1 /* 2 * Copyright (c) 2003-2011 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@backplane.com> 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * 3. Neither the name of The DragonFly Project nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific, prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 */ 34 35 /* 36 * Each cpu in a system has its own self-contained light weight kernel 37 * thread scheduler, which means that generally speaking we only need 38 * to use a critical section to avoid problems. Foreign thread 39 * scheduling is queued via (async) IPIs. 40 */ 41 42 #include <sys/param.h> 43 #include <sys/systm.h> 44 #include <sys/kernel.h> 45 #include <sys/proc.h> 46 #include <sys/rtprio.h> 47 #include <sys/kinfo.h> 48 #include <sys/malloc.h> 49 #include <sys/queue.h> 50 #include <sys/sysctl.h> 51 #include <sys/kthread.h> 52 #include <machine/cpu.h> 53 #include <sys/lock.h> 54 #include <sys/spinlock.h> 55 #include <sys/ktr.h> 56 #include <sys/indefinite.h> 57 58 #include <sys/thread2.h> 59 #include <sys/spinlock2.h> 60 #include <sys/indefinite2.h> 61 62 #include <sys/dsched.h> 63 64 #include <vm/vm.h> 65 #include <vm/vm_param.h> 66 #include <vm/vm_kern.h> 67 #include <vm/vm_object.h> 68 #include <vm/vm_page.h> 69 #include <vm/vm_map.h> 70 #include <vm/vm_pager.h> 71 #include <vm/vm_extern.h> 72 73 #include <machine/stdarg.h> 74 #include <machine/smp.h> 75 #include <machine/clock.h> 76 77 #define LOOPMASK 78 79 #if !defined(KTR_CTXSW) 80 #define KTR_CTXSW KTR_ALL 81 #endif 82 KTR_INFO_MASTER(ctxsw); 83 KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p", int cpu, struct thread *td); 84 KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p", int cpu, struct thread *td); 85 KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s", struct thread *td, char *comm); 86 KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", struct thread *td); 87 88 static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads"); 89 90 #ifdef INVARIANTS 91 static int panic_on_cscount = 0; 92 #endif 93 #ifdef DEBUG_LWKT_THREAD 94 static int64_t switch_count = 0; 95 static int64_t preempt_hit = 0; 96 static int64_t preempt_miss = 0; 97 static int64_t preempt_weird = 0; 98 #endif 99 static int lwkt_use_spin_port; 100 __read_mostly static struct objcache *thread_cache; 101 int cpu_mwait_spin = 0; 102 103 static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame); 104 static void lwkt_setcpu_remote(void *arg); 105 106 /* 107 * We can make all thread ports use the spin backend instead of the thread 108 * backend. This should only be set to debug the spin backend. 109 */ 110 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port); 111 112 #ifdef INVARIANTS 113 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, 114 "Panic if attempting to switch lwkt's while mastering cpusync"); 115 #endif 116 #ifdef DEBUG_LWKT_THREAD 117 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, 118 "Number of switched threads"); 119 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, 120 "Successful preemption events"); 121 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, 122 "Failed preemption events"); 123 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, 124 "Number of preempted threads."); 125 #endif 126 extern int lwkt_sched_debug; 127 int lwkt_sched_debug = 0; 128 SYSCTL_INT(_lwkt, OID_AUTO, sched_debug, CTLFLAG_RW, 129 &lwkt_sched_debug, 0, "Scheduler debug"); 130 __read_mostly static u_int lwkt_spin_loops = 10; 131 SYSCTL_UINT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW, 132 &lwkt_spin_loops, 0, "Scheduler spin loops until sorted decon"); 133 __read_mostly static int preempt_enable = 1; 134 SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW, 135 &preempt_enable, 0, "Enable preemption"); 136 static int lwkt_cache_threads = 0; 137 SYSCTL_INT(_lwkt, OID_AUTO, cache_threads, CTLFLAG_RD, 138 &lwkt_cache_threads, 0, "thread+kstack cache"); 139 140 /* 141 * These helper procedures handle the runq, they can only be called from 142 * within a critical section. 143 * 144 * WARNING! Prior to SMP being brought up it is possible to enqueue and 145 * dequeue threads belonging to other cpus, so be sure to use td->td_gd 146 * instead of 'mycpu' when referencing the globaldata structure. Once 147 * SMP live enqueuing and dequeueing only occurs on the current cpu. 148 */ 149 static __inline 150 void 151 _lwkt_dequeue(thread_t td) 152 { 153 if (td->td_flags & TDF_RUNQ) { 154 struct globaldata *gd = td->td_gd; 155 156 td->td_flags &= ~TDF_RUNQ; 157 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq); 158 --gd->gd_tdrunqcount; 159 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL) 160 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING); 161 } 162 } 163 164 /* 165 * Priority enqueue. 166 * 167 * There are a limited number of lwkt threads runnable since user 168 * processes only schedule one at a time per cpu. However, there can 169 * be many user processes in kernel mode exiting from a tsleep() which 170 * become runnable. 171 * 172 * We scan the queue in both directions to help deal with degenerate 173 * situations when hundreds or thousands (or more) threads are runnable. 174 * 175 * NOTE: lwkt_schedulerclock() will force a round-robin based on td_pri and 176 * will ignore user priority. This is to ensure that user threads in 177 * kernel mode get cpu at some point regardless of what the user 178 * scheduler thinks. 179 */ 180 static __inline 181 void 182 _lwkt_enqueue(thread_t td) 183 { 184 thread_t xtd; /* forward scan */ 185 thread_t rtd; /* reverse scan */ 186 187 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) { 188 struct globaldata *gd = td->td_gd; 189 190 td->td_flags |= TDF_RUNQ; 191 xtd = TAILQ_FIRST(&gd->gd_tdrunq); 192 if (xtd == NULL) { 193 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 194 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING); 195 } else { 196 /* 197 * NOTE: td_upri - higher numbers more desireable, same sense 198 * as td_pri (typically reversed from lwp_upri). 199 * 200 * In the equal priority case we want the best selection 201 * at the beginning so the less desireable selections know 202 * that they have to setrunqueue/go-to-another-cpu, even 203 * though it means switching back to the 'best' selection. 204 * This also avoids degenerate situations when many threads 205 * are runnable or waking up at the same time. 206 * 207 * If upri matches exactly place at end/round-robin. 208 */ 209 rtd = TAILQ_LAST(&gd->gd_tdrunq, lwkt_queue); 210 211 while (xtd && 212 (xtd->td_pri > td->td_pri || 213 (xtd->td_pri == td->td_pri && 214 xtd->td_upri >= td->td_upri))) { 215 xtd = TAILQ_NEXT(xtd, td_threadq); 216 217 /* 218 * Doing a reverse scan at the same time is an optimization 219 * for the insert-closer-to-tail case that avoids having to 220 * scan the entire list. This situation can occur when 221 * thousands of threads are woken up at the same time. 222 */ 223 if (rtd->td_pri > td->td_pri || 224 (rtd->td_pri == td->td_pri && 225 rtd->td_upri >= td->td_upri)) { 226 TAILQ_INSERT_AFTER(&gd->gd_tdrunq, rtd, td, td_threadq); 227 goto skip; 228 } 229 rtd = TAILQ_PREV(rtd, lwkt_queue, td_threadq); 230 } 231 if (xtd) 232 TAILQ_INSERT_BEFORE(xtd, td, td_threadq); 233 else 234 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 235 } 236 skip: 237 ++gd->gd_tdrunqcount; 238 239 /* 240 * Request a LWKT reschedule if we are now at the head of the queue. 241 */ 242 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) 243 need_lwkt_resched(); 244 } 245 } 246 247 static boolean_t 248 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags) 249 { 250 struct thread *td = (struct thread *)obj; 251 252 td->td_kstack = NULL; 253 td->td_kstack_size = 0; 254 td->td_flags = TDF_ALLOCATED_THREAD; 255 td->td_mpflags = 0; 256 return (1); 257 } 258 259 static void 260 _lwkt_thread_dtor(void *obj, void *privdata) 261 { 262 struct thread *td = (struct thread *)obj; 263 264 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD, 265 ("_lwkt_thread_dtor: not allocated from objcache")); 266 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack && 267 td->td_kstack_size > 0, 268 ("_lwkt_thread_dtor: corrupted stack")); 269 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size); 270 td->td_kstack = NULL; 271 td->td_flags = 0; 272 } 273 274 /* 275 * Initialize the lwkt s/system. 276 * 277 * Nominally cache up to 32 thread + kstack structures. Cache more on 278 * systems with a lot of cpu cores. 279 */ 280 static void 281 lwkt_init(void) 282 { 283 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads); 284 if (lwkt_cache_threads == 0) { 285 lwkt_cache_threads = ncpus * 4; 286 if (lwkt_cache_threads < 32) 287 lwkt_cache_threads = 32; 288 } 289 thread_cache = objcache_create_mbacked( 290 M_THREAD, sizeof(struct thread), 291 0, lwkt_cache_threads, 292 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL); 293 } 294 SYSINIT(lwkt_init, SI_BOOT2_LWKT_INIT, SI_ORDER_FIRST, lwkt_init, NULL); 295 296 /* 297 * Schedule a thread to run. As the current thread we can always safely 298 * schedule ourselves, and a shortcut procedure is provided for that 299 * function. 300 * 301 * (non-blocking, self contained on a per cpu basis) 302 */ 303 void 304 lwkt_schedule_self(thread_t td) 305 { 306 KKASSERT((td->td_flags & TDF_MIGRATING) == 0); 307 crit_enter_quick(td); 308 KASSERT(td != &td->td_gd->gd_idlethread, 309 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!")); 310 KKASSERT(td->td_lwp == NULL || 311 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 312 _lwkt_enqueue(td); 313 crit_exit_quick(td); 314 } 315 316 /* 317 * Deschedule a thread. 318 * 319 * (non-blocking, self contained on a per cpu basis) 320 */ 321 void 322 lwkt_deschedule_self(thread_t td) 323 { 324 crit_enter_quick(td); 325 _lwkt_dequeue(td); 326 crit_exit_quick(td); 327 } 328 329 /* 330 * LWKTs operate on a per-cpu basis 331 * 332 * WARNING! Called from early boot, 'mycpu' may not work yet. 333 */ 334 void 335 lwkt_gdinit(struct globaldata *gd) 336 { 337 TAILQ_INIT(&gd->gd_tdrunq); 338 TAILQ_INIT(&gd->gd_tdallq); 339 lockinit(&gd->gd_sysctllock, "sysctl", 0, LK_CANRECURSE); 340 } 341 342 /* 343 * Create a new thread. The thread must be associated with a process context 344 * or LWKT start address before it can be scheduled. If the target cpu is 345 * -1 the thread will be created on the current cpu. 346 * 347 * If you intend to create a thread without a process context this function 348 * does everything except load the startup and switcher function. 349 */ 350 thread_t 351 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags) 352 { 353 static int cpu_rotator; 354 globaldata_t gd = mycpu; 355 void *stack; 356 357 /* 358 * If static thread storage is not supplied allocate a thread. Reuse 359 * a cached free thread if possible. gd_freetd is used to keep an exiting 360 * thread intact through the exit. 361 */ 362 if (td == NULL) { 363 crit_enter_gd(gd); 364 if ((td = gd->gd_freetd) != NULL) { 365 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK| 366 TDF_RUNQ)) == 0); 367 gd->gd_freetd = NULL; 368 } else { 369 td = objcache_get(thread_cache, M_WAITOK); 370 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK| 371 TDF_RUNQ)) == 0); 372 } 373 crit_exit_gd(gd); 374 KASSERT((td->td_flags & 375 (TDF_ALLOCATED_THREAD|TDF_RUNNING|TDF_PREEMPT_LOCK)) == 376 TDF_ALLOCATED_THREAD, 377 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags)); 378 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK); 379 } 380 381 /* 382 * Try to reuse cached stack. 383 */ 384 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) { 385 if (flags & TDF_ALLOCATED_STACK) { 386 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size); 387 stack = NULL; 388 } 389 } 390 if (stack == NULL) { 391 if (cpu < 0) 392 stack = (void *)kmem_alloc_stack(&kernel_map, stksize, 0); 393 else 394 stack = (void *)kmem_alloc_stack(&kernel_map, stksize, 395 KM_CPU(cpu)); 396 flags |= TDF_ALLOCATED_STACK; 397 } 398 if (cpu < 0) { 399 cpu = ++cpu_rotator; 400 cpu_ccfence(); 401 cpu = (uint32_t)cpu % (uint32_t)ncpus; 402 } 403 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu)); 404 return(td); 405 } 406 407 /* 408 * Initialize a preexisting thread structure. This function is used by 409 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread. 410 * 411 * All threads start out in a critical section at a priority of 412 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as 413 * appropriate. This function may send an IPI message when the 414 * requested cpu is not the current cpu and consequently gd_tdallq may 415 * not be initialized synchronously from the point of view of the originating 416 * cpu. 417 * 418 * NOTE! we have to be careful in regards to creating threads for other cpus 419 * if SMP has not yet been activated. 420 */ 421 static void 422 lwkt_init_thread_remote(void *arg) 423 { 424 thread_t td = arg; 425 426 /* 427 * Protected by critical section held by IPI dispatch 428 */ 429 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq); 430 } 431 432 /* 433 * lwkt core thread structural initialization. 434 * 435 * NOTE: All threads are initialized as mpsafe threads. 436 */ 437 void 438 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags, 439 struct globaldata *gd) 440 { 441 globaldata_t mygd = mycpu; 442 443 bzero(td, sizeof(struct thread)); 444 td->td_kstack = stack; 445 td->td_kstack_size = stksize; 446 td->td_flags = flags; 447 td->td_mpflags = 0; 448 td->td_type = TD_TYPE_GENERIC; 449 td->td_gd = gd; 450 td->td_pri = TDPRI_KERN_DAEMON; 451 td->td_critcount = 1; 452 td->td_toks_have = NULL; 453 td->td_toks_stop = &td->td_toks_base; 454 if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT)) { 455 lwkt_initport_spin(&td->td_msgport, td, 456 (flags & TDF_FIXEDCPU) ? TRUE : FALSE); 457 } else { 458 lwkt_initport_thread(&td->td_msgport, td); 459 } 460 pmap_init_thread(td); 461 /* 462 * Normally initializing a thread for a remote cpu requires sending an 463 * IPI. However, the idlethread is setup before the other cpus are 464 * activated so we have to treat it as a special case. XXX manipulation 465 * of gd_tdallq requires the BGL. 466 */ 467 if (gd == mygd || td == &gd->gd_idlethread) { 468 crit_enter_gd(mygd); 469 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); 470 crit_exit_gd(mygd); 471 } else { 472 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td); 473 } 474 dsched_enter_thread(td); 475 } 476 477 void 478 lwkt_set_comm(thread_t td, const char *ctl, ...) 479 { 480 __va_list va; 481 482 __va_start(va, ctl); 483 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va); 484 __va_end(va); 485 KTR_LOG(ctxsw_newtd, td, td->td_comm); 486 } 487 488 /* 489 * Prevent the thread from getting destroyed. Note that unlike PHOLD/PRELE 490 * this does not prevent the thread from migrating to another cpu so the 491 * gd_tdallq state is not protected by this. 492 */ 493 void 494 lwkt_hold(thread_t td) 495 { 496 atomic_add_int(&td->td_refs, 1); 497 } 498 499 void 500 lwkt_rele(thread_t td) 501 { 502 KKASSERT(td->td_refs > 0); 503 atomic_add_int(&td->td_refs, -1); 504 } 505 506 void 507 lwkt_free_thread(thread_t td) 508 { 509 KKASSERT(td->td_refs == 0); 510 KKASSERT((td->td_flags & (TDF_RUNNING | TDF_PREEMPT_LOCK | 511 TDF_RUNQ | TDF_TSLEEPQ)) == 0); 512 if (td->td_flags & TDF_ALLOCATED_THREAD) { 513 objcache_put(thread_cache, td); 514 } else if (td->td_flags & TDF_ALLOCATED_STACK) { 515 /* client-allocated struct with internally allocated stack */ 516 KASSERT(td->td_kstack && td->td_kstack_size > 0, 517 ("lwkt_free_thread: corrupted stack")); 518 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size); 519 td->td_kstack = NULL; 520 td->td_kstack_size = 0; 521 } 522 523 KTR_LOG(ctxsw_deadtd, td); 524 } 525 526 527 /* 528 * Switch to the next runnable lwkt. If no LWKTs are runnable then 529 * switch to the idlethread. Switching must occur within a critical 530 * section to avoid races with the scheduling queue. 531 * 532 * We always have full control over our cpu's run queue. Other cpus 533 * that wish to manipulate our queue must use the cpu_*msg() calls to 534 * talk to our cpu, so a critical section is all that is needed and 535 * the result is very, very fast thread switching. 536 * 537 * The LWKT scheduler uses a fixed priority model and round-robins at 538 * each priority level. User process scheduling is a totally 539 * different beast and LWKT priorities should not be confused with 540 * user process priorities. 541 * 542 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch() 543 * is not called by the current thread in the preemption case, only when 544 * the preempting thread blocks (in order to return to the original thread). 545 * 546 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread 547 * migration and tsleep deschedule the current lwkt thread and call 548 * lwkt_switch(). In particular, the target cpu of the migration fully 549 * expects the thread to become non-runnable and can deadlock against 550 * cpusync operations if we run any IPIs prior to switching the thread out. 551 * 552 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF 553 * THE CURRENT THREAD HAS BEEN DESCHEDULED! 554 */ 555 void 556 lwkt_switch(void) 557 { 558 globaldata_t gd = mycpu; 559 thread_t td = gd->gd_curthread; 560 thread_t ntd; 561 thread_t xtd; 562 int upri; 563 #ifdef LOOPMASK 564 uint64_t tsc_base = rdtsc(); 565 #endif 566 567 KKASSERT(gd->gd_processing_ipiq == 0); 568 KKASSERT(td->td_flags & TDF_RUNNING); 569 570 /* 571 * Switching from within a 'fast' (non thread switched) interrupt or IPI 572 * is illegal. However, we may have to do it anyway if we hit a fatal 573 * kernel trap or we have paniced. 574 * 575 * If this case occurs save and restore the interrupt nesting level. 576 */ 577 if (gd->gd_intr_nesting_level) { 578 int savegdnest; 579 int savegdtrap; 580 581 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) { 582 panic("lwkt_switch: Attempt to switch from a " 583 "fast interrupt, ipi, or hard code section, " 584 "td %p\n", 585 td); 586 } else { 587 savegdnest = gd->gd_intr_nesting_level; 588 savegdtrap = gd->gd_trap_nesting_level; 589 gd->gd_intr_nesting_level = 0; 590 gd->gd_trap_nesting_level = 0; 591 if ((td->td_flags & TDF_PANICWARN) == 0) { 592 td->td_flags |= TDF_PANICWARN; 593 kprintf("Warning: thread switch from interrupt, IPI, " 594 "or hard code section.\n" 595 "thread %p (%s)\n", td, td->td_comm); 596 print_backtrace(-1); 597 } 598 lwkt_switch(); 599 gd->gd_intr_nesting_level = savegdnest; 600 gd->gd_trap_nesting_level = savegdtrap; 601 return; 602 } 603 } 604 605 /* 606 * Release our current user process designation if we are blocking 607 * or if a user reschedule was requested. 608 * 609 * NOTE: This function is NOT called if we are switching into or 610 * returning from a preemption. 611 * 612 * NOTE: Releasing our current user process designation may cause 613 * it to be assigned to another thread, which in turn will 614 * cause us to block in the usched acquire code when we attempt 615 * to return to userland. 616 * 617 * NOTE: On SMP systems this can be very nasty when heavy token 618 * contention is present so we want to be careful not to 619 * release the designation gratuitously. 620 */ 621 if (td->td_release && 622 (user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) { 623 td->td_release(td); 624 } 625 626 /* 627 * Release all tokens. Once we do this we must remain in the critical 628 * section and cannot run IPIs or other interrupts until we switch away 629 * because they may implode if they try to get a token using our thread 630 * context. 631 */ 632 crit_enter_gd(gd); 633 if (TD_TOKS_HELD(td)) 634 lwkt_relalltokens(td); 635 636 /* 637 * We had better not be holding any spin locks, but don't get into an 638 * endless panic loop. 639 */ 640 KASSERT(gd->gd_spinlocks == 0 || panicstr != NULL, 641 ("lwkt_switch: still holding %d exclusive spinlocks!", 642 gd->gd_spinlocks)); 643 644 #ifdef INVARIANTS 645 if (td->td_cscount) { 646 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n", 647 td); 648 if (panic_on_cscount) 649 panic("switching while mastering cpusync"); 650 } 651 #endif 652 653 /* 654 * If we had preempted another thread on this cpu, resume the preempted 655 * thread. This occurs transparently, whether the preempted thread 656 * was scheduled or not (it may have been preempted after descheduling 657 * itself). 658 * 659 * We have to setup the MP lock for the original thread after backing 660 * out the adjustment that was made to curthread when the original 661 * was preempted. 662 */ 663 if ((ntd = td->td_preempted) != NULL) { 664 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK); 665 ntd->td_flags |= TDF_PREEMPT_DONE; 666 ntd->td_contended = 0; /* reset contended */ 667 668 /* 669 * The interrupt may have woken a thread up, we need to properly 670 * set the reschedule flag if the originally interrupted thread is 671 * at a lower priority. 672 * 673 * NOTE: The interrupt may not have descheduled ntd. 674 * 675 * NOTE: We do not reschedule if there are no threads on the runq. 676 * (ntd could be the idlethread). 677 */ 678 xtd = TAILQ_FIRST(&gd->gd_tdrunq); 679 if (xtd && xtd != ntd) 680 need_lwkt_resched(); 681 goto havethread_preempted; 682 } 683 684 /* 685 * Figure out switch target. If we cannot switch to our desired target 686 * look for a thread that we can switch to. 687 * 688 * NOTE! The limited spin loop and related parameters are extremely 689 * important for system performance, particularly for pipes and 690 * concurrent conflicting VM faults. 691 */ 692 clear_lwkt_resched(); 693 ntd = TAILQ_FIRST(&gd->gd_tdrunq); 694 695 if (ntd) { 696 do { 697 if (TD_TOKS_NOT_HELD(ntd) || 698 lwkt_getalltokens(ntd, (ntd->td_contended > lwkt_spin_loops))) 699 { 700 goto havethread; 701 } 702 ++ntd->td_contended; /* overflow ok */ 703 if (gd->gd_indefinite.type == 0) 704 indefinite_init(&gd->gd_indefinite, NULL, 0, 't'); 705 #ifdef LOOPMASK 706 if (tsc_frequency && rdtsc() - tsc_base > tsc_frequency) { 707 kprintf("lwkt_switch: excessive contended %d " 708 "thread %p\n", ntd->td_contended, ntd); 709 tsc_base = rdtsc(); 710 } 711 #endif 712 } while (ntd->td_contended < (lwkt_spin_loops >> 1)); 713 upri = ntd->td_upri; 714 715 /* 716 * Bleh, the thread we wanted to switch to has a contended token. 717 * See if we can switch to another thread. 718 * 719 * We generally don't want to do this because it represents a 720 * priority inversion, but contending tokens on the same cpu can 721 * cause real problems if we don't now that we have an exclusive 722 * priority mechanism over shared for tokens. 723 * 724 * The solution is to allow threads with pending tokens to compete 725 * for them (a lower priority thread will get less cpu once it 726 * returns from the kernel anyway). If a thread does not have 727 * any contending tokens, we go by td_pri and upri. 728 */ 729 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) { 730 if (TD_TOKS_NOT_HELD(ntd) && 731 ntd->td_pri < TDPRI_KERN_LPSCHED && upri > ntd->td_upri) { 732 continue; 733 } 734 if (upri < ntd->td_upri) 735 upri = ntd->td_upri; 736 737 /* 738 * Try this one. 739 */ 740 if (TD_TOKS_NOT_HELD(ntd) || 741 lwkt_getalltokens(ntd, (ntd->td_contended > lwkt_spin_loops))) { 742 goto havethread; 743 } 744 ++ntd->td_contended; /* overflow ok */ 745 } 746 747 /* 748 * Fall through, switch to idle thread to get us out of the current 749 * context. Since we were contended, prevent HLT by flagging a 750 * LWKT reschedule. 751 */ 752 need_lwkt_resched(); 753 } 754 755 /* 756 * We either contended on ntd or the runq is empty. We must switch 757 * through the idle thread to get out of the current context. 758 */ 759 ntd = &gd->gd_idlethread; 760 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) 761 ASSERT_NO_TOKENS_HELD(ntd); 762 cpu_time.cp_msg[0] = 0; 763 goto haveidle; 764 765 havethread: 766 /* 767 * Clear gd_idle_repeat when doing a normal switch to a non-idle 768 * thread. 769 */ 770 ntd->td_wmesg = NULL; 771 ntd->td_contended = 0; /* reset once scheduled */ 772 ++gd->gd_cnt.v_swtch; 773 gd->gd_idle_repeat = 0; 774 775 /* 776 * If we were busy waiting record final disposition 777 */ 778 if (gd->gd_indefinite.type) 779 indefinite_done(&gd->gd_indefinite); 780 781 havethread_preempted: 782 /* 783 * If the new target does not need the MP lock and we are holding it, 784 * release the MP lock. If the new target requires the MP lock we have 785 * already acquired it for the target. 786 */ 787 ; 788 haveidle: 789 KASSERT(ntd->td_critcount, 790 ("priority problem in lwkt_switch %d %d", 791 td->td_critcount, ntd->td_critcount)); 792 793 if (td != ntd) { 794 /* 795 * Execute the actual thread switch operation. This function 796 * returns to the current thread and returns the previous thread 797 * (which may be different from the thread we switched to). 798 * 799 * We are responsible for marking ntd as TDF_RUNNING. 800 */ 801 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0); 802 #ifdef DEBUG_LWKT_THREAD 803 ++switch_count; 804 #endif 805 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd); 806 ntd->td_flags |= TDF_RUNNING; 807 lwkt_switch_return(td->td_switch(ntd)); 808 /* ntd invalid, td_switch() can return a different thread_t */ 809 } 810 811 /* 812 * catch-all. XXX is this strictly needed? 813 */ 814 splz_check(); 815 816 /* NOTE: current cpu may have changed after switch */ 817 crit_exit_quick(td); 818 } 819 820 /* 821 * Called by assembly in the td_switch (thread restore path) for thread 822 * bootstrap cases which do not 'return' to lwkt_switch(). 823 */ 824 void 825 lwkt_switch_return(thread_t otd) 826 { 827 globaldata_t rgd; 828 #ifdef LOOPMASK 829 uint64_t tsc_base = rdtsc(); 830 #endif 831 int exiting; 832 833 exiting = otd->td_flags & TDF_EXITING; 834 cpu_ccfence(); 835 836 /* 837 * Check if otd was migrating. Now that we are on ntd we can finish 838 * up the migration. This is a bit messy but it is the only place 839 * where td is known to be fully descheduled. 840 * 841 * We can only activate the migration if otd was migrating but not 842 * held on the cpu due to a preemption chain. We still have to 843 * clear TDF_RUNNING on the old thread either way. 844 * 845 * We are responsible for clearing the previously running thread's 846 * TDF_RUNNING. 847 */ 848 if ((rgd = otd->td_migrate_gd) != NULL && 849 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) { 850 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) == 851 (TDF_MIGRATING | TDF_RUNNING)); 852 otd->td_migrate_gd = NULL; 853 otd->td_flags &= ~TDF_RUNNING; 854 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd); 855 } else { 856 otd->td_flags &= ~TDF_RUNNING; 857 } 858 859 /* 860 * Final exit validations (see lwp_wait()). Note that otd becomes 861 * invalid the *instant* we set TDF_MP_EXITSIG. 862 * 863 * Use the EXITING status loaded from before we clear TDF_RUNNING, 864 * because if it is not set otd becomes invalid the instant we clear 865 * TDF_RUNNING on it (otherwise, if the system is fast enough, we 866 * might 'steal' TDF_EXITING from another switch-return!). 867 */ 868 while (exiting) { 869 u_int mpflags; 870 871 mpflags = otd->td_mpflags; 872 cpu_ccfence(); 873 874 if (mpflags & TDF_MP_EXITWAIT) { 875 if (atomic_cmpset_int(&otd->td_mpflags, mpflags, 876 mpflags | TDF_MP_EXITSIG)) { 877 wakeup(otd); 878 break; 879 } 880 } else { 881 if (atomic_cmpset_int(&otd->td_mpflags, mpflags, 882 mpflags | TDF_MP_EXITSIG)) { 883 wakeup(otd); 884 break; 885 } 886 } 887 888 #ifdef LOOPMASK 889 if (tsc_frequency && rdtsc() - tsc_base > tsc_frequency) { 890 kprintf("lwkt_switch_return: excessive TDF_EXITING " 891 "thread %p\n", otd); 892 tsc_base = rdtsc(); 893 } 894 #endif 895 } 896 } 897 898 /* 899 * Request that the target thread preempt the current thread. Preemption 900 * can only occur only: 901 * 902 * - If our critical section is the one that we were called with 903 * - The relative priority of the target thread is higher 904 * - The target is not excessively interrupt-nested via td_nest_count 905 * - The target thread holds no tokens. 906 * - The target thread is not already scheduled and belongs to the 907 * current cpu. 908 * - The current thread is not holding any spin-locks. 909 * 910 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically 911 * this is called via lwkt_schedule() through the td_preemptable callback. 912 * critcount is the managed critical priority that we should ignore in order 913 * to determine whether preemption is possible (aka usually just the crit 914 * priority of lwkt_schedule() itself). 915 * 916 * Preemption is typically limited to interrupt threads. 917 * 918 * Operation works in a fairly straight-forward manner. The normal 919 * scheduling code is bypassed and we switch directly to the target 920 * thread. When the target thread attempts to block or switch away 921 * code at the base of lwkt_switch() will switch directly back to our 922 * thread. Our thread is able to retain whatever tokens it holds and 923 * if the target needs one of them the target will switch back to us 924 * and reschedule itself normally. 925 */ 926 void 927 lwkt_preempt(thread_t ntd, int critcount) 928 { 929 struct globaldata *gd = mycpu; 930 thread_t xtd; 931 thread_t td; 932 int save_gd_intr_nesting_level; 933 934 /* 935 * The caller has put us in a critical section. We can only preempt 936 * if the caller of the caller was not in a critical section (basically 937 * a local interrupt), as determined by the 'critcount' parameter. We 938 * also can't preempt if the caller is holding any spinlocks (even if 939 * he isn't in a critical section). This also handles the tokens test. 940 * 941 * YYY The target thread must be in a critical section (else it must 942 * inherit our critical section? I dunno yet). 943 */ 944 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri)); 945 946 td = gd->gd_curthread; 947 if (preempt_enable == 0) { 948 #ifdef DEBUG_LWKT_THREAD 949 ++preempt_miss; 950 #endif 951 return; 952 } 953 if (ntd->td_pri <= td->td_pri) { 954 #ifdef DEBUG_LWKT_THREAD 955 ++preempt_miss; 956 #endif 957 return; 958 } 959 if (td->td_critcount > critcount) { 960 #ifdef DEBUG_LWKT_THREAD 961 ++preempt_miss; 962 #endif 963 return; 964 } 965 if (td->td_nest_count >= 2) { 966 #ifdef DEBUG_LWKT_THREAD 967 ++preempt_miss; 968 #endif 969 return; 970 } 971 if (td->td_cscount) { 972 #ifdef DEBUG_LWKT_THREAD 973 ++preempt_miss; 974 #endif 975 return; 976 } 977 if (ntd->td_gd != gd) { 978 #ifdef DEBUG_LWKT_THREAD 979 ++preempt_miss; 980 #endif 981 return; 982 } 983 984 /* 985 * We don't have to check spinlocks here as they will also bump 986 * td_critcount. 987 * 988 * Do not try to preempt if the target thread is holding any tokens. 989 * We could try to acquire the tokens but this case is so rare there 990 * is no need to support it. 991 */ 992 KKASSERT(gd->gd_spinlocks == 0); 993 994 if (TD_TOKS_HELD(ntd)) { 995 #ifdef DEBUG_LWKT_THREAD 996 ++preempt_miss; 997 #endif 998 return; 999 } 1000 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) { 1001 #ifdef DEBUG_LWKT_THREAD 1002 ++preempt_weird; 1003 #endif 1004 return; 1005 } 1006 if (ntd->td_preempted) { 1007 #ifdef DEBUG_LWKT_THREAD 1008 ++preempt_hit; 1009 #endif 1010 return; 1011 } 1012 KKASSERT(gd->gd_processing_ipiq == 0); 1013 1014 /* 1015 * Since we are able to preempt the current thread, there is no need to 1016 * call need_lwkt_resched(). 1017 * 1018 * We must temporarily clear gd_intr_nesting_level around the switch 1019 * since switchouts from the target thread are allowed (they will just 1020 * return to our thread), and since the target thread has its own stack. 1021 * 1022 * A preemption must switch back to the original thread, assert the 1023 * case. 1024 */ 1025 #ifdef DEBUG_LWKT_THREAD 1026 ++preempt_hit; 1027 #endif 1028 ntd->td_preempted = td; 1029 td->td_flags |= TDF_PREEMPT_LOCK; 1030 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd); 1031 save_gd_intr_nesting_level = gd->gd_intr_nesting_level; 1032 gd->gd_intr_nesting_level = 0; 1033 1034 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0); 1035 ntd->td_flags |= TDF_RUNNING; 1036 xtd = td->td_switch(ntd); 1037 KKASSERT(xtd == ntd); 1038 lwkt_switch_return(xtd); 1039 gd->gd_intr_nesting_level = save_gd_intr_nesting_level; 1040 1041 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE)); 1042 ntd->td_preempted = NULL; 1043 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE); 1044 } 1045 1046 /* 1047 * Conditionally call splz() if gd_reqflags indicates work is pending. 1048 * This will work inside a critical section but not inside a hard code 1049 * section. 1050 * 1051 * (self contained on a per cpu basis) 1052 */ 1053 void 1054 splz_check(void) 1055 { 1056 globaldata_t gd = mycpu; 1057 thread_t td = gd->gd_curthread; 1058 1059 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && 1060 gd->gd_intr_nesting_level == 0 && 1061 td->td_nest_count < 2) 1062 { 1063 splz(); 1064 } 1065 } 1066 1067 /* 1068 * This version is integrated into crit_exit, reqflags has already 1069 * been tested but td_critcount has not. 1070 * 1071 * We only want to execute the splz() on the 1->0 transition of 1072 * critcount and not in a hard code section or if too deeply nested. 1073 * 1074 * NOTE: gd->gd_spinlocks is implied to be 0 when td_critcount is 0. 1075 */ 1076 void 1077 lwkt_maybe_splz(thread_t td) 1078 { 1079 globaldata_t gd = td->td_gd; 1080 1081 if (td->td_critcount == 0 && 1082 gd->gd_intr_nesting_level == 0 && 1083 td->td_nest_count < 2) 1084 { 1085 splz(); 1086 } 1087 } 1088 1089 /* 1090 * Drivers which set up processing co-threads can call this function to 1091 * run the co-thread at a higher priority and to allow it to preempt 1092 * normal threads. 1093 */ 1094 void 1095 lwkt_set_interrupt_support_thread(void) 1096 { 1097 thread_t td = curthread; 1098 1099 lwkt_setpri_self(TDPRI_INT_SUPPORT); 1100 td->td_flags |= TDF_INTTHREAD; 1101 td->td_preemptable = lwkt_preempt; 1102 } 1103 1104 1105 /* 1106 * This function is used to negotiate a passive release of the current 1107 * process/lwp designation with the user scheduler, allowing the user 1108 * scheduler to schedule another user thread. The related kernel thread 1109 * (curthread) continues running in the released state. 1110 */ 1111 void 1112 lwkt_passive_release(struct thread *td) 1113 { 1114 struct lwp *lp = td->td_lwp; 1115 1116 td->td_release = NULL; 1117 lwkt_setpri_self(TDPRI_KERN_USER); 1118 1119 lp->lwp_proc->p_usched->release_curproc(lp); 1120 } 1121 1122 1123 /* 1124 * This implements a LWKT yield, allowing a kernel thread to yield to other 1125 * kernel threads at the same or higher priority. This function can be 1126 * called in a tight loop and will typically only yield once per tick. 1127 * 1128 * Most kernel threads run at the same priority in order to allow equal 1129 * sharing. 1130 * 1131 * (self contained on a per cpu basis) 1132 */ 1133 void 1134 lwkt_yield(void) 1135 { 1136 globaldata_t gd = mycpu; 1137 thread_t td = gd->gd_curthread; 1138 1139 /* 1140 * Should never be called with spinlocks held but there is a path 1141 * via ACPI where it might happen. 1142 */ 1143 if (gd->gd_spinlocks) 1144 return; 1145 1146 /* 1147 * Safe to call splz if we are not too-heavily nested. 1148 */ 1149 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1150 splz(); 1151 1152 /* 1153 * Caller allows switching 1154 */ 1155 if (lwkt_resched_wanted()) { 1156 atomic_set_int(&td->td_mpflags, TDF_MP_DIDYIELD); 1157 lwkt_schedule_self(td); 1158 lwkt_switch(); 1159 } 1160 } 1161 1162 /* 1163 * The quick version processes pending interrupts and higher-priority 1164 * LWKT threads but will not round-robin same-priority LWKT threads. 1165 * 1166 * When called while attempting to return to userland the only same-pri 1167 * threads are the ones which have already tried to become the current 1168 * user process. 1169 */ 1170 void 1171 lwkt_yield_quick(void) 1172 { 1173 globaldata_t gd = mycpu; 1174 thread_t td = gd->gd_curthread; 1175 1176 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1177 splz(); 1178 if (lwkt_resched_wanted()) { 1179 crit_enter(); 1180 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) { 1181 clear_lwkt_resched(); 1182 } else { 1183 atomic_set_int(&td->td_mpflags, TDF_MP_DIDYIELD); 1184 lwkt_schedule_self(curthread); 1185 lwkt_switch(); 1186 } 1187 crit_exit(); 1188 } 1189 } 1190 1191 /* 1192 * This yield is designed for kernel threads with a user context. 1193 * 1194 * The kernel acting on behalf of the user is potentially cpu-bound, 1195 * this function will efficiently allow other threads to run and also 1196 * switch to other processes by releasing. 1197 * 1198 * The lwkt_user_yield() function is designed to have very low overhead 1199 * if no yield is determined to be needed. 1200 */ 1201 void 1202 lwkt_user_yield(void) 1203 { 1204 globaldata_t gd = mycpu; 1205 thread_t td = gd->gd_curthread; 1206 1207 /* 1208 * Should never be called with spinlocks held but there is a path 1209 * via ACPI where it might happen. 1210 */ 1211 if (gd->gd_spinlocks) 1212 return; 1213 1214 /* 1215 * Always run any pending interrupts in case we are in a critical 1216 * section. 1217 */ 1218 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1219 splz(); 1220 1221 /* 1222 * Switch (which forces a release) if another kernel thread needs 1223 * the cpu, if userland wants us to resched, or if our kernel 1224 * quantum has run out. 1225 */ 1226 if (lwkt_resched_wanted() || 1227 user_resched_wanted()) 1228 { 1229 lwkt_switch(); 1230 } 1231 1232 #if 0 1233 /* 1234 * Reacquire the current process if we are released. 1235 * 1236 * XXX not implemented atm. The kernel may be holding locks and such, 1237 * so we want the thread to continue to receive cpu. 1238 */ 1239 if (td->td_release == NULL && lp) { 1240 lp->lwp_proc->p_usched->acquire_curproc(lp); 1241 td->td_release = lwkt_passive_release; 1242 lwkt_setpri_self(TDPRI_USER_NORM); 1243 } 1244 #endif 1245 } 1246 1247 /* 1248 * Generic schedule. Possibly schedule threads belonging to other cpus and 1249 * deal with threads that might be blocked on a wait queue. 1250 * 1251 * We have a little helper inline function which does additional work after 1252 * the thread has been enqueued, including dealing with preemption and 1253 * setting need_lwkt_resched() (which prevents the kernel from returning 1254 * to userland until it has processed higher priority threads). 1255 * 1256 * It is possible for this routine to be called after a failed _enqueue 1257 * (due to the target thread migrating, sleeping, or otherwise blocked). 1258 * We have to check that the thread is actually on the run queue! 1259 */ 1260 static __inline 1261 void 1262 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount) 1263 { 1264 if (ntd->td_flags & TDF_RUNQ) { 1265 if (ntd->td_preemptable) { 1266 ntd->td_preemptable(ntd, ccount); /* YYY +token */ 1267 } 1268 } 1269 } 1270 1271 static __inline 1272 void 1273 _lwkt_schedule(thread_t td) 1274 { 1275 globaldata_t mygd = mycpu; 1276 1277 KASSERT(td != &td->td_gd->gd_idlethread, 1278 ("lwkt_schedule(): scheduling gd_idlethread is illegal!")); 1279 KKASSERT((td->td_flags & TDF_MIGRATING) == 0); 1280 crit_enter_gd(mygd); 1281 KKASSERT(td->td_lwp == NULL || 1282 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 1283 1284 if (td == mygd->gd_curthread) { 1285 _lwkt_enqueue(td); 1286 } else { 1287 /* 1288 * If we own the thread, there is no race (since we are in a 1289 * critical section). If we do not own the thread there might 1290 * be a race but the target cpu will deal with it. 1291 */ 1292 if (td->td_gd == mygd) { 1293 _lwkt_enqueue(td); 1294 _lwkt_schedule_post(mygd, td, 1); 1295 } else { 1296 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0); 1297 } 1298 } 1299 crit_exit_gd(mygd); 1300 } 1301 1302 void 1303 lwkt_schedule(thread_t td) 1304 { 1305 _lwkt_schedule(td); 1306 } 1307 1308 void 1309 lwkt_schedule_noresched(thread_t td) /* XXX not impl */ 1310 { 1311 _lwkt_schedule(td); 1312 } 1313 1314 /* 1315 * When scheduled remotely if frame != NULL the IPIQ is being 1316 * run via doreti or an interrupt then preemption can be allowed. 1317 * 1318 * To allow preemption we have to drop the critical section so only 1319 * one is present in _lwkt_schedule_post. 1320 */ 1321 static void 1322 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame) 1323 { 1324 thread_t td = curthread; 1325 thread_t ntd = arg; 1326 1327 if (frame && ntd->td_preemptable) { 1328 crit_exit_noyield(td); 1329 _lwkt_schedule(ntd); 1330 crit_enter_quick(td); 1331 } else { 1332 _lwkt_schedule(ntd); 1333 } 1334 } 1335 1336 /* 1337 * Thread migration using a 'Pull' method. The thread may or may not be 1338 * the current thread. It MUST be descheduled and in a stable state. 1339 * lwkt_giveaway() must be called on the cpu owning the thread. 1340 * 1341 * At any point after lwkt_giveaway() is called, the target cpu may 1342 * 'pull' the thread by calling lwkt_acquire(). 1343 * 1344 * We have to make sure the thread is not sitting on a per-cpu tsleep 1345 * queue or it will blow up when it moves to another cpu. 1346 * 1347 * MPSAFE - must be called under very specific conditions. 1348 */ 1349 void 1350 lwkt_giveaway(thread_t td) 1351 { 1352 globaldata_t gd = mycpu; 1353 1354 crit_enter_gd(gd); 1355 if (td->td_flags & TDF_TSLEEPQ) 1356 tsleep_remove(td); 1357 KKASSERT(td->td_gd == gd); 1358 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); 1359 td->td_flags |= TDF_MIGRATING; 1360 crit_exit_gd(gd); 1361 } 1362 1363 void 1364 lwkt_acquire(thread_t td) 1365 { 1366 globaldata_t gd; 1367 globaldata_t mygd; 1368 1369 KKASSERT(td->td_flags & TDF_MIGRATING); 1370 gd = td->td_gd; 1371 mygd = mycpu; 1372 if (gd != mycpu) { 1373 #ifdef LOOPMASK 1374 uint64_t tsc_base = rdtsc(); 1375 #endif 1376 cpu_lfence(); 1377 KKASSERT((td->td_flags & TDF_RUNQ) == 0); 1378 crit_enter_gd(mygd); 1379 DEBUG_PUSH_INFO("lwkt_acquire"); 1380 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) { 1381 lwkt_process_ipiq(); 1382 cpu_lfence(); 1383 #ifdef _KERNEL_VIRTUAL 1384 vkernel_yield(); 1385 #endif 1386 #ifdef LOOPMASK 1387 if (tsc_frequency && rdtsc() - tsc_base > tsc_frequency) { 1388 kprintf("lwkt_acquire: stuck td %p td->td_flags %08x\n", 1389 td, td->td_flags); 1390 tsc_base = rdtsc(); 1391 } 1392 #endif 1393 } 1394 DEBUG_POP_INFO(); 1395 cpu_mfence(); 1396 td->td_gd = mygd; 1397 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1398 td->td_flags &= ~TDF_MIGRATING; 1399 crit_exit_gd(mygd); 1400 } else { 1401 crit_enter_gd(mygd); 1402 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1403 td->td_flags &= ~TDF_MIGRATING; 1404 crit_exit_gd(mygd); 1405 } 1406 } 1407 1408 /* 1409 * Generic deschedule. Descheduling threads other then your own should be 1410 * done only in carefully controlled circumstances. Descheduling is 1411 * asynchronous. 1412 * 1413 * This function may block if the cpu has run out of messages. 1414 */ 1415 void 1416 lwkt_deschedule(thread_t td) 1417 { 1418 crit_enter(); 1419 if (td == curthread) { 1420 _lwkt_dequeue(td); 1421 } else { 1422 if (td->td_gd == mycpu) { 1423 _lwkt_dequeue(td); 1424 } else { 1425 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td); 1426 } 1427 } 1428 crit_exit(); 1429 } 1430 1431 /* 1432 * Set the target thread's priority. This routine does not automatically 1433 * switch to a higher priority thread, LWKT threads are not designed for 1434 * continuous priority changes. Yield if you want to switch. 1435 */ 1436 void 1437 lwkt_setpri(thread_t td, int pri) 1438 { 1439 if (td->td_pri != pri) { 1440 KKASSERT(pri >= 0); 1441 crit_enter(); 1442 if (td->td_flags & TDF_RUNQ) { 1443 KKASSERT(td->td_gd == mycpu); 1444 _lwkt_dequeue(td); 1445 td->td_pri = pri; 1446 _lwkt_enqueue(td); 1447 } else { 1448 td->td_pri = pri; 1449 } 1450 crit_exit(); 1451 } 1452 } 1453 1454 /* 1455 * Set the initial priority for a thread prior to it being scheduled for 1456 * the first time. The thread MUST NOT be scheduled before or during 1457 * this call. The thread may be assigned to a cpu other then the current 1458 * cpu. 1459 * 1460 * Typically used after a thread has been created with TDF_STOPPREQ, 1461 * and before the thread is initially scheduled. 1462 */ 1463 void 1464 lwkt_setpri_initial(thread_t td, int pri) 1465 { 1466 KKASSERT(pri >= 0); 1467 KKASSERT((td->td_flags & TDF_RUNQ) == 0); 1468 td->td_pri = pri; 1469 } 1470 1471 void 1472 lwkt_setpri_self(int pri) 1473 { 1474 thread_t td = curthread; 1475 1476 KKASSERT(pri >= 0 && pri <= TDPRI_MAX); 1477 crit_enter(); 1478 if (td->td_flags & TDF_RUNQ) { 1479 _lwkt_dequeue(td); 1480 td->td_pri = pri; 1481 _lwkt_enqueue(td); 1482 } else { 1483 td->td_pri = pri; 1484 } 1485 crit_exit(); 1486 } 1487 1488 /* 1489 * hz tick scheduler clock for LWKT threads 1490 */ 1491 void 1492 lwkt_schedulerclock(thread_t td) 1493 { 1494 globaldata_t gd = td->td_gd; 1495 thread_t xtd; 1496 1497 xtd = TAILQ_FIRST(&gd->gd_tdrunq); 1498 if (xtd == td) { 1499 /* 1500 * If the current thread is at the head of the runq shift it to the 1501 * end of any equal-priority threads and request a LWKT reschedule 1502 * if it moved. 1503 * 1504 * Ignore upri in this situation. There will only be one user thread 1505 * in user mode, all others will be user threads running in kernel 1506 * mode and we have to make sure they get some cpu. 1507 */ 1508 xtd = TAILQ_NEXT(td, td_threadq); 1509 if (xtd && xtd->td_pri == td->td_pri) { 1510 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq); 1511 while (xtd && xtd->td_pri == td->td_pri) 1512 xtd = TAILQ_NEXT(xtd, td_threadq); 1513 if (xtd) 1514 TAILQ_INSERT_BEFORE(xtd, td, td_threadq); 1515 else 1516 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 1517 need_lwkt_resched(); 1518 } 1519 } else if (xtd) { 1520 /* 1521 * If we scheduled a thread other than the one at the head of the 1522 * queue always request a reschedule every tick. 1523 */ 1524 need_lwkt_resched(); 1525 } 1526 /* else curthread probably the idle thread, no need to reschedule */ 1527 } 1528 1529 /* 1530 * Migrate the current thread to the specified cpu. 1531 * 1532 * This is accomplished by descheduling ourselves from the current cpu 1533 * and setting td_migrate_gd. The lwkt_switch() code will detect that the 1534 * 'old' thread wants to migrate after it has been completely switched out 1535 * and will complete the migration. 1536 * 1537 * TDF_MIGRATING prevents scheduling races while the thread is being migrated. 1538 * 1539 * We must be sure to release our current process designation (if a user 1540 * process) before clearing out any tsleepq we are on because the release 1541 * code may re-add us. 1542 * 1543 * We must be sure to remove ourselves from the current cpu's tsleepq 1544 * before potentially moving to another queue. The thread can be on 1545 * a tsleepq due to a left-over tsleep_interlock(). 1546 */ 1547 1548 void 1549 lwkt_setcpu_self(globaldata_t rgd) 1550 { 1551 thread_t td = curthread; 1552 1553 if (td->td_gd != rgd) { 1554 crit_enter_quick(td); 1555 1556 if (td->td_release) 1557 td->td_release(td); 1558 if (td->td_flags & TDF_TSLEEPQ) 1559 tsleep_remove(td); 1560 1561 /* 1562 * Set TDF_MIGRATING to prevent a spurious reschedule while we are 1563 * trying to deschedule ourselves and switch away, then deschedule 1564 * ourself, remove us from tdallq, and set td_migrate_gd. Finally, 1565 * call lwkt_switch() to complete the operation. 1566 */ 1567 td->td_flags |= TDF_MIGRATING; 1568 lwkt_deschedule_self(td); 1569 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1570 td->td_migrate_gd = rgd; 1571 lwkt_switch(); 1572 1573 /* 1574 * We are now on the target cpu 1575 */ 1576 KKASSERT(rgd == mycpu); 1577 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); 1578 crit_exit_quick(td); 1579 } 1580 } 1581 1582 void 1583 lwkt_migratecpu(int cpuid) 1584 { 1585 globaldata_t rgd; 1586 1587 rgd = globaldata_find(cpuid); 1588 lwkt_setcpu_self(rgd); 1589 } 1590 1591 /* 1592 * Remote IPI for cpu migration (called while in a critical section so we 1593 * do not have to enter another one). 1594 * 1595 * The thread (td) has already been completely descheduled from the 1596 * originating cpu and we can simply assert the case. The thread is 1597 * assigned to the new cpu and enqueued. 1598 * 1599 * The thread will re-add itself to tdallq when it resumes execution. 1600 */ 1601 static void 1602 lwkt_setcpu_remote(void *arg) 1603 { 1604 thread_t td = arg; 1605 globaldata_t gd = mycpu; 1606 1607 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0); 1608 td->td_gd = gd; 1609 cpu_mfence(); 1610 td->td_flags &= ~TDF_MIGRATING; 1611 KKASSERT(td->td_migrate_gd == NULL); 1612 KKASSERT(td->td_lwp == NULL || 1613 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 1614 _lwkt_enqueue(td); 1615 } 1616 1617 struct lwp * 1618 lwkt_preempted_proc(void) 1619 { 1620 thread_t td = curthread; 1621 while (td->td_preempted) 1622 td = td->td_preempted; 1623 return(td->td_lwp); 1624 } 1625 1626 /* 1627 * Create a kernel process/thread/whatever. It shares it's address space 1628 * with proc0 - ie: kernel only. 1629 * 1630 * If the cpu is not specified one will be selected. In the future 1631 * specifying a cpu of -1 will enable kernel thread migration between 1632 * cpus. 1633 */ 1634 int 1635 lwkt_create(void (*func)(void *), void *arg, struct thread **tdp, 1636 thread_t template, int tdflags, int cpu, const char *fmt, ...) 1637 { 1638 thread_t td; 1639 __va_list ap; 1640 1641 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu, 1642 tdflags); 1643 if (tdp) 1644 *tdp = td; 1645 cpu_set_thread_handler(td, lwkt_exit, func, arg); 1646 1647 /* 1648 * Set up arg0 for 'ps' etc 1649 */ 1650 __va_start(ap, fmt); 1651 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap); 1652 __va_end(ap); 1653 1654 /* 1655 * Schedule the thread to run 1656 */ 1657 if (td->td_flags & TDF_NOSTART) 1658 td->td_flags &= ~TDF_NOSTART; 1659 else 1660 lwkt_schedule(td); 1661 return 0; 1662 } 1663 1664 /* 1665 * Destroy an LWKT thread. Warning! This function is not called when 1666 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and 1667 * uses a different reaping mechanism. 1668 */ 1669 void 1670 lwkt_exit(void) 1671 { 1672 thread_t td = curthread; 1673 thread_t std; 1674 globaldata_t gd; 1675 1676 /* 1677 * Do any cleanup that might block here 1678 */ 1679 biosched_done(td); 1680 dsched_exit_thread(td); 1681 1682 /* 1683 * Get us into a critical section to interlock gd_freetd and loop 1684 * until we can get it freed. 1685 * 1686 * We have to cache the current td in gd_freetd because objcache_put()ing 1687 * it would rip it out from under us while our thread is still active. 1688 * 1689 * We are the current thread so of course our own TDF_RUNNING bit will 1690 * be set, so unlike the lwp reap code we don't wait for it to clear. 1691 */ 1692 gd = mycpu; 1693 crit_enter_quick(td); 1694 for (;;) { 1695 if (td->td_refs) { 1696 tsleep(td, 0, "tdreap", 1); 1697 continue; 1698 } 1699 if ((std = gd->gd_freetd) != NULL) { 1700 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0); 1701 gd->gd_freetd = NULL; 1702 objcache_put(thread_cache, std); 1703 continue; 1704 } 1705 break; 1706 } 1707 1708 /* 1709 * Remove thread resources from kernel lists and deschedule us for 1710 * the last time. We cannot block after this point or we may end 1711 * up with a stale td on the tsleepq. 1712 * 1713 * None of this may block, the critical section is the only thing 1714 * protecting tdallq and the only thing preventing new lwkt_hold() 1715 * thread refs now. 1716 */ 1717 if (td->td_flags & TDF_TSLEEPQ) 1718 tsleep_remove(td); 1719 lwkt_deschedule_self(td); 1720 lwkt_remove_tdallq(td); 1721 KKASSERT(td->td_refs == 0); 1722 1723 /* 1724 * Final cleanup 1725 */ 1726 KKASSERT(gd->gd_freetd == NULL); 1727 if (td->td_flags & TDF_ALLOCATED_THREAD) 1728 gd->gd_freetd = td; 1729 cpu_thread_exit(); 1730 } 1731 1732 void 1733 lwkt_remove_tdallq(thread_t td) 1734 { 1735 KKASSERT(td->td_gd == mycpu); 1736 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1737 } 1738 1739 /* 1740 * Code reduction and branch prediction improvements. Call/return 1741 * overhead on modern cpus often degenerates into 0 cycles due to 1742 * the cpu's branch prediction hardware and return pc cache. We 1743 * can take advantage of this by not inlining medium-complexity 1744 * functions and we can also reduce the branch prediction impact 1745 * by collapsing perfectly predictable branches into a single 1746 * procedure instead of duplicating it. 1747 * 1748 * Is any of this noticeable? Probably not, so I'll take the 1749 * smaller code size. 1750 */ 1751 void 1752 crit_exit_wrapper(__DEBUG_CRIT_ARG__) 1753 { 1754 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__); 1755 } 1756 1757 void 1758 crit_panic(void) 1759 { 1760 thread_t td = curthread; 1761 int lcrit = td->td_critcount; 1762 1763 td->td_critcount = 0; 1764 cpu_ccfence(); 1765 panic("td_critcount is/would-go negative! %p %d", td, lcrit); 1766 /* NOT REACHED */ 1767 } 1768 1769 /* 1770 * Called from debugger/panic on cpus which have been stopped. We must still 1771 * process the IPIQ while stopped. 1772 * 1773 * If we are dumping also try to process any pending interrupts. This may 1774 * or may not work depending on the state of the cpu at the point it was 1775 * stopped. 1776 */ 1777 void 1778 lwkt_smp_stopped(void) 1779 { 1780 globaldata_t gd = mycpu; 1781 1782 if (dumping) { 1783 lwkt_process_ipiq(); 1784 --gd->gd_intr_nesting_level; 1785 splz(); 1786 ++gd->gd_intr_nesting_level; 1787 } else { 1788 lwkt_process_ipiq(); 1789 } 1790 cpu_smp_stopped(); 1791 } 1792