xref: /freebsd/sys/amd64/amd64/fpu.c (revision a316b26e50bbed7cf655fbba726ab87d8ab7599d)
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 1.17 1994/11/14 14:59:06 bde 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 <sys/devconf.h>
47 #include <sys/ioctl.h>
48 #include <sys/syslog.h>
49 #include <sys/signalvar.h>
50 
51 #include <machine/cpu.h>
52 #include <machine/pcb.h>
53 #include <machine/trap.h>
54 #include <machine/clock.h>
55 #include <machine/specialreg.h>
56 
57 #include <i386/isa/icu.h>
58 #include <i386/isa/isa_device.h>
59 #include <i386/isa/isa.h>
60 
61 /*
62  * 387 and 287 Numeric Coprocessor Extension (NPX) Driver.
63  */
64 
65 #ifdef	__GNUC__
66 
67 #define	fldcw(addr)		__asm("fldcw %0" : : "m" (*(addr)))
68 #define	fnclex()		__asm("fnclex")
69 #define	fninit()		__asm("fninit")
70 #define	fnop()			__asm("fnop")
71 #define	fnsave(addr)		__asm("fnsave %0" : "=m" (*(addr)))
72 #define	fnstcw(addr)		__asm("fnstcw %0" : "=m" (*(addr)))
73 #define	fnstsw(addr)		__asm("fnstsw %0" : "=m" (*(addr)))
74 #define	fp_divide_by_0()	__asm("fldz; fld1; fdiv %st,%st(1); fnop")
75 #define	frstor(addr)		__asm("frstor %0" : : "m" (*(addr)))
76 #define	start_emulating()	__asm("smsw %%ax; orb %0,%%al; lmsw %%ax" \
77 				      : : "n" (CR0_TS) : "ax")
78 #define	stop_emulating()	__asm("clts")
79 
80 #else	/* not __GNUC__ */
81 
82 void	fldcw		__P((caddr_t addr));
83 void	fnclex		__P((void));
84 void	fninit		__P((void));
85 void	fnop		__P((void));
86 void	fnsave		__P((caddr_t addr));
87 void	fnstcw		__P((caddr_t addr));
88 void	fnstsw		__P((caddr_t addr));
89 void	fp_divide_by_0	__P((void));
90 void	frstor		__P((caddr_t addr));
91 void	start_emulating	__P((void));
92 void	stop_emulating	__P((void));
93 
94 #endif	/* __GNUC__ */
95 
96 typedef u_char bool_t;
97 
98 static	int	npxattach	__P((struct isa_device *dvp));
99 static	int	npxprobe	__P((struct isa_device *dvp));
100 static	int	npxprobe1	__P((struct isa_device *dvp));
101 
102 struct	isa_driver npxdriver = {
103 	npxprobe, npxattach, "npx",
104 };
105 
106 int	hw_float;		/* XXX currently just alias for npx_exists */
107 u_int	npx0_imask = SWI_CLOCK_MASK;
108 struct proc	*npxproc;
109 
110 static	bool_t			npx_ex16;
111 static	bool_t			npx_exists;
112 static	struct gate_descriptor	npx_idt_probeintr;
113 static	int			npx_intrno;
114 static	volatile u_int		npx_intrs_while_probing;
115 static	bool_t			npx_irq13;
116 static	volatile u_int		npx_traps_while_probing;
117 
118 /*
119  * Special interrupt handlers.  Someday intr0-intr15 will be used to count
120  * interrupts.  We'll still need a special exception 16 handler.  The busy
121  * latch stuff in probeintr() can be moved to npxprobe().
122  */
123 inthand_t probeintr;
124 asm
125 ("
126 	.text
127 _probeintr:
128 	ss
129 	incl	_npx_intrs_while_probing
130 	pushl	%eax
131 	movb	$0x20,%al	# EOI (asm in strings loses cpp features)
132 	outb	%al,$0xa0	# IO_ICU2
133 	outb	%al,$0x20	# IO_ICU1
134 	movb	$0,%al
135 	outb	%al,$0xf0	# clear BUSY# latch
136 	popl	%eax
137 	iret
138 ");
139 
140 inthand_t probetrap;
141 asm
142 ("
143 	.text
144 _probetrap:
145 	ss
146 	incl	_npx_traps_while_probing
147 	fnclex
148 	iret
149 ");
150 
151 /*
152  * Probe routine.  Initialize cr0 to give correct behaviour for [f]wait
153  * whether the device exists or not (XXX should be elsewhere).  Set flags
154  * to tell npxattach() what to do.  Modify device struct if npx doesn't
155  * need to use interrupts.  Return 1 if device exists.
156  */
157 static int
158 npxprobe(dvp)
159 	struct isa_device *dvp;
160 {
161 	int	result;
162 	u_long	save_eflags;
163 	u_char	save_icu1_mask;
164 	u_char	save_icu2_mask;
165 	struct	gate_descriptor save_idt_npxintr;
166 	struct	gate_descriptor save_idt_npxtrap;
167 	/*
168 	 * This routine is now just a wrapper for npxprobe1(), to install
169 	 * special npx interrupt and trap handlers, to enable npx interrupts
170 	 * and to disable other interrupts.  Someday isa_configure() will
171 	 * install suitable handlers and run with interrupts enabled so we
172 	 * won't need to do so much here.
173 	 */
174 	npx_intrno = NRSVIDT + ffs(dvp->id_irq) - 1;
175 	save_eflags = read_eflags();
176 	disable_intr();
177 	save_icu1_mask = inb(IO_ICU1 + 1);
178 	save_icu2_mask = inb(IO_ICU2 + 1);
179 	save_idt_npxintr = idt[npx_intrno];
180 	save_idt_npxtrap = idt[16];
181 	outb(IO_ICU1 + 1, ~(IRQ_SLAVE | dvp->id_irq));
182 	outb(IO_ICU2 + 1, ~(dvp->id_irq >> 8));
183 	setidt(16, probetrap, SDT_SYS386TGT, SEL_KPL);
184 	setidt(npx_intrno, probeintr, SDT_SYS386IGT, SEL_KPL);
185 	npx_idt_probeintr = idt[npx_intrno];
186 	enable_intr();
187 	result = npxprobe1(dvp);
188 	disable_intr();
189 	outb(IO_ICU1 + 1, save_icu1_mask);
190 	outb(IO_ICU2 + 1, save_icu2_mask);
191 	idt[npx_intrno] = save_idt_npxintr;
192 	idt[16] = save_idt_npxtrap;
193 	write_eflags(save_eflags);
194 	return (result);
195 }
196 
197 static int
198 npxprobe1(dvp)
199 	struct isa_device *dvp;
200 {
201 	u_short control;
202 	u_short status;
203 
204 	/*
205 	 * Partially reset the coprocessor, if any.  Some BIOS's don't reset
206 	 * it after a warm boot.
207 	 */
208 	outb(0xf1, 0);		/* full reset on some systems, NOP on others */
209 	outb(0xf0, 0);		/* clear BUSY# latch */
210 	/*
211 	 * Prepare to trap all ESC (i.e., NPX) instructions and all WAIT
212 	 * instructions.  We must set the CR0_MP bit and use the CR0_TS
213 	 * bit to control the trap, because setting the CR0_EM bit does
214 	 * not cause WAIT instructions to trap.  It's important to trap
215 	 * WAIT instructions - otherwise the "wait" variants of no-wait
216 	 * control instructions would degenerate to the "no-wait" variants
217 	 * after FP context switches but work correctly otherwise.  It's
218 	 * particularly important to trap WAITs when there is no NPX -
219 	 * otherwise the "wait" variants would always degenerate.
220 	 *
221 	 * Try setting CR0_NE to get correct error reporting on 486DX's.
222 	 * Setting it should fail or do nothing on lesser processors.
223 	 */
224 	load_cr0(rcr0() | CR0_MP | CR0_NE);
225 	/*
226 	 * But don't trap while we're probing.
227 	 */
228 	stop_emulating();
229 	/*
230 	 * Finish resetting the coprocessor, if any.  If there is an error
231 	 * pending, then we may get a bogus IRQ13, but probeintr() will handle
232 	 * it OK.  Bogus halts have never been observed, but we enabled
233 	 * IRQ13 and cleared the BUSY# latch early to handle them anyway.
234 	 */
235 	fninit();
236 	fnop();			/* wait for fninit (fwait might hang) */
237 	DELAY(1000);		/* wait for any IRQ13 */
238 #ifdef DIAGNOSTIC
239 	if (npx_intrs_while_probing != 0)
240 		printf("fninit caused %u bogus npx interrupt(s)\n",
241 		       npx_intrs_while_probing);
242 	if (npx_traps_while_probing != 0)
243 		printf("fninit caused %u bogus npx trap(s)\n",
244 		       npx_traps_while_probing);
245 #endif
246 	/*
247 	 * Check for a status of mostly zero.
248 	 */
249 	status = 0x5a5a;
250 	fnstsw(&status);
251 	if ((status & 0xb8ff) == 0) {
252 		/*
253 		 * Good, now check for a proper control word.
254 		 */
255 		control = 0x5a5a;
256 		fnstcw(&control);
257 		if ((control & 0x1f3f) == 0x033f) {
258 			hw_float = npx_exists = 1;
259 			/*
260 			 * We have an npx, now divide by 0 to see if exception
261 			 * 16 works.
262 			 */
263 			control &= ~(1 << 2);	/* enable divide by 0 trap */
264 			fldcw(&control);
265 			npx_traps_while_probing = npx_intrs_while_probing = 0;
266 			fp_divide_by_0();
267 			if (npx_traps_while_probing != 0) {
268 				/*
269 				 * Good, exception 16 works.
270 				 */
271 				npx_ex16 = 1;
272 				dvp->id_irq = 0;	/* zap the interrupt */
273 				/*
274 				 * special return value to flag that we do not
275 				 * actually use any I/O registers
276 				 */
277 				return (-1);
278 			}
279 			if (npx_intrs_while_probing != 0) {
280 				/*
281 				 * Bad, we are stuck with IRQ13.
282 				 */
283 				npx_irq13 = 1;
284 				/*
285 				 * npxattach would be too late to set npx0_imask.
286 				 */
287 				npx0_imask |= dvp->id_irq;
288 				return (IO_NPXSIZE);
289 			}
290 			/*
291 			 * Worse, even IRQ13 is broken.  Use emulator.
292 			 */
293 		}
294 	}
295 	/*
296 	 * Probe failed, but we want to get to npxattach to initialize the
297 	 * emulator and say that it has been installed.  XXX handle devices
298 	 * that aren't really devices better.
299 	 */
300 	dvp->id_irq = 0;
301 	/*
302 	 * special return value to flag that we do not
303 	 * actually use any I/O registers
304 	 */
305 	return (-1);
306 }
307 
308 static struct kern_devconf kdc_npx[NNPX] = { {
309 	0, 0, 0,		/* filled in by dev_attach */
310 	"npx", 0, { MDDT_ISA, 0 },
311 	isa_generic_externalize, 0, 0, ISA_EXTERNALLEN,
312 	&kdc_isa0,		/* parent */
313 	0,			/* parentdata */
314 	DC_BUSY,
315 	"Floating-point unit"
316 } };
317 
318 static inline void
319 npx_registerdev(struct isa_device *id)
320 {
321 	int	unit;
322 
323 	unit = id->id_unit;
324 	if (unit != 0)
325 		kdc_npx[unit] = kdc_npx[0];
326 	kdc_npx[unit].kdc_unit = unit;
327 	kdc_npx[unit].kdc_isa = id;
328 	dev_attach(&kdc_npx[unit]);
329 }
330 
331 /*
332  * Attach routine - announce which it is, and wire into system
333  */
334 int
335 npxattach(dvp)
336 	struct isa_device *dvp;
337 {
338 	printf("npx%d: ", dvp->id_unit);
339 	if (npx_ex16)
340 		printf("INT 16 interface\n");
341 	else if (npx_irq13)
342 		printf("IRQ 13 interface\n");
343 #if defined(MATH_EMULATE) || defined(GPL_MATH_EMULATE)
344 	else if (npx_exists) {
345 		printf("error reporting broken; using 387 emulator\n");
346 		npx_exists = 0;
347 	} else
348 		printf("387 emulator\n");
349 #else
350 	else
351 		printf("no 387 emulator in kernel!\n");
352 #endif
353 	npxinit(__INITIAL_NPXCW__);
354 	if (npx_exists)
355 		npx_registerdev(dvp);
356 	return (1);		/* XXX unused */
357 }
358 
359 /*
360  * Initialize floating point unit.
361  */
362 void
363 npxinit(control)
364 	u_short control;
365 {
366 	struct save87 dummy;
367 
368 	if (!npx_exists)
369 		return;
370 	/*
371 	 * fninit has the same h/w bugs as fnsave.  Use the detoxified
372 	 * fnsave to throw away any junk in the fpu.  npxsave() initializes
373 	 * the fpu and sets npxproc = NULL as important side effects.
374 	 */
375 	npxsave(&dummy);
376 	stop_emulating();
377 	fldcw(&control);
378 	if (curpcb != NULL)
379 		fnsave(&curpcb->pcb_savefpu);
380 	start_emulating();
381 }
382 
383 /*
384  * Free coprocessor (if we have it).
385  */
386 void
387 npxexit(p)
388 	struct proc *p;
389 {
390 
391 	if (p == npxproc)
392 		npxsave(&curpcb->pcb_savefpu);
393 	if (npx_exists) {
394 		u_int	masked_exceptions;
395 
396 		masked_exceptions = curpcb->pcb_savefpu.sv_env.en_cw
397 				    & curpcb->pcb_savefpu.sv_env.en_sw & 0x7f;
398 		/*
399 		 * Overflow, divde by 0, and invalid operand would have
400 		 * caused a trap in 1.1.5.
401 		 */
402 		if (masked_exceptions & 0x0d)
403 			log(LOG_ERR,
404 	"pid %d (%s) exited with masked floating point exceptions 0x%02x\n",
405 			    p->p_pid, p->p_comm, masked_exceptions);
406 	}
407 }
408 
409 /*
410  * Preserve the FP status word, clear FP exceptions, then generate a SIGFPE.
411  *
412  * Clearing exceptions is necessary mainly to avoid IRQ13 bugs.  We now
413  * depend on longjmp() restoring a usable state.  Restoring the state
414  * or examining it might fail if we didn't clear exceptions.
415  *
416  * XXX there is no standard way to tell SIGFPE handlers about the error
417  * state.  The old interface:
418  *
419  *	void handler(int sig, int code, struct sigcontext *scp);
420  *
421  * is broken because it is non-ANSI and because the FP state is not in
422  * struct sigcontext.
423  *
424  * XXX the FP state is not preserved across signal handlers.  So signal
425  * handlers cannot afford to do FP unless they preserve the state or
426  * longjmp() out.  Both preserving the state and longjmp()ing may be
427  * destroyed by IRQ13 bugs.  Clearing FP exceptions is not an acceptable
428  * solution for signals other than SIGFPE.
429  */
430 void
431 npxintr(frame)
432 	struct intrframe frame;
433 {
434 	int code;
435 
436 	if (npxproc == NULL || !npx_exists) {
437 		printf("npxintr: npxproc = %p, curproc = %p, npx_exists = %d\n",
438 		       npxproc, curproc, npx_exists);
439 		panic("npxintr from nowhere");
440 	}
441 	if (npxproc != curproc) {
442 		printf("npxintr: npxproc = %p, curproc = %p, npx_exists = %d\n",
443 		       npxproc, curproc, npx_exists);
444 		panic("npxintr from non-current process");
445 	}
446 
447 	fnstsw(&curpcb->pcb_savefpu.sv_ex_sw);
448 	fnclex();
449 	outb(0xf0, 0);
450 
451 	/*
452 	 * Pass exception to process.
453 	 */
454 	if (ISPL(frame.if_cs) == SEL_UPL) {
455 		/*
456 		 * Interrupt is essentially a trap, so we can afford to call
457 		 * the SIGFPE handler (if any) as soon as the interrupt
458 		 * returns.
459 		 *
460 		 * XXX little or nothing is gained from this, and plenty is
461 		 * lost - the interrupt frame has to contain the trap frame
462 		 * (this is otherwise only necessary for the rescheduling trap
463 		 * in doreti, and the frame for that could easily be set up
464 		 * just before it is used).
465 		 */
466 		curproc->p_md.md_regs = (int *)&frame.if_es;
467 #ifdef notyet
468 		/*
469 		 * Encode the appropriate code for detailed information on
470 		 * this exception.
471 		 */
472 		code = XXX_ENCODE(curpcb->pcb_savefpu.sv_ex_sw);
473 #else
474 		code = 0;	/* XXX */
475 #endif
476 		trapsignal(curproc, SIGFPE, code);
477 	} else {
478 		/*
479 		 * Nested interrupt.  These losers occur when:
480 		 *	o an IRQ13 is bogusly generated at a bogus time, e.g.:
481 		 *		o immediately after an fnsave or frstor of an
482 		 *		  error state.
483 		 *		o a couple of 386 instructions after
484 		 *		  "fstpl _memvar" causes a stack overflow.
485 		 *	  These are especially nasty when combined with a
486 		 *	  trace trap.
487 		 *	o an IRQ13 occurs at the same time as another higher-
488 		 *	  priority interrupt.
489 		 *
490 		 * Treat them like a true async interrupt.
491 		 */
492 		psignal(curproc, SIGFPE);
493 	}
494 }
495 
496 /*
497  * Implement device not available (DNA) exception
498  *
499  * It would be better to switch FP context here (if curproc != npxproc)
500  * and not necessarily for every context switch, but it is too hard to
501  * access foreign pcb's.
502  */
503 int
504 npxdna()
505 {
506 	if (!npx_exists)
507 		return (0);
508 	if (npxproc != NULL) {
509 		printf("npxdna: npxproc = %p, curproc = %p\n",
510 		       npxproc, curproc);
511 		panic("npxdna");
512 	}
513 	stop_emulating();
514 	/*
515 	 * Record new context early in case frstor causes an IRQ13.
516 	 */
517 	npxproc = curproc;
518 	curpcb->pcb_savefpu.sv_ex_sw = 0;
519 	/*
520 	 * The following frstor may cause an IRQ13 when the state being
521 	 * restored has a pending error.  The error will appear to have been
522 	 * triggered by the current (npx) user instruction even when that
523 	 * instruction is a no-wait instruction that should not trigger an
524 	 * error (e.g., fnclex).  On at least one 486 system all of the
525 	 * no-wait instructions are broken the same as frstor, so our
526 	 * treatment does not amplify the breakage.  On at least one
527 	 * 386/Cyrix 387 system, fnclex works correctly while frstor and
528 	 * fnsave are broken, so our treatment breaks fnclex if it is the
529 	 * first FPU instruction after a context switch.
530 	 */
531 	frstor(&curpcb->pcb_savefpu);
532 
533 	return (1);
534 }
535 
536 /*
537  * Wrapper for fnsave instruction to handle h/w bugs.  If there is an error
538  * pending, then fnsave generates a bogus IRQ13 on some systems.  Force
539  * any IRQ13 to be handled immediately, and then ignore it.  This routine is
540  * often called at splhigh so it must not use many system services.  In
541  * particular, it's much easier to install a special handler than to
542  * guarantee that it's safe to use npxintr() and its supporting code.
543  */
544 void
545 npxsave(addr)
546 	struct save87 *addr;
547 {
548 	u_char	icu1_mask;
549 	u_char	icu2_mask;
550 	u_char	old_icu1_mask;
551 	u_char	old_icu2_mask;
552 	struct gate_descriptor	save_idt_npxintr;
553 
554 	disable_intr();
555 	old_icu1_mask = inb(IO_ICU1 + 1);
556 	old_icu2_mask = inb(IO_ICU2 + 1);
557 	save_idt_npxintr = idt[npx_intrno];
558 	outb(IO_ICU1 + 1, old_icu1_mask & ~(IRQ_SLAVE | npx0_imask));
559 	outb(IO_ICU2 + 1, old_icu2_mask & ~(npx0_imask >> 8));
560 	idt[npx_intrno] = npx_idt_probeintr;
561 	enable_intr();
562 	stop_emulating();
563 	fnsave(addr);
564 	fnop();
565 	outb(0xf0, 0);
566 	start_emulating();
567 	npxproc = NULL;
568 	disable_intr();
569 	icu1_mask = inb(IO_ICU1 + 1);	/* masks may have changed */
570 	icu2_mask = inb(IO_ICU2 + 1);
571 	outb(IO_ICU1 + 1,
572 	     (icu1_mask & ~npx0_imask) | (old_icu1_mask & npx0_imask));
573 	outb(IO_ICU2 + 1,
574 	     (icu2_mask & ~(npx0_imask >> 8))
575 	     | (old_icu2_mask & (npx0_imask >> 8)));
576 	idt[npx_intrno] = save_idt_npxintr;
577 	enable_intr();		/* back to usual state */
578 }
579 
580 #endif /* NNPX > 0 */
581