fpu.c revision 4478
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 1.16 1994/11/06 00:58:06 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 fnsave(addr) __asm("fnsave %0" : "=m" (*addr) : "0" (*addr)) 71#define fnstcw(addr) __asm("fnstcw %0" : "=m" (*addr) : "0" (*addr)) 72#define fnstsw(addr) __asm("fnstsw %0" : "=m" (*addr) : "0" (*addr)) 73#define fp_divide_by_0() __asm("fldz; fld1; fdiv %st,%st(1); fwait") 74#define frstor(addr) __asm("frstor %0" : : "m" (*addr)) 75#define fwait() __asm("fwait") 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 82void fldcw __P((caddr_t addr)); 83void fnclex __P((void)); 84void fninit __P((void)); 85void fnsave __P((caddr_t addr)); 86void fnstcw __P((caddr_t addr)); 87void fnstsw __P((caddr_t addr)); 88void fp_divide_by_0 __P((void)); 89void frstor __P((caddr_t addr)); 90void fwait __P((void)); 91void start_emulating __P((void)); 92void stop_emulating __P((void)); 93 94#endif /* __GNUC__ */ 95 96typedef u_char bool_t; 97 98static int npxattach __P((struct isa_device *dvp)); 99static int npxprobe __P((struct isa_device *dvp)); 100static int npxprobe1 __P((struct isa_device *dvp)); 101 102struct isa_driver npxdriver = { 103 npxprobe, npxattach, "npx", 104}; 105 106u_int npx0_imask = SWI_CLOCK_MASK; 107struct proc *npxproc; 108 109static bool_t npx_ex16; 110static bool_t npx_exists; 111int hw_float; 112static struct gate_descriptor npx_idt_probeintr; 113static int npx_intrno; 114static volatile u_int npx_intrs_while_probing; 115static bool_t npx_irq13; 116static 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 probintr() can be moved to npxprobe(). 122 */ 123inthand_t probeintr; 124asm 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 140inthand_t probetrap; 141asm 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 */ 157static int 158npxprobe(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 197static int 198npxprobe1(dvp) 199 struct isa_device *dvp; 200{ 201 int control; 202 int status; 203#ifdef lint 204 npxintr(); 205#endif 206 /* 207 * Partially reset the coprocessor, if any. Some BIOS's don't reset 208 * it after a warm boot. 209 */ 210 outb(0xf1, 0); /* full reset on some systems, NOP on others */ 211 outb(0xf0, 0); /* clear BUSY# latch */ 212 /* 213 * Prepare to trap all ESC (i.e., NPX) instructions and all WAIT 214 * instructions. We must set the CR0_MP bit and use the CR0_TS 215 * bit to control the trap, because setting the CR0_EM bit does 216 * not cause WAIT instructions to trap. It's important to trap 217 * WAIT instructions - otherwise the "wait" variants of no-wait 218 * control instructions would degenerate to the "no-wait" variants 219 * after FP context switches but work correctly otherwise. It's 220 * particularly important to trap WAITs when there is no NPX - 221 * otherwise the "wait" variants would always degenerate. 222 * 223 * Try setting CR0_NE to get correct error reporting on 486DX's. 224 * Setting it should fail or do nothing on lesser processors. 225 */ 226 load_cr0(rcr0() | CR0_MP | CR0_NE); 227 /* 228 * But don't trap while we're probing. 229 */ 230 stop_emulating(); 231 /* 232 * Finish resetting the coprocessor, if any. If there is an error 233 * pending, then we may get a bogus IRQ13, but probeintr() will handle 234 * it OK. Bogus halts have never been observed, but we enabled 235 * IRQ13 and cleared the BUSY# latch early to handle them anyway. 236 */ 237 fninit(); 238 DELAY(1000); /* wait for any IRQ13 (fwait might hang) */ 239#ifdef DIAGNOSTIC 240 if (npx_intrs_while_probing != 0) 241 printf("fninit caused %u bogus npx interrupt(s)\n", 242 npx_intrs_while_probing); 243 if (npx_traps_while_probing != 0) 244 printf("fninit caused %u bogus npx trap(s)\n", 245 npx_traps_while_probing); 246#endif 247 /* 248 * Check for a status of mostly zero. 249 */ 250 status = 0x5a5a; 251 fnstsw(&status); 252 if ((status & 0xb8ff) == 0) { 253 /* 254 * Good, now check for a proper control word. 255 */ 256 control = 0x5a5a; 257 fnstcw(&control); 258 if ((control & 0x1f3f) == 0x033f) { 259 hw_float = npx_exists = 1; 260 /* 261 * We have an npx, now divide by 0 to see if exception 262 * 16 works. 263 */ 264 control &= ~(1 << 2); /* enable divide by 0 trap */ 265 fldcw(&control); 266 npx_traps_while_probing = npx_intrs_while_probing = 0; 267 fp_divide_by_0(); 268 if (npx_traps_while_probing != 0) { 269 /* 270 * Good, exception 16 works. 271 */ 272 npx_ex16 = 1; 273 dvp->id_irq = 0; /* zap the interrupt */ 274 /* 275 * special return value to flag that we do not 276 * actually use any I/O registers 277 */ 278 return (-1); 279 } 280 if (npx_intrs_while_probing != 0) { 281 /* 282 * Bad, we are stuck with IRQ13. 283 */ 284 npx_irq13 = 1; 285 npx0_imask = dvp->id_irq; /* npxattach too late */ 286 return (IO_NPXSIZE); 287 } 288 /* 289 * Worse, even IRQ13 is broken. Use emulator. 290 */ 291 } 292 } 293 /* 294 * Probe failed, but we want to get to npxattach to initialize the 295 * emulator and say that it has been installed. XXX handle devices 296 * that aren't really devices better. 297 */ 298 dvp->id_irq = 0; 299 /* 300 * special return value to flag that we do not 301 * actually use any I/O registers 302 */ 303 return (-1); 304} 305 306static struct kern_devconf kdc_npx[NNPX] = { { 307 0, 0, 0, /* filled in by dev_attach */ 308 "npx", 0, { MDDT_ISA, 0 }, 309 isa_generic_externalize, 0, 0, ISA_EXTERNALLEN, 310 &kdc_isa0, /* parent */ 311 0, /* parentdata */ 312 DC_UNKNOWN, /* not supported */ 313 "Floating-point unit" 314} }; 315 316static inline void 317npx_registerdev(struct isa_device *id) 318{ 319 if(id->id_unit) 320 kdc_npx[id->id_unit] = kdc_npx[0]; 321 kdc_npx[id->id_unit].kdc_unit = id->id_unit; 322 kdc_npx[id->id_unit].kdc_isa = id; 323 dev_attach(&kdc_npx[id->id_unit]); 324} 325 326/* 327 * Attach routine - announce which it is, and wire into system 328 */ 329int 330npxattach(dvp) 331 struct isa_device *dvp; 332{ 333 if (!npx_ex16 && !npx_irq13) { 334 if (npx_exists) { 335 printf("npx%d: Error reporting broken, using 387 emulator\n",dvp->id_unit); 336 hw_float = npx_exists = 0; 337 } else { 338 printf("npx%d: 387 Emulator\n",dvp->id_unit); 339 } 340 } 341 npxinit(__INITIAL_NPXCW__); 342 npx_registerdev(dvp); 343 return (1); /* XXX unused */ 344} 345 346/* 347 * Initialize floating point unit. 348 */ 349void 350npxinit(control) 351 u_int control; 352{ 353 struct save87 dummy; 354 355 if (!npx_exists) 356 return; 357 /* 358 * fninit has the same h/w bugs as fnsave. Use the detoxified 359 * fnsave to throw away any junk in the fpu. fnsave initializes 360 * the fpu and sets npxproc = NULL as important side effects. 361 */ 362 npxsave(&dummy); 363 stop_emulating(); 364 fldcw(&control); 365 if (curpcb != NULL) 366 fnsave(&curpcb->pcb_savefpu); 367 start_emulating(); 368} 369 370/* 371 * Free coprocessor (if we have it). 372 */ 373void 374npxexit(p) 375 struct proc *p; 376{ 377 378 if (p == npxproc) 379 npxsave(&curpcb->pcb_savefpu); 380 if (npx_exists) { 381 u_int masked_exceptions; 382 383 masked_exceptions = curpcb->pcb_savefpu.sv_env.en_cw 384 & curpcb->pcb_savefpu.sv_env.en_sw & 0x7f; 385 /* 386 * Overflow, divde by 0, and invalid operand would have 387 * caused a trap in 1.1.5. 388 */ 389 if (masked_exceptions & 0x0d) 390 log(LOG_ERR, 391 "pid %d (%s) exited with masked floating point exceptions 0x%02x\n", 392 p->p_pid, p->p_comm, masked_exceptions); 393 } 394} 395 396/* 397 * Record the FPU state and reinitialize it all except for the control word. 398 * Then generate a SIGFPE. 399 * 400 * Reinitializing the state allows naive SIGFPE handlers to longjmp without 401 * doing any fixups. 402 * 403 * XXX there is currently no way to pass the full error state to signal 404 * handlers, and if this is a nested interrupt there is no way to pass even 405 * a status code! So there is no way to have a non-naive SIGFPE handler. At 406 * best a handler could do an fninit followed by an fldcw of a static value. 407 * fnclex would be of little use because it would leave junk on the FPU stack. 408 * Returning from the handler would be even less safe than usual because 409 * IRQ13 exception handling makes exceptions even less precise than usual. 410 */ 411void 412npxintr(frame) 413 struct intrframe frame; 414{ 415 int code; 416 417 if (npxproc == NULL || !npx_exists) { 418 /* XXX no %p in stand/printf.c. Cast to quiet gcc -Wall. */ 419 printf("npxintr: npxproc = %lx, curproc = %lx, npx_exists = %d\n", 420 (u_long) npxproc, (u_long) curproc, npx_exists); 421 panic("npxintr from nowhere"); 422 } 423 if (npxproc != curproc) { 424 printf("npxintr: npxproc = %lx, curproc = %lx, npx_exists = %d\n", 425 (u_long) npxproc, (u_long) curproc, npx_exists); 426 panic("npxintr from non-current process"); 427 } 428 /* 429 * Save state. This does an implied fninit. It had better not halt 430 * the cpu or we'll hang. 431 */ 432 outb(0xf0, 0); 433 fnsave(&curpcb->pcb_savefpu); 434 fwait(); 435 /* 436 * Restore control word (was clobbered by fnsave). 437 */ 438 fldcw(&curpcb->pcb_savefpu.sv_env.en_cw); 439 fwait(); 440 /* 441 * Remember the exception status word and tag word. The current 442 * (almost fninit'ed) fpu state is in the fpu and the exception 443 * state just saved will soon be junk. However, the implied fninit 444 * doesn't change the error pointers or register contents, and we 445 * preserved the control word and will copy the status and tag 446 * words, so the complete exception state can be recovered. 447 */ 448 curpcb->pcb_savefpu.sv_ex_sw = curpcb->pcb_savefpu.sv_env.en_sw; 449 curpcb->pcb_savefpu.sv_ex_tw = curpcb->pcb_savefpu.sv_env.en_tw; 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(npxproc, SIGFPE); 493 } 494} 495 496/* 497 * Implement device not available (DNA) exception 498 * 499 * It would be better to switch FP context here (only). This would require 500 * saving the state in the proc table instead of in the pcb. 501 */ 502int 503npxdna() 504{ 505 if (!npx_exists) 506 return (0); 507 if (npxproc != NULL) { 508 printf("npxdna: npxproc = %lx, curproc = %lx\n", 509 (u_long) npxproc, (u_long) curproc); 510 panic("npxdna"); 511 } 512 stop_emulating(); 513 /* 514 * Record new context early in case frstor causes an IRQ13. 515 */ 516 npxproc = curproc; 517 /* 518 * The following frstor may cause an IRQ13 when the state being 519 * restored has a pending error. The error will appear to have been 520 * triggered by the current (npx) user instruction even when that 521 * instruction is a no-wait instruction that should not trigger an 522 * error (e.g., fnclex). On at least one 486 system all of the 523 * no-wait instructions are broken the same as frstor, so our 524 * treatment does not amplify the breakage. On at least one 525 * 386/Cyrix 387 system, fnclex works correctly while frstor and 526 * fnsave are broken, so our treatment breaks fnclex if it is the 527 * first FPU instruction after a context switch. 528 */ 529 frstor(&curpcb->pcb_savefpu); 530 531 return (1); 532} 533 534/* 535 * Wrapper for fnsave instruction to handle h/w bugs. If there is an error 536 * pending, then fnsave generates a bogus IRQ13 on some systems. Force 537 * any IRQ13 to be handled immediately, and then ignore it. This routine is 538 * often called at splhigh so it must not use many system services. In 539 * particular, it's much easier to install a special handler than to 540 * guarantee that it's safe to use npxintr() and its supporting code. 541 */ 542void 543npxsave(addr) 544 struct save87 *addr; 545{ 546 u_char icu1_mask; 547 u_char icu2_mask; 548 u_char old_icu1_mask; 549 u_char old_icu2_mask; 550 struct gate_descriptor save_idt_npxintr; 551 552 disable_intr(); 553 old_icu1_mask = inb(IO_ICU1 + 1); 554 old_icu2_mask = inb(IO_ICU2 + 1); 555 save_idt_npxintr = idt[npx_intrno]; 556 outb(IO_ICU1 + 1, old_icu1_mask & ~(IRQ_SLAVE | npx0_imask)); 557 outb(IO_ICU2 + 1, old_icu2_mask & ~(npx0_imask >> 8)); 558 idt[npx_intrno] = npx_idt_probeintr; 559 enable_intr(); 560 stop_emulating(); 561 fnsave(addr); 562 fwait(); 563 start_emulating(); 564 npxproc = NULL; 565 disable_intr(); 566 icu1_mask = inb(IO_ICU1 + 1); /* masks may have changed */ 567 icu2_mask = inb(IO_ICU2 + 1); 568 outb(IO_ICU1 + 1, 569 (icu1_mask & ~npx0_imask) | (old_icu1_mask & npx0_imask)); 570 outb(IO_ICU2 + 1, 571 (icu2_mask & ~(npx0_imask >> 8)) 572 | (old_icu2_mask & (npx0_imask >> 8))); 573 idt[npx_intrno] = save_idt_npxintr; 574 enable_intr(); /* back to usual state */ 575} 576 577#endif /* NNPX > 0 */ 578