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/queue.h> 49 #include <sys/sysctl.h> 50 #include <sys/kthread.h> 51 #include <machine/cpu.h> 52 #include <sys/lock.h> 53 #include <sys/spinlock.h> 54 #include <sys/ktr.h> 55 56 #include <sys/thread2.h> 57 #include <sys/spinlock2.h> 58 #include <sys/mplock2.h> 59 60 #include <sys/dsched.h> 61 62 #include <vm/vm.h> 63 #include <vm/vm_param.h> 64 #include <vm/vm_kern.h> 65 #include <vm/vm_object.h> 66 #include <vm/vm_page.h> 67 #include <vm/vm_map.h> 68 #include <vm/vm_pager.h> 69 #include <vm/vm_extern.h> 70 71 #include <machine/stdarg.h> 72 #include <machine/smp.h> 73 74 #if !defined(KTR_CTXSW) 75 #define KTR_CTXSW KTR_ALL 76 #endif 77 KTR_INFO_MASTER(ctxsw); 78 KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p", int cpu, struct thread *td); 79 KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p", int cpu, struct thread *td); 80 KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s", struct thread *td, char *comm); 81 KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", struct thread *td); 82 83 static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads"); 84 85 #ifdef INVARIANTS 86 static int panic_on_cscount = 0; 87 #endif 88 static __int64_t switch_count = 0; 89 static __int64_t preempt_hit = 0; 90 static __int64_t preempt_miss = 0; 91 static __int64_t preempt_weird = 0; 92 static int lwkt_use_spin_port; 93 static struct objcache *thread_cache; 94 95 static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame); 96 static void lwkt_setcpu_remote(void *arg); 97 98 extern void cpu_heavy_restore(void); 99 extern void cpu_lwkt_restore(void); 100 extern void cpu_kthread_restore(void); 101 extern void cpu_idle_restore(void); 102 103 /* 104 * We can make all thread ports use the spin backend instead of the thread 105 * backend. This should only be set to debug the spin backend. 106 */ 107 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port); 108 109 #ifdef INVARIANTS 110 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, 111 "Panic if attempting to switch lwkt's while mastering cpusync"); 112 #endif 113 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, 114 "Number of switched threads"); 115 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, 116 "Successful preemption events"); 117 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, 118 "Failed preemption events"); 119 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, 120 "Number of preempted threads."); 121 static int fairq_enable = 0; 122 SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW, 123 &fairq_enable, 0, "Turn on fairq priority accumulators"); 124 static int fairq_bypass = -1; 125 SYSCTL_INT(_lwkt, OID_AUTO, fairq_bypass, CTLFLAG_RW, 126 &fairq_bypass, 0, "Allow fairq to bypass td on token failure"); 127 extern int lwkt_sched_debug; 128 int lwkt_sched_debug = 0; 129 SYSCTL_INT(_lwkt, OID_AUTO, sched_debug, CTLFLAG_RW, 130 &lwkt_sched_debug, 0, "Scheduler debug"); 131 static int lwkt_spin_loops = 10; 132 SYSCTL_INT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW, 133 &lwkt_spin_loops, 0, "Scheduler spin loops until sorted decon"); 134 static int lwkt_spin_reseq = 0; 135 SYSCTL_INT(_lwkt, OID_AUTO, spin_reseq, CTLFLAG_RW, 136 &lwkt_spin_reseq, 0, "Scheduler resequencer enable"); 137 static int lwkt_spin_monitor = 0; 138 SYSCTL_INT(_lwkt, OID_AUTO, spin_monitor, CTLFLAG_RW, 139 &lwkt_spin_monitor, 0, "Scheduler uses monitor/mwait"); 140 static int lwkt_spin_fatal = 0; /* disabled */ 141 SYSCTL_INT(_lwkt, OID_AUTO, spin_fatal, CTLFLAG_RW, 142 &lwkt_spin_fatal, 0, "LWKT scheduler spin loops till fatal panic"); 143 static int preempt_enable = 1; 144 SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW, 145 &preempt_enable, 0, "Enable preemption"); 146 static int lwkt_cache_threads = 0; 147 SYSCTL_INT(_lwkt, OID_AUTO, cache_threads, CTLFLAG_RD, 148 &lwkt_cache_threads, 0, "thread+kstack cache"); 149 150 static __cachealign int lwkt_cseq_rindex; 151 static __cachealign int lwkt_cseq_windex; 152 153 /* 154 * These helper procedures handle the runq, they can only be called from 155 * within a critical section. 156 * 157 * WARNING! Prior to SMP being brought up it is possible to enqueue and 158 * dequeue threads belonging to other cpus, so be sure to use td->td_gd 159 * instead of 'mycpu' when referencing the globaldata structure. Once 160 * SMP live enqueuing and dequeueing only occurs on the current cpu. 161 */ 162 static __inline 163 void 164 _lwkt_dequeue(thread_t td) 165 { 166 if (td->td_flags & TDF_RUNQ) { 167 struct globaldata *gd = td->td_gd; 168 169 td->td_flags &= ~TDF_RUNQ; 170 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq); 171 --gd->gd_tdrunqcount; 172 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL) 173 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING); 174 } 175 } 176 177 /* 178 * Priority enqueue. 179 * 180 * There are a limited number of lwkt threads runnable since user 181 * processes only schedule one at a time per cpu. However, there can 182 * be many user processes in kernel mode exiting from a tsleep() which 183 * become runnable. 184 * 185 * NOTE: lwkt_schedulerclock() will force a round-robin based on td_pri and 186 * will ignore user priority. This is to ensure that user threads in 187 * kernel mode get cpu at some point regardless of what the user 188 * scheduler thinks. 189 */ 190 static __inline 191 void 192 _lwkt_enqueue(thread_t td) 193 { 194 thread_t xtd; 195 196 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) { 197 struct globaldata *gd = td->td_gd; 198 199 td->td_flags |= TDF_RUNQ; 200 xtd = TAILQ_FIRST(&gd->gd_tdrunq); 201 if (xtd == NULL) { 202 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 203 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING); 204 } else { 205 /* 206 * NOTE: td_upri - higher numbers more desireable, same sense 207 * as td_pri (typically reversed from lwp_upri). 208 * 209 * In the equal priority case we want the best selection 210 * at the beginning so the less desireable selections know 211 * that they have to setrunqueue/go-to-another-cpu, even 212 * though it means switching back to the 'best' selection. 213 * This also avoids degenerate situations when many threads 214 * are runnable or waking up at the same time. 215 * 216 * If upri matches exactly place at end/round-robin. 217 */ 218 while (xtd && 219 (xtd->td_pri >= td->td_pri || 220 (xtd->td_pri == td->td_pri && 221 xtd->td_upri >= td->td_upri))) { 222 xtd = TAILQ_NEXT(xtd, td_threadq); 223 } 224 if (xtd) 225 TAILQ_INSERT_BEFORE(xtd, td, td_threadq); 226 else 227 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 228 } 229 ++gd->gd_tdrunqcount; 230 231 /* 232 * Request a LWKT reschedule if we are now at the head of the queue. 233 */ 234 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) 235 need_lwkt_resched(); 236 } 237 } 238 239 static __boolean_t 240 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags) 241 { 242 struct thread *td = (struct thread *)obj; 243 244 td->td_kstack = NULL; 245 td->td_kstack_size = 0; 246 td->td_flags = TDF_ALLOCATED_THREAD; 247 td->td_mpflags = 0; 248 return (1); 249 } 250 251 static void 252 _lwkt_thread_dtor(void *obj, void *privdata) 253 { 254 struct thread *td = (struct thread *)obj; 255 256 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD, 257 ("_lwkt_thread_dtor: not allocated from objcache")); 258 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack && 259 td->td_kstack_size > 0, 260 ("_lwkt_thread_dtor: corrupted stack")); 261 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size); 262 td->td_kstack = NULL; 263 td->td_flags = 0; 264 } 265 266 /* 267 * Initialize the lwkt s/system. 268 * 269 * Nominally cache up to 32 thread + kstack structures. Cache more on 270 * systems with a lot of cpu cores. 271 */ 272 void 273 lwkt_init(void) 274 { 275 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads); 276 if (lwkt_cache_threads == 0) { 277 lwkt_cache_threads = ncpus * 4; 278 if (lwkt_cache_threads < 32) 279 lwkt_cache_threads = 32; 280 } 281 thread_cache = objcache_create_mbacked( 282 M_THREAD, sizeof(struct thread), 283 0, lwkt_cache_threads, 284 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL); 285 } 286 287 /* 288 * Schedule a thread to run. As the current thread we can always safely 289 * schedule ourselves, and a shortcut procedure is provided for that 290 * function. 291 * 292 * (non-blocking, self contained on a per cpu basis) 293 */ 294 void 295 lwkt_schedule_self(thread_t td) 296 { 297 KKASSERT((td->td_flags & TDF_MIGRATING) == 0); 298 crit_enter_quick(td); 299 KASSERT(td != &td->td_gd->gd_idlethread, 300 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!")); 301 KKASSERT(td->td_lwp == NULL || 302 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 303 _lwkt_enqueue(td); 304 crit_exit_quick(td); 305 } 306 307 /* 308 * Deschedule a thread. 309 * 310 * (non-blocking, self contained on a per cpu basis) 311 */ 312 void 313 lwkt_deschedule_self(thread_t td) 314 { 315 crit_enter_quick(td); 316 _lwkt_dequeue(td); 317 crit_exit_quick(td); 318 } 319 320 /* 321 * LWKTs operate on a per-cpu basis 322 * 323 * WARNING! Called from early boot, 'mycpu' may not work yet. 324 */ 325 void 326 lwkt_gdinit(struct globaldata *gd) 327 { 328 TAILQ_INIT(&gd->gd_tdrunq); 329 TAILQ_INIT(&gd->gd_tdallq); 330 } 331 332 /* 333 * Create a new thread. The thread must be associated with a process context 334 * or LWKT start address before it can be scheduled. If the target cpu is 335 * -1 the thread will be created on the current cpu. 336 * 337 * If you intend to create a thread without a process context this function 338 * does everything except load the startup and switcher function. 339 */ 340 thread_t 341 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags) 342 { 343 static int cpu_rotator; 344 globaldata_t gd = mycpu; 345 void *stack; 346 347 /* 348 * If static thread storage is not supplied allocate a thread. Reuse 349 * a cached free thread if possible. gd_freetd is used to keep an exiting 350 * thread intact through the exit. 351 */ 352 if (td == NULL) { 353 crit_enter_gd(gd); 354 if ((td = gd->gd_freetd) != NULL) { 355 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK| 356 TDF_RUNQ)) == 0); 357 gd->gd_freetd = NULL; 358 } else { 359 td = objcache_get(thread_cache, M_WAITOK); 360 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK| 361 TDF_RUNQ)) == 0); 362 } 363 crit_exit_gd(gd); 364 KASSERT((td->td_flags & 365 (TDF_ALLOCATED_THREAD|TDF_RUNNING|TDF_PREEMPT_LOCK)) == 366 TDF_ALLOCATED_THREAD, 367 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags)); 368 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK); 369 } 370 371 /* 372 * Try to reuse cached stack. 373 */ 374 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) { 375 if (flags & TDF_ALLOCATED_STACK) { 376 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size); 377 stack = NULL; 378 } 379 } 380 if (stack == NULL) { 381 stack = (void *)kmem_alloc_stack(&kernel_map, stksize); 382 flags |= TDF_ALLOCATED_STACK; 383 } 384 if (cpu < 0) { 385 cpu = ++cpu_rotator; 386 cpu_ccfence(); 387 cpu %= ncpus; 388 } 389 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu)); 390 return(td); 391 } 392 393 /* 394 * Initialize a preexisting thread structure. This function is used by 395 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread. 396 * 397 * All threads start out in a critical section at a priority of 398 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as 399 * appropriate. This function may send an IPI message when the 400 * requested cpu is not the current cpu and consequently gd_tdallq may 401 * not be initialized synchronously from the point of view of the originating 402 * cpu. 403 * 404 * NOTE! we have to be careful in regards to creating threads for other cpus 405 * if SMP has not yet been activated. 406 */ 407 static void 408 lwkt_init_thread_remote(void *arg) 409 { 410 thread_t td = arg; 411 412 /* 413 * Protected by critical section held by IPI dispatch 414 */ 415 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq); 416 } 417 418 /* 419 * lwkt core thread structural initialization. 420 * 421 * NOTE: All threads are initialized as mpsafe threads. 422 */ 423 void 424 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags, 425 struct globaldata *gd) 426 { 427 globaldata_t mygd = mycpu; 428 429 bzero(td, sizeof(struct thread)); 430 td->td_kstack = stack; 431 td->td_kstack_size = stksize; 432 td->td_flags = flags; 433 td->td_mpflags = 0; 434 td->td_type = TD_TYPE_GENERIC; 435 td->td_gd = gd; 436 td->td_pri = TDPRI_KERN_DAEMON; 437 td->td_critcount = 1; 438 td->td_toks_have = NULL; 439 td->td_toks_stop = &td->td_toks_base; 440 if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT)) 441 lwkt_initport_spin(&td->td_msgport, td); 442 else 443 lwkt_initport_thread(&td->td_msgport, td); 444 pmap_init_thread(td); 445 /* 446 * Normally initializing a thread for a remote cpu requires sending an 447 * IPI. However, the idlethread is setup before the other cpus are 448 * activated so we have to treat it as a special case. XXX manipulation 449 * of gd_tdallq requires the BGL. 450 */ 451 if (gd == mygd || td == &gd->gd_idlethread) { 452 crit_enter_gd(mygd); 453 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); 454 crit_exit_gd(mygd); 455 } else { 456 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td); 457 } 458 dsched_new_thread(td); 459 } 460 461 void 462 lwkt_set_comm(thread_t td, const char *ctl, ...) 463 { 464 __va_list va; 465 466 __va_start(va, ctl); 467 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va); 468 __va_end(va); 469 KTR_LOG(ctxsw_newtd, td, td->td_comm); 470 } 471 472 /* 473 * Prevent the thread from getting destroyed. Note that unlike PHOLD/PRELE 474 * this does not prevent the thread from migrating to another cpu so the 475 * gd_tdallq state is not protected by this. 476 */ 477 void 478 lwkt_hold(thread_t td) 479 { 480 atomic_add_int(&td->td_refs, 1); 481 } 482 483 void 484 lwkt_rele(thread_t td) 485 { 486 KKASSERT(td->td_refs > 0); 487 atomic_add_int(&td->td_refs, -1); 488 } 489 490 void 491 lwkt_free_thread(thread_t td) 492 { 493 KKASSERT(td->td_refs == 0); 494 KKASSERT((td->td_flags & (TDF_RUNNING | TDF_PREEMPT_LOCK | 495 TDF_RUNQ | TDF_TSLEEPQ)) == 0); 496 if (td->td_flags & TDF_ALLOCATED_THREAD) { 497 objcache_put(thread_cache, td); 498 } else if (td->td_flags & TDF_ALLOCATED_STACK) { 499 /* client-allocated struct with internally allocated stack */ 500 KASSERT(td->td_kstack && td->td_kstack_size > 0, 501 ("lwkt_free_thread: corrupted stack")); 502 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size); 503 td->td_kstack = NULL; 504 td->td_kstack_size = 0; 505 } 506 KTR_LOG(ctxsw_deadtd, td); 507 } 508 509 510 /* 511 * Switch to the next runnable lwkt. If no LWKTs are runnable then 512 * switch to the idlethread. Switching must occur within a critical 513 * section to avoid races with the scheduling queue. 514 * 515 * We always have full control over our cpu's run queue. Other cpus 516 * that wish to manipulate our queue must use the cpu_*msg() calls to 517 * talk to our cpu, so a critical section is all that is needed and 518 * the result is very, very fast thread switching. 519 * 520 * The LWKT scheduler uses a fixed priority model and round-robins at 521 * each priority level. User process scheduling is a totally 522 * different beast and LWKT priorities should not be confused with 523 * user process priorities. 524 * 525 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch() 526 * is not called by the current thread in the preemption case, only when 527 * the preempting thread blocks (in order to return to the original thread). 528 * 529 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread 530 * migration and tsleep deschedule the current lwkt thread and call 531 * lwkt_switch(). In particular, the target cpu of the migration fully 532 * expects the thread to become non-runnable and can deadlock against 533 * cpusync operations if we run any IPIs prior to switching the thread out. 534 * 535 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF 536 * THE CURRENT THREAD HAS BEEN DESCHEDULED! 537 */ 538 void 539 lwkt_switch(void) 540 { 541 globaldata_t gd = mycpu; 542 thread_t td = gd->gd_curthread; 543 thread_t ntd; 544 int spinning = 0; 545 546 KKASSERT(gd->gd_processing_ipiq == 0); 547 KKASSERT(td->td_flags & TDF_RUNNING); 548 549 /* 550 * Switching from within a 'fast' (non thread switched) interrupt or IPI 551 * is illegal. However, we may have to do it anyway if we hit a fatal 552 * kernel trap or we have paniced. 553 * 554 * If this case occurs save and restore the interrupt nesting level. 555 */ 556 if (gd->gd_intr_nesting_level) { 557 int savegdnest; 558 int savegdtrap; 559 560 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) { 561 panic("lwkt_switch: Attempt to switch from a " 562 "fast interrupt, ipi, or hard code section, " 563 "td %p\n", 564 td); 565 } else { 566 savegdnest = gd->gd_intr_nesting_level; 567 savegdtrap = gd->gd_trap_nesting_level; 568 gd->gd_intr_nesting_level = 0; 569 gd->gd_trap_nesting_level = 0; 570 if ((td->td_flags & TDF_PANICWARN) == 0) { 571 td->td_flags |= TDF_PANICWARN; 572 kprintf("Warning: thread switch from interrupt, IPI, " 573 "or hard code section.\n" 574 "thread %p (%s)\n", td, td->td_comm); 575 print_backtrace(-1); 576 } 577 lwkt_switch(); 578 gd->gd_intr_nesting_level = savegdnest; 579 gd->gd_trap_nesting_level = savegdtrap; 580 return; 581 } 582 } 583 584 /* 585 * Release our current user process designation if we are blocking 586 * or if a user reschedule was requested. 587 * 588 * NOTE: This function is NOT called if we are switching into or 589 * returning from a preemption. 590 * 591 * NOTE: Releasing our current user process designation may cause 592 * it to be assigned to another thread, which in turn will 593 * cause us to block in the usched acquire code when we attempt 594 * to return to userland. 595 * 596 * NOTE: On SMP systems this can be very nasty when heavy token 597 * contention is present so we want to be careful not to 598 * release the designation gratuitously. 599 */ 600 if (td->td_release && 601 (user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) { 602 td->td_release(td); 603 } 604 605 /* 606 * Release all tokens 607 */ 608 crit_enter_gd(gd); 609 if (TD_TOKS_HELD(td)) 610 lwkt_relalltokens(td); 611 612 /* 613 * We had better not be holding any spin locks, but don't get into an 614 * endless panic loop. 615 */ 616 KASSERT(gd->gd_spinlocks == 0 || panicstr != NULL, 617 ("lwkt_switch: still holding %d exclusive spinlocks!", 618 gd->gd_spinlocks)); 619 620 621 #ifdef INVARIANTS 622 if (td->td_cscount) { 623 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n", 624 td); 625 if (panic_on_cscount) 626 panic("switching while mastering cpusync"); 627 } 628 #endif 629 630 /* 631 * If we had preempted another thread on this cpu, resume the preempted 632 * thread. This occurs transparently, whether the preempted thread 633 * was scheduled or not (it may have been preempted after descheduling 634 * itself). 635 * 636 * We have to setup the MP lock for the original thread after backing 637 * out the adjustment that was made to curthread when the original 638 * was preempted. 639 */ 640 if ((ntd = td->td_preempted) != NULL) { 641 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK); 642 ntd->td_flags |= TDF_PREEMPT_DONE; 643 644 /* 645 * The interrupt may have woken a thread up, we need to properly 646 * set the reschedule flag if the originally interrupted thread is 647 * at a lower priority. 648 * 649 * The interrupt may not have descheduled. 650 */ 651 if (TAILQ_FIRST(&gd->gd_tdrunq) != ntd) 652 need_lwkt_resched(); 653 goto havethread_preempted; 654 } 655 656 /* 657 * If we cannot obtain ownership of the tokens we cannot immediately 658 * schedule the target thread. 659 * 660 * Reminder: Again, we cannot afford to run any IPIs in this path if 661 * the current thread has been descheduled. 662 */ 663 for (;;) { 664 clear_lwkt_resched(); 665 666 /* 667 * Hotpath - pull the head of the run queue and attempt to schedule 668 * it. 669 */ 670 ntd = TAILQ_FIRST(&gd->gd_tdrunq); 671 672 if (ntd == NULL) { 673 /* 674 * Runq is empty, switch to idle to allow it to halt. 675 */ 676 ntd = &gd->gd_idlethread; 677 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) 678 ASSERT_NO_TOKENS_HELD(ntd); 679 cpu_time.cp_msg[0] = 0; 680 cpu_time.cp_stallpc = 0; 681 goto haveidle; 682 } 683 684 /* 685 * Hotpath - schedule ntd. 686 * 687 * NOTE: For UP there is no mplock and lwkt_getalltokens() 688 * always succeeds. 689 */ 690 if (TD_TOKS_NOT_HELD(ntd) || 691 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops))) 692 { 693 goto havethread; 694 } 695 696 /* 697 * Coldpath (SMP only since tokens always succeed on UP) 698 * 699 * We had some contention on the thread we wanted to schedule. 700 * What we do now is try to find a thread that we can schedule 701 * in its stead. 702 * 703 * The coldpath scan does NOT rearrange threads in the run list. 704 * The lwkt_schedulerclock() will assert need_lwkt_resched() on 705 * the next tick whenever the current head is not the current thread. 706 */ 707 #ifdef INVARIANTS 708 ++ntd->td_contended; 709 #endif 710 ++gd->gd_cnt.v_token_colls; 711 712 if (fairq_bypass > 0) 713 goto skip; 714 715 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) { 716 #ifndef NO_LWKT_SPLIT_USERPRI 717 /* 718 * Never schedule threads returning to userland or the 719 * user thread scheduler helper thread when higher priority 720 * threads are present. The runq is sorted by priority 721 * so we can give up traversing it when we find the first 722 * low priority thread. 723 */ 724 if (ntd->td_pri < TDPRI_KERN_LPSCHED) { 725 ntd = NULL; 726 break; 727 } 728 #endif 729 730 /* 731 * Try this one. 732 */ 733 if (TD_TOKS_NOT_HELD(ntd) || 734 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops))) { 735 goto havethread; 736 } 737 #ifdef INVARIANTS 738 ++ntd->td_contended; 739 #endif 740 ++gd->gd_cnt.v_token_colls; 741 } 742 743 skip: 744 /* 745 * We exhausted the run list, meaning that all runnable threads 746 * are contested. 747 */ 748 cpu_pause(); 749 ntd = &gd->gd_idlethread; 750 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) 751 ASSERT_NO_TOKENS_HELD(ntd); 752 /* contention case, do not clear contention mask */ 753 754 /* 755 * We are going to have to retry but if the current thread is not 756 * on the runq we instead switch through the idle thread to get away 757 * from the current thread. We have to flag for lwkt reschedule 758 * to prevent the idle thread from halting. 759 * 760 * NOTE: A non-zero spinning is passed to lwkt_getalltokens() to 761 * instruct it to deal with the potential for deadlocks by 762 * ordering the tokens by address. 763 */ 764 if ((td->td_flags & TDF_RUNQ) == 0) { 765 need_lwkt_resched(); /* prevent hlt */ 766 goto haveidle; 767 } 768 #if defined(INVARIANTS) && defined(__x86_64__) 769 if ((read_rflags() & PSL_I) == 0) { 770 cpu_enable_intr(); 771 panic("lwkt_switch() called with interrupts disabled"); 772 } 773 #endif 774 775 /* 776 * Number iterations so far. After a certain point we switch to 777 * a sorted-address/monitor/mwait version of lwkt_getalltokens() 778 */ 779 if (spinning < 0x7FFFFFFF) 780 ++spinning; 781 782 /* 783 * lwkt_getalltokens() failed in sorted token mode, we can use 784 * monitor/mwait in this case. 785 */ 786 if (spinning >= lwkt_spin_loops && 787 (cpu_mi_feature & CPU_MI_MONITOR) && 788 lwkt_spin_monitor) 789 { 790 cpu_mmw_pause_int(&gd->gd_reqflags, 791 (gd->gd_reqflags | RQF_SPINNING) & 792 ~RQF_IDLECHECK_WK_MASK); 793 } 794 795 /* 796 * We already checked that td is still scheduled so this should be 797 * safe. 798 */ 799 splz_check(); 800 801 /* 802 * This experimental resequencer is used as a fall-back to reduce 803 * hw cache line contention by placing each core's scheduler into a 804 * time-domain-multplexed slot. 805 * 806 * The resequencer is disabled by default. It's functionality has 807 * largely been superceeded by the token algorithm which limits races 808 * to a subset of cores. 809 * 810 * The resequencer algorithm tends to break down when more than 811 * 20 cores are contending. What appears to happen is that new 812 * tokens can be obtained out of address-sorted order by new cores 813 * while existing cores languish in long delays between retries and 814 * wind up being starved-out of the token acquisition. 815 */ 816 if (lwkt_spin_reseq && spinning >= lwkt_spin_reseq) { 817 int cseq = atomic_fetchadd_int(&lwkt_cseq_windex, 1); 818 int oseq; 819 820 while ((oseq = lwkt_cseq_rindex) != cseq) { 821 cpu_ccfence(); 822 #if 1 823 if (cpu_mi_feature & CPU_MI_MONITOR) { 824 cpu_mmw_pause_int(&lwkt_cseq_rindex, oseq); 825 } else { 826 #endif 827 cpu_pause(); 828 cpu_lfence(); 829 #if 1 830 } 831 #endif 832 } 833 DELAY(1); 834 atomic_add_int(&lwkt_cseq_rindex, 1); 835 } 836 /* highest level for(;;) loop */ 837 } 838 839 havethread: 840 /* 841 * Clear gd_idle_repeat when doing a normal switch to a non-idle 842 * thread. 843 */ 844 ntd->td_wmesg = NULL; 845 ++gd->gd_cnt.v_swtch; 846 gd->gd_idle_repeat = 0; 847 848 havethread_preempted: 849 /* 850 * If the new target does not need the MP lock and we are holding it, 851 * release the MP lock. If the new target requires the MP lock we have 852 * already acquired it for the target. 853 */ 854 ; 855 haveidle: 856 KASSERT(ntd->td_critcount, 857 ("priority problem in lwkt_switch %d %d", 858 td->td_critcount, ntd->td_critcount)); 859 860 if (td != ntd) { 861 /* 862 * Execute the actual thread switch operation. This function 863 * returns to the current thread and returns the previous thread 864 * (which may be different from the thread we switched to). 865 * 866 * We are responsible for marking ntd as TDF_RUNNING. 867 */ 868 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0); 869 ++switch_count; 870 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd); 871 ntd->td_flags |= TDF_RUNNING; 872 lwkt_switch_return(td->td_switch(ntd)); 873 /* ntd invalid, td_switch() can return a different thread_t */ 874 } 875 876 /* 877 * catch-all. XXX is this strictly needed? 878 */ 879 splz_check(); 880 881 /* NOTE: current cpu may have changed after switch */ 882 crit_exit_quick(td); 883 } 884 885 /* 886 * Called by assembly in the td_switch (thread restore path) for thread 887 * bootstrap cases which do not 'return' to lwkt_switch(). 888 */ 889 void 890 lwkt_switch_return(thread_t otd) 891 { 892 globaldata_t rgd; 893 894 /* 895 * Check if otd was migrating. Now that we are on ntd we can finish 896 * up the migration. This is a bit messy but it is the only place 897 * where td is known to be fully descheduled. 898 * 899 * We can only activate the migration if otd was migrating but not 900 * held on the cpu due to a preemption chain. We still have to 901 * clear TDF_RUNNING on the old thread either way. 902 * 903 * We are responsible for clearing the previously running thread's 904 * TDF_RUNNING. 905 */ 906 if ((rgd = otd->td_migrate_gd) != NULL && 907 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) { 908 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) == 909 (TDF_MIGRATING | TDF_RUNNING)); 910 otd->td_migrate_gd = NULL; 911 otd->td_flags &= ~TDF_RUNNING; 912 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd); 913 } else { 914 otd->td_flags &= ~TDF_RUNNING; 915 } 916 917 /* 918 * Final exit validations (see lwp_wait()). Note that otd becomes 919 * invalid the *instant* we set TDF_MP_EXITSIG. 920 */ 921 while (otd->td_flags & TDF_EXITING) { 922 u_int mpflags; 923 924 mpflags = otd->td_mpflags; 925 cpu_ccfence(); 926 927 if (mpflags & TDF_MP_EXITWAIT) { 928 if (atomic_cmpset_int(&otd->td_mpflags, mpflags, 929 mpflags | TDF_MP_EXITSIG)) { 930 wakeup(otd); 931 break; 932 } 933 } else { 934 if (atomic_cmpset_int(&otd->td_mpflags, mpflags, 935 mpflags | TDF_MP_EXITSIG)) { 936 wakeup(otd); 937 break; 938 } 939 } 940 } 941 } 942 943 /* 944 * Request that the target thread preempt the current thread. Preemption 945 * can only occur if our only critical section is the one that we were called 946 * with, the relative priority of the target thread is higher, and the target 947 * thread holds no tokens. This also only works if we are not holding any 948 * spinlocks (obviously). 949 * 950 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically 951 * this is called via lwkt_schedule() through the td_preemptable callback. 952 * critcount is the managed critical priority that we should ignore in order 953 * to determine whether preemption is possible (aka usually just the crit 954 * priority of lwkt_schedule() itself). 955 * 956 * Preemption is typically limited to interrupt threads. 957 * 958 * Operation works in a fairly straight-forward manner. The normal 959 * scheduling code is bypassed and we switch directly to the target 960 * thread. When the target thread attempts to block or switch away 961 * code at the base of lwkt_switch() will switch directly back to our 962 * thread. Our thread is able to retain whatever tokens it holds and 963 * if the target needs one of them the target will switch back to us 964 * and reschedule itself normally. 965 */ 966 void 967 lwkt_preempt(thread_t ntd, int critcount) 968 { 969 struct globaldata *gd = mycpu; 970 thread_t xtd; 971 thread_t td; 972 int save_gd_intr_nesting_level; 973 974 /* 975 * The caller has put us in a critical section. We can only preempt 976 * if the caller of the caller was not in a critical section (basically 977 * a local interrupt), as determined by the 'critcount' parameter. We 978 * also can't preempt if the caller is holding any spinlocks (even if 979 * he isn't in a critical section). This also handles the tokens test. 980 * 981 * YYY The target thread must be in a critical section (else it must 982 * inherit our critical section? I dunno yet). 983 */ 984 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri)); 985 986 td = gd->gd_curthread; 987 if (preempt_enable == 0) { 988 ++preempt_miss; 989 return; 990 } 991 if (ntd->td_pri <= td->td_pri) { 992 ++preempt_miss; 993 return; 994 } 995 if (td->td_critcount > critcount) { 996 ++preempt_miss; 997 return; 998 } 999 if (td->td_cscount) { 1000 ++preempt_miss; 1001 return; 1002 } 1003 if (ntd->td_gd != gd) { 1004 ++preempt_miss; 1005 return; 1006 } 1007 /* 1008 * We don't have to check spinlocks here as they will also bump 1009 * td_critcount. 1010 * 1011 * Do not try to preempt if the target thread is holding any tokens. 1012 * We could try to acquire the tokens but this case is so rare there 1013 * is no need to support it. 1014 */ 1015 KKASSERT(gd->gd_spinlocks == 0); 1016 1017 if (TD_TOKS_HELD(ntd)) { 1018 ++preempt_miss; 1019 return; 1020 } 1021 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) { 1022 ++preempt_weird; 1023 return; 1024 } 1025 if (ntd->td_preempted) { 1026 ++preempt_hit; 1027 return; 1028 } 1029 KKASSERT(gd->gd_processing_ipiq == 0); 1030 1031 /* 1032 * Since we are able to preempt the current thread, there is no need to 1033 * call need_lwkt_resched(). 1034 * 1035 * We must temporarily clear gd_intr_nesting_level around the switch 1036 * since switchouts from the target thread are allowed (they will just 1037 * return to our thread), and since the target thread has its own stack. 1038 * 1039 * A preemption must switch back to the original thread, assert the 1040 * case. 1041 */ 1042 ++preempt_hit; 1043 ntd->td_preempted = td; 1044 td->td_flags |= TDF_PREEMPT_LOCK; 1045 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd); 1046 save_gd_intr_nesting_level = gd->gd_intr_nesting_level; 1047 gd->gd_intr_nesting_level = 0; 1048 1049 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0); 1050 ntd->td_flags |= TDF_RUNNING; 1051 xtd = td->td_switch(ntd); 1052 KKASSERT(xtd == ntd); 1053 lwkt_switch_return(xtd); 1054 gd->gd_intr_nesting_level = save_gd_intr_nesting_level; 1055 1056 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE)); 1057 ntd->td_preempted = NULL; 1058 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE); 1059 } 1060 1061 /* 1062 * Conditionally call splz() if gd_reqflags indicates work is pending. 1063 * This will work inside a critical section but not inside a hard code 1064 * section. 1065 * 1066 * (self contained on a per cpu basis) 1067 */ 1068 void 1069 splz_check(void) 1070 { 1071 globaldata_t gd = mycpu; 1072 thread_t td = gd->gd_curthread; 1073 1074 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && 1075 gd->gd_intr_nesting_level == 0 && 1076 td->td_nest_count < 2) 1077 { 1078 splz(); 1079 } 1080 } 1081 1082 /* 1083 * This version is integrated into crit_exit, reqflags has already 1084 * been tested but td_critcount has not. 1085 * 1086 * We only want to execute the splz() on the 1->0 transition of 1087 * critcount and not in a hard code section or if too deeply nested. 1088 * 1089 * NOTE: gd->gd_spinlocks is implied to be 0 when td_critcount is 0. 1090 */ 1091 void 1092 lwkt_maybe_splz(thread_t td) 1093 { 1094 globaldata_t gd = td->td_gd; 1095 1096 if (td->td_critcount == 0 && 1097 gd->gd_intr_nesting_level == 0 && 1098 td->td_nest_count < 2) 1099 { 1100 splz(); 1101 } 1102 } 1103 1104 /* 1105 * Drivers which set up processing co-threads can call this function to 1106 * run the co-thread at a higher priority and to allow it to preempt 1107 * normal threads. 1108 */ 1109 void 1110 lwkt_set_interrupt_support_thread(void) 1111 { 1112 thread_t td = curthread; 1113 1114 lwkt_setpri_self(TDPRI_INT_SUPPORT); 1115 td->td_flags |= TDF_INTTHREAD; 1116 td->td_preemptable = lwkt_preempt; 1117 } 1118 1119 1120 /* 1121 * This function is used to negotiate a passive release of the current 1122 * process/lwp designation with the user scheduler, allowing the user 1123 * scheduler to schedule another user thread. The related kernel thread 1124 * (curthread) continues running in the released state. 1125 */ 1126 void 1127 lwkt_passive_release(struct thread *td) 1128 { 1129 struct lwp *lp = td->td_lwp; 1130 1131 #ifndef NO_LWKT_SPLIT_USERPRI 1132 td->td_release = NULL; 1133 lwkt_setpri_self(TDPRI_KERN_USER); 1134 #endif 1135 1136 lp->lwp_proc->p_usched->release_curproc(lp); 1137 } 1138 1139 1140 /* 1141 * This implements a LWKT yield, allowing a kernel thread to yield to other 1142 * kernel threads at the same or higher priority. This function can be 1143 * called in a tight loop and will typically only yield once per tick. 1144 * 1145 * Most kernel threads run at the same priority in order to allow equal 1146 * sharing. 1147 * 1148 * (self contained on a per cpu basis) 1149 */ 1150 void 1151 lwkt_yield(void) 1152 { 1153 globaldata_t gd = mycpu; 1154 thread_t td = gd->gd_curthread; 1155 1156 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1157 splz(); 1158 if (lwkt_resched_wanted()) { 1159 lwkt_schedule_self(curthread); 1160 lwkt_switch(); 1161 } 1162 } 1163 1164 /* 1165 * The quick version processes pending interrupts and higher-priority 1166 * LWKT threads but will not round-robin same-priority LWKT threads. 1167 * 1168 * When called while attempting to return to userland the only same-pri 1169 * threads are the ones which have already tried to become the current 1170 * user process. 1171 */ 1172 void 1173 lwkt_yield_quick(void) 1174 { 1175 globaldata_t gd = mycpu; 1176 thread_t td = gd->gd_curthread; 1177 1178 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1179 splz(); 1180 if (lwkt_resched_wanted()) { 1181 crit_enter(); 1182 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) { 1183 clear_lwkt_resched(); 1184 } else { 1185 lwkt_schedule_self(curthread); 1186 lwkt_switch(); 1187 } 1188 crit_exit(); 1189 } 1190 } 1191 1192 /* 1193 * This yield is designed for kernel threads with a user context. 1194 * 1195 * The kernel acting on behalf of the user is potentially cpu-bound, 1196 * this function will efficiently allow other threads to run and also 1197 * switch to other processes by releasing. 1198 * 1199 * The lwkt_user_yield() function is designed to have very low overhead 1200 * if no yield is determined to be needed. 1201 */ 1202 void 1203 lwkt_user_yield(void) 1204 { 1205 globaldata_t gd = mycpu; 1206 thread_t td = gd->gd_curthread; 1207 1208 /* 1209 * Always run any pending interrupts in case we are in a critical 1210 * section. 1211 */ 1212 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1213 splz(); 1214 1215 /* 1216 * Switch (which forces a release) if another kernel thread needs 1217 * the cpu, if userland wants us to resched, or if our kernel 1218 * quantum has run out. 1219 */ 1220 if (lwkt_resched_wanted() || 1221 user_resched_wanted()) 1222 { 1223 lwkt_switch(); 1224 } 1225 1226 #if 0 1227 /* 1228 * Reacquire the current process if we are released. 1229 * 1230 * XXX not implemented atm. The kernel may be holding locks and such, 1231 * so we want the thread to continue to receive cpu. 1232 */ 1233 if (td->td_release == NULL && lp) { 1234 lp->lwp_proc->p_usched->acquire_curproc(lp); 1235 td->td_release = lwkt_passive_release; 1236 lwkt_setpri_self(TDPRI_USER_NORM); 1237 } 1238 #endif 1239 } 1240 1241 /* 1242 * Generic schedule. Possibly schedule threads belonging to other cpus and 1243 * deal with threads that might be blocked on a wait queue. 1244 * 1245 * We have a little helper inline function which does additional work after 1246 * the thread has been enqueued, including dealing with preemption and 1247 * setting need_lwkt_resched() (which prevents the kernel from returning 1248 * to userland until it has processed higher priority threads). 1249 * 1250 * It is possible for this routine to be called after a failed _enqueue 1251 * (due to the target thread migrating, sleeping, or otherwise blocked). 1252 * We have to check that the thread is actually on the run queue! 1253 */ 1254 static __inline 1255 void 1256 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount) 1257 { 1258 if (ntd->td_flags & TDF_RUNQ) { 1259 if (ntd->td_preemptable) { 1260 ntd->td_preemptable(ntd, ccount); /* YYY +token */ 1261 } 1262 } 1263 } 1264 1265 static __inline 1266 void 1267 _lwkt_schedule(thread_t td) 1268 { 1269 globaldata_t mygd = mycpu; 1270 1271 KASSERT(td != &td->td_gd->gd_idlethread, 1272 ("lwkt_schedule(): scheduling gd_idlethread is illegal!")); 1273 KKASSERT((td->td_flags & TDF_MIGRATING) == 0); 1274 crit_enter_gd(mygd); 1275 KKASSERT(td->td_lwp == NULL || 1276 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 1277 1278 if (td == mygd->gd_curthread) { 1279 _lwkt_enqueue(td); 1280 } else { 1281 /* 1282 * If we own the thread, there is no race (since we are in a 1283 * critical section). If we do not own the thread there might 1284 * be a race but the target cpu will deal with it. 1285 */ 1286 if (td->td_gd == mygd) { 1287 _lwkt_enqueue(td); 1288 _lwkt_schedule_post(mygd, td, 1); 1289 } else { 1290 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0); 1291 } 1292 } 1293 crit_exit_gd(mygd); 1294 } 1295 1296 void 1297 lwkt_schedule(thread_t td) 1298 { 1299 _lwkt_schedule(td); 1300 } 1301 1302 void 1303 lwkt_schedule_noresched(thread_t td) /* XXX not impl */ 1304 { 1305 _lwkt_schedule(td); 1306 } 1307 1308 /* 1309 * When scheduled remotely if frame != NULL the IPIQ is being 1310 * run via doreti or an interrupt then preemption can be allowed. 1311 * 1312 * To allow preemption we have to drop the critical section so only 1313 * one is present in _lwkt_schedule_post. 1314 */ 1315 static void 1316 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame) 1317 { 1318 thread_t td = curthread; 1319 thread_t ntd = arg; 1320 1321 if (frame && ntd->td_preemptable) { 1322 crit_exit_noyield(td); 1323 _lwkt_schedule(ntd); 1324 crit_enter_quick(td); 1325 } else { 1326 _lwkt_schedule(ntd); 1327 } 1328 } 1329 1330 /* 1331 * Thread migration using a 'Pull' method. The thread may or may not be 1332 * the current thread. It MUST be descheduled and in a stable state. 1333 * lwkt_giveaway() must be called on the cpu owning the thread. 1334 * 1335 * At any point after lwkt_giveaway() is called, the target cpu may 1336 * 'pull' the thread by calling lwkt_acquire(). 1337 * 1338 * We have to make sure the thread is not sitting on a per-cpu tsleep 1339 * queue or it will blow up when it moves to another cpu. 1340 * 1341 * MPSAFE - must be called under very specific conditions. 1342 */ 1343 void 1344 lwkt_giveaway(thread_t td) 1345 { 1346 globaldata_t gd = mycpu; 1347 1348 crit_enter_gd(gd); 1349 if (td->td_flags & TDF_TSLEEPQ) 1350 tsleep_remove(td); 1351 KKASSERT(td->td_gd == gd); 1352 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); 1353 td->td_flags |= TDF_MIGRATING; 1354 crit_exit_gd(gd); 1355 } 1356 1357 void 1358 lwkt_acquire(thread_t td) 1359 { 1360 globaldata_t gd; 1361 globaldata_t mygd; 1362 int retry = 10000000; 1363 1364 KKASSERT(td->td_flags & TDF_MIGRATING); 1365 gd = td->td_gd; 1366 mygd = mycpu; 1367 if (gd != mycpu) { 1368 cpu_lfence(); 1369 KKASSERT((td->td_flags & TDF_RUNQ) == 0); 1370 crit_enter_gd(mygd); 1371 DEBUG_PUSH_INFO("lwkt_acquire"); 1372 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) { 1373 lwkt_process_ipiq(); 1374 cpu_lfence(); 1375 if (--retry == 0) { 1376 kprintf("lwkt_acquire: stuck: td %p td->td_flags %08x\n", 1377 td, td->td_flags); 1378 retry = 10000000; 1379 } 1380 } 1381 DEBUG_POP_INFO(); 1382 cpu_mfence(); 1383 td->td_gd = mygd; 1384 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1385 td->td_flags &= ~TDF_MIGRATING; 1386 crit_exit_gd(mygd); 1387 } else { 1388 crit_enter_gd(mygd); 1389 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1390 td->td_flags &= ~TDF_MIGRATING; 1391 crit_exit_gd(mygd); 1392 } 1393 } 1394 1395 /* 1396 * Generic deschedule. Descheduling threads other then your own should be 1397 * done only in carefully controlled circumstances. Descheduling is 1398 * asynchronous. 1399 * 1400 * This function may block if the cpu has run out of messages. 1401 */ 1402 void 1403 lwkt_deschedule(thread_t td) 1404 { 1405 crit_enter(); 1406 if (td == curthread) { 1407 _lwkt_dequeue(td); 1408 } else { 1409 if (td->td_gd == mycpu) { 1410 _lwkt_dequeue(td); 1411 } else { 1412 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td); 1413 } 1414 } 1415 crit_exit(); 1416 } 1417 1418 /* 1419 * Set the target thread's priority. This routine does not automatically 1420 * switch to a higher priority thread, LWKT threads are not designed for 1421 * continuous priority changes. Yield if you want to switch. 1422 */ 1423 void 1424 lwkt_setpri(thread_t td, int pri) 1425 { 1426 if (td->td_pri != pri) { 1427 KKASSERT(pri >= 0); 1428 crit_enter(); 1429 if (td->td_flags & TDF_RUNQ) { 1430 KKASSERT(td->td_gd == mycpu); 1431 _lwkt_dequeue(td); 1432 td->td_pri = pri; 1433 _lwkt_enqueue(td); 1434 } else { 1435 td->td_pri = pri; 1436 } 1437 crit_exit(); 1438 } 1439 } 1440 1441 /* 1442 * Set the initial priority for a thread prior to it being scheduled for 1443 * the first time. The thread MUST NOT be scheduled before or during 1444 * this call. The thread may be assigned to a cpu other then the current 1445 * cpu. 1446 * 1447 * Typically used after a thread has been created with TDF_STOPPREQ, 1448 * and before the thread is initially scheduled. 1449 */ 1450 void 1451 lwkt_setpri_initial(thread_t td, int pri) 1452 { 1453 KKASSERT(pri >= 0); 1454 KKASSERT((td->td_flags & TDF_RUNQ) == 0); 1455 td->td_pri = pri; 1456 } 1457 1458 void 1459 lwkt_setpri_self(int pri) 1460 { 1461 thread_t td = curthread; 1462 1463 KKASSERT(pri >= 0 && pri <= TDPRI_MAX); 1464 crit_enter(); 1465 if (td->td_flags & TDF_RUNQ) { 1466 _lwkt_dequeue(td); 1467 td->td_pri = pri; 1468 _lwkt_enqueue(td); 1469 } else { 1470 td->td_pri = pri; 1471 } 1472 crit_exit(); 1473 } 1474 1475 /* 1476 * hz tick scheduler clock for LWKT threads 1477 */ 1478 void 1479 lwkt_schedulerclock(thread_t td) 1480 { 1481 globaldata_t gd = td->td_gd; 1482 thread_t xtd; 1483 1484 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) { 1485 /* 1486 * If the current thread is at the head of the runq shift it to the 1487 * end of any equal-priority threads and request a LWKT reschedule 1488 * if it moved. 1489 * 1490 * Ignore upri in this situation. There will only be one user thread 1491 * in user mode, all others will be user threads running in kernel 1492 * mode and we have to make sure they get some cpu. 1493 */ 1494 xtd = TAILQ_NEXT(td, td_threadq); 1495 if (xtd && xtd->td_pri == td->td_pri) { 1496 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq); 1497 while (xtd && xtd->td_pri == td->td_pri) 1498 xtd = TAILQ_NEXT(xtd, td_threadq); 1499 if (xtd) 1500 TAILQ_INSERT_BEFORE(xtd, td, td_threadq); 1501 else 1502 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 1503 need_lwkt_resched(); 1504 } 1505 } else { 1506 /* 1507 * If we scheduled a thread other than the one at the head of the 1508 * queue always request a reschedule every tick. 1509 */ 1510 need_lwkt_resched(); 1511 } 1512 } 1513 1514 /* 1515 * Migrate the current thread to the specified cpu. 1516 * 1517 * This is accomplished by descheduling ourselves from the current cpu 1518 * and setting td_migrate_gd. The lwkt_switch() code will detect that the 1519 * 'old' thread wants to migrate after it has been completely switched out 1520 * and will complete the migration. 1521 * 1522 * TDF_MIGRATING prevents scheduling races while the thread is being migrated. 1523 * 1524 * We must be sure to release our current process designation (if a user 1525 * process) before clearing out any tsleepq we are on because the release 1526 * code may re-add us. 1527 * 1528 * We must be sure to remove ourselves from the current cpu's tsleepq 1529 * before potentially moving to another queue. The thread can be on 1530 * a tsleepq due to a left-over tsleep_interlock(). 1531 */ 1532 1533 void 1534 lwkt_setcpu_self(globaldata_t rgd) 1535 { 1536 thread_t td = curthread; 1537 1538 if (td->td_gd != rgd) { 1539 crit_enter_quick(td); 1540 1541 if (td->td_release) 1542 td->td_release(td); 1543 if (td->td_flags & TDF_TSLEEPQ) 1544 tsleep_remove(td); 1545 1546 /* 1547 * Set TDF_MIGRATING to prevent a spurious reschedule while we are 1548 * trying to deschedule ourselves and switch away, then deschedule 1549 * ourself, remove us from tdallq, and set td_migrate_gd. Finally, 1550 * call lwkt_switch() to complete the operation. 1551 */ 1552 td->td_flags |= TDF_MIGRATING; 1553 lwkt_deschedule_self(td); 1554 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1555 td->td_migrate_gd = rgd; 1556 lwkt_switch(); 1557 1558 /* 1559 * We are now on the target cpu 1560 */ 1561 KKASSERT(rgd == mycpu); 1562 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); 1563 crit_exit_quick(td); 1564 } 1565 } 1566 1567 void 1568 lwkt_migratecpu(int cpuid) 1569 { 1570 globaldata_t rgd; 1571 1572 rgd = globaldata_find(cpuid); 1573 lwkt_setcpu_self(rgd); 1574 } 1575 1576 /* 1577 * Remote IPI for cpu migration (called while in a critical section so we 1578 * do not have to enter another one). 1579 * 1580 * The thread (td) has already been completely descheduled from the 1581 * originating cpu and we can simply assert the case. The thread is 1582 * assigned to the new cpu and enqueued. 1583 * 1584 * The thread will re-add itself to tdallq when it resumes execution. 1585 */ 1586 static void 1587 lwkt_setcpu_remote(void *arg) 1588 { 1589 thread_t td = arg; 1590 globaldata_t gd = mycpu; 1591 1592 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0); 1593 td->td_gd = gd; 1594 cpu_mfence(); 1595 td->td_flags &= ~TDF_MIGRATING; 1596 KKASSERT(td->td_migrate_gd == NULL); 1597 KKASSERT(td->td_lwp == NULL || 1598 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 1599 _lwkt_enqueue(td); 1600 } 1601 1602 struct lwp * 1603 lwkt_preempted_proc(void) 1604 { 1605 thread_t td = curthread; 1606 while (td->td_preempted) 1607 td = td->td_preempted; 1608 return(td->td_lwp); 1609 } 1610 1611 /* 1612 * Create a kernel process/thread/whatever. It shares it's address space 1613 * with proc0 - ie: kernel only. 1614 * 1615 * If the cpu is not specified one will be selected. In the future 1616 * specifying a cpu of -1 will enable kernel thread migration between 1617 * cpus. 1618 */ 1619 int 1620 lwkt_create(void (*func)(void *), void *arg, struct thread **tdp, 1621 thread_t template, int tdflags, int cpu, const char *fmt, ...) 1622 { 1623 thread_t td; 1624 __va_list ap; 1625 1626 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu, 1627 tdflags); 1628 if (tdp) 1629 *tdp = td; 1630 cpu_set_thread_handler(td, lwkt_exit, func, arg); 1631 1632 /* 1633 * Set up arg0 for 'ps' etc 1634 */ 1635 __va_start(ap, fmt); 1636 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap); 1637 __va_end(ap); 1638 1639 /* 1640 * Schedule the thread to run 1641 */ 1642 if (td->td_flags & TDF_NOSTART) 1643 td->td_flags &= ~TDF_NOSTART; 1644 else 1645 lwkt_schedule(td); 1646 return 0; 1647 } 1648 1649 /* 1650 * Destroy an LWKT thread. Warning! This function is not called when 1651 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and 1652 * uses a different reaping mechanism. 1653 */ 1654 void 1655 lwkt_exit(void) 1656 { 1657 thread_t td = curthread; 1658 thread_t std; 1659 globaldata_t gd; 1660 1661 /* 1662 * Do any cleanup that might block here 1663 */ 1664 if (td->td_flags & TDF_VERBOSE) 1665 kprintf("kthread %p %s has exited\n", td, td->td_comm); 1666 biosched_done(td); 1667 dsched_exit_thread(td); 1668 1669 /* 1670 * Get us into a critical section to interlock gd_freetd and loop 1671 * until we can get it freed. 1672 * 1673 * We have to cache the current td in gd_freetd because objcache_put()ing 1674 * it would rip it out from under us while our thread is still active. 1675 * 1676 * We are the current thread so of course our own TDF_RUNNING bit will 1677 * be set, so unlike the lwp reap code we don't wait for it to clear. 1678 */ 1679 gd = mycpu; 1680 crit_enter_quick(td); 1681 for (;;) { 1682 if (td->td_refs) { 1683 tsleep(td, 0, "tdreap", 1); 1684 continue; 1685 } 1686 if ((std = gd->gd_freetd) != NULL) { 1687 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0); 1688 gd->gd_freetd = NULL; 1689 objcache_put(thread_cache, std); 1690 continue; 1691 } 1692 break; 1693 } 1694 1695 /* 1696 * Remove thread resources from kernel lists and deschedule us for 1697 * the last time. We cannot block after this point or we may end 1698 * up with a stale td on the tsleepq. 1699 * 1700 * None of this may block, the critical section is the only thing 1701 * protecting tdallq and the only thing preventing new lwkt_hold() 1702 * thread refs now. 1703 */ 1704 if (td->td_flags & TDF_TSLEEPQ) 1705 tsleep_remove(td); 1706 lwkt_deschedule_self(td); 1707 lwkt_remove_tdallq(td); 1708 KKASSERT(td->td_refs == 0); 1709 1710 /* 1711 * Final cleanup 1712 */ 1713 KKASSERT(gd->gd_freetd == NULL); 1714 if (td->td_flags & TDF_ALLOCATED_THREAD) 1715 gd->gd_freetd = td; 1716 cpu_thread_exit(); 1717 } 1718 1719 void 1720 lwkt_remove_tdallq(thread_t td) 1721 { 1722 KKASSERT(td->td_gd == mycpu); 1723 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1724 } 1725 1726 /* 1727 * Code reduction and branch prediction improvements. Call/return 1728 * overhead on modern cpus often degenerates into 0 cycles due to 1729 * the cpu's branch prediction hardware and return pc cache. We 1730 * can take advantage of this by not inlining medium-complexity 1731 * functions and we can also reduce the branch prediction impact 1732 * by collapsing perfectly predictable branches into a single 1733 * procedure instead of duplicating it. 1734 * 1735 * Is any of this noticeable? Probably not, so I'll take the 1736 * smaller code size. 1737 */ 1738 void 1739 crit_exit_wrapper(__DEBUG_CRIT_ARG__) 1740 { 1741 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__); 1742 } 1743 1744 void 1745 crit_panic(void) 1746 { 1747 thread_t td = curthread; 1748 int lcrit = td->td_critcount; 1749 1750 td->td_critcount = 0; 1751 panic("td_critcount is/would-go negative! %p %d", td, lcrit); 1752 /* NOT REACHED */ 1753 } 1754 1755 /* 1756 * Called from debugger/panic on cpus which have been stopped. We must still 1757 * process the IPIQ while stopped, even if we were stopped while in a critical 1758 * section (XXX). 1759 * 1760 * If we are dumping also try to process any pending interrupts. This may 1761 * or may not work depending on the state of the cpu at the point it was 1762 * stopped. 1763 */ 1764 void 1765 lwkt_smp_stopped(void) 1766 { 1767 globaldata_t gd = mycpu; 1768 1769 crit_enter_gd(gd); 1770 if (dumping) { 1771 lwkt_process_ipiq(); 1772 splz(); 1773 } else { 1774 lwkt_process_ipiq(); 1775 } 1776 crit_exit_gd(gd); 1777 } 1778