xref: /dragonfly/sys/platform/vkernel64/x86_64/npx.c (revision 10cbe914)
1 /*
2  * Copyright (c) 2006 The DragonFly Project.  All rights reserved.
3  * Copyright (c) 1990 William Jolitz.
4  * Copyright (c) 1991 The Regents of the University of California.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The DragonFly Project
8  * by Matthew Dillon <dillon@backplane.com>
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  *
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in
18  *    the documentation and/or other materials provided with the
19  *    distribution.
20  * 3. Neither the name of The DragonFly Project nor the names of its
21  *    contributors may be used to endorse or promote products derived
22  *    from this software without specific, prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
25  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
26  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
27  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
28  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
29  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
30  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
31  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
32  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
33  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
34  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35  * SUCH DAMAGE.
36  *
37  * from: @(#)npx.c	7.2 (Berkeley) 5/12/91
38  * $FreeBSD: src/sys/i386/isa/npx.c,v 1.80.2.3 2001/10/20 19:04:38 tegge Exp $
39  * $DragonFly: src/sys/platform/vkernel/i386/npx.c,v 1.8 2008/01/29 19:54:56 dillon Exp $
40  */
41 
42 #include "opt_debug_npx.h"
43 
44 #include <sys/param.h>
45 #include <sys/systm.h>
46 #include <sys/bus.h>
47 #include <sys/kernel.h>
48 #include <sys/malloc.h>
49 #include <sys/module.h>
50 #include <sys/sysctl.h>
51 #include <sys/proc.h>
52 #include <sys/rman.h>
53 #ifdef NPX_DEBUG
54 #include <sys/syslog.h>
55 #endif
56 #include <sys/signalvar.h>
57 #include <sys/thread2.h>
58 
59 #ifndef SMP
60 #include <machine/asmacros.h>
61 #endif
62 #include <machine/cputypes.h>
63 #include <machine/frame.h>
64 #include <machine/md_var.h>
65 #include <machine/pcb.h>
66 #include <machine/psl.h>
67 #ifndef SMP
68 #include <machine/clock.h>
69 #endif
70 #include <machine/specialreg.h>
71 #include <machine/segments.h>
72 #include <machine/globaldata.h>
73 
74 #define	fldcw(addr)		__asm("fldcw %0" : : "m" (*(addr)))
75 #define	fnclex()		__asm("fnclex")
76 #define	fninit()		__asm("fninit")
77 #define	fnop()			__asm("fnop")
78 #define	fnsave(addr)		__asm __volatile("fnsave %0" : "=m" (*(addr)))
79 #define	fnstcw(addr)		__asm __volatile("fnstcw %0" : "=m" (*(addr)))
80 #define	fnstsw(addr)		__asm __volatile("fnstsw %0" : "=m" (*(addr)))
81 #define	frstor(addr)		__asm("frstor %0" : : "m" (*(addr)))
82 #ifndef CPU_DISABLE_SSE
83 #define	fxrstor(addr)		__asm("fxrstor %0" : : "m" (*(addr)))
84 #define	fxsave(addr)		__asm __volatile("fxsave %0" : "=m" (*(addr)))
85 #endif
86 
87 #ifndef CPU_DISABLE_SSE
88 #define GET_FPU_EXSW_PTR(td) \
89 	(cpu_fxsr ? \
90 		&(td)->td_savefpu->sv_xmm.sv_ex_sw : \
91 		&(td)->td_savefpu->sv_87.sv_ex_sw)
92 #else /* CPU_DISABLE_SSE */
93 #define GET_FPU_EXSW_PTR(td) \
94 	(&(td)->td_savefpu->sv_87.sv_ex_sw)
95 #endif /* CPU_DISABLE_SSE */
96 
97 typedef u_char bool_t;
98 #ifndef CPU_DISABLE_SSE
99 static	void	fpu_clean_state(void);
100 #endif
101 
102 int cpu_fxsr = 0;
103 
104 static struct krate badfprate = { 1 };
105 
106 /*static	int	npx_attach	(device_t dev);*/
107 static	void	fpusave		(union savefpu *);
108 static	void	fpurstor	(union savefpu *);
109 
110 #if (defined(I586_CPU) || defined(I686_CPU)) && !defined(CPU_DISABLE_SSE)
111 int mmxopt = 1;
112 SYSCTL_INT(_kern, OID_AUTO, mmxopt, CTLFLAG_RD, &mmxopt, 0,
113 	"MMX/XMM optimized bcopy/copyin/copyout support");
114 #endif
115 
116 static int      hw_instruction_sse;
117 SYSCTL_INT(_hw, OID_AUTO, instruction_sse, CTLFLAG_RD,
118     &hw_instruction_sse, 0, "SIMD/MMX2 instructions available in CPU");
119 
120 #if 0
121 /*
122  * Attach routine - announce which it is, and wire into system
123  */
124 int
125 npx_attach(device_t dev)
126 {
127 	npxinit(__INITIAL_NPXCW__);
128 	return (0);
129 }
130 #endif
131 
132 void
133 init_fpu(int supports_sse)
134 {
135 	cpu_fxsr = hw_instruction_sse = supports_sse;
136 }
137 
138 /*
139  * Initialize the floating point unit.
140  */
141 void
142 npxinit(u_short control)
143 {
144 	static union savefpu dummy __aligned(16);
145 
146 	/*
147 	 * fninit has the same h/w bugs as fnsave.  Use the detoxified
148 	 * fnsave to throw away any junk in the fpu.  npxsave() initializes
149 	 * the fpu and sets npxthread = NULL as important side effects.
150 	 */
151 	npxsave(&dummy);
152 	crit_enter();
153 	/*stop_emulating();*/
154 	fldcw(&control);
155 	fpusave(curthread->td_savefpu);
156 	mdcpu->gd_npxthread = NULL;
157 	/*start_emulating();*/
158 	crit_exit();
159 }
160 
161 /*
162  * Free coprocessor (if we have it).
163  */
164 void
165 npxexit(void)
166 {
167 	if (curthread == mdcpu->gd_npxthread)
168 		npxsave(curthread->td_savefpu);
169 }
170 
171 #if 0
172 /*
173  * The following mechanism is used to ensure that the FPE_... value
174  * that is passed as a trapcode to the signal handler of the user
175  * process does not have more than one bit set.
176  *
177  * Multiple bits may be set if the user process modifies the control
178  * word while a status word bit is already set.  While this is a sign
179  * of bad coding, we have no choise than to narrow them down to one
180  * bit, since we must not send a trapcode that is not exactly one of
181  * the FPE_ macros.
182  *
183  * The mechanism has a static table with 127 entries.  Each combination
184  * of the 7 FPU status word exception bits directly translates to a
185  * position in this table, where a single FPE_... value is stored.
186  * This FPE_... value stored there is considered the "most important"
187  * of the exception bits and will be sent as the signal code.  The
188  * precedence of the bits is based upon Intel Document "Numerical
189  * Applications", Chapter "Special Computational Situations".
190  *
191  * The macro to choose one of these values does these steps: 1) Throw
192  * away status word bits that cannot be masked.  2) Throw away the bits
193  * currently masked in the control word, assuming the user isn't
194  * interested in them anymore.  3) Reinsert status word bit 7 (stack
195  * fault) if it is set, which cannot be masked but must be presered.
196  * 4) Use the remaining bits to point into the trapcode table.
197  *
198  * The 6 maskable bits in order of their preference, as stated in the
199  * above referenced Intel manual:
200  * 1  Invalid operation (FP_X_INV)
201  * 1a   Stack underflow
202  * 1b   Stack overflow
203  * 1c   Operand of unsupported format
204  * 1d   SNaN operand.
205  * 2  QNaN operand (not an exception, irrelavant here)
206  * 3  Any other invalid-operation not mentioned above or zero divide
207  *      (FP_X_INV, FP_X_DZ)
208  * 4  Denormal operand (FP_X_DNML)
209  * 5  Numeric over/underflow (FP_X_OFL, FP_X_UFL)
210  * 6  Inexact result (FP_X_IMP)
211  */
212 static char fpetable[128] = {
213 	0,
214 	FPE_FLTINV,	/*  1 - INV */
215 	FPE_FLTUND,	/*  2 - DNML */
216 	FPE_FLTINV,	/*  3 - INV | DNML */
217 	FPE_FLTDIV,	/*  4 - DZ */
218 	FPE_FLTINV,	/*  5 - INV | DZ */
219 	FPE_FLTDIV,	/*  6 - DNML | DZ */
220 	FPE_FLTINV,	/*  7 - INV | DNML | DZ */
221 	FPE_FLTOVF,	/*  8 - OFL */
222 	FPE_FLTINV,	/*  9 - INV | OFL */
223 	FPE_FLTUND,	/*  A - DNML | OFL */
224 	FPE_FLTINV,	/*  B - INV | DNML | OFL */
225 	FPE_FLTDIV,	/*  C - DZ | OFL */
226 	FPE_FLTINV,	/*  D - INV | DZ | OFL */
227 	FPE_FLTDIV,	/*  E - DNML | DZ | OFL */
228 	FPE_FLTINV,	/*  F - INV | DNML | DZ | OFL */
229 	FPE_FLTUND,	/* 10 - UFL */
230 	FPE_FLTINV,	/* 11 - INV | UFL */
231 	FPE_FLTUND,	/* 12 - DNML | UFL */
232 	FPE_FLTINV,	/* 13 - INV | DNML | UFL */
233 	FPE_FLTDIV,	/* 14 - DZ | UFL */
234 	FPE_FLTINV,	/* 15 - INV | DZ | UFL */
235 	FPE_FLTDIV,	/* 16 - DNML | DZ | UFL */
236 	FPE_FLTINV,	/* 17 - INV | DNML | DZ | UFL */
237 	FPE_FLTOVF,	/* 18 - OFL | UFL */
238 	FPE_FLTINV,	/* 19 - INV | OFL | UFL */
239 	FPE_FLTUND,	/* 1A - DNML | OFL | UFL */
240 	FPE_FLTINV,	/* 1B - INV | DNML | OFL | UFL */
241 	FPE_FLTDIV,	/* 1C - DZ | OFL | UFL */
242 	FPE_FLTINV,	/* 1D - INV | DZ | OFL | UFL */
243 	FPE_FLTDIV,	/* 1E - DNML | DZ | OFL | UFL */
244 	FPE_FLTINV,	/* 1F - INV | DNML | DZ | OFL | UFL */
245 	FPE_FLTRES,	/* 20 - IMP */
246 	FPE_FLTINV,	/* 21 - INV | IMP */
247 	FPE_FLTUND,	/* 22 - DNML | IMP */
248 	FPE_FLTINV,	/* 23 - INV | DNML | IMP */
249 	FPE_FLTDIV,	/* 24 - DZ | IMP */
250 	FPE_FLTINV,	/* 25 - INV | DZ | IMP */
251 	FPE_FLTDIV,	/* 26 - DNML | DZ | IMP */
252 	FPE_FLTINV,	/* 27 - INV | DNML | DZ | IMP */
253 	FPE_FLTOVF,	/* 28 - OFL | IMP */
254 	FPE_FLTINV,	/* 29 - INV | OFL | IMP */
255 	FPE_FLTUND,	/* 2A - DNML | OFL | IMP */
256 	FPE_FLTINV,	/* 2B - INV | DNML | OFL | IMP */
257 	FPE_FLTDIV,	/* 2C - DZ | OFL | IMP */
258 	FPE_FLTINV,	/* 2D - INV | DZ | OFL | IMP */
259 	FPE_FLTDIV,	/* 2E - DNML | DZ | OFL | IMP */
260 	FPE_FLTINV,	/* 2F - INV | DNML | DZ | OFL | IMP */
261 	FPE_FLTUND,	/* 30 - UFL | IMP */
262 	FPE_FLTINV,	/* 31 - INV | UFL | IMP */
263 	FPE_FLTUND,	/* 32 - DNML | UFL | IMP */
264 	FPE_FLTINV,	/* 33 - INV | DNML | UFL | IMP */
265 	FPE_FLTDIV,	/* 34 - DZ | UFL | IMP */
266 	FPE_FLTINV,	/* 35 - INV | DZ | UFL | IMP */
267 	FPE_FLTDIV,	/* 36 - DNML | DZ | UFL | IMP */
268 	FPE_FLTINV,	/* 37 - INV | DNML | DZ | UFL | IMP */
269 	FPE_FLTOVF,	/* 38 - OFL | UFL | IMP */
270 	FPE_FLTINV,	/* 39 - INV | OFL | UFL | IMP */
271 	FPE_FLTUND,	/* 3A - DNML | OFL | UFL | IMP */
272 	FPE_FLTINV,	/* 3B - INV | DNML | OFL | UFL | IMP */
273 	FPE_FLTDIV,	/* 3C - DZ | OFL | UFL | IMP */
274 	FPE_FLTINV,	/* 3D - INV | DZ | OFL | UFL | IMP */
275 	FPE_FLTDIV,	/* 3E - DNML | DZ | OFL | UFL | IMP */
276 	FPE_FLTINV,	/* 3F - INV | DNML | DZ | OFL | UFL | IMP */
277 	FPE_FLTSUB,	/* 40 - STK */
278 	FPE_FLTSUB,	/* 41 - INV | STK */
279 	FPE_FLTUND,	/* 42 - DNML | STK */
280 	FPE_FLTSUB,	/* 43 - INV | DNML | STK */
281 	FPE_FLTDIV,	/* 44 - DZ | STK */
282 	FPE_FLTSUB,	/* 45 - INV | DZ | STK */
283 	FPE_FLTDIV,	/* 46 - DNML | DZ | STK */
284 	FPE_FLTSUB,	/* 47 - INV | DNML | DZ | STK */
285 	FPE_FLTOVF,	/* 48 - OFL | STK */
286 	FPE_FLTSUB,	/* 49 - INV | OFL | STK */
287 	FPE_FLTUND,	/* 4A - DNML | OFL | STK */
288 	FPE_FLTSUB,	/* 4B - INV | DNML | OFL | STK */
289 	FPE_FLTDIV,	/* 4C - DZ | OFL | STK */
290 	FPE_FLTSUB,	/* 4D - INV | DZ | OFL | STK */
291 	FPE_FLTDIV,	/* 4E - DNML | DZ | OFL | STK */
292 	FPE_FLTSUB,	/* 4F - INV | DNML | DZ | OFL | STK */
293 	FPE_FLTUND,	/* 50 - UFL | STK */
294 	FPE_FLTSUB,	/* 51 - INV | UFL | STK */
295 	FPE_FLTUND,	/* 52 - DNML | UFL | STK */
296 	FPE_FLTSUB,	/* 53 - INV | DNML | UFL | STK */
297 	FPE_FLTDIV,	/* 54 - DZ | UFL | STK */
298 	FPE_FLTSUB,	/* 55 - INV | DZ | UFL | STK */
299 	FPE_FLTDIV,	/* 56 - DNML | DZ | UFL | STK */
300 	FPE_FLTSUB,	/* 57 - INV | DNML | DZ | UFL | STK */
301 	FPE_FLTOVF,	/* 58 - OFL | UFL | STK */
302 	FPE_FLTSUB,	/* 59 - INV | OFL | UFL | STK */
303 	FPE_FLTUND,	/* 5A - DNML | OFL | UFL | STK */
304 	FPE_FLTSUB,	/* 5B - INV | DNML | OFL | UFL | STK */
305 	FPE_FLTDIV,	/* 5C - DZ | OFL | UFL | STK */
306 	FPE_FLTSUB,	/* 5D - INV | DZ | OFL | UFL | STK */
307 	FPE_FLTDIV,	/* 5E - DNML | DZ | OFL | UFL | STK */
308 	FPE_FLTSUB,	/* 5F - INV | DNML | DZ | OFL | UFL | STK */
309 	FPE_FLTRES,	/* 60 - IMP | STK */
310 	FPE_FLTSUB,	/* 61 - INV | IMP | STK */
311 	FPE_FLTUND,	/* 62 - DNML | IMP | STK */
312 	FPE_FLTSUB,	/* 63 - INV | DNML | IMP | STK */
313 	FPE_FLTDIV,	/* 64 - DZ | IMP | STK */
314 	FPE_FLTSUB,	/* 65 - INV | DZ | IMP | STK */
315 	FPE_FLTDIV,	/* 66 - DNML | DZ | IMP | STK */
316 	FPE_FLTSUB,	/* 67 - INV | DNML | DZ | IMP | STK */
317 	FPE_FLTOVF,	/* 68 - OFL | IMP | STK */
318 	FPE_FLTSUB,	/* 69 - INV | OFL | IMP | STK */
319 	FPE_FLTUND,	/* 6A - DNML | OFL | IMP | STK */
320 	FPE_FLTSUB,	/* 6B - INV | DNML | OFL | IMP | STK */
321 	FPE_FLTDIV,	/* 6C - DZ | OFL | IMP | STK */
322 	FPE_FLTSUB,	/* 6D - INV | DZ | OFL | IMP | STK */
323 	FPE_FLTDIV,	/* 6E - DNML | DZ | OFL | IMP | STK */
324 	FPE_FLTSUB,	/* 6F - INV | DNML | DZ | OFL | IMP | STK */
325 	FPE_FLTUND,	/* 70 - UFL | IMP | STK */
326 	FPE_FLTSUB,	/* 71 - INV | UFL | IMP | STK */
327 	FPE_FLTUND,	/* 72 - DNML | UFL | IMP | STK */
328 	FPE_FLTSUB,	/* 73 - INV | DNML | UFL | IMP | STK */
329 	FPE_FLTDIV,	/* 74 - DZ | UFL | IMP | STK */
330 	FPE_FLTSUB,	/* 75 - INV | DZ | UFL | IMP | STK */
331 	FPE_FLTDIV,	/* 76 - DNML | DZ | UFL | IMP | STK */
332 	FPE_FLTSUB,	/* 77 - INV | DNML | DZ | UFL | IMP | STK */
333 	FPE_FLTOVF,	/* 78 - OFL | UFL | IMP | STK */
334 	FPE_FLTSUB,	/* 79 - INV | OFL | UFL | IMP | STK */
335 	FPE_FLTUND,	/* 7A - DNML | OFL | UFL | IMP | STK */
336 	FPE_FLTSUB,	/* 7B - INV | DNML | OFL | UFL | IMP | STK */
337 	FPE_FLTDIV,	/* 7C - DZ | OFL | UFL | IMP | STK */
338 	FPE_FLTSUB,	/* 7D - INV | DZ | OFL | UFL | IMP | STK */
339 	FPE_FLTDIV,	/* 7E - DNML | DZ | OFL | UFL | IMP | STK */
340 	FPE_FLTSUB,	/* 7F - INV | DNML | DZ | OFL | UFL | IMP | STK */
341 };
342 #endif
343 
344 /*
345  * Implement the device not available (DNA) exception.  gd_npxthread had
346  * better be NULL.  Restore the current thread's FP state and set gd_npxthread
347  * to curthread.
348  *
349  * Interrupts are enabled and preemption can occur.  Enter a critical
350  * section to stabilize the FP state.
351  */
352 int
353 npxdna(struct trapframe *frame)
354 {
355 	thread_t td = curthread;
356 	u_long *exstat;
357 	int didinit = 0;
358 
359 	if (mdcpu->gd_npxthread != NULL) {
360 		kprintf("npxdna: npxthread = %p, curthread = %p\n",
361 		       mdcpu->gd_npxthread, td);
362 		panic("npxdna");
363 	}
364 
365 	/*
366 	 * Setup the initial saved state if the thread has never before
367 	 * used the FP unit.  This also occurs when a thread pushes a
368 	 * signal handler and uses FP in the handler.
369 	 */
370 	if ((curthread->td_flags & TDF_USINGFP) == 0) {
371 		curthread->td_flags |= TDF_USINGFP;
372 		npxinit(__INITIAL_NPXCW__);
373 		didinit = 1;
374 	}
375 
376 	/*
377 	 * The setting of gd_npxthread and the call to fpurstor() must not
378 	 * be preempted by an interrupt thread or we will take an npxdna
379 	 * trap and potentially save our current fpstate (which is garbage)
380 	 * and then restore the garbage rather then the originally saved
381 	 * fpstate.
382 	 */
383 	crit_enter();
384 	/*stop_emulating();*/
385 	/*
386 	 * Record new context early in case frstor causes an IRQ13.
387 	 */
388 	mdcpu->gd_npxthread = td;
389 	exstat = GET_FPU_EXSW_PTR(td);
390 	*exstat = 0;
391 	/*
392 	 * The following frstor may cause an IRQ13 when the state being
393 	 * restored has a pending error.  The error will appear to have been
394 	 * triggered by the current (npx) user instruction even when that
395 	 * instruction is a no-wait instruction that should not trigger an
396 	 * error (e.g., fnclex).  On at least one 486 system all of the
397 	 * no-wait instructions are broken the same as frstor, so our
398 	 * treatment does not amplify the breakage.  On at least one
399 	 * 386/Cyrix 387 system, fnclex works correctly while frstor and
400 	 * fnsave are broken, so our treatment breaks fnclex if it is the
401 	 * first FPU instruction after a context switch.
402 	 */
403 	if ((td->td_savefpu->sv_xmm.sv_env.en_mxcsr & ~0xFFBF) && cpu_fxsr) {
404 		krateprintf(&badfprate,
405 			    "FXRSTR: illegal FP MXCSR %08x didinit = %d\n",
406 			    td->td_savefpu->sv_xmm.sv_env.en_mxcsr, didinit);
407 		td->td_savefpu->sv_xmm.sv_env.en_mxcsr &= 0xFFBF;
408 		lwpsignal(curproc, curthread->td_lwp, SIGFPE);
409 	}
410 	fpurstor(curthread->td_savefpu);
411 	crit_exit();
412 
413 	return (1);
414 }
415 
416 /*
417  * Wrapper for the fnsave instruction to handle h/w bugs.  If there is an error
418  * pending, then fnsave generates a bogus IRQ13 on some systems.  Force
419  * any IRQ13 to be handled immediately, and then ignore it.  This routine is
420  * often called at splhigh so it must not use many system services.  In
421  * particular, it's much easier to install a special handler than to
422  * guarantee that it's safe to use npxintr() and its supporting code.
423  *
424  * WARNING!  This call is made during a switch and the MP lock will be
425  * setup for the new target thread rather then the current thread, so we
426  * cannot do anything here that depends on the *_mplock() functions as
427  * we may trip over their assertions.
428  *
429  * WARNING!  When using fxsave we MUST fninit after saving the FP state.  The
430  * kernel will always assume that the FP state is 'safe' (will not cause
431  * exceptions) for mmx/xmm use if npxthread is NULL.  The kernel must still
432  * setup a custom save area before actually using the FP unit, but it will
433  * not bother calling fninit.  This greatly improves kernel performance when
434  * it wishes to use the FP unit.
435  */
436 void
437 npxsave(union savefpu *addr)
438 {
439 	crit_enter();
440 	/*stop_emulating();*/
441 	fpusave(addr);
442 	mdcpu->gd_npxthread = NULL;
443 	fninit();
444 	/*start_emulating();*/
445 	crit_exit();
446 }
447 
448 static void
449 fpusave(union savefpu *addr)
450 {
451 	if (cpu_fxsr)
452 		fxsave(addr);
453 	else
454 		fnsave(addr);
455 }
456 
457 /*
458  * Save the FP state to the mcontext structure.
459  *
460  * WARNING: If you want to try to npxsave() directly to mctx->mc_fpregs,
461  * then it MUST be 16-byte aligned.  Currently this is not guarenteed.
462  */
463 void
464 npxpush(mcontext_t *mctx)
465 {
466 	thread_t td = curthread;
467 
468 	if (td->td_flags & TDF_USINGFP) {
469 		if (mdcpu->gd_npxthread == td) {
470 			/*
471 			 * XXX Note: This is a bit inefficient if the signal
472 			 * handler uses floating point, extra faults will
473 			 * occur.
474 			 */
475 			mctx->mc_ownedfp = _MC_FPOWNED_FPU;
476 			npxsave(td->td_savefpu);
477 		} else {
478 			mctx->mc_ownedfp = _MC_FPOWNED_PCB;
479 		}
480 		bcopy(td->td_savefpu, mctx->mc_fpregs, sizeof(mctx->mc_fpregs));
481 		td->td_flags &= ~TDF_USINGFP;
482 		mctx->mc_fpformat =
483 #ifndef CPU_DISABLE_SSE
484 			(cpu_fxsr) ? _MC_FPFMT_XMM :
485 #endif
486 			_MC_FPFMT_387;
487 	} else {
488 		mctx->mc_ownedfp = _MC_FPOWNED_NONE;
489 		mctx->mc_fpformat = _MC_FPFMT_NODEV;
490 	}
491 }
492 
493 /*
494  * Restore the FP state from the mcontext structure.
495  */
496 void
497 npxpop(mcontext_t *mctx)
498 {
499 	thread_t td = curthread;
500 
501 	switch(mctx->mc_ownedfp) {
502 	case _MC_FPOWNED_NONE:
503 		/*
504 		 * If the signal handler used the FP unit but the interrupted
505 		 * code did not, release the FP unit.  Clear TDF_USINGFP will
506 		 * force the FP unit to reinit so the interrupted code sees
507 		 * a clean slate.
508 		 */
509 		if (td->td_flags & TDF_USINGFP) {
510 			if (td == mdcpu->gd_npxthread)
511 				npxsave(td->td_savefpu);
512 			td->td_flags &= ~TDF_USINGFP;
513 		}
514 		break;
515 	case _MC_FPOWNED_FPU:
516 	case _MC_FPOWNED_PCB:
517 		/*
518 		 * Clear ownership of the FP unit and restore our saved state.
519 		 *
520 		 * NOTE: The signal handler may have set-up some FP state and
521 		 * enabled the FP unit, so we have to restore no matter what.
522 		 *
523 		 * XXX: This is bit inefficient, if the code being returned
524 		 * to is actively using the FP this results in multiple
525 		 * kernel faults.
526 		 *
527 		 * WARNING: The saved state was exposed to userland and may
528 		 * have to be sanitized to avoid a GP fault in the kernel.
529 		 */
530 		if (td == mdcpu->gd_npxthread)
531 			npxsave(td->td_savefpu);
532 		bcopy(mctx->mc_fpregs, td->td_savefpu, sizeof(*td->td_savefpu));
533 		if ((td->td_savefpu->sv_xmm.sv_env.en_mxcsr & ~0xFFBF) &&
534 		    cpu_fxsr) {
535 			krateprintf(&badfprate,
536 				    "pid %d (%s) signal return from user: "
537 				    "illegal FP MXCSR %08x\n",
538 				    td->td_proc->p_pid,
539 				    td->td_proc->p_comm,
540 				    td->td_savefpu->sv_xmm.sv_env.en_mxcsr);
541 			td->td_savefpu->sv_xmm.sv_env.en_mxcsr &= 0xFFBF;
542 		}
543 		td->td_flags |= TDF_USINGFP;
544 		break;
545 	}
546 }
547 
548 
549 #ifndef CPU_DISABLE_SSE
550 /*
551  * On AuthenticAMD processors, the fxrstor instruction does not restore
552  * the x87's stored last instruction pointer, last data pointer, and last
553  * opcode values, except in the rare case in which the exception summary
554  * (ES) bit in the x87 status word is set to 1.
555  *
556  * In order to avoid leaking this information across processes, we clean
557  * these values by performing a dummy load before executing fxrstor().
558  */
559 static	double	dummy_variable = 0.0;
560 static void
561 fpu_clean_state(void)
562 {
563 	u_short status;
564 
565 	/*
566 	 * Clear the ES bit in the x87 status word if it is currently
567 	 * set, in order to avoid causing a fault in the upcoming load.
568 	 */
569 	fnstsw(&status);
570 	if (status & 0x80)
571 		fnclex();
572 
573 	/*
574 	 * Load the dummy variable into the x87 stack.  This mangles
575 	 * the x87 stack, but we don't care since we're about to call
576 	 * fxrstor() anyway.
577 	 */
578 	__asm __volatile("ffree %%st(7); fld %0" : : "m" (dummy_variable));
579 }
580 #endif /* CPU_DISABLE_SSE */
581 
582 static void
583 fpurstor(union savefpu *addr)
584 {
585 #ifndef CPU_DISABLE_SSE
586 	if (cpu_fxsr) {
587 		fpu_clean_state();
588 		fxrstor(addr);
589 	} else {
590 		frstor(addr);
591 	}
592 #else
593 	frstor(addr);
594 #endif
595 }
596