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