fpu.c revision 81493
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 * $FreeBSD: head/sys/amd64/amd64/fpu.c 81493 2001-08-10 22:53:32Z jhb $ 36 */ 37 38#include "opt_cpu.h" 39#include "opt_debug_npx.h" 40#include "opt_math_emulate.h" 41 42#include <sys/param.h> 43#include <sys/systm.h> 44#include <sys/bus.h> 45#include <sys/kernel.h> 46#include <sys/lock.h> 47#include <sys/malloc.h> 48#include <sys/module.h> 49#include <sys/mutex.h> 50#include <sys/mutex.h> 51#include <sys/proc.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#ifndef SMP 62#include <machine/asmacros.h> 63#endif 64#include <machine/cputypes.h> 65#include <machine/frame.h> 66#include <machine/md_var.h> 67#include <machine/pcb.h> 68#include <machine/psl.h> 69#ifndef SMP 70#include <machine/clock.h> 71#endif 72#include <machine/resource.h> 73#include <machine/specialreg.h> 74#include <machine/segments.h> 75 76#ifndef SMP 77#include <i386/isa/icu.h> 78#include <i386/isa/intr_machdep.h> 79#include <i386/isa/isa.h> 80#endif 81#include <isa/isavar.h> 82 83/* 84 * 387 and 287 Numeric Coprocessor Extension (NPX) Driver. 85 */ 86 87/* Configuration flags. */ 88#define NPX_DISABLE_I586_OPTIMIZED_BCOPY (1 << 0) 89#define NPX_DISABLE_I586_OPTIMIZED_BZERO (1 << 1) 90#define NPX_DISABLE_I586_OPTIMIZED_COPYIO (1 << 2) 91#define NPX_PREFER_EMULATOR (1 << 3) 92 93#ifdef __GNUC__ 94 95#define fldcw(addr) __asm("fldcw %0" : : "m" (*(addr))) 96#define fnclex() __asm("fnclex") 97#define fninit() __asm("fninit") 98#define fnsave(addr) __asm __volatile("fnsave %0" : "=m" (*(addr))) 99#define fnstcw(addr) __asm __volatile("fnstcw %0" : "=m" (*(addr))) 100#define fnstsw(addr) __asm __volatile("fnstsw %0" : "=m" (*(addr))) 101#define fp_divide_by_0() __asm("fldz; fld1; fdiv %st,%st(1); fnop") 102#define frstor(addr) __asm("frstor %0" : : "m" (*(addr))) 103#define fxrstor(addr) __asm("fxrstor %0" : : "m" (*(addr))) 104#define fxsave(addr) __asm __volatile("fxsave %0" : "=m" (*(addr))) 105#define start_emulating() __asm("smsw %%ax; orb %0,%%al; lmsw %%ax" \ 106 : : "n" (CR0_TS) : "ax") 107#define stop_emulating() __asm("clts") 108 109#else /* not __GNUC__ */ 110 111void fldcw __P((caddr_t addr)); 112void fnclex __P((void)); 113void fninit __P((void)); 114void fnsave __P((caddr_t addr)); 115void fnstcw __P((caddr_t addr)); 116void fnstsw __P((caddr_t addr)); 117void fp_divide_by_0 __P((void)); 118void frstor __P((caddr_t addr)); 119void fxsave __P((caddr_t addr)); 120void fxrstor __P((caddr_t addr)); 121void start_emulating __P((void)); 122void stop_emulating __P((void)); 123 124#endif /* __GNUC__ */ 125 126#ifdef CPU_ENABLE_SSE 127#define GET_FPU_CW(proc) \ 128 (cpu_fxsr ? \ 129 (proc)->p_addr->u_pcb.pcb_save.sv_xmm.sv_env.en_cw : \ 130 (proc)->p_addr->u_pcb.pcb_save.sv_87.sv_env.en_cw) 131#define GET_FPU_SW(proc) \ 132 (cpu_fxsr ? \ 133 (proc)->p_addr->u_pcb.pcb_save.sv_xmm.sv_env.en_sw : \ 134 (proc)->p_addr->u_pcb.pcb_save.sv_87.sv_env.en_sw) 135#define GET_FPU_EXSW_PTR(pcb) \ 136 (cpu_fxsr ? \ 137 &(pcb)->pcb_save.sv_xmm.sv_ex_sw : \ 138 &(pcb)->pcb_save.sv_87.sv_ex_sw) 139#else /* CPU_ENABLE_SSE */ 140#define GET_FPU_CW(proc) \ 141 (proc->p_addr->u_pcb.pcb_save.sv_87.sv_env.en_cw) 142#define GET_FPU_SW(proc) \ 143 (proc->p_addr->u_pcb.pcb_save.sv_87.sv_env.en_sw) 144#define GET_FPU_EXSW_PTR(pcb) \ 145 (&(pcb)->pcb_save.sv_87.sv_ex_sw) 146#endif /* CPU_ENABLE_SSE */ 147 148typedef u_char bool_t; 149 150static int npx_attach __P((device_t dev)); 151static void npx_identify __P((driver_t *driver, device_t parent)); 152#ifndef SMP 153static void npx_intr __P((void *)); 154#endif 155static int npx_probe __P((device_t dev)); 156static int npx_probe1 __P((device_t dev)); 157static void fpusave __P((union savefpu *)); 158static void fpurstor __P((union savefpu *)); 159#ifdef I586_CPU_XXX 160static long timezero __P((const char *funcname, 161 void (*func)(void *buf, size_t len))); 162#endif /* I586_CPU */ 163 164int hw_float; /* XXX currently just alias for npx_exists */ 165 166SYSCTL_INT(_hw,HW_FLOATINGPT, floatingpoint, 167 CTLFLAG_RD, &hw_float, 0, 168 "Floatingpoint instructions executed in hardware"); 169 170#ifndef SMP 171static volatile u_int npx_intrs_while_probing; 172static volatile u_int npx_traps_while_probing; 173#endif 174 175static bool_t npx_ex16; 176static bool_t npx_exists; 177static bool_t npx_irq13; 178static int npx_irq; /* irq number */ 179 180#ifndef SMP 181/* 182 * Special interrupt handlers. Someday intr0-intr15 will be used to count 183 * interrupts. We'll still need a special exception 16 handler. The busy 184 * latch stuff in probeintr() can be moved to npxprobe(). 185 */ 186inthand_t probeintr; 187__asm(" \n\ 188 .text \n\ 189 .p2align 2,0x90 \n\ 190 .type " __XSTRING(CNAME(probeintr)) ",@function \n\ 191" __XSTRING(CNAME(probeintr)) ": \n\ 192 ss \n\ 193 incl " __XSTRING(CNAME(npx_intrs_while_probing)) " \n\ 194 pushl %eax \n\ 195 movb $0x20,%al # EOI (asm in strings loses cpp features) \n\ 196 outb %al,$0xa0 # IO_ICU2 \n\ 197 outb %al,$0x20 # IO_ICU1 \n\ 198 movb $0,%al \n\ 199 outb %al,$0xf0 # clear BUSY# latch \n\ 200 popl %eax \n\ 201 iret \n\ 202"); 203 204inthand_t probetrap; 205__asm(" \n\ 206 .text \n\ 207 .p2align 2,0x90 \n\ 208 .type " __XSTRING(CNAME(probetrap)) ",@function \n\ 209" __XSTRING(CNAME(probetrap)) ": \n\ 210 ss \n\ 211 incl " __XSTRING(CNAME(npx_traps_while_probing)) " \n\ 212 fnclex \n\ 213 iret \n\ 214"); 215#endif /* SMP */ 216 217/* 218 * Identify routine. Create a connection point on our parent for probing. 219 */ 220static void 221npx_identify(driver, parent) 222 driver_t *driver; 223 device_t parent; 224{ 225 device_t child; 226 227 child = BUS_ADD_CHILD(parent, 0, "npx", 0); 228 if (child == NULL) 229 panic("npx_identify"); 230} 231 232#ifndef SMP 233/* 234 * Do minimal handling of npx interrupts to convert them to traps. 235 */ 236static void 237npx_intr(dummy) 238 void *dummy; 239{ 240 struct proc *p; 241 242 /* 243 * The BUSY# latch must be cleared in all cases so that the next 244 * unmasked npx exception causes an interrupt. 245 */ 246 outb(0xf0, 0); 247 248 /* 249 * npxproc is normally non-null here. In that case, schedule an 250 * AST to finish the exception handling in the correct context 251 * (this interrupt may occur after the process has entered the 252 * kernel via a syscall or an interrupt). Otherwise, the npx 253 * state of the process that caused this interrupt must have been 254 * pushed to the process' pcb, and clearing of the busy latch 255 * above has finished the (essentially null) handling of this 256 * interrupt. Control will eventually return to the instruction 257 * that caused it and it will repeat. We will eventually (usually 258 * soon) win the race to handle the interrupt properly. 259 */ 260 p = PCPU_GET(npxproc); 261 if (p != NULL) { 262 p->p_addr->u_pcb.pcb_flags |= PCB_NPXTRAP; 263 mtx_lock_spin(&sched_lock); 264 p->p_sflag |= PS_ASTPENDING; 265 mtx_unlock_spin(&sched_lock); 266 } 267} 268 269/* 270 * XXX these "local" variables of npx_probe() are non-local so that 271 * npxprobe1() can abuse them. 272 */ 273static int npx_intrno; 274static struct gate_descriptor save_idt_npxintr; 275#endif /* !SMP */ 276 277/* 278 * Probe routine. Initialize cr0 to give correct behaviour for [f]wait 279 * whether the device exists or not (XXX should be elsewhere). Set flags 280 * to tell npxattach() what to do. Modify device struct if npx doesn't 281 * need to use interrupts. Return 1 if device exists. 282 */ 283static int 284npx_probe(dev) 285 device_t dev; 286{ 287#ifdef SMP 288 289 if (resource_int_value("npx", 0, "irq", &npx_irq) != 0) 290 npx_irq = 13; 291 return npx_probe1(dev); 292 293#else /* SMP */ 294 295 int result; 296 critical_t savecrit; 297 u_char save_icu1_mask; 298 u_char save_icu2_mask; 299 struct gate_descriptor save_idt_npxtrap; 300 /* 301 * This routine is now just a wrapper for npxprobe1(), to install 302 * special npx interrupt and trap handlers, to enable npx interrupts 303 * and to disable other interrupts. Someday isa_configure() will 304 * install suitable handlers and run with interrupts enabled so we 305 * won't need to do so much here. 306 */ 307 if (resource_int_value("npx", 0, "irq", &npx_irq) != 0) 308 npx_irq = 13; 309 npx_intrno = NRSVIDT + npx_irq; 310 savecrit = critical_enter(); 311 save_icu1_mask = inb(IO_ICU1 + 1); 312 save_icu2_mask = inb(IO_ICU2 + 1); 313 save_idt_npxintr = idt[npx_intrno]; 314 save_idt_npxtrap = idt[16]; 315 outb(IO_ICU1 + 1, ~IRQ_SLAVE); 316 outb(IO_ICU2 + 1, ~(1 << (npx_irq - 8))); 317 setidt(16, probetrap, SDT_SYS386TGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL)); 318 setidt(npx_intrno, probeintr, SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL)); 319 320 /* 321 * XXX This looks highly bogus, but it appears that npc_probe1 322 * needs interrupts enabled. Does this make any difference 323 * here? 324 */ 325 critical_exit(savecrit); 326 result = npx_probe1(dev); 327 savecrit = critical_enter(); 328 outb(IO_ICU1 + 1, save_icu1_mask); 329 outb(IO_ICU2 + 1, save_icu2_mask); 330 idt[npx_intrno] = save_idt_npxintr; 331 idt[16] = save_idt_npxtrap; 332 critical_exit(savecrit); 333 return (result); 334 335#endif /* SMP */ 336} 337 338static int 339npx_probe1(dev) 340 device_t dev; 341{ 342#ifndef SMP 343 u_short control; 344 u_short status; 345#endif 346 347 /* 348 * Partially reset the coprocessor, if any. Some BIOS's don't reset 349 * it after a warm boot. 350 */ 351 outb(0xf1, 0); /* full reset on some systems, NOP on others */ 352 outb(0xf0, 0); /* clear BUSY# latch */ 353 /* 354 * Prepare to trap all ESC (i.e., NPX) instructions and all WAIT 355 * instructions. We must set the CR0_MP bit and use the CR0_TS 356 * bit to control the trap, because setting the CR0_EM bit does 357 * not cause WAIT instructions to trap. It's important to trap 358 * WAIT instructions - otherwise the "wait" variants of no-wait 359 * control instructions would degenerate to the "no-wait" variants 360 * after FP context switches but work correctly otherwise. It's 361 * particularly important to trap WAITs when there is no NPX - 362 * otherwise the "wait" variants would always degenerate. 363 * 364 * Try setting CR0_NE to get correct error reporting on 486DX's. 365 * Setting it should fail or do nothing on lesser processors. 366 */ 367 load_cr0(rcr0() | CR0_MP | CR0_NE); 368 /* 369 * But don't trap while we're probing. 370 */ 371 stop_emulating(); 372 /* 373 * Finish resetting the coprocessor, if any. If there is an error 374 * pending, then we may get a bogus IRQ13, but probeintr() will handle 375 * it OK. Bogus halts have never been observed, but we enabled 376 * IRQ13 and cleared the BUSY# latch early to handle them anyway. 377 */ 378 fninit(); 379 380#ifdef SMP 381 /* 382 * Exception 16 MUST work for SMP. 383 */ 384 npx_irq13 = 0; 385 npx_ex16 = hw_float = npx_exists = 1; 386 device_set_desc(dev, "math processor"); 387 return (0); 388 389#else /* !SMP */ 390 device_set_desc(dev, "math processor"); 391 392 /* 393 * Don't use fwait here because it might hang. 394 * Don't use fnop here because it usually hangs if there is no FPU. 395 */ 396 DELAY(1000); /* wait for any IRQ13 */ 397#ifdef DIAGNOSTIC 398 if (npx_intrs_while_probing != 0) 399 printf("fninit caused %u bogus npx interrupt(s)\n", 400 npx_intrs_while_probing); 401 if (npx_traps_while_probing != 0) 402 printf("fninit caused %u bogus npx trap(s)\n", 403 npx_traps_while_probing); 404#endif 405 /* 406 * Check for a status of mostly zero. 407 */ 408 status = 0x5a5a; 409 fnstsw(&status); 410 if ((status & 0xb8ff) == 0) { 411 /* 412 * Good, now check for a proper control word. 413 */ 414 control = 0x5a5a; 415 fnstcw(&control); 416 if ((control & 0x1f3f) == 0x033f) { 417 hw_float = npx_exists = 1; 418 /* 419 * We have an npx, now divide by 0 to see if exception 420 * 16 works. 421 */ 422 control &= ~(1 << 2); /* enable divide by 0 trap */ 423 fldcw(&control); 424 npx_traps_while_probing = npx_intrs_while_probing = 0; 425 fp_divide_by_0(); 426 if (npx_traps_while_probing != 0) { 427 /* 428 * Good, exception 16 works. 429 */ 430 npx_ex16 = 1; 431 return (0); 432 } 433 if (npx_intrs_while_probing != 0) { 434 int rid; 435 struct resource *r; 436 void *intr; 437 /* 438 * Bad, we are stuck with IRQ13. 439 */ 440 npx_irq13 = 1; 441 442 /* 443 * We allocate these resources permanently, 444 * so there is no need to keep track of them. 445 */ 446 rid = 0; 447 r = bus_alloc_resource(dev, SYS_RES_IOPORT, 448 &rid, IO_NPX, IO_NPX, 449 IO_NPXSIZE, RF_ACTIVE); 450 if (r == 0) 451 panic("npx: can't get ports"); 452 rid = 0; 453 r = bus_alloc_resource(dev, SYS_RES_IRQ, 454 &rid, npx_irq, npx_irq, 455 1, RF_ACTIVE); 456 if (r == 0) 457 panic("npx: can't get IRQ"); 458 BUS_SETUP_INTR(device_get_parent(dev), 459 dev, r, 460 INTR_TYPE_MISC | INTR_FAST, 461 npx_intr, 0, &intr); 462 if (intr == 0) 463 panic("npx: can't create intr"); 464 465 /* 466 * XXX BUS_SETUP_INTR() has changed 467 * idt[npx_intrno] to point to Xfastintr0 468 * instead of Xfastintr0. Adjust 469 * save_idt_npxintr so that npxprobe() 470 * doesn't undo this. 471 */ 472 save_idt_npxintr = idt[npx_intrno]; 473 474 return (0); 475 } 476 /* 477 * Worse, even IRQ13 is broken. Use emulator. 478 */ 479 } 480 } 481 /* 482 * Probe failed, but we want to get to npxattach to initialize the 483 * emulator and say that it has been installed. XXX handle devices 484 * that aren't really devices better. 485 */ 486 return (0); 487#endif /* SMP */ 488} 489 490/* 491 * Attach routine - announce which it is, and wire into system 492 */ 493int 494npx_attach(dev) 495 device_t dev; 496{ 497 int flags; 498 499 if (resource_int_value("npx", 0, "flags", &flags) != 0) 500 flags = 0; 501 502 if (flags) 503 device_printf(dev, "flags 0x%x ", flags); 504 if (npx_irq13) { 505 device_printf(dev, "using IRQ 13 interface\n"); 506 } else { 507#if defined(MATH_EMULATE) || defined(GPL_MATH_EMULATE) 508 if (npx_ex16) { 509 if (!(flags & NPX_PREFER_EMULATOR)) 510 device_printf(dev, "INT 16 interface\n"); 511 else { 512 device_printf(dev, "FPU exists, but flags request " 513 "emulator\n"); 514 hw_float = npx_exists = 0; 515 } 516 } else if (npx_exists) { 517 device_printf(dev, "error reporting broken; using 387 emulator\n"); 518 hw_float = npx_exists = 0; 519 } else 520 device_printf(dev, "387 emulator\n"); 521#else 522 if (npx_ex16) { 523 device_printf(dev, "INT 16 interface\n"); 524 if (flags & NPX_PREFER_EMULATOR) { 525 device_printf(dev, "emulator requested, but none compiled " 526 "into kernel, using FPU\n"); 527 } 528 } else 529 device_printf(dev, "no 387 emulator in kernel and no FPU!\n"); 530#endif 531 } 532 npxinit(__INITIAL_NPXCW__); 533 534#ifdef I586_CPU_XXX 535 if (cpu_class == CPUCLASS_586 && npx_ex16 && npx_exists && 536 timezero("i586_bzero()", i586_bzero) < 537 timezero("bzero()", bzero) * 4 / 5) { 538 if (!(flags & NPX_DISABLE_I586_OPTIMIZED_BCOPY)) { 539 bcopy_vector = i586_bcopy; 540 ovbcopy_vector = i586_bcopy; 541 } 542 if (!(flags & NPX_DISABLE_I586_OPTIMIZED_BZERO)) 543 bzero = i586_bzero; 544 if (!(flags & NPX_DISABLE_I586_OPTIMIZED_COPYIO)) { 545 copyin_vector = i586_copyin; 546 copyout_vector = i586_copyout; 547 } 548 } 549#endif 550 551 return (0); /* XXX unused */ 552} 553 554/* 555 * Initialize floating point unit. 556 */ 557void 558npxinit(control) 559 u_short control; 560{ 561 static union savefpu dummy; 562 critical_t savecrit; 563 564 if (!npx_exists) 565 return; 566 /* 567 * fninit has the same h/w bugs as fnsave. Use the detoxified 568 * fnsave to throw away any junk in the fpu. npxsave() initializes 569 * the fpu and sets npxproc = NULL as important side effects. 570 */ 571 savecrit = critical_enter(); 572 npxsave(&dummy); 573 stop_emulating(); 574 fldcw(&control); 575 if (PCPU_GET(curpcb) != NULL) 576 fpusave(&PCPU_GET(curpcb)->pcb_save); 577 start_emulating(); 578 critical_exit(savecrit); 579} 580 581/* 582 * Free coprocessor (if we have it). 583 */ 584void 585npxexit(p) 586 struct proc *p; 587{ 588 critical_t savecrit; 589 590 savecrit = critical_enter(); 591 if (p == PCPU_GET(npxproc)) 592 npxsave(&PCPU_GET(curpcb)->pcb_save); 593 critical_exit(savecrit); 594#ifdef NPX_DEBUG 595 if (npx_exists) { 596 u_int masked_exceptions; 597 598 masked_exceptions = PCPU_GET(curpcb)->pcb_save.sv_87.sv_env.en_cw 599 & PCPU_GET(curpcb)->pcb_save.sv_87.sv_env.en_sw & 0x7f; 600 /* 601 * Log exceptions that would have trapped with the old 602 * control word (overflow, divide by 0, and invalid operand). 603 */ 604 if (masked_exceptions & 0x0d) 605 log(LOG_ERR, 606 "pid %d (%s) exited with masked floating point exceptions 0x%02x\n", 607 p->p_pid, p->p_comm, masked_exceptions); 608 } 609#endif 610} 611 612/* 613 * The following mechanism is used to ensure that the FPE_... value 614 * that is passed as a trapcode to the signal handler of the user 615 * process does not have more than one bit set. 616 * 617 * Multiple bits may be set if the user process modifies the control 618 * word while a status word bit is already set. While this is a sign 619 * of bad coding, we have no choise than to narrow them down to one 620 * bit, since we must not send a trapcode that is not exactly one of 621 * the FPE_ macros. 622 * 623 * The mechanism has a static table with 127 entries. Each combination 624 * of the 7 FPU status word exception bits directly translates to a 625 * position in this table, where a single FPE_... value is stored. 626 * This FPE_... value stored there is considered the "most important" 627 * of the exception bits and will be sent as the signal code. The 628 * precedence of the bits is based upon Intel Document "Numerical 629 * Applications", Chapter "Special Computational Situations". 630 * 631 * The macro to choose one of these values does these steps: 1) Throw 632 * away status word bits that cannot be masked. 2) Throw away the bits 633 * currently masked in the control word, assuming the user isn't 634 * interested in them anymore. 3) Reinsert status word bit 7 (stack 635 * fault) if it is set, which cannot be masked but must be presered. 636 * 4) Use the remaining bits to point into the trapcode table. 637 * 638 * The 6 maskable bits in order of their preference, as stated in the 639 * above referenced Intel manual: 640 * 1 Invalid operation (FP_X_INV) 641 * 1a Stack underflow 642 * 1b Stack overflow 643 * 1c Operand of unsupported format 644 * 1d SNaN operand. 645 * 2 QNaN operand (not an exception, irrelavant here) 646 * 3 Any other invalid-operation not mentioned above or zero divide 647 * (FP_X_INV, FP_X_DZ) 648 * 4 Denormal operand (FP_X_DNML) 649 * 5 Numeric over/underflow (FP_X_OFL, FP_X_UFL) 650 * 6 Inexact result (FP_X_IMP) 651 */ 652static char fpetable[128] = { 653 0, 654 FPE_FLTINV, /* 1 - INV */ 655 FPE_FLTUND, /* 2 - DNML */ 656 FPE_FLTINV, /* 3 - INV | DNML */ 657 FPE_FLTDIV, /* 4 - DZ */ 658 FPE_FLTINV, /* 5 - INV | DZ */ 659 FPE_FLTDIV, /* 6 - DNML | DZ */ 660 FPE_FLTINV, /* 7 - INV | DNML | DZ */ 661 FPE_FLTOVF, /* 8 - OFL */ 662 FPE_FLTINV, /* 9 - INV | OFL */ 663 FPE_FLTUND, /* A - DNML | OFL */ 664 FPE_FLTINV, /* B - INV | DNML | OFL */ 665 FPE_FLTDIV, /* C - DZ | OFL */ 666 FPE_FLTINV, /* D - INV | DZ | OFL */ 667 FPE_FLTDIV, /* E - DNML | DZ | OFL */ 668 FPE_FLTINV, /* F - INV | DNML | DZ | OFL */ 669 FPE_FLTUND, /* 10 - UFL */ 670 FPE_FLTINV, /* 11 - INV | UFL */ 671 FPE_FLTUND, /* 12 - DNML | UFL */ 672 FPE_FLTINV, /* 13 - INV | DNML | UFL */ 673 FPE_FLTDIV, /* 14 - DZ | UFL */ 674 FPE_FLTINV, /* 15 - INV | DZ | UFL */ 675 FPE_FLTDIV, /* 16 - DNML | DZ | UFL */ 676 FPE_FLTINV, /* 17 - INV | DNML | DZ | UFL */ 677 FPE_FLTOVF, /* 18 - OFL | UFL */ 678 FPE_FLTINV, /* 19 - INV | OFL | UFL */ 679 FPE_FLTUND, /* 1A - DNML | OFL | UFL */ 680 FPE_FLTINV, /* 1B - INV | DNML | OFL | UFL */ 681 FPE_FLTDIV, /* 1C - DZ | OFL | UFL */ 682 FPE_FLTINV, /* 1D - INV | DZ | OFL | UFL */ 683 FPE_FLTDIV, /* 1E - DNML | DZ | OFL | UFL */ 684 FPE_FLTINV, /* 1F - INV | DNML | DZ | OFL | UFL */ 685 FPE_FLTRES, /* 20 - IMP */ 686 FPE_FLTINV, /* 21 - INV | IMP */ 687 FPE_FLTUND, /* 22 - DNML | IMP */ 688 FPE_FLTINV, /* 23 - INV | DNML | IMP */ 689 FPE_FLTDIV, /* 24 - DZ | IMP */ 690 FPE_FLTINV, /* 25 - INV | DZ | IMP */ 691 FPE_FLTDIV, /* 26 - DNML | DZ | IMP */ 692 FPE_FLTINV, /* 27 - INV | DNML | DZ | IMP */ 693 FPE_FLTOVF, /* 28 - OFL | IMP */ 694 FPE_FLTINV, /* 29 - INV | OFL | IMP */ 695 FPE_FLTUND, /* 2A - DNML | OFL | IMP */ 696 FPE_FLTINV, /* 2B - INV | DNML | OFL | IMP */ 697 FPE_FLTDIV, /* 2C - DZ | OFL | IMP */ 698 FPE_FLTINV, /* 2D - INV | DZ | OFL | IMP */ 699 FPE_FLTDIV, /* 2E - DNML | DZ | OFL | IMP */ 700 FPE_FLTINV, /* 2F - INV | DNML | DZ | OFL | IMP */ 701 FPE_FLTUND, /* 30 - UFL | IMP */ 702 FPE_FLTINV, /* 31 - INV | UFL | IMP */ 703 FPE_FLTUND, /* 32 - DNML | UFL | IMP */ 704 FPE_FLTINV, /* 33 - INV | DNML | UFL | IMP */ 705 FPE_FLTDIV, /* 34 - DZ | UFL | IMP */ 706 FPE_FLTINV, /* 35 - INV | DZ | UFL | IMP */ 707 FPE_FLTDIV, /* 36 - DNML | DZ | UFL | IMP */ 708 FPE_FLTINV, /* 37 - INV | DNML | DZ | UFL | IMP */ 709 FPE_FLTOVF, /* 38 - OFL | UFL | IMP */ 710 FPE_FLTINV, /* 39 - INV | OFL | UFL | IMP */ 711 FPE_FLTUND, /* 3A - DNML | OFL | UFL | IMP */ 712 FPE_FLTINV, /* 3B - INV | DNML | OFL | UFL | IMP */ 713 FPE_FLTDIV, /* 3C - DZ | OFL | UFL | IMP */ 714 FPE_FLTINV, /* 3D - INV | DZ | OFL | UFL | IMP */ 715 FPE_FLTDIV, /* 3E - DNML | DZ | OFL | UFL | IMP */ 716 FPE_FLTINV, /* 3F - INV | DNML | DZ | OFL | UFL | IMP */ 717 FPE_FLTSUB, /* 40 - STK */ 718 FPE_FLTSUB, /* 41 - INV | STK */ 719 FPE_FLTUND, /* 42 - DNML | STK */ 720 FPE_FLTSUB, /* 43 - INV | DNML | STK */ 721 FPE_FLTDIV, /* 44 - DZ | STK */ 722 FPE_FLTSUB, /* 45 - INV | DZ | STK */ 723 FPE_FLTDIV, /* 46 - DNML | DZ | STK */ 724 FPE_FLTSUB, /* 47 - INV | DNML | DZ | STK */ 725 FPE_FLTOVF, /* 48 - OFL | STK */ 726 FPE_FLTSUB, /* 49 - INV | OFL | STK */ 727 FPE_FLTUND, /* 4A - DNML | OFL | STK */ 728 FPE_FLTSUB, /* 4B - INV | DNML | OFL | STK */ 729 FPE_FLTDIV, /* 4C - DZ | OFL | STK */ 730 FPE_FLTSUB, /* 4D - INV | DZ | OFL | STK */ 731 FPE_FLTDIV, /* 4E - DNML | DZ | OFL | STK */ 732 FPE_FLTSUB, /* 4F - INV | DNML | DZ | OFL | STK */ 733 FPE_FLTUND, /* 50 - UFL | STK */ 734 FPE_FLTSUB, /* 51 - INV | UFL | STK */ 735 FPE_FLTUND, /* 52 - DNML | UFL | STK */ 736 FPE_FLTSUB, /* 53 - INV | DNML | UFL | STK */ 737 FPE_FLTDIV, /* 54 - DZ | UFL | STK */ 738 FPE_FLTSUB, /* 55 - INV | DZ | UFL | STK */ 739 FPE_FLTDIV, /* 56 - DNML | DZ | UFL | STK */ 740 FPE_FLTSUB, /* 57 - INV | DNML | DZ | UFL | STK */ 741 FPE_FLTOVF, /* 58 - OFL | UFL | STK */ 742 FPE_FLTSUB, /* 59 - INV | OFL | UFL | STK */ 743 FPE_FLTUND, /* 5A - DNML | OFL | UFL | STK */ 744 FPE_FLTSUB, /* 5B - INV | DNML | OFL | UFL | STK */ 745 FPE_FLTDIV, /* 5C - DZ | OFL | UFL | STK */ 746 FPE_FLTSUB, /* 5D - INV | DZ | OFL | UFL | STK */ 747 FPE_FLTDIV, /* 5E - DNML | DZ | OFL | UFL | STK */ 748 FPE_FLTSUB, /* 5F - INV | DNML | DZ | OFL | UFL | STK */ 749 FPE_FLTRES, /* 60 - IMP | STK */ 750 FPE_FLTSUB, /* 61 - INV | IMP | STK */ 751 FPE_FLTUND, /* 62 - DNML | IMP | STK */ 752 FPE_FLTSUB, /* 63 - INV | DNML | IMP | STK */ 753 FPE_FLTDIV, /* 64 - DZ | IMP | STK */ 754 FPE_FLTSUB, /* 65 - INV | DZ | IMP | STK */ 755 FPE_FLTDIV, /* 66 - DNML | DZ | IMP | STK */ 756 FPE_FLTSUB, /* 67 - INV | DNML | DZ | IMP | STK */ 757 FPE_FLTOVF, /* 68 - OFL | IMP | STK */ 758 FPE_FLTSUB, /* 69 - INV | OFL | IMP | STK */ 759 FPE_FLTUND, /* 6A - DNML | OFL | IMP | STK */ 760 FPE_FLTSUB, /* 6B - INV | DNML | OFL | IMP | STK */ 761 FPE_FLTDIV, /* 6C - DZ | OFL | IMP | STK */ 762 FPE_FLTSUB, /* 6D - INV | DZ | OFL | IMP | STK */ 763 FPE_FLTDIV, /* 6E - DNML | DZ | OFL | IMP | STK */ 764 FPE_FLTSUB, /* 6F - INV | DNML | DZ | OFL | IMP | STK */ 765 FPE_FLTUND, /* 70 - UFL | IMP | STK */ 766 FPE_FLTSUB, /* 71 - INV | UFL | IMP | STK */ 767 FPE_FLTUND, /* 72 - DNML | UFL | IMP | STK */ 768 FPE_FLTSUB, /* 73 - INV | DNML | UFL | IMP | STK */ 769 FPE_FLTDIV, /* 74 - DZ | UFL | IMP | STK */ 770 FPE_FLTSUB, /* 75 - INV | DZ | UFL | IMP | STK */ 771 FPE_FLTDIV, /* 76 - DNML | DZ | UFL | IMP | STK */ 772 FPE_FLTSUB, /* 77 - INV | DNML | DZ | UFL | IMP | STK */ 773 FPE_FLTOVF, /* 78 - OFL | UFL | IMP | STK */ 774 FPE_FLTSUB, /* 79 - INV | OFL | UFL | IMP | STK */ 775 FPE_FLTUND, /* 7A - DNML | OFL | UFL | IMP | STK */ 776 FPE_FLTSUB, /* 7B - INV | DNML | OFL | UFL | IMP | STK */ 777 FPE_FLTDIV, /* 7C - DZ | OFL | UFL | IMP | STK */ 778 FPE_FLTSUB, /* 7D - INV | DZ | OFL | UFL | IMP | STK */ 779 FPE_FLTDIV, /* 7E - DNML | DZ | OFL | UFL | IMP | STK */ 780 FPE_FLTSUB, /* 7F - INV | DNML | DZ | OFL | UFL | IMP | STK */ 781}; 782 783/* 784 * Preserve the FP status word, clear FP exceptions, then generate a SIGFPE. 785 * 786 * Clearing exceptions is necessary mainly to avoid IRQ13 bugs. We now 787 * depend on longjmp() restoring a usable state. Restoring the state 788 * or examining it might fail if we didn't clear exceptions. 789 * 790 * The error code chosen will be one of the FPE_... macros. It will be 791 * sent as the second argument to old BSD-style signal handlers and as 792 * "siginfo_t->si_code" (second argument) to SA_SIGINFO signal handlers. 793 * 794 * XXX the FP state is not preserved across signal handlers. So signal 795 * handlers cannot afford to do FP unless they preserve the state or 796 * longjmp() out. Both preserving the state and longjmp()ing may be 797 * destroyed by IRQ13 bugs. Clearing FP exceptions is not an acceptable 798 * solution for signals other than SIGFPE. 799 */ 800int 801npxtrap() 802{ 803 critical_t savecrit; 804 u_short control, status; 805 u_long *exstat; 806 807 if (!npx_exists) { 808 printf("npxtrap: npxproc = %p, curproc = %p, npx_exists = %d\n", 809 PCPU_GET(npxproc), curproc, npx_exists); 810 panic("npxtrap from nowhere"); 811 } 812 savecrit = critical_enter(); 813 814 /* 815 * Interrupt handling (for another interrupt) may have pushed the 816 * state to memory. Fetch the relevant parts of the state from 817 * wherever they are. 818 */ 819 if (PCPU_GET(npxproc) != curproc) { 820 control = GET_FPU_CW(curproc); 821 status = GET_FPU_SW(curproc); 822 } else { 823 fnstcw(&control); 824 fnstsw(&status); 825 } 826 827 exstat = GET_FPU_EXSW_PTR(&curproc->p_addr->u_pcb); 828 *exstat = status; 829 if (PCPU_GET(npxproc) != curproc) 830 GET_FPU_SW(curproc) &= ~0x80bf; 831 else 832 fnclex(); 833 critical_exit(savecrit); 834 return (fpetable[status & ((~control & 0x3f) | 0x40)]); 835} 836 837/* 838 * Implement device not available (DNA) exception 839 * 840 * It would be better to switch FP context here (if curproc != npxproc) 841 * and not necessarily for every context switch, but it is too hard to 842 * access foreign pcb's. 843 */ 844int 845npxdna() 846{ 847 u_long *exstat; 848 critical_t s; 849 850 if (!npx_exists) 851 return (0); 852 if (PCPU_GET(npxproc) != NULL) { 853 printf("npxdna: npxproc = %p, curproc = %p\n", 854 PCPU_GET(npxproc), curproc); 855 panic("npxdna"); 856 } 857 s = critical_enter(); 858 stop_emulating(); 859 /* 860 * Record new context early in case frstor causes an IRQ13. 861 */ 862 PCPU_SET(npxproc, CURPROC); 863 864 exstat = GET_FPU_EXSW_PTR(PCPU_GET(curpcb)); 865 *exstat = 0; 866 /* 867 * The following frstor may cause an IRQ13 when the state being 868 * restored has a pending error. The error will appear to have been 869 * triggered by the current (npx) user instruction even when that 870 * instruction is a no-wait instruction that should not trigger an 871 * error (e.g., fnclex). On at least one 486 system all of the 872 * no-wait instructions are broken the same as frstor, so our 873 * treatment does not amplify the breakage. On at least one 874 * 386/Cyrix 387 system, fnclex works correctly while frstor and 875 * fnsave are broken, so our treatment breaks fnclex if it is the 876 * first FPU instruction after a context switch. 877 */ 878 fpurstor(&PCPU_GET(curpcb)->pcb_save); 879 critical_exit(s); 880 881 return (1); 882} 883 884/* 885 * Wrapper for fnsave instruction, partly to handle hardware bugs. When npx 886 * exceptions are reported via IRQ13, spurious IRQ13's may be triggered by 887 * no-wait npx instructions. See the Intel application note AP-578 for 888 * details. This doesn't cause any additional complications here. IRQ13's 889 * are inherently asynchronous unless the CPU is frozen to deliver them -- 890 * one that started in userland may be delivered many instructions later, 891 * after the process has entered the kernel. It may even be delivered after 892 * the fnsave here completes. A spurious IRQ13 for the fnsave is handled in 893 * the same way as a very-late-arriving non-spurious IRQ13 from user mode: 894 * it is normally ignored at first because we set npxproc to NULL; it is 895 * normally retriggered in npxdna() after return to user mode. 896 * 897 * npxsave() must be called with interrupts disabled, so that it clears 898 * npxproc atomically with saving the state. We require callers to do the 899 * disabling, since most callers need to disable interrupts anyway to call 900 * npxsave() atomically with checking npxproc. 901 * 902 * A previous version of npxsave() went to great lengths to excecute fnsave 903 * with interrupts enabled in case executing it froze the CPU. This case 904 * can't happen, at least for Intel CPU/NPX's. Spurious IRQ13's don't imply 905 * spurious freezes. 906 */ 907void 908npxsave(addr) 909 union savefpu *addr; 910{ 911 912 stop_emulating(); 913 fpusave(addr); 914 915 start_emulating(); 916 PCPU_SET(npxproc, NULL); 917} 918 919static void 920fpusave(addr) 921 union savefpu *addr; 922{ 923 924 if (!cpu_fxsr) 925 fnsave(addr); 926 else 927 fxsave(addr); 928} 929 930static void 931fpurstor(addr) 932 union savefpu *addr; 933{ 934 if (!cpu_fxsr) 935 frstor(addr); 936 else 937 fxrstor(addr); 938} 939 940#ifdef I586_CPU_XXX 941static long 942timezero(funcname, func) 943 const char *funcname; 944 void (*func) __P((void *buf, size_t len)); 945 946{ 947 void *buf; 948#define BUFSIZE 1048576 949 long usec; 950 struct timeval finish, start; 951 952 buf = malloc(BUFSIZE, M_TEMP, M_NOWAIT); 953 if (buf == NULL) 954 return (BUFSIZE); 955 microtime(&start); 956 (*func)(buf, BUFSIZE); 957 microtime(&finish); 958 usec = 1000000 * (finish.tv_sec - start.tv_sec) + 959 finish.tv_usec - start.tv_usec; 960 if (usec <= 0) 961 usec = 1; 962 if (bootverbose) 963 printf("%s bandwidth = %u kBps\n", funcname, 964 (u_int32_t)(((BUFSIZE >> 10) * 1000000) / usec)); 965 free(buf, M_TEMP); 966 return (usec); 967} 968#endif /* I586_CPU */ 969 970static device_method_t npx_methods[] = { 971 /* Device interface */ 972 DEVMETHOD(device_identify, npx_identify), 973 DEVMETHOD(device_probe, npx_probe), 974 DEVMETHOD(device_attach, npx_attach), 975 DEVMETHOD(device_detach, bus_generic_detach), 976 DEVMETHOD(device_shutdown, bus_generic_shutdown), 977 DEVMETHOD(device_suspend, bus_generic_suspend), 978 DEVMETHOD(device_resume, bus_generic_resume), 979 980 { 0, 0 } 981}; 982 983static driver_t npx_driver = { 984 "npx", 985 npx_methods, 986 1, /* no softc */ 987}; 988 989static devclass_t npx_devclass; 990 991/* 992 * We prefer to attach to the root nexus so that the usual case (exception 16) 993 * doesn't describe the processor as being `on isa'. 994 */ 995DRIVER_MODULE(npx, nexus, npx_driver, npx_devclass, 0, 0); 996 997/* 998 * This sucks up the legacy ISA support assignments from PNPBIOS. 999 */ 1000static struct isa_pnp_id npxisa_ids[] = { 1001 { 0x040cd041, "Legacy ISA coprocessor support" }, /* PNP0C04 */ 1002 { 0 } 1003}; 1004 1005static int 1006npxisa_probe(device_t dev) 1007{ 1008 int result; 1009 if ((result = ISA_PNP_PROBE(device_get_parent(dev), dev, npxisa_ids)) <= 0) { 1010 device_quiet(dev); 1011 } 1012 return(result); 1013} 1014 1015static int 1016npxisa_attach(device_t dev) 1017{ 1018 return (0); 1019} 1020 1021static device_method_t npxisa_methods[] = { 1022 /* Device interface */ 1023 DEVMETHOD(device_probe, npxisa_probe), 1024 DEVMETHOD(device_attach, npxisa_attach), 1025 DEVMETHOD(device_detach, bus_generic_detach), 1026 DEVMETHOD(device_shutdown, bus_generic_shutdown), 1027 DEVMETHOD(device_suspend, bus_generic_suspend), 1028 DEVMETHOD(device_resume, bus_generic_resume), 1029 1030 { 0, 0 } 1031}; 1032 1033static driver_t npxisa_driver = { 1034 "npxisa", 1035 npxisa_methods, 1036 1, /* no softc */ 1037}; 1038 1039static devclass_t npxisa_devclass; 1040 1041DRIVER_MODULE(npxisa, isa, npxisa_driver, npxisa_devclass, 0, 0); 1042 1043