1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright (c) 2001, John Baldwin <jhb@FreeBSD.org>. 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. Redistributions in binary form must reproduce the above copyright 12 * notice, this list of conditions and the following disclaimer in the 13 * documentation and/or other materials provided with the distribution. 14 * 15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 25 * SUCH DAMAGE. 26 */ 27 28 /* 29 * This module holds the global variables and machine independent functions 30 * used for the kernel SMP support. 31 */ 32 33 #include <sys/cdefs.h> 34 __FBSDID("$FreeBSD$"); 35 36 #include <sys/param.h> 37 #include <sys/systm.h> 38 #include <sys/kernel.h> 39 #include <sys/ktr.h> 40 #include <sys/proc.h> 41 #include <sys/bus.h> 42 #include <sys/lock.h> 43 #include <sys/malloc.h> 44 #include <sys/mutex.h> 45 #include <sys/pcpu.h> 46 #include <sys/sched.h> 47 #include <sys/smp.h> 48 #include <sys/sysctl.h> 49 50 #include <machine/cpu.h> 51 #include <machine/smp.h> 52 53 #include "opt_sched.h" 54 55 #ifdef SMP 56 MALLOC_DEFINE(M_TOPO, "toponodes", "SMP topology data"); 57 58 volatile cpuset_t stopped_cpus; 59 volatile cpuset_t started_cpus; 60 volatile cpuset_t suspended_cpus; 61 cpuset_t hlt_cpus_mask; 62 cpuset_t logical_cpus_mask; 63 64 void (*cpustop_restartfunc)(void); 65 #endif 66 67 static int sysctl_kern_smp_active(SYSCTL_HANDLER_ARGS); 68 69 /* This is used in modules that need to work in both SMP and UP. */ 70 cpuset_t all_cpus; 71 72 int mp_ncpus; 73 /* export this for libkvm consumers. */ 74 int mp_maxcpus = MAXCPU; 75 76 volatile int smp_started; 77 u_int mp_maxid; 78 79 static SYSCTL_NODE(_kern, OID_AUTO, smp, CTLFLAG_RD|CTLFLAG_CAPRD, NULL, 80 "Kernel SMP"); 81 82 SYSCTL_INT(_kern_smp, OID_AUTO, maxid, CTLFLAG_RD|CTLFLAG_CAPRD, &mp_maxid, 0, 83 "Max CPU ID."); 84 85 SYSCTL_INT(_kern_smp, OID_AUTO, maxcpus, CTLFLAG_RD|CTLFLAG_CAPRD, &mp_maxcpus, 86 0, "Max number of CPUs that the system was compiled for."); 87 88 SYSCTL_PROC(_kern_smp, OID_AUTO, active, CTLFLAG_RD|CTLTYPE_INT|CTLFLAG_MPSAFE, 89 NULL, 0, sysctl_kern_smp_active, "I", 90 "Indicates system is running in SMP mode"); 91 92 int smp_disabled = 0; /* has smp been disabled? */ 93 SYSCTL_INT(_kern_smp, OID_AUTO, disabled, CTLFLAG_RDTUN|CTLFLAG_CAPRD, 94 &smp_disabled, 0, "SMP has been disabled from the loader"); 95 96 int smp_cpus = 1; /* how many cpu's running */ 97 SYSCTL_INT(_kern_smp, OID_AUTO, cpus, CTLFLAG_RD|CTLFLAG_CAPRD, &smp_cpus, 0, 98 "Number of CPUs online"); 99 100 int smp_threads_per_core = 1; /* how many SMT threads are running per core */ 101 SYSCTL_INT(_kern_smp, OID_AUTO, threads_per_core, CTLFLAG_RD|CTLFLAG_CAPRD, 102 &smp_threads_per_core, 0, "Number of SMT threads online per core"); 103 104 int mp_ncores = -1; /* how many physical cores running */ 105 SYSCTL_INT(_kern_smp, OID_AUTO, cores, CTLFLAG_RD|CTLFLAG_CAPRD, &mp_ncores, 0, 106 "Number of CPUs online"); 107 108 int smp_topology = 0; /* Which topology we're using. */ 109 SYSCTL_INT(_kern_smp, OID_AUTO, topology, CTLFLAG_RDTUN, &smp_topology, 0, 110 "Topology override setting; 0 is default provided by hardware."); 111 112 #ifdef SMP 113 /* Enable forwarding of a signal to a process running on a different CPU */ 114 static int forward_signal_enabled = 1; 115 SYSCTL_INT(_kern_smp, OID_AUTO, forward_signal_enabled, CTLFLAG_RW, 116 &forward_signal_enabled, 0, 117 "Forwarding of a signal to a process on a different CPU"); 118 119 /* Variables needed for SMP rendezvous. */ 120 static volatile int smp_rv_ncpus; 121 static void (*volatile smp_rv_setup_func)(void *arg); 122 static void (*volatile smp_rv_action_func)(void *arg); 123 static void (*volatile smp_rv_teardown_func)(void *arg); 124 static void *volatile smp_rv_func_arg; 125 static volatile int smp_rv_waiters[4]; 126 127 /* 128 * Shared mutex to restrict busywaits between smp_rendezvous() and 129 * smp(_targeted)_tlb_shootdown(). A deadlock occurs if both of these 130 * functions trigger at once and cause multiple CPUs to busywait with 131 * interrupts disabled. 132 */ 133 struct mtx smp_ipi_mtx; 134 135 /* 136 * Let the MD SMP code initialize mp_maxid very early if it can. 137 */ 138 static void 139 mp_setmaxid(void *dummy) 140 { 141 142 cpu_mp_setmaxid(); 143 144 KASSERT(mp_ncpus >= 1, ("%s: CPU count < 1", __func__)); 145 KASSERT(mp_ncpus > 1 || mp_maxid == 0, 146 ("%s: one CPU but mp_maxid is not zero", __func__)); 147 KASSERT(mp_maxid >= mp_ncpus - 1, 148 ("%s: counters out of sync: max %d, count %d", __func__, 149 mp_maxid, mp_ncpus)); 150 } 151 SYSINIT(cpu_mp_setmaxid, SI_SUB_TUNABLES, SI_ORDER_FIRST, mp_setmaxid, NULL); 152 153 /* 154 * Call the MD SMP initialization code. 155 */ 156 static void 157 mp_start(void *dummy) 158 { 159 160 mtx_init(&smp_ipi_mtx, "smp rendezvous", NULL, MTX_SPIN); 161 162 /* Probe for MP hardware. */ 163 if (smp_disabled != 0 || cpu_mp_probe() == 0) { 164 mp_ncores = 1; 165 mp_ncpus = 1; 166 CPU_SETOF(PCPU_GET(cpuid), &all_cpus); 167 return; 168 } 169 170 cpu_mp_start(); 171 printf("FreeBSD/SMP: Multiprocessor System Detected: %d CPUs\n", 172 mp_ncpus); 173 174 /* Provide a default for most architectures that don't have SMT/HTT. */ 175 if (mp_ncores < 0) 176 mp_ncores = mp_ncpus; 177 178 cpu_mp_announce(); 179 } 180 SYSINIT(cpu_mp, SI_SUB_CPU, SI_ORDER_THIRD, mp_start, NULL); 181 182 void 183 forward_signal(struct thread *td) 184 { 185 int id; 186 187 /* 188 * signotify() has already set TDF_ASTPENDING and TDF_NEEDSIGCHECK on 189 * this thread, so all we need to do is poke it if it is currently 190 * executing so that it executes ast(). 191 */ 192 THREAD_LOCK_ASSERT(td, MA_OWNED); 193 KASSERT(TD_IS_RUNNING(td), 194 ("forward_signal: thread is not TDS_RUNNING")); 195 196 CTR1(KTR_SMP, "forward_signal(%p)", td->td_proc); 197 198 if (!smp_started || cold || KERNEL_PANICKED()) 199 return; 200 if (!forward_signal_enabled) 201 return; 202 203 /* No need to IPI ourself. */ 204 if (td == curthread) 205 return; 206 207 id = td->td_oncpu; 208 if (id == NOCPU) 209 return; 210 ipi_cpu(id, IPI_AST); 211 } 212 213 /* 214 * When called the executing CPU will send an IPI to all other CPUs 215 * requesting that they halt execution. 216 * 217 * Usually (but not necessarily) called with 'other_cpus' as its arg. 218 * 219 * - Signals all CPUs in map to stop. 220 * - Waits for each to stop. 221 * 222 * Returns: 223 * -1: error 224 * 0: NA 225 * 1: ok 226 * 227 */ 228 #if defined(__amd64__) || defined(__i386__) 229 #define X86 1 230 #else 231 #define X86 0 232 #endif 233 static int 234 generic_stop_cpus(cpuset_t map, u_int type) 235 { 236 #ifdef KTR 237 char cpusetbuf[CPUSETBUFSIZ]; 238 #endif 239 static volatile u_int stopping_cpu = NOCPU; 240 int i; 241 volatile cpuset_t *cpus; 242 243 KASSERT( 244 type == IPI_STOP || type == IPI_STOP_HARD 245 #if X86 246 || type == IPI_SUSPEND 247 #endif 248 , ("%s: invalid stop type", __func__)); 249 250 if (!smp_started) 251 return (0); 252 253 CTR2(KTR_SMP, "stop_cpus(%s) with %u type", 254 cpusetobj_strprint(cpusetbuf, &map), type); 255 256 #if X86 257 /* 258 * When suspending, ensure there are are no IPIs in progress. 259 * IPIs that have been issued, but not yet delivered (e.g. 260 * not pending on a vCPU when running under virtualization) 261 * will be lost, violating FreeBSD's assumption of reliable 262 * IPI delivery. 263 */ 264 if (type == IPI_SUSPEND) 265 mtx_lock_spin(&smp_ipi_mtx); 266 #endif 267 268 #if X86 269 if (!nmi_is_broadcast || nmi_kdb_lock == 0) { 270 #endif 271 if (stopping_cpu != PCPU_GET(cpuid)) 272 while (atomic_cmpset_int(&stopping_cpu, NOCPU, 273 PCPU_GET(cpuid)) == 0) 274 while (stopping_cpu != NOCPU) 275 cpu_spinwait(); /* spin */ 276 277 /* send the stop IPI to all CPUs in map */ 278 ipi_selected(map, type); 279 #if X86 280 } 281 #endif 282 283 #if X86 284 if (type == IPI_SUSPEND) 285 cpus = &suspended_cpus; 286 else 287 #endif 288 cpus = &stopped_cpus; 289 290 i = 0; 291 while (!CPU_SUBSET(cpus, &map)) { 292 /* spin */ 293 cpu_spinwait(); 294 i++; 295 if (i == 100000000) { 296 printf("timeout stopping cpus\n"); 297 break; 298 } 299 } 300 301 #if X86 302 if (type == IPI_SUSPEND) 303 mtx_unlock_spin(&smp_ipi_mtx); 304 #endif 305 306 stopping_cpu = NOCPU; 307 return (1); 308 } 309 310 int 311 stop_cpus(cpuset_t map) 312 { 313 314 return (generic_stop_cpus(map, IPI_STOP)); 315 } 316 317 int 318 stop_cpus_hard(cpuset_t map) 319 { 320 321 return (generic_stop_cpus(map, IPI_STOP_HARD)); 322 } 323 324 #if X86 325 int 326 suspend_cpus(cpuset_t map) 327 { 328 329 return (generic_stop_cpus(map, IPI_SUSPEND)); 330 } 331 #endif 332 333 /* 334 * Called by a CPU to restart stopped CPUs. 335 * 336 * Usually (but not necessarily) called with 'stopped_cpus' as its arg. 337 * 338 * - Signals all CPUs in map to restart. 339 * - Waits for each to restart. 340 * 341 * Returns: 342 * -1: error 343 * 0: NA 344 * 1: ok 345 */ 346 static int 347 generic_restart_cpus(cpuset_t map, u_int type) 348 { 349 #ifdef KTR 350 char cpusetbuf[CPUSETBUFSIZ]; 351 #endif 352 volatile cpuset_t *cpus; 353 354 #if X86 355 KASSERT(type == IPI_STOP || type == IPI_STOP_HARD 356 || type == IPI_SUSPEND, ("%s: invalid stop type", __func__)); 357 358 if (!smp_started) 359 return (0); 360 361 CTR1(KTR_SMP, "restart_cpus(%s)", cpusetobj_strprint(cpusetbuf, &map)); 362 363 if (type == IPI_SUSPEND) 364 cpus = &resuming_cpus; 365 else 366 cpus = &stopped_cpus; 367 368 /* signal other cpus to restart */ 369 if (type == IPI_SUSPEND) 370 CPU_COPY_STORE_REL(&map, &toresume_cpus); 371 else 372 CPU_COPY_STORE_REL(&map, &started_cpus); 373 374 /* 375 * Wake up any CPUs stopped with MWAIT. From MI code we can't tell if 376 * MONITOR/MWAIT is enabled, but the potentially redundant writes are 377 * relatively inexpensive. 378 */ 379 if (type == IPI_STOP) { 380 struct monitorbuf *mb; 381 u_int id; 382 383 CPU_FOREACH(id) { 384 if (!CPU_ISSET(id, &map)) 385 continue; 386 387 mb = &pcpu_find(id)->pc_monitorbuf; 388 atomic_store_int(&mb->stop_state, 389 MONITOR_STOPSTATE_RUNNING); 390 } 391 } 392 393 if (!nmi_is_broadcast || nmi_kdb_lock == 0) { 394 /* wait for each to clear its bit */ 395 while (CPU_OVERLAP(cpus, &map)) 396 cpu_spinwait(); 397 } 398 #else /* !X86 */ 399 KASSERT(type == IPI_STOP || type == IPI_STOP_HARD, 400 ("%s: invalid stop type", __func__)); 401 402 if (!smp_started) 403 return (0); 404 405 CTR1(KTR_SMP, "restart_cpus(%s)", cpusetobj_strprint(cpusetbuf, &map)); 406 407 cpus = &stopped_cpus; 408 409 /* signal other cpus to restart */ 410 CPU_COPY_STORE_REL(&map, &started_cpus); 411 412 /* wait for each to clear its bit */ 413 while (CPU_OVERLAP(cpus, &map)) 414 cpu_spinwait(); 415 #endif 416 return (1); 417 } 418 419 int 420 restart_cpus(cpuset_t map) 421 { 422 423 return (generic_restart_cpus(map, IPI_STOP)); 424 } 425 426 #if X86 427 int 428 resume_cpus(cpuset_t map) 429 { 430 431 return (generic_restart_cpus(map, IPI_SUSPEND)); 432 } 433 #endif 434 #undef X86 435 436 /* 437 * All-CPU rendezvous. CPUs are signalled, all execute the setup function 438 * (if specified), rendezvous, execute the action function (if specified), 439 * rendezvous again, execute the teardown function (if specified), and then 440 * resume. 441 * 442 * Note that the supplied external functions _must_ be reentrant and aware 443 * that they are running in parallel and in an unknown lock context. 444 */ 445 void 446 smp_rendezvous_action(void) 447 { 448 struct thread *td; 449 void *local_func_arg; 450 void (*local_setup_func)(void*); 451 void (*local_action_func)(void*); 452 void (*local_teardown_func)(void*); 453 #ifdef INVARIANTS 454 int owepreempt; 455 #endif 456 457 /* Ensure we have up-to-date values. */ 458 atomic_add_acq_int(&smp_rv_waiters[0], 1); 459 while (smp_rv_waiters[0] < smp_rv_ncpus) 460 cpu_spinwait(); 461 462 /* Fetch rendezvous parameters after acquire barrier. */ 463 local_func_arg = smp_rv_func_arg; 464 local_setup_func = smp_rv_setup_func; 465 local_action_func = smp_rv_action_func; 466 local_teardown_func = smp_rv_teardown_func; 467 468 /* 469 * Use a nested critical section to prevent any preemptions 470 * from occurring during a rendezvous action routine. 471 * Specifically, if a rendezvous handler is invoked via an IPI 472 * and the interrupted thread was in the critical_exit() 473 * function after setting td_critnest to 0 but before 474 * performing a deferred preemption, this routine can be 475 * invoked with td_critnest set to 0 and td_owepreempt true. 476 * In that case, a critical_exit() during the rendezvous 477 * action would trigger a preemption which is not permitted in 478 * a rendezvous action. To fix this, wrap all of the 479 * rendezvous action handlers in a critical section. We 480 * cannot use a regular critical section however as having 481 * critical_exit() preempt from this routine would also be 482 * problematic (the preemption must not occur before the IPI 483 * has been acknowledged via an EOI). Instead, we 484 * intentionally ignore td_owepreempt when leaving the 485 * critical section. This should be harmless because we do 486 * not permit rendezvous action routines to schedule threads, 487 * and thus td_owepreempt should never transition from 0 to 1 488 * during this routine. 489 */ 490 td = curthread; 491 td->td_critnest++; 492 #ifdef INVARIANTS 493 owepreempt = td->td_owepreempt; 494 #endif 495 496 /* 497 * If requested, run a setup function before the main action 498 * function. Ensure all CPUs have completed the setup 499 * function before moving on to the action function. 500 */ 501 if (local_setup_func != smp_no_rendezvous_barrier) { 502 if (smp_rv_setup_func != NULL) 503 smp_rv_setup_func(smp_rv_func_arg); 504 atomic_add_int(&smp_rv_waiters[1], 1); 505 while (smp_rv_waiters[1] < smp_rv_ncpus) 506 cpu_spinwait(); 507 } 508 509 if (local_action_func != NULL) 510 local_action_func(local_func_arg); 511 512 if (local_teardown_func != smp_no_rendezvous_barrier) { 513 /* 514 * Signal that the main action has been completed. If a 515 * full exit rendezvous is requested, then all CPUs will 516 * wait here until all CPUs have finished the main action. 517 */ 518 atomic_add_int(&smp_rv_waiters[2], 1); 519 while (smp_rv_waiters[2] < smp_rv_ncpus) 520 cpu_spinwait(); 521 522 if (local_teardown_func != NULL) 523 local_teardown_func(local_func_arg); 524 } 525 526 /* 527 * Signal that the rendezvous is fully completed by this CPU. 528 * This means that no member of smp_rv_* pseudo-structure will be 529 * accessed by this target CPU after this point; in particular, 530 * memory pointed by smp_rv_func_arg. 531 * 532 * The release semantic ensures that all accesses performed by 533 * the current CPU are visible when smp_rendezvous_cpus() 534 * returns, by synchronizing with the 535 * atomic_load_acq_int(&smp_rv_waiters[3]). 536 */ 537 atomic_add_rel_int(&smp_rv_waiters[3], 1); 538 539 td->td_critnest--; 540 KASSERT(owepreempt == td->td_owepreempt, 541 ("rendezvous action changed td_owepreempt")); 542 } 543 544 void 545 smp_rendezvous_cpus(cpuset_t map, 546 void (* setup_func)(void *), 547 void (* action_func)(void *), 548 void (* teardown_func)(void *), 549 void *arg) 550 { 551 int curcpumap, i, ncpus = 0; 552 553 /* See comments in the !SMP case. */ 554 if (!smp_started) { 555 spinlock_enter(); 556 if (setup_func != NULL) 557 setup_func(arg); 558 if (action_func != NULL) 559 action_func(arg); 560 if (teardown_func != NULL) 561 teardown_func(arg); 562 spinlock_exit(); 563 return; 564 } 565 566 /* 567 * Make sure we come here with interrupts enabled. Otherwise we 568 * livelock if smp_ipi_mtx is owned by a thread which sent us an IPI. 569 */ 570 MPASS(curthread->td_md.md_spinlock_count == 0); 571 572 CPU_FOREACH(i) { 573 if (CPU_ISSET(i, &map)) 574 ncpus++; 575 } 576 if (ncpus == 0) 577 panic("ncpus is 0 with non-zero map"); 578 579 mtx_lock_spin(&smp_ipi_mtx); 580 581 /* Pass rendezvous parameters via global variables. */ 582 smp_rv_ncpus = ncpus; 583 smp_rv_setup_func = setup_func; 584 smp_rv_action_func = action_func; 585 smp_rv_teardown_func = teardown_func; 586 smp_rv_func_arg = arg; 587 smp_rv_waiters[1] = 0; 588 smp_rv_waiters[2] = 0; 589 smp_rv_waiters[3] = 0; 590 atomic_store_rel_int(&smp_rv_waiters[0], 0); 591 592 /* 593 * Signal other processors, which will enter the IPI with 594 * interrupts off. 595 */ 596 curcpumap = CPU_ISSET(curcpu, &map); 597 CPU_CLR(curcpu, &map); 598 ipi_selected(map, IPI_RENDEZVOUS); 599 600 /* Check if the current CPU is in the map */ 601 if (curcpumap != 0) 602 smp_rendezvous_action(); 603 604 /* 605 * Ensure that the master CPU waits for all the other 606 * CPUs to finish the rendezvous, so that smp_rv_* 607 * pseudo-structure and the arg are guaranteed to not 608 * be in use. 609 * 610 * Load acquire synchronizes with the release add in 611 * smp_rendezvous_action(), which ensures that our caller sees 612 * all memory actions done by the called functions on other 613 * CPUs. 614 */ 615 while (atomic_load_acq_int(&smp_rv_waiters[3]) < ncpus) 616 cpu_spinwait(); 617 618 mtx_unlock_spin(&smp_ipi_mtx); 619 } 620 621 void 622 smp_rendezvous(void (* setup_func)(void *), 623 void (* action_func)(void *), 624 void (* teardown_func)(void *), 625 void *arg) 626 { 627 smp_rendezvous_cpus(all_cpus, setup_func, action_func, teardown_func, arg); 628 } 629 630 static struct cpu_group group[MAXCPU * MAX_CACHE_LEVELS + 1]; 631 632 struct cpu_group * 633 smp_topo(void) 634 { 635 char cpusetbuf[CPUSETBUFSIZ], cpusetbuf2[CPUSETBUFSIZ]; 636 struct cpu_group *top; 637 638 /* 639 * Check for a fake topology request for debugging purposes. 640 */ 641 switch (smp_topology) { 642 case 1: 643 /* Dual core with no sharing. */ 644 top = smp_topo_1level(CG_SHARE_NONE, 2, 0); 645 break; 646 case 2: 647 /* No topology, all cpus are equal. */ 648 top = smp_topo_none(); 649 break; 650 case 3: 651 /* Dual core with shared L2. */ 652 top = smp_topo_1level(CG_SHARE_L2, 2, 0); 653 break; 654 case 4: 655 /* quad core, shared l3 among each package, private l2. */ 656 top = smp_topo_1level(CG_SHARE_L3, 4, 0); 657 break; 658 case 5: 659 /* quad core, 2 dualcore parts on each package share l2. */ 660 top = smp_topo_2level(CG_SHARE_NONE, 2, CG_SHARE_L2, 2, 0); 661 break; 662 case 6: 663 /* Single-core 2xHTT */ 664 top = smp_topo_1level(CG_SHARE_L1, 2, CG_FLAG_HTT); 665 break; 666 case 7: 667 /* quad core with a shared l3, 8 threads sharing L2. */ 668 top = smp_topo_2level(CG_SHARE_L3, 4, CG_SHARE_L2, 8, 669 CG_FLAG_SMT); 670 break; 671 default: 672 /* Default, ask the system what it wants. */ 673 top = cpu_topo(); 674 break; 675 } 676 /* 677 * Verify the returned topology. 678 */ 679 if (top->cg_count != mp_ncpus) 680 panic("Built bad topology at %p. CPU count %d != %d", 681 top, top->cg_count, mp_ncpus); 682 if (CPU_CMP(&top->cg_mask, &all_cpus)) 683 panic("Built bad topology at %p. CPU mask (%s) != (%s)", 684 top, cpusetobj_strprint(cpusetbuf, &top->cg_mask), 685 cpusetobj_strprint(cpusetbuf2, &all_cpus)); 686 687 /* 688 * Collapse nonsense levels that may be created out of convenience by 689 * the MD layers. They cause extra work in the search functions. 690 */ 691 while (top->cg_children == 1) { 692 top = &top->cg_child[0]; 693 top->cg_parent = NULL; 694 } 695 return (top); 696 } 697 698 struct cpu_group * 699 smp_topo_alloc(u_int count) 700 { 701 static u_int index; 702 u_int curr; 703 704 curr = index; 705 index += count; 706 return (&group[curr]); 707 } 708 709 struct cpu_group * 710 smp_topo_none(void) 711 { 712 struct cpu_group *top; 713 714 top = &group[0]; 715 top->cg_parent = NULL; 716 top->cg_child = NULL; 717 top->cg_mask = all_cpus; 718 top->cg_count = mp_ncpus; 719 top->cg_children = 0; 720 top->cg_level = CG_SHARE_NONE; 721 top->cg_flags = 0; 722 723 return (top); 724 } 725 726 static int 727 smp_topo_addleaf(struct cpu_group *parent, struct cpu_group *child, int share, 728 int count, int flags, int start) 729 { 730 char cpusetbuf[CPUSETBUFSIZ], cpusetbuf2[CPUSETBUFSIZ]; 731 cpuset_t mask; 732 int i; 733 734 CPU_ZERO(&mask); 735 for (i = 0; i < count; i++, start++) 736 CPU_SET(start, &mask); 737 child->cg_parent = parent; 738 child->cg_child = NULL; 739 child->cg_children = 0; 740 child->cg_level = share; 741 child->cg_count = count; 742 child->cg_flags = flags; 743 child->cg_mask = mask; 744 parent->cg_children++; 745 for (; parent != NULL; parent = parent->cg_parent) { 746 if (CPU_OVERLAP(&parent->cg_mask, &child->cg_mask)) 747 panic("Duplicate children in %p. mask (%s) child (%s)", 748 parent, 749 cpusetobj_strprint(cpusetbuf, &parent->cg_mask), 750 cpusetobj_strprint(cpusetbuf2, &child->cg_mask)); 751 CPU_OR(&parent->cg_mask, &child->cg_mask); 752 parent->cg_count += child->cg_count; 753 } 754 755 return (start); 756 } 757 758 struct cpu_group * 759 smp_topo_1level(int share, int count, int flags) 760 { 761 struct cpu_group *child; 762 struct cpu_group *top; 763 int packages; 764 int cpu; 765 int i; 766 767 cpu = 0; 768 top = &group[0]; 769 packages = mp_ncpus / count; 770 top->cg_child = child = &group[1]; 771 top->cg_level = CG_SHARE_NONE; 772 for (i = 0; i < packages; i++, child++) 773 cpu = smp_topo_addleaf(top, child, share, count, flags, cpu); 774 return (top); 775 } 776 777 struct cpu_group * 778 smp_topo_2level(int l2share, int l2count, int l1share, int l1count, 779 int l1flags) 780 { 781 struct cpu_group *top; 782 struct cpu_group *l1g; 783 struct cpu_group *l2g; 784 int cpu; 785 int i; 786 int j; 787 788 cpu = 0; 789 top = &group[0]; 790 l2g = &group[1]; 791 top->cg_child = l2g; 792 top->cg_level = CG_SHARE_NONE; 793 top->cg_children = mp_ncpus / (l2count * l1count); 794 l1g = l2g + top->cg_children; 795 for (i = 0; i < top->cg_children; i++, l2g++) { 796 l2g->cg_parent = top; 797 l2g->cg_child = l1g; 798 l2g->cg_level = l2share; 799 for (j = 0; j < l2count; j++, l1g++) 800 cpu = smp_topo_addleaf(l2g, l1g, l1share, l1count, 801 l1flags, cpu); 802 } 803 return (top); 804 } 805 806 struct cpu_group * 807 smp_topo_find(struct cpu_group *top, int cpu) 808 { 809 struct cpu_group *cg; 810 cpuset_t mask; 811 int children; 812 int i; 813 814 CPU_SETOF(cpu, &mask); 815 cg = top; 816 for (;;) { 817 if (!CPU_OVERLAP(&cg->cg_mask, &mask)) 818 return (NULL); 819 if (cg->cg_children == 0) 820 return (cg); 821 children = cg->cg_children; 822 for (i = 0, cg = cg->cg_child; i < children; cg++, i++) 823 if (CPU_OVERLAP(&cg->cg_mask, &mask)) 824 break; 825 } 826 return (NULL); 827 } 828 #else /* !SMP */ 829 830 void 831 smp_rendezvous_cpus(cpuset_t map, 832 void (*setup_func)(void *), 833 void (*action_func)(void *), 834 void (*teardown_func)(void *), 835 void *arg) 836 { 837 /* 838 * In the !SMP case we just need to ensure the same initial conditions 839 * as the SMP case. 840 */ 841 spinlock_enter(); 842 if (setup_func != NULL) 843 setup_func(arg); 844 if (action_func != NULL) 845 action_func(arg); 846 if (teardown_func != NULL) 847 teardown_func(arg); 848 spinlock_exit(); 849 } 850 851 void 852 smp_rendezvous(void (*setup_func)(void *), 853 void (*action_func)(void *), 854 void (*teardown_func)(void *), 855 void *arg) 856 { 857 858 smp_rendezvous_cpus(all_cpus, setup_func, action_func, teardown_func, 859 arg); 860 } 861 862 /* 863 * Provide dummy SMP support for UP kernels. Modules that need to use SMP 864 * APIs will still work using this dummy support. 865 */ 866 static void 867 mp_setvariables_for_up(void *dummy) 868 { 869 mp_ncpus = 1; 870 mp_ncores = 1; 871 mp_maxid = PCPU_GET(cpuid); 872 CPU_SETOF(mp_maxid, &all_cpus); 873 KASSERT(PCPU_GET(cpuid) == 0, ("UP must have a CPU ID of zero")); 874 } 875 SYSINIT(cpu_mp_setvariables, SI_SUB_TUNABLES, SI_ORDER_FIRST, 876 mp_setvariables_for_up, NULL); 877 #endif /* SMP */ 878 879 void 880 smp_no_rendezvous_barrier(void *dummy) 881 { 882 #ifdef SMP 883 KASSERT((!smp_started),("smp_no_rendezvous called and smp is started")); 884 #endif 885 } 886 887 void 888 smp_rendezvous_cpus_retry(cpuset_t map, 889 void (* setup_func)(void *), 890 void (* action_func)(void *), 891 void (* teardown_func)(void *), 892 void (* wait_func)(void *, int), 893 struct smp_rendezvous_cpus_retry_arg *arg) 894 { 895 int cpu; 896 897 /* 898 * Execute an action on all specified CPUs while retrying until they 899 * all acknowledge completion. 900 */ 901 CPU_COPY(&map, &arg->cpus); 902 for (;;) { 903 smp_rendezvous_cpus( 904 arg->cpus, 905 setup_func, 906 action_func, 907 teardown_func, 908 arg); 909 910 if (CPU_EMPTY(&arg->cpus)) 911 break; 912 913 CPU_FOREACH(cpu) { 914 if (!CPU_ISSET(cpu, &arg->cpus)) 915 continue; 916 wait_func(arg, cpu); 917 } 918 } 919 } 920 921 void 922 smp_rendezvous_cpus_done(struct smp_rendezvous_cpus_retry_arg *arg) 923 { 924 925 CPU_CLR_ATOMIC(curcpu, &arg->cpus); 926 } 927 928 /* 929 * Wait for specified idle threads to switch once. This ensures that even 930 * preempted threads have cycled through the switch function once, 931 * exiting their codepaths. This allows us to change global pointers 932 * with no other synchronization. 933 */ 934 int 935 quiesce_cpus(cpuset_t map, const char *wmesg, int prio) 936 { 937 struct pcpu *pcpu; 938 u_int gen[MAXCPU]; 939 int error; 940 int cpu; 941 942 error = 0; 943 for (cpu = 0; cpu <= mp_maxid; cpu++) { 944 if (!CPU_ISSET(cpu, &map) || CPU_ABSENT(cpu)) 945 continue; 946 pcpu = pcpu_find(cpu); 947 gen[cpu] = pcpu->pc_idlethread->td_generation; 948 } 949 for (cpu = 0; cpu <= mp_maxid; cpu++) { 950 if (!CPU_ISSET(cpu, &map) || CPU_ABSENT(cpu)) 951 continue; 952 pcpu = pcpu_find(cpu); 953 thread_lock(curthread); 954 sched_bind(curthread, cpu); 955 thread_unlock(curthread); 956 while (gen[cpu] == pcpu->pc_idlethread->td_generation) { 957 error = tsleep(quiesce_cpus, prio, wmesg, 1); 958 if (error != EWOULDBLOCK) 959 goto out; 960 error = 0; 961 } 962 } 963 out: 964 thread_lock(curthread); 965 sched_unbind(curthread); 966 thread_unlock(curthread); 967 968 return (error); 969 } 970 971 int 972 quiesce_all_cpus(const char *wmesg, int prio) 973 { 974 975 return quiesce_cpus(all_cpus, wmesg, prio); 976 } 977 978 /* 979 * Observe all CPUs not executing in critical section. 980 * We are not in one so the check for us is safe. If the found 981 * thread changes to something else we know the section was 982 * exited as well. 983 */ 984 void 985 quiesce_all_critical(void) 986 { 987 struct thread *td, *newtd; 988 struct pcpu *pcpu; 989 int cpu; 990 991 MPASS(curthread->td_critnest == 0); 992 993 CPU_FOREACH(cpu) { 994 pcpu = cpuid_to_pcpu[cpu]; 995 td = pcpu->pc_curthread; 996 for (;;) { 997 if (td->td_critnest == 0) 998 break; 999 cpu_spinwait(); 1000 newtd = (struct thread *) 1001 atomic_load_acq_ptr((void *)pcpu->pc_curthread); 1002 if (td != newtd) 1003 break; 1004 } 1005 } 1006 } 1007 1008 static void 1009 cpus_fence_seq_cst_issue(void *arg __unused) 1010 { 1011 1012 atomic_thread_fence_seq_cst(); 1013 } 1014 1015 /* 1016 * Send an IPI forcing a sequentially consistent fence. 1017 * 1018 * Allows replacement of an explicitly fence with a compiler barrier. 1019 * Trades speed up during normal execution for a significant slowdown when 1020 * the barrier is needed. 1021 */ 1022 void 1023 cpus_fence_seq_cst(void) 1024 { 1025 1026 #ifdef SMP 1027 smp_rendezvous( 1028 smp_no_rendezvous_barrier, 1029 cpus_fence_seq_cst_issue, 1030 smp_no_rendezvous_barrier, 1031 NULL 1032 ); 1033 #else 1034 cpus_fence_seq_cst_issue(NULL); 1035 #endif 1036 } 1037 1038 /* Extra care is taken with this sysctl because the data type is volatile */ 1039 static int 1040 sysctl_kern_smp_active(SYSCTL_HANDLER_ARGS) 1041 { 1042 int error, active; 1043 1044 active = smp_started; 1045 error = SYSCTL_OUT(req, &active, sizeof(active)); 1046 return (error); 1047 } 1048 1049 #ifdef SMP 1050 void 1051 topo_init_node(struct topo_node *node) 1052 { 1053 1054 bzero(node, sizeof(*node)); 1055 TAILQ_INIT(&node->children); 1056 } 1057 1058 void 1059 topo_init_root(struct topo_node *root) 1060 { 1061 1062 topo_init_node(root); 1063 root->type = TOPO_TYPE_SYSTEM; 1064 } 1065 1066 /* 1067 * Add a child node with the given ID under the given parent. 1068 * Do nothing if there is already a child with that ID. 1069 */ 1070 struct topo_node * 1071 topo_add_node_by_hwid(struct topo_node *parent, int hwid, 1072 topo_node_type type, uintptr_t subtype) 1073 { 1074 struct topo_node *node; 1075 1076 TAILQ_FOREACH_REVERSE(node, &parent->children, 1077 topo_children, siblings) { 1078 if (node->hwid == hwid 1079 && node->type == type && node->subtype == subtype) { 1080 return (node); 1081 } 1082 } 1083 1084 node = malloc(sizeof(*node), M_TOPO, M_WAITOK); 1085 topo_init_node(node); 1086 node->parent = parent; 1087 node->hwid = hwid; 1088 node->type = type; 1089 node->subtype = subtype; 1090 TAILQ_INSERT_TAIL(&parent->children, node, siblings); 1091 parent->nchildren++; 1092 1093 return (node); 1094 } 1095 1096 /* 1097 * Find a child node with the given ID under the given parent. 1098 */ 1099 struct topo_node * 1100 topo_find_node_by_hwid(struct topo_node *parent, int hwid, 1101 topo_node_type type, uintptr_t subtype) 1102 { 1103 1104 struct topo_node *node; 1105 1106 TAILQ_FOREACH(node, &parent->children, siblings) { 1107 if (node->hwid == hwid 1108 && node->type == type && node->subtype == subtype) { 1109 return (node); 1110 } 1111 } 1112 1113 return (NULL); 1114 } 1115 1116 /* 1117 * Given a node change the order of its parent's child nodes such 1118 * that the node becomes the firt child while preserving the cyclic 1119 * order of the children. In other words, the given node is promoted 1120 * by rotation. 1121 */ 1122 void 1123 topo_promote_child(struct topo_node *child) 1124 { 1125 struct topo_node *next; 1126 struct topo_node *node; 1127 struct topo_node *parent; 1128 1129 parent = child->parent; 1130 next = TAILQ_NEXT(child, siblings); 1131 TAILQ_REMOVE(&parent->children, child, siblings); 1132 TAILQ_INSERT_HEAD(&parent->children, child, siblings); 1133 1134 while (next != NULL) { 1135 node = next; 1136 next = TAILQ_NEXT(node, siblings); 1137 TAILQ_REMOVE(&parent->children, node, siblings); 1138 TAILQ_INSERT_AFTER(&parent->children, child, node, siblings); 1139 child = node; 1140 } 1141 } 1142 1143 /* 1144 * Iterate to the next node in the depth-first search (traversal) of 1145 * the topology tree. 1146 */ 1147 struct topo_node * 1148 topo_next_node(struct topo_node *top, struct topo_node *node) 1149 { 1150 struct topo_node *next; 1151 1152 if ((next = TAILQ_FIRST(&node->children)) != NULL) 1153 return (next); 1154 1155 if ((next = TAILQ_NEXT(node, siblings)) != NULL) 1156 return (next); 1157 1158 while (node != top && (node = node->parent) != top) 1159 if ((next = TAILQ_NEXT(node, siblings)) != NULL) 1160 return (next); 1161 1162 return (NULL); 1163 } 1164 1165 /* 1166 * Iterate to the next node in the depth-first search of the topology tree, 1167 * but without descending below the current node. 1168 */ 1169 struct topo_node * 1170 topo_next_nonchild_node(struct topo_node *top, struct topo_node *node) 1171 { 1172 struct topo_node *next; 1173 1174 if ((next = TAILQ_NEXT(node, siblings)) != NULL) 1175 return (next); 1176 1177 while (node != top && (node = node->parent) != top) 1178 if ((next = TAILQ_NEXT(node, siblings)) != NULL) 1179 return (next); 1180 1181 return (NULL); 1182 } 1183 1184 /* 1185 * Assign the given ID to the given topology node that represents a logical 1186 * processor. 1187 */ 1188 void 1189 topo_set_pu_id(struct topo_node *node, cpuid_t id) 1190 { 1191 1192 KASSERT(node->type == TOPO_TYPE_PU, 1193 ("topo_set_pu_id: wrong node type: %u", node->type)); 1194 KASSERT(CPU_EMPTY(&node->cpuset) && node->cpu_count == 0, 1195 ("topo_set_pu_id: cpuset already not empty")); 1196 node->id = id; 1197 CPU_SET(id, &node->cpuset); 1198 node->cpu_count = 1; 1199 node->subtype = 1; 1200 1201 while ((node = node->parent) != NULL) { 1202 KASSERT(!CPU_ISSET(id, &node->cpuset), 1203 ("logical ID %u is already set in node %p", id, node)); 1204 CPU_SET(id, &node->cpuset); 1205 node->cpu_count++; 1206 } 1207 } 1208 1209 static struct topology_spec { 1210 topo_node_type type; 1211 bool match_subtype; 1212 uintptr_t subtype; 1213 } topology_level_table[TOPO_LEVEL_COUNT] = { 1214 [TOPO_LEVEL_PKG] = { .type = TOPO_TYPE_PKG, }, 1215 [TOPO_LEVEL_GROUP] = { .type = TOPO_TYPE_GROUP, }, 1216 [TOPO_LEVEL_CACHEGROUP] = { 1217 .type = TOPO_TYPE_CACHE, 1218 .match_subtype = true, 1219 .subtype = CG_SHARE_L3, 1220 }, 1221 [TOPO_LEVEL_CORE] = { .type = TOPO_TYPE_CORE, }, 1222 [TOPO_LEVEL_THREAD] = { .type = TOPO_TYPE_PU, }, 1223 }; 1224 1225 static bool 1226 topo_analyze_table(struct topo_node *root, int all, enum topo_level level, 1227 struct topo_analysis *results) 1228 { 1229 struct topology_spec *spec; 1230 struct topo_node *node; 1231 int count; 1232 1233 if (level >= TOPO_LEVEL_COUNT) 1234 return (true); 1235 1236 spec = &topology_level_table[level]; 1237 count = 0; 1238 node = topo_next_node(root, root); 1239 1240 while (node != NULL) { 1241 if (node->type != spec->type || 1242 (spec->match_subtype && node->subtype != spec->subtype)) { 1243 node = topo_next_node(root, node); 1244 continue; 1245 } 1246 if (!all && CPU_EMPTY(&node->cpuset)) { 1247 node = topo_next_nonchild_node(root, node); 1248 continue; 1249 } 1250 1251 count++; 1252 1253 if (!topo_analyze_table(node, all, level + 1, results)) 1254 return (false); 1255 1256 node = topo_next_nonchild_node(root, node); 1257 } 1258 1259 /* No explicit subgroups is essentially one subgroup. */ 1260 if (count == 0) { 1261 count = 1; 1262 1263 if (!topo_analyze_table(root, all, level + 1, results)) 1264 return (false); 1265 } 1266 1267 if (results->entities[level] == -1) 1268 results->entities[level] = count; 1269 else if (results->entities[level] != count) 1270 return (false); 1271 1272 return (true); 1273 } 1274 1275 /* 1276 * Check if the topology is uniform, that is, each package has the same number 1277 * of cores in it and each core has the same number of threads (logical 1278 * processors) in it. If so, calculate the number of packages, the number of 1279 * groups per package, the number of cachegroups per group, and the number of 1280 * logical processors per cachegroup. 'all' parameter tells whether to include 1281 * administratively disabled logical processors into the analysis. 1282 */ 1283 int 1284 topo_analyze(struct topo_node *topo_root, int all, 1285 struct topo_analysis *results) 1286 { 1287 1288 results->entities[TOPO_LEVEL_PKG] = -1; 1289 results->entities[TOPO_LEVEL_CORE] = -1; 1290 results->entities[TOPO_LEVEL_THREAD] = -1; 1291 results->entities[TOPO_LEVEL_GROUP] = -1; 1292 results->entities[TOPO_LEVEL_CACHEGROUP] = -1; 1293 1294 if (!topo_analyze_table(topo_root, all, TOPO_LEVEL_PKG, results)) 1295 return (0); 1296 1297 KASSERT(results->entities[TOPO_LEVEL_PKG] > 0, 1298 ("bug in topology or analysis")); 1299 1300 return (1); 1301 } 1302 1303 #endif /* SMP */ 1304 1305