xref: /freebsd/sys/amd64/amd64/fpu.c (revision 325151a3)
1 /*-
2  * Copyright (c) 1990 William Jolitz.
3  * Copyright (c) 1991 The Regents of the University of California.
4  * All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  * 4. Neither the name of the University nor the names of its contributors
15  *    may be used to endorse or promote products derived from this software
16  *    without specific prior written permission.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  *
30  *	from: @(#)npx.c	7.2 (Berkeley) 5/12/91
31  */
32 
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD$");
35 
36 #include <sys/param.h>
37 #include <sys/systm.h>
38 #include <sys/bus.h>
39 #include <sys/kernel.h>
40 #include <sys/lock.h>
41 #include <sys/malloc.h>
42 #include <sys/module.h>
43 #include <sys/mutex.h>
44 #include <sys/mutex.h>
45 #include <sys/proc.h>
46 #include <sys/sysctl.h>
47 #include <machine/bus.h>
48 #include <sys/rman.h>
49 #include <sys/signalvar.h>
50 #include <vm/uma.h>
51 
52 #include <machine/cputypes.h>
53 #include <machine/frame.h>
54 #include <machine/intr_machdep.h>
55 #include <machine/md_var.h>
56 #include <machine/pcb.h>
57 #include <machine/psl.h>
58 #include <machine/resource.h>
59 #include <machine/specialreg.h>
60 #include <machine/segments.h>
61 #include <machine/ucontext.h>
62 
63 /*
64  * Floating point support.
65  */
66 
67 #if defined(__GNUCLIKE_ASM) && !defined(lint)
68 
69 #define	fldcw(cw)		__asm __volatile("fldcw %0" : : "m" (cw))
70 #define	fnclex()		__asm __volatile("fnclex")
71 #define	fninit()		__asm __volatile("fninit")
72 #define	fnstcw(addr)		__asm __volatile("fnstcw %0" : "=m" (*(addr)))
73 #define	fnstsw(addr)		__asm __volatile("fnstsw %0" : "=am" (*(addr)))
74 #define	fxrstor(addr)		__asm __volatile("fxrstor %0" : : "m" (*(addr)))
75 #define	fxsave(addr)		__asm __volatile("fxsave %0" : "=m" (*(addr)))
76 #define	ldmxcsr(csr)		__asm __volatile("ldmxcsr %0" : : "m" (csr))
77 #define	stmxcsr(addr)		__asm __volatile("stmxcsr %0" : : "m" (*(addr)))
78 
79 static __inline void
80 xrstor(char *addr, uint64_t mask)
81 {
82 	uint32_t low, hi;
83 
84 	low = mask;
85 	hi = mask >> 32;
86 	__asm __volatile("xrstor %0" : : "m" (*addr), "a" (low), "d" (hi));
87 }
88 
89 static __inline void
90 xsave(char *addr, uint64_t mask)
91 {
92 	uint32_t low, hi;
93 
94 	low = mask;
95 	hi = mask >> 32;
96 	__asm __volatile("xsave %0" : "=m" (*addr) : "a" (low), "d" (hi) :
97 	    "memory");
98 }
99 
100 #else	/* !(__GNUCLIKE_ASM && !lint) */
101 
102 void	fldcw(u_short cw);
103 void	fnclex(void);
104 void	fninit(void);
105 void	fnstcw(caddr_t addr);
106 void	fnstsw(caddr_t addr);
107 void	fxsave(caddr_t addr);
108 void	fxrstor(caddr_t addr);
109 void	ldmxcsr(u_int csr);
110 void	stmxcsr(u_int *csr);
111 void	xrstor(char *addr, uint64_t mask);
112 void	xsave(char *addr, uint64_t mask);
113 
114 #endif	/* __GNUCLIKE_ASM && !lint */
115 
116 #define	start_emulating()	load_cr0(rcr0() | CR0_TS)
117 #define	stop_emulating()	clts()
118 
119 CTASSERT(sizeof(struct savefpu) == 512);
120 CTASSERT(sizeof(struct xstate_hdr) == 64);
121 CTASSERT(sizeof(struct savefpu_ymm) == 832);
122 
123 /*
124  * This requirement is to make it easier for asm code to calculate
125  * offset of the fpu save area from the pcb address. FPU save area
126  * must be 64-byte aligned.
127  */
128 CTASSERT(sizeof(struct pcb) % XSAVE_AREA_ALIGN == 0);
129 
130 /*
131  * Ensure the copy of XCR0 saved in a core is contained in the padding
132  * area.
133  */
134 CTASSERT(X86_XSTATE_XCR0_OFFSET >= offsetof(struct savefpu, sv_pad) &&
135     X86_XSTATE_XCR0_OFFSET + sizeof(uint64_t) <= sizeof(struct savefpu));
136 
137 static	void	fpu_clean_state(void);
138 
139 SYSCTL_INT(_hw, HW_FLOATINGPT, floatingpoint, CTLFLAG_RD,
140     SYSCTL_NULL_INT_PTR, 1, "Floating point instructions executed in hardware");
141 
142 int use_xsave;			/* non-static for cpu_switch.S */
143 uint64_t xsave_mask;		/* the same */
144 static	uma_zone_t fpu_save_area_zone;
145 static	struct savefpu *fpu_initialstate;
146 
147 struct xsave_area_elm_descr {
148 	u_int	offset;
149 	u_int	size;
150 } *xsave_area_desc;
151 
152 void
153 fpusave(void *addr)
154 {
155 
156 	if (use_xsave)
157 		xsave((char *)addr, xsave_mask);
158 	else
159 		fxsave((char *)addr);
160 }
161 
162 void
163 fpurestore(void *addr)
164 {
165 
166 	if (use_xsave)
167 		xrstor((char *)addr, xsave_mask);
168 	else
169 		fxrstor((char *)addr);
170 }
171 
172 void
173 fpususpend(void *addr)
174 {
175 	u_long cr0;
176 
177 	cr0 = rcr0();
178 	stop_emulating();
179 	fpusave(addr);
180 	load_cr0(cr0);
181 }
182 
183 void
184 fpuresume(void *addr)
185 {
186 	u_long cr0;
187 
188 	cr0 = rcr0();
189 	stop_emulating();
190 	fninit();
191 	if (use_xsave)
192 		load_xcr(XCR0, xsave_mask);
193 	fpurestore(addr);
194 	load_cr0(cr0);
195 }
196 
197 /*
198  * Enable XSAVE if supported and allowed by user.
199  * Calculate the xsave_mask.
200  */
201 static void
202 fpuinit_bsp1(void)
203 {
204 	u_int cp[4];
205 	uint64_t xsave_mask_user;
206 
207 	if ((cpu_feature2 & CPUID2_XSAVE) != 0) {
208 		use_xsave = 1;
209 		TUNABLE_INT_FETCH("hw.use_xsave", &use_xsave);
210 	}
211 	if (!use_xsave)
212 		return;
213 
214 	cpuid_count(0xd, 0x0, cp);
215 	xsave_mask = XFEATURE_ENABLED_X87 | XFEATURE_ENABLED_SSE;
216 	if ((cp[0] & xsave_mask) != xsave_mask)
217 		panic("CPU0 does not support X87 or SSE: %x", cp[0]);
218 	xsave_mask = ((uint64_t)cp[3] << 32) | cp[0];
219 	xsave_mask_user = xsave_mask;
220 	TUNABLE_ULONG_FETCH("hw.xsave_mask", &xsave_mask_user);
221 	xsave_mask_user |= XFEATURE_ENABLED_X87 | XFEATURE_ENABLED_SSE;
222 	xsave_mask &= xsave_mask_user;
223 	if ((xsave_mask & XFEATURE_AVX512) != XFEATURE_AVX512)
224 		xsave_mask &= ~XFEATURE_AVX512;
225 	if ((xsave_mask & XFEATURE_MPX) != XFEATURE_MPX)
226 		xsave_mask &= ~XFEATURE_MPX;
227 
228 	cpuid_count(0xd, 0x1, cp);
229 	if ((cp[0] & CPUID_EXTSTATE_XSAVEOPT) != 0) {
230 		/*
231 		 * Patch the XSAVE instruction in the cpu_switch code
232 		 * to XSAVEOPT.  We assume that XSAVE encoding used
233 		 * REX byte, and set the bit 4 of the r/m byte.
234 		 */
235 		ctx_switch_xsave[3] |= 0x10;
236 	}
237 }
238 
239 /*
240  * Calculate the fpu save area size.
241  */
242 static void
243 fpuinit_bsp2(void)
244 {
245 	u_int cp[4];
246 
247 	if (use_xsave) {
248 		cpuid_count(0xd, 0x0, cp);
249 		cpu_max_ext_state_size = cp[1];
250 
251 		/*
252 		 * Reload the cpu_feature2, since we enabled OSXSAVE.
253 		 */
254 		do_cpuid(1, cp);
255 		cpu_feature2 = cp[2];
256 	} else
257 		cpu_max_ext_state_size = sizeof(struct savefpu);
258 }
259 
260 /*
261  * Initialize the floating point unit.
262  */
263 void
264 fpuinit(void)
265 {
266 	register_t saveintr;
267 	u_int mxcsr;
268 	u_short control;
269 
270 	if (IS_BSP())
271 		fpuinit_bsp1();
272 
273 	if (use_xsave) {
274 		load_cr4(rcr4() | CR4_XSAVE);
275 		load_xcr(XCR0, xsave_mask);
276 	}
277 
278 	/*
279 	 * XCR0 shall be set up before CPU can report the save area size.
280 	 */
281 	if (IS_BSP())
282 		fpuinit_bsp2();
283 
284 	/*
285 	 * It is too early for critical_enter() to work on AP.
286 	 */
287 	saveintr = intr_disable();
288 	stop_emulating();
289 	fninit();
290 	control = __INITIAL_FPUCW__;
291 	fldcw(control);
292 	mxcsr = __INITIAL_MXCSR__;
293 	ldmxcsr(mxcsr);
294 	start_emulating();
295 	intr_restore(saveintr);
296 }
297 
298 /*
299  * On the boot CPU we generate a clean state that is used to
300  * initialize the floating point unit when it is first used by a
301  * process.
302  */
303 static void
304 fpuinitstate(void *arg __unused)
305 {
306 	register_t saveintr;
307 	int cp[4], i, max_ext_n;
308 
309 	fpu_initialstate = malloc(cpu_max_ext_state_size, M_DEVBUF,
310 	    M_WAITOK | M_ZERO);
311 	saveintr = intr_disable();
312 	stop_emulating();
313 
314 	fpusave(fpu_initialstate);
315 	if (fpu_initialstate->sv_env.en_mxcsr_mask)
316 		cpu_mxcsr_mask = fpu_initialstate->sv_env.en_mxcsr_mask;
317 	else
318 		cpu_mxcsr_mask = 0xFFBF;
319 
320 	/*
321 	 * The fninit instruction does not modify XMM registers.  The
322 	 * fpusave call dumped the garbage contained in the registers
323 	 * after reset to the initial state saved.  Clear XMM
324 	 * registers file image to make the startup program state and
325 	 * signal handler XMM register content predictable.
326 	 */
327 	bzero(&fpu_initialstate->sv_xmm[0], sizeof(struct xmmacc));
328 
329 	/*
330 	 * Create a table describing the layout of the CPU Extended
331 	 * Save Area.
332 	 */
333 	if (use_xsave) {
334 		max_ext_n = flsl(xsave_mask);
335 		xsave_area_desc = malloc(max_ext_n * sizeof(struct
336 		    xsave_area_elm_descr), M_DEVBUF, M_WAITOK | M_ZERO);
337 		/* x87 state */
338 		xsave_area_desc[0].offset = 0;
339 		xsave_area_desc[0].size = 160;
340 		/* XMM */
341 		xsave_area_desc[1].offset = 160;
342 		xsave_area_desc[1].size = 288 - 160;
343 
344 		for (i = 2; i < max_ext_n; i++) {
345 			cpuid_count(0xd, i, cp);
346 			xsave_area_desc[i].offset = cp[1];
347 			xsave_area_desc[i].size = cp[0];
348 		}
349 	}
350 
351 	fpu_save_area_zone = uma_zcreate("FPU_save_area",
352 	    cpu_max_ext_state_size, NULL, NULL, NULL, NULL,
353 	    XSAVE_AREA_ALIGN - 1, 0);
354 
355 	start_emulating();
356 	intr_restore(saveintr);
357 }
358 SYSINIT(fpuinitstate, SI_SUB_DRIVERS, SI_ORDER_ANY, fpuinitstate, NULL);
359 
360 /*
361  * Free coprocessor (if we have it).
362  */
363 void
364 fpuexit(struct thread *td)
365 {
366 
367 	critical_enter();
368 	if (curthread == PCPU_GET(fpcurthread)) {
369 		stop_emulating();
370 		fpusave(curpcb->pcb_save);
371 		start_emulating();
372 		PCPU_SET(fpcurthread, NULL);
373 	}
374 	critical_exit();
375 }
376 
377 int
378 fpuformat()
379 {
380 
381 	return (_MC_FPFMT_XMM);
382 }
383 
384 /*
385  * The following mechanism is used to ensure that the FPE_... value
386  * that is passed as a trapcode to the signal handler of the user
387  * process does not have more than one bit set.
388  *
389  * Multiple bits may be set if the user process modifies the control
390  * word while a status word bit is already set.  While this is a sign
391  * of bad coding, we have no choise than to narrow them down to one
392  * bit, since we must not send a trapcode that is not exactly one of
393  * the FPE_ macros.
394  *
395  * The mechanism has a static table with 127 entries.  Each combination
396  * of the 7 FPU status word exception bits directly translates to a
397  * position in this table, where a single FPE_... value is stored.
398  * This FPE_... value stored there is considered the "most important"
399  * of the exception bits and will be sent as the signal code.  The
400  * precedence of the bits is based upon Intel Document "Numerical
401  * Applications", Chapter "Special Computational Situations".
402  *
403  * The macro to choose one of these values does these steps: 1) Throw
404  * away status word bits that cannot be masked.  2) Throw away the bits
405  * currently masked in the control word, assuming the user isn't
406  * interested in them anymore.  3) Reinsert status word bit 7 (stack
407  * fault) if it is set, which cannot be masked but must be presered.
408  * 4) Use the remaining bits to point into the trapcode table.
409  *
410  * The 6 maskable bits in order of their preference, as stated in the
411  * above referenced Intel manual:
412  * 1  Invalid operation (FP_X_INV)
413  * 1a   Stack underflow
414  * 1b   Stack overflow
415  * 1c   Operand of unsupported format
416  * 1d   SNaN operand.
417  * 2  QNaN operand (not an exception, irrelavant here)
418  * 3  Any other invalid-operation not mentioned above or zero divide
419  *      (FP_X_INV, FP_X_DZ)
420  * 4  Denormal operand (FP_X_DNML)
421  * 5  Numeric over/underflow (FP_X_OFL, FP_X_UFL)
422  * 6  Inexact result (FP_X_IMP)
423  */
424 static char fpetable[128] = {
425 	0,
426 	FPE_FLTINV,	/*  1 - INV */
427 	FPE_FLTUND,	/*  2 - DNML */
428 	FPE_FLTINV,	/*  3 - INV | DNML */
429 	FPE_FLTDIV,	/*  4 - DZ */
430 	FPE_FLTINV,	/*  5 - INV | DZ */
431 	FPE_FLTDIV,	/*  6 - DNML | DZ */
432 	FPE_FLTINV,	/*  7 - INV | DNML | DZ */
433 	FPE_FLTOVF,	/*  8 - OFL */
434 	FPE_FLTINV,	/*  9 - INV | OFL */
435 	FPE_FLTUND,	/*  A - DNML | OFL */
436 	FPE_FLTINV,	/*  B - INV | DNML | OFL */
437 	FPE_FLTDIV,	/*  C - DZ | OFL */
438 	FPE_FLTINV,	/*  D - INV | DZ | OFL */
439 	FPE_FLTDIV,	/*  E - DNML | DZ | OFL */
440 	FPE_FLTINV,	/*  F - INV | DNML | DZ | OFL */
441 	FPE_FLTUND,	/* 10 - UFL */
442 	FPE_FLTINV,	/* 11 - INV | UFL */
443 	FPE_FLTUND,	/* 12 - DNML | UFL */
444 	FPE_FLTINV,	/* 13 - INV | DNML | UFL */
445 	FPE_FLTDIV,	/* 14 - DZ | UFL */
446 	FPE_FLTINV,	/* 15 - INV | DZ | UFL */
447 	FPE_FLTDIV,	/* 16 - DNML | DZ | UFL */
448 	FPE_FLTINV,	/* 17 - INV | DNML | DZ | UFL */
449 	FPE_FLTOVF,	/* 18 - OFL | UFL */
450 	FPE_FLTINV,	/* 19 - INV | OFL | UFL */
451 	FPE_FLTUND,	/* 1A - DNML | OFL | UFL */
452 	FPE_FLTINV,	/* 1B - INV | DNML | OFL | UFL */
453 	FPE_FLTDIV,	/* 1C - DZ | OFL | UFL */
454 	FPE_FLTINV,	/* 1D - INV | DZ | OFL | UFL */
455 	FPE_FLTDIV,	/* 1E - DNML | DZ | OFL | UFL */
456 	FPE_FLTINV,	/* 1F - INV | DNML | DZ | OFL | UFL */
457 	FPE_FLTRES,	/* 20 - IMP */
458 	FPE_FLTINV,	/* 21 - INV | IMP */
459 	FPE_FLTUND,	/* 22 - DNML | IMP */
460 	FPE_FLTINV,	/* 23 - INV | DNML | IMP */
461 	FPE_FLTDIV,	/* 24 - DZ | IMP */
462 	FPE_FLTINV,	/* 25 - INV | DZ | IMP */
463 	FPE_FLTDIV,	/* 26 - DNML | DZ | IMP */
464 	FPE_FLTINV,	/* 27 - INV | DNML | DZ | IMP */
465 	FPE_FLTOVF,	/* 28 - OFL | IMP */
466 	FPE_FLTINV,	/* 29 - INV | OFL | IMP */
467 	FPE_FLTUND,	/* 2A - DNML | OFL | IMP */
468 	FPE_FLTINV,	/* 2B - INV | DNML | OFL | IMP */
469 	FPE_FLTDIV,	/* 2C - DZ | OFL | IMP */
470 	FPE_FLTINV,	/* 2D - INV | DZ | OFL | IMP */
471 	FPE_FLTDIV,	/* 2E - DNML | DZ | OFL | IMP */
472 	FPE_FLTINV,	/* 2F - INV | DNML | DZ | OFL | IMP */
473 	FPE_FLTUND,	/* 30 - UFL | IMP */
474 	FPE_FLTINV,	/* 31 - INV | UFL | IMP */
475 	FPE_FLTUND,	/* 32 - DNML | UFL | IMP */
476 	FPE_FLTINV,	/* 33 - INV | DNML | UFL | IMP */
477 	FPE_FLTDIV,	/* 34 - DZ | UFL | IMP */
478 	FPE_FLTINV,	/* 35 - INV | DZ | UFL | IMP */
479 	FPE_FLTDIV,	/* 36 - DNML | DZ | UFL | IMP */
480 	FPE_FLTINV,	/* 37 - INV | DNML | DZ | UFL | IMP */
481 	FPE_FLTOVF,	/* 38 - OFL | UFL | IMP */
482 	FPE_FLTINV,	/* 39 - INV | OFL | UFL | IMP */
483 	FPE_FLTUND,	/* 3A - DNML | OFL | UFL | IMP */
484 	FPE_FLTINV,	/* 3B - INV | DNML | OFL | UFL | IMP */
485 	FPE_FLTDIV,	/* 3C - DZ | OFL | UFL | IMP */
486 	FPE_FLTINV,	/* 3D - INV | DZ | OFL | UFL | IMP */
487 	FPE_FLTDIV,	/* 3E - DNML | DZ | OFL | UFL | IMP */
488 	FPE_FLTINV,	/* 3F - INV | DNML | DZ | OFL | UFL | IMP */
489 	FPE_FLTSUB,	/* 40 - STK */
490 	FPE_FLTSUB,	/* 41 - INV | STK */
491 	FPE_FLTUND,	/* 42 - DNML | STK */
492 	FPE_FLTSUB,	/* 43 - INV | DNML | STK */
493 	FPE_FLTDIV,	/* 44 - DZ | STK */
494 	FPE_FLTSUB,	/* 45 - INV | DZ | STK */
495 	FPE_FLTDIV,	/* 46 - DNML | DZ | STK */
496 	FPE_FLTSUB,	/* 47 - INV | DNML | DZ | STK */
497 	FPE_FLTOVF,	/* 48 - OFL | STK */
498 	FPE_FLTSUB,	/* 49 - INV | OFL | STK */
499 	FPE_FLTUND,	/* 4A - DNML | OFL | STK */
500 	FPE_FLTSUB,	/* 4B - INV | DNML | OFL | STK */
501 	FPE_FLTDIV,	/* 4C - DZ | OFL | STK */
502 	FPE_FLTSUB,	/* 4D - INV | DZ | OFL | STK */
503 	FPE_FLTDIV,	/* 4E - DNML | DZ | OFL | STK */
504 	FPE_FLTSUB,	/* 4F - INV | DNML | DZ | OFL | STK */
505 	FPE_FLTUND,	/* 50 - UFL | STK */
506 	FPE_FLTSUB,	/* 51 - INV | UFL | STK */
507 	FPE_FLTUND,	/* 52 - DNML | UFL | STK */
508 	FPE_FLTSUB,	/* 53 - INV | DNML | UFL | STK */
509 	FPE_FLTDIV,	/* 54 - DZ | UFL | STK */
510 	FPE_FLTSUB,	/* 55 - INV | DZ | UFL | STK */
511 	FPE_FLTDIV,	/* 56 - DNML | DZ | UFL | STK */
512 	FPE_FLTSUB,	/* 57 - INV | DNML | DZ | UFL | STK */
513 	FPE_FLTOVF,	/* 58 - OFL | UFL | STK */
514 	FPE_FLTSUB,	/* 59 - INV | OFL | UFL | STK */
515 	FPE_FLTUND,	/* 5A - DNML | OFL | UFL | STK */
516 	FPE_FLTSUB,	/* 5B - INV | DNML | OFL | UFL | STK */
517 	FPE_FLTDIV,	/* 5C - DZ | OFL | UFL | STK */
518 	FPE_FLTSUB,	/* 5D - INV | DZ | OFL | UFL | STK */
519 	FPE_FLTDIV,	/* 5E - DNML | DZ | OFL | UFL | STK */
520 	FPE_FLTSUB,	/* 5F - INV | DNML | DZ | OFL | UFL | STK */
521 	FPE_FLTRES,	/* 60 - IMP | STK */
522 	FPE_FLTSUB,	/* 61 - INV | IMP | STK */
523 	FPE_FLTUND,	/* 62 - DNML | IMP | STK */
524 	FPE_FLTSUB,	/* 63 - INV | DNML | IMP | STK */
525 	FPE_FLTDIV,	/* 64 - DZ | IMP | STK */
526 	FPE_FLTSUB,	/* 65 - INV | DZ | IMP | STK */
527 	FPE_FLTDIV,	/* 66 - DNML | DZ | IMP | STK */
528 	FPE_FLTSUB,	/* 67 - INV | DNML | DZ | IMP | STK */
529 	FPE_FLTOVF,	/* 68 - OFL | IMP | STK */
530 	FPE_FLTSUB,	/* 69 - INV | OFL | IMP | STK */
531 	FPE_FLTUND,	/* 6A - DNML | OFL | IMP | STK */
532 	FPE_FLTSUB,	/* 6B - INV | DNML | OFL | IMP | STK */
533 	FPE_FLTDIV,	/* 6C - DZ | OFL | IMP | STK */
534 	FPE_FLTSUB,	/* 6D - INV | DZ | OFL | IMP | STK */
535 	FPE_FLTDIV,	/* 6E - DNML | DZ | OFL | IMP | STK */
536 	FPE_FLTSUB,	/* 6F - INV | DNML | DZ | OFL | IMP | STK */
537 	FPE_FLTUND,	/* 70 - UFL | IMP | STK */
538 	FPE_FLTSUB,	/* 71 - INV | UFL | IMP | STK */
539 	FPE_FLTUND,	/* 72 - DNML | UFL | IMP | STK */
540 	FPE_FLTSUB,	/* 73 - INV | DNML | UFL | IMP | STK */
541 	FPE_FLTDIV,	/* 74 - DZ | UFL | IMP | STK */
542 	FPE_FLTSUB,	/* 75 - INV | DZ | UFL | IMP | STK */
543 	FPE_FLTDIV,	/* 76 - DNML | DZ | UFL | IMP | STK */
544 	FPE_FLTSUB,	/* 77 - INV | DNML | DZ | UFL | IMP | STK */
545 	FPE_FLTOVF,	/* 78 - OFL | UFL | IMP | STK */
546 	FPE_FLTSUB,	/* 79 - INV | OFL | UFL | IMP | STK */
547 	FPE_FLTUND,	/* 7A - DNML | OFL | UFL | IMP | STK */
548 	FPE_FLTSUB,	/* 7B - INV | DNML | OFL | UFL | IMP | STK */
549 	FPE_FLTDIV,	/* 7C - DZ | OFL | UFL | IMP | STK */
550 	FPE_FLTSUB,	/* 7D - INV | DZ | OFL | UFL | IMP | STK */
551 	FPE_FLTDIV,	/* 7E - DNML | DZ | OFL | UFL | IMP | STK */
552 	FPE_FLTSUB,	/* 7F - INV | DNML | DZ | OFL | UFL | IMP | STK */
553 };
554 
555 /*
556  * Read the FP status and control words, then generate si_code value
557  * for SIGFPE.  The error code chosen will be one of the
558  * FPE_... macros.  It will be sent as the second argument to old
559  * BSD-style signal handlers and as "siginfo_t->si_code" (second
560  * argument) to SA_SIGINFO signal handlers.
561  *
562  * Some time ago, we cleared the x87 exceptions with FNCLEX there.
563  * Clearing exceptions was necessary mainly to avoid IRQ13 bugs.  The
564  * usermode code which understands the FPU hardware enough to enable
565  * the exceptions, can also handle clearing the exception state in the
566  * handler.  The only consequence of not clearing the exception is the
567  * rethrow of the SIGFPE on return from the signal handler and
568  * reexecution of the corresponding instruction.
569  *
570  * For XMM traps, the exceptions were never cleared.
571  */
572 int
573 fputrap_x87(void)
574 {
575 	struct savefpu *pcb_save;
576 	u_short control, status;
577 
578 	critical_enter();
579 
580 	/*
581 	 * Interrupt handling (for another interrupt) may have pushed the
582 	 * state to memory.  Fetch the relevant parts of the state from
583 	 * wherever they are.
584 	 */
585 	if (PCPU_GET(fpcurthread) != curthread) {
586 		pcb_save = curpcb->pcb_save;
587 		control = pcb_save->sv_env.en_cw;
588 		status = pcb_save->sv_env.en_sw;
589 	} else {
590 		fnstcw(&control);
591 		fnstsw(&status);
592 	}
593 
594 	critical_exit();
595 	return (fpetable[status & ((~control & 0x3f) | 0x40)]);
596 }
597 
598 int
599 fputrap_sse(void)
600 {
601 	u_int mxcsr;
602 
603 	critical_enter();
604 	if (PCPU_GET(fpcurthread) != curthread)
605 		mxcsr = curpcb->pcb_save->sv_env.en_mxcsr;
606 	else
607 		stmxcsr(&mxcsr);
608 	critical_exit();
609 	return (fpetable[(mxcsr & (~mxcsr >> 7)) & 0x3f]);
610 }
611 
612 /*
613  * Device Not Available (DNA, #NM) exception handler.
614  *
615  * It would be better to switch FP context here (if curthread !=
616  * fpcurthread) and not necessarily for every context switch, but it
617  * is too hard to access foreign pcb's.
618  */
619 void
620 fpudna(void)
621 {
622 
623 	/*
624 	 * This handler is entered with interrupts enabled, so context
625 	 * switches may occur before critical_enter() is executed.  If
626 	 * a context switch occurs, then when we regain control, our
627 	 * state will have been completely restored.  The CPU may
628 	 * change underneath us, but the only part of our context that
629 	 * lives in the CPU is CR0.TS and that will be "restored" by
630 	 * setting it on the new CPU.
631 	 */
632 	critical_enter();
633 
634 	if (PCPU_GET(fpcurthread) == curthread) {
635 		printf("fpudna: fpcurthread == curthread\n");
636 		stop_emulating();
637 		critical_exit();
638 		return;
639 	}
640 	if (PCPU_GET(fpcurthread) != NULL) {
641 		panic("fpudna: fpcurthread = %p (%d), curthread = %p (%d)\n",
642 		    PCPU_GET(fpcurthread), PCPU_GET(fpcurthread)->td_tid,
643 		    curthread, curthread->td_tid);
644 	}
645 	stop_emulating();
646 	/*
647 	 * Record new context early in case frstor causes a trap.
648 	 */
649 	PCPU_SET(fpcurthread, curthread);
650 
651 	fpu_clean_state();
652 
653 	if ((curpcb->pcb_flags & PCB_FPUINITDONE) == 0) {
654 		/*
655 		 * This is the first time this thread has used the FPU or
656 		 * the PCB doesn't contain a clean FPU state.  Explicitly
657 		 * load an initial state.
658 		 *
659 		 * We prefer to restore the state from the actual save
660 		 * area in PCB instead of directly loading from
661 		 * fpu_initialstate, to ignite the XSAVEOPT
662 		 * tracking engine.
663 		 */
664 		bcopy(fpu_initialstate, curpcb->pcb_save, cpu_max_ext_state_size);
665 		fpurestore(curpcb->pcb_save);
666 		if (curpcb->pcb_initial_fpucw != __INITIAL_FPUCW__)
667 			fldcw(curpcb->pcb_initial_fpucw);
668 		if (PCB_USER_FPU(curpcb))
669 			set_pcb_flags(curpcb,
670 			    PCB_FPUINITDONE | PCB_USERFPUINITDONE);
671 		else
672 			set_pcb_flags(curpcb, PCB_FPUINITDONE);
673 	} else
674 		fpurestore(curpcb->pcb_save);
675 	critical_exit();
676 }
677 
678 void
679 fpudrop()
680 {
681 	struct thread *td;
682 
683 	td = PCPU_GET(fpcurthread);
684 	KASSERT(td == curthread, ("fpudrop: fpcurthread != curthread"));
685 	CRITICAL_ASSERT(td);
686 	PCPU_SET(fpcurthread, NULL);
687 	clear_pcb_flags(td->td_pcb, PCB_FPUINITDONE);
688 	start_emulating();
689 }
690 
691 /*
692  * Get the user state of the FPU into pcb->pcb_user_save without
693  * dropping ownership (if possible).  It returns the FPU ownership
694  * status.
695  */
696 int
697 fpugetregs(struct thread *td)
698 {
699 	struct pcb *pcb;
700 	uint64_t *xstate_bv, bit;
701 	char *sa;
702 	int max_ext_n, i, owned;
703 
704 	pcb = td->td_pcb;
705 	if ((pcb->pcb_flags & PCB_USERFPUINITDONE) == 0) {
706 		bcopy(fpu_initialstate, get_pcb_user_save_pcb(pcb),
707 		    cpu_max_ext_state_size);
708 		get_pcb_user_save_pcb(pcb)->sv_env.en_cw =
709 		    pcb->pcb_initial_fpucw;
710 		fpuuserinited(td);
711 		return (_MC_FPOWNED_PCB);
712 	}
713 	critical_enter();
714 	if (td == PCPU_GET(fpcurthread) && PCB_USER_FPU(pcb)) {
715 		fpusave(get_pcb_user_save_pcb(pcb));
716 		owned = _MC_FPOWNED_FPU;
717 	} else {
718 		owned = _MC_FPOWNED_PCB;
719 	}
720 	critical_exit();
721 	if (use_xsave) {
722 		/*
723 		 * Handle partially saved state.
724 		 */
725 		sa = (char *)get_pcb_user_save_pcb(pcb);
726 		xstate_bv = (uint64_t *)(sa + sizeof(struct savefpu) +
727 		    offsetof(struct xstate_hdr, xstate_bv));
728 		max_ext_n = flsl(xsave_mask);
729 		for (i = 0; i < max_ext_n; i++) {
730 			bit = 1ULL << i;
731 			if ((xsave_mask & bit) == 0 || (*xstate_bv & bit) != 0)
732 				continue;
733 			bcopy((char *)fpu_initialstate +
734 			    xsave_area_desc[i].offset,
735 			    sa + xsave_area_desc[i].offset,
736 			    xsave_area_desc[i].size);
737 			*xstate_bv |= bit;
738 		}
739 	}
740 	return (owned);
741 }
742 
743 void
744 fpuuserinited(struct thread *td)
745 {
746 	struct pcb *pcb;
747 
748 	pcb = td->td_pcb;
749 	if (PCB_USER_FPU(pcb))
750 		set_pcb_flags(pcb,
751 		    PCB_FPUINITDONE | PCB_USERFPUINITDONE);
752 	else
753 		set_pcb_flags(pcb, PCB_FPUINITDONE);
754 }
755 
756 int
757 fpusetxstate(struct thread *td, char *xfpustate, size_t xfpustate_size)
758 {
759 	struct xstate_hdr *hdr, *ehdr;
760 	size_t len, max_len;
761 	uint64_t bv;
762 
763 	/* XXXKIB should we clear all extended state in xstate_bv instead ? */
764 	if (xfpustate == NULL)
765 		return (0);
766 	if (!use_xsave)
767 		return (EOPNOTSUPP);
768 
769 	len = xfpustate_size;
770 	if (len < sizeof(struct xstate_hdr))
771 		return (EINVAL);
772 	max_len = cpu_max_ext_state_size - sizeof(struct savefpu);
773 	if (len > max_len)
774 		return (EINVAL);
775 
776 	ehdr = (struct xstate_hdr *)xfpustate;
777 	bv = ehdr->xstate_bv;
778 
779 	/*
780 	 * Avoid #gp.
781 	 */
782 	if (bv & ~xsave_mask)
783 		return (EINVAL);
784 
785 	hdr = (struct xstate_hdr *)(get_pcb_user_save_td(td) + 1);
786 
787 	hdr->xstate_bv = bv;
788 	bcopy(xfpustate + sizeof(struct xstate_hdr),
789 	    (char *)(hdr + 1), len - sizeof(struct xstate_hdr));
790 
791 	return (0);
792 }
793 
794 /*
795  * Set the state of the FPU.
796  */
797 int
798 fpusetregs(struct thread *td, struct savefpu *addr, char *xfpustate,
799     size_t xfpustate_size)
800 {
801 	struct pcb *pcb;
802 	int error;
803 
804 	pcb = td->td_pcb;
805 	critical_enter();
806 	if (td == PCPU_GET(fpcurthread) && PCB_USER_FPU(pcb)) {
807 		error = fpusetxstate(td, xfpustate, xfpustate_size);
808 		if (error != 0) {
809 			critical_exit();
810 			return (error);
811 		}
812 		bcopy(addr, get_pcb_user_save_td(td), sizeof(*addr));
813 		fpurestore(get_pcb_user_save_td(td));
814 		critical_exit();
815 		set_pcb_flags(pcb, PCB_FPUINITDONE | PCB_USERFPUINITDONE);
816 	} else {
817 		critical_exit();
818 		error = fpusetxstate(td, xfpustate, xfpustate_size);
819 		if (error != 0)
820 			return (error);
821 		bcopy(addr, get_pcb_user_save_td(td), sizeof(*addr));
822 		fpuuserinited(td);
823 	}
824 	return (0);
825 }
826 
827 /*
828  * On AuthenticAMD processors, the fxrstor instruction does not restore
829  * the x87's stored last instruction pointer, last data pointer, and last
830  * opcode values, except in the rare case in which the exception summary
831  * (ES) bit in the x87 status word is set to 1.
832  *
833  * In order to avoid leaking this information across processes, we clean
834  * these values by performing a dummy load before executing fxrstor().
835  */
836 static void
837 fpu_clean_state(void)
838 {
839 	static float dummy_variable = 0.0;
840 	u_short status;
841 
842 	/*
843 	 * Clear the ES bit in the x87 status word if it is currently
844 	 * set, in order to avoid causing a fault in the upcoming load.
845 	 */
846 	fnstsw(&status);
847 	if (status & 0x80)
848 		fnclex();
849 
850 	/*
851 	 * Load the dummy variable into the x87 stack.  This mangles
852 	 * the x87 stack, but we don't care since we're about to call
853 	 * fxrstor() anyway.
854 	 */
855 	__asm __volatile("ffree %%st(7); flds %0" : : "m" (dummy_variable));
856 }
857 
858 /*
859  * This really sucks.  We want the acpi version only, but it requires
860  * the isa_if.h file in order to get the definitions.
861  */
862 #include "opt_isa.h"
863 #ifdef DEV_ISA
864 #include <isa/isavar.h>
865 /*
866  * This sucks up the legacy ISA support assignments from PNPBIOS/ACPI.
867  */
868 static struct isa_pnp_id fpupnp_ids[] = {
869 	{ 0x040cd041, "Legacy ISA coprocessor support" }, /* PNP0C04 */
870 	{ 0 }
871 };
872 
873 static int
874 fpupnp_probe(device_t dev)
875 {
876 	int result;
877 
878 	result = ISA_PNP_PROBE(device_get_parent(dev), dev, fpupnp_ids);
879 	if (result <= 0)
880 		device_quiet(dev);
881 	return (result);
882 }
883 
884 static int
885 fpupnp_attach(device_t dev)
886 {
887 
888 	return (0);
889 }
890 
891 static device_method_t fpupnp_methods[] = {
892 	/* Device interface */
893 	DEVMETHOD(device_probe,		fpupnp_probe),
894 	DEVMETHOD(device_attach,	fpupnp_attach),
895 	DEVMETHOD(device_detach,	bus_generic_detach),
896 	DEVMETHOD(device_shutdown,	bus_generic_shutdown),
897 	DEVMETHOD(device_suspend,	bus_generic_suspend),
898 	DEVMETHOD(device_resume,	bus_generic_resume),
899 
900 	{ 0, 0 }
901 };
902 
903 static driver_t fpupnp_driver = {
904 	"fpupnp",
905 	fpupnp_methods,
906 	1,			/* no softc */
907 };
908 
909 static devclass_t fpupnp_devclass;
910 
911 DRIVER_MODULE(fpupnp, acpi, fpupnp_driver, fpupnp_devclass, 0, 0);
912 #endif	/* DEV_ISA */
913 
914 static MALLOC_DEFINE(M_FPUKERN_CTX, "fpukern_ctx",
915     "Kernel contexts for FPU state");
916 
917 #define	FPU_KERN_CTX_FPUINITDONE 0x01
918 #define	FPU_KERN_CTX_DUMMY	 0x02	/* avoided save for the kern thread */
919 #define	FPU_KERN_CTX_INUSE	 0x04
920 
921 struct fpu_kern_ctx {
922 	struct savefpu *prev;
923 	uint32_t flags;
924 	char hwstate1[];
925 };
926 
927 struct fpu_kern_ctx *
928 fpu_kern_alloc_ctx(u_int flags)
929 {
930 	struct fpu_kern_ctx *res;
931 	size_t sz;
932 
933 	sz = sizeof(struct fpu_kern_ctx) + XSAVE_AREA_ALIGN +
934 	    cpu_max_ext_state_size;
935 	res = malloc(sz, M_FPUKERN_CTX, ((flags & FPU_KERN_NOWAIT) ?
936 	    M_NOWAIT : M_WAITOK) | M_ZERO);
937 	return (res);
938 }
939 
940 void
941 fpu_kern_free_ctx(struct fpu_kern_ctx *ctx)
942 {
943 
944 	KASSERT((ctx->flags & FPU_KERN_CTX_INUSE) == 0, ("free'ing inuse ctx"));
945 	/* XXXKIB clear the memory ? */
946 	free(ctx, M_FPUKERN_CTX);
947 }
948 
949 static struct savefpu *
950 fpu_kern_ctx_savefpu(struct fpu_kern_ctx *ctx)
951 {
952 	vm_offset_t p;
953 
954 	p = (vm_offset_t)&ctx->hwstate1;
955 	p = roundup2(p, XSAVE_AREA_ALIGN);
956 	return ((struct savefpu *)p);
957 }
958 
959 int
960 fpu_kern_enter(struct thread *td, struct fpu_kern_ctx *ctx, u_int flags)
961 {
962 	struct pcb *pcb;
963 
964 	KASSERT((ctx->flags & FPU_KERN_CTX_INUSE) == 0, ("using inuse ctx"));
965 
966 	if ((flags & FPU_KERN_KTHR) != 0 && is_fpu_kern_thread(0)) {
967 		ctx->flags = FPU_KERN_CTX_DUMMY | FPU_KERN_CTX_INUSE;
968 		return (0);
969 	}
970 	pcb = td->td_pcb;
971 	KASSERT(!PCB_USER_FPU(pcb) || pcb->pcb_save ==
972 	    get_pcb_user_save_pcb(pcb), ("mangled pcb_save"));
973 	ctx->flags = FPU_KERN_CTX_INUSE;
974 	if ((pcb->pcb_flags & PCB_FPUINITDONE) != 0)
975 		ctx->flags |= FPU_KERN_CTX_FPUINITDONE;
976 	fpuexit(td);
977 	ctx->prev = pcb->pcb_save;
978 	pcb->pcb_save = fpu_kern_ctx_savefpu(ctx);
979 	set_pcb_flags(pcb, PCB_KERNFPU);
980 	clear_pcb_flags(pcb, PCB_FPUINITDONE);
981 	return (0);
982 }
983 
984 int
985 fpu_kern_leave(struct thread *td, struct fpu_kern_ctx *ctx)
986 {
987 	struct pcb *pcb;
988 
989 	KASSERT((ctx->flags & FPU_KERN_CTX_INUSE) != 0,
990 	    ("leaving not inuse ctx"));
991 	ctx->flags &= ~FPU_KERN_CTX_INUSE;
992 
993 	if (is_fpu_kern_thread(0) && (ctx->flags & FPU_KERN_CTX_DUMMY) != 0)
994 		return (0);
995 	KASSERT((ctx->flags & FPU_KERN_CTX_DUMMY) == 0, ("dummy ctx"));
996 	pcb = td->td_pcb;
997 	critical_enter();
998 	if (curthread == PCPU_GET(fpcurthread))
999 		fpudrop();
1000 	critical_exit();
1001 	pcb->pcb_save = ctx->prev;
1002 	if (pcb->pcb_save == get_pcb_user_save_pcb(pcb)) {
1003 		if ((pcb->pcb_flags & PCB_USERFPUINITDONE) != 0) {
1004 			set_pcb_flags(pcb, PCB_FPUINITDONE);
1005 			clear_pcb_flags(pcb, PCB_KERNFPU);
1006 		} else
1007 			clear_pcb_flags(pcb, PCB_FPUINITDONE | PCB_KERNFPU);
1008 	} else {
1009 		if ((ctx->flags & FPU_KERN_CTX_FPUINITDONE) != 0)
1010 			set_pcb_flags(pcb, PCB_FPUINITDONE);
1011 		else
1012 			clear_pcb_flags(pcb, PCB_FPUINITDONE);
1013 		KASSERT(!PCB_USER_FPU(pcb), ("unpaired fpu_kern_leave"));
1014 	}
1015 	return (0);
1016 }
1017 
1018 int
1019 fpu_kern_thread(u_int flags)
1020 {
1021 
1022 	KASSERT((curthread->td_pflags & TDP_KTHREAD) != 0,
1023 	    ("Only kthread may use fpu_kern_thread"));
1024 	KASSERT(curpcb->pcb_save == get_pcb_user_save_pcb(curpcb),
1025 	    ("mangled pcb_save"));
1026 	KASSERT(PCB_USER_FPU(curpcb), ("recursive call"));
1027 
1028 	set_pcb_flags(curpcb, PCB_KERNFPU);
1029 	return (0);
1030 }
1031 
1032 int
1033 is_fpu_kern_thread(u_int flags)
1034 {
1035 
1036 	if ((curthread->td_pflags & TDP_KTHREAD) == 0)
1037 		return (0);
1038 	return ((curpcb->pcb_flags & PCB_KERNFPU) != 0);
1039 }
1040 
1041 /*
1042  * FPU save area alloc/free/init utility routines
1043  */
1044 struct savefpu *
1045 fpu_save_area_alloc(void)
1046 {
1047 
1048 	return (uma_zalloc(fpu_save_area_zone, 0));
1049 }
1050 
1051 void
1052 fpu_save_area_free(struct savefpu *fsa)
1053 {
1054 
1055 	uma_zfree(fpu_save_area_zone, fsa);
1056 }
1057 
1058 void
1059 fpu_save_area_reset(struct savefpu *fsa)
1060 {
1061 
1062 	bcopy(fpu_initialstate, fsa, cpu_max_ext_state_size);
1063 }
1064