1 /* $NetBSD: cpu.c,v 1.117 2018/01/13 14:48:13 bouyer Exp $ */ 2 3 /*- 4 * Copyright (c) 2000 The NetBSD Foundation, Inc. 5 * Copyright (c) 2002, 2006, 2007 YAMAMOTO Takashi, 6 * All rights reserved. 7 * 8 * This code is derived from software contributed to The NetBSD Foundation 9 * by RedBack Networks Inc. 10 * 11 * Author: Bill Sommerfeld 12 * 13 * Redistribution and use in source and binary forms, with or without 14 * modification, are permitted provided that the following conditions 15 * are met: 16 * 1. Redistributions of source code must retain the above copyright 17 * notice, this list of conditions and the following disclaimer. 18 * 2. Redistributions in binary form must reproduce the above copyright 19 * notice, this list of conditions and the following disclaimer in the 20 * documentation and/or other materials provided with the distribution. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS 23 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 24 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 25 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS 26 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 27 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 28 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 29 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 30 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 31 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 32 * POSSIBILITY OF SUCH DAMAGE. 33 */ 34 35 /* 36 * Copyright (c) 1999 Stefan Grefen 37 * 38 * Redistribution and use in source and binary forms, with or without 39 * modification, are permitted provided that the following conditions 40 * are met: 41 * 1. Redistributions of source code must retain the above copyright 42 * notice, this list of conditions and the following disclaimer. 43 * 2. Redistributions in binary form must reproduce the above copyright 44 * notice, this list of conditions and the following disclaimer in the 45 * documentation and/or other materials provided with the distribution. 46 * 3. All advertising materials mentioning features or use of this software 47 * must display the following acknowledgement: 48 * This product includes software developed by the NetBSD 49 * Foundation, Inc. and its contributors. 50 * 4. Neither the name of The NetBSD Foundation nor the names of its 51 * contributors may be used to endorse or promote products derived 52 * from this software without specific prior written permission. 53 * 54 * THIS SOFTWARE IS PROVIDED BY AUTHOR AND CONTRIBUTORS ``AS IS'' AND ANY 55 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 56 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 57 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR AND CONTRIBUTORS BE LIABLE 58 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 59 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 60 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 61 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 62 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 63 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 64 * SUCH DAMAGE. 65 */ 66 67 #include <sys/cdefs.h> 68 __KERNEL_RCSID(0, "$NetBSD: cpu.c,v 1.117 2018/01/13 14:48:13 bouyer Exp $"); 69 70 #include "opt_ddb.h" 71 #include "opt_multiprocessor.h" 72 #include "opt_mpbios.h" /* for MPDEBUG */ 73 #include "opt_mtrr.h" 74 #include "opt_xen.h" 75 76 #include "lapic.h" 77 #include "ioapic.h" 78 79 #include <sys/param.h> 80 #include <sys/proc.h> 81 #include <sys/systm.h> 82 #include <sys/device.h> 83 #include <sys/kmem.h> 84 #include <sys/cpu.h> 85 #include <sys/cpufreq.h> 86 #include <sys/atomic.h> 87 #include <sys/reboot.h> 88 #include <sys/idle.h> 89 90 #include <uvm/uvm.h> 91 92 #include <machine/cpu.h> 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 #include <x86/fpu.h> 106 107 #include <xen/xen.h> 108 #include <xen/xen-public/vcpu.h> 109 #include <xen/vcpuvar.h> 110 111 #if NLAPIC > 0 112 #include <machine/apicvar.h> 113 #include <machine/i82489reg.h> 114 #include <machine/i82489var.h> 115 #endif 116 117 #include <dev/ic/mc146818reg.h> 118 #include <dev/isa/isareg.h> 119 120 static int cpu_match(device_t, cfdata_t, void *); 121 static void cpu_attach(device_t, device_t, void *); 122 static void cpu_defer(device_t); 123 static int cpu_rescan(device_t, const char *, const int *); 124 static void cpu_childdetached(device_t, device_t); 125 static int vcpu_match(device_t, cfdata_t, void *); 126 static void vcpu_attach(device_t, device_t, void *); 127 static void cpu_attach_common(device_t, device_t, void *); 128 void cpu_offline_md(void); 129 130 struct cpu_softc { 131 device_t sc_dev; /* device tree glue */ 132 struct cpu_info *sc_info; /* pointer to CPU info */ 133 bool sc_wasonline; 134 }; 135 136 int mp_cpu_start(struct cpu_info *, vaddr_t); 137 void mp_cpu_start_cleanup(struct cpu_info *); 138 const struct cpu_functions mp_cpu_funcs = { mp_cpu_start, NULL, 139 mp_cpu_start_cleanup }; 140 141 CFATTACH_DECL2_NEW(cpu, sizeof(struct cpu_softc), 142 cpu_match, cpu_attach, NULL, NULL, cpu_rescan, cpu_childdetached); 143 144 CFATTACH_DECL_NEW(vcpu, sizeof(struct cpu_softc), 145 vcpu_match, vcpu_attach, NULL, NULL); 146 147 /* 148 * Statically-allocated CPU info for the primary CPU (or the only 149 * CPU, on uniprocessors). The CPU info list is initialized to 150 * point at it. 151 */ 152 struct cpu_info cpu_info_primary __aligned(CACHE_LINE_SIZE) = { 153 .ci_dev = 0, 154 .ci_self = &cpu_info_primary, 155 .ci_idepth = -1, 156 .ci_curlwp = &lwp0, 157 .ci_curldt = -1, 158 }; 159 struct cpu_info phycpu_info_primary __aligned(CACHE_LINE_SIZE) = { 160 .ci_dev = 0, 161 .ci_self = &phycpu_info_primary, 162 }; 163 164 struct cpu_info *cpu_info_list = &cpu_info_primary; 165 struct cpu_info *phycpu_info_list = &phycpu_info_primary; 166 167 uint32_t cpu_feature[7] __read_mostly; /* X86 CPUID feature bits 168 * [0] basic features %edx 169 * [1] basic features %ecx 170 * [2] extended features %edx 171 * [3] extended features %ecx 172 * [4] VIA padlock features 173 * [5] structured extended features cpuid.7:%ebx 174 * [6] structured extended features cpuid.7:%ecx 175 */ 176 177 bool x86_mp_online; 178 paddr_t mp_trampoline_paddr = MP_TRAMPOLINE; 179 180 #if defined(MULTIPROCESSOR) 181 void cpu_hatch(void *); 182 static void cpu_boot_secondary(struct cpu_info *ci); 183 static void cpu_start_secondary(struct cpu_info *ci); 184 #endif /* MULTIPROCESSOR */ 185 186 static int 187 cpu_match(device_t parent, cfdata_t match, void *aux) 188 { 189 190 return 1; 191 } 192 193 static void 194 cpu_attach(device_t parent, device_t self, void *aux) 195 { 196 struct cpu_softc *sc = device_private(self); 197 struct cpu_attach_args *caa = aux; 198 struct cpu_info *ci; 199 uintptr_t ptr; 200 static int nphycpu = 0; 201 202 sc->sc_dev = self; 203 204 /* 205 * If we're an Application Processor, allocate a cpu_info 206 * If we're the first attached CPU use the primary cpu_info, 207 * otherwise allocate a new one 208 */ 209 aprint_naive("\n"); 210 aprint_normal("\n"); 211 if (nphycpu > 0) { 212 struct cpu_info *tmp; 213 ptr = (uintptr_t)kmem_zalloc(sizeof(*ci) + CACHE_LINE_SIZE - 1, 214 KM_SLEEP); 215 ci = (struct cpu_info *)roundup2(ptr, CACHE_LINE_SIZE); 216 ci->ci_curldt = -1; 217 218 tmp = phycpu_info_list; 219 while (tmp->ci_next) 220 tmp = tmp->ci_next; 221 222 tmp->ci_next = ci; 223 } else { 224 ci = &phycpu_info_primary; 225 } 226 227 ci->ci_self = ci; 228 sc->sc_info = ci; 229 230 ci->ci_dev = self; 231 ci->ci_acpiid = caa->cpu_id; 232 ci->ci_cpuid = caa->cpu_number; 233 ci->ci_vcpu = NULL; 234 ci->ci_index = nphycpu++; 235 236 if (!pmf_device_register(self, NULL, NULL)) 237 aprint_error_dev(self, "couldn't establish power handler\n"); 238 239 (void)config_defer(self, cpu_defer); 240 } 241 242 static void 243 cpu_defer(device_t self) 244 { 245 cpu_rescan(self, NULL, NULL); 246 } 247 248 static int 249 cpu_rescan(device_t self, const char *ifattr, const int *locators) 250 { 251 struct cpu_softc *sc = device_private(self); 252 struct cpufeature_attach_args cfaa; 253 struct cpu_info *ci = sc->sc_info; 254 255 memset(&cfaa, 0, sizeof(cfaa)); 256 cfaa.ci = ci; 257 258 if (ifattr_match(ifattr, "cpufeaturebus")) { 259 260 if (ci->ci_frequency == NULL) { 261 cfaa.name = "frequency"; 262 ci->ci_frequency = config_found_ia(self, 263 "cpufeaturebus", &cfaa, NULL); 264 } 265 } 266 267 return 0; 268 } 269 270 static void 271 cpu_childdetached(device_t self, device_t child) 272 { 273 struct cpu_softc *sc = device_private(self); 274 struct cpu_info *ci = sc->sc_info; 275 276 if (ci->ci_frequency == child) 277 ci->ci_frequency = NULL; 278 } 279 280 static int 281 vcpu_match(device_t parent, cfdata_t match, void *aux) 282 { 283 struct vcpu_attach_args *vcaa = aux; 284 struct vcpu_runstate_info vcr; 285 int error; 286 287 if (strcmp(vcaa->vcaa_name, match->cf_name) == 0) { 288 error = HYPERVISOR_vcpu_op(VCPUOP_get_runstate_info, 289 vcaa->vcaa_caa.cpu_number, &vcr); 290 switch (error) { 291 case 0: 292 return 1; 293 case -ENOENT: 294 return 0; 295 default: 296 panic("Unknown hypervisor error %d returned on vcpu runstate probe\n", error); 297 } 298 } 299 300 return 0; 301 } 302 303 static void 304 vcpu_attach(device_t parent, device_t self, void *aux) 305 { 306 struct vcpu_attach_args *vcaa = aux; 307 308 KASSERT(vcaa->vcaa_caa.cpu_func == NULL); 309 vcaa->vcaa_caa.cpu_func = &mp_cpu_funcs; 310 cpu_attach_common(parent, self, &vcaa->vcaa_caa); 311 312 if (!pmf_device_register(self, NULL, NULL)) 313 aprint_error_dev(self, "couldn't establish power handler\n"); 314 } 315 316 static int 317 vcpu_is_up(struct cpu_info *ci) 318 { 319 KASSERT(ci != NULL); 320 return HYPERVISOR_vcpu_op(VCPUOP_is_up, ci->ci_cpuid, NULL); 321 } 322 323 static void 324 cpu_vm_init(struct cpu_info *ci) 325 { 326 int ncolors = 2, i; 327 328 for (i = CAI_ICACHE; i <= CAI_L2CACHE; i++) { 329 struct x86_cache_info *cai; 330 int tcolors; 331 332 cai = &ci->ci_cinfo[i]; 333 334 tcolors = atop(cai->cai_totalsize); 335 switch (cai->cai_associativity) { 336 case 0xff: 337 tcolors = 1; /* fully associative */ 338 break; 339 case 0: 340 case 1: 341 break; 342 default: 343 tcolors /= cai->cai_associativity; 344 } 345 ncolors = max(ncolors, tcolors); 346 } 347 348 /* 349 * Knowing the size of the largest cache on this CPU, potentially 350 * re-color our pages. 351 */ 352 aprint_debug_dev(ci->ci_dev, "%d page colors\n", ncolors); 353 uvm_page_recolor(ncolors); 354 pmap_tlb_cpu_init(ci); 355 #ifndef __HAVE_DIRECT_MAP 356 pmap_vpage_cpu_init(ci); 357 #endif 358 } 359 360 static void 361 cpu_attach_common(device_t parent, device_t self, void *aux) 362 { 363 struct cpu_softc *sc = device_private(self); 364 struct cpu_attach_args *caa = aux; 365 struct cpu_info *ci; 366 uintptr_t ptr; 367 int cpunum = caa->cpu_number; 368 static bool again = false; 369 370 sc->sc_dev = self; 371 372 /* 373 * If we're an Application Processor, allocate a cpu_info 374 * structure, otherwise use the primary's. 375 */ 376 if (caa->cpu_role == CPU_ROLE_AP) { 377 aprint_naive(": Application Processor\n"); 378 ptr = (uintptr_t)kmem_alloc(sizeof(*ci) + CACHE_LINE_SIZE - 1, 379 KM_SLEEP); 380 ci = (struct cpu_info *)roundup2(ptr, CACHE_LINE_SIZE); 381 memset(ci, 0, sizeof(*ci)); 382 cpu_init_tss(ci); 383 } else { 384 aprint_naive(": %s Processor\n", 385 caa->cpu_role == CPU_ROLE_SP ? "Single" : "Boot"); 386 ci = &cpu_info_primary; 387 } 388 389 ci->ci_self = ci; 390 sc->sc_info = ci; 391 ci->ci_dev = self; 392 ci->ci_cpuid = cpunum; 393 394 KASSERT(HYPERVISOR_shared_info != NULL); 395 KASSERT(cpunum < XEN_LEGACY_MAX_VCPUS); 396 ci->ci_vcpu = &HYPERVISOR_shared_info->vcpu_info[cpunum]; 397 398 KASSERT(ci->ci_func == 0); 399 ci->ci_func = caa->cpu_func; 400 aprint_normal("\n"); 401 402 /* Must be called before mi_cpu_attach(). */ 403 cpu_vm_init(ci); 404 405 if (caa->cpu_role == CPU_ROLE_AP) { 406 int error; 407 408 error = mi_cpu_attach(ci); 409 410 KASSERT(ci->ci_data.cpu_idlelwp != NULL); 411 if (error != 0) { 412 aprint_error_dev(self, 413 "mi_cpu_attach failed with %d\n", error); 414 return; 415 } 416 417 } else { 418 KASSERT(ci->ci_data.cpu_idlelwp != NULL); 419 } 420 421 KASSERT(ci->ci_cpuid == ci->ci_index); 422 #ifdef __x86_64__ 423 /* No user PGD mapped for this CPU yet */ 424 ci->ci_xen_current_user_pgd = 0; 425 #endif 426 #if defined(__x86_64__) || defined(PAE) 427 mutex_init(&ci->ci_kpm_mtx, MUTEX_DEFAULT, IPL_VM); 428 #endif 429 pmap_reference(pmap_kernel()); 430 ci->ci_pmap = pmap_kernel(); 431 ci->ci_tlbstate = TLBSTATE_STALE; 432 433 /* 434 * Boot processor may not be attached first, but the below 435 * must be done to allow booting other processors. 436 */ 437 if (!again) { 438 atomic_or_32(&ci->ci_flags, CPUF_PRESENT | CPUF_PRIMARY); 439 /* Basic init. */ 440 cpu_intr_init(ci); 441 cpu_get_tsc_freq(ci); 442 cpu_init(ci); 443 pmap_cpu_init_late(ci); 444 445 /* Every processor needs to init its own ipi h/w (similar to lapic) */ 446 xen_ipi_init(); 447 448 /* Make sure DELAY() is initialized. */ 449 DELAY(1); 450 again = true; 451 } 452 453 /* further PCB init done later. */ 454 455 switch (caa->cpu_role) { 456 case CPU_ROLE_SP: 457 atomic_or_32(&ci->ci_flags, CPUF_SP); 458 cpu_identify(ci); 459 x86_cpu_idle_init(); 460 break; 461 462 case CPU_ROLE_BP: 463 atomic_or_32(&ci->ci_flags, CPUF_BSP); 464 cpu_identify(ci); 465 x86_cpu_idle_init(); 466 break; 467 468 case CPU_ROLE_AP: 469 atomic_or_32(&ci->ci_flags, CPUF_AP); 470 471 /* 472 * report on an AP 473 */ 474 475 #if defined(MULTIPROCESSOR) 476 /* interrupt handler stack */ 477 cpu_intr_init(ci); 478 479 /* Setup per-cpu memory for gdt */ 480 gdt_alloc_cpu(ci); 481 482 pmap_cpu_init_late(ci); 483 cpu_start_secondary(ci); 484 485 if (ci->ci_flags & CPUF_PRESENT) { 486 struct cpu_info *tmp; 487 488 cpu_identify(ci); 489 tmp = cpu_info_list; 490 while (tmp->ci_next) 491 tmp = tmp->ci_next; 492 493 tmp->ci_next = ci; 494 } 495 #else 496 aprint_error_dev(ci->ci_dev, "not started\n"); 497 #endif 498 break; 499 500 default: 501 panic("unknown processor type??\n"); 502 } 503 504 #ifdef MPVERBOSE 505 if (mp_verbose) { 506 struct lwp *l = ci->ci_data.cpu_idlelwp; 507 struct pcb *pcb = lwp_getpcb(l); 508 509 aprint_verbose_dev(self, 510 "idle lwp at %p, idle sp at 0x%p\n", 511 l, 512 #ifdef i386 513 (void *)pcb->pcb_esp 514 #else 515 (void *)pcb->pcb_rsp 516 #endif 517 ); 518 519 } 520 #endif /* MPVERBOSE */ 521 } 522 523 /* 524 * Initialize the processor appropriately. 525 */ 526 527 void 528 cpu_init(struct cpu_info *ci) 529 { 530 531 /* 532 * If we have FXSAVE/FXRESTOR, use them. 533 */ 534 if (cpu_feature[0] & CPUID_FXSR) { 535 lcr4(rcr4() | CR4_OSFXSR); 536 537 /* 538 * If we have SSE/SSE2, enable XMM exceptions. 539 */ 540 if (cpu_feature[0] & (CPUID_SSE|CPUID_SSE2)) 541 lcr4(rcr4() | CR4_OSXMMEXCPT); 542 } 543 544 if (x86_fpu_save >= FPU_SAVE_FXSAVE) { 545 fpuinit_mxcsr_mask(); 546 } 547 548 atomic_or_32(&ci->ci_flags, CPUF_RUNNING); 549 } 550 551 552 #ifdef MULTIPROCESSOR 553 554 void 555 cpu_boot_secondary_processors(void) 556 { 557 struct cpu_info *ci; 558 u_long i; 559 for (i = 0; i < maxcpus; i++) { 560 ci = cpu_lookup(i); 561 if (ci == NULL) 562 continue; 563 if (ci->ci_data.cpu_idlelwp == NULL) 564 continue; 565 if ((ci->ci_flags & CPUF_PRESENT) == 0) 566 continue; 567 if (ci->ci_flags & (CPUF_BSP|CPUF_SP|CPUF_PRIMARY)) 568 continue; 569 cpu_boot_secondary(ci); 570 } 571 572 x86_mp_online = true; 573 } 574 575 static void 576 cpu_init_idle_lwp(struct cpu_info *ci) 577 { 578 struct lwp *l = ci->ci_data.cpu_idlelwp; 579 struct pcb *pcb = lwp_getpcb(l); 580 581 pcb->pcb_cr0 = rcr0(); 582 } 583 584 void 585 cpu_init_idle_lwps(void) 586 { 587 struct cpu_info *ci; 588 u_long i; 589 590 for (i = 0; i < maxcpus; i++) { 591 ci = cpu_lookup(i); 592 if (ci == NULL) 593 continue; 594 if (ci->ci_data.cpu_idlelwp == NULL) 595 continue; 596 if ((ci->ci_flags & CPUF_PRESENT) == 0) 597 continue; 598 cpu_init_idle_lwp(ci); 599 } 600 } 601 602 static void 603 cpu_start_secondary(struct cpu_info *ci) 604 { 605 int i; 606 607 aprint_debug_dev(ci->ci_dev, "starting\n"); 608 609 ci->ci_curlwp = ci->ci_data.cpu_idlelwp; 610 611 if (CPU_STARTUP(ci, (vaddr_t) cpu_hatch) != 0) { 612 return; 613 } 614 615 /* 616 * wait for it to become ready 617 */ 618 for (i = 100000; (!(ci->ci_flags & CPUF_PRESENT)) && i > 0; i--) { 619 delay(10); 620 } 621 if ((ci->ci_flags & CPUF_PRESENT) == 0) { 622 aprint_error_dev(ci->ci_dev, "failed to become ready\n"); 623 #if defined(MPDEBUG) && defined(DDB) 624 printf("dropping into debugger; continue from here to resume boot\n"); 625 Debugger(); 626 #endif 627 } 628 629 CPU_START_CLEANUP(ci); 630 } 631 632 void 633 cpu_boot_secondary(struct cpu_info *ci) 634 { 635 int i; 636 atomic_or_32(&ci->ci_flags, CPUF_GO); 637 for (i = 100000; (!(ci->ci_flags & CPUF_RUNNING)) && i > 0; i--) { 638 delay(10); 639 } 640 if ((ci->ci_flags & CPUF_RUNNING) == 0) { 641 aprint_error_dev(ci->ci_dev, "CPU failed to start\n"); 642 #if defined(MPDEBUG) && defined(DDB) 643 printf("dropping into debugger; continue from here to resume boot\n"); 644 Debugger(); 645 #endif 646 } 647 } 648 649 /* 650 * APs end up here immediately after initialisation and VCPUOP_up in 651 * mp_cpu_start(). 652 * At this point, we are running in the idle pcb/idle stack of the new 653 * CPU. This function jumps to the idle loop and starts looking for 654 * work. 655 */ 656 extern void x86_64_tls_switch(struct lwp *); 657 void 658 cpu_hatch(void *v) 659 { 660 struct cpu_info *ci = (struct cpu_info *)v; 661 struct pcb *pcb; 662 int s, i; 663 664 /* Setup TLS and kernel GS/FS */ 665 cpu_init_msrs(ci, true); 666 cpu_init_idt(); 667 gdt_init_cpu(ci); 668 669 cpu_probe(ci); 670 671 atomic_or_32(&ci->ci_flags, CPUF_PRESENT); 672 673 while ((ci->ci_flags & CPUF_GO) == 0) { 674 /* Don't use delay, boot CPU may be patching the text. */ 675 for (i = 10000; i != 0; i--) 676 x86_pause(); 677 } 678 679 /* Because the text may have been patched in x86_patch(). */ 680 x86_flush(); 681 tlbflushg(); 682 683 KASSERT((ci->ci_flags & CPUF_RUNNING) == 0); 684 685 pcb = lwp_getpcb(curlwp); 686 pcb->pcb_cr3 = pmap_pdirpa(pmap_kernel(), 0); 687 pcb = lwp_getpcb(ci->ci_data.cpu_idlelwp); 688 689 xen_ipi_init(); 690 691 xen_initclocks(); 692 693 #ifdef __x86_64__ 694 fpuinit(ci); 695 #endif 696 697 lldt(GSEL(GLDT_SEL, SEL_KPL)); 698 699 cpu_init(ci); 700 cpu_get_tsc_freq(ci); 701 702 s = splhigh(); 703 x86_enable_intr(); 704 splx(s); 705 706 aprint_debug_dev(ci->ci_dev, "running\n"); 707 708 cpu_switchto(NULL, ci->ci_data.cpu_idlelwp, true); 709 710 idle_loop(NULL); 711 KASSERT(false); 712 } 713 714 #if defined(DDB) 715 716 #include <ddb/db_output.h> 717 #include <machine/db_machdep.h> 718 719 /* 720 * Dump CPU information from ddb. 721 */ 722 void 723 cpu_debug_dump(void) 724 { 725 struct cpu_info *ci; 726 CPU_INFO_ITERATOR cii; 727 728 db_printf("addr dev id flags ipis curlwp fpcurlwp\n"); 729 for (CPU_INFO_FOREACH(cii, ci)) { 730 db_printf("%p %s %ld %x %x %10p %10p\n", 731 ci, 732 ci->ci_dev == NULL ? "BOOT" : device_xname(ci->ci_dev), 733 (long)ci->ci_cpuid, 734 ci->ci_flags, ci->ci_ipis, 735 ci->ci_curlwp, 736 ci->ci_fpcurlwp); 737 } 738 } 739 #endif /* DDB */ 740 741 #endif /* MULTIPROCESSOR */ 742 743 extern void hypervisor_callback(void); 744 extern void failsafe_callback(void); 745 #ifdef __x86_64__ 746 typedef void (vector)(void); 747 extern vector Xsyscall, Xsyscall32; 748 #endif 749 750 /* 751 * Setup the "trampoline". On Xen, we setup nearly all cpu context 752 * outside a trampoline, so we prototype and call targetip like so: 753 * void targetip(struct cpu_info *); 754 */ 755 756 static void 757 gdt_prepframes(paddr_t *frames, vaddr_t base, uint32_t entries) 758 { 759 int i; 760 for (i = 0; i < entries; i++) { 761 frames[i] = ((paddr_t)xpmap_ptetomach( 762 (pt_entry_t *)(base + (i << PAGE_SHIFT)))) >> PAGE_SHIFT; 763 764 /* Mark Read-only */ 765 pmap_pte_clearbits(kvtopte(base + (i << PAGE_SHIFT)), 766 PG_RW); 767 } 768 } 769 770 #ifdef __x86_64__ 771 extern char *ldtstore; 772 773 static void 774 xen_init_amd64_vcpuctxt(struct cpu_info *ci, struct vcpu_guest_context *initctx, 775 void targetrip(struct cpu_info *)) 776 { 777 /* page frames to point at GDT */ 778 extern int gdt_size; 779 paddr_t frames[16]; 780 psize_t gdt_ents; 781 782 struct lwp *l; 783 struct pcb *pcb; 784 785 volatile struct vcpu_info *vci; 786 787 KASSERT(ci != NULL); 788 KASSERT(ci != &cpu_info_primary); 789 KASSERT(initctx != NULL); 790 KASSERT(targetrip != NULL); 791 792 memset(initctx, 0, sizeof(*initctx)); 793 794 gdt_ents = roundup(gdt_size, PAGE_SIZE) >> PAGE_SHIFT; 795 KASSERT(gdt_ents <= 16); 796 797 gdt_prepframes(frames, (vaddr_t)ci->ci_gdt, gdt_ents); 798 799 /* Initialise the vcpu context: We use idle_loop()'s pcb context. */ 800 801 l = ci->ci_data.cpu_idlelwp; 802 803 KASSERT(l != NULL); 804 pcb = lwp_getpcb(l); 805 KASSERT(pcb != NULL); 806 807 /* resume with interrupts off */ 808 vci = ci->ci_vcpu; 809 vci->evtchn_upcall_mask = 1; 810 xen_mb(); 811 812 /* resume in kernel-mode */ 813 initctx->flags = VGCF_in_kernel | VGCF_online; 814 815 /* Stack and entry points: 816 * We arrange for the stack frame for cpu_hatch() to 817 * appear as a callee frame of lwp_trampoline(). Being a 818 * leaf frame prevents trampling on any of the MD stack setup 819 * that x86/vm_machdep.c:cpu_lwp_fork() does for idle_loop() 820 */ 821 822 initctx->user_regs.rdi = (uint64_t) ci; /* targetrip(ci); */ 823 initctx->user_regs.rip = (vaddr_t) targetrip; 824 825 initctx->user_regs.cs = GSEL(GCODE_SEL, SEL_KPL); 826 827 initctx->user_regs.rflags = pcb->pcb_flags; 828 initctx->user_regs.rsp = pcb->pcb_rsp; 829 830 /* Data segments */ 831 initctx->user_regs.ss = GSEL(GDATA_SEL, SEL_KPL); 832 initctx->user_regs.es = GSEL(GDATA_SEL, SEL_KPL); 833 initctx->user_regs.ds = GSEL(GDATA_SEL, SEL_KPL); 834 835 /* GDT */ 836 memcpy(initctx->gdt_frames, frames, sizeof(frames)); 837 initctx->gdt_ents = gdt_ents; 838 839 /* LDT */ 840 initctx->ldt_base = (unsigned long)ldtstore; 841 initctx->ldt_ents = LDT_SIZE >> 3; 842 843 /* Kernel context state */ 844 initctx->kernel_ss = GSEL(GDATA_SEL, SEL_KPL); 845 initctx->kernel_sp = pcb->pcb_rsp0; 846 initctx->ctrlreg[0] = pcb->pcb_cr0; 847 initctx->ctrlreg[1] = 0; /* "resuming" from kernel - no User cr3. */ 848 initctx->ctrlreg[2] = (vaddr_t)targetrip; 849 /* 850 * Use pmap_kernel() L4 PD directly, until we setup the 851 * per-cpu L4 PD in pmap_cpu_init_late() 852 */ 853 initctx->ctrlreg[3] = xen_pfn_to_cr3(x86_btop(xpmap_ptom(ci->ci_kpm_pdirpa))); 854 initctx->ctrlreg[4] = CR4_PAE | CR4_OSFXSR | CR4_OSXMMEXCPT; 855 856 /* Xen callbacks */ 857 initctx->event_callback_eip = (unsigned long)hypervisor_callback; 858 initctx->failsafe_callback_eip = (unsigned long)failsafe_callback; 859 initctx->syscall_callback_eip = (unsigned long)Xsyscall; 860 861 return; 862 } 863 #else /* i386 */ 864 extern union descriptor *ldtstore; 865 extern void Xsyscall(void); 866 867 static void 868 xen_init_i386_vcpuctxt(struct cpu_info *ci, struct vcpu_guest_context *initctx, 869 void targeteip(struct cpu_info *)) 870 { 871 /* page frames to point at GDT */ 872 extern int gdt_size; 873 paddr_t frames[16]; 874 psize_t gdt_ents; 875 876 struct lwp *l; 877 struct pcb *pcb; 878 879 volatile struct vcpu_info *vci; 880 881 KASSERT(ci != NULL); 882 KASSERT(ci != &cpu_info_primary); 883 KASSERT(initctx != NULL); 884 KASSERT(targeteip != NULL); 885 886 memset(initctx, 0, sizeof(*initctx)); 887 888 gdt_ents = roundup(gdt_size, PAGE_SIZE) >> PAGE_SHIFT; 889 KASSERT(gdt_ents <= 16); 890 891 gdt_prepframes(frames, (vaddr_t)ci->ci_gdt, gdt_ents); 892 893 /* 894 * Initialise the vcpu context: 895 * We use this cpu's idle_loop() pcb context. 896 */ 897 898 l = ci->ci_data.cpu_idlelwp; 899 900 KASSERT(l != NULL); 901 pcb = lwp_getpcb(l); 902 KASSERT(pcb != NULL); 903 904 /* resume with interrupts off */ 905 vci = ci->ci_vcpu; 906 vci->evtchn_upcall_mask = 1; 907 xen_mb(); 908 909 /* resume in kernel-mode */ 910 initctx->flags = VGCF_in_kernel | VGCF_online; 911 912 /* Stack frame setup for cpu_hatch(): 913 * We arrange for the stack frame for cpu_hatch() to 914 * appear as a callee frame of lwp_trampoline(). Being a 915 * leaf frame prevents trampling on any of the MD stack setup 916 * that x86/vm_machdep.c:cpu_lwp_fork() does for idle_loop() 917 */ 918 919 initctx->user_regs.esp = pcb->pcb_esp - 4; /* Leave word for 920 arg1 */ 921 { 922 /* targeteip(ci); */ 923 uint32_t *arg = (uint32_t *)initctx->user_regs.esp; 924 arg[1] = (uint32_t)ci; /* arg1 */ 925 } 926 927 initctx->user_regs.eip = (vaddr_t)targeteip; 928 initctx->user_regs.cs = GSEL(GCODE_SEL, SEL_KPL); 929 initctx->user_regs.eflags |= pcb->pcb_iopl; 930 931 /* Data segments */ 932 initctx->user_regs.ss = GSEL(GDATA_SEL, SEL_KPL); 933 initctx->user_regs.es = GSEL(GDATA_SEL, SEL_KPL); 934 initctx->user_regs.ds = GSEL(GDATA_SEL, SEL_KPL); 935 initctx->user_regs.fs = GSEL(GDATA_SEL, SEL_KPL); 936 937 /* GDT */ 938 memcpy(initctx->gdt_frames, frames, sizeof(frames)); 939 initctx->gdt_ents = gdt_ents; 940 941 /* LDT */ 942 initctx->ldt_base = (unsigned long)ldtstore; 943 initctx->ldt_ents = NLDT; 944 945 /* Kernel context state */ 946 initctx->kernel_ss = GSEL(GDATA_SEL, SEL_KPL); 947 initctx->kernel_sp = pcb->pcb_esp0; 948 initctx->ctrlreg[0] = pcb->pcb_cr0; 949 initctx->ctrlreg[1] = 0; /* "resuming" from kernel - no User cr3. */ 950 initctx->ctrlreg[2] = (vaddr_t)targeteip; 951 #ifdef PAE 952 initctx->ctrlreg[3] = xen_pfn_to_cr3(x86_btop(xpmap_ptom(ci->ci_pae_l3_pdirpa))); 953 #else 954 initctx->ctrlreg[3] = xen_pfn_to_cr3(x86_btop(xpmap_ptom(pcb->pcb_cr3))); 955 #endif 956 initctx->ctrlreg[4] = /* CR4_PAE | */CR4_OSFXSR | CR4_OSXMMEXCPT; 957 958 /* Xen callbacks */ 959 initctx->event_callback_eip = (unsigned long)hypervisor_callback; 960 initctx->event_callback_cs = GSEL(GCODE_SEL, SEL_KPL); 961 initctx->failsafe_callback_eip = (unsigned long)failsafe_callback; 962 initctx->failsafe_callback_cs = GSEL(GCODE_SEL, SEL_KPL); 963 964 return; 965 } 966 #endif /* __x86_64__ */ 967 968 int 969 mp_cpu_start(struct cpu_info *ci, vaddr_t target) 970 { 971 int hyperror; 972 struct vcpu_guest_context vcpuctx; 973 974 KASSERT(ci != NULL); 975 KASSERT(ci != &cpu_info_primary); 976 KASSERT(ci->ci_flags & CPUF_AP); 977 978 #ifdef __x86_64__ 979 xen_init_amd64_vcpuctxt(ci, &vcpuctx, (void (*)(struct cpu_info *))target); 980 #else 981 xen_init_i386_vcpuctxt(ci, &vcpuctx, (void (*)(struct cpu_info *))target); 982 #endif 983 984 /* Initialise the given vcpu to execute cpu_hatch(ci); */ 985 if ((hyperror = HYPERVISOR_vcpu_op(VCPUOP_initialise, ci->ci_cpuid, &vcpuctx))) { 986 aprint_error(": context initialisation failed. errno = %d\n", hyperror); 987 return hyperror; 988 } 989 990 /* Start it up */ 991 992 /* First bring it down */ 993 if ((hyperror = HYPERVISOR_vcpu_op(VCPUOP_down, ci->ci_cpuid, NULL))) { 994 aprint_error(": VCPUOP_down hypervisor command failed. errno = %d\n", hyperror); 995 return hyperror; 996 } 997 998 if ((hyperror = HYPERVISOR_vcpu_op(VCPUOP_up, ci->ci_cpuid, NULL))) { 999 aprint_error(": VCPUOP_up hypervisor command failed. errno = %d\n", hyperror); 1000 return hyperror; 1001 } 1002 1003 if (!vcpu_is_up(ci)) { 1004 aprint_error(": did not come up\n"); 1005 return -1; 1006 } 1007 1008 return 0; 1009 } 1010 1011 void 1012 mp_cpu_start_cleanup(struct cpu_info *ci) 1013 { 1014 if (vcpu_is_up(ci)) { 1015 aprint_debug_dev(ci->ci_dev, "is started.\n"); 1016 } else { 1017 aprint_error_dev(ci->ci_dev, "did not start up.\n"); 1018 } 1019 } 1020 1021 void 1022 cpu_init_msrs(struct cpu_info *ci, bool full) 1023 { 1024 #ifdef __x86_64__ 1025 if (full) { 1026 HYPERVISOR_set_segment_base(SEGBASE_FS, 0); 1027 HYPERVISOR_set_segment_base(SEGBASE_GS_KERNEL, (uint64_t)ci); 1028 HYPERVISOR_set_segment_base(SEGBASE_GS_USER, 0); 1029 } 1030 #endif 1031 1032 if (cpu_feature[2] & CPUID_NOX) 1033 wrmsr(MSR_EFER, rdmsr(MSR_EFER) | EFER_NXE); 1034 } 1035 1036 void 1037 cpu_offline_md(void) 1038 { 1039 int s; 1040 1041 s = splhigh(); 1042 fpusave_cpu(true); 1043 splx(s); 1044 } 1045 1046 void 1047 cpu_get_tsc_freq(struct cpu_info *ci) 1048 { 1049 uint32_t vcpu_tversion; 1050 const volatile vcpu_time_info_t *tinfo = &ci->ci_vcpu->time; 1051 1052 vcpu_tversion = tinfo->version; 1053 while (tinfo->version == vcpu_tversion); /* Wait for a time update. XXX: timeout ? */ 1054 1055 uint64_t freq = 1000000000ULL << 32; 1056 freq = freq / (uint64_t)tinfo->tsc_to_system_mul; 1057 if (tinfo->tsc_shift < 0) 1058 freq = freq << -tinfo->tsc_shift; 1059 else 1060 freq = freq >> tinfo->tsc_shift; 1061 ci->ci_data.cpu_cc_freq = freq; 1062 } 1063 1064 void 1065 x86_cpu_idle_xen(void) 1066 { 1067 struct cpu_info *ci = curcpu(); 1068 1069 KASSERT(ci->ci_ilevel == IPL_NONE); 1070 1071 x86_disable_intr(); 1072 if (!__predict_false(ci->ci_want_resched)) { 1073 idle_block(); 1074 } else { 1075 x86_enable_intr(); 1076 } 1077 } 1078 1079 /* 1080 * Loads pmap for the current CPU. 1081 */ 1082 void 1083 cpu_load_pmap(struct pmap *pmap, struct pmap *oldpmap) 1084 { 1085 KASSERT(pmap != pmap_kernel()); 1086 1087 #if defined(__x86_64__) || defined(PAE) 1088 struct cpu_info *ci = curcpu(); 1089 cpuid_t cid = cpu_index(ci); 1090 1091 mutex_enter(&ci->ci_kpm_mtx); 1092 /* make new pmap visible to xen_kpm_sync() */ 1093 kcpuset_atomic_set(pmap->pm_xen_ptp_cpus, cid); 1094 #endif 1095 1096 #ifdef i386 1097 #ifdef PAE 1098 { 1099 int i; 1100 paddr_t l3_pd = xpmap_ptom_masked(ci->ci_pae_l3_pdirpa); 1101 /* don't update the kernel L3 slot */ 1102 for (i = 0 ; i < PDP_SIZE - 1; i++) { 1103 xpq_queue_pte_update(l3_pd + i * sizeof(pd_entry_t), 1104 xpmap_ptom(pmap->pm_pdirpa[i]) | PG_V); 1105 } 1106 tlbflush(); 1107 } 1108 #else /* PAE */ 1109 lcr3(pmap_pdirpa(pmap, 0)); 1110 #endif /* PAE */ 1111 #endif /* i386 */ 1112 1113 #ifdef __x86_64__ 1114 { 1115 int i; 1116 pd_entry_t *new_pgd; 1117 paddr_t l4_pd_ma; 1118 1119 l4_pd_ma = xpmap_ptom_masked(ci->ci_kpm_pdirpa); 1120 1121 /* 1122 * Map user space address in kernel space and load 1123 * user cr3 1124 */ 1125 new_pgd = pmap->pm_pdir; 1126 KASSERT(pmap == ci->ci_pmap); 1127 1128 /* Copy user pmap L4 PDEs (in user addr. range) to per-cpu L4 */ 1129 for (i = 0; i < PDIR_SLOT_PTE; i++) { 1130 KASSERT(pmap != pmap_kernel() || new_pgd[i] == 0); 1131 if (ci->ci_kpm_pdir[i] != new_pgd[i]) { 1132 xpq_queue_pte_update( 1133 l4_pd_ma + i * sizeof(pd_entry_t), 1134 new_pgd[i]); 1135 } 1136 } 1137 1138 xen_set_user_pgd(pmap_pdirpa(pmap, 0)); 1139 ci->ci_xen_current_user_pgd = pmap_pdirpa(pmap, 0); 1140 1141 tlbflush(); 1142 } 1143 #endif /* __x86_64__ */ 1144 1145 #if defined(__x86_64__) || defined(PAE) 1146 /* old pmap no longer visible to xen_kpm_sync() */ 1147 if (oldpmap != pmap_kernel()) { 1148 kcpuset_atomic_clear(oldpmap->pm_xen_ptp_cpus, cid); 1149 } 1150 mutex_exit(&ci->ci_kpm_mtx); 1151 #endif 1152 } 1153 1154 /* 1155 * pmap_cpu_init_late: perform late per-CPU initialization. 1156 * 1157 * Short note about percpu PDIR pages. Both the PAE and __x86_64__ architectures 1158 * have per-cpu PDIR tables, for two different reasons: 1159 * - on PAE, this is to get around Xen's pagetable setup constraints (multiple 1160 * L3[3]s cannot point to the same L2 - Xen will refuse to pin a table set up 1161 * this way). 1162 * - on __x86_64__, this is for multiple CPUs to map in different user pmaps 1163 * (see cpu_load_pmap()). 1164 * 1165 * What this means for us is that the PDIR of the pmap_kernel() is considered 1166 * to be a canonical "SHADOW" PDIR with the following properties: 1167 * - its recursive mapping points to itself 1168 * - per-cpu recursive mappings point to themselves on __x86_64__ 1169 * - per-cpu L4 pages' kernel entries are expected to be in sync with 1170 * the shadow 1171 */ 1172 1173 void 1174 pmap_cpu_init_late(struct cpu_info *ci) 1175 { 1176 #if defined(PAE) || defined(__x86_64__) 1177 /* 1178 * The BP has already its own PD page allocated during early 1179 * MD startup. 1180 */ 1181 1182 #if defined(__x86_64__) 1183 /* Setup per-cpu normal_pdes */ 1184 int i; 1185 extern pd_entry_t * const normal_pdes[]; 1186 for (i = 0;i < PTP_LEVELS - 1;i++) { 1187 ci->ci_normal_pdes[i] = normal_pdes[i]; 1188 } 1189 #endif /* __x86_64__ */ 1190 1191 if (ci == &cpu_info_primary) 1192 return; 1193 1194 KASSERT(ci != NULL); 1195 1196 #if defined(PAE) 1197 cpu_alloc_l3_page(ci); 1198 KASSERT(ci->ci_pae_l3_pdirpa != 0); 1199 1200 /* Initialise L2 entries 0 - 2: Point them to pmap_kernel() */ 1201 int i; 1202 for (i = 0 ; i < PDP_SIZE - 1; i++) { 1203 ci->ci_pae_l3_pdir[i] = 1204 xpmap_ptom_masked(pmap_kernel()->pm_pdirpa[i]) | PG_V; 1205 } 1206 #endif /* PAE */ 1207 1208 ci->ci_kpm_pdir = (pd_entry_t *)uvm_km_alloc(kernel_map, PAGE_SIZE, 0, 1209 UVM_KMF_WIRED | UVM_KMF_ZERO | UVM_KMF_NOWAIT); 1210 1211 if (ci->ci_kpm_pdir == NULL) { 1212 panic("%s: failed to allocate L4 per-cpu PD for CPU %d\n", 1213 __func__, cpu_index(ci)); 1214 } 1215 ci->ci_kpm_pdirpa = vtophys((vaddr_t)ci->ci_kpm_pdir); 1216 KASSERT(ci->ci_kpm_pdirpa != 0); 1217 1218 #if defined(__x86_64__) 1219 extern pt_entry_t xpmap_pg_nx; 1220 1221 /* Copy over the pmap_kernel() shadow L4 entries */ 1222 memcpy(ci->ci_kpm_pdir, pmap_kernel()->pm_pdir, PAGE_SIZE); 1223 1224 /* Recursive kernel mapping */ 1225 ci->ci_kpm_pdir[PDIR_SLOT_PTE] = xpmap_ptom_masked(ci->ci_kpm_pdirpa) 1226 | PG_V | xpmap_pg_nx; 1227 #elif defined(PAE) 1228 /* Copy over the pmap_kernel() shadow L2 entries */ 1229 memcpy(ci->ci_kpm_pdir, pmap_kernel()->pm_pdir + PDIR_SLOT_KERN, 1230 nkptp[PTP_LEVELS - 1] * sizeof(pd_entry_t)); 1231 #endif 1232 1233 /* Xen wants a RO pdir. */ 1234 pmap_protect(pmap_kernel(), (vaddr_t)ci->ci_kpm_pdir, 1235 (vaddr_t)ci->ci_kpm_pdir + PAGE_SIZE, VM_PROT_READ); 1236 pmap_update(pmap_kernel()); 1237 #if defined(PAE) 1238 /* 1239 * Initialize L3 entry 3. This mapping is shared across all pmaps and is 1240 * static, ie: loading a new pmap will not update this entry. 1241 */ 1242 ci->ci_pae_l3_pdir[3] = xpmap_ptom_masked(ci->ci_kpm_pdirpa) | PG_V; 1243 1244 /* Xen wants a RO L3. */ 1245 pmap_protect(pmap_kernel(), (vaddr_t)ci->ci_pae_l3_pdir, 1246 (vaddr_t)ci->ci_pae_l3_pdir + PAGE_SIZE, VM_PROT_READ); 1247 pmap_update(pmap_kernel()); 1248 1249 xpq_queue_pin_l3_table(xpmap_ptom_masked(ci->ci_pae_l3_pdirpa)); 1250 1251 #elif defined(__x86_64__) 1252 xpq_queue_pin_l4_table(xpmap_ptom_masked(ci->ci_kpm_pdirpa)); 1253 #endif /* PAE , __x86_64__ */ 1254 #endif /* defined(PAE) || defined(__x86_64__) */ 1255 } 1256 1257 /* 1258 * Notify all other cpus to halt. 1259 */ 1260 1261 void 1262 cpu_broadcast_halt(void) 1263 { 1264 xen_broadcast_ipi(XEN_IPI_HALT); 1265 } 1266 1267 /* 1268 * Send a dummy ipi to a cpu. 1269 */ 1270 1271 void 1272 cpu_kick(struct cpu_info *ci) 1273 { 1274 (void)xen_send_ipi(ci, XEN_IPI_KICK); 1275 } 1276