FreeBSD/Linux Kernel Cross Reference
sys/i386/isa/npx.c
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: releng/5.3/sys/i386/isa/npx.c 157863 2006-04-19 07:03:14Z cperciva $");
35
36 #include "opt_cpu.h"
37 #include "opt_debug_npx.h"
38 #include "opt_isa.h"
39 #include "opt_npx.h"
40
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/bus.h>
44 #include <sys/kernel.h>
45 #include <sys/lock.h>
46 #include <sys/malloc.h>
47 #include <sys/module.h>
48 #include <sys/mutex.h>
49 #include <sys/mutex.h>
50 #include <sys/proc.h>
51 #include <sys/smp.h>
52 #include <sys/sysctl.h>
53 #include <machine/bus.h>
54 #include <sys/rman.h>
55 #ifdef NPX_DEBUG
56 #include <sys/syslog.h>
57 #endif
58 #include <sys/signalvar.h>
59 #include <sys/user.h>
60
61 #include <machine/asmacros.h>
62 #include <machine/cputypes.h>
63 #include <machine/frame.h>
64 #include <machine/md_var.h>
65 #include <machine/pcb.h>
66 #include <machine/psl.h>
67 #include <machine/clock.h>
68 #include <machine/resource.h>
69 #include <machine/specialreg.h>
70 #include <machine/segments.h>
71 #include <machine/ucontext.h>
72
73 #ifdef PC98
74 #include <pc98/pc98/pc98.h>
75 #else
76 #include <i386/isa/isa.h>
77 #endif
78 #include <machine/intr_machdep.h>
79 #ifdef DEV_ISA
80 #include <isa/isavar.h>
81 #endif
82
83 #if !defined(CPU_ENABLE_SSE) && defined(I686_CPU)
84 #define CPU_ENABLE_SSE
85 #endif
86 #if defined(CPU_DISABLE_SSE)
87 #undef CPU_ENABLE_SSE
88 #endif
89
90 /*
91 * 387 and 287 Numeric Coprocessor Extension (NPX) Driver.
92 */
93
94 /* Configuration flags. */
95 #define NPX_DISABLE_I586_OPTIMIZED_BCOPY (1 << 0)
96 #define NPX_DISABLE_I586_OPTIMIZED_BZERO (1 << 1)
97 #define NPX_DISABLE_I586_OPTIMIZED_COPYIO (1 << 2)
98
99 #if (defined(__GNUC__) && !defined(lint)) || defined(__INTEL_COMPILER)
100
101 #define fldcw(addr) __asm("fldcw %0" : : "m" (*(addr)))
102 #define fnclex() __asm("fnclex")
103 #define fninit() __asm("fninit")
104 #define fnsave(addr) __asm __volatile("fnsave %0" : "=m" (*(addr)))
105 #define fnstcw(addr) __asm __volatile("fnstcw %0" : "=m" (*(addr)))
106 #define fnstsw(addr) __asm __volatile("fnstsw %0" : "=m" (*(addr)))
107 #define fp_divide_by_0() __asm("fldz; fld1; fdiv %st,%st(1); fnop")
108 #define frstor(addr) __asm("frstor %0" : : "m" (*(addr)))
109 #ifdef CPU_ENABLE_SSE
110 #define fxrstor(addr) __asm("fxrstor %0" : : "m" (*(addr)))
111 #define fxsave(addr) __asm __volatile("fxsave %0" : "=m" (*(addr)))
112 #endif
113 #define start_emulating() __asm("smsw %%ax; orb %0,%%al; lmsw %%ax" \
114 : : "n" (CR0_TS) : "ax")
115 #define stop_emulating() __asm("clts")
116
117 #else /* !((__GNUC__ && !lint ) || __INTEL_COMPILER) */
118
119 void fldcw(caddr_t addr);
120 void fnclex(void);
121 void fninit(void);
122 void fnsave(caddr_t addr);
123 void fnstcw(caddr_t addr);
124 void fnstsw(caddr_t addr);
125 void fp_divide_by_0(void);
126 void frstor(caddr_t addr);
127 #ifdef CPU_ENABLE_SSE
128 void fxsave(caddr_t addr);
129 void fxrstor(caddr_t addr);
130 #endif
131 void start_emulating(void);
132 void stop_emulating(void);
133
134 #endif /* (__GNUC__ && !lint ) || __INTEL_COMPILER */
135
136 #ifdef CPU_ENABLE_SSE
137 #define GET_FPU_CW(thread) \
138 (cpu_fxsr ? \
139 (thread)->td_pcb->pcb_save.sv_xmm.sv_env.en_cw : \
140 (thread)->td_pcb->pcb_save.sv_87.sv_env.en_cw)
141 #define GET_FPU_SW(thread) \
142 (cpu_fxsr ? \
143 (thread)->td_pcb->pcb_save.sv_xmm.sv_env.en_sw : \
144 (thread)->td_pcb->pcb_save.sv_87.sv_env.en_sw)
145 #else /* CPU_ENABLE_SSE */
146 #define GET_FPU_CW(thread) \
147 (thread->td_pcb->pcb_save.sv_87.sv_env.en_cw)
148 #define GET_FPU_SW(thread) \
149 (thread->td_pcb->pcb_save.sv_87.sv_env.en_sw)
150 #endif /* CPU_ENABLE_SSE */
151
152 typedef u_char bool_t;
153
154 #ifdef CPU_ENABLE_SSE
155 static void fpu_clean_state(void);
156 #endif
157
158 static void fpusave(union savefpu *);
159 static void fpurstor(union savefpu *);
160 static int npx_attach(device_t dev);
161 static void npx_identify(driver_t *driver, device_t parent);
162 static void npx_intr(void *);
163 static int npx_probe(device_t dev);
164 #ifdef I586_CPU_XXX
165 static long timezero(const char *funcname,
166 void (*func)(void *buf, size_t len));
167 #endif /* I586_CPU */
168
169 int hw_float; /* XXX currently just alias for npx_exists */
170
171 SYSCTL_INT(_hw,HW_FLOATINGPT, floatingpoint,
172 CTLFLAG_RD, &hw_float, 0,
173 "Floatingpoint instructions executed in hardware");
174
175 static volatile u_int npx_intrs_while_probing;
176 static volatile u_int npx_traps_while_probing;
177
178 static union savefpu npx_cleanstate;
179 static bool_t npx_cleanstate_ready;
180 static bool_t npx_ex16;
181 static bool_t npx_exists;
182 static bool_t npx_irq13;
183
184 alias_for_inthand_t probetrap;
185 __asm(" \n\
186 .text \n\
187 .p2align 2,0x90 \n\
188 .type " __XSTRING(CNAME(probetrap)) ",@function \n\
189 " __XSTRING(CNAME(probetrap)) ": \n\
190 ss \n\
191 incl " __XSTRING(CNAME(npx_traps_while_probing)) " \n\
192 fnclex \n\
193 iret \n\
194 ");
195
196 /*
197 * Identify routine. Create a connection point on our parent for probing.
198 */
199 static void
200 npx_identify(driver, parent)
201 driver_t *driver;
202 device_t parent;
203 {
204 device_t child;
205
206 child = BUS_ADD_CHILD(parent, 0, "npx", 0);
207 if (child == NULL)
208 panic("npx_identify");
209 }
210
211 /*
212 * Do minimal handling of npx interrupts to convert them to traps.
213 */
214 static void
215 npx_intr(dummy)
216 void *dummy;
217 {
218 struct thread *td;
219
220 npx_intrs_while_probing++;
221
222 /*
223 * The BUSY# latch must be cleared in all cases so that the next
224 * unmasked npx exception causes an interrupt.
225 */
226 #ifdef PC98
227 outb(0xf8, 0);
228 #else
229 outb(0xf0, 0);
230 #endif
231
232 /*
233 * fpcurthread is normally non-null here. In that case, schedule an
234 * AST to finish the exception handling in the correct context
235 * (this interrupt may occur after the thread has entered the
236 * kernel via a syscall or an interrupt). Otherwise, the npx
237 * state of the thread that caused this interrupt must have been
238 * pushed to the thread's pcb, and clearing of the busy latch
239 * above has finished the (essentially null) handling of this
240 * interrupt. Control will eventually return to the instruction
241 * that caused it and it will repeat. We will eventually (usually
242 * soon) win the race to handle the interrupt properly.
243 */
244 td = PCPU_GET(fpcurthread);
245 if (td != NULL) {
246 td->td_pcb->pcb_flags |= PCB_NPXTRAP;
247 mtx_lock_spin(&sched_lock);
248 td->td_flags |= TDF_ASTPENDING;
249 mtx_unlock_spin(&sched_lock);
250 }
251 }
252
253 /*
254 * Probe routine. Initialize cr0 to give correct behaviour for [f]wait
255 * whether the device exists or not (XXX should be elsewhere). Set flags
256 * to tell npxattach() what to do. Modify device struct if npx doesn't
257 * need to use interrupts. Return 0 if device exists.
258 */
259 static int
260 npx_probe(dev)
261 device_t dev;
262 {
263 struct gate_descriptor save_idt_npxtrap;
264 struct resource *ioport_res, *irq_res;
265 void *irq_cookie;
266 int ioport_rid, irq_num, irq_rid;
267 u_short control;
268 u_short status;
269
270 save_idt_npxtrap = idt[IDT_MF];
271 setidt(IDT_MF, probetrap, SDT_SYS386TGT, SEL_KPL,
272 GSEL(GCODE_SEL, SEL_KPL));
273 ioport_rid = 0;
274 ioport_res = bus_alloc_resource(dev, SYS_RES_IOPORT, &ioport_rid,
275 IO_NPX, IO_NPX, IO_NPXSIZE, RF_ACTIVE);
276 if (ioport_res == NULL)
277 panic("npx: can't get ports");
278 #ifdef PC98
279 if (resource_int_value("npx", 0, "irq", &irq_num) != 0)
280 irq_num = 8;
281 #else
282 if (resource_int_value("npx", 0, "irq", &irq_num) != 0)
283 irq_num = 13;
284 #endif
285 irq_rid = 0;
286 irq_res = bus_alloc_resource(dev, SYS_RES_IRQ, &irq_rid, irq_num,
287 irq_num, 1, RF_ACTIVE);
288 if (irq_res == NULL)
289 panic("npx: can't get IRQ");
290 if (bus_setup_intr(dev, irq_res, INTR_TYPE_MISC | INTR_FAST, npx_intr,
291 NULL, &irq_cookie) != 0)
292 panic("npx: can't create intr");
293
294 /*
295 * Partially reset the coprocessor, if any. Some BIOS's don't reset
296 * it after a warm boot.
297 */
298 #ifdef PC98
299 outb(0xf8,0);
300 #else
301 outb(0xf1, 0); /* full reset on some systems, NOP on others */
302 outb(0xf0, 0); /* clear BUSY# latch */
303 #endif
304 /*
305 * Prepare to trap all ESC (i.e., NPX) instructions and all WAIT
306 * instructions. We must set the CR0_MP bit and use the CR0_TS
307 * bit to control the trap, because setting the CR0_EM bit does
308 * not cause WAIT instructions to trap. It's important to trap
309 * WAIT instructions - otherwise the "wait" variants of no-wait
310 * control instructions would degenerate to the "no-wait" variants
311 * after FP context switches but work correctly otherwise. It's
312 * particularly important to trap WAITs when there is no NPX -
313 * otherwise the "wait" variants would always degenerate.
314 *
315 * Try setting CR0_NE to get correct error reporting on 486DX's.
316 * Setting it should fail or do nothing on lesser processors.
317 */
318 load_cr0(rcr0() | CR0_MP | CR0_NE);
319 /*
320 * But don't trap while we're probing.
321 */
322 stop_emulating();
323 /*
324 * Finish resetting the coprocessor, if any. If there is an error
325 * pending, then we may get a bogus IRQ13, but npx_intr() will handle
326 * it OK. Bogus halts have never been observed, but we enabled
327 * IRQ13 and cleared the BUSY# latch early to handle them anyway.
328 */
329 fninit();
330
331 device_set_desc(dev, "math processor");
332
333 /*
334 * Don't use fwait here because it might hang.
335 * Don't use fnop here because it usually hangs if there is no FPU.
336 */
337 DELAY(1000); /* wait for any IRQ13 */
338 #ifdef DIAGNOSTIC
339 if (npx_intrs_while_probing != 0)
340 printf("fninit caused %u bogus npx interrupt(s)\n",
341 npx_intrs_while_probing);
342 if (npx_traps_while_probing != 0)
343 printf("fninit caused %u bogus npx trap(s)\n",
344 npx_traps_while_probing);
345 #endif
346 /*
347 * Check for a status of mostly zero.
348 */
349 status = 0x5a5a;
350 fnstsw(&status);
351 if ((status & 0xb8ff) == 0) {
352 /*
353 * Good, now check for a proper control word.
354 */
355 control = 0x5a5a;
356 fnstcw(&control);
357 if ((control & 0x1f3f) == 0x033f) {
358 hw_float = npx_exists = 1;
359 /*
360 * We have an npx, now divide by 0 to see if exception
361 * 16 works.
362 */
363 control &= ~(1 << 2); /* enable divide by 0 trap */
364 fldcw(&control);
365 #ifdef FPU_ERROR_BROKEN
366 /*
367 * FPU error signal doesn't work on some CPU
368 * accelerator board.
369 */
370 npx_ex16 = 1;
371 return (0);
372 #endif
373 npx_traps_while_probing = npx_intrs_while_probing = 0;
374 fp_divide_by_0();
375 DELAY(1000); /* wait for any IRQ13 */
376 if (npx_traps_while_probing != 0) {
377 /*
378 * Good, exception 16 works.
379 */
380 npx_ex16 = 1;
381 goto no_irq13;
382 }
383 if (npx_intrs_while_probing != 0) {
384 /*
385 * Bad, we are stuck with IRQ13.
386 */
387 npx_irq13 = 1;
388 idt[IDT_MF] = save_idt_npxtrap;
389 #ifdef SMP
390 if (mp_ncpus > 1)
391 panic("npx0 cannot use IRQ 13 on an SMP system");
392 #endif
393 return (0);
394 }
395 /*
396 * Worse, even IRQ13 is broken. Use emulator.
397 */
398 }
399 }
400 /*
401 * Probe failed, but we want to get to npxattach to initialize the
402 * emulator and say that it has been installed. XXX handle devices
403 * that aren't really devices better.
404 */
405 #ifdef SMP
406 if (mp_ncpus > 1)
407 panic("npx0 cannot be emulated on an SMP system");
408 #endif
409 /* FALLTHROUGH */
410 no_irq13:
411 idt[IDT_MF] = save_idt_npxtrap;
412 bus_teardown_intr(dev, irq_res, irq_cookie);
413 bus_release_resource(dev, SYS_RES_IRQ, irq_rid, irq_res);
414 bus_release_resource(dev, SYS_RES_IOPORT, ioport_rid, ioport_res);
415 return (0);
416 }
417
418 /*
419 * Attach routine - announce which it is, and wire into system
420 */
421 static int
422 npx_attach(dev)
423 device_t dev;
424 {
425 int flags;
426 register_t s;
427
428 if (resource_int_value("npx", 0, "flags", &flags) != 0)
429 flags = 0;
430
431 if (flags)
432 device_printf(dev, "flags 0x%x ", flags);
433 if (npx_irq13) {
434 device_printf(dev, "using IRQ 13 interface\n");
435 } else {
436 if (npx_ex16)
437 device_printf(dev, "INT 16 interface\n");
438 else
439 device_printf(dev, "WARNING: no FPU!\n");
440 }
441 npxinit(__INITIAL_NPXCW__);
442
443 if (npx_cleanstate_ready == 0) {
444 s = intr_disable();
445 stop_emulating();
446 fpusave(&npx_cleanstate);
447 start_emulating();
448 npx_cleanstate_ready = 1;
449 intr_restore(s);
450 }
451 #ifdef I586_CPU_XXX
452 if (cpu_class == CPUCLASS_586 && npx_ex16 && npx_exists &&
453 timezero("i586_bzero()", i586_bzero) <
454 timezero("bzero()", bzero) * 4 / 5) {
455 if (!(flags & NPX_DISABLE_I586_OPTIMIZED_BCOPY))
456 bcopy_vector = i586_bcopy;
457 if (!(flags & NPX_DISABLE_I586_OPTIMIZED_BZERO))
458 bzero_vector = i586_bzero;
459 if (!(flags & NPX_DISABLE_I586_OPTIMIZED_COPYIO)) {
460 copyin_vector = i586_copyin;
461 copyout_vector = i586_copyout;
462 }
463 }
464 #endif
465
466 return (0); /* XXX unused */
467 }
468
469 /*
470 * Initialize floating point unit.
471 */
472 void
473 npxinit(control)
474 u_short control;
475 {
476 static union savefpu dummy;
477 register_t savecrit;
478
479 if (!npx_exists)
480 return;
481 /*
482 * fninit has the same h/w bugs as fnsave. Use the detoxified
483 * fnsave to throw away any junk in the fpu. npxsave() initializes
484 * the fpu and sets fpcurthread = NULL as important side effects.
485 */
486 savecrit = intr_disable();
487 npxsave(&dummy);
488 stop_emulating();
489 #ifdef CPU_ENABLE_SSE
490 /* XXX npxsave() doesn't actually initialize the fpu in the SSE case. */
491 if (cpu_fxsr)
492 fninit();
493 #endif
494 fldcw(&control);
495 start_emulating();
496 intr_restore(savecrit);
497 }
498
499 /*
500 * Free coprocessor (if we have it).
501 */
502 void
503 npxexit(td)
504 struct thread *td;
505 {
506 register_t savecrit;
507
508 savecrit = intr_disable();
509 if (curthread == PCPU_GET(fpcurthread))
510 npxsave(&PCPU_GET(curpcb)->pcb_save);
511 intr_restore(savecrit);
512 #ifdef NPX_DEBUG
513 if (npx_exists) {
514 u_int masked_exceptions;
515
516 masked_exceptions = GET_FPU_CW(td) & GET_FPU_SW(td) & 0x7f;
517 /*
518 * Log exceptions that would have trapped with the old
519 * control word (overflow, divide by 0, and invalid operand).
520 */
521 if (masked_exceptions & 0x0d)
522 log(LOG_ERR,
523 "pid %d (%s) exited with masked floating point exceptions 0x%02x\n",
524 td->td_proc->p_pid, td->td_proc->p_comm,
525 masked_exceptions);
526 }
527 #endif
528 }
529
530 int
531 npxformat()
532 {
533
534 if (!npx_exists)
535 return (_MC_FPFMT_NODEV);
536 #ifdef CPU_ENABLE_SSE
537 if (cpu_fxsr)
538 return (_MC_FPFMT_XMM);
539 #endif
540 return (_MC_FPFMT_387);
541 }
542
543 /*
544 * The following mechanism is used to ensure that the FPE_... value
545 * that is passed as a trapcode to the signal handler of the user
546 * process does not have more than one bit set.
547 *
548 * Multiple bits may be set if the user process modifies the control
549 * word while a status word bit is already set. While this is a sign
550 * of bad coding, we have no choise than to narrow them down to one
551 * bit, since we must not send a trapcode that is not exactly one of
552 * the FPE_ macros.
553 *
554 * The mechanism has a static table with 127 entries. Each combination
555 * of the 7 FPU status word exception bits directly translates to a
556 * position in this table, where a single FPE_... value is stored.
557 * This FPE_... value stored there is considered the "most important"
558 * of the exception bits and will be sent as the signal code. The
559 * precedence of the bits is based upon Intel Document "Numerical
560 * Applications", Chapter "Special Computational Situations".
561 *
562 * The macro to choose one of these values does these steps: 1) Throw
563 * away status word bits that cannot be masked. 2) Throw away the bits
564 * currently masked in the control word, assuming the user isn't
565 * interested in them anymore. 3) Reinsert status word bit 7 (stack
566 * fault) if it is set, which cannot be masked but must be presered.
567 * 4) Use the remaining bits to point into the trapcode table.
568 *
569 * The 6 maskable bits in order of their preference, as stated in the
570 * above referenced Intel manual:
571 * 1 Invalid operation (FP_X_INV)
572 * 1a Stack underflow
573 * 1b Stack overflow
574 * 1c Operand of unsupported format
575 * 1d SNaN operand.
576 * 2 QNaN operand (not an exception, irrelavant here)
577 * 3 Any other invalid-operation not mentioned above or zero divide
578 * (FP_X_INV, FP_X_DZ)
579 * 4 Denormal operand (FP_X_DNML)
580 * 5 Numeric over/underflow (FP_X_OFL, FP_X_UFL)
581 * 6 Inexact result (FP_X_IMP)
582 */
583 static char fpetable[128] = {
584 0,
585 FPE_FLTINV, /* 1 - INV */
586 FPE_FLTUND, /* 2 - DNML */
587 FPE_FLTINV, /* 3 - INV | DNML */
588 FPE_FLTDIV, /* 4 - DZ */
589 FPE_FLTINV, /* 5 - INV | DZ */
590 FPE_FLTDIV, /* 6 - DNML | DZ */
591 FPE_FLTINV, /* 7 - INV | DNML | DZ */
592 FPE_FLTOVF, /* 8 - OFL */
593 FPE_FLTINV, /* 9 - INV | OFL */
594 FPE_FLTUND, /* A - DNML | OFL */
595 FPE_FLTINV, /* B - INV | DNML | OFL */
596 FPE_FLTDIV, /* C - DZ | OFL */
597 FPE_FLTINV, /* D - INV | DZ | OFL */
598 FPE_FLTDIV, /* E - DNML | DZ | OFL */
599 FPE_FLTINV, /* F - INV | DNML | DZ | OFL */
600 FPE_FLTUND, /* 10 - UFL */
601 FPE_FLTINV, /* 11 - INV | UFL */
602 FPE_FLTUND, /* 12 - DNML | UFL */
603 FPE_FLTINV, /* 13 - INV | DNML | UFL */
604 FPE_FLTDIV, /* 14 - DZ | UFL */
605 FPE_FLTINV, /* 15 - INV | DZ | UFL */
606 FPE_FLTDIV, /* 16 - DNML | DZ | UFL */
607 FPE_FLTINV, /* 17 - INV | DNML | DZ | UFL */
608 FPE_FLTOVF, /* 18 - OFL | UFL */
609 FPE_FLTINV, /* 19 - INV | OFL | UFL */
610 FPE_FLTUND, /* 1A - DNML | OFL | UFL */
611 FPE_FLTINV, /* 1B - INV | DNML | OFL | UFL */
612 FPE_FLTDIV, /* 1C - DZ | OFL | UFL */
613 FPE_FLTINV, /* 1D - INV | DZ | OFL | UFL */
614 FPE_FLTDIV, /* 1E - DNML | DZ | OFL | UFL */
615 FPE_FLTINV, /* 1F - INV | DNML | DZ | OFL | UFL */
616 FPE_FLTRES, /* 20 - IMP */
617 FPE_FLTINV, /* 21 - INV | IMP */
618 FPE_FLTUND, /* 22 - DNML | IMP */
619 FPE_FLTINV, /* 23 - INV | DNML | IMP */
620 FPE_FLTDIV, /* 24 - DZ | IMP */
621 FPE_FLTINV, /* 25 - INV | DZ | IMP */
622 FPE_FLTDIV, /* 26 - DNML | DZ | IMP */
623 FPE_FLTINV, /* 27 - INV | DNML | DZ | IMP */
624 FPE_FLTOVF, /* 28 - OFL | IMP */
625 FPE_FLTINV, /* 29 - INV | OFL | IMP */
626 FPE_FLTUND, /* 2A - DNML | OFL | IMP */
627 FPE_FLTINV, /* 2B - INV | DNML | OFL | IMP */
628 FPE_FLTDIV, /* 2C - DZ | OFL | IMP */
629 FPE_FLTINV, /* 2D - INV | DZ | OFL | IMP */
630 FPE_FLTDIV, /* 2E - DNML | DZ | OFL | IMP */
631 FPE_FLTINV, /* 2F - INV | DNML | DZ | OFL | IMP */
632 FPE_FLTUND, /* 30 - UFL | IMP */
633 FPE_FLTINV, /* 31 - INV | UFL | IMP */
634 FPE_FLTUND, /* 32 - DNML | UFL | IMP */
635 FPE_FLTINV, /* 33 - INV | DNML | UFL | IMP */
636 FPE_FLTDIV, /* 34 - DZ | UFL | IMP */
637 FPE_FLTINV, /* 35 - INV | DZ | UFL | IMP */
638 FPE_FLTDIV, /* 36 - DNML | DZ | UFL | IMP */
639 FPE_FLTINV, /* 37 - INV | DNML | DZ | UFL | IMP */
640 FPE_FLTOVF, /* 38 - OFL | UFL | IMP */
641 FPE_FLTINV, /* 39 - INV | OFL | UFL | IMP */
642 FPE_FLTUND, /* 3A - DNML | OFL | UFL | IMP */
643 FPE_FLTINV, /* 3B - INV | DNML | OFL | UFL | IMP */
644 FPE_FLTDIV, /* 3C - DZ | OFL | UFL | IMP */
645 FPE_FLTINV, /* 3D - INV | DZ | OFL | UFL | IMP */
646 FPE_FLTDIV, /* 3E - DNML | DZ | OFL | UFL | IMP */
647 FPE_FLTINV, /* 3F - INV | DNML | DZ | OFL | UFL | IMP */
648 FPE_FLTSUB, /* 40 - STK */
649 FPE_FLTSUB, /* 41 - INV | STK */
650 FPE_FLTUND, /* 42 - DNML | STK */
651 FPE_FLTSUB, /* 43 - INV | DNML | STK */
652 FPE_FLTDIV, /* 44 - DZ | STK */
653 FPE_FLTSUB, /* 45 - INV | DZ | STK */
654 FPE_FLTDIV, /* 46 - DNML | DZ | STK */
655 FPE_FLTSUB, /* 47 - INV | DNML | DZ | STK */
656 FPE_FLTOVF, /* 48 - OFL | STK */
657 FPE_FLTSUB, /* 49 - INV | OFL | STK */
658 FPE_FLTUND, /* 4A - DNML | OFL | STK */
659 FPE_FLTSUB, /* 4B - INV | DNML | OFL | STK */
660 FPE_FLTDIV, /* 4C - DZ | OFL | STK */
661 FPE_FLTSUB, /* 4D - INV | DZ | OFL | STK */
662 FPE_FLTDIV, /* 4E - DNML | DZ | OFL | STK */
663 FPE_FLTSUB, /* 4F - INV | DNML | DZ | OFL | STK */
664 FPE_FLTUND, /* 50 - UFL | STK */
665 FPE_FLTSUB, /* 51 - INV | UFL | STK */
666 FPE_FLTUND, /* 52 - DNML | UFL | STK */
667 FPE_FLTSUB, /* 53 - INV | DNML | UFL | STK */
668 FPE_FLTDIV, /* 54 - DZ | UFL | STK */
669 FPE_FLTSUB, /* 55 - INV | DZ | UFL | STK */
670 FPE_FLTDIV, /* 56 - DNML | DZ | UFL | STK */
671 FPE_FLTSUB, /* 57 - INV | DNML | DZ | UFL | STK */
672 FPE_FLTOVF, /* 58 - OFL | UFL | STK */
673 FPE_FLTSUB, /* 59 - INV | OFL | UFL | STK */
674 FPE_FLTUND, /* 5A - DNML | OFL | UFL | STK */
675 FPE_FLTSUB, /* 5B - INV | DNML | OFL | UFL | STK */
676 FPE_FLTDIV, /* 5C - DZ | OFL | UFL | STK */
677 FPE_FLTSUB, /* 5D - INV | DZ | OFL | UFL | STK */
678 FPE_FLTDIV, /* 5E - DNML | DZ | OFL | UFL | STK */
679 FPE_FLTSUB, /* 5F - INV | DNML | DZ | OFL | UFL | STK */
680 FPE_FLTRES, /* 60 - IMP | STK */
681 FPE_FLTSUB, /* 61 - INV | IMP | STK */
682 FPE_FLTUND, /* 62 - DNML | IMP | STK */
683 FPE_FLTSUB, /* 63 - INV | DNML | IMP | STK */
684 FPE_FLTDIV, /* 64 - DZ | IMP | STK */
685 FPE_FLTSUB, /* 65 - INV | DZ | IMP | STK */
686 FPE_FLTDIV, /* 66 - DNML | DZ | IMP | STK */
687 FPE_FLTSUB, /* 67 - INV | DNML | DZ | IMP | STK */
688 FPE_FLTOVF, /* 68 - OFL | IMP | STK */
689 FPE_FLTSUB, /* 69 - INV | OFL | IMP | STK */
690 FPE_FLTUND, /* 6A - DNML | OFL | IMP | STK */
691 FPE_FLTSUB, /* 6B - INV | DNML | OFL | IMP | STK */
692 FPE_FLTDIV, /* 6C - DZ | OFL | IMP | STK */
693 FPE_FLTSUB, /* 6D - INV | DZ | OFL | IMP | STK */
694 FPE_FLTDIV, /* 6E - DNML | DZ | OFL | IMP | STK */
695 FPE_FLTSUB, /* 6F - INV | DNML | DZ | OFL | IMP | STK */
696 FPE_FLTUND, /* 70 - UFL | IMP | STK */
697 FPE_FLTSUB, /* 71 - INV | UFL | IMP | STK */
698 FPE_FLTUND, /* 72 - DNML | UFL | IMP | STK */
699 FPE_FLTSUB, /* 73 - INV | DNML | UFL | IMP | STK */
700 FPE_FLTDIV, /* 74 - DZ | UFL | IMP | STK */
701 FPE_FLTSUB, /* 75 - INV | DZ | UFL | IMP | STK */
702 FPE_FLTDIV, /* 76 - DNML | DZ | UFL | IMP | STK */
703 FPE_FLTSUB, /* 77 - INV | DNML | DZ | UFL | IMP | STK */
704 FPE_FLTOVF, /* 78 - OFL | UFL | IMP | STK */
705 FPE_FLTSUB, /* 79 - INV | OFL | UFL | IMP | STK */
706 FPE_FLTUND, /* 7A - DNML | OFL | UFL | IMP | STK */
707 FPE_FLTSUB, /* 7B - INV | DNML | OFL | UFL | IMP | STK */
708 FPE_FLTDIV, /* 7C - DZ | OFL | UFL | IMP | STK */
709 FPE_FLTSUB, /* 7D - INV | DZ | OFL | UFL | IMP | STK */
710 FPE_FLTDIV, /* 7E - DNML | DZ | OFL | UFL | IMP | STK */
711 FPE_FLTSUB, /* 7F - INV | DNML | DZ | OFL | UFL | IMP | STK */
712 };
713
714 /*
715 * Preserve the FP status word, clear FP exceptions, then generate a SIGFPE.
716 *
717 * Clearing exceptions is necessary mainly to avoid IRQ13 bugs. We now
718 * depend on longjmp() restoring a usable state. Restoring the state
719 * or examining it might fail if we didn't clear exceptions.
720 *
721 * The error code chosen will be one of the FPE_... macros. It will be
722 * sent as the second argument to old BSD-style signal handlers and as
723 * "siginfo_t->si_code" (second argument) to SA_SIGINFO signal handlers.
724 *
725 * XXX the FP state is not preserved across signal handlers. So signal
726 * handlers cannot afford to do FP unless they preserve the state or
727 * longjmp() out. Both preserving the state and longjmp()ing may be
728 * destroyed by IRQ13 bugs. Clearing FP exceptions is not an acceptable
729 * solution for signals other than SIGFPE.
730 */
731 int
732 npxtrap()
733 {
734 register_t savecrit;
735 u_short control, status;
736
737 if (!npx_exists) {
738 printf("npxtrap: fpcurthread = %p, curthread = %p, npx_exists = %d\n",
739 PCPU_GET(fpcurthread), curthread, npx_exists);
740 panic("npxtrap from nowhere");
741 }
742 savecrit = intr_disable();
743
744 /*
745 * Interrupt handling (for another interrupt) may have pushed the
746 * state to memory. Fetch the relevant parts of the state from
747 * wherever they are.
748 */
749 if (PCPU_GET(fpcurthread) != curthread) {
750 control = GET_FPU_CW(curthread);
751 status = GET_FPU_SW(curthread);
752 } else {
753 fnstcw(&control);
754 fnstsw(&status);
755 }
756
757 if (PCPU_GET(fpcurthread) == curthread)
758 fnclex();
759 intr_restore(savecrit);
760 return (fpetable[status & ((~control & 0x3f) | 0x40)]);
761 }
762
763 /*
764 * Implement device not available (DNA) exception
765 *
766 * It would be better to switch FP context here (if curthread != fpcurthread)
767 * and not necessarily for every context switch, but it is too hard to
768 * access foreign pcb's.
769 */
770
771 static int err_count = 0;
772
773 int
774 npxdna()
775 {
776 struct pcb *pcb;
777 register_t s;
778 u_short control;
779
780 if (!npx_exists)
781 return (0);
782 if (PCPU_GET(fpcurthread) == curthread) {
783 printf("npxdna: fpcurthread == curthread %d times\n",
784 ++err_count);
785 stop_emulating();
786 return (1);
787 }
788 if (PCPU_GET(fpcurthread) != NULL) {
789 printf("npxdna: fpcurthread = %p (%d), curthread = %p (%d)\n",
790 PCPU_GET(fpcurthread),
791 PCPU_GET(fpcurthread)->td_proc->p_pid,
792 curthread, curthread->td_proc->p_pid);
793 panic("npxdna");
794 }
795 s = intr_disable();
796 stop_emulating();
797 /*
798 * Record new context early in case frstor causes an IRQ13.
799 */
800 PCPU_SET(fpcurthread, curthread);
801 pcb = PCPU_GET(curpcb);
802
803 if ((pcb->pcb_flags & PCB_NPXINITDONE) == 0) {
804 /*
805 * This is the first time this thread has used the FPU or
806 * the PCB doesn't contain a clean FPU state. Explicitly
807 * initialize the FPU and load the default control word.
808 */
809 fninit();
810 control = __INITIAL_NPXCW__;
811 fldcw(&control);
812 pcb->pcb_flags |= PCB_NPXINITDONE;
813 } else {
814 /*
815 * The following frstor may cause an IRQ13 when the state
816 * being restored has a pending error. The error will
817 * appear to have been triggered by the current (npx) user
818 * instruction even when that instruction is a no-wait
819 * instruction that should not trigger an error (e.g.,
820 * fnclex). On at least one 486 system all of the no-wait
821 * instructions are broken the same as frstor, so our
822 * treatment does not amplify the breakage. On at least
823 * one 386/Cyrix 387 system, fnclex works correctly while
824 * frstor and fnsave are broken, so our treatment breaks
825 * fnclex if it is the first FPU instruction after a context
826 * switch.
827 */
828 fpurstor(&pcb->pcb_save);
829 }
830 intr_restore(s);
831
832 return (1);
833 }
834
835 /*
836 * Wrapper for fnsave instruction, partly to handle hardware bugs. When npx
837 * exceptions are reported via IRQ13, spurious IRQ13's may be triggered by
838 * no-wait npx instructions. See the Intel application note AP-578 for
839 * details. This doesn't cause any additional complications here. IRQ13's
840 * are inherently asynchronous unless the CPU is frozen to deliver them --
841 * one that started in userland may be delivered many instructions later,
842 * after the process has entered the kernel. It may even be delivered after
843 * the fnsave here completes. A spurious IRQ13 for the fnsave is handled in
844 * the same way as a very-late-arriving non-spurious IRQ13 from user mode:
845 * it is normally ignored at first because we set fpcurthread to NULL; it is
846 * normally retriggered in npxdna() after return to user mode.
847 *
848 * npxsave() must be called with interrupts disabled, so that it clears
849 * fpcurthread atomically with saving the state. We require callers to do the
850 * disabling, since most callers need to disable interrupts anyway to call
851 * npxsave() atomically with checking fpcurthread.
852 *
853 * A previous version of npxsave() went to great lengths to excecute fnsave
854 * with interrupts enabled in case executing it froze the CPU. This case
855 * can't happen, at least for Intel CPU/NPX's. Spurious IRQ13's don't imply
856 * spurious freezes.
857 */
858 void
859 npxsave(addr)
860 union savefpu *addr;
861 {
862
863 stop_emulating();
864 fpusave(addr);
865
866 start_emulating();
867 PCPU_SET(fpcurthread, NULL);
868 }
869
870 /*
871 * This should be called with interrupts disabled and only when the owning
872 * FPU thread is non-null.
873 */
874 void
875 npxdrop()
876 {
877 struct thread *td;
878
879 /*
880 * Discard pending exceptions in the !cpu_fxsr case so that unmasked
881 * ones don't cause a panic on the next frstor.
882 */
883 #ifdef CPU_ENABLE_SSE
884 if (!cpu_fxsr)
885 #endif
886 fnclex();
887
888 td = PCPU_GET(fpcurthread);
889 PCPU_SET(fpcurthread, NULL);
890 td->td_pcb->pcb_flags &= ~PCB_NPXINITDONE;
891 start_emulating();
892 }
893
894 /*
895 * Get the state of the FPU without dropping ownership (if possible).
896 * It returns the FPU ownership status.
897 */
898 int
899 npxgetregs(td, addr)
900 struct thread *td;
901 union savefpu *addr;
902 {
903 register_t s;
904
905 if (!npx_exists)
906 return (_MC_FPOWNED_NONE);
907
908 if ((td->td_pcb->pcb_flags & PCB_NPXINITDONE) == 0) {
909 if (npx_cleanstate_ready)
910 bcopy(&npx_cleanstate, addr, sizeof(npx_cleanstate));
911 else
912 bzero(addr, sizeof(*addr));
913 return (_MC_FPOWNED_NONE);
914 }
915 s = intr_disable();
916 if (td == PCPU_GET(fpcurthread)) {
917 fpusave(addr);
918 #ifdef CPU_ENABLE_SSE
919 if (!cpu_fxsr)
920 #endif
921 /*
922 * fnsave initializes the FPU and destroys whatever
923 * context it contains. Make sure the FPU owner
924 * starts with a clean state next time.
925 */
926 npxdrop();
927 intr_restore(s);
928 return (_MC_FPOWNED_FPU);
929 } else {
930 intr_restore(s);
931 bcopy(&td->td_pcb->pcb_save, addr, sizeof(*addr));
932 return (_MC_FPOWNED_PCB);
933 }
934 }
935
936 /*
937 * Set the state of the FPU.
938 */
939 void
940 npxsetregs(td, addr)
941 struct thread *td;
942 union savefpu *addr;
943 {
944 register_t s;
945
946 if (!npx_exists)
947 return;
948
949 s = intr_disable();
950 if (td == PCPU_GET(fpcurthread)) {
951 #ifdef CPU_ENABLE_SSE
952 if (!cpu_fxsr)
953 #endif
954 fnclex(); /* As in npxdrop(). */
955 fpurstor(addr);
956 intr_restore(s);
957 } else {
958 intr_restore(s);
959 bcopy(addr, &td->td_pcb->pcb_save, sizeof(*addr));
960 }
961 curthread->td_pcb->pcb_flags |= PCB_NPXINITDONE;
962 }
963
964 static void
965 fpusave(addr)
966 union savefpu *addr;
967 {
968
969 #ifdef CPU_ENABLE_SSE
970 if (cpu_fxsr)
971 fxsave(addr);
972 else
973 #endif
974 fnsave(addr);
975 }
976
977 #ifdef CPU_ENABLE_SSE
978 /*
979 * On AuthenticAMD processors, the fxrstor instruction does not restore
980 * the x87's stored last instruction pointer, last data pointer, and last
981 * opcode values, except in the rare case in which the exception summary
982 * (ES) bit in the x87 status word is set to 1.
983 *
984 * In order to avoid leaking this information across processes, we clean
985 * these values by performing a dummy load before executing fxrstor().
986 */
987 static double dummy_variable = 0.0;
988 static void
989 fpu_clean_state(void)
990 {
991 u_short status;
992
993 /*
994 * Clear the ES bit in the x87 status word if it is currently
995 * set, in order to avoid causing a fault in the upcoming load.
996 */
997 fnstsw(&status);
998 if (status & 0x80)
999 fnclex();
1000
1001 /*
1002 * Load the dummy variable into the x87 stack. This mangles
1003 * the x87 stack, but we don't care since we're about to call
1004 * fxrstor() anyway.
1005 */
1006 __asm __volatile("ffree %%st(7); fld %0" : : "m" (dummy_variable));
1007 }
1008 #endif /* CPU_ENABLE_SSE */
1009
1010 static void
1011 fpurstor(addr)
1012 union savefpu *addr;
1013 {
1014
1015 #ifdef CPU_ENABLE_SSE
1016 if (cpu_fxsr) {
1017 fpu_clean_state();
1018 fxrstor(addr);
1019 } else
1020 #endif
1021 frstor(addr);
1022 }
1023
1024 #ifdef I586_CPU_XXX
1025 static long
1026 timezero(funcname, func)
1027 const char *funcname;
1028 void (*func)(void *buf, size_t len);
1029
1030 {
1031 void *buf;
1032 #define BUFSIZE 1048576
1033 long usec;
1034 struct timeval finish, start;
1035
1036 buf = malloc(BUFSIZE, M_TEMP, M_NOWAIT);
1037 if (buf == NULL)
1038 return (BUFSIZE);
1039 microtime(&start);
1040 (*func)(buf, BUFSIZE);
1041 microtime(&finish);
1042 usec = 1000000 * (finish.tv_sec - start.tv_sec) +
1043 finish.tv_usec - start.tv_usec;
1044 if (usec <= 0)
1045 usec = 1;
1046 if (bootverbose)
1047 printf("%s bandwidth = %u kBps\n", funcname,
1048 (u_int32_t)(((BUFSIZE >> 10) * 1000000) / usec));
1049 free(buf, M_TEMP);
1050 return (usec);
1051 }
1052 #endif /* I586_CPU */
1053
1054 static device_method_t npx_methods[] = {
1055 /* Device interface */
1056 DEVMETHOD(device_identify, npx_identify),
1057 DEVMETHOD(device_probe, npx_probe),
1058 DEVMETHOD(device_attach, npx_attach),
1059 DEVMETHOD(device_detach, bus_generic_detach),
1060 DEVMETHOD(device_shutdown, bus_generic_shutdown),
1061 DEVMETHOD(device_suspend, bus_generic_suspend),
1062 DEVMETHOD(device_resume, bus_generic_resume),
1063
1064 { 0, 0 }
1065 };
1066
1067 static driver_t npx_driver = {
1068 "npx",
1069 npx_methods,
1070 1, /* no softc */
1071 };
1072
1073 static devclass_t npx_devclass;
1074
1075 /*
1076 * We prefer to attach to the root nexus so that the usual case (exception 16)
1077 * doesn't describe the processor as being `on isa'.
1078 */
1079 DRIVER_MODULE(npx, nexus, npx_driver, npx_devclass, 0, 0);
1080
1081 #ifdef DEV_ISA
1082 /*
1083 * This sucks up the legacy ISA support assignments from PNPBIOS/ACPI.
1084 */
1085 static struct isa_pnp_id npxisa_ids[] = {
1086 { 0x040cd041, "Legacy ISA coprocessor support" }, /* PNP0C04 */
1087 { 0 }
1088 };
1089
1090 static int
1091 npxisa_probe(device_t dev)
1092 {
1093 int result;
1094 if ((result = ISA_PNP_PROBE(device_get_parent(dev), dev, npxisa_ids)) <= 0) {
1095 device_quiet(dev);
1096 }
1097 return(result);
1098 }
1099
1100 static int
1101 npxisa_attach(device_t dev)
1102 {
1103 return (0);
1104 }
1105
1106 static device_method_t npxisa_methods[] = {
1107 /* Device interface */
1108 DEVMETHOD(device_probe, npxisa_probe),
1109 DEVMETHOD(device_attach, npxisa_attach),
1110 DEVMETHOD(device_detach, bus_generic_detach),
1111 DEVMETHOD(device_shutdown, bus_generic_shutdown),
1112 DEVMETHOD(device_suspend, bus_generic_suspend),
1113 DEVMETHOD(device_resume, bus_generic_resume),
1114
1115 { 0, 0 }
1116 };
1117
1118 static driver_t npxisa_driver = {
1119 "npxisa",
1120 npxisa_methods,
1121 1, /* no softc */
1122 };
1123
1124 static devclass_t npxisa_devclass;
1125
1126 DRIVER_MODULE(npxisa, isa, npxisa_driver, npxisa_devclass, 0, 0);
1127 #ifndef PC98
1128 DRIVER_MODULE(npxisa, acpi, npxisa_driver, npxisa_devclass, 0, 0);
1129 #endif
1130 #endif /* DEV_ISA */
Cache object: d2ecbf61334c635ef552f7afba079f8e
|