subr_smp.c revision 367457
1/*- 2 * Copyright (c) 2001, John Baldwin <jhb@FreeBSD.org>. 3 * 4 * Redistribution and use in source and binary forms, with or without 5 * modification, are permitted provided that the following conditions 6 * are met: 7 * 1. Redistributions of source code must retain the above copyright 8 * notice, this list of conditions and the following disclaimer. 9 * 2. Redistributions in binary form must reproduce the above copyright 10 * notice, this list of conditions and the following disclaimer in the 11 * documentation and/or other materials provided with the distribution. 12 * 13 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 14 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 15 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 16 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 17 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 18 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 19 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 20 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 21 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 22 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 23 * SUCH DAMAGE. 24 */ 25 26/* 27 * This module holds the global variables and machine independent functions 28 * used for the kernel SMP support. 29 */ 30 31#include <sys/cdefs.h> 32__FBSDID("$FreeBSD: stable/11/sys/kern/subr_smp.c 367457 2020-11-07 18:10:59Z dim $"); 33 34#include <sys/param.h> 35#include <sys/systm.h> 36#include <sys/kernel.h> 37#include <sys/ktr.h> 38#include <sys/proc.h> 39#include <sys/bus.h> 40#include <sys/lock.h> 41#include <sys/malloc.h> 42#include <sys/mutex.h> 43#include <sys/pcpu.h> 44#include <sys/sched.h> 45#include <sys/smp.h> 46#include <sys/sysctl.h> 47 48#include <machine/cpu.h> 49#include <machine/smp.h> 50 51#include "opt_sched.h" 52 53#ifdef SMP 54MALLOC_DEFINE(M_TOPO, "toponodes", "SMP topology data"); 55 56volatile cpuset_t stopped_cpus; 57volatile cpuset_t started_cpus; 58volatile cpuset_t suspended_cpus; 59cpuset_t hlt_cpus_mask; 60cpuset_t logical_cpus_mask; 61 62void (*cpustop_restartfunc)(void); 63#endif 64 65static int sysctl_kern_smp_active(SYSCTL_HANDLER_ARGS); 66 67/* This is used in modules that need to work in both SMP and UP. */ 68cpuset_t all_cpus; 69 70int mp_ncpus; 71/* export this for libkvm consumers. */ 72int mp_maxcpus = MAXCPU; 73 74volatile int smp_started; 75u_int mp_maxid; 76 77static SYSCTL_NODE(_kern, OID_AUTO, smp, CTLFLAG_RD|CTLFLAG_CAPRD, NULL, 78 "Kernel SMP"); 79 80SYSCTL_INT(_kern_smp, OID_AUTO, maxid, CTLFLAG_RD|CTLFLAG_CAPRD, &mp_maxid, 0, 81 "Max CPU ID."); 82 83SYSCTL_INT(_kern_smp, OID_AUTO, maxcpus, CTLFLAG_RD|CTLFLAG_CAPRD, &mp_maxcpus, 84 0, "Max number of CPUs that the system was compiled for."); 85 86SYSCTL_PROC(_kern_smp, OID_AUTO, active, CTLFLAG_RD | CTLTYPE_INT, NULL, 0, 87 sysctl_kern_smp_active, "I", "Indicates system is running in SMP mode"); 88 89int smp_disabled = 0; /* has smp been disabled? */ 90SYSCTL_INT(_kern_smp, OID_AUTO, disabled, CTLFLAG_RDTUN|CTLFLAG_CAPRD, 91 &smp_disabled, 0, "SMP has been disabled from the loader"); 92 93int smp_cpus = 1; /* how many cpu's running */ 94SYSCTL_INT(_kern_smp, OID_AUTO, cpus, CTLFLAG_RD|CTLFLAG_CAPRD, &smp_cpus, 0, 95 "Number of CPUs online"); 96 97int smp_topology = 0; /* Which topology we're using. */ 98SYSCTL_INT(_kern_smp, OID_AUTO, topology, CTLFLAG_RDTUN, &smp_topology, 0, 99 "Topology override setting; 0 is default provided by hardware."); 100 101#ifdef SMP 102/* Enable forwarding of a signal to a process running on a different CPU */ 103static int forward_signal_enabled = 1; 104SYSCTL_INT(_kern_smp, OID_AUTO, forward_signal_enabled, CTLFLAG_RW, 105 &forward_signal_enabled, 0, 106 "Forwarding of a signal to a process on a different CPU"); 107 108/* Variables needed for SMP rendezvous. */ 109static volatile int smp_rv_ncpus; 110static void (*volatile smp_rv_setup_func)(void *arg); 111static void (*volatile smp_rv_action_func)(void *arg); 112static void (*volatile smp_rv_teardown_func)(void *arg); 113static void *volatile smp_rv_func_arg; 114static volatile int smp_rv_waiters[4]; 115 116/* 117 * Shared mutex to restrict busywaits between smp_rendezvous() and 118 * smp(_targeted)_tlb_shootdown(). A deadlock occurs if both of these 119 * functions trigger at once and cause multiple CPUs to busywait with 120 * interrupts disabled. 121 */ 122struct mtx smp_ipi_mtx; 123 124/* 125 * Let the MD SMP code initialize mp_maxid very early if it can. 126 */ 127static void 128mp_setmaxid(void *dummy) 129{ 130 131 cpu_mp_setmaxid(); 132 133 KASSERT(mp_ncpus >= 1, ("%s: CPU count < 1", __func__)); 134 KASSERT(mp_ncpus > 1 || mp_maxid == 0, 135 ("%s: one CPU but mp_maxid is not zero", __func__)); 136 KASSERT(mp_maxid >= mp_ncpus - 1, 137 ("%s: counters out of sync: max %d, count %d", __func__, 138 mp_maxid, mp_ncpus)); 139} 140SYSINIT(cpu_mp_setmaxid, SI_SUB_TUNABLES, SI_ORDER_FIRST, mp_setmaxid, NULL); 141 142/* 143 * Call the MD SMP initialization code. 144 */ 145static void 146mp_start(void *dummy) 147{ 148 149 mtx_init(&smp_ipi_mtx, "smp rendezvous", NULL, MTX_SPIN); 150 151 /* Probe for MP hardware. */ 152 if (smp_disabled != 0 || cpu_mp_probe() == 0) { 153 mp_ncpus = 1; 154 CPU_SETOF(PCPU_GET(cpuid), &all_cpus); 155 return; 156 } 157 158 cpu_mp_start(); 159 printf("FreeBSD/SMP: Multiprocessor System Detected: %d CPUs\n", 160 mp_ncpus); 161 cpu_mp_announce(); 162} 163SYSINIT(cpu_mp, SI_SUB_CPU, SI_ORDER_THIRD, mp_start, NULL); 164 165void 166forward_signal(struct thread *td) 167{ 168 int id; 169 170 /* 171 * signotify() has already set TDF_ASTPENDING and TDF_NEEDSIGCHECK on 172 * this thread, so all we need to do is poke it if it is currently 173 * executing so that it executes ast(). 174 */ 175 THREAD_LOCK_ASSERT(td, MA_OWNED); 176 KASSERT(TD_IS_RUNNING(td), 177 ("forward_signal: thread is not TDS_RUNNING")); 178 179 CTR1(KTR_SMP, "forward_signal(%p)", td->td_proc); 180 181 if (!smp_started || cold || panicstr) 182 return; 183 if (!forward_signal_enabled) 184 return; 185 186 /* No need to IPI ourself. */ 187 if (td == curthread) 188 return; 189 190 id = td->td_oncpu; 191 if (id == NOCPU) 192 return; 193 ipi_cpu(id, IPI_AST); 194} 195 196/* 197 * When called the executing CPU will send an IPI to all other CPUs 198 * requesting that they halt execution. 199 * 200 * Usually (but not necessarily) called with 'other_cpus' as its arg. 201 * 202 * - Signals all CPUs in map to stop. 203 * - Waits for each to stop. 204 * 205 * Returns: 206 * -1: error 207 * 0: NA 208 * 1: ok 209 * 210 */ 211#if defined(__amd64__) || defined(__i386__) 212#define X86 1 213#else 214#define X86 0 215#endif 216static int 217generic_stop_cpus(cpuset_t map, u_int type) 218{ 219#ifdef KTR 220 char cpusetbuf[CPUSETBUFSIZ]; 221#endif 222 static volatile u_int stopping_cpu = NOCPU; 223 int i; 224 volatile cpuset_t *cpus; 225 226 KASSERT( 227 type == IPI_STOP || type == IPI_STOP_HARD 228#if X86 229 || type == IPI_SUSPEND 230#endif 231 , ("%s: invalid stop type", __func__)); 232 233 if (!smp_started) 234 return (0); 235 236 CTR2(KTR_SMP, "stop_cpus(%s) with %u type", 237 cpusetobj_strprint(cpusetbuf, &map), type); 238 239#if X86 240 /* 241 * When suspending, ensure there are are no IPIs in progress. 242 * IPIs that have been issued, but not yet delivered (e.g. 243 * not pending on a vCPU when running under virtualization) 244 * will be lost, violating FreeBSD's assumption of reliable 245 * IPI delivery. 246 */ 247 if (type == IPI_SUSPEND) 248 mtx_lock_spin(&smp_ipi_mtx); 249#endif 250 251#if X86 252 if (!nmi_is_broadcast || nmi_kdb_lock == 0) { 253#endif 254 if (stopping_cpu != PCPU_GET(cpuid)) 255 while (atomic_cmpset_int(&stopping_cpu, NOCPU, 256 PCPU_GET(cpuid)) == 0) 257 while (stopping_cpu != NOCPU) 258 cpu_spinwait(); /* spin */ 259 260 /* send the stop IPI to all CPUs in map */ 261 ipi_selected(map, type); 262#if X86 263 } 264#endif 265 266#if X86 267 if (type == IPI_SUSPEND) 268 cpus = &suspended_cpus; 269 else 270#endif 271 cpus = &stopped_cpus; 272 273 i = 0; 274 while (!CPU_SUBSET(cpus, &map)) { 275 /* spin */ 276 cpu_spinwait(); 277 i++; 278 if (i == 100000000) { 279 printf("timeout stopping cpus\n"); 280 break; 281 } 282 } 283 284#if X86 285 if (type == IPI_SUSPEND) 286 mtx_unlock_spin(&smp_ipi_mtx); 287#endif 288 289 stopping_cpu = NOCPU; 290 return (1); 291} 292 293int 294stop_cpus(cpuset_t map) 295{ 296 297 return (generic_stop_cpus(map, IPI_STOP)); 298} 299 300int 301stop_cpus_hard(cpuset_t map) 302{ 303 304 return (generic_stop_cpus(map, IPI_STOP_HARD)); 305} 306 307#if X86 308int 309suspend_cpus(cpuset_t map) 310{ 311 312 return (generic_stop_cpus(map, IPI_SUSPEND)); 313} 314#endif 315 316/* 317 * Called by a CPU to restart stopped CPUs. 318 * 319 * Usually (but not necessarily) called with 'stopped_cpus' as its arg. 320 * 321 * - Signals all CPUs in map to restart. 322 * - Waits for each to restart. 323 * 324 * Returns: 325 * -1: error 326 * 0: NA 327 * 1: ok 328 */ 329static int 330generic_restart_cpus(cpuset_t map, u_int type) 331{ 332#ifdef KTR 333 char cpusetbuf[CPUSETBUFSIZ]; 334#endif 335 volatile cpuset_t *cpus; 336 337 KASSERT(type == IPI_STOP || type == IPI_STOP_HARD 338#if X86 339 || type == IPI_SUSPEND 340#endif 341 , ("%s: invalid stop type", __func__)); 342 343 if (!smp_started) 344 return (0); 345 346 CTR1(KTR_SMP, "restart_cpus(%s)", cpusetobj_strprint(cpusetbuf, &map)); 347 348#if X86 349 if (type == IPI_SUSPEND) 350 cpus = &resuming_cpus; 351 else 352#endif 353 cpus = &stopped_cpus; 354 355 /* signal other cpus to restart */ 356#if X86 357 if (type == IPI_SUSPEND) 358 CPU_COPY_STORE_REL(&map, &toresume_cpus); 359 else 360#endif 361 CPU_COPY_STORE_REL(&map, &started_cpus); 362 363#if X86 364 if (!nmi_is_broadcast || nmi_kdb_lock == 0) { 365#endif 366 /* wait for each to clear its bit */ 367 while (CPU_OVERLAP(cpus, &map)) 368 cpu_spinwait(); 369#if X86 370 } 371#endif 372 373 return (1); 374} 375 376int 377restart_cpus(cpuset_t map) 378{ 379 380 return (generic_restart_cpus(map, IPI_STOP)); 381} 382 383#if X86 384int 385resume_cpus(cpuset_t map) 386{ 387 388 return (generic_restart_cpus(map, IPI_SUSPEND)); 389} 390#endif 391#undef X86 392 393/* 394 * All-CPU rendezvous. CPUs are signalled, all execute the setup function 395 * (if specified), rendezvous, execute the action function (if specified), 396 * rendezvous again, execute the teardown function (if specified), and then 397 * resume. 398 * 399 * Note that the supplied external functions _must_ be reentrant and aware 400 * that they are running in parallel and in an unknown lock context. 401 */ 402void 403smp_rendezvous_action(void) 404{ 405 struct thread *td; 406 void *local_func_arg; 407 void (*local_setup_func)(void*); 408 void (*local_action_func)(void*); 409 void (*local_teardown_func)(void*); 410#ifdef INVARIANTS 411 int owepreempt; 412#endif 413 414 /* Ensure we have up-to-date values. */ 415 atomic_add_acq_int(&smp_rv_waiters[0], 1); 416 while (smp_rv_waiters[0] < smp_rv_ncpus) 417 cpu_spinwait(); 418 419 /* Fetch rendezvous parameters after acquire barrier. */ 420 local_func_arg = smp_rv_func_arg; 421 local_setup_func = smp_rv_setup_func; 422 local_action_func = smp_rv_action_func; 423 local_teardown_func = smp_rv_teardown_func; 424 425 /* 426 * Use a nested critical section to prevent any preemptions 427 * from occurring during a rendezvous action routine. 428 * Specifically, if a rendezvous handler is invoked via an IPI 429 * and the interrupted thread was in the critical_exit() 430 * function after setting td_critnest to 0 but before 431 * performing a deferred preemption, this routine can be 432 * invoked with td_critnest set to 0 and td_owepreempt true. 433 * In that case, a critical_exit() during the rendezvous 434 * action would trigger a preemption which is not permitted in 435 * a rendezvous action. To fix this, wrap all of the 436 * rendezvous action handlers in a critical section. We 437 * cannot use a regular critical section however as having 438 * critical_exit() preempt from this routine would also be 439 * problematic (the preemption must not occur before the IPI 440 * has been acknowledged via an EOI). Instead, we 441 * intentionally ignore td_owepreempt when leaving the 442 * critical section. This should be harmless because we do 443 * not permit rendezvous action routines to schedule threads, 444 * and thus td_owepreempt should never transition from 0 to 1 445 * during this routine. 446 */ 447 td = curthread; 448 td->td_critnest++; 449#ifdef INVARIANTS 450 owepreempt = td->td_owepreempt; 451#endif 452 453 /* 454 * If requested, run a setup function before the main action 455 * function. Ensure all CPUs have completed the setup 456 * function before moving on to the action function. 457 */ 458 if (local_setup_func != smp_no_rendezvous_barrier) { 459 if (smp_rv_setup_func != NULL) 460 smp_rv_setup_func(smp_rv_func_arg); 461 atomic_add_int(&smp_rv_waiters[1], 1); 462 while (smp_rv_waiters[1] < smp_rv_ncpus) 463 cpu_spinwait(); 464 } 465 466 if (local_action_func != NULL) 467 local_action_func(local_func_arg); 468 469 if (local_teardown_func != smp_no_rendezvous_barrier) { 470 /* 471 * Signal that the main action has been completed. If a 472 * full exit rendezvous is requested, then all CPUs will 473 * wait here until all CPUs have finished the main action. 474 */ 475 atomic_add_int(&smp_rv_waiters[2], 1); 476 while (smp_rv_waiters[2] < smp_rv_ncpus) 477 cpu_spinwait(); 478 479 if (local_teardown_func != NULL) 480 local_teardown_func(local_func_arg); 481 } 482 483 /* 484 * Signal that the rendezvous is fully completed by this CPU. 485 * This means that no member of smp_rv_* pseudo-structure will be 486 * accessed by this target CPU after this point; in particular, 487 * memory pointed by smp_rv_func_arg. 488 * 489 * The release semantic ensures that all accesses performed by 490 * the current CPU are visible when smp_rendezvous_cpus() 491 * returns, by synchronizing with the 492 * atomic_load_acq_int(&smp_rv_waiters[3]). 493 */ 494 atomic_add_rel_int(&smp_rv_waiters[3], 1); 495 496 td->td_critnest--; 497 KASSERT(owepreempt == td->td_owepreempt, 498 ("rendezvous action changed td_owepreempt")); 499} 500 501void 502smp_rendezvous_cpus(cpuset_t map, 503 void (* setup_func)(void *), 504 void (* action_func)(void *), 505 void (* teardown_func)(void *), 506 void *arg) 507{ 508 int curcpumap, i, ncpus = 0; 509 510 /* Look comments in the !SMP case. */ 511 if (!smp_started) { 512 spinlock_enter(); 513 if (setup_func != NULL) 514 setup_func(arg); 515 if (action_func != NULL) 516 action_func(arg); 517 if (teardown_func != NULL) 518 teardown_func(arg); 519 spinlock_exit(); 520 return; 521 } 522 523 CPU_FOREACH(i) { 524 if (CPU_ISSET(i, &map)) 525 ncpus++; 526 } 527 if (ncpus == 0) 528 panic("ncpus is 0 with non-zero map"); 529 530 mtx_lock_spin(&smp_ipi_mtx); 531 532 /* Pass rendezvous parameters via global variables. */ 533 smp_rv_ncpus = ncpus; 534 smp_rv_setup_func = setup_func; 535 smp_rv_action_func = action_func; 536 smp_rv_teardown_func = teardown_func; 537 smp_rv_func_arg = arg; 538 smp_rv_waiters[1] = 0; 539 smp_rv_waiters[2] = 0; 540 smp_rv_waiters[3] = 0; 541 atomic_store_rel_int(&smp_rv_waiters[0], 0); 542 543 /* 544 * Signal other processors, which will enter the IPI with 545 * interrupts off. 546 */ 547 curcpumap = CPU_ISSET(curcpu, &map); 548 CPU_CLR(curcpu, &map); 549 ipi_selected(map, IPI_RENDEZVOUS); 550 551 /* Check if the current CPU is in the map */ 552 if (curcpumap != 0) 553 smp_rendezvous_action(); 554 555 /* 556 * Ensure that the master CPU waits for all the other 557 * CPUs to finish the rendezvous, so that smp_rv_* 558 * pseudo-structure and the arg are guaranteed to not 559 * be in use. 560 * 561 * Load acquire synchronizes with the release add in 562 * smp_rendezvous_action(), which ensures that our caller sees 563 * all memory actions done by the called functions on other 564 * CPUs. 565 */ 566 while (atomic_load_acq_int(&smp_rv_waiters[3]) < ncpus) 567 cpu_spinwait(); 568 569 mtx_unlock_spin(&smp_ipi_mtx); 570} 571 572void 573smp_rendezvous(void (* setup_func)(void *), 574 void (* action_func)(void *), 575 void (* teardown_func)(void *), 576 void *arg) 577{ 578 smp_rendezvous_cpus(all_cpus, setup_func, action_func, teardown_func, arg); 579} 580 581static struct cpu_group group[MAXCPU * MAX_CACHE_LEVELS + 1]; 582 583struct cpu_group * 584smp_topo(void) 585{ 586 char cpusetbuf[CPUSETBUFSIZ], cpusetbuf2[CPUSETBUFSIZ]; 587 struct cpu_group *top; 588 589 /* 590 * Check for a fake topology request for debugging purposes. 591 */ 592 switch (smp_topology) { 593 case 1: 594 /* Dual core with no sharing. */ 595 top = smp_topo_1level(CG_SHARE_NONE, 2, 0); 596 break; 597 case 2: 598 /* No topology, all cpus are equal. */ 599 top = smp_topo_none(); 600 break; 601 case 3: 602 /* Dual core with shared L2. */ 603 top = smp_topo_1level(CG_SHARE_L2, 2, 0); 604 break; 605 case 4: 606 /* quad core, shared l3 among each package, private l2. */ 607 top = smp_topo_1level(CG_SHARE_L3, 4, 0); 608 break; 609 case 5: 610 /* quad core, 2 dualcore parts on each package share l2. */ 611 top = smp_topo_2level(CG_SHARE_NONE, 2, CG_SHARE_L2, 2, 0); 612 break; 613 case 6: 614 /* Single-core 2xHTT */ 615 top = smp_topo_1level(CG_SHARE_L1, 2, CG_FLAG_HTT); 616 break; 617 case 7: 618 /* quad core with a shared l3, 8 threads sharing L2. */ 619 top = smp_topo_2level(CG_SHARE_L3, 4, CG_SHARE_L2, 8, 620 CG_FLAG_SMT); 621 break; 622 default: 623 /* Default, ask the system what it wants. */ 624 top = cpu_topo(); 625 break; 626 } 627 /* 628 * Verify the returned topology. 629 */ 630 if (top->cg_count != mp_ncpus) 631 panic("Built bad topology at %p. CPU count %d != %d", 632 top, top->cg_count, mp_ncpus); 633 if (CPU_CMP(&top->cg_mask, &all_cpus)) 634 panic("Built bad topology at %p. CPU mask (%s) != (%s)", 635 top, cpusetobj_strprint(cpusetbuf, &top->cg_mask), 636 cpusetobj_strprint(cpusetbuf2, &all_cpus)); 637 return (top); 638} 639 640struct cpu_group * 641smp_topo_alloc(u_int count) 642{ 643 static u_int index; 644 u_int curr; 645 646 curr = index; 647 index += count; 648 return (&group[curr]); 649} 650 651struct cpu_group * 652smp_topo_none(void) 653{ 654 struct cpu_group *top; 655 656 top = &group[0]; 657 top->cg_parent = NULL; 658 top->cg_child = NULL; 659 top->cg_mask = all_cpus; 660 top->cg_count = mp_ncpus; 661 top->cg_children = 0; 662 top->cg_level = CG_SHARE_NONE; 663 top->cg_flags = 0; 664 665 return (top); 666} 667 668static int 669smp_topo_addleaf(struct cpu_group *parent, struct cpu_group *child, int share, 670 int count, int flags, int start) 671{ 672 char cpusetbuf[CPUSETBUFSIZ], cpusetbuf2[CPUSETBUFSIZ]; 673 cpuset_t mask; 674 int i; 675 676 CPU_ZERO(&mask); 677 for (i = 0; i < count; i++, start++) 678 CPU_SET(start, &mask); 679 child->cg_parent = parent; 680 child->cg_child = NULL; 681 child->cg_children = 0; 682 child->cg_level = share; 683 child->cg_count = count; 684 child->cg_flags = flags; 685 child->cg_mask = mask; 686 parent->cg_children++; 687 for (; parent != NULL; parent = parent->cg_parent) { 688 if (CPU_OVERLAP(&parent->cg_mask, &child->cg_mask)) 689 panic("Duplicate children in %p. mask (%s) child (%s)", 690 parent, 691 cpusetobj_strprint(cpusetbuf, &parent->cg_mask), 692 cpusetobj_strprint(cpusetbuf2, &child->cg_mask)); 693 CPU_OR(&parent->cg_mask, &child->cg_mask); 694 parent->cg_count += child->cg_count; 695 } 696 697 return (start); 698} 699 700struct cpu_group * 701smp_topo_1level(int share, int count, int flags) 702{ 703 struct cpu_group *child; 704 struct cpu_group *top; 705 int packages; 706 int cpu; 707 int i; 708 709 cpu = 0; 710 top = &group[0]; 711 packages = mp_ncpus / count; 712 top->cg_child = child = &group[1]; 713 top->cg_level = CG_SHARE_NONE; 714 for (i = 0; i < packages; i++, child++) 715 cpu = smp_topo_addleaf(top, child, share, count, flags, cpu); 716 return (top); 717} 718 719struct cpu_group * 720smp_topo_2level(int l2share, int l2count, int l1share, int l1count, 721 int l1flags) 722{ 723 struct cpu_group *top; 724 struct cpu_group *l1g; 725 struct cpu_group *l2g; 726 int cpu; 727 int i; 728 int j; 729 730 cpu = 0; 731 top = &group[0]; 732 l2g = &group[1]; 733 top->cg_child = l2g; 734 top->cg_level = CG_SHARE_NONE; 735 top->cg_children = mp_ncpus / (l2count * l1count); 736 l1g = l2g + top->cg_children; 737 for (i = 0; i < top->cg_children; i++, l2g++) { 738 l2g->cg_parent = top; 739 l2g->cg_child = l1g; 740 l2g->cg_level = l2share; 741 for (j = 0; j < l2count; j++, l1g++) 742 cpu = smp_topo_addleaf(l2g, l1g, l1share, l1count, 743 l1flags, cpu); 744 } 745 return (top); 746} 747 748 749struct cpu_group * 750smp_topo_find(struct cpu_group *top, int cpu) 751{ 752 struct cpu_group *cg; 753 cpuset_t mask; 754 int children; 755 int i; 756 757 CPU_SETOF(cpu, &mask); 758 cg = top; 759 for (;;) { 760 if (!CPU_OVERLAP(&cg->cg_mask, &mask)) 761 return (NULL); 762 if (cg->cg_children == 0) 763 return (cg); 764 children = cg->cg_children; 765 for (i = 0, cg = cg->cg_child; i < children; cg++, i++) 766 if (CPU_OVERLAP(&cg->cg_mask, &mask)) 767 break; 768 } 769 return (NULL); 770} 771#else /* !SMP */ 772 773void 774smp_rendezvous_cpus(cpuset_t map, 775 void (*setup_func)(void *), 776 void (*action_func)(void *), 777 void (*teardown_func)(void *), 778 void *arg) 779{ 780 /* 781 * In the !SMP case we just need to ensure the same initial conditions 782 * as the SMP case. 783 */ 784 spinlock_enter(); 785 if (setup_func != NULL) 786 setup_func(arg); 787 if (action_func != NULL) 788 action_func(arg); 789 if (teardown_func != NULL) 790 teardown_func(arg); 791 spinlock_exit(); 792} 793 794void 795smp_rendezvous(void (*setup_func)(void *), 796 void (*action_func)(void *), 797 void (*teardown_func)(void *), 798 void *arg) 799{ 800 801 /* Look comments in the smp_rendezvous_cpus() case. */ 802 spinlock_enter(); 803 if (setup_func != NULL) 804 setup_func(arg); 805 if (action_func != NULL) 806 action_func(arg); 807 if (teardown_func != NULL) 808 teardown_func(arg); 809 spinlock_exit(); 810} 811 812/* 813 * Provide dummy SMP support for UP kernels. Modules that need to use SMP 814 * APIs will still work using this dummy support. 815 */ 816static void 817mp_setvariables_for_up(void *dummy) 818{ 819 mp_ncpus = 1; 820 mp_maxid = PCPU_GET(cpuid); 821 CPU_SETOF(mp_maxid, &all_cpus); 822 KASSERT(PCPU_GET(cpuid) == 0, ("UP must have a CPU ID of zero")); 823} 824SYSINIT(cpu_mp_setvariables, SI_SUB_TUNABLES, SI_ORDER_FIRST, 825 mp_setvariables_for_up, NULL); 826#endif /* SMP */ 827 828/* 829 * smp_no_rendevous_barrier was renamed to smp_no_rendezvous_barrier 830 * in __FreeBSD_version 1101508, with the old name remaining in 11.x 831 * as an alias for compatibility. The old name will be gone in 12.0 832 * (__FreeBSD_version >= 1200028). 833 */ 834__strong_reference(smp_no_rendezvous_barrier, smp_no_rendevous_barrier); 835void 836smp_no_rendezvous_barrier(void *dummy) 837{ 838#ifdef SMP 839 KASSERT((!smp_started),("smp_no_rendezvous called and smp is started")); 840#endif 841} 842 843/* 844 * Wait specified idle threads to switch once. This ensures that even 845 * preempted threads have cycled through the switch function once, 846 * exiting their codepaths. This allows us to change global pointers 847 * with no other synchronization. 848 */ 849int 850quiesce_cpus(cpuset_t map, const char *wmesg, int prio) 851{ 852 struct pcpu *pcpu; 853 u_int gen[MAXCPU]; 854 int error; 855 int cpu; 856 857 error = 0; 858 for (cpu = 0; cpu <= mp_maxid; cpu++) { 859 if (!CPU_ISSET(cpu, &map) || CPU_ABSENT(cpu)) 860 continue; 861 pcpu = pcpu_find(cpu); 862 gen[cpu] = pcpu->pc_idlethread->td_generation; 863 } 864 for (cpu = 0; cpu <= mp_maxid; cpu++) { 865 if (!CPU_ISSET(cpu, &map) || CPU_ABSENT(cpu)) 866 continue; 867 pcpu = pcpu_find(cpu); 868 thread_lock(curthread); 869 sched_bind(curthread, cpu); 870 thread_unlock(curthread); 871 while (gen[cpu] == pcpu->pc_idlethread->td_generation) { 872 error = tsleep(quiesce_cpus, prio, wmesg, 1); 873 if (error != EWOULDBLOCK) 874 goto out; 875 error = 0; 876 } 877 } 878out: 879 thread_lock(curthread); 880 sched_unbind(curthread); 881 thread_unlock(curthread); 882 883 return (error); 884} 885 886int 887quiesce_all_cpus(const char *wmesg, int prio) 888{ 889 890 return quiesce_cpus(all_cpus, wmesg, prio); 891} 892 893/* Extra care is taken with this sysctl because the data type is volatile */ 894static int 895sysctl_kern_smp_active(SYSCTL_HANDLER_ARGS) 896{ 897 int error, active; 898 899 active = smp_started; 900 error = SYSCTL_OUT(req, &active, sizeof(active)); 901 return (error); 902} 903 904 905#ifdef SMP 906void 907topo_init_node(struct topo_node *node) 908{ 909 910 bzero(node, sizeof(*node)); 911 TAILQ_INIT(&node->children); 912} 913 914void 915topo_init_root(struct topo_node *root) 916{ 917 918 topo_init_node(root); 919 root->type = TOPO_TYPE_SYSTEM; 920} 921 922/* 923 * Add a child node with the given ID under the given parent. 924 * Do nothing if there is already a child with that ID. 925 */ 926struct topo_node * 927topo_add_node_by_hwid(struct topo_node *parent, int hwid, 928 topo_node_type type, uintptr_t subtype) 929{ 930 struct topo_node *node; 931 932 TAILQ_FOREACH_REVERSE(node, &parent->children, 933 topo_children, siblings) { 934 if (node->hwid == hwid 935 && node->type == type && node->subtype == subtype) { 936 return (node); 937 } 938 } 939 940 node = malloc(sizeof(*node), M_TOPO, M_WAITOK); 941 topo_init_node(node); 942 node->parent = parent; 943 node->hwid = hwid; 944 node->type = type; 945 node->subtype = subtype; 946 TAILQ_INSERT_TAIL(&parent->children, node, siblings); 947 parent->nchildren++; 948 949 return (node); 950} 951 952/* 953 * Find a child node with the given ID under the given parent. 954 */ 955struct topo_node * 956topo_find_node_by_hwid(struct topo_node *parent, int hwid, 957 topo_node_type type, uintptr_t subtype) 958{ 959 960 struct topo_node *node; 961 962 TAILQ_FOREACH(node, &parent->children, siblings) { 963 if (node->hwid == hwid 964 && node->type == type && node->subtype == subtype) { 965 return (node); 966 } 967 } 968 969 return (NULL); 970} 971 972/* 973 * Given a node change the order of its parent's child nodes such 974 * that the node becomes the firt child while preserving the cyclic 975 * order of the children. In other words, the given node is promoted 976 * by rotation. 977 */ 978void 979topo_promote_child(struct topo_node *child) 980{ 981 struct topo_node *next; 982 struct topo_node *node; 983 struct topo_node *parent; 984 985 parent = child->parent; 986 next = TAILQ_NEXT(child, siblings); 987 TAILQ_REMOVE(&parent->children, child, siblings); 988 TAILQ_INSERT_HEAD(&parent->children, child, siblings); 989 990 while (next != NULL) { 991 node = next; 992 next = TAILQ_NEXT(node, siblings); 993 TAILQ_REMOVE(&parent->children, node, siblings); 994 TAILQ_INSERT_AFTER(&parent->children, child, node, siblings); 995 child = node; 996 } 997} 998 999/* 1000 * Iterate to the next node in the depth-first search (traversal) of 1001 * the topology tree. 1002 */ 1003struct topo_node * 1004topo_next_node(struct topo_node *top, struct topo_node *node) 1005{ 1006 struct topo_node *next; 1007 1008 if ((next = TAILQ_FIRST(&node->children)) != NULL) 1009 return (next); 1010 1011 if ((next = TAILQ_NEXT(node, siblings)) != NULL) 1012 return (next); 1013 1014 while ((node = node->parent) != top) 1015 if ((next = TAILQ_NEXT(node, siblings)) != NULL) 1016 return (next); 1017 1018 return (NULL); 1019} 1020 1021/* 1022 * Iterate to the next node in the depth-first search of the topology tree, 1023 * but without descending below the current node. 1024 */ 1025struct topo_node * 1026topo_next_nonchild_node(struct topo_node *top, struct topo_node *node) 1027{ 1028 struct topo_node *next; 1029 1030 if ((next = TAILQ_NEXT(node, siblings)) != NULL) 1031 return (next); 1032 1033 while ((node = node->parent) != top) 1034 if ((next = TAILQ_NEXT(node, siblings)) != NULL) 1035 return (next); 1036 1037 return (NULL); 1038} 1039 1040/* 1041 * Assign the given ID to the given topology node that represents a logical 1042 * processor. 1043 */ 1044void 1045topo_set_pu_id(struct topo_node *node, cpuid_t id) 1046{ 1047 1048 KASSERT(node->type == TOPO_TYPE_PU, 1049 ("topo_set_pu_id: wrong node type: %u", node->type)); 1050 KASSERT(CPU_EMPTY(&node->cpuset) && node->cpu_count == 0, 1051 ("topo_set_pu_id: cpuset already not empty")); 1052 node->id = id; 1053 CPU_SET(id, &node->cpuset); 1054 node->cpu_count = 1; 1055 node->subtype = 1; 1056 1057 while ((node = node->parent) != NULL) { 1058 KASSERT(!CPU_ISSET(id, &node->cpuset), 1059 ("logical ID %u is already set in node %p", id, node)); 1060 CPU_SET(id, &node->cpuset); 1061 node->cpu_count++; 1062 } 1063} 1064 1065/* 1066 * Check if the topology is uniform, that is, each package has the same number 1067 * of cores in it and each core has the same number of threads (logical 1068 * processors) in it. If so, calculate the number of package, the number of 1069 * cores per package and the number of logical processors per core. 1070 * 'all' parameter tells whether to include administratively disabled logical 1071 * processors into the analysis. 1072 */ 1073int 1074topo_analyze(struct topo_node *topo_root, int all, 1075 int *pkg_count, int *cores_per_pkg, int *thrs_per_core) 1076{ 1077 struct topo_node *pkg_node; 1078 struct topo_node *core_node; 1079 struct topo_node *pu_node; 1080 int thrs_per_pkg; 1081 int cpp_counter; 1082 int tpc_counter; 1083 int tpp_counter; 1084 1085 *pkg_count = 0; 1086 *cores_per_pkg = -1; 1087 *thrs_per_core = -1; 1088 thrs_per_pkg = -1; 1089 pkg_node = topo_root; 1090 while (pkg_node != NULL) { 1091 if (pkg_node->type != TOPO_TYPE_PKG) { 1092 pkg_node = topo_next_node(topo_root, pkg_node); 1093 continue; 1094 } 1095 if (!all && CPU_EMPTY(&pkg_node->cpuset)) { 1096 pkg_node = topo_next_nonchild_node(topo_root, pkg_node); 1097 continue; 1098 } 1099 1100 (*pkg_count)++; 1101 1102 cpp_counter = 0; 1103 tpp_counter = 0; 1104 core_node = pkg_node; 1105 while (core_node != NULL) { 1106 if (core_node->type == TOPO_TYPE_CORE) { 1107 if (!all && CPU_EMPTY(&core_node->cpuset)) { 1108 core_node = 1109 topo_next_nonchild_node(pkg_node, 1110 core_node); 1111 continue; 1112 } 1113 1114 cpp_counter++; 1115 1116 tpc_counter = 0; 1117 pu_node = core_node; 1118 while (pu_node != NULL) { 1119 if (pu_node->type == TOPO_TYPE_PU && 1120 (all || !CPU_EMPTY(&pu_node->cpuset))) 1121 tpc_counter++; 1122 pu_node = topo_next_node(core_node, 1123 pu_node); 1124 } 1125 1126 if (*thrs_per_core == -1) 1127 *thrs_per_core = tpc_counter; 1128 else if (*thrs_per_core != tpc_counter) 1129 return (0); 1130 1131 core_node = topo_next_nonchild_node(pkg_node, 1132 core_node); 1133 } else { 1134 /* PU node directly under PKG. */ 1135 if (core_node->type == TOPO_TYPE_PU && 1136 (all || !CPU_EMPTY(&core_node->cpuset))) 1137 tpp_counter++; 1138 core_node = topo_next_node(pkg_node, 1139 core_node); 1140 } 1141 } 1142 1143 if (*cores_per_pkg == -1) 1144 *cores_per_pkg = cpp_counter; 1145 else if (*cores_per_pkg != cpp_counter) 1146 return (0); 1147 if (thrs_per_pkg == -1) 1148 thrs_per_pkg = tpp_counter; 1149 else if (thrs_per_pkg != tpp_counter) 1150 return (0); 1151 1152 pkg_node = topo_next_nonchild_node(topo_root, pkg_node); 1153 } 1154 1155 KASSERT(*pkg_count > 0, 1156 ("bug in topology or analysis")); 1157 if (*cores_per_pkg == 0) { 1158 KASSERT(*thrs_per_core == -1 && thrs_per_pkg > 0, 1159 ("bug in topology or analysis")); 1160 *thrs_per_core = thrs_per_pkg; 1161 } 1162 1163 return (1); 1164} 1165#endif /* SMP */ 1166 1167