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 #ifdef _KERNEL_VIRTUAL 75 #include <pthread.h> 76 #endif 77 78 #if !defined(KTR_CTXSW) 79 #define KTR_CTXSW KTR_ALL 80 #endif 81 KTR_INFO_MASTER(ctxsw); 82 KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p", int cpu, struct thread *td); 83 KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p", int cpu, struct thread *td); 84 KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s", struct thread *td, char *comm); 85 KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", struct thread *td); 86 87 static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads"); 88 89 #ifdef INVARIANTS 90 static int panic_on_cscount = 0; 91 #endif 92 static __int64_t switch_count = 0; 93 static __int64_t preempt_hit = 0; 94 static __int64_t preempt_miss = 0; 95 static __int64_t preempt_weird = 0; 96 static int lwkt_use_spin_port; 97 static struct objcache *thread_cache; 98 int cpu_mwait_spin = 0; 99 100 static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame); 101 static void lwkt_setcpu_remote(void *arg); 102 103 extern void cpu_heavy_restore(void); 104 extern void cpu_lwkt_restore(void); 105 extern void cpu_kthread_restore(void); 106 extern void cpu_idle_restore(void); 107 108 /* 109 * We can make all thread ports use the spin backend instead of the thread 110 * backend. This should only be set to debug the spin backend. 111 */ 112 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port); 113 114 #ifdef INVARIANTS 115 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, 116 "Panic if attempting to switch lwkt's while mastering cpusync"); 117 #endif 118 SYSCTL_INT(_hw, OID_AUTO, cpu_mwait_spin, CTLFLAG_RW, &cpu_mwait_spin, 0, 119 "monitor/mwait target state"); 120 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, 121 "Number of switched threads"); 122 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, 123 "Successful preemption events"); 124 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, 125 "Failed preemption events"); 126 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, 127 "Number of preempted threads."); 128 static int fairq_enable = 0; 129 SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW, 130 &fairq_enable, 0, "Turn on fairq priority accumulators"); 131 static int fairq_bypass = -1; 132 SYSCTL_INT(_lwkt, OID_AUTO, fairq_bypass, CTLFLAG_RW, 133 &fairq_bypass, 0, "Allow fairq to bypass td on token failure"); 134 extern int lwkt_sched_debug; 135 int lwkt_sched_debug = 0; 136 SYSCTL_INT(_lwkt, OID_AUTO, sched_debug, CTLFLAG_RW, 137 &lwkt_sched_debug, 0, "Scheduler debug"); 138 static int lwkt_spin_loops = 10; 139 SYSCTL_INT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW, 140 &lwkt_spin_loops, 0, "Scheduler spin loops until sorted decon"); 141 static int lwkt_spin_reseq = 0; 142 SYSCTL_INT(_lwkt, OID_AUTO, spin_reseq, CTLFLAG_RW, 143 &lwkt_spin_reseq, 0, "Scheduler resequencer enable"); 144 static int lwkt_spin_monitor = 0; 145 SYSCTL_INT(_lwkt, OID_AUTO, spin_monitor, CTLFLAG_RW, 146 &lwkt_spin_monitor, 0, "Scheduler uses monitor/mwait"); 147 static int lwkt_spin_fatal = 0; /* disabled */ 148 SYSCTL_INT(_lwkt, OID_AUTO, spin_fatal, CTLFLAG_RW, 149 &lwkt_spin_fatal, 0, "LWKT scheduler spin loops till fatal panic"); 150 static int preempt_enable = 1; 151 SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW, 152 &preempt_enable, 0, "Enable preemption"); 153 static int lwkt_cache_threads = 0; 154 SYSCTL_INT(_lwkt, OID_AUTO, cache_threads, CTLFLAG_RD, 155 &lwkt_cache_threads, 0, "thread+kstack cache"); 156 157 #ifndef _KERNEL_VIRTUAL 158 static __cachealign int lwkt_cseq_rindex; 159 static __cachealign int lwkt_cseq_windex; 160 #endif 161 162 /* 163 * These helper procedures handle the runq, they can only be called from 164 * within a critical section. 165 * 166 * WARNING! Prior to SMP being brought up it is possible to enqueue and 167 * dequeue threads belonging to other cpus, so be sure to use td->td_gd 168 * instead of 'mycpu' when referencing the globaldata structure. Once 169 * SMP live enqueuing and dequeueing only occurs on the current cpu. 170 */ 171 static __inline 172 void 173 _lwkt_dequeue(thread_t td) 174 { 175 if (td->td_flags & TDF_RUNQ) { 176 struct globaldata *gd = td->td_gd; 177 178 td->td_flags &= ~TDF_RUNQ; 179 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq); 180 --gd->gd_tdrunqcount; 181 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL) 182 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING); 183 } 184 } 185 186 /* 187 * Priority enqueue. 188 * 189 * There are a limited number of lwkt threads runnable since user 190 * processes only schedule one at a time per cpu. However, there can 191 * be many user processes in kernel mode exiting from a tsleep() which 192 * become runnable. 193 * 194 * NOTE: lwkt_schedulerclock() will force a round-robin based on td_pri and 195 * will ignore user priority. This is to ensure that user threads in 196 * kernel mode get cpu at some point regardless of what the user 197 * scheduler thinks. 198 */ 199 static __inline 200 void 201 _lwkt_enqueue(thread_t td) 202 { 203 thread_t xtd; 204 205 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) { 206 struct globaldata *gd = td->td_gd; 207 208 td->td_flags |= TDF_RUNQ; 209 xtd = TAILQ_FIRST(&gd->gd_tdrunq); 210 if (xtd == NULL) { 211 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 212 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING); 213 } else { 214 /* 215 * NOTE: td_upri - higher numbers more desireable, same sense 216 * as td_pri (typically reversed from lwp_upri). 217 * 218 * In the equal priority case we want the best selection 219 * at the beginning so the less desireable selections know 220 * that they have to setrunqueue/go-to-another-cpu, even 221 * though it means switching back to the 'best' selection. 222 * This also avoids degenerate situations when many threads 223 * are runnable or waking up at the same time. 224 * 225 * If upri matches exactly place at end/round-robin. 226 */ 227 while (xtd && 228 (xtd->td_pri >= td->td_pri || 229 (xtd->td_pri == td->td_pri && 230 xtd->td_upri >= td->td_upri))) { 231 xtd = TAILQ_NEXT(xtd, td_threadq); 232 } 233 if (xtd) 234 TAILQ_INSERT_BEFORE(xtd, td, td_threadq); 235 else 236 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 237 } 238 ++gd->gd_tdrunqcount; 239 240 /* 241 * Request a LWKT reschedule if we are now at the head of the queue. 242 */ 243 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) 244 need_lwkt_resched(); 245 } 246 } 247 248 static __boolean_t 249 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags) 250 { 251 struct thread *td = (struct thread *)obj; 252 253 td->td_kstack = NULL; 254 td->td_kstack_size = 0; 255 td->td_flags = TDF_ALLOCATED_THREAD; 256 td->td_mpflags = 0; 257 return (1); 258 } 259 260 static void 261 _lwkt_thread_dtor(void *obj, void *privdata) 262 { 263 struct thread *td = (struct thread *)obj; 264 265 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD, 266 ("_lwkt_thread_dtor: not allocated from objcache")); 267 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack && 268 td->td_kstack_size > 0, 269 ("_lwkt_thread_dtor: corrupted stack")); 270 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size); 271 td->td_kstack = NULL; 272 td->td_flags = 0; 273 } 274 275 /* 276 * Initialize the lwkt s/system. 277 * 278 * Nominally cache up to 32 thread + kstack structures. Cache more on 279 * systems with a lot of cpu cores. 280 */ 281 void 282 lwkt_init(void) 283 { 284 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads); 285 if (lwkt_cache_threads == 0) { 286 lwkt_cache_threads = ncpus * 4; 287 if (lwkt_cache_threads < 32) 288 lwkt_cache_threads = 32; 289 } 290 thread_cache = objcache_create_mbacked( 291 M_THREAD, sizeof(struct thread), 292 0, lwkt_cache_threads, 293 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL); 294 } 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 } 340 341 /* 342 * Create a new thread. The thread must be associated with a process context 343 * or LWKT start address before it can be scheduled. If the target cpu is 344 * -1 the thread will be created on the current cpu. 345 * 346 * If you intend to create a thread without a process context this function 347 * does everything except load the startup and switcher function. 348 */ 349 thread_t 350 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags) 351 { 352 static int cpu_rotator; 353 globaldata_t gd = mycpu; 354 void *stack; 355 356 /* 357 * If static thread storage is not supplied allocate a thread. Reuse 358 * a cached free thread if possible. gd_freetd is used to keep an exiting 359 * thread intact through the exit. 360 */ 361 if (td == NULL) { 362 crit_enter_gd(gd); 363 if ((td = gd->gd_freetd) != NULL) { 364 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK| 365 TDF_RUNQ)) == 0); 366 gd->gd_freetd = NULL; 367 } else { 368 td = objcache_get(thread_cache, M_WAITOK); 369 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK| 370 TDF_RUNQ)) == 0); 371 } 372 crit_exit_gd(gd); 373 KASSERT((td->td_flags & 374 (TDF_ALLOCATED_THREAD|TDF_RUNNING|TDF_PREEMPT_LOCK)) == 375 TDF_ALLOCATED_THREAD, 376 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags)); 377 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK); 378 } 379 380 /* 381 * Try to reuse cached stack. 382 */ 383 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) { 384 if (flags & TDF_ALLOCATED_STACK) { 385 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size); 386 stack = NULL; 387 } 388 } 389 if (stack == NULL) { 390 stack = (void *)kmem_alloc_stack(&kernel_map, stksize); 391 flags |= TDF_ALLOCATED_STACK; 392 } 393 if (cpu < 0) { 394 cpu = ++cpu_rotator; 395 cpu_ccfence(); 396 cpu %= ncpus; 397 } 398 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu)); 399 return(td); 400 } 401 402 /* 403 * Initialize a preexisting thread structure. This function is used by 404 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread. 405 * 406 * All threads start out in a critical section at a priority of 407 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as 408 * appropriate. This function may send an IPI message when the 409 * requested cpu is not the current cpu and consequently gd_tdallq may 410 * not be initialized synchronously from the point of view of the originating 411 * cpu. 412 * 413 * NOTE! we have to be careful in regards to creating threads for other cpus 414 * if SMP has not yet been activated. 415 */ 416 static void 417 lwkt_init_thread_remote(void *arg) 418 { 419 thread_t td = arg; 420 421 /* 422 * Protected by critical section held by IPI dispatch 423 */ 424 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq); 425 } 426 427 /* 428 * lwkt core thread structural initialization. 429 * 430 * NOTE: All threads are initialized as mpsafe threads. 431 */ 432 void 433 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags, 434 struct globaldata *gd) 435 { 436 globaldata_t mygd = mycpu; 437 438 bzero(td, sizeof(struct thread)); 439 td->td_kstack = stack; 440 td->td_kstack_size = stksize; 441 td->td_flags = flags; 442 td->td_mpflags = 0; 443 td->td_type = TD_TYPE_GENERIC; 444 td->td_gd = gd; 445 td->td_pri = TDPRI_KERN_DAEMON; 446 td->td_critcount = 1; 447 td->td_toks_have = NULL; 448 td->td_toks_stop = &td->td_toks_base; 449 if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT)) 450 lwkt_initport_spin(&td->td_msgport, td); 451 else 452 lwkt_initport_thread(&td->td_msgport, td); 453 pmap_init_thread(td); 454 /* 455 * Normally initializing a thread for a remote cpu requires sending an 456 * IPI. However, the idlethread is setup before the other cpus are 457 * activated so we have to treat it as a special case. XXX manipulation 458 * of gd_tdallq requires the BGL. 459 */ 460 if (gd == mygd || td == &gd->gd_idlethread) { 461 crit_enter_gd(mygd); 462 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); 463 crit_exit_gd(mygd); 464 } else { 465 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td); 466 } 467 dsched_new_thread(td); 468 } 469 470 void 471 lwkt_set_comm(thread_t td, const char *ctl, ...) 472 { 473 __va_list va; 474 475 __va_start(va, ctl); 476 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va); 477 __va_end(va); 478 KTR_LOG(ctxsw_newtd, td, td->td_comm); 479 } 480 481 /* 482 * Prevent the thread from getting destroyed. Note that unlike PHOLD/PRELE 483 * this does not prevent the thread from migrating to another cpu so the 484 * gd_tdallq state is not protected by this. 485 */ 486 void 487 lwkt_hold(thread_t td) 488 { 489 atomic_add_int(&td->td_refs, 1); 490 } 491 492 void 493 lwkt_rele(thread_t td) 494 { 495 KKASSERT(td->td_refs > 0); 496 atomic_add_int(&td->td_refs, -1); 497 } 498 499 void 500 lwkt_free_thread(thread_t td) 501 { 502 KKASSERT(td->td_refs == 0); 503 KKASSERT((td->td_flags & (TDF_RUNNING | TDF_PREEMPT_LOCK | 504 TDF_RUNQ | TDF_TSLEEPQ)) == 0); 505 if (td->td_flags & TDF_ALLOCATED_THREAD) { 506 objcache_put(thread_cache, td); 507 } else if (td->td_flags & TDF_ALLOCATED_STACK) { 508 /* client-allocated struct with internally allocated stack */ 509 KASSERT(td->td_kstack && td->td_kstack_size > 0, 510 ("lwkt_free_thread: corrupted stack")); 511 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size); 512 td->td_kstack = NULL; 513 td->td_kstack_size = 0; 514 } 515 516 KTR_LOG(ctxsw_deadtd, td); 517 } 518 519 520 /* 521 * Switch to the next runnable lwkt. If no LWKTs are runnable then 522 * switch to the idlethread. Switching must occur within a critical 523 * section to avoid races with the scheduling queue. 524 * 525 * We always have full control over our cpu's run queue. Other cpus 526 * that wish to manipulate our queue must use the cpu_*msg() calls to 527 * talk to our cpu, so a critical section is all that is needed and 528 * the result is very, very fast thread switching. 529 * 530 * The LWKT scheduler uses a fixed priority model and round-robins at 531 * each priority level. User process scheduling is a totally 532 * different beast and LWKT priorities should not be confused with 533 * user process priorities. 534 * 535 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch() 536 * is not called by the current thread in the preemption case, only when 537 * the preempting thread blocks (in order to return to the original thread). 538 * 539 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread 540 * migration and tsleep deschedule the current lwkt thread and call 541 * lwkt_switch(). In particular, the target cpu of the migration fully 542 * expects the thread to become non-runnable and can deadlock against 543 * cpusync operations if we run any IPIs prior to switching the thread out. 544 * 545 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF 546 * THE CURRENT THREAD HAS BEEN DESCHEDULED! 547 */ 548 void 549 lwkt_switch(void) 550 { 551 globaldata_t gd = mycpu; 552 thread_t td = gd->gd_curthread; 553 thread_t ntd; 554 int spinning = 0; 555 556 KKASSERT(gd->gd_processing_ipiq == 0); 557 KKASSERT(td->td_flags & TDF_RUNNING); 558 559 /* 560 * Switching from within a 'fast' (non thread switched) interrupt or IPI 561 * is illegal. However, we may have to do it anyway if we hit a fatal 562 * kernel trap or we have paniced. 563 * 564 * If this case occurs save and restore the interrupt nesting level. 565 */ 566 if (gd->gd_intr_nesting_level) { 567 int savegdnest; 568 int savegdtrap; 569 570 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) { 571 panic("lwkt_switch: Attempt to switch from a " 572 "fast interrupt, ipi, or hard code section, " 573 "td %p\n", 574 td); 575 } else { 576 savegdnest = gd->gd_intr_nesting_level; 577 savegdtrap = gd->gd_trap_nesting_level; 578 gd->gd_intr_nesting_level = 0; 579 gd->gd_trap_nesting_level = 0; 580 if ((td->td_flags & TDF_PANICWARN) == 0) { 581 td->td_flags |= TDF_PANICWARN; 582 kprintf("Warning: thread switch from interrupt, IPI, " 583 "or hard code section.\n" 584 "thread %p (%s)\n", td, td->td_comm); 585 print_backtrace(-1); 586 } 587 lwkt_switch(); 588 gd->gd_intr_nesting_level = savegdnest; 589 gd->gd_trap_nesting_level = savegdtrap; 590 return; 591 } 592 } 593 594 /* 595 * Release our current user process designation if we are blocking 596 * or if a user reschedule was requested. 597 * 598 * NOTE: This function is NOT called if we are switching into or 599 * returning from a preemption. 600 * 601 * NOTE: Releasing our current user process designation may cause 602 * it to be assigned to another thread, which in turn will 603 * cause us to block in the usched acquire code when we attempt 604 * to return to userland. 605 * 606 * NOTE: On SMP systems this can be very nasty when heavy token 607 * contention is present so we want to be careful not to 608 * release the designation gratuitously. 609 */ 610 if (td->td_release && 611 (user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) { 612 td->td_release(td); 613 } 614 615 /* 616 * Release all tokens 617 */ 618 crit_enter_gd(gd); 619 if (TD_TOKS_HELD(td)) 620 lwkt_relalltokens(td); 621 622 /* 623 * We had better not be holding any spin locks, but don't get into an 624 * endless panic loop. 625 */ 626 KASSERT(gd->gd_spinlocks == 0 || panicstr != NULL, 627 ("lwkt_switch: still holding %d exclusive spinlocks!", 628 gd->gd_spinlocks)); 629 630 631 #ifdef INVARIANTS 632 if (td->td_cscount) { 633 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n", 634 td); 635 if (panic_on_cscount) 636 panic("switching while mastering cpusync"); 637 } 638 #endif 639 640 /* 641 * If we had preempted another thread on this cpu, resume the preempted 642 * thread. This occurs transparently, whether the preempted thread 643 * was scheduled or not (it may have been preempted after descheduling 644 * itself). 645 * 646 * We have to setup the MP lock for the original thread after backing 647 * out the adjustment that was made to curthread when the original 648 * was preempted. 649 */ 650 if ((ntd = td->td_preempted) != NULL) { 651 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK); 652 ntd->td_flags |= TDF_PREEMPT_DONE; 653 654 /* 655 * The interrupt may have woken a thread up, we need to properly 656 * set the reschedule flag if the originally interrupted thread is 657 * at a lower priority. 658 * 659 * The interrupt may not have descheduled. 660 */ 661 if (TAILQ_FIRST(&gd->gd_tdrunq) != ntd) 662 need_lwkt_resched(); 663 goto havethread_preempted; 664 } 665 666 /* 667 * If we cannot obtain ownership of the tokens we cannot immediately 668 * schedule the target thread. 669 * 670 * Reminder: Again, we cannot afford to run any IPIs in this path if 671 * the current thread has been descheduled. 672 */ 673 for (;;) { 674 clear_lwkt_resched(); 675 676 /* 677 * Hotpath - pull the head of the run queue and attempt to schedule 678 * it. 679 */ 680 ntd = TAILQ_FIRST(&gd->gd_tdrunq); 681 682 if (ntd == NULL) { 683 /* 684 * Runq is empty, switch to idle to allow it to halt. 685 */ 686 ntd = &gd->gd_idlethread; 687 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) 688 ASSERT_NO_TOKENS_HELD(ntd); 689 cpu_time.cp_msg[0] = 0; 690 cpu_time.cp_stallpc = 0; 691 goto haveidle; 692 } 693 694 /* 695 * Hotpath - schedule ntd. 696 * 697 * NOTE: For UP there is no mplock and lwkt_getalltokens() 698 * always succeeds. 699 */ 700 if (TD_TOKS_NOT_HELD(ntd) || 701 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops))) 702 { 703 goto havethread; 704 } 705 706 /* 707 * Coldpath (SMP only since tokens always succeed on UP) 708 * 709 * We had some contention on the thread we wanted to schedule. 710 * What we do now is try to find a thread that we can schedule 711 * in its stead. 712 * 713 * The coldpath scan does NOT rearrange threads in the run list. 714 * The lwkt_schedulerclock() will assert need_lwkt_resched() on 715 * the next tick whenever the current head is not the current thread. 716 */ 717 #ifdef INVARIANTS 718 ++ntd->td_contended; 719 #endif 720 ++gd->gd_cnt.v_token_colls; 721 722 if (fairq_bypass > 0) 723 goto skip; 724 725 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) { 726 #ifndef NO_LWKT_SPLIT_USERPRI 727 /* 728 * Never schedule threads returning to userland or the 729 * user thread scheduler helper thread when higher priority 730 * threads are present. The runq is sorted by priority 731 * so we can give up traversing it when we find the first 732 * low priority thread. 733 */ 734 if (ntd->td_pri < TDPRI_KERN_LPSCHED) { 735 ntd = NULL; 736 break; 737 } 738 #endif 739 740 /* 741 * Try this one. 742 */ 743 if (TD_TOKS_NOT_HELD(ntd) || 744 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops))) { 745 goto havethread; 746 } 747 #ifdef INVARIANTS 748 ++ntd->td_contended; 749 #endif 750 ++gd->gd_cnt.v_token_colls; 751 } 752 753 skip: 754 /* 755 * We exhausted the run list, meaning that all runnable threads 756 * are contested. 757 */ 758 cpu_pause(); 759 #ifdef _KERNEL_VIRTUAL 760 pthread_yield(); 761 #endif 762 ntd = &gd->gd_idlethread; 763 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) 764 ASSERT_NO_TOKENS_HELD(ntd); 765 /* contention case, do not clear contention mask */ 766 767 /* 768 * We are going to have to retry but if the current thread is not 769 * on the runq we instead switch through the idle thread to get away 770 * from the current thread. We have to flag for lwkt reschedule 771 * to prevent the idle thread from halting. 772 * 773 * NOTE: A non-zero spinning is passed to lwkt_getalltokens() to 774 * instruct it to deal with the potential for deadlocks by 775 * ordering the tokens by address. 776 */ 777 if ((td->td_flags & TDF_RUNQ) == 0) { 778 need_lwkt_resched(); /* prevent hlt */ 779 goto haveidle; 780 } 781 #if defined(INVARIANTS) && defined(__x86_64__) 782 if ((read_rflags() & PSL_I) == 0) { 783 cpu_enable_intr(); 784 panic("lwkt_switch() called with interrupts disabled"); 785 } 786 #endif 787 788 /* 789 * Number iterations so far. After a certain point we switch to 790 * a sorted-address/monitor/mwait version of lwkt_getalltokens() 791 */ 792 if (spinning < 0x7FFFFFFF) 793 ++spinning; 794 795 #ifndef _KERNEL_VIRTUAL 796 /* 797 * lwkt_getalltokens() failed in sorted token mode, we can use 798 * monitor/mwait in this case. 799 */ 800 if (spinning >= lwkt_spin_loops && 801 (cpu_mi_feature & CPU_MI_MONITOR) && 802 lwkt_spin_monitor) 803 { 804 cpu_mmw_pause_int(&gd->gd_reqflags, 805 (gd->gd_reqflags | RQF_SPINNING) & 806 ~RQF_IDLECHECK_WK_MASK, 807 cpu_mwait_spin); 808 } 809 #endif 810 811 /* 812 * We already checked that td is still scheduled so this should be 813 * safe. 814 */ 815 splz_check(); 816 817 #ifndef _KERNEL_VIRTUAL 818 /* 819 * This experimental resequencer is used as a fall-back to reduce 820 * hw cache line contention by placing each core's scheduler into a 821 * time-domain-multplexed slot. 822 * 823 * The resequencer is disabled by default. It's functionality has 824 * largely been superceeded by the token algorithm which limits races 825 * to a subset of cores. 826 * 827 * The resequencer algorithm tends to break down when more than 828 * 20 cores are contending. What appears to happen is that new 829 * tokens can be obtained out of address-sorted order by new cores 830 * while existing cores languish in long delays between retries and 831 * wind up being starved-out of the token acquisition. 832 */ 833 if (lwkt_spin_reseq && spinning >= lwkt_spin_reseq) { 834 int cseq = atomic_fetchadd_int(&lwkt_cseq_windex, 1); 835 int oseq; 836 837 while ((oseq = lwkt_cseq_rindex) != cseq) { 838 cpu_ccfence(); 839 #if 1 840 if (cpu_mi_feature & CPU_MI_MONITOR) { 841 cpu_mmw_pause_int(&lwkt_cseq_rindex, oseq, cpu_mwait_spin); 842 } else { 843 #endif 844 cpu_pause(); 845 cpu_lfence(); 846 #if 1 847 } 848 #endif 849 } 850 DELAY(1); 851 atomic_add_int(&lwkt_cseq_rindex, 1); 852 } 853 #endif 854 /* highest level for(;;) loop */ 855 } 856 857 havethread: 858 /* 859 * Clear gd_idle_repeat when doing a normal switch to a non-idle 860 * thread. 861 */ 862 ntd->td_wmesg = NULL; 863 ++gd->gd_cnt.v_swtch; 864 gd->gd_idle_repeat = 0; 865 866 havethread_preempted: 867 /* 868 * If the new target does not need the MP lock and we are holding it, 869 * release the MP lock. If the new target requires the MP lock we have 870 * already acquired it for the target. 871 */ 872 ; 873 haveidle: 874 KASSERT(ntd->td_critcount, 875 ("priority problem in lwkt_switch %d %d", 876 td->td_critcount, ntd->td_critcount)); 877 878 if (td != ntd) { 879 /* 880 * Execute the actual thread switch operation. This function 881 * returns to the current thread and returns the previous thread 882 * (which may be different from the thread we switched to). 883 * 884 * We are responsible for marking ntd as TDF_RUNNING. 885 */ 886 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0); 887 ++switch_count; 888 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd); 889 ntd->td_flags |= TDF_RUNNING; 890 lwkt_switch_return(td->td_switch(ntd)); 891 /* ntd invalid, td_switch() can return a different thread_t */ 892 } 893 894 /* 895 * catch-all. XXX is this strictly needed? 896 */ 897 splz_check(); 898 899 /* NOTE: current cpu may have changed after switch */ 900 crit_exit_quick(td); 901 } 902 903 /* 904 * Called by assembly in the td_switch (thread restore path) for thread 905 * bootstrap cases which do not 'return' to lwkt_switch(). 906 */ 907 void 908 lwkt_switch_return(thread_t otd) 909 { 910 globaldata_t rgd; 911 912 /* 913 * Check if otd was migrating. Now that we are on ntd we can finish 914 * up the migration. This is a bit messy but it is the only place 915 * where td is known to be fully descheduled. 916 * 917 * We can only activate the migration if otd was migrating but not 918 * held on the cpu due to a preemption chain. We still have to 919 * clear TDF_RUNNING on the old thread either way. 920 * 921 * We are responsible for clearing the previously running thread's 922 * TDF_RUNNING. 923 */ 924 if ((rgd = otd->td_migrate_gd) != NULL && 925 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) { 926 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) == 927 (TDF_MIGRATING | TDF_RUNNING)); 928 otd->td_migrate_gd = NULL; 929 otd->td_flags &= ~TDF_RUNNING; 930 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd); 931 } else { 932 otd->td_flags &= ~TDF_RUNNING; 933 } 934 935 /* 936 * Final exit validations (see lwp_wait()). Note that otd becomes 937 * invalid the *instant* we set TDF_MP_EXITSIG. 938 */ 939 while (otd->td_flags & TDF_EXITING) { 940 u_int mpflags; 941 942 mpflags = otd->td_mpflags; 943 cpu_ccfence(); 944 945 if (mpflags & TDF_MP_EXITWAIT) { 946 if (atomic_cmpset_int(&otd->td_mpflags, mpflags, 947 mpflags | TDF_MP_EXITSIG)) { 948 wakeup(otd); 949 break; 950 } 951 } else { 952 if (atomic_cmpset_int(&otd->td_mpflags, mpflags, 953 mpflags | TDF_MP_EXITSIG)) { 954 wakeup(otd); 955 break; 956 } 957 } 958 } 959 } 960 961 /* 962 * Request that the target thread preempt the current thread. Preemption 963 * can only occur if our only critical section is the one that we were called 964 * with, the relative priority of the target thread is higher, and the target 965 * thread holds no tokens. This also only works if we are not holding any 966 * spinlocks (obviously). 967 * 968 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically 969 * this is called via lwkt_schedule() through the td_preemptable callback. 970 * critcount is the managed critical priority that we should ignore in order 971 * to determine whether preemption is possible (aka usually just the crit 972 * priority of lwkt_schedule() itself). 973 * 974 * Preemption is typically limited to interrupt threads. 975 * 976 * Operation works in a fairly straight-forward manner. The normal 977 * scheduling code is bypassed and we switch directly to the target 978 * thread. When the target thread attempts to block or switch away 979 * code at the base of lwkt_switch() will switch directly back to our 980 * thread. Our thread is able to retain whatever tokens it holds and 981 * if the target needs one of them the target will switch back to us 982 * and reschedule itself normally. 983 */ 984 void 985 lwkt_preempt(thread_t ntd, int critcount) 986 { 987 struct globaldata *gd = mycpu; 988 thread_t xtd; 989 thread_t td; 990 int save_gd_intr_nesting_level; 991 992 /* 993 * The caller has put us in a critical section. We can only preempt 994 * if the caller of the caller was not in a critical section (basically 995 * a local interrupt), as determined by the 'critcount' parameter. We 996 * also can't preempt if the caller is holding any spinlocks (even if 997 * he isn't in a critical section). This also handles the tokens test. 998 * 999 * YYY The target thread must be in a critical section (else it must 1000 * inherit our critical section? I dunno yet). 1001 */ 1002 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri)); 1003 1004 td = gd->gd_curthread; 1005 if (preempt_enable == 0) { 1006 ++preempt_miss; 1007 return; 1008 } 1009 if (ntd->td_pri <= td->td_pri) { 1010 ++preempt_miss; 1011 return; 1012 } 1013 if (td->td_critcount > critcount) { 1014 ++preempt_miss; 1015 return; 1016 } 1017 if (td->td_cscount) { 1018 ++preempt_miss; 1019 return; 1020 } 1021 if (ntd->td_gd != gd) { 1022 ++preempt_miss; 1023 return; 1024 } 1025 /* 1026 * We don't have to check spinlocks here as they will also bump 1027 * td_critcount. 1028 * 1029 * Do not try to preempt if the target thread is holding any tokens. 1030 * We could try to acquire the tokens but this case is so rare there 1031 * is no need to support it. 1032 */ 1033 KKASSERT(gd->gd_spinlocks == 0); 1034 1035 if (TD_TOKS_HELD(ntd)) { 1036 ++preempt_miss; 1037 return; 1038 } 1039 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) { 1040 ++preempt_weird; 1041 return; 1042 } 1043 if (ntd->td_preempted) { 1044 ++preempt_hit; 1045 return; 1046 } 1047 KKASSERT(gd->gd_processing_ipiq == 0); 1048 1049 /* 1050 * Since we are able to preempt the current thread, there is no need to 1051 * call need_lwkt_resched(). 1052 * 1053 * We must temporarily clear gd_intr_nesting_level around the switch 1054 * since switchouts from the target thread are allowed (they will just 1055 * return to our thread), and since the target thread has its own stack. 1056 * 1057 * A preemption must switch back to the original thread, assert the 1058 * case. 1059 */ 1060 ++preempt_hit; 1061 ntd->td_preempted = td; 1062 td->td_flags |= TDF_PREEMPT_LOCK; 1063 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd); 1064 save_gd_intr_nesting_level = gd->gd_intr_nesting_level; 1065 gd->gd_intr_nesting_level = 0; 1066 1067 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0); 1068 ntd->td_flags |= TDF_RUNNING; 1069 xtd = td->td_switch(ntd); 1070 KKASSERT(xtd == ntd); 1071 lwkt_switch_return(xtd); 1072 gd->gd_intr_nesting_level = save_gd_intr_nesting_level; 1073 1074 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE)); 1075 ntd->td_preempted = NULL; 1076 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE); 1077 } 1078 1079 /* 1080 * Conditionally call splz() if gd_reqflags indicates work is pending. 1081 * This will work inside a critical section but not inside a hard code 1082 * section. 1083 * 1084 * (self contained on a per cpu basis) 1085 */ 1086 void 1087 splz_check(void) 1088 { 1089 globaldata_t gd = mycpu; 1090 thread_t td = gd->gd_curthread; 1091 1092 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && 1093 gd->gd_intr_nesting_level == 0 && 1094 td->td_nest_count < 2) 1095 { 1096 splz(); 1097 } 1098 } 1099 1100 /* 1101 * This version is integrated into crit_exit, reqflags has already 1102 * been tested but td_critcount has not. 1103 * 1104 * We only want to execute the splz() on the 1->0 transition of 1105 * critcount and not in a hard code section or if too deeply nested. 1106 * 1107 * NOTE: gd->gd_spinlocks is implied to be 0 when td_critcount is 0. 1108 */ 1109 void 1110 lwkt_maybe_splz(thread_t td) 1111 { 1112 globaldata_t gd = td->td_gd; 1113 1114 if (td->td_critcount == 0 && 1115 gd->gd_intr_nesting_level == 0 && 1116 td->td_nest_count < 2) 1117 { 1118 splz(); 1119 } 1120 } 1121 1122 /* 1123 * Drivers which set up processing co-threads can call this function to 1124 * run the co-thread at a higher priority and to allow it to preempt 1125 * normal threads. 1126 */ 1127 void 1128 lwkt_set_interrupt_support_thread(void) 1129 { 1130 thread_t td = curthread; 1131 1132 lwkt_setpri_self(TDPRI_INT_SUPPORT); 1133 td->td_flags |= TDF_INTTHREAD; 1134 td->td_preemptable = lwkt_preempt; 1135 } 1136 1137 1138 /* 1139 * This function is used to negotiate a passive release of the current 1140 * process/lwp designation with the user scheduler, allowing the user 1141 * scheduler to schedule another user thread. The related kernel thread 1142 * (curthread) continues running in the released state. 1143 */ 1144 void 1145 lwkt_passive_release(struct thread *td) 1146 { 1147 struct lwp *lp = td->td_lwp; 1148 1149 #ifndef NO_LWKT_SPLIT_USERPRI 1150 td->td_release = NULL; 1151 lwkt_setpri_self(TDPRI_KERN_USER); 1152 #endif 1153 1154 lp->lwp_proc->p_usched->release_curproc(lp); 1155 } 1156 1157 1158 /* 1159 * This implements a LWKT yield, allowing a kernel thread to yield to other 1160 * kernel threads at the same or higher priority. This function can be 1161 * called in a tight loop and will typically only yield once per tick. 1162 * 1163 * Most kernel threads run at the same priority in order to allow equal 1164 * sharing. 1165 * 1166 * (self contained on a per cpu basis) 1167 */ 1168 void 1169 lwkt_yield(void) 1170 { 1171 globaldata_t gd = mycpu; 1172 thread_t td = gd->gd_curthread; 1173 1174 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1175 splz(); 1176 if (lwkt_resched_wanted()) { 1177 lwkt_schedule_self(curthread); 1178 lwkt_switch(); 1179 } 1180 } 1181 1182 /* 1183 * The quick version processes pending interrupts and higher-priority 1184 * LWKT threads but will not round-robin same-priority LWKT threads. 1185 * 1186 * When called while attempting to return to userland the only same-pri 1187 * threads are the ones which have already tried to become the current 1188 * user process. 1189 */ 1190 void 1191 lwkt_yield_quick(void) 1192 { 1193 globaldata_t gd = mycpu; 1194 thread_t td = gd->gd_curthread; 1195 1196 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1197 splz(); 1198 if (lwkt_resched_wanted()) { 1199 crit_enter(); 1200 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) { 1201 clear_lwkt_resched(); 1202 } else { 1203 lwkt_schedule_self(curthread); 1204 lwkt_switch(); 1205 } 1206 crit_exit(); 1207 } 1208 } 1209 1210 /* 1211 * This yield is designed for kernel threads with a user context. 1212 * 1213 * The kernel acting on behalf of the user is potentially cpu-bound, 1214 * this function will efficiently allow other threads to run and also 1215 * switch to other processes by releasing. 1216 * 1217 * The lwkt_user_yield() function is designed to have very low overhead 1218 * if no yield is determined to be needed. 1219 */ 1220 void 1221 lwkt_user_yield(void) 1222 { 1223 globaldata_t gd = mycpu; 1224 thread_t td = gd->gd_curthread; 1225 1226 /* 1227 * Always run any pending interrupts in case we are in a critical 1228 * section. 1229 */ 1230 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1231 splz(); 1232 1233 /* 1234 * Switch (which forces a release) if another kernel thread needs 1235 * the cpu, if userland wants us to resched, or if our kernel 1236 * quantum has run out. 1237 */ 1238 if (lwkt_resched_wanted() || 1239 user_resched_wanted()) 1240 { 1241 lwkt_switch(); 1242 } 1243 1244 #if 0 1245 /* 1246 * Reacquire the current process if we are released. 1247 * 1248 * XXX not implemented atm. The kernel may be holding locks and such, 1249 * so we want the thread to continue to receive cpu. 1250 */ 1251 if (td->td_release == NULL && lp) { 1252 lp->lwp_proc->p_usched->acquire_curproc(lp); 1253 td->td_release = lwkt_passive_release; 1254 lwkt_setpri_self(TDPRI_USER_NORM); 1255 } 1256 #endif 1257 } 1258 1259 /* 1260 * Generic schedule. Possibly schedule threads belonging to other cpus and 1261 * deal with threads that might be blocked on a wait queue. 1262 * 1263 * We have a little helper inline function which does additional work after 1264 * the thread has been enqueued, including dealing with preemption and 1265 * setting need_lwkt_resched() (which prevents the kernel from returning 1266 * to userland until it has processed higher priority threads). 1267 * 1268 * It is possible for this routine to be called after a failed _enqueue 1269 * (due to the target thread migrating, sleeping, or otherwise blocked). 1270 * We have to check that the thread is actually on the run queue! 1271 */ 1272 static __inline 1273 void 1274 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount) 1275 { 1276 if (ntd->td_flags & TDF_RUNQ) { 1277 if (ntd->td_preemptable) { 1278 ntd->td_preemptable(ntd, ccount); /* YYY +token */ 1279 } 1280 } 1281 } 1282 1283 static __inline 1284 void 1285 _lwkt_schedule(thread_t td) 1286 { 1287 globaldata_t mygd = mycpu; 1288 1289 KASSERT(td != &td->td_gd->gd_idlethread, 1290 ("lwkt_schedule(): scheduling gd_idlethread is illegal!")); 1291 KKASSERT((td->td_flags & TDF_MIGRATING) == 0); 1292 crit_enter_gd(mygd); 1293 KKASSERT(td->td_lwp == NULL || 1294 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 1295 1296 if (td == mygd->gd_curthread) { 1297 _lwkt_enqueue(td); 1298 } else { 1299 /* 1300 * If we own the thread, there is no race (since we are in a 1301 * critical section). If we do not own the thread there might 1302 * be a race but the target cpu will deal with it. 1303 */ 1304 if (td->td_gd == mygd) { 1305 _lwkt_enqueue(td); 1306 _lwkt_schedule_post(mygd, td, 1); 1307 } else { 1308 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0); 1309 } 1310 } 1311 crit_exit_gd(mygd); 1312 } 1313 1314 void 1315 lwkt_schedule(thread_t td) 1316 { 1317 _lwkt_schedule(td); 1318 } 1319 1320 void 1321 lwkt_schedule_noresched(thread_t td) /* XXX not impl */ 1322 { 1323 _lwkt_schedule(td); 1324 } 1325 1326 /* 1327 * When scheduled remotely if frame != NULL the IPIQ is being 1328 * run via doreti or an interrupt then preemption can be allowed. 1329 * 1330 * To allow preemption we have to drop the critical section so only 1331 * one is present in _lwkt_schedule_post. 1332 */ 1333 static void 1334 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame) 1335 { 1336 thread_t td = curthread; 1337 thread_t ntd = arg; 1338 1339 if (frame && ntd->td_preemptable) { 1340 crit_exit_noyield(td); 1341 _lwkt_schedule(ntd); 1342 crit_enter_quick(td); 1343 } else { 1344 _lwkt_schedule(ntd); 1345 } 1346 } 1347 1348 /* 1349 * Thread migration using a 'Pull' method. The thread may or may not be 1350 * the current thread. It MUST be descheduled and in a stable state. 1351 * lwkt_giveaway() must be called on the cpu owning the thread. 1352 * 1353 * At any point after lwkt_giveaway() is called, the target cpu may 1354 * 'pull' the thread by calling lwkt_acquire(). 1355 * 1356 * We have to make sure the thread is not sitting on a per-cpu tsleep 1357 * queue or it will blow up when it moves to another cpu. 1358 * 1359 * MPSAFE - must be called under very specific conditions. 1360 */ 1361 void 1362 lwkt_giveaway(thread_t td) 1363 { 1364 globaldata_t gd = mycpu; 1365 1366 crit_enter_gd(gd); 1367 if (td->td_flags & TDF_TSLEEPQ) 1368 tsleep_remove(td); 1369 KKASSERT(td->td_gd == gd); 1370 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); 1371 td->td_flags |= TDF_MIGRATING; 1372 crit_exit_gd(gd); 1373 } 1374 1375 void 1376 lwkt_acquire(thread_t td) 1377 { 1378 globaldata_t gd; 1379 globaldata_t mygd; 1380 int retry = 10000000; 1381 1382 KKASSERT(td->td_flags & TDF_MIGRATING); 1383 gd = td->td_gd; 1384 mygd = mycpu; 1385 if (gd != mycpu) { 1386 cpu_lfence(); 1387 KKASSERT((td->td_flags & TDF_RUNQ) == 0); 1388 crit_enter_gd(mygd); 1389 DEBUG_PUSH_INFO("lwkt_acquire"); 1390 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) { 1391 lwkt_process_ipiq(); 1392 cpu_lfence(); 1393 if (--retry == 0) { 1394 kprintf("lwkt_acquire: stuck: td %p td->td_flags %08x\n", 1395 td, td->td_flags); 1396 retry = 10000000; 1397 } 1398 #ifdef _KERNEL_VIRTUAL 1399 pthread_yield(); 1400 #endif 1401 } 1402 DEBUG_POP_INFO(); 1403 cpu_mfence(); 1404 td->td_gd = mygd; 1405 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1406 td->td_flags &= ~TDF_MIGRATING; 1407 crit_exit_gd(mygd); 1408 } else { 1409 crit_enter_gd(mygd); 1410 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1411 td->td_flags &= ~TDF_MIGRATING; 1412 crit_exit_gd(mygd); 1413 } 1414 } 1415 1416 /* 1417 * Generic deschedule. Descheduling threads other then your own should be 1418 * done only in carefully controlled circumstances. Descheduling is 1419 * asynchronous. 1420 * 1421 * This function may block if the cpu has run out of messages. 1422 */ 1423 void 1424 lwkt_deschedule(thread_t td) 1425 { 1426 crit_enter(); 1427 if (td == curthread) { 1428 _lwkt_dequeue(td); 1429 } else { 1430 if (td->td_gd == mycpu) { 1431 _lwkt_dequeue(td); 1432 } else { 1433 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td); 1434 } 1435 } 1436 crit_exit(); 1437 } 1438 1439 /* 1440 * Set the target thread's priority. This routine does not automatically 1441 * switch to a higher priority thread, LWKT threads are not designed for 1442 * continuous priority changes. Yield if you want to switch. 1443 */ 1444 void 1445 lwkt_setpri(thread_t td, int pri) 1446 { 1447 if (td->td_pri != pri) { 1448 KKASSERT(pri >= 0); 1449 crit_enter(); 1450 if (td->td_flags & TDF_RUNQ) { 1451 KKASSERT(td->td_gd == mycpu); 1452 _lwkt_dequeue(td); 1453 td->td_pri = pri; 1454 _lwkt_enqueue(td); 1455 } else { 1456 td->td_pri = pri; 1457 } 1458 crit_exit(); 1459 } 1460 } 1461 1462 /* 1463 * Set the initial priority for a thread prior to it being scheduled for 1464 * the first time. The thread MUST NOT be scheduled before or during 1465 * this call. The thread may be assigned to a cpu other then the current 1466 * cpu. 1467 * 1468 * Typically used after a thread has been created with TDF_STOPPREQ, 1469 * and before the thread is initially scheduled. 1470 */ 1471 void 1472 lwkt_setpri_initial(thread_t td, int pri) 1473 { 1474 KKASSERT(pri >= 0); 1475 KKASSERT((td->td_flags & TDF_RUNQ) == 0); 1476 td->td_pri = pri; 1477 } 1478 1479 void 1480 lwkt_setpri_self(int pri) 1481 { 1482 thread_t td = curthread; 1483 1484 KKASSERT(pri >= 0 && pri <= TDPRI_MAX); 1485 crit_enter(); 1486 if (td->td_flags & TDF_RUNQ) { 1487 _lwkt_dequeue(td); 1488 td->td_pri = pri; 1489 _lwkt_enqueue(td); 1490 } else { 1491 td->td_pri = pri; 1492 } 1493 crit_exit(); 1494 } 1495 1496 /* 1497 * hz tick scheduler clock for LWKT threads 1498 */ 1499 void 1500 lwkt_schedulerclock(thread_t td) 1501 { 1502 globaldata_t gd = td->td_gd; 1503 thread_t xtd; 1504 1505 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) { 1506 /* 1507 * If the current thread is at the head of the runq shift it to the 1508 * end of any equal-priority threads and request a LWKT reschedule 1509 * if it moved. 1510 * 1511 * Ignore upri in this situation. There will only be one user thread 1512 * in user mode, all others will be user threads running in kernel 1513 * mode and we have to make sure they get some cpu. 1514 */ 1515 xtd = TAILQ_NEXT(td, td_threadq); 1516 if (xtd && xtd->td_pri == td->td_pri) { 1517 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq); 1518 while (xtd && xtd->td_pri == td->td_pri) 1519 xtd = TAILQ_NEXT(xtd, td_threadq); 1520 if (xtd) 1521 TAILQ_INSERT_BEFORE(xtd, td, td_threadq); 1522 else 1523 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 1524 need_lwkt_resched(); 1525 } 1526 } else { 1527 /* 1528 * If we scheduled a thread other than the one at the head of the 1529 * queue always request a reschedule every tick. 1530 */ 1531 need_lwkt_resched(); 1532 } 1533 } 1534 1535 /* 1536 * Migrate the current thread to the specified cpu. 1537 * 1538 * This is accomplished by descheduling ourselves from the current cpu 1539 * and setting td_migrate_gd. The lwkt_switch() code will detect that the 1540 * 'old' thread wants to migrate after it has been completely switched out 1541 * and will complete the migration. 1542 * 1543 * TDF_MIGRATING prevents scheduling races while the thread is being migrated. 1544 * 1545 * We must be sure to release our current process designation (if a user 1546 * process) before clearing out any tsleepq we are on because the release 1547 * code may re-add us. 1548 * 1549 * We must be sure to remove ourselves from the current cpu's tsleepq 1550 * before potentially moving to another queue. The thread can be on 1551 * a tsleepq due to a left-over tsleep_interlock(). 1552 */ 1553 1554 void 1555 lwkt_setcpu_self(globaldata_t rgd) 1556 { 1557 thread_t td = curthread; 1558 1559 if (td->td_gd != rgd) { 1560 crit_enter_quick(td); 1561 1562 if (td->td_release) 1563 td->td_release(td); 1564 if (td->td_flags & TDF_TSLEEPQ) 1565 tsleep_remove(td); 1566 1567 /* 1568 * Set TDF_MIGRATING to prevent a spurious reschedule while we are 1569 * trying to deschedule ourselves and switch away, then deschedule 1570 * ourself, remove us from tdallq, and set td_migrate_gd. Finally, 1571 * call lwkt_switch() to complete the operation. 1572 */ 1573 td->td_flags |= TDF_MIGRATING; 1574 lwkt_deschedule_self(td); 1575 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1576 td->td_migrate_gd = rgd; 1577 lwkt_switch(); 1578 1579 /* 1580 * We are now on the target cpu 1581 */ 1582 KKASSERT(rgd == mycpu); 1583 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); 1584 crit_exit_quick(td); 1585 } 1586 } 1587 1588 void 1589 lwkt_migratecpu(int cpuid) 1590 { 1591 globaldata_t rgd; 1592 1593 rgd = globaldata_find(cpuid); 1594 lwkt_setcpu_self(rgd); 1595 } 1596 1597 /* 1598 * Remote IPI for cpu migration (called while in a critical section so we 1599 * do not have to enter another one). 1600 * 1601 * The thread (td) has already been completely descheduled from the 1602 * originating cpu and we can simply assert the case. The thread is 1603 * assigned to the new cpu and enqueued. 1604 * 1605 * The thread will re-add itself to tdallq when it resumes execution. 1606 */ 1607 static void 1608 lwkt_setcpu_remote(void *arg) 1609 { 1610 thread_t td = arg; 1611 globaldata_t gd = mycpu; 1612 1613 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0); 1614 td->td_gd = gd; 1615 cpu_mfence(); 1616 td->td_flags &= ~TDF_MIGRATING; 1617 KKASSERT(td->td_migrate_gd == NULL); 1618 KKASSERT(td->td_lwp == NULL || 1619 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 1620 _lwkt_enqueue(td); 1621 } 1622 1623 struct lwp * 1624 lwkt_preempted_proc(void) 1625 { 1626 thread_t td = curthread; 1627 while (td->td_preempted) 1628 td = td->td_preempted; 1629 return(td->td_lwp); 1630 } 1631 1632 /* 1633 * Create a kernel process/thread/whatever. It shares it's address space 1634 * with proc0 - ie: kernel only. 1635 * 1636 * If the cpu is not specified one will be selected. In the future 1637 * specifying a cpu of -1 will enable kernel thread migration between 1638 * cpus. 1639 */ 1640 int 1641 lwkt_create(void (*func)(void *), void *arg, struct thread **tdp, 1642 thread_t template, int tdflags, int cpu, const char *fmt, ...) 1643 { 1644 thread_t td; 1645 __va_list ap; 1646 1647 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu, 1648 tdflags); 1649 if (tdp) 1650 *tdp = td; 1651 cpu_set_thread_handler(td, lwkt_exit, func, arg); 1652 1653 /* 1654 * Set up arg0 for 'ps' etc 1655 */ 1656 __va_start(ap, fmt); 1657 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap); 1658 __va_end(ap); 1659 1660 /* 1661 * Schedule the thread to run 1662 */ 1663 if (td->td_flags & TDF_NOSTART) 1664 td->td_flags &= ~TDF_NOSTART; 1665 else 1666 lwkt_schedule(td); 1667 return 0; 1668 } 1669 1670 /* 1671 * Destroy an LWKT thread. Warning! This function is not called when 1672 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and 1673 * uses a different reaping mechanism. 1674 */ 1675 void 1676 lwkt_exit(void) 1677 { 1678 thread_t td = curthread; 1679 thread_t std; 1680 globaldata_t gd; 1681 1682 /* 1683 * Do any cleanup that might block here 1684 */ 1685 if (td->td_flags & TDF_VERBOSE) 1686 kprintf("kthread %p %s has exited\n", td, td->td_comm); 1687 biosched_done(td); 1688 dsched_exit_thread(td); 1689 1690 /* 1691 * Get us into a critical section to interlock gd_freetd and loop 1692 * until we can get it freed. 1693 * 1694 * We have to cache the current td in gd_freetd because objcache_put()ing 1695 * it would rip it out from under us while our thread is still active. 1696 * 1697 * We are the current thread so of course our own TDF_RUNNING bit will 1698 * be set, so unlike the lwp reap code we don't wait for it to clear. 1699 */ 1700 gd = mycpu; 1701 crit_enter_quick(td); 1702 for (;;) { 1703 if (td->td_refs) { 1704 tsleep(td, 0, "tdreap", 1); 1705 continue; 1706 } 1707 if ((std = gd->gd_freetd) != NULL) { 1708 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0); 1709 gd->gd_freetd = NULL; 1710 objcache_put(thread_cache, std); 1711 continue; 1712 } 1713 break; 1714 } 1715 1716 /* 1717 * Remove thread resources from kernel lists and deschedule us for 1718 * the last time. We cannot block after this point or we may end 1719 * up with a stale td on the tsleepq. 1720 * 1721 * None of this may block, the critical section is the only thing 1722 * protecting tdallq and the only thing preventing new lwkt_hold() 1723 * thread refs now. 1724 */ 1725 if (td->td_flags & TDF_TSLEEPQ) 1726 tsleep_remove(td); 1727 lwkt_deschedule_self(td); 1728 lwkt_remove_tdallq(td); 1729 KKASSERT(td->td_refs == 0); 1730 1731 /* 1732 * Final cleanup 1733 */ 1734 KKASSERT(gd->gd_freetd == NULL); 1735 if (td->td_flags & TDF_ALLOCATED_THREAD) 1736 gd->gd_freetd = td; 1737 cpu_thread_exit(); 1738 } 1739 1740 void 1741 lwkt_remove_tdallq(thread_t td) 1742 { 1743 KKASSERT(td->td_gd == mycpu); 1744 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1745 } 1746 1747 /* 1748 * Code reduction and branch prediction improvements. Call/return 1749 * overhead on modern cpus often degenerates into 0 cycles due to 1750 * the cpu's branch prediction hardware and return pc cache. We 1751 * can take advantage of this by not inlining medium-complexity 1752 * functions and we can also reduce the branch prediction impact 1753 * by collapsing perfectly predictable branches into a single 1754 * procedure instead of duplicating it. 1755 * 1756 * Is any of this noticeable? Probably not, so I'll take the 1757 * smaller code size. 1758 */ 1759 void 1760 crit_exit_wrapper(__DEBUG_CRIT_ARG__) 1761 { 1762 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__); 1763 } 1764 1765 void 1766 crit_panic(void) 1767 { 1768 thread_t td = curthread; 1769 int lcrit = td->td_critcount; 1770 1771 td->td_critcount = 0; 1772 panic("td_critcount is/would-go negative! %p %d", td, lcrit); 1773 /* NOT REACHED */ 1774 } 1775 1776 /* 1777 * Called from debugger/panic on cpus which have been stopped. We must still 1778 * process the IPIQ while stopped, even if we were stopped while in a critical 1779 * section (XXX). 1780 * 1781 * If we are dumping also try to process any pending interrupts. This may 1782 * or may not work depending on the state of the cpu at the point it was 1783 * stopped. 1784 */ 1785 void 1786 lwkt_smp_stopped(void) 1787 { 1788 globaldata_t gd = mycpu; 1789 1790 crit_enter_gd(gd); 1791 if (dumping) { 1792 lwkt_process_ipiq(); 1793 splz(); 1794 } else { 1795 lwkt_process_ipiq(); 1796 } 1797 crit_exit_gd(gd); 1798 } 1799