xref: /freebsd/sys/i386/i386/npx.c (revision 1f1e2261)
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 int use_xsave;
179 uint64_t xsave_mask;
180 static	uma_zone_t fpu_save_area_zone;
181 static	union savefpu *npx_initialstate;
182 
183 static struct xsave_area_elm_descr {
184 	u_int	offset;
185 	u_int	size;
186 } *xsave_area_desc;
187 
188 static	volatile u_int		npx_traps_while_probing;
189 
190 alias_for_inthand_t probetrap;
191 __asm("								\n\
192 	.text							\n\
193 	.p2align 2,0x90						\n\
194 	.type	" __XSTRING(CNAME(probetrap)) ",@function	\n\
195 " __XSTRING(CNAME(probetrap)) ":				\n\
196 	ss							\n\
197 	incl	" __XSTRING(CNAME(npx_traps_while_probing)) "	\n\
198 	fnclex							\n\
199 	iret							\n\
200 ");
201 
202 /*
203  * Determine if an FPU is present and how to use it.
204  */
205 static int
206 npx_probe(void)
207 {
208 	struct gate_descriptor save_idt_npxtrap;
209 	u_short control, status;
210 
211 	/*
212 	 * Modern CPUs all have an FPU that uses the INT16 interface
213 	 * and provide a simple way to verify that, so handle the
214 	 * common case right away.
215 	 */
216 	if (cpu_feature & CPUID_FPU) {
217 		hw_float = 1;
218 		return (1);
219 	}
220 
221 	save_idt_npxtrap = idt[IDT_MF];
222 	setidt(IDT_MF, probetrap, SDT_SYS386TGT, SEL_KPL,
223 	    GSEL(GCODE_SEL, SEL_KPL));
224 
225 	/*
226 	 * Don't trap while we're probing.
227 	 */
228 	stop_emulating();
229 
230 	/*
231 	 * Finish resetting the coprocessor, if any.  If there is an error
232 	 * pending, then we may get a bogus IRQ13, but npx_intr() will handle
233 	 * it OK.  Bogus halts have never been observed, but we enabled
234 	 * IRQ13 and cleared the BUSY# latch early to handle them anyway.
235 	 */
236 	fninit();
237 
238 	/*
239 	 * Don't use fwait here because it might hang.
240 	 * Don't use fnop here because it usually hangs if there is no FPU.
241 	 */
242 	DELAY(1000);		/* wait for any IRQ13 */
243 #ifdef DIAGNOSTIC
244 	if (npx_traps_while_probing != 0)
245 		printf("fninit caused %u bogus npx trap(s)\n",
246 		       npx_traps_while_probing);
247 #endif
248 	/*
249 	 * Check for a status of mostly zero.
250 	 */
251 	status = 0x5a5a;
252 	fnstsw(&status);
253 	if ((status & 0xb8ff) == 0) {
254 		/*
255 		 * Good, now check for a proper control word.
256 		 */
257 		control = 0x5a5a;
258 		fnstcw(&control);
259 		if ((control & 0x1f3f) == 0x033f) {
260 			/*
261 			 * We have an npx, now divide by 0 to see if exception
262 			 * 16 works.
263 			 */
264 			control &= ~(1 << 2);	/* enable divide by 0 trap */
265 			fldcw(control);
266 			npx_traps_while_probing = 0;
267 			fp_divide_by_0();
268 			if (npx_traps_while_probing != 0) {
269 				/*
270 				 * Good, exception 16 works.
271 				 */
272 				hw_float = 1;
273 				goto cleanup;
274 			}
275 			printf(
276 	"FPU does not use exception 16 for error reporting\n");
277 			goto cleanup;
278 		}
279 	}
280 
281 	/*
282 	 * Probe failed.  Floating point simply won't work.
283 	 * Notify user and disable FPU/MMX/SSE instruction execution.
284 	 */
285 	printf("WARNING: no FPU!\n");
286 	__asm __volatile("smsw %%ax; orb %0,%%al; lmsw %%ax" : :
287 	    "n" (CR0_EM | CR0_MP) : "ax");
288 
289 cleanup:
290 	idt[IDT_MF] = save_idt_npxtrap;
291 	return (hw_float);
292 }
293 
294 static void
295 fpusave_xsaveopt(union savefpu *addr)
296 {
297 
298 	xsaveopt((char *)addr, xsave_mask);
299 }
300 
301 static void
302 fpusave_xsave(union savefpu *addr)
303 {
304 
305 	xsave((char *)addr, xsave_mask);
306 }
307 
308 static void
309 fpusave_fxsave(union savefpu *addr)
310 {
311 
312 	fxsave((char *)addr);
313 }
314 
315 static void
316 fpusave_fnsave(union savefpu *addr)
317 {
318 
319 	fnsave((char *)addr);
320 }
321 
322 static void
323 init_xsave(void)
324 {
325 
326 	if (use_xsave)
327 		return;
328 	if (!cpu_fxsr || (cpu_feature2 & CPUID2_XSAVE) == 0)
329 		return;
330 	use_xsave = 1;
331 	TUNABLE_INT_FETCH("hw.use_xsave", &use_xsave);
332 }
333 
334 DEFINE_IFUNC(, void, fpusave, (union savefpu *))
335 {
336 
337 	init_xsave();
338 	if (use_xsave)
339 		return ((cpu_stdext_feature & CPUID_EXTSTATE_XSAVEOPT) != 0 ?
340 		    fpusave_xsaveopt : fpusave_xsave);
341 	if (cpu_fxsr)
342 		return (fpusave_fxsave);
343 	return (fpusave_fnsave);
344 }
345 
346 /*
347  * Enable XSAVE if supported and allowed by user.
348  * Calculate the xsave_mask.
349  */
350 static void
351 npxinit_bsp1(void)
352 {
353 	u_int cp[4];
354 	uint64_t xsave_mask_user;
355 
356 	TUNABLE_INT_FETCH("hw.lazy_fpu_switch", &lazy_fpu_switch);
357 	if (!use_xsave)
358 		return;
359 	cpuid_count(0xd, 0x0, cp);
360 	xsave_mask = XFEATURE_ENABLED_X87 | XFEATURE_ENABLED_SSE;
361 	if ((cp[0] & xsave_mask) != xsave_mask)
362 		panic("CPU0 does not support X87 or SSE: %x", cp[0]);
363 	xsave_mask = ((uint64_t)cp[3] << 32) | cp[0];
364 	xsave_mask_user = xsave_mask;
365 	TUNABLE_QUAD_FETCH("hw.xsave_mask", &xsave_mask_user);
366 	xsave_mask_user |= XFEATURE_ENABLED_X87 | XFEATURE_ENABLED_SSE;
367 	xsave_mask &= xsave_mask_user;
368 	if ((xsave_mask & XFEATURE_AVX512) != XFEATURE_AVX512)
369 		xsave_mask &= ~XFEATURE_AVX512;
370 	if ((xsave_mask & XFEATURE_MPX) != XFEATURE_MPX)
371 		xsave_mask &= ~XFEATURE_MPX;
372 }
373 
374 /*
375  * Calculate the fpu save area size.
376  */
377 static void
378 npxinit_bsp2(void)
379 {
380 	u_int cp[4];
381 
382 	if (use_xsave) {
383 		cpuid_count(0xd, 0x0, cp);
384 		cpu_max_ext_state_size = cp[1];
385 
386 		/*
387 		 * Reload the cpu_feature2, since we enabled OSXSAVE.
388 		 */
389 		do_cpuid(1, cp);
390 		cpu_feature2 = cp[2];
391 	} else
392 		cpu_max_ext_state_size = sizeof(union savefpu);
393 }
394 
395 /*
396  * Initialize floating point unit.
397  */
398 void
399 npxinit(bool bsp)
400 {
401 	static union savefpu dummy;
402 	register_t saveintr;
403 	u_int mxcsr;
404 	u_short control;
405 
406 	if (bsp) {
407 		if (!npx_probe())
408 			return;
409 		npxinit_bsp1();
410 	}
411 
412 	if (use_xsave) {
413 		load_cr4(rcr4() | CR4_XSAVE);
414 		load_xcr(XCR0, xsave_mask);
415 	}
416 
417 	/*
418 	 * XCR0 shall be set up before CPU can report the save area size.
419 	 */
420 	if (bsp)
421 		npxinit_bsp2();
422 
423 	/*
424 	 * fninit has the same h/w bugs as fnsave.  Use the detoxified
425 	 * fnsave to throw away any junk in the fpu.  fpusave() initializes
426 	 * the fpu.
427 	 *
428 	 * It is too early for critical_enter() to work on AP.
429 	 */
430 	saveintr = intr_disable();
431 	stop_emulating();
432 	if (cpu_fxsr)
433 		fninit();
434 	else
435 		fnsave(&dummy);
436 	control = __INITIAL_NPXCW__;
437 	fldcw(control);
438 	if (cpu_fxsr) {
439 		mxcsr = __INITIAL_MXCSR__;
440 		ldmxcsr(mxcsr);
441 	}
442 	start_emulating();
443 	intr_restore(saveintr);
444 }
445 
446 /*
447  * On the boot CPU we generate a clean state that is used to
448  * initialize the floating point unit when it is first used by a
449  * process.
450  */
451 static void
452 npxinitstate(void *arg __unused)
453 {
454 	uint64_t *xstate_bv;
455 	register_t saveintr;
456 	int cp[4], i, max_ext_n;
457 
458 	if (!hw_float)
459 		return;
460 
461 	/* Do potentially blocking operations before disabling interrupts. */
462 	fpu_save_area_zone = uma_zcreate("FPU_save_area",
463 	    cpu_max_ext_state_size, NULL, NULL, NULL, NULL,
464 	    XSAVE_AREA_ALIGN - 1, 0);
465 	npx_initialstate = uma_zalloc(fpu_save_area_zone, M_WAITOK | M_ZERO);
466 	if (use_xsave) {
467 		if (xsave_mask >> 32 != 0)
468 			max_ext_n = fls(xsave_mask >> 32) + 32;
469 		else
470 			max_ext_n = fls(xsave_mask);
471 		xsave_area_desc = malloc(max_ext_n * sizeof(struct
472 		    xsave_area_elm_descr), M_DEVBUF, M_WAITOK | M_ZERO);
473 	}
474 
475 	saveintr = intr_disable();
476 	stop_emulating();
477 
478 	if (cpu_fxsr)
479 		fpusave_fxsave(npx_initialstate);
480 	else
481 		fpusave_fnsave(npx_initialstate);
482 	if (cpu_fxsr) {
483 		if (npx_initialstate->sv_xmm.sv_env.en_mxcsr_mask)
484 			cpu_mxcsr_mask =
485 			    npx_initialstate->sv_xmm.sv_env.en_mxcsr_mask;
486 		else
487 			cpu_mxcsr_mask = 0xFFBF;
488 
489 		/*
490 		 * The fninit instruction does not modify XMM
491 		 * registers or x87 registers (MM/ST).  The fpusave
492 		 * call dumped the garbage contained in the registers
493 		 * after reset to the initial state saved.  Clear XMM
494 		 * and x87 registers file image to make the startup
495 		 * program state and signal handler XMM/x87 register
496 		 * content predictable.
497 		 */
498 		bzero(npx_initialstate->sv_xmm.sv_fp,
499 		    sizeof(npx_initialstate->sv_xmm.sv_fp));
500 		bzero(npx_initialstate->sv_xmm.sv_xmm,
501 		    sizeof(npx_initialstate->sv_xmm.sv_xmm));
502 
503 	} else
504 		bzero(npx_initialstate->sv_87.sv_ac,
505 		    sizeof(npx_initialstate->sv_87.sv_ac));
506 
507 	/*
508 	 * Create a table describing the layout of the CPU Extended
509 	 * Save Area.  See Intel SDM rev. 075 Vol. 1 13.4.1 "Legacy
510 	 * Region of an XSAVE Area" for the source of offsets/sizes.
511 	 * Note that 32bit XSAVE does not use %xmm8-%xmm15, see
512 	 * 10.5.1.2 and 13.5.2 "SSE State".
513 	 */
514 	if (use_xsave) {
515 		xstate_bv = (uint64_t *)((char *)(npx_initialstate + 1) +
516 		    offsetof(struct xstate_hdr, xstate_bv));
517 		*xstate_bv = XFEATURE_ENABLED_X87 | XFEATURE_ENABLED_SSE;
518 
519 		/* x87 state */
520 		xsave_area_desc[0].offset = 0;
521 		xsave_area_desc[0].size = 160;
522 		/* XMM */
523 		xsave_area_desc[1].offset = 160;
524 		xsave_area_desc[1].size = 288 - 160;
525 
526 		for (i = 2; i < max_ext_n; i++) {
527 			cpuid_count(0xd, i, cp);
528 			xsave_area_desc[i].offset = cp[1];
529 			xsave_area_desc[i].size = cp[0];
530 		}
531 	}
532 
533 	start_emulating();
534 	intr_restore(saveintr);
535 }
536 SYSINIT(npxinitstate, SI_SUB_CPU, SI_ORDER_ANY, npxinitstate, NULL);
537 
538 /*
539  * Free coprocessor (if we have it).
540  */
541 void
542 npxexit(struct thread *td)
543 {
544 
545 	critical_enter();
546 	if (curthread == PCPU_GET(fpcurthread)) {
547 		stop_emulating();
548 		fpusave(curpcb->pcb_save);
549 		start_emulating();
550 		PCPU_SET(fpcurthread, NULL);
551 	}
552 	critical_exit();
553 #ifdef NPX_DEBUG
554 	if (hw_float) {
555 		u_int	masked_exceptions;
556 
557 		masked_exceptions = GET_FPU_CW(td) & GET_FPU_SW(td) & 0x7f;
558 		/*
559 		 * Log exceptions that would have trapped with the old
560 		 * control word (overflow, divide by 0, and invalid operand).
561 		 */
562 		if (masked_exceptions & 0x0d)
563 			log(LOG_ERR,
564 	"pid %d (%s) exited with masked floating point exceptions 0x%02x\n",
565 			    td->td_proc->p_pid, td->td_proc->p_comm,
566 			    masked_exceptions);
567 	}
568 #endif
569 }
570 
571 int
572 npxformat(void)
573 {
574 
575 	if (!hw_float)
576 		return (_MC_FPFMT_NODEV);
577 	if (cpu_fxsr)
578 		return (_MC_FPFMT_XMM);
579 	return (_MC_FPFMT_387);
580 }
581 
582 /*
583  * The following mechanism is used to ensure that the FPE_... value
584  * that is passed as a trapcode to the signal handler of the user
585  * process does not have more than one bit set.
586  *
587  * Multiple bits may be set if the user process modifies the control
588  * word while a status word bit is already set.  While this is a sign
589  * of bad coding, we have no choice than to narrow them down to one
590  * bit, since we must not send a trapcode that is not exactly one of
591  * the FPE_ macros.
592  *
593  * The mechanism has a static table with 127 entries.  Each combination
594  * of the 7 FPU status word exception bits directly translates to a
595  * position in this table, where a single FPE_... value is stored.
596  * This FPE_... value stored there is considered the "most important"
597  * of the exception bits and will be sent as the signal code.  The
598  * precedence of the bits is based upon Intel Document "Numerical
599  * Applications", Chapter "Special Computational Situations".
600  *
601  * The macro to choose one of these values does these steps: 1) Throw
602  * away status word bits that cannot be masked.  2) Throw away the bits
603  * currently masked in the control word, assuming the user isn't
604  * interested in them anymore.  3) Reinsert status word bit 7 (stack
605  * fault) if it is set, which cannot be masked but must be presered.
606  * 4) Use the remaining bits to point into the trapcode table.
607  *
608  * The 6 maskable bits in order of their preference, as stated in the
609  * above referenced Intel manual:
610  * 1  Invalid operation (FP_X_INV)
611  * 1a   Stack underflow
612  * 1b   Stack overflow
613  * 1c   Operand of unsupported format
614  * 1d   SNaN operand.
615  * 2  QNaN operand (not an exception, irrelavant here)
616  * 3  Any other invalid-operation not mentioned above or zero divide
617  *      (FP_X_INV, FP_X_DZ)
618  * 4  Denormal operand (FP_X_DNML)
619  * 5  Numeric over/underflow (FP_X_OFL, FP_X_UFL)
620  * 6  Inexact result (FP_X_IMP)
621  */
622 static char fpetable[128] = {
623 	0,
624 	FPE_FLTINV,	/*  1 - INV */
625 	FPE_FLTUND,	/*  2 - DNML */
626 	FPE_FLTINV,	/*  3 - INV | DNML */
627 	FPE_FLTDIV,	/*  4 - DZ */
628 	FPE_FLTINV,	/*  5 - INV | DZ */
629 	FPE_FLTDIV,	/*  6 - DNML | DZ */
630 	FPE_FLTINV,	/*  7 - INV | DNML | DZ */
631 	FPE_FLTOVF,	/*  8 - OFL */
632 	FPE_FLTINV,	/*  9 - INV | OFL */
633 	FPE_FLTUND,	/*  A - DNML | OFL */
634 	FPE_FLTINV,	/*  B - INV | DNML | OFL */
635 	FPE_FLTDIV,	/*  C - DZ | OFL */
636 	FPE_FLTINV,	/*  D - INV | DZ | OFL */
637 	FPE_FLTDIV,	/*  E - DNML | DZ | OFL */
638 	FPE_FLTINV,	/*  F - INV | DNML | DZ | OFL */
639 	FPE_FLTUND,	/* 10 - UFL */
640 	FPE_FLTINV,	/* 11 - INV | UFL */
641 	FPE_FLTUND,	/* 12 - DNML | UFL */
642 	FPE_FLTINV,	/* 13 - INV | DNML | UFL */
643 	FPE_FLTDIV,	/* 14 - DZ | UFL */
644 	FPE_FLTINV,	/* 15 - INV | DZ | UFL */
645 	FPE_FLTDIV,	/* 16 - DNML | DZ | UFL */
646 	FPE_FLTINV,	/* 17 - INV | DNML | DZ | UFL */
647 	FPE_FLTOVF,	/* 18 - OFL | UFL */
648 	FPE_FLTINV,	/* 19 - INV | OFL | UFL */
649 	FPE_FLTUND,	/* 1A - DNML | OFL | UFL */
650 	FPE_FLTINV,	/* 1B - INV | DNML | OFL | UFL */
651 	FPE_FLTDIV,	/* 1C - DZ | OFL | UFL */
652 	FPE_FLTINV,	/* 1D - INV | DZ | OFL | UFL */
653 	FPE_FLTDIV,	/* 1E - DNML | DZ | OFL | UFL */
654 	FPE_FLTINV,	/* 1F - INV | DNML | DZ | OFL | UFL */
655 	FPE_FLTRES,	/* 20 - IMP */
656 	FPE_FLTINV,	/* 21 - INV | IMP */
657 	FPE_FLTUND,	/* 22 - DNML | IMP */
658 	FPE_FLTINV,	/* 23 - INV | DNML | IMP */
659 	FPE_FLTDIV,	/* 24 - DZ | IMP */
660 	FPE_FLTINV,	/* 25 - INV | DZ | IMP */
661 	FPE_FLTDIV,	/* 26 - DNML | DZ | IMP */
662 	FPE_FLTINV,	/* 27 - INV | DNML | DZ | IMP */
663 	FPE_FLTOVF,	/* 28 - OFL | IMP */
664 	FPE_FLTINV,	/* 29 - INV | OFL | IMP */
665 	FPE_FLTUND,	/* 2A - DNML | OFL | IMP */
666 	FPE_FLTINV,	/* 2B - INV | DNML | OFL | IMP */
667 	FPE_FLTDIV,	/* 2C - DZ | OFL | IMP */
668 	FPE_FLTINV,	/* 2D - INV | DZ | OFL | IMP */
669 	FPE_FLTDIV,	/* 2E - DNML | DZ | OFL | IMP */
670 	FPE_FLTINV,	/* 2F - INV | DNML | DZ | OFL | IMP */
671 	FPE_FLTUND,	/* 30 - UFL | IMP */
672 	FPE_FLTINV,	/* 31 - INV | UFL | IMP */
673 	FPE_FLTUND,	/* 32 - DNML | UFL | IMP */
674 	FPE_FLTINV,	/* 33 - INV | DNML | UFL | IMP */
675 	FPE_FLTDIV,	/* 34 - DZ | UFL | IMP */
676 	FPE_FLTINV,	/* 35 - INV | DZ | UFL | IMP */
677 	FPE_FLTDIV,	/* 36 - DNML | DZ | UFL | IMP */
678 	FPE_FLTINV,	/* 37 - INV | DNML | DZ | UFL | IMP */
679 	FPE_FLTOVF,	/* 38 - OFL | UFL | IMP */
680 	FPE_FLTINV,	/* 39 - INV | OFL | UFL | IMP */
681 	FPE_FLTUND,	/* 3A - DNML | OFL | UFL | IMP */
682 	FPE_FLTINV,	/* 3B - INV | DNML | OFL | UFL | IMP */
683 	FPE_FLTDIV,	/* 3C - DZ | OFL | UFL | IMP */
684 	FPE_FLTINV,	/* 3D - INV | DZ | OFL | UFL | IMP */
685 	FPE_FLTDIV,	/* 3E - DNML | DZ | OFL | UFL | IMP */
686 	FPE_FLTINV,	/* 3F - INV | DNML | DZ | OFL | UFL | IMP */
687 	FPE_FLTSUB,	/* 40 - STK */
688 	FPE_FLTSUB,	/* 41 - INV | STK */
689 	FPE_FLTUND,	/* 42 - DNML | STK */
690 	FPE_FLTSUB,	/* 43 - INV | DNML | STK */
691 	FPE_FLTDIV,	/* 44 - DZ | STK */
692 	FPE_FLTSUB,	/* 45 - INV | DZ | STK */
693 	FPE_FLTDIV,	/* 46 - DNML | DZ | STK */
694 	FPE_FLTSUB,	/* 47 - INV | DNML | DZ | STK */
695 	FPE_FLTOVF,	/* 48 - OFL | STK */
696 	FPE_FLTSUB,	/* 49 - INV | OFL | STK */
697 	FPE_FLTUND,	/* 4A - DNML | OFL | STK */
698 	FPE_FLTSUB,	/* 4B - INV | DNML | OFL | STK */
699 	FPE_FLTDIV,	/* 4C - DZ | OFL | STK */
700 	FPE_FLTSUB,	/* 4D - INV | DZ | OFL | STK */
701 	FPE_FLTDIV,	/* 4E - DNML | DZ | OFL | STK */
702 	FPE_FLTSUB,	/* 4F - INV | DNML | DZ | OFL | STK */
703 	FPE_FLTUND,	/* 50 - UFL | STK */
704 	FPE_FLTSUB,	/* 51 - INV | UFL | STK */
705 	FPE_FLTUND,	/* 52 - DNML | UFL | STK */
706 	FPE_FLTSUB,	/* 53 - INV | DNML | UFL | STK */
707 	FPE_FLTDIV,	/* 54 - DZ | UFL | STK */
708 	FPE_FLTSUB,	/* 55 - INV | DZ | UFL | STK */
709 	FPE_FLTDIV,	/* 56 - DNML | DZ | UFL | STK */
710 	FPE_FLTSUB,	/* 57 - INV | DNML | DZ | UFL | STK */
711 	FPE_FLTOVF,	/* 58 - OFL | UFL | STK */
712 	FPE_FLTSUB,	/* 59 - INV | OFL | UFL | STK */
713 	FPE_FLTUND,	/* 5A - DNML | OFL | UFL | STK */
714 	FPE_FLTSUB,	/* 5B - INV | DNML | OFL | UFL | STK */
715 	FPE_FLTDIV,	/* 5C - DZ | OFL | UFL | STK */
716 	FPE_FLTSUB,	/* 5D - INV | DZ | OFL | UFL | STK */
717 	FPE_FLTDIV,	/* 5E - DNML | DZ | OFL | UFL | STK */
718 	FPE_FLTSUB,	/* 5F - INV | DNML | DZ | OFL | UFL | STK */
719 	FPE_FLTRES,	/* 60 - IMP | STK */
720 	FPE_FLTSUB,	/* 61 - INV | IMP | STK */
721 	FPE_FLTUND,	/* 62 - DNML | IMP | STK */
722 	FPE_FLTSUB,	/* 63 - INV | DNML | IMP | STK */
723 	FPE_FLTDIV,	/* 64 - DZ | IMP | STK */
724 	FPE_FLTSUB,	/* 65 - INV | DZ | IMP | STK */
725 	FPE_FLTDIV,	/* 66 - DNML | DZ | IMP | STK */
726 	FPE_FLTSUB,	/* 67 - INV | DNML | DZ | IMP | STK */
727 	FPE_FLTOVF,	/* 68 - OFL | IMP | STK */
728 	FPE_FLTSUB,	/* 69 - INV | OFL | IMP | STK */
729 	FPE_FLTUND,	/* 6A - DNML | OFL | IMP | STK */
730 	FPE_FLTSUB,	/* 6B - INV | DNML | OFL | IMP | STK */
731 	FPE_FLTDIV,	/* 6C - DZ | OFL | IMP | STK */
732 	FPE_FLTSUB,	/* 6D - INV | DZ | OFL | IMP | STK */
733 	FPE_FLTDIV,	/* 6E - DNML | DZ | OFL | IMP | STK */
734 	FPE_FLTSUB,	/* 6F - INV | DNML | DZ | OFL | IMP | STK */
735 	FPE_FLTUND,	/* 70 - UFL | IMP | STK */
736 	FPE_FLTSUB,	/* 71 - INV | UFL | IMP | STK */
737 	FPE_FLTUND,	/* 72 - DNML | UFL | IMP | STK */
738 	FPE_FLTSUB,	/* 73 - INV | DNML | UFL | IMP | STK */
739 	FPE_FLTDIV,	/* 74 - DZ | UFL | IMP | STK */
740 	FPE_FLTSUB,	/* 75 - INV | DZ | UFL | IMP | STK */
741 	FPE_FLTDIV,	/* 76 - DNML | DZ | UFL | IMP | STK */
742 	FPE_FLTSUB,	/* 77 - INV | DNML | DZ | UFL | IMP | STK */
743 	FPE_FLTOVF,	/* 78 - OFL | UFL | IMP | STK */
744 	FPE_FLTSUB,	/* 79 - INV | OFL | UFL | IMP | STK */
745 	FPE_FLTUND,	/* 7A - DNML | OFL | UFL | IMP | STK */
746 	FPE_FLTSUB,	/* 7B - INV | DNML | OFL | UFL | IMP | STK */
747 	FPE_FLTDIV,	/* 7C - DZ | OFL | UFL | IMP | STK */
748 	FPE_FLTSUB,	/* 7D - INV | DZ | OFL | UFL | IMP | STK */
749 	FPE_FLTDIV,	/* 7E - DNML | DZ | OFL | UFL | IMP | STK */
750 	FPE_FLTSUB,	/* 7F - INV | DNML | DZ | OFL | UFL | IMP | STK */
751 };
752 
753 /*
754  * Read the FP status and control words, then generate si_code value
755  * for SIGFPE.  The error code chosen will be one of the
756  * FPE_... macros.  It will be sent as the second argument to old
757  * BSD-style signal handlers and as "siginfo_t->si_code" (second
758  * argument) to SA_SIGINFO signal handlers.
759  *
760  * Some time ago, we cleared the x87 exceptions with FNCLEX there.
761  * Clearing exceptions was necessary mainly to avoid IRQ13 bugs.  The
762  * usermode code which understands the FPU hardware enough to enable
763  * the exceptions, can also handle clearing the exception state in the
764  * handler.  The only consequence of not clearing the exception is the
765  * rethrow of the SIGFPE on return from the signal handler and
766  * reexecution of the corresponding instruction.
767  *
768  * For XMM traps, the exceptions were never cleared.
769  */
770 int
771 npxtrap_x87(void)
772 {
773 	u_short control, status;
774 
775 	if (!hw_float) {
776 		printf(
777 	"npxtrap_x87: fpcurthread = %p, curthread = %p, hw_float = %d\n",
778 		       PCPU_GET(fpcurthread), curthread, hw_float);
779 		panic("npxtrap from nowhere");
780 	}
781 	critical_enter();
782 
783 	/*
784 	 * Interrupt handling (for another interrupt) may have pushed the
785 	 * state to memory.  Fetch the relevant parts of the state from
786 	 * wherever they are.
787 	 */
788 	if (PCPU_GET(fpcurthread) != curthread) {
789 		control = GET_FPU_CW(curthread);
790 		status = GET_FPU_SW(curthread);
791 	} else {
792 		fnstcw(&control);
793 		fnstsw(&status);
794 	}
795 	critical_exit();
796 	return (fpetable[status & ((~control & 0x3f) | 0x40)]);
797 }
798 
799 int
800 npxtrap_sse(void)
801 {
802 	u_int mxcsr;
803 
804 	if (!hw_float) {
805 		printf(
806 	"npxtrap_sse: fpcurthread = %p, curthread = %p, hw_float = %d\n",
807 		       PCPU_GET(fpcurthread), curthread, hw_float);
808 		panic("npxtrap from nowhere");
809 	}
810 	critical_enter();
811 	if (PCPU_GET(fpcurthread) != curthread)
812 		mxcsr = curthread->td_pcb->pcb_save->sv_xmm.sv_env.en_mxcsr;
813 	else
814 		stmxcsr(&mxcsr);
815 	critical_exit();
816 	return (fpetable[(mxcsr & (~mxcsr >> 7)) & 0x3f]);
817 }
818 
819 static void
820 restore_npx_curthread(struct thread *td, struct pcb *pcb)
821 {
822 
823 	/*
824 	 * Record new context early in case frstor causes a trap.
825 	 */
826 	PCPU_SET(fpcurthread, td);
827 
828 	stop_emulating();
829 	if (cpu_fxsr)
830 		fpu_clean_state();
831 
832 	if ((pcb->pcb_flags & PCB_NPXINITDONE) == 0) {
833 		/*
834 		 * This is the first time this thread has used the FPU or
835 		 * the PCB doesn't contain a clean FPU state.  Explicitly
836 		 * load an initial state.
837 		 *
838 		 * We prefer to restore the state from the actual save
839 		 * area in PCB instead of directly loading from
840 		 * npx_initialstate, to ignite the XSAVEOPT
841 		 * tracking engine.
842 		 */
843 		bcopy(npx_initialstate, pcb->pcb_save, cpu_max_ext_state_size);
844 		fpurstor(pcb->pcb_save);
845 		if (pcb->pcb_initial_npxcw != __INITIAL_NPXCW__)
846 			fldcw(pcb->pcb_initial_npxcw);
847 		pcb->pcb_flags |= PCB_NPXINITDONE;
848 		if (PCB_USER_FPU(pcb))
849 			pcb->pcb_flags |= PCB_NPXUSERINITDONE;
850 	} else {
851 		fpurstor(pcb->pcb_save);
852 	}
853 }
854 
855 /*
856  * Implement device not available (DNA) exception
857  *
858  * It would be better to switch FP context here (if curthread != fpcurthread)
859  * and not necessarily for every context switch, but it is too hard to
860  * access foreign pcb's.
861  */
862 int
863 npxdna(void)
864 {
865 	struct thread *td;
866 
867 	if (!hw_float)
868 		return (0);
869 	td = curthread;
870 	critical_enter();
871 
872 	KASSERT((curpcb->pcb_flags & PCB_NPXNOSAVE) == 0,
873 	    ("npxdna while in fpu_kern_enter(FPU_KERN_NOCTX)"));
874 	if (__predict_false(PCPU_GET(fpcurthread) == td)) {
875 		/*
876 		 * Some virtual machines seems to set %cr0.TS at
877 		 * arbitrary moments.  Silently clear the TS bit
878 		 * regardless of the eager/lazy FPU context switch
879 		 * mode.
880 		 */
881 		stop_emulating();
882 	} else {
883 		if (__predict_false(PCPU_GET(fpcurthread) != NULL)) {
884 			printf(
885 		    "npxdna: fpcurthread = %p (%d), curthread = %p (%d)\n",
886 			    PCPU_GET(fpcurthread),
887 			    PCPU_GET(fpcurthread)->td_proc->p_pid,
888 			    td, td->td_proc->p_pid);
889 			panic("npxdna");
890 		}
891 		restore_npx_curthread(td, td->td_pcb);
892 	}
893 	critical_exit();
894 	return (1);
895 }
896 
897 /*
898  * Wrapper for fpusave() called from context switch routines.
899  *
900  * npxsave() must be called with interrupts disabled, so that it clears
901  * fpcurthread atomically with saving the state.  We require callers to do the
902  * disabling, since most callers need to disable interrupts anyway to call
903  * npxsave() atomically with checking fpcurthread.
904  */
905 void
906 npxsave(union savefpu *addr)
907 {
908 
909 	stop_emulating();
910 	fpusave(addr);
911 }
912 
913 void npxswitch(struct thread *td, struct pcb *pcb);
914 void
915 npxswitch(struct thread *td, struct pcb *pcb)
916 {
917 
918 	if (lazy_fpu_switch || (td->td_pflags & TDP_KTHREAD) != 0 ||
919 	    !PCB_USER_FPU(pcb)) {
920 		start_emulating();
921 		PCPU_SET(fpcurthread, NULL);
922 	} else if (PCPU_GET(fpcurthread) != td) {
923 		restore_npx_curthread(td, pcb);
924 	}
925 }
926 
927 /*
928  * Unconditionally save the current co-processor state across suspend and
929  * resume.
930  */
931 void
932 npxsuspend(union savefpu *addr)
933 {
934 	register_t cr0;
935 
936 	if (!hw_float)
937 		return;
938 	if (PCPU_GET(fpcurthread) == NULL) {
939 		bcopy(npx_initialstate, addr, cpu_max_ext_state_size);
940 		return;
941 	}
942 	cr0 = rcr0();
943 	stop_emulating();
944 	fpusave(addr);
945 	load_cr0(cr0);
946 }
947 
948 void
949 npxresume(union savefpu *addr)
950 {
951 	register_t cr0;
952 
953 	if (!hw_float)
954 		return;
955 
956 	cr0 = rcr0();
957 	npxinit(false);
958 	stop_emulating();
959 	fpurstor(addr);
960 	load_cr0(cr0);
961 }
962 
963 void
964 npxdrop(void)
965 {
966 	struct thread *td;
967 
968 	/*
969 	 * Discard pending exceptions in the !cpu_fxsr case so that unmasked
970 	 * ones don't cause a panic on the next frstor.
971 	 */
972 	if (!cpu_fxsr)
973 		fnclex();
974 
975 	td = PCPU_GET(fpcurthread);
976 	KASSERT(td == curthread, ("fpudrop: fpcurthread != curthread"));
977 	CRITICAL_ASSERT(td);
978 	PCPU_SET(fpcurthread, NULL);
979 	td->td_pcb->pcb_flags &= ~PCB_NPXINITDONE;
980 	start_emulating();
981 }
982 
983 /*
984  * Get the user state of the FPU into pcb->pcb_user_save without
985  * dropping ownership (if possible).  It returns the FPU ownership
986  * status.
987  */
988 int
989 npxgetregs(struct thread *td)
990 {
991 	struct pcb *pcb;
992 	uint64_t *xstate_bv, bit;
993 	char *sa;
994 	int max_ext_n, i;
995 	int owned;
996 
997 	if (!hw_float)
998 		return (_MC_FPOWNED_NONE);
999 
1000 	pcb = td->td_pcb;
1001 	critical_enter();
1002 	if ((pcb->pcb_flags & PCB_NPXINITDONE) == 0) {
1003 		bcopy(npx_initialstate, get_pcb_user_save_pcb(pcb),
1004 		    cpu_max_ext_state_size);
1005 		SET_FPU_CW(get_pcb_user_save_pcb(pcb), pcb->pcb_initial_npxcw);
1006 		npxuserinited(td);
1007 		critical_exit();
1008 		return (_MC_FPOWNED_PCB);
1009 	}
1010 	if (td == PCPU_GET(fpcurthread)) {
1011 		fpusave(get_pcb_user_save_pcb(pcb));
1012 		if (!cpu_fxsr)
1013 			/*
1014 			 * fnsave initializes the FPU and destroys whatever
1015 			 * context it contains.  Make sure the FPU owner
1016 			 * starts with a clean state next time.
1017 			 */
1018 			npxdrop();
1019 		owned = _MC_FPOWNED_FPU;
1020 	} else {
1021 		owned = _MC_FPOWNED_PCB;
1022 	}
1023 	if (use_xsave) {
1024 		/*
1025 		 * Handle partially saved state.
1026 		 */
1027 		sa = (char *)get_pcb_user_save_pcb(pcb);
1028 		xstate_bv = (uint64_t *)(sa + sizeof(union savefpu) +
1029 		    offsetof(struct xstate_hdr, xstate_bv));
1030 		if (xsave_mask >> 32 != 0)
1031 			max_ext_n = fls(xsave_mask >> 32) + 32;
1032 		else
1033 			max_ext_n = fls(xsave_mask);
1034 		for (i = 0; i < max_ext_n; i++) {
1035 			bit = 1ULL << i;
1036 			if ((xsave_mask & bit) == 0 || (*xstate_bv & bit) != 0)
1037 				continue;
1038 			bcopy((char *)npx_initialstate +
1039 			    xsave_area_desc[i].offset,
1040 			    sa + xsave_area_desc[i].offset,
1041 			    xsave_area_desc[i].size);
1042 			*xstate_bv |= bit;
1043 		}
1044 	}
1045 	critical_exit();
1046 	return (owned);
1047 }
1048 
1049 void
1050 npxuserinited(struct thread *td)
1051 {
1052 	struct pcb *pcb;
1053 
1054 	CRITICAL_ASSERT(td);
1055 	pcb = td->td_pcb;
1056 	if (PCB_USER_FPU(pcb))
1057 		pcb->pcb_flags |= PCB_NPXINITDONE;
1058 	pcb->pcb_flags |= PCB_NPXUSERINITDONE;
1059 }
1060 
1061 int
1062 npxsetxstate(struct thread *td, char *xfpustate, size_t xfpustate_size)
1063 {
1064 	struct xstate_hdr *hdr, *ehdr;
1065 	size_t len, max_len;
1066 	uint64_t bv;
1067 
1068 	/* XXXKIB should we clear all extended state in xstate_bv instead ? */
1069 	if (xfpustate == NULL)
1070 		return (0);
1071 	if (!use_xsave)
1072 		return (EOPNOTSUPP);
1073 
1074 	len = xfpustate_size;
1075 	if (len < sizeof(struct xstate_hdr))
1076 		return (EINVAL);
1077 	max_len = cpu_max_ext_state_size - sizeof(union savefpu);
1078 	if (len > max_len)
1079 		return (EINVAL);
1080 
1081 	ehdr = (struct xstate_hdr *)xfpustate;
1082 	bv = ehdr->xstate_bv;
1083 
1084 	/*
1085 	 * Avoid #gp.
1086 	 */
1087 	if (bv & ~xsave_mask)
1088 		return (EINVAL);
1089 
1090 	hdr = (struct xstate_hdr *)(get_pcb_user_save_td(td) + 1);
1091 
1092 	hdr->xstate_bv = bv;
1093 	bcopy(xfpustate + sizeof(struct xstate_hdr),
1094 	    (char *)(hdr + 1), len - sizeof(struct xstate_hdr));
1095 
1096 	return (0);
1097 }
1098 
1099 int
1100 npxsetregs(struct thread *td, union savefpu *addr, char *xfpustate,
1101 	size_t xfpustate_size)
1102 {
1103 	struct pcb *pcb;
1104 	int error;
1105 
1106 	if (!hw_float)
1107 		return (ENXIO);
1108 
1109 	if (cpu_fxsr)
1110 		addr->sv_xmm.sv_env.en_mxcsr &= cpu_mxcsr_mask;
1111 	pcb = td->td_pcb;
1112 	error = 0;
1113 	critical_enter();
1114 	if (td == PCPU_GET(fpcurthread) && PCB_USER_FPU(pcb)) {
1115 		error = npxsetxstate(td, xfpustate, xfpustate_size);
1116 		if (error == 0) {
1117 			if (!cpu_fxsr)
1118 				fnclex();	/* As in npxdrop(). */
1119 			bcopy(addr, get_pcb_user_save_td(td), sizeof(*addr));
1120 			fpurstor(get_pcb_user_save_td(td));
1121 			pcb->pcb_flags |= PCB_NPXUSERINITDONE | PCB_NPXINITDONE;
1122 		}
1123 	} else {
1124 		error = npxsetxstate(td, xfpustate, xfpustate_size);
1125 		if (error == 0) {
1126 			bcopy(addr, get_pcb_user_save_td(td), sizeof(*addr));
1127 			npxuserinited(td);
1128 		}
1129 	}
1130 	critical_exit();
1131 	return (error);
1132 }
1133 
1134 static void
1135 npx_fill_fpregs_xmm1(struct savexmm *sv_xmm, struct save87 *sv_87)
1136 {
1137 	struct env87 *penv_87;
1138 	struct envxmm *penv_xmm;
1139 	struct fpacc87 *fx_reg;
1140 	int i, st;
1141 	uint64_t mantissa;
1142 	uint16_t tw, exp;
1143 	uint8_t ab_tw;
1144 
1145 	penv_87 = &sv_87->sv_env;
1146 	penv_xmm = &sv_xmm->sv_env;
1147 
1148 	/* FPU control/status */
1149 	penv_87->en_cw = penv_xmm->en_cw;
1150 	penv_87->en_sw = penv_xmm->en_sw;
1151 	penv_87->en_fip = penv_xmm->en_fip;
1152 	penv_87->en_fcs = penv_xmm->en_fcs;
1153 	penv_87->en_opcode = penv_xmm->en_opcode;
1154 	penv_87->en_foo = penv_xmm->en_foo;
1155 	penv_87->en_fos = penv_xmm->en_fos;
1156 
1157 	/*
1158 	 * FPU registers and tags.
1159 	 * For ST(i), i = fpu_reg - top; we start with fpu_reg=7.
1160 	 */
1161 	st = 7 - ((penv_xmm->en_sw >> 11) & 7);
1162 	ab_tw = penv_xmm->en_tw;
1163 	tw = 0;
1164 	for (i = 0x80; i != 0; i >>= 1) {
1165 		sv_87->sv_ac[st] = sv_xmm->sv_fp[st].fp_acc;
1166 		tw <<= 2;
1167 		if (ab_tw & i) {
1168 			/* Non-empty - we need to check ST(i) */
1169 			fx_reg = &sv_xmm->sv_fp[st].fp_acc;
1170 			/* The first 64 bits contain the mantissa. */
1171 			mantissa = *((uint64_t *)fx_reg->fp_bytes);
1172 			/*
1173 			 * The final 16 bits contain the sign bit and the exponent.
1174 			 * Mask the sign bit since it is of no consequence to these
1175 			 * tests.
1176 			 */
1177 			exp = *((uint16_t *)&fx_reg->fp_bytes[8]) & 0x7fff;
1178 			if (exp == 0) {
1179 				if (mantissa == 0)
1180 					tw |= 1; /* Zero */
1181 				else
1182 					tw |= 2; /* Denormal */
1183 			} else if (exp == 0x7fff)
1184 				tw |= 2; /* Infinity or NaN */
1185 		} else
1186 			tw |= 3; /* Empty */
1187 		st = (st - 1) & 7;
1188 	}
1189 	penv_87->en_tw = tw;
1190 }
1191 
1192 void
1193 npx_fill_fpregs_xmm(struct savexmm *sv_xmm, struct save87 *sv_87)
1194 {
1195 
1196 	bzero(sv_87, sizeof(*sv_87));
1197 	npx_fill_fpregs_xmm1(sv_xmm, sv_87);
1198 }
1199 
1200 void
1201 npx_set_fpregs_xmm(struct save87 *sv_87, struct savexmm *sv_xmm)
1202 {
1203 	struct env87 *penv_87;
1204 	struct envxmm *penv_xmm;
1205 	int i;
1206 
1207 	penv_87 = &sv_87->sv_env;
1208 	penv_xmm = &sv_xmm->sv_env;
1209 
1210 	/* FPU control/status */
1211 	penv_xmm->en_cw = penv_87->en_cw;
1212 	penv_xmm->en_sw = penv_87->en_sw;
1213 	penv_xmm->en_fip = penv_87->en_fip;
1214 	penv_xmm->en_fcs = penv_87->en_fcs;
1215 	penv_xmm->en_opcode = penv_87->en_opcode;
1216 	penv_xmm->en_foo = penv_87->en_foo;
1217 	penv_xmm->en_fos = penv_87->en_fos;
1218 
1219 	/*
1220 	 * FPU registers and tags.
1221 	 * Abridged  /  Full translation (values in binary), see FXSAVE spec.
1222 	 * 0		11
1223 	 * 1		00, 01, 10
1224 	 */
1225 	penv_xmm->en_tw = 0;
1226 	for (i = 0; i < 8; ++i) {
1227 		sv_xmm->sv_fp[i].fp_acc = sv_87->sv_ac[i];
1228 		if ((penv_87->en_tw & (3 << i * 2)) != (3 << i * 2))
1229 			penv_xmm->en_tw |= 1 << i;
1230 	}
1231 }
1232 
1233 void
1234 npx_get_fsave(void *addr)
1235 {
1236 	struct thread *td;
1237 	union savefpu *sv;
1238 
1239 	td = curthread;
1240 	npxgetregs(td);
1241 	sv = get_pcb_user_save_td(td);
1242 	if (cpu_fxsr)
1243 		npx_fill_fpregs_xmm1(&sv->sv_xmm, addr);
1244 	else
1245 		bcopy(sv, addr, sizeof(struct env87) +
1246 		    sizeof(struct fpacc87[8]));
1247 }
1248 
1249 int
1250 npx_set_fsave(void *addr)
1251 {
1252 	union savefpu sv;
1253 	int error;
1254 
1255 	bzero(&sv, sizeof(sv));
1256 	if (cpu_fxsr)
1257 		npx_set_fpregs_xmm(addr, &sv.sv_xmm);
1258 	else
1259 		bcopy(addr, &sv, sizeof(struct env87) +
1260 		    sizeof(struct fpacc87[8]));
1261 	error = npxsetregs(curthread, &sv, NULL, 0);
1262 	return (error);
1263 }
1264 
1265 /*
1266  * On AuthenticAMD processors, the fxrstor instruction does not restore
1267  * the x87's stored last instruction pointer, last data pointer, and last
1268  * opcode values, except in the rare case in which the exception summary
1269  * (ES) bit in the x87 status word is set to 1.
1270  *
1271  * In order to avoid leaking this information across processes, we clean
1272  * these values by performing a dummy load before executing fxrstor().
1273  */
1274 static void
1275 fpu_clean_state(void)
1276 {
1277 	static float dummy_variable = 0.0;
1278 	u_short status;
1279 
1280 	/*
1281 	 * Clear the ES bit in the x87 status word if it is currently
1282 	 * set, in order to avoid causing a fault in the upcoming load.
1283 	 */
1284 	fnstsw(&status);
1285 	if (status & 0x80)
1286 		fnclex();
1287 
1288 	/*
1289 	 * Load the dummy variable into the x87 stack.  This mangles
1290 	 * the x87 stack, but we don't care since we're about to call
1291 	 * fxrstor() anyway.
1292 	 */
1293 	__asm __volatile("ffree %%st(7); flds %0" : : "m" (dummy_variable));
1294 }
1295 
1296 static void
1297 fpurstor(union savefpu *addr)
1298 {
1299 
1300 	if (use_xsave)
1301 		xrstor((char *)addr, xsave_mask);
1302 	else if (cpu_fxsr)
1303 		fxrstor(addr);
1304 	else
1305 		frstor(addr);
1306 }
1307 
1308 #ifdef DEV_ISA
1309 /*
1310  * This sucks up the legacy ISA support assignments from PNPBIOS/ACPI.
1311  */
1312 static struct isa_pnp_id npxisa_ids[] = {
1313 	{ 0x040cd041, "Legacy ISA coprocessor support" }, /* PNP0C04 */
1314 	{ 0 }
1315 };
1316 
1317 static int
1318 npxisa_probe(device_t dev)
1319 {
1320 	int result;
1321 	if ((result = ISA_PNP_PROBE(device_get_parent(dev), dev, npxisa_ids)) <= 0) {
1322 		device_quiet(dev);
1323 	}
1324 	return(result);
1325 }
1326 
1327 static int
1328 npxisa_attach(device_t dev)
1329 {
1330 	return (0);
1331 }
1332 
1333 static device_method_t npxisa_methods[] = {
1334 	/* Device interface */
1335 	DEVMETHOD(device_probe,		npxisa_probe),
1336 	DEVMETHOD(device_attach,	npxisa_attach),
1337 	DEVMETHOD(device_detach,	bus_generic_detach),
1338 	DEVMETHOD(device_shutdown,	bus_generic_shutdown),
1339 	DEVMETHOD(device_suspend,	bus_generic_suspend),
1340 	DEVMETHOD(device_resume,	bus_generic_resume),
1341 	{ 0, 0 }
1342 };
1343 
1344 static driver_t npxisa_driver = {
1345 	"npxisa",
1346 	npxisa_methods,
1347 	1,			/* no softc */
1348 };
1349 
1350 DRIVER_MODULE(npxisa, isa, npxisa_driver, 0, 0);
1351 DRIVER_MODULE(npxisa, acpi, npxisa_driver, 0, 0);
1352 ISA_PNP_INFO(npxisa_ids);
1353 #endif /* DEV_ISA */
1354 
1355 static MALLOC_DEFINE(M_FPUKERN_CTX, "fpukern_ctx",
1356     "Kernel contexts for FPU state");
1357 
1358 #define	FPU_KERN_CTX_NPXINITDONE 0x01
1359 #define	FPU_KERN_CTX_DUMMY	 0x02
1360 #define	FPU_KERN_CTX_INUSE	 0x04
1361 
1362 struct fpu_kern_ctx {
1363 	union savefpu *prev;
1364 	uint32_t flags;
1365 	char hwstate1[];
1366 };
1367 
1368 struct fpu_kern_ctx *
1369 fpu_kern_alloc_ctx(u_int flags)
1370 {
1371 	struct fpu_kern_ctx *res;
1372 	size_t sz;
1373 
1374 	sz = sizeof(struct fpu_kern_ctx) + XSAVE_AREA_ALIGN +
1375 	    cpu_max_ext_state_size;
1376 	res = malloc(sz, M_FPUKERN_CTX, ((flags & FPU_KERN_NOWAIT) ?
1377 	    M_NOWAIT : M_WAITOK) | M_ZERO);
1378 	return (res);
1379 }
1380 
1381 void
1382 fpu_kern_free_ctx(struct fpu_kern_ctx *ctx)
1383 {
1384 
1385 	KASSERT((ctx->flags & FPU_KERN_CTX_INUSE) == 0, ("free'ing inuse ctx"));
1386 	/* XXXKIB clear the memory ? */
1387 	free(ctx, M_FPUKERN_CTX);
1388 }
1389 
1390 static union savefpu *
1391 fpu_kern_ctx_savefpu(struct fpu_kern_ctx *ctx)
1392 {
1393 	vm_offset_t p;
1394 
1395 	p = (vm_offset_t)&ctx->hwstate1;
1396 	p = roundup2(p, XSAVE_AREA_ALIGN);
1397 	return ((union savefpu *)p);
1398 }
1399 
1400 void
1401 fpu_kern_enter(struct thread *td, struct fpu_kern_ctx *ctx, u_int flags)
1402 {
1403 	struct pcb *pcb;
1404 
1405 	pcb = td->td_pcb;
1406 	KASSERT((flags & FPU_KERN_NOCTX) != 0 || ctx != NULL,
1407 	    ("ctx is required when !FPU_KERN_NOCTX"));
1408 	KASSERT(ctx == NULL || (ctx->flags & FPU_KERN_CTX_INUSE) == 0,
1409 	    ("using inuse ctx"));
1410 	KASSERT((pcb->pcb_flags & PCB_NPXNOSAVE) == 0,
1411 	    ("recursive fpu_kern_enter while in PCB_NPXNOSAVE state"));
1412 
1413 	if ((flags & FPU_KERN_NOCTX) != 0) {
1414 		critical_enter();
1415 		stop_emulating();
1416 		if (curthread == PCPU_GET(fpcurthread)) {
1417 			fpusave(curpcb->pcb_save);
1418 			PCPU_SET(fpcurthread, NULL);
1419 		} else {
1420 			KASSERT(PCPU_GET(fpcurthread) == NULL,
1421 			    ("invalid fpcurthread"));
1422 		}
1423 
1424 		/*
1425 		 * This breaks XSAVEOPT tracker, but
1426 		 * PCB_NPXNOSAVE state is supposed to never need to
1427 		 * save FPU context at all.
1428 		 */
1429 		fpurstor(npx_initialstate);
1430 		pcb->pcb_flags |= PCB_KERNNPX | PCB_NPXNOSAVE | PCB_NPXINITDONE;
1431 		return;
1432 	}
1433 	if ((flags & FPU_KERN_KTHR) != 0 && is_fpu_kern_thread(0)) {
1434 		ctx->flags = FPU_KERN_CTX_DUMMY | FPU_KERN_CTX_INUSE;
1435 		return;
1436 	}
1437 	pcb = td->td_pcb;
1438 	critical_enter();
1439 	KASSERT(!PCB_USER_FPU(pcb) || pcb->pcb_save ==
1440 	    get_pcb_user_save_pcb(pcb), ("mangled pcb_save"));
1441 	ctx->flags = FPU_KERN_CTX_INUSE;
1442 	if ((pcb->pcb_flags & PCB_NPXINITDONE) != 0)
1443 		ctx->flags |= FPU_KERN_CTX_NPXINITDONE;
1444 	npxexit(td);
1445 	ctx->prev = pcb->pcb_save;
1446 	pcb->pcb_save = fpu_kern_ctx_savefpu(ctx);
1447 	pcb->pcb_flags |= PCB_KERNNPX;
1448 	pcb->pcb_flags &= ~PCB_NPXINITDONE;
1449 	critical_exit();
1450 }
1451 
1452 int
1453 fpu_kern_leave(struct thread *td, struct fpu_kern_ctx *ctx)
1454 {
1455 	struct pcb *pcb;
1456 
1457 	pcb = td->td_pcb;
1458 
1459 	if ((pcb->pcb_flags & PCB_NPXNOSAVE) != 0) {
1460 		KASSERT(ctx == NULL, ("non-null ctx after FPU_KERN_NOCTX"));
1461 		KASSERT(PCPU_GET(fpcurthread) == NULL,
1462 		    ("non-NULL fpcurthread for PCB_NPXNOSAVE"));
1463 		CRITICAL_ASSERT(td);
1464 
1465 		pcb->pcb_flags &= ~(PCB_NPXNOSAVE | PCB_NPXINITDONE);
1466 		start_emulating();
1467 	} else {
1468 		KASSERT((ctx->flags & FPU_KERN_CTX_INUSE) != 0,
1469 		    ("leaving not inuse ctx"));
1470 		ctx->flags &= ~FPU_KERN_CTX_INUSE;
1471 
1472 		if (is_fpu_kern_thread(0) &&
1473 		    (ctx->flags & FPU_KERN_CTX_DUMMY) != 0)
1474 			return (0);
1475 		KASSERT((ctx->flags & FPU_KERN_CTX_DUMMY) == 0,
1476 		    ("dummy ctx"));
1477 		critical_enter();
1478 		if (curthread == PCPU_GET(fpcurthread))
1479 			npxdrop();
1480 		pcb->pcb_save = ctx->prev;
1481 	}
1482 
1483 	if (pcb->pcb_save == get_pcb_user_save_pcb(pcb)) {
1484 		if ((pcb->pcb_flags & PCB_NPXUSERINITDONE) != 0) {
1485 			pcb->pcb_flags |= PCB_NPXINITDONE;
1486 			if ((pcb->pcb_flags & PCB_KERNNPX_THR) == 0)
1487 				pcb->pcb_flags &= ~PCB_KERNNPX;
1488 		} else if ((pcb->pcb_flags & PCB_KERNNPX_THR) == 0)
1489 			pcb->pcb_flags &= ~(PCB_NPXINITDONE | PCB_KERNNPX);
1490 	} else {
1491 		if ((ctx->flags & FPU_KERN_CTX_NPXINITDONE) != 0)
1492 			pcb->pcb_flags |= PCB_NPXINITDONE;
1493 		else
1494 			pcb->pcb_flags &= ~PCB_NPXINITDONE;
1495 		KASSERT(!PCB_USER_FPU(pcb), ("unpaired fpu_kern_leave"));
1496 	}
1497 	critical_exit();
1498 	return (0);
1499 }
1500 
1501 int
1502 fpu_kern_thread(u_int flags)
1503 {
1504 
1505 	KASSERT((curthread->td_pflags & TDP_KTHREAD) != 0,
1506 	    ("Only kthread may use fpu_kern_thread"));
1507 	KASSERT(curpcb->pcb_save == get_pcb_user_save_pcb(curpcb),
1508 	    ("mangled pcb_save"));
1509 	KASSERT(PCB_USER_FPU(curpcb), ("recursive call"));
1510 
1511 	curpcb->pcb_flags |= PCB_KERNNPX | PCB_KERNNPX_THR;
1512 	return (0);
1513 }
1514 
1515 int
1516 is_fpu_kern_thread(u_int flags)
1517 {
1518 
1519 	if ((curthread->td_pflags & TDP_KTHREAD) == 0)
1520 		return (0);
1521 	return ((curpcb->pcb_flags & PCB_KERNNPX_THR) != 0);
1522 }
1523 
1524 /*
1525  * FPU save area alloc/free/init utility routines
1526  */
1527 union savefpu *
1528 fpu_save_area_alloc(void)
1529 {
1530 
1531 	return (uma_zalloc(fpu_save_area_zone, M_WAITOK));
1532 }
1533 
1534 void
1535 fpu_save_area_free(union savefpu *fsa)
1536 {
1537 
1538 	uma_zfree(fpu_save_area_zone, fsa);
1539 }
1540 
1541 void
1542 fpu_save_area_reset(union savefpu *fsa)
1543 {
1544 
1545 	bcopy(npx_initialstate, fsa, cpu_max_ext_state_size);
1546 }
1547