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 * $FreeBSD$ 36 */ 37 38 #include "opt_cpu.h" 39 #include "opt_debug_npx.h" 40 #include "opt_math_emulate.h" 41 #include "opt_npx.h" 42 43 #include <sys/param.h> 44 #include <sys/systm.h> 45 #include <sys/bus.h> 46 #include <sys/kernel.h> 47 #include <sys/lock.h> 48 #include <sys/malloc.h> 49 #include <sys/module.h> 50 #include <sys/mutex.h> 51 #include <sys/mutex.h> 52 #include <sys/proc.h> 53 #include <sys/sysctl.h> 54 #include <machine/bus.h> 55 #include <sys/rman.h> 56 #ifdef NPX_DEBUG 57 #include <sys/syslog.h> 58 #endif 59 #include <sys/signalvar.h> 60 #include <sys/user.h> 61 62 #ifndef SMP 63 #include <machine/asmacros.h> 64 #endif 65 #include <machine/cputypes.h> 66 #include <machine/frame.h> 67 #include <machine/md_var.h> 68 #include <machine/pcb.h> 69 #include <machine/psl.h> 70 #ifndef SMP 71 #include <machine/clock.h> 72 #endif 73 #include <machine/resource.h> 74 #include <machine/specialreg.h> 75 #include <machine/segments.h> 76 77 #ifndef SMP 78 #include <i386/isa/icu.h> 79 #ifdef PC98 80 #include <pc98/pc98/pc98.h> 81 #else 82 #include <i386/isa/isa.h> 83 #endif 84 #endif 85 #include <isa/isavar.h> 86 87 /* 88 * 387 and 287 Numeric Coprocessor Extension (NPX) Driver. 89 */ 90 91 /* Configuration flags. */ 92 #define NPX_DISABLE_I586_OPTIMIZED_BCOPY (1 << 0) 93 #define NPX_DISABLE_I586_OPTIMIZED_BZERO (1 << 1) 94 #define NPX_DISABLE_I586_OPTIMIZED_COPYIO (1 << 2) 95 #define NPX_PREFER_EMULATOR (1 << 3) 96 97 #ifdef __GNUC__ 98 99 #define fldcw(addr) __asm("fldcw %0" : : "m" (*(addr))) 100 #define fnclex() __asm("fnclex") 101 #define fninit() __asm("fninit") 102 #define fnsave(addr) __asm __volatile("fnsave %0" : "=m" (*(addr))) 103 #define fnstcw(addr) __asm __volatile("fnstcw %0" : "=m" (*(addr))) 104 #define fnstsw(addr) __asm __volatile("fnstsw %0" : "=m" (*(addr))) 105 #define fp_divide_by_0() __asm("fldz; fld1; fdiv %st,%st(1); fnop") 106 #define frstor(addr) __asm("frstor %0" : : "m" (*(addr))) 107 #ifdef CPU_ENABLE_SSE 108 #define fxrstor(addr) __asm("fxrstor %0" : : "m" (*(addr))) 109 #define fxsave(addr) __asm __volatile("fxsave %0" : "=m" (*(addr))) 110 #endif 111 #define start_emulating() __asm("smsw %%ax; orb %0,%%al; lmsw %%ax" \ 112 : : "n" (CR0_TS) : "ax") 113 #define stop_emulating() __asm("clts") 114 115 #else /* not __GNUC__ */ 116 117 void fldcw __P((caddr_t addr)); 118 void fnclex __P((void)); 119 void fninit __P((void)); 120 void fnsave __P((caddr_t addr)); 121 void fnstcw __P((caddr_t addr)); 122 void fnstsw __P((caddr_t addr)); 123 void fp_divide_by_0 __P((void)); 124 void frstor __P((caddr_t addr)); 125 #ifdef CPU_ENABLE_SSE 126 void fxsave __P((caddr_t addr)); 127 void fxrstor __P((caddr_t addr)); 128 #endif 129 void start_emulating __P((void)); 130 void stop_emulating __P((void)); 131 132 #endif /* __GNUC__ */ 133 134 #ifdef CPU_ENABLE_SSE 135 #define GET_FPU_CW(thread) \ 136 (cpu_fxsr ? \ 137 (thread)->td_pcb->pcb_save.sv_xmm.sv_env.en_cw : \ 138 (thread)->td_pcb->pcb_save.sv_87.sv_env.en_cw) 139 #define GET_FPU_SW(thread) \ 140 (cpu_fxsr ? \ 141 (thread)->td_pcb->pcb_save.sv_xmm.sv_env.en_sw : \ 142 (thread)->td_pcb->pcb_save.sv_87.sv_env.en_sw) 143 #define GET_FPU_EXSW_PTR(pcb) \ 144 (cpu_fxsr ? \ 145 &(pcb)->pcb_save.sv_xmm.sv_ex_sw : \ 146 &(pcb)->pcb_save.sv_87.sv_ex_sw) 147 #else /* CPU_ENABLE_SSE */ 148 #define GET_FPU_CW(thread) \ 149 (thread->td_pcb->pcb_save.sv_87.sv_env.en_cw) 150 #define GET_FPU_SW(thread) \ 151 (thread->td_pcb->pcb_save.sv_87.sv_env.en_sw) 152 #define GET_FPU_EXSW_PTR(pcb) \ 153 (&(pcb)->pcb_save.sv_87.sv_ex_sw) 154 #endif /* CPU_ENABLE_SSE */ 155 156 typedef u_char bool_t; 157 158 static int npx_attach __P((device_t dev)); 159 static void npx_identify __P((driver_t *driver, device_t parent)); 160 #ifndef SMP 161 static void npx_intr __P((void *)); 162 #endif 163 static int npx_probe __P((device_t dev)); 164 static void fpusave __P((union savefpu *)); 165 static void fpurstor __P((union savefpu *)); 166 #ifdef I586_CPU_XXX 167 static long timezero __P((const char *funcname, 168 void (*func)(void *buf, size_t len))); 169 #endif /* I586_CPU */ 170 171 int hw_float; /* XXX currently just alias for npx_exists */ 172 173 SYSCTL_INT(_hw,HW_FLOATINGPT, floatingpoint, 174 CTLFLAG_RD, &hw_float, 0, 175 "Floatingpoint instructions executed in hardware"); 176 177 #ifndef SMP 178 static volatile u_int npx_intrs_while_probing; 179 static volatile u_int npx_traps_while_probing; 180 #endif 181 182 static bool_t npx_ex16; 183 static bool_t npx_exists; 184 static bool_t npx_irq13; 185 186 #ifndef SMP 187 alias_for_inthand_t probetrap; 188 __asm(" \n\ 189 .text \n\ 190 .p2align 2,0x90 \n\ 191 .type " __XSTRING(CNAME(probetrap)) ",@function \n\ 192 " __XSTRING(CNAME(probetrap)) ": \n\ 193 ss \n\ 194 incl " __XSTRING(CNAME(npx_traps_while_probing)) " \n\ 195 fnclex \n\ 196 iret \n\ 197 "); 198 #endif /* SMP */ 199 200 /* 201 * Identify routine. Create a connection point on our parent for probing. 202 */ 203 static void 204 npx_identify(driver, parent) 205 driver_t *driver; 206 device_t parent; 207 { 208 device_t child; 209 210 child = BUS_ADD_CHILD(parent, 0, "npx", 0); 211 if (child == NULL) 212 panic("npx_identify"); 213 } 214 215 #ifndef SMP 216 /* 217 * Do minimal handling of npx interrupts to convert them to traps. 218 */ 219 static void 220 npx_intr(dummy) 221 void *dummy; 222 { 223 struct thread *td; 224 225 #ifndef SMP 226 npx_intrs_while_probing++; 227 #endif 228 229 /* 230 * The BUSY# latch must be cleared in all cases so that the next 231 * unmasked npx exception causes an interrupt. 232 */ 233 #ifdef PC98 234 outb(0xf8, 0); 235 #else 236 outb(0xf0, 0); 237 #endif 238 239 /* 240 * npxthread is normally non-null here. In that case, schedule an 241 * AST to finish the exception handling in the correct context 242 * (this interrupt may occur after the thread has entered the 243 * kernel via a syscall or an interrupt). Otherwise, the npx 244 * state of the thread that caused this interrupt must have been 245 * pushed to the thread's pcb, and clearing of the busy latch 246 * above has finished the (essentially null) handling of this 247 * interrupt. Control will eventually return to the instruction 248 * that caused it and it will repeat. We will eventually (usually 249 * soon) win the race to handle the interrupt properly. 250 */ 251 td = PCPU_GET(npxthread); 252 if (td != NULL) { 253 td->td_pcb->pcb_flags |= PCB_NPXTRAP; 254 mtx_lock_spin(&sched_lock); 255 td->td_kse->ke_flags |= KEF_ASTPENDING; 256 mtx_unlock_spin(&sched_lock); 257 } 258 } 259 #endif /* !SMP */ 260 261 /* 262 * Probe routine. Initialize cr0 to give correct behaviour for [f]wait 263 * whether the device exists or not (XXX should be elsewhere). Set flags 264 * to tell npxattach() what to do. Modify device struct if npx doesn't 265 * need to use interrupts. Return 0 if device exists. 266 */ 267 static int 268 npx_probe(dev) 269 device_t dev; 270 { 271 #ifndef SMP 272 struct gate_descriptor save_idt_npxtrap; 273 struct resource *ioport_res, *irq_res; 274 void *irq_cookie; 275 int ioport_rid, irq_num, irq_rid; 276 u_short control; 277 u_short status; 278 279 save_idt_npxtrap = idt[16]; 280 setidt(16, probetrap, SDT_SYS386TGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL)); 281 ioport_rid = 0; 282 ioport_res = bus_alloc_resource(dev, SYS_RES_IOPORT, &ioport_rid, 283 IO_NPX, IO_NPX, IO_NPXSIZE, RF_ACTIVE); 284 if (ioport_res == NULL) 285 panic("npx: can't get ports"); 286 #ifdef PC98 287 if (resource_int_value("npx", 0, "irq", &irq_num) != 0) 288 irq_num = 8; 289 #else 290 if (resource_int_value("npx", 0, "irq", &irq_num) != 0) 291 irq_num = 13; 292 #endif 293 irq_rid = 0; 294 irq_res = bus_alloc_resource(dev, SYS_RES_IRQ, &ioport_rid, irq_num, 295 irq_num, 1, RF_ACTIVE); 296 if (irq_res == NULL) 297 panic("npx: can't get IRQ"); 298 if (bus_setup_intr(dev, irq_res, INTR_TYPE_MISC | INTR_FAST, npx_intr, 299 NULL, &irq_cookie) != 0) 300 panic("npx: can't create intr"); 301 #endif /* !SMP */ 302 303 /* 304 * Partially reset the coprocessor, if any. Some BIOS's don't reset 305 * it after a warm boot. 306 */ 307 #ifdef PC98 308 outb(0xf8,0); 309 #else 310 outb(0xf1, 0); /* full reset on some systems, NOP on others */ 311 outb(0xf0, 0); /* clear BUSY# latch */ 312 #endif 313 /* 314 * Prepare to trap all ESC (i.e., NPX) instructions and all WAIT 315 * instructions. We must set the CR0_MP bit and use the CR0_TS 316 * bit to control the trap, because setting the CR0_EM bit does 317 * not cause WAIT instructions to trap. It's important to trap 318 * WAIT instructions - otherwise the "wait" variants of no-wait 319 * control instructions would degenerate to the "no-wait" variants 320 * after FP context switches but work correctly otherwise. It's 321 * particularly important to trap WAITs when there is no NPX - 322 * otherwise the "wait" variants would always degenerate. 323 * 324 * Try setting CR0_NE to get correct error reporting on 486DX's. 325 * Setting it should fail or do nothing on lesser processors. 326 */ 327 load_cr0(rcr0() | CR0_MP | CR0_NE); 328 /* 329 * But don't trap while we're probing. 330 */ 331 stop_emulating(); 332 /* 333 * Finish resetting the coprocessor, if any. If there is an error 334 * pending, then we may get a bogus IRQ13, but npx_intr() will handle 335 * it OK. Bogus halts have never been observed, but we enabled 336 * IRQ13 and cleared the BUSY# latch early to handle them anyway. 337 */ 338 fninit(); 339 340 device_set_desc(dev, "math processor"); 341 342 #ifdef SMP 343 344 /* 345 * Exception 16 MUST work for SMP. 346 */ 347 npx_ex16 = hw_float = npx_exists = 1; 348 return (0); 349 350 #else /* !SMP */ 351 352 /* 353 * Don't use fwait here because it might hang. 354 * Don't use fnop here because it usually hangs if there is no FPU. 355 */ 356 DELAY(1000); /* wait for any IRQ13 */ 357 #ifdef DIAGNOSTIC 358 if (npx_intrs_while_probing != 0) 359 printf("fninit caused %u bogus npx interrupt(s)\n", 360 npx_intrs_while_probing); 361 if (npx_traps_while_probing != 0) 362 printf("fninit caused %u bogus npx trap(s)\n", 363 npx_traps_while_probing); 364 #endif 365 /* 366 * Check for a status of mostly zero. 367 */ 368 status = 0x5a5a; 369 fnstsw(&status); 370 if ((status & 0xb8ff) == 0) { 371 /* 372 * Good, now check for a proper control word. 373 */ 374 control = 0x5a5a; 375 fnstcw(&control); 376 if ((control & 0x1f3f) == 0x033f) { 377 hw_float = npx_exists = 1; 378 /* 379 * We have an npx, now divide by 0 to see if exception 380 * 16 works. 381 */ 382 control &= ~(1 << 2); /* enable divide by 0 trap */ 383 fldcw(&control); 384 #ifdef FPU_ERROR_BROKEN 385 /* 386 * FPU error signal doesn't work on some CPU 387 * accelerator board. 388 */ 389 npx_ex16 = 1; 390 return (0); 391 #endif 392 npx_traps_while_probing = npx_intrs_while_probing = 0; 393 fp_divide_by_0(); 394 if (npx_traps_while_probing != 0) { 395 /* 396 * Good, exception 16 works. 397 */ 398 npx_ex16 = 1; 399 goto no_irq13; 400 } 401 if (npx_intrs_while_probing != 0) { 402 /* 403 * Bad, we are stuck with IRQ13. 404 */ 405 npx_irq13 = 1; 406 idt[16] = save_idt_npxtrap; 407 return (0); 408 } 409 /* 410 * Worse, even IRQ13 is broken. Use emulator. 411 */ 412 } 413 } 414 /* 415 * Probe failed, but we want to get to npxattach to initialize the 416 * emulator and say that it has been installed. XXX handle devices 417 * that aren't really devices better. 418 */ 419 /* FALLTHROUGH */ 420 no_irq13: 421 idt[16] = save_idt_npxtrap; 422 bus_teardown_intr(dev, irq_res, irq_cookie); 423 424 /* 425 * XXX hack around brokenness of bus_teardown_intr(). If we left the 426 * irq active then we would get it instead of exception 16. 427 */ 428 INTRDIS(1 << irq_num); 429 430 bus_release_resource(dev, SYS_RES_IRQ, irq_rid, irq_res); 431 bus_release_resource(dev, SYS_RES_IOPORT, ioport_rid, ioport_res); 432 return (0); 433 434 #endif /* SMP */ 435 } 436 437 /* 438 * Attach routine - announce which it is, and wire into system 439 */ 440 int 441 npx_attach(dev) 442 device_t dev; 443 { 444 int flags; 445 446 if (resource_int_value("npx", 0, "flags", &flags) != 0) 447 flags = 0; 448 449 if (flags) 450 device_printf(dev, "flags 0x%x ", flags); 451 if (npx_irq13) { 452 device_printf(dev, "using IRQ 13 interface\n"); 453 } else { 454 #if defined(MATH_EMULATE) || defined(GPL_MATH_EMULATE) 455 if (npx_ex16) { 456 if (!(flags & NPX_PREFER_EMULATOR)) 457 device_printf(dev, "INT 16 interface\n"); 458 else { 459 device_printf(dev, "FPU exists, but flags request " 460 "emulator\n"); 461 hw_float = npx_exists = 0; 462 } 463 } else if (npx_exists) { 464 device_printf(dev, "error reporting broken; using 387 emulator\n"); 465 hw_float = npx_exists = 0; 466 } else 467 device_printf(dev, "387 emulator\n"); 468 #else 469 if (npx_ex16) { 470 device_printf(dev, "INT 16 interface\n"); 471 if (flags & NPX_PREFER_EMULATOR) { 472 device_printf(dev, "emulator requested, but none compiled " 473 "into kernel, using FPU\n"); 474 } 475 } else 476 device_printf(dev, "no 387 emulator in kernel and no FPU!\n"); 477 #endif 478 } 479 npxinit(__INITIAL_NPXCW__); 480 481 #ifdef I586_CPU_XXX 482 if (cpu_class == CPUCLASS_586 && npx_ex16 && npx_exists && 483 timezero("i586_bzero()", i586_bzero) < 484 timezero("bzero()", bzero) * 4 / 5) { 485 if (!(flags & NPX_DISABLE_I586_OPTIMIZED_BCOPY)) { 486 bcopy_vector = i586_bcopy; 487 ovbcopy_vector = i586_bcopy; 488 } 489 if (!(flags & NPX_DISABLE_I586_OPTIMIZED_BZERO)) 490 bzero = i586_bzero; 491 if (!(flags & NPX_DISABLE_I586_OPTIMIZED_COPYIO)) { 492 copyin_vector = i586_copyin; 493 copyout_vector = i586_copyout; 494 } 495 } 496 #endif 497 498 return (0); /* XXX unused */ 499 } 500 501 /* 502 * Initialize floating point unit. 503 */ 504 void 505 npxinit(control) 506 u_short control; 507 { 508 static union savefpu dummy; 509 critical_t savecrit; 510 511 if (!npx_exists) 512 return; 513 /* 514 * fninit has the same h/w bugs as fnsave. Use the detoxified 515 * fnsave to throw away any junk in the fpu. npxsave() initializes 516 * the fpu and sets npxthread = NULL as important side effects. 517 */ 518 savecrit = critical_enter(); 519 npxsave(&dummy); 520 stop_emulating(); 521 #ifdef CPU_ENABLE_SSE 522 /* XXX npxsave() doesn't actually initialize the fpu in the SSE case. */ 523 if (cpu_fxsr) 524 fninit(); 525 #endif 526 fldcw(&control); 527 if (PCPU_GET(curpcb) != NULL) 528 fpusave(&PCPU_GET(curpcb)->pcb_save); 529 start_emulating(); 530 critical_exit(savecrit); 531 } 532 533 /* 534 * Free coprocessor (if we have it). 535 */ 536 void 537 npxexit(td) 538 struct thread *td; 539 { 540 critical_t savecrit; 541 542 savecrit = critical_enter(); 543 if (td == PCPU_GET(npxthread)) 544 npxsave(&PCPU_GET(curpcb)->pcb_save); 545 critical_exit(savecrit); 546 #ifdef NPX_DEBUG 547 if (npx_exists) { 548 u_int masked_exceptions; 549 550 masked_exceptions = PCPU_GET(curpcb)->pcb_save.sv_87.sv_env.en_cw 551 & PCPU_GET(curpcb)->pcb_save.sv_87.sv_env.en_sw & 0x7f; 552 /* 553 * Log exceptions that would have trapped with the old 554 * control word (overflow, divide by 0, and invalid operand). 555 */ 556 if (masked_exceptions & 0x0d) 557 log(LOG_ERR, 558 "pid %d (%s) exited with masked floating point exceptions 0x%02x\n", 559 td->td_proc->p_pid, td->td_proc->p_comm, 560 masked_exceptions); 561 } 562 #endif 563 } 564 565 /* 566 * The following mechanism is used to ensure that the FPE_... value 567 * that is passed as a trapcode to the signal handler of the user 568 * process does not have more than one bit set. 569 * 570 * Multiple bits may be set if the user process modifies the control 571 * word while a status word bit is already set. While this is a sign 572 * of bad coding, we have no choise than to narrow them down to one 573 * bit, since we must not send a trapcode that is not exactly one of 574 * the FPE_ macros. 575 * 576 * The mechanism has a static table with 127 entries. Each combination 577 * of the 7 FPU status word exception bits directly translates to a 578 * position in this table, where a single FPE_... value is stored. 579 * This FPE_... value stored there is considered the "most important" 580 * of the exception bits and will be sent as the signal code. The 581 * precedence of the bits is based upon Intel Document "Numerical 582 * Applications", Chapter "Special Computational Situations". 583 * 584 * The macro to choose one of these values does these steps: 1) Throw 585 * away status word bits that cannot be masked. 2) Throw away the bits 586 * currently masked in the control word, assuming the user isn't 587 * interested in them anymore. 3) Reinsert status word bit 7 (stack 588 * fault) if it is set, which cannot be masked but must be presered. 589 * 4) Use the remaining bits to point into the trapcode table. 590 * 591 * The 6 maskable bits in order of their preference, as stated in the 592 * above referenced Intel manual: 593 * 1 Invalid operation (FP_X_INV) 594 * 1a Stack underflow 595 * 1b Stack overflow 596 * 1c Operand of unsupported format 597 * 1d SNaN operand. 598 * 2 QNaN operand (not an exception, irrelavant here) 599 * 3 Any other invalid-operation not mentioned above or zero divide 600 * (FP_X_INV, FP_X_DZ) 601 * 4 Denormal operand (FP_X_DNML) 602 * 5 Numeric over/underflow (FP_X_OFL, FP_X_UFL) 603 * 6 Inexact result (FP_X_IMP) 604 */ 605 static char fpetable[128] = { 606 0, 607 FPE_FLTINV, /* 1 - INV */ 608 FPE_FLTUND, /* 2 - DNML */ 609 FPE_FLTINV, /* 3 - INV | DNML */ 610 FPE_FLTDIV, /* 4 - DZ */ 611 FPE_FLTINV, /* 5 - INV | DZ */ 612 FPE_FLTDIV, /* 6 - DNML | DZ */ 613 FPE_FLTINV, /* 7 - INV | DNML | DZ */ 614 FPE_FLTOVF, /* 8 - OFL */ 615 FPE_FLTINV, /* 9 - INV | OFL */ 616 FPE_FLTUND, /* A - DNML | OFL */ 617 FPE_FLTINV, /* B - INV | DNML | OFL */ 618 FPE_FLTDIV, /* C - DZ | OFL */ 619 FPE_FLTINV, /* D - INV | DZ | OFL */ 620 FPE_FLTDIV, /* E - DNML | DZ | OFL */ 621 FPE_FLTINV, /* F - INV | DNML | DZ | OFL */ 622 FPE_FLTUND, /* 10 - UFL */ 623 FPE_FLTINV, /* 11 - INV | UFL */ 624 FPE_FLTUND, /* 12 - DNML | UFL */ 625 FPE_FLTINV, /* 13 - INV | DNML | UFL */ 626 FPE_FLTDIV, /* 14 - DZ | UFL */ 627 FPE_FLTINV, /* 15 - INV | DZ | UFL */ 628 FPE_FLTDIV, /* 16 - DNML | DZ | UFL */ 629 FPE_FLTINV, /* 17 - INV | DNML | DZ | UFL */ 630 FPE_FLTOVF, /* 18 - OFL | UFL */ 631 FPE_FLTINV, /* 19 - INV | OFL | UFL */ 632 FPE_FLTUND, /* 1A - DNML | OFL | UFL */ 633 FPE_FLTINV, /* 1B - INV | DNML | OFL | UFL */ 634 FPE_FLTDIV, /* 1C - DZ | OFL | UFL */ 635 FPE_FLTINV, /* 1D - INV | DZ | OFL | UFL */ 636 FPE_FLTDIV, /* 1E - DNML | DZ | OFL | UFL */ 637 FPE_FLTINV, /* 1F - INV | DNML | DZ | OFL | UFL */ 638 FPE_FLTRES, /* 20 - IMP */ 639 FPE_FLTINV, /* 21 - INV | IMP */ 640 FPE_FLTUND, /* 22 - DNML | IMP */ 641 FPE_FLTINV, /* 23 - INV | DNML | IMP */ 642 FPE_FLTDIV, /* 24 - DZ | IMP */ 643 FPE_FLTINV, /* 25 - INV | DZ | IMP */ 644 FPE_FLTDIV, /* 26 - DNML | DZ | IMP */ 645 FPE_FLTINV, /* 27 - INV | DNML | DZ | IMP */ 646 FPE_FLTOVF, /* 28 - OFL | IMP */ 647 FPE_FLTINV, /* 29 - INV | OFL | IMP */ 648 FPE_FLTUND, /* 2A - DNML | OFL | IMP */ 649 FPE_FLTINV, /* 2B - INV | DNML | OFL | IMP */ 650 FPE_FLTDIV, /* 2C - DZ | OFL | IMP */ 651 FPE_FLTINV, /* 2D - INV | DZ | OFL | IMP */ 652 FPE_FLTDIV, /* 2E - DNML | DZ | OFL | IMP */ 653 FPE_FLTINV, /* 2F - INV | DNML | DZ | OFL | IMP */ 654 FPE_FLTUND, /* 30 - UFL | IMP */ 655 FPE_FLTINV, /* 31 - INV | UFL | IMP */ 656 FPE_FLTUND, /* 32 - DNML | UFL | IMP */ 657 FPE_FLTINV, /* 33 - INV | DNML | UFL | IMP */ 658 FPE_FLTDIV, /* 34 - DZ | UFL | IMP */ 659 FPE_FLTINV, /* 35 - INV | DZ | UFL | IMP */ 660 FPE_FLTDIV, /* 36 - DNML | DZ | UFL | IMP */ 661 FPE_FLTINV, /* 37 - INV | DNML | DZ | UFL | IMP */ 662 FPE_FLTOVF, /* 38 - OFL | UFL | IMP */ 663 FPE_FLTINV, /* 39 - INV | OFL | UFL | IMP */ 664 FPE_FLTUND, /* 3A - DNML | OFL | UFL | IMP */ 665 FPE_FLTINV, /* 3B - INV | DNML | OFL | UFL | IMP */ 666 FPE_FLTDIV, /* 3C - DZ | OFL | UFL | IMP */ 667 FPE_FLTINV, /* 3D - INV | DZ | OFL | UFL | IMP */ 668 FPE_FLTDIV, /* 3E - DNML | DZ | OFL | UFL | IMP */ 669 FPE_FLTINV, /* 3F - INV | DNML | DZ | OFL | UFL | IMP */ 670 FPE_FLTSUB, /* 40 - STK */ 671 FPE_FLTSUB, /* 41 - INV | STK */ 672 FPE_FLTUND, /* 42 - DNML | STK */ 673 FPE_FLTSUB, /* 43 - INV | DNML | STK */ 674 FPE_FLTDIV, /* 44 - DZ | STK */ 675 FPE_FLTSUB, /* 45 - INV | DZ | STK */ 676 FPE_FLTDIV, /* 46 - DNML | DZ | STK */ 677 FPE_FLTSUB, /* 47 - INV | DNML | DZ | STK */ 678 FPE_FLTOVF, /* 48 - OFL | STK */ 679 FPE_FLTSUB, /* 49 - INV | OFL | STK */ 680 FPE_FLTUND, /* 4A - DNML | OFL | STK */ 681 FPE_FLTSUB, /* 4B - INV | DNML | OFL | STK */ 682 FPE_FLTDIV, /* 4C - DZ | OFL | STK */ 683 FPE_FLTSUB, /* 4D - INV | DZ | OFL | STK */ 684 FPE_FLTDIV, /* 4E - DNML | DZ | OFL | STK */ 685 FPE_FLTSUB, /* 4F - INV | DNML | DZ | OFL | STK */ 686 FPE_FLTUND, /* 50 - UFL | STK */ 687 FPE_FLTSUB, /* 51 - INV | UFL | STK */ 688 FPE_FLTUND, /* 52 - DNML | UFL | STK */ 689 FPE_FLTSUB, /* 53 - INV | DNML | UFL | STK */ 690 FPE_FLTDIV, /* 54 - DZ | UFL | STK */ 691 FPE_FLTSUB, /* 55 - INV | DZ | UFL | STK */ 692 FPE_FLTDIV, /* 56 - DNML | DZ | UFL | STK */ 693 FPE_FLTSUB, /* 57 - INV | DNML | DZ | UFL | STK */ 694 FPE_FLTOVF, /* 58 - OFL | UFL | STK */ 695 FPE_FLTSUB, /* 59 - INV | OFL | UFL | STK */ 696 FPE_FLTUND, /* 5A - DNML | OFL | UFL | STK */ 697 FPE_FLTSUB, /* 5B - INV | DNML | OFL | UFL | STK */ 698 FPE_FLTDIV, /* 5C - DZ | OFL | UFL | STK */ 699 FPE_FLTSUB, /* 5D - INV | DZ | OFL | UFL | STK */ 700 FPE_FLTDIV, /* 5E - DNML | DZ | OFL | UFL | STK */ 701 FPE_FLTSUB, /* 5F - INV | DNML | DZ | OFL | UFL | STK */ 702 FPE_FLTRES, /* 60 - IMP | STK */ 703 FPE_FLTSUB, /* 61 - INV | IMP | STK */ 704 FPE_FLTUND, /* 62 - DNML | IMP | STK */ 705 FPE_FLTSUB, /* 63 - INV | DNML | IMP | STK */ 706 FPE_FLTDIV, /* 64 - DZ | IMP | STK */ 707 FPE_FLTSUB, /* 65 - INV | DZ | IMP | STK */ 708 FPE_FLTDIV, /* 66 - DNML | DZ | IMP | STK */ 709 FPE_FLTSUB, /* 67 - INV | DNML | DZ | IMP | STK */ 710 FPE_FLTOVF, /* 68 - OFL | IMP | STK */ 711 FPE_FLTSUB, /* 69 - INV | OFL | IMP | STK */ 712 FPE_FLTUND, /* 6A - DNML | OFL | IMP | STK */ 713 FPE_FLTSUB, /* 6B - INV | DNML | OFL | IMP | STK */ 714 FPE_FLTDIV, /* 6C - DZ | OFL | IMP | STK */ 715 FPE_FLTSUB, /* 6D - INV | DZ | OFL | IMP | STK */ 716 FPE_FLTDIV, /* 6E - DNML | DZ | OFL | IMP | STK */ 717 FPE_FLTSUB, /* 6F - INV | DNML | DZ | OFL | IMP | STK */ 718 FPE_FLTUND, /* 70 - UFL | IMP | STK */ 719 FPE_FLTSUB, /* 71 - INV | UFL | IMP | STK */ 720 FPE_FLTUND, /* 72 - DNML | UFL | IMP | STK */ 721 FPE_FLTSUB, /* 73 - INV | DNML | UFL | IMP | STK */ 722 FPE_FLTDIV, /* 74 - DZ | UFL | IMP | STK */ 723 FPE_FLTSUB, /* 75 - INV | DZ | UFL | IMP | STK */ 724 FPE_FLTDIV, /* 76 - DNML | DZ | UFL | IMP | STK */ 725 FPE_FLTSUB, /* 77 - INV | DNML | DZ | UFL | IMP | STK */ 726 FPE_FLTOVF, /* 78 - OFL | UFL | IMP | STK */ 727 FPE_FLTSUB, /* 79 - INV | OFL | UFL | IMP | STK */ 728 FPE_FLTUND, /* 7A - DNML | OFL | UFL | IMP | STK */ 729 FPE_FLTSUB, /* 7B - INV | DNML | OFL | UFL | IMP | STK */ 730 FPE_FLTDIV, /* 7C - DZ | OFL | UFL | IMP | STK */ 731 FPE_FLTSUB, /* 7D - INV | DZ | OFL | UFL | IMP | STK */ 732 FPE_FLTDIV, /* 7E - DNML | DZ | OFL | UFL | IMP | STK */ 733 FPE_FLTSUB, /* 7F - INV | DNML | DZ | OFL | UFL | IMP | STK */ 734 }; 735 736 /* 737 * Preserve the FP status word, clear FP exceptions, then generate a SIGFPE. 738 * 739 * Clearing exceptions is necessary mainly to avoid IRQ13 bugs. We now 740 * depend on longjmp() restoring a usable state. Restoring the state 741 * or examining it might fail if we didn't clear exceptions. 742 * 743 * The error code chosen will be one of the FPE_... macros. It will be 744 * sent as the second argument to old BSD-style signal handlers and as 745 * "siginfo_t->si_code" (second argument) to SA_SIGINFO signal handlers. 746 * 747 * XXX the FP state is not preserved across signal handlers. So signal 748 * handlers cannot afford to do FP unless they preserve the state or 749 * longjmp() out. Both preserving the state and longjmp()ing may be 750 * destroyed by IRQ13 bugs. Clearing FP exceptions is not an acceptable 751 * solution for signals other than SIGFPE. 752 */ 753 int 754 npxtrap() 755 { 756 critical_t savecrit; 757 u_short control, status; 758 u_long *exstat; 759 760 if (!npx_exists) { 761 printf("npxtrap: npxthread = %p, curthread = %p, npx_exists = %d\n", 762 PCPU_GET(npxthread), curthread, npx_exists); 763 panic("npxtrap from nowhere"); 764 } 765 savecrit = critical_enter(); 766 767 /* 768 * Interrupt handling (for another interrupt) may have pushed the 769 * state to memory. Fetch the relevant parts of the state from 770 * wherever they are. 771 */ 772 if (PCPU_GET(npxthread) != curthread) { 773 control = GET_FPU_CW(curthread); 774 status = GET_FPU_SW(curthread); 775 } else { 776 fnstcw(&control); 777 fnstsw(&status); 778 } 779 780 exstat = GET_FPU_EXSW_PTR(curthread->td_pcb); 781 *exstat = status; 782 if (PCPU_GET(npxthread) != curthread) 783 GET_FPU_SW(curthread) &= ~0x80bf; 784 else 785 fnclex(); 786 critical_exit(savecrit); 787 return (fpetable[status & ((~control & 0x3f) | 0x40)]); 788 } 789 790 /* 791 * Implement device not available (DNA) exception 792 * 793 * It would be better to switch FP context here (if curthread != npxthread) 794 * and not necessarily for every context switch, but it is too hard to 795 * access foreign pcb's. 796 */ 797 int 798 npxdna() 799 { 800 u_long *exstat; 801 critical_t s; 802 803 if (!npx_exists) 804 return (0); 805 if (PCPU_GET(npxthread) != NULL) { 806 printf("npxdna: npxthread = %p, curthread = %p\n", 807 PCPU_GET(npxthread), curthread); 808 panic("npxdna"); 809 } 810 s = critical_enter(); 811 stop_emulating(); 812 /* 813 * Record new context early in case frstor causes an IRQ13. 814 */ 815 PCPU_SET(npxthread, curthread); 816 817 exstat = GET_FPU_EXSW_PTR(PCPU_GET(curpcb)); 818 *exstat = 0; 819 /* 820 * The following frstor may cause an IRQ13 when the state being 821 * restored has a pending error. The error will appear to have been 822 * triggered by the current (npx) user instruction even when that 823 * instruction is a no-wait instruction that should not trigger an 824 * error (e.g., fnclex). On at least one 486 system all of the 825 * no-wait instructions are broken the same as frstor, so our 826 * treatment does not amplify the breakage. On at least one 827 * 386/Cyrix 387 system, fnclex works correctly while frstor and 828 * fnsave are broken, so our treatment breaks fnclex if it is the 829 * first FPU instruction after a context switch. 830 */ 831 fpurstor(&PCPU_GET(curpcb)->pcb_save); 832 critical_exit(s); 833 834 return (1); 835 } 836 837 /* 838 * Wrapper for fnsave instruction, partly to handle hardware bugs. When npx 839 * exceptions are reported via IRQ13, spurious IRQ13's may be triggered by 840 * no-wait npx instructions. See the Intel application note AP-578 for 841 * details. This doesn't cause any additional complications here. IRQ13's 842 * are inherently asynchronous unless the CPU is frozen to deliver them -- 843 * one that started in userland may be delivered many instructions later, 844 * after the process has entered the kernel. It may even be delivered after 845 * the fnsave here completes. A spurious IRQ13 for the fnsave is handled in 846 * the same way as a very-late-arriving non-spurious IRQ13 from user mode: 847 * it is normally ignored at first because we set npxthread to NULL; it is 848 * normally retriggered in npxdna() after return to user mode. 849 * 850 * npxsave() must be called with interrupts disabled, so that it clears 851 * npxthread atomically with saving the state. We require callers to do the 852 * disabling, since most callers need to disable interrupts anyway to call 853 * npxsave() atomically with checking npxthread. 854 * 855 * A previous version of npxsave() went to great lengths to excecute fnsave 856 * with interrupts enabled in case executing it froze the CPU. This case 857 * can't happen, at least for Intel CPU/NPX's. Spurious IRQ13's don't imply 858 * spurious freezes. 859 */ 860 void 861 npxsave(addr) 862 union savefpu *addr; 863 { 864 865 stop_emulating(); 866 fpusave(addr); 867 868 start_emulating(); 869 PCPU_SET(npxthread, NULL); 870 } 871 872 static void 873 fpusave(addr) 874 union savefpu *addr; 875 { 876 877 #ifdef CPU_ENABLE_SSE 878 if (cpu_fxsr) 879 fxsave(addr); 880 else 881 #endif 882 fnsave(addr); 883 } 884 885 static void 886 fpurstor(addr) 887 union savefpu *addr; 888 { 889 890 #ifdef CPU_ENABLE_SSE 891 if (cpu_fxsr) 892 fxrstor(addr); 893 else 894 #endif 895 frstor(addr); 896 } 897 898 #ifdef I586_CPU_XXX 899 static long 900 timezero(funcname, func) 901 const char *funcname; 902 void (*func) __P((void *buf, size_t len)); 903 904 { 905 void *buf; 906 #define BUFSIZE 1048576 907 long usec; 908 struct timeval finish, start; 909 910 buf = malloc(BUFSIZE, M_TEMP, M_NOWAIT); 911 if (buf == NULL) 912 return (BUFSIZE); 913 microtime(&start); 914 (*func)(buf, BUFSIZE); 915 microtime(&finish); 916 usec = 1000000 * (finish.tv_sec - start.tv_sec) + 917 finish.tv_usec - start.tv_usec; 918 if (usec <= 0) 919 usec = 1; 920 if (bootverbose) 921 printf("%s bandwidth = %u kBps\n", funcname, 922 (u_int32_t)(((BUFSIZE >> 10) * 1000000) / usec)); 923 free(buf, M_TEMP); 924 return (usec); 925 } 926 #endif /* I586_CPU */ 927 928 static device_method_t npx_methods[] = { 929 /* Device interface */ 930 DEVMETHOD(device_identify, npx_identify), 931 DEVMETHOD(device_probe, npx_probe), 932 DEVMETHOD(device_attach, npx_attach), 933 DEVMETHOD(device_detach, bus_generic_detach), 934 DEVMETHOD(device_shutdown, bus_generic_shutdown), 935 DEVMETHOD(device_suspend, bus_generic_suspend), 936 DEVMETHOD(device_resume, bus_generic_resume), 937 938 { 0, 0 } 939 }; 940 941 static driver_t npx_driver = { 942 "npx", 943 npx_methods, 944 1, /* no softc */ 945 }; 946 947 static devclass_t npx_devclass; 948 949 /* 950 * We prefer to attach to the root nexus so that the usual case (exception 16) 951 * doesn't describe the processor as being `on isa'. 952 */ 953 DRIVER_MODULE(npx, nexus, npx_driver, npx_devclass, 0, 0); 954 955 /* 956 * This sucks up the legacy ISA support assignments from PNPBIOS/ACPI. 957 */ 958 static struct isa_pnp_id npxisa_ids[] = { 959 { 0x040cd041, "Legacy ISA coprocessor support" }, /* PNP0C04 */ 960 { 0 } 961 }; 962 963 static int 964 npxisa_probe(device_t dev) 965 { 966 int result; 967 if ((result = ISA_PNP_PROBE(device_get_parent(dev), dev, npxisa_ids)) <= 0) { 968 device_quiet(dev); 969 } 970 return(result); 971 } 972 973 static int 974 npxisa_attach(device_t dev) 975 { 976 return (0); 977 } 978 979 static device_method_t npxisa_methods[] = { 980 /* Device interface */ 981 DEVMETHOD(device_probe, npxisa_probe), 982 DEVMETHOD(device_attach, npxisa_attach), 983 DEVMETHOD(device_detach, bus_generic_detach), 984 DEVMETHOD(device_shutdown, bus_generic_shutdown), 985 DEVMETHOD(device_suspend, bus_generic_suspend), 986 DEVMETHOD(device_resume, bus_generic_resume), 987 988 { 0, 0 } 989 }; 990 991 static driver_t npxisa_driver = { 992 "npxisa", 993 npxisa_methods, 994 1, /* no softc */ 995 }; 996 997 static devclass_t npxisa_devclass; 998 999 DRIVER_MODULE(npxisa, isa, npxisa_driver, npxisa_devclass, 0, 0); 1000 #ifndef PC98 1001 DRIVER_MODULE(npxisa, acpi, npxisa_driver, npxisa_devclass, 0, 0); 1002 #endif 1003 1004