xref: /openbsd/sys/arch/i386/isa/npx.c (revision 0f9e9ec2)
1 /*	$OpenBSD: npx.c,v 1.76 2024/05/13 01:15:50 jsg Exp $	*/
2 /*	$NetBSD: npx.c,v 1.57 1996/05/12 23:12:24 mycroft Exp $	*/
3 
4 #if 0
5 #define IPRINTF(x)	printf x
6 #else
7 #define	IPRINTF(x)
8 #endif
9 
10 /*-
11  * Copyright (c) 1994, 1995 Charles M. Hannum.  All rights reserved.
12  * Copyright (c) 1990 William Jolitz.
13  * Copyright (c) 1991 The Regents of the University of California.
14  * All rights reserved.
15  *
16  * Redistribution and use in source and binary forms, with or without
17  * modification, are permitted provided that the following conditions
18  * are met:
19  * 1. Redistributions of source code must retain the above copyright
20  *    notice, this list of conditions and the following disclaimer.
21  * 2. Redistributions in binary form must reproduce the above copyright
22  *    notice, this list of conditions and the following disclaimer in the
23  *    documentation and/or other materials provided with the distribution.
24  * 3. Neither the name of the University nor the names of its contributors
25  *    may be used to endorse or promote products derived from this software
26  *    without specific prior written permission.
27  *
28  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
29  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
30  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
31  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
32  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
33  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
34  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
35  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
36  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
37  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38  * SUCH DAMAGE.
39  *
40  *	@(#)npx.c	7.2 (Berkeley) 5/12/91
41  */
42 
43 #include <sys/param.h>
44 #include <sys/systm.h>
45 #include <sys/proc.h>
46 #include <sys/signalvar.h>
47 #include <sys/user.h>
48 #include <sys/device.h>
49 
50 #include <uvm/uvm_extern.h>
51 
52 #include <machine/intr.h>
53 #include <machine/npx.h>
54 #include <machine/pio.h>
55 #include <machine/cpufunc.h>
56 #include <machine/pcb.h>
57 #include <machine/trap.h>
58 #include <machine/specialreg.h>
59 #include <machine/i8259.h>
60 
61 #include <dev/isa/isavar.h>
62 
63 /*
64  * 387 and 287 Numeric Coprocessor Extension (NPX) Driver.
65  *
66  * We do lazy initialization and switching using the TS bit in cr0 and the
67  * MDP_USEDFPU bit in mdproc.
68  *
69  * DNA exceptions are handled like this:
70  *
71  * 1) If there is no NPX, return and go to the emulator.
72  * 2) If someone else has used the NPX, save its state into that process's PCB.
73  * 3a) If MDP_USEDFPU is not set, set it and initialize the NPX.
74  * 3b) Otherwise, reload the process's previous NPX state.
75  *
76  * When a process is created or exec()s, its saved cr0 image has the TS bit
77  * set and the MDP_USEDFPU bit clear.  The MDP_USEDFPU bit is set when the
78  * process first gets a DNA and the NPX is initialized.  The TS bit is turned
79  * off when the NPX is used, and turned on again later when the process's NPX
80  * state is saved.
81  */
82 
83 #define	fldcw(addr)		__asm("fldcw %0" : : "m" (*addr))
84 #define	fnclex()		__asm("fnclex")
85 #define	fninit()		__asm("fninit")
86 #define	fnsave(addr)		__asm("fnsave %0" : "=m" (*addr))
87 #define	fnstcw(addr)		__asm("fnstcw %0" : "=m" (*addr))
88 #define	fnstsw(addr)		__asm("fnstsw %0" : "=m" (*addr))
89 #define	fp_divide_by_0()	__asm("fldz; fld1; fdiv %st,%st(1); fwait")
90 #define	frstor(addr)		__asm("frstor %0" : : "m" (*addr))
91 #define	fwait()			__asm("fwait")
92 #define	clts()			__asm("clts")
93 #define	stts()			lcr0(rcr0() | CR0_TS)
94 
95 /*
96  * The mxcsr_mask for this host, taken from fxsave() on the primary CPU
97  */
98 uint32_t	fpu_mxcsr_mask;
99 
100 int npxintr(void *);
101 static int npxprobe1(struct isa_attach_args *);
102 static int x86fpflags_to_siginfo(u_int32_t);
103 
104 
105 struct npx_softc {
106 	struct device sc_dev;
107 	void *sc_ih;
108 };
109 
110 int npxprobe(struct device *, void *, void *);
111 void npxattach(struct device *, struct device *, void *);
112 
113 const struct cfattach npx_ca = {
114 	sizeof(struct npx_softc), npxprobe, npxattach
115 };
116 
117 struct cfdriver npx_cd = {
118 	NULL, "npx", DV_DULL
119 };
120 
121 enum npx_type {
122 	NPX_NONE = 0,
123 	NPX_INTERRUPT,
124 	NPX_EXCEPTION,
125 	NPX_BROKEN,
126 	NPX_CPUID,
127 };
128 
129 static	enum npx_type		npx_type;
130 static	volatile u_int		npx_intrs_while_probing
131 				    __attribute__((section(".kudata")));
132 static	volatile u_int		npx_traps_while_probing
133 				    __attribute__((section(".kudata")));
134 
135 #define fxsave(addr)		__asm("fxsave %0" : "=m" (*addr))
136 #define fxrstor(addr)		__asm("fxrstor %0" : : "m" (*addr))
137 #define ldmxcsr(addr)		__asm("ldmxcsr %0" : : "m" (*addr))
138 
139 static __inline void
fpu_save(union savefpu * addr)140 fpu_save(union savefpu *addr)
141 {
142 
143 	if (i386_use_fxsave) {
144 		fxsave(&addr->sv_xmm);
145 		/* FXSAVE doesn't FNINIT like FNSAVE does -- so do it here. */
146 		fninit();
147 	} else
148 		fnsave(&addr->sv_87);
149 }
150 
151 static int
npxdna_notset(struct cpu_info * ci)152 npxdna_notset(struct cpu_info *ci)
153 {
154 	panic("npxdna vector not initialized");
155 }
156 
157 int    (*npxdna_func)(struct cpu_info *) = npxdna_notset;
158 int    npxdna_s87(struct cpu_info *);
159 int    npxdna_xmm(struct cpu_info *);
160 
161 /*
162  * Special interrupt handlers.  Someday intr0-intr15 will be used to count
163  * interrupts.  We'll still need a special exception 16 handler.  The busy
164  * latch stuff in probintr() can be moved to npxprobe().
165  */
166 void probeintr(void);
167 asm (".text\n\t"
168 "probeintr:\n\t"
169 	"ss\n\t"
170 	"incl	npx_intrs_while_probing\n\t"
171 	"pushl	%eax\n\t"
172 	"movb	$0x20,%al	# EOI (asm in strings loses cpp features)\n\t"
173 	"outb	%al,$0xa0	# IO_ICU2\n\t"
174 	"outb	%al,$0x20	# IO_ICU1\n\t"
175 	"movb	$0,%al\n\t"
176 	"outb	%al,$0xf0	# clear BUSY# latch\n\t"
177 	"popl	%eax\n\t"
178 	"iret\n\t");
179 
180 void probetrap(void);
181 asm (".text\n\t"
182 "probetrap:\n\t"
183 	"ss\n\t"
184 	"incl	npx_traps_while_probing\n\t"
185 	"fnclex\n\t"
186 	"iret\n\t");
187 
188 static inline int
npxprobe1(struct isa_attach_args * ia)189 npxprobe1(struct isa_attach_args *ia)
190 {
191 	int control;
192 	int status;
193 
194 	ia->ia_iosize = 16;
195 	ia->ia_msize = 0;
196 
197 	/*
198 	 * Finish resetting the coprocessor, if any.  If there is an error
199 	 * pending, then we may get a bogus IRQ13, but probeintr() will handle
200 	 * it OK.  Bogus halts have never been observed, but we enabled
201 	 * IRQ13 and cleared the BUSY# latch early to handle them anyway.
202 	 */
203 	fninit();
204 	delay(1000);		/* wait for any IRQ13 (fwait might hang) */
205 
206 	/*
207 	 * Check for a status of mostly zero.
208 	 */
209 	status = 0x5a5a;
210 	fnstsw(&status);
211 	if ((status & 0xb8ff) == 0) {
212 		/*
213 		 * Good, now check for a proper control word.
214 		 */
215 		control = 0x5a5a;
216 		fnstcw(&control);
217 		if ((control & 0x1f3f) == 0x033f) {
218 			/*
219 			 * We have an npx, now divide by 0 to see if exception
220 			 * 16 works.
221 			 */
222 			control &= ~(1 << 2);	/* enable divide by 0 trap */
223 			fldcw(&control);
224 			npx_traps_while_probing = npx_intrs_while_probing = 0;
225 			fp_divide_by_0();
226 			delay(1);
227 			if (npx_traps_while_probing != 0) {
228 				/*
229 				 * Good, exception 16 works.
230 				 */
231 				npx_type = NPX_EXCEPTION;
232 				ia->ia_irq = IRQUNK;	/* zap the interrupt */
233 			} else if (npx_intrs_while_probing != 0) {
234 				/*
235 				 * Bad, we are stuck with IRQ13.
236 				 */
237 				npx_type = NPX_INTERRUPT;
238 			} else {
239 				/*
240 				 * Worse, even IRQ13 is broken.
241 				 */
242 				npx_type = NPX_BROKEN;
243 				ia->ia_irq = IRQUNK;
244 			}
245 			return 1;
246 		}
247 	}
248 
249 	/*
250 	 * Probe failed.  There is no usable FPU.
251 	 */
252 	npx_type = NPX_NONE;
253 	return 0;
254 }
255 
256 /*
257  * Probe routine.  Initialize cr0 to give correct behaviour for [f]wait
258  * whether the device exists or not (XXX should be elsewhere).  Set flags
259  * to tell npxattach() what to do.  Modify device struct if npx doesn't
260  * need to use interrupts.  Return 1 if device exists.
261  */
262 int
npxprobe(struct device * parent,void * match,void * aux)263 npxprobe(struct device *parent, void *match, void *aux)
264 {
265 	struct	isa_attach_args *ia = aux;
266 	int	irq;
267 	int	result;
268 	u_long	s;
269 	unsigned save_imen;
270 	struct	gate_descriptor save_idt_npxintr;
271 	struct	gate_descriptor save_idt_npxtrap;
272 
273 	if (cpu_feature & CPUID_FPU) {
274 		npx_type = NPX_CPUID;
275 		ia->ia_irq = IRQUNK;	/* Don't want the interrupt vector */
276 		ia->ia_iosize = 16;
277 		ia->ia_msize = 0;
278 		return 1;
279 	}
280 
281 	/*
282 	 * This routine is now just a wrapper for npxprobe1(), to install
283 	 * special npx interrupt and trap handlers, to enable npx interrupts
284 	 * and to disable other interrupts.  Someday isa_configure() will
285 	 * install suitable handlers and run with interrupts enabled so we
286 	 * won't need to do so much here.
287 	 */
288 	irq = NRSVIDT + ia->ia_irq;
289 	s = intr_disable();
290 	save_idt_npxintr = idt[irq];
291 	save_idt_npxtrap = idt[16];
292 	setgate(&idt[irq], probeintr, 0, SDT_SYS386IGT, SEL_KPL, GICODE_SEL);
293 	setgate(&idt[16], probetrap, 0, SDT_SYS386TGT, SEL_KPL, GCODE_SEL);
294 	save_imen = imen;
295 	imen = ~((1 << IRQ_SLAVE) | (1 << ia->ia_irq));
296 	SET_ICUS();
297 
298 	/*
299 	 * Partially reset the coprocessor, if any.  Some BIOS's don't reset
300 	 * it after a warm boot.
301 	 */
302 	outb(0xf1, 0);		/* full reset on some systems, NOP on others */
303 	delay(1000);
304 	outb(0xf0, 0);		/* clear BUSY# latch */
305 
306 	/*
307 	 * We set CR0 in locore to trap all ESC and WAIT instructions.
308 	 * We have to turn off the CR0_EM bit temporarily while probing.
309 	 */
310 	lcr0(rcr0() & ~(CR0_EM|CR0_TS));
311 	intr_restore(s);
312 	result = npxprobe1(ia);
313 	s = intr_disable();
314 	lcr0(rcr0() | (CR0_EM|CR0_TS));
315 
316 	imen = save_imen;
317 	SET_ICUS();
318 	idt[irq] = save_idt_npxintr;
319 	idt[16] = save_idt_npxtrap;
320 	intr_restore(s);
321 	return (result);
322 }
323 
324 int npx586bug1(int, int);
325 asm (".text\n\t"
326 "npx586bug1:\n\t"
327 	"fildl	4(%esp)		# x\n\t"
328 	"fildl	8(%esp)		# y\n\t"
329 	"fld	%st(1)\n\t"
330 	"fdiv	%st(1),%st	# x/y\n\t"
331 	"fmulp	%st,%st(1)	# (x/y)*y\n\t"
332 	"fsubrp	%st,%st(1)	# x-(x/y)*y\n\t"
333 	"pushl	$0\n\t"
334 	"fistpl	(%esp)\n\t"
335 	"popl	%eax\n\t"
336 	"ret\n\t");
337 
338 void
npxinit(struct cpu_info * ci)339 npxinit(struct cpu_info *ci)
340 {
341 	lcr0(rcr0() & ~(CR0_EM|CR0_TS));
342 	fninit();
343 	if (npx586bug1(4195835, 3145727) != 0) {
344 		printf("%s: WARNING: Pentium FDIV bug detected!\n",
345 		    ci->ci_dev->dv_xname);
346 	}
347 	if (fpu_mxcsr_mask == 0 && i386_use_fxsave) {
348 		struct savexmm xm __attribute__((aligned(16)));
349 
350 		bzero(&xm, sizeof(xm));
351 		fxsave(&xm);
352 		if (xm.sv_env.en_mxcsr_mask)
353 			fpu_mxcsr_mask = xm.sv_env.en_mxcsr_mask;
354 		else
355 			fpu_mxcsr_mask = __INITIAL_MXCSR_MASK__;
356 	}
357 	lcr0(rcr0() | (CR0_TS));
358 }
359 
360 /*
361  * Attach routine - announce which it is, and wire into system
362  */
363 void
npxattach(struct device * parent,struct device * self,void * aux)364 npxattach(struct device *parent, struct device *self, void *aux)
365 {
366 	struct npx_softc *sc = (void *)self;
367 	struct isa_attach_args *ia = aux;
368 
369 	switch (npx_type) {
370 	case NPX_INTERRUPT:
371 		printf("\n");
372 		lcr0(rcr0() & ~CR0_NE);
373 		sc->sc_ih = isa_intr_establish(ia->ia_ic, ia->ia_irq,
374 		    IST_EDGE, IPL_NONE, npxintr, 0, sc->sc_dev.dv_xname);
375 		break;
376 	case NPX_EXCEPTION:
377 		printf(": using exception 16\n");
378 		break;
379 	case NPX_CPUID:
380 		printf(": reported by CPUID; using exception 16\n");
381 		npx_type = NPX_EXCEPTION;
382 		break;
383 	case NPX_BROKEN:
384 		printf(": error reporting broken; not using\n");
385 		npx_type = NPX_NONE;
386 		return;
387 	case NPX_NONE:
388 		return;
389 	}
390 
391 	npxinit(&cpu_info_primary);
392 
393 	if (i386_use_fxsave)
394 		npxdna_func = npxdna_xmm;
395 	else
396 		npxdna_func = npxdna_s87;
397 }
398 
399 /*
400  * Record the FPU state and reinitialize it all except for the control word.
401  * Then generate a SIGFPE.
402  *
403  * Reinitializing the state allows naive SIGFPE handlers to longjmp without
404  * doing any fixups.
405  *
406  * XXX there is currently no way to pass the full error state to signal
407  * handlers, and if this is a nested interrupt there is no way to pass even
408  * a status code!  So there is no way to have a non-naive SIGFPE handler.  At
409  * best a handler could do an fninit followed by an fldcw of a static value.
410  * fnclex would be of little use because it would leave junk on the FPU stack.
411  * Returning from the handler would be even less safe than usual because
412  * IRQ13 exception handling makes exceptions even less precise than usual.
413  */
414 int
npxintr(void * arg)415 npxintr(void *arg)
416 {
417 	struct cpu_info *ci = curcpu();
418 	struct proc *p = ci->ci_fpcurproc;
419 	union savefpu *addr;
420 	struct intrframe *frame = arg;
421 	int code;
422 	union sigval sv;
423 
424 	uvmexp.traps++;
425 	IPRINTF(("%s: fp intr\n", ci->ci_dev->dv_xname));
426 
427 	if (p == NULL || npx_type == NPX_NONE) {
428 		/* XXX no %p in stand/printf.c.  Cast to quiet gcc -Wall. */
429 		printf("npxintr: p = %lx, curproc = %lx, npx_type = %d\n",
430 		       (u_long) p, (u_long) curproc, npx_type);
431 		panic("npxintr from nowhere");
432 	}
433 	/*
434 	 * Clear the interrupt latch.
435 	 */
436 	outb(0xf0, 0);
437 	/*
438 	 * If we're saving, ignore the interrupt.  The FPU will happily
439 	 * generate another one when we restore the state later.
440 	 */
441 	if (ci->ci_fpsaving)
442 		return (1);
443 
444 #ifdef DIAGNOSTIC
445 	/*
446 	 * At this point, fpcurproc should be curproc.  If it wasn't, the TS
447 	 * bit should be set, and we should have gotten a DNA exception.
448 	 */
449 	if (p != curproc)
450 		panic("npxintr: wrong process");
451 #endif
452 
453 	/*
454 	 * Find the address of fpcurproc's saved FPU state.  (Given the
455 	 * invariant above, this is always the one in curpcb.)
456 	 */
457 	addr = &p->p_addr->u_pcb.pcb_savefpu;
458 	/*
459 	 * Save state.  This does an implied fninit.  It had better not halt
460 	 * the cpu or we'll hang.
461 	 */
462 	fpu_save(addr);
463 	fwait();
464 	/*
465 	 * Restore control word (was clobbered by fpu_save).
466 	 */
467 	if (i386_use_fxsave) {
468 		fldcw(&addr->sv_xmm.sv_env.en_cw);
469 		/*
470 		 * FNINIT doesn't affect MXCSR or the XMM registers;
471 		 * no need to re-load MXCSR here.
472 		 */
473 	} else
474 		fldcw(&addr->sv_87.sv_env.en_cw);
475 	fwait();
476 	/*
477 	 * Remember the exception status word and tag word.  The current
478 	 * (almost fninit'ed) fpu state is in the fpu and the exception
479 	 * state just saved will soon be junk.  However, the implied fninit
480 	 * doesn't change the error pointers or register contents, and we
481 	 * preserved the control word and will copy the status and tag
482 	 * words, so the complete exception state can be recovered.
483 	 */
484 	if (i386_use_fxsave) {
485 	        addr->sv_xmm.sv_ex_sw = addr->sv_xmm.sv_env.en_sw;
486 	        addr->sv_xmm.sv_ex_tw = addr->sv_xmm.sv_env.en_tw;
487 	} else {
488 	        addr->sv_87.sv_ex_sw = addr->sv_87.sv_env.en_sw;
489 	        addr->sv_87.sv_ex_tw = addr->sv_87.sv_env.en_tw;
490 	}
491 
492 	/*
493 	 * Pass exception to process.  If it's the current process, try to do
494 	 * it immediately.
495 	 */
496 	if (p == curproc && USERMODE(frame->if_cs, frame->if_eflags)) {
497 		/*
498 		 * Interrupt is essentially a trap, so we can afford to call
499 		 * the SIGFPE handler (if any) as soon as the interrupt
500 		 * returns.
501 		 *
502 		 * XXX little or nothing is gained from this, and plenty is
503 		 * lost - the interrupt frame has to contain the trap frame
504 		 * (this is otherwise only necessary for the rescheduling trap
505 		 * in doreti, and the frame for that could easily be set up
506 		 * just before it is used).
507 		 */
508 		p->p_md.md_regs = (struct trapframe *)&frame->if_fs;
509 
510 		/*
511 		 * Encode the appropriate code for detailed information on
512 		 * this exception.
513 		 */
514 		if (i386_use_fxsave)
515 			code = x86fpflags_to_siginfo(addr->sv_xmm.sv_ex_sw);
516 		else
517 			code = x86fpflags_to_siginfo(addr->sv_87.sv_ex_sw);
518 		sv.sival_int = frame->if_eip;
519 		trapsignal(p, SIGFPE, T_ARITHTRAP, code, sv);
520 	} else {
521 		/*
522 		 * Nested interrupt.  These losers occur when:
523 		 *	o an IRQ13 is bogusly generated at a bogus time, e.g.:
524 		 *		o immediately after an fnsave or frstor of an
525 		 *		  error state.
526 		 *		o a couple of 386 instructions after
527 		 *		  "fstpl _memvar" causes a stack overflow.
528 		 *	  These are especially nasty when combined with a
529 		 *	  trace trap.
530 		 *	o an IRQ13 occurs at the same time as another higher-
531 		 *	  priority interrupt.
532 		 *
533 		 * Treat them like a true async interrupt.
534 		 */
535 		KERNEL_LOCK();
536 		psignal(p, SIGFPE);
537 		KERNEL_UNLOCK();
538 	}
539 
540 	return (1);
541 }
542 
543 void
npxtrap(struct trapframe * frame)544 npxtrap(struct trapframe *frame)
545 {
546 	struct proc *p = curcpu()->ci_fpcurproc;
547 	union savefpu *addr = &p->p_addr->u_pcb.pcb_savefpu;
548 	u_int32_t mxcsr, statbits;
549 	int code;
550 	union sigval sv;
551 
552 #ifdef DIAGNOSTIC
553 	/*
554 	 * At this point, fpcurproc should be curproc.  If it wasn't, the TS
555 	 * bit should be set, and we should have gotten a DNA exception.
556 	 */
557 	if (p != curproc)
558 		panic("npxtrap: wrong process");
559 #endif
560 
561 	fxsave(&addr->sv_xmm);
562 	mxcsr = addr->sv_xmm.sv_env.en_mxcsr;
563 	statbits = mxcsr;
564 	mxcsr &= ~0x3f;
565 	ldmxcsr(&mxcsr);
566 	addr->sv_xmm.sv_ex_sw = addr->sv_xmm.sv_env.en_sw;
567 	addr->sv_xmm.sv_ex_tw = addr->sv_xmm.sv_env.en_tw;
568 	code = x86fpflags_to_siginfo(statbits);
569 	sv.sival_int = frame->tf_eip;
570 	trapsignal(p, SIGFPE, frame->tf_err, code, sv);
571 }
572 
573 static int
x86fpflags_to_siginfo(u_int32_t flags)574 x86fpflags_to_siginfo(u_int32_t flags)
575 {
576         int i;
577         static int x86fp_siginfo_table[] = {
578                 FPE_FLTINV, /* bit 0 - invalid operation */
579                 FPE_FLTRES, /* bit 1 - denormal operand */
580                 FPE_FLTDIV, /* bit 2 - divide by zero   */
581                 FPE_FLTOVF, /* bit 3 - fp overflow      */
582                 FPE_FLTUND, /* bit 4 - fp underflow     */
583                 FPE_FLTRES, /* bit 5 - fp precision     */
584                 FPE_FLTINV, /* bit 6 - stack fault      */
585         };
586 
587         for (i=0;i < sizeof(x86fp_siginfo_table)/sizeof(int); i++) {
588                 if (flags & (1 << i))
589                         return (x86fp_siginfo_table[i]);
590         }
591         /* punt if flags not set */
592         return (FPE_FLTINV);
593 }
594 
595 /*
596  * Implement device not available (DNA) exception
597  *
598  * If we were the last process to use the FPU, we can simply return.
599  * Otherwise, we save the previous state, if necessary, and restore our last
600  * saved state.
601  */
602 int
npxdna_xmm(struct cpu_info * ci)603 npxdna_xmm(struct cpu_info *ci)
604 {
605 	union savefpu *sfp;
606 	struct proc *p;
607 	int s;
608 
609 	if (ci->ci_fpsaving) {
610 		printf("recursive npx trap; cr0=%x\n", rcr0());
611 		return (0);
612 	}
613 
614 	s = splipi();		/* lock out IPI's while we clean house.. */
615 
616 #ifdef MULTIPROCESSOR
617 	p = ci->ci_curproc;
618 #else
619 	p = curproc;
620 #endif
621 
622 	IPRINTF(("%s: dna for %lx%s\n", ci->ci_dev->dv_xname, (u_long)p,
623 	    (p->p_md.md_flags & MDP_USEDFPU) ? " (used fpu)" : ""));
624 
625 	/*
626 	 * XXX should have a fast-path here when no save/restore is necessary
627 	 */
628 	/*
629 	 * Initialize the FPU state to clear any exceptions.  If someone else
630 	 * was using the FPU, save their state (which does an implicit
631 	 * initialization).
632 	 */
633 	if (ci->ci_fpcurproc != NULL) {
634 		IPRINTF(("%s: fp save %lx\n", ci->ci_dev->dv_xname,
635 		    (u_long)ci->ci_fpcurproc));
636 		npxsave_cpu(ci, ci->ci_fpcurproc != &proc0);
637 	} else {
638 		clts();
639 		IPRINTF(("%s: fp init\n", ci->ci_dev->dv_xname));
640 		fninit();
641 		fwait();
642 		stts();
643 	}
644 	splx(s);
645 
646 	IPRINTF(("%s: done saving\n", ci->ci_dev->dv_xname));
647 	KDASSERT(ci->ci_fpcurproc == NULL);
648 #ifndef MULTIPROCESSOR
649 	KDASSERT(p->p_addr->u_pcb.pcb_fpcpu == NULL);
650 #else
651 	if (p->p_addr->u_pcb.pcb_fpcpu != NULL)
652 		npxsave_proc(p, 1);
653 #endif
654 	p->p_addr->u_pcb.pcb_cr0 &= ~CR0_TS;
655 	clts();
656 	s = splipi();
657 	ci->ci_fpcurproc = p;
658 	p->p_addr->u_pcb.pcb_fpcpu = ci;
659 	splx(s);
660 	uvmexp.fpswtch++;
661 
662 	sfp = &p->p_addr->u_pcb.pcb_savefpu;
663 
664 	if ((p->p_md.md_flags & MDP_USEDFPU) == 0) {
665 		bzero(&sfp->sv_xmm, sizeof(sfp->sv_xmm));
666 		sfp->sv_xmm.sv_env.en_cw = __INITIAL_NPXCW__;
667 		sfp->sv_xmm.sv_env.en_mxcsr = __INITIAL_MXCSR__;
668 		fxrstor(&sfp->sv_xmm);
669 		p->p_md.md_flags |= MDP_USEDFPU;
670 	} else {
671 		static double	zero = 0.0;
672 
673 		/*
674 		 * amd fpu does not restore fip, fdp, fop on fxrstor
675 		 * thus leaking other process's execution history.
676 		 */
677 		fnclex();
678 		__asm volatile("ffree %%st(7)\n\tfldl %0" : : "m" (zero));
679 		fxrstor(&sfp->sv_xmm);
680 	}
681 
682 	return (1);
683 }
684 
685 int
npxdna_s87(struct cpu_info * ci)686 npxdna_s87(struct cpu_info *ci)
687 {
688 	union savefpu *sfp;
689 	struct proc *p;
690 	int s;
691 
692 	KDASSERT(i386_use_fxsave == 0);
693 
694 	if (ci->ci_fpsaving) {
695 		printf("recursive npx trap; cr0=%x\n", rcr0());
696 		return (0);
697 	}
698 
699 	s = splipi();		/* lock out IPI's while we clean house.. */
700 #ifdef MULTIPROCESSOR
701 	p = ci->ci_curproc;
702 #else
703 	p = curproc;
704 #endif
705 
706 	IPRINTF(("%s: dna for %lx%s\n", ci->ci_dev->dv_xname, (u_long)p,
707 	    (p->p_md.md_flags & MDP_USEDFPU) ? " (used fpu)" : ""));
708 
709 	/*
710 	 * If someone else was using our FPU, save their state (which does an
711 	 * implicit initialization); otherwise, initialize the FPU state to
712 	 * clear any exceptions.
713 	 */
714 	if (ci->ci_fpcurproc != NULL) {
715 		IPRINTF(("%s: fp save %lx\n", ci->ci_dev->dv_xname,
716 		    (u_long)ci->ci_fpcurproc));
717 		npxsave_cpu(ci, ci->ci_fpcurproc != &proc0);
718 	} else {
719 		clts();
720 		IPRINTF(("%s: fp init\n", ci->ci_dev->dv_xname));
721 		fninit();
722 		fwait();
723 		stts();
724 	}
725 	splx(s);
726 
727 	IPRINTF(("%s: done saving\n", ci->ci_dev->dv_xname));
728 	KDASSERT(ci->ci_fpcurproc == NULL);
729 #ifndef MULTIPROCESSOR
730 	KDASSERT(p->p_addr->u_pcb.pcb_fpcpu == NULL);
731 #else
732 	if (p->p_addr->u_pcb.pcb_fpcpu != NULL)
733 		npxsave_proc(p, 1);
734 #endif
735 	p->p_addr->u_pcb.pcb_cr0 &= ~CR0_TS;
736 	clts();
737 	s = splipi();
738 	ci->ci_fpcurproc = p;
739 	p->p_addr->u_pcb.pcb_fpcpu = ci;
740 	splx(s);
741 	uvmexp.fpswtch++;
742 
743 	sfp = &p->p_addr->u_pcb.pcb_savefpu;
744 
745 	if ((p->p_md.md_flags & MDP_USEDFPU) == 0) {
746 		bzero(&sfp->sv_87, sizeof(sfp->sv_87));
747 		sfp->sv_87.sv_env.en_cw = __INITIAL_NPXCW__;
748 		sfp->sv_87.sv_env.en_tw = 0xffff;
749 		frstor(&sfp->sv_87);
750 		p->p_md.md_flags |= MDP_USEDFPU;
751 	} else {
752 		/*
753 		 * The following frstor may cause an IRQ13 when the state being
754 		 * restored has a pending error.  The error will appear to have
755 		 * been triggered by the current (npx) user instruction even
756 		 * when that instruction is a no-wait instruction that should
757 		 * not trigger an error (e.g., fnclex).  On at least one 486
758 		 * system all of the no-wait instructions are broken the same
759 		 * as frstor, so our treatment does not amplify the breakage.
760 		 * On at least one 386/Cyrix 387 system, fnclex works correctly
761 		 * while frstor and fnsave are broken, so our treatment breaks
762 		 * fnclex if it is the first FPU instruction after a context
763 		 * switch.
764 		 */
765 		frstor(&sfp->sv_87);
766 	}
767 
768 	return (1);
769 }
770 
771 /*
772  * The FNSAVE instruction clears the FPU state.  Rather than reloading the FPU
773  * immediately, we clear fpcurproc and turn on CR0_TS to force a DNA and a
774  * reload of the FPU state the next time we try to use it.  This routine
775  * is only called when forking, core dumping, or debugging, or swapping,
776  * so the lazy reload at worst forces us to trap once per fork(), and at best
777  * saves us a reload once per fork().
778  */
779 void
npxsave_cpu(struct cpu_info * ci,int save)780 npxsave_cpu(struct cpu_info *ci, int save)
781 {
782 	struct proc *p;
783 	int s;
784 
785 	KDASSERT(ci == curcpu());
786 
787 	p = ci->ci_fpcurproc;
788 	if (p == NULL)
789 		return;
790 
791 	IPRINTF(("%s: fp cpu %s %lx\n", ci->ci_dev->dv_xname,
792 	    save ? "save" : "flush", (u_long)p));
793 
794 	if (save) {
795 #ifdef DIAGNOSTIC
796 		if (ci->ci_fpsaving != 0)
797 			panic("npxsave_cpu: recursive save!");
798 #endif
799 		 /*
800 		  * Set ci->ci_fpsaving, so that any pending exception will be
801 		  * thrown away.  (It will be caught again if/when the FPU
802 		  * state is restored.)
803 		  *
804 		  * XXX on i386 and earlier, this routine should always be
805 		  * called at spl0; if it might called with the NPX interrupt
806 		  * masked, it would be necessary to forcibly unmask the NPX
807 		  * interrupt so that it could succeed.
808 		  * XXX this is irrelevant on 486 and above (systems
809 		  * which report FP failures via traps rather than irq13).
810 		  * XXX punting for now..
811 		  */
812 		clts();
813 		ci->ci_fpsaving = 1;
814 		fpu_save(&p->p_addr->u_pcb.pcb_savefpu);
815 		ci->ci_fpsaving = 0;
816 		/* It is unclear if this is needed. */
817 		fwait();
818 	}
819 
820 	/*
821 	 * We set the TS bit in the saved CR0 for this process, so that it
822 	 * will get a DNA exception on any FPU instruction and force a reload.
823 	 */
824 	stts();
825 	p->p_addr->u_pcb.pcb_cr0 |= CR0_TS;
826 
827 	s = splipi();
828 	p->p_addr->u_pcb.pcb_fpcpu = NULL;
829 	ci->ci_fpcurproc = NULL;
830 	splx(s);
831 }
832 
833 /*
834  * Save p's FPU state, which may be on this processor or another processor.
835  */
836 void
npxsave_proc(struct proc * p,int save)837 npxsave_proc(struct proc *p, int save)
838 {
839 	struct cpu_info *ci = curcpu();
840 	struct cpu_info *oci;
841 
842 	KDASSERT(p->p_addr != NULL);
843 
844 	oci = p->p_addr->u_pcb.pcb_fpcpu;
845 	if (oci == NULL)
846 		return;
847 
848 	IPRINTF(("%s: fp proc %s %lx\n", ci->ci_dev->dv_xname,
849 	    save ? "save" : "flush", (u_long)p));
850 
851 #if defined(MULTIPROCESSOR)
852 	if (oci == ci) {
853 		int s = splipi();
854 		npxsave_cpu(ci, save);
855 		splx(s);
856 	} else {
857 		IPRINTF(("%s: fp ipi to %s %s %lx\n", ci->ci_dev->dv_xname,
858 		    oci->ci_dev->dv_xname, save ? "save" : "flush", (u_long)p));
859 
860 		oci->ci_fpsaveproc = p;
861 		i386_send_ipi(oci,
862 		    save ? I386_IPI_SYNCH_FPU : I386_IPI_FLUSH_FPU);
863 		while (p->p_addr->u_pcb.pcb_fpcpu != NULL)
864 			CPU_BUSY_CYCLE();
865 	}
866 #else
867 	KASSERT(ci->ci_fpcurproc == p);
868 	npxsave_cpu(ci, save);
869 #endif
870 }
871 
872 void
fpu_kernel_enter(void)873 fpu_kernel_enter(void)
874 {
875 	struct cpu_info	*ci = curcpu();
876 	uint32_t	 cw;
877 	int		 s;
878 
879 	/*
880 	 * Fast path.  If the kernel was using the FPU before, there
881 	 * is no work to do besides clearing TS.
882 	 */
883 	if (ci->ci_fpcurproc == &proc0) {
884 		clts();
885 		return;
886 	}
887 
888 	s = splipi();
889 
890 	if (ci->ci_fpcurproc != NULL) {
891 		npxsave_cpu(ci, 1);
892 		uvmexp.fpswtch++;
893 	}
894 
895 	/* Claim the FPU */
896 	ci->ci_fpcurproc = &proc0;
897 
898 	splx(s);
899 
900 	/* Disable DNA exceptions */
901 	clts();
902 
903 	/* Initialize the FPU */
904 	fninit();
905 	cw = __INITIAL_NPXCW__;
906 	fldcw(&cw);
907 	if (i386_has_sse || i386_has_sse2) {
908 		cw = __INITIAL_MXCSR__;
909 		ldmxcsr(&cw);
910 	}
911 }
912 
913 void
fpu_kernel_exit(void)914 fpu_kernel_exit(void)
915 {
916 	/* Enable DNA exceptions */
917 	stts();
918 }
919