xref: /illumos-gate/usr/src/uts/intel/os/fpu.c (revision 1f606c5b8e527ed924f5cbdbbce612887c7dbe32)
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) 1992, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright 2021 Joyent, Inc.
24  * Copyright 2021 RackTop Systems, Inc.
25  * Copyright 2023 Oxide Computer Company
26  */
27 
28 /*	Copyright (c) 1990, 1991 UNIX System Laboratories, Inc. */
29 /*	Copyright (c) 1984, 1986, 1987, 1988, 1989, 1990 AT&T   */
30 /*		All Rights Reserved				*/
31 
32 /*	Copyright (c) 1987, 1988 Microsoft Corporation		*/
33 /*		All Rights Reserved				*/
34 
35 /*
36  * Copyright (c) 2009, Intel Corporation.
37  * All rights reserved.
38  */
39 
40 #include <sys/types.h>
41 #include <sys/param.h>
42 #include <sys/signal.h>
43 #include <sys/regset.h>
44 #include <sys/privregs.h>
45 #include <sys/psw.h>
46 #include <sys/trap.h>
47 #include <sys/fault.h>
48 #include <sys/systm.h>
49 #include <sys/user.h>
50 #include <sys/file.h>
51 #include <sys/proc.h>
52 #include <sys/pcb.h>
53 #include <sys/lwp.h>
54 #include <sys/cpuvar.h>
55 #include <sys/thread.h>
56 #include <sys/disp.h>
57 #include <sys/fp.h>
58 #include <sys/siginfo.h>
59 #include <sys/archsystm.h>
60 #include <sys/kmem.h>
61 #include <sys/debug.h>
62 #include <sys/x86_archext.h>
63 #include <sys/sysmacros.h>
64 #include <sys/cmn_err.h>
65 #include <sys/kfpu.h>
66 #include <sys/stdbool.h>
67 #include <sys/stdalign.h>
68 #include <sys/procfs_isa.h>
69 #include <sys/sunddi.h>
70 
71 /*
72  * FPU Management Overview
73  * -----------------------
74  *
75  * The x86 FPU has evolved substantially since its days as the x87 coprocessor;
76  * however, many aspects of its life as a coprocessor are still around in x86.
77  *
78  * Today, when we refer to the 'FPU', we don't just mean the original x87 FPU.
79  * While that state still exists, there is much more that is covered by the FPU.
80  * Today, this includes not just traditional FPU state, but also supervisor only
81  * state. The following state is currently managed and covered logically by the
82  * idea of the FPU registers and more generally is called the Extended Processor
83  * States:
84  *
85  *    o Traditional x87 FPU
86  *    o Vector Registers (%xmm, %ymm, %zmm)
87  *    o Memory Protection Extensions (MPX) Bounds Registers
88  *    o Protected Key Rights Registers (PKRU)
89  *    o Processor Trace data
90  *    o Control-Flow Enforcement state
91  *    o Hardware Duty Cycle
92  *    o Hardware P-states
93  *
94  * The rest of this covers how the FPU is managed and controlled, how state is
95  * saved and restored between threads, interactions with hypervisors, and other
96  * information exported to userland through aux vectors. A lot of background
97  * information is here to synthesize major parts of the Intel SDM, but
98  * unfortunately, it is not a replacement for reading it.
99  *
100  * FPU Control Registers
101  * ---------------------
102  *
103  * Because the x87 FPU began its life as a co-processor and the FPU was
104  * optional there are several bits that show up in %cr0 that we have to
105  * manipulate when dealing with the FPU. These are:
106  *
107  *   o CR0.ET	The 'extension type' bit. This was used originally to indicate
108  *		that the FPU co-processor was present. Now it is forced on for
109  *		compatibility. This is often used to verify whether or not the
110  *		FPU is present.
111  *
112  *   o CR0.NE	The 'native error' bit. Used to indicate that native error
113  *		mode should be enabled. This indicates that we should take traps
114  *		on FPU errors. The OS enables this early in boot.
115  *
116  *   o CR0.MP	The 'Monitor Coprocessor' bit. Used to control whether or not
117  *		wait/fwait instructions generate a #NM if CR0.TS is set.
118  *
119  *   o CR0.EM	The 'Emulation' bit. This is used to cause floating point
120  *		operations (x87 through SSE4) to trap with a #UD so they can be
121  *		emulated. The system never sets this bit, but makes sure it is
122  *		clear on processor start up.
123  *
124  *   o CR0.TS	The 'Task Switched' bit. When this is turned on, a floating
125  *		point operation will generate a #NM. An fwait will as well,
126  *		depending on the value in CR0.MP.
127  *
128  * Our general policy is that CR0.ET, CR0.NE, and CR0.MP are always set by
129  * the system. Similarly CR0.EM is always unset by the system. CR0.TS has a more
130  * complicated role. Historically it has been used to allow running systems to
131  * restore the FPU registers lazily. This will be discussed in greater depth
132  * later on.
133  *
134  * %cr4 is also used as part of the FPU control. Specifically we need to worry
135  * about the following bits in the system:
136  *
137  *   o CR4.OSFXSR	This bit is used to indicate that the OS understands and
138  *			supports the execution of the fxsave and fxrstor
139  *			instructions. This bit is required to be set to enable
140  *			the use of the SSE->SSE4 instructions.
141  *
142  *   o CR4.OSXMMEXCPT	This bit is used to indicate that the OS can understand
143  *			and take a SIMD floating point exception (#XM). This bit
144  *			is always enabled by the system.
145  *
146  *   o CR4.OSXSAVE	This bit is used to indicate that the OS understands and
147  *			supports the execution of the xsave and xrstor family of
148  *			instructions. This bit is required to use any of the AVX
149  *			and newer feature sets.
150  *
151  * Because all supported processors are 64-bit, they'll always support the XMM
152  * extensions and we will enable both CR4.OXFXSR and CR4.OSXMMEXCPT in boot.
153  * CR4.OSXSAVE will be enabled and used whenever xsave is reported in cpuid.
154  *
155  * %xcr0 is used to manage the behavior of the xsave feature set and is only
156  * present on the system if xsave is supported. %xcr0 is read and written to
157  * through by the xgetbv and xsetbv instructions. This register is present
158  * whenever the xsave feature set is supported. Each bit in %xcr0 refers to a
159  * different component of the xsave state and controls whether or not that
160  * information is saved and restored. For newer feature sets like AVX and MPX,
161  * it also controls whether or not the corresponding instructions can be
162  * executed (much like CR0.OSFXSR does for the SSE feature sets).
163  *
164  * Everything in %xcr0 is around features available to users. There is also the
165  * IA32_XSS MSR which is used to control supervisor-only features that are still
166  * part of the xsave state. Bits that can be set in %xcr0 are reserved in
167  * IA32_XSS and vice versa. This is an important property that is particularly
168  * relevant to how the xsave instructions operate.
169  *
170  * Save Mechanisms
171  * ---------------
172  *
173  * When switching between running threads the FPU state needs to be saved and
174  * restored by the OS. If this state was not saved, users would rightfully
175  * complain about corrupt state. There are three mechanisms that exist on the
176  * processor for saving and restoring these state images:
177  *
178  *   o fsave
179  *   o fxsave
180  *   o xsave
181  *
182  * fsave saves and restores only the x87 FPU and is the oldest of these
183  * mechanisms. This mechanism is never used in the kernel today because we are
184  * always running on systems that support fxsave.
185  *
186  * The fxsave and fxrstor mechanism allows the x87 FPU and the SSE register
187  * state to be saved and restored to and from a struct fxsave_state. This is the
188  * default mechanism that is used to save and restore the FPU on amd64. An
189  * important aspect of fxsave that was different from the original i386 fsave
190  * mechanism is that the restoring of FPU state with pending exceptions will not
191  * generate an exception, it will be deferred to the next use of the FPU.
192  *
193  * The final and by far the most complex mechanism is that of the xsave set.
194  * xsave allows for saving and restoring all of the traditional x86 pieces (x87
195  * and SSE), while allowing for extensions that will save the %ymm, %zmm, etc.
196  * registers.
197  *
198  * Data is saved and restored into and out of a struct xsave_state. The first
199  * part of the struct xsave_state is equivalent to the struct fxsave_state.
200  * After that, there is a header which is used to describe the remaining
201  * portions of the state. The header is a 64-byte value of which the first two
202  * uint64_t values are defined and the rest are reserved and must be zero. The
203  * first uint64_t is the xstate_bv member. This describes which values in the
204  * xsave_state are actually valid and present. This is updated on a save and
205  * used on restore. The second member is the xcomp_bv member. Its last bit
206  * determines whether or not a compressed version of the structure is used.
207  *
208  * When the uncompressed structure is used (currently the only format we
209  * support), then each state component is at a fixed offset in the structure,
210  * even if it is not being used. For example, if you only saved the AVX related
211  * state, but did not save the MPX related state, the offset would not change
212  * for any component. With the compressed format, components that aren't used
213  * are all elided (though the x87 and SSE state are always there).
214  *
215  * Unlike fxsave which saves all state, the xsave family does not always save
216  * and restore all the state that could be covered by the xsave_state. The
217  * instructions all take an argument which is a mask of what to consider. This
218  * is the same mask that will be used in the xstate_bv vector and it is also the
219  * same values that are present in %xcr0 and IA32_XSS. Though IA32_XSS is only
220  * considered with the xsaves and xrstors instructions.
221  *
222  * When a save or restore is requested, a bitwise and is performed between the
223  * requested bits and those that have been enabled in %xcr0. Only the bits that
224  * match that are then saved or restored. Others will be silently ignored by
225  * the processor. This idea is used often in the OS. We will always request that
226  * we save and restore all of the state, but only those portions that are
227  * actually enabled in %xcr0 will be touched.
228  *
229  * If a feature has been asked to be restored that is not set in the xstate_bv
230  * feature vector of the save state, then it will be set to its initial state by
231  * the processor (usually zeros). Also, when asked to save state, the processor
232  * may not write out data that is in its initial state as an optimization. This
233  * optimization only applies to saving data and not to restoring data.
234  *
235  * There are a few different variants of the xsave and xrstor instruction. They
236  * are:
237  *
238  *   o xsave	This is the original save instruction. It will save all of the
239  *		requested data in the xsave state structure. It only saves data
240  *		in the uncompressed (xcomp_bv[63] is zero) format. It may be
241  *		executed at all privilege levels.
242  *
243  *   o xrstor	This is the original restore instruction. It will restore all of
244  *		the requested data. The xrstor function can handle both the
245  *		compressed and uncompressed formats. It may be executed at all
246  *		privilege levels.
247  *
248  *   o xsaveopt	This is a variant of the xsave instruction that employs
249  *		optimizations to try and only write out state that has been
250  *		modified since the last time an xrstor instruction was called.
251  *		The processor tracks a tuple of information about the last
252  *		xrstor and tries to ensure that the same buffer is being used
253  *		when this optimization is being used. However, because of the
254  *		way that it tracks the xrstor buffer based on the address of it,
255  *		it is not suitable for use if that buffer can be easily reused.
256  *		The most common case is trying to save data to the stack in
257  *		rtld. It may be executed at all privilege levels.
258  *
259  *   o xsavec	This is a variant of the xsave instruction that writes out the
260  *		compressed form of the xsave_state. Otherwise it behaves as
261  *		xsave. It may be executed at all privilege levels.
262  *
263  *   o xsaves	This is a variant of the xsave instruction. It is similar to
264  *		xsavec in that it always writes the compressed form of the
265  *		buffer. Unlike all the other forms, this instruction looks at
266  *		both the user (%xcr0) and supervisor (IA32_XSS MSR) to determine
267  *		what to save and restore. xsaves also implements the same
268  *		optimization that xsaveopt does around modified pieces. User
269  *		land may not execute the instruction.
270  *
271  *   o xrstors	This is a variant of the xrstor instruction. Similar to xsaves
272  *		it can save and restore both the user and privileged states.
273  *		Unlike xrstor it can only operate on the compressed form.
274  *		User land may not execute the instruction.
275  *
276  * Based on all of these, the kernel has a precedence for what it will use.
277  * Basically, xsaves (not supported) is preferred to xsaveopt, which is
278  * preferred to xsave. A similar scheme is used when informing rtld (more later)
279  * about what it should use. xsavec is preferred to xsave. xsaveopt is not
280  * recommended due to the modified optimization not being appropriate for this
281  * use.
282  *
283  * Finally, there is one last gotcha with the xsave state. Importantly some AMD
284  * processors did not always save and restore some of the FPU exception state in
285  * some cases like Intel did. In those cases the OS will make up for this fact
286  * itself.
287  *
288  * FPU Initialization
289  * ------------------
290  *
291  * One difference with the FPU registers is that not all threads have FPU state,
292  * only those that have an lwp. Generally this means kernel threads, which all
293  * share p0 and its lwp, do not have FPU state. Though there are definitely
294  * exceptions such as kcfpoold. In the rest of this discussion we'll use thread
295  * and lwp interchangeably, just think of thread meaning a thread that has a
296  * lwp.
297  *
298  * Each lwp has its FPU state allocated in its pcb (process control block). The
299  * actual storage comes from the fpsave_cachep kmem cache. This cache is sized
300  * dynamically at start up based on the save mechanism that we're using and the
301  * amount of memory required for it. This is dynamic because the xsave_state
302  * size varies based on the supported feature set.
303  *
304  * The hardware side of the FPU is initialized early in boot before we mount the
305  * root file system. This is effectively done in fpu_probe(). This is where we
306  * make the final decision about what the save and restore mechanisms we should
307  * use are, create the fpsave_cachep kmem cache, and initialize a number of
308  * function pointers that use save and restoring logic.
309  *
310  * The thread/lwp side is a a little more involved. There are two different
311  * things that we need to concern ourselves with. The first is how the FPU
312  * resources are allocated and the second is how the FPU state is initialized
313  * for a given lwp.
314  *
315  * We allocate the FPU save state from our kmem cache as part of lwp_fp_init().
316  * This is always called unconditionally by the system as part of creating an
317  * LWP.
318  *
319  * There are three different initialization paths that we deal with. The first
320  * is when we are executing a new process. As part of exec all of the register
321  * state is reset. The exec case is particularly important because init is born
322  * like Athena, sprouting from the head of the kernel, without any true parent
323  * to fork from. The second is used whenever we fork or create a new lwp.  The
324  * third is to deal with special lwps like the agent lwp.
325  *
326  * During exec, we will call fp_exec() which will initialize and set up the FPU
327  * state for the process. That will fill in the initial state for the FPU and
328  * also set that state in the FPU itself. As part of fp_exec() we also install a
329  * thread context operations vector that takes care of dealing with the saving
330  * and restoring of the FPU. These context handlers will also be called whenever
331  * an lwp is created or forked. In those cases, to initialize the FPU we will
332  * call fp_new_lwp(). Like fp_exec(), fp_new_lwp() will install a context
333  * operations vector for the new thread.
334  *
335  * Next we'll end up in the context operation fp_new_lwp(). This saves the
336  * current thread's state, initializes the new thread's state, and copies over
337  * the relevant parts of the originating thread's state. It's as this point that
338  * we also install the FPU context operations into the new thread, which ensures
339  * that all future threads that are descendants of the current one get the
340  * thread context operations (unless they call exec).
341  *
342  * To deal with some things like the agent lwp, we double check the state of the
343  * FPU in sys_rtt_common() to make sure that it has been enabled before
344  * returning to userland. In general, this path should be rare, but it's useful
345  * for the odd lwp here and there.
346  *
347  * The FPU state will remain valid most of the time. There are times that
348  * the state will be rewritten. For example in restorecontext, due to /proc, or
349  * the lwp calls exec(). Whether the context is being freed or we are resetting
350  * the state, we will call fp_free() to disable the FPU and our context.
351  *
352  * Finally, when the lwp is destroyed, it will actually destroy and free the FPU
353  * state by calling fp_lwp_cleanup().
354  *
355  * Kernel FPU Multiplexing
356  * -----------------------
357  *
358  * Just as the kernel has to maintain all of the general purpose registers when
359  * switching between scheduled threads, the same is true of the FPU registers.
360  *
361  * When a thread has FPU state, it also has a set of context operations
362  * installed. These context operations take care of making sure that the FPU is
363  * properly saved and restored during a context switch (fpsave_ctxt and
364  * fprestore_ctxt respectively). This means that the current implementation of
365  * the FPU is 'eager', when a thread is running the CPU will have its FPU state
366  * loaded. While this is always true when executing in userland, there are a few
367  * cases where this is not true in the kernel.
368  *
369  * This was not always the case. Traditionally on x86 a 'lazy' FPU restore was
370  * employed. This meant that the FPU would be saved on a context switch and the
371  * CR0.TS bit would be set. When a thread next tried to use the FPU, it would
372  * then take a #NM trap, at which point we would restore the FPU from the save
373  * area and return to userland. Given the frequency of use of the FPU alone by
374  * libc, there's no point returning to userland just to trap again.
375  *
376  * There are a few cases though where the FPU state may need to be changed for a
377  * thread on its behalf. The most notable cases are in the case of processes
378  * using /proc, restorecontext, forking, etc. In all of these cases the kernel
379  * will force a threads FPU state to be saved into the PCB through the fp_save()
380  * function. Whenever the FPU is saved, then the FPU_VALID flag is set on the
381  * pcb. This indicates that the save state holds currently valid data. As a side
382  * effect of this, CR0.TS will be set. To make sure that all of the state is
383  * updated before returning to userland, in these cases, we set a flag on the
384  * PCB that says the FPU needs to be updated. This will make sure that we take
385  * the slow path out of a system call to fix things up for the thread. Due to
386  * the fact that this is a rather rare case, effectively setting the equivalent
387  * of t_postsys is acceptable.
388  *
389  * CR0.TS will be set after a save occurs and cleared when a restore occurs.
390  * Generally this means it will be cleared immediately by the new thread that is
391  * running in a context switch. However, this isn't the case for kernel threads.
392  * They currently operate with CR0.TS set as no kernel state is restored for
393  * them. This means that using the FPU will cause a #NM and panic.
394  *
395  * The FPU_VALID flag on the currently executing thread's pcb is meant to track
396  * what the value of CR0.TS should be. If it is set, then CR0.TS will be set.
397  * However, because we eagerly restore, the only time that CR0.TS should be set
398  * for a non-kernel thread is during operations where it will be cleared before
399  * returning to userland and importantly, the only data that is in it is its
400  * own.
401  *
402  * Kernel FPU Usage
403  * ----------------
404  *
405  * Traditionally the kernel never used the FPU since it had no need for
406  * floating point operations. However, modern FPU hardware supports a variety
407  * of SIMD extensions which can speed up code such as parity calculations or
408  * encryption.
409  *
410  * To allow the kernel to take advantage of these features, the
411  * kernel_fpu_begin() and kernel_fpu_end() functions should be wrapped
412  * around any usage of the FPU by the kernel to ensure that user-level context
413  * is properly saved/restored, as well as to properly setup the FPU for use by
414  * the kernel. There are a variety of ways this wrapping can be used, as
415  * discussed in this section below.
416  *
417  * When kernel_fpu_begin() and kernel_fpu_end() are used for extended
418  * operations, the kernel_fpu_alloc() function should be used to allocate a
419  * kfpu_state_t structure that is used to save/restore the thread's kernel FPU
420  * state. This structure is not tied to any thread. That is, different threads
421  * can reuse the same kfpu_state_t structure, although not concurrently. A
422  * kfpu_state_t structure is freed by the kernel_fpu_free() function.
423  *
424  * In some cases, the kernel may need to use the FPU for a short operation
425  * without the overhead to manage a kfpu_state_t structure and without
426  * allowing for a context switch off the FPU. In this case the KFPU_NO_STATE
427  * bit can be set in the kernel_fpu_begin() and kernel_fpu_end() flags
428  * parameter. This indicates that there is no kfpu_state_t. When used this way,
429  * kernel preemption should be disabled by the caller (kpreempt_disable) before
430  * calling kernel_fpu_begin(), and re-enabled after calling kernel_fpu_end().
431  * For this usage, it is important to limit the kernel's FPU use to short
432  * operations. The tradeoff between using the FPU without a kfpu_state_t
433  * structure vs. the overhead of allowing a context switch while using the FPU
434  * should be carefully considered on a case by case basis.
435  *
436  * In other cases, kernel threads have an LWP, but never execute in user space.
437  * In this situation, the LWP's pcb_fpu area can be used to save/restore the
438  * kernel's FPU state if the thread is context switched, instead of having to
439  * allocate and manage a kfpu_state_t structure. The KFPU_USE_LWP bit in the
440  * kernel_fpu_begin() and kernel_fpu_end() flags parameter is used to
441  * enable this behavior. It is the caller's responsibility to ensure that this
442  * is only used for a kernel thread which never executes in user space.
443  *
444  * FPU Exceptions
445  * --------------
446  *
447  * Certain operations can cause the kernel to take traps due to FPU activity.
448  * Generally these events will cause a user process to receive a SIGFPU and if
449  * the kernel receives it in kernel context, we will die. Traditionally the #NM
450  * (Device Not Available / No Math) exception generated by CR0.TS would have
451  * caused us to restore the FPU. Now it is a fatal event regardless of whether
452  * or not userland causes it.
453  *
454  * While there are some cases where the kernel uses the FPU, it is up to the
455  * kernel to use the FPU in a way such that it cannot receive a trap or to use
456  * the appropriate trap protection mechanisms.
457  *
458  * Hypervisors
459  * -----------
460  *
461  * When providing support for hypervisors things are a little bit more
462  * complicated because the FPU is not virtualized at all. This means that they
463  * need to save and restore the FPU and %xcr0 across entry and exit to the
464  * guest. To facilitate this, we provide a series of APIs in <sys/hma.h>. These
465  * allow us to use the full native state to make sure that we are always saving
466  * and restoring the full FPU that the host sees, even when the guest is using a
467  * subset.
468  *
469  * One tricky aspect of this is that the guest may be using a subset of %xcr0
470  * and therefore changing our %xcr0 on the fly. It is vital that when we're
471  * saving and restoring the FPU that we always use the largest %xcr0 contents
472  * otherwise we will end up leaving behind data in it.
473  *
474  * ELF PLT Support
475  * ---------------
476  *
477  * rtld has to preserve a subset of the FPU when it is saving and restoring
478  * registers due to the amd64 SYS V ABI. See cmd/sgs/rtld/amd64/boot_elf.s for
479  * more information. As a result, we set up an aux vector that contains
480  * information about what save and restore mechanisms it should be using and
481  * the sizing thereof based on what the kernel supports. This is passed down in
482  * a series of aux vectors SUN_AT_FPTYPE and SUN_AT_FPSIZE. This information is
483  * initialized in fpu_subr.c.
484  *
485  * Signal Handling and the ucontext_t
486  * ----------------------------------
487  *
488  * One of the many gifts that signals give us is the twofold fact that when a
489  * signal occurs, the signal handler is allowed to change the CPU's state
490  * arbitrarily and when the signal handler is done executing, we must restore it
491  * back to the original state. However, the second part of this is that the
492  * signal handler is actually allowed to modify the state that the thread will
493  * return to! To create this facade, the kernel will create a full ucontext_t
494  * state, effectively calling getcontext(2) on the thread's behalf, and a
495  * pointer to that is given to the signal handler (the void * argument for the
496  * sa_sigaction function pointer in sigaction(2)). When libc is done with a
497  * signal, it will call setcontext(2) with that same ucontext_t.
498  *
499  * Now, the ucontext_t has a fixed ABI for both ILP32 and LP64 environments and
500  * it's often declared on the stack itself, with the signal handler spilling all
501  * this state to the stack. The ucontext_t machine portion was broken into the
502  * general purpose and floating point registers. In 64-bit code, the floating
503  * point registers were mostly the same as the results of the fxsave instruction
504  * (i.e. struct fxsave_state). While the 64-bit kernel still uses the equivalent
505  * starting point for information, it is transformed into a different shape to
506  * deal with the history of the 32-bit SYS V ABI.
507  *
508  * While this worked, if you're reading this, you're aware that the x86 FPU and
509  * extended register states didn't stop at the initial 16 128-bit %xmm
510  * registers. Since then we have added 256-bit %ymm, 512-bit %zmm, and the %k
511  * opmask registers. None of these fit inside the standard ucontext_t; however,
512  * they must all be preserved and restored across a signal. While the various
513  * x86 platform-specific ABIs all suggest that these registers are not preserved
514  * across a function call, receiving a signal is not a function call and must be
515  * thought of like a process receiving an interrupt. In other words, this
516  * extended state must be preserved.
517  *
518  * To facilitate this, we have extended the ucontext_t structure with an
519  * additional flag, UC_XSAVE, which indicates that the traditional padding
520  * member, uc_xsave, actually is a pointer to the extended state. While this is
521  * accessible outside of a signal handling context through the combination of
522  * ucontext_alloc(3C) and getcontext_extd(2), our design around saving this
523  * state is focused on signal handling. Signal handling spills all this state to
524  * the stack and if we cannot spill the entire state to the stack then our
525  * inability to deliver the signal results in the process being killed! While
526  * there are separate efforts to ensure that the signal stack sizing that is
527  * used for the minimum and maximum signal sizes are sufficient, we still need
528  * to do our part to minimize the likelihood here.
529  *
530  * In designing this, we make the following observations which have helped us
531  * focus our design:
532  *
533  *   o While the start of an xsave area is the traditional 512-byte fxsave XMM
534  *     region, we already have that in the fpregs. Thus there is no reason to
535  *     duplicate it. This not only saves 512 bytes of additional stack space,
536  *     but it also means we don't have to ask which of the version of it to take
537  *     if they were to differ.
538  *
539  *   o Many applications out there aren't necessarily using the extended vectors
540  *     and even when we do make libc and others take advantage of it, it will
541  *     behoove us to ensure that they are put back into their initial state
542  *     after use. This leads us to expect that in a number of cases, the actual
543  *     extended register state will be in its initial state.
544  *
545  *   o While the signal handler does allow contents to be modified, we are
546  *     starting with making the interface private and thus allowing us to excise
547  *     components that are in their initial state.
548  *
549  *   o There are similarities to what we want to create with the compressed
550  *     xsave format; however, because we don't always have support for the
551  *     compressed format, we can't just arbitrarily say let's do a compressed
552  *     save to the user stack.
553  *
554  *   o Because we are not handing this state directly to and from hardware, we
555  *     don't need to meet some of the constraints of the compressed xsave format
556  *     around wanting alignment for the initial save or additional components.
557  *
558  * All of the above lead us to our own unique format for this data. When the
559  * UC_XSAVE flag is set in the ucontext_t, the uc_xsave member points to a
560  * uc_xsave_t structure which has a magic version number, a 32-bit length of the
561  * overall structure, and the 64-bit state bit-vector to represent which
562  * components are valid. Following this 8-byte header, each component that is
563  * present in the bit vector is immediately written out in roughly ascending bit
564  * order (the order is determined based on the order of the fpu_xsave_info
565  * array).
566  *
567  * This makes the rough logic that we have here when taking a signal and writing
568  * out this state as:
569  *
570  *   1. Ensure that the FPU is saved and that the contents of the pcb save area
571  *      are valid. That is, call fp_save() if the state is not already flagged
572  *      with FPU_VALID.
573  *
574  *   2. Copy the bit-vector from the save area and remove the XFEATURE_LEGACY_FP
575  *      and XFEATURE_SSE bits as these will be placed in the xsave area.
576  *
577  *   3. Initialize the uc_xsave_t by setting our version field, initializing the
578  *      length to the length of the current structure, and then setting the
579  *      modified bit vector above.
580  *
581  *   4. Walk each remaining bit of the bit-vector. For each set bit, copy out
582  *      its extended state starting at the current length in the header and then
583  *      increase the header size by that length.
584  *
585  *   5. Finally write out the final uc_xsave_t structure.
586  *
587  * The above process is also used when someone manually calls getcontext_extd(2)
588  * to get this state. The main difference between the two is which copyout
589  * function we use. This deserves some explanation. Our main starting point for
590  * all the logic here is fpu_signal_copyout(). It takes a copyfunc that allows
591  * the signal handling context to operate with a different copyout than we
592  * normally use in say getcontext_extd(2).
593  *
594  * When we've received a signal, we're at the intersection of several different
595  * gotchas. Normal copyout (or ddi_copyout()) will trigger watchpoints. That is,
596  * the watchpoints effectively set a copyout override function (t_copyops) that
597  * we end up vectoring to rather than a normal copyout. This allows the data to
598  * be modified and for the watchpoint to fire. While this is all well and good
599  * normally, it is problematic if we are trying to handle a signal. The signal
600  * deliver logic, sendsig(), goes through and disables the watchpoint for the
601  * region of the stack that we are copying out to. However, disabling
602  * watchpoints is not sufficient, we also need to use the copyout_noerr
603  * variants.
604  *
605  * These variants also require the use of on_fault() and no_fault() for error
606  * handling. While it is tempting to try and on_fault() the entire
607  * fpu_signal_copyout() operation, that is actually fraught for a few reasons.
608  * The first is that we don't want to disable faults during the entire operation
609  * as if the kernel messes up we will treat that as a user error. That isn't
610  * theoretical and happened during development. The second and perhaps more
611  * important issue is that correctly bounding the on_fault() / no_fault() means
612  * being careful about state. For example, kernel pre-emption is often disabled
613  * during parts of these operations, but it needs to be re-enabled when we're
614  * done. This would require tracking in some volatile variable that this had
615  * been enabled and disabled and tracking that.
616  *
617  * Instead, this is why fpu_signal_copyout() takes a copy out function as an
618  * argument. When we're in signal handling context, the function will use
619  * coypout_noerr() and wrap it in the appropriate on_fault() mechanisms.
620  *
621  * RESTORING STATE
622  *
623  * Copying out our current state is the easier half of this problem. When the
624  * kernel is done with a signal it calls setcontext(2) with the ucontext_t we
625  * assembled for it as described above. setcontext(2) isn't just used for
626  * returning from signals.
627  *
628  * The process for this goes in two steps. The first step is to copy in,
629  * validate, and transform the ucontext_t UC_XSAVE that we created above into an
630  * equivalent xsave format that we can use the appropriate xrstor function on.
631  * This first phase is implemented in fpu_signal_copyin(). Once that is done, we
632  * come back through a second phase that is driven out of restorecontext() and
633  * is implemented in fpu_set_xsave().
634  *
635  * Let's start by discussing the second part of this, which is more
636  * straightforward. In particular, the second phase assumes that all of the
637  * validation and error handling has been done by the first phase. This means
638  * here, we have a buffer that is already the appropriate size
639  * (cpuid_get_xsave_size()) and all we need to do is make sure that we can
640  * replace the actual save state with the current one.
641  *
642  * The only piece of shenanigans we have to do is around the kernel provided
643  * notion of 'status' and 'xstatus', which are cached versions of the x87 and
644  * SSE exception vectors. These are part of the fpregset ABI and therefore we
645  * need to propagate them from the temporary storage that part 1 sets up in the
646  * ignored region of the fxsave data. We use that because it is not persisted by
647  * the CPU, so clobbering it is generally alright.
648  *
649  * Once that is done, we simply note that we need a PCB update to occur to
650  * refresh the FPU state before we return to userland. Given that someone has
651  * called setcontext(2), this was always going to happen because we have to
652  * update segment registers and related, so this isn't so bad. With that, let's
653  * move onto the more nuanced part (1).
654  *
655  * When we're handling a setcontext(2) we have, in userland, a data structure
656  * that should match one we serialized out, though we cannot assume that a user
657  * has not modified it either accidentally or maliciously. Our goal is to set up
658  * the appropriate xsave state that can be passed to the CPU's xrstor. The first
659  * problem we have to deal with is where do we actually put this state?
660  *
661  * While not many programs actually call setcontext(2) of their own volition,
662  * this is going to get hit every time we take a signal. The first thought was
663  * to re-use the existing thread's save area; however, that's a bit challenging
664  * for a few reasons. In particular, we would need to ensure that we don't go
665  * off-CPU for any reason, which we cannot assume with a copyin from a user
666  * address space. In particular, it is trivial for us to hit a case where the
667  * stack has been paged out for some reason, which eschews that path.
668  *
669  * Instead, whenever a thread first calls setcontext(2), generally from signal
670  * context, we will at that time allocate another entry from the 'fpsave_cachep'
671  * kmem cache, giving us a buffer of the appropriate space to handle this. Once
672  * this buffer has been allocated, we leave it assigned to the thread's pcb and
673  * only tear it down when the thread itself finally exits. We reason that a
674  * thread that takes a signal once is either going to have the process exit
675  * shortly thereafter or is much more likely to take a signal again in the
676  * future. Many daemons and other processes set things up so signals are
677  * dispatched via one location, masking signals in other thread, using
678  * sigsuspend(2), signalfd(3C), or something similar.
679  *
680  * With this buffer in hand, we begin our task of reassembling state. Note, all
681  * of this is conditional on UC_XSAVE being set in the uc_flags member of the
682  * ucontext_t. If it is not set, then we assume that there is no extended state
683  * and will use the traditional path of setting the fpregset_t into the system
684  * via setfpregs().
685  *
686  * We first will copyin and validate the uc_xsave_t. In particular, we need to
687  * make sure the version makes sense, that the xsave component bit-vector
688  * doesn't have anything unexpected and more importantly unsupported in it, and
689  * that the addresses we've been given are within the user address space. At
690  * this point we can walk through our table of implemented bits and process
691  * them.
692  *
693  * For most components in here, the processing is straightforward. We continue
694  * walking our cursor and copy data into the kernel and place it in the
695  * appropriate place in our xsave state. If a xsave state component bit-vector
696  * isn't set, then we must ensure that we have the item in the initial state,
697  * which for everything other than the x87/SSE state is the memory being zeroed.
698  *
699  * The most unique case in the copyin state is that of the x87/SSE state. You
700  * might recall that we didn't copy it out explicitly as part of the uc_xsave_t,
701  * but instead have opted to use the single definition in the fpregset_t. Thus
702  * here, we copy it out of the fpregset_t, which the kernel has helpfully
703  * already unified into the 64-bit fxsave version prior to calling us, and
704  * install that into the save area we're building up.
705  *
706  * As part of this, there are two important pieces to be aware of. The first is
707  * that because the fpregset_t has both the status and xstatus members
708  * mentioned earlier, we temporarily copy them to the software-usable ignored
709  * areas of the fxsave state so we can corral this extra state into part (2)
710  * without needing to allocate additional space. The second piece is that when
711  * we're done processing this we explicitly remove the UC_FPU flag that would
712  * tell the kernel to proceed with updating that region. The problem is that
713  * that goes directly into the pcb's save area and not to the intermediate
714  * buffer as it uses the same entry point as /proc, mainly setfpregs().
715  *
716  * We don't do much validation of the actual contents of the registers that are
717  * being set with the exception of ensuring that no reserved bits of the mxcsr
718  * are used. This is not as strict as /proc, but failure here means the process
719  * is likely going to die (returning from setcontext() in a signal handler is
720  * fatal).
721  *
722  * /proc xregs
723  * -----------
724  *
725  * Observability of the state of the extended registers is important for
726  * understanding the system. While on the surface this is similar to signal
727  * handling, it is crucially different in a number of ways:
728  *
729  *   o In signal handling, we're trying to conserve every byte of stack that we
730  *     can.
731  *   o The /proc xregs file will end up in core files, which means that we need
732  *     a way of knowing what components are present and not present in it,
733  *     because this will vary from CPU to CPU due to the addition of
734  *     architectural features. For example, some CPUs support AVX-512, but
735  *     others do not.
736  *
737  *   o The signal handling structure (uc_xsave_t) is private and we're not
738  *     trying to have software modify it, on the other hand, the /proc
739  *     interfaces that we support we do want software to be able to interrogate
740  *     and manipulate. These need to be something that we can introduce
741  *     additional components into and make other changes that still allow it to
742  *     work.
743  *
744  * The x86 xregs format is documented in proc(5). The short form is that the
745  * prxregset_hdr_t has a number of information entries, which are of the type
746  * prxregset_info_t. Each of the information headers has a type, size, and
747  * offset which indicate where to find the additional data.
748  *
749  * Each entry is described as one of the entries in the fpu_xsave_info[]. These
750  * items either are a 1:1 correspondence with a xsave related feature (e.g.
751  * there is one entry for each of the three AVX-512 components) or it is
752  * something synthetic that we provide as additional information such as the
753  * PRX_INFO_XCR, which is a way of getting information about the system such as
754  * what is enabled in %xcr0 out there.
755  *
756  * Unlike signal handling, we are given the buffer to place everything that
757  * needs to be written out. This is partially the design of the /proc APIs. That
758  * is, we will always assemble everything into the entire buffer that /proc asks
759  * us to, and then it will use as much or as little of it as is required.
760  * Similarly, when setting things, we don't have to worry about copying in
761  * information in the same way as signal handling does, because /proc takes care
762  * of it and always hands us a full buffer. Sizing that is a little nuanced, but
763  * is all handled in prmachdep.c.
764  *
765  * When someone performs a read of the xregs and thus is asking us for the
766  * current state, there is a little bit of nuance that we need to deal with.
767  * The first, is whether or not the FPU is enabled and the second is if the FPU
768  * is enabled, whether a given component is noted as being in its initial state.
769  * This basically gives us three possible states for a given component:
770  *
771  *   1. FPU_EN is not set and FPU_VALID is not set. This means we need to take
772  *      the illumos FPU default for an item. More on that in a moment.
773  *   2. The saved xsave state indicates that the bit for a given component is
774  *      zero -- specifically the xsh_xstate_bv member of the struct xsave_state.
775  *      In this case, we must take the CPU's default for an item. This is
776  *      usually the same as illumos, but not always.
777  *   3. The saved xsave state indicates that a given component's state bit is
778  *      valid. The simplest of our cases. We can just take what we have from the
779  *      xsave state.
780  *
781  * The CPU's default state for most components other than the x87/SSE state is
782  * to have it be zeroed. This is what we treat as our default state as well. The
783  * primary difference is in the initialization of the x87/SSE state. The SYS V
784  * ABI requires that we enable a different floating point control word then the
785  * hardware default. This means that when we're dealing with case (1) for
786  * x87/SSE we have to be more careful than the other components. Thankfully for
787  * everything else this is just keeping it zeroed.
788  *
789  * A reasonable question would be why not just skip components that aren't
790  * marked as present. There are a few reasons we take a different approach and
791  * always include them. Both of these are to make lives simpler for consumers.
792  * In the first case, when someone is performing a read and wants to reassemble
793  * and answer the question of 'what is the value of %ymm0 or %zmm15', they have
794  * to combine multiple disparate parts. If one knows that the data we put into
795  * there is always valid and represents what is in hardware and doesn't have to
796  * keep track of what are the defaults in different circumstances, then that
797  * greatly simplifies consumers lives. It also helps us for core files and other
798  * observability cases because the answer to what is the operating system's
799  * default may change over time.
800  *
801  * Similarly, including all the possible structures means that we have
802  * simplified writes. Writes are always setting the full state of a thread,
803  * meaning that if someone wants to modify only a single register they must do a
804  * read, modify, and write. By including everything that they might need, it
805  * makes it easier for consumers to do this and not have to cons up the whole
806  * structure on their own.
807  *
808  * When we're setting state, things change around a little bit. We have a few
809  * constraints that are laid out in proc(5). In particular, we require that the
810  * PRX_INFO_XSAVE component always be present to tell us which other components
811  * we expect to be here and which ones we don't. We also are much stricter about
812  * writes in several ways. Of all the components, the PRX_INFO_XCR is read-only
813  * and may not be modified by a calling process. In addition, when we have
814  * 32-bit applications which have reserved registers in the %ymm, %zmm, etc.
815  * components, if they are being written to and have modifications, then we will
816  * indicate an error there.
817  *
818  * Because we are given the entire buffer from userland and don't need to have
819  * an intermediate place to copy it in, we will validate the entire thing in
820  * advance. Once it has been validated and we consider it legal, then we will
821  * translate each entry into its corresponding entry in pcb's normal floating
822  * point state. This is different from signal handling mostly because of the
823  * fact that we are not using copyin, and once we get to this point, there is
824  * no more validation, so we don't have the same concerns around blocking while
825  * pre-emption is disabled.
826  *
827  * The Wrinkle with fpregs
828  * -----------------------
829  *
830  * When we instead turn our attention to the fpregs, whether we're gathering
831  * them as part of the ucontext_t or as part of /proc, there are a few
832  * complications that we need to be aware of when we're operating on a kernel
833  * that is using xsave as the save mechanism. When we're using fxsave as the
834  * save mechanism, the CPU will always save the entire 512-byte fxsave region.
835  * The fpregs ABI that the kernel expects is basically this structure itself,
836  * which is transformed into a 32-bit compatible form in archdep.c.
837  *
838  * But xsave makes this much more complex and has historically been a source of
839  * bugs in the system. In particular, unlike fxsave, xsave has its component bit
840  * vector that is written out to indicate validity. This means that blindly
841  * copying the fxsave area without checking those bits will lead us to do the
842  * wrong thing. The XMM state flag mostly covers the 16 128-bit %xmm registers,
843  * while the x87 legacy fp flag covers the rest of the state. This is all good,
844  * aside from the MCXSR.
845  *
846  * One of the more complicated pieces of xsave state management is correctly
847  * answering the question of when the MXCSR is written out to xsave_state. In
848  * practice, this is rather convoluted and varies. If either the XMM or AVX
849  * feature bits are set then the CPU will write out the MXCSR and its mask
850  * register into the traditional fxsave state region. This behavior is dependent
851  * on the type of save function that we use. xsave and xsaveopt will look at the
852  * AVX feature bit; however, xsavec does not and only considers the SSE feature
853  * bit. This means that when we're retrieving things, we need to check both of
854  * those bits to determine if we should use the initial state or the value
855  * written out.
856  *
857  * When we come to someone trying to set the fpregs through /proc, the main
858  * question we have is what happens to the extended registers. We have opted to
859  * implement and document it such that a write to the fpregs only impacts the
860  * fpregs. Put differently, we will save the FPU state with fp_save() ahead of
861  * copying the data into the save area, set the state bits for x87 and XMM
862  * state, and then set the FPU to be restored. All in all, this basically means
863  * that writing to fpregs does not touch any of the %ymm, %zmm, or other state
864  * that we might have present.
865  *
866  * Forward Looking: Adding Intel AMX Support
867  * -----------------------------------------
868  *
869  * Nothing can stop the march of features being added into the FPU. One of the
870  * larger chunks that we will need to wrangle with is Intel's Advanced Matrix
871  * Extensions (AMX), which add a large chunk of xsave state to each process.
872  * While things like AVX and AVX-512 have been enabled by default, the broader
873  * OS community has not been wanting to do this for AMX ,because of the size of
874  * the state which exceeds 8 KiB. While the signal handling state went out of
875  * its way to minimize the size it wrote to the stack, if this is used, it would
876  * need to be preserved.
877  *
878  * To deal with this reality and the fact that folks don't really want to
879  * enable it by default for all purposes when its use will be quite special
880  * purpose, Intel has also added a MSR around extended feature disable or xfd.
881  * This is what we represent in the PRX_INFO_XCR prx_xfd member. Our starting
882  * assumption, and the reason that so much of the /proc and signal logic ensures
883  * that we have the thread and process around, taking as an example the unused
884  * process argument in fpu_proc_xregs_info(), is that we will follow suit and
885  * default to having support disabled, but that a process will be able to opt
886  * into it, which will result in several different assumptions around signal
887  * stack sizing and cause us to reallocate and extend the pcb's FPU save state.
888  *
889  * The following is a list of items to pay attention to for future folks who
890  * work on this:
891  *
892  *   o We will want to confirm whether other systems have opted to make this
893  *     process-wide or thread-wide. Assuming process-wide, we will need to do a
894  *     hold of all lwps while making a change. The interface for that probably
895  *     doesn't want to be /proc, as a process probably doesn't want to write to
896  *     its own control file. Changing it for another process could be done
897  *     through the agent-lwp.
898  *   o Opting into this should probably be a one-way street.
899  *   o Opting into this will need to evaluate all threads and in particular
900  *     stack sizes to confirm they adhere to the new minimum.
901  *   o We will need to make sure that setting and clearing the xfd MSR is part
902  *     of the FPU context ops and something we set by default on every CPU.
903  *   o We will need to add a new interface to allow opting into this feature.
904  *   o We will need to ensure that all subsequently created signal stacks adhere
905  *     to a required minimum size that we communicate through libc.
906  *   o We will need to make sure that both rtld and libc no longer rely on a
907  *     static value of the AT_SUN_FPSIZE, but rather realize that this can be
908  *     dynamic. At that time, we should evaluate if we can get away with not
909  *     needing to save this for rtld, even though signal handlers should assume
910  *     they will.
911  *   o The various components (because there is more than one) will want to be
912  *     added to the fpu_xsave_info[]. Consulting the processes's xfd will be
913  *     required and probably require logic changes.
914  *
915  * The above is not exhaustive. We'll probably have some other issues and fun
916  * while doing this.
917  */
918 
919 /*
920  * The kind of FPU we advertise to rtld so it knows what to do when working
921  * through the PLT.
922  */
923 int fp_elf = AT_386_FPINFO_FXSAVE;
924 
925 /*
926  * Mechanism to save FPU state.
927  */
928 int fp_save_mech = FP_FXSAVE;
929 
930 kmem_cache_t *fpsave_cachep;
931 
932 /* Legacy fxsave layout + xsave header + ymm */
933 #define	AVX_XSAVE_SIZE		(512 + 64 + 256)
934 
935 /*
936  * Various sanity checks.
937  */
938 CTASSERT(sizeof (struct fxsave_state) == 512);
939 CTASSERT(sizeof (struct fnsave_state) == 108);
940 CTASSERT((offsetof(struct fxsave_state, fx_xmm[0]) & 0xf) == 0);
941 CTASSERT(sizeof (struct xsave_state) >= AVX_XSAVE_SIZE);
942 
943 /*
944  * Basic architectural alignment information.
945  */
946 #define	FPU_ALIGN_XMM	16
947 #define	FPU_ALIGN_YMM	32
948 #define	FPU_ALIGN_ZMM	64
949 
950 /*
951  * This structure is the x86 implementation of the kernel FPU that is defined in
952  * uts/common/sys/kfpu.h.
953  */
954 
955 typedef enum kfpu_flags {
956 	/*
957 	 * This indicates that the save state has initial FPU data.
958 	 */
959 	KFPU_F_INITIALIZED = 0x01
960 } kfpu_flags_t;
961 
962 struct kfpu_state {
963 	fpu_ctx_t	kfpu_ctx;
964 	kfpu_flags_t	kfpu_flags;
965 	kthread_t	*kfpu_curthread;
966 };
967 
968 /*
969  * Initial kfpu state for SSE/SSE2 used by fpinit()
970  */
971 const struct fxsave_state sse_initial = {
972 	FPU_CW_INIT,	/* fx_fcw */
973 	0,		/* fx_fsw */
974 	0,		/* fx_fctw */
975 	0,		/* fx_fop */
976 	0,		/* fx_rip */
977 	0,		/* fx_rdp */
978 	SSE_MXCSR_INIT	/* fx_mxcsr */
979 	/* rest of structure is zero */
980 };
981 
982 /*
983  * Initial kfpu state for AVX used by fpinit()
984  */
985 const struct xsave_state avx_initial = {
986 	/*
987 	 * The definition below needs to be identical with sse_initial
988 	 * defined above.
989 	 */
990 	.xs_fxsave = {
991 		.fx_fcw = FPU_CW_INIT,
992 		.fx_mxcsr = SSE_MXCSR_INIT,
993 	},
994 	.xs_header = {
995 		/*
996 		 * bit0 = 1 for XSTATE_BV to indicate that legacy fields are
997 		 * valid, and CPU should initialize XMM/YMM.
998 		 */
999 		.xsh_xstate_bv = 1,
1000 		.xsh_xcomp_bv = 0,
1001 	},
1002 };
1003 
1004 /*
1005  * mxcsr_mask value (possibly reset in fpu_probe); used to avoid
1006  * the #gp exception caused by setting unsupported bits in the
1007  * MXCSR register
1008  */
1009 uint32_t sse_mxcsr_mask = SSE_MXCSR_MASK_DEFAULT;
1010 
1011 /*
1012  * This vector is patched to xsave_ctxt() or xsaveopt_ctxt() if we discover we
1013  * have an XSAVE-capable chip in fpu_probe.
1014  */
1015 void (*fpsave_ctxt)(void *) = fpxsave_ctxt;
1016 void (*fprestore_ctxt)(void *) = fpxrestore_ctxt;
1017 
1018 /*
1019  * This function pointer is changed to xsaveopt if the CPU is xsaveopt capable.
1020  */
1021 void (*xsavep)(struct xsave_state *, uint64_t) = xsave;
1022 
1023 static int fpe_sicode(uint_t);
1024 static int fpe_simd_sicode(uint_t);
1025 static void fp_new_lwp(void *, void *);
1026 static void fp_free_ctx(void *, int);
1027 
1028 static struct ctxop *
1029 fp_ctxop_allocate(struct fpu_ctx *fp)
1030 {
1031 	const struct ctxop_template tpl = {
1032 		.ct_rev		= CTXOP_TPL_REV,
1033 		.ct_save	= fpsave_ctxt,
1034 		.ct_restore	= fprestore_ctxt,
1035 		.ct_fork	= fp_new_lwp,
1036 		.ct_lwp_create	= fp_new_lwp,
1037 		.ct_free	= fp_free_ctx,
1038 	};
1039 	return (ctxop_allocate(&tpl, fp));
1040 }
1041 
1042 /*
1043  * Copy the state of parent lwp's floating point context into the new lwp.
1044  * Invoked for both fork() and lwp_create().
1045  *
1046  * Note that we inherit -only- the control state (e.g. exception masks,
1047  * rounding, precision control, etc.); the FPU registers are otherwise
1048  * reset to their initial state.
1049  */
1050 static void
1051 fp_new_lwp(void *parent, void *child)
1052 {
1053 	kthread_id_t t = parent, ct = child;
1054 	struct fpu_ctx *fp;		/* parent fpu context */
1055 	struct fpu_ctx *cfp;		/* new fpu context */
1056 	struct fxsave_state *fx, *cfx;
1057 	struct xsave_state *cxs;
1058 
1059 	ASSERT(fp_kind != FP_NO);
1060 
1061 	fp = &t->t_lwp->lwp_pcb.pcb_fpu;
1062 	cfp = &ct->t_lwp->lwp_pcb.pcb_fpu;
1063 
1064 	/*
1065 	 * If the parent FPU state is still in the FPU hw then save it;
1066 	 * conveniently, fp_save() already does this for us nicely.
1067 	 */
1068 	fp_save(fp);
1069 
1070 	cfp->fpu_flags = FPU_EN | FPU_VALID;
1071 	cfp->fpu_regs.kfpu_status = 0;
1072 	cfp->fpu_regs.kfpu_xstatus = 0;
1073 
1074 	/*
1075 	 * Make sure that the child's FPU is cleaned up and made ready for user
1076 	 * land.
1077 	 */
1078 	PCB_SET_UPDATE_FPU(&ct->t_lwp->lwp_pcb);
1079 
1080 	switch (fp_save_mech) {
1081 	case FP_FXSAVE:
1082 		fx = fp->fpu_regs.kfpu_u.kfpu_fx;
1083 		cfx = cfp->fpu_regs.kfpu_u.kfpu_fx;
1084 		bcopy(&sse_initial, cfx, sizeof (*cfx));
1085 		cfx->fx_mxcsr = fx->fx_mxcsr & ~SSE_MXCSR_EFLAGS;
1086 		cfx->fx_fcw = fx->fx_fcw;
1087 		break;
1088 
1089 	case FP_XSAVE:
1090 		cfp->fpu_xsave_mask = fp->fpu_xsave_mask;
1091 
1092 		VERIFY(fp->fpu_regs.kfpu_u.kfpu_xs != NULL);
1093 
1094 		fx = &fp->fpu_regs.kfpu_u.kfpu_xs->xs_fxsave;
1095 		cxs = cfp->fpu_regs.kfpu_u.kfpu_xs;
1096 		cfx = &cxs->xs_fxsave;
1097 
1098 		bcopy(&avx_initial, cxs, sizeof (*cxs));
1099 		cfx->fx_mxcsr = fx->fx_mxcsr & ~SSE_MXCSR_EFLAGS;
1100 		cfx->fx_fcw = fx->fx_fcw;
1101 		cxs->xs_header.xsh_xstate_bv |=
1102 		    (get_xcr(XFEATURE_ENABLED_MASK) & XFEATURE_FP_INITIAL);
1103 		break;
1104 	default:
1105 		panic("Invalid fp_save_mech");
1106 		/*NOTREACHED*/
1107 	}
1108 
1109 	/*
1110 	 * Mark that both the parent and child need to have the FPU cleaned up
1111 	 * before returning to userland.
1112 	 */
1113 
1114 	ctxop_attach(ct, fp_ctxop_allocate(cfp));
1115 }
1116 
1117 /*
1118  * Free any state associated with floating point context.
1119  * Fp_free can be called in three cases:
1120  * 1) from reaper -> thread_free -> freectx-> fp_free
1121  *	fp context belongs to a thread on deathrow
1122  *	nothing to do,  thread will never be resumed
1123  *	thread calling ctxfree is reaper
1124  *
1125  * 2) from exec -> freectx -> fp_free
1126  *	fp context belongs to the current thread
1127  *	must disable fpu, thread calling ctxfree is curthread
1128  *
1129  * 3) from restorecontext -> setfpregs -> fp_free
1130  *	we have a modified context in the memory (lwp->pcb_fpu)
1131  *	disable fpu and release the fp context for the CPU
1132  *
1133  */
1134 void
1135 fp_free(struct fpu_ctx *fp)
1136 {
1137 	ASSERT(fp_kind != FP_NO);
1138 
1139 	if (fp->fpu_flags & FPU_VALID)
1140 		return;
1141 
1142 	kpreempt_disable();
1143 	/*
1144 	 * We want to do fpsave rather than fpdisable so that we can
1145 	 * keep the fpu_flags as FPU_VALID tracking the CR0_TS bit
1146 	 */
1147 	fp->fpu_flags |= FPU_VALID;
1148 	/* If for current thread disable FP to track FPU_VALID */
1149 	if (curthread->t_lwp && fp == &curthread->t_lwp->lwp_pcb.pcb_fpu) {
1150 		/* Clear errors if any to prevent frstor from complaining */
1151 		(void) fperr_reset();
1152 		if (fp_kind & __FP_SSE)
1153 			(void) fpxerr_reset();
1154 		fpdisable();
1155 	}
1156 	kpreempt_enable();
1157 }
1158 
1159 /*
1160  * Wrapper for freectx to make the types line up for fp_free()
1161  */
1162 static void
1163 fp_free_ctx(void *arg, int isexec __unused)
1164 {
1165 	fp_free((struct fpu_ctx *)arg);
1166 }
1167 
1168 /*
1169  * Store the floating point state and disable the floating point unit.
1170  */
1171 void
1172 fp_save(struct fpu_ctx *fp)
1173 {
1174 	ASSERT(fp_kind != FP_NO);
1175 
1176 	kpreempt_disable();
1177 	if (!fp || fp->fpu_flags & FPU_VALID ||
1178 	    (fp->fpu_flags & FPU_EN) == 0) {
1179 		kpreempt_enable();
1180 		return;
1181 	}
1182 	ASSERT(curthread->t_lwp && fp == &curthread->t_lwp->lwp_pcb.pcb_fpu);
1183 
1184 	switch (fp_save_mech) {
1185 	case FP_FXSAVE:
1186 		fpxsave(fp->fpu_regs.kfpu_u.kfpu_fx);
1187 		break;
1188 
1189 	case FP_XSAVE:
1190 		xsavep(fp->fpu_regs.kfpu_u.kfpu_xs, fp->fpu_xsave_mask);
1191 		break;
1192 	default:
1193 		panic("Invalid fp_save_mech");
1194 		/*NOTREACHED*/
1195 	}
1196 
1197 	fp->fpu_flags |= FPU_VALID;
1198 
1199 	/*
1200 	 * We save the FPU as part of forking, execing, modifications via /proc,
1201 	 * restorecontext, etc. As such, we need to make sure that we return to
1202 	 * userland with valid state in the FPU. If we're context switched out
1203 	 * before we hit sys_rtt_common() we'll end up having restored the FPU
1204 	 * as part of the context ops operations. The restore logic always makes
1205 	 * sure that FPU_VALID is set before doing a restore so we don't restore
1206 	 * it a second time.
1207 	 */
1208 	PCB_SET_UPDATE_FPU(&curthread->t_lwp->lwp_pcb);
1209 
1210 	kpreempt_enable();
1211 }
1212 
1213 /*
1214  * Restore the FPU context for the thread:
1215  * The possibilities are:
1216  *	1. No active FPU context: Load the new context into the FPU hw
1217  *	   and enable the FPU.
1218  */
1219 void
1220 fp_restore(struct fpu_ctx *fp)
1221 {
1222 	switch (fp_save_mech) {
1223 	case FP_FXSAVE:
1224 		fpxrestore(fp->fpu_regs.kfpu_u.kfpu_fx);
1225 		break;
1226 
1227 	case FP_XSAVE:
1228 		xrestore(fp->fpu_regs.kfpu_u.kfpu_xs, fp->fpu_xsave_mask);
1229 		break;
1230 	default:
1231 		panic("Invalid fp_save_mech");
1232 		/*NOTREACHED*/
1233 	}
1234 
1235 	fp->fpu_flags &= ~FPU_VALID;
1236 }
1237 
1238 /*
1239  * Reset the FPU such that it is in a valid state for a new thread that is
1240  * coming out of exec. The FPU will be in a usable state at this point. At this
1241  * point we know that the FPU state has already been allocated and if this
1242  * wasn't an init process, then it will have had fp_free() previously called.
1243  */
1244 void
1245 fp_exec(void)
1246 {
1247 	struct fpu_ctx *fp = &ttolwp(curthread)->lwp_pcb.pcb_fpu;
1248 
1249 	if (fp_save_mech == FP_XSAVE) {
1250 		fp->fpu_xsave_mask = XFEATURE_FP_ALL;
1251 	}
1252 
1253 	struct ctxop *ctx = fp_ctxop_allocate(fp);
1254 	/*
1255 	 * Make sure that we're not preempted in the middle of initializing the
1256 	 * FPU on CPU.
1257 	 */
1258 	kpreempt_disable();
1259 	ctxop_attach(curthread, ctx);
1260 	fpinit();
1261 	fp->fpu_flags = FPU_EN;
1262 	kpreempt_enable();
1263 }
1264 
1265 
1266 /*
1267  * Seeds the initial state for the current thread.  The possibilities are:
1268  *      1. Another process has modified the FPU state before we have done any
1269  *         initialization: Load the FPU state from the LWP state.
1270  *      2. The FPU state has not been externally modified:  Load a clean state.
1271  */
1272 void
1273 fp_seed(void)
1274 {
1275 	struct fpu_ctx *fp = &ttolwp(curthread)->lwp_pcb.pcb_fpu;
1276 
1277 	ASSERT(curthread->t_preempt >= 1);
1278 	ASSERT((fp->fpu_flags & FPU_EN) == 0);
1279 
1280 	/*
1281 	 * Always initialize a new context and initialize the hardware.
1282 	 */
1283 	if (fp_save_mech == FP_XSAVE) {
1284 		fp->fpu_xsave_mask = XFEATURE_FP_ALL;
1285 	}
1286 
1287 	ctxop_attach(curthread, fp_ctxop_allocate(fp));
1288 	fpinit();
1289 
1290 	/*
1291 	 * If FPU_VALID is set, it means someone has modified registers via
1292 	 * /proc.  In this case, restore the current lwp's state.
1293 	 */
1294 	if (fp->fpu_flags & FPU_VALID)
1295 		fp_restore(fp);
1296 
1297 	ASSERT((fp->fpu_flags & FPU_VALID) == 0);
1298 	fp->fpu_flags = FPU_EN;
1299 }
1300 
1301 /*
1302  * When using xsave/xrstor, these three functions are used by the lwp code to
1303  * manage the memory for the xsave area.
1304  */
1305 void
1306 fp_lwp_init(klwp_t *lwp)
1307 {
1308 	struct fpu_ctx *fp = &lwp->lwp_pcb.pcb_fpu;
1309 
1310 	/*
1311 	 * We keep a copy of the pointer in lwp_fpu so that we can restore the
1312 	 * value in forklwp() after we duplicate the parent's LWP state.
1313 	 */
1314 	lwp->lwp_fpu = fp->fpu_regs.kfpu_u.kfpu_generic =
1315 	    kmem_cache_alloc(fpsave_cachep, KM_SLEEP);
1316 	fp->fpu_signal = NULL;
1317 
1318 	if (fp_save_mech == FP_XSAVE) {
1319 		/*
1320 		 *
1321 		 * We bzero since the fpinit() code path will only
1322 		 * partially initialize the xsave area using avx_inital.
1323 		 */
1324 		ASSERT(cpuid_get_xsave_size() >= sizeof (struct xsave_state));
1325 		bzero(fp->fpu_regs.kfpu_u.kfpu_xs, cpuid_get_xsave_size());
1326 	}
1327 }
1328 
1329 void
1330 fp_lwp_cleanup(klwp_t *lwp)
1331 {
1332 	struct fpu_ctx *fp = &lwp->lwp_pcb.pcb_fpu;
1333 
1334 	if (fp->fpu_regs.kfpu_u.kfpu_generic != NULL) {
1335 		kmem_cache_free(fpsave_cachep,
1336 		    fp->fpu_regs.kfpu_u.kfpu_generic);
1337 		lwp->lwp_fpu = fp->fpu_regs.kfpu_u.kfpu_generic = NULL;
1338 	}
1339 
1340 	if (fp->fpu_signal != NULL) {
1341 		kmem_cache_free(fpsave_cachep, fp->fpu_signal);
1342 		fp->fpu_signal = NULL;
1343 	}
1344 }
1345 
1346 /*
1347  * Called during the process of forklwp(). The kfpu_u pointer will have been
1348  * overwritten while copying the parent's LWP structure. We have a valid copy
1349  * stashed in the child's lwp_fpu which we use to restore the correct value.
1350  */
1351 void
1352 fp_lwp_dup(klwp_t *lwp)
1353 {
1354 	void *xp = lwp->lwp_fpu;
1355 	size_t sz;
1356 
1357 	switch (fp_save_mech) {
1358 	case FP_FXSAVE:
1359 		sz = sizeof (struct fxsave_state);
1360 		break;
1361 	case FP_XSAVE:
1362 		sz = cpuid_get_xsave_size();
1363 		break;
1364 	default:
1365 		panic("Invalid fp_save_mech");
1366 		/*NOTREACHED*/
1367 	}
1368 
1369 	/* copy the parent's values into the new lwp's struct */
1370 	bcopy(lwp->lwp_pcb.pcb_fpu.fpu_regs.kfpu_u.kfpu_generic, xp, sz);
1371 	/* now restore the pointer */
1372 	lwp->lwp_pcb.pcb_fpu.fpu_regs.kfpu_u.kfpu_generic = xp;
1373 	/* Ensure that we don't inherit our parent's signal state */
1374 	lwp->lwp_pcb.pcb_fpu.fpu_signal = NULL;
1375 }
1376 
1377 /*
1378  * Handle a processor extension error fault
1379  * Returns non zero for error.
1380  */
1381 
1382 /*ARGSUSED*/
1383 int
1384 fpexterrflt(struct regs *rp)
1385 {
1386 	uint32_t fpcw, fpsw;
1387 	fpu_ctx_t *fp = &ttolwp(curthread)->lwp_pcb.pcb_fpu;
1388 
1389 	ASSERT(fp_kind != FP_NO);
1390 
1391 	/*
1392 	 * Now we can enable the interrupts.
1393 	 * (NOTE: x87 fp exceptions come thru interrupt gate)
1394 	 */
1395 	sti();
1396 
1397 	if (!fpu_exists)
1398 		return (FPE_FLTINV);
1399 
1400 	/*
1401 	 * Do an unconditional save of the FP state.  If it's dirty (TS=0),
1402 	 * it'll be saved into the fpu context area passed in (that of the
1403 	 * current thread).  If it's not dirty (it may not be, due to
1404 	 * an intervening save due to a context switch between the sti(),
1405 	 * above and here, then it's safe to just use the stored values in
1406 	 * the context save area to determine the cause of the fault.
1407 	 */
1408 	fp_save(fp);
1409 
1410 	/* clear exception flags in saved state, as if by fnclex */
1411 	switch (fp_save_mech) {
1412 	case FP_FXSAVE:
1413 		fpsw = fp->fpu_regs.kfpu_u.kfpu_fx->fx_fsw;
1414 		fpcw = fp->fpu_regs.kfpu_u.kfpu_fx->fx_fcw;
1415 		fp->fpu_regs.kfpu_u.kfpu_fx->fx_fsw &= ~FPS_SW_EFLAGS;
1416 		break;
1417 
1418 	case FP_XSAVE:
1419 		fpsw = fp->fpu_regs.kfpu_u.kfpu_xs->xs_fxsave.fx_fsw;
1420 		fpcw = fp->fpu_regs.kfpu_u.kfpu_xs->xs_fxsave.fx_fcw;
1421 		fp->fpu_regs.kfpu_u.kfpu_xs->xs_fxsave.fx_fsw &= ~FPS_SW_EFLAGS;
1422 		/*
1423 		 * Always set LEGACY_FP as it may have been cleared by XSAVE
1424 		 * instruction
1425 		 */
1426 		fp->fpu_regs.kfpu_u.kfpu_xs->xs_header.xsh_xstate_bv |=
1427 		    XFEATURE_LEGACY_FP;
1428 		break;
1429 	default:
1430 		panic("Invalid fp_save_mech");
1431 		/*NOTREACHED*/
1432 	}
1433 
1434 	fp->fpu_regs.kfpu_status = fpsw;
1435 
1436 	if ((fpsw & FPS_ES) == 0)
1437 		return (0);		/* No exception */
1438 
1439 	/*
1440 	 * "and" the exception flags with the complement of the mask
1441 	 * bits to determine which exception occurred
1442 	 */
1443 	return (fpe_sicode(fpsw & ~fpcw & 0x3f));
1444 }
1445 
1446 /*
1447  * Handle an SSE/SSE2 precise exception.
1448  * Returns a non-zero sicode for error.
1449  */
1450 /*ARGSUSED*/
1451 int
1452 fpsimderrflt(struct regs *rp)
1453 {
1454 	uint32_t mxcsr, xmask;
1455 	fpu_ctx_t *fp = &ttolwp(curthread)->lwp_pcb.pcb_fpu;
1456 
1457 	ASSERT(fp_kind & __FP_SSE);
1458 
1459 	/*
1460 	 * NOTE: Interrupts are disabled during execution of this
1461 	 * function.  They are enabled by the caller in trap.c.
1462 	 */
1463 
1464 	/*
1465 	 * The only way we could have gotten here if there is no FP unit
1466 	 * is via a user executing an INT $19 instruction, so there is
1467 	 * no fault in that case.
1468 	 */
1469 	if (!fpu_exists)
1470 		return (0);
1471 
1472 	/*
1473 	 * Do an unconditional save of the FP state.  If it's dirty (TS=0),
1474 	 * it'll be saved into the fpu context area passed in (that of the
1475 	 * current thread).  If it's not dirty, then it's safe to just use
1476 	 * the stored values in the context save area to determine the
1477 	 * cause of the fault.
1478 	 */
1479 	fp_save(fp);		/* save the FPU state */
1480 
1481 	if (fp_save_mech == FP_XSAVE) {
1482 		mxcsr = fp->fpu_regs.kfpu_u.kfpu_xs->xs_fxsave.fx_mxcsr;
1483 		fp->fpu_regs.kfpu_status =
1484 		    fp->fpu_regs.kfpu_u.kfpu_xs->xs_fxsave.fx_fsw;
1485 	} else {
1486 		mxcsr = fp->fpu_regs.kfpu_u.kfpu_fx->fx_mxcsr;
1487 		fp->fpu_regs.kfpu_status = fp->fpu_regs.kfpu_u.kfpu_fx->fx_fsw;
1488 	}
1489 	fp->fpu_regs.kfpu_xstatus = mxcsr;
1490 
1491 	/*
1492 	 * compute the mask that determines which conditions can cause
1493 	 * a #xm exception, and use this to clean the status bits so that
1494 	 * we can identify the true cause of this one.
1495 	 */
1496 	xmask = (mxcsr >> 7) & SSE_MXCSR_EFLAGS;
1497 	return (fpe_simd_sicode((mxcsr & SSE_MXCSR_EFLAGS) & ~xmask));
1498 }
1499 
1500 /*
1501  * In the unlikely event that someone is relying on this subcode being
1502  * FPE_FLTILL for denormalize exceptions, it can always be patched back
1503  * again to restore old behaviour.
1504  */
1505 int fpe_fltden = FPE_FLTDEN;
1506 
1507 /*
1508  * Map from the FPU status word to the FP exception si_code.
1509  */
1510 static int
1511 fpe_sicode(uint_t sw)
1512 {
1513 	if (sw & FPS_IE)
1514 		return (FPE_FLTINV);
1515 	if (sw & FPS_ZE)
1516 		return (FPE_FLTDIV);
1517 	if (sw & FPS_DE)
1518 		return (fpe_fltden);
1519 	if (sw & FPS_OE)
1520 		return (FPE_FLTOVF);
1521 	if (sw & FPS_UE)
1522 		return (FPE_FLTUND);
1523 	if (sw & FPS_PE)
1524 		return (FPE_FLTRES);
1525 	return (FPE_FLTINV);	/* default si_code for other exceptions */
1526 }
1527 
1528 /*
1529  * Map from the SSE status word to the FP exception si_code.
1530  */
1531 static int
1532 fpe_simd_sicode(uint_t sw)
1533 {
1534 	if (sw & SSE_IE)
1535 		return (FPE_FLTINV);
1536 	if (sw & SSE_ZE)
1537 		return (FPE_FLTDIV);
1538 	if (sw & SSE_DE)
1539 		return (FPE_FLTDEN);
1540 	if (sw & SSE_OE)
1541 		return (FPE_FLTOVF);
1542 	if (sw & SSE_UE)
1543 		return (FPE_FLTUND);
1544 	if (sw & SSE_PE)
1545 		return (FPE_FLTRES);
1546 	return (FPE_FLTINV);	/* default si_code for other exceptions */
1547 }
1548 
1549 /*
1550  * This routine is invoked as part of libc's __fpstart implementation
1551  * via sysi86(2).
1552  *
1553  * It may be called -before- any context has been assigned in which case
1554  * we try and avoid touching the hardware.  Or it may be invoked well
1555  * after the context has been assigned and fiddled with, in which case
1556  * just tweak it directly.
1557  */
1558 void
1559 fpsetcw(uint16_t fcw, uint32_t mxcsr)
1560 {
1561 	struct fpu_ctx *fp = &curthread->t_lwp->lwp_pcb.pcb_fpu;
1562 	struct fxsave_state *fx;
1563 
1564 	if (!fpu_exists || fp_kind == FP_NO)
1565 		return;
1566 
1567 	if ((fp->fpu_flags & FPU_EN) == 0) {
1568 		if (fcw == FPU_CW_INIT && mxcsr == SSE_MXCSR_INIT) {
1569 			/*
1570 			 * Common case.  Floating point unit not yet
1571 			 * enabled, and kernel already intends to initialize
1572 			 * the hardware the way the caller wants.
1573 			 */
1574 			return;
1575 		}
1576 		/*
1577 		 * Hmm.  Userland wants a different default.
1578 		 * Do a fake "first trap" to establish the context, then
1579 		 * handle as if we already had a context before we came in.
1580 		 */
1581 		kpreempt_disable();
1582 		fp_seed();
1583 		kpreempt_enable();
1584 	}
1585 
1586 	/*
1587 	 * Ensure that the current hardware state is flushed back to the
1588 	 * pcb, then modify that copy.  Next use of the fp will
1589 	 * restore the context.
1590 	 */
1591 	fp_save(fp);
1592 
1593 	switch (fp_save_mech) {
1594 	case FP_FXSAVE:
1595 		fx = fp->fpu_regs.kfpu_u.kfpu_fx;
1596 		fx->fx_fcw = fcw;
1597 		fx->fx_mxcsr = sse_mxcsr_mask & mxcsr;
1598 		break;
1599 
1600 	case FP_XSAVE:
1601 		fx = &fp->fpu_regs.kfpu_u.kfpu_xs->xs_fxsave;
1602 		fx->fx_fcw = fcw;
1603 		fx->fx_mxcsr = sse_mxcsr_mask & mxcsr;
1604 		/*
1605 		 * Always set LEGACY_FP as it may have been cleared by XSAVE
1606 		 * instruction
1607 		 */
1608 		fp->fpu_regs.kfpu_u.kfpu_xs->xs_header.xsh_xstate_bv |=
1609 		    XFEATURE_LEGACY_FP;
1610 		break;
1611 	default:
1612 		panic("Invalid fp_save_mech");
1613 		/*NOTREACHED*/
1614 	}
1615 }
1616 
1617 static void
1618 kernel_fpu_fpstate_init(kfpu_state_t *kfpu)
1619 {
1620 	struct xsave_state *xs;
1621 
1622 	switch (fp_save_mech) {
1623 	case FP_FXSAVE:
1624 		bcopy(&sse_initial, kfpu->kfpu_ctx.fpu_regs.kfpu_u.kfpu_fx,
1625 		    sizeof (struct fxsave_state));
1626 		kfpu->kfpu_ctx.fpu_xsave_mask = 0;
1627 		break;
1628 	case FP_XSAVE:
1629 		xs = kfpu->kfpu_ctx.fpu_regs.kfpu_u.kfpu_xs;
1630 		bzero(xs, cpuid_get_xsave_size());
1631 		bcopy(&avx_initial, xs, sizeof (*xs));
1632 		xs->xs_header.xsh_xstate_bv = XFEATURE_LEGACY_FP | XFEATURE_SSE;
1633 		kfpu->kfpu_ctx.fpu_xsave_mask = XFEATURE_FP_ALL;
1634 		break;
1635 	default:
1636 		panic("invalid fp_save_mech");
1637 	}
1638 
1639 	/*
1640 	 * Set the corresponding flags that the system expects on the FPU state
1641 	 * to indicate that this is our state. The FPU_EN flag is required to
1642 	 * indicate that FPU usage is allowed. The FPU_KERN flag is explicitly
1643 	 * not set below as it represents that this state is being suppressed
1644 	 * by the kernel.
1645 	 */
1646 	kfpu->kfpu_ctx.fpu_flags = FPU_EN | FPU_VALID;
1647 	kfpu->kfpu_flags |= KFPU_F_INITIALIZED;
1648 }
1649 
1650 kfpu_state_t *
1651 kernel_fpu_alloc(int kmflags)
1652 {
1653 	kfpu_state_t *kfpu;
1654 
1655 	if ((kfpu = kmem_zalloc(sizeof (kfpu_state_t), kmflags)) == NULL) {
1656 		return (NULL);
1657 	}
1658 
1659 	kfpu->kfpu_ctx.fpu_regs.kfpu_u.kfpu_generic =
1660 	    kmem_cache_alloc(fpsave_cachep, kmflags);
1661 	if (kfpu->kfpu_ctx.fpu_regs.kfpu_u.kfpu_generic == NULL) {
1662 		kmem_free(kfpu, sizeof (kfpu_state_t));
1663 		return (NULL);
1664 	}
1665 
1666 	kernel_fpu_fpstate_init(kfpu);
1667 
1668 	return (kfpu);
1669 }
1670 
1671 void
1672 kernel_fpu_free(kfpu_state_t *kfpu)
1673 {
1674 	kmem_cache_free(fpsave_cachep,
1675 	    kfpu->kfpu_ctx.fpu_regs.kfpu_u.kfpu_generic);
1676 	kmem_free(kfpu, sizeof (kfpu_state_t));
1677 }
1678 
1679 static void
1680 kernel_fpu_ctx_save(void *arg)
1681 {
1682 	kfpu_state_t *kfpu = arg;
1683 	fpu_ctx_t *pf;
1684 
1685 	if (kfpu == NULL) {
1686 		/*
1687 		 * A NULL kfpu implies this is a kernel thread with an LWP and
1688 		 * no user-level FPU usage. Use the lwp fpu save area.
1689 		 */
1690 		pf = &curthread->t_lwp->lwp_pcb.pcb_fpu;
1691 
1692 		ASSERT(curthread->t_procp->p_flag & SSYS);
1693 		ASSERT3U(pf->fpu_flags & FPU_VALID, ==, 0);
1694 
1695 		fp_save(pf);
1696 	} else {
1697 		pf = &kfpu->kfpu_ctx;
1698 
1699 		ASSERT3P(kfpu->kfpu_curthread, ==, curthread);
1700 		ASSERT3U(pf->fpu_flags & FPU_VALID, ==, 0);
1701 
1702 		/*
1703 		 * Note, we can't use fp_save because it assumes that we're
1704 		 * saving to the thread's PCB and not somewhere else. Because
1705 		 * this is a different FPU context, we instead have to do this
1706 		 * ourselves.
1707 		 */
1708 		switch (fp_save_mech) {
1709 		case FP_FXSAVE:
1710 			fpxsave(pf->fpu_regs.kfpu_u.kfpu_fx);
1711 			break;
1712 		case FP_XSAVE:
1713 			xsavep(pf->fpu_regs.kfpu_u.kfpu_xs, pf->fpu_xsave_mask);
1714 			break;
1715 		default:
1716 			panic("Invalid fp_save_mech");
1717 		}
1718 
1719 		/*
1720 		 * Because we have saved context here, our save state is no
1721 		 * longer valid and therefore needs to be reinitialized.
1722 		 */
1723 		kfpu->kfpu_flags &= ~KFPU_F_INITIALIZED;
1724 	}
1725 
1726 	pf->fpu_flags |= FPU_VALID;
1727 
1728 	/*
1729 	 * Clear KFPU flag. This allows swtch to check for improper kernel
1730 	 * usage of the FPU (i.e. switching to a new thread while the old
1731 	 * thread was in the kernel and using the FPU, but did not perform a
1732 	 * context save).
1733 	 */
1734 	curthread->t_flag &= ~T_KFPU;
1735 }
1736 
1737 static void
1738 kernel_fpu_ctx_restore(void *arg)
1739 {
1740 	kfpu_state_t *kfpu = arg;
1741 	fpu_ctx_t *pf;
1742 
1743 	if (kfpu == NULL) {
1744 		/*
1745 		 * A NULL kfpu implies this is a kernel thread with an LWP and
1746 		 * no user-level FPU usage. Use the lwp fpu save area.
1747 		 */
1748 		pf = &curthread->t_lwp->lwp_pcb.pcb_fpu;
1749 
1750 		ASSERT(curthread->t_procp->p_flag & SSYS);
1751 		ASSERT3U(pf->fpu_flags & FPU_VALID, !=, 0);
1752 	} else {
1753 		pf = &kfpu->kfpu_ctx;
1754 
1755 		ASSERT3P(kfpu->kfpu_curthread, ==, curthread);
1756 		ASSERT3U(pf->fpu_flags & FPU_VALID, !=, 0);
1757 	}
1758 
1759 	fp_restore(pf);
1760 	curthread->t_flag |= T_KFPU;
1761 }
1762 
1763 /*
1764  * Validate that the thread is not switching off-cpu while actively using the
1765  * FPU within the kernel.
1766  */
1767 void
1768 kernel_fpu_no_swtch(void)
1769 {
1770 	if ((curthread->t_flag & T_KFPU) != 0) {
1771 		panic("curthread swtch-ing while the kernel is using the FPU");
1772 	}
1773 }
1774 
1775 static const struct ctxop_template kfpu_ctxop_tpl = {
1776 	.ct_rev		= CTXOP_TPL_REV,
1777 	.ct_save	= kernel_fpu_ctx_save,
1778 	.ct_restore	= kernel_fpu_ctx_restore,
1779 };
1780 
1781 void
1782 kernel_fpu_begin(kfpu_state_t *kfpu, uint_t flags)
1783 {
1784 	klwp_t *pl = curthread->t_lwp;
1785 	struct ctxop *ctx;
1786 
1787 	if ((curthread->t_flag & T_KFPU) != 0) {
1788 		panic("curthread attempting to nest kernel FPU states");
1789 	}
1790 
1791 	/* KFPU_USE_LWP and KFPU_NO_STATE are mutually exclusive. */
1792 	ASSERT((flags & (KFPU_USE_LWP | KFPU_NO_STATE)) !=
1793 	    (KFPU_USE_LWP | KFPU_NO_STATE));
1794 
1795 	if ((flags & KFPU_NO_STATE) == KFPU_NO_STATE) {
1796 		/*
1797 		 * Since we don't have a kfpu_state or usable lwp pcb_fpu to
1798 		 * hold our kernel FPU context, we depend on the caller doing
1799 		 * kpreempt_disable for the duration of our FPU usage. This
1800 		 * should only be done for very short periods of time.
1801 		 */
1802 		ASSERT(curthread->t_preempt > 0);
1803 		ASSERT(kfpu == NULL);
1804 
1805 		if (pl != NULL) {
1806 			/*
1807 			 * We might have already saved once so FPU_VALID could
1808 			 * be set. This is handled in fp_save.
1809 			 */
1810 			fp_save(&pl->lwp_pcb.pcb_fpu);
1811 			pl->lwp_pcb.pcb_fpu.fpu_flags |= FPU_KERNEL;
1812 		}
1813 
1814 		curthread->t_flag |= T_KFPU;
1815 
1816 		/* Always restore the fpu to the initial state. */
1817 		fpinit();
1818 
1819 		return;
1820 	}
1821 
1822 	/*
1823 	 * We either have a kfpu, or are using the LWP pcb_fpu for context ops.
1824 	 */
1825 
1826 	if ((flags & KFPU_USE_LWP) == 0) {
1827 		if (kfpu->kfpu_curthread != NULL)
1828 			panic("attempting to reuse kernel FPU state at %p when "
1829 			    "another thread already is using", kfpu);
1830 
1831 		if ((kfpu->kfpu_flags & KFPU_F_INITIALIZED) == 0)
1832 			kernel_fpu_fpstate_init(kfpu);
1833 
1834 		kfpu->kfpu_curthread = curthread;
1835 	}
1836 
1837 	/*
1838 	 * Not all threads may have an active LWP. If they do and we're not
1839 	 * going to re-use the LWP, then we should go ahead and save the state.
1840 	 * We must also note that the fpu is now being used by the kernel and
1841 	 * therefore we do not want to manage the fpu state via the user-level
1842 	 * thread's context handlers.
1843 	 *
1844 	 * We might have already saved once (due to a prior use of the kernel
1845 	 * FPU or another code path) so FPU_VALID could be set. This is handled
1846 	 * by fp_save, as is the FPU_EN check.
1847 	 */
1848 	ctx = ctxop_allocate(&kfpu_ctxop_tpl, kfpu);
1849 	kpreempt_disable();
1850 	if (pl != NULL) {
1851 		if ((flags & KFPU_USE_LWP) == 0)
1852 			fp_save(&pl->lwp_pcb.pcb_fpu);
1853 		pl->lwp_pcb.pcb_fpu.fpu_flags |= FPU_KERNEL;
1854 	}
1855 
1856 	/*
1857 	 * Set the context operations for kernel FPU usage.  Because kernel FPU
1858 	 * setup and ctxop attachment needs to happen under the protection of
1859 	 * kpreempt_disable(), we allocate the ctxop outside the guard so its
1860 	 * sleeping allocation will not cause a voluntary swtch().  This allows
1861 	 * the rest of the initialization to proceed, ensuring valid state for
1862 	 * the ctxop handlers.
1863 	 */
1864 	ctxop_attach(curthread, ctx);
1865 	curthread->t_flag |= T_KFPU;
1866 
1867 	if ((flags & KFPU_USE_LWP) == KFPU_USE_LWP) {
1868 		/*
1869 		 * For pure kernel threads with an LWP, we can use the LWP's
1870 		 * pcb_fpu to save/restore context.
1871 		 */
1872 		fpu_ctx_t *pf = &pl->lwp_pcb.pcb_fpu;
1873 
1874 		VERIFY(curthread->t_procp->p_flag & SSYS);
1875 		VERIFY(kfpu == NULL);
1876 		ASSERT((pf->fpu_flags & FPU_EN) == 0);
1877 
1878 		/* Always restore the fpu to the initial state. */
1879 		if (fp_save_mech == FP_XSAVE)
1880 			pf->fpu_xsave_mask = XFEATURE_FP_ALL;
1881 		fpinit();
1882 		pf->fpu_flags = FPU_EN | FPU_KERNEL;
1883 	} else {
1884 		/* initialize the kfpu state */
1885 		kernel_fpu_ctx_restore(kfpu);
1886 	}
1887 	kpreempt_enable();
1888 }
1889 
1890 void
1891 kernel_fpu_end(kfpu_state_t *kfpu, uint_t flags)
1892 {
1893 	if ((curthread->t_flag & T_KFPU) == 0) {
1894 		panic("curthread attempting to clear kernel FPU state "
1895 		    "without using it");
1896 	}
1897 
1898 	/*
1899 	 * General comments on why the rest of this function is structured the
1900 	 * way it is. Be aware that there is a lot of subtlety here.
1901 	 *
1902 	 * If a user-level thread ever uses the fpu while in the kernel, then
1903 	 * we cannot call fpdisable since that does STTS. That will set the
1904 	 * ts bit in %cr0 which will cause an exception if anything touches the
1905 	 * fpu. However, the user-level context switch handler (fpsave_ctxt)
1906 	 * needs to access the fpu to save the registers into the pcb.
1907 	 * fpsave_ctxt relies on CLTS having been done to clear the ts bit in
1908 	 * fprestore_ctxt when the thread context switched onto the CPU.
1909 	 *
1910 	 * Calling fpdisable only effects the current CPU's %cr0 register.
1911 	 *
1912 	 * During ctxop_remove and kpreempt_enable, we can voluntarily context
1913 	 * switch, so the CPU we were on when we entered this function might
1914 	 * not be the same one we're on when we return from ctxop_remove or end
1915 	 * the function. Note there can be user-level context switch handlers
1916 	 * still installed if this is a user-level thread.
1917 	 *
1918 	 * We also must be careful in the unlikely chance we're running in an
1919 	 * interrupt thread, since we can't leave the CPU's %cr0 TS state set
1920 	 * incorrectly for the "real" thread to resume on this CPU.
1921 	 */
1922 
1923 	if ((flags & KFPU_NO_STATE) == 0) {
1924 		kpreempt_disable();
1925 	} else {
1926 		ASSERT(curthread->t_preempt > 0);
1927 	}
1928 
1929 	curthread->t_flag &= ~T_KFPU;
1930 
1931 	/*
1932 	 * When we are ending things, we explicitly don't save the current
1933 	 * kernel FPU state back to the temporary state. The kfpu API is not
1934 	 * intended to be a permanent save location.
1935 	 *
1936 	 * If this is a user-level thread and we were to context switch
1937 	 * before returning to user-land, fpsave_ctxt will be a no-op since we
1938 	 * already saved the user-level FPU state the first time we run
1939 	 * kernel_fpu_begin (i.e. we won't save the bad kernel fpu state over
1940 	 * the user-level fpu state). The fpsave_ctxt functions only save if
1941 	 * FPU_VALID is not already set. fp_save also set PCB_SET_UPDATE_FPU so
1942 	 * fprestore_ctxt will be done in sys_rtt_common when the thread
1943 	 * finally returns to user-land.
1944 	 */
1945 
1946 	if ((curthread->t_procp->p_flag & SSYS) != 0 &&
1947 	    curthread->t_intr == NULL) {
1948 		/*
1949 		 * A kernel thread which is not an interrupt thread, so we
1950 		 * STTS now.
1951 		 */
1952 		fpdisable();
1953 	}
1954 
1955 	if ((flags & KFPU_NO_STATE) == 0) {
1956 		ctxop_remove(curthread, &kfpu_ctxop_tpl, kfpu);
1957 
1958 		if (kfpu != NULL) {
1959 			if (kfpu->kfpu_curthread != curthread) {
1960 				panic("attempting to end kernel FPU state "
1961 				    "for %p, but active thread is not "
1962 				    "curthread", kfpu);
1963 			} else {
1964 				kfpu->kfpu_curthread = NULL;
1965 			}
1966 		}
1967 
1968 		kpreempt_enable();
1969 	}
1970 
1971 	if (curthread->t_lwp != NULL) {
1972 		uint_t f;
1973 
1974 		if (flags & KFPU_USE_LWP) {
1975 			f = FPU_EN | FPU_KERNEL;
1976 		} else {
1977 			f = FPU_KERNEL;
1978 		}
1979 		curthread->t_lwp->lwp_pcb.pcb_fpu.fpu_flags &= ~f;
1980 	}
1981 }
1982 
1983 /*
1984  * Fill in FPU information that is required by exec.
1985  */
1986 void
1987 fpu_auxv_info(int *typep, size_t *lenp)
1988 {
1989 	*typep = fp_elf;
1990 	switch (fp_save_mech) {
1991 	case FP_FXSAVE:
1992 		*lenp = sizeof (struct fxsave_state);
1993 		break;
1994 	case FP_XSAVE:
1995 		*lenp = cpuid_get_xsave_size();
1996 		break;
1997 	default:
1998 		*lenp = 0;
1999 		break;
2000 	}
2001 }
2002 
2003 /*
2004  * This function exists to transform an xsave_state into an fxsave_state. The
2005  * way that we have to do this is nuanced. We assume that callers have already
2006  * handled FPU_EN and thus we only need to consider the xsave_state and its
2007  * component vector itself. This results in the following cases that we need to
2008  * consider:
2009  *
2010  *   o Neither the x87 / XMM state bits are set. We use the hardware default and
2011  *     need to ensure to copy the xsave header.
2012  *   o Both x87 / XMM state bits are set. We can copy everything.
2013  *   o Only the x87 bit is set. We need to copy the x87 state but make the XMM
2014  *     state be in the initial case.
2015  *   o Only the XMM bit is set. The reverse of the above case.
2016  *
2017  * The illumos and hardware defaults in 'sse_initial' and 'avx_initial' are
2018  * generally the same; however, the default floating point control word is
2019  * different.
2020  *
2021  * Finally, we have the complication of the MXCSR and MCXSR_MASK registers.
2022  * Because we are using xsave and xsaveopt in the kernel right now and not
2023  * xsavec, the hardware may write out the MXCSR and MXCSR_MASK registers if the
2024  * XFEATURE_AVX bit is set. Therefore if we don't have the XMM bit set but AVX
2025  * is set, we must also come back and copy out the MXCSR register. Sorry, we
2026  * don't make the rules.
2027  */
2028 static void
2029 fpu_xsave_to_fxsave(const struct xsave_state *xsave, struct fxsave_state *fx)
2030 {
2031 	const uint64_t comps = xsave->xs_header.xsh_xstate_bv;
2032 
2033 	switch (comps & (XFEATURE_LEGACY_FP | XFEATURE_SSE)) {
2034 	case XFEATURE_LEGACY_FP | XFEATURE_SSE:
2035 		bcopy(xsave, fx, sizeof (*fx));
2036 		return;
2037 	case XFEATURE_LEGACY_FP:
2038 		bcopy(xsave, fx, offsetof(struct fxsave_state, fx_xmm));
2039 		fx->fx_mxcsr = SSE_MXCSR_INIT;
2040 		fx->fx_mxcsr_mask = 0;
2041 		break;
2042 	case XFEATURE_SSE:
2043 		bcopy(&sse_initial, fx, offsetof(struct fxsave_state,
2044 		    fx_mxcsr));
2045 
2046 		fx->fx_fcw = FPU_CW_INIT_HW;
2047 		fx->fx_mxcsr = xsave->xs_fxsave.fx_mxcsr;
2048 		fx->fx_mxcsr_mask = xsave->xs_fxsave.fx_mxcsr_mask;
2049 		bcopy(xsave->xs_fxsave.fx_xmm, fx->fx_xmm, sizeof (fx->fx_xmm));
2050 		break;
2051 	default:
2052 		bcopy(&sse_initial, fx, sizeof (*fx));
2053 		fx->fx_fcw = FPU_CW_INIT_HW;
2054 		break;
2055 	}
2056 
2057 	/*
2058 	 * Account for the AVX causing MXCSR to be valid.
2059 	 */
2060 	if ((xsave->xs_header.xsh_xstate_bv & XFEATURE_AVX) != 0 &&
2061 	    (xsave->xs_header.xsh_xstate_bv & XFEATURE_SSE) == 0) {
2062 		fx->fx_mxcsr = xsave->xs_fxsave.fx_mxcsr;
2063 		fx->fx_mxcsr_mask = xsave->xs_fxsave.fx_mxcsr_mask;
2064 	}
2065 }
2066 
2067 /*
2068  * This function is designed to answer the question of are we using any xsave
2069  * family of instructions in context switch and therefore we have this state.
2070  * This should still remain true if we are using xsavec or xsaves in the kernel
2071  * in the future.
2072  */
2073 boolean_t
2074 fpu_xsave_enabled(void)
2075 {
2076 	return (fp_save_mech == FP_XSAVE);
2077 }
2078 
2079 /*
2080  * The following structure is used to track and manage the programmatic
2081  * construction of /proc and signal stack spilling of xsave information. All
2082  * known xsave types that the kernel supports must be included here.
2083  */
2084 typedef struct xsave_proc_info {
2085 	/*
2086 	 * This matches the /proc xregs type that this data represents. This s
2087 	 * used for /proc only.
2088 	 */
2089 	uint32_t xi_type;
2090 	/*
2091 	 * This indicates the size of the /proc data that we're operating on.
2092 	 * This is only used for /proc.
2093 	 */
2094 	size_t	xi_size;
2095 	/*
2096 	 * This indicates the alignment that we want to have for the member when
2097 	 * we're writing out. This is not used when setting data. This is only
2098 	 * used for /proc.
2099 	 */
2100 	size_t	xi_align;
2101 	/*
2102 	 * This indicates whether this member must always be considered or not.
2103 	 * This is used in both /proc and context/signal handling.
2104 	 */
2105 	bool	xi_always;
2106 	/*
2107 	 * This contains the corresponding bits in the xsave bit vector that
2108 	 * corresponds to this entry. This is used for both /proc and
2109 	 * context/signal handling.
2110 	 */
2111 	uint64_t xi_bits;
2112 	/*
2113 	 * The xi_fill function pointer is used to write out the /proc regset
2114 	 * data (e.g. when a user reads xregs). This is only used for the /proc
2115 	 * handling. The xi_valid function pointer is used instead to validate a
2116 	 * given set of data that we've read in, while the xi_set pointer is
2117 	 * used to actually transform the data in the underlying fpu save area.
2118 	 */
2119 	void	(*xi_fill)(const fpu_ctx_t *, const struct xsave_proc_info *,
2120 	    void *);
2121 	bool	(*xi_valid)(model_t, const void *);
2122 	void	(*xi_set)(fpu_ctx_t *, const struct xsave_proc_info *,
2123 	    uint64_t, const void *);
2124 	/*
2125 	 * The xi_signal_in and xi_signal_out function pointers are used for
2126 	 * extended context and signal handling information. They are used when
2127 	 * reading in data from a ucontext_t and writing it out respectively.
2128 	 * These are only used for context/signal handling.
2129 	 */
2130 	int	(*xi_signal_in)(const struct xsave_proc_info *,
2131 	    const ucontext_t *, const uc_xsave_t *, void *, uintptr_t *,
2132 	    const uintptr_t);
2133 	int	(*xi_signal_out)(const struct xsave_proc_info *, fpu_copyout_f,
2134 	    uc_xsave_t *, const void *fpup, uintptr_t);
2135 } xsave_proc_info_t;
2136 
2137 static bool
2138 fpu_proc_xregs_initial_state(const fpu_ctx_t *fpu, uint64_t feats)
2139 {
2140 	const struct xsave_state *xs = fpu->fpu_regs.kfpu_u.kfpu_xs;
2141 
2142 	if ((fpu->fpu_flags & (FPU_EN | FPU_VALID)) == 0) {
2143 		return (true);
2144 	}
2145 
2146 	return ((xs->xs_header.xsh_xstate_bv & feats) == 0);
2147 }
2148 
2149 static void
2150 fpu_proc_xregs_xcr_fill(const fpu_ctx_t *fpu, const xsave_proc_info_t *info,
2151     void *datap)
2152 {
2153 	prxregset_xcr_t *xcr = datap;
2154 
2155 	xcr->prx_xcr_xcr0 = xsave_bv_all;
2156 }
2157 
2158 /*
2159  * Unlike other instruction portions, we treat the xsave header and the legacy
2160  * XMM section together as both are somewhat tied at the instruction hip. Unlike
2161  * the when dealing with other xsave regions like the ymm and zmm components,
2162  * the initial state here is much more nuanced as it has to match what we actual
2163  * do in the OS and depends on the components that are present.
2164  */
2165 static void
2166 fpu_proc_xregs_xsave_fill(const fpu_ctx_t *fpu, const xsave_proc_info_t *info,
2167     void *datap)
2168 {
2169 	prxregset_xsave_t *prxsave = datap;
2170 	const struct xsave_state *xsave = fpu->fpu_regs.kfpu_u.kfpu_xs;
2171 	size_t hdr_off;
2172 
2173 	/*
2174 	 * In the x87/XMM case, the no device vs. initial state is different
2175 	 * because the initial state case still wants us to copy the real xsave
2176 	 * header. It's also worth calling out that the actual illumos default
2177 	 * fxsave state is not the same as what Intel documents. The main
2178 	 * difference is in what the x87 FPU control word is. This results in
2179 	 * the following different cases that we need to think about:
2180 	 *
2181 	 *   o FPU_EN is not set. So we use the illumos default.
2182 	 */
2183 	if ((fpu->fpu_flags & FPU_EN) == 0) {
2184 		bcopy(&avx_initial, prxsave, sizeof (*prxsave));
2185 		return;
2186 	}
2187 
2188 	/*
2189 	 * Convert all the fxsave region while taking into account the validity
2190 	 * of the xsave bits. The prxregset_xsave_t structure is the same as the
2191 	 * xsave structure in our ABI and Intel designed the xsave header to
2192 	 * begin with the 512-bit fxsave structure.
2193 	 */
2194 	fpu_xsave_to_fxsave(xsave, (struct fxsave_state *)prxsave);
2195 
2196 	/*
2197 	 * Now that we've dealt with the x87 and XMM state, take care of the
2198 	 * header.
2199 	 */
2200 	hdr_off = offsetof(prxregset_xsave_t, prx_xsh_xstate_bv);
2201 	bcopy((const void *)((uintptr_t)xsave + hdr_off),
2202 	    (void *)((uintptr_t)prxsave + hdr_off),
2203 	    sizeof (struct xsave_header));
2204 }
2205 
2206 static void
2207 fpu_proc_xregs_std_fill(const fpu_ctx_t *fpu, const xsave_proc_info_t *info,
2208     void *datap)
2209 {
2210 	if (!fpu_proc_xregs_initial_state(fpu, info->xi_bits)) {
2211 		size_t size, off;
2212 		const void *xsave_off;
2213 
2214 		cpuid_get_xsave_info(info->xi_bits, &size, &off);
2215 		ASSERT3U(size, ==, info->xi_size);
2216 		xsave_off = (void *)((uintptr_t)fpu->fpu_regs.kfpu_u.kfpu_xs +
2217 		    off);
2218 		bcopy(xsave_off, datap, info->xi_size);
2219 	}
2220 }
2221 
2222 /*
2223  * Users are not allowed to actually set the xcr information this way. However,
2224  * to make it easier for someone to just do a read, modify, write, of the xregs
2225  * data, if it is identical, then we will accept it (and do nothing).
2226  */
2227 static bool
2228 fpu_proc_xregs_xcr_valid(model_t model, const void *datap)
2229 {
2230 	const prxregset_xcr_t *xcr = datap;
2231 
2232 	return (xcr->prx_xcr_xcr0 == xsave_bv_all && xcr->prx_xcr_xfd == 0 &&
2233 	    xcr->prx_xcr_pad[0] == 0 && xcr->prx_xcr_pad[1] == 0);
2234 }
2235 
2236 /*
2237  * To match traditional /proc semantics, we do not error if reserved bits of
2238  * MXCSR are set, they will be masked off when writing data. We do not allow
2239  * someone to indicate that they are asking for compressed xsave data, hence the
2240  * check that prx_xsh_comp_bv is zero. Separately, in fpu_proc_xregs_set() we
2241  * check that each component that was indicated in the xstate_bv is actually
2242  * present.
2243  */
2244 static bool
2245 fpu_proc_xregs_xsave_valid(model_t model, const void *datap)
2246 {
2247 	const prxregset_xsave_t *xsave = datap;
2248 	uint64_t rsvd[6] = { 0 };
2249 
2250 	if (bcmp(rsvd, xsave->prx_xsh_reserved, sizeof (rsvd)) != 0 ||
2251 	    xsave->prx_xsh_xcomp_bv != 0) {
2252 		return (false);
2253 	}
2254 
2255 	if ((xsave->prx_xsh_xstate_bv & ~xsave_bv_all) != 0) {
2256 		return (false);
2257 	}
2258 
2259 	return (true);
2260 }
2261 
2262 /*
2263  * The YMM, ZMM, and Hi-ZMM registers are all valid when in an LP64 environment
2264  * on x86; however, when operating in ILP32, subsets are reserved. We require
2265  * that all reserved portions are set to zero.
2266  */
2267 static bool
2268 fpu_proc_xregs_ymm_valid(model_t model, const void *datap)
2269 {
2270 	upad128_t ymm_zero[8];
2271 	const prxregset_ymm_t *ymm = datap;
2272 
2273 	if (model == DATAMODEL_LP64) {
2274 		return (true);
2275 	}
2276 
2277 	bzero(&ymm_zero, sizeof (ymm_zero));
2278 	return (bcmp(&ymm->prx_ymm[8], &ymm_zero, sizeof (ymm_zero)) == 0);
2279 }
2280 
2281 static bool
2282 fpu_proc_xregs_zmm_valid(model_t model, const void *datap)
2283 {
2284 	upad256_t zmm_zero[8];
2285 	const prxregset_zmm_t *zmm = datap;
2286 
2287 	if (model == DATAMODEL_LP64) {
2288 		return (true);
2289 	}
2290 
2291 	bzero(&zmm_zero, sizeof (zmm_zero));
2292 	return (bcmp(&zmm->prx_zmm[8], &zmm_zero, sizeof (zmm_zero)) == 0);
2293 }
2294 
2295 static bool
2296 fpu_proc_xregs_hi_zmm_valid(model_t model, const void *datap)
2297 {
2298 	prxregset_hi_zmm_t hi_zmm_zero;
2299 	const prxregset_hi_zmm_t *hi_zmm = datap;
2300 
2301 	if (model == DATAMODEL_LP64) {
2302 		return (true);
2303 	}
2304 
2305 	bzero(&hi_zmm_zero, sizeof (hi_zmm_zero));
2306 	return (bcmp(hi_zmm, &hi_zmm_zero, sizeof (hi_zmm_zero)) == 0);
2307 }
2308 
2309 /*
2310  * The xsave state consists of the first 512 bytes of the XMM state and then the
2311  * xsave header itself. Because of the xsave header, this structure is marked
2312  * with xi_always, so we must always process and consider it.
2313  *
2314  * Semantically if either of the bits around SSE / x87 is set, then we will copy
2315  * the entire thing. This may mean that we end up copying a region that is not
2316  * valid into the save area; however, that should be OK as we still have the
2317  * specific bit flags that indicate what we should consider or not.
2318  *
2319  * There is one additional wrinkle we need to consider and honor here. The CPU
2320  * will load the MXCSR values if the AVX bit is set in an xrstor regardless of
2321  * anything else. So if this is set and we do not have a valid x87/XMM bits
2322  * set then we will set the MXCSR to its default state in case the processor
2323  * tries to load it. For reference see:
2324  *
2325  *   o Intel SDM Volume 1: 13.8.1 Standard Form of XRSTOR
2326  *   o AMD64 Volume 2: Section 11.5.9 MXCSR State Management
2327  *
2328  * Note, the behavior around this changes depending on whether using the
2329  * compressed xrstor or not. We are not, but it's worth being aware of. We do
2330  * not worry about MXCSR_MASK because the instructions ignore it.
2331  */
2332 static void
2333 fpu_proc_xregs_xsave_set(fpu_ctx_t *fpu, const xsave_proc_info_t *info,
2334     uint64_t xsave_bv, const void *datap)
2335 {
2336 	const struct xsave_state *src_xs = datap;
2337 	struct xsave_state *targ_xs = fpu->fpu_regs.kfpu_u.kfpu_xs;
2338 
2339 	if ((xsave_bv & info->xi_bits) != 0) {
2340 		bcopy(&src_xs->xs_fxsave, &targ_xs->xs_fxsave,
2341 		    sizeof (struct fxsave_state));
2342 	} else if ((xsave_bv & XFEATURE_AVX) != 0) {
2343 		targ_xs->xs_fxsave.fx_mxcsr = SSE_MXCSR_INIT;
2344 	}
2345 
2346 	bcopy(&src_xs->xs_header, &targ_xs->xs_header,
2347 	    sizeof (struct xsave_header));
2348 	targ_xs->xs_fxsave.fx_mxcsr &= sse_mxcsr_mask;
2349 }
2350 
2351 static void
2352 fpu_proc_xregs_std_set(fpu_ctx_t *fpu, const xsave_proc_info_t *info,
2353     uint64_t xsave_bv, const void *datap)
2354 {
2355 	size_t size, off;
2356 	void *xsave_off;
2357 
2358 	cpuid_get_xsave_info(info->xi_bits, &size, &off);
2359 	xsave_off = (void *)((uintptr_t)fpu->fpu_regs.kfpu_u.kfpu_xs +
2360 	    off);
2361 	bcopy(datap, xsave_off, size);
2362 }
2363 
2364 /*
2365  * Dealing with XMM data is a little more annoying in signal context. If UC_FPU
2366  * is set, the ucontext_t's fpregset_t contains a copy of the XMM region. That
2367  * must take priority over an XMM region that showed up in the uc_xsave_t data.
2368  * In the signal copyout code we do not save XMM region in the uc_xsave_t or set
2369  * it as a present component because of it being kept in the fpregset_t. Because
2370  * of this behavior, if we find the XMM (or x87) state bits present, we treat
2371  * that as an error.
2372  *
2373  * The system has always gone through and cleaned up the reserved bits in the
2374  * fxsave state when someone calls setcontext(). Therefore we need to do the
2375  * same thing which is why you see the masking of the mxcsr below.
2376  *
2377  * Finally, there is one last wrinkle here that we need to consider. The
2378  * fpregset_t has two private words which cache the status/exception
2379  * information. Therefore, we well...  cheat. Intel has left bytes 464 (0x1d0)
2380  * through 511 (0x1ff) available for us to do what we want. So we will pass this
2381  * through that for the moment to help us pass this state around without too
2382  * much extra allocation.
2383  */
2384 static int
2385 fpu_signal_copyin_xmm(const xsave_proc_info_t *info, const ucontext_t *kuc,
2386     const uc_xsave_t *ucx, void *fpup, uintptr_t *udatap,
2387     const uintptr_t max_udata)
2388 {
2389 	struct xsave_state *xsave = fpup;
2390 
2391 	if ((ucx->ucx_bv & info->xi_bits) != 0) {
2392 		return (EINVAL);
2393 	}
2394 
2395 	if ((kuc->uc_flags & UC_FPU) != 0) {
2396 		bcopy(&kuc->uc_mcontext.fpregs, &xsave->xs_fxsave,
2397 		    sizeof (struct fxsave_state));
2398 		xsave->xs_fxsave.__fx_ign2[3]._l[0] =
2399 		    kuc->uc_mcontext.fpregs.fp_reg_set.fpchip_state.status;
2400 		xsave->xs_fxsave.__fx_ign2[3]._l[1] =
2401 		    kuc->uc_mcontext.fpregs.fp_reg_set.fpchip_state.xstatus;
2402 		xsave->xs_fxsave.fx_mxcsr &= sse_mxcsr_mask;
2403 		xsave->xs_header.xsh_xstate_bv |= info->xi_bits;
2404 	}
2405 
2406 	return (0);
2407 }
2408 
2409 static int
2410 fpu_signal_copyin_std(const xsave_proc_info_t *info, const ucontext_t *kuc,
2411     const uc_xsave_t *ucx, void *fpup, uintptr_t *udatap,
2412     const uintptr_t max_udata)
2413 {
2414 	size_t len, xsave_off;
2415 	void *copy_to;
2416 	struct xsave_state *xsave = fpup;
2417 
2418 	cpuid_get_xsave_info(info->xi_bits, &len, &xsave_off);
2419 	if (*udatap + len > max_udata) {
2420 		return (EOVERFLOW);
2421 	}
2422 
2423 	copy_to = (void *)((uintptr_t)fpup + xsave_off);
2424 	if (ddi_copyin((void *)*udatap, copy_to, len, 0) != 0) {
2425 		return (EFAULT);
2426 	}
2427 
2428 	xsave->xs_header.xsh_xstate_bv |= info->xi_bits;
2429 	*udatap = *udatap + len;
2430 
2431 	return (0);
2432 }
2433 
2434 static int
2435 fpu_signal_copyout_std(const xsave_proc_info_t *info, fpu_copyout_f copyfunc,
2436     uc_xsave_t *ucx, const void *fpup, uintptr_t udatap)
2437 {
2438 	size_t len, xsave_off;
2439 	const void *copy_from;
2440 	void *copy_to;
2441 	int ret;
2442 
2443 	cpuid_get_xsave_info(info->xi_bits, &len, &xsave_off);
2444 	copy_from = (void *)(uintptr_t)fpup + xsave_off;
2445 	copy_to = (void *)(udatap + ucx->ucx_len);
2446 
2447 	ret = copyfunc(copy_from, copy_to, len);
2448 	if (ret != 0) {
2449 		return (ret);
2450 	}
2451 
2452 	ucx->ucx_len += len;
2453 	ucx->ucx_bv |= info->xi_bits;
2454 	return (0);
2455 }
2456 
2457 /*
2458  * This table contains information about the extended FPU states and synthetic
2459  * information we create for /proc, the ucontext_t, and signal handling. The
2460  * definition of the xsave_proc_info_t describes how each member is used.
2461  *
2462  * In general, this table is expected to be in the order of the xsave data
2463  * structure itself. Synthetic elements that we create can go anywhere and new
2464  * ones should be inserted at the end. This structure is walked in order to
2465  * produce the /proc and signal handling logic, so changing the order is
2466  * meaningful for those and should not be done lightly.
2467  */
2468 static const xsave_proc_info_t fpu_xsave_info[] = { {
2469 	.xi_type = PRX_INFO_XCR,
2470 	.xi_size = sizeof (prxregset_xcr_t),
2471 	.xi_align = alignof (prxregset_xcr_t),
2472 	.xi_always = true,
2473 	.xi_bits = 0,
2474 	.xi_fill = fpu_proc_xregs_xcr_fill,
2475 	.xi_valid = fpu_proc_xregs_xcr_valid
2476 }, {
2477 	/*
2478 	 * The XSAVE entry covers both the xsave header and the %xmm registers.
2479 	 * Note, there is no signal copyout information for the %xmm registers
2480 	 * because it is expected that that data is already in the fpregset_t.
2481 	 */
2482 	.xi_type = PRX_INFO_XSAVE,
2483 	.xi_size = sizeof (prxregset_xsave_t),
2484 	.xi_align = FPU_ALIGN_XMM,
2485 	.xi_always = true,
2486 	.xi_bits = XFEATURE_LEGACY_FP | XFEATURE_SSE,
2487 	.xi_fill = fpu_proc_xregs_xsave_fill,
2488 	.xi_set = fpu_proc_xregs_xsave_set,
2489 	.xi_valid = fpu_proc_xregs_xsave_valid,
2490 	.xi_signal_in = fpu_signal_copyin_xmm
2491 }, {
2492 	.xi_type = PRX_INFO_YMM,
2493 	.xi_size = sizeof (prxregset_ymm_t),
2494 	.xi_align = FPU_ALIGN_YMM,
2495 	.xi_always = false,
2496 	.xi_bits = XFEATURE_AVX,
2497 	.xi_fill = fpu_proc_xregs_std_fill,
2498 	.xi_set = fpu_proc_xregs_std_set,
2499 	.xi_signal_in = fpu_signal_copyin_std,
2500 	.xi_valid = fpu_proc_xregs_ymm_valid,
2501 	.xi_signal_out = fpu_signal_copyout_std
2502 }, {
2503 	/*
2504 	 * There is no /proc validation function for the mask registers because
2505 	 * they are the same in ILP32 / LP64 and there is nothing for us to
2506 	 * actually validate.
2507 	 */
2508 	.xi_type = PRX_INFO_OPMASK,
2509 	.xi_size = sizeof (prxregset_opmask_t),
2510 	.xi_align = alignof (prxregset_opmask_t),
2511 	.xi_always = false,
2512 	.xi_bits = XFEATURE_AVX512_OPMASK,
2513 	.xi_fill = fpu_proc_xregs_std_fill,
2514 	.xi_set = fpu_proc_xregs_std_set,
2515 	.xi_signal_in = fpu_signal_copyin_std,
2516 	.xi_signal_out = fpu_signal_copyout_std
2517 }, {
2518 	.xi_type = PRX_INFO_ZMM,
2519 	.xi_size = sizeof (prxregset_zmm_t),
2520 	.xi_align = FPU_ALIGN_ZMM,
2521 	.xi_always = false,
2522 	.xi_bits = XFEATURE_AVX512_ZMM,
2523 	.xi_fill = fpu_proc_xregs_std_fill,
2524 	.xi_set = fpu_proc_xregs_std_set,
2525 	.xi_valid = fpu_proc_xregs_zmm_valid,
2526 	.xi_signal_in = fpu_signal_copyin_std,
2527 	.xi_signal_out = fpu_signal_copyout_std
2528 }, {
2529 	.xi_type = PRX_INFO_HI_ZMM,
2530 	.xi_size = sizeof (prxregset_hi_zmm_t),
2531 	.xi_align = FPU_ALIGN_ZMM,
2532 	.xi_always = false,
2533 	.xi_bits = XFEATURE_AVX512_HI_ZMM,
2534 	.xi_fill = fpu_proc_xregs_std_fill,
2535 	.xi_set = fpu_proc_xregs_std_set,
2536 	.xi_valid = fpu_proc_xregs_hi_zmm_valid,
2537 	.xi_signal_in = fpu_signal_copyin_std,
2538 	.xi_signal_out = fpu_signal_copyout_std
2539 } };
2540 
2541 static bool
2542 fpu_proc_xregs_include(const xsave_proc_info_t *infop)
2543 {
2544 	return (infop->xi_always || (xsave_bv_all & infop->xi_bits) != 0);
2545 }
2546 
2547 void
2548 fpu_proc_xregs_info(struct proc *p __unused, uint32_t *ninfop, uint32_t *sizep,
2549     uint32_t *dstart)
2550 {
2551 	size_t ret = sizeof (prxregset_hdr_t);
2552 	uint32_t ninfo = 0;
2553 
2554 	ASSERT(fpu_xsave_enabled());
2555 
2556 	/*
2557 	 * Right now the set of flags that are enabled in the FPU is global.
2558 	 * That is, while the pcb's fcpu_ctx_t has the fpu_xsave_mask, the
2559 	 * actual things that might show up and we care about are all about what
2560 	 * is set up in %xcr0 which is stored in the global xsave_bv_all. If we
2561 	 * move to per-process FPU enablement which is likely to come with AMX,
2562 	 * then this will need the proc_t to look at, hence why we've set things
2563 	 * up with the unused variable above.
2564 	 *
2565 	 * We take two passes through the array. The first is just to count up
2566 	 * how many informational entries we need.
2567 	 */
2568 	for (size_t i = 0; i < ARRAY_SIZE(fpu_xsave_info); i++) {
2569 		if (!fpu_proc_xregs_include(&fpu_xsave_info[i]))
2570 			continue;
2571 		ninfo++;
2572 	}
2573 
2574 	ASSERT3U(ninfo, >, 0);
2575 	ret += sizeof (prxregset_info_t) * ninfo;
2576 
2577 	for (size_t i = 0; i < ARRAY_SIZE(fpu_xsave_info); i++) {
2578 		size_t curphase;
2579 		if (!fpu_proc_xregs_include(&fpu_xsave_info[i]))
2580 			continue;
2581 
2582 		curphase = ret % fpu_xsave_info[i].xi_align;
2583 		if (ret < fpu_xsave_info[i].xi_align) {
2584 			ret = fpu_xsave_info[i].xi_align;
2585 		} else if (curphase != 0) {
2586 			ret += curphase;
2587 		}
2588 
2589 		if (i == 0 && dstart != NULL) {
2590 			*dstart = ret;
2591 		}
2592 
2593 		ret += fpu_xsave_info[i].xi_size;
2594 	}
2595 
2596 	VERIFY3U(ret, <=, UINT32_MAX);
2597 	if (sizep != NULL) {
2598 		*sizep = ret;
2599 	}
2600 
2601 	if (ninfop != NULL) {
2602 		*ninfop = ninfo;
2603 	}
2604 }
2605 
2606 /*
2607  * This function supports /proc. Because /proc does not have a process locked
2608  * while processing a PCSXREG, this tries to establish an upper bound that we
2609  * will validate later in fpu_proc_xregs_set(). We basically say that if you
2610  * take the maximum xsave size and add 1 KiB that is a good enough approximation
2611  * for the maximum size. The 1 KiB is us basically trying to rationalize the
2612  * overhead of our structures that we're adding right, while being cognisant of
2613  * differing alignments and the fact that the full xsave size is in some cases
2614  * (when supervisor states or features we don't support are present) going to be
2615  * larger than we would need for this.
2616  */
2617 size_t
2618 fpu_proc_xregs_max_size(void)
2619 {
2620 	VERIFY(fpu_xsave_enabled());
2621 	return (cpuid_get_xsave_size() + 0x1000);
2622 }
2623 
2624 /*
2625  * This functions supports /proc. In particular, it's meant to perform the
2626  * following:
2627  *
2628  *  o Potentially save the current thread's registers.
2629  *  o Write out the x86 xsave /proc xregs format data from the xsave data we
2630  *    actually have. Note, this can be a little weird for cases where the FPU is
2631  *    not actually enabled, which happens for system processes.
2632  */
2633 void
2634 fpu_proc_xregs_get(klwp_t *lwp, void *buf)
2635 {
2636 	uint32_t size, ninfo, curinfo, dstart;
2637 	fpu_ctx_t *fpu = &lwp->lwp_pcb.pcb_fpu;
2638 	prxregset_hdr_t *hdr = buf;
2639 
2640 	ASSERT(fpu_xsave_enabled());
2641 	fpu_proc_xregs_info(lwp->lwp_procp, &ninfo, &size, &dstart);
2642 
2643 	/*
2644 	 * Before we get going, defensively zero out all the data buffer so that
2645 	 * the rest of the fill functions can assume a specific base.
2646 	 */
2647 	bzero(buf, size);
2648 
2649 	kpreempt_disable();
2650 	if ((fpu->fpu_flags & (FPU_EN | FPU_VALID)) == FPU_EN) {
2651 		/*
2652 		 * This case suggests that thread in question doesn't have a
2653 		 * valid FPU save state which should only happen when it is on
2654 		 * CPU. If this is the case, we must ensure that we save the
2655 		 * current FPU state before proceeding. We also sanity check
2656 		 * several things here before doing this as using /proc on
2657 		 * yourself is always exciting. fp_save() will ensure that the
2658 		 * thread is flagged to go back to being an eager FPU before
2659 		 * returning back to userland.
2660 		 */
2661 		VERIFY3P(curthread, ==, lwptot(lwp));
2662 		VERIFY0(lwptot(lwp)->t_flag & T_KFPU);
2663 		fp_save(fpu);
2664 	}
2665 	kpreempt_enable();
2666 
2667 	hdr->pr_type = PR_TYPE_XSAVE;
2668 	hdr->pr_size = size;
2669 	hdr->pr_flags = hdr->pr_pad[0] = hdr->pr_pad[1] = hdr->pr_pad[2] =
2670 	    hdr->pr_pad[3] = 0;
2671 	hdr->pr_ninfo = ninfo;
2672 
2673 	curinfo = 0;
2674 	for (size_t i = 0; i < ARRAY_SIZE(fpu_xsave_info); i++) {
2675 		void *startp;
2676 		uint32_t phase;
2677 
2678 		if (!fpu_proc_xregs_include(&fpu_xsave_info[i]))
2679 			continue;
2680 
2681 		phase = dstart % fpu_xsave_info[i].xi_align;
2682 		if (dstart < fpu_xsave_info[i].xi_align) {
2683 			ASSERT3U(i, !=, 0);
2684 			dstart = fpu_xsave_info[i].xi_align;
2685 		} else if (phase != 0) {
2686 			ASSERT3U(i, !=, 0);
2687 			dstart += phase;
2688 		}
2689 
2690 		hdr->pr_info[curinfo].pri_type = fpu_xsave_info[i].xi_type;
2691 		hdr->pr_info[curinfo].pri_flags = 0;
2692 		hdr->pr_info[curinfo].pri_size = fpu_xsave_info[i].xi_size;
2693 		hdr->pr_info[curinfo].pri_offset = dstart;
2694 
2695 		startp = (void *)((uintptr_t)buf + dstart);
2696 		fpu_xsave_info[i].xi_fill(fpu, &fpu_xsave_info[i], startp);
2697 		dstart += fpu_xsave_info[i].xi_size;
2698 		ASSERT3U(curinfo, <=, ninfo);
2699 		curinfo++;
2700 	}
2701 }
2702 
2703 /*
2704  * We have been asked to set the data in the FPU for a given thread. Our
2705  * prmachdep code has already validated that the raw semantics of the data that
2706  * we have are valid (that is the appropriate sizes, offsets, and flags). We now
2707  * apply additional checking here:
2708  *
2709  *   o The xsave structure is present and only valid bits are set.
2710  *   o If the xsave component bit-vector is set, we have the corresponding proc
2711  *     info item.
2712  *   o Read-only items are ignored if and only if they actually match what we
2713  *     gave the user mostly as a courtesy to simplify things here.
2714  *   o ILP32 processes which can't support many of the regions are allowed to
2715  *     have the items here (as we likely gave them to them), but they must be
2716  *     zero if they are set.
2717  *
2718  * We take a first pass through all the data, validating it makes sense for the
2719  * FPU. Only after that point do we ensure that we have the FPU data in question
2720  * and then we clobber all the FPU data. Part of the semantics of setting this
2721  * is that we're setting the entire extended FPU.
2722  */
2723 int
2724 fpu_proc_xregs_set(klwp_t *lwp, void *buf)
2725 {
2726 	prxregset_hdr_t *prx = buf;
2727 	model_t model = lwp_getdatamodel(lwp);
2728 	uint64_t bv_found = 0;
2729 	const prxregset_xsave_t *xsave = NULL;
2730 	fpu_ctx_t *fpu = &lwp->lwp_pcb.pcb_fpu;
2731 
2732 	VERIFY(fpu_xsave_enabled());
2733 
2734 	/*
2735 	 * First, walk each note info header that we have from the user and
2736 	 * proceed to validate it. The prmachdep code has already validated that
2737 	 * the size, type, and offset information is valid, but it has not
2738 	 * validated the semantic contents of this or if someone is trying to
2739 	 * write something they shouldn't.
2740 	 *
2741 	 * While we walk this, we keep track of where the xsave header is. We
2742 	 * also track all of the bits that we have found along the way so we can
2743 	 * match up and ensure that everything that was set has a corresponding
2744 	 * bit in the xsave bitmap. If we have something in the xsave bitmap,
2745 	 * but not its corresponding data, then that is an error. However, we
2746 	 * allow folks to write data regions without the bit set in the xsave
2747 	 * data to make the read, modify, write process simpler.
2748 	 */
2749 	for (uint32_t i = 0; i < prx->pr_ninfo; i++) {
2750 		const prxregset_info_t *info = &prx->pr_info[i];
2751 		bool found = false;
2752 
2753 		for (size_t pt = 0; pt < ARRAY_SIZE(fpu_xsave_info); pt++) {
2754 			void *data;
2755 			if (info->pri_type != fpu_xsave_info[pt].xi_type)
2756 				continue;
2757 
2758 			found = true;
2759 			data = (void *)((uintptr_t)buf + info->pri_offset);
2760 			if (fpu_xsave_info[pt].xi_valid != NULL &&
2761 			    !fpu_xsave_info[pt].xi_valid(model, data)) {
2762 				return (EINVAL);
2763 			}
2764 
2765 			if (info->pri_type == PRX_INFO_XSAVE) {
2766 				xsave = data;
2767 			}
2768 			bv_found |= fpu_xsave_info[pt].xi_bits;
2769 			break;
2770 		}
2771 
2772 		if (!found) {
2773 			return (EINVAL);
2774 		}
2775 	}
2776 
2777 	/*
2778 	 * No xsave data, no dice.
2779 	 */
2780 	if (xsave == NULL) {
2781 		return (EINVAL);
2782 	}
2783 
2784 	/*
2785 	 * If anything is set in the xsave header that was not found as we
2786 	 * walked structures, then that is an error. The opposite is not true as
2787 	 * discussed above.
2788 	 */
2789 	if ((xsave->prx_xsh_xstate_bv & ~bv_found) != 0) {
2790 		return (EINVAL);
2791 	}
2792 
2793 	/*
2794 	 * At this point, we consider all the data actually valid. Now we must
2795 	 * set up this information in the save area. If this is our own lwp, we
2796 	 * must disable it first. Otherwise, we expect that it is already valid.
2797 	 * To try to sanitize this, we will defensively zero the entire region
2798 	 * as we are setting everything that will result in here.
2799 	 */
2800 	kpreempt_disable();
2801 	if ((fpu->fpu_flags & (FPU_EN | FPU_VALID)) == FPU_EN) {
2802 		/*
2803 		 * This case suggests that thread in question doesn't have a
2804 		 * valid FPU save state which should only happen when it is on
2805 		 * CPU. If this is the case, we explicitly disable the FPU, but
2806 		 * do not save it before proceeding. We also sanity check
2807 		 * several things here before doing this as using /proc on
2808 		 * yourself is always exciting. Unlike fp_save(), fp_free() does
2809 		 * not signal that an update is required, so we unconditionally
2810 		 * set that for all threads.
2811 		 */
2812 		VERIFY3P(curthread, ==, lwptot(lwp));
2813 		VERIFY0(lwptot(lwp)->t_flag & T_KFPU);
2814 		fp_free(fpu);
2815 	}
2816 	PCB_SET_UPDATE_FPU(&lwp->lwp_pcb);
2817 	bzero(lwp->lwp_pcb.pcb_fpu.fpu_regs.kfpu_u.kfpu_generic,
2818 	    cpuid_get_xsave_size());
2819 
2820 	for (uint32_t i = 0; i < prx->pr_ninfo; i++) {
2821 		const prxregset_info_t *info = &prx->pr_info[i];
2822 		bool found = false;
2823 
2824 		for (size_t pt = 0; pt < ARRAY_SIZE(fpu_xsave_info); pt++) {
2825 			const void *data;
2826 			if (info->pri_type != fpu_xsave_info[pt].xi_type)
2827 				continue;
2828 
2829 			/*
2830 			 * Check if we have a set function and if we should
2831 			 * include this. We may not if this is something like
2832 			 * PRX_INFO_XCR which is read-only.
2833 			 *
2834 			 * We may not include a given entry as it may not have
2835 			 * been set in the actual xsave state that we have been
2836 			 * asked to restore, in which case to not break the
2837 			 * xsaveopt logic, we must leave it in its initial
2838 			 * state, e.g. zeroed (generally). XMM data initial
2839 			 * state is not zeroed, but is marked with xi_always to
2840 			 * help account for this.
2841 			 */
2842 			found = true;
2843 			if (fpu_xsave_info[pt].xi_set == NULL)
2844 				break;
2845 			if (!fpu_xsave_info[pt].xi_always &&
2846 			    (xsave->prx_xsh_xstate_bv &
2847 			    fpu_xsave_info[pt].xi_bits) !=
2848 			    fpu_xsave_info[pt].xi_bits) {
2849 				break;
2850 			}
2851 
2852 			data = (void *)((uintptr_t)buf + info->pri_offset);
2853 			fpu_xsave_info[pt].xi_set(fpu, &fpu_xsave_info[pt],
2854 			    xsave->prx_xsh_xstate_bv, data);
2855 		}
2856 
2857 		VERIFY(found);
2858 	}
2859 	kpreempt_enable();
2860 
2861 	return (0);
2862 }
2863 
2864 /*
2865  * To be included in the signal copyout logic we must have a copy function and
2866  * the bit in question must be included. Note, we don't consult xi_always here
2867  * as that is really part of what is always present for xsave logic and
2868  * therefore isn't really pertinent here because of our custom format. See the
2869  * big theory statement for more info.
2870  */
2871 static bool
2872 fpu_signal_include(const xsave_proc_info_t *infop, uint64_t xs_bv)
2873 {
2874 	return ((infop->xi_bits & xs_bv) == infop->xi_bits &&
2875 	    infop->xi_signal_out != NULL);
2876 }
2877 
2878 /*
2879  * We need to fill out the xsave related data into the ucontext_t that we've
2880  * been given. We should have a valid user pointer at this point in the uc_xsave
2881  * member. This is much simpler than the copyin that we have. Here are the
2882  * current assumptions:
2883  *
2884  *   o This is being called for the current thread. This is not meant to operate
2885  *     on an arbitrary thread's state.
2886  *   o We cannot assume whether the FPU is valid in the pcb or not. While most
2887  *     callers will have just called getfpregs() which saved the state, don't
2888  *     assume that.
2889  *   o We assume that the user address has the requisite required space for this
2890  *     to be copied out.
2891  *   o We assume that copyfunc() will ensure we are not copying into a kernel
2892  *     address.
2893  *
2894  * For more information on the format of the data, see the 'Signal Handling and
2895  * the ucontext_t' portion of the big theory statement. We copy out all the
2896  * constituent parts and then come back and write out the actual final header
2897  * information.
2898  */
2899 int
2900 fpu_signal_copyout(klwp_t *lwp, uintptr_t uaddr, fpu_copyout_f copyfunc)
2901 {
2902 	struct fpu_ctx *fpu = &lwp->lwp_pcb.pcb_fpu;
2903 	uint64_t xs_bv;
2904 	uc_xsave_t ucx;
2905 	int ret;
2906 
2907 	VERIFY3P(curthread, ==, lwptot(lwp));
2908 	VERIFY0(lwptot(lwp)->t_flag & T_KFPU);
2909 	VERIFY3U(fpu->fpu_flags & FPU_EN, ==, FPU_EN);
2910 
2911 	if (!fpu_xsave_enabled()) {
2912 		return (ENOTSUP);
2913 	}
2914 
2915 	/*
2916 	 * Unlike when we're dealing with /proc, we can unconditionally call
2917 	 * fp_save() because this is always called in the context where the lwp
2918 	 * we're operating on is always the one on CPU (which is what fp_save()
2919 	 * asserts).
2920 	 */
2921 	fp_save(fpu);
2922 
2923 	bzero(&ucx, sizeof (ucx));
2924 	ucx.ucx_vers = UC_XSAVE_VERS;
2925 	ucx.ucx_len += sizeof (uc_xsave_t);
2926 
2927 	xs_bv = fpu->fpu_regs.kfpu_u.kfpu_xs->xs_header.xsh_xstate_bv;
2928 	for (size_t i = 0; i < ARRAY_SIZE(fpu_xsave_info); i++) {
2929 		const xsave_proc_info_t *info = &fpu_xsave_info[i];
2930 
2931 		if (!fpu_signal_include(&fpu_xsave_info[i], xs_bv))
2932 			continue;
2933 		ret = info->xi_signal_out(info, copyfunc, &ucx,
2934 		    lwp->lwp_pcb.pcb_fpu.fpu_regs.kfpu_u.kfpu_generic,
2935 		    uaddr);
2936 		if (ret != 0) {
2937 			kpreempt_enable();
2938 			return (ret);
2939 		}
2940 	}
2941 
2942 	/*
2943 	 * Now that everything has been copied out, we should have an accurate
2944 	 * value in the uc_xsave_t header and we can copy that out at the start
2945 	 * of the user data.
2946 	 */
2947 	ret = copyfunc(&ucx, (void *)uaddr, sizeof (ucx));
2948 	return (ret);
2949 }
2950 
2951 /*
2952  * Here we've been given a ucontext_t which potentially has a user pointer to
2953  * xsave state that we've copied out previously. In this case we need to do the
2954  * following, assuming UC_XSAVE is present:
2955  *
2956  *   o Copy in our header and validate it.
2957  *   o Allocate an fpu context to use as a holding ground for all this data.
2958  *   o If UC_FPU is set, override the xsave structure with the saved XMM state,
2959  *     clear UC_FPU, and make sure that the correct xsave_bv bits are set.
2960  *
2961  * Currently we always allocate the additional state as a holding ground for the
2962  * FPU. What we're copying in may not be valid and we don't want to clobber the
2963  * existing FPU state or deal with merging it until we believe it's reasonable
2964  * enough. The proc_t is here to set us up for when we have per-process settings
2965  * in the extended feature disable MSRs.
2966  */
2967 int
2968 fpu_signal_copyin(klwp_t *lwp, ucontext_t *kuc)
2969 {
2970 	uc_xsave_t ucx;
2971 	uint64_t bv;
2972 	uintptr_t data, max_data;
2973 	void *fpu;
2974 	proc_t *p = lwp->lwp_procp;
2975 	size_t ksize;
2976 
2977 	/*
2978 	 * Because this has been opaque filler and the kernel has never
2979 	 * historically looked at it, we don't really care about the uc_xsave
2980 	 * pointer being garbage in the case that the flag is not set. While
2981 	 * this isn't perhaps the most sporting choice in some cases, this is on
2982 	 * the other hand, pragmatic.
2983 	 */
2984 	if ((kuc->uc_flags & UC_XSAVE) != 0) {
2985 		if (kuc->uc_xsave == 0) {
2986 			return (EINVAL);
2987 		}
2988 
2989 		if (!fpu_xsave_enabled()) {
2990 			return (ENOTSUP);
2991 		}
2992 	} else {
2993 		return (0);
2994 	}
2995 
2996 	if (ddi_copyin((const void *)kuc->uc_xsave, &ucx, sizeof (ucx), 0) !=
2997 	    0) {
2998 		return (EFAULT);
2999 	}
3000 
3001 	ksize = cpuid_get_xsave_size();
3002 	if (ucx.ucx_vers != UC_XSAVE_VERS || ucx.ucx_len < sizeof (ucx) ||
3003 	    ucx.ucx_len > ksize ||
3004 	    (ucx.ucx_bv & ~xsave_bv_all) != 0 ||
3005 	    (uintptr_t)p->p_as->a_userlimit - ucx.ucx_len <
3006 	    (uintptr_t)kuc->uc_xsave) {
3007 		return (EINVAL);
3008 	}
3009 
3010 	/*
3011 	 * OK, our goal right now is to recreate a valid xsave_state structure
3012 	 * that we'll ultimately end up having to merge with our existing one in
3013 	 * the FPU save state. The reason we describe this as a merge is to help
3014 	 * future us when we want to retain supervisor state which will never be
3015 	 * part of userland signal state. The design of the userland signal
3016 	 * state is basically to compress it as much as we can. This is done for
3017 	 * two reasons:
3018 	 *
3019 	 *   1) We currently consider this a private interface.
3020 	 *   2) We really want to minimize the actual amount of stack space we
3021 	 *	use as much as possible. Most applications aren't using AVX-512
3022 	 *	right now, so doing our own compression style is worthwhile. If
3023 	 *	libc adopts AVX-512 routines, we may want to change this.
3024 	 *
3025 	 * On the allocation below, our assumption is that if a thread has taken
3026 	 * a signal, then it is likely to take a signal again in the future (or
3027 	 * be shortly headed to its demise). As such, when that happens we will
3028 	 * leave the allocated signal stack around for the process. Most
3029 	 * applications don't allow all threads to take signals, so this should
3030 	 * hopefully help amortize the cost of the allocation.
3031 	 */
3032 	max_data = (uintptr_t)kuc->uc_xsave + ucx.ucx_len;
3033 	data = (uintptr_t)kuc->uc_xsave + sizeof (ucx);
3034 	bv = ucx.ucx_bv;
3035 	if (lwp->lwp_pcb.pcb_fpu.fpu_signal == NULL) {
3036 		lwp->lwp_pcb.pcb_fpu.fpu_signal =
3037 		    kmem_cache_alloc(fpsave_cachep, KM_SLEEP);
3038 	}
3039 	fpu = lwp->lwp_pcb.pcb_fpu.fpu_signal;
3040 
3041 	/*
3042 	 * Unconditionally initialize the memory we get in here to ensure that
3043 	 * it is in a reasonable state for ourselves. This ensures that unused
3044 	 * regions are mostly left in their initial state (the main exception
3045 	 * here is the x87/XMM state, but that should be OK). We don't fill in
3046 	 * the initial xsave state as we expect that to happen as part of our
3047 	 * processing.
3048 	 */
3049 	bzero(fpu, ksize);
3050 
3051 	for (size_t i = 0; i < ARRAY_SIZE(fpu_xsave_info); i++) {
3052 		int ret;
3053 		const xsave_proc_info_t *info = &fpu_xsave_info[i];
3054 		if (!info->xi_always && (info->xi_bits & bv) == 0)
3055 			continue;
3056 		bv &= ~info->xi_bits;
3057 
3058 		if (info->xi_signal_in == NULL)
3059 			continue;
3060 		ret = info->xi_signal_in(info, kuc, &ucx, fpu, &data, max_data);
3061 		if (ret != 0) {
3062 			return (ret);
3063 		}
3064 	}
3065 	ASSERT0(bv);
3066 
3067 	/*
3068 	 * As described in the big theory statement section 'Signal Handling and
3069 	 * the ucontext_t', we always remove UC_FPU from here as we've taken
3070 	 * care of reassembling it ourselves.
3071 	 */
3072 	kuc->uc_flags &= ~UC_FPU;
3073 	kuc->uc_xsave = (uintptr_t)fpu;
3074 
3075 	return (0);
3076 }
3077 
3078 /*
3079  * This determines the size of the signal stack that we need for our custom form
3080  * of the xsave state.
3081  */
3082 size_t
3083 fpu_signal_size(klwp_t *lwp)
3084 {
3085 	struct fpu_ctx *fpu = &lwp->lwp_pcb.pcb_fpu;
3086 	size_t len = sizeof (uc_xsave_t);
3087 	uint64_t xs_bv;
3088 
3089 	VERIFY3P(curthread, ==, lwptot(lwp));
3090 	VERIFY0(lwptot(lwp)->t_flag & T_KFPU);
3091 	VERIFY3U(fpu->fpu_flags & FPU_EN, ==, FPU_EN);
3092 
3093 	if (!fpu_xsave_enabled()) {
3094 		return (0);
3095 	}
3096 
3097 	kpreempt_disable();
3098 	if ((fpu->fpu_flags & (FPU_EN | FPU_VALID)) == FPU_EN) {
3099 		fp_save(fpu);
3100 	}
3101 
3102 	xs_bv = fpu->fpu_regs.kfpu_u.kfpu_xs->xs_header.xsh_xstate_bv;
3103 	for (size_t i = 0; i < ARRAY_SIZE(fpu_xsave_info); i++) {
3104 		size_t comp_size;
3105 
3106 		if (!fpu_signal_include(&fpu_xsave_info[i], xs_bv))
3107 			continue;
3108 
3109 		cpuid_get_xsave_info(fpu_xsave_info[i].xi_bits, &comp_size,
3110 		    NULL);
3111 		len += comp_size;
3112 	}
3113 
3114 	kpreempt_enable();
3115 	return (len);
3116 }
3117 
3118 /*
3119  * This function is used in service of restorecontext() to set the specified
3120  * thread's extended FPU state to the passed in data. Our assumptions at this
3121  * point from the system are:
3122  *
3123  *   o Someone has already verified that the actual xsave header is correct.
3124  *   o Any traditional XMM state that causes a #gp has been clamped.
3125  *   o That data is basically the correct sized xsave state structure. Right now
3126  *     that means it is not compressed and follows the CPUID-based rules for
3127  *     constructing and laying out data.
3128  *   o That the lwp argument refers to the current thread.
3129  *
3130  * Our primary purpose here is to merge the current FPU state with what exists
3131  * here. Right now, "merge", strictly speaking is just "replace". We can get
3132  * away with just replacing everything because all we currently save are user
3133  * states. If we start saving kernel states in here, this will get more nuanced
3134  * and we will need to be more careful about how we store data here.
3135  */
3136 void
3137 fpu_set_xsave(klwp_t *lwp, const void *data)
3138 {
3139 	struct fpu_ctx *fpu = &lwp->lwp_pcb.pcb_fpu;
3140 	uint32_t status, xstatus;
3141 	struct xsave_state *dst_xsave;
3142 
3143 	VERIFY(fpu_xsave_enabled());
3144 	VERIFY3P(curthread, ==, lwptot(lwp));
3145 	VERIFY0(lwptot(lwp)->t_flag & T_KFPU);
3146 	ASSERT3U(fpu->fpu_flags & FPU_EN, ==, FPU_EN);
3147 
3148 	/*
3149 	 * We use fp_save() here rather than a stock fpdisable() so we can
3150 	 * attempt to honor our invariants that when the thread state has been
3151 	 * saved, the valid flag is set, even though we're going to be
3152 	 * overwriting it shortly. If we just called fpdisable() then we would
3153 	 * basically be asking for trouble.
3154 	 *
3155 	 * Because we are modifying the state here and we don't want the system
3156 	 * to end up in an odd state, we are being a little paranoid and
3157 	 * disabling preemption across this operation. In particular, once the
3158 	 * state is properly tagged with FPU_VALID, there should be no other way
3159 	 * that this thread can return to userland and get cleared out because
3160 	 * we're resetting its context; however, we let paranoia win out.
3161 	 */
3162 	kpreempt_disable();
3163 	if ((fpu->fpu_flags & (FPU_EN | FPU_VALID)) == FPU_EN) {
3164 		fp_save(fpu);
3165 	}
3166 
3167 	bcopy(data, lwp->lwp_pcb.pcb_fpu.fpu_regs.kfpu_u.kfpu_generic,
3168 	    cpuid_get_xsave_size());
3169 	dst_xsave = lwp->lwp_pcb.pcb_fpu.fpu_regs.kfpu_u.kfpu_generic;
3170 	status = dst_xsave->xs_fxsave.__fx_ign2[3]._l[0];
3171 	xstatus = dst_xsave->xs_fxsave.__fx_ign2[3]._l[1];
3172 	dst_xsave->xs_fxsave.__fx_ign2[3]._l[0] = 0;
3173 	dst_xsave->xs_fxsave.__fx_ign2[3]._l[1] = 0;
3174 
3175 	/*
3176 	 * These two status words are information that the kernel itself uses to
3177 	 * track additional information and is part of the traditional fpregset,
3178 	 * but is not part of our xregs information. Because we are setting this
3179 	 * state, we leave it up to the rest of the kernel to determine whether
3180 	 * this came from an fpregset_t or is being reset to the default of 0.
3181 	 */
3182 	fpu->fpu_regs.kfpu_status = status;
3183 	fpu->fpu_regs.kfpu_xstatus = xstatus;
3184 
3185 	fpu->fpu_flags |= FPU_VALID;
3186 	PCB_SET_UPDATE_FPU(&lwp->lwp_pcb);
3187 	kpreempt_enable();
3188 }
3189 
3190 /*
3191  * Convert the current FPU state to the traditional fpregset_t. In the 64-bit
3192  * kernel, this is just an fxsave_state with additional values for the status
3193  * and xstatus members.
3194  *
3195  * This has the same nuance as the xregs cases discussed above, but is simpler
3196  * in that we only need to handle the fxsave state, but more complicated because
3197  * we need to check our save mechanism.
3198  */
3199 void
3200 fpu_get_fpregset(klwp_t *lwp, fpregset_t *fp)
3201 {
3202 	struct fpu_ctx *fpu = &lwp->lwp_pcb.pcb_fpu;
3203 
3204 	kpreempt_disable();
3205 	fp->fp_reg_set.fpchip_state.status = fpu->fpu_regs.kfpu_status;
3206 	fp->fp_reg_set.fpchip_state.xstatus = fpu->fpu_regs.kfpu_xstatus;
3207 
3208 	if ((fpu->fpu_flags & (FPU_EN | FPU_VALID)) == FPU_EN) {
3209 		/*
3210 		 * If we're requesting the fpregs of a thread that isn't
3211 		 * currently valid and isn't the one that we're executing, then
3212 		 * we consider getting this information to be a best-effort and
3213 		 * we will not stop the thread in question to serialize it,
3214 		 * which means possibly getting stale data. This is the
3215 		 * traditional semantics that the system has used to service
3216 		 * this for /proc.
3217 		 */
3218 		if (curthread == lwptot(lwp)) {
3219 			VERIFY0(lwptot(lwp)->t_flag & T_KFPU);
3220 			fp_save(fpu);
3221 		}
3222 	}
3223 
3224 	/*
3225 	 * If the FPU is not enabled and the state isn't valid (due to someone
3226 	 * else setting it), just copy the initial state.
3227 	 */
3228 	if ((fpu->fpu_flags & (FPU_EN | FPU_VALID)) == 0) {
3229 		bcopy(&sse_initial, fp, sizeof (sse_initial));
3230 		kpreempt_enable();
3231 		return;
3232 	}
3233 
3234 	/*
3235 	 * Given that we have an enabled FPU, we must look at the type of FPU
3236 	 * save mechanism to clean this up. In particular, while we can just
3237 	 * copy the save area with FXSAVE, with XSAVE we must carefully copy
3238 	 * only the bits that are valid and reset the rest to their default
3239 	 * state.
3240 	 */
3241 	switch (fp_save_mech) {
3242 	case FP_FXSAVE:
3243 		bcopy(fpu->fpu_regs.kfpu_u.kfpu_fx, fp,
3244 		    sizeof (struct fxsave_state));
3245 		break;
3246 	case FP_XSAVE:
3247 		fpu_xsave_to_fxsave(fpu->fpu_regs.kfpu_u.kfpu_xs,
3248 		    (struct fxsave_state *)fp);
3249 		break;
3250 	default:
3251 		panic("Invalid fp_save_mech");
3252 	}
3253 
3254 	kpreempt_enable();
3255 }
3256 
3257 /*
3258  * This is a request to set the ABI fpregset_t into our actual hardware state.
3259  * In the 64-bit kernel the first 512 bytes of the fpregset_t is the same as the
3260  * 512-byte fxsave area.
3261  */
3262 void
3263 fpu_set_fpregset(klwp_t *lwp, const fpregset_t *fp)
3264 {
3265 	struct fpu_ctx *fpu = &lwp->lwp_pcb.pcb_fpu;
3266 
3267 	kpreempt_disable();
3268 	if ((fpu->fpu_flags & (FPU_EN | FPU_VALID)) == FPU_EN) {
3269 		/*
3270 		 * We always save the entire FPU. This is required if we're
3271 		 * using xsave. If we're using fxsave, we could skip the
3272 		 * 512-byte write and instead just disable the FPU since we'd be
3273 		 * replacing it all. For now we don't bother with more
3274 		 * conditional logic.
3275 		 */
3276 		VERIFY3P(curthread, ==, lwptot(lwp));
3277 		VERIFY0(lwptot(lwp)->t_flag & T_KFPU);
3278 		fp_save(fpu);
3279 	}
3280 
3281 	fpu->fpu_regs.kfpu_xstatus = fp->fp_reg_set.fpchip_state.xstatus;
3282 	fpu->fpu_regs.kfpu_status = fp->fp_reg_set.fpchip_state.status;
3283 	switch (fp_save_mech) {
3284 	case FP_FXSAVE:
3285 		bcopy(fp, fpu->fpu_regs.kfpu_u.kfpu_fx,
3286 		    sizeof (struct fxsave_state));
3287 		break;
3288 	case FP_XSAVE:
3289 		bcopy(fp, fpu->fpu_regs.kfpu_u.kfpu_xs,
3290 		    sizeof (struct fxsave_state));
3291 		fpu->fpu_regs.kfpu_u.kfpu_xs->xs_header.xsh_xstate_bv |=
3292 		    XFEATURE_LEGACY_FP | XFEATURE_SSE;
3293 		break;
3294 	default:
3295 		panic("Invalid fp_save_mech");
3296 	}
3297 
3298 	fpu->fpu_flags |= FPU_VALID;
3299 	PCB_SET_UPDATE_FPU(&lwp->lwp_pcb);
3300 	kpreempt_enable();
3301 }
3302