1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause 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/param.h> 34 #include <sys/systm.h> 35 #include <sys/kernel.h> 36 #include <sys/ktr.h> 37 #include <sys/proc.h> 38 #include <sys/bus.h> 39 #include <sys/lock.h> 40 #include <sys/malloc.h> 41 #include <sys/mutex.h> 42 #include <sys/pcpu.h> 43 #include <sys/sched.h> 44 #include <sys/smp.h> 45 #include <sys/sysctl.h> 46 47 #include <machine/cpu.h> 48 #include <machine/pcb.h> 49 #include <machine/smp.h> 50 51 #include "opt_sched.h" 52 53 #ifdef SMP 54 MALLOC_DEFINE(M_TOPO, "toponodes", "SMP topology data"); 55 56 volatile cpuset_t stopped_cpus; 57 volatile cpuset_t started_cpus; 58 volatile cpuset_t suspended_cpus; 59 cpuset_t hlt_cpus_mask; 60 cpuset_t logical_cpus_mask; 61 62 void (*cpustop_restartfunc)(void); 63 #endif 64 65 static int sysctl_kern_smp_active(SYSCTL_HANDLER_ARGS); 66 67 /* This is used in modules that need to work in both SMP and UP. */ 68 cpuset_t all_cpus; 69 70 int mp_ncpus; 71 /* export this for libkvm consumers. */ 72 int mp_maxcpus = MAXCPU; 73 74 volatile int smp_started; 75 u_int mp_maxid; 76 77 /* Array of CPU contexts saved during a panic. */ 78 struct pcb *stoppcbs; 79 80 static SYSCTL_NODE(_kern, OID_AUTO, smp, 81 CTLFLAG_RD | CTLFLAG_CAPRD | CTLFLAG_MPSAFE, NULL, 82 "Kernel SMP"); 83 84 SYSCTL_INT(_kern_smp, OID_AUTO, maxid, CTLFLAG_RD|CTLFLAG_CAPRD, &mp_maxid, 0, 85 "Max CPU ID."); 86 87 SYSCTL_INT(_kern_smp, OID_AUTO, maxcpus, CTLFLAG_RD|CTLFLAG_CAPRD, &mp_maxcpus, 88 0, "Max number of CPUs that the system was compiled for."); 89 90 SYSCTL_PROC(_kern_smp, OID_AUTO, active, CTLFLAG_RD|CTLTYPE_INT|CTLFLAG_MPSAFE, 91 NULL, 0, sysctl_kern_smp_active, "I", 92 "Indicates system is running in SMP mode"); 93 94 int smp_disabled = 0; /* has smp been disabled? */ 95 SYSCTL_INT(_kern_smp, OID_AUTO, disabled, CTLFLAG_RDTUN|CTLFLAG_CAPRD, 96 &smp_disabled, 0, "SMP has been disabled from the loader"); 97 98 int smp_cpus = 1; /* how many cpu's running */ 99 SYSCTL_INT(_kern_smp, OID_AUTO, cpus, CTLFLAG_RD|CTLFLAG_CAPRD, &smp_cpus, 0, 100 "Number of CPUs online"); 101 102 int smp_threads_per_core = 1; /* how many SMT threads are running per core */ 103 SYSCTL_INT(_kern_smp, OID_AUTO, threads_per_core, CTLFLAG_RD|CTLFLAG_CAPRD, 104 &smp_threads_per_core, 0, "Number of SMT threads online per core"); 105 106 int mp_ncores = -1; /* how many physical cores running */ 107 SYSCTL_INT(_kern_smp, OID_AUTO, cores, CTLFLAG_RD|CTLFLAG_CAPRD, &mp_ncores, 0, 108 "Number of physical cores online"); 109 110 int smp_topology = 0; /* Which topology we're using. */ 111 SYSCTL_INT(_kern_smp, OID_AUTO, topology, CTLFLAG_RDTUN, &smp_topology, 0, 112 "Topology override setting; 0 is default provided by hardware."); 113 114 #ifdef SMP 115 /* Variables needed for SMP rendezvous. */ 116 static volatile int smp_rv_ncpus; 117 static void (*volatile smp_rv_setup_func)(void *arg); 118 static void (*volatile smp_rv_action_func)(void *arg); 119 static void (*volatile smp_rv_teardown_func)(void *arg); 120 static void *volatile smp_rv_func_arg; 121 static volatile int smp_rv_waiters[4]; 122 123 /* 124 * Shared mutex to restrict busywaits between smp_rendezvous() and 125 * smp(_targeted)_tlb_shootdown(). A deadlock occurs if both of these 126 * functions trigger at once and cause multiple CPUs to busywait with 127 * interrupts disabled. 128 */ 129 struct mtx smp_ipi_mtx; 130 131 /* 132 * Let the MD SMP code initialize mp_maxid very early if it can. 133 */ 134 static void 135 mp_setmaxid(void *dummy) 136 { 137 138 cpu_mp_setmaxid(); 139 140 KASSERT(mp_ncpus >= 1, ("%s: CPU count < 1", __func__)); 141 KASSERT(mp_ncpus > 1 || mp_maxid == 0, 142 ("%s: one CPU but mp_maxid is not zero", __func__)); 143 KASSERT(mp_maxid >= mp_ncpus - 1, 144 ("%s: counters out of sync: max %d, count %d", __func__, 145 mp_maxid, mp_ncpus)); 146 147 cpusetsizemin = howmany(mp_maxid + 1, NBBY); 148 } 149 SYSINIT(cpu_mp_setmaxid, SI_SUB_TUNABLES, SI_ORDER_FIRST, mp_setmaxid, NULL); 150 151 /* 152 * Call the MD SMP initialization code. 153 */ 154 static void 155 mp_start(void *dummy) 156 { 157 158 mtx_init(&smp_ipi_mtx, "smp rendezvous", NULL, MTX_SPIN); 159 160 /* Probe for MP hardware. */ 161 if (smp_disabled != 0 || cpu_mp_probe() == 0) { 162 mp_ncores = 1; 163 mp_ncpus = 1; 164 CPU_SETOF(PCPU_GET(cpuid), &all_cpus); 165 return; 166 } 167 168 cpu_mp_start(); 169 printf("FreeBSD/SMP: Multiprocessor System Detected: %d CPUs\n", 170 mp_ncpus); 171 172 /* Provide a default for most architectures that don't have SMT/HTT. */ 173 if (mp_ncores < 0) 174 mp_ncores = mp_ncpus; 175 176 stoppcbs = mallocarray(mp_maxid + 1, sizeof(struct pcb), M_DEVBUF, 177 M_WAITOK | M_ZERO); 178 179 cpu_mp_announce(); 180 } 181 SYSINIT(cpu_mp, SI_SUB_CPU, SI_ORDER_THIRD, mp_start, NULL); 182 183 void 184 forward_signal(struct thread *td) 185 { 186 int id; 187 188 /* 189 * signotify() has already set TDA_AST and TDA_SIG on td_ast for 190 * this thread, so all we need to do is poke it if it is currently 191 * executing so that it executes ast(). 192 */ 193 THREAD_LOCK_ASSERT(td, MA_OWNED); 194 KASSERT(TD_IS_RUNNING(td), 195 ("forward_signal: thread is not TDS_RUNNING")); 196 197 CTR1(KTR_SMP, "forward_signal(%p)", td->td_proc); 198 199 if (!smp_started || cold || KERNEL_PANICKED()) 200 return; 201 202 /* No need to IPI ourself. */ 203 if (td == curthread) 204 return; 205 206 id = td->td_oncpu; 207 if (id == NOCPU) 208 return; 209 ipi_cpu(id, IPI_AST); 210 } 211 212 /* 213 * When called the executing CPU will send an IPI to all other CPUs 214 * requesting that they halt execution. 215 * 216 * Usually (but not necessarily) called with 'other_cpus' as its arg. 217 * 218 * - Signals all CPUs in map to stop. 219 * - Waits for each to stop. 220 * 221 * Returns: 222 * -1: error 223 * 0: NA 224 * 1: ok 225 * 226 */ 227 #if defined(__amd64__) || defined(__i386__) 228 #define X86 1 229 #else 230 #define X86 0 231 #endif 232 static int 233 generic_stop_cpus(cpuset_t map, u_int type) 234 { 235 #ifdef KTR 236 char cpusetbuf[CPUSETBUFSIZ]; 237 #endif 238 static volatile u_int stopping_cpu = NOCPU; 239 int i; 240 volatile cpuset_t *cpus; 241 242 KASSERT( 243 type == IPI_STOP || type == IPI_STOP_HARD 244 #if X86 245 || type == IPI_SUSPEND || type == IPI_OFF 246 #endif 247 , ("%s: invalid stop type", __func__)); 248 249 if (!smp_started) 250 return (0); 251 252 CTR2(KTR_SMP, "stop_cpus(%s) with %u type", 253 cpusetobj_strprint(cpusetbuf, &map), type); 254 255 #if X86 256 /* 257 * When suspending, ensure there are are no IPIs in progress. 258 * IPIs that have been issued, but not yet delivered (e.g. 259 * not pending on a vCPU when running under virtualization) 260 * will be lost, violating FreeBSD's assumption of reliable 261 * IPI delivery. 262 */ 263 if (type == IPI_SUSPEND || type == IPI_OFF) 264 mtx_lock_spin(&smp_ipi_mtx); 265 #endif 266 267 #if X86 268 if (!nmi_is_broadcast || nmi_kdb_lock == 0) { 269 #endif 270 if (stopping_cpu != PCPU_GET(cpuid)) 271 while (atomic_cmpset_int(&stopping_cpu, NOCPU, 272 PCPU_GET(cpuid)) == 0) 273 while (stopping_cpu != NOCPU) 274 cpu_spinwait(); /* spin */ 275 276 /* send the stop IPI to all CPUs in map */ 277 ipi_selected(map, type); 278 #if X86 279 } 280 #endif 281 282 #if X86 283 if (type == IPI_SUSPEND || type == IPI_OFF) 284 cpus = &suspended_cpus; 285 else 286 #endif 287 cpus = &stopped_cpus; 288 289 i = 0; 290 while (!CPU_SUBSET(cpus, &map)) { 291 /* spin */ 292 cpu_spinwait(); 293 i++; 294 if (i == 100000000) { 295 printf("timeout stopping cpus\n"); 296 break; 297 } 298 } 299 300 #if X86 301 if (type == IPI_SUSPEND || type == IPI_OFF) 302 mtx_unlock_spin(&smp_ipi_mtx); 303 #endif 304 305 stopping_cpu = NOCPU; 306 return (1); 307 } 308 309 int 310 stop_cpus(cpuset_t map) 311 { 312 313 return (generic_stop_cpus(map, IPI_STOP)); 314 } 315 316 int 317 stop_cpus_hard(cpuset_t map) 318 { 319 320 return (generic_stop_cpus(map, IPI_STOP_HARD)); 321 } 322 323 #if X86 324 int 325 suspend_cpus(cpuset_t map) 326 { 327 328 return (generic_stop_cpus(map, IPI_SUSPEND)); 329 } 330 331 int 332 offline_cpus(cpuset_t map) 333 { 334 335 return (generic_stop_cpus(map, IPI_OFF)); 336 } 337 #endif 338 339 /* 340 * Called by a CPU to restart stopped CPUs. 341 * 342 * Usually (but not necessarily) called with 'stopped_cpus' as its arg. 343 * 344 * - Signals all CPUs in map to restart. 345 * - Waits for each to restart. 346 * 347 * Returns: 348 * -1: error 349 * 0: NA 350 * 1: ok 351 */ 352 static int 353 generic_restart_cpus(cpuset_t map, u_int type) 354 { 355 #ifdef KTR 356 char cpusetbuf[CPUSETBUFSIZ]; 357 #endif 358 volatile cpuset_t *cpus; 359 360 #if X86 361 KASSERT(type == IPI_STOP || type == IPI_STOP_HARD 362 || type == IPI_SUSPEND, ("%s: invalid stop type", __func__)); 363 364 if (!smp_started) 365 return (0); 366 367 CTR1(KTR_SMP, "restart_cpus(%s)", cpusetobj_strprint(cpusetbuf, &map)); 368 369 if (type == IPI_SUSPEND) 370 cpus = &resuming_cpus; 371 else 372 cpus = &stopped_cpus; 373 374 /* signal other cpus to restart */ 375 if (type == IPI_SUSPEND) 376 CPU_COPY_STORE_REL(&map, &toresume_cpus); 377 else 378 CPU_COPY_STORE_REL(&map, &started_cpus); 379 380 /* 381 * Wake up any CPUs stopped with MWAIT. From MI code we can't tell if 382 * MONITOR/MWAIT is enabled, but the potentially redundant writes are 383 * relatively inexpensive. 384 */ 385 if (type == IPI_STOP) { 386 struct monitorbuf *mb; 387 u_int id; 388 389 CPU_FOREACH(id) { 390 if (!CPU_ISSET(id, &map)) 391 continue; 392 393 mb = &pcpu_find(id)->pc_monitorbuf; 394 atomic_store_int(&mb->stop_state, 395 MONITOR_STOPSTATE_RUNNING); 396 } 397 } 398 399 if (!nmi_is_broadcast || nmi_kdb_lock == 0) { 400 /* wait for each to clear its bit */ 401 while (CPU_OVERLAP(cpus, &map)) 402 cpu_spinwait(); 403 } 404 #else /* !X86 */ 405 KASSERT(type == IPI_STOP || type == IPI_STOP_HARD, 406 ("%s: invalid stop type", __func__)); 407 408 if (!smp_started) 409 return (0); 410 411 CTR1(KTR_SMP, "restart_cpus(%s)", cpusetobj_strprint(cpusetbuf, &map)); 412 413 cpus = &stopped_cpus; 414 415 /* signal other cpus to restart */ 416 CPU_COPY_STORE_REL(&map, &started_cpus); 417 418 /* wait for each to clear its bit */ 419 while (CPU_OVERLAP(cpus, &map)) 420 cpu_spinwait(); 421 #endif 422 return (1); 423 } 424 425 int 426 restart_cpus(cpuset_t map) 427 { 428 429 return (generic_restart_cpus(map, IPI_STOP)); 430 } 431 432 #if X86 433 int 434 resume_cpus(cpuset_t map) 435 { 436 437 return (generic_restart_cpus(map, IPI_SUSPEND)); 438 } 439 #endif 440 #undef X86 441 442 /* 443 * All-CPU rendezvous. CPUs are signalled, all execute the setup function 444 * (if specified), rendezvous, execute the action function (if specified), 445 * rendezvous again, execute the teardown function (if specified), and then 446 * resume. 447 * 448 * Note that the supplied external functions _must_ be reentrant and aware 449 * that they are running in parallel and in an unknown lock context. 450 */ 451 void 452 smp_rendezvous_action(void) 453 { 454 struct thread *td; 455 void *local_func_arg; 456 void (*local_setup_func)(void*); 457 void (*local_action_func)(void*); 458 void (*local_teardown_func)(void*); 459 #ifdef INVARIANTS 460 int owepreempt; 461 #endif 462 463 /* Ensure we have up-to-date values. */ 464 atomic_add_acq_int(&smp_rv_waiters[0], 1); 465 while (smp_rv_waiters[0] < smp_rv_ncpus) 466 cpu_spinwait(); 467 468 /* Fetch rendezvous parameters after acquire barrier. */ 469 local_func_arg = smp_rv_func_arg; 470 local_setup_func = smp_rv_setup_func; 471 local_action_func = smp_rv_action_func; 472 local_teardown_func = smp_rv_teardown_func; 473 474 /* 475 * Use a nested critical section to prevent any preemptions 476 * from occurring during a rendezvous action routine. 477 * Specifically, if a rendezvous handler is invoked via an IPI 478 * and the interrupted thread was in the critical_exit() 479 * function after setting td_critnest to 0 but before 480 * performing a deferred preemption, this routine can be 481 * invoked with td_critnest set to 0 and td_owepreempt true. 482 * In that case, a critical_exit() during the rendezvous 483 * action would trigger a preemption which is not permitted in 484 * a rendezvous action. To fix this, wrap all of the 485 * rendezvous action handlers in a critical section. We 486 * cannot use a regular critical section however as having 487 * critical_exit() preempt from this routine would also be 488 * problematic (the preemption must not occur before the IPI 489 * has been acknowledged via an EOI). Instead, we 490 * intentionally ignore td_owepreempt when leaving the 491 * critical section. This should be harmless because we do 492 * not permit rendezvous action routines to schedule threads, 493 * and thus td_owepreempt should never transition from 0 to 1 494 * during this routine. 495 */ 496 td = curthread; 497 td->td_critnest++; 498 #ifdef INVARIANTS 499 owepreempt = td->td_owepreempt; 500 #endif 501 502 /* 503 * If requested, run a setup function before the main action 504 * function. Ensure all CPUs have completed the setup 505 * function before moving on to the action function. 506 */ 507 if (local_setup_func != smp_no_rendezvous_barrier) { 508 if (local_setup_func != NULL) 509 local_setup_func(local_func_arg); 510 atomic_add_int(&smp_rv_waiters[1], 1); 511 while (smp_rv_waiters[1] < smp_rv_ncpus) 512 cpu_spinwait(); 513 } 514 515 if (local_action_func != NULL) 516 local_action_func(local_func_arg); 517 518 if (local_teardown_func != smp_no_rendezvous_barrier) { 519 /* 520 * Signal that the main action has been completed. If a 521 * full exit rendezvous is requested, then all CPUs will 522 * wait here until all CPUs have finished the main action. 523 */ 524 atomic_add_int(&smp_rv_waiters[2], 1); 525 while (smp_rv_waiters[2] < smp_rv_ncpus) 526 cpu_spinwait(); 527 528 if (local_teardown_func != NULL) 529 local_teardown_func(local_func_arg); 530 } 531 532 /* 533 * Signal that the rendezvous is fully completed by this CPU. 534 * This means that no member of smp_rv_* pseudo-structure will be 535 * accessed by this target CPU after this point; in particular, 536 * memory pointed by smp_rv_func_arg. 537 * 538 * The release semantic ensures that all accesses performed by 539 * the current CPU are visible when smp_rendezvous_cpus() 540 * returns, by synchronizing with the 541 * atomic_load_acq_int(&smp_rv_waiters[3]). 542 */ 543 atomic_add_rel_int(&smp_rv_waiters[3], 1); 544 545 td->td_critnest--; 546 KASSERT(owepreempt == td->td_owepreempt, 547 ("rendezvous action changed td_owepreempt")); 548 } 549 550 void 551 smp_rendezvous_cpus(cpuset_t map, 552 void (* setup_func)(void *), 553 void (* action_func)(void *), 554 void (* teardown_func)(void *), 555 void *arg) 556 { 557 int curcpumap, i, ncpus = 0; 558 559 /* See comments in the !SMP case. */ 560 if (!smp_started) { 561 spinlock_enter(); 562 if (setup_func != NULL) 563 setup_func(arg); 564 if (action_func != NULL) 565 action_func(arg); 566 if (teardown_func != NULL) 567 teardown_func(arg); 568 spinlock_exit(); 569 return; 570 } 571 572 /* 573 * Make sure we come here with interrupts enabled. Otherwise we 574 * livelock if smp_ipi_mtx is owned by a thread which sent us an IPI. 575 */ 576 MPASS(curthread->td_md.md_spinlock_count == 0); 577 578 CPU_FOREACH(i) { 579 if (CPU_ISSET(i, &map)) 580 ncpus++; 581 } 582 if (ncpus == 0) 583 panic("ncpus is 0 with non-zero map"); 584 585 mtx_lock_spin(&smp_ipi_mtx); 586 587 /* Pass rendezvous parameters via global variables. */ 588 smp_rv_ncpus = ncpus; 589 smp_rv_setup_func = setup_func; 590 smp_rv_action_func = action_func; 591 smp_rv_teardown_func = teardown_func; 592 smp_rv_func_arg = arg; 593 smp_rv_waiters[1] = 0; 594 smp_rv_waiters[2] = 0; 595 smp_rv_waiters[3] = 0; 596 atomic_store_rel_int(&smp_rv_waiters[0], 0); 597 598 /* 599 * Signal other processors, which will enter the IPI with 600 * interrupts off. 601 */ 602 curcpumap = CPU_ISSET(curcpu, &map); 603 CPU_CLR(curcpu, &map); 604 ipi_selected(map, IPI_RENDEZVOUS); 605 606 /* Check if the current CPU is in the map */ 607 if (curcpumap != 0) 608 smp_rendezvous_action(); 609 610 /* 611 * Ensure that the master CPU waits for all the other 612 * CPUs to finish the rendezvous, so that smp_rv_* 613 * pseudo-structure and the arg are guaranteed to not 614 * be in use. 615 * 616 * Load acquire synchronizes with the release add in 617 * smp_rendezvous_action(), which ensures that our caller sees 618 * all memory actions done by the called functions on other 619 * CPUs. 620 */ 621 while (atomic_load_acq_int(&smp_rv_waiters[3]) < ncpus) 622 cpu_spinwait(); 623 624 mtx_unlock_spin(&smp_ipi_mtx); 625 } 626 627 void 628 smp_rendezvous(void (* setup_func)(void *), 629 void (* action_func)(void *), 630 void (* teardown_func)(void *), 631 void *arg) 632 { 633 smp_rendezvous_cpus(all_cpus, setup_func, action_func, teardown_func, arg); 634 } 635 636 static void 637 smp_topo_fill(struct cpu_group *cg) 638 { 639 int c; 640 641 for (c = 0; c < cg->cg_children; c++) 642 smp_topo_fill(&cg->cg_child[c]); 643 cg->cg_first = CPU_FFS(&cg->cg_mask) - 1; 644 cg->cg_last = CPU_FLS(&cg->cg_mask) - 1; 645 } 646 647 struct cpu_group * 648 smp_topo(void) 649 { 650 char cpusetbuf[CPUSETBUFSIZ], cpusetbuf2[CPUSETBUFSIZ]; 651 static struct cpu_group *top = NULL; 652 653 /* 654 * The first call to smp_topo() is guaranteed to occur 655 * during the kernel boot while we are still single-threaded. 656 */ 657 if (top != NULL) 658 return (top); 659 660 /* 661 * Check for a fake topology request for debugging purposes. 662 */ 663 switch (smp_topology) { 664 case 1: 665 /* Dual core with no sharing. */ 666 top = smp_topo_1level(CG_SHARE_NONE, 2, 0); 667 break; 668 case 2: 669 /* No topology, all cpus are equal. */ 670 top = smp_topo_none(); 671 break; 672 case 3: 673 /* Dual core with shared L2. */ 674 top = smp_topo_1level(CG_SHARE_L2, 2, 0); 675 break; 676 case 4: 677 /* quad core, shared l3 among each package, private l2. */ 678 top = smp_topo_1level(CG_SHARE_L3, 4, 0); 679 break; 680 case 5: 681 /* quad core, 2 dualcore parts on each package share l2. */ 682 top = smp_topo_2level(CG_SHARE_NONE, 2, CG_SHARE_L2, 2, 0); 683 break; 684 case 6: 685 /* Single-core 2xHTT */ 686 top = smp_topo_1level(CG_SHARE_L1, 2, CG_FLAG_HTT); 687 break; 688 case 7: 689 /* quad core with a shared l3, 8 threads sharing L2. */ 690 top = smp_topo_2level(CG_SHARE_L3, 4, CG_SHARE_L2, 8, 691 CG_FLAG_SMT); 692 break; 693 default: 694 /* Default, ask the system what it wants. */ 695 top = cpu_topo(); 696 break; 697 } 698 /* 699 * Verify the returned topology. 700 */ 701 if (top->cg_count != mp_ncpus) 702 panic("Built bad topology at %p. CPU count %d != %d", 703 top, top->cg_count, mp_ncpus); 704 if (CPU_CMP(&top->cg_mask, &all_cpus)) 705 panic("Built bad topology at %p. CPU mask (%s) != (%s)", 706 top, cpusetobj_strprint(cpusetbuf, &top->cg_mask), 707 cpusetobj_strprint(cpusetbuf2, &all_cpus)); 708 709 /* 710 * Collapse nonsense levels that may be created out of convenience by 711 * the MD layers. They cause extra work in the search functions. 712 */ 713 while (top->cg_children == 1) { 714 top = &top->cg_child[0]; 715 top->cg_parent = NULL; 716 } 717 smp_topo_fill(top); 718 return (top); 719 } 720 721 struct cpu_group * 722 smp_topo_alloc(u_int count) 723 { 724 static struct cpu_group *group = NULL; 725 static u_int index; 726 u_int curr; 727 728 if (group == NULL) { 729 group = mallocarray((mp_maxid + 1) * MAX_CACHE_LEVELS + 1, 730 sizeof(*group), M_DEVBUF, M_WAITOK | M_ZERO); 731 } 732 curr = index; 733 index += count; 734 return (&group[curr]); 735 } 736 737 struct cpu_group * 738 smp_topo_none(void) 739 { 740 struct cpu_group *top; 741 742 top = smp_topo_alloc(1); 743 top->cg_parent = NULL; 744 top->cg_child = NULL; 745 top->cg_mask = all_cpus; 746 top->cg_count = mp_ncpus; 747 top->cg_children = 0; 748 top->cg_level = CG_SHARE_NONE; 749 top->cg_flags = 0; 750 751 return (top); 752 } 753 754 static int 755 smp_topo_addleaf(struct cpu_group *parent, struct cpu_group *child, int share, 756 int count, int flags, int start) 757 { 758 char cpusetbuf[CPUSETBUFSIZ], cpusetbuf2[CPUSETBUFSIZ]; 759 cpuset_t mask; 760 int i; 761 762 CPU_ZERO(&mask); 763 for (i = 0; i < count; i++, start++) 764 CPU_SET(start, &mask); 765 child->cg_parent = parent; 766 child->cg_child = NULL; 767 child->cg_children = 0; 768 child->cg_level = share; 769 child->cg_count = count; 770 child->cg_flags = flags; 771 child->cg_mask = mask; 772 parent->cg_children++; 773 for (; parent != NULL; parent = parent->cg_parent) { 774 if (CPU_OVERLAP(&parent->cg_mask, &child->cg_mask)) 775 panic("Duplicate children in %p. mask (%s) child (%s)", 776 parent, 777 cpusetobj_strprint(cpusetbuf, &parent->cg_mask), 778 cpusetobj_strprint(cpusetbuf2, &child->cg_mask)); 779 CPU_OR(&parent->cg_mask, &parent->cg_mask, &child->cg_mask); 780 parent->cg_count += child->cg_count; 781 } 782 783 return (start); 784 } 785 786 struct cpu_group * 787 smp_topo_1level(int share, int count, int flags) 788 { 789 struct cpu_group *child; 790 struct cpu_group *top; 791 int packages; 792 int cpu; 793 int i; 794 795 cpu = 0; 796 packages = mp_ncpus / count; 797 top = smp_topo_alloc(1 + packages); 798 top->cg_child = child = top + 1; 799 top->cg_level = CG_SHARE_NONE; 800 for (i = 0; i < packages; i++, child++) 801 cpu = smp_topo_addleaf(top, child, share, count, flags, cpu); 802 return (top); 803 } 804 805 struct cpu_group * 806 smp_topo_2level(int l2share, int l2count, int l1share, int l1count, 807 int l1flags) 808 { 809 struct cpu_group *top; 810 struct cpu_group *l1g; 811 struct cpu_group *l2g; 812 int cpu; 813 int i; 814 int j; 815 816 cpu = 0; 817 top = smp_topo_alloc(1 + mp_ncpus / (l2count * l1count) + 818 mp_ncpus / l1count); 819 l2g = top + 1; 820 top->cg_child = l2g; 821 top->cg_level = CG_SHARE_NONE; 822 top->cg_children = mp_ncpus / (l2count * l1count); 823 l1g = l2g + top->cg_children; 824 for (i = 0; i < top->cg_children; i++, l2g++) { 825 l2g->cg_parent = top; 826 l2g->cg_child = l1g; 827 l2g->cg_level = l2share; 828 for (j = 0; j < l2count; j++, l1g++) 829 cpu = smp_topo_addleaf(l2g, l1g, l1share, l1count, 830 l1flags, cpu); 831 } 832 return (top); 833 } 834 835 struct cpu_group * 836 smp_topo_find(struct cpu_group *top, int cpu) 837 { 838 struct cpu_group *cg; 839 cpuset_t mask; 840 int children; 841 int i; 842 843 CPU_SETOF(cpu, &mask); 844 cg = top; 845 for (;;) { 846 if (!CPU_OVERLAP(&cg->cg_mask, &mask)) 847 return (NULL); 848 if (cg->cg_children == 0) 849 return (cg); 850 children = cg->cg_children; 851 for (i = 0, cg = cg->cg_child; i < children; cg++, i++) 852 if (CPU_OVERLAP(&cg->cg_mask, &mask)) 853 break; 854 } 855 return (NULL); 856 } 857 #else /* !SMP */ 858 859 void 860 smp_rendezvous_cpus(cpuset_t map, 861 void (*setup_func)(void *), 862 void (*action_func)(void *), 863 void (*teardown_func)(void *), 864 void *arg) 865 { 866 /* 867 * In the !SMP case we just need to ensure the same initial conditions 868 * as the SMP case. 869 */ 870 spinlock_enter(); 871 if (setup_func != NULL) 872 setup_func(arg); 873 if (action_func != NULL) 874 action_func(arg); 875 if (teardown_func != NULL) 876 teardown_func(arg); 877 spinlock_exit(); 878 } 879 880 void 881 smp_rendezvous(void (*setup_func)(void *), 882 void (*action_func)(void *), 883 void (*teardown_func)(void *), 884 void *arg) 885 { 886 887 smp_rendezvous_cpus(all_cpus, setup_func, action_func, teardown_func, 888 arg); 889 } 890 891 /* 892 * Provide dummy SMP support for UP kernels. Modules that need to use SMP 893 * APIs will still work using this dummy support. 894 */ 895 static void 896 mp_setvariables_for_up(void *dummy) 897 { 898 mp_ncpus = 1; 899 mp_ncores = 1; 900 mp_maxid = PCPU_GET(cpuid); 901 CPU_SETOF(mp_maxid, &all_cpus); 902 KASSERT(PCPU_GET(cpuid) == 0, ("UP must have a CPU ID of zero")); 903 } 904 SYSINIT(cpu_mp_setvariables, SI_SUB_TUNABLES, SI_ORDER_FIRST, 905 mp_setvariables_for_up, NULL); 906 #endif /* SMP */ 907 908 void 909 smp_no_rendezvous_barrier(void *dummy) 910 { 911 #ifdef SMP 912 KASSERT((!smp_started),("smp_no_rendezvous called and smp is started")); 913 #endif 914 } 915 916 void 917 smp_rendezvous_cpus_retry(cpuset_t map, 918 void (* setup_func)(void *), 919 void (* action_func)(void *), 920 void (* teardown_func)(void *), 921 void (* wait_func)(void *, int), 922 struct smp_rendezvous_cpus_retry_arg *arg) 923 { 924 int cpu; 925 926 CPU_COPY(&map, &arg->cpus); 927 928 /* 929 * Only one CPU to execute on. 930 */ 931 if (!smp_started) { 932 spinlock_enter(); 933 if (setup_func != NULL) 934 setup_func(arg); 935 if (action_func != NULL) 936 action_func(arg); 937 if (teardown_func != NULL) 938 teardown_func(arg); 939 spinlock_exit(); 940 return; 941 } 942 943 /* 944 * Execute an action on all specified CPUs while retrying until they 945 * all acknowledge completion. 946 */ 947 for (;;) { 948 smp_rendezvous_cpus( 949 arg->cpus, 950 setup_func, 951 action_func, 952 teardown_func, 953 arg); 954 955 if (CPU_EMPTY(&arg->cpus)) 956 break; 957 958 CPU_FOREACH(cpu) { 959 if (!CPU_ISSET(cpu, &arg->cpus)) 960 continue; 961 wait_func(arg, cpu); 962 } 963 } 964 } 965 966 void 967 smp_rendezvous_cpus_done(struct smp_rendezvous_cpus_retry_arg *arg) 968 { 969 970 CPU_CLR_ATOMIC(curcpu, &arg->cpus); 971 } 972 973 /* 974 * If (prio & PDROP) == 0: 975 * Wait for specified idle threads to switch once. This ensures that even 976 * preempted threads have cycled through the switch function once, 977 * exiting their codepaths. This allows us to change global pointers 978 * with no other synchronization. 979 * If (prio & PDROP) != 0: 980 * Force the specified CPUs to switch context at least once. 981 */ 982 int 983 quiesce_cpus(cpuset_t map, const char *wmesg, int prio) 984 { 985 struct pcpu *pcpu; 986 u_int *gen; 987 int error; 988 int cpu; 989 990 error = 0; 991 if ((prio & PDROP) == 0) { 992 gen = mallocarray(sizeof(u_int), mp_maxid + 1, M_TEMP, 993 M_WAITOK); 994 for (cpu = 0; cpu <= mp_maxid; cpu++) { 995 if (!CPU_ISSET(cpu, &map) || CPU_ABSENT(cpu)) 996 continue; 997 pcpu = pcpu_find(cpu); 998 gen[cpu] = pcpu->pc_idlethread->td_generation; 999 } 1000 } 1001 for (cpu = 0; cpu <= mp_maxid; cpu++) { 1002 if (!CPU_ISSET(cpu, &map) || CPU_ABSENT(cpu)) 1003 continue; 1004 pcpu = pcpu_find(cpu); 1005 thread_lock(curthread); 1006 sched_bind(curthread, cpu); 1007 thread_unlock(curthread); 1008 if ((prio & PDROP) != 0) 1009 continue; 1010 while (gen[cpu] == pcpu->pc_idlethread->td_generation) { 1011 error = tsleep(quiesce_cpus, prio & ~PDROP, wmesg, 1); 1012 if (error != EWOULDBLOCK) 1013 goto out; 1014 error = 0; 1015 } 1016 } 1017 out: 1018 thread_lock(curthread); 1019 sched_unbind(curthread); 1020 thread_unlock(curthread); 1021 if ((prio & PDROP) == 0) 1022 free(gen, M_TEMP); 1023 1024 return (error); 1025 } 1026 1027 int 1028 quiesce_all_cpus(const char *wmesg, int prio) 1029 { 1030 1031 return quiesce_cpus(all_cpus, wmesg, prio); 1032 } 1033 1034 /* 1035 * Observe all CPUs not executing in critical section. 1036 * We are not in one so the check for us is safe. If the found 1037 * thread changes to something else we know the section was 1038 * exited as well. 1039 */ 1040 void 1041 quiesce_all_critical(void) 1042 { 1043 struct thread *td, *newtd; 1044 struct pcpu *pcpu; 1045 int cpu; 1046 1047 MPASS(curthread->td_critnest == 0); 1048 1049 CPU_FOREACH(cpu) { 1050 pcpu = cpuid_to_pcpu[cpu]; 1051 td = pcpu->pc_curthread; 1052 for (;;) { 1053 if (td->td_critnest == 0) 1054 break; 1055 cpu_spinwait(); 1056 newtd = (struct thread *) 1057 atomic_load_acq_ptr((void *)pcpu->pc_curthread); 1058 if (td != newtd) 1059 break; 1060 } 1061 } 1062 } 1063 1064 static void 1065 cpus_fence_seq_cst_issue(void *arg __unused) 1066 { 1067 1068 atomic_thread_fence_seq_cst(); 1069 } 1070 1071 /* 1072 * Send an IPI forcing a sequentially consistent fence. 1073 * 1074 * Allows replacement of an explicitly fence with a compiler barrier. 1075 * Trades speed up during normal execution for a significant slowdown when 1076 * the barrier is needed. 1077 */ 1078 void 1079 cpus_fence_seq_cst(void) 1080 { 1081 1082 #ifdef SMP 1083 smp_rendezvous( 1084 smp_no_rendezvous_barrier, 1085 cpus_fence_seq_cst_issue, 1086 smp_no_rendezvous_barrier, 1087 NULL 1088 ); 1089 #else 1090 cpus_fence_seq_cst_issue(NULL); 1091 #endif 1092 } 1093 1094 /* Extra care is taken with this sysctl because the data type is volatile */ 1095 static int 1096 sysctl_kern_smp_active(SYSCTL_HANDLER_ARGS) 1097 { 1098 int error, active; 1099 1100 active = smp_started; 1101 error = SYSCTL_OUT(req, &active, sizeof(active)); 1102 return (error); 1103 } 1104 1105 #ifdef SMP 1106 void 1107 topo_init_node(struct topo_node *node) 1108 { 1109 1110 bzero(node, sizeof(*node)); 1111 TAILQ_INIT(&node->children); 1112 } 1113 1114 void 1115 topo_init_root(struct topo_node *root) 1116 { 1117 1118 topo_init_node(root); 1119 root->type = TOPO_TYPE_SYSTEM; 1120 } 1121 1122 /* 1123 * Add a child node with the given ID under the given parent. 1124 * Do nothing if there is already a child with that ID. 1125 */ 1126 struct topo_node * 1127 topo_add_node_by_hwid(struct topo_node *parent, int hwid, 1128 topo_node_type type, uintptr_t subtype) 1129 { 1130 struct topo_node *node; 1131 1132 TAILQ_FOREACH_REVERSE(node, &parent->children, 1133 topo_children, siblings) { 1134 if (node->hwid == hwid 1135 && node->type == type && node->subtype == subtype) { 1136 return (node); 1137 } 1138 } 1139 1140 node = malloc(sizeof(*node), M_TOPO, M_WAITOK); 1141 topo_init_node(node); 1142 node->parent = parent; 1143 node->hwid = hwid; 1144 node->type = type; 1145 node->subtype = subtype; 1146 TAILQ_INSERT_TAIL(&parent->children, node, siblings); 1147 parent->nchildren++; 1148 1149 return (node); 1150 } 1151 1152 /* 1153 * Find a child node with the given ID under the given parent. 1154 */ 1155 struct topo_node * 1156 topo_find_node_by_hwid(struct topo_node *parent, int hwid, 1157 topo_node_type type, uintptr_t subtype) 1158 { 1159 1160 struct topo_node *node; 1161 1162 TAILQ_FOREACH(node, &parent->children, siblings) { 1163 if (node->hwid == hwid 1164 && node->type == type && node->subtype == subtype) { 1165 return (node); 1166 } 1167 } 1168 1169 return (NULL); 1170 } 1171 1172 /* 1173 * Given a node change the order of its parent's child nodes such 1174 * that the node becomes the firt child while preserving the cyclic 1175 * order of the children. In other words, the given node is promoted 1176 * by rotation. 1177 */ 1178 void 1179 topo_promote_child(struct topo_node *child) 1180 { 1181 struct topo_node *next; 1182 struct topo_node *node; 1183 struct topo_node *parent; 1184 1185 parent = child->parent; 1186 next = TAILQ_NEXT(child, siblings); 1187 TAILQ_REMOVE(&parent->children, child, siblings); 1188 TAILQ_INSERT_HEAD(&parent->children, child, siblings); 1189 1190 while (next != NULL) { 1191 node = next; 1192 next = TAILQ_NEXT(node, siblings); 1193 TAILQ_REMOVE(&parent->children, node, siblings); 1194 TAILQ_INSERT_AFTER(&parent->children, child, node, siblings); 1195 child = node; 1196 } 1197 } 1198 1199 /* 1200 * Iterate to the next node in the depth-first search (traversal) of 1201 * the topology tree. 1202 */ 1203 struct topo_node * 1204 topo_next_node(struct topo_node *top, struct topo_node *node) 1205 { 1206 struct topo_node *next; 1207 1208 if ((next = TAILQ_FIRST(&node->children)) != NULL) 1209 return (next); 1210 1211 if ((next = TAILQ_NEXT(node, siblings)) != NULL) 1212 return (next); 1213 1214 while (node != top && (node = node->parent) != top) 1215 if ((next = TAILQ_NEXT(node, siblings)) != NULL) 1216 return (next); 1217 1218 return (NULL); 1219 } 1220 1221 /* 1222 * Iterate to the next node in the depth-first search of the topology tree, 1223 * but without descending below the current node. 1224 */ 1225 struct topo_node * 1226 topo_next_nonchild_node(struct topo_node *top, struct topo_node *node) 1227 { 1228 struct topo_node *next; 1229 1230 if ((next = TAILQ_NEXT(node, siblings)) != NULL) 1231 return (next); 1232 1233 while (node != top && (node = node->parent) != top) 1234 if ((next = TAILQ_NEXT(node, siblings)) != NULL) 1235 return (next); 1236 1237 return (NULL); 1238 } 1239 1240 /* 1241 * Assign the given ID to the given topology node that represents a logical 1242 * processor. 1243 */ 1244 void 1245 topo_set_pu_id(struct topo_node *node, cpuid_t id) 1246 { 1247 1248 KASSERT(node->type == TOPO_TYPE_PU, 1249 ("topo_set_pu_id: wrong node type: %u", node->type)); 1250 KASSERT(CPU_EMPTY(&node->cpuset) && node->cpu_count == 0, 1251 ("topo_set_pu_id: cpuset already not empty")); 1252 node->id = id; 1253 CPU_SET(id, &node->cpuset); 1254 node->cpu_count = 1; 1255 node->subtype = 1; 1256 1257 while ((node = node->parent) != NULL) { 1258 KASSERT(!CPU_ISSET(id, &node->cpuset), 1259 ("logical ID %u is already set in node %p", id, node)); 1260 CPU_SET(id, &node->cpuset); 1261 node->cpu_count++; 1262 } 1263 } 1264 1265 static struct topology_spec { 1266 topo_node_type type; 1267 bool match_subtype; 1268 uintptr_t subtype; 1269 } topology_level_table[TOPO_LEVEL_COUNT] = { 1270 [TOPO_LEVEL_PKG] = { .type = TOPO_TYPE_PKG, }, 1271 [TOPO_LEVEL_GROUP] = { .type = TOPO_TYPE_GROUP, }, 1272 [TOPO_LEVEL_CACHEGROUP] = { 1273 .type = TOPO_TYPE_CACHE, 1274 .match_subtype = true, 1275 .subtype = CG_SHARE_L3, 1276 }, 1277 [TOPO_LEVEL_CORE] = { .type = TOPO_TYPE_CORE, }, 1278 [TOPO_LEVEL_THREAD] = { .type = TOPO_TYPE_PU, }, 1279 }; 1280 1281 static bool 1282 topo_analyze_table(struct topo_node *root, int all, enum topo_level level, 1283 struct topo_analysis *results) 1284 { 1285 struct topology_spec *spec; 1286 struct topo_node *node; 1287 int count; 1288 1289 if (level >= TOPO_LEVEL_COUNT) 1290 return (true); 1291 1292 spec = &topology_level_table[level]; 1293 count = 0; 1294 node = topo_next_node(root, root); 1295 1296 while (node != NULL) { 1297 if (node->type != spec->type || 1298 (spec->match_subtype && node->subtype != spec->subtype)) { 1299 node = topo_next_node(root, node); 1300 continue; 1301 } 1302 if (!all && CPU_EMPTY(&node->cpuset)) { 1303 node = topo_next_nonchild_node(root, node); 1304 continue; 1305 } 1306 1307 count++; 1308 1309 if (!topo_analyze_table(node, all, level + 1, results)) 1310 return (false); 1311 1312 node = topo_next_nonchild_node(root, node); 1313 } 1314 1315 /* No explicit subgroups is essentially one subgroup. */ 1316 if (count == 0) { 1317 count = 1; 1318 1319 if (!topo_analyze_table(root, all, level + 1, results)) 1320 return (false); 1321 } 1322 1323 if (results->entities[level] == -1) 1324 results->entities[level] = count; 1325 else if (results->entities[level] != count) 1326 return (false); 1327 1328 return (true); 1329 } 1330 1331 /* 1332 * Check if the topology is uniform, that is, each package has the same number 1333 * of cores in it and each core has the same number of threads (logical 1334 * processors) in it. If so, calculate the number of packages, the number of 1335 * groups per package, the number of cachegroups per group, and the number of 1336 * logical processors per cachegroup. 'all' parameter tells whether to include 1337 * administratively disabled logical processors into the analysis. 1338 */ 1339 int 1340 topo_analyze(struct topo_node *topo_root, int all, 1341 struct topo_analysis *results) 1342 { 1343 1344 results->entities[TOPO_LEVEL_PKG] = -1; 1345 results->entities[TOPO_LEVEL_CORE] = -1; 1346 results->entities[TOPO_LEVEL_THREAD] = -1; 1347 results->entities[TOPO_LEVEL_GROUP] = -1; 1348 results->entities[TOPO_LEVEL_CACHEGROUP] = -1; 1349 1350 if (!topo_analyze_table(topo_root, all, TOPO_LEVEL_PKG, results)) 1351 return (0); 1352 1353 KASSERT(results->entities[TOPO_LEVEL_PKG] > 0, 1354 ("bug in topology or analysis")); 1355 1356 return (1); 1357 } 1358 1359 #endif /* SMP */ 1360