xref: /freebsd/sys/i386/i386/npx.c (revision bdd1243d)
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  * 3. 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 "opt_cpu.h"
37 #include "opt_isa.h"
38 #include "opt_npx.h"
39 
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/bus.h>
43 #include <sys/kernel.h>
44 #include <sys/lock.h>
45 #include <sys/malloc.h>
46 #include <sys/module.h>
47 #include <sys/mutex.h>
48 #include <sys/mutex.h>
49 #include <sys/proc.h>
50 #include <sys/smp.h>
51 #include <sys/sysctl.h>
52 #include <machine/bus.h>
53 #include <sys/rman.h>
54 #ifdef NPX_DEBUG
55 #include <sys/syslog.h>
56 #endif
57 #include <sys/signalvar.h>
58 #include <vm/uma.h>
59 
60 #include <machine/asmacros.h>
61 #include <machine/cputypes.h>
62 #include <machine/frame.h>
63 #include <machine/md_var.h>
64 #include <machine/pcb.h>
65 #include <machine/psl.h>
66 #include <machine/resource.h>
67 #include <machine/specialreg.h>
68 #include <machine/segments.h>
69 #include <machine/ucontext.h>
70 #include <x86/ifunc.h>
71 
72 #include <machine/intr_machdep.h>
73 
74 #ifdef DEV_ISA
75 #include <isa/isavar.h>
76 #endif
77 
78 /*
79  * 387 and 287 Numeric Coprocessor Extension (NPX) Driver.
80  */
81 
82 #define	fldcw(cw)		__asm __volatile("fldcw %0" : : "m" (cw))
83 #define	fnclex()		__asm __volatile("fnclex")
84 #define	fninit()		__asm __volatile("fninit")
85 #define	fnsave(addr)		__asm __volatile("fnsave %0" : "=m" (*(addr)))
86 #define	fnstcw(addr)		__asm __volatile("fnstcw %0" : "=m" (*(addr)))
87 #define	fnstsw(addr)		__asm __volatile("fnstsw %0" : "=am" (*(addr)))
88 #define	fp_divide_by_0()	__asm __volatile( \
89 				    "fldz; fld1; fdiv %st,%st(1); fnop")
90 #define	frstor(addr)		__asm __volatile("frstor %0" : : "m" (*(addr)))
91 #define	fxrstor(addr)		__asm __volatile("fxrstor %0" : : "m" (*(addr)))
92 #define	fxsave(addr)		__asm __volatile("fxsave %0" : "=m" (*(addr)))
93 #define	ldmxcsr(csr)		__asm __volatile("ldmxcsr %0" : : "m" (csr))
94 #define	stmxcsr(addr)		__asm __volatile("stmxcsr %0" : : "m" (*(addr)))
95 
96 static __inline void
97 xrstor(char *addr, uint64_t mask)
98 {
99 	uint32_t low, hi;
100 
101 	low = mask;
102 	hi = mask >> 32;
103 	__asm __volatile("xrstor %0" : : "m" (*addr), "a" (low), "d" (hi));
104 }
105 
106 static __inline void
107 xsave(char *addr, uint64_t mask)
108 {
109 	uint32_t low, hi;
110 
111 	low = mask;
112 	hi = mask >> 32;
113 	__asm __volatile("xsave %0" : "=m" (*addr) : "a" (low), "d" (hi) :
114 	    "memory");
115 }
116 
117 static __inline void
118 xsaveopt(char *addr, uint64_t mask)
119 {
120 	uint32_t low, hi;
121 
122 	low = mask;
123 	hi = mask >> 32;
124 	__asm __volatile("xsaveopt %0" : "=m" (*addr) : "a" (low), "d" (hi) :
125 	    "memory");
126 }
127 
128 #define	start_emulating()	load_cr0(rcr0() | CR0_TS)
129 #define	stop_emulating()	clts()
130 
131 #define GET_FPU_CW(thread) \
132 	(cpu_fxsr ? \
133 		(thread)->td_pcb->pcb_save->sv_xmm.sv_env.en_cw : \
134 		(thread)->td_pcb->pcb_save->sv_87.sv_env.en_cw)
135 #define GET_FPU_SW(thread) \
136 	(cpu_fxsr ? \
137 		(thread)->td_pcb->pcb_save->sv_xmm.sv_env.en_sw : \
138 		(thread)->td_pcb->pcb_save->sv_87.sv_env.en_sw)
139 #define SET_FPU_CW(savefpu, value) do { \
140 	if (cpu_fxsr) \
141 		(savefpu)->sv_xmm.sv_env.en_cw = (value); \
142 	else \
143 		(savefpu)->sv_87.sv_env.en_cw = (value); \
144 } while (0)
145 
146 CTASSERT(sizeof(union savefpu) == 512);
147 CTASSERT(sizeof(struct xstate_hdr) == 64);
148 CTASSERT(sizeof(struct savefpu_ymm) == 832);
149 
150 /*
151  * This requirement is to make it easier for asm code to calculate
152  * offset of the fpu save area from the pcb address. FPU save area
153  * must be 64-byte aligned.
154  */
155 CTASSERT(sizeof(struct pcb) % XSAVE_AREA_ALIGN == 0);
156 
157 /*
158  * Ensure the copy of XCR0 saved in a core is contained in the padding
159  * area.
160  */
161 CTASSERT(X86_XSTATE_XCR0_OFFSET >= offsetof(struct savexmm, sv_pad) &&
162     X86_XSTATE_XCR0_OFFSET + sizeof(uint64_t) <= sizeof(struct savexmm));
163 
164 static	void	fpu_clean_state(void);
165 
166 static	void	fpurstor(union savefpu *);
167 
168 int	hw_float;
169 
170 SYSCTL_INT(_hw, HW_FLOATINGPT, floatingpoint, CTLFLAG_RD,
171     &hw_float, 0, "Floating point instructions executed in hardware");
172 
173 int lazy_fpu_switch = 0;
174 SYSCTL_INT(_hw, OID_AUTO, lazy_fpu_switch, CTLFLAG_RWTUN | CTLFLAG_NOFETCH,
175     &lazy_fpu_switch, 0,
176     "Lazily load FPU context after context switch");
177 
178 u_int cpu_fxsr;		/* SSE enabled */
179 int use_xsave;
180 uint64_t xsave_mask;
181 static	uma_zone_t fpu_save_area_zone;
182 static	union savefpu *npx_initialstate;
183 
184 static struct xsave_area_elm_descr {
185 	u_int	offset;
186 	u_int	size;
187 } *xsave_area_desc;
188 
189 static	volatile u_int		npx_traps_while_probing;
190 
191 alias_for_inthand_t probetrap;
192 __asm("								\n\
193 	.text							\n\
194 	.p2align 2,0x90						\n\
195 	.type	" __XSTRING(CNAME(probetrap)) ",@function	\n\
196 " __XSTRING(CNAME(probetrap)) ":				\n\
197 	ss							\n\
198 	incl	" __XSTRING(CNAME(npx_traps_while_probing)) "	\n\
199 	fnclex							\n\
200 	iret							\n\
201 ");
202 
203 /*
204  * Determine if an FPU is present and how to use it.
205  */
206 static int
207 npx_probe(void)
208 {
209 	struct gate_descriptor save_idt_npxtrap;
210 	u_short control, status;
211 
212 	/*
213 	 * Modern CPUs all have an FPU that uses the INT16 interface
214 	 * and provide a simple way to verify that, so handle the
215 	 * common case right away.
216 	 */
217 	if (cpu_feature & CPUID_FPU) {
218 		hw_float = 1;
219 		return (1);
220 	}
221 
222 	save_idt_npxtrap = idt[IDT_MF];
223 	setidt(IDT_MF, probetrap, SDT_SYS386TGT, SEL_KPL,
224 	    GSEL(GCODE_SEL, SEL_KPL));
225 
226 	/*
227 	 * Don't trap while we're probing.
228 	 */
229 	stop_emulating();
230 
231 	/*
232 	 * Finish resetting the coprocessor, if any.  If there is an error
233 	 * pending, then we may get a bogus IRQ13, but npx_intr() will handle
234 	 * it OK.  Bogus halts have never been observed, but we enabled
235 	 * IRQ13 and cleared the BUSY# latch early to handle them anyway.
236 	 */
237 	fninit();
238 
239 	/*
240 	 * Don't use fwait here because it might hang.
241 	 * Don't use fnop here because it usually hangs if there is no FPU.
242 	 */
243 	DELAY(1000);		/* wait for any IRQ13 */
244 #ifdef DIAGNOSTIC
245 	if (npx_traps_while_probing != 0)
246 		printf("fninit caused %u bogus npx trap(s)\n",
247 		       npx_traps_while_probing);
248 #endif
249 	/*
250 	 * Check for a status of mostly zero.
251 	 */
252 	status = 0x5a5a;
253 	fnstsw(&status);
254 	if ((status & 0xb8ff) == 0) {
255 		/*
256 		 * Good, now check for a proper control word.
257 		 */
258 		control = 0x5a5a;
259 		fnstcw(&control);
260 		if ((control & 0x1f3f) == 0x033f) {
261 			/*
262 			 * We have an npx, now divide by 0 to see if exception
263 			 * 16 works.
264 			 */
265 			control &= ~(1 << 2);	/* enable divide by 0 trap */
266 			fldcw(control);
267 			npx_traps_while_probing = 0;
268 			fp_divide_by_0();
269 			if (npx_traps_while_probing != 0) {
270 				/*
271 				 * Good, exception 16 works.
272 				 */
273 				hw_float = 1;
274 				goto cleanup;
275 			}
276 			printf(
277 	"FPU does not use exception 16 for error reporting\n");
278 			goto cleanup;
279 		}
280 	}
281 
282 	/*
283 	 * Probe failed.  Floating point simply won't work.
284 	 * Notify user and disable FPU/MMX/SSE instruction execution.
285 	 */
286 	printf("WARNING: no FPU!\n");
287 	__asm __volatile("smsw %%ax; orb %0,%%al; lmsw %%ax" : :
288 	    "n" (CR0_EM | CR0_MP) : "ax");
289 
290 cleanup:
291 	idt[IDT_MF] = save_idt_npxtrap;
292 	return (hw_float);
293 }
294 
295 static void
296 fpusave_xsaveopt(union savefpu *addr)
297 {
298 
299 	xsaveopt((char *)addr, xsave_mask);
300 }
301 
302 static void
303 fpusave_xsave(union savefpu *addr)
304 {
305 
306 	xsave((char *)addr, xsave_mask);
307 }
308 
309 static void
310 fpusave_fxsave(union savefpu *addr)
311 {
312 
313 	fxsave((char *)addr);
314 }
315 
316 static void
317 fpusave_fnsave(union savefpu *addr)
318 {
319 
320 	fnsave((char *)addr);
321 }
322 
323 DEFINE_IFUNC(, void, fpusave, (union savefpu *))
324 {
325 	if (use_xsave)
326 		return ((cpu_stdext_feature & CPUID_EXTSTATE_XSAVEOPT) != 0 ?
327 		    fpusave_xsaveopt : fpusave_xsave);
328 	if (cpu_fxsr)
329 		return (fpusave_fxsave);
330 	return (fpusave_fnsave);
331 }
332 
333 /*
334  * Enable XSAVE if supported and allowed by user.
335  * Calculate the xsave_mask.
336  */
337 static void
338 npxinit_bsp1(void)
339 {
340 	u_int cp[4];
341 	uint64_t xsave_mask_user;
342 
343 	TUNABLE_INT_FETCH("hw.lazy_fpu_switch", &lazy_fpu_switch);
344 	if (!use_xsave)
345 		return;
346 	cpuid_count(0xd, 0x0, cp);
347 	xsave_mask = XFEATURE_ENABLED_X87 | XFEATURE_ENABLED_SSE;
348 	if ((cp[0] & xsave_mask) != xsave_mask)
349 		panic("CPU0 does not support X87 or SSE: %x", cp[0]);
350 	xsave_mask = ((uint64_t)cp[3] << 32) | cp[0];
351 	xsave_mask_user = xsave_mask;
352 	TUNABLE_QUAD_FETCH("hw.xsave_mask", &xsave_mask_user);
353 	xsave_mask_user |= XFEATURE_ENABLED_X87 | XFEATURE_ENABLED_SSE;
354 	xsave_mask &= xsave_mask_user;
355 	if ((xsave_mask & XFEATURE_AVX512) != XFEATURE_AVX512)
356 		xsave_mask &= ~XFEATURE_AVX512;
357 	if ((xsave_mask & XFEATURE_MPX) != XFEATURE_MPX)
358 		xsave_mask &= ~XFEATURE_MPX;
359 }
360 
361 /*
362  * Calculate the fpu save area size.
363  */
364 static void
365 npxinit_bsp2(void)
366 {
367 	u_int cp[4];
368 
369 	if (use_xsave) {
370 		cpuid_count(0xd, 0x0, cp);
371 		cpu_max_ext_state_size = cp[1];
372 
373 		/*
374 		 * Reload the cpu_feature2, since we enabled OSXSAVE.
375 		 */
376 		do_cpuid(1, cp);
377 		cpu_feature2 = cp[2];
378 	} else
379 		cpu_max_ext_state_size = sizeof(union savefpu);
380 }
381 
382 /*
383  * Initialize floating point unit.
384  */
385 void
386 npxinit(bool bsp)
387 {
388 	static union savefpu dummy;
389 	register_t saveintr;
390 	u_int mxcsr;
391 	u_short control;
392 
393 	if (bsp) {
394 		if (!npx_probe())
395 			return;
396 		npxinit_bsp1();
397 	}
398 
399 	if (use_xsave) {
400 		load_cr4(rcr4() | CR4_XSAVE);
401 		load_xcr(XCR0, xsave_mask);
402 	}
403 
404 	/*
405 	 * XCR0 shall be set up before CPU can report the save area size.
406 	 */
407 	if (bsp)
408 		npxinit_bsp2();
409 
410 	/*
411 	 * fninit has the same h/w bugs as fnsave.  Use the detoxified
412 	 * fnsave to throw away any junk in the fpu.  fpusave() initializes
413 	 * the fpu.
414 	 *
415 	 * It is too early for critical_enter() to work on AP.
416 	 */
417 	saveintr = intr_disable();
418 	stop_emulating();
419 	if (cpu_fxsr)
420 		fninit();
421 	else
422 		fnsave(&dummy);
423 	control = __INITIAL_NPXCW__;
424 	fldcw(control);
425 	if (cpu_fxsr) {
426 		mxcsr = __INITIAL_MXCSR__;
427 		ldmxcsr(mxcsr);
428 	}
429 	start_emulating();
430 	intr_restore(saveintr);
431 }
432 
433 /*
434  * On the boot CPU we generate a clean state that is used to
435  * initialize the floating point unit when it is first used by a
436  * process.
437  */
438 static void
439 npxinitstate(void *arg __unused)
440 {
441 	uint64_t *xstate_bv;
442 	register_t saveintr;
443 	int cp[4], i, max_ext_n;
444 
445 	if (!hw_float)
446 		return;
447 
448 	/* Do potentially blocking operations before disabling interrupts. */
449 	fpu_save_area_zone = uma_zcreate("FPU_save_area",
450 	    cpu_max_ext_state_size, NULL, NULL, NULL, NULL,
451 	    XSAVE_AREA_ALIGN - 1, 0);
452 	npx_initialstate = uma_zalloc(fpu_save_area_zone, M_WAITOK | M_ZERO);
453 	if (use_xsave) {
454 		if (xsave_mask >> 32 != 0)
455 			max_ext_n = fls(xsave_mask >> 32) + 32;
456 		else
457 			max_ext_n = fls(xsave_mask);
458 		xsave_area_desc = malloc(max_ext_n * sizeof(struct
459 		    xsave_area_elm_descr), M_DEVBUF, M_WAITOK | M_ZERO);
460 	}
461 
462 	saveintr = intr_disable();
463 	stop_emulating();
464 
465 	if (cpu_fxsr)
466 		fpusave_fxsave(npx_initialstate);
467 	else
468 		fpusave_fnsave(npx_initialstate);
469 	if (cpu_fxsr) {
470 		if (npx_initialstate->sv_xmm.sv_env.en_mxcsr_mask)
471 			cpu_mxcsr_mask =
472 			    npx_initialstate->sv_xmm.sv_env.en_mxcsr_mask;
473 		else
474 			cpu_mxcsr_mask = 0xFFBF;
475 
476 		/*
477 		 * The fninit instruction does not modify XMM
478 		 * registers or x87 registers (MM/ST).  The fpusave
479 		 * call dumped the garbage contained in the registers
480 		 * after reset to the initial state saved.  Clear XMM
481 		 * and x87 registers file image to make the startup
482 		 * program state and signal handler XMM/x87 register
483 		 * content predictable.
484 		 */
485 		bzero(npx_initialstate->sv_xmm.sv_fp,
486 		    sizeof(npx_initialstate->sv_xmm.sv_fp));
487 		bzero(npx_initialstate->sv_xmm.sv_xmm,
488 		    sizeof(npx_initialstate->sv_xmm.sv_xmm));
489 
490 	} else
491 		bzero(npx_initialstate->sv_87.sv_ac,
492 		    sizeof(npx_initialstate->sv_87.sv_ac));
493 
494 	/*
495 	 * Create a table describing the layout of the CPU Extended
496 	 * Save Area.  See Intel SDM rev. 075 Vol. 1 13.4.1 "Legacy
497 	 * Region of an XSAVE Area" for the source of offsets/sizes.
498 	 * Note that 32bit XSAVE does not use %xmm8-%xmm15, see
499 	 * 10.5.1.2 and 13.5.2 "SSE State".
500 	 */
501 	if (use_xsave) {
502 		xstate_bv = (uint64_t *)((char *)(npx_initialstate + 1) +
503 		    offsetof(struct xstate_hdr, xstate_bv));
504 		*xstate_bv = XFEATURE_ENABLED_X87 | XFEATURE_ENABLED_SSE;
505 
506 		/* x87 state */
507 		xsave_area_desc[0].offset = 0;
508 		xsave_area_desc[0].size = 160;
509 		/* XMM */
510 		xsave_area_desc[1].offset = 160;
511 		xsave_area_desc[1].size = 288 - 160;
512 
513 		for (i = 2; i < max_ext_n; i++) {
514 			cpuid_count(0xd, i, cp);
515 			xsave_area_desc[i].offset = cp[1];
516 			xsave_area_desc[i].size = cp[0];
517 		}
518 	}
519 
520 	start_emulating();
521 	intr_restore(saveintr);
522 }
523 SYSINIT(npxinitstate, SI_SUB_CPU, SI_ORDER_ANY, npxinitstate, NULL);
524 
525 /*
526  * Free coprocessor (if we have it).
527  */
528 void
529 npxexit(struct thread *td)
530 {
531 
532 	critical_enter();
533 	if (curthread == PCPU_GET(fpcurthread)) {
534 		stop_emulating();
535 		fpusave(curpcb->pcb_save);
536 		start_emulating();
537 		PCPU_SET(fpcurthread, NULL);
538 	}
539 	critical_exit();
540 #ifdef NPX_DEBUG
541 	if (hw_float) {
542 		u_int	masked_exceptions;
543 
544 		masked_exceptions = GET_FPU_CW(td) & GET_FPU_SW(td) & 0x7f;
545 		/*
546 		 * Log exceptions that would have trapped with the old
547 		 * control word (overflow, divide by 0, and invalid operand).
548 		 */
549 		if (masked_exceptions & 0x0d)
550 			log(LOG_ERR,
551 	"pid %d (%s) exited with masked floating point exceptions 0x%02x\n",
552 			    td->td_proc->p_pid, td->td_proc->p_comm,
553 			    masked_exceptions);
554 	}
555 #endif
556 }
557 
558 int
559 npxformat(void)
560 {
561 
562 	if (!hw_float)
563 		return (_MC_FPFMT_NODEV);
564 	if (cpu_fxsr)
565 		return (_MC_FPFMT_XMM);
566 	return (_MC_FPFMT_387);
567 }
568 
569 /*
570  * The following mechanism is used to ensure that the FPE_... value
571  * that is passed as a trapcode to the signal handler of the user
572  * process does not have more than one bit set.
573  *
574  * Multiple bits may be set if the user process modifies the control
575  * word while a status word bit is already set.  While this is a sign
576  * of bad coding, we have no choice than to narrow them down to one
577  * bit, since we must not send a trapcode that is not exactly one of
578  * the FPE_ macros.
579  *
580  * The mechanism has a static table with 127 entries.  Each combination
581  * of the 7 FPU status word exception bits directly translates to a
582  * position in this table, where a single FPE_... value is stored.
583  * This FPE_... value stored there is considered the "most important"
584  * of the exception bits and will be sent as the signal code.  The
585  * precedence of the bits is based upon Intel Document "Numerical
586  * Applications", Chapter "Special Computational Situations".
587  *
588  * The macro to choose one of these values does these steps: 1) Throw
589  * away status word bits that cannot be masked.  2) Throw away the bits
590  * currently masked in the control word, assuming the user isn't
591  * interested in them anymore.  3) Reinsert status word bit 7 (stack
592  * fault) if it is set, which cannot be masked but must be presered.
593  * 4) Use the remaining bits to point into the trapcode table.
594  *
595  * The 6 maskable bits in order of their preference, as stated in the
596  * above referenced Intel manual:
597  * 1  Invalid operation (FP_X_INV)
598  * 1a   Stack underflow
599  * 1b   Stack overflow
600  * 1c   Operand of unsupported format
601  * 1d   SNaN operand.
602  * 2  QNaN operand (not an exception, irrelavant here)
603  * 3  Any other invalid-operation not mentioned above or zero divide
604  *      (FP_X_INV, FP_X_DZ)
605  * 4  Denormal operand (FP_X_DNML)
606  * 5  Numeric over/underflow (FP_X_OFL, FP_X_UFL)
607  * 6  Inexact result (FP_X_IMP)
608  */
609 static char fpetable[128] = {
610 	0,
611 	FPE_FLTINV,	/*  1 - INV */
612 	FPE_FLTUND,	/*  2 - DNML */
613 	FPE_FLTINV,	/*  3 - INV | DNML */
614 	FPE_FLTDIV,	/*  4 - DZ */
615 	FPE_FLTINV,	/*  5 - INV | DZ */
616 	FPE_FLTDIV,	/*  6 - DNML | DZ */
617 	FPE_FLTINV,	/*  7 - INV | DNML | DZ */
618 	FPE_FLTOVF,	/*  8 - OFL */
619 	FPE_FLTINV,	/*  9 - INV | OFL */
620 	FPE_FLTUND,	/*  A - DNML | OFL */
621 	FPE_FLTINV,	/*  B - INV | DNML | OFL */
622 	FPE_FLTDIV,	/*  C - DZ | OFL */
623 	FPE_FLTINV,	/*  D - INV | DZ | OFL */
624 	FPE_FLTDIV,	/*  E - DNML | DZ | OFL */
625 	FPE_FLTINV,	/*  F - INV | DNML | DZ | OFL */
626 	FPE_FLTUND,	/* 10 - UFL */
627 	FPE_FLTINV,	/* 11 - INV | UFL */
628 	FPE_FLTUND,	/* 12 - DNML | UFL */
629 	FPE_FLTINV,	/* 13 - INV | DNML | UFL */
630 	FPE_FLTDIV,	/* 14 - DZ | UFL */
631 	FPE_FLTINV,	/* 15 - INV | DZ | UFL */
632 	FPE_FLTDIV,	/* 16 - DNML | DZ | UFL */
633 	FPE_FLTINV,	/* 17 - INV | DNML | DZ | UFL */
634 	FPE_FLTOVF,	/* 18 - OFL | UFL */
635 	FPE_FLTINV,	/* 19 - INV | OFL | UFL */
636 	FPE_FLTUND,	/* 1A - DNML | OFL | UFL */
637 	FPE_FLTINV,	/* 1B - INV | DNML | OFL | UFL */
638 	FPE_FLTDIV,	/* 1C - DZ | OFL | UFL */
639 	FPE_FLTINV,	/* 1D - INV | DZ | OFL | UFL */
640 	FPE_FLTDIV,	/* 1E - DNML | DZ | OFL | UFL */
641 	FPE_FLTINV,	/* 1F - INV | DNML | DZ | OFL | UFL */
642 	FPE_FLTRES,	/* 20 - IMP */
643 	FPE_FLTINV,	/* 21 - INV | IMP */
644 	FPE_FLTUND,	/* 22 - DNML | IMP */
645 	FPE_FLTINV,	/* 23 - INV | DNML | IMP */
646 	FPE_FLTDIV,	/* 24 - DZ | IMP */
647 	FPE_FLTINV,	/* 25 - INV | DZ | IMP */
648 	FPE_FLTDIV,	/* 26 - DNML | DZ | IMP */
649 	FPE_FLTINV,	/* 27 - INV | DNML | DZ | IMP */
650 	FPE_FLTOVF,	/* 28 - OFL | IMP */
651 	FPE_FLTINV,	/* 29 - INV | OFL | IMP */
652 	FPE_FLTUND,	/* 2A - DNML | OFL | IMP */
653 	FPE_FLTINV,	/* 2B - INV | DNML | OFL | IMP */
654 	FPE_FLTDIV,	/* 2C - DZ | OFL | IMP */
655 	FPE_FLTINV,	/* 2D - INV | DZ | OFL | IMP */
656 	FPE_FLTDIV,	/* 2E - DNML | DZ | OFL | IMP */
657 	FPE_FLTINV,	/* 2F - INV | DNML | DZ | OFL | IMP */
658 	FPE_FLTUND,	/* 30 - UFL | IMP */
659 	FPE_FLTINV,	/* 31 - INV | UFL | IMP */
660 	FPE_FLTUND,	/* 32 - DNML | UFL | IMP */
661 	FPE_FLTINV,	/* 33 - INV | DNML | UFL | IMP */
662 	FPE_FLTDIV,	/* 34 - DZ | UFL | IMP */
663 	FPE_FLTINV,	/* 35 - INV | DZ | UFL | IMP */
664 	FPE_FLTDIV,	/* 36 - DNML | DZ | UFL | IMP */
665 	FPE_FLTINV,	/* 37 - INV | DNML | DZ | UFL | IMP */
666 	FPE_FLTOVF,	/* 38 - OFL | UFL | IMP */
667 	FPE_FLTINV,	/* 39 - INV | OFL | UFL | IMP */
668 	FPE_FLTUND,	/* 3A - DNML | OFL | UFL | IMP */
669 	FPE_FLTINV,	/* 3B - INV | DNML | OFL | UFL | IMP */
670 	FPE_FLTDIV,	/* 3C - DZ | OFL | UFL | IMP */
671 	FPE_FLTINV,	/* 3D - INV | DZ | OFL | UFL | IMP */
672 	FPE_FLTDIV,	/* 3E - DNML | DZ | OFL | UFL | IMP */
673 	FPE_FLTINV,	/* 3F - INV | DNML | DZ | OFL | UFL | IMP */
674 	FPE_FLTSUB,	/* 40 - STK */
675 	FPE_FLTSUB,	/* 41 - INV | STK */
676 	FPE_FLTUND,	/* 42 - DNML | STK */
677 	FPE_FLTSUB,	/* 43 - INV | DNML | STK */
678 	FPE_FLTDIV,	/* 44 - DZ | STK */
679 	FPE_FLTSUB,	/* 45 - INV | DZ | STK */
680 	FPE_FLTDIV,	/* 46 - DNML | DZ | STK */
681 	FPE_FLTSUB,	/* 47 - INV | DNML | DZ | STK */
682 	FPE_FLTOVF,	/* 48 - OFL | STK */
683 	FPE_FLTSUB,	/* 49 - INV | OFL | STK */
684 	FPE_FLTUND,	/* 4A - DNML | OFL | STK */
685 	FPE_FLTSUB,	/* 4B - INV | DNML | OFL | STK */
686 	FPE_FLTDIV,	/* 4C - DZ | OFL | STK */
687 	FPE_FLTSUB,	/* 4D - INV | DZ | OFL | STK */
688 	FPE_FLTDIV,	/* 4E - DNML | DZ | OFL | STK */
689 	FPE_FLTSUB,	/* 4F - INV | DNML | DZ | OFL | STK */
690 	FPE_FLTUND,	/* 50 - UFL | STK */
691 	FPE_FLTSUB,	/* 51 - INV | UFL | STK */
692 	FPE_FLTUND,	/* 52 - DNML | UFL | STK */
693 	FPE_FLTSUB,	/* 53 - INV | DNML | UFL | STK */
694 	FPE_FLTDIV,	/* 54 - DZ | UFL | STK */
695 	FPE_FLTSUB,	/* 55 - INV | DZ | UFL | STK */
696 	FPE_FLTDIV,	/* 56 - DNML | DZ | UFL | STK */
697 	FPE_FLTSUB,	/* 57 - INV | DNML | DZ | UFL | STK */
698 	FPE_FLTOVF,	/* 58 - OFL | UFL | STK */
699 	FPE_FLTSUB,	/* 59 - INV | OFL | UFL | STK */
700 	FPE_FLTUND,	/* 5A - DNML | OFL | UFL | STK */
701 	FPE_FLTSUB,	/* 5B - INV | DNML | OFL | UFL | STK */
702 	FPE_FLTDIV,	/* 5C - DZ | OFL | UFL | STK */
703 	FPE_FLTSUB,	/* 5D - INV | DZ | OFL | UFL | STK */
704 	FPE_FLTDIV,	/* 5E - DNML | DZ | OFL | UFL | STK */
705 	FPE_FLTSUB,	/* 5F - INV | DNML | DZ | OFL | UFL | STK */
706 	FPE_FLTRES,	/* 60 - IMP | STK */
707 	FPE_FLTSUB,	/* 61 - INV | IMP | STK */
708 	FPE_FLTUND,	/* 62 - DNML | IMP | STK */
709 	FPE_FLTSUB,	/* 63 - INV | DNML | IMP | STK */
710 	FPE_FLTDIV,	/* 64 - DZ | IMP | STK */
711 	FPE_FLTSUB,	/* 65 - INV | DZ | IMP | STK */
712 	FPE_FLTDIV,	/* 66 - DNML | DZ | IMP | STK */
713 	FPE_FLTSUB,	/* 67 - INV | DNML | DZ | IMP | STK */
714 	FPE_FLTOVF,	/* 68 - OFL | IMP | STK */
715 	FPE_FLTSUB,	/* 69 - INV | OFL | IMP | STK */
716 	FPE_FLTUND,	/* 6A - DNML | OFL | IMP | STK */
717 	FPE_FLTSUB,	/* 6B - INV | DNML | OFL | IMP | STK */
718 	FPE_FLTDIV,	/* 6C - DZ | OFL | IMP | STK */
719 	FPE_FLTSUB,	/* 6D - INV | DZ | OFL | IMP | STK */
720 	FPE_FLTDIV,	/* 6E - DNML | DZ | OFL | IMP | STK */
721 	FPE_FLTSUB,	/* 6F - INV | DNML | DZ | OFL | IMP | STK */
722 	FPE_FLTUND,	/* 70 - UFL | IMP | STK */
723 	FPE_FLTSUB,	/* 71 - INV | UFL | IMP | STK */
724 	FPE_FLTUND,	/* 72 - DNML | UFL | IMP | STK */
725 	FPE_FLTSUB,	/* 73 - INV | DNML | UFL | IMP | STK */
726 	FPE_FLTDIV,	/* 74 - DZ | UFL | IMP | STK */
727 	FPE_FLTSUB,	/* 75 - INV | DZ | UFL | IMP | STK */
728 	FPE_FLTDIV,	/* 76 - DNML | DZ | UFL | IMP | STK */
729 	FPE_FLTSUB,	/* 77 - INV | DNML | DZ | UFL | IMP | STK */
730 	FPE_FLTOVF,	/* 78 - OFL | UFL | IMP | STK */
731 	FPE_FLTSUB,	/* 79 - INV | OFL | UFL | IMP | STK */
732 	FPE_FLTUND,	/* 7A - DNML | OFL | UFL | IMP | STK */
733 	FPE_FLTSUB,	/* 7B - INV | DNML | OFL | UFL | IMP | STK */
734 	FPE_FLTDIV,	/* 7C - DZ | OFL | UFL | IMP | STK */
735 	FPE_FLTSUB,	/* 7D - INV | DZ | OFL | UFL | IMP | STK */
736 	FPE_FLTDIV,	/* 7E - DNML | DZ | OFL | UFL | IMP | STK */
737 	FPE_FLTSUB,	/* 7F - INV | DNML | DZ | OFL | UFL | IMP | STK */
738 };
739 
740 /*
741  * Read the FP status and control words, then generate si_code value
742  * for SIGFPE.  The error code chosen will be one of the
743  * FPE_... macros.  It will be sent as the second argument to old
744  * BSD-style signal handlers and as "siginfo_t->si_code" (second
745  * argument) to SA_SIGINFO signal handlers.
746  *
747  * Some time ago, we cleared the x87 exceptions with FNCLEX there.
748  * Clearing exceptions was necessary mainly to avoid IRQ13 bugs.  The
749  * usermode code which understands the FPU hardware enough to enable
750  * the exceptions, can also handle clearing the exception state in the
751  * handler.  The only consequence of not clearing the exception is the
752  * rethrow of the SIGFPE on return from the signal handler and
753  * reexecution of the corresponding instruction.
754  *
755  * For XMM traps, the exceptions were never cleared.
756  */
757 int
758 npxtrap_x87(void)
759 {
760 	u_short control, status;
761 
762 	if (!hw_float) {
763 		printf(
764 	"npxtrap_x87: fpcurthread = %p, curthread = %p, hw_float = %d\n",
765 		       PCPU_GET(fpcurthread), curthread, hw_float);
766 		panic("npxtrap from nowhere");
767 	}
768 	critical_enter();
769 
770 	/*
771 	 * Interrupt handling (for another interrupt) may have pushed the
772 	 * state to memory.  Fetch the relevant parts of the state from
773 	 * wherever they are.
774 	 */
775 	if (PCPU_GET(fpcurthread) != curthread) {
776 		control = GET_FPU_CW(curthread);
777 		status = GET_FPU_SW(curthread);
778 	} else {
779 		fnstcw(&control);
780 		fnstsw(&status);
781 	}
782 	critical_exit();
783 	return (fpetable[status & ((~control & 0x3f) | 0x40)]);
784 }
785 
786 int
787 npxtrap_sse(void)
788 {
789 	u_int mxcsr;
790 
791 	if (!hw_float) {
792 		printf(
793 	"npxtrap_sse: fpcurthread = %p, curthread = %p, hw_float = %d\n",
794 		       PCPU_GET(fpcurthread), curthread, hw_float);
795 		panic("npxtrap from nowhere");
796 	}
797 	critical_enter();
798 	if (PCPU_GET(fpcurthread) != curthread)
799 		mxcsr = curthread->td_pcb->pcb_save->sv_xmm.sv_env.en_mxcsr;
800 	else
801 		stmxcsr(&mxcsr);
802 	critical_exit();
803 	return (fpetable[(mxcsr & (~mxcsr >> 7)) & 0x3f]);
804 }
805 
806 static void
807 restore_npx_curthread(struct thread *td, struct pcb *pcb)
808 {
809 
810 	/*
811 	 * Record new context early in case frstor causes a trap.
812 	 */
813 	PCPU_SET(fpcurthread, td);
814 
815 	stop_emulating();
816 	if (cpu_fxsr)
817 		fpu_clean_state();
818 
819 	if ((pcb->pcb_flags & PCB_NPXINITDONE) == 0) {
820 		/*
821 		 * This is the first time this thread has used the FPU or
822 		 * the PCB doesn't contain a clean FPU state.  Explicitly
823 		 * load an initial state.
824 		 *
825 		 * We prefer to restore the state from the actual save
826 		 * area in PCB instead of directly loading from
827 		 * npx_initialstate, to ignite the XSAVEOPT
828 		 * tracking engine.
829 		 */
830 		bcopy(npx_initialstate, pcb->pcb_save, cpu_max_ext_state_size);
831 		fpurstor(pcb->pcb_save);
832 		if (pcb->pcb_initial_npxcw != __INITIAL_NPXCW__)
833 			fldcw(pcb->pcb_initial_npxcw);
834 		pcb->pcb_flags |= PCB_NPXINITDONE;
835 		if (PCB_USER_FPU(pcb))
836 			pcb->pcb_flags |= PCB_NPXUSERINITDONE;
837 	} else {
838 		fpurstor(pcb->pcb_save);
839 	}
840 }
841 
842 /*
843  * Implement device not available (DNA) exception
844  *
845  * It would be better to switch FP context here (if curthread != fpcurthread)
846  * and not necessarily for every context switch, but it is too hard to
847  * access foreign pcb's.
848  */
849 int
850 npxdna(void)
851 {
852 	struct thread *td;
853 
854 	if (!hw_float)
855 		return (0);
856 	td = curthread;
857 	critical_enter();
858 
859 	KASSERT((curpcb->pcb_flags & PCB_NPXNOSAVE) == 0,
860 	    ("npxdna while in fpu_kern_enter(FPU_KERN_NOCTX)"));
861 	if (__predict_false(PCPU_GET(fpcurthread) == td)) {
862 		/*
863 		 * Some virtual machines seems to set %cr0.TS at
864 		 * arbitrary moments.  Silently clear the TS bit
865 		 * regardless of the eager/lazy FPU context switch
866 		 * mode.
867 		 */
868 		stop_emulating();
869 	} else {
870 		if (__predict_false(PCPU_GET(fpcurthread) != NULL)) {
871 			printf(
872 		    "npxdna: fpcurthread = %p (%d), curthread = %p (%d)\n",
873 			    PCPU_GET(fpcurthread),
874 			    PCPU_GET(fpcurthread)->td_proc->p_pid,
875 			    td, td->td_proc->p_pid);
876 			panic("npxdna");
877 		}
878 		restore_npx_curthread(td, td->td_pcb);
879 	}
880 	critical_exit();
881 	return (1);
882 }
883 
884 /*
885  * Wrapper for fpusave() called from context switch routines.
886  *
887  * npxsave() must be called with interrupts disabled, so that it clears
888  * fpcurthread atomically with saving the state.  We require callers to do the
889  * disabling, since most callers need to disable interrupts anyway to call
890  * npxsave() atomically with checking fpcurthread.
891  */
892 void
893 npxsave(union savefpu *addr)
894 {
895 
896 	stop_emulating();
897 	fpusave(addr);
898 }
899 
900 void npxswitch(struct thread *td, struct pcb *pcb);
901 void
902 npxswitch(struct thread *td, struct pcb *pcb)
903 {
904 
905 	if (lazy_fpu_switch || (td->td_pflags & TDP_KTHREAD) != 0 ||
906 	    !PCB_USER_FPU(pcb)) {
907 		start_emulating();
908 		PCPU_SET(fpcurthread, NULL);
909 	} else if (PCPU_GET(fpcurthread) != td) {
910 		restore_npx_curthread(td, pcb);
911 	}
912 }
913 
914 /*
915  * Unconditionally save the current co-processor state across suspend and
916  * resume.
917  */
918 void
919 npxsuspend(union savefpu *addr)
920 {
921 	register_t cr0;
922 
923 	if (!hw_float)
924 		return;
925 	if (PCPU_GET(fpcurthread) == NULL) {
926 		bcopy(npx_initialstate, addr, cpu_max_ext_state_size);
927 		return;
928 	}
929 	cr0 = rcr0();
930 	stop_emulating();
931 	fpusave(addr);
932 	load_cr0(cr0);
933 }
934 
935 void
936 npxresume(union savefpu *addr)
937 {
938 	register_t cr0;
939 
940 	if (!hw_float)
941 		return;
942 
943 	cr0 = rcr0();
944 	npxinit(false);
945 	stop_emulating();
946 	fpurstor(addr);
947 	load_cr0(cr0);
948 }
949 
950 void
951 npxdrop(void)
952 {
953 	struct thread *td;
954 
955 	/*
956 	 * Discard pending exceptions in the !cpu_fxsr case so that unmasked
957 	 * ones don't cause a panic on the next frstor.
958 	 */
959 	if (!cpu_fxsr)
960 		fnclex();
961 
962 	td = PCPU_GET(fpcurthread);
963 	KASSERT(td == curthread, ("fpudrop: fpcurthread != curthread"));
964 	CRITICAL_ASSERT(td);
965 	PCPU_SET(fpcurthread, NULL);
966 	td->td_pcb->pcb_flags &= ~PCB_NPXINITDONE;
967 	start_emulating();
968 }
969 
970 /*
971  * Get the user state of the FPU into pcb->pcb_user_save without
972  * dropping ownership (if possible).  It returns the FPU ownership
973  * status.
974  */
975 int
976 npxgetregs(struct thread *td)
977 {
978 	struct pcb *pcb;
979 	uint64_t *xstate_bv, bit;
980 	char *sa;
981 	int max_ext_n, i;
982 	int owned;
983 
984 	if (!hw_float)
985 		return (_MC_FPOWNED_NONE);
986 
987 	pcb = td->td_pcb;
988 	critical_enter();
989 	if ((pcb->pcb_flags & PCB_NPXINITDONE) == 0) {
990 		bcopy(npx_initialstate, get_pcb_user_save_pcb(pcb),
991 		    cpu_max_ext_state_size);
992 		SET_FPU_CW(get_pcb_user_save_pcb(pcb), pcb->pcb_initial_npxcw);
993 		npxuserinited(td);
994 		critical_exit();
995 		return (_MC_FPOWNED_PCB);
996 	}
997 	if (td == PCPU_GET(fpcurthread)) {
998 		fpusave(get_pcb_user_save_pcb(pcb));
999 		if (!cpu_fxsr)
1000 			/*
1001 			 * fnsave initializes the FPU and destroys whatever
1002 			 * context it contains.  Make sure the FPU owner
1003 			 * starts with a clean state next time.
1004 			 */
1005 			npxdrop();
1006 		owned = _MC_FPOWNED_FPU;
1007 	} else {
1008 		owned = _MC_FPOWNED_PCB;
1009 	}
1010 	if (use_xsave) {
1011 		/*
1012 		 * Handle partially saved state.
1013 		 */
1014 		sa = (char *)get_pcb_user_save_pcb(pcb);
1015 		xstate_bv = (uint64_t *)(sa + sizeof(union savefpu) +
1016 		    offsetof(struct xstate_hdr, xstate_bv));
1017 		if (xsave_mask >> 32 != 0)
1018 			max_ext_n = fls(xsave_mask >> 32) + 32;
1019 		else
1020 			max_ext_n = fls(xsave_mask);
1021 		for (i = 0; i < max_ext_n; i++) {
1022 			bit = 1ULL << i;
1023 			if ((xsave_mask & bit) == 0 || (*xstate_bv & bit) != 0)
1024 				continue;
1025 			bcopy((char *)npx_initialstate +
1026 			    xsave_area_desc[i].offset,
1027 			    sa + xsave_area_desc[i].offset,
1028 			    xsave_area_desc[i].size);
1029 			*xstate_bv |= bit;
1030 		}
1031 	}
1032 	critical_exit();
1033 	return (owned);
1034 }
1035 
1036 void
1037 npxuserinited(struct thread *td)
1038 {
1039 	struct pcb *pcb;
1040 
1041 	CRITICAL_ASSERT(td);
1042 	pcb = td->td_pcb;
1043 	if (PCB_USER_FPU(pcb))
1044 		pcb->pcb_flags |= PCB_NPXINITDONE;
1045 	pcb->pcb_flags |= PCB_NPXUSERINITDONE;
1046 }
1047 
1048 int
1049 npxsetxstate(struct thread *td, char *xfpustate, size_t xfpustate_size)
1050 {
1051 	struct xstate_hdr *hdr, *ehdr;
1052 	size_t len, max_len;
1053 	uint64_t bv;
1054 
1055 	/* XXXKIB should we clear all extended state in xstate_bv instead ? */
1056 	if (xfpustate == NULL)
1057 		return (0);
1058 	if (!use_xsave)
1059 		return (EOPNOTSUPP);
1060 
1061 	len = xfpustate_size;
1062 	if (len < sizeof(struct xstate_hdr))
1063 		return (EINVAL);
1064 	max_len = cpu_max_ext_state_size - sizeof(union savefpu);
1065 	if (len > max_len)
1066 		return (EINVAL);
1067 
1068 	ehdr = (struct xstate_hdr *)xfpustate;
1069 	bv = ehdr->xstate_bv;
1070 
1071 	/*
1072 	 * Avoid #gp.
1073 	 */
1074 	if (bv & ~xsave_mask)
1075 		return (EINVAL);
1076 
1077 	hdr = (struct xstate_hdr *)(get_pcb_user_save_td(td) + 1);
1078 
1079 	hdr->xstate_bv = bv;
1080 	bcopy(xfpustate + sizeof(struct xstate_hdr),
1081 	    (char *)(hdr + 1), len - sizeof(struct xstate_hdr));
1082 
1083 	return (0);
1084 }
1085 
1086 int
1087 npxsetregs(struct thread *td, union savefpu *addr, char *xfpustate,
1088 	size_t xfpustate_size)
1089 {
1090 	struct pcb *pcb;
1091 	int error;
1092 
1093 	if (!hw_float)
1094 		return (ENXIO);
1095 
1096 	if (cpu_fxsr)
1097 		addr->sv_xmm.sv_env.en_mxcsr &= cpu_mxcsr_mask;
1098 	pcb = td->td_pcb;
1099 	error = 0;
1100 	critical_enter();
1101 	if (td == PCPU_GET(fpcurthread) && PCB_USER_FPU(pcb)) {
1102 		error = npxsetxstate(td, xfpustate, xfpustate_size);
1103 		if (error == 0) {
1104 			if (!cpu_fxsr)
1105 				fnclex();	/* As in npxdrop(). */
1106 			bcopy(addr, get_pcb_user_save_td(td), sizeof(*addr));
1107 			fpurstor(get_pcb_user_save_td(td));
1108 			pcb->pcb_flags |= PCB_NPXUSERINITDONE | PCB_NPXINITDONE;
1109 		}
1110 	} else {
1111 		error = npxsetxstate(td, xfpustate, xfpustate_size);
1112 		if (error == 0) {
1113 			bcopy(addr, get_pcb_user_save_td(td), sizeof(*addr));
1114 			npxuserinited(td);
1115 		}
1116 	}
1117 	critical_exit();
1118 	return (error);
1119 }
1120 
1121 static void
1122 npx_fill_fpregs_xmm1(struct savexmm *sv_xmm, struct save87 *sv_87)
1123 {
1124 	struct env87 *penv_87;
1125 	struct envxmm *penv_xmm;
1126 	struct fpacc87 *fx_reg;
1127 	int i, st;
1128 	uint64_t mantissa;
1129 	uint16_t tw, exp;
1130 	uint8_t ab_tw;
1131 
1132 	penv_87 = &sv_87->sv_env;
1133 	penv_xmm = &sv_xmm->sv_env;
1134 
1135 	/* FPU control/status */
1136 	penv_87->en_cw = penv_xmm->en_cw;
1137 	penv_87->en_sw = penv_xmm->en_sw;
1138 	penv_87->en_fip = penv_xmm->en_fip;
1139 	penv_87->en_fcs = penv_xmm->en_fcs;
1140 	penv_87->en_opcode = penv_xmm->en_opcode;
1141 	penv_87->en_foo = penv_xmm->en_foo;
1142 	penv_87->en_fos = penv_xmm->en_fos;
1143 
1144 	/*
1145 	 * FPU registers and tags.
1146 	 * For ST(i), i = fpu_reg - top; we start with fpu_reg=7.
1147 	 */
1148 	st = 7 - ((penv_xmm->en_sw >> 11) & 7);
1149 	ab_tw = penv_xmm->en_tw;
1150 	tw = 0;
1151 	for (i = 0x80; i != 0; i >>= 1) {
1152 		sv_87->sv_ac[st] = sv_xmm->sv_fp[st].fp_acc;
1153 		tw <<= 2;
1154 		if (ab_tw & i) {
1155 			/* Non-empty - we need to check ST(i) */
1156 			fx_reg = &sv_xmm->sv_fp[st].fp_acc;
1157 			/* The first 64 bits contain the mantissa. */
1158 			mantissa = *((uint64_t *)fx_reg->fp_bytes);
1159 			/*
1160 			 * The final 16 bits contain the sign bit and the exponent.
1161 			 * Mask the sign bit since it is of no consequence to these
1162 			 * tests.
1163 			 */
1164 			exp = *((uint16_t *)&fx_reg->fp_bytes[8]) & 0x7fff;
1165 			if (exp == 0) {
1166 				if (mantissa == 0)
1167 					tw |= 1; /* Zero */
1168 				else
1169 					tw |= 2; /* Denormal */
1170 			} else if (exp == 0x7fff)
1171 				tw |= 2; /* Infinity or NaN */
1172 		} else
1173 			tw |= 3; /* Empty */
1174 		st = (st - 1) & 7;
1175 	}
1176 	penv_87->en_tw = tw;
1177 }
1178 
1179 void
1180 npx_fill_fpregs_xmm(struct savexmm *sv_xmm, struct save87 *sv_87)
1181 {
1182 
1183 	bzero(sv_87, sizeof(*sv_87));
1184 	npx_fill_fpregs_xmm1(sv_xmm, sv_87);
1185 }
1186 
1187 void
1188 npx_set_fpregs_xmm(struct save87 *sv_87, struct savexmm *sv_xmm)
1189 {
1190 	struct env87 *penv_87;
1191 	struct envxmm *penv_xmm;
1192 	int i;
1193 
1194 	penv_87 = &sv_87->sv_env;
1195 	penv_xmm = &sv_xmm->sv_env;
1196 
1197 	/* FPU control/status */
1198 	penv_xmm->en_cw = penv_87->en_cw;
1199 	penv_xmm->en_sw = penv_87->en_sw;
1200 	penv_xmm->en_fip = penv_87->en_fip;
1201 	penv_xmm->en_fcs = penv_87->en_fcs;
1202 	penv_xmm->en_opcode = penv_87->en_opcode;
1203 	penv_xmm->en_foo = penv_87->en_foo;
1204 	penv_xmm->en_fos = penv_87->en_fos;
1205 
1206 	/*
1207 	 * FPU registers and tags.
1208 	 * Abridged  /  Full translation (values in binary), see FXSAVE spec.
1209 	 * 0		11
1210 	 * 1		00, 01, 10
1211 	 */
1212 	penv_xmm->en_tw = 0;
1213 	for (i = 0; i < 8; ++i) {
1214 		sv_xmm->sv_fp[i].fp_acc = sv_87->sv_ac[i];
1215 		if ((penv_87->en_tw & (3 << i * 2)) != (3 << i * 2))
1216 			penv_xmm->en_tw |= 1 << i;
1217 	}
1218 }
1219 
1220 void
1221 npx_get_fsave(void *addr)
1222 {
1223 	struct thread *td;
1224 	union savefpu *sv;
1225 
1226 	td = curthread;
1227 	npxgetregs(td);
1228 	sv = get_pcb_user_save_td(td);
1229 	if (cpu_fxsr)
1230 		npx_fill_fpregs_xmm1(&sv->sv_xmm, addr);
1231 	else
1232 		bcopy(sv, addr, sizeof(struct env87) +
1233 		    sizeof(struct fpacc87[8]));
1234 }
1235 
1236 int
1237 npx_set_fsave(void *addr)
1238 {
1239 	union savefpu sv;
1240 	int error;
1241 
1242 	bzero(&sv, sizeof(sv));
1243 	if (cpu_fxsr)
1244 		npx_set_fpregs_xmm(addr, &sv.sv_xmm);
1245 	else
1246 		bcopy(addr, &sv, sizeof(struct env87) +
1247 		    sizeof(struct fpacc87[8]));
1248 	error = npxsetregs(curthread, &sv, NULL, 0);
1249 	return (error);
1250 }
1251 
1252 /*
1253  * On AuthenticAMD processors, the fxrstor instruction does not restore
1254  * the x87's stored last instruction pointer, last data pointer, and last
1255  * opcode values, except in the rare case in which the exception summary
1256  * (ES) bit in the x87 status word is set to 1.
1257  *
1258  * In order to avoid leaking this information across processes, we clean
1259  * these values by performing a dummy load before executing fxrstor().
1260  */
1261 static void
1262 fpu_clean_state(void)
1263 {
1264 	static float dummy_variable = 0.0;
1265 	u_short status;
1266 
1267 	/*
1268 	 * Clear the ES bit in the x87 status word if it is currently
1269 	 * set, in order to avoid causing a fault in the upcoming load.
1270 	 */
1271 	fnstsw(&status);
1272 	if (status & 0x80)
1273 		fnclex();
1274 
1275 	/*
1276 	 * Load the dummy variable into the x87 stack.  This mangles
1277 	 * the x87 stack, but we don't care since we're about to call
1278 	 * fxrstor() anyway.
1279 	 */
1280 	__asm __volatile("ffree %%st(7); flds %0" : : "m" (dummy_variable));
1281 }
1282 
1283 static void
1284 fpurstor(union savefpu *addr)
1285 {
1286 
1287 	if (use_xsave)
1288 		xrstor((char *)addr, xsave_mask);
1289 	else if (cpu_fxsr)
1290 		fxrstor(addr);
1291 	else
1292 		frstor(addr);
1293 }
1294 
1295 #ifdef DEV_ISA
1296 /*
1297  * This sucks up the legacy ISA support assignments from PNPBIOS/ACPI.
1298  */
1299 static struct isa_pnp_id npxisa_ids[] = {
1300 	{ 0x040cd041, "Legacy ISA coprocessor support" }, /* PNP0C04 */
1301 	{ 0 }
1302 };
1303 
1304 static int
1305 npxisa_probe(device_t dev)
1306 {
1307 	int result;
1308 	if ((result = ISA_PNP_PROBE(device_get_parent(dev), dev, npxisa_ids)) <= 0) {
1309 		device_quiet(dev);
1310 	}
1311 	return(result);
1312 }
1313 
1314 static int
1315 npxisa_attach(device_t dev)
1316 {
1317 	return (0);
1318 }
1319 
1320 static device_method_t npxisa_methods[] = {
1321 	/* Device interface */
1322 	DEVMETHOD(device_probe,		npxisa_probe),
1323 	DEVMETHOD(device_attach,	npxisa_attach),
1324 	DEVMETHOD(device_detach,	bus_generic_detach),
1325 	DEVMETHOD(device_shutdown,	bus_generic_shutdown),
1326 	DEVMETHOD(device_suspend,	bus_generic_suspend),
1327 	DEVMETHOD(device_resume,	bus_generic_resume),
1328 	{ 0, 0 }
1329 };
1330 
1331 static driver_t npxisa_driver = {
1332 	"npxisa",
1333 	npxisa_methods,
1334 	1,			/* no softc */
1335 };
1336 
1337 DRIVER_MODULE(npxisa, isa, npxisa_driver, 0, 0);
1338 DRIVER_MODULE(npxisa, acpi, npxisa_driver, 0, 0);
1339 ISA_PNP_INFO(npxisa_ids);
1340 #endif /* DEV_ISA */
1341 
1342 static MALLOC_DEFINE(M_FPUKERN_CTX, "fpukern_ctx",
1343     "Kernel contexts for FPU state");
1344 
1345 #define	FPU_KERN_CTX_NPXINITDONE 0x01
1346 #define	FPU_KERN_CTX_DUMMY	 0x02
1347 #define	FPU_KERN_CTX_INUSE	 0x04
1348 
1349 struct fpu_kern_ctx {
1350 	union savefpu *prev;
1351 	uint32_t flags;
1352 	char hwstate1[];
1353 };
1354 
1355 struct fpu_kern_ctx *
1356 fpu_kern_alloc_ctx(u_int flags)
1357 {
1358 	struct fpu_kern_ctx *res;
1359 	size_t sz;
1360 
1361 	sz = sizeof(struct fpu_kern_ctx) + XSAVE_AREA_ALIGN +
1362 	    cpu_max_ext_state_size;
1363 	res = malloc(sz, M_FPUKERN_CTX, ((flags & FPU_KERN_NOWAIT) ?
1364 	    M_NOWAIT : M_WAITOK) | M_ZERO);
1365 	return (res);
1366 }
1367 
1368 void
1369 fpu_kern_free_ctx(struct fpu_kern_ctx *ctx)
1370 {
1371 
1372 	KASSERT((ctx->flags & FPU_KERN_CTX_INUSE) == 0, ("free'ing inuse ctx"));
1373 	/* XXXKIB clear the memory ? */
1374 	free(ctx, M_FPUKERN_CTX);
1375 }
1376 
1377 static union savefpu *
1378 fpu_kern_ctx_savefpu(struct fpu_kern_ctx *ctx)
1379 {
1380 	vm_offset_t p;
1381 
1382 	p = (vm_offset_t)&ctx->hwstate1;
1383 	p = roundup2(p, XSAVE_AREA_ALIGN);
1384 	return ((union savefpu *)p);
1385 }
1386 
1387 void
1388 fpu_kern_enter(struct thread *td, struct fpu_kern_ctx *ctx, u_int flags)
1389 {
1390 	struct pcb *pcb;
1391 
1392 	pcb = td->td_pcb;
1393 	KASSERT((flags & FPU_KERN_NOCTX) != 0 || ctx != NULL,
1394 	    ("ctx is required when !FPU_KERN_NOCTX"));
1395 	KASSERT(ctx == NULL || (ctx->flags & FPU_KERN_CTX_INUSE) == 0,
1396 	    ("using inuse ctx"));
1397 	KASSERT((pcb->pcb_flags & PCB_NPXNOSAVE) == 0,
1398 	    ("recursive fpu_kern_enter while in PCB_NPXNOSAVE state"));
1399 
1400 	if ((flags & FPU_KERN_NOCTX) != 0) {
1401 		critical_enter();
1402 		stop_emulating();
1403 		if (curthread == PCPU_GET(fpcurthread)) {
1404 			fpusave(curpcb->pcb_save);
1405 			PCPU_SET(fpcurthread, NULL);
1406 		} else {
1407 			KASSERT(PCPU_GET(fpcurthread) == NULL,
1408 			    ("invalid fpcurthread"));
1409 		}
1410 
1411 		/*
1412 		 * This breaks XSAVEOPT tracker, but
1413 		 * PCB_NPXNOSAVE state is supposed to never need to
1414 		 * save FPU context at all.
1415 		 */
1416 		fpurstor(npx_initialstate);
1417 		pcb->pcb_flags |= PCB_KERNNPX | PCB_NPXNOSAVE | PCB_NPXINITDONE;
1418 		return;
1419 	}
1420 	if ((flags & FPU_KERN_KTHR) != 0 && is_fpu_kern_thread(0)) {
1421 		ctx->flags = FPU_KERN_CTX_DUMMY | FPU_KERN_CTX_INUSE;
1422 		return;
1423 	}
1424 	pcb = td->td_pcb;
1425 	critical_enter();
1426 	KASSERT(!PCB_USER_FPU(pcb) || pcb->pcb_save ==
1427 	    get_pcb_user_save_pcb(pcb), ("mangled pcb_save"));
1428 	ctx->flags = FPU_KERN_CTX_INUSE;
1429 	if ((pcb->pcb_flags & PCB_NPXINITDONE) != 0)
1430 		ctx->flags |= FPU_KERN_CTX_NPXINITDONE;
1431 	npxexit(td);
1432 	ctx->prev = pcb->pcb_save;
1433 	pcb->pcb_save = fpu_kern_ctx_savefpu(ctx);
1434 	pcb->pcb_flags |= PCB_KERNNPX;
1435 	pcb->pcb_flags &= ~PCB_NPXINITDONE;
1436 	critical_exit();
1437 }
1438 
1439 int
1440 fpu_kern_leave(struct thread *td, struct fpu_kern_ctx *ctx)
1441 {
1442 	struct pcb *pcb;
1443 
1444 	pcb = td->td_pcb;
1445 
1446 	if ((pcb->pcb_flags & PCB_NPXNOSAVE) != 0) {
1447 		KASSERT(ctx == NULL, ("non-null ctx after FPU_KERN_NOCTX"));
1448 		KASSERT(PCPU_GET(fpcurthread) == NULL,
1449 		    ("non-NULL fpcurthread for PCB_NPXNOSAVE"));
1450 		CRITICAL_ASSERT(td);
1451 
1452 		pcb->pcb_flags &= ~(PCB_NPXNOSAVE | PCB_NPXINITDONE);
1453 		start_emulating();
1454 	} else {
1455 		KASSERT((ctx->flags & FPU_KERN_CTX_INUSE) != 0,
1456 		    ("leaving not inuse ctx"));
1457 		ctx->flags &= ~FPU_KERN_CTX_INUSE;
1458 
1459 		if (is_fpu_kern_thread(0) &&
1460 		    (ctx->flags & FPU_KERN_CTX_DUMMY) != 0)
1461 			return (0);
1462 		KASSERT((ctx->flags & FPU_KERN_CTX_DUMMY) == 0,
1463 		    ("dummy ctx"));
1464 		critical_enter();
1465 		if (curthread == PCPU_GET(fpcurthread))
1466 			npxdrop();
1467 		pcb->pcb_save = ctx->prev;
1468 	}
1469 
1470 	if (pcb->pcb_save == get_pcb_user_save_pcb(pcb)) {
1471 		if ((pcb->pcb_flags & PCB_NPXUSERINITDONE) != 0) {
1472 			pcb->pcb_flags |= PCB_NPXINITDONE;
1473 			if ((pcb->pcb_flags & PCB_KERNNPX_THR) == 0)
1474 				pcb->pcb_flags &= ~PCB_KERNNPX;
1475 		} else if ((pcb->pcb_flags & PCB_KERNNPX_THR) == 0)
1476 			pcb->pcb_flags &= ~(PCB_NPXINITDONE | PCB_KERNNPX);
1477 	} else {
1478 		if ((ctx->flags & FPU_KERN_CTX_NPXINITDONE) != 0)
1479 			pcb->pcb_flags |= PCB_NPXINITDONE;
1480 		else
1481 			pcb->pcb_flags &= ~PCB_NPXINITDONE;
1482 		KASSERT(!PCB_USER_FPU(pcb), ("unpaired fpu_kern_leave"));
1483 	}
1484 	critical_exit();
1485 	return (0);
1486 }
1487 
1488 int
1489 fpu_kern_thread(u_int flags)
1490 {
1491 
1492 	KASSERT((curthread->td_pflags & TDP_KTHREAD) != 0,
1493 	    ("Only kthread may use fpu_kern_thread"));
1494 	KASSERT(curpcb->pcb_save == get_pcb_user_save_pcb(curpcb),
1495 	    ("mangled pcb_save"));
1496 	KASSERT(PCB_USER_FPU(curpcb), ("recursive call"));
1497 
1498 	curpcb->pcb_flags |= PCB_KERNNPX | PCB_KERNNPX_THR;
1499 	return (0);
1500 }
1501 
1502 int
1503 is_fpu_kern_thread(u_int flags)
1504 {
1505 
1506 	if ((curthread->td_pflags & TDP_KTHREAD) == 0)
1507 		return (0);
1508 	return ((curpcb->pcb_flags & PCB_KERNNPX_THR) != 0);
1509 }
1510 
1511 /*
1512  * FPU save area alloc/free/init utility routines
1513  */
1514 union savefpu *
1515 fpu_save_area_alloc(void)
1516 {
1517 
1518 	return (uma_zalloc(fpu_save_area_zone, M_WAITOK));
1519 }
1520 
1521 void
1522 fpu_save_area_free(union savefpu *fsa)
1523 {
1524 
1525 	uma_zfree(fpu_save_area_zone, fsa);
1526 }
1527 
1528 void
1529 fpu_save_area_reset(union savefpu *fsa)
1530 {
1531 
1532 	bcopy(npx_initialstate, fsa, cpu_max_ext_state_size);
1533 }
1534