cpu.c revision 1.127
1 /* $NetBSD: cpu.c,v 1.127 2017/05/31 10:08:24 pgoyette Exp $ */ 2 3 /*- 4 * Copyright (c) 2000-2012 NetBSD Foundation, Inc. 5 * All rights reserved. 6 * 7 * This code is derived from software contributed to The NetBSD Foundation 8 * by Bill Sommerfeld of RedBack Networks Inc, and by Andrew Doran. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 19 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS 20 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 21 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 22 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS 23 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 24 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 25 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 26 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 27 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 28 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 29 * POSSIBILITY OF SUCH DAMAGE. 30 */ 31 32 /* 33 * Copyright (c) 1999 Stefan Grefen 34 * 35 * Redistribution and use in source and binary forms, with or without 36 * modification, are permitted provided that the following conditions 37 * are met: 38 * 1. Redistributions of source code must retain the above copyright 39 * notice, this list of conditions and the following disclaimer. 40 * 2. Redistributions in binary form must reproduce the above copyright 41 * notice, this list of conditions and the following disclaimer in the 42 * documentation and/or other materials provided with the distribution. 43 * 3. All advertising materials mentioning features or use of this software 44 * must display the following acknowledgement: 45 * This product includes software developed by the NetBSD 46 * Foundation, Inc. and its contributors. 47 * 4. Neither the name of The NetBSD Foundation nor the names of its 48 * contributors may be used to endorse or promote products derived 49 * from this software without specific prior written permission. 50 * 51 * THIS SOFTWARE IS PROVIDED BY AUTHOR AND CONTRIBUTORS ``AS IS'' AND ANY 52 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 53 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 54 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR AND CONTRIBUTORS BE LIABLE 55 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 56 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 57 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 58 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 59 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 60 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 61 * SUCH DAMAGE. 62 */ 63 64 #include <sys/cdefs.h> 65 __KERNEL_RCSID(0, "$NetBSD: cpu.c,v 1.127 2017/05/31 10:08:24 pgoyette Exp $"); 66 67 #include "opt_ddb.h" 68 #include "opt_mpbios.h" /* for MPDEBUG */ 69 #include "opt_mtrr.h" 70 #include "opt_multiprocessor.h" 71 72 #include "lapic.h" 73 #include "ioapic.h" 74 75 #include <sys/param.h> 76 #include <sys/proc.h> 77 #include <sys/systm.h> 78 #include <sys/device.h> 79 #include <sys/kmem.h> 80 #include <sys/cpu.h> 81 #include <sys/cpufreq.h> 82 #include <sys/idle.h> 83 #include <sys/atomic.h> 84 #include <sys/reboot.h> 85 86 #include <uvm/uvm.h> 87 88 #include "acpica.h" /* for NACPICA, for mp_verbose */ 89 90 #include <machine/cpufunc.h> 91 #include <machine/cpuvar.h> 92 #include <machine/pmap.h> 93 #include <machine/vmparam.h> 94 #if defined(MULTIPROCESSOR) 95 #include <machine/mpbiosvar.h> 96 #endif 97 #include <machine/mpconfig.h> /* for mp_verbose */ 98 #include <machine/pcb.h> 99 #include <machine/specialreg.h> 100 #include <machine/segments.h> 101 #include <machine/gdt.h> 102 #include <machine/mtrr.h> 103 #include <machine/pio.h> 104 #include <machine/cpu_counter.h> 105 106 #include <x86/fpu.h> 107 108 #ifdef i386 109 #include <machine/tlog.h> 110 #endif 111 112 #if NLAPIC > 0 113 #include <machine/apicvar.h> 114 #include <machine/i82489reg.h> 115 #include <machine/i82489var.h> 116 #endif 117 118 #include <dev/ic/mc146818reg.h> 119 #include <i386/isa/nvram.h> 120 #include <dev/isa/isareg.h> 121 122 #include "tsc.h" 123 124 static int cpu_match(device_t, cfdata_t, void *); 125 static void cpu_attach(device_t, device_t, void *); 126 static void cpu_defer(device_t); 127 static int cpu_rescan(device_t, const char *, const int *); 128 static void cpu_childdetached(device_t, device_t); 129 static bool cpu_stop(device_t); 130 static bool cpu_suspend(device_t, const pmf_qual_t *); 131 static bool cpu_resume(device_t, const pmf_qual_t *); 132 static bool cpu_shutdown(device_t, int); 133 134 struct cpu_softc { 135 device_t sc_dev; /* device tree glue */ 136 struct cpu_info *sc_info; /* pointer to CPU info */ 137 bool sc_wasonline; 138 }; 139 140 #ifdef MULTIPROCESSOR 141 int mp_cpu_start(struct cpu_info *, paddr_t); 142 void mp_cpu_start_cleanup(struct cpu_info *); 143 const struct cpu_functions mp_cpu_funcs = { mp_cpu_start, NULL, 144 mp_cpu_start_cleanup }; 145 #endif 146 147 148 CFATTACH_DECL2_NEW(cpu, sizeof(struct cpu_softc), 149 cpu_match, cpu_attach, NULL, NULL, cpu_rescan, cpu_childdetached); 150 151 /* 152 * Statically-allocated CPU info for the primary CPU (or the only 153 * CPU, on uniprocessors). The CPU info list is initialized to 154 * point at it. 155 */ 156 #ifdef TRAPLOG 157 struct tlog tlog_primary; 158 #endif 159 struct cpu_info cpu_info_primary __aligned(CACHE_LINE_SIZE) = { 160 .ci_dev = 0, 161 .ci_self = &cpu_info_primary, 162 .ci_idepth = -1, 163 .ci_curlwp = &lwp0, 164 .ci_curldt = -1, 165 #ifdef TRAPLOG 166 .ci_tlog_base = &tlog_primary, 167 #endif /* !TRAPLOG */ 168 }; 169 170 struct cpu_info *cpu_info_list = &cpu_info_primary; 171 172 static void cpu_set_tss_gates(struct cpu_info *); 173 174 #ifdef i386 175 static void tss_init(struct i386tss *, void *, void *); 176 #endif 177 178 static void cpu_init_idle_lwp(struct cpu_info *); 179 180 uint32_t cpu_feature[7] __read_mostly; /* X86 CPUID feature bits */ 181 /* [0] basic features cpuid.1:%edx 182 * [1] basic features cpuid.1:%ecx (CPUID2_xxx bits) 183 * [2] extended features cpuid:80000001:%edx 184 * [3] extended features cpuid:80000001:%ecx 185 * [4] VIA padlock features 186 * [5] structured extended features cpuid.7:%ebx 187 * [6] structured extended features cpuid.7:%ecx 188 */ 189 190 extern char x86_64_doubleflt_stack[]; 191 192 #ifdef MULTIPROCESSOR 193 bool x86_mp_online; 194 paddr_t mp_trampoline_paddr = MP_TRAMPOLINE; 195 #endif 196 #if NLAPIC > 0 197 static vaddr_t cmos_data_mapping; 198 #endif 199 struct cpu_info *cpu_starting; 200 201 #ifdef MULTIPROCESSOR 202 void cpu_hatch(void *); 203 static void cpu_boot_secondary(struct cpu_info *ci); 204 static void cpu_start_secondary(struct cpu_info *ci); 205 #endif 206 #if NLAPIC > 0 207 static void cpu_copy_trampoline(void); 208 #endif 209 210 /* 211 * Runs once per boot once multiprocessor goo has been detected and 212 * the local APIC on the boot processor has been mapped. 213 * 214 * Called from lapic_boot_init() (from mpbios_scan()). 215 */ 216 #if NLAPIC > 0 217 void 218 cpu_init_first(void) 219 { 220 221 cpu_info_primary.ci_cpuid = lapic_cpu_number(); 222 cpu_copy_trampoline(); 223 224 cmos_data_mapping = uvm_km_alloc(kernel_map, PAGE_SIZE, 0, UVM_KMF_VAONLY); 225 if (cmos_data_mapping == 0) 226 panic("No KVA for page 0"); 227 pmap_kenter_pa(cmos_data_mapping, 0, VM_PROT_READ|VM_PROT_WRITE, 0); 228 pmap_update(pmap_kernel()); 229 } 230 #endif 231 232 static int 233 cpu_match(device_t parent, cfdata_t match, void *aux) 234 { 235 236 return 1; 237 } 238 239 static void 240 cpu_vm_init(struct cpu_info *ci) 241 { 242 int ncolors = 2, i; 243 244 for (i = CAI_ICACHE; i <= CAI_L2CACHE; i++) { 245 struct x86_cache_info *cai; 246 int tcolors; 247 248 cai = &ci->ci_cinfo[i]; 249 250 tcolors = atop(cai->cai_totalsize); 251 switch(cai->cai_associativity) { 252 case 0xff: 253 tcolors = 1; /* fully associative */ 254 break; 255 case 0: 256 case 1: 257 break; 258 default: 259 tcolors /= cai->cai_associativity; 260 } 261 ncolors = max(ncolors, tcolors); 262 /* 263 * If the desired number of colors is not a power of 264 * two, it won't be good. Find the greatest power of 265 * two which is an even divisor of the number of colors, 266 * to preserve even coloring of pages. 267 */ 268 if (ncolors & (ncolors - 1) ) { 269 int try, picked = 1; 270 for (try = 1; try < ncolors; try *= 2) { 271 if (ncolors % try == 0) picked = try; 272 } 273 if (picked == 1) { 274 panic("desired number of cache colors %d is " 275 " > 1, but not even!", ncolors); 276 } 277 ncolors = picked; 278 } 279 } 280 281 /* 282 * Knowing the size of the largest cache on this CPU, potentially 283 * re-color our pages. 284 */ 285 aprint_debug_dev(ci->ci_dev, "%d page colors\n", ncolors); 286 uvm_page_recolor(ncolors); 287 288 pmap_tlb_cpu_init(ci); 289 #ifndef __HAVE_DIRECT_MAP 290 pmap_vpage_cpu_init(ci); 291 #endif 292 } 293 294 static void 295 cpu_attach(device_t parent, device_t self, void *aux) 296 { 297 struct cpu_softc *sc = device_private(self); 298 struct cpu_attach_args *caa = aux; 299 struct cpu_info *ci; 300 uintptr_t ptr; 301 #if NLAPIC > 0 302 int cpunum = caa->cpu_number; 303 #endif 304 static bool again; 305 306 sc->sc_dev = self; 307 308 if (ncpu == maxcpus) { 309 #ifndef _LP64 310 aprint_error(": too many CPUs, please use NetBSD/amd64\n"); 311 #else 312 aprint_error(": too many CPUs\n"); 313 #endif 314 return; 315 } 316 317 /* 318 * If we're an Application Processor, allocate a cpu_info 319 * structure, otherwise use the primary's. 320 */ 321 if (caa->cpu_role == CPU_ROLE_AP) { 322 if ((boothowto & RB_MD1) != 0) { 323 aprint_error(": multiprocessor boot disabled\n"); 324 if (!pmf_device_register(self, NULL, NULL)) 325 aprint_error_dev(self, 326 "couldn't establish power handler\n"); 327 return; 328 } 329 aprint_naive(": Application Processor\n"); 330 ptr = (uintptr_t)kmem_zalloc(sizeof(*ci) + CACHE_LINE_SIZE - 1, 331 KM_SLEEP); 332 ci = (struct cpu_info *)roundup2(ptr, CACHE_LINE_SIZE); 333 ci->ci_curldt = -1; 334 #ifdef TRAPLOG 335 ci->ci_tlog_base = kmem_zalloc(sizeof(struct tlog), KM_SLEEP); 336 #endif 337 } else { 338 aprint_naive(": %s Processor\n", 339 caa->cpu_role == CPU_ROLE_SP ? "Single" : "Boot"); 340 ci = &cpu_info_primary; 341 #if NLAPIC > 0 342 if (cpunum != lapic_cpu_number()) { 343 /* XXX should be done earlier. */ 344 uint32_t reg; 345 aprint_verbose("\n"); 346 aprint_verbose_dev(self, "running CPU at apic %d" 347 " instead of at expected %d", lapic_cpu_number(), 348 cpunum); 349 reg = lapic_readreg(LAPIC_ID); 350 lapic_writereg(LAPIC_ID, (reg & ~LAPIC_ID_MASK) | 351 (cpunum << LAPIC_ID_SHIFT)); 352 } 353 if (cpunum != lapic_cpu_number()) { 354 aprint_error_dev(self, "unable to reset apic id\n"); 355 } 356 #endif 357 } 358 359 ci->ci_self = ci; 360 sc->sc_info = ci; 361 ci->ci_dev = self; 362 ci->ci_acpiid = caa->cpu_id; 363 ci->ci_cpuid = caa->cpu_number; 364 ci->ci_func = caa->cpu_func; 365 aprint_normal("\n"); 366 367 /* Must be before mi_cpu_attach(). */ 368 cpu_vm_init(ci); 369 370 if (caa->cpu_role == CPU_ROLE_AP) { 371 int error; 372 373 error = mi_cpu_attach(ci); 374 if (error != 0) { 375 aprint_error_dev(self, 376 "mi_cpu_attach failed with %d\n", error); 377 return; 378 } 379 cpu_init_tss(ci); 380 } else { 381 KASSERT(ci->ci_data.cpu_idlelwp != NULL); 382 } 383 384 pmap_reference(pmap_kernel()); 385 ci->ci_pmap = pmap_kernel(); 386 ci->ci_tlbstate = TLBSTATE_STALE; 387 388 /* 389 * Boot processor may not be attached first, but the below 390 * must be done to allow booting other processors. 391 */ 392 if (!again) { 393 atomic_or_32(&ci->ci_flags, CPUF_PRESENT | CPUF_PRIMARY); 394 /* Basic init. */ 395 cpu_intr_init(ci); 396 cpu_get_tsc_freq(ci); 397 cpu_init(ci); 398 cpu_set_tss_gates(ci); 399 pmap_cpu_init_late(ci); 400 #if NLAPIC > 0 401 if (caa->cpu_role != CPU_ROLE_SP) { 402 /* Enable lapic. */ 403 lapic_enable(); 404 lapic_set_lvt(); 405 lapic_calibrate_timer(ci); 406 } 407 #endif 408 /* Make sure DELAY() is initialized. */ 409 DELAY(1); 410 again = true; 411 } 412 413 /* further PCB init done later. */ 414 415 switch (caa->cpu_role) { 416 case CPU_ROLE_SP: 417 atomic_or_32(&ci->ci_flags, CPUF_SP); 418 cpu_identify(ci); 419 x86_errata(); 420 x86_cpu_idle_init(); 421 break; 422 423 case CPU_ROLE_BP: 424 atomic_or_32(&ci->ci_flags, CPUF_BSP); 425 cpu_identify(ci); 426 x86_errata(); 427 x86_cpu_idle_init(); 428 break; 429 430 #ifdef MULTIPROCESSOR 431 case CPU_ROLE_AP: 432 /* 433 * report on an AP 434 */ 435 cpu_intr_init(ci); 436 gdt_alloc_cpu(ci); 437 cpu_set_tss_gates(ci); 438 pmap_cpu_init_late(ci); 439 cpu_start_secondary(ci); 440 if (ci->ci_flags & CPUF_PRESENT) { 441 struct cpu_info *tmp; 442 443 cpu_identify(ci); 444 tmp = cpu_info_list; 445 while (tmp->ci_next) 446 tmp = tmp->ci_next; 447 448 tmp->ci_next = ci; 449 } 450 break; 451 #endif 452 453 default: 454 panic("unknown processor type??\n"); 455 } 456 457 pat_init(ci); 458 459 if (!pmf_device_register1(self, cpu_suspend, cpu_resume, cpu_shutdown)) 460 aprint_error_dev(self, "couldn't establish power handler\n"); 461 462 #ifdef MULTIPROCESSOR 463 if (mp_verbose) { 464 struct lwp *l = ci->ci_data.cpu_idlelwp; 465 struct pcb *pcb = lwp_getpcb(l); 466 467 aprint_verbose_dev(self, 468 "idle lwp at %p, idle sp at %p\n", 469 l, 470 #ifdef i386 471 (void *)pcb->pcb_esp 472 #else 473 (void *)pcb->pcb_rsp 474 #endif 475 ); 476 } 477 #endif 478 479 /* 480 * Postpone the "cpufeaturebus" scan. 481 * It is safe to scan the pseudo-bus 482 * only after all CPUs have attached. 483 */ 484 (void)config_defer(self, cpu_defer); 485 } 486 487 static void 488 cpu_defer(device_t self) 489 { 490 cpu_rescan(self, NULL, NULL); 491 } 492 493 static int 494 cpu_rescan(device_t self, const char *ifattr, const int *locators) 495 { 496 struct cpu_softc *sc = device_private(self); 497 struct cpufeature_attach_args cfaa; 498 struct cpu_info *ci = sc->sc_info; 499 500 memset(&cfaa, 0, sizeof(cfaa)); 501 cfaa.ci = ci; 502 503 if (ifattr_match(ifattr, "cpufeaturebus")) { 504 505 if (ci->ci_frequency == NULL) { 506 cfaa.name = "frequency"; 507 ci->ci_frequency = config_found_ia(self, 508 "cpufeaturebus", &cfaa, NULL); 509 } 510 511 if (ci->ci_padlock == NULL) { 512 cfaa.name = "padlock"; 513 ci->ci_padlock = config_found_ia(self, 514 "cpufeaturebus", &cfaa, NULL); 515 } 516 517 if (ci->ci_temperature == NULL) { 518 cfaa.name = "temperature"; 519 ci->ci_temperature = config_found_ia(self, 520 "cpufeaturebus", &cfaa, NULL); 521 } 522 523 if (ci->ci_vm == NULL) { 524 cfaa.name = "vm"; 525 ci->ci_vm = config_found_ia(self, 526 "cpufeaturebus", &cfaa, NULL); 527 } 528 } 529 530 return 0; 531 } 532 533 static void 534 cpu_childdetached(device_t self, device_t child) 535 { 536 struct cpu_softc *sc = device_private(self); 537 struct cpu_info *ci = sc->sc_info; 538 539 if (ci->ci_frequency == child) 540 ci->ci_frequency = NULL; 541 542 if (ci->ci_padlock == child) 543 ci->ci_padlock = NULL; 544 545 if (ci->ci_temperature == child) 546 ci->ci_temperature = NULL; 547 548 if (ci->ci_vm == child) 549 ci->ci_vm = NULL; 550 } 551 552 /* 553 * Initialize the processor appropriately. 554 */ 555 556 void 557 cpu_init(struct cpu_info *ci) 558 { 559 uint32_t cr4 = 0; 560 561 lcr0(rcr0() | CR0_WP); 562 563 /* 564 * On a P6 or above, enable global TLB caching if the 565 * hardware supports it. 566 */ 567 if (cpu_feature[0] & CPUID_PGE) 568 cr4 |= CR4_PGE; /* enable global TLB caching */ 569 570 /* 571 * If we have FXSAVE/FXRESTOR, use them. 572 */ 573 if (cpu_feature[0] & CPUID_FXSR) { 574 cr4 |= CR4_OSFXSR; 575 576 /* 577 * If we have SSE/SSE2, enable XMM exceptions. 578 */ 579 if (cpu_feature[0] & (CPUID_SSE|CPUID_SSE2)) 580 cr4 |= CR4_OSXMMEXCPT; 581 } 582 583 /* If xsave is supported, enable it */ 584 if (cpu_feature[1] & CPUID2_XSAVE) 585 cr4 |= CR4_OSXSAVE; 586 587 /* If SMEP is supported, enable it */ 588 if (cpu_feature[5] & CPUID_SEF_SMEP) 589 cr4 |= CR4_SMEP; 590 591 if (cr4) { 592 cr4 |= rcr4(); 593 lcr4(cr4); 594 } 595 596 /* If xsave is enabled, enable all fpu features */ 597 if (cr4 & CR4_OSXSAVE) 598 wrxcr(0, x86_xsave_features & XCR0_FPU); 599 600 #ifdef MTRR 601 /* 602 * On a P6 or above, initialize MTRR's if the hardware supports them. 603 */ 604 if (cpu_feature[0] & CPUID_MTRR) { 605 if ((ci->ci_flags & CPUF_AP) == 0) 606 i686_mtrr_init_first(); 607 mtrr_init_cpu(ci); 608 } 609 610 #ifdef i386 611 if (strcmp((char *)(ci->ci_vendor), "AuthenticAMD") == 0) { 612 /* 613 * Must be a K6-2 Step >= 7 or a K6-III. 614 */ 615 if (CPUID_TO_FAMILY(ci->ci_signature) == 5) { 616 if (CPUID_TO_MODEL(ci->ci_signature) > 8 || 617 (CPUID_TO_MODEL(ci->ci_signature) == 8 && 618 CPUID_TO_STEPPING(ci->ci_signature) >= 7)) { 619 mtrr_funcs = &k6_mtrr_funcs; 620 k6_mtrr_init_first(); 621 mtrr_init_cpu(ci); 622 } 623 } 624 } 625 #endif /* i386 */ 626 #endif /* MTRR */ 627 628 if (ci != &cpu_info_primary) { 629 /* Synchronize TSC again, and check for drift. */ 630 wbinvd(); 631 atomic_or_32(&ci->ci_flags, CPUF_RUNNING); 632 tsc_sync_ap(ci); 633 } else { 634 atomic_or_32(&ci->ci_flags, CPUF_RUNNING); 635 } 636 } 637 638 #ifdef MULTIPROCESSOR 639 void 640 cpu_boot_secondary_processors(void) 641 { 642 struct cpu_info *ci; 643 kcpuset_t *cpus; 644 u_long i; 645 646 /* Now that we know the number of CPUs, patch the text segment. */ 647 x86_patch(false); 648 649 kcpuset_create(&cpus, true); 650 kcpuset_set(cpus, cpu_index(curcpu())); 651 for (i = 0; i < maxcpus; i++) { 652 ci = cpu_lookup(i); 653 if (ci == NULL) 654 continue; 655 if (ci->ci_data.cpu_idlelwp == NULL) 656 continue; 657 if ((ci->ci_flags & CPUF_PRESENT) == 0) 658 continue; 659 if (ci->ci_flags & (CPUF_BSP|CPUF_SP|CPUF_PRIMARY)) 660 continue; 661 cpu_boot_secondary(ci); 662 kcpuset_set(cpus, cpu_index(ci)); 663 } 664 while (!kcpuset_match(cpus, kcpuset_running)) 665 ; 666 kcpuset_destroy(cpus); 667 668 x86_mp_online = true; 669 670 /* Now that we know about the TSC, attach the timecounter. */ 671 tsc_tc_init(); 672 673 /* Enable zeroing of pages in the idle loop if we have SSE2. */ 674 vm_page_zero_enable = ((cpu_feature[0] & CPUID_SSE2) != 0); 675 } 676 #endif 677 678 static void 679 cpu_init_idle_lwp(struct cpu_info *ci) 680 { 681 struct lwp *l = ci->ci_data.cpu_idlelwp; 682 struct pcb *pcb = lwp_getpcb(l); 683 684 pcb->pcb_cr0 = rcr0(); 685 } 686 687 void 688 cpu_init_idle_lwps(void) 689 { 690 struct cpu_info *ci; 691 u_long i; 692 693 for (i = 0; i < maxcpus; i++) { 694 ci = cpu_lookup(i); 695 if (ci == NULL) 696 continue; 697 if (ci->ci_data.cpu_idlelwp == NULL) 698 continue; 699 if ((ci->ci_flags & CPUF_PRESENT) == 0) 700 continue; 701 cpu_init_idle_lwp(ci); 702 } 703 } 704 705 #ifdef MULTIPROCESSOR 706 void 707 cpu_start_secondary(struct cpu_info *ci) 708 { 709 extern paddr_t mp_pdirpa; 710 u_long psl; 711 int i; 712 713 mp_pdirpa = pmap_init_tmp_pgtbl(mp_trampoline_paddr); 714 atomic_or_32(&ci->ci_flags, CPUF_AP); 715 ci->ci_curlwp = ci->ci_data.cpu_idlelwp; 716 if (CPU_STARTUP(ci, mp_trampoline_paddr) != 0) { 717 return; 718 } 719 720 /* 721 * Wait for it to become ready. Setting cpu_starting opens the 722 * initial gate and allows the AP to start soft initialization. 723 */ 724 KASSERT(cpu_starting == NULL); 725 cpu_starting = ci; 726 for (i = 100000; (!(ci->ci_flags & CPUF_PRESENT)) && i > 0; i--) { 727 #ifdef MPDEBUG 728 extern int cpu_trace[3]; 729 static int otrace[3]; 730 if (memcmp(otrace, cpu_trace, sizeof(otrace)) != 0) { 731 aprint_debug_dev(ci->ci_dev, "trace %02x %02x %02x\n", 732 cpu_trace[0], cpu_trace[1], cpu_trace[2]); 733 memcpy(otrace, cpu_trace, sizeof(otrace)); 734 } 735 #endif 736 i8254_delay(10); 737 } 738 739 if ((ci->ci_flags & CPUF_PRESENT) == 0) { 740 aprint_error_dev(ci->ci_dev, "failed to become ready\n"); 741 #if defined(MPDEBUG) && defined(DDB) 742 printf("dropping into debugger; continue from here to resume boot\n"); 743 Debugger(); 744 #endif 745 } else { 746 /* 747 * Synchronize time stamp counters. Invalidate cache and do 748 * twice to try and minimize possible cache effects. Disable 749 * interrupts to try and rule out any external interference. 750 */ 751 psl = x86_read_psl(); 752 x86_disable_intr(); 753 wbinvd(); 754 tsc_sync_bp(ci); 755 x86_write_psl(psl); 756 } 757 758 CPU_START_CLEANUP(ci); 759 cpu_starting = NULL; 760 } 761 762 void 763 cpu_boot_secondary(struct cpu_info *ci) 764 { 765 int64_t drift; 766 u_long psl; 767 int i; 768 769 atomic_or_32(&ci->ci_flags, CPUF_GO); 770 for (i = 100000; (!(ci->ci_flags & CPUF_RUNNING)) && i > 0; i--) { 771 i8254_delay(10); 772 } 773 if ((ci->ci_flags & CPUF_RUNNING) == 0) { 774 aprint_error_dev(ci->ci_dev, "failed to start\n"); 775 #if defined(MPDEBUG) && defined(DDB) 776 printf("dropping into debugger; continue from here to resume boot\n"); 777 Debugger(); 778 #endif 779 } else { 780 /* Synchronize TSC again, check for drift. */ 781 drift = ci->ci_data.cpu_cc_skew; 782 psl = x86_read_psl(); 783 x86_disable_intr(); 784 wbinvd(); 785 tsc_sync_bp(ci); 786 x86_write_psl(psl); 787 drift -= ci->ci_data.cpu_cc_skew; 788 aprint_debug_dev(ci->ci_dev, "TSC skew=%lld drift=%lld\n", 789 (long long)ci->ci_data.cpu_cc_skew, (long long)drift); 790 tsc_sync_drift(drift); 791 } 792 } 793 794 /* 795 * The CPU ends up here when it's ready to run. 796 * This is called from code in mptramp.s; at this point, we are running 797 * in the idle pcb/idle stack of the new CPU. When this function returns, 798 * this processor will enter the idle loop and start looking for work. 799 */ 800 void 801 cpu_hatch(void *v) 802 { 803 struct cpu_info *ci = (struct cpu_info *)v; 804 struct pcb *pcb; 805 int s; 806 807 cpu_init_msrs(ci, true); 808 cpu_probe(ci); 809 810 ci->ci_data.cpu_cc_freq = cpu_info_primary.ci_data.cpu_cc_freq; 811 /* cpu_get_tsc_freq(ci); */ 812 813 KDASSERT((ci->ci_flags & CPUF_PRESENT) == 0); 814 815 /* 816 * Synchronize time stamp counters. Invalidate cache and do twice 817 * to try and minimize possible cache effects. Note that interrupts 818 * are off at this point. 819 */ 820 wbinvd(); 821 atomic_or_32(&ci->ci_flags, CPUF_PRESENT); 822 tsc_sync_ap(ci); 823 824 /* 825 * Wait to be brought online. Use 'monitor/mwait' if available, 826 * in order to make the TSC drift as much as possible. so that 827 * we can detect it later. If not available, try 'pause'. 828 * We'd like to use 'hlt', but we have interrupts off. 829 */ 830 while ((ci->ci_flags & CPUF_GO) == 0) { 831 if ((cpu_feature[1] & CPUID2_MONITOR) != 0) { 832 x86_monitor(&ci->ci_flags, 0, 0); 833 if ((ci->ci_flags & CPUF_GO) != 0) { 834 continue; 835 } 836 x86_mwait(0, 0); 837 } else { 838 for (i = 100; i != 0; i--) { 839 x86_pause(); 840 } 841 } 842 } 843 844 /* Because the text may have been patched in x86_patch(). */ 845 wbinvd(); 846 x86_flush(); 847 tlbflushg(); 848 849 KASSERT((ci->ci_flags & CPUF_RUNNING) == 0); 850 851 #ifdef PAE 852 pd_entry_t * l3_pd = ci->ci_pae_l3_pdir; 853 for (i = 0 ; i < PDP_SIZE; i++) { 854 l3_pd[i] = pmap_kernel()->pm_pdirpa[i] | PG_V; 855 } 856 lcr3(ci->ci_pae_l3_pdirpa); 857 #else 858 lcr3(pmap_pdirpa(pmap_kernel(), 0)); 859 #endif 860 861 pcb = lwp_getpcb(curlwp); 862 pcb->pcb_cr3 = rcr3(); 863 pcb = lwp_getpcb(ci->ci_data.cpu_idlelwp); 864 lcr0(pcb->pcb_cr0); 865 866 cpu_init_idt(); 867 gdt_init_cpu(ci); 868 #if NLAPIC > 0 869 lapic_enable(); 870 lapic_set_lvt(); 871 lapic_initclocks(); 872 #endif 873 874 fpuinit(ci); 875 lldt(GSYSSEL(GLDT_SEL, SEL_KPL)); 876 ltr(ci->ci_tss_sel); 877 878 cpu_init(ci); 879 cpu_get_tsc_freq(ci); 880 881 s = splhigh(); 882 lapic_write_tpri(0); 883 x86_enable_intr(); 884 splx(s); 885 x86_errata(); 886 887 aprint_debug_dev(ci->ci_dev, "running\n"); 888 889 idle_loop(NULL); 890 KASSERT(false); 891 } 892 #endif 893 894 #if defined(DDB) 895 896 #include <ddb/db_output.h> 897 #include <machine/db_machdep.h> 898 899 /* 900 * Dump CPU information from ddb. 901 */ 902 void 903 cpu_debug_dump(void) 904 { 905 struct cpu_info *ci; 906 CPU_INFO_ITERATOR cii; 907 908 db_printf("addr dev id flags ipis curlwp fpcurlwp\n"); 909 for (CPU_INFO_FOREACH(cii, ci)) { 910 db_printf("%p %s %ld %x %x %10p %10p\n", 911 ci, 912 ci->ci_dev == NULL ? "BOOT" : device_xname(ci->ci_dev), 913 (long)ci->ci_cpuid, 914 ci->ci_flags, ci->ci_ipis, 915 ci->ci_curlwp, 916 ci->ci_fpcurlwp); 917 } 918 } 919 #endif 920 921 #if NLAPIC > 0 922 static void 923 cpu_copy_trampoline(void) 924 { 925 /* 926 * Copy boot code. 927 */ 928 extern u_char cpu_spinup_trampoline[]; 929 extern u_char cpu_spinup_trampoline_end[]; 930 931 vaddr_t mp_trampoline_vaddr; 932 933 mp_trampoline_vaddr = uvm_km_alloc(kernel_map, PAGE_SIZE, 0, 934 UVM_KMF_VAONLY); 935 936 pmap_kenter_pa(mp_trampoline_vaddr, mp_trampoline_paddr, 937 VM_PROT_READ | VM_PROT_WRITE, 0); 938 pmap_update(pmap_kernel()); 939 memcpy((void *)mp_trampoline_vaddr, 940 cpu_spinup_trampoline, 941 cpu_spinup_trampoline_end - cpu_spinup_trampoline); 942 943 pmap_kremove(mp_trampoline_vaddr, PAGE_SIZE); 944 pmap_update(pmap_kernel()); 945 uvm_km_free(kernel_map, mp_trampoline_vaddr, PAGE_SIZE, UVM_KMF_VAONLY); 946 } 947 #endif 948 949 #ifdef i386 950 static void 951 tss_init(struct i386tss *tss, void *stack, void *func) 952 { 953 KASSERT(curcpu()->ci_pmap == pmap_kernel()); 954 955 memset(tss, 0, sizeof *tss); 956 tss->tss_esp0 = tss->tss_esp = (int)((char *)stack + USPACE - 16); 957 tss->tss_ss0 = GSEL(GDATA_SEL, SEL_KPL); 958 tss->__tss_cs = GSEL(GCODE_SEL, SEL_KPL); 959 tss->tss_fs = GSEL(GCPU_SEL, SEL_KPL); 960 tss->tss_gs = tss->__tss_es = tss->__tss_ds = 961 tss->__tss_ss = GSEL(GDATA_SEL, SEL_KPL); 962 /* %cr3 contains the value associated to pmap_kernel */ 963 tss->tss_cr3 = rcr3(); 964 tss->tss_esp = (int)((char *)stack + USPACE - 16); 965 tss->tss_ldt = GSEL(GLDT_SEL, SEL_KPL); 966 tss->__tss_eflags = PSL_MBO | PSL_NT; /* XXX not needed? */ 967 tss->__tss_eip = (int)func; 968 } 969 970 /* XXX */ 971 #define IDTVEC(name) __CONCAT(X, name) 972 typedef void (vector)(void); 973 extern vector IDTVEC(tss_trap08); 974 #if defined(DDB) && defined(MULTIPROCESSOR) 975 extern vector Xintrddbipi, Xx2apic_intrddbipi; 976 extern int ddb_vec; 977 #endif 978 979 static void 980 cpu_set_tss_gates(struct cpu_info *ci) 981 { 982 struct segment_descriptor sd; 983 984 ci->ci_doubleflt_stack = (char *)uvm_km_alloc(kernel_map, USPACE, 0, 985 UVM_KMF_WIRED); 986 tss_init(&ci->ci_doubleflt_tss, ci->ci_doubleflt_stack, 987 IDTVEC(tss_trap08)); 988 setsegment(&sd, &ci->ci_doubleflt_tss, sizeof(struct i386tss) - 1, 989 SDT_SYS386TSS, SEL_KPL, 0, 0); 990 ci->ci_gdt[GTRAPTSS_SEL].sd = sd; 991 setgate(&idt[8], NULL, 0, SDT_SYSTASKGT, SEL_KPL, 992 GSEL(GTRAPTSS_SEL, SEL_KPL)); 993 994 #if defined(DDB) && defined(MULTIPROCESSOR) 995 /* 996 * Set up separate handler for the DDB IPI, so that it doesn't 997 * stomp on a possibly corrupted stack. 998 * 999 * XXX overwriting the gate set in db_machine_init. 1000 * Should rearrange the code so that it's set only once. 1001 */ 1002 ci->ci_ddbipi_stack = (char *)uvm_km_alloc(kernel_map, USPACE, 0, 1003 UVM_KMF_WIRED); 1004 tss_init(&ci->ci_ddbipi_tss, ci->ci_ddbipi_stack, 1005 x2apic_mode ? Xx2apic_intrddbipi : Xintrddbipi); 1006 1007 setsegment(&sd, &ci->ci_ddbipi_tss, sizeof(struct i386tss) - 1, 1008 SDT_SYS386TSS, SEL_KPL, 0, 0); 1009 ci->ci_gdt[GIPITSS_SEL].sd = sd; 1010 1011 setgate(&idt[ddb_vec], NULL, 0, SDT_SYSTASKGT, SEL_KPL, 1012 GSEL(GIPITSS_SEL, SEL_KPL)); 1013 #endif 1014 } 1015 #else 1016 static void 1017 cpu_set_tss_gates(struct cpu_info *ci) 1018 { 1019 1020 } 1021 #endif /* i386 */ 1022 1023 #ifdef MULTIPROCESSOR 1024 int 1025 mp_cpu_start(struct cpu_info *ci, paddr_t target) 1026 { 1027 unsigned short dwordptr[2]; 1028 int error; 1029 1030 /* 1031 * Bootstrap code must be addressable in real mode 1032 * and it must be page aligned. 1033 */ 1034 KASSERT(target < 0x10000 && target % PAGE_SIZE == 0); 1035 1036 /* 1037 * "The BSP must initialize CMOS shutdown code to 0Ah ..." 1038 */ 1039 1040 outb(IO_RTC, NVRAM_RESET); 1041 outb(IO_RTC+1, NVRAM_RESET_JUMP); 1042 1043 /* 1044 * "and the warm reset vector (DWORD based at 40:67) to point 1045 * to the AP startup code ..." 1046 */ 1047 1048 dwordptr[0] = 0; 1049 dwordptr[1] = target >> 4; 1050 1051 #if NLAPIC > 0 1052 memcpy((uint8_t *)cmos_data_mapping + 0x467, dwordptr, 4); 1053 #endif 1054 1055 if ((cpu_feature[0] & CPUID_APIC) == 0) { 1056 aprint_error("mp_cpu_start: CPU does not have APIC\n"); 1057 return ENODEV; 1058 } 1059 1060 /* 1061 * ... prior to executing the following sequence:". We'll also add in 1062 * local cache flush, in case the BIOS has left the AP with its cache 1063 * disabled. It may not be able to cope with MP coherency. 1064 */ 1065 wbinvd(); 1066 1067 if (ci->ci_flags & CPUF_AP) { 1068 error = x86_ipi_init(ci->ci_cpuid); 1069 if (error != 0) { 1070 aprint_error_dev(ci->ci_dev, "%s: IPI not taken (1)\n", 1071 __func__); 1072 return error; 1073 } 1074 i8254_delay(10000); 1075 1076 error = x86_ipi_startup(ci->ci_cpuid, target / PAGE_SIZE); 1077 if (error != 0) { 1078 aprint_error_dev(ci->ci_dev, "%s: IPI not taken (2)\n", 1079 __func__); 1080 return error; 1081 } 1082 i8254_delay(200); 1083 1084 error = x86_ipi_startup(ci->ci_cpuid, target / PAGE_SIZE); 1085 if (error != 0) { 1086 aprint_error_dev(ci->ci_dev, "%s: IPI not taken (3)\n", 1087 __func__); 1088 return error; 1089 } 1090 i8254_delay(200); 1091 } 1092 1093 return 0; 1094 } 1095 1096 void 1097 mp_cpu_start_cleanup(struct cpu_info *ci) 1098 { 1099 /* 1100 * Ensure the NVRAM reset byte contains something vaguely sane. 1101 */ 1102 1103 outb(IO_RTC, NVRAM_RESET); 1104 outb(IO_RTC+1, NVRAM_RESET_RST); 1105 } 1106 #endif 1107 1108 #ifdef __x86_64__ 1109 typedef void (vector)(void); 1110 extern vector Xsyscall, Xsyscall32; 1111 #endif 1112 1113 void 1114 cpu_init_msrs(struct cpu_info *ci, bool full) 1115 { 1116 #ifdef __x86_64__ 1117 wrmsr(MSR_STAR, 1118 ((uint64_t)GSEL(GCODE_SEL, SEL_KPL) << 32) | 1119 ((uint64_t)LSEL(LSYSRETBASE_SEL, SEL_UPL) << 48)); 1120 wrmsr(MSR_LSTAR, (uint64_t)Xsyscall); 1121 wrmsr(MSR_CSTAR, (uint64_t)Xsyscall32); 1122 wrmsr(MSR_SFMASK, PSL_NT|PSL_T|PSL_I|PSL_C); 1123 1124 if (full) { 1125 wrmsr(MSR_FSBASE, 0); 1126 wrmsr(MSR_GSBASE, (uint64_t)ci); 1127 wrmsr(MSR_KERNELGSBASE, 0); 1128 } 1129 #endif /* __x86_64__ */ 1130 1131 if (cpu_feature[2] & CPUID_NOX) 1132 wrmsr(MSR_EFER, rdmsr(MSR_EFER) | EFER_NXE); 1133 } 1134 1135 void 1136 cpu_offline_md(void) 1137 { 1138 int s; 1139 1140 s = splhigh(); 1141 fpusave_cpu(true); 1142 splx(s); 1143 } 1144 1145 /* XXX joerg restructure and restart CPUs individually */ 1146 static bool 1147 cpu_stop(device_t dv) 1148 { 1149 struct cpu_softc *sc = device_private(dv); 1150 struct cpu_info *ci = sc->sc_info; 1151 int err; 1152 1153 KASSERT((ci->ci_flags & CPUF_PRESENT) != 0); 1154 1155 if ((ci->ci_flags & CPUF_PRIMARY) != 0) 1156 return true; 1157 1158 if (ci->ci_data.cpu_idlelwp == NULL) 1159 return true; 1160 1161 sc->sc_wasonline = !(ci->ci_schedstate.spc_flags & SPCF_OFFLINE); 1162 1163 if (sc->sc_wasonline) { 1164 mutex_enter(&cpu_lock); 1165 err = cpu_setstate(ci, false); 1166 mutex_exit(&cpu_lock); 1167 1168 if (err != 0) 1169 return false; 1170 } 1171 1172 return true; 1173 } 1174 1175 static bool 1176 cpu_suspend(device_t dv, const pmf_qual_t *qual) 1177 { 1178 struct cpu_softc *sc = device_private(dv); 1179 struct cpu_info *ci = sc->sc_info; 1180 1181 if ((ci->ci_flags & CPUF_PRESENT) == 0) 1182 return true; 1183 else { 1184 cpufreq_suspend(ci); 1185 } 1186 1187 return cpu_stop(dv); 1188 } 1189 1190 static bool 1191 cpu_resume(device_t dv, const pmf_qual_t *qual) 1192 { 1193 struct cpu_softc *sc = device_private(dv); 1194 struct cpu_info *ci = sc->sc_info; 1195 int err = 0; 1196 1197 if ((ci->ci_flags & CPUF_PRESENT) == 0) 1198 return true; 1199 1200 if ((ci->ci_flags & CPUF_PRIMARY) != 0) 1201 goto out; 1202 1203 if (ci->ci_data.cpu_idlelwp == NULL) 1204 goto out; 1205 1206 if (sc->sc_wasonline) { 1207 mutex_enter(&cpu_lock); 1208 err = cpu_setstate(ci, true); 1209 mutex_exit(&cpu_lock); 1210 } 1211 1212 out: 1213 if (err != 0) 1214 return false; 1215 1216 cpufreq_resume(ci); 1217 1218 return true; 1219 } 1220 1221 static bool 1222 cpu_shutdown(device_t dv, int how) 1223 { 1224 struct cpu_softc *sc = device_private(dv); 1225 struct cpu_info *ci = sc->sc_info; 1226 1227 if ((ci->ci_flags & CPUF_BSP) != 0) 1228 return false; 1229 1230 if ((ci->ci_flags & CPUF_PRESENT) == 0) 1231 return true; 1232 1233 return cpu_stop(dv); 1234 } 1235 1236 void 1237 cpu_get_tsc_freq(struct cpu_info *ci) 1238 { 1239 uint64_t last_tsc; 1240 1241 if (cpu_hascounter()) { 1242 last_tsc = cpu_counter_serializing(); 1243 i8254_delay(100000); 1244 ci->ci_data.cpu_cc_freq = 1245 (cpu_counter_serializing() - last_tsc) * 10; 1246 } 1247 } 1248 1249 void 1250 x86_cpu_idle_mwait(void) 1251 { 1252 struct cpu_info *ci = curcpu(); 1253 1254 KASSERT(ci->ci_ilevel == IPL_NONE); 1255 1256 x86_monitor(&ci->ci_want_resched, 0, 0); 1257 if (__predict_false(ci->ci_want_resched)) { 1258 return; 1259 } 1260 x86_mwait(0, 0); 1261 } 1262 1263 void 1264 x86_cpu_idle_halt(void) 1265 { 1266 struct cpu_info *ci = curcpu(); 1267 1268 KASSERT(ci->ci_ilevel == IPL_NONE); 1269 1270 x86_disable_intr(); 1271 if (!__predict_false(ci->ci_want_resched)) { 1272 x86_stihlt(); 1273 } else { 1274 x86_enable_intr(); 1275 } 1276 } 1277 1278 /* 1279 * Loads pmap for the current CPU. 1280 */ 1281 void 1282 cpu_load_pmap(struct pmap *pmap, struct pmap *oldpmap) 1283 { 1284 #ifdef PAE 1285 struct cpu_info *ci = curcpu(); 1286 bool interrupts_enabled; 1287 pd_entry_t *l3_pd = ci->ci_pae_l3_pdir; 1288 int i; 1289 1290 /* 1291 * disable interrupts to block TLB shootdowns, which can reload cr3. 1292 * while this doesn't block NMIs, it's probably ok as NMIs unlikely 1293 * reload cr3. 1294 */ 1295 interrupts_enabled = (x86_read_flags() & PSL_I) != 0; 1296 if (interrupts_enabled) 1297 x86_disable_intr(); 1298 1299 for (i = 0 ; i < PDP_SIZE; i++) { 1300 l3_pd[i] = pmap->pm_pdirpa[i] | PG_V; 1301 } 1302 1303 if (interrupts_enabled) 1304 x86_enable_intr(); 1305 tlbflush(); 1306 #else /* PAE */ 1307 lcr3(pmap_pdirpa(pmap, 0)); 1308 #endif /* PAE */ 1309 } 1310 1311 /* 1312 * Notify all other cpus to halt. 1313 */ 1314 1315 void 1316 cpu_broadcast_halt(void) 1317 { 1318 x86_broadcast_ipi(X86_IPI_HALT); 1319 } 1320 1321 /* 1322 * Send a dummy ipi to a cpu to force it to run splraise()/spllower() 1323 */ 1324 1325 void 1326 cpu_kick(struct cpu_info *ci) 1327 { 1328 x86_send_ipi(ci, 0); 1329 } 1330
Indexes created Tue Sep 23 14:10:03 GMT 2025