xref: /freebsd/sys/amd64/amd64/fpu.c (revision 0c43d89a0d8e976ca494d4837f4c1f3734d2c300)
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. All advertising materials mentioning features or use of this software
15  *    must display the following acknowledgement:
16  *	This product includes software developed by the University of
17  *	California, Berkeley and its contributors.
18  * 4. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *	from: @(#)npx.c	7.2 (Berkeley) 5/12/91
35  *	$Id: npx.c,v 1.9 1994/06/11 05:17:15 davidg Exp $
36  */
37 
38 #include "npx.h"
39 #if NNPX > 0
40 
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/conf.h>
44 #include <sys/file.h>
45 #include <sys/proc.h>
46 #include <machine/cpu.h>
47 #include <machine/pcb.h>
48 #include <machine/trap.h>
49 #include <sys/ioctl.h>
50 #include <machine/specialreg.h>
51 #include <i386/isa/icu.h>
52 #include <i386/isa/isa_device.h>
53 #include <i386/isa/isa.h>
54 
55 /*
56  * 387 and 287 Numeric Coprocessor Extension (NPX) Driver.
57  */
58 
59 #ifdef	__GNUC__
60 
61 #define	disable_intr()		__asm("cli")
62 #define	enable_intr()		__asm("sti")
63 #define	fldcw(addr)		__asm("fldcw %0" : : "m" (*addr))
64 #define	fnclex()		__asm("fnclex")
65 #define	fninit()		__asm("fninit")
66 #define	fnsave(addr)		__asm("fnsave %0" : "=m" (*addr) : "0" (*addr))
67 #define	fnstcw(addr)		__asm("fnstcw %0" : "=m" (*addr) : "0" (*addr))
68 #define	fnstsw(addr)		__asm("fnstsw %0" : "=m" (*addr) : "0" (*addr))
69 #define	fp_divide_by_0()	__asm("fldz; fld1; fdiv %st,%st(1); fwait")
70 #define	frstor(addr)		__asm("frstor %0" : : "m" (*addr))
71 #define	fwait()			__asm("fwait")
72 #define	read_eflags()		({u_long ef; \
73 				  __asm("pushf; popl %0" : "=a" (ef)); \
74 				  ef; })
75 #define	start_emulating()	__asm("smsw %%ax; orb %0,%%al; lmsw %%ax" \
76 				      : : "n" (CR0_TS) : "ax")
77 #define	stop_emulating()	__asm("clts")
78 #define	write_eflags(ef)	__asm("pushl %0; popf" : : "a" ((u_long) ef))
79 
80 #else	/* not __GNUC__ */
81 
82 void	disable_intr	__P((void));
83 void	enable_intr	__P((void));
84 void	fldcw		__P((caddr_t addr));
85 void	fnclex		__P((void));
86 void	fninit		__P((void));
87 void	fnsave		__P((caddr_t addr));
88 void	fnstcw		__P((caddr_t addr));
89 void	fnstsw		__P((caddr_t addr));
90 void	fp_divide_by_0	__P((void));
91 void	frstor		__P((caddr_t addr));
92 void	fwait		__P((void));
93 u_long	read_eflags	__P((void));
94 void	start_emulating	__P((void));
95 void	stop_emulating	__P((void));
96 void	write_eflags	__P((u_long ef));
97 
98 #endif	/* __GNUC__ */
99 
100 typedef u_char bool_t;
101 
102 extern	struct gate_descriptor idt[];
103 
104 int	npxdna		__P((void));
105 void	npxexit		__P((struct proc *p));
106 void	npxinit		__P((u_int control));
107 void	npxintr		__P((struct intrframe frame));
108 void	npxsave		__P((struct save87 *addr));
109 static	int	npxattach	__P((struct isa_device *dvp));
110 static	int	npxprobe	__P((struct isa_device *dvp));
111 static	int	npxprobe1	__P((struct isa_device *dvp));
112 
113 struct	isa_driver npxdriver = {
114 	npxprobe, npxattach, "npx",
115 };
116 
117 u_int	npx0_imask;
118 struct proc	*npxproc;
119 
120 static	bool_t			npx_ex16;
121 static	bool_t			npx_exists;
122 static	struct gate_descriptor	npx_idt_probeintr;
123 static	int			npx_intrno;
124 static	volatile u_int		npx_intrs_while_probing;
125 static	bool_t			npx_irq13;
126 static	volatile u_int		npx_traps_while_probing;
127 
128 /*
129  * Special interrupt handlers.  Someday intr0-intr15 will be used to count
130  * interrupts.  We'll still need a special exception 16 handler.  The busy
131  * latch stuff in probintr() can be moved to npxprobe().
132  */
133 void probeintr(void);
134 asm
135 ("
136 	.text
137 _probeintr:
138 	ss
139 	incl	_npx_intrs_while_probing
140 	pushl	%eax
141 	movb	$0x20,%al	# EOI (asm in strings loses cpp features)
142 	outb	%al,$0xa0	# IO_ICU2
143 	outb	%al,$0x20	#IO_ICU1
144 	movb	$0,%al
145 	outb	%al,$0xf0	# clear BUSY# latch
146 	popl	%eax
147 	iret
148 ");
149 
150 void probetrap(void);
151 asm
152 ("
153 	.text
154 _probetrap:
155 	ss
156 	incl	_npx_traps_while_probing
157 	fnclex
158 	iret
159 ");
160 
161 /*
162  * Probe routine.  Initialize cr0 to give correct behaviour for [f]wait
163  * whether the device exists or not (XXX should be elsewhere).  Set flags
164  * to tell npxattach() what to do.  Modify device struct if npx doesn't
165  * need to use interrupts.  Return 1 if device exists.
166  */
167 static int
168 npxprobe(dvp)
169 	struct isa_device *dvp;
170 {
171 	int	result;
172 	u_long	save_eflags;
173 	u_char	save_icu1_mask;
174 	u_char	save_icu2_mask;
175 	struct	gate_descriptor save_idt_npxintr;
176 	struct	gate_descriptor save_idt_npxtrap;
177 	/*
178 	 * This routine is now just a wrapper for npxprobe1(), to install
179 	 * special npx interrupt and trap handlers, to enable npx interrupts
180 	 * and to disable other interrupts.  Someday isa_configure() will
181 	 * install suitable handlers and run with interrupts enabled so we
182 	 * won't need to do so much here.
183 	 */
184 	npx_intrno = NRSVIDT + ffs(dvp->id_irq) - 1;
185 	save_eflags = read_eflags();
186 	disable_intr();
187 	save_icu1_mask = inb(IO_ICU1 + 1);
188 	save_icu2_mask = inb(IO_ICU2 + 1);
189 	save_idt_npxintr = idt[npx_intrno];
190 	save_idt_npxtrap = idt[16];
191 	outb(IO_ICU1 + 1, ~(IRQ_SLAVE | dvp->id_irq));
192 	outb(IO_ICU2 + 1, ~(dvp->id_irq >> 8));
193 	setidt(16, probetrap, SDT_SYS386TGT, SEL_KPL);
194 	setidt(npx_intrno, probeintr, SDT_SYS386IGT, SEL_KPL);
195 	npx_idt_probeintr = idt[npx_intrno];
196 	enable_intr();
197 	result = npxprobe1(dvp);
198 	disable_intr();
199 	outb(IO_ICU1 + 1, save_icu1_mask);
200 	outb(IO_ICU2 + 1, save_icu2_mask);
201 	idt[npx_intrno] = save_idt_npxintr;
202 	idt[16] = save_idt_npxtrap;
203 	write_eflags(save_eflags);
204 	return (result);
205 }
206 
207 static int
208 npxprobe1(dvp)
209 	struct isa_device *dvp;
210 {
211 	int control;
212 	int status;
213 #ifdef lint
214 	npxintr();
215 #endif
216 	/*
217 	 * Partially reset the coprocessor, if any.  Some BIOS's don't reset
218 	 * it after a warm boot.
219 	 */
220 	outb(0xf1, 0);		/* full reset on some systems, NOP on others */
221 	outb(0xf0, 0);		/* clear BUSY# latch */
222 	/*
223 	 * Prepare to trap all ESC (i.e., NPX) instructions and all WAIT
224 	 * instructions.  We must set the CR0_MP bit and use the CR0_TS
225 	 * bit to control the trap, because setting the CR0_EM bit does
226 	 * not cause WAIT instructions to trap.  It's important to trap
227 	 * WAIT instructions - otherwise the "wait" variants of no-wait
228 	 * control instructions would degenerate to the "no-wait" variants
229 	 * after FP context switches but work correctly otherwise.  It's
230 	 * particularly important to trap WAITs when there is no NPX -
231 	 * otherwise the "wait" variants would always degenerate.
232 	 *
233 	 * Try setting CR0_NE to get correct error reporting on 486DX's.
234 	 * Setting it should fail or do nothing on lesser processors.
235 	 */
236 	load_cr0(rcr0() | CR0_MP | CR0_NE);
237 	/*
238 	 * But don't trap while we're probing.
239 	 */
240 	stop_emulating();
241 	/*
242 	 * Finish resetting the coprocessor, if any.  If there is an error
243 	 * pending, then we may get a bogus IRQ13, but probeintr() will handle
244 	 * it OK.  Bogus halts have never been observed, but we enabled
245 	 * IRQ13 and cleared the BUSY# latch early to handle them anyway.
246 	 */
247 	fninit();
248 	DELAY(1000);		/* wait for any IRQ13 (fwait might hang) */
249 #ifdef DIAGNOSTIC
250 	if (npx_intrs_while_probing != 0)
251 		printf("fninit caused %u bogus npx interrupt(s)\n",
252 		       npx_intrs_while_probing);
253 	if (npx_traps_while_probing != 0)
254 		printf("fninit caused %u bogus npx trap(s)\n",
255 		       npx_traps_while_probing);
256 #endif
257 	/*
258 	 * Check for a status of mostly zero.
259 	 */
260 	status = 0x5a5a;
261 	fnstsw(&status);
262 	if ((status & 0xb8ff) == 0) {
263 		/*
264 		 * Good, now check for a proper control word.
265 		 */
266 		control = 0x5a5a;
267 		fnstcw(&control);
268 		if ((control & 0x1f3f) == 0x033f) {
269 			npx_exists = 1;
270 			/*
271 			 * We have an npx, now divide by 0 to see if exception
272 			 * 16 works.
273 			 */
274 			control &= ~(1 << 2);	/* enable divide by 0 trap */
275 			fldcw(&control);
276 			npx_traps_while_probing = npx_intrs_while_probing = 0;
277 			fp_divide_by_0();
278 			if (npx_traps_while_probing != 0) {
279 				/*
280 				 * Good, exception 16 works.
281 				 */
282 				npx_ex16 = 1;
283 				dvp->id_irq = 0;	/* zap the interrupt */
284 				/*
285 				 * special return value to flag that we do not
286 				 * actually use any I/O registers
287 				 */
288 				return (-1);
289 			}
290 			if (npx_intrs_while_probing != 0) {
291 				/*
292 				 * Bad, we are stuck with IRQ13.
293 				 */
294 				npx_irq13 = 1;
295 				npx0_imask = dvp->id_irq;	/* npxattach too late */
296 				return (IO_NPXSIZE);
297 			}
298 			/*
299 			 * Worse, even IRQ13 is broken.  Use emulator.
300 			 */
301 		}
302 	}
303 	/*
304 	 * Probe failed, but we want to get to npxattach to initialize the
305 	 * emulator and say that it has been installed.  XXX handle devices
306 	 * that aren't really devices better.
307 	 */
308 	dvp->id_irq = 0;
309 	/*
310 	 * special return value to flag that we do not
311 	 * actually use any I/O registers
312 	 */
313 	return (-1);
314 }
315 
316 /*
317  * Attach routine - announce which it is, and wire into system
318  */
319 int
320 npxattach(dvp)
321 	struct isa_device *dvp;
322 {
323 	if (!npx_ex16 && !npx_irq13) {
324 		if (npx_exists) {
325 			printf("npx%d: Error reporting broken, using 387 emulator\n",dvp->id_unit);
326 			npx_exists = 0;
327 		} else {
328 			printf("npx%d: 387 Emulator\n",dvp->id_unit);
329 		}
330 	}
331 	npxinit(__INITIAL_NPXCW__);
332 	return (1);		/* XXX unused */
333 }
334 
335 /*
336  * Initialize floating point unit.
337  */
338 void
339 npxinit(control)
340 	u_int control;
341 {
342 	struct save87 dummy;
343 
344 	if (!npx_exists)
345 		return;
346 	/*
347 	 * fninit has the same h/w bugs as fnsave.  Use the detoxified
348 	 * fnsave to throw away any junk in the fpu.  fnsave initializes
349 	 * the fpu and sets npxproc = NULL as important side effects.
350 	 */
351 	npxsave(&dummy);
352 	stop_emulating();
353 	fldcw(&control);
354 	if (curpcb != NULL)
355 		fnsave(&curpcb->pcb_savefpu);
356 	start_emulating();
357 }
358 
359 /*
360  * Free coprocessor (if we have it).
361  */
362 void
363 npxexit(p)
364 	struct proc *p;
365 {
366 
367 	if (p == npxproc) {
368 		start_emulating();
369 		npxproc = NULL;
370 	}
371 }
372 
373 /*
374  * Record the FPU state and reinitialize it all except for the control word.
375  * Then generate a SIGFPE.
376  *
377  * Reinitializing the state allows naive SIGFPE handlers to longjmp without
378  * doing any fixups.
379  *
380  * XXX there is currently no way to pass the full error state to signal
381  * handlers, and if this is a nested interrupt there is no way to pass even
382  * a status code!  So there is no way to have a non-naive SIGFPE handler.  At
383  * best a handler could do an fninit followed by an fldcw of a static value.
384  * fnclex would be of little use because it would leave junk on the FPU stack.
385  * Returning from the handler would be even less safe than usual because
386  * IRQ13 exception handling makes exceptions even less precise than usual.
387  */
388 void
389 npxintr(frame)
390 	struct intrframe frame;
391 {
392 	int code;
393 
394 	if (npxproc == NULL || !npx_exists) {
395 		/* XXX no %p in stand/printf.c.  Cast to quiet gcc -Wall. */
396 		printf("npxintr: npxproc = %lx, curproc = %lx, npx_exists = %d\n",
397 		       (u_long) npxproc, (u_long) curproc, npx_exists);
398 		panic("npxintr from nowhere");
399 	}
400 	if (npxproc != curproc) {
401 		printf("npxintr: npxproc = %lx, curproc = %lx, npx_exists = %d\n",
402 		       (u_long) npxproc, (u_long) curproc, npx_exists);
403 		panic("npxintr from non-current process");
404 	}
405 	/*
406 	 * Save state.  This does an implied fninit.  It had better not halt
407 	 * the cpu or we'll hang.
408 	 */
409 	outb(0xf0, 0);
410 	fnsave(&curpcb->pcb_savefpu);
411 	fwait();
412 	/*
413 	 * Restore control word (was clobbered by fnsave).
414 	 */
415 	fldcw(&curpcb->pcb_savefpu.sv_env.en_cw);
416 	fwait();
417 	/*
418 	 * Remember the exception status word and tag word.  The current
419 	 * (almost fninit'ed) fpu state is in the fpu and the exception
420 	 * state just saved will soon be junk.  However, the implied fninit
421 	 * doesn't change the error pointers or register contents, and we
422 	 * preserved the control word and will copy the status and tag
423 	 * words, so the complete exception state can be recovered.
424 	 */
425 	curpcb->pcb_savefpu.sv_ex_sw = curpcb->pcb_savefpu.sv_env.en_sw;
426 	curpcb->pcb_savefpu.sv_ex_tw = curpcb->pcb_savefpu.sv_env.en_tw;
427 
428 	/*
429 	 * Pass exception to process.
430 	 */
431 	if (ISPL(frame.if_cs) == SEL_UPL) {
432 		/*
433 		 * Interrupt is essentially a trap, so we can afford to call
434 		 * the SIGFPE handler (if any) as soon as the interrupt
435 		 * returns.
436 		 *
437 		 * XXX little or nothing is gained from this, and plenty is
438 		 * lost - the interrupt frame has to contain the trap frame
439 		 * (this is otherwise only necessary for the rescheduling trap
440 		 * in doreti, and the frame for that could easily be set up
441 		 * just before it is used).
442 		 */
443 		curproc->p_md.md_regs = (int *)&frame.if_es;
444 #ifdef notyet
445 		/*
446 		 * Encode the appropriate code for detailed information on
447 		 * this exception.
448 		 */
449 		code = XXX_ENCODE(curpcb->pcb_savefpu.sv_ex_sw);
450 #else
451 		code = 0;	/* XXX */
452 #endif
453 		trapsignal(curproc, SIGFPE, code);
454 	} else {
455 		/*
456 		 * Nested interrupt.  These losers occur when:
457 		 *	o an IRQ13 is bogusly generated at a bogus time, e.g.:
458 		 *		o immediately after an fnsave or frstor of an
459 		 *		  error state.
460 		 *		o a couple of 386 instructions after
461 		 *		  "fstpl _memvar" causes a stack overflow.
462 		 *	  These are especially nasty when combined with a
463 		 *	  trace trap.
464 		 *	o an IRQ13 occurs at the same time as another higher-
465 		 *	  priority interrupt.
466 		 *
467 		 * Treat them like a true async interrupt.
468 		 */
469 		psignal(npxproc, SIGFPE);
470 	}
471 }
472 
473 /*
474  * Implement device not available (DNA) exception
475  *
476  * It would be better to switch FP context here (only).  This would require
477  * saving the state in the proc table instead of in the pcb.
478  */
479 int
480 npxdna()
481 {
482 	if (!npx_exists)
483 		return (0);
484 	if (npxproc != NULL) {
485 		printf("npxdna: npxproc = %lx, curproc = %lx\n",
486 		       (u_long) npxproc, (u_long) curproc);
487 		panic("npxdna");
488 	}
489 	stop_emulating();
490 	/*
491 	 * Record new context early in case frstor causes an IRQ13.
492 	 */
493 	npxproc = curproc;
494 	/*
495 	 * The following frstor may cause an IRQ13 when the state being
496 	 * restored has a pending error.  The error will appear to have been
497 	 * triggered by the current (npx) user instruction even when that
498 	 * instruction is a no-wait instruction that should not trigger an
499 	 * error (e.g., fnclex).  On at least one 486 system all of the
500 	 * no-wait instructions are broken the same as frstor, so our
501 	 * treatment does not amplify the breakage.  On at least one
502 	 * 386/Cyrix 387 system, fnclex works correctly while frstor and
503 	 * fnsave are broken, so our treatment breaks fnclex if it is the
504 	 * first FPU instruction after a context switch.
505 	 */
506 	frstor(&curpcb->pcb_savefpu);
507 
508 	return (1);
509 }
510 
511 /*
512  * Wrapper for fnsave instruction to handle h/w bugs.  If there is an error
513  * pending, then fnsave generates a bogus IRQ13 on some systems.  Force
514  * any IRQ13 to be handled immediately, and then ignore it.  This routine is
515  * often called at splhigh so it must not use many system services.  In
516  * particular, it's much easier to install a special handler than to
517  * guarantee that it's safe to use npxintr() and its supporting code.
518  */
519 void
520 npxsave(addr)
521 	struct save87 *addr;
522 {
523 	u_char	icu1_mask;
524 	u_char	icu2_mask;
525 	u_char	old_icu1_mask;
526 	u_char	old_icu2_mask;
527 	struct gate_descriptor	save_idt_npxintr;
528 
529 	disable_intr();
530 	old_icu1_mask = inb(IO_ICU1 + 1);
531 	old_icu2_mask = inb(IO_ICU2 + 1);
532 	save_idt_npxintr = idt[npx_intrno];
533 	outb(IO_ICU1 + 1, old_icu1_mask & ~(IRQ_SLAVE | npx0_imask));
534 	outb(IO_ICU2 + 1, old_icu2_mask & ~(npx0_imask >> 8));
535 	idt[npx_intrno] = npx_idt_probeintr;
536 	enable_intr();
537 	stop_emulating();
538 	fnsave(addr);
539 	fwait();
540 	start_emulating();
541 	npxproc = NULL;
542 	disable_intr();
543 	icu1_mask = inb(IO_ICU1 + 1);	/* masks may have changed */
544 	icu2_mask = inb(IO_ICU2 + 1);
545 	outb(IO_ICU1 + 1,
546 	     (icu1_mask & ~npx0_imask) | (old_icu1_mask & npx0_imask));
547 	outb(IO_ICU2 + 1,
548 	     (icu2_mask & ~(npx0_imask >> 8))
549 	     | (old_icu2_mask & (npx0_imask >> 8)));
550 	idt[npx_intrno] = save_idt_npxintr;
551 	enable_intr();		/* back to usual state */
552 }
553 
554 #endif /* NNPX > 0 */
555