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