1 /*- 2 * Copyright (c) 2001, John Baldwin <jhb@FreeBSD.org>. 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice, this list of conditions and the following disclaimer. 10 * 2. Redistributions in binary form must reproduce the above copyright 11 * notice, this list of conditions and the following disclaimer in the 12 * documentation and/or other materials provided with the distribution. 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 /* 28 * This module holds the global variables and machine independent functions 29 * used for the kernel SMP support. 30 */ 31 32 #include <sys/cdefs.h> 33 __FBSDID("$FreeBSD$"); 34 35 #include <sys/param.h> 36 #include <sys/systm.h> 37 #include <sys/kernel.h> 38 #include <sys/ktr.h> 39 #include <sys/proc.h> 40 #include <sys/bus.h> 41 #include <sys/lock.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 54 volatile cpuset_t stopped_cpus; 55 volatile cpuset_t started_cpus; 56 volatile cpuset_t suspended_cpus; 57 cpuset_t hlt_cpus_mask; 58 cpuset_t logical_cpus_mask; 59 60 void (*cpustop_restartfunc)(void); 61 #endif 62 /* This is used in modules that need to work in both SMP and UP. */ 63 cpuset_t all_cpus; 64 65 int mp_ncpus; 66 /* export this for libkvm consumers. */ 67 int mp_maxcpus = MAXCPU; 68 69 volatile int smp_started; 70 u_int mp_maxid; 71 72 static SYSCTL_NODE(_kern, OID_AUTO, smp, CTLFLAG_RD|CTLFLAG_CAPRD, NULL, 73 "Kernel SMP"); 74 75 SYSCTL_INT(_kern_smp, OID_AUTO, maxid, CTLFLAG_RD|CTLFLAG_CAPRD, &mp_maxid, 0, 76 "Max CPU ID."); 77 78 SYSCTL_INT(_kern_smp, OID_AUTO, maxcpus, CTLFLAG_RD|CTLFLAG_CAPRD, &mp_maxcpus, 79 0, "Max number of CPUs that the system was compiled for."); 80 81 int smp_active = 0; /* are the APs allowed to run? */ 82 SYSCTL_INT(_kern_smp, OID_AUTO, active, CTLFLAG_RW, &smp_active, 0, 83 "Number of Auxillary Processors (APs) that were successfully started"); 84 85 int smp_disabled = 0; /* has smp been disabled? */ 86 SYSCTL_INT(_kern_smp, OID_AUTO, disabled, CTLFLAG_RDTUN|CTLFLAG_CAPRD, 87 &smp_disabled, 0, "SMP has been disabled from the loader"); 88 TUNABLE_INT("kern.smp.disabled", &smp_disabled); 89 90 int smp_cpus = 1; /* how many cpu's running */ 91 SYSCTL_INT(_kern_smp, OID_AUTO, cpus, CTLFLAG_RD|CTLFLAG_CAPRD, &smp_cpus, 0, 92 "Number of CPUs online"); 93 94 int smp_topology = 0; /* Which topology we're using. */ 95 SYSCTL_INT(_kern_smp, OID_AUTO, topology, CTLFLAG_RD, &smp_topology, 0, 96 "Topology override setting; 0 is default provided by hardware."); 97 TUNABLE_INT("kern.smp.topology", &smp_topology); 98 99 #ifdef SMP 100 /* Enable forwarding of a signal to a process running on a different CPU */ 101 static int forward_signal_enabled = 1; 102 SYSCTL_INT(_kern_smp, OID_AUTO, forward_signal_enabled, CTLFLAG_RW, 103 &forward_signal_enabled, 0, 104 "Forwarding of a signal to a process on a different CPU"); 105 106 /* Variables needed for SMP rendezvous. */ 107 static volatile int smp_rv_ncpus; 108 static void (*volatile smp_rv_setup_func)(void *arg); 109 static void (*volatile smp_rv_action_func)(void *arg); 110 static void (*volatile smp_rv_teardown_func)(void *arg); 111 static void *volatile smp_rv_func_arg; 112 static volatile int smp_rv_waiters[4]; 113 114 /* 115 * Shared mutex to restrict busywaits between smp_rendezvous() and 116 * smp(_targeted)_tlb_shootdown(). A deadlock occurs if both of these 117 * functions trigger at once and cause multiple CPUs to busywait with 118 * interrupts disabled. 119 */ 120 struct mtx smp_ipi_mtx; 121 122 /* 123 * Let the MD SMP code initialize mp_maxid very early if it can. 124 */ 125 static void 126 mp_setmaxid(void *dummy) 127 { 128 cpu_mp_setmaxid(); 129 } 130 SYSINIT(cpu_mp_setmaxid, SI_SUB_TUNABLES, SI_ORDER_FIRST, mp_setmaxid, NULL); 131 132 /* 133 * Call the MD SMP initialization code. 134 */ 135 static void 136 mp_start(void *dummy) 137 { 138 139 mtx_init(&smp_ipi_mtx, "smp rendezvous", NULL, MTX_SPIN); 140 141 /* Probe for MP hardware. */ 142 if (smp_disabled != 0 || cpu_mp_probe() == 0) { 143 mp_ncpus = 1; 144 CPU_SETOF(PCPU_GET(cpuid), &all_cpus); 145 return; 146 } 147 148 cpu_mp_start(); 149 printf("FreeBSD/SMP: Multiprocessor System Detected: %d CPUs\n", 150 mp_ncpus); 151 cpu_mp_announce(); 152 } 153 SYSINIT(cpu_mp, SI_SUB_CPU, SI_ORDER_THIRD, mp_start, NULL); 154 155 void 156 forward_signal(struct thread *td) 157 { 158 int id; 159 160 /* 161 * signotify() has already set TDF_ASTPENDING and TDF_NEEDSIGCHECK on 162 * this thread, so all we need to do is poke it if it is currently 163 * executing so that it executes ast(). 164 */ 165 THREAD_LOCK_ASSERT(td, MA_OWNED); 166 KASSERT(TD_IS_RUNNING(td), 167 ("forward_signal: thread is not TDS_RUNNING")); 168 169 CTR1(KTR_SMP, "forward_signal(%p)", td->td_proc); 170 171 if (!smp_started || cold || panicstr) 172 return; 173 if (!forward_signal_enabled) 174 return; 175 176 /* No need to IPI ourself. */ 177 if (td == curthread) 178 return; 179 180 id = td->td_oncpu; 181 if (id == NOCPU) 182 return; 183 ipi_cpu(id, IPI_AST); 184 } 185 186 /* 187 * When called the executing CPU will send an IPI to all other CPUs 188 * requesting that they halt execution. 189 * 190 * Usually (but not necessarily) called with 'other_cpus' as its arg. 191 * 192 * - Signals all CPUs in map to stop. 193 * - Waits for each to stop. 194 * 195 * Returns: 196 * -1: error 197 * 0: NA 198 * 1: ok 199 * 200 */ 201 static int 202 generic_stop_cpus(cpuset_t map, u_int type) 203 { 204 #ifdef KTR 205 char cpusetbuf[CPUSETBUFSIZ]; 206 #endif 207 static volatile u_int stopping_cpu = NOCPU; 208 int i; 209 volatile cpuset_t *cpus; 210 211 KASSERT( 212 #if defined(__amd64__) || defined(__i386__) 213 type == IPI_STOP || type == IPI_STOP_HARD || type == IPI_SUSPEND, 214 #else 215 type == IPI_STOP || type == IPI_STOP_HARD, 216 #endif 217 ("%s: invalid stop type", __func__)); 218 219 if (!smp_started) 220 return (0); 221 222 CTR2(KTR_SMP, "stop_cpus(%s) with %u type", 223 cpusetobj_strprint(cpusetbuf, &map), type); 224 225 #if defined(__amd64__) || defined(__i386__) 226 /* 227 * When suspending, ensure there are are no IPIs in progress. 228 * IPIs that have been issued, but not yet delivered (e.g. 229 * not pending on a vCPU when running under virtualization) 230 * will be lost, violating FreeBSD's assumption of reliable 231 * IPI delivery. 232 */ 233 if (type == IPI_SUSPEND) 234 mtx_lock_spin(&smp_ipi_mtx); 235 #endif 236 237 if (stopping_cpu != PCPU_GET(cpuid)) 238 while (atomic_cmpset_int(&stopping_cpu, NOCPU, 239 PCPU_GET(cpuid)) == 0) 240 while (stopping_cpu != NOCPU) 241 cpu_spinwait(); /* spin */ 242 243 /* send the stop IPI to all CPUs in map */ 244 ipi_selected(map, type); 245 246 #if defined(__amd64__) || defined(__i386__) 247 if (type == IPI_SUSPEND) 248 cpus = &suspended_cpus; 249 else 250 #endif 251 cpus = &stopped_cpus; 252 253 i = 0; 254 while (!CPU_SUBSET(cpus, &map)) { 255 /* spin */ 256 cpu_spinwait(); 257 i++; 258 if (i == 100000000) { 259 printf("timeout stopping cpus\n"); 260 break; 261 } 262 } 263 264 #if defined(__amd64__) || defined(__i386__) 265 if (type == IPI_SUSPEND) 266 mtx_unlock_spin(&smp_ipi_mtx); 267 #endif 268 269 stopping_cpu = NOCPU; 270 return (1); 271 } 272 273 int 274 stop_cpus(cpuset_t map) 275 { 276 277 return (generic_stop_cpus(map, IPI_STOP)); 278 } 279 280 int 281 stop_cpus_hard(cpuset_t map) 282 { 283 284 return (generic_stop_cpus(map, IPI_STOP_HARD)); 285 } 286 287 #if defined(__amd64__) || defined(__i386__) 288 int 289 suspend_cpus(cpuset_t map) 290 { 291 292 return (generic_stop_cpus(map, IPI_SUSPEND)); 293 } 294 #endif 295 296 /* 297 * Called by a CPU to restart stopped CPUs. 298 * 299 * Usually (but not necessarily) called with 'stopped_cpus' as its arg. 300 * 301 * - Signals all CPUs in map to restart. 302 * - Waits for each to restart. 303 * 304 * Returns: 305 * -1: error 306 * 0: NA 307 * 1: ok 308 */ 309 static int 310 generic_restart_cpus(cpuset_t map, u_int type) 311 { 312 #ifdef KTR 313 char cpusetbuf[CPUSETBUFSIZ]; 314 #endif 315 volatile cpuset_t *cpus; 316 317 KASSERT( 318 #if defined(__amd64__) || defined(__i386__) 319 type == IPI_STOP || type == IPI_STOP_HARD || type == IPI_SUSPEND, 320 #else 321 type == IPI_STOP || type == IPI_STOP_HARD, 322 #endif 323 ("%s: invalid stop type", __func__)); 324 325 if (!smp_started) 326 return 0; 327 328 CTR1(KTR_SMP, "restart_cpus(%s)", cpusetobj_strprint(cpusetbuf, &map)); 329 330 #if defined(__amd64__) || defined(__i386__) 331 if (type == IPI_SUSPEND) 332 cpus = &suspended_cpus; 333 else 334 #endif 335 cpus = &stopped_cpus; 336 337 /* signal other cpus to restart */ 338 CPU_COPY_STORE_REL(&map, &started_cpus); 339 340 /* wait for each to clear its bit */ 341 while (CPU_OVERLAP(cpus, &map)) 342 cpu_spinwait(); 343 344 return 1; 345 } 346 347 int 348 restart_cpus(cpuset_t map) 349 { 350 351 return (generic_restart_cpus(map, IPI_STOP)); 352 } 353 354 #if defined(__amd64__) || defined(__i386__) 355 int 356 resume_cpus(cpuset_t map) 357 { 358 359 return (generic_restart_cpus(map, IPI_SUSPEND)); 360 } 361 #endif 362 363 /* 364 * All-CPU rendezvous. CPUs are signalled, all execute the setup function 365 * (if specified), rendezvous, execute the action function (if specified), 366 * rendezvous again, execute the teardown function (if specified), and then 367 * resume. 368 * 369 * Note that the supplied external functions _must_ be reentrant and aware 370 * that they are running in parallel and in an unknown lock context. 371 */ 372 void 373 smp_rendezvous_action(void) 374 { 375 struct thread *td; 376 void *local_func_arg; 377 void (*local_setup_func)(void*); 378 void (*local_action_func)(void*); 379 void (*local_teardown_func)(void*); 380 #ifdef INVARIANTS 381 int owepreempt; 382 #endif 383 384 /* Ensure we have up-to-date values. */ 385 atomic_add_acq_int(&smp_rv_waiters[0], 1); 386 while (smp_rv_waiters[0] < smp_rv_ncpus) 387 cpu_spinwait(); 388 389 /* Fetch rendezvous parameters after acquire barrier. */ 390 local_func_arg = smp_rv_func_arg; 391 local_setup_func = smp_rv_setup_func; 392 local_action_func = smp_rv_action_func; 393 local_teardown_func = smp_rv_teardown_func; 394 395 /* 396 * Use a nested critical section to prevent any preemptions 397 * from occurring during a rendezvous action routine. 398 * Specifically, if a rendezvous handler is invoked via an IPI 399 * and the interrupted thread was in the critical_exit() 400 * function after setting td_critnest to 0 but before 401 * performing a deferred preemption, this routine can be 402 * invoked with td_critnest set to 0 and td_owepreempt true. 403 * In that case, a critical_exit() during the rendezvous 404 * action would trigger a preemption which is not permitted in 405 * a rendezvous action. To fix this, wrap all of the 406 * rendezvous action handlers in a critical section. We 407 * cannot use a regular critical section however as having 408 * critical_exit() preempt from this routine would also be 409 * problematic (the preemption must not occur before the IPI 410 * has been acknowledged via an EOI). Instead, we 411 * intentionally ignore td_owepreempt when leaving the 412 * critical section. This should be harmless because we do 413 * not permit rendezvous action routines to schedule threads, 414 * and thus td_owepreempt should never transition from 0 to 1 415 * during this routine. 416 */ 417 td = curthread; 418 td->td_critnest++; 419 #ifdef INVARIANTS 420 owepreempt = td->td_owepreempt; 421 #endif 422 423 /* 424 * If requested, run a setup function before the main action 425 * function. Ensure all CPUs have completed the setup 426 * function before moving on to the action function. 427 */ 428 if (local_setup_func != smp_no_rendevous_barrier) { 429 if (smp_rv_setup_func != NULL) 430 smp_rv_setup_func(smp_rv_func_arg); 431 atomic_add_int(&smp_rv_waiters[1], 1); 432 while (smp_rv_waiters[1] < smp_rv_ncpus) 433 cpu_spinwait(); 434 } 435 436 if (local_action_func != NULL) 437 local_action_func(local_func_arg); 438 439 if (local_teardown_func != smp_no_rendevous_barrier) { 440 /* 441 * Signal that the main action has been completed. If a 442 * full exit rendezvous is requested, then all CPUs will 443 * wait here until all CPUs have finished the main action. 444 */ 445 atomic_add_int(&smp_rv_waiters[2], 1); 446 while (smp_rv_waiters[2] < smp_rv_ncpus) 447 cpu_spinwait(); 448 449 if (local_teardown_func != NULL) 450 local_teardown_func(local_func_arg); 451 } 452 453 /* 454 * Signal that the rendezvous is fully completed by this CPU. 455 * This means that no member of smp_rv_* pseudo-structure will be 456 * accessed by this target CPU after this point; in particular, 457 * memory pointed by smp_rv_func_arg. 458 */ 459 atomic_add_int(&smp_rv_waiters[3], 1); 460 461 td->td_critnest--; 462 KASSERT(owepreempt == td->td_owepreempt, 463 ("rendezvous action changed td_owepreempt")); 464 } 465 466 void 467 smp_rendezvous_cpus(cpuset_t map, 468 void (* setup_func)(void *), 469 void (* action_func)(void *), 470 void (* teardown_func)(void *), 471 void *arg) 472 { 473 int curcpumap, i, ncpus = 0; 474 475 /* Look comments in the !SMP case. */ 476 if (!smp_started) { 477 spinlock_enter(); 478 if (setup_func != NULL) 479 setup_func(arg); 480 if (action_func != NULL) 481 action_func(arg); 482 if (teardown_func != NULL) 483 teardown_func(arg); 484 spinlock_exit(); 485 return; 486 } 487 488 CPU_FOREACH(i) { 489 if (CPU_ISSET(i, &map)) 490 ncpus++; 491 } 492 if (ncpus == 0) 493 panic("ncpus is 0 with non-zero map"); 494 495 mtx_lock_spin(&smp_ipi_mtx); 496 497 /* Pass rendezvous parameters via global variables. */ 498 smp_rv_ncpus = ncpus; 499 smp_rv_setup_func = setup_func; 500 smp_rv_action_func = action_func; 501 smp_rv_teardown_func = teardown_func; 502 smp_rv_func_arg = arg; 503 smp_rv_waiters[1] = 0; 504 smp_rv_waiters[2] = 0; 505 smp_rv_waiters[3] = 0; 506 atomic_store_rel_int(&smp_rv_waiters[0], 0); 507 508 /* 509 * Signal other processors, which will enter the IPI with 510 * interrupts off. 511 */ 512 curcpumap = CPU_ISSET(curcpu, &map); 513 CPU_CLR(curcpu, &map); 514 ipi_selected(map, IPI_RENDEZVOUS); 515 516 /* Check if the current CPU is in the map */ 517 if (curcpumap != 0) 518 smp_rendezvous_action(); 519 520 /* 521 * Ensure that the master CPU waits for all the other 522 * CPUs to finish the rendezvous, so that smp_rv_* 523 * pseudo-structure and the arg are guaranteed to not 524 * be in use. 525 */ 526 while (atomic_load_acq_int(&smp_rv_waiters[3]) < ncpus) 527 cpu_spinwait(); 528 529 mtx_unlock_spin(&smp_ipi_mtx); 530 } 531 532 void 533 smp_rendezvous(void (* setup_func)(void *), 534 void (* action_func)(void *), 535 void (* teardown_func)(void *), 536 void *arg) 537 { 538 smp_rendezvous_cpus(all_cpus, setup_func, action_func, teardown_func, arg); 539 } 540 541 static struct cpu_group group[MAXCPU]; 542 543 struct cpu_group * 544 smp_topo(void) 545 { 546 char cpusetbuf[CPUSETBUFSIZ], cpusetbuf2[CPUSETBUFSIZ]; 547 struct cpu_group *top; 548 549 /* 550 * Check for a fake topology request for debugging purposes. 551 */ 552 switch (smp_topology) { 553 case 1: 554 /* Dual core with no sharing. */ 555 top = smp_topo_1level(CG_SHARE_NONE, 2, 0); 556 break; 557 case 2: 558 /* No topology, all cpus are equal. */ 559 top = smp_topo_none(); 560 break; 561 case 3: 562 /* Dual core with shared L2. */ 563 top = smp_topo_1level(CG_SHARE_L2, 2, 0); 564 break; 565 case 4: 566 /* quad core, shared l3 among each package, private l2. */ 567 top = smp_topo_1level(CG_SHARE_L3, 4, 0); 568 break; 569 case 5: 570 /* quad core, 2 dualcore parts on each package share l2. */ 571 top = smp_topo_2level(CG_SHARE_NONE, 2, CG_SHARE_L2, 2, 0); 572 break; 573 case 6: 574 /* Single-core 2xHTT */ 575 top = smp_topo_1level(CG_SHARE_L1, 2, CG_FLAG_HTT); 576 break; 577 case 7: 578 /* quad core with a shared l3, 8 threads sharing L2. */ 579 top = smp_topo_2level(CG_SHARE_L3, 4, CG_SHARE_L2, 8, 580 CG_FLAG_SMT); 581 break; 582 default: 583 /* Default, ask the system what it wants. */ 584 top = cpu_topo(); 585 break; 586 } 587 /* 588 * Verify the returned topology. 589 */ 590 if (top->cg_count != mp_ncpus) 591 panic("Built bad topology at %p. CPU count %d != %d", 592 top, top->cg_count, mp_ncpus); 593 if (CPU_CMP(&top->cg_mask, &all_cpus)) 594 panic("Built bad topology at %p. CPU mask (%s) != (%s)", 595 top, cpusetobj_strprint(cpusetbuf, &top->cg_mask), 596 cpusetobj_strprint(cpusetbuf2, &all_cpus)); 597 return (top); 598 } 599 600 struct cpu_group * 601 smp_topo_none(void) 602 { 603 struct cpu_group *top; 604 605 top = &group[0]; 606 top->cg_parent = NULL; 607 top->cg_child = NULL; 608 top->cg_mask = all_cpus; 609 top->cg_count = mp_ncpus; 610 top->cg_children = 0; 611 top->cg_level = CG_SHARE_NONE; 612 top->cg_flags = 0; 613 614 return (top); 615 } 616 617 static int 618 smp_topo_addleaf(struct cpu_group *parent, struct cpu_group *child, int share, 619 int count, int flags, int start) 620 { 621 char cpusetbuf[CPUSETBUFSIZ], cpusetbuf2[CPUSETBUFSIZ]; 622 cpuset_t mask; 623 int i; 624 625 CPU_ZERO(&mask); 626 for (i = 0; i < count; i++, start++) 627 CPU_SET(start, &mask); 628 child->cg_parent = parent; 629 child->cg_child = NULL; 630 child->cg_children = 0; 631 child->cg_level = share; 632 child->cg_count = count; 633 child->cg_flags = flags; 634 child->cg_mask = mask; 635 parent->cg_children++; 636 for (; parent != NULL; parent = parent->cg_parent) { 637 if (CPU_OVERLAP(&parent->cg_mask, &child->cg_mask)) 638 panic("Duplicate children in %p. mask (%s) child (%s)", 639 parent, 640 cpusetobj_strprint(cpusetbuf, &parent->cg_mask), 641 cpusetobj_strprint(cpusetbuf2, &child->cg_mask)); 642 CPU_OR(&parent->cg_mask, &child->cg_mask); 643 parent->cg_count += child->cg_count; 644 } 645 646 return (start); 647 } 648 649 struct cpu_group * 650 smp_topo_1level(int share, int count, int flags) 651 { 652 struct cpu_group *child; 653 struct cpu_group *top; 654 int packages; 655 int cpu; 656 int i; 657 658 cpu = 0; 659 top = &group[0]; 660 packages = mp_ncpus / count; 661 top->cg_child = child = &group[1]; 662 top->cg_level = CG_SHARE_NONE; 663 for (i = 0; i < packages; i++, child++) 664 cpu = smp_topo_addleaf(top, child, share, count, flags, cpu); 665 return (top); 666 } 667 668 struct cpu_group * 669 smp_topo_2level(int l2share, int l2count, int l1share, int l1count, 670 int l1flags) 671 { 672 struct cpu_group *top; 673 struct cpu_group *l1g; 674 struct cpu_group *l2g; 675 int cpu; 676 int i; 677 int j; 678 679 cpu = 0; 680 top = &group[0]; 681 l2g = &group[1]; 682 top->cg_child = l2g; 683 top->cg_level = CG_SHARE_NONE; 684 top->cg_children = mp_ncpus / (l2count * l1count); 685 l1g = l2g + top->cg_children; 686 for (i = 0; i < top->cg_children; i++, l2g++) { 687 l2g->cg_parent = top; 688 l2g->cg_child = l1g; 689 l2g->cg_level = l2share; 690 for (j = 0; j < l2count; j++, l1g++) 691 cpu = smp_topo_addleaf(l2g, l1g, l1share, l1count, 692 l1flags, cpu); 693 } 694 return (top); 695 } 696 697 698 struct cpu_group * 699 smp_topo_find(struct cpu_group *top, int cpu) 700 { 701 struct cpu_group *cg; 702 cpuset_t mask; 703 int children; 704 int i; 705 706 CPU_SETOF(cpu, &mask); 707 cg = top; 708 for (;;) { 709 if (!CPU_OVERLAP(&cg->cg_mask, &mask)) 710 return (NULL); 711 if (cg->cg_children == 0) 712 return (cg); 713 children = cg->cg_children; 714 for (i = 0, cg = cg->cg_child; i < children; cg++, i++) 715 if (CPU_OVERLAP(&cg->cg_mask, &mask)) 716 break; 717 } 718 return (NULL); 719 } 720 #else /* !SMP */ 721 722 void 723 smp_rendezvous_cpus(cpuset_t map, 724 void (*setup_func)(void *), 725 void (*action_func)(void *), 726 void (*teardown_func)(void *), 727 void *arg) 728 { 729 /* 730 * In the !SMP case we just need to ensure the same initial conditions 731 * as the SMP case. 732 */ 733 spinlock_enter(); 734 if (setup_func != NULL) 735 setup_func(arg); 736 if (action_func != NULL) 737 action_func(arg); 738 if (teardown_func != NULL) 739 teardown_func(arg); 740 spinlock_exit(); 741 } 742 743 void 744 smp_rendezvous(void (*setup_func)(void *), 745 void (*action_func)(void *), 746 void (*teardown_func)(void *), 747 void *arg) 748 { 749 750 /* Look comments in the smp_rendezvous_cpus() case. */ 751 spinlock_enter(); 752 if (setup_func != NULL) 753 setup_func(arg); 754 if (action_func != NULL) 755 action_func(arg); 756 if (teardown_func != NULL) 757 teardown_func(arg); 758 spinlock_exit(); 759 } 760 761 /* 762 * Provide dummy SMP support for UP kernels. Modules that need to use SMP 763 * APIs will still work using this dummy support. 764 */ 765 static void 766 mp_setvariables_for_up(void *dummy) 767 { 768 mp_ncpus = 1; 769 mp_maxid = PCPU_GET(cpuid); 770 CPU_SETOF(mp_maxid, &all_cpus); 771 KASSERT(PCPU_GET(cpuid) == 0, ("UP must have a CPU ID of zero")); 772 } 773 SYSINIT(cpu_mp_setvariables, SI_SUB_TUNABLES, SI_ORDER_FIRST, 774 mp_setvariables_for_up, NULL); 775 #endif /* SMP */ 776 777 void 778 smp_no_rendevous_barrier(void *dummy) 779 { 780 #ifdef SMP 781 KASSERT((!smp_started),("smp_no_rendevous called and smp is started")); 782 #endif 783 } 784 785 /* 786 * Wait specified idle threads to switch once. This ensures that even 787 * preempted threads have cycled through the switch function once, 788 * exiting their codepaths. This allows us to change global pointers 789 * with no other synchronization. 790 */ 791 int 792 quiesce_cpus(cpuset_t map, const char *wmesg, int prio) 793 { 794 struct pcpu *pcpu; 795 u_int gen[MAXCPU]; 796 int error; 797 int cpu; 798 799 error = 0; 800 for (cpu = 0; cpu <= mp_maxid; cpu++) { 801 if (!CPU_ISSET(cpu, &map) || CPU_ABSENT(cpu)) 802 continue; 803 pcpu = pcpu_find(cpu); 804 gen[cpu] = pcpu->pc_idlethread->td_generation; 805 } 806 for (cpu = 0; cpu <= mp_maxid; cpu++) { 807 if (!CPU_ISSET(cpu, &map) || CPU_ABSENT(cpu)) 808 continue; 809 pcpu = pcpu_find(cpu); 810 thread_lock(curthread); 811 sched_bind(curthread, cpu); 812 thread_unlock(curthread); 813 while (gen[cpu] == pcpu->pc_idlethread->td_generation) { 814 error = tsleep(quiesce_cpus, prio, wmesg, 1); 815 if (error != EWOULDBLOCK) 816 goto out; 817 error = 0; 818 } 819 } 820 out: 821 thread_lock(curthread); 822 sched_unbind(curthread); 823 thread_unlock(curthread); 824 825 return (error); 826 } 827 828 int 829 quiesce_all_cpus(const char *wmesg, int prio) 830 { 831 832 return quiesce_cpus(all_cpus, wmesg, prio); 833 } 834