mp_machdep.c revision 271999
1/*- 2 * Copyright (c) 1996, by Steve Passe 3 * Copyright (c) 2003, by Peter Wemm 4 * All rights reserved. 5 * 6 * Redistribution and use in source and binary forms, with or without 7 * modification, are permitted provided that the following conditions 8 * are met: 9 * 1. Redistributions of source code must retain the above copyright 10 * notice, this list of conditions and the following disclaimer. 11 * 2. The name of the developer may NOT be used to endorse or promote products 12 * derived from this software without specific prior written permission. 13 * 14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 24 * SUCH DAMAGE. 25 */ 26 27#include <sys/cdefs.h> 28__FBSDID("$FreeBSD: stable/10/sys/amd64/amd64/mp_machdep.c 271999 2014-09-22 20:34:36Z jhb $"); 29 30#include "opt_cpu.h" 31#include "opt_ddb.h" 32#include "opt_kstack_pages.h" 33#include "opt_sched.h" 34#include "opt_smp.h" 35 36#include <sys/param.h> 37#include <sys/systm.h> 38#include <sys/bus.h> 39#include <sys/cpuset.h> 40#ifdef GPROF 41#include <sys/gmon.h> 42#endif 43#include <sys/kernel.h> 44#include <sys/ktr.h> 45#include <sys/lock.h> 46#include <sys/malloc.h> 47#include <sys/memrange.h> 48#include <sys/mutex.h> 49#include <sys/pcpu.h> 50#include <sys/proc.h> 51#include <sys/sched.h> 52#include <sys/smp.h> 53#include <sys/sysctl.h> 54 55#include <vm/vm.h> 56#include <vm/vm_param.h> 57#include <vm/pmap.h> 58#include <vm/vm_kern.h> 59#include <vm/vm_extern.h> 60 61#include <x86/apicreg.h> 62#include <machine/clock.h> 63#include <machine/cputypes.h> 64#include <machine/cpufunc.h> 65#include <x86/mca.h> 66#include <machine/md_var.h> 67#include <machine/pcb.h> 68#include <machine/psl.h> 69#include <machine/smp.h> 70#include <machine/specialreg.h> 71#include <machine/tss.h> 72#include <machine/cpu.h> 73 74#define WARMBOOT_TARGET 0 75#define WARMBOOT_OFF (KERNBASE + 0x0467) 76#define WARMBOOT_SEG (KERNBASE + 0x0469) 77 78#define CMOS_REG (0x70) 79#define CMOS_DATA (0x71) 80#define BIOS_RESET (0x0f) 81#define BIOS_WARM (0x0a) 82 83/* lock region used by kernel profiling */ 84int mcount_lock; 85 86int mp_naps; /* # of Applications processors */ 87int boot_cpu_id = -1; /* designated BSP */ 88 89extern struct pcpu __pcpu[]; 90 91/* AP uses this during bootstrap. Do not staticize. */ 92char *bootSTK; 93static int bootAP; 94 95/* Free these after use */ 96void *bootstacks[MAXCPU]; 97 98/* Temporary variables for init_secondary() */ 99char *doublefault_stack; 100char *nmi_stack; 101void *dpcpu; 102 103struct pcb stoppcbs[MAXCPU]; 104struct susppcb **susppcbs; 105 106/* Variables needed for SMP tlb shootdown. */ 107vm_offset_t smp_tlb_addr2; 108struct invpcid_descr smp_tlb_invpcid; 109volatile int smp_tlb_wait; 110uint64_t pcid_cr3; 111pmap_t smp_tlb_pmap; 112extern int invpcid_works; 113 114#ifdef COUNT_IPIS 115/* Interrupt counts. */ 116static u_long *ipi_preempt_counts[MAXCPU]; 117static u_long *ipi_ast_counts[MAXCPU]; 118u_long *ipi_invltlb_counts[MAXCPU]; 119u_long *ipi_invlrng_counts[MAXCPU]; 120u_long *ipi_invlpg_counts[MAXCPU]; 121u_long *ipi_invlcache_counts[MAXCPU]; 122u_long *ipi_rendezvous_counts[MAXCPU]; 123static u_long *ipi_hardclock_counts[MAXCPU]; 124#endif 125 126/* Default cpu_ops implementation. */ 127struct cpu_ops cpu_ops = { 128 .ipi_vectored = lapic_ipi_vectored 129}; 130 131extern inthand_t IDTVEC(fast_syscall), IDTVEC(fast_syscall32); 132 133extern int pmap_pcid_enabled; 134 135/* 136 * Local data and functions. 137 */ 138 139static volatile cpuset_t ipi_nmi_pending; 140 141/* used to hold the AP's until we are ready to release them */ 142static struct mtx ap_boot_mtx; 143 144/* Set to 1 once we're ready to let the APs out of the pen. */ 145static volatile int aps_ready = 0; 146 147/* 148 * Store data from cpu_add() until later in the boot when we actually setup 149 * the APs. 150 */ 151struct cpu_info { 152 int cpu_present:1; 153 int cpu_bsp:1; 154 int cpu_disabled:1; 155 int cpu_hyperthread:1; 156} static cpu_info[MAX_APIC_ID + 1]; 157int cpu_apic_ids[MAXCPU]; 158int apic_cpuids[MAX_APIC_ID + 1]; 159 160/* Holds pending bitmap based IPIs per CPU */ 161volatile u_int cpu_ipi_pending[MAXCPU]; 162 163static u_int boot_address; 164static int cpu_logical; /* logical cpus per core */ 165static int cpu_cores; /* cores per package */ 166 167static void assign_cpu_ids(void); 168static void set_interrupt_apic_ids(void); 169static int start_all_aps(void); 170static int start_ap(int apic_id); 171static void release_aps(void *dummy); 172 173static u_int hyperthreading_cpus; /* logical cpus sharing L1 cache */ 174static int hyperthreading_allowed = 1; 175static u_int bootMP_size; 176 177static void 178mem_range_AP_init(void) 179{ 180 if (mem_range_softc.mr_op && mem_range_softc.mr_op->initAP) 181 mem_range_softc.mr_op->initAP(&mem_range_softc); 182} 183 184static void 185topo_probe_amd(void) 186{ 187 int core_id_bits; 188 int id; 189 190 /* AMD processors do not support HTT. */ 191 cpu_logical = 1; 192 193 if ((amd_feature2 & AMDID2_CMP) == 0) { 194 cpu_cores = 1; 195 return; 196 } 197 198 core_id_bits = (cpu_procinfo2 & AMDID_COREID_SIZE) >> 199 AMDID_COREID_SIZE_SHIFT; 200 if (core_id_bits == 0) { 201 cpu_cores = (cpu_procinfo2 & AMDID_CMP_CORES) + 1; 202 return; 203 } 204 205 /* Fam 10h and newer should get here. */ 206 for (id = 0; id <= MAX_APIC_ID; id++) { 207 /* Check logical CPU availability. */ 208 if (!cpu_info[id].cpu_present || cpu_info[id].cpu_disabled) 209 continue; 210 /* Check if logical CPU has the same package ID. */ 211 if ((id >> core_id_bits) != (boot_cpu_id >> core_id_bits)) 212 continue; 213 cpu_cores++; 214 } 215} 216 217/* 218 * Round up to the next power of two, if necessary, and then 219 * take log2. 220 * Returns -1 if argument is zero. 221 */ 222static __inline int 223mask_width(u_int x) 224{ 225 226 return (fls(x << (1 - powerof2(x))) - 1); 227} 228 229static void 230topo_probe_0x4(void) 231{ 232 u_int p[4]; 233 int pkg_id_bits; 234 int core_id_bits; 235 int max_cores; 236 int max_logical; 237 int id; 238 239 /* Both zero and one here mean one logical processor per package. */ 240 max_logical = (cpu_feature & CPUID_HTT) != 0 ? 241 (cpu_procinfo & CPUID_HTT_CORES) >> 16 : 1; 242 if (max_logical <= 1) 243 return; 244 245 /* 246 * Because of uniformity assumption we examine only 247 * those logical processors that belong to the same 248 * package as BSP. Further, we count number of 249 * logical processors that belong to the same core 250 * as BSP thus deducing number of threads per core. 251 */ 252 if (cpu_high >= 0x4) { 253 cpuid_count(0x04, 0, p); 254 max_cores = ((p[0] >> 26) & 0x3f) + 1; 255 } else 256 max_cores = 1; 257 core_id_bits = mask_width(max_logical/max_cores); 258 if (core_id_bits < 0) 259 return; 260 pkg_id_bits = core_id_bits + mask_width(max_cores); 261 262 for (id = 0; id <= MAX_APIC_ID; id++) { 263 /* Check logical CPU availability. */ 264 if (!cpu_info[id].cpu_present || cpu_info[id].cpu_disabled) 265 continue; 266 /* Check if logical CPU has the same package ID. */ 267 if ((id >> pkg_id_bits) != (boot_cpu_id >> pkg_id_bits)) 268 continue; 269 cpu_cores++; 270 /* Check if logical CPU has the same package and core IDs. */ 271 if ((id >> core_id_bits) == (boot_cpu_id >> core_id_bits)) 272 cpu_logical++; 273 } 274 275 KASSERT(cpu_cores >= 1 && cpu_logical >= 1, 276 ("topo_probe_0x4 couldn't find BSP")); 277 278 cpu_cores /= cpu_logical; 279 hyperthreading_cpus = cpu_logical; 280} 281 282static void 283topo_probe_0xb(void) 284{ 285 u_int p[4]; 286 int bits; 287 int cnt; 288 int i; 289 int logical; 290 int type; 291 int x; 292 293 /* We only support three levels for now. */ 294 for (i = 0; i < 3; i++) { 295 cpuid_count(0x0b, i, p); 296 297 /* Fall back if CPU leaf 11 doesn't really exist. */ 298 if (i == 0 && p[1] == 0) { 299 topo_probe_0x4(); 300 return; 301 } 302 303 bits = p[0] & 0x1f; 304 logical = p[1] &= 0xffff; 305 type = (p[2] >> 8) & 0xff; 306 if (type == 0 || logical == 0) 307 break; 308 /* 309 * Because of uniformity assumption we examine only 310 * those logical processors that belong to the same 311 * package as BSP. 312 */ 313 for (cnt = 0, x = 0; x <= MAX_APIC_ID; x++) { 314 if (!cpu_info[x].cpu_present || 315 cpu_info[x].cpu_disabled) 316 continue; 317 if (x >> bits == boot_cpu_id >> bits) 318 cnt++; 319 } 320 if (type == CPUID_TYPE_SMT) 321 cpu_logical = cnt; 322 else if (type == CPUID_TYPE_CORE) 323 cpu_cores = cnt; 324 } 325 if (cpu_logical == 0) 326 cpu_logical = 1; 327 cpu_cores /= cpu_logical; 328} 329 330/* 331 * Both topology discovery code and code that consumes topology 332 * information assume top-down uniformity of the topology. 333 * That is, all physical packages must be identical and each 334 * core in a package must have the same number of threads. 335 * Topology information is queried only on BSP, on which this 336 * code runs and for which it can query CPUID information. 337 * Then topology is extrapolated on all packages using the 338 * uniformity assumption. 339 */ 340static void 341topo_probe(void) 342{ 343 static int cpu_topo_probed = 0; 344 345 if (cpu_topo_probed) 346 return; 347 348 CPU_ZERO(&logical_cpus_mask); 349 if (mp_ncpus <= 1) 350 cpu_cores = cpu_logical = 1; 351 else if (cpu_vendor_id == CPU_VENDOR_AMD) 352 topo_probe_amd(); 353 else if (cpu_vendor_id == CPU_VENDOR_INTEL) { 354 /* 355 * See Intel(R) 64 Architecture Processor 356 * Topology Enumeration article for details. 357 * 358 * Note that 0x1 <= cpu_high < 4 case should be 359 * compatible with topo_probe_0x4() logic when 360 * CPUID.1:EBX[23:16] > 0 (cpu_cores will be 1) 361 * or it should trigger the fallback otherwise. 362 */ 363 if (cpu_high >= 0xb) 364 topo_probe_0xb(); 365 else if (cpu_high >= 0x1) 366 topo_probe_0x4(); 367 } 368 369 /* 370 * Fallback: assume each logical CPU is in separate 371 * physical package. That is, no multi-core, no SMT. 372 */ 373 if (cpu_cores == 0 || cpu_logical == 0) 374 cpu_cores = cpu_logical = 1; 375 cpu_topo_probed = 1; 376} 377 378struct cpu_group * 379cpu_topo(void) 380{ 381 int cg_flags; 382 383 /* 384 * Determine whether any threading flags are 385 * necessry. 386 */ 387 topo_probe(); 388 if (cpu_logical > 1 && hyperthreading_cpus) 389 cg_flags = CG_FLAG_HTT; 390 else if (cpu_logical > 1) 391 cg_flags = CG_FLAG_SMT; 392 else 393 cg_flags = 0; 394 if (mp_ncpus % (cpu_cores * cpu_logical) != 0) { 395 printf("WARNING: Non-uniform processors.\n"); 396 printf("WARNING: Using suboptimal topology.\n"); 397 return (smp_topo_none()); 398 } 399 /* 400 * No multi-core or hyper-threaded. 401 */ 402 if (cpu_logical * cpu_cores == 1) 403 return (smp_topo_none()); 404 /* 405 * Only HTT no multi-core. 406 */ 407 if (cpu_logical > 1 && cpu_cores == 1) 408 return (smp_topo_1level(CG_SHARE_L1, cpu_logical, cg_flags)); 409 /* 410 * Only multi-core no HTT. 411 */ 412 if (cpu_cores > 1 && cpu_logical == 1) 413 return (smp_topo_1level(CG_SHARE_L2, cpu_cores, cg_flags)); 414 /* 415 * Both HTT and multi-core. 416 */ 417 return (smp_topo_2level(CG_SHARE_L2, cpu_cores, 418 CG_SHARE_L1, cpu_logical, cg_flags)); 419} 420 421/* 422 * Calculate usable address in base memory for AP trampoline code. 423 */ 424u_int 425mp_bootaddress(u_int basemem) 426{ 427 428 bootMP_size = mptramp_end - mptramp_start; 429 boot_address = trunc_page(basemem * 1024); /* round down to 4k boundary */ 430 if (((basemem * 1024) - boot_address) < bootMP_size) 431 boot_address -= PAGE_SIZE; /* not enough, lower by 4k */ 432 /* 3 levels of page table pages */ 433 mptramp_pagetables = boot_address - (PAGE_SIZE * 3); 434 435 return mptramp_pagetables; 436} 437 438void 439cpu_add(u_int apic_id, char boot_cpu) 440{ 441 442 if (apic_id > MAX_APIC_ID) { 443 panic("SMP: APIC ID %d too high", apic_id); 444 return; 445 } 446 KASSERT(cpu_info[apic_id].cpu_present == 0, ("CPU %d added twice", 447 apic_id)); 448 cpu_info[apic_id].cpu_present = 1; 449 if (boot_cpu) { 450 KASSERT(boot_cpu_id == -1, 451 ("CPU %d claims to be BSP, but CPU %d already is", apic_id, 452 boot_cpu_id)); 453 boot_cpu_id = apic_id; 454 cpu_info[apic_id].cpu_bsp = 1; 455 } 456 if (mp_ncpus < MAXCPU) { 457 mp_ncpus++; 458 mp_maxid = mp_ncpus - 1; 459 } 460 if (bootverbose) 461 printf("SMP: Added CPU %d (%s)\n", apic_id, boot_cpu ? "BSP" : 462 "AP"); 463} 464 465void 466cpu_mp_setmaxid(void) 467{ 468 469 /* 470 * mp_maxid should be already set by calls to cpu_add(). 471 * Just sanity check its value here. 472 */ 473 if (mp_ncpus == 0) 474 KASSERT(mp_maxid == 0, 475 ("%s: mp_ncpus is zero, but mp_maxid is not", __func__)); 476 else if (mp_ncpus == 1) 477 mp_maxid = 0; 478 else 479 KASSERT(mp_maxid >= mp_ncpus - 1, 480 ("%s: counters out of sync: max %d, count %d", __func__, 481 mp_maxid, mp_ncpus)); 482} 483 484int 485cpu_mp_probe(void) 486{ 487 488 /* 489 * Always record BSP in CPU map so that the mbuf init code works 490 * correctly. 491 */ 492 CPU_SETOF(0, &all_cpus); 493 if (mp_ncpus == 0) { 494 /* 495 * No CPUs were found, so this must be a UP system. Setup 496 * the variables to represent a system with a single CPU 497 * with an id of 0. 498 */ 499 mp_ncpus = 1; 500 return (0); 501 } 502 503 /* At least one CPU was found. */ 504 if (mp_ncpus == 1) { 505 /* 506 * One CPU was found, so this must be a UP system with 507 * an I/O APIC. 508 */ 509 mp_maxid = 0; 510 return (0); 511 } 512 513 /* At least two CPUs were found. */ 514 return (1); 515} 516 517/* 518 * Initialize the IPI handlers and start up the AP's. 519 */ 520void 521cpu_mp_start(void) 522{ 523 int i; 524 525 /* Initialize the logical ID to APIC ID table. */ 526 for (i = 0; i < MAXCPU; i++) { 527 cpu_apic_ids[i] = -1; 528 cpu_ipi_pending[i] = 0; 529 } 530 531 /* Install an inter-CPU IPI for TLB invalidation */ 532 if (pmap_pcid_enabled) { 533 setidt(IPI_INVLTLB, IDTVEC(invltlb_pcid), SDT_SYSIGT, 534 SEL_KPL, 0); 535 setidt(IPI_INVLPG, IDTVEC(invlpg_pcid), SDT_SYSIGT, 536 SEL_KPL, 0); 537 } else { 538 setidt(IPI_INVLTLB, IDTVEC(invltlb), SDT_SYSIGT, SEL_KPL, 0); 539 setidt(IPI_INVLPG, IDTVEC(invlpg), SDT_SYSIGT, SEL_KPL, 0); 540 } 541 setidt(IPI_INVLRNG, IDTVEC(invlrng), SDT_SYSIGT, SEL_KPL, 0); 542 543 /* Install an inter-CPU IPI for cache invalidation. */ 544 setidt(IPI_INVLCACHE, IDTVEC(invlcache), SDT_SYSIGT, SEL_KPL, 0); 545 546 /* Install an inter-CPU IPI for all-CPU rendezvous */ 547 setidt(IPI_RENDEZVOUS, IDTVEC(rendezvous), SDT_SYSIGT, SEL_KPL, 0); 548 549 /* Install generic inter-CPU IPI handler */ 550 setidt(IPI_BITMAP_VECTOR, IDTVEC(ipi_intr_bitmap_handler), 551 SDT_SYSIGT, SEL_KPL, 0); 552 553 /* Install an inter-CPU IPI for CPU stop/restart */ 554 setidt(IPI_STOP, IDTVEC(cpustop), SDT_SYSIGT, SEL_KPL, 0); 555 556 /* Install an inter-CPU IPI for CPU suspend/resume */ 557 setidt(IPI_SUSPEND, IDTVEC(cpususpend), SDT_SYSIGT, SEL_KPL, 0); 558 559 /* Set boot_cpu_id if needed. */ 560 if (boot_cpu_id == -1) { 561 boot_cpu_id = PCPU_GET(apic_id); 562 cpu_info[boot_cpu_id].cpu_bsp = 1; 563 } else 564 KASSERT(boot_cpu_id == PCPU_GET(apic_id), 565 ("BSP's APIC ID doesn't match boot_cpu_id")); 566 567 /* Probe logical/physical core configuration. */ 568 topo_probe(); 569 570 assign_cpu_ids(); 571 572 /* Start each Application Processor */ 573 start_all_aps(); 574 575 set_interrupt_apic_ids(); 576} 577 578 579/* 580 * Print various information about the SMP system hardware and setup. 581 */ 582void 583cpu_mp_announce(void) 584{ 585 const char *hyperthread; 586 int i; 587 588 printf("FreeBSD/SMP: %d package(s) x %d core(s)", 589 mp_ncpus / (cpu_cores * cpu_logical), cpu_cores); 590 if (hyperthreading_cpus > 1) 591 printf(" x %d HTT threads", cpu_logical); 592 else if (cpu_logical > 1) 593 printf(" x %d SMT threads", cpu_logical); 594 printf("\n"); 595 596 /* List active CPUs first. */ 597 printf(" cpu0 (BSP): APIC ID: %2d\n", boot_cpu_id); 598 for (i = 1; i < mp_ncpus; i++) { 599 if (cpu_info[cpu_apic_ids[i]].cpu_hyperthread) 600 hyperthread = "/HT"; 601 else 602 hyperthread = ""; 603 printf(" cpu%d (AP%s): APIC ID: %2d\n", i, hyperthread, 604 cpu_apic_ids[i]); 605 } 606 607 /* List disabled CPUs last. */ 608 for (i = 0; i <= MAX_APIC_ID; i++) { 609 if (!cpu_info[i].cpu_present || !cpu_info[i].cpu_disabled) 610 continue; 611 if (cpu_info[i].cpu_hyperthread) 612 hyperthread = "/HT"; 613 else 614 hyperthread = ""; 615 printf(" cpu (AP%s): APIC ID: %2d (disabled)\n", hyperthread, 616 i); 617 } 618} 619 620/* 621 * AP CPU's call this to initialize themselves. 622 */ 623void 624init_secondary(void) 625{ 626 struct pcpu *pc; 627 struct nmi_pcpu *np; 628 u_int64_t msr, cr0; 629 u_int cpuid; 630 int cpu, gsel_tss, x; 631 struct region_descriptor ap_gdt; 632 633 /* Set by the startup code for us to use */ 634 cpu = bootAP; 635 636 /* Init tss */ 637 common_tss[cpu] = common_tss[0]; 638 common_tss[cpu].tss_rsp0 = 0; /* not used until after switch */ 639 common_tss[cpu].tss_iobase = sizeof(struct amd64tss) + 640 IOPAGES * PAGE_SIZE; 641 common_tss[cpu].tss_ist1 = (long)&doublefault_stack[PAGE_SIZE]; 642 643 /* The NMI stack runs on IST2. */ 644 np = ((struct nmi_pcpu *) &nmi_stack[PAGE_SIZE]) - 1; 645 common_tss[cpu].tss_ist2 = (long) np; 646 647 /* Prepare private GDT */ 648 gdt_segs[GPROC0_SEL].ssd_base = (long) &common_tss[cpu]; 649 for (x = 0; x < NGDT; x++) { 650 if (x != GPROC0_SEL && x != (GPROC0_SEL + 1) && 651 x != GUSERLDT_SEL && x != (GUSERLDT_SEL + 1)) 652 ssdtosd(&gdt_segs[x], &gdt[NGDT * cpu + x]); 653 } 654 ssdtosyssd(&gdt_segs[GPROC0_SEL], 655 (struct system_segment_descriptor *)&gdt[NGDT * cpu + GPROC0_SEL]); 656 ap_gdt.rd_limit = NGDT * sizeof(gdt[0]) - 1; 657 ap_gdt.rd_base = (long) &gdt[NGDT * cpu]; 658 lgdt(&ap_gdt); /* does magic intra-segment return */ 659 660 /* Get per-cpu data */ 661 pc = &__pcpu[cpu]; 662 663 /* prime data page for it to use */ 664 pcpu_init(pc, cpu, sizeof(struct pcpu)); 665 dpcpu_init(dpcpu, cpu); 666 pc->pc_apic_id = cpu_apic_ids[cpu]; 667 pc->pc_prvspace = pc; 668 pc->pc_curthread = 0; 669 pc->pc_tssp = &common_tss[cpu]; 670 pc->pc_commontssp = &common_tss[cpu]; 671 pc->pc_rsp0 = 0; 672 pc->pc_tss = (struct system_segment_descriptor *)&gdt[NGDT * cpu + 673 GPROC0_SEL]; 674 pc->pc_fs32p = &gdt[NGDT * cpu + GUFS32_SEL]; 675 pc->pc_gs32p = &gdt[NGDT * cpu + GUGS32_SEL]; 676 pc->pc_ldt = (struct system_segment_descriptor *)&gdt[NGDT * cpu + 677 GUSERLDT_SEL]; 678 679 /* Save the per-cpu pointer for use by the NMI handler. */ 680 np->np_pcpu = (register_t) pc; 681 682 wrmsr(MSR_FSBASE, 0); /* User value */ 683 wrmsr(MSR_GSBASE, (u_int64_t)pc); 684 wrmsr(MSR_KGSBASE, (u_int64_t)pc); /* XXX User value while we're in the kernel */ 685 686 lidt(&r_idt); 687 688 gsel_tss = GSEL(GPROC0_SEL, SEL_KPL); 689 ltr(gsel_tss); 690 691 /* 692 * Set to a known state: 693 * Set by mpboot.s: CR0_PG, CR0_PE 694 * Set by cpu_setregs: CR0_NE, CR0_MP, CR0_TS, CR0_WP, CR0_AM 695 */ 696 cr0 = rcr0(); 697 cr0 &= ~(CR0_CD | CR0_NW | CR0_EM); 698 load_cr0(cr0); 699 700 /* Set up the fast syscall stuff */ 701 msr = rdmsr(MSR_EFER) | EFER_SCE; 702 wrmsr(MSR_EFER, msr); 703 wrmsr(MSR_LSTAR, (u_int64_t)IDTVEC(fast_syscall)); 704 wrmsr(MSR_CSTAR, (u_int64_t)IDTVEC(fast_syscall32)); 705 msr = ((u_int64_t)GSEL(GCODE_SEL, SEL_KPL) << 32) | 706 ((u_int64_t)GSEL(GUCODE32_SEL, SEL_UPL) << 48); 707 wrmsr(MSR_STAR, msr); 708 wrmsr(MSR_SF_MASK, PSL_NT|PSL_T|PSL_I|PSL_C|PSL_D); 709 710 /* Disable local APIC just to be sure. */ 711 lapic_disable(); 712 713 /* signal our startup to the BSP. */ 714 mp_naps++; 715 716 /* Spin until the BSP releases the AP's. */ 717 while (!aps_ready) 718 ia32_pause(); 719 720 /* Initialize the PAT MSR. */ 721 pmap_init_pat(); 722 723 /* set up CPU registers and state */ 724 cpu_setregs(); 725 726 /* set up SSE/NX registers */ 727 initializecpu(); 728 729 /* set up FPU state on the AP */ 730 fpuinit(); 731 732 if (cpu_ops.cpu_init) 733 cpu_ops.cpu_init(); 734 735 /* A quick check from sanity claus */ 736 cpuid = PCPU_GET(cpuid); 737 if (PCPU_GET(apic_id) != lapic_id()) { 738 printf("SMP: cpuid = %d\n", cpuid); 739 printf("SMP: actual apic_id = %d\n", lapic_id()); 740 printf("SMP: correct apic_id = %d\n", PCPU_GET(apic_id)); 741 panic("cpuid mismatch! boom!!"); 742 } 743 744 /* Initialize curthread. */ 745 KASSERT(PCPU_GET(idlethread) != NULL, ("no idle thread")); 746 PCPU_SET(curthread, PCPU_GET(idlethread)); 747 748 mca_init(); 749 750 mtx_lock_spin(&ap_boot_mtx); 751 752 /* Init local apic for irq's */ 753 lapic_setup(1); 754 755 /* Set memory range attributes for this CPU to match the BSP */ 756 mem_range_AP_init(); 757 758 smp_cpus++; 759 760 CTR1(KTR_SMP, "SMP: AP CPU #%d Launched", cpuid); 761 printf("SMP: AP CPU #%d Launched!\n", cpuid); 762 763 /* Determine if we are a logical CPU. */ 764 /* XXX Calculation depends on cpu_logical being a power of 2, e.g. 2 */ 765 if (cpu_logical > 1 && PCPU_GET(apic_id) % cpu_logical != 0) 766 CPU_SET(cpuid, &logical_cpus_mask); 767 768 if (bootverbose) 769 lapic_dump("AP"); 770 771 if (smp_cpus == mp_ncpus) { 772 /* enable IPI's, tlb shootdown, freezes etc */ 773 atomic_store_rel_int(&smp_started, 1); 774 } 775 776 /* 777 * Enable global pages TLB extension 778 * This also implicitly flushes the TLB 779 */ 780 781 load_cr4(rcr4() | CR4_PGE); 782 if (pmap_pcid_enabled) 783 load_cr4(rcr4() | CR4_PCIDE); 784 load_ds(_udatasel); 785 load_es(_udatasel); 786 load_fs(_ufssel); 787 mtx_unlock_spin(&ap_boot_mtx); 788 789 /* Wait until all the AP's are up. */ 790 while (smp_started == 0) 791 ia32_pause(); 792 793 /* Start per-CPU event timers. */ 794 cpu_initclocks_ap(); 795 796 sched_throw(NULL); 797 798 panic("scheduler returned us to %s", __func__); 799 /* NOTREACHED */ 800} 801 802/******************************************************************* 803 * local functions and data 804 */ 805 806/* 807 * We tell the I/O APIC code about all the CPUs we want to receive 808 * interrupts. If we don't want certain CPUs to receive IRQs we 809 * can simply not tell the I/O APIC code about them in this function. 810 * We also do not tell it about the BSP since it tells itself about 811 * the BSP internally to work with UP kernels and on UP machines. 812 */ 813static void 814set_interrupt_apic_ids(void) 815{ 816 u_int i, apic_id; 817 818 for (i = 0; i < MAXCPU; i++) { 819 apic_id = cpu_apic_ids[i]; 820 if (apic_id == -1) 821 continue; 822 if (cpu_info[apic_id].cpu_bsp) 823 continue; 824 if (cpu_info[apic_id].cpu_disabled) 825 continue; 826 827 /* Don't let hyperthreads service interrupts. */ 828 if (hyperthreading_cpus > 1 && 829 apic_id % hyperthreading_cpus != 0) 830 continue; 831 832 intr_add_cpu(i); 833 } 834} 835 836/* 837 * Assign logical CPU IDs to local APICs. 838 */ 839static void 840assign_cpu_ids(void) 841{ 842 u_int i; 843 844 TUNABLE_INT_FETCH("machdep.hyperthreading_allowed", 845 &hyperthreading_allowed); 846 847 /* Check for explicitly disabled CPUs. */ 848 for (i = 0; i <= MAX_APIC_ID; i++) { 849 if (!cpu_info[i].cpu_present || cpu_info[i].cpu_bsp) 850 continue; 851 852 if (hyperthreading_cpus > 1 && i % hyperthreading_cpus != 0) { 853 cpu_info[i].cpu_hyperthread = 1; 854 855 /* 856 * Don't use HT CPU if it has been disabled by a 857 * tunable. 858 */ 859 if (hyperthreading_allowed == 0) { 860 cpu_info[i].cpu_disabled = 1; 861 continue; 862 } 863 } 864 865 /* Don't use this CPU if it has been disabled by a tunable. */ 866 if (resource_disabled("lapic", i)) { 867 cpu_info[i].cpu_disabled = 1; 868 continue; 869 } 870 } 871 872 if (hyperthreading_allowed == 0 && hyperthreading_cpus > 1) { 873 hyperthreading_cpus = 0; 874 cpu_logical = 1; 875 } 876 877 /* 878 * Assign CPU IDs to local APIC IDs and disable any CPUs 879 * beyond MAXCPU. CPU 0 is always assigned to the BSP. 880 * 881 * To minimize confusion for userland, we attempt to number 882 * CPUs such that all threads and cores in a package are 883 * grouped together. For now we assume that the BSP is always 884 * the first thread in a package and just start adding APs 885 * starting with the BSP's APIC ID. 886 */ 887 mp_ncpus = 1; 888 cpu_apic_ids[0] = boot_cpu_id; 889 apic_cpuids[boot_cpu_id] = 0; 890 for (i = boot_cpu_id + 1; i != boot_cpu_id; 891 i == MAX_APIC_ID ? i = 0 : i++) { 892 if (!cpu_info[i].cpu_present || cpu_info[i].cpu_bsp || 893 cpu_info[i].cpu_disabled) 894 continue; 895 896 if (mp_ncpus < MAXCPU) { 897 cpu_apic_ids[mp_ncpus] = i; 898 apic_cpuids[i] = mp_ncpus; 899 mp_ncpus++; 900 } else 901 cpu_info[i].cpu_disabled = 1; 902 } 903 KASSERT(mp_maxid >= mp_ncpus - 1, 904 ("%s: counters out of sync: max %d, count %d", __func__, mp_maxid, 905 mp_ncpus)); 906} 907 908/* 909 * start each AP in our list 910 */ 911static int 912start_all_aps(void) 913{ 914 vm_offset_t va = boot_address + KERNBASE; 915 u_int64_t *pt4, *pt3, *pt2; 916 u_int32_t mpbioswarmvec; 917 int apic_id, cpu, i; 918 u_char mpbiosreason; 919 920 mtx_init(&ap_boot_mtx, "ap boot", NULL, MTX_SPIN); 921 922 /* install the AP 1st level boot code */ 923 pmap_kenter(va, boot_address); 924 pmap_invalidate_page(kernel_pmap, va); 925 bcopy(mptramp_start, (void *)va, bootMP_size); 926 927 /* Locate the page tables, they'll be below the trampoline */ 928 pt4 = (u_int64_t *)(uintptr_t)(mptramp_pagetables + KERNBASE); 929 pt3 = pt4 + (PAGE_SIZE) / sizeof(u_int64_t); 930 pt2 = pt3 + (PAGE_SIZE) / sizeof(u_int64_t); 931 932 /* Create the initial 1GB replicated page tables */ 933 for (i = 0; i < 512; i++) { 934 /* Each slot of the level 4 pages points to the same level 3 page */ 935 pt4[i] = (u_int64_t)(uintptr_t)(mptramp_pagetables + PAGE_SIZE); 936 pt4[i] |= PG_V | PG_RW | PG_U; 937 938 /* Each slot of the level 3 pages points to the same level 2 page */ 939 pt3[i] = (u_int64_t)(uintptr_t)(mptramp_pagetables + (2 * PAGE_SIZE)); 940 pt3[i] |= PG_V | PG_RW | PG_U; 941 942 /* The level 2 page slots are mapped with 2MB pages for 1GB. */ 943 pt2[i] = i * (2 * 1024 * 1024); 944 pt2[i] |= PG_V | PG_RW | PG_PS | PG_U; 945 } 946 947 /* save the current value of the warm-start vector */ 948 mpbioswarmvec = *((u_int32_t *) WARMBOOT_OFF); 949 outb(CMOS_REG, BIOS_RESET); 950 mpbiosreason = inb(CMOS_DATA); 951 952 /* setup a vector to our boot code */ 953 *((volatile u_short *) WARMBOOT_OFF) = WARMBOOT_TARGET; 954 *((volatile u_short *) WARMBOOT_SEG) = (boot_address >> 4); 955 outb(CMOS_REG, BIOS_RESET); 956 outb(CMOS_DATA, BIOS_WARM); /* 'warm-start' */ 957 958 /* start each AP */ 959 for (cpu = 1; cpu < mp_ncpus; cpu++) { 960 apic_id = cpu_apic_ids[cpu]; 961 962 /* allocate and set up an idle stack data page */ 963 bootstacks[cpu] = (void *)kmem_malloc(kernel_arena, 964 KSTACK_PAGES * PAGE_SIZE, M_WAITOK | M_ZERO); 965 doublefault_stack = (char *)kmem_malloc(kernel_arena, 966 PAGE_SIZE, M_WAITOK | M_ZERO); 967 nmi_stack = (char *)kmem_malloc(kernel_arena, PAGE_SIZE, 968 M_WAITOK | M_ZERO); 969 dpcpu = (void *)kmem_malloc(kernel_arena, DPCPU_SIZE, 970 M_WAITOK | M_ZERO); 971 972 bootSTK = (char *)bootstacks[cpu] + KSTACK_PAGES * PAGE_SIZE - 8; 973 bootAP = cpu; 974 975 /* attempt to start the Application Processor */ 976 if (!start_ap(apic_id)) { 977 /* restore the warmstart vector */ 978 *(u_int32_t *) WARMBOOT_OFF = mpbioswarmvec; 979 panic("AP #%d (PHY# %d) failed!", cpu, apic_id); 980 } 981 982 CPU_SET(cpu, &all_cpus); /* record AP in CPU map */ 983 } 984 985 /* restore the warmstart vector */ 986 *(u_int32_t *) WARMBOOT_OFF = mpbioswarmvec; 987 988 outb(CMOS_REG, BIOS_RESET); 989 outb(CMOS_DATA, mpbiosreason); 990 991 /* number of APs actually started */ 992 return mp_naps; 993} 994 995 996/* 997 * This function starts the AP (application processor) identified 998 * by the APIC ID 'physicalCpu'. It does quite a "song and dance" 999 * to accomplish this. This is necessary because of the nuances 1000 * of the different hardware we might encounter. It isn't pretty, 1001 * but it seems to work. 1002 */ 1003static int 1004start_ap(int apic_id) 1005{ 1006 int vector, ms; 1007 int cpus; 1008 1009 /* calculate the vector */ 1010 vector = (boot_address >> 12) & 0xff; 1011 1012 /* used as a watchpoint to signal AP startup */ 1013 cpus = mp_naps; 1014 1015 ipi_startup(apic_id, vector); 1016 1017 /* Wait up to 5 seconds for it to start. */ 1018 for (ms = 0; ms < 5000; ms++) { 1019 if (mp_naps > cpus) 1020 return 1; /* return SUCCESS */ 1021 DELAY(1000); 1022 } 1023 return 0; /* return FAILURE */ 1024} 1025 1026#ifdef COUNT_XINVLTLB_HITS 1027u_int xhits_gbl[MAXCPU]; 1028u_int xhits_pg[MAXCPU]; 1029u_int xhits_rng[MAXCPU]; 1030static SYSCTL_NODE(_debug, OID_AUTO, xhits, CTLFLAG_RW, 0, ""); 1031SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, global, CTLFLAG_RW, &xhits_gbl, 1032 sizeof(xhits_gbl), "IU", ""); 1033SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, page, CTLFLAG_RW, &xhits_pg, 1034 sizeof(xhits_pg), "IU", ""); 1035SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, range, CTLFLAG_RW, &xhits_rng, 1036 sizeof(xhits_rng), "IU", ""); 1037 1038u_int ipi_global; 1039u_int ipi_page; 1040u_int ipi_range; 1041u_int ipi_range_size; 1042SYSCTL_UINT(_debug_xhits, OID_AUTO, ipi_global, CTLFLAG_RW, &ipi_global, 0, ""); 1043SYSCTL_UINT(_debug_xhits, OID_AUTO, ipi_page, CTLFLAG_RW, &ipi_page, 0, ""); 1044SYSCTL_UINT(_debug_xhits, OID_AUTO, ipi_range, CTLFLAG_RW, &ipi_range, 0, ""); 1045SYSCTL_UINT(_debug_xhits, OID_AUTO, ipi_range_size, CTLFLAG_RW, 1046 &ipi_range_size, 0, ""); 1047 1048u_int ipi_masked_global; 1049u_int ipi_masked_page; 1050u_int ipi_masked_range; 1051u_int ipi_masked_range_size; 1052SYSCTL_UINT(_debug_xhits, OID_AUTO, ipi_masked_global, CTLFLAG_RW, 1053 &ipi_masked_global, 0, ""); 1054SYSCTL_UINT(_debug_xhits, OID_AUTO, ipi_masked_page, CTLFLAG_RW, 1055 &ipi_masked_page, 0, ""); 1056SYSCTL_UINT(_debug_xhits, OID_AUTO, ipi_masked_range, CTLFLAG_RW, 1057 &ipi_masked_range, 0, ""); 1058SYSCTL_UINT(_debug_xhits, OID_AUTO, ipi_masked_range_size, CTLFLAG_RW, 1059 &ipi_masked_range_size, 0, ""); 1060#endif /* COUNT_XINVLTLB_HITS */ 1061 1062/* 1063 * Init and startup IPI. 1064 */ 1065void 1066ipi_startup(int apic_id, int vector) 1067{ 1068 1069 /* 1070 * first we do an INIT IPI: this INIT IPI might be run, resetting 1071 * and running the target CPU. OR this INIT IPI might be latched (P5 1072 * bug), CPU waiting for STARTUP IPI. OR this INIT IPI might be 1073 * ignored. 1074 */ 1075 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE | 1076 APIC_LEVEL_ASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_INIT, apic_id); 1077 lapic_ipi_wait(-1); 1078 DELAY(10000); /* wait ~10mS */ 1079 1080 /* 1081 * next we do a STARTUP IPI: the previous INIT IPI might still be 1082 * latched, (P5 bug) this 1st STARTUP would then terminate 1083 * immediately, and the previously started INIT IPI would continue. OR 1084 * the previous INIT IPI has already run. and this STARTUP IPI will 1085 * run. OR the previous INIT IPI was ignored. and this STARTUP IPI 1086 * will run. 1087 */ 1088 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE | 1089 APIC_LEVEL_DEASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_STARTUP | 1090 vector, apic_id); 1091 lapic_ipi_wait(-1); 1092 DELAY(200); /* wait ~200uS */ 1093 1094 /* 1095 * finally we do a 2nd STARTUP IPI: this 2nd STARTUP IPI should run IF 1096 * the previous STARTUP IPI was cancelled by a latched INIT IPI. OR 1097 * this STARTUP IPI will be ignored, as only ONE STARTUP IPI is 1098 * recognized after hardware RESET or INIT IPI. 1099 */ 1100 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE | 1101 APIC_LEVEL_DEASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_STARTUP | 1102 vector, apic_id); 1103 lapic_ipi_wait(-1); 1104 DELAY(200); /* wait ~200uS */ 1105} 1106 1107/* 1108 * Send an IPI to specified CPU handling the bitmap logic. 1109 */ 1110static void 1111ipi_send_cpu(int cpu, u_int ipi) 1112{ 1113 u_int bitmap, old_pending, new_pending; 1114 1115 KASSERT(cpu_apic_ids[cpu] != -1, ("IPI to non-existent CPU %d", cpu)); 1116 1117 if (IPI_IS_BITMAPED(ipi)) { 1118 bitmap = 1 << ipi; 1119 ipi = IPI_BITMAP_VECTOR; 1120 do { 1121 old_pending = cpu_ipi_pending[cpu]; 1122 new_pending = old_pending | bitmap; 1123 } while (!atomic_cmpset_int(&cpu_ipi_pending[cpu], 1124 old_pending, new_pending)); 1125 if (old_pending) 1126 return; 1127 } 1128 cpu_ops.ipi_vectored(ipi, cpu_apic_ids[cpu]); 1129} 1130 1131/* 1132 * Flush the TLB on all other CPU's 1133 */ 1134static void 1135smp_tlb_shootdown(u_int vector, pmap_t pmap, vm_offset_t addr1, 1136 vm_offset_t addr2) 1137{ 1138 u_int ncpu; 1139 1140 ncpu = mp_ncpus - 1; /* does not shootdown self */ 1141 if (ncpu < 1) 1142 return; /* no other cpus */ 1143 if (!(read_rflags() & PSL_I)) 1144 panic("%s: interrupts disabled", __func__); 1145 mtx_lock_spin(&smp_ipi_mtx); 1146 smp_tlb_invpcid.addr = addr1; 1147 if (pmap == NULL) { 1148 smp_tlb_invpcid.pcid = 0; 1149 } else { 1150 smp_tlb_invpcid.pcid = pmap->pm_pcid; 1151 pcid_cr3 = pmap->pm_cr3; 1152 } 1153 smp_tlb_addr2 = addr2; 1154 smp_tlb_pmap = pmap; 1155 atomic_store_rel_int(&smp_tlb_wait, 0); 1156 ipi_all_but_self(vector); 1157 while (smp_tlb_wait < ncpu) 1158 ia32_pause(); 1159 mtx_unlock_spin(&smp_ipi_mtx); 1160} 1161 1162static void 1163smp_targeted_tlb_shootdown(cpuset_t mask, u_int vector, pmap_t pmap, 1164 vm_offset_t addr1, vm_offset_t addr2) 1165{ 1166 int cpu, ncpu, othercpus; 1167 1168 othercpus = mp_ncpus - 1; 1169 if (CPU_ISFULLSET(&mask)) { 1170 if (othercpus < 1) 1171 return; 1172 } else { 1173 CPU_CLR(PCPU_GET(cpuid), &mask); 1174 if (CPU_EMPTY(&mask)) 1175 return; 1176 } 1177 if (!(read_rflags() & PSL_I)) 1178 panic("%s: interrupts disabled", __func__); 1179 mtx_lock_spin(&smp_ipi_mtx); 1180 smp_tlb_invpcid.addr = addr1; 1181 if (pmap == NULL) { 1182 smp_tlb_invpcid.pcid = 0; 1183 } else { 1184 smp_tlb_invpcid.pcid = pmap->pm_pcid; 1185 pcid_cr3 = pmap->pm_cr3; 1186 } 1187 smp_tlb_addr2 = addr2; 1188 smp_tlb_pmap = pmap; 1189 atomic_store_rel_int(&smp_tlb_wait, 0); 1190 if (CPU_ISFULLSET(&mask)) { 1191 ncpu = othercpus; 1192 ipi_all_but_self(vector); 1193 } else { 1194 ncpu = 0; 1195 while ((cpu = CPU_FFS(&mask)) != 0) { 1196 cpu--; 1197 CPU_CLR(cpu, &mask); 1198 CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__, 1199 cpu, vector); 1200 ipi_send_cpu(cpu, vector); 1201 ncpu++; 1202 } 1203 } 1204 while (smp_tlb_wait < ncpu) 1205 ia32_pause(); 1206 mtx_unlock_spin(&smp_ipi_mtx); 1207} 1208 1209void 1210smp_cache_flush(void) 1211{ 1212 1213 if (smp_started) 1214 smp_tlb_shootdown(IPI_INVLCACHE, NULL, 0, 0); 1215} 1216 1217void 1218smp_invltlb(pmap_t pmap) 1219{ 1220 1221 if (smp_started) { 1222 smp_tlb_shootdown(IPI_INVLTLB, pmap, 0, 0); 1223#ifdef COUNT_XINVLTLB_HITS 1224 ipi_global++; 1225#endif 1226 } 1227} 1228 1229void 1230smp_invlpg(pmap_t pmap, vm_offset_t addr) 1231{ 1232 1233 if (smp_started) { 1234 smp_tlb_shootdown(IPI_INVLPG, pmap, addr, 0); 1235#ifdef COUNT_XINVLTLB_HITS 1236 ipi_page++; 1237#endif 1238 } 1239} 1240 1241void 1242smp_invlpg_range(pmap_t pmap, vm_offset_t addr1, vm_offset_t addr2) 1243{ 1244 1245 if (smp_started) { 1246 smp_tlb_shootdown(IPI_INVLRNG, pmap, addr1, addr2); 1247#ifdef COUNT_XINVLTLB_HITS 1248 ipi_range++; 1249 ipi_range_size += (addr2 - addr1) / PAGE_SIZE; 1250#endif 1251 } 1252} 1253 1254void 1255smp_masked_invltlb(cpuset_t mask, pmap_t pmap) 1256{ 1257 1258 if (smp_started) { 1259 smp_targeted_tlb_shootdown(mask, IPI_INVLTLB, pmap, 0, 0); 1260#ifdef COUNT_XINVLTLB_HITS 1261 ipi_masked_global++; 1262#endif 1263 } 1264} 1265 1266void 1267smp_masked_invlpg(cpuset_t mask, pmap_t pmap, vm_offset_t addr) 1268{ 1269 1270 if (smp_started) { 1271 smp_targeted_tlb_shootdown(mask, IPI_INVLPG, pmap, addr, 0); 1272#ifdef COUNT_XINVLTLB_HITS 1273 ipi_masked_page++; 1274#endif 1275 } 1276} 1277 1278void 1279smp_masked_invlpg_range(cpuset_t mask, pmap_t pmap, vm_offset_t addr1, 1280 vm_offset_t addr2) 1281{ 1282 1283 if (smp_started) { 1284 smp_targeted_tlb_shootdown(mask, IPI_INVLRNG, pmap, addr1, 1285 addr2); 1286#ifdef COUNT_XINVLTLB_HITS 1287 ipi_masked_range++; 1288 ipi_masked_range_size += (addr2 - addr1) / PAGE_SIZE; 1289#endif 1290 } 1291} 1292 1293void 1294ipi_bitmap_handler(struct trapframe frame) 1295{ 1296 struct trapframe *oldframe; 1297 struct thread *td; 1298 int cpu = PCPU_GET(cpuid); 1299 u_int ipi_bitmap; 1300 1301 critical_enter(); 1302 td = curthread; 1303 td->td_intr_nesting_level++; 1304 oldframe = td->td_intr_frame; 1305 td->td_intr_frame = &frame; 1306 ipi_bitmap = atomic_readandclear_int(&cpu_ipi_pending[cpu]); 1307 if (ipi_bitmap & (1 << IPI_PREEMPT)) { 1308#ifdef COUNT_IPIS 1309 (*ipi_preempt_counts[cpu])++; 1310#endif 1311 sched_preempt(td); 1312 } 1313 if (ipi_bitmap & (1 << IPI_AST)) { 1314#ifdef COUNT_IPIS 1315 (*ipi_ast_counts[cpu])++; 1316#endif 1317 /* Nothing to do for AST */ 1318 } 1319 if (ipi_bitmap & (1 << IPI_HARDCLOCK)) { 1320#ifdef COUNT_IPIS 1321 (*ipi_hardclock_counts[cpu])++; 1322#endif 1323 hardclockintr(); 1324 } 1325 td->td_intr_frame = oldframe; 1326 td->td_intr_nesting_level--; 1327 critical_exit(); 1328} 1329 1330/* 1331 * send an IPI to a set of cpus. 1332 */ 1333void 1334ipi_selected(cpuset_t cpus, u_int ipi) 1335{ 1336 int cpu; 1337 1338 /* 1339 * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit 1340 * of help in order to understand what is the source. 1341 * Set the mask of receiving CPUs for this purpose. 1342 */ 1343 if (ipi == IPI_STOP_HARD) 1344 CPU_OR_ATOMIC(&ipi_nmi_pending, &cpus); 1345 1346 while ((cpu = CPU_FFS(&cpus)) != 0) { 1347 cpu--; 1348 CPU_CLR(cpu, &cpus); 1349 CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__, cpu, ipi); 1350 ipi_send_cpu(cpu, ipi); 1351 } 1352} 1353 1354/* 1355 * send an IPI to a specific CPU. 1356 */ 1357void 1358ipi_cpu(int cpu, u_int ipi) 1359{ 1360 1361 /* 1362 * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit 1363 * of help in order to understand what is the source. 1364 * Set the mask of receiving CPUs for this purpose. 1365 */ 1366 if (ipi == IPI_STOP_HARD) 1367 CPU_SET_ATOMIC(cpu, &ipi_nmi_pending); 1368 1369 CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__, cpu, ipi); 1370 ipi_send_cpu(cpu, ipi); 1371} 1372 1373/* 1374 * send an IPI to all CPUs EXCEPT myself 1375 */ 1376void 1377ipi_all_but_self(u_int ipi) 1378{ 1379 cpuset_t other_cpus; 1380 1381 other_cpus = all_cpus; 1382 CPU_CLR(PCPU_GET(cpuid), &other_cpus); 1383 1384 if (IPI_IS_BITMAPED(ipi)) { 1385 ipi_selected(other_cpus, ipi); 1386 return; 1387 } 1388 1389 /* 1390 * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit 1391 * of help in order to understand what is the source. 1392 * Set the mask of receiving CPUs for this purpose. 1393 */ 1394 if (ipi == IPI_STOP_HARD) 1395 CPU_OR_ATOMIC(&ipi_nmi_pending, &other_cpus); 1396 1397 CTR2(KTR_SMP, "%s: ipi: %x", __func__, ipi); 1398 cpu_ops.ipi_vectored(ipi, APIC_IPI_DEST_OTHERS); 1399} 1400 1401int 1402ipi_nmi_handler() 1403{ 1404 u_int cpuid; 1405 1406 /* 1407 * As long as there is not a simple way to know about a NMI's 1408 * source, if the bitmask for the current CPU is present in 1409 * the global pending bitword an IPI_STOP_HARD has been issued 1410 * and should be handled. 1411 */ 1412 cpuid = PCPU_GET(cpuid); 1413 if (!CPU_ISSET(cpuid, &ipi_nmi_pending)) 1414 return (1); 1415 1416 CPU_CLR_ATOMIC(cpuid, &ipi_nmi_pending); 1417 cpustop_handler(); 1418 return (0); 1419} 1420 1421/* 1422 * Handle an IPI_STOP by saving our current context and spinning until we 1423 * are resumed. 1424 */ 1425void 1426cpustop_handler(void) 1427{ 1428 u_int cpu; 1429 1430 cpu = PCPU_GET(cpuid); 1431 1432 savectx(&stoppcbs[cpu]); 1433 1434 /* Indicate that we are stopped */ 1435 CPU_SET_ATOMIC(cpu, &stopped_cpus); 1436 1437 /* Wait for restart */ 1438 while (!CPU_ISSET(cpu, &started_cpus)) 1439 ia32_pause(); 1440 1441 CPU_CLR_ATOMIC(cpu, &started_cpus); 1442 CPU_CLR_ATOMIC(cpu, &stopped_cpus); 1443 1444#ifdef DDB 1445 amd64_db_resume_dbreg(); 1446#endif 1447 1448 if (cpu == 0 && cpustop_restartfunc != NULL) { 1449 cpustop_restartfunc(); 1450 cpustop_restartfunc = NULL; 1451 } 1452} 1453 1454/* 1455 * Handle an IPI_SUSPEND by saving our current context and spinning until we 1456 * are resumed. 1457 */ 1458void 1459cpususpend_handler(void) 1460{ 1461 u_int cpu; 1462 1463 mtx_assert(&smp_ipi_mtx, MA_NOTOWNED); 1464 1465 cpu = PCPU_GET(cpuid); 1466 if (savectx(&susppcbs[cpu]->sp_pcb)) { 1467 fpususpend(susppcbs[cpu]->sp_fpususpend); 1468 wbinvd(); 1469 CPU_SET_ATOMIC(cpu, &suspended_cpus); 1470 } else { 1471 fpuresume(susppcbs[cpu]->sp_fpususpend); 1472 pmap_init_pat(); 1473 initializecpu(); 1474 PCPU_SET(switchtime, 0); 1475 PCPU_SET(switchticks, ticks); 1476 1477 /* Indicate that we are resumed */ 1478 CPU_CLR_ATOMIC(cpu, &suspended_cpus); 1479 } 1480 1481 /* Wait for resume */ 1482 while (!CPU_ISSET(cpu, &started_cpus)) 1483 ia32_pause(); 1484 1485 if (cpu_ops.cpu_resume) 1486 cpu_ops.cpu_resume(); 1487 if (vmm_resume_p) 1488 vmm_resume_p(); 1489 1490 /* Resume MCA and local APIC */ 1491 mca_resume(); 1492 lapic_setup(0); 1493 1494 CPU_CLR_ATOMIC(cpu, &started_cpus); 1495 /* Indicate that we are resumed */ 1496 CPU_CLR_ATOMIC(cpu, &suspended_cpus); 1497} 1498 1499/* 1500 * Handlers for TLB related IPIs 1501 */ 1502void 1503invltlb_handler(void) 1504{ 1505#ifdef COUNT_XINVLTLB_HITS 1506 xhits_gbl[PCPU_GET(cpuid)]++; 1507#endif /* COUNT_XINVLTLB_HITS */ 1508#ifdef COUNT_IPIS 1509 (*ipi_invltlb_counts[PCPU_GET(cpuid)])++; 1510#endif /* COUNT_IPIS */ 1511 1512 invltlb(); 1513 atomic_add_int(&smp_tlb_wait, 1); 1514} 1515 1516void 1517invltlb_pcid_handler(void) 1518{ 1519 uint64_t cr3; 1520 u_int cpuid; 1521#ifdef COUNT_XINVLTLB_HITS 1522 xhits_gbl[PCPU_GET(cpuid)]++; 1523#endif /* COUNT_XINVLTLB_HITS */ 1524#ifdef COUNT_IPIS 1525 (*ipi_invltlb_counts[PCPU_GET(cpuid)])++; 1526#endif /* COUNT_IPIS */ 1527 1528 if (smp_tlb_invpcid.pcid != (uint64_t)-1 && 1529 smp_tlb_invpcid.pcid != 0) { 1530 if (invpcid_works) { 1531 invpcid(&smp_tlb_invpcid, INVPCID_CTX); 1532 } else { 1533 /* Otherwise reload %cr3 twice. */ 1534 cr3 = rcr3(); 1535 if (cr3 != pcid_cr3) { 1536 load_cr3(pcid_cr3); 1537 cr3 |= CR3_PCID_SAVE; 1538 } 1539 load_cr3(cr3); 1540 } 1541 } else { 1542 invltlb_globpcid(); 1543 } 1544 if (smp_tlb_pmap != NULL) { 1545 cpuid = PCPU_GET(cpuid); 1546 if (!CPU_ISSET(cpuid, &smp_tlb_pmap->pm_active)) 1547 CPU_CLR_ATOMIC(cpuid, &smp_tlb_pmap->pm_save); 1548 } 1549 1550 atomic_add_int(&smp_tlb_wait, 1); 1551} 1552 1553void 1554invlpg_handler(void) 1555{ 1556#ifdef COUNT_XINVLTLB_HITS 1557 xhits_pg[PCPU_GET(cpuid)]++; 1558#endif /* COUNT_XINVLTLB_HITS */ 1559#ifdef COUNT_IPIS 1560 (*ipi_invlpg_counts[PCPU_GET(cpuid)])++; 1561#endif /* COUNT_IPIS */ 1562 1563 invlpg(smp_tlb_invpcid.addr); 1564 atomic_add_int(&smp_tlb_wait, 1); 1565} 1566 1567void 1568invlpg_pcid_handler(void) 1569{ 1570 uint64_t cr3; 1571#ifdef COUNT_XINVLTLB_HITS 1572 xhits_pg[PCPU_GET(cpuid)]++; 1573#endif /* COUNT_XINVLTLB_HITS */ 1574#ifdef COUNT_IPIS 1575 (*ipi_invlpg_counts[PCPU_GET(cpuid)])++; 1576#endif /* COUNT_IPIS */ 1577 1578 if (smp_tlb_invpcid.pcid == (uint64_t)-1) { 1579 invltlb_globpcid(); 1580 } else if (smp_tlb_invpcid.pcid == 0) { 1581 invlpg(smp_tlb_invpcid.addr); 1582 } else if (invpcid_works) { 1583 invpcid(&smp_tlb_invpcid, INVPCID_ADDR); 1584 } else { 1585 /* 1586 * PCID supported, but INVPCID is not. 1587 * Temporarily switch to the target address 1588 * space and do INVLPG. 1589 */ 1590 cr3 = rcr3(); 1591 if (cr3 != pcid_cr3) 1592 load_cr3(pcid_cr3 | CR3_PCID_SAVE); 1593 invlpg(smp_tlb_invpcid.addr); 1594 load_cr3(cr3 | CR3_PCID_SAVE); 1595 } 1596 1597 atomic_add_int(&smp_tlb_wait, 1); 1598} 1599 1600static inline void 1601invlpg_range(vm_offset_t start, vm_offset_t end) 1602{ 1603 1604 do { 1605 invlpg(start); 1606 start += PAGE_SIZE; 1607 } while (start < end); 1608} 1609 1610void 1611invlrng_handler(void) 1612{ 1613 struct invpcid_descr d; 1614 vm_offset_t addr; 1615 uint64_t cr3; 1616 u_int cpuid; 1617#ifdef COUNT_XINVLTLB_HITS 1618 xhits_rng[PCPU_GET(cpuid)]++; 1619#endif /* COUNT_XINVLTLB_HITS */ 1620#ifdef COUNT_IPIS 1621 (*ipi_invlrng_counts[PCPU_GET(cpuid)])++; 1622#endif /* COUNT_IPIS */ 1623 1624 addr = smp_tlb_invpcid.addr; 1625 if (pmap_pcid_enabled) { 1626 if (smp_tlb_invpcid.pcid == 0) { 1627 /* 1628 * kernel pmap - use invlpg to invalidate 1629 * global mapping. 1630 */ 1631 invlpg_range(addr, smp_tlb_addr2); 1632 } else if (smp_tlb_invpcid.pcid == (uint64_t)-1) { 1633 invltlb_globpcid(); 1634 if (smp_tlb_pmap != NULL) { 1635 cpuid = PCPU_GET(cpuid); 1636 if (!CPU_ISSET(cpuid, &smp_tlb_pmap->pm_active)) 1637 CPU_CLR_ATOMIC(cpuid, 1638 &smp_tlb_pmap->pm_save); 1639 } 1640 } else if (invpcid_works) { 1641 d = smp_tlb_invpcid; 1642 do { 1643 invpcid(&d, INVPCID_ADDR); 1644 d.addr += PAGE_SIZE; 1645 } while (d.addr <= smp_tlb_addr2); 1646 } else { 1647 cr3 = rcr3(); 1648 if (cr3 != pcid_cr3) 1649 load_cr3(pcid_cr3 | CR3_PCID_SAVE); 1650 invlpg_range(addr, smp_tlb_addr2); 1651 load_cr3(cr3 | CR3_PCID_SAVE); 1652 } 1653 } else { 1654 invlpg_range(addr, smp_tlb_addr2); 1655 } 1656 1657 atomic_add_int(&smp_tlb_wait, 1); 1658} 1659 1660void 1661invlcache_handler(void) 1662{ 1663#ifdef COUNT_IPIS 1664 (*ipi_invlcache_counts[PCPU_GET(cpuid)])++; 1665#endif /* COUNT_IPIS */ 1666 1667 wbinvd(); 1668 atomic_add_int(&smp_tlb_wait, 1); 1669} 1670 1671/* 1672 * This is called once the rest of the system is up and running and we're 1673 * ready to let the AP's out of the pen. 1674 */ 1675static void 1676release_aps(void *dummy __unused) 1677{ 1678 1679 if (mp_ncpus == 1) 1680 return; 1681 atomic_store_rel_int(&aps_ready, 1); 1682 while (smp_started == 0) 1683 ia32_pause(); 1684} 1685SYSINIT(start_aps, SI_SUB_SMP, SI_ORDER_FIRST, release_aps, NULL); 1686 1687#ifdef COUNT_IPIS 1688/* 1689 * Setup interrupt counters for IPI handlers. 1690 */ 1691static void 1692mp_ipi_intrcnt(void *dummy) 1693{ 1694 char buf[64]; 1695 int i; 1696 1697 CPU_FOREACH(i) { 1698 snprintf(buf, sizeof(buf), "cpu%d:invltlb", i); 1699 intrcnt_add(buf, &ipi_invltlb_counts[i]); 1700 snprintf(buf, sizeof(buf), "cpu%d:invlrng", i); 1701 intrcnt_add(buf, &ipi_invlrng_counts[i]); 1702 snprintf(buf, sizeof(buf), "cpu%d:invlpg", i); 1703 intrcnt_add(buf, &ipi_invlpg_counts[i]); 1704 snprintf(buf, sizeof(buf), "cpu%d:invlcache", i); 1705 intrcnt_add(buf, &ipi_invlcache_counts[i]); 1706 snprintf(buf, sizeof(buf), "cpu%d:preempt", i); 1707 intrcnt_add(buf, &ipi_preempt_counts[i]); 1708 snprintf(buf, sizeof(buf), "cpu%d:ast", i); 1709 intrcnt_add(buf, &ipi_ast_counts[i]); 1710 snprintf(buf, sizeof(buf), "cpu%d:rendezvous", i); 1711 intrcnt_add(buf, &ipi_rendezvous_counts[i]); 1712 snprintf(buf, sizeof(buf), "cpu%d:hardclock", i); 1713 intrcnt_add(buf, &ipi_hardclock_counts[i]); 1714 } 1715} 1716SYSINIT(mp_ipi_intrcnt, SI_SUB_INTR, SI_ORDER_MIDDLE, mp_ipi_intrcnt, NULL); 1717#endif 1718 1719