		case 4:
5224 			/*
5225 			 * Extract the brand string
5226 			 */
5227 			*iptr++ = cp->cp_eax;
5228 			*iptr++ = cp->cp_ebx;
5229 			*iptr++ = cp->cp_ecx;
5230 			*iptr++ = cp->cp_edx;
5231 			break;
5232 		case 5:
5233 			switch (cpi->cpi_vendor) {
5234 			case X86_VENDOR_AMD:
5235 				/*
5236 				 * The Athlon and Duron were the first
5237 				 * parts to report the sizes of the
5238 				 * TLB for large pages. Before then,
5239 				 * we don't trust the data.
5240 				 */
5241 				if (cpi->cpi_family < 6 ||
5242 				    (cpi->cpi_family == 6 &&
5243 				    cpi->cpi_model < 1))
5244 					cp->cp_eax = 0;
5245 				break;
5246 			default:
5247 				break;
5248 			}
5249 			break;
5250 		case 6:
5251 			switch (cpi->cpi_vendor) {
5252 			case X86_VENDOR_AMD:
5253 				/*
5254 				 * The Athlon and Duron were the first
5255 				 * AMD parts with L2 TLB's.
5256 				 * Before then, don't trust the data.
5257 				 */
5258 				if (cpi->cpi_family < 6 ||
5259 				    (cpi->cpi_family == 6 &&
5260 				    cpi->cpi_model < 1))
5261 					cp->cp_eax = cp->cp_ebx = 0;
5262 				/*
5263 				 * AMD Duron rev A0 reports L2
5264 				 * cache size incorrectly as 1K
5265 				 * when it is really 64K
5266 				 */
5267 				if (cpi->cpi_family == 6 &&
5268 				    cpi->cpi_model == 3 &&
5269 				    cpi->cpi_step == 0) {
5270 					cp->cp_ecx &= 0xffff;
5271 					cp->cp_ecx |= 0x400000;
5272 				}
5273 				break;
5274 			case X86_VENDOR_Cyrix:	/* VIA C3 */
5275 				/*
5276 				 * VIA C3 processors are a bit messed
5277 				 * up w.r.t. encoding cache sizes in %ecx
5278 				 */
5279 				if (cpi->cpi_family != 6)
5280 					break;
5281 				/*
5282 				 * model 7 and 8 were incorrectly encoded
5283 				 *
5284 				 * xxx is model 8 really broken?
5285 				 */
5286 				if (cpi->cpi_model == 7 ||
5287 				    cpi->cpi_model == 8)
5288 					cp->cp_ecx =
5289 					    BITX(cp->cp_ecx, 31, 24) << 16 |
5290 					    BITX(cp->cp_ecx, 23, 16) << 12 |
5291 					    BITX(cp->cp_ecx, 15, 8) << 8 |
5292 					    BITX(cp->cp_ecx, 7, 0);
5293 				/*
5294 				 * model 9 stepping 1 has wrong associativity
5295 				 */
5296 				if (cpi->cpi_model == 9 && cpi->cpi_step == 1)
5297 					cp->cp_ecx |= 8 << 12;
5298 				break;
5299 			case X86_VENDOR_Intel:
5300 				/*
5301 				 * Extended L2 Cache features function.
5302 				 * First appeared on Prescott.
5303 				 */
5304 			default:
5305 				break;
5306 			}
5307 			break;
5308 		default:
5309 			break;
5310 		}
5311 	}
5312 }
5313 
5314 static const char *
5315 intel_cpubrand(const struct cpuid_info *cpi)
5316 {
5317 	int i;
5318 
5319 	ASSERT(is_x86_feature(x86_featureset, X86FSET_CPUID));
5320 
5321 	switch (cpi->cpi_family) {
5322 	case 5:
5323 		return ("Intel Pentium(r)");
5324 	case 6:
5325 		switch (cpi->cpi_model) {
5326 			uint_t celeron, xeon;
5327 			const struct cpuid_regs *cp;
5328 		case 0:
5329 		case 1:
5330 		case 2:
5331 			return ("Intel Pentium(r) Pro");
5332 		case 3:
5333 		case 4:
5334 			return ("Intel Pentium(r) II");
5335 		case 6:
5336 			return ("Intel Celeron(r)");
5337 		case 5:
5338 		case 7:
5339 			celeron = xeon = 0;
5340 			cp = &cpi->cpi_std[2];	/* cache info */
5341 
5342 			for (i = 1; i < 4; i++) {
5343 				uint_t tmp;
5344 
5345 				tmp = (cp->cp_eax >> (8 * i)) & 0xff;
5346 				if (tmp == 0x40)
5347 					celeron++;
5348 				if (tmp >= 0x44 && tmp <= 0x45)
5349 					xeon++;
5350 			}
5351 
5352 			for (i = 0; i < 2; i++) {
5353 				uint_t tmp;
5354 
5355 				tmp = (cp->cp_ebx >> (8 * i)) & 0xff;
5356 				if (tmp == 0x40)
5357 					celeron++;
5358 				else if (tmp >= 0x44 && tmp <= 0x45)
5359 					xeon++;
5360 			}
5361 
5362 			for (i = 0; i < 4; i++) {
5363 				uint_t tmp;
5364 
5365 				tmp = (cp->cp_ecx >> (8 * i)) & 0xff;
5366 				if (tmp == 0x40)
5367 					celeron++;
5368 				else if (tmp >= 0x44 && tmp <= 0x45)
5369 					xeon++;
5370 			}
5371 
5372 			for (i = 0; i < 4; i++) {
5373 				uint_t tmp;
5374 
5375 				tmp = (cp->cp_edx >> (8 * i)) & 0xff;
5376 				if (tmp == 0x40)
5377 					celeron++;
5378 				else if (tmp >= 0x44 && tmp <= 0x45)
5379 					xeon++;
5380 			}
5381 
5382 			if (celeron)
5383 				return ("Intel Celeron(r)");
5384 			if (xeon)
5385 				return (cpi->cpi_model == 5 ?
5386 				    "Intel Pentium(r) II Xeon(tm)" :
5387 				    "Intel Pentium(r) III Xeon(tm)");
5388 			return (cpi->cpi_model == 5 ?
5389 			    "Intel Pentium(r) II or Pentium(r) II Xeon(tm)" :
5390 			    "Intel Pentium(r) III or Pentium(r) III Xeon(tm)");
5391 		default:
5392 			break;
5393 		}
5394 	default:
5395 		break;
5396 	}
5397 
5398 	/* BrandID is present if the field is nonzero */
5399 	if (cpi->cpi_brandid != 0) {
5400 		static const struct {
5401 			uint_t bt_bid;
5402 			const char *bt_str;
5403 		} brand_tbl[] = {
5404 			{ 0x1,	"Intel(r) Celeron(r)" },
5405 			{ 0x2,	"Intel(r) Pentium(r) III" },
5406 			{ 0x3,	"Intel(r) Pentium(r) III Xeon(tm)" },
5407 			{ 0x4,	"Intel(r) Pentium(r) III" },
5408 			{ 0x6,	"Mobile Intel(r) Pentium(r) III" },
5409 			{ 0x7,	"Mobile Intel(r) Celeron(r)" },
5410 			{ 0x8,	"Intel(r) Pentium(r) 4" },
5411 			{ 0x9,	"Intel(r) Pentium(r) 4" },
5412 			{ 0xa,	"Intel(r) Celeron(r)" },
5413 			{ 0xb,	"Intel(r) Xeon(tm)" },
5414 			{ 0xc,	"Intel(r) Xeon(tm) MP" },
5415 			{ 0xe,	"Mobile Intel(r) Pentium(r) 4" },
5416 			{ 0xf,	"Mobile Intel(r) Celeron(r)" },
5417 			{ 0x11, "Mobile Genuine Intel(r)" },
5418 			{ 0x12, "Intel(r) Celeron(r) M" },
5419 			{ 0x13, "Mobile Intel(r) Celeron(r)" },
5420 			{ 0x14, "Intel(r) Celeron(r)" },
5421 			{ 0x15, "Mobile Genuine Intel(r)" },
5422 			{ 0x16,	"Intel(r) Pentium(r) M" },
5423 			{ 0x17, "Mobile Intel(r) Celeron(r)" }
5424 		};
5425 		uint_t btblmax = sizeof (brand_tbl) / sizeof (brand_tbl[0]);
5426 		uint_t sgn;
5427 
5428 		sgn = (cpi->cpi_family << 8) |
5429 		    (cpi->cpi_model << 4) | cpi->cpi_step;
5430 
5431 		for (i = 0; i < btblmax; i++)
5432 			if (brand_tbl[i].bt_bid == cpi->cpi_brandid)
5433 				break;
5434 		if (i < btblmax) {
5435 			if (sgn == 0x6b1 && cpi->cpi_brandid == 3)
5436 				return ("Intel(r) Celeron(r)");
5437 			if (sgn < 0xf13 && cpi->cpi_brandid == 0xb)
5438 				return ("Intel(r) Xeon(tm) MP");
5439 			if (sgn < 0xf13 && cpi->cpi_brandid == 0xe)
5440 				return ("Intel(r) Xeon(tm)");
5441 			return (brand_tbl[i].bt_str);
5442 		}
5443 	}
5444 
5445 	return (NULL);
5446 }
5447 
5448 static const char *
5449 amd_cpubrand(const struct cpuid_info *cpi)
5450 {
5451 	ASSERT(is_x86_feature(x86_featureset, X86FSET_CPUID));
5452 
5453 	switch (cpi->cpi_family) {
5454 	case 5:
5455 		switch (cpi->cpi_model) {
5456 		case 0:
5457 		case 1:
5458 		case 2:
5459 		case 3:
5460 		case 4:
5461 		case 5:
5462 			return ("AMD-K5(r)");
5463 		case 6:
5464 		case 7:
5465 			return ("AMD-K6(r)");
5466 		case 8:
5467 			return ("AMD-K6(r)-2");
5468 		case 9:
5469 			return ("AMD-K6(r)-III");
5470 		default:
5471 			return ("AMD (family 5)");
5472 		}
5473 	case 6:
5474 		switch (cpi->cpi_model) {
5475 		case 1:
5476 			return ("AMD-K7(tm)");
5477 		case 0:
5478 		case 2:
5479 		case 4:
5480 			return ("AMD Athlon(tm)");
5481 		case 3:
5482 		case 7:
5483 			return ("AMD Duron(tm)");
5484 		case 6:
5485 		case 8:
5486 		case 10:
5487 			/*
5488 			 * Use the L2 cache size to distinguish
5489 			 */
5490 			return ((cpi->cpi_extd[6].cp_ecx >> 16) >= 256 ?
5491 			    "AMD Athlon(tm)" : "AMD Duron(tm)");
5492 		default:
5493 			return ("AMD (family 6)");
5494 		}
5495 	default:
5496 		break;
5497 	}
5498 
5499 	if (cpi->cpi_family == 0xf && cpi->cpi_model == 5 &&
5500 	    cpi->cpi_brandid != 0) {
5501 		switch (BITX(cpi->cpi_brandid, 7, 5)) {
5502 		case 3:
5503 			return ("AMD Opteron(tm) UP 1xx");
5504 		case 4:
5505 			return ("AMD Opteron(tm) DP 2xx");
5506 		case 5:
5507 			return ("AMD Opteron(tm) MP 8xx");
5508 		default:
5509 			return ("AMD Opteron(tm)");
5510 		}
5511 	}
5512 
5513 	return (NULL);
5514 }
5515 
5516 static const char *
5517 cyrix_cpubrand(struct cpuid_info *cpi, uint_t type)
5518 {
5519 	ASSERT(is_x86_feature(x86_featureset, X86FSET_CPUID));
5520 
5521 	switch (type) {
5522 	case X86_TYPE_CYRIX_6x86:
5523 		return ("Cyrix 6x86");
5524 	case X86_TYPE_CYRIX_6x86L:
5525 		return ("Cyrix 6x86L");
5526 	case X86_TYPE_CYRIX_6x86MX:
5527 		return ("Cyrix 6x86MX");
5528 	case X86_TYPE_CYRIX_GXm:
5529 		return ("Cyrix GXm");
5530 	case X86_TYPE_CYRIX_MediaGX:
5531 		return ("Cyrix MediaGX");
5532 	case X86_TYPE_CYRIX_MII:
5533 		return ("Cyrix M2");
5534 	case X86_TYPE_VIA_CYRIX_III:
5535 		return ("VIA Cyrix M3");
5536 	default:
5537 		/*
5538 		 * Have another wild guess ..
5539 		 */
5540 		if (cpi->cpi_family == 4 && cpi->cpi_model == 9)
5541 			return ("Cyrix 5x86");
5542 		else if (cpi->cpi_family == 5) {
5543 			switch (cpi->cpi_model) {
5544 			case 2:
5545 				return ("Cyrix 6x86");	/* Cyrix M1 */
5546 			case 4:
5547 				return ("Cyrix MediaGX");
5548 			default:
5549 				break;
5550 			}
5551 		} else if (cpi->cpi_family == 6) {
5552 			switch (cpi->cpi_model) {
5553 			case 0:
5554 				return ("Cyrix 6x86MX"); /* Cyrix M2? */
5555 			case 5:
5556 			case 6:
5557 			case 7:
5558 			case 8:
5559 			case 9:
5560 				return ("VIA C3");
5561 			default:
5562 				break;
5563 			}
5564 		}
5565 		break;
5566 	}
5567 	return (NULL);
5568 }
5569 
5570 /*
5571  * This only gets called in the case that the CPU extended
5572  * feature brand string (0x80000002, 0x80000003, 0x80000004)
5573  * aren't available, or contain null bytes for some reason.
5574  */
5575 static void
5576 fabricate_brandstr(struct cpuid_info *cpi)
5577 {
5578 	const char *brand = NULL;
5579 
5580 	switch (cpi->cpi_vendor) {
5581 	case X86_VENDOR_Intel:
5582 		brand = intel_cpubrand(cpi);
5583 		break;
5584 	case X86_VENDOR_AMD:
5585 		brand = amd_cpubrand(cpi);
5586 		break;
5587 	case X86_VENDOR_Cyrix:
5588 		brand = cyrix_cpubrand(cpi, x86_type);
5589 		break;
5590 	case X86_VENDOR_NexGen:
5591 		if (cpi->cpi_family == 5 && cpi->cpi_model == 0)
5592 			brand = "NexGen Nx586";
5593 		break;
5594 	case X86_VENDOR_Centaur:
5595 		if (cpi->cpi_family == 5)
5596 			switch (cpi->cpi_model) {
5597 			case 4:
5598 				brand = "Centaur C6";
5599 				break;
5600 			case 8:
5601 				brand = "Centaur C2";
5602 				break;
5603 			case 9:
5604 				brand = "Centaur C3";
5605 				break;
5606 			default:
5607 				break;
5608 			}
5609 		break;
5610 	case X86_VENDOR_Rise:
5611 		if (cpi->cpi_family == 5 &&
5612 		    (cpi->cpi_model == 0 || cpi->cpi_model == 2))
5613 			brand = "Rise mP6";
5614 		break;
5615 	case X86_VENDOR_SiS:
5616 		if (cpi->cpi_family == 5 && cpi->cpi_model == 0)
5617 			brand = "SiS 55x";
5618 		break;
5619 	case X86_VENDOR_TM:
5620 		if (cpi->cpi_family == 5 && cpi->cpi_model == 4)
5621 			brand = "Transmeta Crusoe TM3x00 or TM5x00";
5622 		break;
5623 	case X86_VENDOR_NSC:
5624 	case X86_VENDOR_UMC:
5625 	default:
5626 		break;
5627 	}
5628 	if (brand) {
5629 		(void) strcpy((char *)cpi->cpi_brandstr, brand);
5630 		return;
5631 	}
5632 
5633 	/*
5634 	 * If all else fails ...
5635 	 */
5636 	(void) snprintf(cpi->cpi_brandstr, sizeof (cpi->cpi_brandstr),
5637 	    "%s %d.%d.%d", cpi->cpi_vendorstr, cpi->cpi_family,
5638 	    cpi->cpi_model, cpi->cpi_step);
5639 }
5640 
5641 /*
5642  * This routine is called just after kernel memory allocation
5643  * becomes available on cpu0, and as part of mp_startup() on
5644  * the other cpus.
5645  *
5646  * Fixup the brand string, and collect any information from cpuid
5647  * that requires dynamically allocated storage to represent.
5648  */
5649 
5650 static void
5651 cpuid_pass_dynamic(cpu_t *cpu, void *_arg __unused)
5652 {
5653 	int	i, max, shft, level, size;
5654 	struct cpuid_regs regs;
5655 	struct cpuid_regs *cp;
5656 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
5657 
5658 	/*
5659 	 * Deterministic cache parameters
5660 	 *
5661 	 * Intel uses leaf 0x4 for this, while AMD uses leaf 0x8000001d. The
5662 	 * values that are present are currently defined to be the same. This
5663 	 * means we can use the same logic to parse it as long as we use the
5664 	 * appropriate leaf to get the data. If you're updating this, make sure
5665 	 * you're careful about which vendor supports which aspect.
5666 	 *
5667 	 * Take this opportunity to detect the number of threads sharing the
5668 	 * last level cache, and construct a corresponding cache id. The
5669 	 * respective cpuid_info members are initialized to the default case of
5670 	 * "no last level cache sharing".
5671 	 */
5672 	cpi->cpi_ncpu_shr_last_cache = 1;
5673 	cpi->cpi_last_lvl_cacheid = cpu->cpu_id;
5674 
5675 	if ((cpi->cpi_maxeax >= 4 && cpi->cpi_vendor == X86_VENDOR_Intel) ||
5676 	    ((cpi->cpi_vendor == X86_VENDOR_AMD ||
5677 	    cpi->cpi_vendor == X86_VENDOR_HYGON) &&
5678 	    cpi->cpi_xmaxeax >= CPUID_LEAF_EXT_1d &&
5679 	    is_x86_feature(x86_featureset, X86FSET_TOPOEXT))) {
5680 		uint32_t leaf;
5681 
5682 		if (cpi->cpi_vendor == X86_VENDOR_Intel) {
5683 			leaf = 4;
5684 		} else {
5685 			leaf = CPUID_LEAF_EXT_1d;
5686 		}
5687 
5688 		/*
5689 		 * Find the # of elements (size) returned by the leaf and along
5690 		 * the way detect last level cache sharing details.
5691 		 */
5692 		bzero(&regs, sizeof (regs));
5693 		cp = &regs;
5694 		for (i = 0, max = 0; i < CPI_FN4_ECX_MAX; i++) {
5695 			cp->cp_eax = leaf;
5696 			cp->cp_ecx = i;
5697 
5698 			(void) __cpuid_insn(cp);
5699 
5700 			if (CPI_CACHE_TYPE(cp) == 0)
5701 				break;
5702 			level = CPI_CACHE_LVL(cp);
5703 			if (level > max) {
5704 				max = level;
5705 				cpi->cpi_ncpu_shr_last_cache =
5706 				    CPI_NTHR_SHR_CACHE(cp) + 1;
5707 			}
5708 		}
5709 		cpi->cpi_cache_leaf_size = size = i;
5710 
5711 		/*
5712 		 * Allocate the cpi_cache_leaves array. The first element
5713 		 * references the regs for the corresponding leaf with %ecx set
5714 		 * to 0. This was gathered in cpuid_pass_extended().
5715 		 */
5716 		if (size > 0) {
5717 			cpi->cpi_cache_leaves =
5718 			    kmem_alloc(size * sizeof (cp), KM_SLEEP);
5719 			if (cpi->cpi_vendor == X86_VENDOR_Intel) {
5720 				cpi->cpi_cache_leaves[0] = &cpi->cpi_std[4];
5721 			} else {
5722 				cpi->cpi_cache_leaves[0] = &cpi->cpi_extd[0x1d];
5723 			}
5724 
5725 			/*
5726 			 * Allocate storage to hold the additional regs
5727 			 * for the leaf, %ecx == 1 .. cpi_cache_leaf_size.
5728 			 *
5729 			 * The regs for the leaf, %ecx == 0 has already
5730 			 * been allocated as indicated above.
5731 			 */
5732 			for (i = 1; i < size; i++) {
5733 				cp = cpi->cpi_cache_leaves[i] =
5734 				    kmem_zalloc(sizeof (regs), KM_SLEEP);
5735 				cp->cp_eax = leaf;
5736 				cp->cp_ecx = i;
5737 
5738 				(void) __cpuid_insn(cp);
5739 			}
5740 		}
5741 		/*
5742 		 * Determine the number of bits needed to represent
5743 		 * the number of CPUs sharing the last level cache.
5744 		 *
5745 		 * Shift off that number of bits from the APIC id to
5746 		 * derive the cache id.
5747 		 */
5748 		shft = 0;
5749 		for (i = 1; i < cpi->cpi_ncpu_shr_last_cache; i <<= 1)
5750 			shft++;
5751 		cpi->cpi_last_lvl_cacheid = cpi->cpi_apicid >> shft;
5752 	}
5753 
5754 	/*
5755 	 * Now fixup the brand string
5756 	 */
5757 	if ((cpi->cpi_xmaxeax & CPUID_LEAF_EXT_0) == 0) {
5758 		fabricate_brandstr(cpi);
5759 	} else {
5760 
5761 		/*
5762 		 * If we successfully extracted a brand string from the cpuid
5763 		 * instruction, clean it up by removing leading spaces and
5764 		 * similar junk.
5765 		 */
5766 		if (cpi->cpi_brandstr[0]) {
5767 			size_t maxlen = sizeof (cpi->cpi_brandstr);
5768 			char *src, *dst;
5769 
5770 			dst = src = (char *)cpi->cpi_brandstr;
5771 			src[maxlen - 1] = '\0';
5772 			/*
5773 			 * strip leading spaces
5774 			 */
5775 			while (*src == ' ')
5776 				src++;
5777 			/*
5778 			 * Remove any 'Genuine' or "Authentic" prefixes
5779 			 */
5780 			if (strncmp(src, "Genuine ", 8) == 0)
5781 				src += 8;
5782 			if (strncmp(src, "Authentic ", 10) == 0)
5783 				src += 10;
5784 
5785 			/*
5786 			 * Now do an in-place copy.
5787 			 * Map (R) to (r) and (TM) to (tm).
5788 			 * The era of teletypes is long gone, and there's
5789 			 * -really- no need to shout.
5790 			 */
5791 			while (*src != '\0') {
5792 				if (src[0] == '(') {
5793 					if (strncmp(src + 1, "R)", 2) == 0) {
5794 						(void) strncpy(dst, "(r)", 3);
5795 						src += 3;
5796 						dst += 3;
5797 						continue;
5798 					}
5799 					if (strncmp(src + 1, "TM)", 3) == 0) {
5800 						(void) strncpy(dst, "(tm)", 4);
5801 						src += 4;
5802 						dst += 4;
5803 						continue;
5804 					}
5805 				}
5806 				*dst++ = *src++;
5807 			}
5808 			*dst = '\0';
5809 
5810 			/*
5811 			 * Finally, remove any trailing spaces
5812 			 */
5813 			while (--dst > cpi->cpi_brandstr)
5814 				if (*dst == ' ')
5815 					*dst = '\0';
5816 				else
5817 					break;
5818 		} else
5819 			fabricate_brandstr(cpi);
5820 	}
5821 }
5822 
5823 typedef struct {
5824 	uint32_t avm_av;
5825 	uint32_t avm_feat;
5826 } av_feat_map_t;
5827 
5828 /*
5829  * These arrays are used to map features that we should add based on x86
5830  * features that are present. As a large number depend on kernel features,
5831  * rather than rechecking and clearing CPUID everywhere, we simply map these.
5832  * There is an array of these for each hwcap word. Some features aren't tracked
5833  * in the kernel x86 featureset and that's ok. They will not show up in here.
5834  */
5835 static const av_feat_map_t x86fset_to_av1[] = {
5836 	{ AV_386_CX8, X86FSET_CX8 },
5837 	{ AV_386_SEP, X86FSET_SEP },
5838 	{ AV_386_AMD_SYSC, X86FSET_ASYSC },
5839 	{ AV_386_CMOV, X86FSET_CMOV },
5840 	{ AV_386_FXSR, X86FSET_SSE },
5841 	{ AV_386_SSE, X86FSET_SSE },
5842 	{ AV_386_SSE2, X86FSET_SSE2 },
5843 	{ AV_386_SSE3, X86FSET_SSE3 },
5844 	{ AV_386_CX16, X86FSET_CX16 },
5845 	{ AV_386_TSCP, X86FSET_TSCP },
5846 	{ AV_386_AMD_SSE4A, X86FSET_SSE4A },
5847 	{ AV_386_SSSE3, X86FSET_SSSE3 },
5848 	{ AV_386_SSE4_1, X86FSET_SSE4_1 },
5849 	{ AV_386_SSE4_2, X86FSET_SSE4_2 },
5850 	{ AV_386_AES, X86FSET_AES },
5851 	{ AV_386_PCLMULQDQ, X86FSET_PCLMULQDQ },
5852 	{ AV_386_XSAVE, X86FSET_XSAVE },
5853 	{ AV_386_AVX, X86FSET_AVX },
5854 	{ AV_386_VMX, X86FSET_VMX },
5855 	{ AV_386_AMD_SVM, X86FSET_SVM }
5856 };
5857 
5858 static const av_feat_map_t x86fset_to_av2[] = {
5859 	{ AV_386_2_F16C, X86FSET_F16C },
5860 	{ AV_386_2_RDRAND, X86FSET_RDRAND },
5861 	{ AV_386_2_BMI1, X86FSET_BMI1 },
5862 	{ AV_386_2_BMI2, X86FSET_BMI2 },
5863 	{ AV_386_2_FMA, X86FSET_FMA },
5864 	{ AV_386_2_AVX2, X86FSET_AVX2 },
5865 	{ AV_386_2_ADX, X86FSET_ADX },
5866 	{ AV_386_2_RDSEED, X86FSET_RDSEED },
5867 	{ AV_386_2_AVX512F, X86FSET_AVX512F },
5868 	{ AV_386_2_AVX512DQ, X86FSET_AVX512DQ },
5869 	{ AV_386_2_AVX512IFMA, X86FSET_AVX512FMA },
5870 	{ AV_386_2_AVX512PF, X86FSET_AVX512PF },
5871 	{ AV_386_2_AVX512ER, X86FSET_AVX512ER },
5872 	{ AV_386_2_AVX512CD, X86FSET_AVX512CD },
5873 	{ AV_386_2_AVX512BW, X86FSET_AVX512BW },
5874 	{ AV_386_2_AVX512VL, X86FSET_AVX512VL },
5875 	{ AV_386_2_AVX512VBMI, X86FSET_AVX512VBMI },
5876 	{ AV_386_2_AVX512VPOPCDQ, X86FSET_AVX512VPOPCDQ },
5877 	{ AV_386_2_SHA, X86FSET_SHA },
5878 	{ AV_386_2_FSGSBASE, X86FSET_FSGSBASE },
5879 	{ AV_386_2_CLFLUSHOPT, X86FSET_CLFLUSHOPT },
5880 	{ AV_386_2_CLWB, X86FSET_CLWB },
5881 	{ AV_386_2_MONITORX, X86FSET_MONITORX },
5882 	{ AV_386_2_CLZERO, X86FSET_CLZERO },
5883 	{ AV_386_2_AVX512_VNNI, X86FSET_AVX512VNNI },
5884 	{ AV_386_2_VPCLMULQDQ, X86FSET_VPCLMULQDQ },
5885 	{ AV_386_2_VAES, X86FSET_VAES },
5886 	{ AV_386_2_GFNI, X86FSET_GFNI },
5887 	{ AV_386_2_AVX512_VP2INT, X86FSET_AVX512_VP2INT },
5888 	{ AV_386_2_AVX512_BITALG, X86FSET_AVX512_BITALG }
5889 };
5890 
5891 static const av_feat_map_t x86fset_to_av3[] = {
5892 	{ AV_386_3_AVX512_VBMI2, X86FSET_AVX512_VBMI2 },
5893 	{ AV_386_3_AVX512_BF16, X86FSET_AVX512_BF16 }
5894 };
5895 
5896 /*
5897  * This routine is called out of bind_hwcap() much later in the life
5898  * of the kernel (post_startup()).  The job of this routine is to resolve
5899  * the hardware feature support and kernel support for those features into
5900  * what we're actually going to tell applications via the aux vector.
5901  *
5902  * Most of the aux vector is derived from the x86_featureset array vector where
5903  * a given feature indicates that an aux vector should be plumbed through. This
5904  * allows the kernel to use one tracking mechanism for these based on whether or
5905  * not it has the required hardware support (most often xsave). Most newer
5906  * features are added there in case we need them in the kernel. Otherwise,
5907  * features are evaluated based on looking at the cpuid features that remain. If
5908  * you find yourself wanting to clear out cpuid features for some reason, they
5909  * should instead be driven by the feature set so we have a consistent view.
5910  */
5911 
5912 static void
5913 cpuid_pass_resolve(cpu_t *cpu, void *arg)
5914 {
5915 	uint_t *hwcap_out = (uint_t *)arg;
5916 	struct cpuid_info *cpi;
5917 	uint_t hwcap_flags = 0, hwcap_flags_2 = 0, hwcap_flags_3 = 0;
5918 
5919 	cpi = cpu->cpu_m.mcpu_cpi;
5920 
5921 	for (uint_t i = 0; i < ARRAY_SIZE(x86fset_to_av1); i++) {
5922 		if (is_x86_feature(x86_featureset,
5923 		    x86fset_to_av1[i].avm_feat)) {
5924 			hwcap_flags |= x86fset_to_av1[i].avm_av;
5925 		}
5926 	}
5927 
5928 	for (uint_t i = 0; i < ARRAY_SIZE(x86fset_to_av2); i++) {
5929 		if (is_x86_feature(x86_featureset,
5930 		    x86fset_to_av2[i].avm_feat)) {
5931 			hwcap_flags_2 |= x86fset_to_av2[i].avm_av;
5932 		}
5933 	}
5934 
5935 	for (uint_t i = 0; i < ARRAY_SIZE(x86fset_to_av3); i++) {
5936 		if (is_x86_feature(x86_featureset,
5937 		    x86fset_to_av3[i].avm_feat)) {
5938 			hwcap_flags_3 |= x86fset_to_av3[i].avm_av;
5939 		}
5940 	}
5941 
5942 	/*
5943 	 * From here on out we're working through features that don't have
5944 	 * corresponding kernel feature flags for various reasons that are
5945 	 * mostly just due to the historical implementation.
5946 	 */
5947 	if (cpi->cpi_maxeax >= 1) {
5948 		uint32_t *edx = &cpi->cpi_support[STD_EDX_FEATURES];
5949 		uint32_t *ecx = &cpi->cpi_support[STD_ECX_FEATURES];
5950 
5951 		*edx = CPI_FEATURES_EDX(cpi);
5952 		*ecx = CPI_FEATURES_ECX(cpi);
5953 
5954 		/*
5955 		 * [no explicit support required beyond x87 fp context]
5956 		 */
5957 		if (!fpu_exists)
5958 			*edx &= ~(CPUID_INTC_EDX_FPU | CPUID_INTC_EDX_MMX);
5959 
5960 		/*
5961 		 * Now map the supported feature vector to things that we
5962 		 * think userland will care about.
5963 		 */
5964 		if (*ecx & CPUID_INTC_ECX_MOVBE)
5965 			hwcap_flags |= AV_386_MOVBE;
5966 
5967 		if (*ecx & CPUID_INTC_ECX_POPCNT)
5968 			hwcap_flags |= AV_386_POPCNT;
5969 		if (*edx & CPUID_INTC_EDX_FPU)
5970 			hwcap_flags |= AV_386_FPU;
5971 		if (*edx & CPUID_INTC_EDX_MMX)
5972 			hwcap_flags |= AV_386_MMX;
5973 		if (*edx & CPUID_INTC_EDX_TSC)
5974 			hwcap_flags |= AV_386_TSC;
5975 	}
5976 
5977 	/*
5978 	 * Check a few miscellaneous features.
5979 	 */
5980 	if (cpi->cpi_xmaxeax < 0x80000001)
5981 		goto resolve_done;
5982 
5983 	switch (cpi->cpi_vendor) {
5984 		uint32_t *edx, *ecx;
5985 
5986 	case X86_VENDOR_Intel:
5987 		/*
5988 		 * Seems like Intel duplicated what we necessary
5989 		 * here to make the initial crop of 64-bit OS's work.
5990 		 * Hopefully, those are the only "extended" bits
5991 		 * they'll add.
5992 		 */
5993 		/*FALLTHROUGH*/
5994 
5995 	case X86_VENDOR_AMD:
5996 	case X86_VENDOR_HYGON:
5997 		edx = &cpi->cpi_support[AMD_EDX_FEATURES];
5998 		ecx = &cpi->cpi_support[AMD_ECX_FEATURES];
5999 
6000 		*edx = CPI_FEATURES_XTD_EDX(cpi);
6001 		*ecx = CPI_FEATURES_XTD_ECX(cpi);
6002 
6003 		/*
6004 		 * [no explicit support required beyond
6005 		 * x87 fp context and exception handlers]
6006 		 */
6007 		if (!fpu_exists)
6008 			*edx &= ~(CPUID_AMD_EDX_MMXamd |
6009 			    CPUID_AMD_EDX_3DNow | CPUID_AMD_EDX_3DNowx);
6010 
6011 		/*
6012 		 * Now map the supported feature vector to
6013 		 * things that we think userland will care about.
6014 		 */
6015 		if (*edx & CPUID_AMD_EDX_MMXamd)
6016 			hwcap_flags |= AV_386_AMD_MMX;
6017 		if (*edx & CPUID_AMD_EDX_3DNow)
6018 			hwcap_flags |= AV_386_AMD_3DNow;
6019 		if (*edx & CPUID_AMD_EDX_3DNowx)
6020 			hwcap_flags |= AV_386_AMD_3DNowx;
6021 
6022 		switch (cpi->cpi_vendor) {
6023 		case X86_VENDOR_AMD:
6024 		case X86_VENDOR_HYGON:
6025 			if (*ecx & CPUID_AMD_ECX_AHF64)
6026 				hwcap_flags |= AV_386_AHF;
6027 			if (*ecx & CPUID_AMD_ECX_LZCNT)
6028 				hwcap_flags |= AV_386_AMD_LZCNT;
6029 			break;
6030 
6031 		case X86_VENDOR_Intel:
6032 			if (*ecx & CPUID_AMD_ECX_LZCNT)
6033 				hwcap_flags |= AV_386_AMD_LZCNT;
6034 			/*
6035 			 * Aarrgh.
6036 			 * Intel uses a different bit in the same word.
6037 			 */
6038 			if (*ecx & CPUID_INTC_ECX_AHF64)
6039 				hwcap_flags |= AV_386_AHF;
6040 			break;
6041 		default:
6042 			break;
6043 		}
6044 		break;
6045 
6046 	default:
6047 		break;
6048 	}
6049 
6050 resolve_done:
6051 	if (hwcap_out != NULL) {
6052 		hwcap_out[0] = hwcap_flags;
6053 		hwcap_out[1] = hwcap_flags_2;
6054 		hwcap_out[2] = hwcap_flags_3;
6055 	}
6056 }
6057 
6058 
6059 /*
6060  * Simulate the cpuid instruction using the data we previously
6061  * captured about this CPU.  We try our best to return the truth
6062  * about the hardware, independently of kernel support.
6063  */
6064 uint32_t
6065 cpuid_insn(cpu_t *cpu, struct cpuid_regs *cp)
6066 {
6067 	struct cpuid_info *cpi;
6068 	struct cpuid_regs *xcp;
6069 
6070 	if (cpu == NULL)
6071 		cpu = CPU;
6072 	cpi = cpu->cpu_m.mcpu_cpi;
6073 
6074 	ASSERT(cpuid_checkpass(cpu, CPUID_PASS_DYNAMIC));
6075 
6076 	/*
6077 	 * CPUID data is cached in two separate places: cpi_std for standard
6078 	 * CPUID leaves , and cpi_extd for extended CPUID leaves.
6079 	 */
6080 	if (cp->cp_eax <= cpi->cpi_maxeax && cp->cp_eax < NMAX_CPI_STD) {
6081 		xcp = &cpi->cpi_std[cp->cp_eax];
6082 	} else if (cp->cp_eax >= CPUID_LEAF_EXT_0 &&
6083 	    cp->cp_eax <= cpi->cpi_xmaxeax &&
6084 	    cp->cp_eax < CPUID_LEAF_EXT_0 + NMAX_CPI_EXTD) {
6085 		xcp = &cpi->cpi_extd[cp->cp_eax - CPUID_LEAF_EXT_0];
6086 	} else {
6087 		/*
6088 		 * The caller is asking for data from an input parameter which
6089 		 * the kernel has not cached.  In this case we go fetch from
6090 		 * the hardware and return the data directly to the user.
6091 		 */
6092 		return (__cpuid_insn(cp));
6093 	}
6094 
6095 	cp->cp_eax = xcp->cp_eax;
6096 	cp->cp_ebx = xcp->cp_ebx;
6097 	cp->cp_ecx = xcp->cp_ecx;
6098 	cp->cp_edx = xcp->cp_edx;
6099 	return (cp->cp_eax);
6100 }
6101 
6102 boolean_t
6103 cpuid_checkpass(const cpu_t *const cpu, const cpuid_pass_t pass)
6104 {
6105 	return (cpu != NULL && cpu->cpu_m.mcpu_cpi != NULL &&
6106 	    cpu->cpu_m.mcpu_cpi->cpi_pass >= pass);
6107 }
6108 
6109 int
6110 cpuid_getbrandstr(cpu_t *cpu, char *s, size_t n)
6111 {
6112 	ASSERT(cpuid_checkpass(cpu, CPUID_PASS_DYNAMIC));
6113 
6114 	return (snprintf(s, n, "%s", cpu->cpu_m.mcpu_cpi->cpi_brandstr));
6115 }
6116 
6117 int
6118 cpuid_is_cmt(cpu_t *cpu)
6119 {
6120 	if (cpu == NULL)
6121 		cpu = CPU;
6122 
6123 	ASSERT(cpuid_checkpass(cpu, CPUID_PASS_BASIC));
6124 
6125 	return (cpu->cpu_m.mcpu_cpi->cpi_chipid >= 0);
6126 }
6127 
6128 /*
6129  * AMD and Intel both implement the 64-bit variant of the syscall
6130  * instruction (syscallq), so if there's -any- support for syscall,
6131  * cpuid currently says "yes, we support this".
6132  *
6133  * However, Intel decided to -not- implement the 32-bit variant of the
6134  * syscall instruction, so we provide a predicate to allow our caller
6135  * to test that subtlety here.
6136  *
6137  * XXPV	Currently, 32-bit syscall instructions don't work via the hypervisor,
6138  *	even in the case where the hardware would in fact support it.
6139  */
6140 /*ARGSUSED*/
6141 int
6142 cpuid_syscall32_insn(cpu_t *cpu)
6143 {
6144 	ASSERT(cpuid_checkpass((cpu == NULL ? CPU : cpu), CPUID_PASS_BASIC));
6145 
6146 #if !defined(__xpv)
6147 	if (cpu == NULL)
6148 		cpu = CPU;
6149 
6150 	/*CSTYLED*/
6151 	{
6152 		struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
6153 
6154 		if ((cpi->cpi_vendor == X86_VENDOR_AMD ||
6155 		    cpi->cpi_vendor == X86_VENDOR_HYGON) &&
6156 		    cpi->cpi_xmaxeax >= 0x80000001 &&
6157 		    (CPI_FEATURES_XTD_EDX(cpi) & CPUID_AMD_EDX_SYSC))
6158 			return (1);
6159 	}
6160 #endif
6161 	return (0);
6162 }
6163 
6164 int
6165 cpuid_getidstr(cpu_t *cpu, char *s, size_t n)
6166 {
6167 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
6168 
6169 	static const char fmt[] =
6170 	    "x86 (%s %X family %d model %d step %d clock %d MHz)";
6171 	static const char fmt_ht[] =
6172 	    "x86 (chipid 0x%x %s %X family %d model %d step %d clock %d MHz)";
6173 
6174 	ASSERT(cpuid_checkpass(cpu, CPUID_PASS_BASIC));
6175 
6176 	if (cpuid_is_cmt(cpu))
6177 		return (snprintf(s, n, fmt_ht, cpi->cpi_chipid,
6178 		    cpi->cpi_vendorstr, cpi->cpi_std[1].cp_eax,
6179 		    cpi->cpi_family, cpi->cpi_model,
6180 		    cpi->cpi_step, cpu->cpu_type_info.pi_clock));
6181 	return (snprintf(s, n, fmt,
6182 	    cpi->cpi_vendorstr, cpi->cpi_std[1].cp_eax,
6183 	    cpi->cpi_family, cpi->cpi_model,
6184 	    cpi->cpi_step, cpu->cpu_type_info.pi_clock));
6185 }
6186 
6187 const char *
6188 cpuid_getvendorstr(cpu_t *cpu)
6189 {
6190 	ASSERT(cpuid_checkpass(cpu, CPUID_PASS_IDENT));
6191 	return ((const char *)cpu->cpu_m.mcpu_cpi->cpi_vendorstr);
6192 }
6193 
6194 uint_t
6195 cpuid_getvendor(cpu_t *cpu)
6196 {
6197 	ASSERT(cpuid_checkpass(cpu, CPUID_PASS_IDENT));
6198 	return (cpu->cpu_m.mcpu_cpi->cpi_vendor);
6199 }
6200 
6201 uint_t
6202 cpuid_getfamily(cpu_t *cpu)
6203 {
6204 	ASSERT(cpuid_checkpass(cpu, CPUID_PASS_IDENT));
6205 	return (cpu->cpu_m.mcpu_cpi->cpi_family);
6206 }
6207 
6208 uint_t
6209 cpuid_getmodel(cpu_t *cpu)
6210 {
6211 	ASSERT(cpuid_checkpass(cpu, CPUID_PASS_IDENT));
6212 	return (cpu->cpu_m.mcpu_cpi->cpi_model);
6213 }
6214 
6215 uint_t
6216 cpuid_get_ncpu_per_chip(cpu_t *cpu)
6217 {
6218 	ASSERT(cpuid_checkpass(cpu, CPUID_PASS_BASIC));
6219 	return (cpu->cpu_m.mcpu_cpi->cpi_ncpu_per_chip);
6220 }
6221 
6222 uint_t
6223 cpuid_get_ncore_per_chip(cpu_t *cpu)
6224 {
6225 	ASSERT(cpuid_checkpass(cpu, CPUID_PASS_BASIC));
6226 	return (cpu->cpu_m.mcpu_cpi->cpi_ncore_per_chip);
6227 }
6228 
6229 uint_t
6230 cpuid_get_ncpu_sharing_last_cache(cpu_t *cpu)
6231 {
6232 	ASSERT(cpuid_checkpass(cpu, CPUID_PASS_EXTENDED));
6233 	return (cpu->cpu_m.mcpu_cpi->cpi_ncpu_shr_last_cache);
6234 }
6235 
6236 id_t
6237 cpuid_get_last_lvl_cacheid(cpu_t *cpu)
6238 {
6239 	ASSERT(cpuid_checkpass(cpu, CPUID_PASS_EXTENDED));
6240 	return (cpu->cpu_m.mcpu_cpi->cpi_last_lvl_cacheid);
6241 }
6242 
6243 uint_t
6244 cpuid_getstep(cpu_t *cpu)
6245 {
6246 	ASSERT(cpuid_checkpass(cpu, CPUID_PASS_IDENT));
6247 	return (cpu->cpu_m.mcpu_cpi->cpi_step);
6248 }
6249 
6250 uint_t
6251 cpuid_getsig(struct cpu *cpu)
6252 {
6253 	ASSERT(cpuid_checkpass(cpu, CPUID_PASS_BASIC));
6254 	return (cpu->cpu_m.mcpu_cpi->cpi_std[1].cp_eax);
6255 }
6256 
6257 x86_chiprev_t
6258 cpuid_getchiprev(struct cpu *cpu)
6259 {
6260 	ASSERT(cpuid_checkpass(cpu, CPUID_PASS_IDENT));
6261 	return (cpu->cpu_m.mcpu_cpi->cpi_chiprev);
6262 }
6263 
6264 const char *
6265 cpuid_getchiprevstr(struct cpu *cpu)
6266 {
6267 	ASSERT(cpuid_checkpass(cpu, CPUID_PASS_IDENT));
6268 	return (cpu->cpu_m.mcpu_cpi->cpi_chiprevstr);
6269 }
6270 
6271 uint32_t
6272 cpuid_getsockettype(struct cpu *cpu)
6273 {
6274 	ASSERT(cpuid_checkpass(cpu, CPUID_PASS_IDENT));
6275 	return (cpu->cpu_m.mcpu_cpi->cpi_socket);
6276 }
6277 
6278 const char *
6279 cpuid_getsocketstr(cpu_t *cpu)
6280 {
6281 	static const char *socketstr = NULL;
6282 	struct cpuid_info *cpi;
6283 
6284 	ASSERT(cpuid_checkpass(cpu, CPUID_PASS_IDENT));
6285 	cpi = cpu->cpu_m.mcpu_cpi;
6286 
6287 	/* Assume that socket types are the same across the system */
6288 	if (socketstr == NULL)
6289 		socketstr = _cpuid_sktstr(cpi->cpi_vendor, cpi->cpi_family,
6290 		    cpi->cpi_model, cpi->cpi_step);
6291 
6292 
6293 	return (socketstr);
6294 }
6295 
6296 x86_uarchrev_t
6297 cpuid_getuarchrev(cpu_t *cpu)
6298 {
6299 	return (cpu->cpu_m.mcpu_cpi->cpi_uarchrev);
6300 }
6301 
6302 int
6303 cpuid_get_chipid(cpu_t *cpu)
6304 {
6305 	ASSERT(cpuid_checkpass(cpu, CPUID_PASS_BASIC));
6306 
6307 	if (cpuid_is_cmt(cpu))
6308 		return (cpu->cpu_m.mcpu_cpi->cpi_chipid);
6309 	return (cpu->cpu_id);
6310 }
6311 
6312 id_t
6313 cpuid_get_coreid(cpu_t *cpu)
6314 {
6315 	ASSERT(cpuid_checkpass(cpu, CPUID_PASS_BASIC));
6316 	return (cpu->cpu_m.mcpu_cpi->cpi_coreid);
6317 }
6318 
6319 int
6320 cpuid_get_pkgcoreid(cpu_t *cpu)
6321 {
6322 	ASSERT(cpuid_checkpass(cpu, CPUID_PASS_BASIC));
6323 	return (cpu->cpu_m.mcpu_cpi->cpi_pkgcoreid);
6324 }
6325 
6326 int
6327 cpuid_get_clogid(cpu_t *cpu)
6328 {
6329 	ASSERT(cpuid_checkpass(cpu, CPUID_PASS_BASIC));
6330 	return (cpu->cpu_m.mcpu_cpi->cpi_clogid);
6331 }
6332 
6333 int
6334 cpuid_get_cacheid(cpu_t *cpu)
6335 {
6336 	ASSERT(cpuid_checkpass(cpu, CPUID_PASS_BASIC));
6337 	return (cpu->cpu_m.mcpu_cpi->cpi_last_lvl_cacheid);
6338 }
6339 
6340 uint_t
6341 cpuid_get_procnodeid(cpu_t *cpu)
6342 {
6343 	ASSERT(cpuid_checkpass(cpu, CPUID_PASS_BASIC));
6344 	return (cpu->cpu_m.mcpu_cpi->cpi_procnodeid);
6345 }
6346 
6347 uint_t
6348 cpuid_get_procnodes_per_pkg(cpu_t *cpu)
6349 {
6350 	ASSERT(cpuid_checkpass(cpu, CPUID_PASS_BASIC));
6351 	return (cpu->cpu_m.mcpu_cpi->cpi_procnodes_per_pkg);
6352 }
6353 
6354 uint_t
6355 cpuid_get_compunitid(cpu_t *cpu)
6356 {
6357 	ASSERT(cpuid_checkpass(cpu, CPUID_PASS_BASIC));
6358 	return (cpu->cpu_m.mcpu_cpi->cpi_compunitid);
6359 }
6360 
6361 uint_t
6362 cpuid_get_cores_per_compunit(cpu_t *cpu)
6363 {
6364 	ASSERT(cpuid_checkpass(cpu, CPUID_PASS_BASIC));
6365 	return (cpu->cpu_m.mcpu_cpi->cpi_cores_per_compunit);
6366 }
6367 
6368 uint32_t
6369 cpuid_get_apicid(cpu_t *cpu)
6370 {
6371 	ASSERT(cpuid_checkpass(cpu, CPUID_PASS_BASIC));
6372 	if (cpu->cpu_m.mcpu_cpi->cpi_maxeax < 1) {
6373 		return (UINT32_MAX);
6374 	} else {
6375 		return (cpu->cpu_m.mcpu_cpi->cpi_apicid);
6376 	}
6377 }
6378 
6379 void
6380 cpuid_get_addrsize(cpu_t *cpu, uint_t *pabits, uint_t *vabits)
6381 {
6382 	struct cpuid_info *cpi;
6383 
6384 	if (cpu == NULL)
6385 		cpu = CPU;
6386 	cpi = cpu->cpu_m.mcpu_cpi;
6387 
6388 	ASSERT(cpuid_checkpass(cpu, CPUID_PASS_BASIC));
6389 
6390 	if (pabits)
6391 		*pabits = cpi->cpi_pabits;
6392 	if (vabits)
6393 		*vabits = cpi->cpi_vabits;
6394 }
6395 
6396 size_t
6397 cpuid_get_xsave_size(void)
6398 {
6399 	return (MAX(cpuid_info0.cpi_xsave.xsav_max_size,
6400 	    sizeof (struct xsave_state)));
6401 }
6402 
6403 /*
6404  * Export information about known offsets to the kernel. We only care about
6405  * things we have actually enabled support for in %xcr0.
6406  */
6407 void
6408 cpuid_get_xsave_info(uint64_t bit, size_t *sizep, size_t *offp)
6409 {
6410 	size_t size, off;
6411 
6412 	VERIFY3U(bit & xsave_bv_all, !=, 0);
6413 
6414 	if (sizep == NULL)
6415 		sizep = &size;
6416 	if (offp == NULL)
6417 		offp = &off;
6418 
6419 	switch (bit) {
6420 	case XFEATURE_LEGACY_FP:
6421 	case XFEATURE_SSE:
6422 		*sizep = sizeof (struct fxsave_state);
6423 		*offp = 0;
6424 		break;
6425 	case XFEATURE_AVX:
6426 		*sizep = cpuid_info0.cpi_xsave.ymm_size;
6427 		*offp = cpuid_info0.cpi_xsave.ymm_offset;
6428 		break;
6429 	case XFEATURE_AVX512_OPMASK:
6430 		*sizep = cpuid_info0.cpi_xsave.opmask_size;
6431 		*offp = cpuid_info0.cpi_xsave.opmask_offset;
6432 		break;
6433 	case XFEATURE_AVX512_ZMM:
6434 		*sizep = cpuid_info0.cpi_xsave.zmmlo_size;
6435 		*offp = cpuid_info0.cpi_xsave.zmmlo_offset;
6436 		break;
6437 	case XFEATURE_AVX512_HI_ZMM:
6438 		*sizep = cpuid_info0.cpi_xsave.zmmhi_size;
6439 		*offp = cpuid_info0.cpi_xsave.zmmhi_offset;
6440 		break;
6441 	default:
6442 		panic("asked for unsupported xsave feature: 0x%lx", bit);
6443 	}
6444 }
6445 
6446 /*
6447  * Return true if the CPUs on this system require 'pointer clearing' for the
6448  * floating point error pointer exception handling. In the past, this has been
6449  * true for all AMD K7 & K8 CPUs, although newer AMD CPUs have been changed to
6450  * behave the same as Intel. This is checked via the CPUID_AMD_EBX_ERR_PTR_ZERO
6451  * feature bit and is reflected in the cpi_fp_amd_save member.
6452  */
6453 boolean_t
6454 cpuid_need_fp_excp_handling(void)
6455 {
6456 	return (cpuid_info0.cpi_vendor == X86_VENDOR_AMD &&
6457 	    cpuid_info0.cpi_fp_amd_save != 0);
6458 }
6459 
6460 /*
6461  * Returns the number of data TLB entries for a corresponding
6462  * pagesize.  If it can't be computed, or isn't known, the
6463  * routine returns zero.  If you ask about an architecturally
6464  * impossible pagesize, the routine will panic (so that the
6465  * hat implementor knows that things are inconsistent.)
6466  */
6467 uint_t
6468 cpuid_get_dtlb_nent(cpu_t *cpu, size_t pagesize)
6469 {
6470 	struct cpuid_info *cpi;
6471 	uint_t dtlb_nent = 0;
6472 
6473 	if (cpu == NULL)
6474 		cpu = CPU;
6475 	cpi = cpu->cpu_m.mcpu_cpi;
6476 
6477 	ASSERT(cpuid_checkpass(cpu, CPUID_PASS_BASIC));
6478 
6479 	/*
6480 	 * Check the L2 TLB info
6481 	 */
6482 	if (cpi->cpi_xmaxeax >= 0x80000006) {
6483 		struct cpuid_regs *cp = &cpi->cpi_extd[6];
6484 
6485 		switch (pagesize) {
6486 
6487 		case 4 * 1024:
6488 			/*
6489 			 * All zero in the top 16 bits of the register
6490 			 * indicates a unified TLB. Size is in low 16 bits.
6491 			 */
6492 			if ((cp->cp_ebx & 0xffff0000) == 0)
6493 				dtlb_nent = cp->cp_ebx & 0x0000ffff;
6494 			else
6495 				dtlb_nent = BITX(cp->cp_ebx, 27, 16);
6496 			break;
6497 
6498 		case 2 * 1024 * 1024:
6499 			if ((cp->cp_eax & 0xffff0000) == 0)
6500 				dtlb_nent = cp->cp_eax & 0x0000ffff;
6501 			else
6502 				dtlb_nent = BITX(cp->cp_eax, 27, 16);
6503 			break;
6504 
6505 		default:
6506 			panic("unknown L2 pagesize");
6507 			/*NOTREACHED*/
6508 		}
6509 	}
6510 
6511 	if (dtlb_nent != 0)
6512 		return (dtlb_nent);
6513 
6514 	/*
6515 	 * No L2 TLB support for this size, try L1.
6516 	 */
6517 	if (cpi->cpi_xmaxeax >= 0x80000005) {
6518 		struct cpuid_regs *cp = &cpi->cpi_extd[5];
6519 
6520 		switch (pagesize) {
6521 		case 4 * 1024:
6522 			dtlb_nent = BITX(cp->cp_ebx, 23, 16);
6523 			break;
6524 		case 2 * 1024 * 1024:
6525 			dtlb_nent = BITX(cp->cp_eax, 23, 16);
6526 			break;
6527 		default:
6528 			panic("unknown L1 d-TLB pagesize");
6529 			/*NOTREACHED*/
6530 		}
6531 	}
6532 
6533 	return (dtlb_nent);
6534 }
6535 
6536 /*
6537  * Return 0 if the erratum is not present or not applicable, positive
6538  * if it is, and negative if the status of the erratum is unknown.
6539  *
6540  * See "Revision Guide for AMD Athlon(tm) 64 and AMD Opteron(tm)
6541  * Processors" #25759, Rev 3.57, August 2005
6542  */
6543 int
6544 cpuid_opteron_erratum(cpu_t *cpu, uint_t erratum)
6545 {
6546 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
6547 	uint_t eax;
6548 
6549 	/*
6550 	 * Bail out if this CPU isn't an AMD CPU, or if it's
6551 	 * a legacy (32-bit) AMD CPU.
6552 	 */
6553 	if (cpi->cpi_vendor != X86_VENDOR_AMD ||
6554 	    cpi->cpi_family == 4 || cpi->cpi_family == 5 ||
6555 	    cpi->cpi_family == 6) {
6556 		return (0);
6557 	}
6558 
6559 	eax = cpi->cpi_std[1].cp_eax;
6560 
6561 #define	SH_B0(eax)	(eax == 0xf40 || eax == 0xf50)
6562 #define	SH_B3(eax)	(eax == 0xf51)
6563 #define	B(eax)		(SH_B0(eax) || SH_B3(eax))
6564 
6565 #define	SH_C0(eax)	(eax == 0xf48 || eax == 0xf58)
6566 
6567 #define	SH_CG(eax)	(eax == 0xf4a || eax == 0xf5a || eax == 0xf7a)
6568 #define	DH_CG(eax)	(eax == 0xfc0 || eax == 0xfe0 || eax == 0xff0)
6569 #define	CH_CG(eax)	(eax == 0xf82 || eax == 0xfb2)
6570 #define	CG(eax)		(SH_CG(eax) || DH_CG(eax) || CH_CG(eax))
6571 
6572 #define	SH_D0(eax)	(eax == 0x10f40 || eax == 0x10f50 || eax == 0x10f70)
6573 #define	DH_D0(eax)	(eax == 0x10fc0 || eax == 0x10ff0)
6574 #define	CH_D0(eax)	(eax == 0x10f80 || eax == 0x10fb0)
6575 #define	D0(eax)		(SH_D0(eax) || DH_D0(eax) || CH_D0(eax))
6576 
6577 #define	SH_E0(eax)	(eax == 0x20f50 || eax == 0x20f40 || eax == 0x20f70)
6578 #define	JH_E1(eax)	(eax == 0x20f10)	/* JH8_E0 had 0x20f30 */
6579 #define	DH_E3(eax)	(eax == 0x20fc0 || eax == 0x20ff0)
6580 #define	SH_E4(eax)	(eax == 0x20f51 || eax == 0x20f71)
6581 #define	BH_E4(eax)	(eax == 0x20fb1)
6582 #define	SH_E5(eax)	(eax == 0x20f42)
6583 #define	DH_E6(eax)	(eax == 0x20ff2 || eax == 0x20fc2)
6584 #define	JH_E6(eax)	(eax == 0x20f12 || eax == 0x20f32)
6585 #define	EX(eax)		(SH_E0(eax) || JH_E1(eax) || DH_E3(eax) || \
6586 			    SH_E4(eax) || BH_E4(eax) || SH_E5(eax) || \
6587 			    DH_E6(eax) || JH_E6(eax))
6588 
6589 #define	DR_AX(eax)	(eax == 0x100f00 || eax == 0x100f01 || eax == 0x100f02)
6590 #define	DR_B0(eax)	(eax == 0x100f20)
6591 #define	DR_B1(eax)	(eax == 0x100f21)
6592 #define	DR_BA(eax)	(eax == 0x100f2a)
6593 #define	DR_B2(eax)	(eax == 0x100f22)
6594 #define	DR_B3(eax)	(eax == 0x100f23)
6595 #define	RB_C0(eax)	(eax == 0x100f40)
6596 
6597 	switch (erratum) {
6598 	case 1:
6599 		return (cpi->cpi_family < 0x10);
6600 	case 51:	/* what does the asterisk mean? */
6601 		return (B(eax) || SH_C0(eax) || CG(eax));
6602 	case 52:
6603 		return (B(eax));
6604 	case 57:
6605 		return (cpi->cpi_family <= 0x11);
6606 	case 58:
6607 		return (B(eax));
6608 	case 60:
6609 		return (cpi->cpi_family <= 0x11);
6610 	case 61:
6611 	case 62:
6612 	case 63:
6613 	case 64:
6614 	case 65:
6615 	case 66:
6616 	case 68:
6617 	case 69:
6618 	case 70:
6619 	case 71:
6620 		return (B(eax));
6621 	case 72:
6622 		return (SH_B0(eax));
6623 	case 74:
6624 		return (B(eax));
6625 	case 75:
6626 		return (cpi->cpi_family < 0x10);
6627 	case 76:
6628 		return (B(eax));
6629 	case 77:
6630 		return (cpi->cpi_family <= 0x11);
6631 	case 78:
6632 		return (B(eax) || SH_C0(eax));
6633 	case 79:
6634 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
6635 	case 80:
6636 	case 81:
6637 	case 82:
6638 		return (B(eax));
6639 	case 83:
6640 		return (B(eax) || SH_C0(eax) || CG(eax));
6641 	case 85:
6642 		return (cpi->cpi_family < 0x10);
6643 	case 86:
6644 		return (SH_C0(eax) || CG(eax));
6645 	case 88:
6646 		return (B(eax) || SH_C0(eax));
6647 	case 89:
6648 		return (cpi->cpi_family < 0x10);
6649 	case 90:
6650 		return (B(eax) || SH_C0(eax) || CG(eax));
6651 	case 91:
6652 	case 92:
6653 		return (B(eax) || SH_C0(eax));
6654 	case 93:
6655 		return (SH_C0(eax));
6656 	case 94:
6657 		return (B(eax) || SH_C0(eax) || CG(eax));
6658 	case 95:
6659 		return (B(eax) || SH_C0(eax));
6660 	case 96:
6661 		return (B(eax) || SH_C0(eax) || CG(eax));
6662 	case 97:
6663 	case 98:
6664 		return (SH_C0(eax) || CG(eax));
6665 	case 99:
6666 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
6667 	case 100:
6668 		return (B(eax) || SH_C0(eax));
6669 	case 101:
6670 	case 103:
6671 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
6672 	case 104:
6673 		return (SH_C0(eax) || CG(eax) || D0(eax));
6674 	case 105:
6675 	case 106:
6676 	case 107:
6677 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
6678 	case 108:
6679 		return (DH_CG(eax));
6680 	case 109:
6681 		return (SH_C0(eax) || CG(eax) || D0(eax));
6682 	case 110:
6683 		return (D0(eax) || EX(eax));
6684 	case 111:
6685 		return (CG(eax));
6686 	case 112:
6687 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
6688 	case 113:
6689 		return (eax == 0x20fc0);
6690 	case 114:
6691 		return (SH_E0(eax) || JH_E1(eax) || DH_E3(eax));
6692 	case 115:
6693 		return (SH_E0(eax) || JH_E1(eax));
6694 	case 116:
6695 		return (SH_E0(eax) || JH_E1(eax) || DH_E3(eax));
6696 	case 117:
6697 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
6698 	case 118:
6699 		return (SH_E0(eax) || JH_E1(eax) || SH_E4(eax) || BH_E4(eax) ||
6700 		    JH_E6(eax));
6701 	case 121:
6702 		return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
6703 	case 122:
6704 		return (cpi->cpi_family < 0x10 || cpi->cpi_family == 0x11);
6705 	case 123:
6706 		return (JH_E1(eax) || BH_E4(eax) || JH_E6(eax));
6707 	case 131:
6708 		return (cpi->cpi_family < 0x10);
6709 	case 6336786:
6710 
6711 		/*
6712 		 * Test for AdvPowerMgmtInfo.TscPStateInvariant
6713 		 * if this is a K8 family or newer processor. We're testing for
6714 		 * this 'erratum' to determine whether or not we have a constant
6715 		 * TSC.
6716 		 *
6717 		 * Our current fix for this is to disable the C1-Clock ramping.
6718 		 * However, this doesn't work on newer processor families nor
6719 		 * does it work when virtualized as those devices don't exist.
6720 		 */
6721 		if (cpi->cpi_family >= 0x12 || get_hwenv() != HW_NATIVE) {
6722 			return (0);
6723 		}
6724 
6725 		if (CPI_FAMILY(cpi) == 0xf) {
6726 			struct cpuid_regs regs;
6727 			regs.cp_eax = 0x80000007;
6728 			(void) __cpuid_insn(&regs);
6729 			return (!(regs.cp_edx & 0x100));
6730 		}
6731 		return (0);
6732 	case 147:
6733 		/*
6734 		 * This erratum (K8 #147) is not present on family 10 and newer.
6735 		 */
6736 		if (cpi->cpi_family >= 0x10) {
6737 			return (0);
6738 		}
6739 		return (((((eax >> 12) & 0xff00) + (eax & 0xf00)) |
6740 		    (((eax >> 4) & 0xf) | ((eax >> 12) & 0xf0))) < 0xf40);
6741 
6742 	case 6671130:
6743 		/*
6744 		 * check for processors (pre-Shanghai) that do not provide
6745 		 * optimal management of 1gb ptes in its tlb.
6746 		 */
6747 		return (cpi->cpi_family == 0x10 && cpi->cpi_model < 4);
6748 
6749 	case 298:
6750 		return (DR_AX(eax) || DR_B0(eax) || DR_B1(eax) || DR_BA(eax) ||
6751 		    DR_B2(eax) || RB_C0(eax));
6752 
6753 	case 721:
6754 		return (cpi->cpi_family == 0x10 || cpi->cpi_family == 0x12);
6755 
6756 	default:
6757 		return (-1);
6758 
6759 	}
6760 }
6761 
6762 /*
6763  * Determine if specified erratum is present via OSVW (OS Visible Workaround).
6764  * Return 1 if erratum is present, 0 if not present and -1 if indeterminate.
6765  */
6766 int
6767 osvw_opteron_erratum(cpu_t *cpu, uint_t erratum)
6768 {
6769 	struct cpuid_info	*cpi;
6770 	uint_t			osvwid;
6771 	static int		osvwfeature = -1;
6772 	uint64_t		osvwlength;
6773 
6774 
6775 	cpi = cpu->cpu_m.mcpu_cpi;
6776 
6777 	/* confirm OSVW supported */
6778 	if (osvwfeature == -1) {
6779 		osvwfeature = cpi->cpi_extd[1].cp_ecx & CPUID_AMD_ECX_OSVW;
6780 	} else {
6781 		/* assert that osvw feature setting is consistent on all cpus */
6782 		ASSERT(osvwfeature ==
6783 		    (cpi->cpi_extd[1].cp_ecx & CPUID_AMD_ECX_OSVW));
6784 	}
6785 	if (!osvwfeature)
6786 		return (-1);
6787 
6788 	osvwlength = rdmsr(MSR_AMD_OSVW_ID_LEN) & OSVW_ID_LEN_MASK;
6789 
6790 	switch (erratum) {
6791 	case 298:	/* osvwid is 0 */
6792 		osvwid = 0;
6793 		if (osvwlength <= (uint64_t)osvwid) {
6794 			/* osvwid 0 is unknown */
6795 			return (-1);
6796 		}
6797 
6798 		/*
6799 		 * Check the OSVW STATUS MSR to determine the state
6800 		 * of the erratum where:
6801 		 *   0 - fixed by HW
6802 		 *   1 - BIOS has applied the workaround when BIOS
6803 		 *   workaround is available. (Or for other errata,
6804 		 *   OS workaround is required.)
6805 		 * For a value of 1, caller will confirm that the
6806 		 * erratum 298 workaround has indeed been applied by BIOS.
6807 		 *
6808 		 * A 1 may be set in cpus that have a HW fix
6809 		 * in a mixed cpu system. Regarding erratum 298:
6810 		 *   In a multiprocessor platform, the workaround above
6811 		 *   should be applied to all processors regardless of
6812 		 *   silicon revision when an affected processor is
6813 		 *   present.
6814 		 */
6815 
6816 		return (rdmsr(MSR_AMD_OSVW_STATUS +
6817 		    (osvwid / OSVW_ID_CNT_PER_MSR)) &
6818 		    (1ULL << (osvwid % OSVW_ID_CNT_PER_MSR)));
6819 
6820 	default:
6821 		return (-1);
6822 	}
6823 }
6824 
6825 static const char assoc_str[] = "associativity";
6826 static const char line_str[] = "line-size";
6827 static const char size_str[] = "size";
6828 
6829 static void
6830 add_cache_prop(dev_info_t *devi, const char *label, const char *type,
6831     uint32_t val)
6832 {
6833 	char buf[128];
6834 
6835 	/*
6836 	 * ndi_prop_update_int() is used because it is desirable for
6837 	 * DDI_PROP_HW_DEF and DDI_PROP_DONTSLEEP to be set.
6838 	 */
6839 	if (snprintf(buf, sizeof (buf), "%s-%s", label, type) < sizeof (buf))
6840 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, devi, buf, val);
6841 }
6842 
6843 /*
6844  * Intel-style cache/tlb description
6845  *
6846  * Standard cpuid level 2 gives a randomly ordered
6847  * selection of tags that index into a table that describes
6848  * cache and tlb properties.
6849  */
6850 
6851 static const char l1_icache_str[] = "l1-icache";
6852 static const char l1_dcache_str[] = "l1-dcache";
6853 static const char l2_cache_str[] = "l2-cache";
6854 static const char l3_cache_str[] = "l3-cache";
6855 static const char itlb4k_str[] = "itlb-4K";
6856 static const char dtlb4k_str[] = "dtlb-4K";
6857 static const char itlb2M_str[] = "itlb-2M";
6858 static const char itlb4M_str[] = "itlb-4M";
6859 static const char dtlb4M_str[] = "dtlb-4M";
6860 static const char dtlb24_str[] = "dtlb0-2M-4M";
6861 static const char itlb424_str[] = "itlb-4K-2M-4M";
6862 static const char itlb24_str[] = "itlb-2M-4M";
6863 static const char dtlb44_str[] = "dtlb-4K-4M";
6864 static const char sl1_dcache_str[] = "sectored-l1-dcache";
6865 static const char sl2_cache_str[] = "sectored-l2-cache";
6866 static const char itrace_str[] = "itrace-cache";
6867 static const char sl3_cache_str[] = "sectored-l3-cache";
6868 static const char sh_l2_tlb4k_str[] = "shared-l2-tlb-4k";
6869 
6870 static const struct cachetab {
6871 	uint8_t		ct_code;
6872 	uint8_t		ct_assoc;
6873 	uint16_t	ct_line_size;
6874 	size_t		ct_size;
6875 	const char	*ct_label;
6876 } intel_ctab[] = {
6877 	/*
6878 	 * maintain descending order!
6879 	 *
6880 	 * Codes ignored - Reason
6881 	 * ----------------------
6882 	 * 40H - intel_cpuid_4_cache_info() disambiguates l2/l3 cache
6883 	 * f0H/f1H - Currently we do not interpret prefetch size by design
6884 	 */
6885 	{ 0xe4, 16, 64, 8*1024*1024, l3_cache_str},
6886 	{ 0xe3, 16, 64, 4*1024*1024, l3_cache_str},
6887 	{ 0xe2, 16, 64, 2*1024*1024, l3_cache_str},
6888 	{ 0xde, 12, 64, 6*1024*1024, l3_cache_str},
6889 	{ 0xdd, 12, 64, 3*1024*1024, l3_cache_str},
6890 	{ 0xdc, 12, 64, ((1*1024*1024)+(512*1024)), l3_cache_str},
6891 	{ 0xd8, 8, 64, 4*1024*1024, l3_cache_str},
6892 	{ 0xd7, 8, 64, 2*1024*1024, l3_cache_str},
6893 	{ 0xd6, 8, 64, 1*1024*1024, l3_cache_str},
6894 	{ 0xd2, 4, 64, 2*1024*1024, l3_cache_str},
6895 	{ 0xd1, 4, 64, 1*1024*1024, l3_cache_str},
6896 	{ 0xd0, 4, 64, 512*1024, l3_cache_str},
6897 	{ 0xca, 4, 0, 512, sh_l2_tlb4k_str},
6898 	{ 0xc0, 4, 0, 8, dtlb44_str },
6899 	{ 0xba, 4, 0, 64, dtlb4k_str },
6900 	{ 0xb4, 4, 0, 256, dtlb4k_str },
6901 	{ 0xb3, 4, 0, 128, dtlb4k_str },
6902 	{ 0xb2, 4, 0, 64, itlb4k_str },
6903 	{ 0xb0, 4, 0, 128, itlb4k_str },
6904 	{ 0x87, 8, 64, 1024*1024, l2_cache_str},
6905 	{ 0x86, 4, 64, 512*1024, l2_cache_str},
6906 	{ 0x85, 8, 32, 2*1024*1024, l2_cache_str},
6907 	{ 0x84, 8, 32, 1024*1024, l2_cache_str},
6908 	{ 0x83, 8, 32, 512*1024, l2_cache_str},
6909 	{ 0x82, 8, 32, 256*1024, l2_cache_str},
6910 	{ 0x80, 8, 64, 512*1024, l2_cache_str},
6911 	{ 0x7f, 2, 64, 512*1024, l2_cache_str},
6912 	{ 0x7d, 8, 64, 2*1024*1024, sl2_cache_str},
6913 	{ 0x7c, 8, 64, 1024*1024, sl2_cache_str},
6914 	{ 0x7b, 8, 64, 512*1024, sl2_cache_str},
6915 	{ 0x7a, 8, 64, 256*1024, sl2_cache_str},
6916 	{ 0x79, 8, 64, 128*1024, sl2_cache_str},
6917 	{ 0x78, 8, 64, 1024*1024, l2_cache_str},
6918 	{ 0x73, 8, 0, 64*1024, itrace_str},
6919 	{ 0x72, 8, 0, 32*1024, itrace_str},
6920 	{ 0x71, 8, 0, 16*1024, itrace_str},
6921 	{ 0x70, 8, 0, 12*1024, itrace_str},
6922 	{ 0x68, 4, 64, 32*1024, sl1_dcache_str},
6923 	{ 0x67, 4, 64, 16*1024, sl1_dcache_str},
6924 	{ 0x66, 4, 64, 8*1024, sl1_dcache_str},
6925 	{ 0x60, 8, 64, 16*1024, sl1_dcache_str},
6926 	{ 0x5d, 0, 0, 256, dtlb44_str},
6927 	{ 0x5c, 0, 0, 128, dtlb44_str},
6928 	{ 0x5b, 0, 0, 64, dtlb44_str},
6929 	{ 0x5a, 4, 0, 32, dtlb24_str},
6930 	{ 0x59, 0, 0, 16, dtlb4k_str},
6931 	{ 0x57, 4, 0, 16, dtlb4k_str},
6932 	{ 0x56, 4, 0, 16, dtlb4M_str},
6933 	{ 0x55, 0, 0, 7, itlb24_str},
6934 	{ 0x52, 0, 0, 256, itlb424_str},
6935 	{ 0x51, 0, 0, 128, itlb424_str},
6936 	{ 0x50, 0, 0, 64, itlb424_str},
6937 	{ 0x4f, 0, 0, 32, itlb4k_str},
6938 	{ 0x4e, 24, 64, 6*1024*1024, l2_cache_str},
6939 	{ 0x4d, 16, 64, 16*1024*1024, l3_cache_str},
6940 	{ 0x4c, 12, 64, 12*1024*1024, l3_cache_str},
6941 	{ 0x4b, 16, 64, 8*1024*1024, l3_cache_str},
6942 	{ 0x4a, 12, 64, 6*1024*1024, l3_cache_str},
6943 	{ 0x49, 16, 64, 4*1024*1024, l3_cache_str},
6944 	{ 0x48, 12, 64, 3*1024*1024, l2_cache_str},
6945 	{ 0x47, 8, 64, 8*1024*1024, l3_cache_str},
6946 	{ 0x46, 4, 64, 4*1024*1024, l3_cache_str},
6947 	{ 0x45, 4, 32, 2*1024*1024, l2_cache_str},
6948 	{ 0x44, 4, 32, 1024*1024, l2_cache_str},
6949 	{ 0x43, 4, 32, 512*1024, l2_cache_str},
6950 	{ 0x42, 4, 32, 256*1024, l2_cache_str},
6951 	{ 0x41, 4, 32, 128*1024, l2_cache_str},
6952 	{ 0x3e, 4, 64, 512*1024, sl2_cache_str},
6953 	{ 0x3d, 6, 64, 384*1024, sl2_cache_str},
6954 	{ 0x3c, 4, 64, 256*1024, sl2_cache_str},
6955 	{ 0x3b, 2, 64, 128*1024, sl2_cache_str},
6956 	{ 0x3a, 6, 64, 192*1024, sl2_cache_str},
6957 	{ 0x39, 4, 64, 128*1024, sl2_cache_str},
6958 	{ 0x30, 8, 64, 32*1024, l1_icache_str},
6959 	{ 0x2c, 8, 64, 32*1024, l1_dcache_str},
6960 	{ 0x29, 8, 64, 4096*1024, sl3_cache_str},
6961 	{ 0x25, 8, 64, 2048*1024, sl3_cache_str},
6962 	{ 0x23, 8, 64, 1024*1024, sl3_cache_str},
6963 	{ 0x22, 4, 64, 512*1024, sl3_cache_str},
6964 	{ 0x0e, 6, 64, 24*1024, l1_dcache_str},
6965 	{ 0x0d, 4, 32, 16*1024, l1_dcache_str},
6966 	{ 0x0c, 4, 32, 16*1024, l1_dcache_str},
6967 	{ 0x0b, 4, 0, 4, itlb4M_str},
6968 	{ 0x0a, 2, 32, 8*1024, l1_dcache_str},
6969 	{ 0x08, 4, 32, 16*1024, l1_icache_str},
6970 	{ 0x06, 4, 32, 8*1024, l1_icache_str},
6971 	{ 0x05, 4, 0, 32, dtlb4M_str},
6972 	{ 0x04, 4, 0, 8, dtlb4M_str},
6973 	{ 0x03, 4, 0, 64, dtlb4k_str},
6974 	{ 0x02, 4, 0, 2, itlb4M_str},
6975 	{ 0x01, 4, 0, 32, itlb4k_str},
6976 	{ 0 }
6977 };
6978 
6979 static const struct cachetab cyrix_ctab[] = {
6980 	{ 0x70, 4, 0, 32, "tlb-4K" },
6981 	{ 0x80, 4, 16, 16*1024, "l1-cache" },
6982 	{ 0 }
6983 };
6984 
6985 /*
6986  * Search a cache table for a matching entry
6987  */
6988 static const struct cachetab *
6989 find_cacheent(const struct cachetab *ct, uint_t code)
6990 {
6991 	if (code != 0) {
6992 		for (; ct->ct_code != 0; ct++)
6993 			if (ct->ct_code <= code)
6994 				break;
6995 		if (ct->ct_code == code)
6996 			return (ct);
6997 	}
6998 	return (NULL);
6999 }
7000 
7001 /*
7002  * Populate cachetab entry with L2 or L3 cache-information using
7003  * cpuid function 4. This function is called from intel_walk_cacheinfo()
7004  * when descriptor 0x49 is encountered. It returns 0 if no such cache
7005  * information is found.
7006  */
7007 static int
7008 intel_cpuid_4_cache_info(struct cachetab *ct, struct cpuid_info *cpi)
7009 {
7010 	uint32_t level, i;
7011 	int ret = 0;
7012 
7013 	for (i = 0; i < cpi->cpi_cache_leaf_size; i++) {
7014 		level = CPI_CACHE_LVL(cpi->cpi_cache_leaves[i]);
7015 
7016 		if (level == 2 || level == 3) {
7017 			ct->ct_assoc =
7018 			    CPI_CACHE_WAYS(cpi->cpi_cache_leaves[i]) + 1;
7019 			ct->ct_line_size =
7020 			    CPI_CACHE_COH_LN_SZ(cpi->cpi_cache_leaves[i]) + 1;
7021 			ct->ct_size = ct->ct_assoc *
7022 			    (CPI_CACHE_PARTS(cpi->cpi_cache_leaves[i]) + 1) *
7023 			    ct->ct_line_size *
7024 			    (cpi->cpi_cache_leaves[i]->cp_ecx + 1);
7025 
7026 			if (level == 2) {
7027 				ct->ct_label = l2_cache_str;
7028 			} else if (level == 3) {
7029 				ct->ct_label = l3_cache_str;
7030 			}
7031 			ret = 1;
7032 		}
7033 	}
7034 
7035 	return (ret);
7036 }
7037 
7038 /*
7039  * Walk the cacheinfo descriptor, applying 'func' to every valid element
7040  * The walk is terminated if the walker returns non-zero.
7041  */
7042 static void
7043 intel_walk_cacheinfo(struct cpuid_info *cpi,
7044     void *arg, int (*func)(void *, const struct cachetab *))
7045 {
7046 	const struct cachetab *ct;
7047 	struct cachetab des_49_ct, des_b1_ct;
7048 	uint8_t *dp;
7049 	int i;
7050 
7051 	if ((dp = cpi->cpi_cacheinfo) == NULL)
7052 		return;
7053 	for (i = 0; i < cpi->cpi_ncache; i++, dp++) {
7054 		/*
7055 		 * For overloaded descriptor 0x49 we use cpuid function 4
7056 		 * if supported by the current processor, to create
7057 		 * cache information.
7058 		 * For overloaded descriptor 0xb1 we use X86_PAE flag
7059 		 * to disambiguate the cache information.
7060 		 */
7061 		if (*dp == 0x49 && cpi->cpi_maxeax >= 0x4 &&
7062 		    intel_cpuid_4_cache_info(&des_49_ct, cpi) == 1) {
7063 				ct = &des_49_ct;
7064 		} else if (*dp == 0xb1) {
7065 			des_b1_ct.ct_code = 0xb1;
7066 			des_b1_ct.ct_assoc = 4;
7067 			des_b1_ct.ct_line_size = 0;
7068 			if (is_x86_feature(x86_featureset, X86FSET_PAE)) {
7069 				des_b1_ct.ct_size = 8;
7070 				des_b1_ct.ct_label = itlb2M_str;
7071 			} else {
7072 				des_b1_ct.ct_size = 4;
7073 				des_b1_ct.ct_label = itlb4M_str;
7074 			}
7075 			ct = &des_b1_ct;
7076 		} else {
7077 			if ((ct = find_cacheent(intel_ctab, *dp)) == NULL) {
7078 				continue;
7079 			}
7080 		}
7081 
7082 		if (func(arg, ct) != 0) {
7083 			break;
7084 		}
7085 	}
7086 }
7087 
7088 /*
7089  * (Like the Intel one, except for Cyrix CPUs)
7090  */
7091 static void
7092 cyrix_walk_cacheinfo(struct cpuid_info *cpi,
7093     void *arg, int (*func)(void *, const struct cachetab *))
7094 {
7095 	const struct cachetab *ct;
7096 	uint8_t *dp;
7097 	int i;
7098 
7099 	if ((dp = cpi->cpi_cacheinfo) == NULL)
7100 		return;
7101 	for (i = 0; i < cpi->cpi_ncache; i++, dp++) {
7102 		/*
7103 		 * Search Cyrix-specific descriptor table first ..
7104 		 */
7105 		if ((ct = find_cacheent(cyrix_ctab, *dp)) != NULL) {
7106 			if (func(arg, ct) != 0)
7107 				break;
7108 			continue;
7109 		}
7110 		/*
7111 		 * .. else fall back to the Intel one
7112 		 */
7113 		if ((ct = find_cacheent(intel_ctab, *dp)) != NULL) {
7114 			if (func(arg, ct) != 0)
7115 				break;
7116 			continue;
7117 		}
7118 	}
7119 }
7120 
7121 /*
7122  * A cacheinfo walker that adds associativity, line-size, and size properties
7123  * to the devinfo node it is passed as an argument.
7124  */
7125 static int
7126 add_cacheent_props(void *arg, const struct cachetab *ct)
7127 {
7128 	dev_info_t *devi = arg;
7129 
7130 	add_cache_prop(devi, ct->ct_label, assoc_str, ct->ct_assoc);
7131 	if (ct->ct_line_size != 0)
7132 		add_cache_prop(devi, ct->ct_label, line_str,
7133 		    ct->ct_line_size);
7134 	add_cache_prop(devi, ct->ct_label, size_str, ct->ct_size);
7135 	return (0);
7136 }
7137 
7138 
7139 static const char fully_assoc[] = "fully-associative?";
7140 
7141 /*
7142  * AMD style cache/tlb description
7143  *
7144  * Extended functions 5 and 6 directly describe properties of
7145  * tlbs and various cache levels.
7146  */
7147 static void
7148 add_amd_assoc(dev_info_t *devi, const char *label, uint_t assoc)
7149 {
7150 	switch (assoc) {
7151 	case 0:	/* reserved; ignore */
7152 		break;
7153 	default:
7154 		add_cache_prop(devi, label, assoc_str, assoc);
7155 		break;
7156 	case 0xff:
7157 		add_cache_prop(devi, label, fully_assoc, 1);
7158 		break;
7159 	}
7160 }
7161 
7162 static void
7163 add_amd_tlb(dev_info_t *devi, const char *label, uint_t assoc, uint_t size)
7164 {
7165 	if (size == 0)
7166 		return;
7167 	add_cache_prop(devi, label, size_str, size);
7168 	add_amd_assoc(devi, label, assoc);
7169 }
7170 
7171 static void
7172 add_amd_cache(dev_info_t *devi, const char *label,
7173     uint_t size, uint_t assoc, uint_t lines_per_tag, uint_t line_size)
7174 {
7175 	if (size == 0 || line_size == 0)
7176 		return;
7177 	add_amd_assoc(devi, label, assoc);
7178 	/*
7179 	 * Most AMD parts have a sectored cache. Multiple cache lines are
7180 	 * associated with each tag. A sector consists of all cache lines
7181 	 * associated with a tag. For example, the AMD K6-III has a sector
7182 	 * size of 2 cache lines per tag.
7183 	 */
7184 	if (lines_per_tag != 0)
7185 		add_cache_prop(devi, label, "lines-per-tag", lines_per_tag);
7186 	add_cache_prop(devi, label, line_str, line_size);
7187 	add_cache_prop(devi, label, size_str, size * 1024);
7188 }
7189 
7190 static void
7191 add_amd_l2_assoc(dev_info_t *devi, const char *label, uint_t assoc)
7192 {
7193 	switch (assoc) {
7194 	case 0:	/* off */
7195 		break;
7196 	case 1:
7197 	case 2:
7198 	case 4:
7199 		add_cache_prop(devi, label, assoc_str, assoc);
7200 		break;
7201 	case 6:
7202 		add_cache_prop(devi, label, assoc_str, 8);
7203 		break;
7204 	case 8:
7205 		add_cache_prop(devi, label, assoc_str, 16);
7206 		break;
7207 	case 0xf:
7208 		add_cache_prop(devi, label, fully_assoc, 1);
7209 		break;
7210 	default: /* reserved; ignore */
7211 		break;
7212 	}
7213 }
7214 
7215 static void
7216 add_amd_l2_tlb(dev_info_t *devi, const char *label, uint_t assoc, uint_t size)
7217 {
7218 	if (size == 0 || assoc == 0)
7219 		return;
7220 	add_amd_l2_assoc(devi, label, assoc);
7221 	add_cache_prop(devi, label, size_str, size);
7222 }
7223 
7224 static void
7225 add_amd_l2_cache(dev_info_t *devi, const char *label,
7226     uint_t size, uint_t assoc, uint_t lines_per_tag, uint_t line_size)
7227 {
7228 	if (size == 0 || assoc == 0 || line_size == 0)
7229 		return;
7230 	add_amd_l2_assoc(devi, label, assoc);
7231 	if (lines_per_tag != 0)
7232 		add_cache_prop(devi, label, "lines-per-tag", lines_per_tag);
7233 	add_cache_prop(devi, label, line_str, line_size);
7234 	add_cache_prop(devi, label, size_str, size * 1024);
7235 }
7236 
7237 static void
7238 amd_cache_info(struct cpuid_info *cpi, dev_info_t *devi)
7239 {
7240 	struct cpuid_regs *cp;
7241 
7242 	if (cpi->cpi_xmaxeax < 0x80000005)
7243 		return;
7244 	cp = &cpi->cpi_extd[5];
7245 
7246 	/*
7247 	 * 4M/2M L1 TLB configuration
7248 	 *
7249 	 * We report the size for 2M pages because AMD uses two
7250 	 * TLB entries for one 4M page.
7251 	 */
7252 	add_amd_tlb(devi, "dtlb-2M",
7253 	    BITX(cp->cp_eax, 31, 24), BITX(cp->cp_eax, 23, 16));
7254 	add_amd_tlb(devi, "itlb-2M",
7255 	    BITX(cp->cp_eax, 15, 8), BITX(cp->cp_eax, 7, 0));
7256 
7257 	/*
7258 	 * 4K L1 TLB configuration
7259 	 */
7260 
7261 	switch (cpi->cpi_vendor) {
7262 		uint_t nentries;
7263 	case X86_VENDOR_TM:
7264 		if (cpi->cpi_family >= 5) {
7265 			/*
7266 			 * Crusoe processors have 256 TLB entries, but
7267 			 * cpuid data format constrains them to only
7268 			 * reporting 255 of them.
7269 			 */
7270 			if ((nentries = BITX(cp->cp_ebx, 23, 16)) == 255)
7271 				nentries = 256;
7272 			/*
7273 			 * Crusoe processors also have a unified TLB
7274 			 */
7275 			add_amd_tlb(devi, "tlb-4K", BITX(cp->cp_ebx, 31, 24),
7276 			    nentries);
7277 			break;
7278 		}
7279 		/*FALLTHROUGH*/
7280 	default:
7281 		add_amd_tlb(devi, itlb4k_str,
7282 		    BITX(cp->cp_ebx, 31, 24), BITX(cp->cp_ebx, 23, 16));
7283 		add_amd_tlb(devi, dtlb4k_str,
7284 		    BITX(cp->cp_ebx, 15, 8), BITX(cp->cp_ebx, 7, 0));
7285 		break;
7286 	}
7287 
7288 	/*
7289 	 * data L1 cache configuration
7290 	 */
7291 
7292 	add_amd_cache(devi, l1_dcache_str,
7293 	    BITX(cp->cp_ecx, 31, 24), BITX(cp->cp_ecx, 23, 16),
7294 	    BITX(cp->cp_ecx, 15, 8), BITX(cp->cp_ecx, 7, 0));
7295 
7296 	/*
7297 	 * code L1 cache configuration
7298 	 */
7299 
7300 	add_amd_cache(devi, l1_icache_str,
7301 	    BITX(cp->cp_edx, 31, 24), BITX(cp->cp_edx, 23, 16),
7302 	    BITX(cp->cp_edx, 15, 8), BITX(cp->cp_edx, 7, 0));
7303 
7304 	if (cpi->cpi_xmaxeax < 0x80000006)
7305 		return;
7306 	cp = &cpi->cpi_extd[6];
7307 
7308 	/* Check for a unified L2 TLB for large pages */
7309 
7310 	if (BITX(cp->cp_eax, 31, 16) == 0)
7311 		add_amd_l2_tlb(devi, "l2-tlb-2M",
7312 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
7313 	else {
7314 		add_amd_l2_tlb(devi, "l2-dtlb-2M",
7315 		    BITX(cp->cp_eax, 31, 28), BITX(cp->cp_eax, 27, 16));
7316 		add_amd_l2_tlb(devi, "l2-itlb-2M",
7317 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
7318 	}
7319 
7320 	/* Check for a unified L2 TLB for 4K pages */
7321 
7322 	if (BITX(cp->cp_ebx, 31, 16) == 0) {
7323 		add_amd_l2_tlb(devi, "l2-tlb-4K",
7324 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
7325 	} else {
7326 		add_amd_l2_tlb(devi, "l2-dtlb-4K",
7327 		    BITX(cp->cp_eax, 31, 28), BITX(cp->cp_eax, 27, 16));
7328 		add_amd_l2_tlb(devi, "l2-itlb-4K",
7329 		    BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
7330 	}
7331 
7332 	add_amd_l2_cache(devi, l2_cache_str,
7333 	    BITX(cp->cp_ecx, 31, 16), BITX(cp->cp_ecx, 15, 12),
7334 	    BITX(cp->cp_ecx, 11, 8), BITX(cp->cp_ecx, 7, 0));
7335 }
7336 
7337 /*
7338  * There are two basic ways that the x86 world describes it cache
7339  * and tlb architecture - Intel's way and AMD's way.
7340  *
7341  * Return which flavor of cache architecture we should use
7342  */
7343 static int
7344 x86_which_cacheinfo(struct cpuid_info *cpi)
7345 {
7346 	switch (cpi->cpi_vendor) {
7347 	case X86_VENDOR_Intel:
7348 		if (cpi->cpi_maxeax >= 2)
7349 			return (X86_VENDOR_Intel);
7350 		break;
7351 	case X86_VENDOR_AMD:
7352 		/*
7353 		 * The K5 model 1 was the first part from AMD that reported
7354 		 * cache sizes via extended cpuid functions.
7355 		 */
7356 		if (cpi->cpi_family > 5 ||
7357 		    (cpi->cpi_family == 5 && cpi->cpi_model >= 1))
7358 			return (X86_VENDOR_AMD);
7359 		break;
7360 	case X86_VENDOR_HYGON:
7361 		return (X86_VENDOR_AMD);
7362 	case X86_VENDOR_TM:
7363 		if (cpi->cpi_family >= 5)
7364 			return (X86_VENDOR_AMD);
7365 		/*FALLTHROUGH*/
7366 	default:
7367 		/*
7368 		 * If they have extended CPU data for 0x80000005
7369 		 * then we assume they have AMD-format cache
7370 		 * information.
7371 		 *
7372 		 * If not, and the vendor happens to be Cyrix,
7373 		 * then try our-Cyrix specific handler.
7374 		 *
7375 		 * If we're not Cyrix, then assume we're using Intel's
7376 		 * table-driven format instead.
7377 		 */
7378 		if (cpi->cpi_xmaxeax >= 0x80000005)
7379 			return (X86_VENDOR_AMD);
7380 		else if (cpi->cpi_vendor == X86_VENDOR_Cyrix)
7381 			return (X86_VENDOR_Cyrix);
7382 		else if (cpi->cpi_maxeax >= 2)
7383 			return (X86_VENDOR_Intel);
7384 		break;
7385 	}
7386 	return (-1);
7387 }
7388 
7389 void
7390 cpuid_set_cpu_properties(void *dip, processorid_t cpu_id,
7391     struct cpuid_info *cpi)
7392 {
7393 	dev_info_t *cpu_devi;
7394 	int create;
7395 
7396 	cpu_devi = (dev_info_t *)dip;
7397 
7398 	/* device_type */
7399 	(void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi,
7400 	    "device_type", "cpu");
7401 
7402 	/* reg */
7403 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7404 	    "reg", cpu_id);
7405 
7406 	/* cpu-mhz, and clock-frequency */
7407 	if (cpu_freq > 0) {
7408 		long long mul;
7409 
7410 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7411 		    "cpu-mhz", cpu_freq);
7412 		if ((mul = cpu_freq * 1000000LL) <= INT_MAX)
7413 			(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7414 			    "clock-frequency", (int)mul);
7415 	}
7416 
7417 	ASSERT(is_x86_feature(x86_featureset, X86FSET_CPUID));
7418 
7419 	/* vendor-id */
7420 	(void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi,
7421 	    "vendor-id", cpi->cpi_vendorstr);
7422 
7423 	if (cpi->cpi_maxeax == 0) {
7424 		return;
7425 	}
7426 
7427 	/*
7428 	 * family, model, and step
7429 	 */
7430 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7431 	    "family", CPI_FAMILY(cpi));
7432 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7433 	    "cpu-model", CPI_MODEL(cpi));
7434 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7435 	    "stepping-id", CPI_STEP(cpi));
7436 
7437 	/* type */
7438 	switch (cpi->cpi_vendor) {
7439 	case X86_VENDOR_Intel:
7440 		create = 1;
7441 		break;
7442 	default:
7443 		create = 0;
7444 		break;
7445 	}
7446 	if (create)
7447 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7448 		    "type", CPI_TYPE(cpi));
7449 
7450 	/* ext-family */
7451 	switch (cpi->cpi_vendor) {
7452 	case X86_VENDOR_Intel:
7453 	case X86_VENDOR_AMD:
7454 		create = cpi->cpi_family >= 0xf;
7455 		break;
7456 	case X86_VENDOR_HYGON:
7457 		create = 1;
7458 		break;
7459 	default:
7460 		create = 0;
7461 		break;
7462 	}
7463 	if (create)
7464 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7465 		    "ext-family", CPI_FAMILY_XTD(cpi));
7466 
7467 	/* ext-model */
7468 	switch (cpi->cpi_vendor) {
7469 	case X86_VENDOR_Intel:
7470 		create = IS_EXTENDED_MODEL_INTEL(cpi);
7471 		break;
7472 	case X86_VENDOR_AMD:
7473 		create = CPI_FAMILY(cpi) == 0xf;
7474 		break;
7475 	case X86_VENDOR_HYGON:
7476 		create = 1;
7477 		break;
7478 	default:
7479 		create = 0;
7480 		break;
7481 	}
7482 	if (create)
7483 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7484 		    "ext-model", CPI_MODEL_XTD(cpi));
7485 
7486 	/* generation */
7487 	switch (cpi->cpi_vendor) {
7488 	case X86_VENDOR_AMD:
7489 	case X86_VENDOR_HYGON:
7490 		/*
7491 		 * AMD K5 model 1 was the first part to support this
7492 		 */
7493 		create = cpi->cpi_xmaxeax >= 0x80000001;
7494 		break;
7495 	default:
7496 		create = 0;
7497 		break;
7498 	}
7499 	if (create)
7500 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7501 		    "generation", BITX((cpi)->cpi_extd[1].cp_eax, 11, 8));
7502 
7503 	/* brand-id */
7504 	switch (cpi->cpi_vendor) {
7505 	case X86_VENDOR_Intel:
7506 		/*
7507 		 * brand id first appeared on Pentium III Xeon model 8,
7508 		 * and Celeron model 8 processors and Opteron
7509 		 */
7510 		create = cpi->cpi_family > 6 ||
7511 		    (cpi->cpi_family == 6 && cpi->cpi_model >= 8);
7512 		break;
7513 	case X86_VENDOR_AMD:
7514 		create = cpi->cpi_family >= 0xf;
7515 		break;
7516 	case X86_VENDOR_HYGON:
7517 		create = 1;
7518 		break;
7519 	default:
7520 		create = 0;
7521 		break;
7522 	}
7523 	if (create && cpi->cpi_brandid != 0) {
7524 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7525 		    "brand-id", cpi->cpi_brandid);
7526 	}
7527 
7528 	/* chunks, and apic-id */
7529 	switch (cpi->cpi_vendor) {
7530 		/*
7531 		 * first available on Pentium IV and Opteron (K8)
7532 		 */
7533 	case X86_VENDOR_Intel:
7534 		create = IS_NEW_F6(cpi) || cpi->cpi_family >= 0xf;
7535 		break;
7536 	case X86_VENDOR_AMD:
7537 		create = cpi->cpi_family >= 0xf;
7538 		break;
7539 	case X86_VENDOR_HYGON:
7540 		create = 1;
7541 		break;
7542 	default:
7543 		create = 0;
7544 		break;
7545 	}
7546 	if (create) {
7547 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7548 		    "chunks", CPI_CHUNKS(cpi));
7549 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7550 		    "apic-id", cpi->cpi_apicid);
7551 		if (cpi->cpi_chipid >= 0) {
7552 			(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7553 			    "chip#", cpi->cpi_chipid);
7554 			(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7555 			    "clog#", cpi->cpi_clogid);
7556 		}
7557 	}
7558 
7559 	/* cpuid-features */
7560 	(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7561 	    "cpuid-features", CPI_FEATURES_EDX(cpi));
7562 
7563 
7564 	/* cpuid-features-ecx */
7565 	switch (cpi->cpi_vendor) {
7566 	case X86_VENDOR_Intel:
7567 		create = IS_NEW_F6(cpi) || cpi->cpi_family >= 0xf;
7568 		break;
7569 	case X86_VENDOR_AMD:
7570 		create = cpi->cpi_family >= 0xf;
7571 		break;
7572 	case X86_VENDOR_HYGON:
7573 		create = 1;
7574 		break;
7575 	default:
7576 		create = 0;
7577 		break;
7578 	}
7579 	if (create)
7580 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7581 		    "cpuid-features-ecx", CPI_FEATURES_ECX(cpi));
7582 
7583 	/* ext-cpuid-features */
7584 	switch (cpi->cpi_vendor) {
7585 	case X86_VENDOR_Intel:
7586 	case X86_VENDOR_AMD:
7587 	case X86_VENDOR_HYGON:
7588 	case X86_VENDOR_Cyrix:
7589 	case X86_VENDOR_TM:
7590 	case X86_VENDOR_Centaur:
7591 		create = cpi->cpi_xmaxeax >= 0x80000001;
7592 		break;
7593 	default:
7594 		create = 0;
7595 		break;
7596 	}
7597 	if (create) {
7598 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7599 		    "ext-cpuid-features", CPI_FEATURES_XTD_EDX(cpi));
7600 		(void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7601 		    "ext-cpuid-features-ecx", CPI_FEATURES_XTD_ECX(cpi));
7602 	}
7603 
7604 	/*
7605 	 * Brand String first appeared in Intel Pentium IV, AMD K5
7606 	 * model 1, and Cyrix GXm.  On earlier models we try and
7607 	 * simulate something similar .. so this string should always
7608 	 * same -something- about the processor, however lame.
7609 	 */
7610 	(void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi,
7611 	    "brand-string", cpi->cpi_brandstr);
7612 
7613 	/*
7614 	 * Finally, cache and tlb information
7615 	 */
7616 	switch (x86_which_cacheinfo(cpi)) {
7617 	case X86_VENDOR_Intel:
7618 		intel_walk_cacheinfo(cpi, cpu_devi, add_cacheent_props);
7619 		break;
7620 	case X86_VENDOR_Cyrix:
7621 		cyrix_walk_cacheinfo(cpi, cpu_devi, add_cacheent_props);
7622 		break;
7623 	case X86_VENDOR_AMD:
7624 		amd_cache_info(cpi, cpu_devi);
7625 		break;
7626 	default:
7627 		break;
7628 	}
7629 }
7630 
7631 struct l2info {
7632 	int *l2i_csz;
7633 	int *l2i_lsz;
7634 	int *l2i_assoc;
7635 	int l2i_ret;
7636 };
7637 
7638 /*
7639  * A cacheinfo walker that fetches the size, line-size and associativity
7640  * of the L2 cache
7641  */
7642 static int
7643 intel_l2cinfo(void *arg, const struct cachetab *ct)
7644 {
7645 	struct l2info *l2i = arg;
7646 	int *ip;
7647 
7648 	if (ct->ct_label != l2_cache_str &&
7649 	    ct->ct_label != sl2_cache_str)
7650 		return (0);	/* not an L2 -- keep walking */
7651 
7652 	if ((ip = l2i->l2i_csz) != NULL)
7653 		*ip = ct->ct_size;
7654 	if ((ip = l2i->l2i_lsz) != NULL)
7655 		*ip = ct->ct_line_size;
7656 	if ((ip = l2i->l2i_assoc) != NULL)
7657 		*ip = ct->ct_assoc;
7658 	l2i->l2i_ret = ct->ct_size;
7659 	return (1);		/* was an L2 -- terminate walk */
7660 }
7661 
7662 /*
7663  * AMD L2/L3 Cache and TLB Associativity Field Definition:
7664  *
7665  *	Unlike the associativity for the L1 cache and tlb where the 8 bit
7666  *	value is the associativity, the associativity for the L2 cache and
7667  *	tlb is encoded in the following table. The 4 bit L2 value serves as
7668  *	an index into the amd_afd[] array to determine the associativity.
7669  *	-1 is undefined. 0 is fully associative.
7670  */
7671 
7672 static int amd_afd[] =
7673 	{-1, 1, 2, -1, 4, -1, 8, -1, 16, -1, 32, 48, 64, 96, 128, 0};
7674 
7675 static void
7676 amd_l2cacheinfo(struct cpuid_info *cpi, struct l2info *l2i)
7677 {
7678 	struct cpuid_regs *cp;
7679 	uint_t size, assoc;
7680 	int i;
7681 	int *ip;
7682 
7683 	if (cpi->cpi_xmaxeax < 0x80000006)
7684 		return;
7685 	cp = &cpi->cpi_extd[6];
7686 
7687 	if ((i = BITX(cp->cp_ecx, 15, 12)) != 0 &&
7688 	    (size = BITX(cp->cp_ecx, 31, 16)) != 0) {
7689 		uint_t cachesz = size * 1024;
7690 		assoc = amd_afd[i];
7691 
7692 		ASSERT(assoc != -1);
7693 
7694 		if ((ip = l2i->l2i_csz) != NULL)
7695 			*ip = cachesz;
7696 		if ((ip = l2i->l2i_lsz) != NULL)
7697 			*ip = BITX(cp->cp_ecx, 7, 0);
7698 		if ((ip = l2i->l2i_assoc) != NULL)
7699 			*ip = assoc;
7700 		l2i->l2i_ret = cachesz;
7701 	}
7702 }
7703 
7704 int
7705 getl2cacheinfo(cpu_t *cpu, int *csz, int *lsz, int *assoc)
7706 {
7707 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
7708 	struct l2info __l2info, *l2i = &__l2info;
7709 
7710 	l2i->l2i_csz = csz;
7711 	l2i->l2i_lsz = lsz;
7712 	l2i->l2i_assoc = assoc;
7713 	l2i->l2i_ret = -1;
7714 
7715 	switch (x86_which_cacheinfo(cpi)) {
7716 	case X86_VENDOR_Intel:
7717 		intel_walk_cacheinfo(cpi, l2i, intel_l2cinfo);
7718 		break;
7719 	case X86_VENDOR_Cyrix:
7720 		cyrix_walk_cacheinfo(cpi, l2i, intel_l2cinfo);
7721 		break;
7722 	case X86_VENDOR_AMD:
7723 		amd_l2cacheinfo(cpi, l2i);
7724 		break;
7725 	default:
7726 		break;
7727 	}
7728 	return (l2i->l2i_ret);
7729 }
7730 
7731 #if !defined(__xpv)
7732 
7733 uint32_t *
7734 cpuid_mwait_alloc(cpu_t *cpu)
7735 {
7736 	uint32_t	*ret;
7737 	size_t		mwait_size;
7738 
7739 	ASSERT(cpuid_checkpass(CPU, CPUID_PASS_EXTENDED));
7740 
7741 	mwait_size = CPU->cpu_m.mcpu_cpi->cpi_mwait.mon_max;
7742 	if (mwait_size == 0)
7743 		return (NULL);
7744 
7745 	/*
7746 	 * kmem_alloc() returns cache line size aligned data for mwait_size
7747 	 * allocations.  mwait_size is currently cache line sized.  Neither
7748 	 * of these implementation details are guarantied to be true in the
7749 	 * future.
7750 	 *
7751 	 * First try allocating mwait_size as kmem_alloc() currently returns
7752 	 * correctly aligned memory.  If kmem_alloc() does not return
7753 	 * mwait_size aligned memory, then use mwait_size ROUNDUP.
7754 	 *
7755 	 * Set cpi_mwait.buf_actual and cpi_mwait.size_actual in case we
7756 	 * decide to free this memory.
7757 	 */
7758 	ret = kmem_zalloc(mwait_size, KM_SLEEP);
7759 	if (ret == (uint32_t *)P2ROUNDUP((uintptr_t)ret, mwait_size)) {
7760 		cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual = ret;
7761 		cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual = mwait_size;
7762 		*ret = MWAIT_RUNNING;
7763 		return (ret);
7764 	} else {
7765 		kmem_free(ret, mwait_size);
7766 		ret = kmem_zalloc(mwait_size * 2, KM_SLEEP);
7767 		cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual = ret;
7768 		cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual = mwait_size * 2;
7769 		ret = (uint32_t *)P2ROUNDUP((uintptr_t)ret, mwait_size);
7770 		*ret = MWAIT_RUNNING;
7771 		return (ret);
7772 	}
7773 }
7774 
7775 void
7776 cpuid_mwait_free(cpu_t *cpu)
7777 {
7778 	if (cpu->cpu_m.mcpu_cpi == NULL) {
7779 		return;
7780 	}
7781 
7782 	if (cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual != NULL &&
7783 	    cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual > 0) {
7784 		kmem_free(cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual,
7785 		    cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual);
7786 	}
7787 
7788 	cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual = NULL;
7789 	cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual = 0;
7790 }
7791 
7792 void
7793 patch_tsc_read(int flag)
7794 {
7795 	size_t cnt;
7796 
7797 	switch (flag) {
7798 	case TSC_NONE:
7799 		cnt = &_no_rdtsc_end - &_no_rdtsc_start;
7800 		(void) memcpy((void *)tsc_read, (void *)&_no_rdtsc_start, cnt);
7801 		break;
7802 	case TSC_RDTSC_LFENCE:
7803 		cnt = &_tsc_lfence_end - &_tsc_lfence_start;
7804 		(void) memcpy((void *)tsc_read,
7805 		    (void *)&_tsc_lfence_start, cnt);
7806 		break;
7807 	case TSC_TSCP:
7808 		cnt = &_tscp_end - &_tscp_start;
7809 		(void) memcpy((void *)tsc_read, (void *)&_tscp_start, cnt);
7810 		break;
7811 	default:
7812 		/* Bail for unexpected TSC types. (TSC_NONE covers 0) */
7813 		cmn_err(CE_PANIC, "Unrecogized TSC type: %d", flag);
7814 		break;
7815 	}
7816 	tsc_type = flag;
7817 }
7818 
7819 int
7820 cpuid_deep_cstates_supported(void)
7821 {
7822 	struct cpuid_info *cpi;
7823 	struct cpuid_regs regs;
7824 
7825 	ASSERT(cpuid_checkpass(CPU, CPUID_PASS_BASIC));
7826 	ASSERT(is_x86_feature(x86_featureset, X86FSET_CPUID));
7827 
7828 	cpi = CPU->cpu_m.mcpu_cpi;
7829 
7830 	switch (cpi->cpi_vendor) {
7831 	case X86_VENDOR_Intel:
7832 		if (cpi->cpi_xmaxeax < 0x80000007)
7833 			return (0);
7834 
7835 		/*
7836 		 * Does TSC run at a constant rate in all C-states?
7837 		 */
7838 		regs.cp_eax = 0x80000007;
7839 		(void) __cpuid_insn(&regs);
7840 		return (regs.cp_edx & CPUID_TSC_CSTATE_INVARIANCE);
7841 
7842 	default:
7843 		return (0);
7844 	}
7845 }
7846 
7847 #endif	/* !__xpv */
7848 
7849 void
7850 post_startup_cpu_fixups(void)
7851 {
7852 #ifndef __xpv
7853 	/*
7854 	 * Some AMD processors support C1E state. Entering this state will
7855 	 * cause the local APIC timer to stop, which we can't deal with at
7856 	 * this time.
7857 	 */
7858 	if (cpuid_getvendor(CPU) == X86_VENDOR_AMD) {
7859 		on_trap_data_t otd;
7860 		uint64_t reg;
7861 
7862 		if (!on_trap(&otd, OT_DATA_ACCESS)) {
7863 			reg = rdmsr(MSR_AMD_INT_PENDING_CMP_HALT);
7864 			/* Disable C1E state if it is enabled by BIOS */
7865 			if ((reg >> AMD_ACTONCMPHALT_SHIFT) &
7866 			    AMD_ACTONCMPHALT_MASK) {
7867 				reg &= ~(AMD_ACTONCMPHALT_MASK <<
7868 				    AMD_ACTONCMPHALT_SHIFT);
7869 				wrmsr(MSR_AMD_INT_PENDING_CMP_HALT, reg);
7870 			}
7871 		}
7872 		no_trap();
7873 	}
7874 #endif	/* !__xpv */
7875 }
7876 
7877 void
7878 enable_pcid(void)
7879 {
7880 	if (x86_use_pcid == -1)
7881 		x86_use_pcid = is_x86_feature(x86_featureset, X86FSET_PCID);
7882 
7883 	if (x86_use_invpcid == -1) {
7884 		x86_use_invpcid = is_x86_feature(x86_featureset,
7885 		    X86FSET_INVPCID);
7886 	}
7887 
7888 	if (!x86_use_pcid)
7889 		return;
7890 
7891 	/*
7892 	 * Intel say that on setting PCIDE, it immediately starts using the PCID
7893 	 * bits; better make sure there's nothing there.
7894 	 */
7895 	ASSERT((getcr3() & MMU_PAGEOFFSET) == PCID_NONE);
7896 
7897 	setcr4(getcr4() | CR4_PCIDE);
7898 }
7899 
7900 /*
7901  * Setup necessary registers to enable XSAVE feature on this processor.
7902  * This function needs to be called early enough, so that no xsave/xrstor
7903  * ops will execute on the processor before the MSRs are properly set up.
7904  *
7905  * Current implementation has the following assumption:
7906  * - cpuid_pass_basic() is done, so that X86 features are known.
7907  * - fpu_probe() is done, so that fp_save_mech is chosen.
7908  */
7909 void
7910 xsave_setup_msr(cpu_t *cpu)
7911 {
7912 	ASSERT(cpuid_checkpass(cpu, CPUID_PASS_BASIC));
7913 	ASSERT(fp_save_mech == FP_XSAVE);
7914 	ASSERT(is_x86_feature(x86_featureset, X86FSET_XSAVE));
7915 
7916 	/* Enable OSXSAVE in CR4. */
7917 	setcr4(getcr4() | CR4_OSXSAVE);
7918 	/*
7919 	 * Update SW copy of ECX, so that /dev/cpu/self/cpuid will report
7920 	 * correct value.
7921 	 */
7922 	cpu->cpu_m.mcpu_cpi->cpi_std[1].cp_ecx |= CPUID_INTC_ECX_OSXSAVE;
7923 	setup_xfem();
7924 }
7925 
7926 /*
7927  * Starting with the Westmere processor the local
7928  * APIC timer will continue running in all C-states,
7929  * including the deepest C-states.
7930  */
7931 int
7932 cpuid_arat_supported(void)
7933 {
7934 	struct cpuid_info *cpi;
7935 	struct cpuid_regs regs;
7936 
7937 	ASSERT(cpuid_checkpass(CPU, CPUID_PASS_BASIC));
7938 	ASSERT(is_x86_feature(x86_featureset, X86FSET_CPUID));
7939 
7940 	cpi = CPU->cpu_m.mcpu_cpi;
7941 
7942 	switch (cpi->cpi_vendor) {
7943 	case X86_VENDOR_Intel:
7944 		/*
7945 		 * Always-running Local APIC Timer is
7946 		 * indicated by CPUID.6.EAX[2].
7947 		 */
7948 		if (cpi->cpi_maxeax >= 6) {
7949 			regs.cp_eax = 6;
7950 			(void) cpuid_insn(NULL, &regs);
7951 			return (regs.cp_eax & CPUID_INTC_EAX_ARAT);
7952 		} else {
7953 			return (0);
7954 		}
7955 	default:
7956 		return (0);
7957 	}
7958 }
7959 
7960 /*
7961  * Check support for Intel ENERGY_PERF_BIAS feature
7962  */
7963 int
7964 cpuid_iepb_supported(struct cpu *cp)
7965 {
7966 	struct cpuid_info *cpi = cp->cpu_m.mcpu_cpi;
7967 	struct cpuid_regs regs;
7968 
7969 	ASSERT(cpuid_checkpass(cp, CPUID_PASS_BASIC));
7970 	ASSERT(is_x86_feature(x86_featureset, X86FSET_CPUID));
7971 
7972 	if (!(is_x86_feature(x86_featureset, X86FSET_MSR))) {
7973 		return (0);
7974 	}
7975 
7976 	/*
7977 	 * Intel ENERGY_PERF_BIAS MSR is indicated by
7978 	 * capability bit CPUID.6.ECX.3
7979 	 */
7980 	if ((cpi->cpi_vendor != X86_VENDOR_Intel) || (cpi->cpi_maxeax < 6))
7981 		return (0);
7982 
7983 	regs.cp_eax = 0x6;
7984 	(void) cpuid_insn(NULL, &regs);
7985 	return (regs.cp_ecx & CPUID_INTC_ECX_PERFBIAS);
7986 }
7987 
7988 /*
7989  * Check support for TSC deadline timer
7990  *
7991  * TSC deadline timer provides a superior software programming
7992  * model over local APIC timer that eliminates "time drifts".
7993  * Instead of specifying a relative time, software specifies an
7994  * absolute time as the target at which the processor should
7995  * generate a timer event.
7996  */
7997 int
7998 cpuid_deadline_tsc_supported(void)
7999 {
8000 	struct cpuid_info *cpi = CPU->cpu_m.mcpu_cpi;
8001 	struct cpuid_regs regs;
8002 
8003 	ASSERT(cpuid_checkpass(CPU, CPUID_PASS_BASIC));
8004 	ASSERT(is_x86_feature(x86_featureset, X86FSET_CPUID));
8005 
8006 	switch (cpi->cpi_vendor) {
8007 	case X86_VENDOR_Intel:
8008 		if (cpi->cpi_maxeax >= 1) {
8009 			regs.cp_eax = 1;
8010 			(void) cpuid_insn(NULL, &regs);
8011 			return (regs.cp_ecx & CPUID_DEADLINE_TSC);
8012 		} else {
8013 			return (0);
8014 		}
8015 	default:
8016 		return (0);
8017 	}
8018 }
8019 
8020 #if !defined(__xpv)
8021 /*
8022  * Patch in versions of bcopy for high performance Intel Nhm processors
8023  * and later...
8024  */
8025 void
8026 patch_memops(uint_t vendor)
8027 {
8028 	size_t cnt, i;
8029 	caddr_t to, from;
8030 
8031 	if ((vendor == X86_VENDOR_Intel) &&
8032 	    is_x86_feature(x86_featureset, X86FSET_SSE4_2)) {
8033 		cnt = &bcopy_patch_end - &bcopy_patch_start;
8034 		to = &bcopy_ck_size;
8035 		from = &bcopy_patch_start;
8036 		for (i = 0; i < cnt; i++) {
8037 			*to++ = *from++;
8038 		}
8039 	}
8040 }
8041 #endif  /*  !__xpv */
8042 
8043 /*
8044  * We're being asked to tell the system how many bits are required to represent
8045  * the various thread and strand IDs. While it's tempting to derive this based
8046  * on the values in cpi_ncore_per_chip and cpi_ncpu_per_chip, that isn't quite
8047  * correct. Instead, this needs to be based on the number of bits that the APIC
8048  * allows for these different configurations. We only update these to a larger
8049  * value if we find one.
8050  */
8051 void
8052 cpuid_get_ext_topo(cpu_t *cpu, uint_t *core_nbits, uint_t *strand_nbits)
8053 {
8054 	struct cpuid_info *cpi;
8055 
8056 	VERIFY(cpuid_checkpass(CPU, CPUID_PASS_BASIC));
8057 	cpi = cpu->cpu_m.mcpu_cpi;
8058 
8059 	if (cpi->cpi_ncore_bits > *core_nbits) {
8060 		*core_nbits = cpi->cpi_ncore_bits;
8061 	}
8062 
8063 	if (cpi->cpi_nthread_bits > *strand_nbits) {
8064 		*strand_nbits = cpi->cpi_nthread_bits;
8065 	}
8066 }
8067 
8068 void
8069 cpuid_pass_ucode(cpu_t *cpu, uchar_t *fset)
8070 {
8071 	struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
8072 	struct cpuid_regs cp;
8073 
8074 	/*
8075 	 * Reread the CPUID portions that we need for various security
8076 	 * information.
8077 	 */
8078 	switch (cpi->cpi_vendor) {
8079 	case X86_VENDOR_Intel:
8080 		/*
8081 		 * Check if we now have leaf 7 available to us.
8082 		 */
8083 		if (cpi->cpi_maxeax < 7) {
8084 			bzero(&cp, sizeof (cp));
8085 			cp.cp_eax = 0;
8086 			cpi->cpi_maxeax = __cpuid_insn(&cp);
8087 			if (cpi->cpi_maxeax < 7)
8088 				break;
8089 		}
8090 
8091 		bzero(&cp, sizeof (cp));
8092 		cp.cp_eax = 7;
8093 		cp.cp_ecx = 0;
8094 		(void) __cpuid_insn(&cp);
8095 		cpi->cpi_std[7] = cp;
8096 		break;
8097 
8098 	case X86_VENDOR_AMD:
8099 	case X86_VENDOR_HYGON:
8100 		/* No xcpuid support */
8101 		if (cpi->cpi_family < 5 ||
8102 		    (cpi->cpi_family == 5 && cpi->cpi_model < 1))
8103 			break;
8104 
8105 		if (cpi->cpi_xmaxeax < CPUID_LEAF_EXT_8) {
8106 			bzero(&cp, sizeof (cp));
8107 			cp.cp_eax = CPUID_LEAF_EXT_0;
8108 			cpi->cpi_xmaxeax = __cpuid_insn(&cp);
8109 			if (cpi->cpi_xmaxeax < CPUID_LEAF_EXT_8)
8110 				break;
8111 		}
8112 
8113 		/*
8114 		 * Most AMD features are in leaf 8. Automatic IBRS was added in
8115 		 * leaf 0x21. So we also check that.
8116 		 */
8117 		bzero(&cp, sizeof (cp));
8118 		cp.cp_eax = CPUID_LEAF_EXT_8;
8119 		(void) __cpuid_insn(&cp);
8120 		platform_cpuid_mangle(cpi->cpi_vendor, CPUID_LEAF_EXT_8, &cp);
8121 		cpi->cpi_extd[8] = cp;
8122 
8123 		if (cpi->cpi_xmaxeax < CPUID_LEAF_EXT_21)
8124 			break;
8125 
8126 		bzero(&cp, sizeof (cp));
8127 		cp.cp_eax = CPUID_LEAF_EXT_21;
8128 		(void) __cpuid_insn(&cp);
8129 		platform_cpuid_mangle(cpi->cpi_vendor, CPUID_LEAF_EXT_21, &cp);
8130 		cpi->cpi_extd[0x21] = cp;
8131 		break;
8132 
8133 	default:
8134 		/*
8135 		 * Nothing to do here. Return an empty set which has already
8136 		 * been zeroed for us.
8137 		 */
8138 		return;
8139 	}
8140 
8141 	cpuid_scan_security(cpu, fset);
8142 }
8143 
8144 /* ARGSUSED */
8145 static int
8146 cpuid_post_ucodeadm_xc(xc_arg_t arg0, xc_arg_t arg1, xc_arg_t arg2)
8147 {
8148 	uchar_t *fset;
8149 	boolean_t first_pass = (boolean_t)arg1;
8150 
8151 	fset = (uchar_t *)(arg0 + sizeof (x86_featureset) * CPU->cpu_id);
8152 	if (first_pass && CPU->cpu_id != 0)
8153 		return (0);
8154 	if (!first_pass && CPU->cpu_id == 0)
8155 		return (0);
8156 	cpuid_pass_ucode(CPU, fset);
8157 
8158 	return (0);
8159 }
8160 
8161 /*
8162  * After a microcode update where the version has changed, then we need to
8163  * rescan CPUID. To do this we check every CPU to make sure that they have the
8164  * same microcode. Then we perform a cross call to all such CPUs. It's the
8165  * caller's job to make sure that no one else can end up doing an update while
8166  * this is going on.
8167  *
8168  * We assume that the system is microcode capable if we're called.
8169  */
8170 void
8171 cpuid_post_ucodeadm(void)
8172 {
8173 	uint32_t rev;
8174 	int i;
8175 	struct cpu *cpu;
8176 	cpuset_t cpuset;
8177 	void *argdata;
8178 	uchar_t *f0;
8179 
8180 	argdata = kmem_zalloc(sizeof (x86_featureset) * NCPU, KM_SLEEP);
8181 
8182 	mutex_enter(&cpu_lock);
8183 	cpu = cpu_get(0);
8184 	rev = cpu->cpu_m.mcpu_ucode_info->cui_rev;
8185 	CPUSET_ONLY(cpuset, 0);
8186 	for (i = 1; i < max_ncpus; i++) {
8187 		if ((cpu = cpu_get(i)) == NULL)
8188 			continue;
8189 
8190 		if (cpu->cpu_m.mcpu_ucode_info->cui_rev != rev) {
8191 			panic("post microcode update CPU %d has differing "
8192 			    "microcode revision (%u) from CPU 0 (%u)",
8193 			    i, cpu->cpu_m.mcpu_ucode_info->cui_rev, rev);
8194 		}
8195 		CPUSET_ADD(cpuset, i);
8196 	}
8197 
8198 	/*
8199 	 * We do the cross calls in two passes. The first pass is only for the
8200 	 * boot CPU. The second pass is for all of the other CPUs. This allows
8201 	 * the boot CPU to go through and change behavior related to patching or
8202 	 * whether or not Enhanced IBRS needs to be enabled and then allow all
8203 	 * other CPUs to follow suit.
8204 	 */
8205 	kpreempt_disable();
8206 	xc_sync((xc_arg_t)argdata, B_TRUE, 0, CPUSET2BV(cpuset),
8207 	    cpuid_post_ucodeadm_xc);
8208 	xc_sync((xc_arg_t)argdata, B_FALSE, 0, CPUSET2BV(cpuset),
8209 	    cpuid_post_ucodeadm_xc);
8210 	kpreempt_enable();
8211 
8212 	/*
8213 	 * OK, now look at each CPU and see if their feature sets are equal.
8214 	 */
8215 	f0 = argdata;
8216 	for (i = 1; i < max_ncpus; i++) {
8217 		uchar_t *fset;
8218 		if (!CPU_IN_SET(cpuset, i))
8219 			continue;
8220 
8221 		fset = (uchar_t *)((uintptr_t)argdata +
8222 		    sizeof (x86_featureset) * i);
8223 
8224 		if (!compare_x86_featureset(f0, fset)) {
8225 			panic("Post microcode update CPU %d has "
8226 			    "differing security feature (%p) set from CPU 0 "
8227 			    "(%p), not appending to feature set", i,
8228 			    (void *)fset, (void *)f0);
8229 		}
8230 	}
8231 
8232 	mutex_exit(&cpu_lock);
8233 
8234 	for (i = 0; i < NUM_X86_FEATURES; i++) {
8235 		cmn_err(CE_CONT, "?post-ucode x86_feature: %s\n",
8236 		    x86_feature_names[i]);
8237 		if (is_x86_feature(f0, i)) {
8238 			add_x86_feature(x86_featureset, i);
8239 		}
8240 	}
8241 	kmem_free(argdata, sizeof (x86_featureset) * NCPU);
8242 }
8243 
8244 typedef void (*cpuid_pass_f)(cpu_t *, void *);
8245 
8246 typedef struct cpuid_pass_def {
8247 	cpuid_pass_t cpd_pass;
8248 	cpuid_pass_f cpd_func;
8249 } cpuid_pass_def_t;
8250 
8251 /*
8252  * See block comment at the top; note that cpuid_pass_ucode is not a pass in the
8253  * normal sense and should not appear here.
8254  */
8255 static const cpuid_pass_def_t cpuid_pass_defs[] = {
8256 	{ CPUID_PASS_PRELUDE, cpuid_pass_prelude },
8257 	{ CPUID_PASS_IDENT, cpuid_pass_ident },
8258 	{ CPUID_PASS_BASIC, cpuid_pass_basic },
8259 	{ CPUID_PASS_EXTENDED, cpuid_pass_extended },
8260 	{ CPUID_PASS_DYNAMIC, cpuid_pass_dynamic },
8261 	{ CPUID_PASS_RESOLVE, cpuid_pass_resolve },
8262 };
8263 
8264 void
8265 cpuid_execpass(cpu_t *cp, cpuid_pass_t pass, void *arg)
8266 {
8267 	VERIFY3S(pass, !=, CPUID_PASS_NONE);
8268 
8269 	if (cp == NULL)
8270 		cp = CPU;
8271 
8272 	/*
8273 	 * Space statically allocated for BSP, ensure pointer is set
8274 	 */
8275 	if (cp->cpu_id == 0 && cp->cpu_m.mcpu_cpi == NULL)
8276 		cp->cpu_m.mcpu_cpi = &cpuid_info0;
8277 
8278 	ASSERT(cpuid_checkpass(cp, pass - 1));
8279 
8280 	for (uint_t i = 0; i < ARRAY_SIZE(cpuid_pass_defs); i++) {
8281 		if (cpuid_pass_defs[i].cpd_pass == pass) {
8282 			cpuid_pass_defs[i].cpd_func(cp, arg);
8283 			cp->cpu_m.mcpu_cpi->cpi_pass = pass;
8284 			return;
8285 		}
8286 	}
8287 
8288 	panic("unable to execute invalid cpuid pass %d on cpu%d\n",
8289 	    pass, cp->cpu_id);
8290 }
8291 
8292 /*
8293  * Extract the processor family from a chiprev.  Processor families are not the
8294  * same as cpuid families; see comments above and in x86_archext.h.
8295  */
8296 x86_processor_family_t
8297 chiprev_family(const x86_chiprev_t cr)
8298 {
8299 	return ((x86_processor_family_t)_X86_CHIPREV_FAMILY(cr));
8300 }
8301 
8302 /*
8303  * A chiprev matches its template if the vendor and family are identical and the
8304  * revision of the chiprev matches one of the bits set in the template.  Callers
8305  * may bitwise-OR together chiprevs of the same vendor and family to form the
8306  * template, or use the _ANY variant.  It is not possible to match chiprevs of
8307  * multiple vendors or processor families with a single call.  Note that this
8308  * function operates on processor families, not cpuid families.
8309  */
8310 boolean_t
8311 chiprev_matches(const x86_chiprev_t cr, const x86_chiprev_t template)
8312 {
8313 	return (_X86_CHIPREV_VENDOR(cr) == _X86_CHIPREV_VENDOR(template) &&
8314 	    _X86_CHIPREV_FAMILY(cr) == _X86_CHIPREV_FAMILY(template) &&
8315 	    (_X86_CHIPREV_REV(cr) & _X86_CHIPREV_REV(template)) != 0);
8316 }
8317 
8318 /*
8319  * A chiprev is at least min if the vendor and family are identical and the
8320  * revision of the chiprev is at least as recent as that of min.  Processor
8321  * families are considered unordered and cannot be compared using this function.
8322  * Note that this function operates on processor families, not cpuid families.
8323  * Use of the _ANY chiprev variant with this function is not useful; it will
8324  * always return B_FALSE if the _ANY variant is supplied as the minimum
8325  * revision.  To determine only whether a chiprev is of a given processor
8326  * family, test the return value of chiprev_family() instead.
8327  */
8328 boolean_t
8329 chiprev_at_least(const x86_chiprev_t cr, const x86_chiprev_t min)
8330 {
8331 	return (_X86_CHIPREV_VENDOR(cr) == _X86_CHIPREV_VENDOR(min) &&
8332 	    _X86_CHIPREV_FAMILY(cr) == _X86_CHIPREV_FAMILY(min) &&
8333 	    _X86_CHIPREV_REV(cr) >= _X86_CHIPREV_REV(min));
8334 }
8335 
8336 /*
8337  * The uarch functions operate in a manner similar to the chiprev functions
8338  * above.  While it is tempting to allow these to operate on microarchitectures
8339  * produced by a specific vendor in an ordered fashion (e.g., ZEN3 is "newer"
8340  * than ZEN2), we elect not to do so because a manufacturer may supply
8341  * processors of multiple different microarchitecture families each of which may
8342  * be internally ordered but unordered with respect to those of other families.
8343  */
8344 x86_uarch_t
8345 uarchrev_uarch(const x86_uarchrev_t ur)
8346 {
8347 	return ((x86_uarch_t)_X86_UARCHREV_UARCH(ur));
8348 }
8349 
8350 boolean_t
8351 uarchrev_matches(const x86_uarchrev_t ur, const x86_uarchrev_t template)
8352 {
8353 	return (_X86_UARCHREV_VENDOR(ur) == _X86_UARCHREV_VENDOR(template) &&
8354 	    _X86_UARCHREV_UARCH(ur) == _X86_UARCHREV_UARCH(template) &&
8355 	    (_X86_UARCHREV_REV(ur) & _X86_UARCHREV_REV(template)) != 0);
8356 }
8357 
8358 boolean_t
8359 uarchrev_at_least(const x86_uarchrev_t ur, const x86_uarchrev_t min)
8360 {
8361 	return (_X86_UARCHREV_VENDOR(ur) == _X86_UARCHREV_VENDOR(min) &&
8362 	    _X86_UARCHREV_UARCH(ur) == _X86_UARCHREV_UARCH(min) &&
8363 	    _X86_UARCHREV_REV(ur) >= _X86_UARCHREV_REV(min));
8364 }
8365 
8366 /*
8367  * Topology cache related information. This is yet another cache interface that
8368  * we're exposing out intended to be used when we have either Intel Leaf 4 or
8369  * AMD Leaf 8x1D (introduced with Zen 1).
8370  */
8371 static boolean_t
8372 cpuid_cache_topo_sup(const struct cpuid_info *cpi)
8373 {
8374 	switch (cpi->cpi_vendor) {
8375 	case X86_VENDOR_Intel:
8376 		if (cpi->cpi_maxeax >= 4) {
8377 			return (B_TRUE);
8378 		}
8379 		break;
8380 	case X86_VENDOR_AMD:
8381 	case X86_VENDOR_HYGON:
8382 		if (cpi->cpi_xmaxeax >= CPUID_LEAF_EXT_1d &&
8383 		    is_x86_feature(x86_featureset, X86FSET_TOPOEXT)) {
8384 			return (B_TRUE);
8385 		}
8386 		break;
8387 	default:
8388 		break;
8389 	}
8390 
8391 	return (B_FALSE);
8392 }
8393 
8394 int
8395 cpuid_getncaches(struct cpu *cpu, uint32_t *ncache)
8396 {
8397 	const struct cpuid_info *cpi;
8398 
8399 	ASSERT(cpuid_checkpass(cpu, CPUID_PASS_DYNAMIC));
8400 	cpi = cpu->cpu_m.mcpu_cpi;
8401 
8402 	if (!cpuid_cache_topo_sup(cpi)) {
8403 		return (ENOTSUP);
8404 	}
8405 
8406 	*ncache = cpi->cpi_cache_leaf_size;
8407 	return (0);
8408 }
8409 
8410 int
8411 cpuid_getcache(struct cpu *cpu, uint32_t cno, x86_cache_t *cache)
8412 {
8413 	const struct cpuid_info *cpi;
8414 	const struct cpuid_regs *cp;
8415 
8416 	ASSERT(cpuid_checkpass(cpu, CPUID_PASS_DYNAMIC));
8417 	cpi = cpu->cpu_m.mcpu_cpi;
8418 
8419 	if (!cpuid_cache_topo_sup(cpi)) {
8420 		return (ENOTSUP);
8421 	}
8422 
8423 	if (cno >= cpi->cpi_cache_leaf_size) {
8424 		return (EINVAL);
8425 	}
8426 
8427 	bzero(cache, sizeof (cache));
8428 	cp = cpi->cpi_cache_leaves[cno];
8429 	switch (CPI_CACHE_TYPE(cp)) {
8430 	case CPI_CACHE_TYPE_DATA:
8431 		cache->xc_type = X86_CACHE_TYPE_DATA;
8432 		break;
8433 	case CPI_CACHE_TYPE_INSTR:
8434 		cache->xc_type = X86_CACHE_TYPE_INST;
8435 		break;
8436 	case CPI_CACHE_TYPE_UNIFIED:
8437 		cache->xc_type = X86_CACHE_TYPE_UNIFIED;
8438 		break;
8439 	case CPI_CACHE_TYPE_DONE:
8440 	default:
8441 		return (EINVAL);
8442 	}
8443 	cache->xc_level = CPI_CACHE_LVL(cp);
8444 	if (CPI_FULL_ASSOC_CACHE(cp) != 0) {
8445 		cache->xc_flags |= X86_CACHE_F_FULL_ASSOC;
8446 	}
8447 	cache->xc_nparts = CPI_CACHE_PARTS(cp) + 1;
8448 	/*
8449 	 * The number of sets is reserved on AMD if the CPU is tagged as fully
8450 	 * associative, where as it is considered valid on Intel.
8451 	 */
8452 	if (cpi->cpi_vendor == X86_VENDOR_AMD &&
8453 	    CPI_FULL_ASSOC_CACHE(cp) != 0) {
8454 		cache->xc_nsets = 1;
8455 	} else {
8456 		cache->xc_nsets = CPI_CACHE_SETS(cp) + 1;
8457 	}
8458 	cache->xc_nways = CPI_CACHE_WAYS(cp) + 1;
8459 	cache->xc_line_size = CPI_CACHE_COH_LN_SZ(cp) + 1;
8460 	cache->xc_size = cache->xc_nparts * cache->xc_nsets * cache->xc_nways *
8461 	    cache->xc_line_size;
8462 	/*
8463 	 * We're looking for the number of bits to cover the number of CPUs that
8464 	 * are being shared. Normally this would be the value - 1, but the CPUID
8465 	 * value is encoded as the actual value minus one, so we don't modify
8466 	 * this at all.
8467 	 */
8468 	cache->xc_apic_shift = highbit(CPI_NTHR_SHR_CACHE(cp));
8469 
8470 	/*
8471 	 * To construct a unique ID we construct a uint64_t that looks as
8472 	 * follows:
8473 	 *
8474 	 * [47:40] cache level
8475 	 * [39:32] CPUID cache type
8476 	 * [31:00] shifted APIC ID
8477 	 *
8478 	 * The shifted APIC ID gives us a guarantee that a given cache entry is
8479 	 * unique within its peers. The other two numbers give us something that
8480 	 * ensures that something is unique within the CPU. If we just had the
8481 	 * APIC ID shifted over by the indicated number of bits we'd end up with
8482 	 * an ID of zero for the L1I, L1D, L2, and L3.
8483 	 *
8484 	 * The format of this ID is private to the system and can change across
8485 	 * a reboot for the time being.
8486 	 */
8487 	cache->xc_id = (uint64_t)cache->xc_level << 40;
8488 	cache->xc_id |= (uint64_t)cache->xc_type << 32;
8489 	cache->xc_id |= (uint64_t)cpi->cpi_apicid >> cache->xc_apic_shift;
8490 
8491 	return (0);
8492 }
8493