npx.c revision 302408
1/*-
2 * Copyright (c) 1990 William Jolitz.
3 * Copyright (c) 1991 The Regents of the University of California.
4 * All rights reserved.
5 *
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
8 * are met:
9 * 1. Redistributions of source code must retain the above copyright
10 *    notice, this list of conditions and the following disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 *    notice, this list of conditions and the following disclaimer in the
13 *    documentation and/or other materials provided with the distribution.
14 * 4. Neither the name of the University nor the names of its contributors
15 *    may be used to endorse or promote products derived from this software
16 *    without specific prior written permission.
17 *
18 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
19 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21 * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
22 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28 * SUCH DAMAGE.
29 *
30 *	from: @(#)npx.c	7.2 (Berkeley) 5/12/91
31 */
32
33#include <sys/cdefs.h>
34__FBSDID("$FreeBSD: stable/11/sys/i386/isa/npx.c 294313 2016-01-19 08:09:09Z kib $");
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
71#include <machine/intr_machdep.h>
72
73#ifdef DEV_ISA
74#include <isa/isavar.h>
75#endif
76
77#if !defined(CPU_DISABLE_SSE) && defined(I686_CPU)
78#define CPU_ENABLE_SSE
79#endif
80
81/*
82 * 387 and 287 Numeric Coprocessor Extension (NPX) Driver.
83 */
84
85#if defined(__GNUCLIKE_ASM) && !defined(lint)
86
87#define	fldcw(cw)		__asm __volatile("fldcw %0" : : "m" (cw))
88#define	fnclex()		__asm __volatile("fnclex")
89#define	fninit()		__asm __volatile("fninit")
90#define	fnsave(addr)		__asm __volatile("fnsave %0" : "=m" (*(addr)))
91#define	fnstcw(addr)		__asm __volatile("fnstcw %0" : "=m" (*(addr)))
92#define	fnstsw(addr)		__asm __volatile("fnstsw %0" : "=am" (*(addr)))
93#define	fp_divide_by_0()	__asm __volatile( \
94				    "fldz; fld1; fdiv %st,%st(1); fnop")
95#define	frstor(addr)		__asm __volatile("frstor %0" : : "m" (*(addr)))
96#ifdef CPU_ENABLE_SSE
97#define	fxrstor(addr)		__asm __volatile("fxrstor %0" : : "m" (*(addr)))
98#define	fxsave(addr)		__asm __volatile("fxsave %0" : "=m" (*(addr)))
99#define	ldmxcsr(csr)		__asm __volatile("ldmxcsr %0" : : "m" (csr))
100#define	stmxcsr(addr)		__asm __volatile("stmxcsr %0" : : "m" (*(addr)))
101
102static __inline void
103xrstor(char *addr, uint64_t mask)
104{
105	uint32_t low, hi;
106
107	low = mask;
108	hi = mask >> 32;
109	__asm __volatile("xrstor %0" : : "m" (*addr), "a" (low), "d" (hi));
110}
111
112static __inline void
113xsave(char *addr, uint64_t mask)
114{
115	uint32_t low, hi;
116
117	low = mask;
118	hi = mask >> 32;
119	__asm __volatile("xsave %0" : "=m" (*addr) : "a" (low), "d" (hi) :
120	    "memory");
121}
122
123static __inline void
124xsaveopt(char *addr, uint64_t mask)
125{
126	uint32_t low, hi;
127
128	low = mask;
129	hi = mask >> 32;
130	__asm __volatile("xsaveopt %0" : "=m" (*addr) : "a" (low), "d" (hi) :
131	    "memory");
132}
133#endif
134#else	/* !(__GNUCLIKE_ASM && !lint) */
135
136void	fldcw(u_short cw);
137void	fnclex(void);
138void	fninit(void);
139void	fnsave(caddr_t addr);
140void	fnstcw(caddr_t addr);
141void	fnstsw(caddr_t addr);
142void	fp_divide_by_0(void);
143void	frstor(caddr_t addr);
144#ifdef CPU_ENABLE_SSE
145void	fxsave(caddr_t addr);
146void	fxrstor(caddr_t addr);
147void	ldmxcsr(u_int csr);
148void	stmxcsr(u_int *csr);
149void	xrstor(char *addr, uint64_t mask);
150void	xsave(char *addr, uint64_t mask);
151void	xsaveopt(char *addr, uint64_t mask);
152#endif
153
154#endif	/* __GNUCLIKE_ASM && !lint */
155
156#define	start_emulating()	load_cr0(rcr0() | CR0_TS)
157#define	stop_emulating()	clts()
158
159#ifdef CPU_ENABLE_SSE
160#define GET_FPU_CW(thread) \
161	(cpu_fxsr ? \
162		(thread)->td_pcb->pcb_save->sv_xmm.sv_env.en_cw : \
163		(thread)->td_pcb->pcb_save->sv_87.sv_env.en_cw)
164#define GET_FPU_SW(thread) \
165	(cpu_fxsr ? \
166		(thread)->td_pcb->pcb_save->sv_xmm.sv_env.en_sw : \
167		(thread)->td_pcb->pcb_save->sv_87.sv_env.en_sw)
168#define SET_FPU_CW(savefpu, value) do { \
169	if (cpu_fxsr) \
170		(savefpu)->sv_xmm.sv_env.en_cw = (value); \
171	else \
172		(savefpu)->sv_87.sv_env.en_cw = (value); \
173} while (0)
174#else /* CPU_ENABLE_SSE */
175#define GET_FPU_CW(thread) \
176	(thread->td_pcb->pcb_save->sv_87.sv_env.en_cw)
177#define GET_FPU_SW(thread) \
178	(thread->td_pcb->pcb_save->sv_87.sv_env.en_sw)
179#define SET_FPU_CW(savefpu, value) \
180	(savefpu)->sv_87.sv_env.en_cw = (value)
181#endif /* CPU_ENABLE_SSE */
182
183#ifdef CPU_ENABLE_SSE
184CTASSERT(sizeof(union savefpu) == 512);
185CTASSERT(sizeof(struct xstate_hdr) == 64);
186CTASSERT(sizeof(struct savefpu_ymm) == 832);
187
188/*
189 * This requirement is to make it easier for asm code to calculate
190 * offset of the fpu save area from the pcb address. FPU save area
191 * must be 64-byte aligned.
192 */
193CTASSERT(sizeof(struct pcb) % XSAVE_AREA_ALIGN == 0);
194
195/*
196 * Ensure the copy of XCR0 saved in a core is contained in the padding
197 * area.
198 */
199CTASSERT(X86_XSTATE_XCR0_OFFSET >= offsetof(struct savexmm, sv_pad) &&
200    X86_XSTATE_XCR0_OFFSET + sizeof(uint64_t) <= sizeof(struct savexmm));
201
202static	void	fpu_clean_state(void);
203#endif
204
205static	void	fpusave(union savefpu *);
206static	void	fpurstor(union savefpu *);
207
208int	hw_float;
209
210SYSCTL_INT(_hw, HW_FLOATINGPT, floatingpoint, CTLFLAG_RD,
211    &hw_float, 0, "Floating point instructions executed in hardware");
212
213#ifdef CPU_ENABLE_SSE
214int use_xsave;
215uint64_t xsave_mask;
216#endif
217static	uma_zone_t fpu_save_area_zone;
218static	union savefpu *npx_initialstate;
219
220#ifdef CPU_ENABLE_SSE
221struct xsave_area_elm_descr {
222	u_int	offset;
223	u_int	size;
224} *xsave_area_desc;
225
226static int use_xsaveopt;
227#endif
228
229static	volatile u_int		npx_traps_while_probing;
230
231alias_for_inthand_t probetrap;
232__asm("								\n\
233	.text							\n\
234	.p2align 2,0x90						\n\
235	.type	" __XSTRING(CNAME(probetrap)) ",@function	\n\
236" __XSTRING(CNAME(probetrap)) ":				\n\
237	ss							\n\
238	incl	" __XSTRING(CNAME(npx_traps_while_probing)) "	\n\
239	fnclex							\n\
240	iret							\n\
241");
242
243/*
244 * Determine if an FPU is present and how to use it.
245 */
246static int
247npx_probe(void)
248{
249	struct gate_descriptor save_idt_npxtrap;
250	u_short control, status;
251
252	/*
253	 * Modern CPUs all have an FPU that uses the INT16 interface
254	 * and provide a simple way to verify that, so handle the
255	 * common case right away.
256	 */
257	if (cpu_feature & CPUID_FPU) {
258		hw_float = 1;
259		return (1);
260	}
261
262	save_idt_npxtrap = idt[IDT_MF];
263	setidt(IDT_MF, probetrap, SDT_SYS386TGT, SEL_KPL,
264	    GSEL(GCODE_SEL, SEL_KPL));
265
266	/*
267	 * Don't trap while we're probing.
268	 */
269	stop_emulating();
270
271	/*
272	 * Finish resetting the coprocessor, if any.  If there is an error
273	 * pending, then we may get a bogus IRQ13, but npx_intr() will handle
274	 * it OK.  Bogus halts have never been observed, but we enabled
275	 * IRQ13 and cleared the BUSY# latch early to handle them anyway.
276	 */
277	fninit();
278
279	/*
280	 * Don't use fwait here because it might hang.
281	 * Don't use fnop here because it usually hangs if there is no FPU.
282	 */
283	DELAY(1000);		/* wait for any IRQ13 */
284#ifdef DIAGNOSTIC
285	if (npx_traps_while_probing != 0)
286		printf("fninit caused %u bogus npx trap(s)\n",
287		       npx_traps_while_probing);
288#endif
289	/*
290	 * Check for a status of mostly zero.
291	 */
292	status = 0x5a5a;
293	fnstsw(&status);
294	if ((status & 0xb8ff) == 0) {
295		/*
296		 * Good, now check for a proper control word.
297		 */
298		control = 0x5a5a;
299		fnstcw(&control);
300		if ((control & 0x1f3f) == 0x033f) {
301			/*
302			 * We have an npx, now divide by 0 to see if exception
303			 * 16 works.
304			 */
305			control &= ~(1 << 2);	/* enable divide by 0 trap */
306			fldcw(control);
307#ifdef FPU_ERROR_BROKEN
308			/*
309			 * FPU error signal doesn't work on some CPU
310			 * accelerator board.
311			 */
312			hw_float = 1;
313			return (1);
314#endif
315			npx_traps_while_probing = 0;
316			fp_divide_by_0();
317			if (npx_traps_while_probing != 0) {
318				/*
319				 * Good, exception 16 works.
320				 */
321				hw_float = 1;
322				goto cleanup;
323			}
324			printf(
325	"FPU does not use exception 16 for error reporting\n");
326			goto cleanup;
327		}
328	}
329
330	/*
331	 * Probe failed.  Floating point simply won't work.
332	 * Notify user and disable FPU/MMX/SSE instruction execution.
333	 */
334	printf("WARNING: no FPU!\n");
335	__asm __volatile("smsw %%ax; orb %0,%%al; lmsw %%ax" : :
336	    "n" (CR0_EM | CR0_MP) : "ax");
337
338cleanup:
339	idt[IDT_MF] = save_idt_npxtrap;
340	return (hw_float);
341}
342
343#ifdef CPU_ENABLE_SSE
344/*
345 * Enable XSAVE if supported and allowed by user.
346 * Calculate the xsave_mask.
347 */
348static void
349npxinit_bsp1(void)
350{
351	u_int cp[4];
352	uint64_t xsave_mask_user;
353
354	if (cpu_fxsr && (cpu_feature2 & CPUID2_XSAVE) != 0) {
355		use_xsave = 1;
356		TUNABLE_INT_FETCH("hw.use_xsave", &use_xsave);
357	}
358	if (!use_xsave)
359		return;
360
361	cpuid_count(0xd, 0x0, cp);
362	xsave_mask = XFEATURE_ENABLED_X87 | XFEATURE_ENABLED_SSE;
363	if ((cp[0] & xsave_mask) != xsave_mask)
364		panic("CPU0 does not support X87 or SSE: %x", cp[0]);
365	xsave_mask = ((uint64_t)cp[3] << 32) | cp[0];
366	xsave_mask_user = xsave_mask;
367	TUNABLE_QUAD_FETCH("hw.xsave_mask", &xsave_mask_user);
368	xsave_mask_user |= XFEATURE_ENABLED_X87 | XFEATURE_ENABLED_SSE;
369	xsave_mask &= xsave_mask_user;
370	if ((xsave_mask & XFEATURE_AVX512) != XFEATURE_AVX512)
371		xsave_mask &= ~XFEATURE_AVX512;
372	if ((xsave_mask & XFEATURE_MPX) != XFEATURE_MPX)
373		xsave_mask &= ~XFEATURE_MPX;
374
375	cpuid_count(0xd, 0x1, cp);
376	if ((cp[0] & CPUID_EXTSTATE_XSAVEOPT) != 0)
377		use_xsaveopt = 1;
378}
379#endif
380/*
381
382 * Calculate the fpu save area size.
383 */
384static void
385npxinit_bsp2(void)
386{
387#ifdef CPU_ENABLE_SSE
388	u_int cp[4];
389
390	if (use_xsave) {
391		cpuid_count(0xd, 0x0, cp);
392		cpu_max_ext_state_size = cp[1];
393
394		/*
395		 * Reload the cpu_feature2, since we enabled OSXSAVE.
396		 */
397		do_cpuid(1, cp);
398		cpu_feature2 = cp[2];
399	} else
400#endif
401		cpu_max_ext_state_size = sizeof(union savefpu);
402}
403
404/*
405 * Initialize floating point unit.
406 */
407void
408npxinit(bool bsp)
409{
410	static union savefpu dummy;
411	register_t saveintr;
412#ifdef CPU_ENABLE_SSE
413	u_int mxcsr;
414#endif
415	u_short control;
416
417	if (bsp) {
418		if (!npx_probe())
419			return;
420#ifdef CPU_ENABLE_SSE
421		npxinit_bsp1();
422#endif
423	}
424
425#ifdef CPU_ENABLE_SSE
426	if (use_xsave) {
427		load_cr4(rcr4() | CR4_XSAVE);
428		load_xcr(XCR0, xsave_mask);
429	}
430#endif
431
432	/*
433	 * XCR0 shall be set up before CPU can report the save area size.
434	 */
435	if (bsp)
436		npxinit_bsp2();
437
438	/*
439	 * fninit has the same h/w bugs as fnsave.  Use the detoxified
440	 * fnsave to throw away any junk in the fpu.  fpusave() initializes
441	 * the fpu.
442	 *
443	 * It is too early for critical_enter() to work on AP.
444	 */
445	saveintr = intr_disable();
446	stop_emulating();
447#ifdef CPU_ENABLE_SSE
448	if (cpu_fxsr)
449		fninit();
450	else
451#endif
452		fnsave(&dummy);
453	control = __INITIAL_NPXCW__;
454	fldcw(control);
455#ifdef CPU_ENABLE_SSE
456	if (cpu_fxsr) {
457		mxcsr = __INITIAL_MXCSR__;
458		ldmxcsr(mxcsr);
459	}
460#endif
461	start_emulating();
462	intr_restore(saveintr);
463}
464
465/*
466 * On the boot CPU we generate a clean state that is used to
467 * initialize the floating point unit when it is first used by a
468 * process.
469 */
470static void
471npxinitstate(void *arg __unused)
472{
473	register_t saveintr;
474#ifdef CPU_ENABLE_SSE
475	int cp[4], i, max_ext_n;
476#endif
477
478	if (!hw_float)
479		return;
480
481	npx_initialstate = malloc(cpu_max_ext_state_size, M_DEVBUF,
482	    M_WAITOK | M_ZERO);
483	saveintr = intr_disable();
484	stop_emulating();
485
486	fpusave(npx_initialstate);
487#ifdef CPU_ENABLE_SSE
488	if (cpu_fxsr) {
489		if (npx_initialstate->sv_xmm.sv_env.en_mxcsr_mask)
490			cpu_mxcsr_mask =
491			    npx_initialstate->sv_xmm.sv_env.en_mxcsr_mask;
492		else
493			cpu_mxcsr_mask = 0xFFBF;
494
495		/*
496		 * The fninit instruction does not modify XMM
497		 * registers or x87 registers (MM/ST).  The fpusave
498		 * call dumped the garbage contained in the registers
499		 * after reset to the initial state saved.  Clear XMM
500		 * and x87 registers file image to make the startup
501		 * program state and signal handler XMM/x87 register
502		 * content predictable.
503		 */
504		bzero(npx_initialstate->sv_xmm.sv_fp,
505		    sizeof(npx_initialstate->sv_xmm.sv_fp));
506		bzero(npx_initialstate->sv_xmm.sv_xmm,
507		    sizeof(npx_initialstate->sv_xmm.sv_xmm));
508	} else
509#endif
510		bzero(npx_initialstate->sv_87.sv_ac,
511		    sizeof(npx_initialstate->sv_87.sv_ac));
512
513#ifdef CPU_ENABLE_SSE
514	/*
515	 * Create a table describing the layout of the CPU Extended
516	 * Save Area.
517	 */
518	if (use_xsave) {
519		if (xsave_mask >> 32 != 0)
520			max_ext_n = fls(xsave_mask >> 32) + 32;
521		else
522			max_ext_n = fls(xsave_mask);
523		xsave_area_desc = malloc(max_ext_n * sizeof(struct
524		    xsave_area_elm_descr), M_DEVBUF, M_WAITOK | M_ZERO);
525		/* x87 state */
526		xsave_area_desc[0].offset = 0;
527		xsave_area_desc[0].size = 160;
528		/* XMM */
529		xsave_area_desc[1].offset = 160;
530		xsave_area_desc[1].size = 288 - 160;
531
532		for (i = 2; i < max_ext_n; i++) {
533			cpuid_count(0xd, i, cp);
534			xsave_area_desc[i].offset = cp[1];
535			xsave_area_desc[i].size = cp[0];
536		}
537	}
538#endif
539
540	fpu_save_area_zone = uma_zcreate("FPU_save_area",
541	    cpu_max_ext_state_size, NULL, NULL, NULL, NULL,
542	    XSAVE_AREA_ALIGN - 1, 0);
543
544	start_emulating();
545	intr_restore(saveintr);
546}
547SYSINIT(npxinitstate, SI_SUB_DRIVERS, SI_ORDER_ANY, npxinitstate, NULL);
548
549/*
550 * Free coprocessor (if we have it).
551 */
552void
553npxexit(td)
554	struct thread *td;
555{
556
557	critical_enter();
558	if (curthread == PCPU_GET(fpcurthread)) {
559		stop_emulating();
560		fpusave(curpcb->pcb_save);
561		start_emulating();
562		PCPU_SET(fpcurthread, NULL);
563	}
564	critical_exit();
565#ifdef NPX_DEBUG
566	if (hw_float) {
567		u_int	masked_exceptions;
568
569		masked_exceptions = GET_FPU_CW(td) & GET_FPU_SW(td) & 0x7f;
570		/*
571		 * Log exceptions that would have trapped with the old
572		 * control word (overflow, divide by 0, and invalid operand).
573		 */
574		if (masked_exceptions & 0x0d)
575			log(LOG_ERR,
576	"pid %d (%s) exited with masked floating point exceptions 0x%02x\n",
577			    td->td_proc->p_pid, td->td_proc->p_comm,
578			    masked_exceptions);
579	}
580#endif
581}
582
583int
584npxformat()
585{
586
587	if (!hw_float)
588		return (_MC_FPFMT_NODEV);
589#ifdef	CPU_ENABLE_SSE
590	if (cpu_fxsr)
591		return (_MC_FPFMT_XMM);
592#endif
593	return (_MC_FPFMT_387);
594}
595
596/*
597 * The following mechanism is used to ensure that the FPE_... value
598 * that is passed as a trapcode to the signal handler of the user
599 * process does not have more than one bit set.
600 *
601 * Multiple bits may be set if the user process modifies the control
602 * word while a status word bit is already set.  While this is a sign
603 * of bad coding, we have no choise than to narrow them down to one
604 * bit, since we must not send a trapcode that is not exactly one of
605 * the FPE_ macros.
606 *
607 * The mechanism has a static table with 127 entries.  Each combination
608 * of the 7 FPU status word exception bits directly translates to a
609 * position in this table, where a single FPE_... value is stored.
610 * This FPE_... value stored there is considered the "most important"
611 * of the exception bits and will be sent as the signal code.  The
612 * precedence of the bits is based upon Intel Document "Numerical
613 * Applications", Chapter "Special Computational Situations".
614 *
615 * The macro to choose one of these values does these steps: 1) Throw
616 * away status word bits that cannot be masked.  2) Throw away the bits
617 * currently masked in the control word, assuming the user isn't
618 * interested in them anymore.  3) Reinsert status word bit 7 (stack
619 * fault) if it is set, which cannot be masked but must be presered.
620 * 4) Use the remaining bits to point into the trapcode table.
621 *
622 * The 6 maskable bits in order of their preference, as stated in the
623 * above referenced Intel manual:
624 * 1  Invalid operation (FP_X_INV)
625 * 1a   Stack underflow
626 * 1b   Stack overflow
627 * 1c   Operand of unsupported format
628 * 1d   SNaN operand.
629 * 2  QNaN operand (not an exception, irrelavant here)
630 * 3  Any other invalid-operation not mentioned above or zero divide
631 *      (FP_X_INV, FP_X_DZ)
632 * 4  Denormal operand (FP_X_DNML)
633 * 5  Numeric over/underflow (FP_X_OFL, FP_X_UFL)
634 * 6  Inexact result (FP_X_IMP)
635 */
636static char fpetable[128] = {
637	0,
638	FPE_FLTINV,	/*  1 - INV */
639	FPE_FLTUND,	/*  2 - DNML */
640	FPE_FLTINV,	/*  3 - INV | DNML */
641	FPE_FLTDIV,	/*  4 - DZ */
642	FPE_FLTINV,	/*  5 - INV | DZ */
643	FPE_FLTDIV,	/*  6 - DNML | DZ */
644	FPE_FLTINV,	/*  7 - INV | DNML | DZ */
645	FPE_FLTOVF,	/*  8 - OFL */
646	FPE_FLTINV,	/*  9 - INV | OFL */
647	FPE_FLTUND,	/*  A - DNML | OFL */
648	FPE_FLTINV,	/*  B - INV | DNML | OFL */
649	FPE_FLTDIV,	/*  C - DZ | OFL */
650	FPE_FLTINV,	/*  D - INV | DZ | OFL */
651	FPE_FLTDIV,	/*  E - DNML | DZ | OFL */
652	FPE_FLTINV,	/*  F - INV | DNML | DZ | OFL */
653	FPE_FLTUND,	/* 10 - UFL */
654	FPE_FLTINV,	/* 11 - INV | UFL */
655	FPE_FLTUND,	/* 12 - DNML | UFL */
656	FPE_FLTINV,	/* 13 - INV | DNML | UFL */
657	FPE_FLTDIV,	/* 14 - DZ | UFL */
658	FPE_FLTINV,	/* 15 - INV | DZ | UFL */
659	FPE_FLTDIV,	/* 16 - DNML | DZ | UFL */
660	FPE_FLTINV,	/* 17 - INV | DNML | DZ | UFL */
661	FPE_FLTOVF,	/* 18 - OFL | UFL */
662	FPE_FLTINV,	/* 19 - INV | OFL | UFL */
663	FPE_FLTUND,	/* 1A - DNML | OFL | UFL */
664	FPE_FLTINV,	/* 1B - INV | DNML | OFL | UFL */
665	FPE_FLTDIV,	/* 1C - DZ | OFL | UFL */
666	FPE_FLTINV,	/* 1D - INV | DZ | OFL | UFL */
667	FPE_FLTDIV,	/* 1E - DNML | DZ | OFL | UFL */
668	FPE_FLTINV,	/* 1F - INV | DNML | DZ | OFL | UFL */
669	FPE_FLTRES,	/* 20 - IMP */
670	FPE_FLTINV,	/* 21 - INV | IMP */
671	FPE_FLTUND,	/* 22 - DNML | IMP */
672	FPE_FLTINV,	/* 23 - INV | DNML | IMP */
673	FPE_FLTDIV,	/* 24 - DZ | IMP */
674	FPE_FLTINV,	/* 25 - INV | DZ | IMP */
675	FPE_FLTDIV,	/* 26 - DNML | DZ | IMP */
676	FPE_FLTINV,	/* 27 - INV | DNML | DZ | IMP */
677	FPE_FLTOVF,	/* 28 - OFL | IMP */
678	FPE_FLTINV,	/* 29 - INV | OFL | IMP */
679	FPE_FLTUND,	/* 2A - DNML | OFL | IMP */
680	FPE_FLTINV,	/* 2B - INV | DNML | OFL | IMP */
681	FPE_FLTDIV,	/* 2C - DZ | OFL | IMP */
682	FPE_FLTINV,	/* 2D - INV | DZ | OFL | IMP */
683	FPE_FLTDIV,	/* 2E - DNML | DZ | OFL | IMP */
684	FPE_FLTINV,	/* 2F - INV | DNML | DZ | OFL | IMP */
685	FPE_FLTUND,	/* 30 - UFL | IMP */
686	FPE_FLTINV,	/* 31 - INV | UFL | IMP */
687	FPE_FLTUND,	/* 32 - DNML | UFL | IMP */
688	FPE_FLTINV,	/* 33 - INV | DNML | UFL | IMP */
689	FPE_FLTDIV,	/* 34 - DZ | UFL | IMP */
690	FPE_FLTINV,	/* 35 - INV | DZ | UFL | IMP */
691	FPE_FLTDIV,	/* 36 - DNML | DZ | UFL | IMP */
692	FPE_FLTINV,	/* 37 - INV | DNML | DZ | UFL | IMP */
693	FPE_FLTOVF,	/* 38 - OFL | UFL | IMP */
694	FPE_FLTINV,	/* 39 - INV | OFL | UFL | IMP */
695	FPE_FLTUND,	/* 3A - DNML | OFL | UFL | IMP */
696	FPE_FLTINV,	/* 3B - INV | DNML | OFL | UFL | IMP */
697	FPE_FLTDIV,	/* 3C - DZ | OFL | UFL | IMP */
698	FPE_FLTINV,	/* 3D - INV | DZ | OFL | UFL | IMP */
699	FPE_FLTDIV,	/* 3E - DNML | DZ | OFL | UFL | IMP */
700	FPE_FLTINV,	/* 3F - INV | DNML | DZ | OFL | UFL | IMP */
701	FPE_FLTSUB,	/* 40 - STK */
702	FPE_FLTSUB,	/* 41 - INV | STK */
703	FPE_FLTUND,	/* 42 - DNML | STK */
704	FPE_FLTSUB,	/* 43 - INV | DNML | STK */
705	FPE_FLTDIV,	/* 44 - DZ | STK */
706	FPE_FLTSUB,	/* 45 - INV | DZ | STK */
707	FPE_FLTDIV,	/* 46 - DNML | DZ | STK */
708	FPE_FLTSUB,	/* 47 - INV | DNML | DZ | STK */
709	FPE_FLTOVF,	/* 48 - OFL | STK */
710	FPE_FLTSUB,	/* 49 - INV | OFL | STK */
711	FPE_FLTUND,	/* 4A - DNML | OFL | STK */
712	FPE_FLTSUB,	/* 4B - INV | DNML | OFL | STK */
713	FPE_FLTDIV,	/* 4C - DZ | OFL | STK */
714	FPE_FLTSUB,	/* 4D - INV | DZ | OFL | STK */
715	FPE_FLTDIV,	/* 4E - DNML | DZ | OFL | STK */
716	FPE_FLTSUB,	/* 4F - INV | DNML | DZ | OFL | STK */
717	FPE_FLTUND,	/* 50 - UFL | STK */
718	FPE_FLTSUB,	/* 51 - INV | UFL | STK */
719	FPE_FLTUND,	/* 52 - DNML | UFL | STK */
720	FPE_FLTSUB,	/* 53 - INV | DNML | UFL | STK */
721	FPE_FLTDIV,	/* 54 - DZ | UFL | STK */
722	FPE_FLTSUB,	/* 55 - INV | DZ | UFL | STK */
723	FPE_FLTDIV,	/* 56 - DNML | DZ | UFL | STK */
724	FPE_FLTSUB,	/* 57 - INV | DNML | DZ | UFL | STK */
725	FPE_FLTOVF,	/* 58 - OFL | UFL | STK */
726	FPE_FLTSUB,	/* 59 - INV | OFL | UFL | STK */
727	FPE_FLTUND,	/* 5A - DNML | OFL | UFL | STK */
728	FPE_FLTSUB,	/* 5B - INV | DNML | OFL | UFL | STK */
729	FPE_FLTDIV,	/* 5C - DZ | OFL | UFL | STK */
730	FPE_FLTSUB,	/* 5D - INV | DZ | OFL | UFL | STK */
731	FPE_FLTDIV,	/* 5E - DNML | DZ | OFL | UFL | STK */
732	FPE_FLTSUB,	/* 5F - INV | DNML | DZ | OFL | UFL | STK */
733	FPE_FLTRES,	/* 60 - IMP | STK */
734	FPE_FLTSUB,	/* 61 - INV | IMP | STK */
735	FPE_FLTUND,	/* 62 - DNML | IMP | STK */
736	FPE_FLTSUB,	/* 63 - INV | DNML | IMP | STK */
737	FPE_FLTDIV,	/* 64 - DZ | IMP | STK */
738	FPE_FLTSUB,	/* 65 - INV | DZ | IMP | STK */
739	FPE_FLTDIV,	/* 66 - DNML | DZ | IMP | STK */
740	FPE_FLTSUB,	/* 67 - INV | DNML | DZ | IMP | STK */
741	FPE_FLTOVF,	/* 68 - OFL | IMP | STK */
742	FPE_FLTSUB,	/* 69 - INV | OFL | IMP | STK */
743	FPE_FLTUND,	/* 6A - DNML | OFL | IMP | STK */
744	FPE_FLTSUB,	/* 6B - INV | DNML | OFL | IMP | STK */
745	FPE_FLTDIV,	/* 6C - DZ | OFL | IMP | STK */
746	FPE_FLTSUB,	/* 6D - INV | DZ | OFL | IMP | STK */
747	FPE_FLTDIV,	/* 6E - DNML | DZ | OFL | IMP | STK */
748	FPE_FLTSUB,	/* 6F - INV | DNML | DZ | OFL | IMP | STK */
749	FPE_FLTUND,	/* 70 - UFL | IMP | STK */
750	FPE_FLTSUB,	/* 71 - INV | UFL | IMP | STK */
751	FPE_FLTUND,	/* 72 - DNML | UFL | IMP | STK */
752	FPE_FLTSUB,	/* 73 - INV | DNML | UFL | IMP | STK */
753	FPE_FLTDIV,	/* 74 - DZ | UFL | IMP | STK */
754	FPE_FLTSUB,	/* 75 - INV | DZ | UFL | IMP | STK */
755	FPE_FLTDIV,	/* 76 - DNML | DZ | UFL | IMP | STK */
756	FPE_FLTSUB,	/* 77 - INV | DNML | DZ | UFL | IMP | STK */
757	FPE_FLTOVF,	/* 78 - OFL | UFL | IMP | STK */
758	FPE_FLTSUB,	/* 79 - INV | OFL | UFL | IMP | STK */
759	FPE_FLTUND,	/* 7A - DNML | OFL | UFL | IMP | STK */
760	FPE_FLTSUB,	/* 7B - INV | DNML | OFL | UFL | IMP | STK */
761	FPE_FLTDIV,	/* 7C - DZ | OFL | UFL | IMP | STK */
762	FPE_FLTSUB,	/* 7D - INV | DZ | OFL | UFL | IMP | STK */
763	FPE_FLTDIV,	/* 7E - DNML | DZ | OFL | UFL | IMP | STK */
764	FPE_FLTSUB,	/* 7F - INV | DNML | DZ | OFL | UFL | IMP | STK */
765};
766
767/*
768 * Read the FP status and control words, then generate si_code value
769 * for SIGFPE.  The error code chosen will be one of the
770 * FPE_... macros.  It will be sent as the second argument to old
771 * BSD-style signal handlers and as "siginfo_t->si_code" (second
772 * argument) to SA_SIGINFO signal handlers.
773 *
774 * Some time ago, we cleared the x87 exceptions with FNCLEX there.
775 * Clearing exceptions was necessary mainly to avoid IRQ13 bugs.  The
776 * usermode code which understands the FPU hardware enough to enable
777 * the exceptions, can also handle clearing the exception state in the
778 * handler.  The only consequence of not clearing the exception is the
779 * rethrow of the SIGFPE on return from the signal handler and
780 * reexecution of the corresponding instruction.
781 *
782 * For XMM traps, the exceptions were never cleared.
783 */
784int
785npxtrap_x87(void)
786{
787	u_short control, status;
788
789	if (!hw_float) {
790		printf(
791	"npxtrap_x87: fpcurthread = %p, curthread = %p, hw_float = %d\n",
792		       PCPU_GET(fpcurthread), curthread, hw_float);
793		panic("npxtrap from nowhere");
794	}
795	critical_enter();
796
797	/*
798	 * Interrupt handling (for another interrupt) may have pushed the
799	 * state to memory.  Fetch the relevant parts of the state from
800	 * wherever they are.
801	 */
802	if (PCPU_GET(fpcurthread) != curthread) {
803		control = GET_FPU_CW(curthread);
804		status = GET_FPU_SW(curthread);
805	} else {
806		fnstcw(&control);
807		fnstsw(&status);
808	}
809	critical_exit();
810	return (fpetable[status & ((~control & 0x3f) | 0x40)]);
811}
812
813#ifdef CPU_ENABLE_SSE
814int
815npxtrap_sse(void)
816{
817	u_int mxcsr;
818
819	if (!hw_float) {
820		printf(
821	"npxtrap_sse: fpcurthread = %p, curthread = %p, hw_float = %d\n",
822		       PCPU_GET(fpcurthread), curthread, hw_float);
823		panic("npxtrap from nowhere");
824	}
825	critical_enter();
826	if (PCPU_GET(fpcurthread) != curthread)
827		mxcsr = curthread->td_pcb->pcb_save->sv_xmm.sv_env.en_mxcsr;
828	else
829		stmxcsr(&mxcsr);
830	critical_exit();
831	return (fpetable[(mxcsr & (~mxcsr >> 7)) & 0x3f]);
832}
833#endif
834
835/*
836 * Implement device not available (DNA) exception
837 *
838 * It would be better to switch FP context here (if curthread != fpcurthread)
839 * and not necessarily for every context switch, but it is too hard to
840 * access foreign pcb's.
841 */
842
843static int err_count = 0;
844
845int
846npxdna(void)
847{
848
849	if (!hw_float)
850		return (0);
851	critical_enter();
852	if (PCPU_GET(fpcurthread) == curthread) {
853		printf("npxdna: fpcurthread == curthread %d times\n",
854		    ++err_count);
855		stop_emulating();
856		critical_exit();
857		return (1);
858	}
859	if (PCPU_GET(fpcurthread) != NULL) {
860		printf("npxdna: fpcurthread = %p (%d), curthread = %p (%d)\n",
861		       PCPU_GET(fpcurthread),
862		       PCPU_GET(fpcurthread)->td_proc->p_pid,
863		       curthread, curthread->td_proc->p_pid);
864		panic("npxdna");
865	}
866	stop_emulating();
867	/*
868	 * Record new context early in case frstor causes a trap.
869	 */
870	PCPU_SET(fpcurthread, curthread);
871
872#ifdef CPU_ENABLE_SSE
873	if (cpu_fxsr)
874		fpu_clean_state();
875#endif
876
877	if ((curpcb->pcb_flags & PCB_NPXINITDONE) == 0) {
878		/*
879		 * This is the first time this thread has used the FPU or
880		 * the PCB doesn't contain a clean FPU state.  Explicitly
881		 * load an initial state.
882		 *
883		 * We prefer to restore the state from the actual save
884		 * area in PCB instead of directly loading from
885		 * npx_initialstate, to ignite the XSAVEOPT
886		 * tracking engine.
887		 */
888		bcopy(npx_initialstate, curpcb->pcb_save, cpu_max_ext_state_size);
889		fpurstor(curpcb->pcb_save);
890		if (curpcb->pcb_initial_npxcw != __INITIAL_NPXCW__)
891			fldcw(curpcb->pcb_initial_npxcw);
892		curpcb->pcb_flags |= PCB_NPXINITDONE;
893		if (PCB_USER_FPU(curpcb))
894			curpcb->pcb_flags |= PCB_NPXUSERINITDONE;
895	} else {
896		fpurstor(curpcb->pcb_save);
897	}
898	critical_exit();
899
900	return (1);
901}
902
903/*
904 * Wrapper for fpusave() called from context switch routines.
905 *
906 * npxsave() must be called with interrupts disabled, so that it clears
907 * fpcurthread atomically with saving the state.  We require callers to do the
908 * disabling, since most callers need to disable interrupts anyway to call
909 * npxsave() atomically with checking fpcurthread.
910 */
911void
912npxsave(addr)
913	union savefpu *addr;
914{
915
916	stop_emulating();
917#ifdef CPU_ENABLE_SSE
918	if (use_xsaveopt)
919		xsaveopt((char *)addr, xsave_mask);
920	else
921#endif
922		fpusave(addr);
923	start_emulating();
924	PCPU_SET(fpcurthread, NULL);
925}
926
927/*
928 * Unconditionally save the current co-processor state across suspend and
929 * resume.
930 */
931void
932npxsuspend(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
948void
949npxresume(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
963void
964npxdrop()
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#ifdef CPU_ENABLE_SSE
973	if (!cpu_fxsr)
974#endif
975		fnclex();
976
977	td = PCPU_GET(fpcurthread);
978	KASSERT(td == curthread, ("fpudrop: fpcurthread != curthread"));
979	CRITICAL_ASSERT(td);
980	PCPU_SET(fpcurthread, NULL);
981	td->td_pcb->pcb_flags &= ~PCB_NPXINITDONE;
982	start_emulating();
983}
984
985/*
986 * Get the user state of the FPU into pcb->pcb_user_save without
987 * dropping ownership (if possible).  It returns the FPU ownership
988 * status.
989 */
990int
991npxgetregs(struct thread *td)
992{
993	struct pcb *pcb;
994#ifdef CPU_ENABLE_SSE
995	uint64_t *xstate_bv, bit;
996	char *sa;
997	int max_ext_n, i;
998#endif
999	int owned;
1000
1001	if (!hw_float)
1002		return (_MC_FPOWNED_NONE);
1003
1004	pcb = td->td_pcb;
1005	if ((pcb->pcb_flags & PCB_NPXINITDONE) == 0) {
1006		bcopy(npx_initialstate, get_pcb_user_save_pcb(pcb),
1007		    cpu_max_ext_state_size);
1008		SET_FPU_CW(get_pcb_user_save_pcb(pcb), pcb->pcb_initial_npxcw);
1009		npxuserinited(td);
1010		return (_MC_FPOWNED_PCB);
1011	}
1012	critical_enter();
1013	if (td == PCPU_GET(fpcurthread)) {
1014		fpusave(get_pcb_user_save_pcb(pcb));
1015#ifdef CPU_ENABLE_SSE
1016		if (!cpu_fxsr)
1017#endif
1018			/*
1019			 * fnsave initializes the FPU and destroys whatever
1020			 * context it contains.  Make sure the FPU owner
1021			 * starts with a clean state next time.
1022			 */
1023			npxdrop();
1024		owned = _MC_FPOWNED_FPU;
1025	} else {
1026		owned = _MC_FPOWNED_PCB;
1027	}
1028	critical_exit();
1029#ifdef CPU_ENABLE_SSE
1030	if (use_xsave) {
1031		/*
1032		 * Handle partially saved state.
1033		 */
1034		sa = (char *)get_pcb_user_save_pcb(pcb);
1035		xstate_bv = (uint64_t *)(sa + sizeof(union savefpu) +
1036		    offsetof(struct xstate_hdr, xstate_bv));
1037		if (xsave_mask >> 32 != 0)
1038			max_ext_n = fls(xsave_mask >> 32) + 32;
1039		else
1040			max_ext_n = fls(xsave_mask);
1041		for (i = 0; i < max_ext_n; i++) {
1042			bit = 1ULL << i;
1043			if ((xsave_mask & bit) == 0 || (*xstate_bv & bit) != 0)
1044				continue;
1045			bcopy((char *)npx_initialstate +
1046			    xsave_area_desc[i].offset,
1047			    sa + xsave_area_desc[i].offset,
1048			    xsave_area_desc[i].size);
1049			*xstate_bv |= bit;
1050		}
1051	}
1052#endif
1053	return (owned);
1054}
1055
1056void
1057npxuserinited(struct thread *td)
1058{
1059	struct pcb *pcb;
1060
1061	pcb = td->td_pcb;
1062	if (PCB_USER_FPU(pcb))
1063		pcb->pcb_flags |= PCB_NPXINITDONE;
1064	pcb->pcb_flags |= PCB_NPXUSERINITDONE;
1065}
1066
1067#ifdef CPU_ENABLE_SSE
1068int
1069npxsetxstate(struct thread *td, char *xfpustate, size_t xfpustate_size)
1070{
1071	struct xstate_hdr *hdr, *ehdr;
1072	size_t len, max_len;
1073	uint64_t bv;
1074
1075	/* XXXKIB should we clear all extended state in xstate_bv instead ? */
1076	if (xfpustate == NULL)
1077		return (0);
1078	if (!use_xsave)
1079		return (EOPNOTSUPP);
1080
1081	len = xfpustate_size;
1082	if (len < sizeof(struct xstate_hdr))
1083		return (EINVAL);
1084	max_len = cpu_max_ext_state_size - sizeof(union savefpu);
1085	if (len > max_len)
1086		return (EINVAL);
1087
1088	ehdr = (struct xstate_hdr *)xfpustate;
1089	bv = ehdr->xstate_bv;
1090
1091	/*
1092	 * Avoid #gp.
1093	 */
1094	if (bv & ~xsave_mask)
1095		return (EINVAL);
1096
1097	hdr = (struct xstate_hdr *)(get_pcb_user_save_td(td) + 1);
1098
1099	hdr->xstate_bv = bv;
1100	bcopy(xfpustate + sizeof(struct xstate_hdr),
1101	    (char *)(hdr + 1), len - sizeof(struct xstate_hdr));
1102
1103	return (0);
1104}
1105#endif
1106
1107int
1108npxsetregs(struct thread *td, union savefpu *addr, char *xfpustate,
1109	size_t xfpustate_size)
1110{
1111	struct pcb *pcb;
1112#ifdef CPU_ENABLE_SSE
1113	int error;
1114#endif
1115
1116	if (!hw_float)
1117		return (ENXIO);
1118
1119	pcb = td->td_pcb;
1120	critical_enter();
1121	if (td == PCPU_GET(fpcurthread) && PCB_USER_FPU(pcb)) {
1122#ifdef CPU_ENABLE_SSE
1123		error = npxsetxstate(td, xfpustate, xfpustate_size);
1124		if (error != 0) {
1125			critical_exit();
1126			return (error);
1127		}
1128		if (!cpu_fxsr)
1129#endif
1130			fnclex();	/* As in npxdrop(). */
1131		bcopy(addr, get_pcb_user_save_td(td), sizeof(*addr));
1132		fpurstor(get_pcb_user_save_td(td));
1133		critical_exit();
1134		pcb->pcb_flags |= PCB_NPXUSERINITDONE | PCB_NPXINITDONE;
1135	} else {
1136		critical_exit();
1137#ifdef CPU_ENABLE_SSE
1138		error = npxsetxstate(td, xfpustate, xfpustate_size);
1139		if (error != 0)
1140			return (error);
1141#endif
1142		bcopy(addr, get_pcb_user_save_td(td), sizeof(*addr));
1143		npxuserinited(td);
1144	}
1145	return (0);
1146}
1147
1148static void
1149fpusave(addr)
1150	union savefpu *addr;
1151{
1152
1153#ifdef CPU_ENABLE_SSE
1154	if (use_xsave)
1155		xsave((char *)addr, xsave_mask);
1156	else if (cpu_fxsr)
1157		fxsave(addr);
1158	else
1159#endif
1160		fnsave(addr);
1161}
1162
1163#ifdef CPU_ENABLE_SSE
1164static void
1165npx_fill_fpregs_xmm1(struct savexmm *sv_xmm, struct save87 *sv_87)
1166{
1167	struct env87 *penv_87;
1168	struct envxmm *penv_xmm;
1169	int i;
1170
1171	penv_87 = &sv_87->sv_env;
1172	penv_xmm = &sv_xmm->sv_env;
1173
1174	/* FPU control/status */
1175	penv_87->en_cw = penv_xmm->en_cw;
1176	penv_87->en_sw = penv_xmm->en_sw;
1177	penv_87->en_fip = penv_xmm->en_fip;
1178	penv_87->en_fcs = penv_xmm->en_fcs;
1179	penv_87->en_opcode = penv_xmm->en_opcode;
1180	penv_87->en_foo = penv_xmm->en_foo;
1181	penv_87->en_fos = penv_xmm->en_fos;
1182
1183	/* FPU registers and tags */
1184	penv_87->en_tw = 0xffff;
1185	for (i = 0; i < 8; ++i) {
1186		sv_87->sv_ac[i] = sv_xmm->sv_fp[i].fp_acc;
1187		if ((penv_xmm->en_tw & (1 << i)) != 0)
1188			/* zero and special are set as valid */
1189			penv_87->en_tw &= ~(3 << i);
1190	}
1191}
1192
1193void
1194npx_fill_fpregs_xmm(struct savexmm *sv_xmm, struct save87 *sv_87)
1195{
1196
1197	bzero(sv_87, sizeof(*sv_87));
1198	npx_fill_fpregs_xmm1(sv_xmm, sv_87);
1199}
1200
1201void
1202npx_set_fpregs_xmm(struct save87 *sv_87, struct savexmm *sv_xmm)
1203{
1204	struct env87 *penv_87;
1205	struct envxmm *penv_xmm;
1206	int i;
1207
1208	penv_87 = &sv_87->sv_env;
1209	penv_xmm = &sv_xmm->sv_env;
1210
1211	/* FPU control/status */
1212	penv_xmm->en_cw = penv_87->en_cw;
1213	penv_xmm->en_sw = penv_87->en_sw;
1214	penv_xmm->en_fip = penv_87->en_fip;
1215	penv_xmm->en_fcs = penv_87->en_fcs;
1216	penv_xmm->en_opcode = penv_87->en_opcode;
1217	penv_xmm->en_foo = penv_87->en_foo;
1218	penv_xmm->en_fos = penv_87->en_fos;
1219
1220	/* FPU registers and tags */
1221	penv_xmm->en_tw = 0;
1222	for (i = 0; i < 8; ++i) {
1223		sv_xmm->sv_fp[i].fp_acc = sv_87->sv_ac[i];
1224		if ((penv_87->en_tw && (3 << i)) != (3 << i))
1225		    penv_xmm->en_tw |= 1 << i;
1226	}
1227}
1228#endif /* CPU_ENABLE_SSE */
1229
1230void
1231npx_get_fsave(void *addr)
1232{
1233	struct thread *td;
1234	union savefpu *sv;
1235
1236	td = curthread;
1237	npxgetregs(td);
1238	sv = get_pcb_user_save_td(td);
1239#ifdef CPU_ENABLE_SSE
1240	if (cpu_fxsr)
1241		npx_fill_fpregs_xmm1(&sv->sv_xmm, addr);
1242	else
1243#endif
1244		bcopy(sv, addr, sizeof(struct env87) +
1245		    sizeof(struct fpacc87[8]));
1246}
1247
1248int
1249npx_set_fsave(void *addr)
1250{
1251	union savefpu sv;
1252	int error;
1253
1254	bzero(&sv, sizeof(sv));
1255#ifdef CPU_ENABLE_SSE
1256	if (cpu_fxsr)
1257		npx_set_fpregs_xmm(addr, &sv.sv_xmm);
1258	else
1259#endif
1260		bcopy(addr, &sv, sizeof(struct env87) +
1261		    sizeof(struct fpacc87[8]));
1262	error = npxsetregs(curthread, &sv, NULL, 0);
1263	return (error);
1264}
1265
1266#ifdef CPU_ENABLE_SSE
1267/*
1268 * On AuthenticAMD processors, the fxrstor instruction does not restore
1269 * the x87's stored last instruction pointer, last data pointer, and last
1270 * opcode values, except in the rare case in which the exception summary
1271 * (ES) bit in the x87 status word is set to 1.
1272 *
1273 * In order to avoid leaking this information across processes, we clean
1274 * these values by performing a dummy load before executing fxrstor().
1275 */
1276static void
1277fpu_clean_state(void)
1278{
1279	static float dummy_variable = 0.0;
1280	u_short status;
1281
1282	/*
1283	 * Clear the ES bit in the x87 status word if it is currently
1284	 * set, in order to avoid causing a fault in the upcoming load.
1285	 */
1286	fnstsw(&status);
1287	if (status & 0x80)
1288		fnclex();
1289
1290	/*
1291	 * Load the dummy variable into the x87 stack.  This mangles
1292	 * the x87 stack, but we don't care since we're about to call
1293	 * fxrstor() anyway.
1294	 */
1295	__asm __volatile("ffree %%st(7); flds %0" : : "m" (dummy_variable));
1296}
1297#endif /* CPU_ENABLE_SSE */
1298
1299static void
1300fpurstor(addr)
1301	union savefpu *addr;
1302{
1303
1304#ifdef CPU_ENABLE_SSE
1305	if (use_xsave)
1306		xrstor((char *)addr, xsave_mask);
1307	else if (cpu_fxsr)
1308		fxrstor(addr);
1309	else
1310#endif
1311		frstor(addr);
1312}
1313
1314#ifdef DEV_ISA
1315/*
1316 * This sucks up the legacy ISA support assignments from PNPBIOS/ACPI.
1317 */
1318static struct isa_pnp_id npxisa_ids[] = {
1319	{ 0x040cd041, "Legacy ISA coprocessor support" }, /* PNP0C04 */
1320	{ 0 }
1321};
1322
1323static int
1324npxisa_probe(device_t dev)
1325{
1326	int result;
1327	if ((result = ISA_PNP_PROBE(device_get_parent(dev), dev, npxisa_ids)) <= 0) {
1328		device_quiet(dev);
1329	}
1330	return(result);
1331}
1332
1333static int
1334npxisa_attach(device_t dev)
1335{
1336	return (0);
1337}
1338
1339static device_method_t npxisa_methods[] = {
1340	/* Device interface */
1341	DEVMETHOD(device_probe,		npxisa_probe),
1342	DEVMETHOD(device_attach,	npxisa_attach),
1343	DEVMETHOD(device_detach,	bus_generic_detach),
1344	DEVMETHOD(device_shutdown,	bus_generic_shutdown),
1345	DEVMETHOD(device_suspend,	bus_generic_suspend),
1346	DEVMETHOD(device_resume,	bus_generic_resume),
1347
1348	{ 0, 0 }
1349};
1350
1351static driver_t npxisa_driver = {
1352	"npxisa",
1353	npxisa_methods,
1354	1,			/* no softc */
1355};
1356
1357static devclass_t npxisa_devclass;
1358
1359DRIVER_MODULE(npxisa, isa, npxisa_driver, npxisa_devclass, 0, 0);
1360#ifndef PC98
1361DRIVER_MODULE(npxisa, acpi, npxisa_driver, npxisa_devclass, 0, 0);
1362#endif
1363#endif /* DEV_ISA */
1364
1365static MALLOC_DEFINE(M_FPUKERN_CTX, "fpukern_ctx",
1366    "Kernel contexts for FPU state");
1367
1368#define	FPU_KERN_CTX_NPXINITDONE 0x01
1369#define	FPU_KERN_CTX_DUMMY	 0x02
1370#define	FPU_KERN_CTX_INUSE	 0x04
1371
1372struct fpu_kern_ctx {
1373	union savefpu *prev;
1374	uint32_t flags;
1375	char hwstate1[];
1376};
1377
1378struct fpu_kern_ctx *
1379fpu_kern_alloc_ctx(u_int flags)
1380{
1381	struct fpu_kern_ctx *res;
1382	size_t sz;
1383
1384	sz = sizeof(struct fpu_kern_ctx) + XSAVE_AREA_ALIGN +
1385	    cpu_max_ext_state_size;
1386	res = malloc(sz, M_FPUKERN_CTX, ((flags & FPU_KERN_NOWAIT) ?
1387	    M_NOWAIT : M_WAITOK) | M_ZERO);
1388	return (res);
1389}
1390
1391void
1392fpu_kern_free_ctx(struct fpu_kern_ctx *ctx)
1393{
1394
1395	KASSERT((ctx->flags & FPU_KERN_CTX_INUSE) == 0, ("free'ing inuse ctx"));
1396	/* XXXKIB clear the memory ? */
1397	free(ctx, M_FPUKERN_CTX);
1398}
1399
1400static union savefpu *
1401fpu_kern_ctx_savefpu(struct fpu_kern_ctx *ctx)
1402{
1403	vm_offset_t p;
1404
1405	p = (vm_offset_t)&ctx->hwstate1;
1406	p = roundup2(p, XSAVE_AREA_ALIGN);
1407	return ((union savefpu *)p);
1408}
1409
1410int
1411fpu_kern_enter(struct thread *td, struct fpu_kern_ctx *ctx, u_int flags)
1412{
1413	struct pcb *pcb;
1414
1415	KASSERT((ctx->flags & FPU_KERN_CTX_INUSE) == 0, ("using inuse ctx"));
1416
1417	if ((flags & FPU_KERN_KTHR) != 0 && is_fpu_kern_thread(0)) {
1418		ctx->flags = FPU_KERN_CTX_DUMMY | FPU_KERN_CTX_INUSE;
1419		return (0);
1420	}
1421	pcb = td->td_pcb;
1422	KASSERT(!PCB_USER_FPU(pcb) || pcb->pcb_save ==
1423	    get_pcb_user_save_pcb(pcb), ("mangled pcb_save"));
1424	ctx->flags = FPU_KERN_CTX_INUSE;
1425	if ((pcb->pcb_flags & PCB_NPXINITDONE) != 0)
1426		ctx->flags |= FPU_KERN_CTX_NPXINITDONE;
1427	npxexit(td);
1428	ctx->prev = pcb->pcb_save;
1429	pcb->pcb_save = fpu_kern_ctx_savefpu(ctx);
1430	pcb->pcb_flags |= PCB_KERNNPX;
1431	pcb->pcb_flags &= ~PCB_NPXINITDONE;
1432	return (0);
1433}
1434
1435int
1436fpu_kern_leave(struct thread *td, struct fpu_kern_ctx *ctx)
1437{
1438	struct pcb *pcb;
1439
1440	KASSERT((ctx->flags & FPU_KERN_CTX_INUSE) != 0,
1441	    ("leaving not inuse ctx"));
1442	ctx->flags &= ~FPU_KERN_CTX_INUSE;
1443
1444	if (is_fpu_kern_thread(0) && (ctx->flags & FPU_KERN_CTX_DUMMY) != 0)
1445		return (0);
1446	pcb = td->td_pcb;
1447	critical_enter();
1448	if (curthread == PCPU_GET(fpcurthread))
1449		npxdrop();
1450	critical_exit();
1451	pcb->pcb_save = ctx->prev;
1452	if (pcb->pcb_save == get_pcb_user_save_pcb(pcb)) {
1453		if ((pcb->pcb_flags & PCB_NPXUSERINITDONE) != 0)
1454			pcb->pcb_flags |= PCB_NPXINITDONE;
1455		else
1456			pcb->pcb_flags &= ~PCB_NPXINITDONE;
1457		pcb->pcb_flags &= ~PCB_KERNNPX;
1458	} else {
1459		if ((ctx->flags & FPU_KERN_CTX_NPXINITDONE) != 0)
1460			pcb->pcb_flags |= PCB_NPXINITDONE;
1461		else
1462			pcb->pcb_flags &= ~PCB_NPXINITDONE;
1463		KASSERT(!PCB_USER_FPU(pcb), ("unpaired fpu_kern_leave"));
1464	}
1465	return (0);
1466}
1467
1468int
1469fpu_kern_thread(u_int flags)
1470{
1471
1472	KASSERT((curthread->td_pflags & TDP_KTHREAD) != 0,
1473	    ("Only kthread may use fpu_kern_thread"));
1474	KASSERT(curpcb->pcb_save == get_pcb_user_save_pcb(curpcb),
1475	    ("mangled pcb_save"));
1476	KASSERT(PCB_USER_FPU(curpcb), ("recursive call"));
1477
1478	curpcb->pcb_flags |= PCB_KERNNPX;
1479	return (0);
1480}
1481
1482int
1483is_fpu_kern_thread(u_int flags)
1484{
1485
1486	if ((curthread->td_pflags & TDP_KTHREAD) == 0)
1487		return (0);
1488	return ((curpcb->pcb_flags & PCB_KERNNPX) != 0);
1489}
1490
1491/*
1492 * FPU save area alloc/free/init utility routines
1493 */
1494union savefpu *
1495fpu_save_area_alloc(void)
1496{
1497
1498	return (uma_zalloc(fpu_save_area_zone, 0));
1499}
1500
1501void
1502fpu_save_area_free(union savefpu *fsa)
1503{
1504
1505	uma_zfree(fpu_save_area_zone, fsa);
1506}
1507
1508void
1509fpu_save_area_reset(union savefpu *fsa)
1510{
1511
1512	bcopy(npx_initialstate, fsa, cpu_max_ext_state_size);
1513}
1514