1 /* $NetBSD: cpu.c,v 1.94 2013/10/23 20:18:50 drochner Exp $ */ 2 /* NetBSD: cpu.c,v 1.18 2004/02/20 17:35:01 yamt Exp */ 3 4 /*- 5 * Copyright (c) 2000 The NetBSD Foundation, Inc. 6 * Copyright (c) 2002, 2006, 2007 YAMAMOTO Takashi, 7 * All rights reserved. 8 * 9 * This code is derived from software contributed to The NetBSD Foundation 10 * by RedBack Networks Inc. 11 * 12 * Author: Bill Sommerfeld 13 * 14 * Redistribution and use in source and binary forms, with or without 15 * modification, are permitted provided that the following conditions 16 * are met: 17 * 1. Redistributions of source code must retain the above copyright 18 * notice, this list of conditions and the following disclaimer. 19 * 2. Redistributions in binary form must reproduce the above copyright 20 * notice, this list of conditions and the following disclaimer in the 21 * documentation and/or other materials provided with the distribution. 22 * 23 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS 24 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 25 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 26 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS 27 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 28 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 29 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 30 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 31 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 32 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 33 * POSSIBILITY OF SUCH DAMAGE. 34 */ 35 36 /* 37 * Copyright (c) 1999 Stefan Grefen 38 * 39 * Redistribution and use in source and binary forms, with or without 40 * modification, are permitted provided that the following conditions 41 * are met: 42 * 1. Redistributions of source code must retain the above copyright 43 * notice, this list of conditions and the following disclaimer. 44 * 2. Redistributions in binary form must reproduce the above copyright 45 * notice, this list of conditions and the following disclaimer in the 46 * documentation and/or other materials provided with the distribution. 47 * 3. All advertising materials mentioning features or use of this software 48 * must display the following acknowledgement: 49 * This product includes software developed by the NetBSD 50 * Foundation, Inc. and its contributors. 51 * 4. Neither the name of The NetBSD Foundation nor the names of its 52 * contributors may be used to endorse or promote products derived 53 * from this software without specific prior written permission. 54 * 55 * THIS SOFTWARE IS PROVIDED BY AUTHOR AND CONTRIBUTORS ``AS IS'' AND ANY 56 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 57 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 58 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR AND CONTRIBUTORS BE LIABLE 59 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 60 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 61 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 62 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 63 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 64 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 65 * SUCH DAMAGE. 66 */ 67 68 #include <sys/cdefs.h> 69 __KERNEL_RCSID(0, "$NetBSD: cpu.c,v 1.94 2013/10/23 20:18:50 drochner Exp $"); 70 71 #include "opt_ddb.h" 72 #include "opt_multiprocessor.h" 73 #include "opt_mpbios.h" /* for MPDEBUG */ 74 #include "opt_mtrr.h" 75 #include "opt_xen.h" 76 77 #include "lapic.h" 78 #include "ioapic.h" 79 80 #include <sys/param.h> 81 #include <sys/proc.h> 82 #include <sys/systm.h> 83 #include <sys/device.h> 84 #include <sys/kmem.h> 85 #include <sys/cpu.h> 86 #include <sys/cpufreq.h> 87 #include <sys/atomic.h> 88 #include <sys/reboot.h> 89 #include <sys/idle.h> 90 91 #include <uvm/uvm.h> 92 93 #include <machine/cpufunc.h> 94 #include <machine/cpuvar.h> 95 #include <machine/pmap.h> 96 #include <machine/vmparam.h> 97 #include <machine/mpbiosvar.h> 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> 104 105 #ifdef i386 106 #include <machine/npx.h> 107 #else 108 #include <machine/fpu.h> 109 #endif 110 111 #include <xen/xen.h> 112 #include <xen/xen-public/vcpu.h> 113 #include <xen/vcpuvar.h> 114 115 #if NLAPIC > 0 116 #include <machine/apicvar.h> 117 #include <machine/i82489reg.h> 118 #include <machine/i82489var.h> 119 #endif 120 121 #include <dev/ic/mc146818reg.h> 122 #include <dev/isa/isareg.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 int vcpu_match(device_t, cfdata_t, void *); 130 static void vcpu_attach(device_t, device_t, void *); 131 static void cpu_attach_common(device_t, device_t, void *); 132 void cpu_offline_md(void); 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 int mp_cpu_start(struct cpu_info *, vaddr_t); 141 void mp_cpu_start_cleanup(struct cpu_info *); 142 const struct cpu_functions mp_cpu_funcs = { mp_cpu_start, NULL, 143 mp_cpu_start_cleanup }; 144 145 CFATTACH_DECL2_NEW(cpu, sizeof(struct cpu_softc), 146 cpu_match, cpu_attach, NULL, NULL, cpu_rescan, cpu_childdetached); 147 148 CFATTACH_DECL_NEW(vcpu, sizeof(struct cpu_softc), 149 vcpu_match, vcpu_attach, NULL, NULL); 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 #include <machine/tlog.h> 158 struct tlog tlog_primary; 159 #endif 160 struct cpu_info cpu_info_primary __aligned(CACHE_LINE_SIZE) = { 161 .ci_dev = 0, 162 .ci_self = &cpu_info_primary, 163 .ci_idepth = -1, 164 .ci_curlwp = &lwp0, 165 .ci_curldt = -1, 166 #ifdef TRAPLOG 167 .ci_tlog = &tlog_primary, 168 #endif 169 170 }; 171 struct cpu_info phycpu_info_primary __aligned(CACHE_LINE_SIZE) = { 172 .ci_dev = 0, 173 .ci_self = &phycpu_info_primary, 174 }; 175 176 struct cpu_info *cpu_info_list = &cpu_info_primary; 177 struct cpu_info *phycpu_info_list = &phycpu_info_primary; 178 179 uint32_t cpu_feature[5]; /* X86 CPUID feature bits 180 * [0] basic features %edx 181 * [1] basic features %ecx 182 * [2] extended features %edx 183 * [3] extended features %ecx 184 * [4] VIA padlock features 185 */ 186 187 bool x86_mp_online; 188 paddr_t mp_trampoline_paddr = MP_TRAMPOLINE; 189 190 #if defined(MULTIPROCESSOR) 191 void cpu_hatch(void *); 192 static void cpu_boot_secondary(struct cpu_info *ci); 193 static void cpu_start_secondary(struct cpu_info *ci); 194 #endif /* MULTIPROCESSOR */ 195 196 static int 197 cpu_match(device_t parent, cfdata_t match, void *aux) 198 { 199 200 return 1; 201 } 202 203 static void 204 cpu_attach(device_t parent, device_t self, void *aux) 205 { 206 struct cpu_softc *sc = device_private(self); 207 struct cpu_attach_args *caa = aux; 208 struct cpu_info *ci; 209 uintptr_t ptr; 210 static int nphycpu = 0; 211 212 sc->sc_dev = self; 213 214 /* 215 * If we're an Application Processor, allocate a cpu_info 216 * If we're the first attached CPU use the primary cpu_info, 217 * otherwise allocate a new one 218 */ 219 aprint_naive("\n"); 220 aprint_normal("\n"); 221 if (nphycpu > 0) { 222 struct cpu_info *tmp; 223 ptr = (uintptr_t)kmem_zalloc(sizeof(*ci) + CACHE_LINE_SIZE - 1, 224 KM_SLEEP); 225 ci = (struct cpu_info *)roundup2(ptr, CACHE_LINE_SIZE); 226 ci->ci_curldt = -1; 227 228 tmp = phycpu_info_list; 229 while (tmp->ci_next) 230 tmp = tmp->ci_next; 231 232 tmp->ci_next = ci; 233 } else { 234 ci = &phycpu_info_primary; 235 } 236 237 ci->ci_self = ci; 238 sc->sc_info = ci; 239 240 ci->ci_dev = self; 241 ci->ci_acpiid = caa->cpu_id; 242 ci->ci_cpuid = caa->cpu_number; 243 ci->ci_vcpu = NULL; 244 ci->ci_index = nphycpu++; 245 246 if (!pmf_device_register(self, NULL, NULL)) 247 aprint_error_dev(self, "couldn't establish power handler\n"); 248 249 (void)config_defer(self, cpu_defer); 250 } 251 252 static void 253 cpu_defer(device_t self) 254 { 255 cpu_rescan(self, NULL, NULL); 256 } 257 258 static int 259 cpu_rescan(device_t self, const char *ifattr, const int *locators) 260 { 261 struct cpu_softc *sc = device_private(self); 262 struct cpufeature_attach_args cfaa; 263 struct cpu_info *ci = sc->sc_info; 264 265 memset(&cfaa, 0, sizeof(cfaa)); 266 cfaa.ci = ci; 267 268 if (ifattr_match(ifattr, "cpufeaturebus")) { 269 270 if (ci->ci_frequency == NULL) { 271 cfaa.name = "frequency"; 272 ci->ci_frequency = config_found_ia(self, 273 "cpufeaturebus", &cfaa, NULL); 274 } 275 } 276 277 return 0; 278 } 279 280 static void 281 cpu_childdetached(device_t self, device_t child) 282 { 283 struct cpu_softc *sc = device_private(self); 284 struct cpu_info *ci = sc->sc_info; 285 286 if (ci->ci_frequency == child) 287 ci->ci_frequency = NULL; 288 } 289 290 static int 291 vcpu_match(device_t parent, cfdata_t match, void *aux) 292 { 293 struct vcpu_attach_args *vcaa = aux; 294 struct vcpu_runstate_info vcr; 295 int error; 296 297 if (strcmp(vcaa->vcaa_name, match->cf_name) == 0) { 298 error = HYPERVISOR_vcpu_op(VCPUOP_get_runstate_info, 299 vcaa->vcaa_caa.cpu_number, 300 &vcr); 301 switch (error) { 302 case 0: 303 return 1; 304 case -ENOENT: 305 return 0; 306 default: 307 panic("Unknown hypervisor error %d returned on vcpu runstate probe\n", error); 308 } 309 } 310 311 return 0; 312 } 313 314 static void 315 vcpu_attach(device_t parent, device_t self, void *aux) 316 { 317 struct vcpu_attach_args *vcaa = aux; 318 319 KASSERT(vcaa->vcaa_caa.cpu_func == NULL); 320 vcaa->vcaa_caa.cpu_func = &mp_cpu_funcs; 321 cpu_attach_common(parent, self, &vcaa->vcaa_caa); 322 323 if (!pmf_device_register(self, NULL, NULL)) 324 aprint_error_dev(self, "couldn't establish power handler\n"); 325 } 326 327 static int 328 vcpu_is_up(struct cpu_info *ci) 329 { 330 KASSERT(ci != NULL); 331 return HYPERVISOR_vcpu_op(VCPUOP_is_up, ci->ci_cpuid, NULL); 332 } 333 334 static void 335 cpu_vm_init(struct cpu_info *ci) 336 { 337 int ncolors = 2, i; 338 339 for (i = CAI_ICACHE; i <= CAI_L2CACHE; i++) { 340 struct x86_cache_info *cai; 341 int tcolors; 342 343 cai = &ci->ci_cinfo[i]; 344 345 tcolors = atop(cai->cai_totalsize); 346 switch(cai->cai_associativity) { 347 case 0xff: 348 tcolors = 1; /* fully associative */ 349 break; 350 case 0: 351 case 1: 352 break; 353 default: 354 tcolors /= cai->cai_associativity; 355 } 356 ncolors = max(ncolors, tcolors); 357 } 358 359 /* 360 * Knowing the size of the largest cache on this CPU, potentially 361 * re-color our pages. 362 */ 363 aprint_debug_dev(ci->ci_dev, "%d page colors\n", ncolors); 364 uvm_page_recolor(ncolors); 365 pmap_tlb_cpu_init(ci); 366 } 367 368 static void 369 cpu_attach_common(device_t parent, device_t self, void *aux) 370 { 371 struct cpu_softc *sc = device_private(self); 372 struct cpu_attach_args *caa = aux; 373 struct cpu_info *ci; 374 uintptr_t ptr; 375 int cpunum = caa->cpu_number; 376 static bool again = false; 377 378 sc->sc_dev = self; 379 380 /* 381 * If we're an Application Processor, allocate a cpu_info 382 * structure, otherwise use the primary's. 383 */ 384 if (caa->cpu_role == CPU_ROLE_AP) { 385 aprint_naive(": Application Processor\n"); 386 ptr = (uintptr_t)kmem_alloc(sizeof(*ci) + CACHE_LINE_SIZE - 1, 387 KM_SLEEP); 388 ci = (struct cpu_info *)roundup2(ptr, CACHE_LINE_SIZE); 389 memset(ci, 0, sizeof(*ci)); 390 #ifdef TRAPLOG 391 ci->ci_tlog_base = kmem_zalloc(sizeof(struct tlog), KM_SLEEP); 392 #endif 393 } else { 394 aprint_naive(": %s Processor\n", 395 caa->cpu_role == CPU_ROLE_SP ? "Single" : "Boot"); 396 ci = &cpu_info_primary; 397 } 398 399 ci->ci_self = ci; 400 sc->sc_info = ci; 401 ci->ci_dev = self; 402 ci->ci_cpuid = cpunum; 403 404 KASSERT(HYPERVISOR_shared_info != NULL); 405 KASSERT(cpunum < XEN_LEGACY_MAX_VCPUS); 406 ci->ci_vcpu = &HYPERVISOR_shared_info->vcpu_info[cpunum]; 407 408 KASSERT(ci->ci_func == 0); 409 ci->ci_func = caa->cpu_func; 410 411 /* Must be called before mi_cpu_attach(). */ 412 cpu_vm_init(ci); 413 414 if (caa->cpu_role == CPU_ROLE_AP) { 415 int error; 416 417 error = mi_cpu_attach(ci); 418 419 KASSERT(ci->ci_data.cpu_idlelwp != NULL); 420 if (error != 0) { 421 aprint_normal("\n"); 422 aprint_error_dev(self, 423 "mi_cpu_attach failed with %d\n", error); 424 return; 425 } 426 427 } else { 428 KASSERT(ci->ci_data.cpu_idlelwp != NULL); 429 } 430 431 KASSERT(ci->ci_cpuid == ci->ci_index); 432 pmap_reference(pmap_kernel()); 433 ci->ci_pmap = pmap_kernel(); 434 ci->ci_tlbstate = TLBSTATE_STALE; 435 436 /* 437 * Boot processor may not be attached first, but the below 438 * must be done to allow booting other processors. 439 */ 440 if (!again) { 441 atomic_or_32(&ci->ci_flags, CPUF_PRESENT | CPUF_PRIMARY); 442 /* Basic init. */ 443 cpu_intr_init(ci); 444 cpu_get_tsc_freq(ci); 445 cpu_init(ci); 446 pmap_cpu_init_late(ci); 447 448 /* Every processor needs to init it's own ipi h/w (similar to lapic) */ 449 xen_ipi_init(); 450 451 /* Make sure DELAY() is initialized. */ 452 DELAY(1); 453 again = true; 454 } 455 456 /* further PCB init done later. */ 457 458 switch (caa->cpu_role) { 459 case CPU_ROLE_SP: 460 atomic_or_32(&ci->ci_flags, CPUF_SP); 461 cpu_identify(ci); 462 x86_cpu_idle_init(); 463 464 break; 465 466 case CPU_ROLE_BP: 467 atomic_or_32(&ci->ci_flags, CPUF_BSP); 468 cpu_identify(ci); 469 x86_cpu_idle_init(); 470 471 break; 472 473 case CPU_ROLE_AP: 474 atomic_or_32(&ci->ci_flags, CPUF_AP); 475 476 /* 477 * report on an AP 478 */ 479 480 #if defined(MULTIPROCESSOR) 481 /* interrupt handler stack */ 482 cpu_intr_init(ci); 483 484 /* Setup per-cpu memory for gdt */ 485 gdt_alloc_cpu(ci); 486 487 pmap_cpu_init_late(ci); 488 cpu_start_secondary(ci); 489 490 if (ci->ci_flags & CPUF_PRESENT) { 491 struct cpu_info *tmp; 492 493 cpu_identify(ci); 494 tmp = cpu_info_list; 495 while (tmp->ci_next) 496 tmp = tmp->ci_next; 497 498 tmp->ci_next = ci; 499 } 500 #else 501 aprint_error(": not started\n"); 502 #endif 503 break; 504 505 default: 506 aprint_normal("\n"); 507 panic("unknown processor type??\n"); 508 } 509 510 #ifdef MPVERBOSE 511 if (mp_verbose) { 512 struct lwp *l = ci->ci_data.cpu_idlelwp; 513 struct pcb *pcb = lwp_getpcb(l); 514 515 aprint_verbose_dev(self, 516 "idle lwp at %p, idle sp at 0x%p\n", 517 l, 518 #ifdef i386 519 (void *)pcb->pcb_esp 520 #else /* i386 */ 521 (void *)pcb->pcb_rsp 522 #endif /* i386 */ 523 ); 524 525 } 526 #endif /* MPVERBOSE */ 527 } 528 529 /* 530 * Initialize the processor appropriately. 531 */ 532 533 void 534 cpu_init(struct cpu_info *ci) 535 { 536 537 /* 538 * If we have FXSAVE/FXRESTOR, use them. 539 */ 540 if (cpu_feature[0] & CPUID_FXSR) { 541 lcr4(rcr4() | CR4_OSFXSR); 542 543 /* 544 * If we have SSE/SSE2, enable XMM exceptions. 545 */ 546 if (cpu_feature[0] & (CPUID_SSE|CPUID_SSE2)) 547 lcr4(rcr4() | CR4_OSXMMEXCPT); 548 } 549 550 #ifdef __x86_64__ 551 /* No user PGD mapped for this CPU yet */ 552 ci->ci_xen_current_user_pgd = 0; 553 #endif 554 #if defined(__x86_64__) || defined(PAE) 555 mutex_init(&ci->ci_kpm_mtx, MUTEX_DEFAULT, IPL_VM); 556 #endif 557 558 atomic_or_32(&ci->ci_flags, CPUF_RUNNING); 559 } 560 561 562 #ifdef MULTIPROCESSOR 563 564 void 565 cpu_boot_secondary_processors(void) 566 { 567 struct cpu_info *ci; 568 u_long i; 569 for (i = 0; i < maxcpus; i++) { 570 ci = cpu_lookup(i); 571 if (ci == NULL) 572 continue; 573 if (ci->ci_data.cpu_idlelwp == NULL) 574 continue; 575 if ((ci->ci_flags & CPUF_PRESENT) == 0) 576 continue; 577 if (ci->ci_flags & (CPUF_BSP|CPUF_SP|CPUF_PRIMARY)) 578 continue; 579 cpu_boot_secondary(ci); 580 } 581 582 x86_mp_online = true; 583 } 584 585 static void 586 cpu_init_idle_lwp(struct cpu_info *ci) 587 { 588 struct lwp *l = ci->ci_data.cpu_idlelwp; 589 struct pcb *pcb = lwp_getpcb(l); 590 591 pcb->pcb_cr0 = rcr0(); 592 } 593 594 void 595 cpu_init_idle_lwps(void) 596 { 597 struct cpu_info *ci; 598 u_long i; 599 600 for (i = 0; i < maxcpus; i++) { 601 ci = cpu_lookup(i); 602 if (ci == NULL) 603 continue; 604 if (ci->ci_data.cpu_idlelwp == NULL) 605 continue; 606 if ((ci->ci_flags & CPUF_PRESENT) == 0) 607 continue; 608 cpu_init_idle_lwp(ci); 609 } 610 } 611 612 static void 613 cpu_start_secondary(struct cpu_info *ci) 614 { 615 int i; 616 617 aprint_debug_dev(ci->ci_dev, "starting\n"); 618 619 ci->ci_curlwp = ci->ci_data.cpu_idlelwp; 620 621 if (CPU_STARTUP(ci, (vaddr_t) cpu_hatch) != 0) { 622 return; 623 } 624 625 /* 626 * wait for it to become ready 627 */ 628 for (i = 100000; (!(ci->ci_flags & CPUF_PRESENT)) && i > 0; i--) { 629 delay(10); 630 } 631 if ((ci->ci_flags & CPUF_PRESENT) == 0) { 632 aprint_error_dev(ci->ci_dev, "failed to become ready\n"); 633 #if defined(MPDEBUG) && defined(DDB) 634 printf("dropping into debugger; continue from here to resume boot\n"); 635 Debugger(); 636 #endif 637 } 638 639 CPU_START_CLEANUP(ci); 640 } 641 642 void 643 cpu_boot_secondary(struct cpu_info *ci) 644 { 645 int i; 646 atomic_or_32(&ci->ci_flags, CPUF_GO); 647 for (i = 100000; (!(ci->ci_flags & CPUF_RUNNING)) && i > 0; i--) { 648 delay(10); 649 } 650 if ((ci->ci_flags & CPUF_RUNNING) == 0) { 651 aprint_error_dev(ci->ci_dev, "CPU failed to start\n"); 652 #if defined(MPDEBUG) && defined(DDB) 653 printf("dropping into debugger; continue from here to resume boot\n"); 654 Debugger(); 655 #endif 656 } 657 } 658 659 /* 660 * APs end up here immediately after initialisation and VCPUOP_up in 661 * mp_cpu_start(). 662 * At this point, we are running in the idle pcb/idle stack of the new 663 * CPU. This function jumps to the idle loop and starts looking for 664 * work. 665 */ 666 extern void x86_64_tls_switch(struct lwp *); 667 void 668 cpu_hatch(void *v) 669 { 670 struct cpu_info *ci = (struct cpu_info *)v; 671 struct pcb *pcb; 672 int s, i; 673 674 /* Setup TLS and kernel GS/FS */ 675 cpu_init_msrs(ci, true); 676 cpu_init_idt(); 677 gdt_init_cpu(ci); 678 679 cpu_probe(ci); 680 681 atomic_or_32(&ci->ci_flags, CPUF_PRESENT); 682 683 while ((ci->ci_flags & CPUF_GO) == 0) { 684 /* Don't use delay, boot CPU may be patching the text. */ 685 for (i = 10000; i != 0; i--) 686 x86_pause(); 687 } 688 689 /* Because the text may have been patched in x86_patch(). */ 690 x86_flush(); 691 tlbflushg(); 692 693 KASSERT((ci->ci_flags & CPUF_RUNNING) == 0); 694 695 pcb = lwp_getpcb(curlwp); 696 pcb->pcb_cr3 = pmap_pdirpa(pmap_kernel(), 0); 697 pcb = lwp_getpcb(ci->ci_data.cpu_idlelwp); 698 699 xen_ipi_init(); 700 701 xen_initclocks(); 702 703 #ifdef __x86_64__ 704 fpuinit(ci); 705 #endif 706 707 lldt(GSEL(GLDT_SEL, SEL_KPL)); 708 709 cpu_init(ci); 710 cpu_get_tsc_freq(ci); 711 712 s = splhigh(); 713 x86_enable_intr(); 714 splx(s); 715 716 aprint_debug_dev(ci->ci_dev, "running\n"); 717 718 cpu_switchto(NULL, ci->ci_data.cpu_idlelwp, true); 719 720 idle_loop(NULL); 721 KASSERT(false); 722 } 723 724 #if defined(DDB) 725 726 #include <ddb/db_output.h> 727 #include <machine/db_machdep.h> 728 729 /* 730 * Dump CPU information from ddb. 731 */ 732 void 733 cpu_debug_dump(void) 734 { 735 struct cpu_info *ci; 736 CPU_INFO_ITERATOR cii; 737 738 db_printf("addr dev id flags ipis curlwp\n"); 739 for (CPU_INFO_FOREACH(cii, ci)) { 740 db_printf("%p %s %ld %x %x %10p\n", 741 ci, 742 ci->ci_dev == NULL ? "BOOT" : device_xname(ci->ci_dev), 743 (long)ci->ci_cpuid, 744 ci->ci_flags, ci->ci_ipis, 745 ci->ci_curlwp); 746 } 747 } 748 #endif /* DDB */ 749 750 #endif /* MULTIPROCESSOR */ 751 752 extern void hypervisor_callback(void); 753 extern void failsafe_callback(void); 754 #ifdef __x86_64__ 755 typedef void (vector)(void); 756 extern vector Xsyscall, Xsyscall32; 757 #endif 758 759 /* 760 * Setup the "trampoline". On Xen, we setup nearly all cpu context 761 * outside a trampoline, so we prototype and call targetip like so: 762 * void targetip(struct cpu_info *); 763 */ 764 765 static void 766 gdt_prepframes(paddr_t *frames, vaddr_t base, uint32_t entries) 767 { 768 int i; 769 for (i = 0; i < roundup(entries, PAGE_SIZE) >> PAGE_SHIFT; i++) { 770 771 frames[i] = ((paddr_t) xpmap_ptetomach( 772 (pt_entry_t *) (base + (i << PAGE_SHIFT)))) 773 >> PAGE_SHIFT; 774 775 /* Mark Read-only */ 776 pmap_pte_clearbits(kvtopte(base + (i << PAGE_SHIFT)), 777 PG_RW); 778 } 779 } 780 781 #ifdef __x86_64__ 782 extern char *ldtstore; 783 784 static void 785 xen_init_amd64_vcpuctxt(struct cpu_info *ci, 786 struct vcpu_guest_context *initctx, 787 void targetrip(struct cpu_info *)) 788 { 789 /* page frames to point at GDT */ 790 extern int gdt_size; 791 paddr_t frames[16]; 792 psize_t gdt_ents; 793 794 struct lwp *l; 795 struct pcb *pcb; 796 797 volatile struct vcpu_info *vci; 798 799 KASSERT(ci != NULL); 800 KASSERT(ci != &cpu_info_primary); 801 KASSERT(initctx != NULL); 802 KASSERT(targetrip != NULL); 803 804 memset(initctx, 0, sizeof *initctx); 805 806 gdt_ents = roundup(gdt_size, PAGE_SIZE) >> PAGE_SHIFT; 807 KASSERT(gdt_ents <= 16); 808 809 gdt_prepframes(frames, (vaddr_t) ci->ci_gdt, gdt_ents); 810 811 /* Initialise the vcpu context: We use idle_loop()'s pcb context. */ 812 813 l = ci->ci_data.cpu_idlelwp; 814 815 KASSERT(l != NULL); 816 pcb = lwp_getpcb(l); 817 KASSERT(pcb != NULL); 818 819 /* resume with interrupts off */ 820 vci = ci->ci_vcpu; 821 vci->evtchn_upcall_mask = 1; 822 xen_mb(); 823 824 /* resume in kernel-mode */ 825 initctx->flags = VGCF_in_kernel | VGCF_online; 826 827 /* Stack and entry points: 828 * We arrange for the stack frame for cpu_hatch() to 829 * appear as a callee frame of lwp_trampoline(). Being a 830 * leaf frame prevents trampling on any of the MD stack setup 831 * that x86/vm_machdep.c:cpu_lwp_fork() does for idle_loop() 832 */ 833 834 initctx->user_regs.rdi = (uint64_t) ci; /* targetrip(ci); */ 835 initctx->user_regs.rip = (vaddr_t) targetrip; 836 837 initctx->user_regs.cs = GSEL(GCODE_SEL, SEL_KPL); 838 839 initctx->user_regs.rflags = pcb->pcb_flags; 840 initctx->user_regs.rsp = pcb->pcb_rsp; 841 842 /* Data segments */ 843 initctx->user_regs.ss = GSEL(GDATA_SEL, SEL_KPL); 844 initctx->user_regs.es = GSEL(GDATA_SEL, SEL_KPL); 845 initctx->user_regs.ds = GSEL(GDATA_SEL, SEL_KPL); 846 847 /* GDT */ 848 memcpy(initctx->gdt_frames, frames, sizeof frames); 849 initctx->gdt_ents = gdt_ents; 850 851 /* LDT */ 852 initctx->ldt_base = (unsigned long) ldtstore; 853 initctx->ldt_ents = LDT_SIZE >> 3; 854 855 /* Kernel context state */ 856 initctx->kernel_ss = GSEL(GDATA_SEL, SEL_KPL); 857 initctx->kernel_sp = pcb->pcb_rsp0; 858 initctx->ctrlreg[0] = pcb->pcb_cr0; 859 initctx->ctrlreg[1] = 0; /* "resuming" from kernel - no User cr3. */ 860 initctx->ctrlreg[2] = (vaddr_t) targetrip; 861 /* 862 * Use pmap_kernel() L4 PD directly, until we setup the 863 * per-cpu L4 PD in pmap_cpu_init_late() 864 */ 865 initctx->ctrlreg[3] = xen_pfn_to_cr3(x86_btop(xpmap_ptom(ci->ci_kpm_pdirpa))); 866 initctx->ctrlreg[4] = CR4_PAE | CR4_OSFXSR | CR4_OSXMMEXCPT; 867 868 869 /* Xen callbacks */ 870 initctx->event_callback_eip = (unsigned long) hypervisor_callback; 871 initctx->failsafe_callback_eip = (unsigned long) failsafe_callback; 872 initctx->syscall_callback_eip = (unsigned long) Xsyscall; 873 874 return; 875 } 876 #else /* i386 */ 877 extern union descriptor *ldt; 878 extern void Xsyscall(void); 879 880 static void 881 xen_init_i386_vcpuctxt(struct cpu_info *ci, 882 struct vcpu_guest_context *initctx, 883 void targeteip(struct cpu_info *)) 884 { 885 /* page frames to point at GDT */ 886 extern int gdt_size; 887 paddr_t frames[16]; 888 psize_t gdt_ents; 889 890 struct lwp *l; 891 struct pcb *pcb; 892 893 volatile struct vcpu_info *vci; 894 895 KASSERT(ci != NULL); 896 KASSERT(ci != &cpu_info_primary); 897 KASSERT(initctx != NULL); 898 KASSERT(targeteip != NULL); 899 900 memset(initctx, 0, sizeof *initctx); 901 902 gdt_ents = roundup(gdt_size, PAGE_SIZE) >> PAGE_SHIFT; 903 KASSERT(gdt_ents <= 16); 904 905 gdt_prepframes(frames, (vaddr_t) ci->ci_gdt, gdt_ents); 906 907 /* 908 * Initialise the vcpu context: 909 * We use this cpu's idle_loop() pcb context. 910 */ 911 912 l = ci->ci_data.cpu_idlelwp; 913 914 KASSERT(l != NULL); 915 pcb = lwp_getpcb(l); 916 KASSERT(pcb != NULL); 917 918 /* resume with interrupts off */ 919 vci = ci->ci_vcpu; 920 vci->evtchn_upcall_mask = 1; 921 xen_mb(); 922 923 /* resume in kernel-mode */ 924 initctx->flags = VGCF_in_kernel | VGCF_online; 925 926 /* Stack frame setup for cpu_hatch(): 927 * We arrange for the stack frame for cpu_hatch() to 928 * appear as a callee frame of lwp_trampoline(). Being a 929 * leaf frame prevents trampling on any of the MD stack setup 930 * that x86/vm_machdep.c:cpu_lwp_fork() does for idle_loop() 931 */ 932 933 initctx->user_regs.esp = pcb->pcb_esp - 4; /* Leave word for 934 arg1 */ 935 { /* targeteip(ci); */ 936 uint32_t *arg = (uint32_t *) initctx->user_regs.esp; 937 arg[1] = (uint32_t) ci; /* arg1 */ 938 939 } 940 941 initctx->user_regs.eip = (vaddr_t) targeteip; 942 initctx->user_regs.cs = GSEL(GCODE_SEL, SEL_KPL); 943 initctx->user_regs.eflags |= pcb->pcb_iopl; 944 945 /* Data segments */ 946 initctx->user_regs.ss = GSEL(GDATA_SEL, SEL_KPL); 947 initctx->user_regs.es = GSEL(GDATA_SEL, SEL_KPL); 948 initctx->user_regs.ds = GSEL(GDATA_SEL, SEL_KPL); 949 initctx->user_regs.fs = GSEL(GDATA_SEL, SEL_KPL); 950 951 /* GDT */ 952 memcpy(initctx->gdt_frames, frames, sizeof frames); 953 initctx->gdt_ents = gdt_ents; 954 955 /* LDT */ 956 initctx->ldt_base = (unsigned long) ldt; 957 initctx->ldt_ents = NLDT; 958 959 /* Kernel context state */ 960 initctx->kernel_ss = GSEL(GDATA_SEL, SEL_KPL); 961 initctx->kernel_sp = pcb->pcb_esp0; 962 initctx->ctrlreg[0] = pcb->pcb_cr0; 963 initctx->ctrlreg[1] = 0; /* "resuming" from kernel - no User cr3. */ 964 initctx->ctrlreg[2] = (vaddr_t) targeteip; 965 #ifdef PAE 966 initctx->ctrlreg[3] = xen_pfn_to_cr3(x86_btop(xpmap_ptom(ci->ci_pae_l3_pdirpa))); 967 #else /* PAE */ 968 initctx->ctrlreg[3] = xen_pfn_to_cr3(x86_btop(xpmap_ptom(pcb->pcb_cr3))); 969 #endif /* PAE */ 970 initctx->ctrlreg[4] = /* CR4_PAE | */CR4_OSFXSR | CR4_OSXMMEXCPT; 971 972 973 /* Xen callbacks */ 974 initctx->event_callback_eip = (unsigned long) hypervisor_callback; 975 initctx->event_callback_cs = GSEL(GCODE_SEL, SEL_KPL); 976 initctx->failsafe_callback_eip = (unsigned long) failsafe_callback; 977 initctx->failsafe_callback_cs = GSEL(GCODE_SEL, SEL_KPL); 978 979 return; 980 } 981 #endif /* __x86_64__ */ 982 983 int 984 mp_cpu_start(struct cpu_info *ci, vaddr_t target) 985 { 986 987 int hyperror; 988 struct vcpu_guest_context vcpuctx; 989 990 KASSERT(ci != NULL); 991 KASSERT(ci != &cpu_info_primary); 992 KASSERT(ci->ci_flags & CPUF_AP); 993 994 #ifdef __x86_64__ 995 xen_init_amd64_vcpuctxt(ci, &vcpuctx, (void (*)(struct cpu_info *))target); 996 #else /* i386 */ 997 xen_init_i386_vcpuctxt(ci, &vcpuctx, (void (*)(struct cpu_info *))target); 998 #endif /* __x86_64__ */ 999 1000 /* Initialise the given vcpu to execute cpu_hatch(ci); */ 1001 if ((hyperror = HYPERVISOR_vcpu_op(VCPUOP_initialise, ci->ci_cpuid, &vcpuctx))) { 1002 aprint_error(": context initialisation failed. errno = %d\n", hyperror); 1003 return hyperror; 1004 } 1005 1006 /* Start it up */ 1007 1008 /* First bring it down */ 1009 if ((hyperror = HYPERVISOR_vcpu_op(VCPUOP_down, ci->ci_cpuid, NULL))) { 1010 aprint_error(": VCPUOP_down hypervisor command failed. errno = %d\n", hyperror); 1011 return hyperror; 1012 } 1013 1014 if ((hyperror = HYPERVISOR_vcpu_op(VCPUOP_up, ci->ci_cpuid, NULL))) { 1015 aprint_error(": VCPUOP_up hypervisor command failed. errno = %d\n", hyperror); 1016 return hyperror; 1017 } 1018 1019 if (!vcpu_is_up(ci)) { 1020 aprint_error(": did not come up\n"); 1021 return -1; 1022 } 1023 1024 return 0; 1025 } 1026 1027 void 1028 mp_cpu_start_cleanup(struct cpu_info *ci) 1029 { 1030 if (vcpu_is_up(ci)) { 1031 aprint_debug_dev(ci->ci_dev, "is started.\n"); 1032 } 1033 else { 1034 aprint_error_dev(ci->ci_dev, "did not start up.\n"); 1035 } 1036 1037 } 1038 1039 void 1040 cpu_init_msrs(struct cpu_info *ci, bool full) 1041 { 1042 #ifdef __x86_64__ 1043 if (full) { 1044 HYPERVISOR_set_segment_base (SEGBASE_FS, 0); 1045 HYPERVISOR_set_segment_base (SEGBASE_GS_KERNEL, (uint64_t) ci); 1046 HYPERVISOR_set_segment_base (SEGBASE_GS_USER, 0); 1047 } 1048 #endif /* __x86_64__ */ 1049 1050 if (cpu_feature[2] & CPUID_NOX) 1051 wrmsr(MSR_EFER, rdmsr(MSR_EFER) | EFER_NXE); 1052 1053 } 1054 1055 void 1056 cpu_get_tsc_freq(struct cpu_info *ci) 1057 { 1058 uint32_t vcpu_tversion; 1059 const volatile vcpu_time_info_t *tinfo = &ci->ci_vcpu->time; 1060 1061 vcpu_tversion = tinfo->version; 1062 while (tinfo->version == vcpu_tversion); /* Wait for a time update. XXX: timeout ? */ 1063 1064 uint64_t freq = 1000000000ULL << 32; 1065 freq = freq / (uint64_t)tinfo->tsc_to_system_mul; 1066 if ( tinfo->tsc_shift < 0 ) 1067 freq = freq << -tinfo->tsc_shift; 1068 else 1069 freq = freq >> tinfo->tsc_shift; 1070 ci->ci_data.cpu_cc_freq = freq; 1071 } 1072 1073 void 1074 x86_cpu_idle_xen(void) 1075 { 1076 struct cpu_info *ci = curcpu(); 1077 1078 KASSERT(ci->ci_ilevel == IPL_NONE); 1079 1080 x86_disable_intr(); 1081 if (!__predict_false(ci->ci_want_resched)) { 1082 idle_block(); 1083 } else { 1084 x86_enable_intr(); 1085 } 1086 } 1087 1088 /* 1089 * Loads pmap for the current CPU. 1090 */ 1091 void 1092 cpu_load_pmap(struct pmap *pmap, struct pmap *oldpmap) 1093 { 1094 KASSERT(pmap != pmap_kernel()); 1095 1096 #if defined(__x86_64__) || defined(PAE) 1097 struct cpu_info *ci = curcpu(); 1098 cpuid_t cid = cpu_index(ci); 1099 1100 mutex_enter(&ci->ci_kpm_mtx); 1101 /* make new pmap visible to xen_kpm_sync() */ 1102 kcpuset_atomic_set(pmap->pm_xen_ptp_cpus, cid); 1103 #endif 1104 #ifdef i386 1105 #ifdef PAE 1106 { 1107 int i; 1108 paddr_t l3_pd = xpmap_ptom_masked(ci->ci_pae_l3_pdirpa); 1109 /* don't update the kernel L3 slot */ 1110 for (i = 0 ; i < PDP_SIZE - 1; i++) { 1111 xpq_queue_pte_update(l3_pd + i * sizeof(pd_entry_t), 1112 xpmap_ptom(pmap->pm_pdirpa[i]) | PG_V); 1113 } 1114 tlbflush(); 1115 } 1116 #else /* PAE */ 1117 lcr3(pmap_pdirpa(pmap, 0)); 1118 #endif /* PAE */ 1119 #endif /* i386 */ 1120 1121 #ifdef __x86_64__ 1122 { 1123 int i; 1124 pd_entry_t *new_pgd; 1125 paddr_t l4_pd_ma; 1126 1127 l4_pd_ma = xpmap_ptom_masked(ci->ci_kpm_pdirpa); 1128 1129 /* 1130 * Map user space address in kernel space and load 1131 * user cr3 1132 */ 1133 new_pgd = pmap->pm_pdir; 1134 KASSERT(pmap == ci->ci_pmap); 1135 1136 /* Copy user pmap L4 PDEs (in user addr. range) to per-cpu L4 */ 1137 for (i = 0; i < PDIR_SLOT_PTE; i++) { 1138 KASSERT(pmap != pmap_kernel() || new_pgd[i] == 0); 1139 if (ci->ci_kpm_pdir[i] != new_pgd[i]) { 1140 xpq_queue_pte_update( 1141 l4_pd_ma + i * sizeof(pd_entry_t), 1142 new_pgd[i]); 1143 } 1144 } 1145 1146 xen_set_user_pgd(pmap_pdirpa(pmap, 0)); 1147 ci->ci_xen_current_user_pgd = pmap_pdirpa(pmap, 0); 1148 1149 tlbflush(); 1150 } 1151 1152 #endif /* __x86_64__ */ 1153 #if defined(__x86_64__) || defined(PAE) 1154 /* old pmap no longer visible to xen_kpm_sync() */ 1155 if (oldpmap != pmap_kernel()) { 1156 kcpuset_atomic_clear(oldpmap->pm_xen_ptp_cpus, cid); 1157 } 1158 mutex_exit(&ci->ci_kpm_mtx); 1159 #endif 1160 } 1161 1162 /* 1163 * pmap_cpu_init_late: perform late per-CPU initialization. 1164 * Short note about percpu PDIR pages: 1165 * Both the PAE and __x86_64__ architectures have per-cpu PDIR 1166 * tables. This is to get around Xen's pagetable setup constraints for 1167 * PAE (multiple L3[3]s cannot point to the same L2 - Xen 1168 * will refuse to pin a table setup this way.) and for multiple cpus 1169 * to map in different user pmaps on __x86_64__ (see: cpu_load_pmap()) 1170 * 1171 * What this means for us is that the PDIR of the pmap_kernel() is 1172 * considered to be a canonical "SHADOW" PDIR with the following 1173 * properties: 1174 * - Its recursive mapping points to itself 1175 * - per-cpu recursive mappings point to themselves on __x86_64__ 1176 * - per-cpu L4 pages' kernel entries are expected to be in sync with 1177 * the shadow 1178 */ 1179 1180 void 1181 pmap_cpu_init_late(struct cpu_info *ci) 1182 { 1183 #if defined(PAE) || defined(__x86_64__) 1184 /* 1185 * The BP has already its own PD page allocated during early 1186 * MD startup. 1187 */ 1188 1189 #if defined(__x86_64__) 1190 /* Setup per-cpu normal_pdes */ 1191 int i; 1192 extern pd_entry_t * const normal_pdes[]; 1193 for (i = 0;i < PTP_LEVELS - 1;i++) { 1194 ci->ci_normal_pdes[i] = normal_pdes[i]; 1195 } 1196 #endif /* __x86_64__ */ 1197 1198 if (ci == &cpu_info_primary) 1199 return; 1200 1201 KASSERT(ci != NULL); 1202 1203 #if defined(PAE) 1204 cpu_alloc_l3_page(ci); 1205 KASSERT(ci->ci_pae_l3_pdirpa != 0); 1206 1207 /* Initialise L2 entries 0 - 2: Point them to pmap_kernel() */ 1208 int i; 1209 for (i = 0 ; i < PDP_SIZE - 1; i++) { 1210 ci->ci_pae_l3_pdir[i] = 1211 xpmap_ptom_masked(pmap_kernel()->pm_pdirpa[i]) | PG_V; 1212 } 1213 #endif /* PAE */ 1214 1215 ci->ci_kpm_pdir = (pd_entry_t *)uvm_km_alloc(kernel_map, PAGE_SIZE, 0, 1216 UVM_KMF_WIRED | UVM_KMF_ZERO | UVM_KMF_NOWAIT); 1217 1218 if (ci->ci_kpm_pdir == NULL) { 1219 panic("%s: failed to allocate L4 per-cpu PD for CPU %d\n", 1220 __func__, cpu_index(ci)); 1221 } 1222 ci->ci_kpm_pdirpa = vtophys((vaddr_t) ci->ci_kpm_pdir); 1223 KASSERT(ci->ci_kpm_pdirpa != 0); 1224 1225 #if defined(__x86_64__) 1226 /* 1227 * Copy over the pmap_kernel() shadow L4 entries 1228 */ 1229 1230 memcpy(ci->ci_kpm_pdir, pmap_kernel()->pm_pdir, PAGE_SIZE); 1231 1232 /* Recursive kernel mapping */ 1233 ci->ci_kpm_pdir[PDIR_SLOT_PTE] = xpmap_ptom_masked(ci->ci_kpm_pdirpa) | PG_k | PG_V; 1234 #elif defined(PAE) 1235 /* Copy over the pmap_kernel() shadow L2 entries that map the kernel */ 1236 memcpy(ci->ci_kpm_pdir, pmap_kernel()->pm_pdir + PDIR_SLOT_KERN, nkptp[PTP_LEVELS - 1] * sizeof(pd_entry_t)); 1237 #endif /* __x86_64__ else PAE */ 1238 1239 /* Xen wants R/O */ 1240 pmap_protect(pmap_kernel(), (vaddr_t)ci->ci_kpm_pdir, 1241 (vaddr_t)ci->ci_kpm_pdir + PAGE_SIZE, VM_PROT_READ); 1242 pmap_update(pmap_kernel()); 1243 #if defined(PAE) 1244 /* Initialise L3 entry 3. This mapping is shared across all 1245 * pmaps and is static, ie; loading a new pmap will not update 1246 * this entry. 1247 */ 1248 1249 ci->ci_pae_l3_pdir[3] = xpmap_ptom_masked(ci->ci_kpm_pdirpa) | PG_k | PG_V; 1250 1251 /* Mark L3 R/O (Xen wants this) */ 1252 pmap_protect(pmap_kernel(), (vaddr_t)ci->ci_pae_l3_pdir, 1253 (vaddr_t)ci->ci_pae_l3_pdir + PAGE_SIZE, VM_PROT_READ); 1254 pmap_update(pmap_kernel()); 1255 1256 xpq_queue_pin_l3_table(xpmap_ptom_masked(ci->ci_pae_l3_pdirpa)); 1257 1258 #elif defined(__x86_64__) 1259 xpq_queue_pin_l4_table(xpmap_ptom_masked(ci->ci_kpm_pdirpa)); 1260 #endif /* PAE , __x86_64__ */ 1261 #endif /* defined(PAE) || defined(__x86_64__) */ 1262 } 1263 1264 /* 1265 * Notify all other cpus to halt. 1266 */ 1267 1268 void 1269 cpu_broadcast_halt(void) 1270 { 1271 xen_broadcast_ipi(XEN_IPI_HALT); 1272 } 1273 1274 /* 1275 * Send a dummy ipi to a cpu. 1276 */ 1277 1278 void 1279 cpu_kick(struct cpu_info *ci) 1280 { 1281 (void)xen_send_ipi(ci, XEN_IPI_KICK); 1282 } 1283