Now available: The Design and Implementation of the FreeBSD Operating System (Second Edition)
[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]
FreeBSD/Linux Kernel Cross Reference
sys/i86pc/ml/locore.s
Version:
- FREEBSD - FREEBSD-13-STABLE - FREEBSD-13-0 - FREEBSD-12-STABLE - FREEBSD-12-0 - FREEBSD-11-STABLE - FREEBSD-11-0 - FREEBSD-10-STABLE - FREEBSD-10-0 - FREEBSD-9-STABLE - FREEBSD-9-0 - FREEBSD-8-STABLE - FREEBSD-8-0 - FREEBSD-7-STABLE - FREEBSD-7-0 - FREEBSD-6-STABLE - FREEBSD-6-0 - FREEBSD-5-STABLE - FREEBSD-5-0 - FREEBSD-4-STABLE - FREEBSD-3-STABLE - FREEBSD22 - l41 - OPENBSD - linux-2.6 - MK84 - PLAN9 - xnu-8792
SearchContext: - none - 3 - 10
SearchContext: - none - 3 - 10
1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21 22 /* 23 * Copyright (c) 1992, 2010, Oracle and/or its affiliates. All rights reserved. 24 */ 25 26 /* Copyright (c) 1990, 1991 UNIX System Laboratories, Inc. */ 27 /* Copyright (c) 1984, 1986, 1987, 1988, 1989, 1990 AT&T */ 28 /* All Rights Reserved */ 29 30 /* Copyright (c) 1987, 1988 Microsoft Corporation */ 31 /* All Rights Reserved */ 32 33 34 #include <sys/asm_linkage.h> 35 #include <sys/asm_misc.h> 36 #include <sys/regset.h> 37 #include <sys/privregs.h> 38 #include <sys/psw.h> 39 #include <sys/reboot.h> 40 #include <sys/x86_archext.h> 41 #include <sys/machparam.h> 42 43 #if defined(__lint) 44 45 #include <sys/types.h> 46 #include <sys/thread.h> 47 #include <sys/systm.h> 48 #include <sys/lgrp.h> 49 #include <sys/regset.h> 50 #include <sys/link.h> 51 #include <sys/bootconf.h> 52 #include <sys/bootsvcs.h> 53 54 #else /* __lint */ 55 56 #include <sys/segments.h> 57 #include <sys/pcb.h> 58 #include <sys/trap.h> 59 #include <sys/ftrace.h> 60 #include <sys/traptrace.h> 61 #include <sys/clock.h> 62 #include <sys/cmn_err.h> 63 #include <sys/pit.h> 64 #include <sys/panic.h> 65 66 #if defined(__xpv) 67 #include <sys/hypervisor.h> 68 #endif 69 70 #include "assym.h" 71 72 /* 73 * Our assumptions: 74 * - We are running in protected-paged mode. 75 * - Interrupts are disabled. 76 * - The GDT and IDT are the callers; we need our copies. 77 * - The kernel's text, initialized data and bss are mapped. 78 * 79 * Our actions: 80 * - Save arguments 81 * - Initialize our stack pointer to the thread 0 stack (t0stack) 82 * and leave room for a phony "struct regs". 83 * - Our GDT and IDT need to get munged. 84 * - Since we are using the boot's GDT descriptors, we need 85 * to copy them into our GDT before we switch to ours. 86 * - We start using our GDT by loading correct values in the 87 * selector registers (cs=KCS_SEL, ds=es=ss=KDS_SEL, fs=KFS_SEL, 88 * gs=KGS_SEL). 89 * - The default LDT entry for syscall is set. 90 * - We load the default LDT into the hardware LDT register. 91 * - We load the default TSS into the hardware task register. 92 * - Check for cpu type, i.e. 486 vs. P5 vs. P6 etc. 93 * - mlsetup(%esp) gets called. 94 * - We change our appearance to look like the real thread 0. 95 * (NOTE: making ourselves to be a real thread may be a noop) 96 * - main() gets called. (NOTE: main() never returns). 97 * 98 * NOW, the real code! 99 */ 100 /* 101 * The very first thing in the kernel's text segment must be a jump 102 * to the os/fakebop.c startup code. 103 */ 104 .text 105 jmp _start 106 107 /* 108 * Globals: 109 */ 110 .globl _locore_start 111 .globl mlsetup 112 .globl main 113 .globl panic 114 .globl t0stack 115 .globl t0 116 .globl sysp 117 .globl edata 118 119 /* 120 * call back into boot - sysp (bootsvcs.h) and bootops (bootconf.h) 121 */ 122 .globl bootops 123 .globl bootopsp 124 125 /* 126 * NOTE: t0stack should be the first thing in the data section so that 127 * if it ever overflows, it will fault on the last kernel text page. 128 */ 129 .data 130 .comm t0stack, DEFAULTSTKSZ, 32 131 .comm t0, 4094, 32 132 133 #endif /* __lint */ 134 135 136 #if defined(__amd64) 137 138 #if defined(__lint) 139 140 /* ARGSUSED */ 141 void 142 _locore_start(struct boot_syscalls *sysp, ulong_t rsi, struct bootops *bop) 143 {} 144 145 #else /* __lint */ 146 147 /* 148 * kobj_init() vectors us back to here with (note) a slightly different 149 * set of arguments than _start is given (see lint prototypes above). 150 * 151 * XXX Make this less vile, please. 152 */ 153 ENTRY_NP(_locore_start) 154 155 /* 156 * %rdi = boot services (should die someday) 157 * %rdx = bootops 158 * end 159 */ 160 161 leaq edata(%rip), %rbp /* reference edata for ksyms */ 162 movq $0, (%rbp) /* limit stack back trace */ 163 164 /* 165 * Initialize our stack pointer to the thread 0 stack (t0stack) 166 * and leave room for a "struct regs" for lwp0. Note that the 167 * stack doesn't actually align to a 16-byte boundary until just 168 * before we call mlsetup because we want to use %rsp to point at 169 * our regs structure. 170 */ 171 leaq t0stack(%rip), %rsp 172 addq $_CONST(DEFAULTSTKSZ - REGSIZE), %rsp 173 #if (REGSIZE & 15) == 0 174 subq $8, %rsp 175 #endif 176 /* 177 * Save call back for special x86 boot services vector 178 */ 179 movq %rdi, sysp(%rip) 180 181 movq %rdx, bootops(%rip) /* save bootops */ 182 movq $bootops, bootopsp(%rip) 183 184 /* 185 * Save arguments and flags, if only for debugging .. 186 */ 187 movq %rdi, REGOFF_RDI(%rsp) 188 movq %rsi, REGOFF_RSI(%rsp) 189 movq %rdx, REGOFF_RDX(%rsp) 190 movq %rcx, REGOFF_RCX(%rsp) 191 movq %r8, REGOFF_R8(%rsp) 192 movq %r9, REGOFF_R9(%rsp) 193 pushf 194 popq %r11 195 movq %r11, REGOFF_RFL(%rsp) 196 197 #if !defined(__xpv) 198 /* 199 * Enable write protect and alignment check faults. 200 */ 201 movq %cr0, %rax 202 orq $_CONST(CR0_WP|CR0_AM), %rax 203 andq $_BITNOT(CR0_WT|CR0_CE), %rax 204 movq %rax, %cr0 205 #endif /* __xpv */ 206 207 /* 208 * (We just assert this works by virtue of being here) 209 */ 210 bts $X86FSET_CPUID, x86_featureset(%rip) 211 212 /* 213 * mlsetup() gets called with a struct regs as argument, while 214 * main takes no args and should never return. 215 */ 216 xorl %ebp, %ebp 217 movq %rsp, %rdi 218 pushq %rbp 219 /* (stack pointer now aligned on 16-byte boundary right here) */ 220 movq %rsp, %rbp 221 call mlsetup 222 call main 223 /* NOTREACHED */ 224 leaq __return_from_main(%rip), %rdi 225 xorl %eax, %eax 226 call panic 227 SET_SIZE(_locore_start) 228 229 #endif /* __amd64 */ 230 #endif /* __lint */ 231 232 #if !defined(__lint) 233 234 __return_from_main: 235 .string "main() returned" 236 __unsupported_cpu: 237 .string "486 style cpu detected - no longer supported!" 238 239 #endif /* !__lint */ 240 241 #if !defined(__amd64) 242 243 #if defined(__lint) 244 245 /* ARGSUSED */ 246 void 247 _locore_start(struct boot_syscalls *sysp, struct bootops *bop) 248 {} 249 250 #else /* __lint */ 251 252 /* 253 * kobj_init() vectors us back to here with (note) a slightly different 254 * set of arguments than _start is given (see lint prototypes above). 255 * 256 * XXX Make this less vile, please. 257 */ 258 ENTRY_NP(_locore_start) 259 260 /* 261 * %ecx = boot services (should die someday) 262 * %ebx = bootops 263 */ 264 mov $edata, %ebp / edata needs to be defined for ksyms 265 movl $0, (%ebp) / limit stack back trace 266 267 /* 268 * Initialize our stack pointer to the thread 0 stack (t0stack) 269 * and leave room for a phony "struct regs". 270 */ 271 movl $t0stack + DEFAULTSTKSZ - REGSIZE, %esp 272 273 /* 274 * Save call back for special x86 boot services vector 275 */ 276 mov %ecx, sysp / save call back for boot services 277 278 mov %ebx, bootops / save bootops 279 movl $bootops, bootopsp 280 281 282 /* 283 * Save all registers and flags 284 */ 285 pushal 286 pushfl 287 288 #if !defined(__xpv) 289 /* 290 * Override bios settings and enable write protect and 291 * alignment check faults. 292 */ 293 movl %cr0, %eax 294 295 /* 296 * enable WP for detecting faults, and enable alignment checking. 297 */ 298 orl $_CONST(CR0_WP|CR0_AM), %eax 299 andl $_BITNOT(CR0_WT|CR0_CE), %eax 300 movl %eax, %cr0 / set the cr0 register correctly and 301 / override the BIOS setup 302 303 /* 304 * If bit 21 of eflags can be flipped, then cpuid is present 305 * and enabled. 306 */ 307 pushfl 308 popl %ecx 309 movl %ecx, %eax 310 xorl $PS_ID, %eax / try complemented bit 311 pushl %eax 312 popfl 313 pushfl 314 popl %eax 315 cmpl %eax, %ecx 316 jne have_cpuid 317 318 /* 319 * cpuid may be disabled on Cyrix, try to detect Cyrix by the 5/2 test 320 * div does not modify the cc flags on Cyrix, even though this may 321 * also be true for other vendors, this is generally true only for 322 * newer models from those vendors that support and do not disable 323 * cpuid (usually because cpuid cannot be disabled) 324 */ 325 326 /* 327 * clear cc flags 328 */ 329 xorb %ah, %ah 330 sahf 331 332 /* 333 * perform 5/2 test 334 */ 335 movw $5, %ax 336 movb $2, %bl 337 divb %bl 338 339 lahf 340 cmpb $2, %ah 341 jne cpu_486 342 343 /* 344 * div did not modify the cc flags, chances are the vendor is Cyrix 345 * assume the vendor is Cyrix and use the CCR's to enable cpuid 346 */ 347 .set CYRIX_CRI, 0x22 / CR Index Register 348 .set CYRIX_CRD, 0x23 / CR Data Register 349 350 .set CYRIX_CCR3, 0xc3 / Config Control Reg 3 351 .set CYRIX_CCR4, 0xe8 / Config Control Reg 4 352 .set CYRIX_DIR0, 0xfe / Device Identification Reg 0 353 .set CYRIX_DIR1, 0xff / Device Identification Reg 1 354 355 /* 356 * even if the cpu vendor is Cyrix and the motherboard/chipset 357 * vendor decided to ignore lines A1-A4 for I/O addresses, I/O port 358 * 0x21 corresponds with 0x23 and since 0x22 is still untouched, 359 * the reads and writes of 0x21 are guaranteed to be off-chip of 360 * the cpu 361 */ 362 363 /* 364 * enable read of ISR at I/O port 0x20 365 */ 366 movb $0xb, %al 367 outb $MCMD_PORT 368 369 /* 370 * read IMR and store in %bl 371 */ 372 inb $MIMR_PORT 373 movb %al, %bl 374 375 /* 376 * mask out all interrupts so that ISR will not change 377 */ 378 movb $0xff, %al 379 outb $MIMR_PORT 380 381 /* 382 * reads of I/O port 0x22 on Cyrix are always directed off-chip 383 * make use of I/O pull-up to test for an unknown device on 0x22 384 */ 385 inb $CYRIX_CRI 386 cmpb $0xff, %al 387 je port_22_free 388 389 /* 390 * motherboard/chipset vendor may be ignoring line A1 of I/O address 391 */ 392 movb %al, %cl 393 394 /* 395 * if the ISR and the value read from 0x22 do not match then we have 396 * detected some unknown device, probably a chipset, at 0x22 397 */ 398 inb $MCMD_PORT 399 cmpb %al, %cl 400 jne restore_IMR 401 402 port_22_free: 403 /* 404 * now test to see if some unknown device is using I/O port 0x23 405 * 406 * read the external I/O port at 0x23 407 */ 408 inb $CYRIX_CRD 409 410 /* 411 * Test for pull-up at 0x23 or if I/O address line A1 is being ignored. 412 * IMR is 0xff so both tests are performed simultaneously. 413 */ 414 cmpb $0xff, %al 415 jne restore_IMR 416 417 /* 418 * We are a Cyrix part. In case we are some model of Cx486 or a Cx586, 419 * record the type and fix it later if not. 420 */ 421 movl $X86_VENDOR_Cyrix, x86_vendor 422 movl $X86_TYPE_CYRIX_486, x86_type 423 424 /* 425 * Try to read CCR3. All Cyrix cpu's which support cpuid have CCR3. 426 * 427 * load CCR3 index into CCR index register 428 */ 429 430 movb $CYRIX_CCR3, %al 431 outb $CYRIX_CRI 432 433 /* 434 * If we are not a Cyrix cpu, then we have performed an external I/O 435 * cycle. If the CCR index was not valid for this Cyrix model, we may 436 * have performed an external I/O cycle as well. In these cases and 437 * if the motherboard/chipset vendor ignores I/O address line A1, 438 * then the PIC will have IRQ3 set at the lowest priority as a side 439 * effect of the above outb. We are reasonalbly confident that there 440 * is not an unknown device on I/O port 0x22, so there should have been 441 * no unpredictable side-effect of the above outb. 442 */ 443 444 /* 445 * read CCR3 446 */ 447 inb $CYRIX_CRD 448 449 /* 450 * If we are not a Cyrix cpu the inb above produced an external I/O 451 * cycle. If we are a Cyrix model that does not support CCR3 wex 452 * produced an external I/O cycle. In all known Cyrix models 6x86 and 453 * above, bit 3 of CCR3 is reserved and cannot be set to 1. In all 454 * Cyrix models prior to the 6x86 that supported CCR3, bits 4-7 are 455 * reserved as well. It is highly unlikely that CCR3 contains the value 456 * 0xff. We test to see if I/O port 0x23 is pull-up or the IMR and 457 * deduce we are not a Cyrix with support for cpuid if so. 458 */ 459 cmpb $0xff, %al 460 je restore_PIC 461 462 /* 463 * There exist 486 ISA Cyrix chips that support CCR3 but do not support 464 * DIR0 and DIR1. If we try to read DIR0, we may generate external I/O 465 * cycles, the exact behavior is model specific and undocumented. 466 * Unfortunately these external I/O cycles may confuse some PIC's beyond 467 * recovery. Fortunatetly we can use the following undocumented trick: 468 * if bit 4 of CCR3 can be toggled, then DIR0 and DIR1 are supported. 469 * Pleasantly MAPEN contains bit 4 of CCR3, so this trick is guaranteed 470 * to work on all Cyrix cpu's which support cpuid. 471 */ 472 movb %al, %dl 473 xorb $0x10, %dl 474 movb %al, %cl 475 476 /* 477 * write back CRR3 with toggled bit 4 to CCR3 478 */ 479 movb $CYRIX_CCR3, %al 480 outb $CYRIX_CRI 481 482 movb %dl, %al 483 outb $CYRIX_CRD 484 485 /* 486 * read CCR3 487 */ 488 movb $CYRIX_CCR3, %al 489 outb $CYRIX_CRI 490 inb $CYRIX_CRD 491 movb %al, %dl 492 493 /* 494 * restore CCR3 495 */ 496 movb $CYRIX_CCR3, %al 497 outb $CYRIX_CRI 498 499 movb %cl, %al 500 outb $CYRIX_CRD 501 502 /* 503 * if bit 4 was not toggled DIR0 and DIR1 are not supported in which 504 * case we do not have cpuid anyway 505 */ 506 andb $0x10, %al 507 andb $0x10, %dl 508 cmpb %al, %dl 509 je restore_PIC 510 511 /* 512 * read DIR0 513 */ 514 movb $CYRIX_DIR0, %al 515 outb $CYRIX_CRI 516 inb $CYRIX_CRD 517 518 /* 519 * test for pull-up 520 */ 521 cmpb $0xff, %al 522 je restore_PIC 523 524 /* 525 * Values of 0x20-0x27 in DIR0 are currently reserved by Cyrix for 526 * future use. If Cyrix ever produces a cpu that supports cpuid with 527 * these ids, the following test will have to change. For now we remain 528 * pessimistic since the formats of the CRR's may be different then. 529 * 530 * test for at least a 6x86, to see if we support both MAPEN and CPUID 531 */ 532 cmpb $0x30, %al 533 jb restore_IMR 534 535 /* 536 * enable MAPEN 537 */ 538 movb $CYRIX_CCR3, %al 539 outb $CYRIX_CRI 540 541 andb $0xf, %cl 542 movb %cl, %al 543 orb $0x10, %al 544 outb $CYRIX_CRD 545 546 /* 547 * select CCR4 548 */ 549 movb $CYRIX_CCR4, %al 550 outb $CYRIX_CRI 551 552 /* 553 * read CCR4 554 */ 555 inb $CYRIX_CRD 556 557 /* 558 * enable cpuid 559 */ 560 orb $0x80, %al 561 movb %al, %dl 562 563 /* 564 * select CCR4 565 */ 566 movb $CYRIX_CCR4, %al 567 outb $CYRIX_CRI 568 569 /* 570 * write CCR4 571 */ 572 movb %dl, %al 573 outb $CYRIX_CRD 574 575 /* 576 * select CCR3 577 */ 578 movb $CYRIX_CCR3, %al 579 outb $CYRIX_CRI 580 581 /* 582 * disable MAPEN and write CCR3 583 */ 584 movb %cl, %al 585 outb $CYRIX_CRD 586 587 /* 588 * restore IMR 589 */ 590 movb %bl, %al 591 outb $MIMR_PORT 592 593 /* 594 * test to see if cpuid available 595 */ 596 pushfl 597 popl %ecx 598 movl %ecx, %eax 599 xorl $PS_ID, %eax / try complemented bit 600 pushl %eax 601 popfl 602 pushfl 603 popl %eax 604 cmpl %eax, %ecx 605 jne have_cpuid 606 jmp cpu_486 607 608 restore_PIC: 609 /* 610 * In case the motherboard/chipset vendor is ignoring line A1 of the 611 * I/O address, we set the PIC priorities to sane values. 612 */ 613 movb $0xc7, %al / irq 7 lowest priority 614 outb $MCMD_PORT 615 616 restore_IMR: 617 movb %bl, %al 618 outb $MIMR_PORT 619 jmp cpu_486 620 621 have_cpuid: 622 /* 623 * cpuid instruction present 624 */ 625 bts $X86FSET_CPUID, x86_featureset / Just to set; Ignore the CF 626 movl $0, %eax 627 cpuid 628 629 movl %ebx, cpu_vendor 630 movl %edx, cpu_vendor+4 631 movl %ecx, cpu_vendor+8 632 633 /* 634 * early cyrix cpus are somewhat strange and need to be 635 * probed in curious ways to determine their identity 636 */ 637 638 leal cpu_vendor, %esi 639 leal CyrixInstead, %edi 640 movl $12, %ecx 641 repz 642 cmpsb 643 je vendor_is_cyrix 644 645 / let mlsetup()/cpuid_pass1() handle everything else in C 646 647 jmp cpu_done 648 649 is486: 650 /* 651 * test to see if a useful cpuid 652 */ 653 testl %eax, %eax 654 jz isa486 655 656 movl $1, %eax 657 cpuid 658 659 movl %eax, %ebx 660 andl $0xF00, %ebx 661 cmpl $0x400, %ebx 662 je isa486 663 664 rep; ret /* use 2 byte return instruction */ 665 /* AMD Software Optimization Guide - Section 6.2 */ 666 isa486: 667 /* 668 * lose the return address 669 */ 670 popl %eax 671 jmp cpu_486 672 673 vendor_is_cyrix: 674 call is486 675 676 /* 677 * Processor signature and feature flags for Cyrix are insane. 678 * BIOS can play with semi-documented registers, so cpuid must be used 679 * cautiously. Since we are Cyrix that has cpuid, we have DIR0 and DIR1 680 * Keep the family in %ebx and feature flags in %edx until not needed 681 */ 682 683 /* 684 * read DIR0 685 */ 686 movb $CYRIX_DIR0, %al 687 outb $CYRIX_CRI 688 inb $CYRIX_CRD 689 690 /* 691 * First we handle the cases where we are a 6x86 or 6x86L. 692 * The 6x86 is basically a 486, the only reliable bit in the 693 * feature flags is for FPU. The 6x86L is better, unfortunately 694 * there is no really good way to distinguish between these two 695 * cpu's. We are pessimistic and when in doubt assume 6x86. 696 */ 697 698 cmpb $0x40, %al 699 jae maybeGX 700 701 /* 702 * We are an M1, either a 6x86 or 6x86L. 703 */ 704 cmpb $0x30, %al 705 je maybe6x86L 706 cmpb $0x31, %al 707 je maybe6x86L 708 cmpb $0x34, %al 709 je maybe6x86L 710 cmpb $0x35, %al 711 je maybe6x86L 712 713 /* 714 * although it is possible that we are a 6x86L, the cpu and 715 * documentation are so buggy, we just do not care. 716 */ 717 jmp likely6x86 718 719 maybe6x86L: 720 /* 721 * read DIR1 722 */ 723 movb $CYRIX_DIR1, %al 724 outb $CYRIX_CRI 725 inb $CYRIX_CRD 726 cmpb $0x22, %al 727 jb likely6x86 728 729 /* 730 * We are a 6x86L, or at least a 6x86 with honest cpuid feature flags 731 */ 732 movl $X86_TYPE_CYRIX_6x86L, x86_type 733 jmp coma_bug 734 735 likely6x86: 736 /* 737 * We are likely a 6x86, or a 6x86L without a way of knowing 738 * 739 * The 6x86 has NO Pentium or Pentium Pro compatible features even 740 * though it claims to be a Pentium Pro compatible! 741 * 742 * The 6x86 core used in the 6x86 may have most of the Pentium system 743 * registers and largely conform to the Pentium System Programming 744 * Reference. Documentation on these parts is long gone. Treat it as 745 * a crippled Pentium and hope for the best. 746 */ 747 748 movl $X86_TYPE_CYRIX_6x86, x86_type 749 jmp coma_bug 750 751 maybeGX: 752 /* 753 * Now we check whether we are a MediaGX or GXm. We have particular 754 * reason for concern here. Even though most of the GXm's 755 * report having TSC in the cpuid feature flags, the TSC may be 756 * horribly broken. What is worse, is that MediaGX's are basically 757 * 486's while the good GXm's are more like Pentium Pro's! 758 */ 759 760 cmpb $0x50, %al 761 jae maybeM2 762 763 /* 764 * We are either a MediaGX (sometimes called a Gx86) or GXm 765 */ 766 767 cmpb $41, %al 768 je maybeMediaGX 769 770 cmpb $44, %al 771 jb maybeGXm 772 773 cmpb $47, %al 774 jbe maybeMediaGX 775 776 /* 777 * We do not honestly know what we are, so assume a MediaGX 778 */ 779 jmp media_gx 780 781 maybeGXm: 782 /* 783 * It is still possible we are either a MediaGX or GXm, trust cpuid 784 * family should be 5 on a GXm 785 */ 786 cmpl $0x500, %ebx 787 je GXm 788 789 /* 790 * BIOS/Cyrix might set family to 6 on a GXm 791 */ 792 cmpl $0x600, %ebx 793 jne media_gx 794 795 GXm: 796 movl $X86_TYPE_CYRIX_GXm, x86_type 797 jmp cpu_done 798 799 maybeMediaGX: 800 /* 801 * read DIR1 802 */ 803 movb $CYRIX_DIR1, %al 804 outb $CYRIX_CRI 805 inb $CYRIX_CRD 806 807 cmpb $0x30, %al 808 jae maybeGXm 809 810 /* 811 * we are a MediaGX for which we do not trust cpuid 812 */ 813 media_gx: 814 movl $X86_TYPE_CYRIX_MediaGX, x86_type 815 jmp cpu_486 816 817 maybeM2: 818 /* 819 * Now we check whether we are a 6x86MX or MII. These cpu's are 820 * virtually identical, but we care because for the 6x86MX, we 821 * must work around the coma bug. Also for 6x86MX prior to revision 822 * 1.4, the TSC may have serious bugs. 823 */ 824 825 cmpb $0x60, %al 826 jae maybeM3 827 828 /* 829 * family should be 6, but BIOS/Cyrix might set it to 5 830 */ 831 cmpl $0x600, %ebx 832 ja cpu_486 833 834 /* 835 * read DIR1 836 */ 837 movb $CYRIX_DIR1, %al 838 outb $CYRIX_CRI 839 inb $CYRIX_CRD 840 841 cmpb $0x8, %al 842 jb cyrix6x86MX 843 cmpb $0x80, %al 844 jb MII 845 846 cyrix6x86MX: 847 /* 848 * It is altogether unclear how the revision stamped on the cpu 849 * maps to the values in DIR0 and DIR1. Just assume TSC is broken. 850 */ 851 movl $X86_TYPE_CYRIX_6x86MX, x86_type 852 jmp coma_bug 853 854 MII: 855 movl $X86_TYPE_CYRIX_MII, x86_type 856 likeMII: 857 jmp cpu_done 858 859 maybeM3: 860 /* 861 * We are some chip that we cannot identify yet, an MIII perhaps. 862 * We will be optimistic and hope that the chip is much like an MII, 863 * and that cpuid is sane. Cyrix seemed to have gotten it right in 864 * time for the MII, we can only hope it stayed that way. 865 * Maybe the BIOS or Cyrix is trying to hint at something 866 */ 867 cmpl $0x500, %ebx 868 je GXm 869 870 cmpb $0x80, %al 871 jae likelyM3 872 873 /* 874 * Just test for the features Cyrix is known for 875 */ 876 877 jmp MII 878 879 likelyM3: 880 /* 881 * DIR0 with values from 0x80 to 0x8f indicates a VIA Cyrix III, aka 882 * the Cyrix MIII. There may be parts later that use the same ranges 883 * for DIR0 with special values in DIR1, maybe the VIA CIII, but for 884 * now we will call anything with a DIR0 of 0x80 or higher an MIII. 885 * The MIII is supposed to support large pages, but we will believe 886 * it when we see it. For now we just enable and test for MII features. 887 */ 888 movl $X86_TYPE_VIA_CYRIX_III, x86_type 889 jmp likeMII 890 891 coma_bug: 892 893 /* 894 * With NO_LOCK set to 0 in CCR1, the usual state that BIOS enforces, some 895 * bus cycles are issued with LOCK# asserted. With NO_LOCK set to 1, all bus 896 * cycles except page table accesses and interrupt ACK cycles do not assert 897 * LOCK#. xchgl is an instruction that asserts LOCK# if NO_LOCK is set to 0. 898 * Due to a bug in the cpu core involving over-optimization of branch 899 * prediction, register renaming, and execution of instructions down both the 900 * X and Y pipes for the xchgl instruction, short loops can be written that 901 * never de-assert LOCK# from one invocation of the loop to the next, ad 902 * infinitum. The undesirable effect of this situation is that interrupts are 903 * not serviced. The ideal workaround to this bug would be to set NO_LOCK to 904 * 1. Unfortunately bus cycles that would otherwise have asserted LOCK# no 905 * longer do, unless they are page table accesses or interrupt ACK cycles. 906 * With LOCK# not asserted, these bus cycles are now cached. This can cause 907 * undesirable behaviour if the ARR's are not configured correctly. Solaris 908 * does not configure the ARR's, nor does it provide any useful mechanism for 909 * doing so, thus the ideal workaround is not viable. Fortunately, the only 910 * known exploits for this bug involve the xchgl instruction specifically. 911 * There is a group of undocumented registers on Cyrix 6x86, 6x86L, and 912 * 6x86MX cpu's which can be used to specify one instruction as a serializing 913 * instruction. With the xchgl instruction serialized, LOCK# is still 914 * asserted, but it is the sole instruction for which LOCK# is asserted. 915 * There is now some added penalty for the xchgl instruction, but the usual 916 * bus locking is preserved. This ingenious workaround was discovered by 917 * disassembling a binary provided by Cyrix as a workaround for this bug on 918 * Windows, but its not documented anywhere by Cyrix, nor is the bug actually 919 * mentioned in any public errata! The only concern for this workaround is 920 * that there may be similar undiscovered bugs with other instructions that 921 * assert LOCK# that may be leveraged to similar ends. The fact that Cyrix 922 * fixed this bug sometime late in 1997 and no other exploits other than 923 * xchgl have been discovered is good indication that this workaround is 924 * reasonable. 925 */ 926 927 .set CYRIX_DBR0, 0x30 / Debug Register 0 928 .set CYRIX_DBR1, 0x31 / Debug Register 1 929 .set CYRIX_DBR2, 0x32 / Debug Register 2 930 .set CYRIX_DBR3, 0x33 / Debug Register 3 931 .set CYRIX_DOR, 0x3c / Debug Opcode Register 932 933 /* 934 * What is known about DBR1, DBR2, DBR3, and DOR is that for normal 935 * cpu execution DBR1, DBR2, and DBR3 are set to 0. To obtain opcode 936 * serialization, DBR1, DBR2, and DBR3 are loaded with 0xb8, 0x7f, 937 * and 0xff. Then, DOR is loaded with the one byte opcode. 938 */ 939 940 /* 941 * select CCR3 942 */ 943 movb $CYRIX_CCR3, %al 944 outb $CYRIX_CRI 945 946 /* 947 * read CCR3 and mask out MAPEN 948 */ 949 inb $CYRIX_CRD 950 andb $0xf, %al 951 952 /* 953 * save masked CCR3 in %ah 954 */ 955 movb %al, %ah 956 957 /* 958 * select CCR3 959 */ 960 movb $CYRIX_CCR3, %al 961 outb $CYRIX_CRI 962 963 /* 964 * enable MAPEN 965 */ 966 movb %ah, %al 967 orb $0x10, %al 968 outb $CYRIX_CRD 969 970 /* 971 * read DBR0 972 */ 973 movb $CYRIX_DBR0, %al 974 outb $CYRIX_CRI 975 inb $CYRIX_CRD 976 977 /* 978 * disable MATCH and save in %bh 979 */ 980 orb $0x80, %al 981 movb %al, %bh 982 983 /* 984 * write DBR0 985 */ 986 movb $CYRIX_DBR0, %al 987 outb $CYRIX_CRI 988 movb %bh, %al 989 outb $CYRIX_CRD 990 991 /* 992 * write DBR1 993 */ 994 movb $CYRIX_DBR1, %al 995 outb $CYRIX_CRI 996 movb $0xf8, %al 997 outb $CYRIX_CRD 998 999 /* 1000 * write DBR2 1001 */ 1002 movb $CYRIX_DBR2, %al 1003 outb $CYRIX_CRI 1004 movb $0x7f, %al 1005 outb $CYRIX_CRD 1006 1007 /* 1008 * write DBR3 1009 */ 1010 movb $CYRIX_DBR3, %al 1011 outb $CYRIX_CRI 1012 xorb %al, %al 1013 outb $CYRIX_CRD 1014 1015 /* 1016 * write DOR 1017 */ 1018 movb $CYRIX_DOR, %al 1019 outb $CYRIX_CRI 1020 movb $0x87, %al 1021 outb $CYRIX_CRD 1022 1023 /* 1024 * enable MATCH 1025 */ 1026 movb $CYRIX_DBR0, %al 1027 outb $CYRIX_CRI 1028 movb %bh, %al 1029 andb $0x7f, %al 1030 outb $CYRIX_CRD 1031 1032 /* 1033 * disable MAPEN 1034 */ 1035 movb $0xc3, %al 1036 outb $CYRIX_CRI 1037 movb %ah, %al 1038 outb $CYRIX_CRD 1039 1040 jmp cpu_done 1041 1042 cpu_done: 1043 1044 popfl /* Restore original FLAGS */ 1045 popal /* Restore all registers */ 1046 1047 #endif /* !__xpv */ 1048 1049 /* 1050 * mlsetup(%esp) gets called. 1051 */ 1052 pushl %esp 1053 call mlsetup 1054 addl $4, %esp 1055 1056 /* 1057 * We change our appearance to look like the real thread 0. 1058 * (NOTE: making ourselves to be a real thread may be a noop) 1059 * main() gets called. (NOTE: main() never returns). 1060 */ 1061 call main 1062 /* NOTREACHED */ 1063 pushl $__return_from_main 1064 call panic 1065 1066 /* NOTREACHED */ 1067 cpu_486: 1068 pushl $__unsupported_cpu 1069 call panic 1070 SET_SIZE(_locore_start) 1071 1072 #endif /* __lint */ 1073 #endif /* !__amd64 */ 1074 1075 1076 /* 1077 * For stack layout, see privregs.h 1078 * When cmntrap gets called, the error code and trap number have been pushed. 1079 * When cmntrap_pushed gets called, the entire struct regs has been pushed. 1080 */ 1081 1082 #if defined(__lint) 1083 1084 /* ARGSUSED */ 1085 void 1086 cmntrap() 1087 {} 1088 1089 #else /* __lint */ 1090 1091 .globl trap /* C handler called below */ 1092 1093 #if defined(__amd64) 1094 1095 ENTRY_NP2(cmntrap, _cmntrap) 1096 1097 INTR_PUSH 1098 1099 ALTENTRY(cmntrap_pushed) 1100 1101 movq %rsp, %rbp 1102 1103 /* 1104 * - if this is a #pf i.e. T_PGFLT, %r15 is live 1105 * and contains the faulting address i.e. a copy of %cr2 1106 * 1107 * - if this is a #db i.e. T_SGLSTP, %r15 is live 1108 * and contains the value of %db6 1109 */ 1110 1111 TRACE_PTR(%rdi, %rbx, %ebx, %rcx, $TT_TRAP) /* Uses labels 8 and 9 */ 1112 TRACE_REGS(%rdi, %rsp, %rbx, %rcx) /* Uses label 9 */ 1113 TRACE_STAMP(%rdi) /* Clobbers %eax, %edx, uses 9 */ 1114 1115 /* 1116 * We must first check if DTrace has set its NOFAULT bit. This 1117 * regrettably must happen before the trap stack is recorded, because 1118 * this requires a call to getpcstack() and may induce recursion if an 1119 * fbt::getpcstack: enabling is inducing the bad load. 1120 */ 1121 movl %gs:CPU_ID, %eax 1122 shlq $CPU_CORE_SHIFT, %rax 1123 leaq cpu_core(%rip), %r8 1124 addq %r8, %rax 1125 movw CPUC_DTRACE_FLAGS(%rax), %cx 1126 testw $CPU_DTRACE_NOFAULT, %cx 1127 jnz .dtrace_induced 1128 1129 TRACE_STACK(%rdi) 1130 1131 movq %rbp, %rdi 1132 movq %r15, %rsi 1133 movl %gs:CPU_ID, %edx 1134 1135 /* 1136 * We know that this isn't a DTrace non-faulting load; we can now safely 1137 * reenable interrupts. (In the case of pagefaults, we enter through an 1138 * interrupt gate.) 1139 */ 1140 ENABLE_INTR_FLAGS 1141 1142 call trap /* trap(rp, addr, cpuid) handles all traps */ 1143 jmp _sys_rtt 1144 1145 .dtrace_induced: 1146 cmpw $KCS_SEL, REGOFF_CS(%rbp) /* test CS for user-mode trap */ 1147 jne 2f /* if from user, panic */ 1148 1149 cmpl $T_PGFLT, REGOFF_TRAPNO(%rbp) 1150 je 0f 1151 1152 cmpl $T_GPFLT, REGOFF_TRAPNO(%rbp) 1153 jne 3f /* if not PF or GP, panic */ 1154 1155 /* 1156 * If we've taken a GPF, we don't (unfortunately) have the address that 1157 * induced the fault. So instead of setting the fault to BADADDR, 1158 * we'll set the fault to ILLOP. 1159 */ 1160 orw $CPU_DTRACE_ILLOP, %cx 1161 movw %cx, CPUC_DTRACE_FLAGS(%rax) 1162 jmp 1f 1163 0: 1164 orw $CPU_DTRACE_BADADDR, %cx 1165 movw %cx, CPUC_DTRACE_FLAGS(%rax) /* set fault to bad addr */ 1166 movq %r15, CPUC_DTRACE_ILLVAL(%rax) 1167 /* fault addr is illegal value */ 1168 1: 1169 movq REGOFF_RIP(%rbp), %rdi 1170 movq %rdi, %r12 1171 call dtrace_instr_size 1172 addq %rax, %r12 1173 movq %r12, REGOFF_RIP(%rbp) 1174 INTR_POP 1175 IRET 1176 /*NOTREACHED*/ 1177 2: 1178 leaq dtrace_badflags(%rip), %rdi 1179 xorl %eax, %eax 1180 call panic 1181 3: 1182 leaq dtrace_badtrap(%rip), %rdi 1183 xorl %eax, %eax 1184 call panic 1185 SET_SIZE(cmntrap) 1186 SET_SIZE(_cmntrap) 1187 1188 #elif defined(__i386) 1189 1190 1191 ENTRY_NP2(cmntrap, _cmntrap) 1192 1193 INTR_PUSH 1194 1195 ALTENTRY(cmntrap_pushed) 1196 1197 movl %esp, %ebp 1198 1199 /* 1200 * - if this is a #pf i.e. T_PGFLT, %esi is live 1201 * and contains the faulting address i.e. a copy of %cr2 1202 * 1203 * - if this is a #db i.e. T_SGLSTP, %esi is live 1204 * and contains the value of %db6 1205 */ 1206 1207 TRACE_PTR(%edi, %ebx, %ebx, %ecx, $TT_TRAP) /* Uses labels 8 and 9 */ 1208 TRACE_REGS(%edi, %esp, %ebx, %ecx) /* Uses label 9 */ 1209 TRACE_STAMP(%edi) /* Clobbers %eax, %edx, uses 9 */ 1210 1211 /* 1212 * We must first check if DTrace has set its NOFAULT bit. This 1213 * regrettably must happen before the trap stack is recorded, because 1214 * this requires a call to getpcstack() and may induce recursion if an 1215 * fbt::getpcstack: enabling is inducing the bad load. 1216 */ 1217 movl %gs:CPU_ID, %eax 1218 shll $CPU_CORE_SHIFT, %eax 1219 addl $cpu_core, %eax 1220 movw CPUC_DTRACE_FLAGS(%eax), %cx 1221 testw $CPU_DTRACE_NOFAULT, %cx 1222 jnz .dtrace_induced 1223 1224 TRACE_STACK(%edi) 1225 1226 pushl %gs:CPU_ID 1227 pushl %esi /* fault address for PGFLTs */ 1228 pushl %ebp /* ®s */ 1229 1230 /* 1231 * We know that this isn't a DTrace non-faulting load; we can now safely 1232 * reenable interrupts. (In the case of pagefaults, we enter through an 1233 * interrupt gate.) 1234 */ 1235 ENABLE_INTR_FLAGS 1236 1237 call trap /* trap(rp, addr, cpuid) handles all traps */ 1238 addl $12, %esp /* get argument off stack */ 1239 jmp _sys_rtt 1240 1241 .dtrace_induced: 1242 cmpw $KCS_SEL, REGOFF_CS(%ebp) /* test CS for user-mode trap */ 1243 jne 2f /* if from user, panic */ 1244 1245 cmpl $T_PGFLT, REGOFF_TRAPNO(%ebp) 1246 je 0f 1247 1248 cmpl $T_GPFLT, REGOFF_TRAPNO(%ebp) 1249 jne 3f /* if not PF or GP, panic */ 1250 1251 /* 1252 * If we've taken a GPF, we don't (unfortunately) have the address that 1253 * induced the fault. So instead of setting the fault to BADADDR, 1254 * we'll set the fault to ILLOP. 1255 */ 1256 orw $CPU_DTRACE_ILLOP, %cx 1257 movw %cx, CPUC_DTRACE_FLAGS(%eax) 1258 jmp 1f 1259 0: 1260 orw $CPU_DTRACE_BADADDR, %cx 1261 movw %cx, CPUC_DTRACE_FLAGS(%eax) /* set fault to bad addr */ 1262 movl %esi, CPUC_DTRACE_ILLVAL(%eax) 1263 /* fault addr is illegal value */ 1264 1: 1265 pushl REGOFF_EIP(%ebp) 1266 call dtrace_instr_size 1267 addl $4, %esp 1268 movl REGOFF_EIP(%ebp), %ecx 1269 addl %eax, %ecx 1270 movl %ecx, REGOFF_EIP(%ebp) 1271 INTR_POP_KERNEL 1272 IRET 1273 /*NOTREACHED*/ 1274 2: 1275 pushl $dtrace_badflags 1276 call panic 1277 3: 1278 pushl $dtrace_badtrap 1279 call panic 1280 SET_SIZE(cmntrap) 1281 SET_SIZE(_cmntrap) 1282 1283 #endif /* __i386 */ 1284 1285 /* 1286 * Declare a uintptr_t which has the size of _cmntrap to enable stack 1287 * traceback code to know when a regs structure is on the stack. 1288 */ 1289 .globl _cmntrap_size 1290 .align CLONGSIZE 1291 _cmntrap_size: 1292 .NWORD . - _cmntrap 1293 .type _cmntrap_size, @object 1294 1295 dtrace_badflags: 1296 .string "bad DTrace flags" 1297 1298 dtrace_badtrap: 1299 .string "bad DTrace trap" 1300 1301 #endif /* __lint */ 1302 1303 #if defined(__lint) 1304 1305 /* ARGSUSED */ 1306 void 1307 cmninttrap() 1308 {} 1309 1310 #if !defined(__xpv) 1311 void 1312 bop_trap_handler(void) 1313 {} 1314 #endif 1315 1316 #else /* __lint */ 1317 1318 .globl trap /* C handler called below */ 1319 1320 #if defined(__amd64) 1321 1322 ENTRY_NP(cmninttrap) 1323 1324 INTR_PUSH 1325 INTGATE_INIT_KERNEL_FLAGS 1326 1327 TRACE_PTR(%rdi, %rbx, %ebx, %rcx, $TT_TRAP) /* Uses labels 8 and 9 */ 1328 TRACE_REGS(%rdi, %rsp, %rbx, %rcx) /* Uses label 9 */ 1329 TRACE_STAMP(%rdi) /* Clobbers %eax, %edx, uses 9 */ 1330 1331 movq %rsp, %rbp 1332 1333 movl %gs:CPU_ID, %edx 1334 xorl %esi, %esi 1335 movq %rsp, %rdi 1336 call trap /* trap(rp, addr, cpuid) handles all traps */ 1337 jmp _sys_rtt 1338 SET_SIZE(cmninttrap) 1339 1340 #if !defined(__xpv) 1341 /* 1342 * Handle traps early in boot. Just revectors into C quickly as 1343 * these are always fatal errors. 1344 * 1345 * Adjust %rsp to get same stack layout as in 32bit mode for bop_trap(). 1346 */ 1347 ENTRY(bop_trap_handler) 1348 movq %rsp, %rdi 1349 sub $8, %rsp 1350 call bop_trap 1351 SET_SIZE(bop_trap_handler) 1352 #endif 1353 1354 #elif defined(__i386) 1355 1356 ENTRY_NP(cmninttrap) 1357 1358 INTR_PUSH 1359 INTGATE_INIT_KERNEL_FLAGS 1360 1361 TRACE_PTR(%edi, %ebx, %ebx, %ecx, $TT_TRAP) /* Uses labels 8 and 9 */ 1362 TRACE_REGS(%edi, %esp, %ebx, %ecx) /* Uses label 9 */ 1363 TRACE_STAMP(%edi) /* Clobbers %eax, %edx, uses 9 */ 1364 1365 movl %esp, %ebp 1366 1367 TRACE_STACK(%edi) 1368 1369 pushl %gs:CPU_ID 1370 pushl $0 1371 pushl %ebp 1372 call trap /* trap(rp, addr, cpuid) handles all traps */ 1373 addl $12, %esp 1374 jmp _sys_rtt 1375 SET_SIZE(cmninttrap) 1376 1377 #if !defined(__xpv) 1378 /* 1379 * Handle traps early in boot. Just revectors into C quickly as 1380 * these are always fatal errors. 1381 */ 1382 ENTRY(bop_trap_handler) 1383 movl %esp, %eax 1384 pushl %eax 1385 call bop_trap 1386 SET_SIZE(bop_trap_handler) 1387 #endif 1388 1389 #endif /* __i386 */ 1390 1391 #endif /* __lint */ 1392 1393 #if defined(__lint) 1394 1395 /* ARGSUSED */ 1396 void 1397 dtrace_trap() 1398 {} 1399 1400 #else /* __lint */ 1401 1402 .globl dtrace_user_probe 1403 1404 #if defined(__amd64) 1405 1406 ENTRY_NP(dtrace_trap) 1407 1408 INTR_PUSH 1409 1410 TRACE_PTR(%rdi, %rbx, %ebx, %rcx, $TT_TRAP) /* Uses labels 8 and 9 */ 1411 TRACE_REGS(%rdi, %rsp, %rbx, %rcx) /* Uses label 9 */ 1412 TRACE_STAMP(%rdi) /* Clobbers %eax, %edx, uses 9 */ 1413 1414 movq %rsp, %rbp 1415 1416 movl %gs:CPU_ID, %edx 1417 #if defined(__xpv) 1418 movq %gs:CPU_VCPU_INFO, %rsi 1419 movq VCPU_INFO_ARCH_CR2(%rsi), %rsi 1420 #else 1421 movq %cr2, %rsi 1422 #endif 1423 movq %rsp, %rdi 1424 1425 ENABLE_INTR_FLAGS 1426 1427 call dtrace_user_probe /* dtrace_user_probe(rp, addr, cpuid) */ 1428 jmp _sys_rtt 1429 1430 SET_SIZE(dtrace_trap) 1431 1432 #elif defined(__i386) 1433 1434 ENTRY_NP(dtrace_trap) 1435 1436 INTR_PUSH 1437 1438 TRACE_PTR(%edi, %ebx, %ebx, %ecx, $TT_TRAP) /* Uses labels 8 and 9 */ 1439 TRACE_REGS(%edi, %esp, %ebx, %ecx) /* Uses label 9 */ 1440 TRACE_STAMP(%edi) /* Clobbers %eax, %edx, uses 9 */ 1441 1442 movl %esp, %ebp 1443 1444 pushl %gs:CPU_ID 1445 #if defined(__xpv) 1446 movl %gs:CPU_VCPU_INFO, %eax 1447 movl VCPU_INFO_ARCH_CR2(%eax), %eax 1448 #else 1449 movl %cr2, %eax 1450 #endif 1451 pushl %eax 1452 pushl %ebp 1453 1454 ENABLE_INTR_FLAGS 1455 1456 call dtrace_user_probe /* dtrace_user_probe(rp, addr, cpuid) */ 1457 addl $12, %esp /* get argument off stack */ 1458 1459 jmp _sys_rtt 1460 SET_SIZE(dtrace_trap) 1461 1462 #endif /* __i386 */ 1463 1464 #endif /* __lint */ 1465 1466 /* 1467 * Return from _sys_trap routine. 1468 */ 1469 1470 #if defined(__lint) 1471 1472 void 1473 lwp_rtt_initial(void) 1474 {} 1475 1476 void 1477 lwp_rtt(void) 1478 {} 1479 1480 void 1481 _sys_rtt(void) 1482 {} 1483 1484 #else /* __lint */ 1485 1486 #if defined(__amd64) 1487 1488 ENTRY_NP(lwp_rtt_initial) 1489 movq %gs:CPU_THREAD, %r15 1490 movq T_STACK(%r15), %rsp /* switch to the thread stack */ 1491 movq %rsp, %rbp 1492 call __dtrace_probe___proc_start 1493 jmp _lwp_rtt 1494 1495 ENTRY_NP(lwp_rtt) 1496 1497 /* 1498 * r14 lwp 1499 * rdx lwp->lwp_procp 1500 * r15 curthread 1501 */ 1502 1503 movq %gs:CPU_THREAD, %r15 1504 movq T_STACK(%r15), %rsp /* switch to the thread stack */ 1505 movq %rsp, %rbp 1506 _lwp_rtt: 1507 call __dtrace_probe___proc_lwp__start 1508 movq %gs:CPU_LWP, %r14 1509 movq LWP_PROCP(%r14), %rdx 1510 1511 /* 1512 * XX64 Is the stack misaligned correctly at this point? 1513 * If not, we need to do a push before calling anything .. 1514 */ 1515 1516 #if defined(DEBUG) 1517 /* 1518 * If we were to run lwp_savectx at this point -without- 1519 * pcb_rupdate being set to 1, we'd end up sampling the hardware 1520 * state left by the previous running lwp, rather than setting 1521 * the values requested by the lwp creator. Bad. 1522 */ 1523 testb $0x1, PCB_RUPDATE(%r14) 1524 jne 1f 1525 leaq _no_pending_updates(%rip), %rdi 1526 movl $__LINE__, %esi 1527 movq %r14, %rdx 1528 xorl %eax, %eax 1529 call panic 1530 _no_pending_updates: 1531 .string "locore.s:%d lwp_rtt(lwp %p) but pcb_rupdate != 1" 1532 1: 1533 #endif 1534 1535 /* 1536 * If agent lwp, clear %fs and %gs 1537 */ 1538 cmpq %r15, P_AGENTTP(%rdx) 1539 jne 1f 1540 xorl %ecx, %ecx 1541 movq %rcx, REGOFF_FS(%rsp) 1542 movq %rcx, REGOFF_GS(%rsp) 1543 movw %cx, LWP_PCB_FS(%r14) 1544 movw %cx, LWP_PCB_GS(%r14) 1545 1: 1546 call dtrace_systrace_rtt 1547 movq REGOFF_RDX(%rsp), %rsi 1548 movq REGOFF_RAX(%rsp), %rdi 1549 call post_syscall /* post_syscall(rval1, rval2) */ 1550 1551 /* 1552 * set up to take fault on first use of fp 1553 */ 1554 STTS(%rdi) 1555 1556 /* 1557 * XXX - may want a fast path that avoids sys_rtt_common in the 1558 * most common case. 1559 */ 1560 ALTENTRY(_sys_rtt) 1561 CLI(%rax) /* disable interrupts */ 1562 ALTENTRY(_sys_rtt_ints_disabled) 1563 movq %rsp, %rdi /* pass rp to sys_rtt_common */ 1564 call sys_rtt_common /* do common sys_rtt tasks */ 1565 testq %rax, %rax /* returning to userland? */ 1566 jz sr_sup 1567 1568 /* 1569 * Return to user 1570 */ 1571 ASSERT_UPCALL_MASK_IS_SET 1572 cmpw $UCS_SEL, REGOFF_CS(%rsp) /* test for native (64-bit) lwp? */ 1573 je sys_rtt_syscall 1574 1575 /* 1576 * Return to 32-bit userland 1577 */ 1578 ALTENTRY(sys_rtt_syscall32) 1579 USER32_POP 1580 IRET 1581 /*NOTREACHED*/ 1582 1583 ALTENTRY(sys_rtt_syscall) 1584 /* 1585 * Return to 64-bit userland 1586 */ 1587 USER_POP 1588 ALTENTRY(nopop_sys_rtt_syscall) 1589 IRET 1590 /*NOTREACHED*/ 1591 SET_SIZE(nopop_sys_rtt_syscall) 1592 1593 /* 1594 * Return to supervisor 1595 * NOTE: to make the check in trap() that tests if we are executing 1596 * segment register fixup/restore code work properly, sr_sup MUST be 1597 * after _sys_rtt . 1598 */ 1599 ALTENTRY(sr_sup) 1600 /* 1601 * Restore regs before doing iretq to kernel mode 1602 */ 1603 INTR_POP 1604 IRET 1605 .globl _sys_rtt_end 1606 _sys_rtt_end: 1607 /*NOTREACHED*/ 1608 SET_SIZE(sr_sup) 1609 SET_SIZE(_sys_rtt_end) 1610 SET_SIZE(lwp_rtt) 1611 SET_SIZE(lwp_rtt_initial) 1612 SET_SIZE(_sys_rtt_ints_disabled) 1613 SET_SIZE(_sys_rtt) 1614 SET_SIZE(sys_rtt_syscall) 1615 SET_SIZE(sys_rtt_syscall32) 1616 1617 #elif defined(__i386) 1618 1619 ENTRY_NP(lwp_rtt_initial) 1620 movl %gs:CPU_THREAD, %eax 1621 movl T_STACK(%eax), %esp /* switch to the thread stack */ 1622 movl %esp, %ebp 1623 call __dtrace_probe___proc_start 1624 jmp _lwp_rtt 1625 1626 ENTRY_NP(lwp_rtt) 1627 movl %gs:CPU_THREAD, %eax 1628 movl T_STACK(%eax), %esp /* switch to the thread stack */ 1629 movl %esp, %ebp 1630 _lwp_rtt: 1631 call __dtrace_probe___proc_lwp__start 1632 1633 /* 1634 * If agent lwp, clear %fs and %gs. 1635 */ 1636 movl %gs:CPU_LWP, %eax 1637 movl LWP_PROCP(%eax), %edx 1638 1639 cmpl %eax, P_AGENTTP(%edx) 1640 jne 1f 1641 movl $0, REGOFF_FS(%esp) 1642 movl $0, REGOFF_GS(%esp) 1643 1: 1644 call dtrace_systrace_rtt 1645 movl REGOFF_EDX(%esp), %edx 1646 movl REGOFF_EAX(%esp), %eax 1647 pushl %edx 1648 pushl %eax 1649 call post_syscall /* post_syscall(rval1, rval2) */ 1650 addl $8, %esp 1651 1652 /* 1653 * set up to take fault on first use of fp 1654 */ 1655 STTS(%eax) 1656 1657 /* 1658 * XXX - may want a fast path that avoids sys_rtt_common in the 1659 * most common case. 1660 */ 1661 ALTENTRY(_sys_rtt) 1662 CLI(%eax) /* disable interrupts */ 1663 ALTENTRY(_sys_rtt_ints_disabled) 1664 pushl %esp /* pass rp to sys_rtt_common */ 1665 call sys_rtt_common 1666 addl $4, %esp /* pop arg */ 1667 testl %eax, %eax /* test for return to user mode */ 1668 jz sr_sup 1669 1670 /* 1671 * Return to User. 1672 */ 1673 ALTENTRY(sys_rtt_syscall) 1674 INTR_POP_USER 1675 1676 /* 1677 * There can be no instructions between this label and IRET or 1678 * we could end up breaking linux brand support. See label usage 1679 * in lx_brand_int80_callback for an example. 1680 */ 1681 ALTENTRY(nopop_sys_rtt_syscall) 1682 IRET 1683 /*NOTREACHED*/ 1684 SET_SIZE(nopop_sys_rtt_syscall) 1685 1686 ALTENTRY(_sys_rtt_end) 1687 1688 /* 1689 * Return to supervisor 1690 */ 1691 ALTENTRY(sr_sup) 1692 1693 /* 1694 * Restore regs before doing iret to kernel mode 1695 */ 1696 INTR_POP_KERNEL 1697 IRET 1698 /*NOTREACHED*/ 1699 1700 SET_SIZE(sr_sup) 1701 SET_SIZE(_sys_rtt_end) 1702 SET_SIZE(lwp_rtt) 1703 SET_SIZE(lwp_rtt_initial) 1704 SET_SIZE(_sys_rtt_ints_disabled) 1705 SET_SIZE(_sys_rtt) 1706 SET_SIZE(sys_rtt_syscall) 1707 1708 #endif /* __i386 */ 1709 1710 #endif /* __lint */ 1711 1712 #if defined(__lint) 1713 1714 /* 1715 * So why do we have to deal with all this crud in the world of ia32? 1716 * 1717 * Basically there are four classes of ia32 implementations, those that do not 1718 * have a TSC, those that have a marginal TSC that is broken to the extent 1719 * that it is useless, those that have a marginal TSC that is not quite so 1720 * horribly broken and can be used with some care, and those that have a 1721 * reliable TSC. This crud has to be here in order to sift through all the 1722 * variants. 1723 */ 1724 1725 /*ARGSUSED*/ 1726 uint64_t 1727 freq_tsc(uint32_t *pit_counter) 1728 { 1729 return (0); 1730 } 1731 1732 #else /* __lint */ 1733 1734 #if defined(__amd64) 1735 1736 /* 1737 * XX64 quick and dirty port from the i386 version. Since we 1738 * believe the amd64 tsc is more reliable, could this code be 1739 * simpler? 1740 */ 1741 ENTRY_NP(freq_tsc) 1742 pushq %rbp 1743 movq %rsp, %rbp 1744 movq %rdi, %r9 /* save pit_counter */ 1745 pushq %rbx 1746 1747 / We have a TSC, but we have no way in general to know how reliable it is. 1748 / Usually a marginal TSC behaves appropriately unless not enough time 1749 / elapses between reads. A reliable TSC can be read as often and as rapidly 1750 / as desired. The simplistic approach of reading the TSC counter and 1751 / correlating to the PIT counter cannot be naively followed. Instead estimates 1752 / have to be taken to successively refine a guess at the speed of the cpu 1753 / and then the TSC and PIT counter are correlated. In practice very rarely 1754 / is more than one quick loop required for an estimate. Measures have to be 1755 / taken to prevent the PIT counter from wrapping beyond its resolution and for 1756 / measuring the clock rate of very fast processors. 1757 / 1758 / The following constant can be tuned. It should be such that the loop does 1759 / not take too many nor too few PIT counts to execute. If this value is too 1760 / large, then on slow machines the loop will take a long time, or the PIT 1761 / counter may even wrap. If this value is too small, then on fast machines 1762 / the PIT counter may count so few ticks that the resolution of the PIT 1763 / itself causes a bad guess. Because this code is used in machines with 1764 / marginal TSC's and/or IO, if this value is too small on those, it may 1765 / cause the calculated cpu frequency to vary slightly from boot to boot. 1766 / 1767 / In all cases even if this constant is set inappropriately, the algorithm 1768 / will still work and the caller should be able to handle variances in the 1769 / calculation of cpu frequency, but the calculation will be inefficient and 1770 / take a disproportionate amount of time relative to a well selected value. 1771 / As the slowest supported cpu becomes faster, this constant should be 1772 / carefully increased. 1773 1774 movl $0x8000, %ecx 1775 1776 / to make sure the instruction cache has been warmed 1777 clc 1778 1779 jmp freq_tsc_loop 1780 1781 / The following block of code up to and including the latching of the PIT 1782 / counter after freq_tsc_perf_loop is very critical and very carefully 1783 / written, it should only be modified with great care. freq_tsc_loop to 1784 / freq_tsc_perf_loop fits exactly in 16 bytes as do the instructions in 1785 / freq_tsc_perf_loop up to the unlatching of the PIT counter. 1786 1787 .align 32 1788 freq_tsc_loop: 1789 / save the loop count in %ebx 1790 movl %ecx, %ebx 1791 1792 / initialize the PIT counter and start a count down 1793 movb $PIT_LOADMODE, %al 1794 outb $PITCTL_PORT 1795 movb $0xff, %al 1796 outb $PITCTR0_PORT 1797 outb $PITCTR0_PORT 1798 1799 / read the TSC and store the TS in %edi:%esi 1800 rdtsc 1801 movl %eax, %esi 1802 1803 freq_tsc_perf_loop: 1804 movl %edx, %edi 1805 movl %eax, %esi 1806 movl %edx, %edi 1807 loop freq_tsc_perf_loop 1808 1809 / read the TSC and store the LSW in %ecx 1810 rdtsc 1811 movl %eax, %ecx 1812 1813 / latch the PIT counter and status 1814 movb $_CONST(PIT_READBACK|PIT_READBACKC0), %al 1815 outb $PITCTL_PORT 1816 1817 / remember if the icache has been warmed 1818 setc %ah 1819 1820 / read the PIT status 1821 inb $PITCTR0_PORT 1822 shll $8, %eax 1823 1824 / read PIT count 1825 inb $PITCTR0_PORT 1826 shll $8, %eax 1827 inb $PITCTR0_PORT 1828 bswap %eax 1829 1830 / check to see if the PIT count was loaded into the CE 1831 btw $_CONST(PITSTAT_NULLCNT+8), %ax 1832 jc freq_tsc_increase_count 1833 1834 / check to see if PIT counter wrapped 1835 btw $_CONST(PITSTAT_OUTPUT+8), %ax 1836 jnc freq_tsc_pit_did_not_wrap 1837 1838 / halve count 1839 shrl $1, %ebx 1840 movl %ebx, %ecx 1841 1842 / the instruction cache has been warmed 1843 stc 1844 1845 jmp freq_tsc_loop 1846 1847 freq_tsc_increase_count: 1848 shll $1, %ebx 1849 jc freq_tsc_too_fast 1850 1851 movl %ebx, %ecx 1852 1853 / the instruction cache has been warmed 1854 stc 1855 1856 jmp freq_tsc_loop 1857 1858 freq_tsc_pit_did_not_wrap: 1859 roll $16, %eax 1860 1861 cmpw $0x2000, %ax 1862 notw %ax 1863 jb freq_tsc_sufficient_duration 1864 1865 freq_tsc_calculate: 1866 / in mode 0, the PIT loads the count into the CE on the first CLK pulse, 1867 / then on the second CLK pulse the CE is decremented, therefore mode 0 1868 / is really a (count + 1) counter, ugh 1869 xorl %esi, %esi 1870 movw %ax, %si 1871 incl %esi 1872 1873 movl $0xf000, %eax 1874 mull %ebx 1875 1876 / tuck away (target_pit_count * loop_count) 1877 movl %edx, %ecx 1878 movl %eax, %ebx 1879 1880 movl %esi, %eax 1881 movl $0xffffffff, %edx 1882 mull %edx 1883 1884 addl %esi, %eax 1885 adcl $0, %edx 1886 1887 cmpl %ecx, %edx 1888 ja freq_tsc_div_safe 1889 jb freq_tsc_too_fast 1890 1891 cmpl %ebx, %eax 1892 jbe freq_tsc_too_fast 1893 1894 freq_tsc_div_safe: 1895 movl %ecx, %edx 1896 movl %ebx, %eax 1897 1898 movl %esi, %ecx 1899 divl %ecx 1900 1901 movl %eax, %ecx 1902 1903 / the instruction cache has been warmed 1904 stc 1905 1906 jmp freq_tsc_loop 1907 1908 freq_tsc_sufficient_duration: 1909 / test to see if the icache has been warmed 1910 btl $16, %eax 1911 jnc freq_tsc_calculate 1912 1913 / recall mode 0 is a (count + 1) counter 1914 andl $0xffff, %eax 1915 incl %eax 1916 1917 / save the number of PIT counts 1918 movl %eax, (%r9) 1919 1920 / calculate the number of TS's that elapsed 1921 movl %ecx, %eax 1922 subl %esi, %eax 1923 sbbl %edi, %edx 1924 1925 jmp freq_tsc_end 1926 1927 freq_tsc_too_fast: 1928 / return 0 as a 64 bit quantity 1929 xorl %eax, %eax 1930 xorl %edx, %edx 1931 1932 freq_tsc_end: 1933 shlq $32, %rdx 1934 orq %rdx, %rax 1935 1936 popq %rbx 1937 leaveq 1938 ret 1939 SET_SIZE(freq_tsc) 1940 1941 #elif defined(__i386) 1942 1943 ENTRY_NP(freq_tsc) 1944 pushl %ebp 1945 movl %esp, %ebp 1946 pushl %edi 1947 pushl %esi 1948 pushl %ebx 1949 1950 / We have a TSC, but we have no way in general to know how reliable it is. 1951 / Usually a marginal TSC behaves appropriately unless not enough time 1952 / elapses between reads. A reliable TSC can be read as often and as rapidly 1953 / as desired. The simplistic approach of reading the TSC counter and 1954 / correlating to the PIT counter cannot be naively followed. Instead estimates 1955 / have to be taken to successively refine a guess at the speed of the cpu 1956 / and then the TSC and PIT counter are correlated. In practice very rarely 1957 / is more than one quick loop required for an estimate. Measures have to be 1958 / taken to prevent the PIT counter from wrapping beyond its resolution and for 1959 / measuring the clock rate of very fast processors. 1960 / 1961 / The following constant can be tuned. It should be such that the loop does 1962 / not take too many nor too few PIT counts to execute. If this value is too 1963 / large, then on slow machines the loop will take a long time, or the PIT 1964 / counter may even wrap. If this value is too small, then on fast machines 1965 / the PIT counter may count so few ticks that the resolution of the PIT 1966 / itself causes a bad guess. Because this code is used in machines with 1967 / marginal TSC's and/or IO, if this value is too small on those, it may 1968 / cause the calculated cpu frequency to vary slightly from boot to boot. 1969 / 1970 / In all cases even if this constant is set inappropriately, the algorithm 1971 / will still work and the caller should be able to handle variances in the 1972 / calculation of cpu frequency, but the calculation will be inefficient and 1973 / take a disproportionate amount of time relative to a well selected value. 1974 / As the slowest supported cpu becomes faster, this constant should be 1975 / carefully increased. 1976 1977 movl $0x8000, %ecx 1978 1979 / to make sure the instruction cache has been warmed 1980 clc 1981 1982 jmp freq_tsc_loop 1983 1984 / The following block of code up to and including the latching of the PIT 1985 / counter after freq_tsc_perf_loop is very critical and very carefully 1986 / written, it should only be modified with great care. freq_tsc_loop to 1987 / freq_tsc_perf_loop fits exactly in 16 bytes as do the instructions in 1988 / freq_tsc_perf_loop up to the unlatching of the PIT counter. 1989 1990 .align 32 1991 freq_tsc_loop: 1992 / save the loop count in %ebx 1993 movl %ecx, %ebx 1994 1995 / initialize the PIT counter and start a count down 1996 movb $PIT_LOADMODE, %al 1997 outb $PITCTL_PORT 1998 movb $0xff, %al 1999 outb $PITCTR0_PORT 2000 outb $PITCTR0_PORT 2001 2002 / read the TSC and store the TS in %edi:%esi 2003 rdtsc 2004 movl %eax, %esi 2005 2006 freq_tsc_perf_loop: 2007 movl %edx, %edi 2008 movl %eax, %esi 2009 movl %edx, %edi 2010 loop freq_tsc_perf_loop 2011 2012 / read the TSC and store the LSW in %ecx 2013 rdtsc 2014 movl %eax, %ecx 2015 2016 / latch the PIT counter and status 2017 movb $_CONST(PIT_READBACK|PIT_READBACKC0), %al 2018 outb $PITCTL_PORT 2019 2020 / remember if the icache has been warmed 2021 setc %ah 2022 2023 / read the PIT status 2024 inb $PITCTR0_PORT 2025 shll $8, %eax 2026 2027 / read PIT count 2028 inb $PITCTR0_PORT 2029 shll $8, %eax 2030 inb $PITCTR0_PORT 2031 bswap %eax 2032 2033 / check to see if the PIT count was loaded into the CE 2034 btw $_CONST(PITSTAT_NULLCNT+8), %ax 2035 jc freq_tsc_increase_count 2036 2037 / check to see if PIT counter wrapped 2038 btw $_CONST(PITSTAT_OUTPUT+8), %ax 2039 jnc freq_tsc_pit_did_not_wrap 2040 2041 / halve count 2042 shrl $1, %ebx 2043 movl %ebx, %ecx 2044 2045 / the instruction cache has been warmed 2046 stc 2047 2048 jmp freq_tsc_loop 2049 2050 freq_tsc_increase_count: 2051 shll $1, %ebx 2052 jc freq_tsc_too_fast 2053 2054 movl %ebx, %ecx 2055 2056 / the instruction cache has been warmed 2057 stc 2058 2059 jmp freq_tsc_loop 2060 2061 freq_tsc_pit_did_not_wrap: 2062 roll $16, %eax 2063 2064 cmpw $0x2000, %ax 2065 notw %ax 2066 jb freq_tsc_sufficient_duration 2067 2068 freq_tsc_calculate: 2069 / in mode 0, the PIT loads the count into the CE on the first CLK pulse, 2070 / then on the second CLK pulse the CE is decremented, therefore mode 0 2071 / is really a (count + 1) counter, ugh 2072 xorl %esi, %esi 2073 movw %ax, %si 2074 incl %esi 2075 2076 movl $0xf000, %eax 2077 mull %ebx 2078 2079 / tuck away (target_pit_count * loop_count) 2080 movl %edx, %ecx 2081 movl %eax, %ebx 2082 2083 movl %esi, %eax 2084 movl $0xffffffff, %edx 2085 mull %edx 2086 2087 addl %esi, %eax 2088 adcl $0, %edx 2089 2090 cmpl %ecx, %edx 2091 ja freq_tsc_div_safe 2092 jb freq_tsc_too_fast 2093 2094 cmpl %ebx, %eax 2095 jbe freq_tsc_too_fast 2096 2097 freq_tsc_div_safe: 2098 movl %ecx, %edx 2099 movl %ebx, %eax 2100 2101 movl %esi, %ecx 2102 divl %ecx 2103 2104 movl %eax, %ecx 2105 2106 / the instruction cache has been warmed 2107 stc 2108 2109 jmp freq_tsc_loop 2110 2111 freq_tsc_sufficient_duration: 2112 / test to see if the icache has been warmed 2113 btl $16, %eax 2114 jnc freq_tsc_calculate 2115 2116 / recall mode 0 is a (count + 1) counter 2117 andl $0xffff, %eax 2118 incl %eax 2119 2120 / save the number of PIT counts 2121 movl 8(%ebp), %ebx 2122 movl %eax, (%ebx) 2123 2124 / calculate the number of TS's that elapsed 2125 movl %ecx, %eax 2126 subl %esi, %eax 2127 sbbl %edi, %edx 2128 2129 jmp freq_tsc_end 2130 2131 freq_tsc_too_fast: 2132 / return 0 as a 64 bit quantity 2133 xorl %eax, %eax 2134 xorl %edx, %edx 2135 2136 freq_tsc_end: 2137 popl %ebx 2138 popl %esi 2139 popl %edi 2140 popl %ebp 2141 ret 2142 SET_SIZE(freq_tsc) 2143 2144 #endif /* __i386 */ 2145 #endif /* __lint */ 2146 2147 #if !defined(__amd64) 2148 #if defined(__lint) 2149 2150 /* 2151 * We do not have a TSC so we use a block of instructions with well known 2152 * timings. 2153 */ 2154 2155 /*ARGSUSED*/ 2156 uint64_t 2157 freq_notsc(uint32_t *pit_counter) 2158 { 2159 return (0); 2160 } 2161 2162 #else /* __lint */ 2163 ENTRY_NP(freq_notsc) 2164 pushl %ebp 2165 movl %esp, %ebp 2166 pushl %edi 2167 pushl %esi 2168 pushl %ebx 2169 2170 / initial count for the idivl loop 2171 movl $0x1000, %ecx 2172 2173 / load the divisor 2174 movl $1, %ebx 2175 2176 jmp freq_notsc_loop 2177 2178 .align 16 2179 freq_notsc_loop: 2180 / set high 32 bits of dividend to zero 2181 xorl %edx, %edx 2182 2183 / save the loop count in %edi 2184 movl %ecx, %edi 2185 2186 / initialize the PIT counter and start a count down 2187 movb $PIT_LOADMODE, %al 2188 outb $PITCTL_PORT 2189 movb $0xff, %al 2190 outb $PITCTR0_PORT 2191 outb $PITCTR0_PORT 2192 2193 / set low 32 bits of dividend to zero 2194 xorl %eax, %eax 2195 2196 / It is vital that the arguments to idivl be set appropriately because on some 2197 / cpu's this instruction takes more or less clock ticks depending on its 2198 / arguments. 2199 freq_notsc_perf_loop: 2200 idivl %ebx 2201 idivl %ebx 2202 idivl %ebx 2203 idivl %ebx 2204 idivl %ebx 2205 loop freq_notsc_perf_loop 2206 2207 / latch the PIT counter and status 2208 movb $_CONST(PIT_READBACK|PIT_READBACKC0), %al 2209 outb $PITCTL_PORT 2210 2211 / read the PIT status 2212 inb $PITCTR0_PORT 2213 shll $8, %eax 2214 2215 / read PIT count 2216 inb $PITCTR0_PORT 2217 shll $8, %eax 2218 inb $PITCTR0_PORT 2219 bswap %eax 2220 2221 / check to see if the PIT count was loaded into the CE 2222 btw $_CONST(PITSTAT_NULLCNT+8), %ax 2223 jc freq_notsc_increase_count 2224 2225 / check to see if PIT counter wrapped 2226 btw $_CONST(PITSTAT_OUTPUT+8), %ax 2227 jnc freq_notsc_pit_did_not_wrap 2228 2229 / halve count 2230 shrl $1, %edi 2231 movl %edi, %ecx 2232 2233 jmp freq_notsc_loop 2234 2235 freq_notsc_increase_count: 2236 shll $1, %edi 2237 jc freq_notsc_too_fast 2238 2239 movl %edi, %ecx 2240 2241 jmp freq_notsc_loop 2242 2243 freq_notsc_pit_did_not_wrap: 2244 shrl $16, %eax 2245 2246 cmpw $0x2000, %ax 2247 notw %ax 2248 jb freq_notsc_sufficient_duration 2249 2250 freq_notsc_calculate: 2251 / in mode 0, the PIT loads the count into the CE on the first CLK pulse, 2252 / then on the second CLK pulse the CE is decremented, therefore mode 0 2253 / is really a (count + 1) counter, ugh 2254 xorl %esi, %esi 2255 movw %ax, %si 2256 incl %esi 2257 2258 movl %edi, %eax 2259 movl $0xf000, %ecx 2260 mull %ecx 2261 2262 / tuck away (target_pit_count * loop_count) 2263 movl %edx, %edi 2264 movl %eax, %ecx 2265 2266 movl %esi, %eax 2267 movl $0xffffffff, %edx 2268 mull %edx 2269 2270 addl %esi, %eax 2271 adcl $0, %edx 2272 2273 cmpl %edi, %edx 2274 ja freq_notsc_div_safe 2275 jb freq_notsc_too_fast 2276 2277 cmpl %ecx, %eax 2278 jbe freq_notsc_too_fast 2279 2280 freq_notsc_div_safe: 2281 movl %edi, %edx 2282 movl %ecx, %eax 2283 2284 movl %esi, %ecx 2285 divl %ecx 2286 2287 movl %eax, %ecx 2288 2289 jmp freq_notsc_loop 2290 2291 freq_notsc_sufficient_duration: 2292 / recall mode 0 is a (count + 1) counter 2293 incl %eax 2294 2295 / save the number of PIT counts 2296 movl 8(%ebp), %ebx 2297 movl %eax, (%ebx) 2298 2299 / calculate the number of cpu clock ticks that elapsed 2300 cmpl $X86_VENDOR_Cyrix, x86_vendor 2301 jz freq_notsc_notcyrix 2302 2303 / freq_notsc_perf_loop takes 86 clock cycles on Cyrix 6x86 cores 2304 movl $86, %eax 2305 jmp freq_notsc_calculate_tsc 2306 2307 freq_notsc_notcyrix: 2308 / freq_notsc_perf_loop takes 237 clock cycles on Intel Pentiums 2309 movl $237, %eax 2310 2311 freq_notsc_calculate_tsc: 2312 mull %edi 2313 2314 jmp freq_notsc_end 2315 2316 freq_notsc_too_fast: 2317 / return 0 as a 64 bit quantity 2318 xorl %eax, %eax 2319 xorl %edx, %edx 2320 2321 freq_notsc_end: 2322 popl %ebx 2323 popl %esi 2324 popl %edi 2325 popl %ebp 2326 2327 ret 2328 SET_SIZE(freq_notsc) 2329 2330 #endif /* __lint */ 2331 #endif /* !__amd64 */ 2332 2333 #if !defined(__lint) 2334 .data 2335 #if !defined(__amd64) 2336 .align 4 2337 cpu_vendor: 2338 .long 0, 0, 0 /* Vendor ID string returned */ 2339 2340 .globl CyrixInstead 2341 2342 .globl x86_featureset 2343 .globl x86_type 2344 .globl x86_vendor 2345 #endif 2346 2347 #endif /* __lint */
Cache object: b40b2f5ea9858d52873a8dbccc1f62e9
[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]