npx.c revision 313463
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: stable/11/sys/i386/isa/npx.c 313463 2017-02-09 04:42:21Z kib $"); 35 36#include "opt_cpu.h" 37#include "opt_isa.h" 38#include "opt_npx.h" 39 40#include <sys/param.h> 41#include <sys/systm.h> 42#include <sys/bus.h> 43#include <sys/kernel.h> 44#include <sys/lock.h> 45#include <sys/malloc.h> 46#include <sys/module.h> 47#include <sys/mutex.h> 48#include <sys/mutex.h> 49#include <sys/proc.h> 50#include <sys/smp.h> 51#include <sys/sysctl.h> 52#include <machine/bus.h> 53#include <sys/rman.h> 54#ifdef NPX_DEBUG 55#include <sys/syslog.h> 56#endif 57#include <sys/signalvar.h> 58#include <vm/uma.h> 59 60#include <machine/asmacros.h> 61#include <machine/cputypes.h> 62#include <machine/frame.h> 63#include <machine/md_var.h> 64#include <machine/pcb.h> 65#include <machine/psl.h> 66#include <machine/resource.h> 67#include <machine/specialreg.h> 68#include <machine/segments.h> 69#include <machine/ucontext.h> 70 71#include <machine/intr_machdep.h> 72 73#ifdef DEV_ISA 74#include <isa/isavar.h> 75#endif 76 77#if !defined(CPU_DISABLE_SSE) && defined(I686_CPU) 78#define CPU_ENABLE_SSE 79#endif 80 81/* 82 * 387 and 287 Numeric Coprocessor Extension (NPX) Driver. 83 */ 84 85#if defined(__GNUCLIKE_ASM) && !defined(lint) 86 87#define fldcw(cw) __asm __volatile("fldcw %0" : : "m" (cw)) 88#define fnclex() __asm __volatile("fnclex") 89#define fninit() __asm __volatile("fninit") 90#define fnsave(addr) __asm __volatile("fnsave %0" : "=m" (*(addr))) 91#define fnstcw(addr) __asm __volatile("fnstcw %0" : "=m" (*(addr))) 92#define fnstsw(addr) __asm __volatile("fnstsw %0" : "=am" (*(addr))) 93#define fp_divide_by_0() __asm __volatile( \ 94 "fldz; fld1; fdiv %st,%st(1); fnop") 95#define frstor(addr) __asm __volatile("frstor %0" : : "m" (*(addr))) 96#ifdef CPU_ENABLE_SSE 97#define fxrstor(addr) __asm __volatile("fxrstor %0" : : "m" (*(addr))) 98#define fxsave(addr) __asm __volatile("fxsave %0" : "=m" (*(addr))) 99#define ldmxcsr(csr) __asm __volatile("ldmxcsr %0" : : "m" (csr)) 100#define stmxcsr(addr) __asm __volatile("stmxcsr %0" : : "m" (*(addr))) 101 102static __inline void 103xrstor(char *addr, uint64_t mask) 104{ 105 uint32_t low, hi; 106 107 low = mask; 108 hi = mask >> 32; 109 __asm __volatile("xrstor %0" : : "m" (*addr), "a" (low), "d" (hi)); 110} 111 112static __inline void 113xsave(char *addr, uint64_t mask) 114{ 115 uint32_t low, hi; 116 117 low = mask; 118 hi = mask >> 32; 119 __asm __volatile("xsave %0" : "=m" (*addr) : "a" (low), "d" (hi) : 120 "memory"); 121} 122 123static __inline void 124xsaveopt(char *addr, uint64_t mask) 125{ 126 uint32_t low, hi; 127 128 low = mask; 129 hi = mask >> 32; 130 __asm __volatile("xsaveopt %0" : "=m" (*addr) : "a" (low), "d" (hi) : 131 "memory"); 132} 133#endif 134#else /* !(__GNUCLIKE_ASM && !lint) */ 135 136void fldcw(u_short cw); 137void fnclex(void); 138void fninit(void); 139void fnsave(caddr_t addr); 140void fnstcw(caddr_t addr); 141void fnstsw(caddr_t addr); 142void fp_divide_by_0(void); 143void frstor(caddr_t addr); 144#ifdef CPU_ENABLE_SSE 145void fxsave(caddr_t addr); 146void fxrstor(caddr_t addr); 147void ldmxcsr(u_int csr); 148void stmxcsr(u_int *csr); 149void xrstor(char *addr, uint64_t mask); 150void xsave(char *addr, uint64_t mask); 151void xsaveopt(char *addr, uint64_t mask); 152#endif 153 154#endif /* __GNUCLIKE_ASM && !lint */ 155 156#define start_emulating() load_cr0(rcr0() | CR0_TS) 157#define stop_emulating() clts() 158 159#ifdef CPU_ENABLE_SSE 160#define GET_FPU_CW(thread) \ 161 (cpu_fxsr ? \ 162 (thread)->td_pcb->pcb_save->sv_xmm.sv_env.en_cw : \ 163 (thread)->td_pcb->pcb_save->sv_87.sv_env.en_cw) 164#define GET_FPU_SW(thread) \ 165 (cpu_fxsr ? \ 166 (thread)->td_pcb->pcb_save->sv_xmm.sv_env.en_sw : \ 167 (thread)->td_pcb->pcb_save->sv_87.sv_env.en_sw) 168#define SET_FPU_CW(savefpu, value) do { \ 169 if (cpu_fxsr) \ 170 (savefpu)->sv_xmm.sv_env.en_cw = (value); \ 171 else \ 172 (savefpu)->sv_87.sv_env.en_cw = (value); \ 173} while (0) 174#else /* CPU_ENABLE_SSE */ 175#define GET_FPU_CW(thread) \ 176 (thread->td_pcb->pcb_save->sv_87.sv_env.en_cw) 177#define GET_FPU_SW(thread) \ 178 (thread->td_pcb->pcb_save->sv_87.sv_env.en_sw) 179#define SET_FPU_CW(savefpu, value) \ 180 (savefpu)->sv_87.sv_env.en_cw = (value) 181#endif /* CPU_ENABLE_SSE */ 182 183#ifdef CPU_ENABLE_SSE 184CTASSERT(sizeof(union savefpu) == 512); 185CTASSERT(sizeof(struct xstate_hdr) == 64); 186CTASSERT(sizeof(struct savefpu_ymm) == 832); 187 188/* 189 * This requirement is to make it easier for asm code to calculate 190 * offset of the fpu save area from the pcb address. FPU save area 191 * must be 64-byte aligned. 192 */ 193CTASSERT(sizeof(struct pcb) % XSAVE_AREA_ALIGN == 0); 194 195/* 196 * Ensure the copy of XCR0 saved in a core is contained in the padding 197 * area. 198 */ 199CTASSERT(X86_XSTATE_XCR0_OFFSET >= offsetof(struct savexmm, sv_pad) && 200 X86_XSTATE_XCR0_OFFSET + sizeof(uint64_t) <= sizeof(struct savexmm)); 201 202static void fpu_clean_state(void); 203#endif 204 205static void fpusave(union savefpu *); 206static void fpurstor(union savefpu *); 207 208int hw_float; 209 210SYSCTL_INT(_hw, HW_FLOATINGPT, floatingpoint, CTLFLAG_RD, 211 &hw_float, 0, "Floating point instructions executed in hardware"); 212 213#ifdef CPU_ENABLE_SSE 214int use_xsave; 215uint64_t xsave_mask; 216#endif 217static uma_zone_t fpu_save_area_zone; 218static union savefpu *npx_initialstate; 219 220#ifdef CPU_ENABLE_SSE 221struct xsave_area_elm_descr { 222 u_int offset; 223 u_int size; 224} *xsave_area_desc; 225 226static int use_xsaveopt; 227#endif 228 229static volatile u_int npx_traps_while_probing; 230 231alias_for_inthand_t probetrap; 232__asm(" \n\ 233 .text \n\ 234 .p2align 2,0x90 \n\ 235 .type " __XSTRING(CNAME(probetrap)) ",@function \n\ 236" __XSTRING(CNAME(probetrap)) ": \n\ 237 ss \n\ 238 incl " __XSTRING(CNAME(npx_traps_while_probing)) " \n\ 239 fnclex \n\ 240 iret \n\ 241"); 242 243/* 244 * Determine if an FPU is present and how to use it. 245 */ 246static int 247npx_probe(void) 248{ 249 struct gate_descriptor save_idt_npxtrap; 250 u_short control, status; 251 252 /* 253 * Modern CPUs all have an FPU that uses the INT16 interface 254 * and provide a simple way to verify that, so handle the 255 * common case right away. 256 */ 257 if (cpu_feature & CPUID_FPU) { 258 hw_float = 1; 259 return (1); 260 } 261 262 save_idt_npxtrap = idt[IDT_MF]; 263 setidt(IDT_MF, probetrap, SDT_SYS386TGT, SEL_KPL, 264 GSEL(GCODE_SEL, SEL_KPL)); 265 266 /* 267 * Don't trap while we're probing. 268 */ 269 stop_emulating(); 270 271 /* 272 * Finish resetting the coprocessor, if any. If there is an error 273 * pending, then we may get a bogus IRQ13, but npx_intr() will handle 274 * it OK. Bogus halts have never been observed, but we enabled 275 * IRQ13 and cleared the BUSY# latch early to handle them anyway. 276 */ 277 fninit(); 278 279 /* 280 * Don't use fwait here because it might hang. 281 * Don't use fnop here because it usually hangs if there is no FPU. 282 */ 283 DELAY(1000); /* wait for any IRQ13 */ 284#ifdef DIAGNOSTIC 285 if (npx_traps_while_probing != 0) 286 printf("fninit caused %u bogus npx trap(s)\n", 287 npx_traps_while_probing); 288#endif 289 /* 290 * Check for a status of mostly zero. 291 */ 292 status = 0x5a5a; 293 fnstsw(&status); 294 if ((status & 0xb8ff) == 0) { 295 /* 296 * Good, now check for a proper control word. 297 */ 298 control = 0x5a5a; 299 fnstcw(&control); 300 if ((control & 0x1f3f) == 0x033f) { 301 /* 302 * We have an npx, now divide by 0 to see if exception 303 * 16 works. 304 */ 305 control &= ~(1 << 2); /* enable divide by 0 trap */ 306 fldcw(control); 307#ifdef FPU_ERROR_BROKEN 308 /* 309 * FPU error signal doesn't work on some CPU 310 * accelerator board. 311 */ 312 hw_float = 1; 313 return (1); 314#endif 315 npx_traps_while_probing = 0; 316 fp_divide_by_0(); 317 if (npx_traps_while_probing != 0) { 318 /* 319 * Good, exception 16 works. 320 */ 321 hw_float = 1; 322 goto cleanup; 323 } 324 printf( 325 "FPU does not use exception 16 for error reporting\n"); 326 goto cleanup; 327 } 328 } 329 330 /* 331 * Probe failed. Floating point simply won't work. 332 * Notify user and disable FPU/MMX/SSE instruction execution. 333 */ 334 printf("WARNING: no FPU!\n"); 335 __asm __volatile("smsw %%ax; orb %0,%%al; lmsw %%ax" : : 336 "n" (CR0_EM | CR0_MP) : "ax"); 337 338cleanup: 339 idt[IDT_MF] = save_idt_npxtrap; 340 return (hw_float); 341} 342 343#ifdef CPU_ENABLE_SSE 344/* 345 * Enable XSAVE if supported and allowed by user. 346 * Calculate the xsave_mask. 347 */ 348static void 349npxinit_bsp1(void) 350{ 351 u_int cp[4]; 352 uint64_t xsave_mask_user; 353 354 if (cpu_fxsr && (cpu_feature2 & CPUID2_XSAVE) != 0) { 355 use_xsave = 1; 356 TUNABLE_INT_FETCH("hw.use_xsave", &use_xsave); 357 } 358 if (!use_xsave) 359 return; 360 361 cpuid_count(0xd, 0x0, cp); 362 xsave_mask = XFEATURE_ENABLED_X87 | XFEATURE_ENABLED_SSE; 363 if ((cp[0] & xsave_mask) != xsave_mask) 364 panic("CPU0 does not support X87 or SSE: %x", cp[0]); 365 xsave_mask = ((uint64_t)cp[3] << 32) | cp[0]; 366 xsave_mask_user = xsave_mask; 367 TUNABLE_QUAD_FETCH("hw.xsave_mask", &xsave_mask_user); 368 xsave_mask_user |= XFEATURE_ENABLED_X87 | XFEATURE_ENABLED_SSE; 369 xsave_mask &= xsave_mask_user; 370 if ((xsave_mask & XFEATURE_AVX512) != XFEATURE_AVX512) 371 xsave_mask &= ~XFEATURE_AVX512; 372 if ((xsave_mask & XFEATURE_MPX) != XFEATURE_MPX) 373 xsave_mask &= ~XFEATURE_MPX; 374 375 cpuid_count(0xd, 0x1, cp); 376 if ((cp[0] & CPUID_EXTSTATE_XSAVEOPT) != 0) 377 use_xsaveopt = 1; 378} 379#endif 380/* 381 382 * Calculate the fpu save area size. 383 */ 384static void 385npxinit_bsp2(void) 386{ 387#ifdef CPU_ENABLE_SSE 388 u_int cp[4]; 389 390 if (use_xsave) { 391 cpuid_count(0xd, 0x0, cp); 392 cpu_max_ext_state_size = cp[1]; 393 394 /* 395 * Reload the cpu_feature2, since we enabled OSXSAVE. 396 */ 397 do_cpuid(1, cp); 398 cpu_feature2 = cp[2]; 399 } else 400#endif 401 cpu_max_ext_state_size = sizeof(union savefpu); 402} 403 404/* 405 * Initialize floating point unit. 406 */ 407void 408npxinit(bool bsp) 409{ 410 static union savefpu dummy; 411 register_t saveintr; 412#ifdef CPU_ENABLE_SSE 413 u_int mxcsr; 414#endif 415 u_short control; 416 417 if (bsp) { 418 if (!npx_probe()) 419 return; 420#ifdef CPU_ENABLE_SSE 421 npxinit_bsp1(); 422#endif 423 } 424 425#ifdef CPU_ENABLE_SSE 426 if (use_xsave) { 427 load_cr4(rcr4() | CR4_XSAVE); 428 load_xcr(XCR0, xsave_mask); 429 } 430#endif 431 432 /* 433 * XCR0 shall be set up before CPU can report the save area size. 434 */ 435 if (bsp) 436 npxinit_bsp2(); 437 438 /* 439 * fninit has the same h/w bugs as fnsave. Use the detoxified 440 * fnsave to throw away any junk in the fpu. fpusave() initializes 441 * the fpu. 442 * 443 * It is too early for critical_enter() to work on AP. 444 */ 445 saveintr = intr_disable(); 446 stop_emulating(); 447#ifdef CPU_ENABLE_SSE 448 if (cpu_fxsr) 449 fninit(); 450 else 451#endif 452 fnsave(&dummy); 453 control = __INITIAL_NPXCW__; 454 fldcw(control); 455#ifdef CPU_ENABLE_SSE 456 if (cpu_fxsr) { 457 mxcsr = __INITIAL_MXCSR__; 458 ldmxcsr(mxcsr); 459 } 460#endif 461 start_emulating(); 462 intr_restore(saveintr); 463} 464 465/* 466 * On the boot CPU we generate a clean state that is used to 467 * initialize the floating point unit when it is first used by a 468 * process. 469 */ 470static void 471npxinitstate(void *arg __unused) 472{ 473 register_t saveintr; 474#ifdef CPU_ENABLE_SSE 475 int cp[4], i, max_ext_n; 476#endif 477 478 if (!hw_float) 479 return; 480 481 npx_initialstate = malloc(cpu_max_ext_state_size, M_DEVBUF, 482 M_WAITOK | M_ZERO); 483 saveintr = intr_disable(); 484 stop_emulating(); 485 486 fpusave(npx_initialstate); 487#ifdef CPU_ENABLE_SSE 488 if (cpu_fxsr) { 489 if (npx_initialstate->sv_xmm.sv_env.en_mxcsr_mask) 490 cpu_mxcsr_mask = 491 npx_initialstate->sv_xmm.sv_env.en_mxcsr_mask; 492 else 493 cpu_mxcsr_mask = 0xFFBF; 494 495 /* 496 * The fninit instruction does not modify XMM 497 * registers or x87 registers (MM/ST). The fpusave 498 * call dumped the garbage contained in the registers 499 * after reset to the initial state saved. Clear XMM 500 * and x87 registers file image to make the startup 501 * program state and signal handler XMM/x87 register 502 * content predictable. 503 */ 504 bzero(npx_initialstate->sv_xmm.sv_fp, 505 sizeof(npx_initialstate->sv_xmm.sv_fp)); 506 bzero(npx_initialstate->sv_xmm.sv_xmm, 507 sizeof(npx_initialstate->sv_xmm.sv_xmm)); 508 } else 509#endif 510 bzero(npx_initialstate->sv_87.sv_ac, 511 sizeof(npx_initialstate->sv_87.sv_ac)); 512 513#ifdef CPU_ENABLE_SSE 514 /* 515 * Create a table describing the layout of the CPU Extended 516 * Save Area. 517 */ 518 if (use_xsave) { 519 if (xsave_mask >> 32 != 0) 520 max_ext_n = fls(xsave_mask >> 32) + 32; 521 else 522 max_ext_n = fls(xsave_mask); 523 xsave_area_desc = malloc(max_ext_n * sizeof(struct 524 xsave_area_elm_descr), M_DEVBUF, M_WAITOK | M_ZERO); 525 /* x87 state */ 526 xsave_area_desc[0].offset = 0; 527 xsave_area_desc[0].size = 160; 528 /* XMM */ 529 xsave_area_desc[1].offset = 160; 530 xsave_area_desc[1].size = 288 - 160; 531 532 for (i = 2; i < max_ext_n; i++) { 533 cpuid_count(0xd, i, cp); 534 xsave_area_desc[i].offset = cp[1]; 535 xsave_area_desc[i].size = cp[0]; 536 } 537 } 538#endif 539 540 fpu_save_area_zone = uma_zcreate("FPU_save_area", 541 cpu_max_ext_state_size, NULL, NULL, NULL, NULL, 542 XSAVE_AREA_ALIGN - 1, 0); 543 544 start_emulating(); 545 intr_restore(saveintr); 546} 547SYSINIT(npxinitstate, SI_SUB_DRIVERS, SI_ORDER_ANY, npxinitstate, NULL); 548 549/* 550 * Free coprocessor (if we have it). 551 */ 552void 553npxexit(struct thread *td) 554{ 555 556 critical_enter(); 557 if (curthread == PCPU_GET(fpcurthread)) { 558 stop_emulating(); 559 fpusave(curpcb->pcb_save); 560 start_emulating(); 561 PCPU_SET(fpcurthread, NULL); 562 } 563 critical_exit(); 564#ifdef NPX_DEBUG 565 if (hw_float) { 566 u_int masked_exceptions; 567 568 masked_exceptions = GET_FPU_CW(td) & GET_FPU_SW(td) & 0x7f; 569 /* 570 * Log exceptions that would have trapped with the old 571 * control word (overflow, divide by 0, and invalid operand). 572 */ 573 if (masked_exceptions & 0x0d) 574 log(LOG_ERR, 575 "pid %d (%s) exited with masked floating point exceptions 0x%02x\n", 576 td->td_proc->p_pid, td->td_proc->p_comm, 577 masked_exceptions); 578 } 579#endif 580} 581 582int 583npxformat(void) 584{ 585 586 if (!hw_float) 587 return (_MC_FPFMT_NODEV); 588#ifdef CPU_ENABLE_SSE 589 if (cpu_fxsr) 590 return (_MC_FPFMT_XMM); 591#endif 592 return (_MC_FPFMT_387); 593} 594 595/* 596 * The following mechanism is used to ensure that the FPE_... value 597 * that is passed as a trapcode to the signal handler of the user 598 * process does not have more than one bit set. 599 * 600 * Multiple bits may be set if the user process modifies the control 601 * word while a status word bit is already set. While this is a sign 602 * of bad coding, we have no choise than to narrow them down to one 603 * bit, since we must not send a trapcode that is not exactly one of 604 * the FPE_ macros. 605 * 606 * The mechanism has a static table with 127 entries. Each combination 607 * of the 7 FPU status word exception bits directly translates to a 608 * position in this table, where a single FPE_... value is stored. 609 * This FPE_... value stored there is considered the "most important" 610 * of the exception bits and will be sent as the signal code. The 611 * precedence of the bits is based upon Intel Document "Numerical 612 * Applications", Chapter "Special Computational Situations". 613 * 614 * The macro to choose one of these values does these steps: 1) Throw 615 * away status word bits that cannot be masked. 2) Throw away the bits 616 * currently masked in the control word, assuming the user isn't 617 * interested in them anymore. 3) Reinsert status word bit 7 (stack 618 * fault) if it is set, which cannot be masked but must be presered. 619 * 4) Use the remaining bits to point into the trapcode table. 620 * 621 * The 6 maskable bits in order of their preference, as stated in the 622 * above referenced Intel manual: 623 * 1 Invalid operation (FP_X_INV) 624 * 1a Stack underflow 625 * 1b Stack overflow 626 * 1c Operand of unsupported format 627 * 1d SNaN operand. 628 * 2 QNaN operand (not an exception, irrelavant here) 629 * 3 Any other invalid-operation not mentioned above or zero divide 630 * (FP_X_INV, FP_X_DZ) 631 * 4 Denormal operand (FP_X_DNML) 632 * 5 Numeric over/underflow (FP_X_OFL, FP_X_UFL) 633 * 6 Inexact result (FP_X_IMP) 634 */ 635static char fpetable[128] = { 636 0, 637 FPE_FLTINV, /* 1 - INV */ 638 FPE_FLTUND, /* 2 - DNML */ 639 FPE_FLTINV, /* 3 - INV | DNML */ 640 FPE_FLTDIV, /* 4 - DZ */ 641 FPE_FLTINV, /* 5 - INV | DZ */ 642 FPE_FLTDIV, /* 6 - DNML | DZ */ 643 FPE_FLTINV, /* 7 - INV | DNML | DZ */ 644 FPE_FLTOVF, /* 8 - OFL */ 645 FPE_FLTINV, /* 9 - INV | OFL */ 646 FPE_FLTUND, /* A - DNML | OFL */ 647 FPE_FLTINV, /* B - INV | DNML | OFL */ 648 FPE_FLTDIV, /* C - DZ | OFL */ 649 FPE_FLTINV, /* D - INV | DZ | OFL */ 650 FPE_FLTDIV, /* E - DNML | DZ | OFL */ 651 FPE_FLTINV, /* F - INV | DNML | DZ | OFL */ 652 FPE_FLTUND, /* 10 - UFL */ 653 FPE_FLTINV, /* 11 - INV | UFL */ 654 FPE_FLTUND, /* 12 - DNML | UFL */ 655 FPE_FLTINV, /* 13 - INV | DNML | UFL */ 656 FPE_FLTDIV, /* 14 - DZ | UFL */ 657 FPE_FLTINV, /* 15 - INV | DZ | UFL */ 658 FPE_FLTDIV, /* 16 - DNML | DZ | UFL */ 659 FPE_FLTINV, /* 17 - INV | DNML | DZ | UFL */ 660 FPE_FLTOVF, /* 18 - OFL | UFL */ 661 FPE_FLTINV, /* 19 - INV | OFL | UFL */ 662 FPE_FLTUND, /* 1A - DNML | OFL | UFL */ 663 FPE_FLTINV, /* 1B - INV | DNML | OFL | UFL */ 664 FPE_FLTDIV, /* 1C - DZ | OFL | UFL */ 665 FPE_FLTINV, /* 1D - INV | DZ | OFL | UFL */ 666 FPE_FLTDIV, /* 1E - DNML | DZ | OFL | UFL */ 667 FPE_FLTINV, /* 1F - INV | DNML | DZ | OFL | UFL */ 668 FPE_FLTRES, /* 20 - IMP */ 669 FPE_FLTINV, /* 21 - INV | IMP */ 670 FPE_FLTUND, /* 22 - DNML | IMP */ 671 FPE_FLTINV, /* 23 - INV | DNML | IMP */ 672 FPE_FLTDIV, /* 24 - DZ | IMP */ 673 FPE_FLTINV, /* 25 - INV | DZ | IMP */ 674 FPE_FLTDIV, /* 26 - DNML | DZ | IMP */ 675 FPE_FLTINV, /* 27 - INV | DNML | DZ | IMP */ 676 FPE_FLTOVF, /* 28 - OFL | IMP */ 677 FPE_FLTINV, /* 29 - INV | OFL | IMP */ 678 FPE_FLTUND, /* 2A - DNML | OFL | IMP */ 679 FPE_FLTINV, /* 2B - INV | DNML | OFL | IMP */ 680 FPE_FLTDIV, /* 2C - DZ | OFL | IMP */ 681 FPE_FLTINV, /* 2D - INV | DZ | OFL | IMP */ 682 FPE_FLTDIV, /* 2E - DNML | DZ | OFL | IMP */ 683 FPE_FLTINV, /* 2F - INV | DNML | DZ | OFL | IMP */ 684 FPE_FLTUND, /* 30 - UFL | IMP */ 685 FPE_FLTINV, /* 31 - INV | UFL | IMP */ 686 FPE_FLTUND, /* 32 - DNML | UFL | IMP */ 687 FPE_FLTINV, /* 33 - INV | DNML | UFL | IMP */ 688 FPE_FLTDIV, /* 34 - DZ | UFL | IMP */ 689 FPE_FLTINV, /* 35 - INV | DZ | UFL | IMP */ 690 FPE_FLTDIV, /* 36 - DNML | DZ | UFL | IMP */ 691 FPE_FLTINV, /* 37 - INV | DNML | DZ | UFL | IMP */ 692 FPE_FLTOVF, /* 38 - OFL | UFL | IMP */ 693 FPE_FLTINV, /* 39 - INV | OFL | UFL | IMP */ 694 FPE_FLTUND, /* 3A - DNML | OFL | UFL | IMP */ 695 FPE_FLTINV, /* 3B - INV | DNML | OFL | UFL | IMP */ 696 FPE_FLTDIV, /* 3C - DZ | OFL | UFL | IMP */ 697 FPE_FLTINV, /* 3D - INV | DZ | OFL | UFL | IMP */ 698 FPE_FLTDIV, /* 3E - DNML | DZ | OFL | UFL | IMP */ 699 FPE_FLTINV, /* 3F - INV | DNML | DZ | OFL | UFL | IMP */ 700 FPE_FLTSUB, /* 40 - STK */ 701 FPE_FLTSUB, /* 41 - INV | STK */ 702 FPE_FLTUND, /* 42 - DNML | STK */ 703 FPE_FLTSUB, /* 43 - INV | DNML | STK */ 704 FPE_FLTDIV, /* 44 - DZ | STK */ 705 FPE_FLTSUB, /* 45 - INV | DZ | STK */ 706 FPE_FLTDIV, /* 46 - DNML | DZ | STK */ 707 FPE_FLTSUB, /* 47 - INV | DNML | DZ | STK */ 708 FPE_FLTOVF, /* 48 - OFL | STK */ 709 FPE_FLTSUB, /* 49 - INV | OFL | STK */ 710 FPE_FLTUND, /* 4A - DNML | OFL | STK */ 711 FPE_FLTSUB, /* 4B - INV | DNML | OFL | STK */ 712 FPE_FLTDIV, /* 4C - DZ | OFL | STK */ 713 FPE_FLTSUB, /* 4D - INV | DZ | OFL | STK */ 714 FPE_FLTDIV, /* 4E - DNML | DZ | OFL | STK */ 715 FPE_FLTSUB, /* 4F - INV | DNML | DZ | OFL | STK */ 716 FPE_FLTUND, /* 50 - UFL | STK */ 717 FPE_FLTSUB, /* 51 - INV | UFL | STK */ 718 FPE_FLTUND, /* 52 - DNML | UFL | STK */ 719 FPE_FLTSUB, /* 53 - INV | DNML | UFL | STK */ 720 FPE_FLTDIV, /* 54 - DZ | UFL | STK */ 721 FPE_FLTSUB, /* 55 - INV | DZ | UFL | STK */ 722 FPE_FLTDIV, /* 56 - DNML | DZ | UFL | STK */ 723 FPE_FLTSUB, /* 57 - INV | DNML | DZ | UFL | STK */ 724 FPE_FLTOVF, /* 58 - OFL | UFL | STK */ 725 FPE_FLTSUB, /* 59 - INV | OFL | UFL | STK */ 726 FPE_FLTUND, /* 5A - DNML | OFL | UFL | STK */ 727 FPE_FLTSUB, /* 5B - INV | DNML | OFL | UFL | STK */ 728 FPE_FLTDIV, /* 5C - DZ | OFL | UFL | STK */ 729 FPE_FLTSUB, /* 5D - INV | DZ | OFL | UFL | STK */ 730 FPE_FLTDIV, /* 5E - DNML | DZ | OFL | UFL | STK */ 731 FPE_FLTSUB, /* 5F - INV | DNML | DZ | OFL | UFL | STK */ 732 FPE_FLTRES, /* 60 - IMP | STK */ 733 FPE_FLTSUB, /* 61 - INV | IMP | STK */ 734 FPE_FLTUND, /* 62 - DNML | IMP | STK */ 735 FPE_FLTSUB, /* 63 - INV | DNML | IMP | STK */ 736 FPE_FLTDIV, /* 64 - DZ | IMP | STK */ 737 FPE_FLTSUB, /* 65 - INV | DZ | IMP | STK */ 738 FPE_FLTDIV, /* 66 - DNML | DZ | IMP | STK */ 739 FPE_FLTSUB, /* 67 - INV | DNML | DZ | IMP | STK */ 740 FPE_FLTOVF, /* 68 - OFL | IMP | STK */ 741 FPE_FLTSUB, /* 69 - INV | OFL | IMP | STK */ 742 FPE_FLTUND, /* 6A - DNML | OFL | IMP | STK */ 743 FPE_FLTSUB, /* 6B - INV | DNML | OFL | IMP | STK */ 744 FPE_FLTDIV, /* 6C - DZ | OFL | IMP | STK */ 745 FPE_FLTSUB, /* 6D - INV | DZ | OFL | IMP | STK */ 746 FPE_FLTDIV, /* 6E - DNML | DZ | OFL | IMP | STK */ 747 FPE_FLTSUB, /* 6F - INV | DNML | DZ | OFL | IMP | STK */ 748 FPE_FLTUND, /* 70 - UFL | IMP | STK */ 749 FPE_FLTSUB, /* 71 - INV | UFL | IMP | STK */ 750 FPE_FLTUND, /* 72 - DNML | UFL | IMP | STK */ 751 FPE_FLTSUB, /* 73 - INV | DNML | UFL | IMP | STK */ 752 FPE_FLTDIV, /* 74 - DZ | UFL | IMP | STK */ 753 FPE_FLTSUB, /* 75 - INV | DZ | UFL | IMP | STK */ 754 FPE_FLTDIV, /* 76 - DNML | DZ | UFL | IMP | STK */ 755 FPE_FLTSUB, /* 77 - INV | DNML | DZ | UFL | IMP | STK */ 756 FPE_FLTOVF, /* 78 - OFL | UFL | IMP | STK */ 757 FPE_FLTSUB, /* 79 - INV | OFL | UFL | IMP | STK */ 758 FPE_FLTUND, /* 7A - DNML | OFL | UFL | IMP | STK */ 759 FPE_FLTSUB, /* 7B - INV | DNML | OFL | UFL | IMP | STK */ 760 FPE_FLTDIV, /* 7C - DZ | OFL | UFL | IMP | STK */ 761 FPE_FLTSUB, /* 7D - INV | DZ | OFL | UFL | IMP | STK */ 762 FPE_FLTDIV, /* 7E - DNML | DZ | OFL | UFL | IMP | STK */ 763 FPE_FLTSUB, /* 7F - INV | DNML | DZ | OFL | UFL | IMP | STK */ 764}; 765 766/* 767 * Read the FP status and control words, then generate si_code value 768 * for SIGFPE. The error code chosen will be one of the 769 * FPE_... macros. It will be sent as the second argument to old 770 * BSD-style signal handlers and as "siginfo_t->si_code" (second 771 * argument) to SA_SIGINFO signal handlers. 772 * 773 * Some time ago, we cleared the x87 exceptions with FNCLEX there. 774 * Clearing exceptions was necessary mainly to avoid IRQ13 bugs. The 775 * usermode code which understands the FPU hardware enough to enable 776 * the exceptions, can also handle clearing the exception state in the 777 * handler. The only consequence of not clearing the exception is the 778 * rethrow of the SIGFPE on return from the signal handler and 779 * reexecution of the corresponding instruction. 780 * 781 * For XMM traps, the exceptions were never cleared. 782 */ 783int 784npxtrap_x87(void) 785{ 786 u_short control, status; 787 788 if (!hw_float) { 789 printf( 790 "npxtrap_x87: fpcurthread = %p, curthread = %p, hw_float = %d\n", 791 PCPU_GET(fpcurthread), curthread, hw_float); 792 panic("npxtrap from nowhere"); 793 } 794 critical_enter(); 795 796 /* 797 * Interrupt handling (for another interrupt) may have pushed the 798 * state to memory. Fetch the relevant parts of the state from 799 * wherever they are. 800 */ 801 if (PCPU_GET(fpcurthread) != curthread) { 802 control = GET_FPU_CW(curthread); 803 status = GET_FPU_SW(curthread); 804 } else { 805 fnstcw(&control); 806 fnstsw(&status); 807 } 808 critical_exit(); 809 return (fpetable[status & ((~control & 0x3f) | 0x40)]); 810} 811 812#ifdef CPU_ENABLE_SSE 813int 814npxtrap_sse(void) 815{ 816 u_int mxcsr; 817 818 if (!hw_float) { 819 printf( 820 "npxtrap_sse: fpcurthread = %p, curthread = %p, hw_float = %d\n", 821 PCPU_GET(fpcurthread), curthread, hw_float); 822 panic("npxtrap from nowhere"); 823 } 824 critical_enter(); 825 if (PCPU_GET(fpcurthread) != curthread) 826 mxcsr = curthread->td_pcb->pcb_save->sv_xmm.sv_env.en_mxcsr; 827 else 828 stmxcsr(&mxcsr); 829 critical_exit(); 830 return (fpetable[(mxcsr & (~mxcsr >> 7)) & 0x3f]); 831} 832#endif 833 834/* 835 * Implement device not available (DNA) exception 836 * 837 * It would be better to switch FP context here (if curthread != fpcurthread) 838 * and not necessarily for every context switch, but it is too hard to 839 * access foreign pcb's. 840 */ 841 842static int err_count = 0; 843 844int 845npxdna(void) 846{ 847 848 if (!hw_float) 849 return (0); 850 critical_enter(); 851 if (PCPU_GET(fpcurthread) == curthread) { 852 printf("npxdna: fpcurthread == curthread %d times\n", 853 ++err_count); 854 stop_emulating(); 855 critical_exit(); 856 return (1); 857 } 858 if (PCPU_GET(fpcurthread) != NULL) { 859 printf("npxdna: fpcurthread = %p (%d), curthread = %p (%d)\n", 860 PCPU_GET(fpcurthread), 861 PCPU_GET(fpcurthread)->td_proc->p_pid, 862 curthread, curthread->td_proc->p_pid); 863 panic("npxdna"); 864 } 865 stop_emulating(); 866 /* 867 * Record new context early in case frstor causes a trap. 868 */ 869 PCPU_SET(fpcurthread, curthread); 870 871#ifdef CPU_ENABLE_SSE 872 if (cpu_fxsr) 873 fpu_clean_state(); 874#endif 875 876 if ((curpcb->pcb_flags & PCB_NPXINITDONE) == 0) { 877 /* 878 * This is the first time this thread has used the FPU or 879 * the PCB doesn't contain a clean FPU state. Explicitly 880 * load an initial state. 881 * 882 * We prefer to restore the state from the actual save 883 * area in PCB instead of directly loading from 884 * npx_initialstate, to ignite the XSAVEOPT 885 * tracking engine. 886 */ 887 bcopy(npx_initialstate, curpcb->pcb_save, cpu_max_ext_state_size); 888 fpurstor(curpcb->pcb_save); 889 if (curpcb->pcb_initial_npxcw != __INITIAL_NPXCW__) 890 fldcw(curpcb->pcb_initial_npxcw); 891 curpcb->pcb_flags |= PCB_NPXINITDONE; 892 if (PCB_USER_FPU(curpcb)) 893 curpcb->pcb_flags |= PCB_NPXUSERINITDONE; 894 } else { 895 fpurstor(curpcb->pcb_save); 896 } 897 critical_exit(); 898 899 return (1); 900} 901 902/* 903 * Wrapper for fpusave() called from context switch routines. 904 * 905 * npxsave() must be called with interrupts disabled, so that it clears 906 * fpcurthread atomically with saving the state. We require callers to do the 907 * disabling, since most callers need to disable interrupts anyway to call 908 * npxsave() atomically with checking fpcurthread. 909 */ 910void 911npxsave(addr) 912 union savefpu *addr; 913{ 914 915 stop_emulating(); 916#ifdef CPU_ENABLE_SSE 917 if (use_xsaveopt) 918 xsaveopt((char *)addr, xsave_mask); 919 else 920#endif 921 fpusave(addr); 922 start_emulating(); 923 PCPU_SET(fpcurthread, NULL); 924} 925 926/* 927 * Unconditionally save the current co-processor state across suspend and 928 * resume. 929 */ 930void 931npxsuspend(union savefpu *addr) 932{ 933 register_t cr0; 934 935 if (!hw_float) 936 return; 937 if (PCPU_GET(fpcurthread) == NULL) { 938 bcopy(npx_initialstate, addr, cpu_max_ext_state_size); 939 return; 940 } 941 cr0 = rcr0(); 942 stop_emulating(); 943 fpusave(addr); 944 load_cr0(cr0); 945} 946 947void 948npxresume(union savefpu *addr) 949{ 950 register_t cr0; 951 952 if (!hw_float) 953 return; 954 955 cr0 = rcr0(); 956 npxinit(false); 957 stop_emulating(); 958 fpurstor(addr); 959 load_cr0(cr0); 960} 961 962void 963npxdrop(void) 964{ 965 struct thread *td; 966 967 /* 968 * Discard pending exceptions in the !cpu_fxsr case so that unmasked 969 * ones don't cause a panic on the next frstor. 970 */ 971#ifdef CPU_ENABLE_SSE 972 if (!cpu_fxsr) 973#endif 974 fnclex(); 975 976 td = PCPU_GET(fpcurthread); 977 KASSERT(td == curthread, ("fpudrop: fpcurthread != curthread")); 978 CRITICAL_ASSERT(td); 979 PCPU_SET(fpcurthread, NULL); 980 td->td_pcb->pcb_flags &= ~PCB_NPXINITDONE; 981 start_emulating(); 982} 983 984/* 985 * Get the user state of the FPU into pcb->pcb_user_save without 986 * dropping ownership (if possible). It returns the FPU ownership 987 * status. 988 */ 989int 990npxgetregs(struct thread *td) 991{ 992 struct pcb *pcb; 993#ifdef CPU_ENABLE_SSE 994 uint64_t *xstate_bv, bit; 995 char *sa; 996 int max_ext_n, i; 997#endif 998 int owned; 999 1000 if (!hw_float) 1001 return (_MC_FPOWNED_NONE); 1002 1003 pcb = td->td_pcb; 1004 if ((pcb->pcb_flags & PCB_NPXINITDONE) == 0) { 1005 bcopy(npx_initialstate, get_pcb_user_save_pcb(pcb), 1006 cpu_max_ext_state_size); 1007 SET_FPU_CW(get_pcb_user_save_pcb(pcb), pcb->pcb_initial_npxcw); 1008 npxuserinited(td); 1009 return (_MC_FPOWNED_PCB); 1010 } 1011 critical_enter(); 1012 if (td == PCPU_GET(fpcurthread)) { 1013 fpusave(get_pcb_user_save_pcb(pcb)); 1014#ifdef CPU_ENABLE_SSE 1015 if (!cpu_fxsr) 1016#endif 1017 /* 1018 * fnsave initializes the FPU and destroys whatever 1019 * context it contains. Make sure the FPU owner 1020 * starts with a clean state next time. 1021 */ 1022 npxdrop(); 1023 owned = _MC_FPOWNED_FPU; 1024 } else { 1025 owned = _MC_FPOWNED_PCB; 1026 } 1027 critical_exit(); 1028#ifdef CPU_ENABLE_SSE 1029 if (use_xsave) { 1030 /* 1031 * Handle partially saved state. 1032 */ 1033 sa = (char *)get_pcb_user_save_pcb(pcb); 1034 xstate_bv = (uint64_t *)(sa + sizeof(union savefpu) + 1035 offsetof(struct xstate_hdr, xstate_bv)); 1036 if (xsave_mask >> 32 != 0) 1037 max_ext_n = fls(xsave_mask >> 32) + 32; 1038 else 1039 max_ext_n = fls(xsave_mask); 1040 for (i = 0; i < max_ext_n; i++) { 1041 bit = 1ULL << i; 1042 if ((xsave_mask & bit) == 0 || (*xstate_bv & bit) != 0) 1043 continue; 1044 bcopy((char *)npx_initialstate + 1045 xsave_area_desc[i].offset, 1046 sa + xsave_area_desc[i].offset, 1047 xsave_area_desc[i].size); 1048 *xstate_bv |= bit; 1049 } 1050 } 1051#endif 1052 return (owned); 1053} 1054 1055void 1056npxuserinited(struct thread *td) 1057{ 1058 struct pcb *pcb; 1059 1060 pcb = td->td_pcb; 1061 if (PCB_USER_FPU(pcb)) 1062 pcb->pcb_flags |= PCB_NPXINITDONE; 1063 pcb->pcb_flags |= PCB_NPXUSERINITDONE; 1064} 1065 1066#ifdef CPU_ENABLE_SSE 1067int 1068npxsetxstate(struct thread *td, char *xfpustate, size_t xfpustate_size) 1069{ 1070 struct xstate_hdr *hdr, *ehdr; 1071 size_t len, max_len; 1072 uint64_t bv; 1073 1074 /* XXXKIB should we clear all extended state in xstate_bv instead ? */ 1075 if (xfpustate == NULL) 1076 return (0); 1077 if (!use_xsave) 1078 return (EOPNOTSUPP); 1079 1080 len = xfpustate_size; 1081 if (len < sizeof(struct xstate_hdr)) 1082 return (EINVAL); 1083 max_len = cpu_max_ext_state_size - sizeof(union savefpu); 1084 if (len > max_len) 1085 return (EINVAL); 1086 1087 ehdr = (struct xstate_hdr *)xfpustate; 1088 bv = ehdr->xstate_bv; 1089 1090 /* 1091 * Avoid #gp. 1092 */ 1093 if (bv & ~xsave_mask) 1094 return (EINVAL); 1095 1096 hdr = (struct xstate_hdr *)(get_pcb_user_save_td(td) + 1); 1097 1098 hdr->xstate_bv = bv; 1099 bcopy(xfpustate + sizeof(struct xstate_hdr), 1100 (char *)(hdr + 1), len - sizeof(struct xstate_hdr)); 1101 1102 return (0); 1103} 1104#endif 1105 1106int 1107npxsetregs(struct thread *td, union savefpu *addr, char *xfpustate, 1108 size_t xfpustate_size) 1109{ 1110 struct pcb *pcb; 1111#ifdef CPU_ENABLE_SSE 1112 int error; 1113#endif 1114 1115 if (!hw_float) 1116 return (ENXIO); 1117 1118 pcb = td->td_pcb; 1119 critical_enter(); 1120 if (td == PCPU_GET(fpcurthread) && PCB_USER_FPU(pcb)) { 1121#ifdef CPU_ENABLE_SSE 1122 error = npxsetxstate(td, xfpustate, xfpustate_size); 1123 if (error != 0) { 1124 critical_exit(); 1125 return (error); 1126 } 1127 if (!cpu_fxsr) 1128#endif 1129 fnclex(); /* As in npxdrop(). */ 1130 bcopy(addr, get_pcb_user_save_td(td), sizeof(*addr)); 1131 fpurstor(get_pcb_user_save_td(td)); 1132 critical_exit(); 1133 pcb->pcb_flags |= PCB_NPXUSERINITDONE | PCB_NPXINITDONE; 1134 } else { 1135 critical_exit(); 1136#ifdef CPU_ENABLE_SSE 1137 error = npxsetxstate(td, xfpustate, xfpustate_size); 1138 if (error != 0) 1139 return (error); 1140#endif 1141 bcopy(addr, get_pcb_user_save_td(td), sizeof(*addr)); 1142 npxuserinited(td); 1143 } 1144 return (0); 1145} 1146 1147static void 1148fpusave(addr) 1149 union savefpu *addr; 1150{ 1151 1152#ifdef CPU_ENABLE_SSE 1153 if (use_xsave) 1154 xsave((char *)addr, xsave_mask); 1155 else if (cpu_fxsr) 1156 fxsave(addr); 1157 else 1158#endif 1159 fnsave(addr); 1160} 1161 1162#ifdef CPU_ENABLE_SSE 1163static void 1164npx_fill_fpregs_xmm1(struct savexmm *sv_xmm, struct save87 *sv_87) 1165{ 1166 struct env87 *penv_87; 1167 struct envxmm *penv_xmm; 1168 int i; 1169 1170 penv_87 = &sv_87->sv_env; 1171 penv_xmm = &sv_xmm->sv_env; 1172 1173 /* FPU control/status */ 1174 penv_87->en_cw = penv_xmm->en_cw; 1175 penv_87->en_sw = penv_xmm->en_sw; 1176 penv_87->en_fip = penv_xmm->en_fip; 1177 penv_87->en_fcs = penv_xmm->en_fcs; 1178 penv_87->en_opcode = penv_xmm->en_opcode; 1179 penv_87->en_foo = penv_xmm->en_foo; 1180 penv_87->en_fos = penv_xmm->en_fos; 1181 1182 /* FPU registers and tags */ 1183 penv_87->en_tw = 0xffff; 1184 for (i = 0; i < 8; ++i) { 1185 sv_87->sv_ac[i] = sv_xmm->sv_fp[i].fp_acc; 1186 if ((penv_xmm->en_tw & (1 << i)) != 0) 1187 /* zero and special are set as valid */ 1188 penv_87->en_tw &= ~(3 << i); 1189 } 1190} 1191 1192void 1193npx_fill_fpregs_xmm(struct savexmm *sv_xmm, struct save87 *sv_87) 1194{ 1195 1196 bzero(sv_87, sizeof(*sv_87)); 1197 npx_fill_fpregs_xmm1(sv_xmm, sv_87); 1198} 1199 1200void 1201npx_set_fpregs_xmm(struct save87 *sv_87, struct savexmm *sv_xmm) 1202{ 1203 struct env87 *penv_87; 1204 struct envxmm *penv_xmm; 1205 int i; 1206 1207 penv_87 = &sv_87->sv_env; 1208 penv_xmm = &sv_xmm->sv_env; 1209 1210 /* FPU control/status */ 1211 penv_xmm->en_cw = penv_87->en_cw; 1212 penv_xmm->en_sw = penv_87->en_sw; 1213 penv_xmm->en_fip = penv_87->en_fip; 1214 penv_xmm->en_fcs = penv_87->en_fcs; 1215 penv_xmm->en_opcode = penv_87->en_opcode; 1216 penv_xmm->en_foo = penv_87->en_foo; 1217 penv_xmm->en_fos = penv_87->en_fos; 1218 1219 /* FPU registers and tags */ 1220 penv_xmm->en_tw = 0; 1221 for (i = 0; i < 8; ++i) { 1222 sv_xmm->sv_fp[i].fp_acc = sv_87->sv_ac[i]; 1223 if ((penv_87->en_tw && (3 << i)) != (3 << i)) 1224 penv_xmm->en_tw |= 1 << i; 1225 } 1226} 1227#endif /* CPU_ENABLE_SSE */ 1228 1229void 1230npx_get_fsave(void *addr) 1231{ 1232 struct thread *td; 1233 union savefpu *sv; 1234 1235 td = curthread; 1236 npxgetregs(td); 1237 sv = get_pcb_user_save_td(td); 1238#ifdef CPU_ENABLE_SSE 1239 if (cpu_fxsr) 1240 npx_fill_fpregs_xmm1(&sv->sv_xmm, addr); 1241 else 1242#endif 1243 bcopy(sv, addr, sizeof(struct env87) + 1244 sizeof(struct fpacc87[8])); 1245} 1246 1247int 1248npx_set_fsave(void *addr) 1249{ 1250 union savefpu sv; 1251 int error; 1252 1253 bzero(&sv, sizeof(sv)); 1254#ifdef CPU_ENABLE_SSE 1255 if (cpu_fxsr) 1256 npx_set_fpregs_xmm(addr, &sv.sv_xmm); 1257 else 1258#endif 1259 bcopy(addr, &sv, sizeof(struct env87) + 1260 sizeof(struct fpacc87[8])); 1261 error = npxsetregs(curthread, &sv, NULL, 0); 1262 return (error); 1263} 1264 1265#ifdef CPU_ENABLE_SSE 1266/* 1267 * On AuthenticAMD processors, the fxrstor instruction does not restore 1268 * the x87's stored last instruction pointer, last data pointer, and last 1269 * opcode values, except in the rare case in which the exception summary 1270 * (ES) bit in the x87 status word is set to 1. 1271 * 1272 * In order to avoid leaking this information across processes, we clean 1273 * these values by performing a dummy load before executing fxrstor(). 1274 */ 1275static void 1276fpu_clean_state(void) 1277{ 1278 static float dummy_variable = 0.0; 1279 u_short status; 1280 1281 /* 1282 * Clear the ES bit in the x87 status word if it is currently 1283 * set, in order to avoid causing a fault in the upcoming load. 1284 */ 1285 fnstsw(&status); 1286 if (status & 0x80) 1287 fnclex(); 1288 1289 /* 1290 * Load the dummy variable into the x87 stack. This mangles 1291 * the x87 stack, but we don't care since we're about to call 1292 * fxrstor() anyway. 1293 */ 1294 __asm __volatile("ffree %%st(7); flds %0" : : "m" (dummy_variable)); 1295} 1296#endif /* CPU_ENABLE_SSE */ 1297 1298static void 1299fpurstor(union savefpu *addr) 1300{ 1301 1302#ifdef CPU_ENABLE_SSE 1303 if (use_xsave) 1304 xrstor((char *)addr, xsave_mask); 1305 else if (cpu_fxsr) 1306 fxrstor(addr); 1307 else 1308#endif 1309 frstor(addr); 1310} 1311 1312#ifdef DEV_ISA 1313/* 1314 * This sucks up the legacy ISA support assignments from PNPBIOS/ACPI. 1315 */ 1316static struct isa_pnp_id npxisa_ids[] = { 1317 { 0x040cd041, "Legacy ISA coprocessor support" }, /* PNP0C04 */ 1318 { 0 } 1319}; 1320 1321static int 1322npxisa_probe(device_t dev) 1323{ 1324 int result; 1325 if ((result = ISA_PNP_PROBE(device_get_parent(dev), dev, npxisa_ids)) <= 0) { 1326 device_quiet(dev); 1327 } 1328 return(result); 1329} 1330 1331static int 1332npxisa_attach(device_t dev) 1333{ 1334 return (0); 1335} 1336 1337static device_method_t npxisa_methods[] = { 1338 /* Device interface */ 1339 DEVMETHOD(device_probe, npxisa_probe), 1340 DEVMETHOD(device_attach, npxisa_attach), 1341 DEVMETHOD(device_detach, bus_generic_detach), 1342 DEVMETHOD(device_shutdown, bus_generic_shutdown), 1343 DEVMETHOD(device_suspend, bus_generic_suspend), 1344 DEVMETHOD(device_resume, bus_generic_resume), 1345 1346 { 0, 0 } 1347}; 1348 1349static driver_t npxisa_driver = { 1350 "npxisa", 1351 npxisa_methods, 1352 1, /* no softc */ 1353}; 1354 1355static devclass_t npxisa_devclass; 1356 1357DRIVER_MODULE(npxisa, isa, npxisa_driver, npxisa_devclass, 0, 0); 1358#ifndef PC98 1359DRIVER_MODULE(npxisa, acpi, npxisa_driver, npxisa_devclass, 0, 0); 1360#endif 1361#endif /* DEV_ISA */ 1362 1363static MALLOC_DEFINE(M_FPUKERN_CTX, "fpukern_ctx", 1364 "Kernel contexts for FPU state"); 1365 1366#define FPU_KERN_CTX_NPXINITDONE 0x01 1367#define FPU_KERN_CTX_DUMMY 0x02 1368#define FPU_KERN_CTX_INUSE 0x04 1369 1370struct fpu_kern_ctx { 1371 union savefpu *prev; 1372 uint32_t flags; 1373 char hwstate1[]; 1374}; 1375 1376struct fpu_kern_ctx * 1377fpu_kern_alloc_ctx(u_int flags) 1378{ 1379 struct fpu_kern_ctx *res; 1380 size_t sz; 1381 1382 sz = sizeof(struct fpu_kern_ctx) + XSAVE_AREA_ALIGN + 1383 cpu_max_ext_state_size; 1384 res = malloc(sz, M_FPUKERN_CTX, ((flags & FPU_KERN_NOWAIT) ? 1385 M_NOWAIT : M_WAITOK) | M_ZERO); 1386 return (res); 1387} 1388 1389void 1390fpu_kern_free_ctx(struct fpu_kern_ctx *ctx) 1391{ 1392 1393 KASSERT((ctx->flags & FPU_KERN_CTX_INUSE) == 0, ("free'ing inuse ctx")); 1394 /* XXXKIB clear the memory ? */ 1395 free(ctx, M_FPUKERN_CTX); 1396} 1397 1398static union savefpu * 1399fpu_kern_ctx_savefpu(struct fpu_kern_ctx *ctx) 1400{ 1401 vm_offset_t p; 1402 1403 p = (vm_offset_t)&ctx->hwstate1; 1404 p = roundup2(p, XSAVE_AREA_ALIGN); 1405 return ((union savefpu *)p); 1406} 1407 1408int 1409fpu_kern_enter(struct thread *td, struct fpu_kern_ctx *ctx, u_int flags) 1410{ 1411 struct pcb *pcb; 1412 1413 KASSERT((ctx->flags & FPU_KERN_CTX_INUSE) == 0, ("using inuse ctx")); 1414 1415 if ((flags & FPU_KERN_KTHR) != 0 && is_fpu_kern_thread(0)) { 1416 ctx->flags = FPU_KERN_CTX_DUMMY | FPU_KERN_CTX_INUSE; 1417 return (0); 1418 } 1419 pcb = td->td_pcb; 1420 KASSERT(!PCB_USER_FPU(pcb) || pcb->pcb_save == 1421 get_pcb_user_save_pcb(pcb), ("mangled pcb_save")); 1422 ctx->flags = FPU_KERN_CTX_INUSE; 1423 if ((pcb->pcb_flags & PCB_NPXINITDONE) != 0) 1424 ctx->flags |= FPU_KERN_CTX_NPXINITDONE; 1425 npxexit(td); 1426 ctx->prev = pcb->pcb_save; 1427 pcb->pcb_save = fpu_kern_ctx_savefpu(ctx); 1428 pcb->pcb_flags |= PCB_KERNNPX; 1429 pcb->pcb_flags &= ~PCB_NPXINITDONE; 1430 return (0); 1431} 1432 1433int 1434fpu_kern_leave(struct thread *td, struct fpu_kern_ctx *ctx) 1435{ 1436 struct pcb *pcb; 1437 1438 KASSERT((ctx->flags & FPU_KERN_CTX_INUSE) != 0, 1439 ("leaving not inuse ctx")); 1440 ctx->flags &= ~FPU_KERN_CTX_INUSE; 1441 1442 if (is_fpu_kern_thread(0) && (ctx->flags & FPU_KERN_CTX_DUMMY) != 0) 1443 return (0); 1444 pcb = td->td_pcb; 1445 critical_enter(); 1446 if (curthread == PCPU_GET(fpcurthread)) 1447 npxdrop(); 1448 critical_exit(); 1449 pcb->pcb_save = ctx->prev; 1450 if (pcb->pcb_save == get_pcb_user_save_pcb(pcb)) { 1451 if ((pcb->pcb_flags & PCB_NPXUSERINITDONE) != 0) 1452 pcb->pcb_flags |= PCB_NPXINITDONE; 1453 else 1454 pcb->pcb_flags &= ~PCB_NPXINITDONE; 1455 pcb->pcb_flags &= ~PCB_KERNNPX; 1456 } else { 1457 if ((ctx->flags & FPU_KERN_CTX_NPXINITDONE) != 0) 1458 pcb->pcb_flags |= PCB_NPXINITDONE; 1459 else 1460 pcb->pcb_flags &= ~PCB_NPXINITDONE; 1461 KASSERT(!PCB_USER_FPU(pcb), ("unpaired fpu_kern_leave")); 1462 } 1463 return (0); 1464} 1465 1466int 1467fpu_kern_thread(u_int flags) 1468{ 1469 1470 KASSERT((curthread->td_pflags & TDP_KTHREAD) != 0, 1471 ("Only kthread may use fpu_kern_thread")); 1472 KASSERT(curpcb->pcb_save == get_pcb_user_save_pcb(curpcb), 1473 ("mangled pcb_save")); 1474 KASSERT(PCB_USER_FPU(curpcb), ("recursive call")); 1475 1476 curpcb->pcb_flags |= PCB_KERNNPX; 1477 return (0); 1478} 1479 1480int 1481is_fpu_kern_thread(u_int flags) 1482{ 1483 1484 if ((curthread->td_pflags & TDP_KTHREAD) == 0) 1485 return (0); 1486 return ((curpcb->pcb_flags & PCB_KERNNPX) != 0); 1487} 1488 1489/* 1490 * FPU save area alloc/free/init utility routines 1491 */ 1492union savefpu * 1493fpu_save_area_alloc(void) 1494{ 1495 1496 return (uma_zalloc(fpu_save_area_zone, 0)); 1497} 1498 1499void 1500fpu_save_area_free(union savefpu *fsa) 1501{ 1502 1503 uma_zfree(fpu_save_area_zone, fsa); 1504} 1505 1506void 1507fpu_save_area_reset(union savefpu *fsa) 1508{ 1509 1510 bcopy(npx_initialstate, fsa, cpu_max_ext_state_size); 1511} 1512