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
is_x86_feature(void * featureset,uint_t feature)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
add_x86_feature(void * featureset,uint_t feature)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
remove_x86_feature(void * featureset,uint_t feature)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
compare_x86_featureset(void * setA,void * setB)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
print_x86_featureset(void * featureset)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
platform_cpuid_mangle(uint_t vendor,uint32_t eax,struct cpuid_regs * cp)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
cpuid_alloc_space(cpu_t * cpu)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
cpuid_free_space(cpu_t * cpu)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
determine_platform(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 } else {
2255 /*
2256 * Check older VMware hardware versions. VMware hypervisor is
2257 * detected by performing an IN operation to VMware hypervisor
2258 * port and checking that value returned in %ebx is VMware
2259 * hypervisor magic value.
2260 *
2261 * References: http://kb.vmware.com/kb/1009458
2262 */
2263 vmware_port(VMWARE_HVCMD_GETVERSION, regs);
2264 if (regs[1] == VMWARE_HVMAGIC) {
2265 platform_type = HW_VMWARE;
2266 return;
2267 }
2268 }
2269
2270 /*
2271 * Check Xen hypervisor. In a fully virtualized domain,
2272 * Xen's pseudo-cpuid function returns a string representing the
2273 * Xen signature in %ebx, %ecx, and %edx. %eax contains the maximum
2274 * supported cpuid function. We need at least a (base + 2) leaf value
2275 * to do what we want to do. Try different base values, since the
2276 * hypervisor might use a different one depending on whether Hyper-V
2277 * emulation is switched on by default or not.
2278 */
2279 for (base = 0x40000000; base < 0x40010000; base += 0x100) {
2280 cp.cp_eax = base;
2281 (void) __cpuid_insn(&cp);
2282 regs[0] = cp.cp_ebx;
2283 regs[1] = cp.cp_ecx;
2284 regs[2] = cp.cp_edx;
2285 regs[3] = 0;
2286 if (strcmp(hvstr, HVSIG_XEN_HVM) == 0 &&
2287 cp.cp_eax >= (base + 2)) {
2288 platform_type &= ~HW_NATIVE;
2289 platform_type |= HW_XEN_HVM;
2290 return;
2291 }
2292 }
2293 }
2294
2295 int
get_hwenv(void)2296 get_hwenv(void)
2297 {
2298 ASSERT(platform_type != -1);
2299 return (platform_type);
2300 }
2301
2302 int
is_controldom(void)2303 is_controldom(void)
2304 {
2305 return (0);
2306 }
2307
2308 #else
2309
2310 int
get_hwenv(void)2311 get_hwenv(void)
2312 {
2313 return (HW_XEN_PV);
2314 }
2315
2316 int
is_controldom(void)2317 is_controldom(void)
2318 {
2319 return (DOMAIN_IS_INITDOMAIN(xen_info));
2320 }
2321
2322 #endif /* __xpv */
2323
2324 /*
2325 * Gather the extended topology information. This should be the same for both
2326 * AMD leaf 8X26 and Intel leaf 0x1F (though the data interpretation varies).
2327 */
2328 static void
cpuid_gather_ext_topo_leaf(struct cpuid_info * cpi,uint32_t leaf)2329 cpuid_gather_ext_topo_leaf(struct cpuid_info *cpi, uint32_t leaf)
2330 {
2331 uint_t i;
2332
2333 for (i = 0; i < ARRAY_SIZE(cpi->cpi_topo); i++) {
2334 struct cpuid_regs *regs = &cpi->cpi_topo[i];
2335
2336 bzero(regs, sizeof (struct cpuid_regs));
2337 regs->cp_eax = leaf;
2338 regs->cp_ecx = i;
2339
2340 (void) __cpuid_insn(regs);
2341 if (CPUID_AMD_8X26_ECX_TYPE(regs->cp_ecx) ==
2342 CPUID_AMD_8X26_TYPE_DONE) {
2343 break;
2344 }
2345 }
2346
2347 cpi->cpi_topo_nleaves = i;
2348 }
2349
2350 /*
2351 * Make sure that we have gathered all of the CPUID leaves that we might need to
2352 * determine topology. We assume that the standard leaf 1 has already been done
2353 * and that xmaxeax has already been calculated.
2354 */
2355 static void
cpuid_gather_amd_topology_leaves(cpu_t * cpu)2356 cpuid_gather_amd_topology_leaves(cpu_t *cpu)
2357 {
2358 struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2359
2360 if (cpi->cpi_xmaxeax >= CPUID_LEAF_EXT_8) {
2361 struct cpuid_regs *cp;
2362
2363 cp = &cpi->cpi_extd[8];
2364 cp->cp_eax = CPUID_LEAF_EXT_8;
2365 (void) __cpuid_insn(cp);
2366 platform_cpuid_mangle(cpi->cpi_vendor, CPUID_LEAF_EXT_8, cp);
2367 }
2368
2369 if (is_x86_feature(x86_featureset, X86FSET_TOPOEXT) &&
2370 cpi->cpi_xmaxeax >= CPUID_LEAF_EXT_1e) {
2371 struct cpuid_regs *cp;
2372
2373 cp = &cpi->cpi_extd[0x1e];
2374 cp->cp_eax = CPUID_LEAF_EXT_1e;
2375 (void) __cpuid_insn(cp);
2376 }
2377
2378 if (cpi->cpi_xmaxeax >= CPUID_LEAF_EXT_26) {
2379 cpuid_gather_ext_topo_leaf(cpi, CPUID_LEAF_EXT_26);
2380 }
2381 }
2382
2383 /*
2384 * Get the APIC ID for this processor. If Leaf B is present and valid, we prefer
2385 * it to everything else. If not, and we're on an AMD system where 8000001e is
2386 * valid, then we use that. Othewrise, we fall back to the default value for the
2387 * APIC ID in leaf 1.
2388 */
2389 static uint32_t
cpuid_gather_apicid(struct cpuid_info * cpi)2390 cpuid_gather_apicid(struct cpuid_info *cpi)
2391 {
2392 /*
2393 * Leaf B changes based on the arguments to it. Because we don't cache
2394 * it, we need to gather it again.
2395 */
2396 if (cpi->cpi_maxeax >= 0xB) {
2397 struct cpuid_regs regs;
2398 struct cpuid_regs *cp;
2399
2400 cp = ®s;
2401 cp->cp_eax = 0xB;
2402 cp->cp_edx = cp->cp_ebx = cp->cp_ecx = 0;
2403 (void) __cpuid_insn(cp);
2404
2405 if (cp->cp_ebx != 0) {
2406 return (cp->cp_edx);
2407 }
2408 }
2409
2410 if ((cpi->cpi_vendor == X86_VENDOR_AMD ||
2411 cpi->cpi_vendor == X86_VENDOR_HYGON) &&
2412 is_x86_feature(x86_featureset, X86FSET_TOPOEXT) &&
2413 cpi->cpi_xmaxeax >= CPUID_LEAF_EXT_1e) {
2414 return (cpi->cpi_extd[0x1e].cp_eax);
2415 }
2416
2417 return (CPI_APIC_ID(cpi));
2418 }
2419
2420 /*
2421 * For AMD processors, attempt to calculate the number of chips and cores that
2422 * exist. The way that we do this varies based on the generation, because the
2423 * generations themselves have changed dramatically.
2424 *
2425 * If cpuid leaf 0x80000008 exists, that generally tells us the number of cores.
2426 * However, with the advent of family 17h (Zen) it actually tells us the number
2427 * of threads, so we need to look at leaf 0x8000001e if available to determine
2428 * its value. Otherwise, for all prior families, the number of enabled cores is
2429 * the same as threads.
2430 *
2431 * If we do not have leaf 0x80000008, then we assume that this processor does
2432 * not have anything. AMD's older CPUID specification says there's no reason to
2433 * fall back to leaf 1.
2434 *
2435 * In some virtualization cases we will not have leaf 8000001e or it will be
2436 * zero. When that happens we assume the number of threads is one.
2437 */
2438 static void
cpuid_amd_ncores(struct cpuid_info * cpi,uint_t * ncpus,uint_t * ncores)2439 cpuid_amd_ncores(struct cpuid_info *cpi, uint_t *ncpus, uint_t *ncores)
2440 {
2441 uint_t nthreads, nthread_per_core;
2442
2443 nthreads = nthread_per_core = 1;
2444
2445 if (cpi->cpi_xmaxeax >= CPUID_LEAF_EXT_8) {
2446 nthreads = BITX(cpi->cpi_extd[8].cp_ecx, 7, 0) + 1;
2447 } else if ((cpi->cpi_std[1].cp_edx & CPUID_INTC_EDX_HTT) != 0) {
2448 nthreads = CPI_CPU_COUNT(cpi);
2449 }
2450
2451 /*
2452 * For us to have threads, and know about it, we have to be at least at
2453 * family 17h and have the cpuid bit that says we have extended
2454 * topology.
2455 */
2456 if (cpi->cpi_family >= 0x17 &&
2457 is_x86_feature(x86_featureset, X86FSET_TOPOEXT) &&
2458 cpi->cpi_xmaxeax >= CPUID_LEAF_EXT_1e) {
2459 nthread_per_core = BITX(cpi->cpi_extd[0x1e].cp_ebx, 15, 8) + 1;
2460 }
2461
2462 *ncpus = nthreads;
2463 *ncores = nthreads / nthread_per_core;
2464 }
2465
2466 /*
2467 * Seed the initial values for the cores and threads for an Intel based
2468 * processor. These values will be overwritten if we detect that the processor
2469 * supports CPUID leaf 0xb.
2470 */
2471 static void
cpuid_intel_ncores(struct cpuid_info * cpi,uint_t * ncpus,uint_t * ncores)2472 cpuid_intel_ncores(struct cpuid_info *cpi, uint_t *ncpus, uint_t *ncores)
2473 {
2474 /*
2475 * Only seed the number of physical cores from the first level leaf 4
2476 * information. The number of threads there indicate how many share the
2477 * L1 cache, which may or may not have anything to do with the number of
2478 * logical CPUs per core.
2479 */
2480 if (cpi->cpi_maxeax >= 4) {
2481 *ncores = BITX(cpi->cpi_std[4].cp_eax, 31, 26) + 1;
2482 } else {
2483 *ncores = 1;
2484 }
2485
2486 if ((cpi->cpi_std[1].cp_edx & CPUID_INTC_EDX_HTT) != 0) {
2487 *ncpus = CPI_CPU_COUNT(cpi);
2488 } else {
2489 *ncpus = *ncores;
2490 }
2491 }
2492
2493 static boolean_t
cpuid_leafB_getids(cpu_t * cpu)2494 cpuid_leafB_getids(cpu_t *cpu)
2495 {
2496 struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2497 struct cpuid_regs regs;
2498 struct cpuid_regs *cp;
2499
2500 if (cpi->cpi_maxeax < 0xB)
2501 return (B_FALSE);
2502
2503 cp = ®s;
2504 cp->cp_eax = 0xB;
2505 cp->cp_edx = cp->cp_ebx = cp->cp_ecx = 0;
2506
2507 (void) __cpuid_insn(cp);
2508
2509 /*
2510 * Check CPUID.EAX=0BH, ECX=0H:EBX is non-zero, which
2511 * indicates that the extended topology enumeration leaf is
2512 * available.
2513 */
2514 if (cp->cp_ebx != 0) {
2515 uint32_t x2apic_id = 0;
2516 uint_t coreid_shift = 0;
2517 uint_t ncpu_per_core = 1;
2518 uint_t chipid_shift = 0;
2519 uint_t ncpu_per_chip = 1;
2520 uint_t i;
2521 uint_t level;
2522
2523 for (i = 0; i < CPI_FNB_ECX_MAX; i++) {
2524 cp->cp_eax = 0xB;
2525 cp->cp_ecx = i;
2526
2527 (void) __cpuid_insn(cp);
2528 level = CPI_CPU_LEVEL_TYPE(cp);
2529
2530 if (level == 1) {
2531 x2apic_id = cp->cp_edx;
2532 coreid_shift = BITX(cp->cp_eax, 4, 0);
2533 ncpu_per_core = BITX(cp->cp_ebx, 15, 0);
2534 } else if (level == 2) {
2535 x2apic_id = cp->cp_edx;
2536 chipid_shift = BITX(cp->cp_eax, 4, 0);
2537 ncpu_per_chip = BITX(cp->cp_ebx, 15, 0);
2538 }
2539 }
2540
2541 /*
2542 * cpi_apicid is taken care of in cpuid_gather_apicid.
2543 */
2544 cpi->cpi_ncpu_per_chip = ncpu_per_chip;
2545 cpi->cpi_ncore_per_chip = ncpu_per_chip /
2546 ncpu_per_core;
2547 cpi->cpi_chipid = x2apic_id >> chipid_shift;
2548 cpi->cpi_clogid = x2apic_id & ((1 << chipid_shift) - 1);
2549 cpi->cpi_coreid = x2apic_id >> coreid_shift;
2550 cpi->cpi_pkgcoreid = cpi->cpi_clogid >> coreid_shift;
2551 cpi->cpi_procnodeid = cpi->cpi_chipid;
2552 cpi->cpi_compunitid = cpi->cpi_coreid;
2553
2554 if (coreid_shift > 0 && chipid_shift > coreid_shift) {
2555 cpi->cpi_nthread_bits = coreid_shift;
2556 cpi->cpi_ncore_bits = chipid_shift - coreid_shift;
2557 }
2558
2559 return (B_TRUE);
2560 } else {
2561 return (B_FALSE);
2562 }
2563 }
2564
2565 static void
cpuid_intel_getids(cpu_t * cpu,void * feature)2566 cpuid_intel_getids(cpu_t *cpu, void *feature)
2567 {
2568 uint_t i;
2569 uint_t chipid_shift = 0;
2570 uint_t coreid_shift = 0;
2571 struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2572
2573 /*
2574 * There are no compute units or processor nodes currently on Intel.
2575 * Always set these to one.
2576 */
2577 cpi->cpi_procnodes_per_pkg = 1;
2578 cpi->cpi_cores_per_compunit = 1;
2579
2580 /*
2581 * If cpuid Leaf B is present, use that to try and get this information.
2582 * It will be the most accurate for Intel CPUs.
2583 */
2584 if (cpuid_leafB_getids(cpu))
2585 return;
2586
2587 /*
2588 * In this case, we have the leaf 1 and leaf 4 values for ncpu_per_chip
2589 * and ncore_per_chip. These represent the largest power of two values
2590 * that we need to cover all of the IDs in the system. Therefore, we use
2591 * those values to seed the number of bits needed to cover information
2592 * in the case when leaf B is not available. These values will probably
2593 * be larger than required, but that's OK.
2594 */
2595 cpi->cpi_nthread_bits = ddi_fls(cpi->cpi_ncpu_per_chip);
2596 cpi->cpi_ncore_bits = ddi_fls(cpi->cpi_ncore_per_chip);
2597
2598 for (i = 1; i < cpi->cpi_ncpu_per_chip; i <<= 1)
2599 chipid_shift++;
2600
2601 cpi->cpi_chipid = cpi->cpi_apicid >> chipid_shift;
2602 cpi->cpi_clogid = cpi->cpi_apicid & ((1 << chipid_shift) - 1);
2603
2604 if (is_x86_feature(feature, X86FSET_CMP)) {
2605 /*
2606 * Multi-core (and possibly multi-threaded)
2607 * processors.
2608 */
2609 uint_t ncpu_per_core = 0;
2610
2611 if (cpi->cpi_ncore_per_chip == 1)
2612 ncpu_per_core = cpi->cpi_ncpu_per_chip;
2613 else if (cpi->cpi_ncore_per_chip > 1)
2614 ncpu_per_core = cpi->cpi_ncpu_per_chip /
2615 cpi->cpi_ncore_per_chip;
2616 /*
2617 * 8bit APIC IDs on dual core Pentiums
2618 * look like this:
2619 *
2620 * +-----------------------+------+------+
2621 * | Physical Package ID | MC | HT |
2622 * +-----------------------+------+------+
2623 * <------- chipid -------->
2624 * <------- coreid --------------->
2625 * <--- clogid -->
2626 * <------>
2627 * pkgcoreid
2628 *
2629 * Where the number of bits necessary to
2630 * represent MC and HT fields together equals
2631 * to the minimum number of bits necessary to
2632 * store the value of cpi->cpi_ncpu_per_chip.
2633 * Of those bits, the MC part uses the number
2634 * of bits necessary to store the value of
2635 * cpi->cpi_ncore_per_chip.
2636 */
2637 for (i = 1; i < ncpu_per_core; i <<= 1)
2638 coreid_shift++;
2639 cpi->cpi_coreid = cpi->cpi_apicid >> coreid_shift;
2640 cpi->cpi_pkgcoreid = cpi->cpi_clogid >> coreid_shift;
2641 } else if (is_x86_feature(feature, X86FSET_HTT)) {
2642 /*
2643 * Single-core multi-threaded processors.
2644 */
2645 cpi->cpi_coreid = cpi->cpi_chipid;
2646 cpi->cpi_pkgcoreid = 0;
2647 } else {
2648 /*
2649 * Single-core single-thread processors.
2650 */
2651 cpi->cpi_coreid = cpu->cpu_id;
2652 cpi->cpi_pkgcoreid = 0;
2653 }
2654 cpi->cpi_procnodeid = cpi->cpi_chipid;
2655 cpi->cpi_compunitid = cpi->cpi_coreid;
2656 }
2657
2658 /*
2659 * Historically, AMD has had CMP chips with only a single thread per core.
2660 * However, starting in family 17h (Zen), this has changed and they now have
2661 * multiple threads. Our internal core id needs to be a unique value.
2662 *
2663 * To determine the core id of an AMD system, if we're from a family before 17h,
2664 * then we just use the cpu id, as that gives us a good value that will be
2665 * unique for each core. If instead, we're on family 17h or later, then we need
2666 * to do something more complicated. CPUID leaf 0x8000001e can tell us
2667 * how many threads are in the system. Based on that, we'll shift the APIC ID.
2668 * We can't use the normal core id in that leaf as it's only unique within the
2669 * socket, which is perfect for cpi_pkgcoreid, but not us.
2670 */
2671 static id_t
cpuid_amd_get_coreid(cpu_t * cpu)2672 cpuid_amd_get_coreid(cpu_t *cpu)
2673 {
2674 struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2675
2676 if (cpi->cpi_family >= 0x17 &&
2677 is_x86_feature(x86_featureset, X86FSET_TOPOEXT) &&
2678 cpi->cpi_xmaxeax >= CPUID_LEAF_EXT_1e) {
2679 uint_t nthreads = BITX(cpi->cpi_extd[0x1e].cp_ebx, 15, 8) + 1;
2680 if (nthreads > 1) {
2681 VERIFY3U(nthreads, ==, 2);
2682 return (cpi->cpi_apicid >> 1);
2683 }
2684 }
2685
2686 return (cpu->cpu_id);
2687 }
2688
2689 /*
2690 * IDs on AMD is a more challenging task. This is notable because of the
2691 * following two facts:
2692 *
2693 * 1. Before family 0x17 (Zen), there was no support for SMT and there was
2694 * also no way to get an actual unique core id from the system. As such, we
2695 * synthesize this case by using cpu->cpu_id. This scheme does not,
2696 * however, guarantee that sibling cores of a chip will have sequential
2697 * coreids starting at a multiple of the number of cores per chip - that is
2698 * usually the case, but if the APIC IDs have been set up in a different
2699 * order then we need to perform a few more gymnastics for the pkgcoreid.
2700 *
2701 * 2. In families 0x15 and 16x (Bulldozer and co.) the cores came in groups
2702 * called compute units. These compute units share the L1I cache, L2 cache,
2703 * and the FPU. To deal with this, a new topology leaf was added in
2704 * 0x8000001e. However, parts of this leaf have different meanings
2705 * once we get to family 0x17.
2706 */
2707
2708 static void
cpuid_amd_getids(cpu_t * cpu,uchar_t * features)2709 cpuid_amd_getids(cpu_t *cpu, uchar_t *features)
2710 {
2711 int i, first_half, coreidsz;
2712 uint32_t nb_caps_reg;
2713 uint_t node2_1;
2714 struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2715 struct cpuid_regs *cp;
2716
2717 /*
2718 * Calculate the core id (this comes from hardware in family 0x17 if it
2719 * hasn't been stripped by virtualization). We always set the compute
2720 * unit id to the same value. Also, initialize the default number of
2721 * cores per compute unit and nodes per package. This will be
2722 * overwritten when we know information about a particular family.
2723 */
2724 cpi->cpi_coreid = cpuid_amd_get_coreid(cpu);
2725 cpi->cpi_compunitid = cpi->cpi_coreid;
2726 cpi->cpi_cores_per_compunit = 1;
2727 cpi->cpi_procnodes_per_pkg = 1;
2728
2729 /*
2730 * To construct the logical ID, we need to determine how many APIC IDs
2731 * are dedicated to the cores and threads. This is provided for us in
2732 * 0x80000008. However, if it's not present (say due to virtualization),
2733 * then we assume it's one. This should be present on all 64-bit AMD
2734 * processors. It was added in family 0xf (Hammer).
2735 */
2736 if (cpi->cpi_xmaxeax >= CPUID_LEAF_EXT_8) {
2737 coreidsz = BITX((cpi)->cpi_extd[8].cp_ecx, 15, 12);
2738
2739 /*
2740 * In AMD parlance chip is really a node while illumos
2741 * uses chip as equivalent to socket/package.
2742 */
2743 if (coreidsz == 0) {
2744 /* Use legacy method */
2745 for (i = 1; i < cpi->cpi_ncore_per_chip; i <<= 1)
2746 coreidsz++;
2747 if (coreidsz == 0)
2748 coreidsz = 1;
2749 }
2750 } else {
2751 /* Assume single-core part */
2752 coreidsz = 1;
2753 }
2754 cpi->cpi_clogid = cpi->cpi_apicid & ((1 << coreidsz) - 1);
2755
2756 /*
2757 * The package core ID varies depending on the family. While it may be
2758 * tempting to use the CPUID_LEAF_EXT_1e %ebx core id, unfortunately,
2759 * this value is the core id in the given node. For non-virtualized
2760 * family 17h, we need to take the logical core id and shift off the
2761 * threads like we do when getting the core id. Otherwise, we can use
2762 * the clogid as is. When family 17h is virtualized, the clogid should
2763 * be sufficient as if we don't have valid data in the leaf, then we
2764 * won't think we have SMT, in which case the cpi_clogid should be
2765 * sufficient.
2766 */
2767 if (cpi->cpi_family >= 0x17 &&
2768 is_x86_feature(x86_featureset, X86FSET_TOPOEXT) &&
2769 cpi->cpi_xmaxeax >= CPUID_LEAF_EXT_1e &&
2770 cpi->cpi_extd[0x1e].cp_ebx != 0) {
2771 uint_t nthreads = BITX(cpi->cpi_extd[0x1e].cp_ebx, 15, 8) + 1;
2772 if (nthreads > 1) {
2773 VERIFY3U(nthreads, ==, 2);
2774 cpi->cpi_pkgcoreid = cpi->cpi_clogid >> 1;
2775 } else {
2776 cpi->cpi_pkgcoreid = cpi->cpi_clogid;
2777 }
2778 } else {
2779 cpi->cpi_pkgcoreid = cpi->cpi_clogid;
2780 }
2781
2782 /*
2783 * Obtain the node ID and compute unit IDs. If we're on family 0x15
2784 * (bulldozer) or newer, then we can derive all of this from leaf
2785 * CPUID_LEAF_EXT_1e. Otherwise, the method varies by family.
2786 */
2787 if (is_x86_feature(x86_featureset, X86FSET_TOPOEXT) &&
2788 cpi->cpi_xmaxeax >= CPUID_LEAF_EXT_1e) {
2789 cp = &cpi->cpi_extd[0x1e];
2790
2791 cpi->cpi_procnodes_per_pkg = BITX(cp->cp_ecx, 10, 8) + 1;
2792 cpi->cpi_procnodeid = BITX(cp->cp_ecx, 7, 0);
2793
2794 /*
2795 * For Bulldozer-era CPUs, recalculate the compute unit
2796 * information.
2797 */
2798 if (cpi->cpi_family >= 0x15 && cpi->cpi_family < 0x17) {
2799 cpi->cpi_cores_per_compunit =
2800 BITX(cp->cp_ebx, 15, 8) + 1;
2801 cpi->cpi_compunitid = BITX(cp->cp_ebx, 7, 0) +
2802 (cpi->cpi_ncore_per_chip /
2803 cpi->cpi_cores_per_compunit) *
2804 (cpi->cpi_procnodeid /
2805 cpi->cpi_procnodes_per_pkg);
2806 }
2807 } else if (cpi->cpi_family == 0xf || cpi->cpi_family >= 0x11) {
2808 cpi->cpi_procnodeid = (cpi->cpi_apicid >> coreidsz) & 7;
2809 } else if (cpi->cpi_family == 0x10) {
2810 /*
2811 * See if we are a multi-node processor.
2812 * All processors in the system have the same number of nodes
2813 */
2814 nb_caps_reg = pci_getl_func(0, 24, 3, 0xe8);
2815 if ((cpi->cpi_model < 8) || BITX(nb_caps_reg, 29, 29) == 0) {
2816 /* Single-node */
2817 cpi->cpi_procnodeid = BITX(cpi->cpi_apicid, 5,
2818 coreidsz);
2819 } else {
2820
2821 /*
2822 * Multi-node revision D (2 nodes per package
2823 * are supported)
2824 */
2825 cpi->cpi_procnodes_per_pkg = 2;
2826
2827 first_half = (cpi->cpi_pkgcoreid <=
2828 (cpi->cpi_ncore_per_chip/2 - 1));
2829
2830 if (cpi->cpi_apicid == cpi->cpi_pkgcoreid) {
2831 /* We are BSP */
2832 cpi->cpi_procnodeid = (first_half ? 0 : 1);
2833 } else {
2834
2835 /* We are AP */
2836 /* NodeId[2:1] bits to use for reading F3xe8 */
2837 node2_1 = BITX(cpi->cpi_apicid, 5, 4) << 1;
2838
2839 nb_caps_reg =
2840 pci_getl_func(0, 24 + node2_1, 3, 0xe8);
2841
2842 /*
2843 * Check IntNodeNum bit (31:30, but bit 31 is
2844 * always 0 on dual-node processors)
2845 */
2846 if (BITX(nb_caps_reg, 30, 30) == 0)
2847 cpi->cpi_procnodeid = node2_1 +
2848 !first_half;
2849 else
2850 cpi->cpi_procnodeid = node2_1 +
2851 first_half;
2852 }
2853 }
2854 } else {
2855 cpi->cpi_procnodeid = 0;
2856 }
2857
2858 cpi->cpi_chipid =
2859 cpi->cpi_procnodeid / cpi->cpi_procnodes_per_pkg;
2860
2861 cpi->cpi_ncore_bits = coreidsz;
2862 cpi->cpi_nthread_bits = ddi_fls(cpi->cpi_ncpu_per_chip /
2863 cpi->cpi_ncore_per_chip);
2864 }
2865
2866 static void
spec_uarch_flush_noop(void)2867 spec_uarch_flush_noop(void)
2868 {
2869 }
2870
2871 /*
2872 * When microcode is present that mitigates MDS, this wrmsr will also flush the
2873 * MDS-related micro-architectural state that would normally happen by calling
2874 * x86_md_clear().
2875 */
2876 static void
spec_uarch_flush_msr(void)2877 spec_uarch_flush_msr(void)
2878 {
2879 wrmsr(MSR_IA32_FLUSH_CMD, IA32_FLUSH_CMD_L1D);
2880 }
2881
2882 /*
2883 * This function points to a function that will flush certain
2884 * micro-architectural state on the processor. This flush is used to mitigate
2885 * three different classes of Intel CPU vulnerabilities: L1TF, MDS, and RFDS.
2886 * This function can point to one of three functions:
2887 *
2888 * - A noop which is done because we either are vulnerable, but do not have
2889 * microcode available to help deal with a fix, or because we aren't
2890 * vulnerable.
2891 *
2892 * - spec_uarch_flush_msr which will issue an L1D flush and if microcode to
2893 * mitigate MDS is present, also perform the equivalent of the MDS flush;
2894 * however, it only flushes the MDS related micro-architectural state on the
2895 * current hyperthread, it does not do anything for the twin.
2896 *
2897 * - x86_md_clear which will flush the MDS related state. This is done when we
2898 * have a processor that is vulnerable to MDS, but is not vulnerable to L1TF
2899 * (RDCL_NO is set); or if the CPU is vulnerable to RFDS and indicates VERW
2900 * can clear it (RFDS_CLEAR is set).
2901 */
2902 void (*spec_uarch_flush)(void) = spec_uarch_flush_noop;
2903
2904 static void
cpuid_update_md_clear(cpu_t * cpu,uchar_t * featureset)2905 cpuid_update_md_clear(cpu_t *cpu, uchar_t *featureset)
2906 {
2907 struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2908
2909 /* Non-Intel doesn't concern us here. */
2910 if (cpi->cpi_vendor != X86_VENDOR_Intel)
2911 return;
2912
2913 /*
2914 * While RDCL_NO indicates that one of the MDS vulnerabilities (MSBDS)
2915 * has been fixed in hardware, it doesn't cover everything related to
2916 * MDS. Therefore we can only rely on MDS_NO to determine that we don't
2917 * need to mitigate this.
2918 *
2919 * We must ALSO check the case of RFDS_NO and if RFDS_CLEAR is set,
2920 * because of the small cases of RFDS.
2921 */
2922
2923 if ((!is_x86_feature(featureset, X86FSET_MDS_NO) &&
2924 is_x86_feature(featureset, X86FSET_MD_CLEAR)) ||
2925 (!is_x86_feature(featureset, X86FSET_RFDS_NO) &&
2926 is_x86_feature(featureset, X86FSET_RFDS_CLEAR))) {
2927 const uint8_t nop = NOP_INSTR;
2928 uint8_t *md = (uint8_t *)x86_md_clear;
2929
2930 *md = nop;
2931 }
2932
2933 membar_producer();
2934 }
2935
2936 static void
cpuid_update_l1d_flush(cpu_t * cpu,uchar_t * featureset)2937 cpuid_update_l1d_flush(cpu_t *cpu, uchar_t *featureset)
2938 {
2939 boolean_t need_l1d, need_mds, need_rfds;
2940 struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
2941
2942 /*
2943 * If we're not on Intel or we've mitigated all of RDCL, MDS, and RFDS
2944 * in hardware, then there's nothing left for us to do for enabling
2945 * the flush. We can also go ahead and say that SMT exclusion is
2946 * unnecessary.
2947 */
2948 if (cpi->cpi_vendor != X86_VENDOR_Intel ||
2949 (is_x86_feature(featureset, X86FSET_RDCL_NO) &&
2950 is_x86_feature(featureset, X86FSET_MDS_NO) &&
2951 is_x86_feature(featureset, X86FSET_RFDS_NO))) {
2952 extern int smt_exclusion;
2953 smt_exclusion = 0;
2954 spec_uarch_flush = spec_uarch_flush_noop;
2955 membar_producer();
2956 return;
2957 }
2958
2959 /*
2960 * The locations where we need to perform an L1D flush are required both
2961 * for mitigating L1TF and MDS. When verw support is present in
2962 * microcode, then the L1D flush will take care of doing that as well.
2963 * However, if we have a system where RDCL_NO is present, but we don't
2964 * have MDS_NO, then we need to do a verw (x86_md_clear) and not a full
2965 * L1D flush.
2966 */
2967 if (!is_x86_feature(featureset, X86FSET_RDCL_NO) &&
2968 is_x86_feature(featureset, X86FSET_FLUSH_CMD) &&
2969 !is_x86_feature(featureset, X86FSET_L1D_VM_NO)) {
2970 need_l1d = B_TRUE;
2971 } else {
2972 need_l1d = B_FALSE;
2973 }
2974
2975 if (!is_x86_feature(featureset, X86FSET_MDS_NO) &&
2976 is_x86_feature(featureset, X86FSET_MD_CLEAR)) {
2977 need_mds = B_TRUE;
2978 } else {
2979 need_mds = B_FALSE;
2980 }
2981
2982 if (!is_x86_feature(featureset, X86FSET_RFDS_NO) &&
2983 is_x86_feature(featureset, X86FSET_RFDS_CLEAR)) {
2984 need_rfds = B_TRUE;
2985 } else {
2986 need_rfds = B_FALSE;
2987 }
2988
2989 if (need_l1d) {
2990 /*
2991 * As of Feb, 2024, no CPU needs L1D *and* RFDS mitigation
2992 * together. If the following VERIFY trips, we need to add
2993 * further fixes here.
2994 */
2995 VERIFY(!need_rfds);
2996 spec_uarch_flush = spec_uarch_flush_msr;
2997 } else if (need_mds || need_rfds) {
2998 spec_uarch_flush = x86_md_clear;
2999 } else {
3000 /*
3001 * We have no hardware mitigations available to us.
3002 */
3003 spec_uarch_flush = spec_uarch_flush_noop;
3004 }
3005 membar_producer();
3006 }
3007
3008 /*
3009 * Branch History Injection (BHI) mitigations.
3010 *
3011 * Intel has provided a software sequence that will scrub the BHB. Like RSB
3012 * (below) we can scribble a return at the beginning to avoid if if the CPU
3013 * is modern enough. We can also scribble a return if the CPU is old enough
3014 * to not have an RSB (pre-eIBRS).
3015 */
3016 typedef enum {
3017 X86_BHI_TOO_OLD_OR_DISABLED, /* Pre-eIBRS or disabled */
3018 X86_BHI_NEW_ENOUGH, /* AMD, or Intel with BHI_NO set */
3019 X86_BHI_DIS_S, /* BHI_NO == 0, but BHI_DIS_S avail. */
3020 /* NOTE: BHI_DIS_S above will still need the software sequence. */
3021 X86_BHI_SOFTWARE_SEQUENCE, /* Use software sequence */
3022 } x86_native_bhi_mitigation_t;
3023
3024 x86_native_bhi_mitigation_t x86_bhi_mitigation = X86_BHI_SOFTWARE_SEQUENCE;
3025
3026 static void
cpuid_enable_bhi_dis_s(void)3027 cpuid_enable_bhi_dis_s(void)
3028 {
3029 uint64_t val;
3030
3031 val = rdmsr(MSR_IA32_SPEC_CTRL);
3032 val |= IA32_SPEC_CTRL_BHI_DIS_S;
3033 wrmsr(MSR_IA32_SPEC_CTRL, val);
3034 }
3035
3036 /*
3037 * This function scribbles RET into the first instruction of x86_bhb_clear()
3038 * if SPECTREV2 mitigations are disabled, the CPU is too old, the CPU is new
3039 * enough to fix (which includes non-Intel CPUs), or the CPU has an explicit
3040 * disable-Branch-History control.
3041 */
3042 static x86_native_bhi_mitigation_t
cpuid_learn_and_patch_bhi(x86_spectrev2_mitigation_t v2mit,cpu_t * cpu,uchar_t * featureset)3043 cpuid_learn_and_patch_bhi(x86_spectrev2_mitigation_t v2mit, cpu_t *cpu,
3044 uchar_t *featureset)
3045 {
3046 struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
3047 const uint8_t ret = RET_INSTR;
3048 uint8_t *bhb_clear = (uint8_t *)x86_bhb_clear;
3049
3050 ASSERT0(cpu->cpu_id);
3051
3052 /* First check for explicitly disabled... */
3053 if (v2mit == X86_SPECTREV2_DISABLED) {
3054 *bhb_clear = ret;
3055 return (X86_BHI_TOO_OLD_OR_DISABLED);
3056 }
3057
3058 /*
3059 * Then check for BHI_NO, which means the CPU doesn't have this bug,
3060 * or if it's non-Intel, in which case this mitigation mechanism
3061 * doesn't apply.
3062 */
3063 if (cpi->cpi_vendor != X86_VENDOR_Intel ||
3064 is_x86_feature(featureset, X86FSET_BHI_NO)) {
3065 *bhb_clear = ret;
3066 return (X86_BHI_NEW_ENOUGH);
3067 }
3068
3069 /*
3070 * Now check for the BHI_CTRL MSR, and then set it if available.
3071 * We will still need to use the software sequence, however.
3072 */
3073 if (is_x86_feature(featureset, X86FSET_BHI_CTRL)) {
3074 cpuid_enable_bhi_dis_s();
3075 return (X86_BHI_DIS_S);
3076 }
3077
3078 /*
3079 * Finally, check if we are too old to bother with RSB:
3080 */
3081 if (v2mit == X86_SPECTREV2_RETPOLINE) {
3082 *bhb_clear = ret;
3083 return (X86_BHI_TOO_OLD_OR_DISABLED);
3084 }
3085
3086 ASSERT(*bhb_clear != ret);
3087 return (X86_BHI_SOFTWARE_SEQUENCE);
3088 }
3089
3090 /*
3091 * We default to enabling Return Stack Buffer (RSB) mitigations.
3092 *
3093 * We used to skip RSB mitigations with Intel eIBRS, but developments around
3094 * post-barrier RSB (PBRSB) guessing suggests we should enable Intel RSB
3095 * mitigations always unless explicitly bypassed, or unless hardware indicates
3096 * the bug has been fixed.
3097 *
3098 * The current decisions for using, or ignoring, a RSB software stuffing
3099 * sequence are expressed by the following table:
3100 *
3101 * +-------+------------+-----------------+--------+
3102 * | eIBRS | PBRSB_NO | context switch | vmexit |
3103 * +-------+------------+-----------------+--------+
3104 * | Yes | No | stuff | stuff |
3105 * | Yes | Yes | ignore | ignore |
3106 * | No | No | stuff | ignore |
3107 * +-------+------------+-----------------+--------+
3108 *
3109 * Note that if an Intel CPU has no eIBRS, it will never enumerate PBRSB_NO,
3110 * because machines with no eIBRS do not have a problem with PBRSB overflow.
3111 * See the Intel document cited below for details.
3112 *
3113 * Also note that AMD AUTO_IBRS has no PBRSB problem, so it is not included in
3114 * the table above, and that there is no situation where vmexit stuffing is
3115 * needed, but context-switch stuffing isn't.
3116 */
3117
3118 /* BEGIN CSTYLED */
3119 /*
3120 * https://www.intel.com/content/www/us/en/developer/articles/technical/software-security-guidance/advisory-guidance/post-barrier-return-stack-buffer-predictions.html
3121 */
3122 /* END CSTYLED */
3123
3124 /*
3125 * AMD indicates that when Automatic IBRS is enabled we do not need to implement
3126 * return stack buffer clearing for VMEXIT as it takes care of it. The manual
3127 * also states that as long as SMEP and we maintain at least one page between
3128 * the kernel and user space (we have much more of a red zone), then we do not
3129 * need to clear the RSB. We constrain this to only when Automatic IRBS is
3130 * present.
3131 */
3132 static void
cpuid_patch_rsb(x86_spectrev2_mitigation_t mit,bool intel_pbrsb_no)3133 cpuid_patch_rsb(x86_spectrev2_mitigation_t mit, bool intel_pbrsb_no)
3134 {
3135 const uint8_t ret = RET_INSTR;
3136 uint8_t *stuff = (uint8_t *)x86_rsb_stuff;
3137 uint8_t *vmx_stuff = (uint8_t *)x86_rsb_stuff_vmexit;
3138
3139 switch (mit) {
3140 case X86_SPECTREV2_AUTO_IBRS:
3141 case X86_SPECTREV2_DISABLED:
3142 /* Don't bother with any RSB stuffing! */
3143 *stuff = ret;
3144 *vmx_stuff = ret;
3145 break;
3146 case X86_SPECTREV2_RETPOLINE:
3147 /*
3148 * The Intel document on Post-Barrier RSB says that processors
3149 * without eIBRS do not have PBRSB problems upon VMEXIT.
3150 */
3151 VERIFY(!intel_pbrsb_no);
3152 VERIFY3U(*stuff, !=, ret);
3153 *vmx_stuff = ret;
3154 break;
3155 default:
3156 /*
3157 * eIBRS is all that's left. If CPU claims PBRSB is fixed,
3158 * don't use the RSB mitigation in either case. Otherwise
3159 * both vmexit and context-switching require the software
3160 * mitigation.
3161 */
3162 if (intel_pbrsb_no) {
3163 /* CPU claims PBRSB problems are fixed. */
3164 *stuff = ret;
3165 *vmx_stuff = ret;
3166 }
3167 VERIFY3U(*stuff, ==, *vmx_stuff);
3168 break;
3169 }
3170 }
3171
3172 static void
cpuid_patch_retpolines(x86_spectrev2_mitigation_t mit)3173 cpuid_patch_retpolines(x86_spectrev2_mitigation_t mit)
3174 {
3175 const char *thunks[] = { "_rax", "_rbx", "_rcx", "_rdx", "_rdi",
3176 "_rsi", "_rbp", "_r8", "_r9", "_r10", "_r11", "_r12", "_r13",
3177 "_r14", "_r15" };
3178 const uint_t nthunks = ARRAY_SIZE(thunks);
3179 const char *type;
3180 uint_t i;
3181
3182 if (mit == x86_spectrev2_mitigation)
3183 return;
3184
3185 switch (mit) {
3186 case X86_SPECTREV2_RETPOLINE:
3187 type = "gen";
3188 break;
3189 case X86_SPECTREV2_AUTO_IBRS:
3190 case X86_SPECTREV2_ENHANCED_IBRS:
3191 case X86_SPECTREV2_DISABLED:
3192 type = "jmp";
3193 break;
3194 default:
3195 panic("asked to update retpoline state with unknown state!");
3196 }
3197
3198 for (i = 0; i < nthunks; i++) {
3199 uintptr_t source, dest;
3200 int ssize, dsize;
3201 char sourcebuf[64], destbuf[64];
3202
3203 (void) snprintf(destbuf, sizeof (destbuf),
3204 "__x86_indirect_thunk%s", thunks[i]);
3205 (void) snprintf(sourcebuf, sizeof (sourcebuf),
3206 "__x86_indirect_thunk_%s%s", type, thunks[i]);
3207
3208 source = kobj_getelfsym(sourcebuf, NULL, &ssize);
3209 dest = kobj_getelfsym(destbuf, NULL, &dsize);
3210 VERIFY3U(source, !=, 0);
3211 VERIFY3U(dest, !=, 0);
3212 VERIFY3S(dsize, >=, ssize);
3213 bcopy((void *)source, (void *)dest, ssize);
3214 }
3215 }
3216
3217 static void
cpuid_enable_enhanced_ibrs(void)3218 cpuid_enable_enhanced_ibrs(void)
3219 {
3220 uint64_t val;
3221
3222 val = rdmsr(MSR_IA32_SPEC_CTRL);
3223 val |= IA32_SPEC_CTRL_IBRS;
3224 wrmsr(MSR_IA32_SPEC_CTRL, val);
3225 }
3226
3227 static void
cpuid_enable_auto_ibrs(void)3228 cpuid_enable_auto_ibrs(void)
3229 {
3230 uint64_t val;
3231
3232 val = rdmsr(MSR_AMD_EFER);
3233 val |= AMD_EFER_AIBRSE;
3234 wrmsr(MSR_AMD_EFER, val);
3235 }
3236
3237 /*
3238 * Determine how we should mitigate TAA or if we need to. Regardless of TAA, if
3239 * we can disable TSX, we do so.
3240 *
3241 * This determination is done only on the boot CPU, potentially after loading
3242 * updated microcode.
3243 */
3244 static void
cpuid_update_tsx(cpu_t * cpu,uchar_t * featureset)3245 cpuid_update_tsx(cpu_t *cpu, uchar_t *featureset)
3246 {
3247 struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
3248
3249 VERIFY(cpu->cpu_id == 0);
3250
3251 if (cpi->cpi_vendor != X86_VENDOR_Intel) {
3252 x86_taa_mitigation = X86_TAA_HW_MITIGATED;
3253 return;
3254 }
3255
3256 if (x86_disable_taa) {
3257 x86_taa_mitigation = X86_TAA_DISABLED;
3258 return;
3259 }
3260
3261 /*
3262 * If we do not have the ability to disable TSX, then our only
3263 * mitigation options are in hardware (TAA_NO), or by using our existing
3264 * MDS mitigation as described above. The latter relies upon us having
3265 * configured MDS mitigations correctly! This includes disabling SMT if
3266 * we want to cross-CPU-thread protection.
3267 */
3268 if (!is_x86_feature(featureset, X86FSET_TSX_CTRL)) {
3269 /*
3270 * It's not clear whether any parts will enumerate TAA_NO
3271 * *without* TSX_CTRL, but let's mark it as such if we see this.
3272 */
3273 if (is_x86_feature(featureset, X86FSET_TAA_NO)) {
3274 x86_taa_mitigation = X86_TAA_HW_MITIGATED;
3275 return;
3276 }
3277
3278 if (is_x86_feature(featureset, X86FSET_MD_CLEAR) &&
3279 !is_x86_feature(featureset, X86FSET_MDS_NO)) {
3280 x86_taa_mitigation = X86_TAA_MD_CLEAR;
3281 } else {
3282 x86_taa_mitigation = X86_TAA_NOTHING;
3283 }
3284 return;
3285 }
3286
3287 /*
3288 * We have TSX_CTRL, but we can only fully disable TSX if we're early
3289 * enough in boot.
3290 *
3291 * Otherwise, we'll fall back to causing transactions to abort as our
3292 * mitigation. TSX-using code will always take the fallback path.
3293 */
3294 if (cpi->cpi_pass < 4) {
3295 x86_taa_mitigation = X86_TAA_TSX_DISABLE;
3296 } else {
3297 x86_taa_mitigation = X86_TAA_TSX_FORCE_ABORT;
3298 }
3299 }
3300
3301 /*
3302 * As mentioned, we should only touch the MSR when we've got a suitable
3303 * microcode loaded on this CPU.
3304 */
3305 static void
cpuid_apply_tsx(x86_taa_mitigation_t taa,uchar_t * featureset)3306 cpuid_apply_tsx(x86_taa_mitigation_t taa, uchar_t *featureset)
3307 {
3308 uint64_t val;
3309
3310 switch (taa) {
3311 case X86_TAA_TSX_DISABLE:
3312 if (!is_x86_feature(featureset, X86FSET_TSX_CTRL))
3313 return;
3314 val = rdmsr(MSR_IA32_TSX_CTRL);
3315 val |= IA32_TSX_CTRL_CPUID_CLEAR | IA32_TSX_CTRL_RTM_DISABLE;
3316 wrmsr(MSR_IA32_TSX_CTRL, val);
3317 break;
3318 case X86_TAA_TSX_FORCE_ABORT:
3319 if (!is_x86_feature(featureset, X86FSET_TSX_CTRL))
3320 return;
3321 val = rdmsr(MSR_IA32_TSX_CTRL);
3322 val |= IA32_TSX_CTRL_RTM_DISABLE;
3323 wrmsr(MSR_IA32_TSX_CTRL, val);
3324 break;
3325 case X86_TAA_HW_MITIGATED:
3326 case X86_TAA_MD_CLEAR:
3327 case X86_TAA_DISABLED:
3328 case X86_TAA_NOTHING:
3329 break;
3330 }
3331 }
3332
3333 static void
cpuid_scan_security(cpu_t * cpu,uchar_t * featureset)3334 cpuid_scan_security(cpu_t *cpu, uchar_t *featureset)
3335 {
3336 struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
3337 x86_spectrev2_mitigation_t v2mit;
3338
3339 if ((cpi->cpi_vendor == X86_VENDOR_AMD ||
3340 cpi->cpi_vendor == X86_VENDOR_HYGON) &&
3341 cpi->cpi_xmaxeax >= CPUID_LEAF_EXT_8) {
3342 if (cpi->cpi_extd[8].cp_ebx & CPUID_AMD_EBX_IBPB)
3343 add_x86_feature(featureset, X86FSET_IBPB);
3344 if (cpi->cpi_extd[8].cp_ebx & CPUID_AMD_EBX_IBRS)
3345 add_x86_feature(featureset, X86FSET_IBRS);
3346 if (cpi->cpi_extd[8].cp_ebx & CPUID_AMD_EBX_STIBP)
3347 add_x86_feature(featureset, X86FSET_STIBP);
3348 if (cpi->cpi_extd[8].cp_ebx & CPUID_AMD_EBX_STIBP_ALL)
3349 add_x86_feature(featureset, X86FSET_STIBP_ALL);
3350 if (cpi->cpi_extd[8].cp_ebx & CPUID_AMD_EBX_SSBD)
3351 add_x86_feature(featureset, X86FSET_SSBD);
3352 if (cpi->cpi_extd[8].cp_ebx & CPUID_AMD_EBX_VIRT_SSBD)
3353 add_x86_feature(featureset, X86FSET_SSBD_VIRT);
3354 if (cpi->cpi_extd[8].cp_ebx & CPUID_AMD_EBX_SSB_NO)
3355 add_x86_feature(featureset, X86FSET_SSB_NO);
3356
3357 /*
3358 * Rather than Enhanced IBRS, AMD has a different feature that
3359 * is a bit in EFER that can be enabled and will basically do
3360 * the right thing while executing in the kernel.
3361 */
3362 if (cpi->cpi_vendor == X86_VENDOR_AMD &&
3363 (cpi->cpi_extd[8].cp_ebx & CPUID_AMD_EBX_PREFER_IBRS) &&
3364 cpi->cpi_xmaxeax >= CPUID_LEAF_EXT_21 &&
3365 (cpi->cpi_extd[0x21].cp_eax & CPUID_AMD_8X21_EAX_AIBRS)) {
3366 add_x86_feature(featureset, X86FSET_AUTO_IBRS);
3367 }
3368
3369 } else if (cpi->cpi_vendor == X86_VENDOR_Intel &&
3370 cpi->cpi_maxeax >= 7) {
3371 struct cpuid_regs *ecp;
3372 ecp = &cpi->cpi_std[7];
3373
3374 if (ecp->cp_edx & CPUID_INTC_EDX_7_0_MD_CLEAR) {
3375 add_x86_feature(featureset, X86FSET_MD_CLEAR);
3376 }
3377
3378 if (ecp->cp_edx & CPUID_INTC_EDX_7_0_SPEC_CTRL) {
3379 add_x86_feature(featureset, X86FSET_IBRS);
3380 add_x86_feature(featureset, X86FSET_IBPB);
3381 }
3382
3383 if (ecp->cp_edx & CPUID_INTC_EDX_7_0_STIBP) {
3384 add_x86_feature(featureset, X86FSET_STIBP);
3385 }
3386
3387 /*
3388 * Some prediction controls are enumerated by subleaf 2 of
3389 * leaf 7.
3390 */
3391 if (CPI_FEATURES_7_2_EDX(cpi) & CPUID_INTC_EDX_7_2_BHI_CTRL) {
3392 add_x86_feature(featureset, X86FSET_BHI_CTRL);
3393 }
3394
3395 /*
3396 * Don't read the arch caps MSR on xpv where we lack the
3397 * on_trap().
3398 */
3399 #ifndef __xpv
3400 if (ecp->cp_edx & CPUID_INTC_EDX_7_0_ARCH_CAPS) {
3401 on_trap_data_t otd;
3402
3403 /*
3404 * Be paranoid and assume we'll get a #GP.
3405 */
3406 if (!on_trap(&otd, OT_DATA_ACCESS)) {
3407 uint64_t reg;
3408
3409 reg = rdmsr(MSR_IA32_ARCH_CAPABILITIES);
3410 if (reg & IA32_ARCH_CAP_RDCL_NO) {
3411 add_x86_feature(featureset,
3412 X86FSET_RDCL_NO);
3413 }
3414 if (reg & IA32_ARCH_CAP_IBRS_ALL) {
3415 add_x86_feature(featureset,
3416 X86FSET_IBRS_ALL);
3417 }
3418 if (reg & IA32_ARCH_CAP_RSBA) {
3419 add_x86_feature(featureset,
3420 X86FSET_RSBA);
3421 }
3422 if (reg & IA32_ARCH_CAP_SKIP_L1DFL_VMENTRY) {
3423 add_x86_feature(featureset,
3424 X86FSET_L1D_VM_NO);
3425 }
3426 if (reg & IA32_ARCH_CAP_SSB_NO) {
3427 add_x86_feature(featureset,
3428 X86FSET_SSB_NO);
3429 }
3430 if (reg & IA32_ARCH_CAP_MDS_NO) {
3431 add_x86_feature(featureset,
3432 X86FSET_MDS_NO);
3433 }
3434 if (reg & IA32_ARCH_CAP_TSX_CTRL) {
3435 add_x86_feature(featureset,
3436 X86FSET_TSX_CTRL);
3437 }
3438 if (reg & IA32_ARCH_CAP_TAA_NO) {
3439 add_x86_feature(featureset,
3440 X86FSET_TAA_NO);
3441 }
3442 if (reg & IA32_ARCH_CAP_RFDS_NO) {
3443 add_x86_feature(featureset,
3444 X86FSET_RFDS_NO);
3445 }
3446 if (reg & IA32_ARCH_CAP_RFDS_CLEAR) {
3447 add_x86_feature(featureset,
3448 X86FSET_RFDS_CLEAR);
3449 }
3450 if (reg & IA32_ARCH_CAP_PBRSB_NO) {
3451 add_x86_feature(featureset,
3452 X86FSET_PBRSB_NO);
3453 }
3454 if (reg & IA32_ARCH_CAP_BHI_NO) {
3455 add_x86_feature(featureset,
3456 X86FSET_BHI_NO);
3457 }
3458 }
3459 no_trap();
3460 }
3461 #endif /* !__xpv */
3462
3463 if (ecp->cp_edx & CPUID_INTC_EDX_7_0_SSBD)
3464 add_x86_feature(featureset, X86FSET_SSBD);
3465
3466 if (ecp->cp_edx & CPUID_INTC_EDX_7_0_FLUSH_CMD)
3467 add_x86_feature(featureset, X86FSET_FLUSH_CMD);
3468 }
3469
3470 /*
3471 * Take care of certain mitigations on the non-boot CPU. The boot CPU
3472 * will have already run this function and determined what we need to
3473 * do. This gives us a hook for per-HW thread mitigations such as
3474 * enhanced IBRS, or disabling TSX.
3475 */
3476 if (cpu->cpu_id != 0) {
3477 switch (x86_spectrev2_mitigation) {
3478 case X86_SPECTREV2_ENHANCED_IBRS:
3479 cpuid_enable_enhanced_ibrs();
3480 break;
3481 case X86_SPECTREV2_AUTO_IBRS:
3482 cpuid_enable_auto_ibrs();
3483 break;
3484 default:
3485 break;
3486 }
3487
3488 /* If we're committed to BHI_DIS_S, set it for this core. */
3489 if (x86_bhi_mitigation == X86_BHI_DIS_S)
3490 cpuid_enable_bhi_dis_s();
3491
3492 cpuid_apply_tsx(x86_taa_mitigation, featureset);
3493 return;
3494 }
3495
3496 /*
3497 * Go through and initialize various security mechanisms that we should
3498 * only do on a single CPU. This includes Spectre V2, L1TF, MDS, and
3499 * TAA.
3500 */
3501
3502 /*
3503 * By default we've come in with retpolines enabled. Check whether we
3504 * should disable them or enable enhanced or automatic IBRS.
3505 *
3506 * Note, we do not allow the use of AMD optimized retpolines as it was
3507 * disclosed by AMD in March 2022 that they were still
3508 * vulnerable. Prior to that point, we used them.
3509 */
3510 if (x86_disable_spectrev2 != 0) {
3511 v2mit = X86_SPECTREV2_DISABLED;
3512 } else if (is_x86_feature(featureset, X86FSET_AUTO_IBRS)) {
3513 cpuid_enable_auto_ibrs();
3514 v2mit = X86_SPECTREV2_AUTO_IBRS;
3515 } else if (is_x86_feature(featureset, X86FSET_IBRS_ALL)) {
3516 cpuid_enable_enhanced_ibrs();
3517 v2mit = X86_SPECTREV2_ENHANCED_IBRS;
3518 } else {
3519 v2mit = X86_SPECTREV2_RETPOLINE;
3520 }
3521
3522 cpuid_patch_retpolines(v2mit);
3523 cpuid_patch_rsb(v2mit, is_x86_feature(featureset, X86FSET_PBRSB_NO));
3524 x86_bhi_mitigation = cpuid_learn_and_patch_bhi(v2mit, cpu, featureset);
3525 x86_spectrev2_mitigation = v2mit;
3526 membar_producer();
3527
3528 /*
3529 * We need to determine what changes are required for mitigating L1TF
3530 * and MDS. If the CPU suffers from either of them, then SMT exclusion
3531 * is required.
3532 *
3533 * If any of these are present, then we need to flush u-arch state at
3534 * various points. For MDS, we need to do so whenever we change to a
3535 * lesser privilege level or we are halting the CPU. For L1TF we need to
3536 * flush the L1D cache at VM entry. When we have microcode that handles
3537 * MDS, the L1D flush also clears the other u-arch state that the
3538 * md_clear does.
3539 */
3540
3541 /*
3542 * Update whether or not we need to be taking explicit action against
3543 * MDS or RFDS.
3544 */
3545 cpuid_update_md_clear(cpu, featureset);
3546
3547 /*
3548 * Determine whether SMT exclusion is required and whether or not we
3549 * need to perform an l1d flush.
3550 */
3551 cpuid_update_l1d_flush(cpu, featureset);
3552
3553 /*
3554 * Determine what our mitigation strategy should be for TAA and then
3555 * also apply TAA mitigations.
3556 */
3557 cpuid_update_tsx(cpu, featureset);
3558 cpuid_apply_tsx(x86_taa_mitigation, featureset);
3559 }
3560
3561 /*
3562 * Setup XFeature_Enabled_Mask register. Required by xsave feature.
3563 */
3564 void
setup_xfem(void)3565 setup_xfem(void)
3566 {
3567 uint64_t flags = XFEATURE_LEGACY_FP;
3568
3569 ASSERT(is_x86_feature(x86_featureset, X86FSET_XSAVE));
3570
3571 if (is_x86_feature(x86_featureset, X86FSET_SSE))
3572 flags |= XFEATURE_SSE;
3573
3574 if (is_x86_feature(x86_featureset, X86FSET_AVX))
3575 flags |= XFEATURE_AVX;
3576
3577 if (is_x86_feature(x86_featureset, X86FSET_AVX512F))
3578 flags |= XFEATURE_AVX512;
3579
3580 set_xcr(XFEATURE_ENABLED_MASK, flags);
3581
3582 xsave_bv_all = flags;
3583 }
3584
3585 static void
cpuid_basic_topology(cpu_t * cpu,uchar_t * featureset)3586 cpuid_basic_topology(cpu_t *cpu, uchar_t *featureset)
3587 {
3588 struct cpuid_info *cpi;
3589
3590 cpi = cpu->cpu_m.mcpu_cpi;
3591
3592 if (cpi->cpi_vendor == X86_VENDOR_AMD ||
3593 cpi->cpi_vendor == X86_VENDOR_HYGON) {
3594 cpuid_gather_amd_topology_leaves(cpu);
3595 }
3596
3597 cpi->cpi_apicid = cpuid_gather_apicid(cpi);
3598
3599 /*
3600 * Before we can calculate the IDs that we should assign to this
3601 * processor, we need to understand how many cores and threads it has.
3602 */
3603 switch (cpi->cpi_vendor) {
3604 case X86_VENDOR_Intel:
3605 cpuid_intel_ncores(cpi, &cpi->cpi_ncpu_per_chip,
3606 &cpi->cpi_ncore_per_chip);
3607 break;
3608 case X86_VENDOR_AMD:
3609 case X86_VENDOR_HYGON:
3610 cpuid_amd_ncores(cpi, &cpi->cpi_ncpu_per_chip,
3611 &cpi->cpi_ncore_per_chip);
3612 break;
3613 default:
3614 /*
3615 * If we have some other x86 compatible chip, it's not clear how
3616 * they would behave. The most common case is virtualization
3617 * today, though there are also 64-bit VIA chips. Assume that
3618 * all we can get is the basic Leaf 1 HTT information.
3619 */
3620 if ((cpi->cpi_std[1].cp_edx & CPUID_INTC_EDX_HTT) != 0) {
3621 cpi->cpi_ncore_per_chip = 1;
3622 cpi->cpi_ncpu_per_chip = CPI_CPU_COUNT(cpi);
3623 }
3624 break;
3625 }
3626
3627 /*
3628 * Based on the calculated number of threads and cores, potentially
3629 * assign the HTT and CMT features.
3630 */
3631 if (cpi->cpi_ncore_per_chip > 1) {
3632 add_x86_feature(featureset, X86FSET_CMP);
3633 }
3634
3635 if (cpi->cpi_ncpu_per_chip > 1 &&
3636 cpi->cpi_ncpu_per_chip != cpi->cpi_ncore_per_chip) {
3637 add_x86_feature(featureset, X86FSET_HTT);
3638 }
3639
3640 /*
3641 * Now that has been set up, we need to go through and calculate all of
3642 * the rest of the parameters that exist. If we think the CPU doesn't
3643 * have either SMT (HTT) or CMP, then we basically go through and fake
3644 * up information in some way. The most likely case for this is
3645 * virtualization where we have a lot of partial topology information.
3646 */
3647 if (!is_x86_feature(featureset, X86FSET_HTT) &&
3648 !is_x86_feature(featureset, X86FSET_CMP)) {
3649 /*
3650 * This is a single core, single-threaded processor.
3651 */
3652 cpi->cpi_procnodes_per_pkg = 1;
3653 cpi->cpi_cores_per_compunit = 1;
3654 cpi->cpi_compunitid = 0;
3655 cpi->cpi_chipid = -1;
3656 cpi->cpi_clogid = 0;
3657 cpi->cpi_coreid = cpu->cpu_id;
3658 cpi->cpi_pkgcoreid = 0;
3659 if (cpi->cpi_vendor == X86_VENDOR_AMD ||
3660 cpi->cpi_vendor == X86_VENDOR_HYGON) {
3661 cpi->cpi_procnodeid = BITX(cpi->cpi_apicid, 3, 0);
3662 } else {
3663 cpi->cpi_procnodeid = cpi->cpi_chipid;
3664 }
3665 } else {
3666 switch (cpi->cpi_vendor) {
3667 case X86_VENDOR_Intel:
3668 cpuid_intel_getids(cpu, featureset);
3669 break;
3670 case X86_VENDOR_AMD:
3671 case X86_VENDOR_HYGON:
3672 cpuid_amd_getids(cpu, featureset);
3673 break;
3674 default:
3675 /*
3676 * In this case, it's hard to say what we should do.
3677 * We're going to model them to the OS as single core
3678 * threads. We don't have a good identifier for them, so
3679 * we're just going to use the cpu id all on a single
3680 * chip.
3681 *
3682 * This case has historically been different from the
3683 * case above where we don't have HTT or CMP. While they
3684 * could be combined, we've opted to keep it separate to
3685 * minimize the risk of topology changes in weird cases.
3686 */
3687 cpi->cpi_procnodes_per_pkg = 1;
3688 cpi->cpi_cores_per_compunit = 1;
3689 cpi->cpi_chipid = 0;
3690 cpi->cpi_coreid = cpu->cpu_id;
3691 cpi->cpi_clogid = cpu->cpu_id;
3692 cpi->cpi_pkgcoreid = cpu->cpu_id;
3693 cpi->cpi_procnodeid = cpi->cpi_chipid;
3694 cpi->cpi_compunitid = cpi->cpi_coreid;
3695 break;
3696 }
3697 }
3698 }
3699
3700 /*
3701 * Gather relevant CPU features from leaf 6 which covers thermal information. We
3702 * always gather leaf 6 if it's supported; however, we only look for features on
3703 * Intel systems as AMD does not currently define any of the features we look
3704 * for below.
3705 */
3706 static void
cpuid_basic_thermal(cpu_t * cpu,uchar_t * featureset)3707 cpuid_basic_thermal(cpu_t *cpu, uchar_t *featureset)
3708 {
3709 struct cpuid_regs *cp;
3710 struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
3711
3712 if (cpi->cpi_maxeax < 6) {
3713 return;
3714 }
3715
3716 cp = &cpi->cpi_std[6];
3717 cp->cp_eax = 6;
3718 cp->cp_ebx = cp->cp_ecx = cp->cp_edx = 0;
3719 (void) __cpuid_insn(cp);
3720 platform_cpuid_mangle(cpi->cpi_vendor, 6, cp);
3721
3722 if (cpi->cpi_vendor != X86_VENDOR_Intel) {
3723 return;
3724 }
3725
3726 if ((cp->cp_eax & CPUID_INTC_EAX_DTS) != 0) {
3727 add_x86_feature(featureset, X86FSET_CORE_THERMAL);
3728 }
3729
3730 if ((cp->cp_eax & CPUID_INTC_EAX_PTM) != 0) {
3731 add_x86_feature(featureset, X86FSET_PKG_THERMAL);
3732 }
3733 }
3734
3735 /*
3736 * This is used when we discover that we have AVX support in cpuid. This
3737 * proceeds to scan for the rest of the AVX derived features.
3738 */
3739 static void
cpuid_basic_avx(cpu_t * cpu,uchar_t * featureset)3740 cpuid_basic_avx(cpu_t *cpu, uchar_t *featureset)
3741 {
3742 struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
3743
3744 /*
3745 * If we don't have AVX, don't bother with most of this.
3746 */
3747 if ((cpi->cpi_std[1].cp_ecx & CPUID_INTC_ECX_AVX) == 0)
3748 return;
3749
3750 add_x86_feature(featureset, X86FSET_AVX);
3751
3752 /*
3753 * Intel says we can't check these without also
3754 * checking AVX.
3755 */
3756 if (cpi->cpi_std[1].cp_ecx & CPUID_INTC_ECX_F16C)
3757 add_x86_feature(featureset, X86FSET_F16C);
3758
3759 if (cpi->cpi_std[1].cp_ecx & CPUID_INTC_ECX_FMA)
3760 add_x86_feature(featureset, X86FSET_FMA);
3761
3762 if (cpi->cpi_std[7].cp_ebx & CPUID_INTC_EBX_7_0_BMI1)
3763 add_x86_feature(featureset, X86FSET_BMI1);
3764
3765 if (cpi->cpi_std[7].cp_ebx & CPUID_INTC_EBX_7_0_BMI2)
3766 add_x86_feature(featureset, X86FSET_BMI2);
3767
3768 if (cpi->cpi_std[7].cp_ebx & CPUID_INTC_EBX_7_0_AVX2)
3769 add_x86_feature(featureset, X86FSET_AVX2);
3770
3771 if (cpi->cpi_std[7].cp_ecx & CPUID_INTC_ECX_7_0_VAES)
3772 add_x86_feature(featureset, X86FSET_VAES);
3773
3774 if (cpi->cpi_std[7].cp_ecx & CPUID_INTC_ECX_7_0_VPCLMULQDQ)
3775 add_x86_feature(featureset, X86FSET_VPCLMULQDQ);
3776
3777 /*
3778 * The rest of the AVX features require AVX512. Do not check them unless
3779 * it is present.
3780 */
3781 if ((cpi->cpi_std[7].cp_ebx & CPUID_INTC_EBX_7_0_AVX512F) == 0)
3782 return;
3783 add_x86_feature(featureset, X86FSET_AVX512F);
3784
3785 if (cpi->cpi_std[7].cp_ebx & CPUID_INTC_EBX_7_0_AVX512DQ)
3786 add_x86_feature(featureset, X86FSET_AVX512DQ);
3787
3788 if (cpi->cpi_std[7].cp_ebx & CPUID_INTC_EBX_7_0_AVX512IFMA)
3789 add_x86_feature(featureset, X86FSET_AVX512FMA);
3790
3791 if (cpi->cpi_std[7].cp_ebx & CPUID_INTC_EBX_7_0_AVX512PF)
3792 add_x86_feature(featureset, X86FSET_AVX512PF);
3793
3794 if (cpi->cpi_std[7].cp_ebx & CPUID_INTC_EBX_7_0_AVX512ER)
3795 add_x86_feature(featureset, X86FSET_AVX512ER);
3796
3797 if (cpi->cpi_std[7].cp_ebx & CPUID_INTC_EBX_7_0_AVX512CD)
3798 add_x86_feature(featureset, X86FSET_AVX512CD);
3799
3800 if (cpi->cpi_std[7].cp_ebx & CPUID_INTC_EBX_7_0_AVX512BW)
3801 add_x86_feature(featureset, X86FSET_AVX512BW);
3802
3803 if (cpi->cpi_std[7].cp_ebx & CPUID_INTC_EBX_7_0_AVX512VL)
3804 add_x86_feature(featureset, X86FSET_AVX512VL);
3805
3806 if (cpi->cpi_std[7].cp_ecx & CPUID_INTC_ECX_7_0_AVX512VBMI)
3807 add_x86_feature(featureset, X86FSET_AVX512VBMI);
3808
3809 if (cpi->cpi_std[7].cp_ecx & CPUID_INTC_ECX_7_0_AVX512VBMI2)
3810 add_x86_feature(featureset, X86FSET_AVX512_VBMI2);
3811
3812 if (cpi->cpi_std[7].cp_ecx & CPUID_INTC_ECX_7_0_AVX512VNNI)
3813 add_x86_feature(featureset, X86FSET_AVX512VNNI);
3814
3815 if (cpi->cpi_std[7].cp_ecx & CPUID_INTC_ECX_7_0_AVX512BITALG)
3816 add_x86_feature(featureset, X86FSET_AVX512_BITALG);
3817
3818 if (cpi->cpi_std[7].cp_ecx & CPUID_INTC_ECX_7_0_AVX512VPOPCDQ)
3819 add_x86_feature(featureset, X86FSET_AVX512VPOPCDQ);
3820
3821 if (cpi->cpi_std[7].cp_edx & CPUID_INTC_EDX_7_0_AVX5124NNIW)
3822 add_x86_feature(featureset, X86FSET_AVX512NNIW);
3823
3824 if (cpi->cpi_std[7].cp_edx & CPUID_INTC_EDX_7_0_AVX5124FMAPS)
3825 add_x86_feature(featureset, X86FSET_AVX512FMAPS);
3826
3827 /*
3828 * More features here are in Leaf 7, subleaf 1. Don't bother checking if
3829 * we don't need to.
3830 */
3831 if (cpi->cpi_std[7].cp_eax < 1)
3832 return;
3833
3834 if (cpi->cpi_sub7[0].cp_eax & CPUID_INTC_EAX_7_1_AVX512_BF16)
3835 add_x86_feature(featureset, X86FSET_AVX512_BF16);
3836 }
3837
3838 /*
3839 * PPIN is the protected processor inventory number. On AMD this is an actual
3840 * feature bit. However, on Intel systems we need to read the platform
3841 * information MSR if we're on a specific model.
3842 */
3843 #if !defined(__xpv)
3844 static void
cpuid_basic_ppin(cpu_t * cpu,uchar_t * featureset)3845 cpuid_basic_ppin(cpu_t *cpu, uchar_t *featureset)
3846 {
3847 on_trap_data_t otd;
3848 struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
3849
3850 switch (cpi->cpi_vendor) {
3851 case X86_VENDOR_AMD:
3852 /*
3853 * This leaf will have already been gathered in the topology
3854 * functions.
3855 */
3856 if (cpi->cpi_xmaxeax >= CPUID_LEAF_EXT_8) {
3857 if (cpi->cpi_extd[8].cp_ebx & CPUID_AMD_EBX_PPIN) {
3858 add_x86_feature(featureset, X86FSET_PPIN);
3859 }
3860 }
3861 break;
3862 case X86_VENDOR_Intel:
3863 if (cpi->cpi_family != 6)
3864 break;
3865 switch (cpi->cpi_model) {
3866 case INTC_MODEL_IVYBRIDGE_XEON:
3867 case INTC_MODEL_HASWELL_XEON:
3868 case INTC_MODEL_BROADWELL_XEON:
3869 case INTC_MODEL_BROADWELL_XEON_D:
3870 case INTC_MODEL_SKYLAKE_XEON:
3871 case INTC_MODEL_ICELAKE_XEON:
3872 if (!on_trap(&otd, OT_DATA_ACCESS)) {
3873 uint64_t value;
3874
3875 value = rdmsr(MSR_PLATFORM_INFO);
3876 if ((value & MSR_PLATFORM_INFO_PPIN) != 0) {
3877 add_x86_feature(featureset,
3878 X86FSET_PPIN);
3879 }
3880 }
3881 no_trap();
3882 break;
3883 default:
3884 break;
3885 }
3886 break;
3887 default:
3888 break;
3889 }
3890 }
3891 #endif /* ! __xpv */
3892
3893 static void
cpuid_pass_prelude(cpu_t * cpu,void * arg)3894 cpuid_pass_prelude(cpu_t *cpu, void *arg)
3895 {
3896 uchar_t *featureset = (uchar_t *)arg;
3897
3898 /*
3899 * We don't run on any processor that doesn't have cpuid, and could not
3900 * possibly have arrived here.
3901 */
3902 add_x86_feature(featureset, X86FSET_CPUID);
3903 }
3904
3905 static void
cpuid_pass_ident(cpu_t * cpu,void * arg __unused)3906 cpuid_pass_ident(cpu_t *cpu, void *arg __unused)
3907 {
3908 struct cpuid_info *cpi;
3909 struct cpuid_regs *cp;
3910
3911 /*
3912 * We require that virtual/native detection be complete and that PCI
3913 * config space access has been set up; at present there is no reliable
3914 * way to determine the latter.
3915 */
3916 #if !defined(__xpv)
3917 ASSERT3S(platform_type, !=, -1);
3918 #endif /* !__xpv */
3919
3920 cpi = cpu->cpu_m.mcpu_cpi;
3921 ASSERT(cpi != NULL);
3922
3923 cp = &cpi->cpi_std[0];
3924 cp->cp_eax = 0;
3925 cpi->cpi_maxeax = __cpuid_insn(cp);
3926 {
3927 uint32_t *iptr = (uint32_t *)cpi->cpi_vendorstr;
3928 *iptr++ = cp->cp_ebx;
3929 *iptr++ = cp->cp_edx;
3930 *iptr++ = cp->cp_ecx;
3931 *(char *)&cpi->cpi_vendorstr[12] = '\0';
3932 }
3933
3934 cpi->cpi_vendor = _cpuid_vendorstr_to_vendorcode(cpi->cpi_vendorstr);
3935 x86_vendor = cpi->cpi_vendor; /* for compatibility */
3936
3937 /*
3938 * Limit the range in case of weird hardware
3939 */
3940 if (cpi->cpi_maxeax > CPI_MAXEAX_MAX)
3941 cpi->cpi_maxeax = CPI_MAXEAX_MAX;
3942 if (cpi->cpi_maxeax < 1)
3943 return;
3944
3945 cp = &cpi->cpi_std[1];
3946 cp->cp_eax = 1;
3947 (void) __cpuid_insn(cp);
3948
3949 /*
3950 * Extract identifying constants for easy access.
3951 */
3952 cpi->cpi_model = CPI_MODEL(cpi);
3953 cpi->cpi_family = CPI_FAMILY(cpi);
3954
3955 if (cpi->cpi_family == 0xf)
3956 cpi->cpi_family += CPI_FAMILY_XTD(cpi);
3957
3958 /*
3959 * Beware: AMD uses "extended model" iff base *FAMILY* == 0xf.
3960 * Intel, and presumably everyone else, uses model == 0xf, as
3961 * one would expect (max value means possible overflow). Sigh.
3962 */
3963
3964 switch (cpi->cpi_vendor) {
3965 case X86_VENDOR_Intel:
3966 if (IS_EXTENDED_MODEL_INTEL(cpi))
3967 cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4;
3968 break;
3969 case X86_VENDOR_AMD:
3970 if (CPI_FAMILY(cpi) == 0xf)
3971 cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4;
3972 break;
3973 case X86_VENDOR_HYGON:
3974 cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4;
3975 break;
3976 default:
3977 if (cpi->cpi_model == 0xf)
3978 cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4;
3979 break;
3980 }
3981
3982 cpi->cpi_step = CPI_STEP(cpi);
3983 cpi->cpi_brandid = CPI_BRANDID(cpi);
3984
3985 /*
3986 * Synthesize chip "revision" and socket type
3987 */
3988 cpi->cpi_chiprev = _cpuid_chiprev(cpi->cpi_vendor, cpi->cpi_family,
3989 cpi->cpi_model, cpi->cpi_step);
3990 cpi->cpi_chiprevstr = _cpuid_chiprevstr(cpi->cpi_vendor,
3991 cpi->cpi_family, cpi->cpi_model, cpi->cpi_step);
3992 cpi->cpi_socket = _cpuid_skt(cpi->cpi_vendor, cpi->cpi_family,
3993 cpi->cpi_model, cpi->cpi_step);
3994 cpi->cpi_uarchrev = _cpuid_uarchrev(cpi->cpi_vendor, cpi->cpi_family,
3995 cpi->cpi_model, cpi->cpi_step);
3996 }
3997
3998 static void
cpuid_pass_basic(cpu_t * cpu,void * arg)3999 cpuid_pass_basic(cpu_t *cpu, void *arg)
4000 {
4001 uchar_t *featureset = (uchar_t *)arg;
4002 uint32_t mask_ecx, mask_edx;
4003 struct cpuid_info *cpi;
4004 struct cpuid_regs *cp;
4005 int xcpuid;
4006 #if !defined(__xpv)
4007 extern int idle_cpu_prefer_mwait;
4008 #endif
4009
4010 cpi = cpu->cpu_m.mcpu_cpi;
4011 ASSERT(cpi != NULL);
4012
4013 if (cpi->cpi_maxeax < 1)
4014 return;
4015
4016 /*
4017 * This was filled during the identification pass.
4018 */
4019 cp = &cpi->cpi_std[1];
4020
4021 /*
4022 * *default* assumptions:
4023 * - believe %edx feature word
4024 * - ignore %ecx feature word
4025 * - 32-bit virtual and physical addressing
4026 */
4027 mask_edx = 0xffffffff;
4028 mask_ecx = 0;
4029
4030 cpi->cpi_pabits = cpi->cpi_vabits = 32;
4031
4032 switch (cpi->cpi_vendor) {
4033 case X86_VENDOR_Intel:
4034 if (cpi->cpi_family == 5)
4035 x86_type = X86_TYPE_P5;
4036 else if (IS_LEGACY_P6(cpi)) {
4037 x86_type = X86_TYPE_P6;
4038 pentiumpro_bug4046376 = 1;
4039 /*
4040 * Clear the SEP bit when it was set erroneously
4041 */
4042 if (cpi->cpi_model < 3 && cpi->cpi_step < 3)
4043 cp->cp_edx &= ~CPUID_INTC_EDX_SEP;
4044 } else if (IS_NEW_F6(cpi) || cpi->cpi_family == 0xf) {
4045 x86_type = X86_TYPE_P4;
4046 /*
4047 * We don't currently depend on any of the %ecx
4048 * features until Prescott, so we'll only check
4049 * this from P4 onwards. We might want to revisit
4050 * that idea later.
4051 */
4052 mask_ecx = 0xffffffff;
4053 } else if (cpi->cpi_family > 0xf)
4054 mask_ecx = 0xffffffff;
4055 /*
4056 * We don't support MONITOR/MWAIT if leaf 5 is not available
4057 * to obtain the monitor linesize.
4058 */
4059 if (cpi->cpi_maxeax < 5)
4060 mask_ecx &= ~CPUID_INTC_ECX_MON;
4061 break;
4062 case X86_VENDOR_IntelClone:
4063 default:
4064 break;
4065 case X86_VENDOR_AMD:
4066 #if defined(OPTERON_ERRATUM_108)
4067 if (cpi->cpi_family == 0xf && cpi->cpi_model == 0xe) {
4068 cp->cp_eax = (0xf0f & cp->cp_eax) | 0xc0;
4069 cpi->cpi_model = 0xc;
4070 } else
4071 #endif
4072 if (cpi->cpi_family == 5) {
4073 /*
4074 * AMD K5 and K6
4075 *
4076 * These CPUs have an incomplete implementation
4077 * of MCA/MCE which we mask away.
4078 */
4079 mask_edx &= ~(CPUID_INTC_EDX_MCE | CPUID_INTC_EDX_MCA);
4080
4081 /*
4082 * Model 0 uses the wrong (APIC) bit
4083 * to indicate PGE. Fix it here.
4084 */
4085 if (cpi->cpi_model == 0) {
4086 if (cp->cp_edx & 0x200) {
4087 cp->cp_edx &= ~0x200;
4088 cp->cp_edx |= CPUID_INTC_EDX_PGE;
4089 }
4090 }
4091
4092 /*
4093 * Early models had problems w/ MMX; disable.
4094 */
4095 if (cpi->cpi_model < 6)
4096 mask_edx &= ~CPUID_INTC_EDX_MMX;
4097 }
4098
4099 /*
4100 * For newer families, SSE3 and CX16, at least, are valid;
4101 * enable all
4102 */
4103 if (cpi->cpi_family >= 0xf)
4104 mask_ecx = 0xffffffff;
4105 /*
4106 * We don't support MONITOR/MWAIT if leaf 5 is not available
4107 * to obtain the monitor linesize.
4108 */
4109 if (cpi->cpi_maxeax < 5)
4110 mask_ecx &= ~CPUID_INTC_ECX_MON;
4111
4112 #if !defined(__xpv)
4113 /*
4114 * AMD has not historically used MWAIT in the CPU's idle loop.
4115 * Pre-family-10h Opterons do not have the MWAIT instruction. We
4116 * know for certain that in at least family 17h, per AMD, mwait
4117 * is preferred. Families in-between are less certain.
4118 */
4119 if (cpi->cpi_family < 0x17) {
4120 idle_cpu_prefer_mwait = 0;
4121 }
4122 #endif
4123
4124 break;
4125 case X86_VENDOR_HYGON:
4126 /* Enable all for Hygon Dhyana CPU */
4127 mask_ecx = 0xffffffff;
4128 break;
4129 case X86_VENDOR_TM:
4130 /*
4131 * workaround the NT workaround in CMS 4.1
4132 */
4133 if (cpi->cpi_family == 5 && cpi->cpi_model == 4 &&
4134 (cpi->cpi_step == 2 || cpi->cpi_step == 3))
4135 cp->cp_edx |= CPUID_INTC_EDX_CX8;
4136 break;
4137 case X86_VENDOR_Centaur:
4138 /*
4139 * workaround the NT workarounds again
4140 */
4141 if (cpi->cpi_family == 6)
4142 cp->cp_edx |= CPUID_INTC_EDX_CX8;
4143 break;
4144 case X86_VENDOR_Cyrix:
4145 /*
4146 * We rely heavily on the probing in locore
4147 * to actually figure out what parts, if any,
4148 * of the Cyrix cpuid instruction to believe.
4149 */
4150 switch (x86_type) {
4151 case X86_TYPE_CYRIX_486:
4152 mask_edx = 0;
4153 break;
4154 case X86_TYPE_CYRIX_6x86:
4155 mask_edx = 0;
4156 break;
4157 case X86_TYPE_CYRIX_6x86L:
4158 mask_edx =
4159 CPUID_INTC_EDX_DE |
4160 CPUID_INTC_EDX_CX8;
4161 break;
4162 case X86_TYPE_CYRIX_6x86MX:
4163 mask_edx =
4164 CPUID_INTC_EDX_DE |
4165 CPUID_INTC_EDX_MSR |
4166 CPUID_INTC_EDX_CX8 |
4167 CPUID_INTC_EDX_PGE |
4168 CPUID_INTC_EDX_CMOV |
4169 CPUID_INTC_EDX_MMX;
4170 break;
4171 case X86_TYPE_CYRIX_GXm:
4172 mask_edx =
4173 CPUID_INTC_EDX_MSR |
4174 CPUID_INTC_EDX_CX8 |
4175 CPUID_INTC_EDX_CMOV |
4176 CPUID_INTC_EDX_MMX;
4177 break;
4178 case X86_TYPE_CYRIX_MediaGX:
4179 break;
4180 case X86_TYPE_CYRIX_MII:
4181 case X86_TYPE_VIA_CYRIX_III:
4182 mask_edx =
4183 CPUID_INTC_EDX_DE |
4184 CPUID_INTC_EDX_TSC |
4185 CPUID_INTC_EDX_MSR |
4186 CPUID_INTC_EDX_CX8 |
4187 CPUID_INTC_EDX_PGE |
4188 CPUID_INTC_EDX_CMOV |
4189 CPUID_INTC_EDX_MMX;
4190 break;
4191 default:
4192 break;
4193 }
4194 break;
4195 }
4196
4197 #if defined(__xpv)
4198 /*
4199 * Do not support MONITOR/MWAIT under a hypervisor
4200 */
4201 mask_ecx &= ~CPUID_INTC_ECX_MON;
4202 /*
4203 * Do not support XSAVE under a hypervisor for now
4204 */
4205 xsave_force_disable = B_TRUE;
4206
4207 #endif /* __xpv */
4208
4209 if (xsave_force_disable) {
4210 mask_ecx &= ~CPUID_INTC_ECX_XSAVE;
4211 mask_ecx &= ~CPUID_INTC_ECX_AVX;
4212 mask_ecx &= ~CPUID_INTC_ECX_F16C;
4213 mask_ecx &= ~CPUID_INTC_ECX_FMA;
4214 }
4215
4216 /*
4217 * Now we've figured out the masks that determine
4218 * which bits we choose to believe, apply the masks
4219 * to the feature words, then map the kernel's view
4220 * of these feature words into its feature word.
4221 */
4222 cp->cp_edx &= mask_edx;
4223 cp->cp_ecx &= mask_ecx;
4224
4225 /*
4226 * apply any platform restrictions (we don't call this
4227 * immediately after __cpuid_insn here, because we need the
4228 * workarounds applied above first)
4229 */
4230 platform_cpuid_mangle(cpi->cpi_vendor, 1, cp);
4231
4232 /*
4233 * In addition to ecx and edx, Intel and AMD are storing a bunch of
4234 * instruction set extensions in leaf 7's ebx, ecx, and edx. Note, leaf
4235 * 7 has sub-leaves determined by ecx.
4236 */
4237 if (cpi->cpi_maxeax >= 7) {
4238 struct cpuid_regs *ecp;
4239 ecp = &cpi->cpi_std[7];
4240 ecp->cp_eax = 7;
4241 ecp->cp_ecx = 0;
4242 (void) __cpuid_insn(ecp);
4243
4244 /*
4245 * If XSAVE has been disabled, just ignore all of the
4246 * extended-save-area dependent flags here. By removing most of
4247 * the leaf 7, sub-leaf 0 flags, that will ensure tha we don't
4248 * end up looking at additional xsave dependent leaves right
4249 * now.
4250 */
4251 if (xsave_force_disable) {
4252 ecp->cp_ebx &= ~CPUID_INTC_EBX_7_0_BMI1;
4253 ecp->cp_ebx &= ~CPUID_INTC_EBX_7_0_BMI2;
4254 ecp->cp_ebx &= ~CPUID_INTC_EBX_7_0_AVX2;
4255 ecp->cp_ebx &= ~CPUID_INTC_EBX_7_0_MPX;
4256 ecp->cp_ebx &= ~CPUID_INTC_EBX_7_0_ALL_AVX512;
4257 ecp->cp_ecx &= ~CPUID_INTC_ECX_7_0_ALL_AVX512;
4258 ecp->cp_edx &= ~CPUID_INTC_EDX_7_0_ALL_AVX512;
4259 ecp->cp_ecx &= ~CPUID_INTC_ECX_7_0_VAES;
4260 ecp->cp_ecx &= ~CPUID_INTC_ECX_7_0_VPCLMULQDQ;
4261 ecp->cp_ecx &= ~CPUID_INTC_ECX_7_0_GFNI;
4262 }
4263
4264 if (ecp->cp_ebx & CPUID_INTC_EBX_7_0_SMEP)
4265 add_x86_feature(featureset, X86FSET_SMEP);
4266
4267 /*
4268 * We check disable_smap here in addition to in startup_smap()
4269 * to ensure CPUs that aren't the boot CPU don't accidentally
4270 * include it in the feature set and thus generate a mismatched
4271 * x86 feature set across CPUs.
4272 */
4273 if (ecp->cp_ebx & CPUID_INTC_EBX_7_0_SMAP &&
4274 disable_smap == 0)
4275 add_x86_feature(featureset, X86FSET_SMAP);
4276
4277 if (ecp->cp_ebx & CPUID_INTC_EBX_7_0_RDSEED)
4278 add_x86_feature(featureset, X86FSET_RDSEED);
4279
4280 if (ecp->cp_ebx & CPUID_INTC_EBX_7_0_ADX)
4281 add_x86_feature(featureset, X86FSET_ADX);
4282
4283 if (ecp->cp_ebx & CPUID_INTC_EBX_7_0_FSGSBASE)
4284 add_x86_feature(featureset, X86FSET_FSGSBASE);
4285
4286 if (ecp->cp_ebx & CPUID_INTC_EBX_7_0_CLFLUSHOPT)
4287 add_x86_feature(featureset, X86FSET_CLFLUSHOPT);
4288
4289 if (ecp->cp_ebx & CPUID_INTC_EBX_7_0_INVPCID)
4290 add_x86_feature(featureset, X86FSET_INVPCID);
4291
4292 if (ecp->cp_ecx & CPUID_INTC_ECX_7_0_UMIP)
4293 add_x86_feature(featureset, X86FSET_UMIP);
4294 if (ecp->cp_ecx & CPUID_INTC_ECX_7_0_PKU)
4295 add_x86_feature(featureset, X86FSET_PKU);
4296 if (ecp->cp_ecx & CPUID_INTC_ECX_7_0_OSPKE)
4297 add_x86_feature(featureset, X86FSET_OSPKE);
4298 if (ecp->cp_ecx & CPUID_INTC_ECX_7_0_GFNI)
4299 add_x86_feature(featureset, X86FSET_GFNI);
4300
4301 if (ecp->cp_ebx & CPUID_INTC_EBX_7_0_CLWB)
4302 add_x86_feature(featureset, X86FSET_CLWB);
4303
4304 if (cpi->cpi_vendor == X86_VENDOR_Intel) {
4305 if (ecp->cp_ebx & CPUID_INTC_EBX_7_0_MPX)
4306 add_x86_feature(featureset, X86FSET_MPX);
4307 }
4308
4309 /*
4310 * If we have subleaf 1 or 2 available, grab and store
4311 * that. This is used for more AVX and related features.
4312 */
4313 if (ecp->cp_eax >= 1) {
4314 struct cpuid_regs *c71;
4315 c71 = &cpi->cpi_sub7[0];
4316 c71->cp_eax = 7;
4317 c71->cp_ecx = 1;
4318 (void) __cpuid_insn(c71);
4319 }
4320
4321 /* Subleaf 2 has certain security indicators in it. */
4322 if (ecp->cp_eax >= 2) {
4323 struct cpuid_regs *c72;
4324 c72 = &cpi->cpi_sub7[1];
4325 c72->cp_eax = 7;
4326 c72->cp_ecx = 2;
4327 (void) __cpuid_insn(c72);
4328 }
4329 }
4330
4331 /*
4332 * fold in overrides from the "eeprom" mechanism
4333 */
4334 cp->cp_edx |= cpuid_feature_edx_include;
4335 cp->cp_edx &= ~cpuid_feature_edx_exclude;
4336
4337 cp->cp_ecx |= cpuid_feature_ecx_include;
4338 cp->cp_ecx &= ~cpuid_feature_ecx_exclude;
4339
4340 if (cp->cp_edx & CPUID_INTC_EDX_PSE) {
4341 add_x86_feature(featureset, X86FSET_LARGEPAGE);
4342 }
4343 if (cp->cp_edx & CPUID_INTC_EDX_TSC) {
4344 add_x86_feature(featureset, X86FSET_TSC);
4345 }
4346 if (cp->cp_edx & CPUID_INTC_EDX_MSR) {
4347 add_x86_feature(featureset, X86FSET_MSR);
4348 }
4349 if (cp->cp_edx & CPUID_INTC_EDX_MTRR) {
4350 add_x86_feature(featureset, X86FSET_MTRR);
4351 }
4352 if (cp->cp_edx & CPUID_INTC_EDX_PGE) {
4353 add_x86_feature(featureset, X86FSET_PGE);
4354 }
4355 if (cp->cp_edx & CPUID_INTC_EDX_CMOV) {
4356 add_x86_feature(featureset, X86FSET_CMOV);
4357 }
4358 if (cp->cp_edx & CPUID_INTC_EDX_MMX) {
4359 add_x86_feature(featureset, X86FSET_MMX);
4360 }
4361 if ((cp->cp_edx & CPUID_INTC_EDX_MCE) != 0 &&
4362 (cp->cp_edx & CPUID_INTC_EDX_MCA) != 0) {
4363 add_x86_feature(featureset, X86FSET_MCA);
4364 }
4365 if (cp->cp_edx & CPUID_INTC_EDX_PAE) {
4366 add_x86_feature(featureset, X86FSET_PAE);
4367 }
4368 if (cp->cp_edx & CPUID_INTC_EDX_CX8) {
4369 add_x86_feature(featureset, X86FSET_CX8);
4370 }
4371 if (cp->cp_ecx & CPUID_INTC_ECX_CX16) {
4372 add_x86_feature(featureset, X86FSET_CX16);
4373 }
4374 if (cp->cp_edx & CPUID_INTC_EDX_PAT) {
4375 add_x86_feature(featureset, X86FSET_PAT);
4376 }
4377 if (cp->cp_edx & CPUID_INTC_EDX_SEP) {
4378 add_x86_feature(featureset, X86FSET_SEP);
4379 }
4380 if (cp->cp_edx & CPUID_INTC_EDX_FXSR) {
4381 /*
4382 * In our implementation, fxsave/fxrstor
4383 * are prerequisites before we'll even
4384 * try and do SSE things.
4385 */
4386 if (cp->cp_edx & CPUID_INTC_EDX_SSE) {
4387 add_x86_feature(featureset, X86FSET_SSE);
4388 }
4389 if (cp->cp_edx & CPUID_INTC_EDX_SSE2) {
4390 add_x86_feature(featureset, X86FSET_SSE2);
4391 }
4392 if (cp->cp_ecx & CPUID_INTC_ECX_SSE3) {
4393 add_x86_feature(featureset, X86FSET_SSE3);
4394 }
4395 if (cp->cp_ecx & CPUID_INTC_ECX_SSSE3) {
4396 add_x86_feature(featureset, X86FSET_SSSE3);
4397 }
4398 if (cp->cp_ecx & CPUID_INTC_ECX_SSE4_1) {
4399 add_x86_feature(featureset, X86FSET_SSE4_1);
4400 }
4401 if (cp->cp_ecx & CPUID_INTC_ECX_SSE4_2) {
4402 add_x86_feature(featureset, X86FSET_SSE4_2);
4403 }
4404 if (cp->cp_ecx & CPUID_INTC_ECX_AES) {
4405 add_x86_feature(featureset, X86FSET_AES);
4406 }
4407 if (cp->cp_ecx & CPUID_INTC_ECX_PCLMULQDQ) {
4408 add_x86_feature(featureset, X86FSET_PCLMULQDQ);
4409 }
4410
4411 if (cpi->cpi_std[7].cp_ebx & CPUID_INTC_EBX_7_0_SHA)
4412 add_x86_feature(featureset, X86FSET_SHA);
4413
4414 if (cp->cp_ecx & CPUID_INTC_ECX_XSAVE) {
4415 add_x86_feature(featureset, X86FSET_XSAVE);
4416
4417 /* We only test AVX & AVX512 when there is XSAVE */
4418 cpuid_basic_avx(cpu, featureset);
4419 }
4420 }
4421
4422 if (cp->cp_ecx & CPUID_INTC_ECX_PCID) {
4423 add_x86_feature(featureset, X86FSET_PCID);
4424 }
4425
4426 if (cp->cp_ecx & CPUID_INTC_ECX_X2APIC) {
4427 add_x86_feature(featureset, X86FSET_X2APIC);
4428 }
4429 if (cp->cp_edx & CPUID_INTC_EDX_DE) {
4430 add_x86_feature(featureset, X86FSET_DE);
4431 }
4432 #if !defined(__xpv)
4433 if (cp->cp_ecx & CPUID_INTC_ECX_MON) {
4434
4435 /*
4436 * We require the CLFLUSH instruction for erratum workaround
4437 * to use MONITOR/MWAIT.
4438 */
4439 if (cp->cp_edx & CPUID_INTC_EDX_CLFSH) {
4440 cpi->cpi_mwait.support |= MWAIT_SUPPORT;
4441 add_x86_feature(featureset, X86FSET_MWAIT);
4442 } else {
4443 extern int idle_cpu_assert_cflush_monitor;
4444
4445 /*
4446 * All processors we are aware of which have
4447 * MONITOR/MWAIT also have CLFLUSH.
4448 */
4449 if (idle_cpu_assert_cflush_monitor) {
4450 ASSERT((cp->cp_ecx & CPUID_INTC_ECX_MON) &&
4451 (cp->cp_edx & CPUID_INTC_EDX_CLFSH));
4452 }
4453 }
4454 }
4455 #endif /* __xpv */
4456
4457 if (cp->cp_ecx & CPUID_INTC_ECX_VMX) {
4458 add_x86_feature(featureset, X86FSET_VMX);
4459 }
4460
4461 if (cp->cp_ecx & CPUID_INTC_ECX_RDRAND)
4462 add_x86_feature(featureset, X86FSET_RDRAND);
4463
4464 /*
4465 * Only need it first time, rest of the cpus would follow suit.
4466 * we only capture this for the bootcpu.
4467 */
4468 if (cp->cp_edx & CPUID_INTC_EDX_CLFSH) {
4469 add_x86_feature(featureset, X86FSET_CLFSH);
4470 x86_clflush_size = (BITX(cp->cp_ebx, 15, 8) * 8);
4471 }
4472 if (is_x86_feature(featureset, X86FSET_PAE))
4473 cpi->cpi_pabits = 36;
4474
4475 if (cpi->cpi_maxeax >= 0xD && !xsave_force_disable) {
4476 struct cpuid_regs r, *ecp;
4477
4478 ecp = &r;
4479 ecp->cp_eax = 0xD;
4480 ecp->cp_ecx = 1;
4481 ecp->cp_edx = ecp->cp_ebx = 0;
4482 (void) __cpuid_insn(ecp);
4483
4484 if (ecp->cp_eax & CPUID_INTC_EAX_D_1_XSAVEOPT)
4485 add_x86_feature(featureset, X86FSET_XSAVEOPT);
4486 if (ecp->cp_eax & CPUID_INTC_EAX_D_1_XSAVEC)
4487 add_x86_feature(featureset, X86FSET_XSAVEC);
4488 if (ecp->cp_eax & CPUID_INTC_EAX_D_1_XSAVES)
4489 add_x86_feature(featureset, X86FSET_XSAVES);
4490
4491 /*
4492 * Zen 2 family processors suffer from erratum 1386 that causes
4493 * xsaves to not function correctly in some circumstances. There
4494 * are no supervisor states in Zen 2 and earlier. Practically
4495 * speaking this has no impact for us as we currently do not
4496 * leverage compressed xsave formats. To safeguard against
4497 * issues in the future where we may opt to using it, we remove
4498 * it from the feature set now. While Matisse has a microcode
4499 * update available with a fix, not all Zen 2 CPUs do so it's
4500 * simpler for the moment to unconditionally remove it.
4501 */
4502 if (cpi->cpi_vendor == X86_VENDOR_AMD &&
4503 uarchrev_uarch(cpi->cpi_uarchrev) <= X86_UARCH_AMD_ZEN2) {
4504 remove_x86_feature(featureset, X86FSET_XSAVES);
4505 }
4506 }
4507
4508 /*
4509 * Work on the "extended" feature information, doing
4510 * some basic initialization to be used in the extended pass.
4511 */
4512 xcpuid = 0;
4513 switch (cpi->cpi_vendor) {
4514 case X86_VENDOR_Intel:
4515 /*
4516 * On KVM we know we will have proper support for extended
4517 * cpuid.
4518 */
4519 if (IS_NEW_F6(cpi) || cpi->cpi_family >= 0xf ||
4520 (get_hwenv() == HW_KVM && cpi->cpi_family == 6 &&
4521 (cpi->cpi_model == 6 || cpi->cpi_model == 2)))
4522 xcpuid++;
4523 break;
4524 case X86_VENDOR_AMD:
4525 if (cpi->cpi_family > 5 ||
4526 (cpi->cpi_family == 5 && cpi->cpi_model >= 1))
4527 xcpuid++;
4528 break;
4529 case X86_VENDOR_Cyrix:
4530 /*
4531 * Only these Cyrix CPUs are -known- to support
4532 * extended cpuid operations.
4533 */
4534 if (x86_type == X86_TYPE_VIA_CYRIX_III ||
4535 x86_type == X86_TYPE_CYRIX_GXm)
4536 xcpuid++;
4537 break;
4538 case X86_VENDOR_HYGON:
4539 case X86_VENDOR_Centaur:
4540 case X86_VENDOR_TM:
4541 default:
4542 xcpuid++;
4543 break;
4544 }
4545
4546 if (xcpuid) {
4547 cp = &cpi->cpi_extd[0];
4548 cp->cp_eax = CPUID_LEAF_EXT_0;
4549 cpi->cpi_xmaxeax = __cpuid_insn(cp);
4550 }
4551
4552 if (cpi->cpi_xmaxeax & CPUID_LEAF_EXT_0) {
4553
4554 if (cpi->cpi_xmaxeax > CPI_XMAXEAX_MAX)
4555 cpi->cpi_xmaxeax = CPI_XMAXEAX_MAX;
4556
4557 switch (cpi->cpi_vendor) {
4558 case X86_VENDOR_Intel:
4559 case X86_VENDOR_AMD:
4560 case X86_VENDOR_HYGON:
4561 if (cpi->cpi_xmaxeax < 0x80000001)
4562 break;
4563 cp = &cpi->cpi_extd[1];
4564 cp->cp_eax = 0x80000001;
4565 (void) __cpuid_insn(cp);
4566
4567 if (cpi->cpi_vendor == X86_VENDOR_AMD &&
4568 cpi->cpi_family == 5 &&
4569 cpi->cpi_model == 6 &&
4570 cpi->cpi_step == 6) {
4571 /*
4572 * K6 model 6 uses bit 10 to indicate SYSC
4573 * Later models use bit 11. Fix it here.
4574 */
4575 if (cp->cp_edx & 0x400) {
4576 cp->cp_edx &= ~0x400;
4577 cp->cp_edx |= CPUID_AMD_EDX_SYSC;
4578 }
4579 }
4580
4581 platform_cpuid_mangle(cpi->cpi_vendor, 0x80000001, cp);
4582
4583 /*
4584 * Compute the additions to the kernel's feature word.
4585 */
4586 if (cp->cp_edx & CPUID_AMD_EDX_NX) {
4587 add_x86_feature(featureset, X86FSET_NX);
4588 }
4589
4590 /*
4591 * Regardless whether or not we boot 64-bit,
4592 * we should have a way to identify whether
4593 * the CPU is capable of running 64-bit.
4594 */
4595 if (cp->cp_edx & CPUID_AMD_EDX_LM) {
4596 add_x86_feature(featureset, X86FSET_64);
4597 }
4598
4599 /* 1 GB large page - enable only for 64 bit kernel */
4600 if (cp->cp_edx & CPUID_AMD_EDX_1GPG) {
4601 add_x86_feature(featureset, X86FSET_1GPG);
4602 }
4603
4604 if ((cpi->cpi_vendor == X86_VENDOR_AMD ||
4605 cpi->cpi_vendor == X86_VENDOR_HYGON) &&
4606 (cpi->cpi_std[1].cp_edx & CPUID_INTC_EDX_FXSR) &&
4607 (cp->cp_ecx & CPUID_AMD_ECX_SSE4A)) {
4608 add_x86_feature(featureset, X86FSET_SSE4A);
4609 }
4610
4611 /*
4612 * It's really tricky to support syscall/sysret in
4613 * the i386 kernel; we rely on sysenter/sysexit
4614 * instead. In the amd64 kernel, things are -way-
4615 * better.
4616 */
4617 if (cp->cp_edx & CPUID_AMD_EDX_SYSC) {
4618 add_x86_feature(featureset, X86FSET_ASYSC);
4619 }
4620
4621 /*
4622 * While we're thinking about system calls, note
4623 * that AMD processors don't support sysenter
4624 * in long mode at all, so don't try to program them.
4625 */
4626 if (x86_vendor == X86_VENDOR_AMD ||
4627 x86_vendor == X86_VENDOR_HYGON) {
4628 remove_x86_feature(featureset, X86FSET_SEP);
4629 }
4630
4631 if (cp->cp_edx & CPUID_AMD_EDX_TSCP) {
4632 add_x86_feature(featureset, X86FSET_TSCP);
4633 }
4634
4635 if (cp->cp_ecx & CPUID_AMD_ECX_SVM) {
4636 add_x86_feature(featureset, X86FSET_SVM);
4637 }
4638
4639 if (cp->cp_ecx & CPUID_AMD_ECX_TOPOEXT) {
4640 add_x86_feature(featureset, X86FSET_TOPOEXT);
4641 }
4642
4643 if (cp->cp_ecx & CPUID_AMD_ECX_PCEC) {
4644 add_x86_feature(featureset, X86FSET_AMD_PCEC);
4645 }
4646
4647 if (cp->cp_ecx & CPUID_AMD_ECX_XOP) {
4648 add_x86_feature(featureset, X86FSET_XOP);
4649 }
4650
4651 if (cp->cp_ecx & CPUID_AMD_ECX_FMA4) {
4652 add_x86_feature(featureset, X86FSET_FMA4);
4653 }
4654
4655 if (cp->cp_ecx & CPUID_AMD_ECX_TBM) {
4656 add_x86_feature(featureset, X86FSET_TBM);
4657 }
4658
4659 if (cp->cp_ecx & CPUID_AMD_ECX_MONITORX) {
4660 add_x86_feature(featureset, X86FSET_MONITORX);
4661 }
4662 break;
4663 default:
4664 break;
4665 }
4666
4667 /*
4668 * Get CPUID data about processor cores and hyperthreads.
4669 */
4670 switch (cpi->cpi_vendor) {
4671 case X86_VENDOR_Intel:
4672 if (cpi->cpi_maxeax >= 4) {
4673 cp = &cpi->cpi_std[4];
4674 cp->cp_eax = 4;
4675 cp->cp_ecx = 0;
4676 (void) __cpuid_insn(cp);
4677 platform_cpuid_mangle(cpi->cpi_vendor, 4, cp);
4678 }
4679 /*FALLTHROUGH*/
4680 case X86_VENDOR_AMD:
4681 case X86_VENDOR_HYGON:
4682 if (cpi->cpi_xmaxeax < CPUID_LEAF_EXT_8)
4683 break;
4684 cp = &cpi->cpi_extd[8];
4685 cp->cp_eax = CPUID_LEAF_EXT_8;
4686 (void) __cpuid_insn(cp);
4687 platform_cpuid_mangle(cpi->cpi_vendor, CPUID_LEAF_EXT_8,
4688 cp);
4689
4690 /*
4691 * AMD uses ebx for some extended functions.
4692 */
4693 if (cpi->cpi_vendor == X86_VENDOR_AMD ||
4694 cpi->cpi_vendor == X86_VENDOR_HYGON) {
4695 /*
4696 * While we're here, check for the AMD "Error
4697 * Pointer Zero/Restore" feature. This can be
4698 * used to setup the FP save handlers
4699 * appropriately.
4700 */
4701 if (cp->cp_ebx & CPUID_AMD_EBX_ERR_PTR_ZERO) {
4702 cpi->cpi_fp_amd_save = 0;
4703 } else {
4704 cpi->cpi_fp_amd_save = 1;
4705 }
4706
4707 if (cp->cp_ebx & CPUID_AMD_EBX_CLZERO) {
4708 add_x86_feature(featureset,
4709 X86FSET_CLZERO);
4710 }
4711 }
4712
4713 /*
4714 * Virtual and physical address limits from
4715 * cpuid override previously guessed values.
4716 */
4717 cpi->cpi_pabits = BITX(cp->cp_eax, 7, 0);
4718 cpi->cpi_vabits = BITX(cp->cp_eax, 15, 8);
4719 break;
4720 default:
4721 break;
4722 }
4723
4724 /*
4725 * Get CPUID data about TSC Invariance in Deep C-State.
4726 */
4727 switch (cpi->cpi_vendor) {
4728 case X86_VENDOR_Intel:
4729 case X86_VENDOR_AMD:
4730 case X86_VENDOR_HYGON:
4731 if (cpi->cpi_maxeax >= 7) {
4732 cp = &cpi->cpi_extd[7];
4733 cp->cp_eax = 0x80000007;
4734 cp->cp_ecx = 0;
4735 (void) __cpuid_insn(cp);
4736 }
4737 break;
4738 default:
4739 break;
4740 }
4741 }
4742
4743 /*
4744 * cpuid_basic_ppin assumes that cpuid_basic_topology has already been
4745 * run and thus gathered some of its dependent leaves.
4746 */
4747 cpuid_basic_topology(cpu, featureset);
4748 cpuid_basic_thermal(cpu, featureset);
4749 #if !defined(__xpv)
4750 cpuid_basic_ppin(cpu, featureset);
4751 #endif
4752
4753 if (cpi->cpi_vendor == X86_VENDOR_AMD ||
4754 cpi->cpi_vendor == X86_VENDOR_HYGON) {
4755 if (cpi->cpi_xmaxeax >= CPUID_LEAF_EXT_8 &&
4756 cpi->cpi_extd[8].cp_ebx & CPUID_AMD_EBX_ERR_PTR_ZERO) {
4757 /* Special handling for AMD FP not necessary. */
4758 cpi->cpi_fp_amd_save = 0;
4759 } else {
4760 cpi->cpi_fp_amd_save = 1;
4761 }
4762 }
4763
4764 /*
4765 * Check (and potentially set) if lfence is serializing.
4766 * This is useful for accurate rdtsc measurements and AMD retpolines.
4767 */
4768 if ((cpi->cpi_vendor == X86_VENDOR_AMD ||
4769 cpi->cpi_vendor == X86_VENDOR_HYGON) &&
4770 is_x86_feature(featureset, X86FSET_SSE2)) {
4771 /*
4772 * The AMD white paper Software Techniques For Managing
4773 * Speculation on AMD Processors details circumstances for when
4774 * lfence instructions are serializing.
4775 *
4776 * On family 0xf and 0x11, it is inherently so. On family 0x10
4777 * and later (excluding 0x11), a bit in the DE_CFG MSR
4778 * determines the lfence behavior. Per that whitepaper, AMD has
4779 * committed to supporting that MSR on all later CPUs.
4780 */
4781 if (cpi->cpi_family == 0xf || cpi->cpi_family == 0x11) {
4782 add_x86_feature(featureset, X86FSET_LFENCE_SER);
4783 } else if (cpi->cpi_family >= 0x10) {
4784 #if !defined(__xpv)
4785 uint64_t val;
4786
4787 /*
4788 * Be careful when attempting to enable the bit, and
4789 * verify that it was actually set in case we are
4790 * running in a hypervisor which is less than faithful
4791 * about its emulation of this feature.
4792 */
4793 on_trap_data_t otd;
4794 if (!on_trap(&otd, OT_DATA_ACCESS)) {
4795 val = rdmsr(MSR_AMD_DE_CFG);
4796 val |= AMD_DE_CFG_LFENCE_DISPATCH;
4797 wrmsr(MSR_AMD_DE_CFG, val);
4798 val = rdmsr(MSR_AMD_DE_CFG);
4799 } else {
4800 val = 0;
4801 }
4802 no_trap();
4803
4804 if ((val & AMD_DE_CFG_LFENCE_DISPATCH) != 0) {
4805 add_x86_feature(featureset, X86FSET_LFENCE_SER);
4806 }
4807 #endif
4808 }
4809 } else if (cpi->cpi_vendor == X86_VENDOR_Intel &&
4810 is_x86_feature(featureset, X86FSET_SSE2)) {
4811 /*
4812 * Documentation and other OSes indicate that lfence is always
4813 * serializing on Intel CPUs.
4814 */
4815 add_x86_feature(featureset, X86FSET_LFENCE_SER);
4816 }
4817
4818
4819 /*
4820 * Check the processor leaves that are used for security features. Grab
4821 * any additional processor-specific leaves that we may not have yet.
4822 */
4823 switch (cpi->cpi_vendor) {
4824 case X86_VENDOR_AMD:
4825 case X86_VENDOR_HYGON:
4826 if (cpi->cpi_xmaxeax >= CPUID_LEAF_EXT_21) {
4827 cp = &cpi->cpi_extd[7];
4828 cp->cp_eax = CPUID_LEAF_EXT_21;
4829 cp->cp_ecx = 0;
4830 (void) __cpuid_insn(cp);
4831 }
4832 break;
4833 default:
4834 break;
4835 }
4836
4837 cpuid_scan_security(cpu, featureset);
4838 }
4839
4840 /*
4841 * Make copies of the cpuid table entries we depend on, in
4842 * part for ease of parsing now, in part so that we have only
4843 * one place to correct any of it, in part for ease of
4844 * later export to userland, and in part so we can look at
4845 * this stuff in a crash dump.
4846 */
4847
4848 static void
cpuid_pass_extended(cpu_t * cpu,void * _arg __unused)4849 cpuid_pass_extended(cpu_t *cpu, void *_arg __unused)
4850 {
4851 uint_t n, nmax;
4852 int i;
4853 struct cpuid_regs *cp;
4854 uint8_t *dp;
4855 uint32_t *iptr;
4856 struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
4857
4858 if (cpi->cpi_maxeax < 1)
4859 return;
4860
4861 if ((nmax = cpi->cpi_maxeax + 1) > NMAX_CPI_STD)
4862 nmax = NMAX_CPI_STD;
4863 /*
4864 * (We already handled n == 0 and n == 1 in the basic pass)
4865 */
4866 for (n = 2, cp = &cpi->cpi_std[2]; n < nmax; n++, cp++) {
4867 /*
4868 * leaves 6 and 7 were handled in the basic pass
4869 */
4870 if (n == 6 || n == 7)
4871 continue;
4872
4873 cp->cp_eax = n;
4874
4875 /*
4876 * CPUID function 4 expects %ecx to be initialized
4877 * with an index which indicates which cache to return
4878 * information about. The OS is expected to call function 4
4879 * with %ecx set to 0, 1, 2, ... until it returns with
4880 * EAX[4:0] set to 0, which indicates there are no more
4881 * caches.
4882 *
4883 * Here, populate cpi_std[4] with the information returned by
4884 * function 4 when %ecx == 0, and do the rest in a later pass
4885 * when dynamic memory allocation becomes available.
4886 *
4887 * Note: we need to explicitly initialize %ecx here, since
4888 * function 4 may have been previously invoked.
4889 */
4890 if (n == 4)
4891 cp->cp_ecx = 0;
4892
4893 (void) __cpuid_insn(cp);
4894 platform_cpuid_mangle(cpi->cpi_vendor, n, cp);
4895 switch (n) {
4896 case 2:
4897 /*
4898 * "the lower 8 bits of the %eax register
4899 * contain a value that identifies the number
4900 * of times the cpuid [instruction] has to be
4901 * executed to obtain a complete image of the
4902 * processor's caching systems."
4903 *
4904 * How *do* they make this stuff up?
4905 */
4906 cpi->cpi_ncache = sizeof (*cp) *
4907 BITX(cp->cp_eax, 7, 0);
4908 if (cpi->cpi_ncache == 0)
4909 break;
4910 cpi->cpi_ncache--; /* skip count byte */
4911
4912 /*
4913 * Well, for now, rather than attempt to implement
4914 * this slightly dubious algorithm, we just look
4915 * at the first 15 ..
4916 */
4917 if (cpi->cpi_ncache > (sizeof (*cp) - 1))
4918 cpi->cpi_ncache = sizeof (*cp) - 1;
4919
4920 dp = cpi->cpi_cacheinfo;
4921 if (BITX(cp->cp_eax, 31, 31) == 0) {
4922 uint8_t *p = (void *)&cp->cp_eax;
4923 for (i = 1; i < 4; i++)
4924 if (p[i] != 0)
4925 *dp++ = p[i];
4926 }
4927 if (BITX(cp->cp_ebx, 31, 31) == 0) {
4928 uint8_t *p = (void *)&cp->cp_ebx;
4929 for (i = 0; i < 4; i++)
4930 if (p[i] != 0)
4931 *dp++ = p[i];
4932 }
4933 if (BITX(cp->cp_ecx, 31, 31) == 0) {
4934 uint8_t *p = (void *)&cp->cp_ecx;
4935 for (i = 0; i < 4; i++)
4936 if (p[i] != 0)
4937 *dp++ = p[i];
4938 }
4939 if (BITX(cp->cp_edx, 31, 31) == 0) {
4940 uint8_t *p = (void *)&cp->cp_edx;
4941 for (i = 0; i < 4; i++)
4942 if (p[i] != 0)
4943 *dp++ = p[i];
4944 }
4945 break;
4946
4947 case 3: /* Processor serial number, if PSN supported */
4948 break;
4949
4950 case 4: /* Deterministic cache parameters */
4951 break;
4952
4953 case 5: /* Monitor/Mwait parameters */
4954 {
4955 size_t mwait_size;
4956
4957 /*
4958 * check cpi_mwait.support which was set in
4959 * cpuid_pass_basic()
4960 */
4961 if (!(cpi->cpi_mwait.support & MWAIT_SUPPORT))
4962 break;
4963
4964 /*
4965 * Protect ourself from insane mwait line size.
4966 * Workaround for incomplete hardware emulator(s).
4967 */
4968 mwait_size = (size_t)MWAIT_SIZE_MAX(cpi);
4969 if (mwait_size < sizeof (uint32_t) ||
4970 !ISP2(mwait_size)) {
4971 #if DEBUG
4972 cmn_err(CE_NOTE, "Cannot handle cpu %d mwait "
4973 "size %ld", cpu->cpu_id, (long)mwait_size);
4974 #endif
4975 break;
4976 }
4977
4978 cpi->cpi_mwait.mon_min = (size_t)MWAIT_SIZE_MIN(cpi);
4979 cpi->cpi_mwait.mon_max = mwait_size;
4980 if (MWAIT_EXTENSION(cpi)) {
4981 cpi->cpi_mwait.support |= MWAIT_EXTENSIONS;
4982 if (MWAIT_INT_ENABLE(cpi))
4983 cpi->cpi_mwait.support |=
4984 MWAIT_ECX_INT_ENABLE;
4985 }
4986 break;
4987 }
4988 default:
4989 break;
4990 }
4991 }
4992
4993 /*
4994 * XSAVE enumeration
4995 */
4996 if (cpi->cpi_maxeax >= 0xD) {
4997 struct cpuid_regs regs;
4998 boolean_t cpuid_d_valid = B_TRUE;
4999
5000 cp = ®s;
5001 cp->cp_eax = 0xD;
5002 cp->cp_edx = cp->cp_ebx = cp->cp_ecx = 0;
5003
5004 (void) __cpuid_insn(cp);
5005
5006 /*
5007 * Sanity checks for debug
5008 */
5009 if ((cp->cp_eax & XFEATURE_LEGACY_FP) == 0 ||
5010 (cp->cp_eax & XFEATURE_SSE) == 0) {
5011 cpuid_d_valid = B_FALSE;
5012 }
5013
5014 cpi->cpi_xsave.xsav_hw_features_low = cp->cp_eax;
5015 cpi->cpi_xsave.xsav_hw_features_high = cp->cp_edx;
5016 cpi->cpi_xsave.xsav_max_size = cp->cp_ecx;
5017
5018 /*
5019 * If the hw supports AVX, get the size and offset in the save
5020 * area for the ymm state.
5021 */
5022 if (cpi->cpi_xsave.xsav_hw_features_low & XFEATURE_AVX) {
5023 cp->cp_eax = 0xD;
5024 cp->cp_ecx = 2;
5025 cp->cp_edx = cp->cp_ebx = 0;
5026
5027 (void) __cpuid_insn(cp);
5028
5029 if (cp->cp_ebx != CPUID_LEAFD_2_YMM_OFFSET ||
5030 cp->cp_eax != CPUID_LEAFD_2_YMM_SIZE) {
5031 cpuid_d_valid = B_FALSE;
5032 }
5033
5034 cpi->cpi_xsave.ymm_size = cp->cp_eax;
5035 cpi->cpi_xsave.ymm_offset = cp->cp_ebx;
5036 }
5037
5038 /*
5039 * If the hw supports MPX, get the size and offset in the
5040 * save area for BNDREGS and BNDCSR.
5041 */
5042 if (cpi->cpi_xsave.xsav_hw_features_low & XFEATURE_MPX) {
5043 cp->cp_eax = 0xD;
5044 cp->cp_ecx = 3;
5045 cp->cp_edx = cp->cp_ebx = 0;
5046
5047 (void) __cpuid_insn(cp);
5048
5049 cpi->cpi_xsave.bndregs_size = cp->cp_eax;
5050 cpi->cpi_xsave.bndregs_offset = cp->cp_ebx;
5051
5052 cp->cp_eax = 0xD;
5053 cp->cp_ecx = 4;
5054 cp->cp_edx = cp->cp_ebx = 0;
5055
5056 (void) __cpuid_insn(cp);
5057
5058 cpi->cpi_xsave.bndcsr_size = cp->cp_eax;
5059 cpi->cpi_xsave.bndcsr_offset = cp->cp_ebx;
5060 }
5061
5062 /*
5063 * If the hw supports AVX512, get the size and offset in the
5064 * save area for the opmask registers and zmm state.
5065 */
5066 if (cpi->cpi_xsave.xsav_hw_features_low & XFEATURE_AVX512) {
5067 cp->cp_eax = 0xD;
5068 cp->cp_ecx = 5;
5069 cp->cp_edx = cp->cp_ebx = 0;
5070
5071 (void) __cpuid_insn(cp);
5072
5073 cpi->cpi_xsave.opmask_size = cp->cp_eax;
5074 cpi->cpi_xsave.opmask_offset = cp->cp_ebx;
5075
5076 cp->cp_eax = 0xD;
5077 cp->cp_ecx = 6;
5078 cp->cp_edx = cp->cp_ebx = 0;
5079
5080 (void) __cpuid_insn(cp);
5081
5082 cpi->cpi_xsave.zmmlo_size = cp->cp_eax;
5083 cpi->cpi_xsave.zmmlo_offset = cp->cp_ebx;
5084
5085 cp->cp_eax = 0xD;
5086 cp->cp_ecx = 7;
5087 cp->cp_edx = cp->cp_ebx = 0;
5088
5089 (void) __cpuid_insn(cp);
5090
5091 cpi->cpi_xsave.zmmhi_size = cp->cp_eax;
5092 cpi->cpi_xsave.zmmhi_offset = cp->cp_ebx;
5093 }
5094
5095 if (is_x86_feature(x86_featureset, X86FSET_XSAVE)) {
5096 xsave_state_size = 0;
5097 } else if (cpuid_d_valid) {
5098 xsave_state_size = cpi->cpi_xsave.xsav_max_size;
5099 } else {
5100 /* Broken CPUID 0xD, probably in HVM */
5101 cmn_err(CE_WARN, "cpu%d: CPUID.0xD returns invalid "
5102 "value: hw_low = %d, hw_high = %d, xsave_size = %d"
5103 ", ymm_size = %d, ymm_offset = %d\n",
5104 cpu->cpu_id, cpi->cpi_xsave.xsav_hw_features_low,
5105 cpi->cpi_xsave.xsav_hw_features_high,
5106 (int)cpi->cpi_xsave.xsav_max_size,
5107 (int)cpi->cpi_xsave.ymm_size,
5108 (int)cpi->cpi_xsave.ymm_offset);
5109
5110 if (xsave_state_size != 0) {
5111 /*
5112 * This must be a non-boot CPU. We cannot
5113 * continue, because boot cpu has already
5114 * enabled XSAVE.
5115 */
5116 ASSERT(cpu->cpu_id != 0);
5117 cmn_err(CE_PANIC, "cpu%d: we have already "
5118 "enabled XSAVE on boot cpu, cannot "
5119 "continue.", cpu->cpu_id);
5120 } else {
5121 /*
5122 * If we reached here on the boot CPU, it's also
5123 * almost certain that we'll reach here on the
5124 * non-boot CPUs. When we're here on a boot CPU
5125 * we should disable the feature, on a non-boot
5126 * CPU we need to confirm that we have.
5127 */
5128 if (cpu->cpu_id == 0) {
5129 remove_x86_feature(x86_featureset,
5130 X86FSET_XSAVE);
5131 remove_x86_feature(x86_featureset,
5132 X86FSET_AVX);
5133 remove_x86_feature(x86_featureset,
5134 X86FSET_F16C);
5135 remove_x86_feature(x86_featureset,
5136 X86FSET_BMI1);
5137 remove_x86_feature(x86_featureset,
5138 X86FSET_BMI2);
5139 remove_x86_feature(x86_featureset,
5140 X86FSET_FMA);
5141 remove_x86_feature(x86_featureset,
5142 X86FSET_AVX2);
5143 remove_x86_feature(x86_featureset,
5144 X86FSET_MPX);
5145 remove_x86_feature(x86_featureset,
5146 X86FSET_AVX512F);
5147 remove_x86_feature(x86_featureset,
5148 X86FSET_AVX512DQ);
5149 remove_x86_feature(x86_featureset,
5150 X86FSET_AVX512PF);
5151 remove_x86_feature(x86_featureset,
5152 X86FSET_AVX512ER);
5153 remove_x86_feature(x86_featureset,
5154 X86FSET_AVX512CD);
5155 remove_x86_feature(x86_featureset,
5156 X86FSET_AVX512BW);
5157 remove_x86_feature(x86_featureset,
5158 X86FSET_AVX512VL);
5159 remove_x86_feature(x86_featureset,
5160 X86FSET_AVX512FMA);
5161 remove_x86_feature(x86_featureset,
5162 X86FSET_AVX512VBMI);
5163 remove_x86_feature(x86_featureset,
5164 X86FSET_AVX512VNNI);
5165 remove_x86_feature(x86_featureset,
5166 X86FSET_AVX512VPOPCDQ);
5167 remove_x86_feature(x86_featureset,
5168 X86FSET_AVX512NNIW);
5169 remove_x86_feature(x86_featureset,
5170 X86FSET_AVX512FMAPS);
5171 remove_x86_feature(x86_featureset,
5172 X86FSET_VAES);
5173 remove_x86_feature(x86_featureset,
5174 X86FSET_VPCLMULQDQ);
5175 remove_x86_feature(x86_featureset,
5176 X86FSET_GFNI);
5177 remove_x86_feature(x86_featureset,
5178 X86FSET_AVX512_VP2INT);
5179 remove_x86_feature(x86_featureset,
5180 X86FSET_AVX512_BITALG);
5181 remove_x86_feature(x86_featureset,
5182 X86FSET_AVX512_VBMI2);
5183 remove_x86_feature(x86_featureset,
5184 X86FSET_AVX512_BF16);
5185
5186 xsave_force_disable = B_TRUE;
5187 } else {
5188 VERIFY(is_x86_feature(x86_featureset,
5189 X86FSET_XSAVE) == B_FALSE);
5190 }
5191 }
5192 }
5193 }
5194
5195
5196 if ((cpi->cpi_xmaxeax & CPUID_LEAF_EXT_0) == 0)
5197 return;
5198
5199 if ((nmax = cpi->cpi_xmaxeax - CPUID_LEAF_EXT_0 + 1) > NMAX_CPI_EXTD)
5200 nmax = NMAX_CPI_EXTD;
5201 /*
5202 * Copy the extended properties, fixing them as we go. While we start at
5203 * 2 because we've already handled a few cases in the basic pass, the
5204 * rest we let ourselves just grab again (e.g. 0x8, 0x21).
5205 */
5206 iptr = (void *)cpi->cpi_brandstr;
5207 for (n = 2, cp = &cpi->cpi_extd[2]; n < nmax; cp++, n++) {
5208 cp->cp_eax = CPUID_LEAF_EXT_0 + n;
5209 (void) __cpuid_insn(cp);
5210 platform_cpuid_mangle(cpi->cpi_vendor, CPUID_LEAF_EXT_0 + n,
5211 cp);
5212 switch (n) {
5213 case 2:
5214 case 3:
5215 case 4:
5216 /*
5217 * Extract the brand string
5218 */
5219 *iptr++ = cp->cp_eax;
5220 *iptr++ = cp->cp_ebx;
5221 *iptr++ = cp->cp_ecx;
5222 *iptr++ = cp->cp_edx;
5223 break;
5224 case 5:
5225 switch (cpi->cpi_vendor) {
5226 case X86_VENDOR_AMD:
5227 /*
5228 * The Athlon and Duron were the first
5229 * parts to report the sizes of the
5230 * TLB for large pages. Before then,
5231 * we don't trust the data.
5232 */
5233 if (cpi->cpi_family < 6 ||
5234 (cpi->cpi_family == 6 &&
5235 cpi->cpi_model < 1))
5236 cp->cp_eax = 0;
5237 break;
5238 default:
5239 break;
5240 }
5241 break;
5242 case 6:
5243 switch (cpi->cpi_vendor) {
5244 case X86_VENDOR_AMD:
5245 /*
5246 * The Athlon and Duron were the first
5247 * AMD parts with L2 TLB's.
5248 * Before then, don't trust the data.
5249 */
5250 if (cpi->cpi_family < 6 ||
5251 (cpi->cpi_family == 6 &&
5252 cpi->cpi_model < 1))
5253 cp->cp_eax = cp->cp_ebx = 0;
5254 /*
5255 * AMD Duron rev A0 reports L2
5256 * cache size incorrectly as 1K
5257 * when it is really 64K
5258 */
5259 if (cpi->cpi_family == 6 &&
5260 cpi->cpi_model == 3 &&
5261 cpi->cpi_step == 0) {
5262 cp->cp_ecx &= 0xffff;
5263 cp->cp_ecx |= 0x400000;
5264 }
5265 break;
5266 case X86_VENDOR_Cyrix: /* VIA C3 */
5267 /*
5268 * VIA C3 processors are a bit messed
5269 * up w.r.t. encoding cache sizes in %ecx
5270 */
5271 if (cpi->cpi_family != 6)
5272 break;
5273 /*
5274 * model 7 and 8 were incorrectly encoded
5275 *
5276 * xxx is model 8 really broken?
5277 */
5278 if (cpi->cpi_model == 7 ||
5279 cpi->cpi_model == 8)
5280 cp->cp_ecx =
5281 BITX(cp->cp_ecx, 31, 24) << 16 |
5282 BITX(cp->cp_ecx, 23, 16) << 12 |
5283 BITX(cp->cp_ecx, 15, 8) << 8 |
5284 BITX(cp->cp_ecx, 7, 0);
5285 /*
5286 * model 9 stepping 1 has wrong associativity
5287 */
5288 if (cpi->cpi_model == 9 && cpi->cpi_step == 1)
5289 cp->cp_ecx |= 8 << 12;
5290 break;
5291 case X86_VENDOR_Intel:
5292 /*
5293 * Extended L2 Cache features function.
5294 * First appeared on Prescott.
5295 */
5296 default:
5297 break;
5298 }
5299 break;
5300 default:
5301 break;
5302 }
5303 }
5304 }
5305
5306 static const char *
intel_cpubrand(const struct cpuid_info * cpi)5307 intel_cpubrand(const struct cpuid_info *cpi)
5308 {
5309 int i;
5310
5311 ASSERT(is_x86_feature(x86_featureset, X86FSET_CPUID));
5312
5313 switch (cpi->cpi_family) {
5314 case 5:
5315 return ("Intel Pentium(r)");
5316 case 6:
5317 switch (cpi->cpi_model) {
5318 uint_t celeron, xeon;
5319 const struct cpuid_regs *cp;
5320 case 0:
5321 case 1:
5322 case 2:
5323 return ("Intel Pentium(r) Pro");
5324 case 3:
5325 case 4:
5326 return ("Intel Pentium(r) II");
5327 case 6:
5328 return ("Intel Celeron(r)");
5329 case 5:
5330 case 7:
5331 celeron = xeon = 0;
5332 cp = &cpi->cpi_std[2]; /* cache info */
5333
5334 for (i = 1; i < 4; i++) {
5335 uint_t tmp;
5336
5337 tmp = (cp->cp_eax >> (8 * i)) & 0xff;
5338 if (tmp == 0x40)
5339 celeron++;
5340 if (tmp >= 0x44 && tmp <= 0x45)
5341 xeon++;
5342 }
5343
5344 for (i = 0; i < 2; i++) {
5345 uint_t tmp;
5346
5347 tmp = (cp->cp_ebx >> (8 * i)) & 0xff;
5348 if (tmp == 0x40)
5349 celeron++;
5350 else if (tmp >= 0x44 && tmp <= 0x45)
5351 xeon++;
5352 }
5353
5354 for (i = 0; i < 4; i++) {
5355 uint_t tmp;
5356
5357 tmp = (cp->cp_ecx >> (8 * i)) & 0xff;
5358 if (tmp == 0x40)
5359 celeron++;
5360 else if (tmp >= 0x44 && tmp <= 0x45)
5361 xeon++;
5362 }
5363
5364 for (i = 0; i < 4; i++) {
5365 uint_t tmp;
5366
5367 tmp = (cp->cp_edx >> (8 * i)) & 0xff;
5368 if (tmp == 0x40)
5369 celeron++;
5370 else if (tmp >= 0x44 && tmp <= 0x45)
5371 xeon++;
5372 }
5373
5374 if (celeron)
5375 return ("Intel Celeron(r)");
5376 if (xeon)
5377 return (cpi->cpi_model == 5 ?
5378 "Intel Pentium(r) II Xeon(tm)" :
5379 "Intel Pentium(r) III Xeon(tm)");
5380 return (cpi->cpi_model == 5 ?
5381 "Intel Pentium(r) II or Pentium(r) II Xeon(tm)" :
5382 "Intel Pentium(r) III or Pentium(r) III Xeon(tm)");
5383 default:
5384 break;
5385 }
5386 default:
5387 break;
5388 }
5389
5390 /* BrandID is present if the field is nonzero */
5391 if (cpi->cpi_brandid != 0) {
5392 static const struct {
5393 uint_t bt_bid;
5394 const char *bt_str;
5395 } brand_tbl[] = {
5396 { 0x1, "Intel(r) Celeron(r)" },
5397 { 0x2, "Intel(r) Pentium(r) III" },
5398 { 0x3, "Intel(r) Pentium(r) III Xeon(tm)" },
5399 { 0x4, "Intel(r) Pentium(r) III" },
5400 { 0x6, "Mobile Intel(r) Pentium(r) III" },
5401 { 0x7, "Mobile Intel(r) Celeron(r)" },
5402 { 0x8, "Intel(r) Pentium(r) 4" },
5403 { 0x9, "Intel(r) Pentium(r) 4" },
5404 { 0xa, "Intel(r) Celeron(r)" },
5405 { 0xb, "Intel(r) Xeon(tm)" },
5406 { 0xc, "Intel(r) Xeon(tm) MP" },
5407 { 0xe, "Mobile Intel(r) Pentium(r) 4" },
5408 { 0xf, "Mobile Intel(r) Celeron(r)" },
5409 { 0x11, "Mobile Genuine Intel(r)" },
5410 { 0x12, "Intel(r) Celeron(r) M" },
5411 { 0x13, "Mobile Intel(r) Celeron(r)" },
5412 { 0x14, "Intel(r) Celeron(r)" },
5413 { 0x15, "Mobile Genuine Intel(r)" },
5414 { 0x16, "Intel(r) Pentium(r) M" },
5415 { 0x17, "Mobile Intel(r) Celeron(r)" }
5416 };
5417 uint_t btblmax = sizeof (brand_tbl) / sizeof (brand_tbl[0]);
5418 uint_t sgn;
5419
5420 sgn = (cpi->cpi_family << 8) |
5421 (cpi->cpi_model << 4) | cpi->cpi_step;
5422
5423 for (i = 0; i < btblmax; i++)
5424 if (brand_tbl[i].bt_bid == cpi->cpi_brandid)
5425 break;
5426 if (i < btblmax) {
5427 if (sgn == 0x6b1 && cpi->cpi_brandid == 3)
5428 return ("Intel(r) Celeron(r)");
5429 if (sgn < 0xf13 && cpi->cpi_brandid == 0xb)
5430 return ("Intel(r) Xeon(tm) MP");
5431 if (sgn < 0xf13 && cpi->cpi_brandid == 0xe)
5432 return ("Intel(r) Xeon(tm)");
5433 return (brand_tbl[i].bt_str);
5434 }
5435 }
5436
5437 return (NULL);
5438 }
5439
5440 static const char *
amd_cpubrand(const struct cpuid_info * cpi)5441 amd_cpubrand(const struct cpuid_info *cpi)
5442 {
5443 ASSERT(is_x86_feature(x86_featureset, X86FSET_CPUID));
5444
5445 switch (cpi->cpi_family) {
5446 case 5:
5447 switch (cpi->cpi_model) {
5448 case 0:
5449 case 1:
5450 case 2:
5451 case 3:
5452 case 4:
5453 case 5:
5454 return ("AMD-K5(r)");
5455 case 6:
5456 case 7:
5457 return ("AMD-K6(r)");
5458 case 8:
5459 return ("AMD-K6(r)-2");
5460 case 9:
5461 return ("AMD-K6(r)-III");
5462 default:
5463 return ("AMD (family 5)");
5464 }
5465 case 6:
5466 switch (cpi->cpi_model) {
5467 case 1:
5468 return ("AMD-K7(tm)");
5469 case 0:
5470 case 2:
5471 case 4:
5472 return ("AMD Athlon(tm)");
5473 case 3:
5474 case 7:
5475 return ("AMD Duron(tm)");
5476 case 6:
5477 case 8:
5478 case 10:
5479 /*
5480 * Use the L2 cache size to distinguish
5481 */
5482 return ((cpi->cpi_extd[6].cp_ecx >> 16) >= 256 ?
5483 "AMD Athlon(tm)" : "AMD Duron(tm)");
5484 default:
5485 return ("AMD (family 6)");
5486 }
5487 default:
5488 break;
5489 }
5490
5491 if (cpi->cpi_family == 0xf && cpi->cpi_model == 5 &&
5492 cpi->cpi_brandid != 0) {
5493 switch (BITX(cpi->cpi_brandid, 7, 5)) {
5494 case 3:
5495 return ("AMD Opteron(tm) UP 1xx");
5496 case 4:
5497 return ("AMD Opteron(tm) DP 2xx");
5498 case 5:
5499 return ("AMD Opteron(tm) MP 8xx");
5500 default:
5501 return ("AMD Opteron(tm)");
5502 }
5503 }
5504
5505 return (NULL);
5506 }
5507
5508 static const char *
cyrix_cpubrand(struct cpuid_info * cpi,uint_t type)5509 cyrix_cpubrand(struct cpuid_info *cpi, uint_t type)
5510 {
5511 ASSERT(is_x86_feature(x86_featureset, X86FSET_CPUID));
5512
5513 switch (type) {
5514 case X86_TYPE_CYRIX_6x86:
5515 return ("Cyrix 6x86");
5516 case X86_TYPE_CYRIX_6x86L:
5517 return ("Cyrix 6x86L");
5518 case X86_TYPE_CYRIX_6x86MX:
5519 return ("Cyrix 6x86MX");
5520 case X86_TYPE_CYRIX_GXm:
5521 return ("Cyrix GXm");
5522 case X86_TYPE_CYRIX_MediaGX:
5523 return ("Cyrix MediaGX");
5524 case X86_TYPE_CYRIX_MII:
5525 return ("Cyrix M2");
5526 case X86_TYPE_VIA_CYRIX_III:
5527 return ("VIA Cyrix M3");
5528 default:
5529 /*
5530 * Have another wild guess ..
5531 */
5532 if (cpi->cpi_family == 4 && cpi->cpi_model == 9)
5533 return ("Cyrix 5x86");
5534 else if (cpi->cpi_family == 5) {
5535 switch (cpi->cpi_model) {
5536 case 2:
5537 return ("Cyrix 6x86"); /* Cyrix M1 */
5538 case 4:
5539 return ("Cyrix MediaGX");
5540 default:
5541 break;
5542 }
5543 } else if (cpi->cpi_family == 6) {
5544 switch (cpi->cpi_model) {
5545 case 0:
5546 return ("Cyrix 6x86MX"); /* Cyrix M2? */
5547 case 5:
5548 case 6:
5549 case 7:
5550 case 8:
5551 case 9:
5552 return ("VIA C3");
5553 default:
5554 break;
5555 }
5556 }
5557 break;
5558 }
5559 return (NULL);
5560 }
5561
5562 /*
5563 * This only gets called in the case that the CPU extended
5564 * feature brand string (0x80000002, 0x80000003, 0x80000004)
5565 * aren't available, or contain null bytes for some reason.
5566 */
5567 static void
fabricate_brandstr(struct cpuid_info * cpi)5568 fabricate_brandstr(struct cpuid_info *cpi)
5569 {
5570 const char *brand = NULL;
5571
5572 switch (cpi->cpi_vendor) {
5573 case X86_VENDOR_Intel:
5574 brand = intel_cpubrand(cpi);
5575 break;
5576 case X86_VENDOR_AMD:
5577 brand = amd_cpubrand(cpi);
5578 break;
5579 case X86_VENDOR_Cyrix:
5580 brand = cyrix_cpubrand(cpi, x86_type);
5581 break;
5582 case X86_VENDOR_NexGen:
5583 if (cpi->cpi_family == 5 && cpi->cpi_model == 0)
5584 brand = "NexGen Nx586";
5585 break;
5586 case X86_VENDOR_Centaur:
5587 if (cpi->cpi_family == 5)
5588 switch (cpi->cpi_model) {
5589 case 4:
5590 brand = "Centaur C6";
5591 break;
5592 case 8:
5593 brand = "Centaur C2";
5594 break;
5595 case 9:
5596 brand = "Centaur C3";
5597 break;
5598 default:
5599 break;
5600 }
5601 break;
5602 case X86_VENDOR_Rise:
5603 if (cpi->cpi_family == 5 &&
5604 (cpi->cpi_model == 0 || cpi->cpi_model == 2))
5605 brand = "Rise mP6";
5606 break;
5607 case X86_VENDOR_SiS:
5608 if (cpi->cpi_family == 5 && cpi->cpi_model == 0)
5609 brand = "SiS 55x";
5610 break;
5611 case X86_VENDOR_TM:
5612 if (cpi->cpi_family == 5 && cpi->cpi_model == 4)
5613 brand = "Transmeta Crusoe TM3x00 or TM5x00";
5614 break;
5615 case X86_VENDOR_NSC:
5616 case X86_VENDOR_UMC:
5617 default:
5618 break;
5619 }
5620 if (brand) {
5621 (void) strcpy((char *)cpi->cpi_brandstr, brand);
5622 return;
5623 }
5624
5625 /*
5626 * If all else fails ...
5627 */
5628 (void) snprintf(cpi->cpi_brandstr, sizeof (cpi->cpi_brandstr),
5629 "%s %d.%d.%d", cpi->cpi_vendorstr, cpi->cpi_family,
5630 cpi->cpi_model, cpi->cpi_step);
5631 }
5632
5633 /*
5634 * This routine is called just after kernel memory allocation
5635 * becomes available on cpu0, and as part of mp_startup() on
5636 * the other cpus.
5637 *
5638 * Fixup the brand string, and collect any information from cpuid
5639 * that requires dynamically allocated storage to represent.
5640 */
5641
5642 static void
cpuid_pass_dynamic(cpu_t * cpu,void * _arg __unused)5643 cpuid_pass_dynamic(cpu_t *cpu, void *_arg __unused)
5644 {
5645 int i, max, shft, level, size;
5646 struct cpuid_regs regs;
5647 struct cpuid_regs *cp;
5648 struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
5649
5650 /*
5651 * Deterministic cache parameters
5652 *
5653 * Intel uses leaf 0x4 for this, while AMD uses leaf 0x8000001d. The
5654 * values that are present are currently defined to be the same. This
5655 * means we can use the same logic to parse it as long as we use the
5656 * appropriate leaf to get the data. If you're updating this, make sure
5657 * you're careful about which vendor supports which aspect.
5658 *
5659 * Take this opportunity to detect the number of threads sharing the
5660 * last level cache, and construct a corresponding cache id. The
5661 * respective cpuid_info members are initialized to the default case of
5662 * "no last level cache sharing".
5663 */
5664 cpi->cpi_ncpu_shr_last_cache = 1;
5665 cpi->cpi_last_lvl_cacheid = cpu->cpu_id;
5666
5667 if ((cpi->cpi_maxeax >= 4 && cpi->cpi_vendor == X86_VENDOR_Intel) ||
5668 ((cpi->cpi_vendor == X86_VENDOR_AMD ||
5669 cpi->cpi_vendor == X86_VENDOR_HYGON) &&
5670 cpi->cpi_xmaxeax >= CPUID_LEAF_EXT_1d &&
5671 is_x86_feature(x86_featureset, X86FSET_TOPOEXT))) {
5672 uint32_t leaf;
5673
5674 if (cpi->cpi_vendor == X86_VENDOR_Intel) {
5675 leaf = 4;
5676 } else {
5677 leaf = CPUID_LEAF_EXT_1d;
5678 }
5679
5680 /*
5681 * Find the # of elements (size) returned by the leaf and along
5682 * the way detect last level cache sharing details.
5683 */
5684 bzero(®s, sizeof (regs));
5685 cp = ®s;
5686 for (i = 0, max = 0; i < CPI_FN4_ECX_MAX; i++) {
5687 cp->cp_eax = leaf;
5688 cp->cp_ecx = i;
5689
5690 (void) __cpuid_insn(cp);
5691
5692 if (CPI_CACHE_TYPE(cp) == 0)
5693 break;
5694 level = CPI_CACHE_LVL(cp);
5695 if (level > max) {
5696 max = level;
5697 cpi->cpi_ncpu_shr_last_cache =
5698 CPI_NTHR_SHR_CACHE(cp) + 1;
5699 }
5700 }
5701 cpi->cpi_cache_leaf_size = size = i;
5702
5703 /*
5704 * Allocate the cpi_cache_leaves array. The first element
5705 * references the regs for the corresponding leaf with %ecx set
5706 * to 0. This was gathered in cpuid_pass_extended().
5707 */
5708 if (size > 0) {
5709 cpi->cpi_cache_leaves =
5710 kmem_alloc(size * sizeof (cp), KM_SLEEP);
5711 if (cpi->cpi_vendor == X86_VENDOR_Intel) {
5712 cpi->cpi_cache_leaves[0] = &cpi->cpi_std[4];
5713 } else {
5714 cpi->cpi_cache_leaves[0] = &cpi->cpi_extd[0x1d];
5715 }
5716
5717 /*
5718 * Allocate storage to hold the additional regs
5719 * for the leaf, %ecx == 1 .. cpi_cache_leaf_size.
5720 *
5721 * The regs for the leaf, %ecx == 0 has already
5722 * been allocated as indicated above.
5723 */
5724 for (i = 1; i < size; i++) {
5725 cp = cpi->cpi_cache_leaves[i] =
5726 kmem_zalloc(sizeof (regs), KM_SLEEP);
5727 cp->cp_eax = leaf;
5728 cp->cp_ecx = i;
5729
5730 (void) __cpuid_insn(cp);
5731 }
5732 }
5733 /*
5734 * Determine the number of bits needed to represent
5735 * the number of CPUs sharing the last level cache.
5736 *
5737 * Shift off that number of bits from the APIC id to
5738 * derive the cache id.
5739 */
5740 shft = 0;
5741 for (i = 1; i < cpi->cpi_ncpu_shr_last_cache; i <<= 1)
5742 shft++;
5743 cpi->cpi_last_lvl_cacheid = cpi->cpi_apicid >> shft;
5744 }
5745
5746 /*
5747 * Now fixup the brand string
5748 */
5749 if ((cpi->cpi_xmaxeax & CPUID_LEAF_EXT_0) == 0) {
5750 fabricate_brandstr(cpi);
5751 } else {
5752
5753 /*
5754 * If we successfully extracted a brand string from the cpuid
5755 * instruction, clean it up by removing leading spaces and
5756 * similar junk.
5757 */
5758 if (cpi->cpi_brandstr[0]) {
5759 size_t maxlen = sizeof (cpi->cpi_brandstr);
5760 char *src, *dst;
5761
5762 dst = src = (char *)cpi->cpi_brandstr;
5763 src[maxlen - 1] = '\0';
5764 /*
5765 * strip leading spaces
5766 */
5767 while (*src == ' ')
5768 src++;
5769 /*
5770 * Remove any 'Genuine' or "Authentic" prefixes
5771 */
5772 if (strncmp(src, "Genuine ", 8) == 0)
5773 src += 8;
5774 if (strncmp(src, "Authentic ", 10) == 0)
5775 src += 10;
5776
5777 /*
5778 * Now do an in-place copy.
5779 * Map (R) to (r) and (TM) to (tm).
5780 * The era of teletypes is long gone, and there's
5781 * -really- no need to shout.
5782 */
5783 while (*src != '\0') {
5784 if (src[0] == '(') {
5785 if (strncmp(src + 1, "R)", 2) == 0) {
5786 (void) strncpy(dst, "(r)", 3);
5787 src += 3;
5788 dst += 3;
5789 continue;
5790 }
5791 if (strncmp(src + 1, "TM)", 3) == 0) {
5792 (void) strncpy(dst, "(tm)", 4);
5793 src += 4;
5794 dst += 4;
5795 continue;
5796 }
5797 }
5798 *dst++ = *src++;
5799 }
5800 *dst = '\0';
5801
5802 /*
5803 * Finally, remove any trailing spaces
5804 */
5805 while (--dst > cpi->cpi_brandstr)
5806 if (*dst == ' ')
5807 *dst = '\0';
5808 else
5809 break;
5810 } else
5811 fabricate_brandstr(cpi);
5812 }
5813 }
5814
5815 typedef struct {
5816 uint32_t avm_av;
5817 uint32_t avm_feat;
5818 } av_feat_map_t;
5819
5820 /*
5821 * These arrays are used to map features that we should add based on x86
5822 * features that are present. As a large number depend on kernel features,
5823 * rather than rechecking and clearing CPUID everywhere, we simply map these.
5824 * There is an array of these for each hwcap word. Some features aren't tracked
5825 * in the kernel x86 featureset and that's ok. They will not show up in here.
5826 */
5827 static const av_feat_map_t x86fset_to_av1[] = {
5828 { AV_386_CX8, X86FSET_CX8 },
5829 { AV_386_SEP, X86FSET_SEP },
5830 { AV_386_AMD_SYSC, X86FSET_ASYSC },
5831 { AV_386_CMOV, X86FSET_CMOV },
5832 { AV_386_FXSR, X86FSET_SSE },
5833 { AV_386_SSE, X86FSET_SSE },
5834 { AV_386_SSE2, X86FSET_SSE2 },
5835 { AV_386_SSE3, X86FSET_SSE3 },
5836 { AV_386_CX16, X86FSET_CX16 },
5837 { AV_386_TSCP, X86FSET_TSCP },
5838 { AV_386_AMD_SSE4A, X86FSET_SSE4A },
5839 { AV_386_SSSE3, X86FSET_SSSE3 },
5840 { AV_386_SSE4_1, X86FSET_SSE4_1 },
5841 { AV_386_SSE4_2, X86FSET_SSE4_2 },
5842 { AV_386_AES, X86FSET_AES },
5843 { AV_386_PCLMULQDQ, X86FSET_PCLMULQDQ },
5844 { AV_386_XSAVE, X86FSET_XSAVE },
5845 { AV_386_AVX, X86FSET_AVX },
5846 { AV_386_VMX, X86FSET_VMX },
5847 { AV_386_AMD_SVM, X86FSET_SVM }
5848 };
5849
5850 static const av_feat_map_t x86fset_to_av2[] = {
5851 { AV_386_2_F16C, X86FSET_F16C },
5852 { AV_386_2_RDRAND, X86FSET_RDRAND },
5853 { AV_386_2_BMI1, X86FSET_BMI1 },
5854 { AV_386_2_BMI2, X86FSET_BMI2 },
5855 { AV_386_2_FMA, X86FSET_FMA },
5856 { AV_386_2_AVX2, X86FSET_AVX2 },
5857 { AV_386_2_ADX, X86FSET_ADX },
5858 { AV_386_2_RDSEED, X86FSET_RDSEED },
5859 { AV_386_2_AVX512F, X86FSET_AVX512F },
5860 { AV_386_2_AVX512DQ, X86FSET_AVX512DQ },
5861 { AV_386_2_AVX512IFMA, X86FSET_AVX512FMA },
5862 { AV_386_2_AVX512PF, X86FSET_AVX512PF },
5863 { AV_386_2_AVX512ER, X86FSET_AVX512ER },
5864 { AV_386_2_AVX512CD, X86FSET_AVX512CD },
5865 { AV_386_2_AVX512BW, X86FSET_AVX512BW },
5866 { AV_386_2_AVX512VL, X86FSET_AVX512VL },
5867 { AV_386_2_AVX512VBMI, X86FSET_AVX512VBMI },
5868 { AV_386_2_AVX512VPOPCDQ, X86FSET_AVX512VPOPCDQ },
5869 { AV_386_2_SHA, X86FSET_SHA },
5870 { AV_386_2_FSGSBASE, X86FSET_FSGSBASE },
5871 { AV_386_2_CLFLUSHOPT, X86FSET_CLFLUSHOPT },
5872 { AV_386_2_CLWB, X86FSET_CLWB },
5873 { AV_386_2_MONITORX, X86FSET_MONITORX },
5874 { AV_386_2_CLZERO, X86FSET_CLZERO },
5875 { AV_386_2_AVX512_VNNI, X86FSET_AVX512VNNI },
5876 { AV_386_2_VPCLMULQDQ, X86FSET_VPCLMULQDQ },
5877 { AV_386_2_VAES, X86FSET_VAES },
5878 { AV_386_2_GFNI, X86FSET_GFNI },
5879 { AV_386_2_AVX512_VP2INT, X86FSET_AVX512_VP2INT },
5880 { AV_386_2_AVX512_BITALG, X86FSET_AVX512_BITALG }
5881 };
5882
5883 static const av_feat_map_t x86fset_to_av3[] = {
5884 { AV_386_3_AVX512_VBMI2, X86FSET_AVX512_VBMI2 },
5885 { AV_386_3_AVX512_BF16, X86FSET_AVX512_BF16 }
5886 };
5887
5888 /*
5889 * This routine is called out of bind_hwcap() much later in the life
5890 * of the kernel (post_startup()). The job of this routine is to resolve
5891 * the hardware feature support and kernel support for those features into
5892 * what we're actually going to tell applications via the aux vector.
5893 *
5894 * Most of the aux vector is derived from the x86_featureset array vector where
5895 * a given feature indicates that an aux vector should be plumbed through. This
5896 * allows the kernel to use one tracking mechanism for these based on whether or
5897 * not it has the required hardware support (most often xsave). Most newer
5898 * features are added there in case we need them in the kernel. Otherwise,
5899 * features are evaluated based on looking at the cpuid features that remain. If
5900 * you find yourself wanting to clear out cpuid features for some reason, they
5901 * should instead be driven by the feature set so we have a consistent view.
5902 */
5903
5904 static void
cpuid_pass_resolve(cpu_t * cpu,void * arg)5905 cpuid_pass_resolve(cpu_t *cpu, void *arg)
5906 {
5907 uint_t *hwcap_out = (uint_t *)arg;
5908 struct cpuid_info *cpi;
5909 uint_t hwcap_flags = 0, hwcap_flags_2 = 0, hwcap_flags_3 = 0;
5910
5911 cpi = cpu->cpu_m.mcpu_cpi;
5912
5913 for (uint_t i = 0; i < ARRAY_SIZE(x86fset_to_av1); i++) {
5914 if (is_x86_feature(x86_featureset,
5915 x86fset_to_av1[i].avm_feat)) {
5916 hwcap_flags |= x86fset_to_av1[i].avm_av;
5917 }
5918 }
5919
5920 for (uint_t i = 0; i < ARRAY_SIZE(x86fset_to_av2); i++) {
5921 if (is_x86_feature(x86_featureset,
5922 x86fset_to_av2[i].avm_feat)) {
5923 hwcap_flags_2 |= x86fset_to_av2[i].avm_av;
5924 }
5925 }
5926
5927 for (uint_t i = 0; i < ARRAY_SIZE(x86fset_to_av3); i++) {
5928 if (is_x86_feature(x86_featureset,
5929 x86fset_to_av3[i].avm_feat)) {
5930 hwcap_flags_3 |= x86fset_to_av3[i].avm_av;
5931 }
5932 }
5933
5934 /*
5935 * From here on out we're working through features that don't have
5936 * corresponding kernel feature flags for various reasons that are
5937 * mostly just due to the historical implementation.
5938 */
5939 if (cpi->cpi_maxeax >= 1) {
5940 uint32_t *edx = &cpi->cpi_support[STD_EDX_FEATURES];
5941 uint32_t *ecx = &cpi->cpi_support[STD_ECX_FEATURES];
5942
5943 *edx = CPI_FEATURES_EDX(cpi);
5944 *ecx = CPI_FEATURES_ECX(cpi);
5945
5946 /*
5947 * [no explicit support required beyond x87 fp context]
5948 */
5949 if (!fpu_exists)
5950 *edx &= ~(CPUID_INTC_EDX_FPU | CPUID_INTC_EDX_MMX);
5951
5952 /*
5953 * Now map the supported feature vector to things that we
5954 * think userland will care about.
5955 */
5956 if (*ecx & CPUID_INTC_ECX_MOVBE)
5957 hwcap_flags |= AV_386_MOVBE;
5958
5959 if (*ecx & CPUID_INTC_ECX_POPCNT)
5960 hwcap_flags |= AV_386_POPCNT;
5961 if (*edx & CPUID_INTC_EDX_FPU)
5962 hwcap_flags |= AV_386_FPU;
5963 if (*edx & CPUID_INTC_EDX_MMX)
5964 hwcap_flags |= AV_386_MMX;
5965 if (*edx & CPUID_INTC_EDX_TSC)
5966 hwcap_flags |= AV_386_TSC;
5967 }
5968
5969 /*
5970 * Check a few miscellaneous features.
5971 */
5972 if (cpi->cpi_xmaxeax < 0x80000001)
5973 goto resolve_done;
5974
5975 switch (cpi->cpi_vendor) {
5976 uint32_t *edx, *ecx;
5977
5978 case X86_VENDOR_Intel:
5979 /*
5980 * Seems like Intel duplicated what we necessary
5981 * here to make the initial crop of 64-bit OS's work.
5982 * Hopefully, those are the only "extended" bits
5983 * they'll add.
5984 */
5985 /*FALLTHROUGH*/
5986
5987 case X86_VENDOR_AMD:
5988 case X86_VENDOR_HYGON:
5989 edx = &cpi->cpi_support[AMD_EDX_FEATURES];
5990 ecx = &cpi->cpi_support[AMD_ECX_FEATURES];
5991
5992 *edx = CPI_FEATURES_XTD_EDX(cpi);
5993 *ecx = CPI_FEATURES_XTD_ECX(cpi);
5994
5995 /*
5996 * [no explicit support required beyond
5997 * x87 fp context and exception handlers]
5998 */
5999 if (!fpu_exists)
6000 *edx &= ~(CPUID_AMD_EDX_MMXamd |
6001 CPUID_AMD_EDX_3DNow | CPUID_AMD_EDX_3DNowx);
6002
6003 /*
6004 * Now map the supported feature vector to
6005 * things that we think userland will care about.
6006 */
6007 if (*edx & CPUID_AMD_EDX_MMXamd)
6008 hwcap_flags |= AV_386_AMD_MMX;
6009 if (*edx & CPUID_AMD_EDX_3DNow)
6010 hwcap_flags |= AV_386_AMD_3DNow;
6011 if (*edx & CPUID_AMD_EDX_3DNowx)
6012 hwcap_flags |= AV_386_AMD_3DNowx;
6013
6014 switch (cpi->cpi_vendor) {
6015 case X86_VENDOR_AMD:
6016 case X86_VENDOR_HYGON:
6017 if (*ecx & CPUID_AMD_ECX_AHF64)
6018 hwcap_flags |= AV_386_AHF;
6019 if (*ecx & CPUID_AMD_ECX_LZCNT)
6020 hwcap_flags |= AV_386_AMD_LZCNT;
6021 break;
6022
6023 case X86_VENDOR_Intel:
6024 if (*ecx & CPUID_AMD_ECX_LZCNT)
6025 hwcap_flags |= AV_386_AMD_LZCNT;
6026 /*
6027 * Aarrgh.
6028 * Intel uses a different bit in the same word.
6029 */
6030 if (*ecx & CPUID_INTC_ECX_AHF64)
6031 hwcap_flags |= AV_386_AHF;
6032 break;
6033 default:
6034 break;
6035 }
6036 break;
6037
6038 default:
6039 break;
6040 }
6041
6042 resolve_done:
6043 if (hwcap_out != NULL) {
6044 hwcap_out[0] = hwcap_flags;
6045 hwcap_out[1] = hwcap_flags_2;
6046 hwcap_out[2] = hwcap_flags_3;
6047 }
6048 }
6049
6050
6051 /*
6052 * Simulate the cpuid instruction using the data we previously
6053 * captured about this CPU. We try our best to return the truth
6054 * about the hardware, independently of kernel support.
6055 */
6056 uint32_t
cpuid_insn(cpu_t * cpu,struct cpuid_regs * cp)6057 cpuid_insn(cpu_t *cpu, struct cpuid_regs *cp)
6058 {
6059 struct cpuid_info *cpi;
6060 struct cpuid_regs *xcp;
6061
6062 if (cpu == NULL)
6063 cpu = CPU;
6064 cpi = cpu->cpu_m.mcpu_cpi;
6065
6066 ASSERT(cpuid_checkpass(cpu, CPUID_PASS_DYNAMIC));
6067
6068 /*
6069 * CPUID data is cached in two separate places: cpi_std for standard
6070 * CPUID leaves , and cpi_extd for extended CPUID leaves.
6071 */
6072 if (cp->cp_eax <= cpi->cpi_maxeax && cp->cp_eax < NMAX_CPI_STD) {
6073 xcp = &cpi->cpi_std[cp->cp_eax];
6074 } else if (cp->cp_eax >= CPUID_LEAF_EXT_0 &&
6075 cp->cp_eax <= cpi->cpi_xmaxeax &&
6076 cp->cp_eax < CPUID_LEAF_EXT_0 + NMAX_CPI_EXTD) {
6077 xcp = &cpi->cpi_extd[cp->cp_eax - CPUID_LEAF_EXT_0];
6078 } else {
6079 /*
6080 * The caller is asking for data from an input parameter which
6081 * the kernel has not cached. In this case we go fetch from
6082 * the hardware and return the data directly to the user.
6083 */
6084 return (__cpuid_insn(cp));
6085 }
6086
6087 cp->cp_eax = xcp->cp_eax;
6088 cp->cp_ebx = xcp->cp_ebx;
6089 cp->cp_ecx = xcp->cp_ecx;
6090 cp->cp_edx = xcp->cp_edx;
6091 return (cp->cp_eax);
6092 }
6093
6094 boolean_t
cpuid_checkpass(const cpu_t * const cpu,const cpuid_pass_t pass)6095 cpuid_checkpass(const cpu_t *const cpu, const cpuid_pass_t pass)
6096 {
6097 return (cpu != NULL && cpu->cpu_m.mcpu_cpi != NULL &&
6098 cpu->cpu_m.mcpu_cpi->cpi_pass >= pass);
6099 }
6100
6101 int
cpuid_getbrandstr(cpu_t * cpu,char * s,size_t n)6102 cpuid_getbrandstr(cpu_t *cpu, char *s, size_t n)
6103 {
6104 ASSERT(cpuid_checkpass(cpu, CPUID_PASS_DYNAMIC));
6105
6106 return (snprintf(s, n, "%s", cpu->cpu_m.mcpu_cpi->cpi_brandstr));
6107 }
6108
6109 int
cpuid_is_cmt(cpu_t * cpu)6110 cpuid_is_cmt(cpu_t *cpu)
6111 {
6112 if (cpu == NULL)
6113 cpu = CPU;
6114
6115 ASSERT(cpuid_checkpass(cpu, CPUID_PASS_BASIC));
6116
6117 return (cpu->cpu_m.mcpu_cpi->cpi_chipid >= 0);
6118 }
6119
6120 /*
6121 * AMD and Intel both implement the 64-bit variant of the syscall
6122 * instruction (syscallq), so if there's -any- support for syscall,
6123 * cpuid currently says "yes, we support this".
6124 *
6125 * However, Intel decided to -not- implement the 32-bit variant of the
6126 * syscall instruction, so we provide a predicate to allow our caller
6127 * to test that subtlety here.
6128 *
6129 * XXPV Currently, 32-bit syscall instructions don't work via the hypervisor,
6130 * even in the case where the hardware would in fact support it.
6131 */
6132 /*ARGSUSED*/
6133 int
cpuid_syscall32_insn(cpu_t * cpu)6134 cpuid_syscall32_insn(cpu_t *cpu)
6135 {
6136 ASSERT(cpuid_checkpass((cpu == NULL ? CPU : cpu), CPUID_PASS_BASIC));
6137
6138 #if !defined(__xpv)
6139 if (cpu == NULL)
6140 cpu = CPU;
6141
6142 /*CSTYLED*/
6143 {
6144 struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
6145
6146 if ((cpi->cpi_vendor == X86_VENDOR_AMD ||
6147 cpi->cpi_vendor == X86_VENDOR_HYGON) &&
6148 cpi->cpi_xmaxeax >= 0x80000001 &&
6149 (CPI_FEATURES_XTD_EDX(cpi) & CPUID_AMD_EDX_SYSC))
6150 return (1);
6151 }
6152 #endif
6153 return (0);
6154 }
6155
6156 int
cpuid_getidstr(cpu_t * cpu,char * s,size_t n)6157 cpuid_getidstr(cpu_t *cpu, char *s, size_t n)
6158 {
6159 struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
6160
6161 static const char fmt[] =
6162 "x86 (%s %X family %d model %d step %d clock %d MHz)";
6163 static const char fmt_ht[] =
6164 "x86 (chipid 0x%x %s %X family %d model %d step %d clock %d MHz)";
6165
6166 ASSERT(cpuid_checkpass(cpu, CPUID_PASS_BASIC));
6167
6168 if (cpuid_is_cmt(cpu))
6169 return (snprintf(s, n, fmt_ht, cpi->cpi_chipid,
6170 cpi->cpi_vendorstr, cpi->cpi_std[1].cp_eax,
6171 cpi->cpi_family, cpi->cpi_model,
6172 cpi->cpi_step, cpu->cpu_type_info.pi_clock));
6173 return (snprintf(s, n, fmt,
6174 cpi->cpi_vendorstr, cpi->cpi_std[1].cp_eax,
6175 cpi->cpi_family, cpi->cpi_model,
6176 cpi->cpi_step, cpu->cpu_type_info.pi_clock));
6177 }
6178
6179 const char *
cpuid_getvendorstr(cpu_t * cpu)6180 cpuid_getvendorstr(cpu_t *cpu)
6181 {
6182 ASSERT(cpuid_checkpass(cpu, CPUID_PASS_IDENT));
6183 return ((const char *)cpu->cpu_m.mcpu_cpi->cpi_vendorstr);
6184 }
6185
6186 uint_t
cpuid_getvendor(cpu_t * cpu)6187 cpuid_getvendor(cpu_t *cpu)
6188 {
6189 ASSERT(cpuid_checkpass(cpu, CPUID_PASS_IDENT));
6190 return (cpu->cpu_m.mcpu_cpi->cpi_vendor);
6191 }
6192
6193 uint_t
cpuid_getfamily(cpu_t * cpu)6194 cpuid_getfamily(cpu_t *cpu)
6195 {
6196 ASSERT(cpuid_checkpass(cpu, CPUID_PASS_IDENT));
6197 return (cpu->cpu_m.mcpu_cpi->cpi_family);
6198 }
6199
6200 uint_t
cpuid_getmodel(cpu_t * cpu)6201 cpuid_getmodel(cpu_t *cpu)
6202 {
6203 ASSERT(cpuid_checkpass(cpu, CPUID_PASS_IDENT));
6204 return (cpu->cpu_m.mcpu_cpi->cpi_model);
6205 }
6206
6207 uint_t
cpuid_get_ncpu_per_chip(cpu_t * cpu)6208 cpuid_get_ncpu_per_chip(cpu_t *cpu)
6209 {
6210 ASSERT(cpuid_checkpass(cpu, CPUID_PASS_BASIC));
6211 return (cpu->cpu_m.mcpu_cpi->cpi_ncpu_per_chip);
6212 }
6213
6214 uint_t
cpuid_get_ncore_per_chip(cpu_t * cpu)6215 cpuid_get_ncore_per_chip(cpu_t *cpu)
6216 {
6217 ASSERT(cpuid_checkpass(cpu, CPUID_PASS_BASIC));
6218 return (cpu->cpu_m.mcpu_cpi->cpi_ncore_per_chip);
6219 }
6220
6221 uint_t
cpuid_get_ncpu_sharing_last_cache(cpu_t * cpu)6222 cpuid_get_ncpu_sharing_last_cache(cpu_t *cpu)
6223 {
6224 ASSERT(cpuid_checkpass(cpu, CPUID_PASS_EXTENDED));
6225 return (cpu->cpu_m.mcpu_cpi->cpi_ncpu_shr_last_cache);
6226 }
6227
6228 id_t
cpuid_get_last_lvl_cacheid(cpu_t * cpu)6229 cpuid_get_last_lvl_cacheid(cpu_t *cpu)
6230 {
6231 ASSERT(cpuid_checkpass(cpu, CPUID_PASS_EXTENDED));
6232 return (cpu->cpu_m.mcpu_cpi->cpi_last_lvl_cacheid);
6233 }
6234
6235 uint_t
cpuid_getstep(cpu_t * cpu)6236 cpuid_getstep(cpu_t *cpu)
6237 {
6238 ASSERT(cpuid_checkpass(cpu, CPUID_PASS_IDENT));
6239 return (cpu->cpu_m.mcpu_cpi->cpi_step);
6240 }
6241
6242 uint_t
cpuid_getsig(struct cpu * cpu)6243 cpuid_getsig(struct cpu *cpu)
6244 {
6245 ASSERT(cpuid_checkpass(cpu, CPUID_PASS_BASIC));
6246 return (cpu->cpu_m.mcpu_cpi->cpi_std[1].cp_eax);
6247 }
6248
6249 uint32_t
cpuid_getchiprev(struct cpu * cpu)6250 cpuid_getchiprev(struct cpu *cpu)
6251 {
6252 ASSERT(cpuid_checkpass(cpu, CPUID_PASS_IDENT));
6253 return (cpu->cpu_m.mcpu_cpi->cpi_chiprev);
6254 }
6255
6256 const char *
cpuid_getchiprevstr(struct cpu * cpu)6257 cpuid_getchiprevstr(struct cpu *cpu)
6258 {
6259 ASSERT(cpuid_checkpass(cpu, CPUID_PASS_IDENT));
6260 return (cpu->cpu_m.mcpu_cpi->cpi_chiprevstr);
6261 }
6262
6263 uint32_t
cpuid_getsockettype(struct cpu * cpu)6264 cpuid_getsockettype(struct cpu *cpu)
6265 {
6266 ASSERT(cpuid_checkpass(cpu, CPUID_PASS_IDENT));
6267 return (cpu->cpu_m.mcpu_cpi->cpi_socket);
6268 }
6269
6270 const char *
cpuid_getsocketstr(cpu_t * cpu)6271 cpuid_getsocketstr(cpu_t *cpu)
6272 {
6273 static const char *socketstr = NULL;
6274 struct cpuid_info *cpi;
6275
6276 ASSERT(cpuid_checkpass(cpu, CPUID_PASS_IDENT));
6277 cpi = cpu->cpu_m.mcpu_cpi;
6278
6279 /* Assume that socket types are the same across the system */
6280 if (socketstr == NULL)
6281 socketstr = _cpuid_sktstr(cpi->cpi_vendor, cpi->cpi_family,
6282 cpi->cpi_model, cpi->cpi_step);
6283
6284
6285 return (socketstr);
6286 }
6287
6288 x86_uarchrev_t
cpuid_getuarchrev(cpu_t * cpu)6289 cpuid_getuarchrev(cpu_t *cpu)
6290 {
6291 return (cpu->cpu_m.mcpu_cpi->cpi_uarchrev);
6292 }
6293
6294 int
cpuid_get_chipid(cpu_t * cpu)6295 cpuid_get_chipid(cpu_t *cpu)
6296 {
6297 ASSERT(cpuid_checkpass(cpu, CPUID_PASS_BASIC));
6298
6299 if (cpuid_is_cmt(cpu))
6300 return (cpu->cpu_m.mcpu_cpi->cpi_chipid);
6301 return (cpu->cpu_id);
6302 }
6303
6304 id_t
cpuid_get_coreid(cpu_t * cpu)6305 cpuid_get_coreid(cpu_t *cpu)
6306 {
6307 ASSERT(cpuid_checkpass(cpu, CPUID_PASS_BASIC));
6308 return (cpu->cpu_m.mcpu_cpi->cpi_coreid);
6309 }
6310
6311 int
cpuid_get_pkgcoreid(cpu_t * cpu)6312 cpuid_get_pkgcoreid(cpu_t *cpu)
6313 {
6314 ASSERT(cpuid_checkpass(cpu, CPUID_PASS_BASIC));
6315 return (cpu->cpu_m.mcpu_cpi->cpi_pkgcoreid);
6316 }
6317
6318 int
cpuid_get_clogid(cpu_t * cpu)6319 cpuid_get_clogid(cpu_t *cpu)
6320 {
6321 ASSERT(cpuid_checkpass(cpu, CPUID_PASS_BASIC));
6322 return (cpu->cpu_m.mcpu_cpi->cpi_clogid);
6323 }
6324
6325 int
cpuid_get_cacheid(cpu_t * cpu)6326 cpuid_get_cacheid(cpu_t *cpu)
6327 {
6328 ASSERT(cpuid_checkpass(cpu, CPUID_PASS_BASIC));
6329 return (cpu->cpu_m.mcpu_cpi->cpi_last_lvl_cacheid);
6330 }
6331
6332 uint_t
cpuid_get_procnodeid(cpu_t * cpu)6333 cpuid_get_procnodeid(cpu_t *cpu)
6334 {
6335 ASSERT(cpuid_checkpass(cpu, CPUID_PASS_BASIC));
6336 return (cpu->cpu_m.mcpu_cpi->cpi_procnodeid);
6337 }
6338
6339 uint_t
cpuid_get_procnodes_per_pkg(cpu_t * cpu)6340 cpuid_get_procnodes_per_pkg(cpu_t *cpu)
6341 {
6342 ASSERT(cpuid_checkpass(cpu, CPUID_PASS_BASIC));
6343 return (cpu->cpu_m.mcpu_cpi->cpi_procnodes_per_pkg);
6344 }
6345
6346 uint_t
cpuid_get_compunitid(cpu_t * cpu)6347 cpuid_get_compunitid(cpu_t *cpu)
6348 {
6349 ASSERT(cpuid_checkpass(cpu, CPUID_PASS_BASIC));
6350 return (cpu->cpu_m.mcpu_cpi->cpi_compunitid);
6351 }
6352
6353 uint_t
cpuid_get_cores_per_compunit(cpu_t * cpu)6354 cpuid_get_cores_per_compunit(cpu_t *cpu)
6355 {
6356 ASSERT(cpuid_checkpass(cpu, CPUID_PASS_BASIC));
6357 return (cpu->cpu_m.mcpu_cpi->cpi_cores_per_compunit);
6358 }
6359
6360 uint32_t
cpuid_get_apicid(cpu_t * cpu)6361 cpuid_get_apicid(cpu_t *cpu)
6362 {
6363 ASSERT(cpuid_checkpass(cpu, CPUID_PASS_BASIC));
6364 if (cpu->cpu_m.mcpu_cpi->cpi_maxeax < 1) {
6365 return (UINT32_MAX);
6366 } else {
6367 return (cpu->cpu_m.mcpu_cpi->cpi_apicid);
6368 }
6369 }
6370
6371 void
cpuid_get_addrsize(cpu_t * cpu,uint_t * pabits,uint_t * vabits)6372 cpuid_get_addrsize(cpu_t *cpu, uint_t *pabits, uint_t *vabits)
6373 {
6374 struct cpuid_info *cpi;
6375
6376 if (cpu == NULL)
6377 cpu = CPU;
6378 cpi = cpu->cpu_m.mcpu_cpi;
6379
6380 ASSERT(cpuid_checkpass(cpu, CPUID_PASS_BASIC));
6381
6382 if (pabits)
6383 *pabits = cpi->cpi_pabits;
6384 if (vabits)
6385 *vabits = cpi->cpi_vabits;
6386 }
6387
6388 size_t
cpuid_get_xsave_size(void)6389 cpuid_get_xsave_size(void)
6390 {
6391 return (MAX(cpuid_info0.cpi_xsave.xsav_max_size,
6392 sizeof (struct xsave_state)));
6393 }
6394
6395 /*
6396 * Export information about known offsets to the kernel. We only care about
6397 * things we have actually enabled support for in %xcr0.
6398 */
6399 void
cpuid_get_xsave_info(uint64_t bit,size_t * sizep,size_t * offp)6400 cpuid_get_xsave_info(uint64_t bit, size_t *sizep, size_t *offp)
6401 {
6402 size_t size, off;
6403
6404 VERIFY3U(bit & xsave_bv_all, !=, 0);
6405
6406 if (sizep == NULL)
6407 sizep = &size;
6408 if (offp == NULL)
6409 offp = &off;
6410
6411 switch (bit) {
6412 case XFEATURE_LEGACY_FP:
6413 case XFEATURE_SSE:
6414 *sizep = sizeof (struct fxsave_state);
6415 *offp = 0;
6416 break;
6417 case XFEATURE_AVX:
6418 *sizep = cpuid_info0.cpi_xsave.ymm_size;
6419 *offp = cpuid_info0.cpi_xsave.ymm_offset;
6420 break;
6421 case XFEATURE_AVX512_OPMASK:
6422 *sizep = cpuid_info0.cpi_xsave.opmask_size;
6423 *offp = cpuid_info0.cpi_xsave.opmask_offset;
6424 break;
6425 case XFEATURE_AVX512_ZMM:
6426 *sizep = cpuid_info0.cpi_xsave.zmmlo_size;
6427 *offp = cpuid_info0.cpi_xsave.zmmlo_offset;
6428 break;
6429 case XFEATURE_AVX512_HI_ZMM:
6430 *sizep = cpuid_info0.cpi_xsave.zmmhi_size;
6431 *offp = cpuid_info0.cpi_xsave.zmmhi_offset;
6432 break;
6433 default:
6434 panic("asked for unsupported xsave feature: 0x%lx", bit);
6435 }
6436 }
6437
6438 /*
6439 * Return true if the CPUs on this system require 'pointer clearing' for the
6440 * floating point error pointer exception handling. In the past, this has been
6441 * true for all AMD K7 & K8 CPUs, although newer AMD CPUs have been changed to
6442 * behave the same as Intel. This is checked via the CPUID_AMD_EBX_ERR_PTR_ZERO
6443 * feature bit and is reflected in the cpi_fp_amd_save member.
6444 */
6445 boolean_t
cpuid_need_fp_excp_handling(void)6446 cpuid_need_fp_excp_handling(void)
6447 {
6448 return (cpuid_info0.cpi_vendor == X86_VENDOR_AMD &&
6449 cpuid_info0.cpi_fp_amd_save != 0);
6450 }
6451
6452 /*
6453 * Returns the number of data TLB entries for a corresponding
6454 * pagesize. If it can't be computed, or isn't known, the
6455 * routine returns zero. If you ask about an architecturally
6456 * impossible pagesize, the routine will panic (so that the
6457 * hat implementor knows that things are inconsistent.)
6458 */
6459 uint_t
cpuid_get_dtlb_nent(cpu_t * cpu,size_t pagesize)6460 cpuid_get_dtlb_nent(cpu_t *cpu, size_t pagesize)
6461 {
6462 struct cpuid_info *cpi;
6463 uint_t dtlb_nent = 0;
6464
6465 if (cpu == NULL)
6466 cpu = CPU;
6467 cpi = cpu->cpu_m.mcpu_cpi;
6468
6469 ASSERT(cpuid_checkpass(cpu, CPUID_PASS_BASIC));
6470
6471 /*
6472 * Check the L2 TLB info
6473 */
6474 if (cpi->cpi_xmaxeax >= 0x80000006) {
6475 struct cpuid_regs *cp = &cpi->cpi_extd[6];
6476
6477 switch (pagesize) {
6478
6479 case 4 * 1024:
6480 /*
6481 * All zero in the top 16 bits of the register
6482 * indicates a unified TLB. Size is in low 16 bits.
6483 */
6484 if ((cp->cp_ebx & 0xffff0000) == 0)
6485 dtlb_nent = cp->cp_ebx & 0x0000ffff;
6486 else
6487 dtlb_nent = BITX(cp->cp_ebx, 27, 16);
6488 break;
6489
6490 case 2 * 1024 * 1024:
6491 if ((cp->cp_eax & 0xffff0000) == 0)
6492 dtlb_nent = cp->cp_eax & 0x0000ffff;
6493 else
6494 dtlb_nent = BITX(cp->cp_eax, 27, 16);
6495 break;
6496
6497 default:
6498 panic("unknown L2 pagesize");
6499 /*NOTREACHED*/
6500 }
6501 }
6502
6503 if (dtlb_nent != 0)
6504 return (dtlb_nent);
6505
6506 /*
6507 * No L2 TLB support for this size, try L1.
6508 */
6509 if (cpi->cpi_xmaxeax >= 0x80000005) {
6510 struct cpuid_regs *cp = &cpi->cpi_extd[5];
6511
6512 switch (pagesize) {
6513 case 4 * 1024:
6514 dtlb_nent = BITX(cp->cp_ebx, 23, 16);
6515 break;
6516 case 2 * 1024 * 1024:
6517 dtlb_nent = BITX(cp->cp_eax, 23, 16);
6518 break;
6519 default:
6520 panic("unknown L1 d-TLB pagesize");
6521 /*NOTREACHED*/
6522 }
6523 }
6524
6525 return (dtlb_nent);
6526 }
6527
6528 /*
6529 * Return 0 if the erratum is not present or not applicable, positive
6530 * if it is, and negative if the status of the erratum is unknown.
6531 *
6532 * See "Revision Guide for AMD Athlon(tm) 64 and AMD Opteron(tm)
6533 * Processors" #25759, Rev 3.57, August 2005
6534 */
6535 int
cpuid_opteron_erratum(cpu_t * cpu,uint_t erratum)6536 cpuid_opteron_erratum(cpu_t *cpu, uint_t erratum)
6537 {
6538 struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
6539 uint_t eax;
6540
6541 /*
6542 * Bail out if this CPU isn't an AMD CPU, or if it's
6543 * a legacy (32-bit) AMD CPU.
6544 */
6545 if (cpi->cpi_vendor != X86_VENDOR_AMD ||
6546 cpi->cpi_family == 4 || cpi->cpi_family == 5 ||
6547 cpi->cpi_family == 6) {
6548 return (0);
6549 }
6550
6551 eax = cpi->cpi_std[1].cp_eax;
6552
6553 #define SH_B0(eax) (eax == 0xf40 || eax == 0xf50)
6554 #define SH_B3(eax) (eax == 0xf51)
6555 #define B(eax) (SH_B0(eax) || SH_B3(eax))
6556
6557 #define SH_C0(eax) (eax == 0xf48 || eax == 0xf58)
6558
6559 #define SH_CG(eax) (eax == 0xf4a || eax == 0xf5a || eax == 0xf7a)
6560 #define DH_CG(eax) (eax == 0xfc0 || eax == 0xfe0 || eax == 0xff0)
6561 #define CH_CG(eax) (eax == 0xf82 || eax == 0xfb2)
6562 #define CG(eax) (SH_CG(eax) || DH_CG(eax) || CH_CG(eax))
6563
6564 #define SH_D0(eax) (eax == 0x10f40 || eax == 0x10f50 || eax == 0x10f70)
6565 #define DH_D0(eax) (eax == 0x10fc0 || eax == 0x10ff0)
6566 #define CH_D0(eax) (eax == 0x10f80 || eax == 0x10fb0)
6567 #define D0(eax) (SH_D0(eax) || DH_D0(eax) || CH_D0(eax))
6568
6569 #define SH_E0(eax) (eax == 0x20f50 || eax == 0x20f40 || eax == 0x20f70)
6570 #define JH_E1(eax) (eax == 0x20f10) /* JH8_E0 had 0x20f30 */
6571 #define DH_E3(eax) (eax == 0x20fc0 || eax == 0x20ff0)
6572 #define SH_E4(eax) (eax == 0x20f51 || eax == 0x20f71)
6573 #define BH_E4(eax) (eax == 0x20fb1)
6574 #define SH_E5(eax) (eax == 0x20f42)
6575 #define DH_E6(eax) (eax == 0x20ff2 || eax == 0x20fc2)
6576 #define JH_E6(eax) (eax == 0x20f12 || eax == 0x20f32)
6577 #define EX(eax) (SH_E0(eax) || JH_E1(eax) || DH_E3(eax) || \
6578 SH_E4(eax) || BH_E4(eax) || SH_E5(eax) || \
6579 DH_E6(eax) || JH_E6(eax))
6580
6581 #define DR_AX(eax) (eax == 0x100f00 || eax == 0x100f01 || eax == 0x100f02)
6582 #define DR_B0(eax) (eax == 0x100f20)
6583 #define DR_B1(eax) (eax == 0x100f21)
6584 #define DR_BA(eax) (eax == 0x100f2a)
6585 #define DR_B2(eax) (eax == 0x100f22)
6586 #define DR_B3(eax) (eax == 0x100f23)
6587 #define RB_C0(eax) (eax == 0x100f40)
6588
6589 switch (erratum) {
6590 case 1:
6591 return (cpi->cpi_family < 0x10);
6592 case 51: /* what does the asterisk mean? */
6593 return (B(eax) || SH_C0(eax) || CG(eax));
6594 case 52:
6595 return (B(eax));
6596 case 57:
6597 return (cpi->cpi_family <= 0x11);
6598 case 58:
6599 return (B(eax));
6600 case 60:
6601 return (cpi->cpi_family <= 0x11);
6602 case 61:
6603 case 62:
6604 case 63:
6605 case 64:
6606 case 65:
6607 case 66:
6608 case 68:
6609 case 69:
6610 case 70:
6611 case 71:
6612 return (B(eax));
6613 case 72:
6614 return (SH_B0(eax));
6615 case 74:
6616 return (B(eax));
6617 case 75:
6618 return (cpi->cpi_family < 0x10);
6619 case 76:
6620 return (B(eax));
6621 case 77:
6622 return (cpi->cpi_family <= 0x11);
6623 case 78:
6624 return (B(eax) || SH_C0(eax));
6625 case 79:
6626 return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
6627 case 80:
6628 case 81:
6629 case 82:
6630 return (B(eax));
6631 case 83:
6632 return (B(eax) || SH_C0(eax) || CG(eax));
6633 case 85:
6634 return (cpi->cpi_family < 0x10);
6635 case 86:
6636 return (SH_C0(eax) || CG(eax));
6637 case 88:
6638 return (B(eax) || SH_C0(eax));
6639 case 89:
6640 return (cpi->cpi_family < 0x10);
6641 case 90:
6642 return (B(eax) || SH_C0(eax) || CG(eax));
6643 case 91:
6644 case 92:
6645 return (B(eax) || SH_C0(eax));
6646 case 93:
6647 return (SH_C0(eax));
6648 case 94:
6649 return (B(eax) || SH_C0(eax) || CG(eax));
6650 case 95:
6651 return (B(eax) || SH_C0(eax));
6652 case 96:
6653 return (B(eax) || SH_C0(eax) || CG(eax));
6654 case 97:
6655 case 98:
6656 return (SH_C0(eax) || CG(eax));
6657 case 99:
6658 return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
6659 case 100:
6660 return (B(eax) || SH_C0(eax));
6661 case 101:
6662 case 103:
6663 return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
6664 case 104:
6665 return (SH_C0(eax) || CG(eax) || D0(eax));
6666 case 105:
6667 case 106:
6668 case 107:
6669 return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
6670 case 108:
6671 return (DH_CG(eax));
6672 case 109:
6673 return (SH_C0(eax) || CG(eax) || D0(eax));
6674 case 110:
6675 return (D0(eax) || EX(eax));
6676 case 111:
6677 return (CG(eax));
6678 case 112:
6679 return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
6680 case 113:
6681 return (eax == 0x20fc0);
6682 case 114:
6683 return (SH_E0(eax) || JH_E1(eax) || DH_E3(eax));
6684 case 115:
6685 return (SH_E0(eax) || JH_E1(eax));
6686 case 116:
6687 return (SH_E0(eax) || JH_E1(eax) || DH_E3(eax));
6688 case 117:
6689 return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax));
6690 case 118:
6691 return (SH_E0(eax) || JH_E1(eax) || SH_E4(eax) || BH_E4(eax) ||
6692 JH_E6(eax));
6693 case 121:
6694 return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax));
6695 case 122:
6696 return (cpi->cpi_family < 0x10 || cpi->cpi_family == 0x11);
6697 case 123:
6698 return (JH_E1(eax) || BH_E4(eax) || JH_E6(eax));
6699 case 131:
6700 return (cpi->cpi_family < 0x10);
6701 case 6336786:
6702
6703 /*
6704 * Test for AdvPowerMgmtInfo.TscPStateInvariant
6705 * if this is a K8 family or newer processor. We're testing for
6706 * this 'erratum' to determine whether or not we have a constant
6707 * TSC.
6708 *
6709 * Our current fix for this is to disable the C1-Clock ramping.
6710 * However, this doesn't work on newer processor families nor
6711 * does it work when virtualized as those devices don't exist.
6712 */
6713 if (cpi->cpi_family >= 0x12 || get_hwenv() != HW_NATIVE) {
6714 return (0);
6715 }
6716
6717 if (CPI_FAMILY(cpi) == 0xf) {
6718 struct cpuid_regs regs;
6719 regs.cp_eax = 0x80000007;
6720 (void) __cpuid_insn(®s);
6721 return (!(regs.cp_edx & 0x100));
6722 }
6723 return (0);
6724 case 147:
6725 /*
6726 * This erratum (K8 #147) is not present on family 10 and newer.
6727 */
6728 if (cpi->cpi_family >= 0x10) {
6729 return (0);
6730 }
6731 return (((((eax >> 12) & 0xff00) + (eax & 0xf00)) |
6732 (((eax >> 4) & 0xf) | ((eax >> 12) & 0xf0))) < 0xf40);
6733
6734 case 6671130:
6735 /*
6736 * check for processors (pre-Shanghai) that do not provide
6737 * optimal management of 1gb ptes in its tlb.
6738 */
6739 return (cpi->cpi_family == 0x10 && cpi->cpi_model < 4);
6740
6741 case 298:
6742 return (DR_AX(eax) || DR_B0(eax) || DR_B1(eax) || DR_BA(eax) ||
6743 DR_B2(eax) || RB_C0(eax));
6744
6745 case 721:
6746 return (cpi->cpi_family == 0x10 || cpi->cpi_family == 0x12);
6747
6748 default:
6749 return (-1);
6750
6751 }
6752 }
6753
6754 /*
6755 * Determine if specified erratum is present via OSVW (OS Visible Workaround).
6756 * Return 1 if erratum is present, 0 if not present and -1 if indeterminate.
6757 */
6758 int
osvw_opteron_erratum(cpu_t * cpu,uint_t erratum)6759 osvw_opteron_erratum(cpu_t *cpu, uint_t erratum)
6760 {
6761 struct cpuid_info *cpi;
6762 uint_t osvwid;
6763 static int osvwfeature = -1;
6764 uint64_t osvwlength;
6765
6766
6767 cpi = cpu->cpu_m.mcpu_cpi;
6768
6769 /* confirm OSVW supported */
6770 if (osvwfeature == -1) {
6771 osvwfeature = cpi->cpi_extd[1].cp_ecx & CPUID_AMD_ECX_OSVW;
6772 } else {
6773 /* assert that osvw feature setting is consistent on all cpus */
6774 ASSERT(osvwfeature ==
6775 (cpi->cpi_extd[1].cp_ecx & CPUID_AMD_ECX_OSVW));
6776 }
6777 if (!osvwfeature)
6778 return (-1);
6779
6780 osvwlength = rdmsr(MSR_AMD_OSVW_ID_LEN) & OSVW_ID_LEN_MASK;
6781
6782 switch (erratum) {
6783 case 298: /* osvwid is 0 */
6784 osvwid = 0;
6785 if (osvwlength <= (uint64_t)osvwid) {
6786 /* osvwid 0 is unknown */
6787 return (-1);
6788 }
6789
6790 /*
6791 * Check the OSVW STATUS MSR to determine the state
6792 * of the erratum where:
6793 * 0 - fixed by HW
6794 * 1 - BIOS has applied the workaround when BIOS
6795 * workaround is available. (Or for other errata,
6796 * OS workaround is required.)
6797 * For a value of 1, caller will confirm that the
6798 * erratum 298 workaround has indeed been applied by BIOS.
6799 *
6800 * A 1 may be set in cpus that have a HW fix
6801 * in a mixed cpu system. Regarding erratum 298:
6802 * In a multiprocessor platform, the workaround above
6803 * should be applied to all processors regardless of
6804 * silicon revision when an affected processor is
6805 * present.
6806 */
6807
6808 return (rdmsr(MSR_AMD_OSVW_STATUS +
6809 (osvwid / OSVW_ID_CNT_PER_MSR)) &
6810 (1ULL << (osvwid % OSVW_ID_CNT_PER_MSR)));
6811
6812 default:
6813 return (-1);
6814 }
6815 }
6816
6817 static const char assoc_str[] = "associativity";
6818 static const char line_str[] = "line-size";
6819 static const char size_str[] = "size";
6820
6821 static void
add_cache_prop(dev_info_t * devi,const char * label,const char * type,uint32_t val)6822 add_cache_prop(dev_info_t *devi, const char *label, const char *type,
6823 uint32_t val)
6824 {
6825 char buf[128];
6826
6827 /*
6828 * ndi_prop_update_int() is used because it is desirable for
6829 * DDI_PROP_HW_DEF and DDI_PROP_DONTSLEEP to be set.
6830 */
6831 if (snprintf(buf, sizeof (buf), "%s-%s", label, type) < sizeof (buf))
6832 (void) ndi_prop_update_int(DDI_DEV_T_NONE, devi, buf, val);
6833 }
6834
6835 /*
6836 * Intel-style cache/tlb description
6837 *
6838 * Standard cpuid level 2 gives a randomly ordered
6839 * selection of tags that index into a table that describes
6840 * cache and tlb properties.
6841 */
6842
6843 static const char l1_icache_str[] = "l1-icache";
6844 static const char l1_dcache_str[] = "l1-dcache";
6845 static const char l2_cache_str[] = "l2-cache";
6846 static const char l3_cache_str[] = "l3-cache";
6847 static const char itlb4k_str[] = "itlb-4K";
6848 static const char dtlb4k_str[] = "dtlb-4K";
6849 static const char itlb2M_str[] = "itlb-2M";
6850 static const char itlb4M_str[] = "itlb-4M";
6851 static const char dtlb4M_str[] = "dtlb-4M";
6852 static const char dtlb24_str[] = "dtlb0-2M-4M";
6853 static const char itlb424_str[] = "itlb-4K-2M-4M";
6854 static const char itlb24_str[] = "itlb-2M-4M";
6855 static const char dtlb44_str[] = "dtlb-4K-4M";
6856 static const char sl1_dcache_str[] = "sectored-l1-dcache";
6857 static const char sl2_cache_str[] = "sectored-l2-cache";
6858 static const char itrace_str[] = "itrace-cache";
6859 static const char sl3_cache_str[] = "sectored-l3-cache";
6860 static const char sh_l2_tlb4k_str[] = "shared-l2-tlb-4k";
6861
6862 static const struct cachetab {
6863 uint8_t ct_code;
6864 uint8_t ct_assoc;
6865 uint16_t ct_line_size;
6866 size_t ct_size;
6867 const char *ct_label;
6868 } intel_ctab[] = {
6869 /*
6870 * maintain descending order!
6871 *
6872 * Codes ignored - Reason
6873 * ----------------------
6874 * 40H - intel_cpuid_4_cache_info() disambiguates l2/l3 cache
6875 * f0H/f1H - Currently we do not interpret prefetch size by design
6876 */
6877 { 0xe4, 16, 64, 8*1024*1024, l3_cache_str},
6878 { 0xe3, 16, 64, 4*1024*1024, l3_cache_str},
6879 { 0xe2, 16, 64, 2*1024*1024, l3_cache_str},
6880 { 0xde, 12, 64, 6*1024*1024, l3_cache_str},
6881 { 0xdd, 12, 64, 3*1024*1024, l3_cache_str},
6882 { 0xdc, 12, 64, ((1*1024*1024)+(512*1024)), l3_cache_str},
6883 { 0xd8, 8, 64, 4*1024*1024, l3_cache_str},
6884 { 0xd7, 8, 64, 2*1024*1024, l3_cache_str},
6885 { 0xd6, 8, 64, 1*1024*1024, l3_cache_str},
6886 { 0xd2, 4, 64, 2*1024*1024, l3_cache_str},
6887 { 0xd1, 4, 64, 1*1024*1024, l3_cache_str},
6888 { 0xd0, 4, 64, 512*1024, l3_cache_str},
6889 { 0xca, 4, 0, 512, sh_l2_tlb4k_str},
6890 { 0xc0, 4, 0, 8, dtlb44_str },
6891 { 0xba, 4, 0, 64, dtlb4k_str },
6892 { 0xb4, 4, 0, 256, dtlb4k_str },
6893 { 0xb3, 4, 0, 128, dtlb4k_str },
6894 { 0xb2, 4, 0, 64, itlb4k_str },
6895 { 0xb0, 4, 0, 128, itlb4k_str },
6896 { 0x87, 8, 64, 1024*1024, l2_cache_str},
6897 { 0x86, 4, 64, 512*1024, l2_cache_str},
6898 { 0x85, 8, 32, 2*1024*1024, l2_cache_str},
6899 { 0x84, 8, 32, 1024*1024, l2_cache_str},
6900 { 0x83, 8, 32, 512*1024, l2_cache_str},
6901 { 0x82, 8, 32, 256*1024, l2_cache_str},
6902 { 0x80, 8, 64, 512*1024, l2_cache_str},
6903 { 0x7f, 2, 64, 512*1024, l2_cache_str},
6904 { 0x7d, 8, 64, 2*1024*1024, sl2_cache_str},
6905 { 0x7c, 8, 64, 1024*1024, sl2_cache_str},
6906 { 0x7b, 8, 64, 512*1024, sl2_cache_str},
6907 { 0x7a, 8, 64, 256*1024, sl2_cache_str},
6908 { 0x79, 8, 64, 128*1024, sl2_cache_str},
6909 { 0x78, 8, 64, 1024*1024, l2_cache_str},
6910 { 0x73, 8, 0, 64*1024, itrace_str},
6911 { 0x72, 8, 0, 32*1024, itrace_str},
6912 { 0x71, 8, 0, 16*1024, itrace_str},
6913 { 0x70, 8, 0, 12*1024, itrace_str},
6914 { 0x68, 4, 64, 32*1024, sl1_dcache_str},
6915 { 0x67, 4, 64, 16*1024, sl1_dcache_str},
6916 { 0x66, 4, 64, 8*1024, sl1_dcache_str},
6917 { 0x60, 8, 64, 16*1024, sl1_dcache_str},
6918 { 0x5d, 0, 0, 256, dtlb44_str},
6919 { 0x5c, 0, 0, 128, dtlb44_str},
6920 { 0x5b, 0, 0, 64, dtlb44_str},
6921 { 0x5a, 4, 0, 32, dtlb24_str},
6922 { 0x59, 0, 0, 16, dtlb4k_str},
6923 { 0x57, 4, 0, 16, dtlb4k_str},
6924 { 0x56, 4, 0, 16, dtlb4M_str},
6925 { 0x55, 0, 0, 7, itlb24_str},
6926 { 0x52, 0, 0, 256, itlb424_str},
6927 { 0x51, 0, 0, 128, itlb424_str},
6928 { 0x50, 0, 0, 64, itlb424_str},
6929 { 0x4f, 0, 0, 32, itlb4k_str},
6930 { 0x4e, 24, 64, 6*1024*1024, l2_cache_str},
6931 { 0x4d, 16, 64, 16*1024*1024, l3_cache_str},
6932 { 0x4c, 12, 64, 12*1024*1024, l3_cache_str},
6933 { 0x4b, 16, 64, 8*1024*1024, l3_cache_str},
6934 { 0x4a, 12, 64, 6*1024*1024, l3_cache_str},
6935 { 0x49, 16, 64, 4*1024*1024, l3_cache_str},
6936 { 0x48, 12, 64, 3*1024*1024, l2_cache_str},
6937 { 0x47, 8, 64, 8*1024*1024, l3_cache_str},
6938 { 0x46, 4, 64, 4*1024*1024, l3_cache_str},
6939 { 0x45, 4, 32, 2*1024*1024, l2_cache_str},
6940 { 0x44, 4, 32, 1024*1024, l2_cache_str},
6941 { 0x43, 4, 32, 512*1024, l2_cache_str},
6942 { 0x42, 4, 32, 256*1024, l2_cache_str},
6943 { 0x41, 4, 32, 128*1024, l2_cache_str},
6944 { 0x3e, 4, 64, 512*1024, sl2_cache_str},
6945 { 0x3d, 6, 64, 384*1024, sl2_cache_str},
6946 { 0x3c, 4, 64, 256*1024, sl2_cache_str},
6947 { 0x3b, 2, 64, 128*1024, sl2_cache_str},
6948 { 0x3a, 6, 64, 192*1024, sl2_cache_str},
6949 { 0x39, 4, 64, 128*1024, sl2_cache_str},
6950 { 0x30, 8, 64, 32*1024, l1_icache_str},
6951 { 0x2c, 8, 64, 32*1024, l1_dcache_str},
6952 { 0x29, 8, 64, 4096*1024, sl3_cache_str},
6953 { 0x25, 8, 64, 2048*1024, sl3_cache_str},
6954 { 0x23, 8, 64, 1024*1024, sl3_cache_str},
6955 { 0x22, 4, 64, 512*1024, sl3_cache_str},
6956 { 0x0e, 6, 64, 24*1024, l1_dcache_str},
6957 { 0x0d, 4, 32, 16*1024, l1_dcache_str},
6958 { 0x0c, 4, 32, 16*1024, l1_dcache_str},
6959 { 0x0b, 4, 0, 4, itlb4M_str},
6960 { 0x0a, 2, 32, 8*1024, l1_dcache_str},
6961 { 0x08, 4, 32, 16*1024, l1_icache_str},
6962 { 0x06, 4, 32, 8*1024, l1_icache_str},
6963 { 0x05, 4, 0, 32, dtlb4M_str},
6964 { 0x04, 4, 0, 8, dtlb4M_str},
6965 { 0x03, 4, 0, 64, dtlb4k_str},
6966 { 0x02, 4, 0, 2, itlb4M_str},
6967 { 0x01, 4, 0, 32, itlb4k_str},
6968 { 0 }
6969 };
6970
6971 static const struct cachetab cyrix_ctab[] = {
6972 { 0x70, 4, 0, 32, "tlb-4K" },
6973 { 0x80, 4, 16, 16*1024, "l1-cache" },
6974 { 0 }
6975 };
6976
6977 /*
6978 * Search a cache table for a matching entry
6979 */
6980 static const struct cachetab *
find_cacheent(const struct cachetab * ct,uint_t code)6981 find_cacheent(const struct cachetab *ct, uint_t code)
6982 {
6983 if (code != 0) {
6984 for (; ct->ct_code != 0; ct++)
6985 if (ct->ct_code <= code)
6986 break;
6987 if (ct->ct_code == code)
6988 return (ct);
6989 }
6990 return (NULL);
6991 }
6992
6993 /*
6994 * Populate cachetab entry with L2 or L3 cache-information using
6995 * cpuid function 4. This function is called from intel_walk_cacheinfo()
6996 * when descriptor 0x49 is encountered. It returns 0 if no such cache
6997 * information is found.
6998 */
6999 static int
intel_cpuid_4_cache_info(struct cachetab * ct,struct cpuid_info * cpi)7000 intel_cpuid_4_cache_info(struct cachetab *ct, struct cpuid_info *cpi)
7001 {
7002 uint32_t level, i;
7003 int ret = 0;
7004
7005 for (i = 0; i < cpi->cpi_cache_leaf_size; i++) {
7006 level = CPI_CACHE_LVL(cpi->cpi_cache_leaves[i]);
7007
7008 if (level == 2 || level == 3) {
7009 ct->ct_assoc =
7010 CPI_CACHE_WAYS(cpi->cpi_cache_leaves[i]) + 1;
7011 ct->ct_line_size =
7012 CPI_CACHE_COH_LN_SZ(cpi->cpi_cache_leaves[i]) + 1;
7013 ct->ct_size = ct->ct_assoc *
7014 (CPI_CACHE_PARTS(cpi->cpi_cache_leaves[i]) + 1) *
7015 ct->ct_line_size *
7016 (cpi->cpi_cache_leaves[i]->cp_ecx + 1);
7017
7018 if (level == 2) {
7019 ct->ct_label = l2_cache_str;
7020 } else if (level == 3) {
7021 ct->ct_label = l3_cache_str;
7022 }
7023 ret = 1;
7024 }
7025 }
7026
7027 return (ret);
7028 }
7029
7030 /*
7031 * Walk the cacheinfo descriptor, applying 'func' to every valid element
7032 * The walk is terminated if the walker returns non-zero.
7033 */
7034 static void
intel_walk_cacheinfo(struct cpuid_info * cpi,void * arg,int (* func)(void *,const struct cachetab *))7035 intel_walk_cacheinfo(struct cpuid_info *cpi,
7036 void *arg, int (*func)(void *, const struct cachetab *))
7037 {
7038 const struct cachetab *ct;
7039 struct cachetab des_49_ct, des_b1_ct;
7040 uint8_t *dp;
7041 int i;
7042
7043 if ((dp = cpi->cpi_cacheinfo) == NULL)
7044 return;
7045 for (i = 0; i < cpi->cpi_ncache; i++, dp++) {
7046 /*
7047 * For overloaded descriptor 0x49 we use cpuid function 4
7048 * if supported by the current processor, to create
7049 * cache information.
7050 * For overloaded descriptor 0xb1 we use X86_PAE flag
7051 * to disambiguate the cache information.
7052 */
7053 if (*dp == 0x49 && cpi->cpi_maxeax >= 0x4 &&
7054 intel_cpuid_4_cache_info(&des_49_ct, cpi) == 1) {
7055 ct = &des_49_ct;
7056 } else if (*dp == 0xb1) {
7057 des_b1_ct.ct_code = 0xb1;
7058 des_b1_ct.ct_assoc = 4;
7059 des_b1_ct.ct_line_size = 0;
7060 if (is_x86_feature(x86_featureset, X86FSET_PAE)) {
7061 des_b1_ct.ct_size = 8;
7062 des_b1_ct.ct_label = itlb2M_str;
7063 } else {
7064 des_b1_ct.ct_size = 4;
7065 des_b1_ct.ct_label = itlb4M_str;
7066 }
7067 ct = &des_b1_ct;
7068 } else {
7069 if ((ct = find_cacheent(intel_ctab, *dp)) == NULL) {
7070 continue;
7071 }
7072 }
7073
7074 if (func(arg, ct) != 0) {
7075 break;
7076 }
7077 }
7078 }
7079
7080 /*
7081 * (Like the Intel one, except for Cyrix CPUs)
7082 */
7083 static void
cyrix_walk_cacheinfo(struct cpuid_info * cpi,void * arg,int (* func)(void *,const struct cachetab *))7084 cyrix_walk_cacheinfo(struct cpuid_info *cpi,
7085 void *arg, int (*func)(void *, const struct cachetab *))
7086 {
7087 const struct cachetab *ct;
7088 uint8_t *dp;
7089 int i;
7090
7091 if ((dp = cpi->cpi_cacheinfo) == NULL)
7092 return;
7093 for (i = 0; i < cpi->cpi_ncache; i++, dp++) {
7094 /*
7095 * Search Cyrix-specific descriptor table first ..
7096 */
7097 if ((ct = find_cacheent(cyrix_ctab, *dp)) != NULL) {
7098 if (func(arg, ct) != 0)
7099 break;
7100 continue;
7101 }
7102 /*
7103 * .. else fall back to the Intel one
7104 */
7105 if ((ct = find_cacheent(intel_ctab, *dp)) != NULL) {
7106 if (func(arg, ct) != 0)
7107 break;
7108 continue;
7109 }
7110 }
7111 }
7112
7113 /*
7114 * A cacheinfo walker that adds associativity, line-size, and size properties
7115 * to the devinfo node it is passed as an argument.
7116 */
7117 static int
add_cacheent_props(void * arg,const struct cachetab * ct)7118 add_cacheent_props(void *arg, const struct cachetab *ct)
7119 {
7120 dev_info_t *devi = arg;
7121
7122 add_cache_prop(devi, ct->ct_label, assoc_str, ct->ct_assoc);
7123 if (ct->ct_line_size != 0)
7124 add_cache_prop(devi, ct->ct_label, line_str,
7125 ct->ct_line_size);
7126 add_cache_prop(devi, ct->ct_label, size_str, ct->ct_size);
7127 return (0);
7128 }
7129
7130
7131 static const char fully_assoc[] = "fully-associative?";
7132
7133 /*
7134 * AMD style cache/tlb description
7135 *
7136 * Extended functions 5 and 6 directly describe properties of
7137 * tlbs and various cache levels.
7138 */
7139 static void
add_amd_assoc(dev_info_t * devi,const char * label,uint_t assoc)7140 add_amd_assoc(dev_info_t *devi, const char *label, uint_t assoc)
7141 {
7142 switch (assoc) {
7143 case 0: /* reserved; ignore */
7144 break;
7145 default:
7146 add_cache_prop(devi, label, assoc_str, assoc);
7147 break;
7148 case 0xff:
7149 add_cache_prop(devi, label, fully_assoc, 1);
7150 break;
7151 }
7152 }
7153
7154 static void
add_amd_tlb(dev_info_t * devi,const char * label,uint_t assoc,uint_t size)7155 add_amd_tlb(dev_info_t *devi, const char *label, uint_t assoc, uint_t size)
7156 {
7157 if (size == 0)
7158 return;
7159 add_cache_prop(devi, label, size_str, size);
7160 add_amd_assoc(devi, label, assoc);
7161 }
7162
7163 static void
add_amd_cache(dev_info_t * devi,const char * label,uint_t size,uint_t assoc,uint_t lines_per_tag,uint_t line_size)7164 add_amd_cache(dev_info_t *devi, const char *label,
7165 uint_t size, uint_t assoc, uint_t lines_per_tag, uint_t line_size)
7166 {
7167 if (size == 0 || line_size == 0)
7168 return;
7169 add_amd_assoc(devi, label, assoc);
7170 /*
7171 * Most AMD parts have a sectored cache. Multiple cache lines are
7172 * associated with each tag. A sector consists of all cache lines
7173 * associated with a tag. For example, the AMD K6-III has a sector
7174 * size of 2 cache lines per tag.
7175 */
7176 if (lines_per_tag != 0)
7177 add_cache_prop(devi, label, "lines-per-tag", lines_per_tag);
7178 add_cache_prop(devi, label, line_str, line_size);
7179 add_cache_prop(devi, label, size_str, size * 1024);
7180 }
7181
7182 static void
add_amd_l2_assoc(dev_info_t * devi,const char * label,uint_t assoc)7183 add_amd_l2_assoc(dev_info_t *devi, const char *label, uint_t assoc)
7184 {
7185 switch (assoc) {
7186 case 0: /* off */
7187 break;
7188 case 1:
7189 case 2:
7190 case 4:
7191 add_cache_prop(devi, label, assoc_str, assoc);
7192 break;
7193 case 6:
7194 add_cache_prop(devi, label, assoc_str, 8);
7195 break;
7196 case 8:
7197 add_cache_prop(devi, label, assoc_str, 16);
7198 break;
7199 case 0xf:
7200 add_cache_prop(devi, label, fully_assoc, 1);
7201 break;
7202 default: /* reserved; ignore */
7203 break;
7204 }
7205 }
7206
7207 static void
add_amd_l2_tlb(dev_info_t * devi,const char * label,uint_t assoc,uint_t size)7208 add_amd_l2_tlb(dev_info_t *devi, const char *label, uint_t assoc, uint_t size)
7209 {
7210 if (size == 0 || assoc == 0)
7211 return;
7212 add_amd_l2_assoc(devi, label, assoc);
7213 add_cache_prop(devi, label, size_str, size);
7214 }
7215
7216 static void
add_amd_l2_cache(dev_info_t * devi,const char * label,uint_t size,uint_t assoc,uint_t lines_per_tag,uint_t line_size)7217 add_amd_l2_cache(dev_info_t *devi, const char *label,
7218 uint_t size, uint_t assoc, uint_t lines_per_tag, uint_t line_size)
7219 {
7220 if (size == 0 || assoc == 0 || line_size == 0)
7221 return;
7222 add_amd_l2_assoc(devi, label, assoc);
7223 if (lines_per_tag != 0)
7224 add_cache_prop(devi, label, "lines-per-tag", lines_per_tag);
7225 add_cache_prop(devi, label, line_str, line_size);
7226 add_cache_prop(devi, label, size_str, size * 1024);
7227 }
7228
7229 static void
amd_cache_info(struct cpuid_info * cpi,dev_info_t * devi)7230 amd_cache_info(struct cpuid_info *cpi, dev_info_t *devi)
7231 {
7232 struct cpuid_regs *cp;
7233
7234 if (cpi->cpi_xmaxeax < 0x80000005)
7235 return;
7236 cp = &cpi->cpi_extd[5];
7237
7238 /*
7239 * 4M/2M L1 TLB configuration
7240 *
7241 * We report the size for 2M pages because AMD uses two
7242 * TLB entries for one 4M page.
7243 */
7244 add_amd_tlb(devi, "dtlb-2M",
7245 BITX(cp->cp_eax, 31, 24), BITX(cp->cp_eax, 23, 16));
7246 add_amd_tlb(devi, "itlb-2M",
7247 BITX(cp->cp_eax, 15, 8), BITX(cp->cp_eax, 7, 0));
7248
7249 /*
7250 * 4K L1 TLB configuration
7251 */
7252
7253 switch (cpi->cpi_vendor) {
7254 uint_t nentries;
7255 case X86_VENDOR_TM:
7256 if (cpi->cpi_family >= 5) {
7257 /*
7258 * Crusoe processors have 256 TLB entries, but
7259 * cpuid data format constrains them to only
7260 * reporting 255 of them.
7261 */
7262 if ((nentries = BITX(cp->cp_ebx, 23, 16)) == 255)
7263 nentries = 256;
7264 /*
7265 * Crusoe processors also have a unified TLB
7266 */
7267 add_amd_tlb(devi, "tlb-4K", BITX(cp->cp_ebx, 31, 24),
7268 nentries);
7269 break;
7270 }
7271 /*FALLTHROUGH*/
7272 default:
7273 add_amd_tlb(devi, itlb4k_str,
7274 BITX(cp->cp_ebx, 31, 24), BITX(cp->cp_ebx, 23, 16));
7275 add_amd_tlb(devi, dtlb4k_str,
7276 BITX(cp->cp_ebx, 15, 8), BITX(cp->cp_ebx, 7, 0));
7277 break;
7278 }
7279
7280 /*
7281 * data L1 cache configuration
7282 */
7283
7284 add_amd_cache(devi, l1_dcache_str,
7285 BITX(cp->cp_ecx, 31, 24), BITX(cp->cp_ecx, 23, 16),
7286 BITX(cp->cp_ecx, 15, 8), BITX(cp->cp_ecx, 7, 0));
7287
7288 /*
7289 * code L1 cache configuration
7290 */
7291
7292 add_amd_cache(devi, l1_icache_str,
7293 BITX(cp->cp_edx, 31, 24), BITX(cp->cp_edx, 23, 16),
7294 BITX(cp->cp_edx, 15, 8), BITX(cp->cp_edx, 7, 0));
7295
7296 if (cpi->cpi_xmaxeax < 0x80000006)
7297 return;
7298 cp = &cpi->cpi_extd[6];
7299
7300 /* Check for a unified L2 TLB for large pages */
7301
7302 if (BITX(cp->cp_eax, 31, 16) == 0)
7303 add_amd_l2_tlb(devi, "l2-tlb-2M",
7304 BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
7305 else {
7306 add_amd_l2_tlb(devi, "l2-dtlb-2M",
7307 BITX(cp->cp_eax, 31, 28), BITX(cp->cp_eax, 27, 16));
7308 add_amd_l2_tlb(devi, "l2-itlb-2M",
7309 BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
7310 }
7311
7312 /* Check for a unified L2 TLB for 4K pages */
7313
7314 if (BITX(cp->cp_ebx, 31, 16) == 0) {
7315 add_amd_l2_tlb(devi, "l2-tlb-4K",
7316 BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
7317 } else {
7318 add_amd_l2_tlb(devi, "l2-dtlb-4K",
7319 BITX(cp->cp_eax, 31, 28), BITX(cp->cp_eax, 27, 16));
7320 add_amd_l2_tlb(devi, "l2-itlb-4K",
7321 BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0));
7322 }
7323
7324 add_amd_l2_cache(devi, l2_cache_str,
7325 BITX(cp->cp_ecx, 31, 16), BITX(cp->cp_ecx, 15, 12),
7326 BITX(cp->cp_ecx, 11, 8), BITX(cp->cp_ecx, 7, 0));
7327 }
7328
7329 /*
7330 * There are two basic ways that the x86 world describes it cache
7331 * and tlb architecture - Intel's way and AMD's way.
7332 *
7333 * Return which flavor of cache architecture we should use
7334 */
7335 static int
x86_which_cacheinfo(struct cpuid_info * cpi)7336 x86_which_cacheinfo(struct cpuid_info *cpi)
7337 {
7338 switch (cpi->cpi_vendor) {
7339 case X86_VENDOR_Intel:
7340 if (cpi->cpi_maxeax >= 2)
7341 return (X86_VENDOR_Intel);
7342 break;
7343 case X86_VENDOR_AMD:
7344 /*
7345 * The K5 model 1 was the first part from AMD that reported
7346 * cache sizes via extended cpuid functions.
7347 */
7348 if (cpi->cpi_family > 5 ||
7349 (cpi->cpi_family == 5 && cpi->cpi_model >= 1))
7350 return (X86_VENDOR_AMD);
7351 break;
7352 case X86_VENDOR_HYGON:
7353 return (X86_VENDOR_AMD);
7354 case X86_VENDOR_TM:
7355 if (cpi->cpi_family >= 5)
7356 return (X86_VENDOR_AMD);
7357 /*FALLTHROUGH*/
7358 default:
7359 /*
7360 * If they have extended CPU data for 0x80000005
7361 * then we assume they have AMD-format cache
7362 * information.
7363 *
7364 * If not, and the vendor happens to be Cyrix,
7365 * then try our-Cyrix specific handler.
7366 *
7367 * If we're not Cyrix, then assume we're using Intel's
7368 * table-driven format instead.
7369 */
7370 if (cpi->cpi_xmaxeax >= 0x80000005)
7371 return (X86_VENDOR_AMD);
7372 else if (cpi->cpi_vendor == X86_VENDOR_Cyrix)
7373 return (X86_VENDOR_Cyrix);
7374 else if (cpi->cpi_maxeax >= 2)
7375 return (X86_VENDOR_Intel);
7376 break;
7377 }
7378 return (-1);
7379 }
7380
7381 void
cpuid_set_cpu_properties(void * dip,processorid_t cpu_id,struct cpuid_info * cpi)7382 cpuid_set_cpu_properties(void *dip, processorid_t cpu_id,
7383 struct cpuid_info *cpi)
7384 {
7385 dev_info_t *cpu_devi;
7386 int create;
7387
7388 cpu_devi = (dev_info_t *)dip;
7389
7390 /* device_type */
7391 (void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi,
7392 "device_type", "cpu");
7393
7394 /* reg */
7395 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7396 "reg", cpu_id);
7397
7398 /* cpu-mhz, and clock-frequency */
7399 if (cpu_freq > 0) {
7400 long long mul;
7401
7402 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7403 "cpu-mhz", cpu_freq);
7404 if ((mul = cpu_freq * 1000000LL) <= INT_MAX)
7405 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7406 "clock-frequency", (int)mul);
7407 }
7408
7409 ASSERT(is_x86_feature(x86_featureset, X86FSET_CPUID));
7410
7411 /* vendor-id */
7412 (void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi,
7413 "vendor-id", cpi->cpi_vendorstr);
7414
7415 if (cpi->cpi_maxeax == 0) {
7416 return;
7417 }
7418
7419 /*
7420 * family, model, and step
7421 */
7422 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7423 "family", CPI_FAMILY(cpi));
7424 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7425 "cpu-model", CPI_MODEL(cpi));
7426 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7427 "stepping-id", CPI_STEP(cpi));
7428
7429 /* type */
7430 switch (cpi->cpi_vendor) {
7431 case X86_VENDOR_Intel:
7432 create = 1;
7433 break;
7434 default:
7435 create = 0;
7436 break;
7437 }
7438 if (create)
7439 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7440 "type", CPI_TYPE(cpi));
7441
7442 /* ext-family */
7443 switch (cpi->cpi_vendor) {
7444 case X86_VENDOR_Intel:
7445 case X86_VENDOR_AMD:
7446 create = cpi->cpi_family >= 0xf;
7447 break;
7448 case X86_VENDOR_HYGON:
7449 create = 1;
7450 break;
7451 default:
7452 create = 0;
7453 break;
7454 }
7455 if (create)
7456 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7457 "ext-family", CPI_FAMILY_XTD(cpi));
7458
7459 /* ext-model */
7460 switch (cpi->cpi_vendor) {
7461 case X86_VENDOR_Intel:
7462 create = IS_EXTENDED_MODEL_INTEL(cpi);
7463 break;
7464 case X86_VENDOR_AMD:
7465 create = CPI_FAMILY(cpi) == 0xf;
7466 break;
7467 case X86_VENDOR_HYGON:
7468 create = 1;
7469 break;
7470 default:
7471 create = 0;
7472 break;
7473 }
7474 if (create)
7475 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7476 "ext-model", CPI_MODEL_XTD(cpi));
7477
7478 /* generation */
7479 switch (cpi->cpi_vendor) {
7480 case X86_VENDOR_AMD:
7481 case X86_VENDOR_HYGON:
7482 /*
7483 * AMD K5 model 1 was the first part to support this
7484 */
7485 create = cpi->cpi_xmaxeax >= 0x80000001;
7486 break;
7487 default:
7488 create = 0;
7489 break;
7490 }
7491 if (create)
7492 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7493 "generation", BITX((cpi)->cpi_extd[1].cp_eax, 11, 8));
7494
7495 /* brand-id */
7496 switch (cpi->cpi_vendor) {
7497 case X86_VENDOR_Intel:
7498 /*
7499 * brand id first appeared on Pentium III Xeon model 8,
7500 * and Celeron model 8 processors and Opteron
7501 */
7502 create = cpi->cpi_family > 6 ||
7503 (cpi->cpi_family == 6 && cpi->cpi_model >= 8);
7504 break;
7505 case X86_VENDOR_AMD:
7506 create = cpi->cpi_family >= 0xf;
7507 break;
7508 case X86_VENDOR_HYGON:
7509 create = 1;
7510 break;
7511 default:
7512 create = 0;
7513 break;
7514 }
7515 if (create && cpi->cpi_brandid != 0) {
7516 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7517 "brand-id", cpi->cpi_brandid);
7518 }
7519
7520 /* chunks, and apic-id */
7521 switch (cpi->cpi_vendor) {
7522 /*
7523 * first available on Pentium IV and Opteron (K8)
7524 */
7525 case X86_VENDOR_Intel:
7526 create = IS_NEW_F6(cpi) || cpi->cpi_family >= 0xf;
7527 break;
7528 case X86_VENDOR_AMD:
7529 create = cpi->cpi_family >= 0xf;
7530 break;
7531 case X86_VENDOR_HYGON:
7532 create = 1;
7533 break;
7534 default:
7535 create = 0;
7536 break;
7537 }
7538 if (create) {
7539 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7540 "chunks", CPI_CHUNKS(cpi));
7541 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7542 "apic-id", cpi->cpi_apicid);
7543 if (cpi->cpi_chipid >= 0) {
7544 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7545 "chip#", cpi->cpi_chipid);
7546 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7547 "clog#", cpi->cpi_clogid);
7548 }
7549 }
7550
7551 /* cpuid-features */
7552 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7553 "cpuid-features", CPI_FEATURES_EDX(cpi));
7554
7555
7556 /* cpuid-features-ecx */
7557 switch (cpi->cpi_vendor) {
7558 case X86_VENDOR_Intel:
7559 create = IS_NEW_F6(cpi) || cpi->cpi_family >= 0xf;
7560 break;
7561 case X86_VENDOR_AMD:
7562 create = cpi->cpi_family >= 0xf;
7563 break;
7564 case X86_VENDOR_HYGON:
7565 create = 1;
7566 break;
7567 default:
7568 create = 0;
7569 break;
7570 }
7571 if (create)
7572 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7573 "cpuid-features-ecx", CPI_FEATURES_ECX(cpi));
7574
7575 /* ext-cpuid-features */
7576 switch (cpi->cpi_vendor) {
7577 case X86_VENDOR_Intel:
7578 case X86_VENDOR_AMD:
7579 case X86_VENDOR_HYGON:
7580 case X86_VENDOR_Cyrix:
7581 case X86_VENDOR_TM:
7582 case X86_VENDOR_Centaur:
7583 create = cpi->cpi_xmaxeax >= 0x80000001;
7584 break;
7585 default:
7586 create = 0;
7587 break;
7588 }
7589 if (create) {
7590 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7591 "ext-cpuid-features", CPI_FEATURES_XTD_EDX(cpi));
7592 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi,
7593 "ext-cpuid-features-ecx", CPI_FEATURES_XTD_ECX(cpi));
7594 }
7595
7596 /*
7597 * Brand String first appeared in Intel Pentium IV, AMD K5
7598 * model 1, and Cyrix GXm. On earlier models we try and
7599 * simulate something similar .. so this string should always
7600 * same -something- about the processor, however lame.
7601 */
7602 (void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi,
7603 "brand-string", cpi->cpi_brandstr);
7604
7605 /*
7606 * Finally, cache and tlb information
7607 */
7608 switch (x86_which_cacheinfo(cpi)) {
7609 case X86_VENDOR_Intel:
7610 intel_walk_cacheinfo(cpi, cpu_devi, add_cacheent_props);
7611 break;
7612 case X86_VENDOR_Cyrix:
7613 cyrix_walk_cacheinfo(cpi, cpu_devi, add_cacheent_props);
7614 break;
7615 case X86_VENDOR_AMD:
7616 amd_cache_info(cpi, cpu_devi);
7617 break;
7618 default:
7619 break;
7620 }
7621 }
7622
7623 struct l2info {
7624 int *l2i_csz;
7625 int *l2i_lsz;
7626 int *l2i_assoc;
7627 int l2i_ret;
7628 };
7629
7630 /*
7631 * A cacheinfo walker that fetches the size, line-size and associativity
7632 * of the L2 cache
7633 */
7634 static int
intel_l2cinfo(void * arg,const struct cachetab * ct)7635 intel_l2cinfo(void *arg, const struct cachetab *ct)
7636 {
7637 struct l2info *l2i = arg;
7638 int *ip;
7639
7640 if (ct->ct_label != l2_cache_str &&
7641 ct->ct_label != sl2_cache_str)
7642 return (0); /* not an L2 -- keep walking */
7643
7644 if ((ip = l2i->l2i_csz) != NULL)
7645 *ip = ct->ct_size;
7646 if ((ip = l2i->l2i_lsz) != NULL)
7647 *ip = ct->ct_line_size;
7648 if ((ip = l2i->l2i_assoc) != NULL)
7649 *ip = ct->ct_assoc;
7650 l2i->l2i_ret = ct->ct_size;
7651 return (1); /* was an L2 -- terminate walk */
7652 }
7653
7654 /*
7655 * AMD L2/L3 Cache and TLB Associativity Field Definition:
7656 *
7657 * Unlike the associativity for the L1 cache and tlb where the 8 bit
7658 * value is the associativity, the associativity for the L2 cache and
7659 * tlb is encoded in the following table. The 4 bit L2 value serves as
7660 * an index into the amd_afd[] array to determine the associativity.
7661 * -1 is undefined. 0 is fully associative.
7662 */
7663
7664 static int amd_afd[] =
7665 {-1, 1, 2, -1, 4, -1, 8, -1, 16, -1, 32, 48, 64, 96, 128, 0};
7666
7667 static void
amd_l2cacheinfo(struct cpuid_info * cpi,struct l2info * l2i)7668 amd_l2cacheinfo(struct cpuid_info *cpi, struct l2info *l2i)
7669 {
7670 struct cpuid_regs *cp;
7671 uint_t size, assoc;
7672 int i;
7673 int *ip;
7674
7675 if (cpi->cpi_xmaxeax < 0x80000006)
7676 return;
7677 cp = &cpi->cpi_extd[6];
7678
7679 if ((i = BITX(cp->cp_ecx, 15, 12)) != 0 &&
7680 (size = BITX(cp->cp_ecx, 31, 16)) != 0) {
7681 uint_t cachesz = size * 1024;
7682 assoc = amd_afd[i];
7683
7684 ASSERT(assoc != -1);
7685
7686 if ((ip = l2i->l2i_csz) != NULL)
7687 *ip = cachesz;
7688 if ((ip = l2i->l2i_lsz) != NULL)
7689 *ip = BITX(cp->cp_ecx, 7, 0);
7690 if ((ip = l2i->l2i_assoc) != NULL)
7691 *ip = assoc;
7692 l2i->l2i_ret = cachesz;
7693 }
7694 }
7695
7696 int
getl2cacheinfo(cpu_t * cpu,int * csz,int * lsz,int * assoc)7697 getl2cacheinfo(cpu_t *cpu, int *csz, int *lsz, int *assoc)
7698 {
7699 struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
7700 struct l2info __l2info, *l2i = &__l2info;
7701
7702 l2i->l2i_csz = csz;
7703 l2i->l2i_lsz = lsz;
7704 l2i->l2i_assoc = assoc;
7705 l2i->l2i_ret = -1;
7706
7707 switch (x86_which_cacheinfo(cpi)) {
7708 case X86_VENDOR_Intel:
7709 intel_walk_cacheinfo(cpi, l2i, intel_l2cinfo);
7710 break;
7711 case X86_VENDOR_Cyrix:
7712 cyrix_walk_cacheinfo(cpi, l2i, intel_l2cinfo);
7713 break;
7714 case X86_VENDOR_AMD:
7715 amd_l2cacheinfo(cpi, l2i);
7716 break;
7717 default:
7718 break;
7719 }
7720 return (l2i->l2i_ret);
7721 }
7722
7723 #if !defined(__xpv)
7724
7725 uint32_t *
cpuid_mwait_alloc(cpu_t * cpu)7726 cpuid_mwait_alloc(cpu_t *cpu)
7727 {
7728 uint32_t *ret;
7729 size_t mwait_size;
7730
7731 ASSERT(cpuid_checkpass(CPU, CPUID_PASS_EXTENDED));
7732
7733 mwait_size = CPU->cpu_m.mcpu_cpi->cpi_mwait.mon_max;
7734 if (mwait_size == 0)
7735 return (NULL);
7736
7737 /*
7738 * kmem_alloc() returns cache line size aligned data for mwait_size
7739 * allocations. mwait_size is currently cache line sized. Neither
7740 * of these implementation details are guarantied to be true in the
7741 * future.
7742 *
7743 * First try allocating mwait_size as kmem_alloc() currently returns
7744 * correctly aligned memory. If kmem_alloc() does not return
7745 * mwait_size aligned memory, then use mwait_size ROUNDUP.
7746 *
7747 * Set cpi_mwait.buf_actual and cpi_mwait.size_actual in case we
7748 * decide to free this memory.
7749 */
7750 ret = kmem_zalloc(mwait_size, KM_SLEEP);
7751 if (ret == (uint32_t *)P2ROUNDUP((uintptr_t)ret, mwait_size)) {
7752 cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual = ret;
7753 cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual = mwait_size;
7754 *ret = MWAIT_RUNNING;
7755 return (ret);
7756 } else {
7757 kmem_free(ret, mwait_size);
7758 ret = kmem_zalloc(mwait_size * 2, KM_SLEEP);
7759 cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual = ret;
7760 cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual = mwait_size * 2;
7761 ret = (uint32_t *)P2ROUNDUP((uintptr_t)ret, mwait_size);
7762 *ret = MWAIT_RUNNING;
7763 return (ret);
7764 }
7765 }
7766
7767 void
cpuid_mwait_free(cpu_t * cpu)7768 cpuid_mwait_free(cpu_t *cpu)
7769 {
7770 if (cpu->cpu_m.mcpu_cpi == NULL) {
7771 return;
7772 }
7773
7774 if (cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual != NULL &&
7775 cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual > 0) {
7776 kmem_free(cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual,
7777 cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual);
7778 }
7779
7780 cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual = NULL;
7781 cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual = 0;
7782 }
7783
7784 void
patch_tsc_read(int flag)7785 patch_tsc_read(int flag)
7786 {
7787 size_t cnt;
7788
7789 switch (flag) {
7790 case TSC_NONE:
7791 cnt = &_no_rdtsc_end - &_no_rdtsc_start;
7792 (void) memcpy((void *)tsc_read, (void *)&_no_rdtsc_start, cnt);
7793 break;
7794 case TSC_RDTSC_LFENCE:
7795 cnt = &_tsc_lfence_end - &_tsc_lfence_start;
7796 (void) memcpy((void *)tsc_read,
7797 (void *)&_tsc_lfence_start, cnt);
7798 break;
7799 case TSC_TSCP:
7800 cnt = &_tscp_end - &_tscp_start;
7801 (void) memcpy((void *)tsc_read, (void *)&_tscp_start, cnt);
7802 break;
7803 default:
7804 /* Bail for unexpected TSC types. (TSC_NONE covers 0) */
7805 cmn_err(CE_PANIC, "Unrecogized TSC type: %d", flag);
7806 break;
7807 }
7808 tsc_type = flag;
7809 }
7810
7811 int
cpuid_deep_cstates_supported(void)7812 cpuid_deep_cstates_supported(void)
7813 {
7814 struct cpuid_info *cpi;
7815 struct cpuid_regs regs;
7816
7817 ASSERT(cpuid_checkpass(CPU, CPUID_PASS_BASIC));
7818 ASSERT(is_x86_feature(x86_featureset, X86FSET_CPUID));
7819
7820 cpi = CPU->cpu_m.mcpu_cpi;
7821
7822 switch (cpi->cpi_vendor) {
7823 case X86_VENDOR_Intel:
7824 if (cpi->cpi_xmaxeax < 0x80000007)
7825 return (0);
7826
7827 /*
7828 * Does TSC run at a constant rate in all C-states?
7829 */
7830 regs.cp_eax = 0x80000007;
7831 (void) __cpuid_insn(®s);
7832 return (regs.cp_edx & CPUID_TSC_CSTATE_INVARIANCE);
7833
7834 default:
7835 return (0);
7836 }
7837 }
7838
7839 #endif /* !__xpv */
7840
7841 void
post_startup_cpu_fixups(void)7842 post_startup_cpu_fixups(void)
7843 {
7844 #ifndef __xpv
7845 /*
7846 * Some AMD processors support C1E state. Entering this state will
7847 * cause the local APIC timer to stop, which we can't deal with at
7848 * this time.
7849 */
7850 if (cpuid_getvendor(CPU) == X86_VENDOR_AMD) {
7851 on_trap_data_t otd;
7852 uint64_t reg;
7853
7854 if (!on_trap(&otd, OT_DATA_ACCESS)) {
7855 reg = rdmsr(MSR_AMD_INT_PENDING_CMP_HALT);
7856 /* Disable C1E state if it is enabled by BIOS */
7857 if ((reg >> AMD_ACTONCMPHALT_SHIFT) &
7858 AMD_ACTONCMPHALT_MASK) {
7859 reg &= ~(AMD_ACTONCMPHALT_MASK <<
7860 AMD_ACTONCMPHALT_SHIFT);
7861 wrmsr(MSR_AMD_INT_PENDING_CMP_HALT, reg);
7862 }
7863 }
7864 no_trap();
7865 }
7866 #endif /* !__xpv */
7867 }
7868
7869 void
enable_pcid(void)7870 enable_pcid(void)
7871 {
7872 if (x86_use_pcid == -1)
7873 x86_use_pcid = is_x86_feature(x86_featureset, X86FSET_PCID);
7874
7875 if (x86_use_invpcid == -1) {
7876 x86_use_invpcid = is_x86_feature(x86_featureset,
7877 X86FSET_INVPCID);
7878 }
7879
7880 if (!x86_use_pcid)
7881 return;
7882
7883 /*
7884 * Intel say that on setting PCIDE, it immediately starts using the PCID
7885 * bits; better make sure there's nothing there.
7886 */
7887 ASSERT((getcr3() & MMU_PAGEOFFSET) == PCID_NONE);
7888
7889 setcr4(getcr4() | CR4_PCIDE);
7890 }
7891
7892 /*
7893 * Setup necessary registers to enable XSAVE feature on this processor.
7894 * This function needs to be called early enough, so that no xsave/xrstor
7895 * ops will execute on the processor before the MSRs are properly set up.
7896 *
7897 * Current implementation has the following assumption:
7898 * - cpuid_pass_basic() is done, so that X86 features are known.
7899 * - fpu_probe() is done, so that fp_save_mech is chosen.
7900 */
7901 void
xsave_setup_msr(cpu_t * cpu)7902 xsave_setup_msr(cpu_t *cpu)
7903 {
7904 ASSERT(cpuid_checkpass(cpu, CPUID_PASS_BASIC));
7905 ASSERT(fp_save_mech == FP_XSAVE);
7906 ASSERT(is_x86_feature(x86_featureset, X86FSET_XSAVE));
7907
7908 /* Enable OSXSAVE in CR4. */
7909 setcr4(getcr4() | CR4_OSXSAVE);
7910 /*
7911 * Update SW copy of ECX, so that /dev/cpu/self/cpuid will report
7912 * correct value.
7913 */
7914 cpu->cpu_m.mcpu_cpi->cpi_std[1].cp_ecx |= CPUID_INTC_ECX_OSXSAVE;
7915 setup_xfem();
7916 }
7917
7918 /*
7919 * Starting with the Westmere processor the local
7920 * APIC timer will continue running in all C-states,
7921 * including the deepest C-states.
7922 */
7923 int
cpuid_arat_supported(void)7924 cpuid_arat_supported(void)
7925 {
7926 struct cpuid_info *cpi;
7927 struct cpuid_regs regs;
7928
7929 ASSERT(cpuid_checkpass(CPU, CPUID_PASS_BASIC));
7930 ASSERT(is_x86_feature(x86_featureset, X86FSET_CPUID));
7931
7932 cpi = CPU->cpu_m.mcpu_cpi;
7933
7934 switch (cpi->cpi_vendor) {
7935 case X86_VENDOR_Intel:
7936 /*
7937 * Always-running Local APIC Timer is
7938 * indicated by CPUID.6.EAX[2].
7939 */
7940 if (cpi->cpi_maxeax >= 6) {
7941 regs.cp_eax = 6;
7942 (void) cpuid_insn(NULL, ®s);
7943 return (regs.cp_eax & CPUID_INTC_EAX_ARAT);
7944 } else {
7945 return (0);
7946 }
7947 default:
7948 return (0);
7949 }
7950 }
7951
7952 /*
7953 * Check support for Intel ENERGY_PERF_BIAS feature
7954 */
7955 int
cpuid_iepb_supported(struct cpu * cp)7956 cpuid_iepb_supported(struct cpu *cp)
7957 {
7958 struct cpuid_info *cpi = cp->cpu_m.mcpu_cpi;
7959 struct cpuid_regs regs;
7960
7961 ASSERT(cpuid_checkpass(cp, CPUID_PASS_BASIC));
7962 ASSERT(is_x86_feature(x86_featureset, X86FSET_CPUID));
7963
7964 if (!(is_x86_feature(x86_featureset, X86FSET_MSR))) {
7965 return (0);
7966 }
7967
7968 /*
7969 * Intel ENERGY_PERF_BIAS MSR is indicated by
7970 * capability bit CPUID.6.ECX.3
7971 */
7972 if ((cpi->cpi_vendor != X86_VENDOR_Intel) || (cpi->cpi_maxeax < 6))
7973 return (0);
7974
7975 regs.cp_eax = 0x6;
7976 (void) cpuid_insn(NULL, ®s);
7977 return (regs.cp_ecx & CPUID_INTC_ECX_PERFBIAS);
7978 }
7979
7980 /*
7981 * Check support for TSC deadline timer
7982 *
7983 * TSC deadline timer provides a superior software programming
7984 * model over local APIC timer that eliminates "time drifts".
7985 * Instead of specifying a relative time, software specifies an
7986 * absolute time as the target at which the processor should
7987 * generate a timer event.
7988 */
7989 int
cpuid_deadline_tsc_supported(void)7990 cpuid_deadline_tsc_supported(void)
7991 {
7992 struct cpuid_info *cpi = CPU->cpu_m.mcpu_cpi;
7993 struct cpuid_regs regs;
7994
7995 ASSERT(cpuid_checkpass(CPU, CPUID_PASS_BASIC));
7996 ASSERT(is_x86_feature(x86_featureset, X86FSET_CPUID));
7997
7998 switch (cpi->cpi_vendor) {
7999 case X86_VENDOR_Intel:
8000 if (cpi->cpi_maxeax >= 1) {
8001 regs.cp_eax = 1;
8002 (void) cpuid_insn(NULL, ®s);
8003 return (regs.cp_ecx & CPUID_DEADLINE_TSC);
8004 } else {
8005 return (0);
8006 }
8007 default:
8008 return (0);
8009 }
8010 }
8011
8012 #if !defined(__xpv)
8013 /*
8014 * Patch in versions of bcopy for high performance Intel Nhm processors
8015 * and later...
8016 */
8017 void
patch_memops(uint_t vendor)8018 patch_memops(uint_t vendor)
8019 {
8020 size_t cnt, i;
8021 caddr_t to, from;
8022
8023 if ((vendor == X86_VENDOR_Intel) &&
8024 is_x86_feature(x86_featureset, X86FSET_SSE4_2)) {
8025 cnt = &bcopy_patch_end - &bcopy_patch_start;
8026 to = &bcopy_ck_size;
8027 from = &bcopy_patch_start;
8028 for (i = 0; i < cnt; i++) {
8029 *to++ = *from++;
8030 }
8031 }
8032 }
8033 #endif /* !__xpv */
8034
8035 /*
8036 * We're being asked to tell the system how many bits are required to represent
8037 * the various thread and strand IDs. While it's tempting to derive this based
8038 * on the values in cpi_ncore_per_chip and cpi_ncpu_per_chip, that isn't quite
8039 * correct. Instead, this needs to be based on the number of bits that the APIC
8040 * allows for these different configurations. We only update these to a larger
8041 * value if we find one.
8042 */
8043 void
cpuid_get_ext_topo(cpu_t * cpu,uint_t * core_nbits,uint_t * strand_nbits)8044 cpuid_get_ext_topo(cpu_t *cpu, uint_t *core_nbits, uint_t *strand_nbits)
8045 {
8046 struct cpuid_info *cpi;
8047
8048 VERIFY(cpuid_checkpass(CPU, CPUID_PASS_BASIC));
8049 cpi = cpu->cpu_m.mcpu_cpi;
8050
8051 if (cpi->cpi_ncore_bits > *core_nbits) {
8052 *core_nbits = cpi->cpi_ncore_bits;
8053 }
8054
8055 if (cpi->cpi_nthread_bits > *strand_nbits) {
8056 *strand_nbits = cpi->cpi_nthread_bits;
8057 }
8058 }
8059
8060 void
cpuid_pass_ucode(cpu_t * cpu,uchar_t * fset)8061 cpuid_pass_ucode(cpu_t *cpu, uchar_t *fset)
8062 {
8063 struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi;
8064 struct cpuid_regs cp;
8065
8066 /*
8067 * Reread the CPUID portions that we need for various security
8068 * information.
8069 */
8070 switch (cpi->cpi_vendor) {
8071 case X86_VENDOR_Intel:
8072 /*
8073 * Check if we now have leaf 7 available to us.
8074 */
8075 if (cpi->cpi_maxeax < 7) {
8076 bzero(&cp, sizeof (cp));
8077 cp.cp_eax = 0;
8078 cpi->cpi_maxeax = __cpuid_insn(&cp);
8079 if (cpi->cpi_maxeax < 7)
8080 break;
8081 }
8082
8083 bzero(&cp, sizeof (cp));
8084 cp.cp_eax = 7;
8085 cp.cp_ecx = 0;
8086 (void) __cpuid_insn(&cp);
8087 cpi->cpi_std[7] = cp;
8088 break;
8089
8090 case X86_VENDOR_AMD:
8091 case X86_VENDOR_HYGON:
8092 /* No xcpuid support */
8093 if (cpi->cpi_family < 5 ||
8094 (cpi->cpi_family == 5 && cpi->cpi_model < 1))
8095 break;
8096
8097 if (cpi->cpi_xmaxeax < CPUID_LEAF_EXT_8) {
8098 bzero(&cp, sizeof (cp));
8099 cp.cp_eax = CPUID_LEAF_EXT_0;
8100 cpi->cpi_xmaxeax = __cpuid_insn(&cp);
8101 if (cpi->cpi_xmaxeax < CPUID_LEAF_EXT_8)
8102 break;
8103 }
8104
8105 /*
8106 * Most AMD features are in leaf 8. Automatic IBRS was added in
8107 * leaf 0x21. So we also check that.
8108 */
8109 bzero(&cp, sizeof (cp));
8110 cp.cp_eax = CPUID_LEAF_EXT_8;
8111 (void) __cpuid_insn(&cp);
8112 platform_cpuid_mangle(cpi->cpi_vendor, CPUID_LEAF_EXT_8, &cp);
8113 cpi->cpi_extd[8] = cp;
8114
8115 if (cpi->cpi_xmaxeax < CPUID_LEAF_EXT_21)
8116 break;
8117
8118 bzero(&cp, sizeof (cp));
8119 cp.cp_eax = CPUID_LEAF_EXT_21;
8120 (void) __cpuid_insn(&cp);
8121 platform_cpuid_mangle(cpi->cpi_vendor, CPUID_LEAF_EXT_21, &cp);
8122 cpi->cpi_extd[0x21] = cp;
8123 break;
8124
8125 default:
8126 /*
8127 * Nothing to do here. Return an empty set which has already
8128 * been zeroed for us.
8129 */
8130 return;
8131 }
8132
8133 cpuid_scan_security(cpu, fset);
8134 }
8135
8136 /* ARGSUSED */
8137 static int
cpuid_post_ucodeadm_xc(xc_arg_t arg0,xc_arg_t arg1,xc_arg_t arg2)8138 cpuid_post_ucodeadm_xc(xc_arg_t arg0, xc_arg_t arg1, xc_arg_t arg2)
8139 {
8140 uchar_t *fset;
8141 boolean_t first_pass = (boolean_t)arg1;
8142
8143 fset = (uchar_t *)(arg0 + sizeof (x86_featureset) * CPU->cpu_id);
8144 if (first_pass && CPU->cpu_id != 0)
8145 return (0);
8146 if (!first_pass && CPU->cpu_id == 0)
8147 return (0);
8148 cpuid_pass_ucode(CPU, fset);
8149
8150 return (0);
8151 }
8152
8153 /*
8154 * After a microcode update where the version has changed, then we need to
8155 * rescan CPUID. To do this we check every CPU to make sure that they have the
8156 * same microcode. Then we perform a cross call to all such CPUs. It's the
8157 * caller's job to make sure that no one else can end up doing an update while
8158 * this is going on.
8159 *
8160 * We assume that the system is microcode capable if we're called.
8161 */
8162 void
cpuid_post_ucodeadm(void)8163 cpuid_post_ucodeadm(void)
8164 {
8165 uint32_t rev;
8166 int i;
8167 struct cpu *cpu;
8168 cpuset_t cpuset;
8169 void *argdata;
8170 uchar_t *f0;
8171
8172 argdata = kmem_zalloc(sizeof (x86_featureset) * NCPU, KM_SLEEP);
8173
8174 mutex_enter(&cpu_lock);
8175 cpu = cpu_get(0);
8176 rev = cpu->cpu_m.mcpu_ucode_info->cui_rev;
8177 CPUSET_ONLY(cpuset, 0);
8178 for (i = 1; i < max_ncpus; i++) {
8179 if ((cpu = cpu_get(i)) == NULL)
8180 continue;
8181
8182 if (cpu->cpu_m.mcpu_ucode_info->cui_rev != rev) {
8183 panic("post microcode update CPU %d has differing "
8184 "microcode revision (%u) from CPU 0 (%u)",
8185 i, cpu->cpu_m.mcpu_ucode_info->cui_rev, rev);
8186 }
8187 CPUSET_ADD(cpuset, i);
8188 }
8189
8190 /*
8191 * We do the cross calls in two passes. The first pass is only for the
8192 * boot CPU. The second pass is for all of the other CPUs. This allows
8193 * the boot CPU to go through and change behavior related to patching or
8194 * whether or not Enhanced IBRS needs to be enabled and then allow all
8195 * other CPUs to follow suit.
8196 */
8197 kpreempt_disable();
8198 xc_sync((xc_arg_t)argdata, B_TRUE, 0, CPUSET2BV(cpuset),
8199 cpuid_post_ucodeadm_xc);
8200 xc_sync((xc_arg_t)argdata, B_FALSE, 0, CPUSET2BV(cpuset),
8201 cpuid_post_ucodeadm_xc);
8202 kpreempt_enable();
8203
8204 /*
8205 * OK, now look at each CPU and see if their feature sets are equal.
8206 */
8207 f0 = argdata;
8208 for (i = 1; i < max_ncpus; i++) {
8209 uchar_t *fset;
8210 if (!CPU_IN_SET(cpuset, i))
8211 continue;
8212
8213 fset = (uchar_t *)((uintptr_t)argdata +
8214 sizeof (x86_featureset) * i);
8215
8216 if (!compare_x86_featureset(f0, fset)) {
8217 panic("Post microcode update CPU %d has "
8218 "differing security feature (%p) set from CPU 0 "
8219 "(%p), not appending to feature set", i,
8220 (void *)fset, (void *)f0);
8221 }
8222 }
8223
8224 mutex_exit(&cpu_lock);
8225
8226 for (i = 0; i < NUM_X86_FEATURES; i++) {
8227 cmn_err(CE_CONT, "?post-ucode x86_feature: %s\n",
8228 x86_feature_names[i]);
8229 if (is_x86_feature(f0, i)) {
8230 add_x86_feature(x86_featureset, i);
8231 }
8232 }
8233 kmem_free(argdata, sizeof (x86_featureset) * NCPU);
8234 }
8235
8236 typedef void (*cpuid_pass_f)(cpu_t *, void *);
8237
8238 typedef struct cpuid_pass_def {
8239 cpuid_pass_t cpd_pass;
8240 cpuid_pass_f cpd_func;
8241 } cpuid_pass_def_t;
8242
8243 /*
8244 * See block comment at the top; note that cpuid_pass_ucode is not a pass in the
8245 * normal sense and should not appear here.
8246 */
8247 static const cpuid_pass_def_t cpuid_pass_defs[] = {
8248 { CPUID_PASS_PRELUDE, cpuid_pass_prelude },
8249 { CPUID_PASS_IDENT, cpuid_pass_ident },
8250 { CPUID_PASS_BASIC, cpuid_pass_basic },
8251 { CPUID_PASS_EXTENDED, cpuid_pass_extended },
8252 { CPUID_PASS_DYNAMIC, cpuid_pass_dynamic },
8253 { CPUID_PASS_RESOLVE, cpuid_pass_resolve },
8254 };
8255
8256 void
cpuid_execpass(cpu_t * cp,cpuid_pass_t pass,void * arg)8257 cpuid_execpass(cpu_t *cp, cpuid_pass_t pass, void *arg)
8258 {
8259 VERIFY3S(pass, !=, CPUID_PASS_NONE);
8260
8261 if (cp == NULL)
8262 cp = CPU;
8263
8264 /*
8265 * Space statically allocated for BSP, ensure pointer is set
8266 */
8267 if (cp->cpu_id == 0 && cp->cpu_m.mcpu_cpi == NULL)
8268 cp->cpu_m.mcpu_cpi = &cpuid_info0;
8269
8270 ASSERT(cpuid_checkpass(cp, pass - 1));
8271
8272 for (uint_t i = 0; i < ARRAY_SIZE(cpuid_pass_defs); i++) {
8273 if (cpuid_pass_defs[i].cpd_pass == pass) {
8274 cpuid_pass_defs[i].cpd_func(cp, arg);
8275 cp->cpu_m.mcpu_cpi->cpi_pass = pass;
8276 return;
8277 }
8278 }
8279
8280 panic("unable to execute invalid cpuid pass %d on cpu%d\n",
8281 pass, cp->cpu_id);
8282 }
8283
8284 /*
8285 * Extract the processor family from a chiprev. Processor families are not the
8286 * same as cpuid families; see comments above and in x86_archext.h.
8287 */
8288 x86_processor_family_t
chiprev_family(const x86_chiprev_t cr)8289 chiprev_family(const x86_chiprev_t cr)
8290 {
8291 return ((x86_processor_family_t)_X86_CHIPREV_FAMILY(cr));
8292 }
8293
8294 /*
8295 * A chiprev matches its template if the vendor and family are identical and the
8296 * revision of the chiprev matches one of the bits set in the template. Callers
8297 * may bitwise-OR together chiprevs of the same vendor and family to form the
8298 * template, or use the _ANY variant. It is not possible to match chiprevs of
8299 * multiple vendors or processor families with a single call. Note that this
8300 * function operates on processor families, not cpuid families.
8301 */
8302 boolean_t
chiprev_matches(const x86_chiprev_t cr,const x86_chiprev_t template)8303 chiprev_matches(const x86_chiprev_t cr, const x86_chiprev_t template)
8304 {
8305 return (_X86_CHIPREV_VENDOR(cr) == _X86_CHIPREV_VENDOR(template) &&
8306 _X86_CHIPREV_FAMILY(cr) == _X86_CHIPREV_FAMILY(template) &&
8307 (_X86_CHIPREV_REV(cr) & _X86_CHIPREV_REV(template)) != 0);
8308 }
8309
8310 /*
8311 * A chiprev is at least min if the vendor and family are identical and the
8312 * revision of the chiprev is at least as recent as that of min. Processor
8313 * families are considered unordered and cannot be compared using this function.
8314 * Note that this function operates on processor families, not cpuid families.
8315 * Use of the _ANY chiprev variant with this function is not useful; it will
8316 * always return B_FALSE if the _ANY variant is supplied as the minimum
8317 * revision. To determine only whether a chiprev is of a given processor
8318 * family, test the return value of chiprev_family() instead.
8319 */
8320 boolean_t
chiprev_at_least(const x86_chiprev_t cr,const x86_chiprev_t min)8321 chiprev_at_least(const x86_chiprev_t cr, const x86_chiprev_t min)
8322 {
8323 return (_X86_CHIPREV_VENDOR(cr) == _X86_CHIPREV_VENDOR(min) &&
8324 _X86_CHIPREV_FAMILY(cr) == _X86_CHIPREV_FAMILY(min) &&
8325 _X86_CHIPREV_REV(cr) >= _X86_CHIPREV_REV(min));
8326 }
8327
8328 /*
8329 * The uarch functions operate in a manner similar to the chiprev functions
8330 * above. While it is tempting to allow these to operate on microarchitectures
8331 * produced by a specific vendor in an ordered fashion (e.g., ZEN3 is "newer"
8332 * than ZEN2), we elect not to do so because a manufacturer may supply
8333 * processors of multiple different microarchitecture families each of which may
8334 * be internally ordered but unordered with respect to those of other families.
8335 */
8336 x86_uarch_t
uarchrev_uarch(const x86_uarchrev_t ur)8337 uarchrev_uarch(const x86_uarchrev_t ur)
8338 {
8339 return ((x86_uarch_t)_X86_UARCHREV_UARCH(ur));
8340 }
8341
8342 boolean_t
uarchrev_matches(const x86_uarchrev_t ur,const x86_uarchrev_t template)8343 uarchrev_matches(const x86_uarchrev_t ur, const x86_uarchrev_t template)
8344 {
8345 return (_X86_UARCHREV_VENDOR(ur) == _X86_UARCHREV_VENDOR(template) &&
8346 _X86_UARCHREV_UARCH(ur) == _X86_UARCHREV_UARCH(template) &&
8347 (_X86_UARCHREV_REV(ur) & _X86_UARCHREV_REV(template)) != 0);
8348 }
8349
8350 boolean_t
uarchrev_at_least(const x86_uarchrev_t ur,const x86_uarchrev_t min)8351 uarchrev_at_least(const x86_uarchrev_t ur, const x86_uarchrev_t min)
8352 {
8353 return (_X86_UARCHREV_VENDOR(ur) == _X86_UARCHREV_VENDOR(min) &&
8354 _X86_UARCHREV_UARCH(ur) == _X86_UARCHREV_UARCH(min) &&
8355 _X86_UARCHREV_REV(ur) >= _X86_UARCHREV_REV(min));
8356 }
8357
8358 /*
8359 * Topology cache related information. This is yet another cache interface that
8360 * we're exposing out intended to be used when we have either Intel Leaf 4 or
8361 * AMD Leaf 8x1D (introduced with Zen 1).
8362 */
8363 static boolean_t
cpuid_cache_topo_sup(const struct cpuid_info * cpi)8364 cpuid_cache_topo_sup(const struct cpuid_info *cpi)
8365 {
8366 switch (cpi->cpi_vendor) {
8367 case X86_VENDOR_Intel:
8368 if (cpi->cpi_maxeax >= 4) {
8369 return (B_TRUE);
8370 }
8371 break;
8372 case X86_VENDOR_AMD:
8373 case X86_VENDOR_HYGON:
8374 if (cpi->cpi_xmaxeax >= CPUID_LEAF_EXT_1d &&
8375 is_x86_feature(x86_featureset, X86FSET_TOPOEXT)) {
8376 return (B_TRUE);
8377 }
8378 break;
8379 default:
8380 break;
8381 }
8382
8383 return (B_FALSE);
8384 }
8385
8386 int
cpuid_getncaches(struct cpu * cpu,uint32_t * ncache)8387 cpuid_getncaches(struct cpu *cpu, uint32_t *ncache)
8388 {
8389 const struct cpuid_info *cpi;
8390
8391 ASSERT(cpuid_checkpass(cpu, CPUID_PASS_DYNAMIC));
8392 cpi = cpu->cpu_m.mcpu_cpi;
8393
8394 if (!cpuid_cache_topo_sup(cpi)) {
8395 return (ENOTSUP);
8396 }
8397
8398 *ncache = cpi->cpi_cache_leaf_size;
8399 return (0);
8400 }
8401
8402 int
cpuid_getcache(struct cpu * cpu,uint32_t cno,x86_cache_t * cache)8403 cpuid_getcache(struct cpu *cpu, uint32_t cno, x86_cache_t *cache)
8404 {
8405 const struct cpuid_info *cpi;
8406 const struct cpuid_regs *cp;
8407
8408 ASSERT(cpuid_checkpass(cpu, CPUID_PASS_DYNAMIC));
8409 cpi = cpu->cpu_m.mcpu_cpi;
8410
8411 if (!cpuid_cache_topo_sup(cpi)) {
8412 return (ENOTSUP);
8413 }
8414
8415 if (cno >= cpi->cpi_cache_leaf_size) {
8416 return (EINVAL);
8417 }
8418
8419 bzero(cache, sizeof (cache));
8420 cp = cpi->cpi_cache_leaves[cno];
8421 switch (CPI_CACHE_TYPE(cp)) {
8422 case CPI_CACHE_TYPE_DATA:
8423 cache->xc_type = X86_CACHE_TYPE_DATA;
8424 break;
8425 case CPI_CACHE_TYPE_INSTR:
8426 cache->xc_type = X86_CACHE_TYPE_INST;
8427 break;
8428 case CPI_CACHE_TYPE_UNIFIED:
8429 cache->xc_type = X86_CACHE_TYPE_UNIFIED;
8430 break;
8431 case CPI_CACHE_TYPE_DONE:
8432 default:
8433 return (EINVAL);
8434 }
8435 cache->xc_level = CPI_CACHE_LVL(cp);
8436 if (CPI_FULL_ASSOC_CACHE(cp) != 0) {
8437 cache->xc_flags |= X86_CACHE_F_FULL_ASSOC;
8438 }
8439 cache->xc_nparts = CPI_CACHE_PARTS(cp) + 1;
8440 /*
8441 * The number of sets is reserved on AMD if the CPU is tagged as fully
8442 * associative, where as it is considered valid on Intel.
8443 */
8444 if (cpi->cpi_vendor == X86_VENDOR_AMD &&
8445 CPI_FULL_ASSOC_CACHE(cp) != 0) {
8446 cache->xc_nsets = 1;
8447 } else {
8448 cache->xc_nsets = CPI_CACHE_SETS(cp) + 1;
8449 }
8450 cache->xc_nways = CPI_CACHE_WAYS(cp) + 1;
8451 cache->xc_line_size = CPI_CACHE_COH_LN_SZ(cp) + 1;
8452 cache->xc_size = cache->xc_nparts * cache->xc_nsets * cache->xc_nways *
8453 cache->xc_line_size;
8454 /*
8455 * We're looking for the number of bits to cover the number of CPUs that
8456 * are being shared. Normally this would be the value - 1, but the CPUID
8457 * value is encoded as the actual value minus one, so we don't modify
8458 * this at all.
8459 */
8460 cache->xc_apic_shift = highbit(CPI_NTHR_SHR_CACHE(cp));
8461
8462 /*
8463 * To construct a unique ID we construct a uint64_t that looks as
8464 * follows:
8465 *
8466 * [47:40] cache level
8467 * [39:32] CPUID cache type
8468 * [31:00] shifted APIC ID
8469 *
8470 * The shifted APIC ID gives us a guarantee that a given cache entry is
8471 * unique within its peers. The other two numbers give us something that
8472 * ensures that something is unique within the CPU. If we just had the
8473 * APIC ID shifted over by the indicated number of bits we'd end up with
8474 * an ID of zero for the L1I, L1D, L2, and L3.
8475 *
8476 * The format of this ID is private to the system and can change across
8477 * a reboot for the time being.
8478 */
8479 cache->xc_id = (uint64_t)cache->xc_level << 40;
8480 cache->xc_id |= (uint64_t)cache->xc_type << 32;
8481 cache->xc_id |= (uint64_t)cpi->cpi_apicid >> cache->xc_apic_shift;
8482
8483 return (0);
8484 }
8485