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