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/caps.h> 54 #include <sys/spinlock.h> 55 #include <sys/ktr.h> 56 57 #include <sys/thread2.h> 58 #include <sys/spinlock2.h> 59 #include <sys/mplock2.h> 60 61 #include <sys/dsched.h> 62 63 #include <vm/vm.h> 64 #include <vm/vm_param.h> 65 #include <vm/vm_kern.h> 66 #include <vm/vm_object.h> 67 #include <vm/vm_page.h> 68 #include <vm/vm_map.h> 69 #include <vm/vm_pager.h> 70 #include <vm/vm_extern.h> 71 72 #include <machine/stdarg.h> 73 #include <machine/smp.h> 74 75 #if !defined(KTR_CTXSW) 76 #define KTR_CTXSW KTR_ALL 77 #endif 78 KTR_INFO_MASTER(ctxsw); 79 KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p", int cpu, struct thread *td); 80 KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p", int cpu, struct thread *td); 81 KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s", struct thread *td, char *comm); 82 KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", struct thread *td); 83 84 static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads"); 85 86 #ifdef INVARIANTS 87 static int panic_on_cscount = 0; 88 #endif 89 static __int64_t switch_count = 0; 90 static __int64_t preempt_hit = 0; 91 static __int64_t preempt_miss = 0; 92 static __int64_t preempt_weird = 0; 93 static int lwkt_use_spin_port; 94 static struct objcache *thread_cache; 95 96 #ifdef SMP 97 static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame); 98 static void lwkt_setcpu_remote(void *arg); 99 #endif 100 101 extern void cpu_heavy_restore(void); 102 extern void cpu_lwkt_restore(void); 103 extern void cpu_kthread_restore(void); 104 extern void cpu_idle_restore(void); 105 106 /* 107 * We can make all thread ports use the spin backend instead of the thread 108 * backend. This should only be set to debug the spin backend. 109 */ 110 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port); 111 112 #ifdef INVARIANTS 113 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, 114 "Panic if attempting to switch lwkt's while mastering cpusync"); 115 #endif 116 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, 117 "Number of switched threads"); 118 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, 119 "Successful preemption events"); 120 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, 121 "Failed preemption events"); 122 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, 123 "Number of preempted threads."); 124 static int fairq_enable = 0; 125 SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW, 126 &fairq_enable, 0, "Turn on fairq priority accumulators"); 127 static int fairq_bypass = -1; 128 SYSCTL_INT(_lwkt, OID_AUTO, fairq_bypass, CTLFLAG_RW, 129 &fairq_bypass, 0, "Allow fairq to bypass td on token failure"); 130 extern int lwkt_sched_debug; 131 int lwkt_sched_debug = 0; 132 SYSCTL_INT(_lwkt, OID_AUTO, sched_debug, CTLFLAG_RW, 133 &lwkt_sched_debug, 0, "Scheduler debug"); 134 static int lwkt_spin_loops = 10; 135 SYSCTL_INT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW, 136 &lwkt_spin_loops, 0, "Scheduler spin loops until sorted decon"); 137 static int lwkt_spin_reseq = 0; 138 SYSCTL_INT(_lwkt, OID_AUTO, spin_reseq, CTLFLAG_RW, 139 &lwkt_spin_reseq, 0, "Scheduler resequencer enable"); 140 static int lwkt_spin_monitor = 0; 141 SYSCTL_INT(_lwkt, OID_AUTO, spin_monitor, CTLFLAG_RW, 142 &lwkt_spin_monitor, 0, "Scheduler uses monitor/mwait"); 143 static int lwkt_spin_fatal = 0; /* disabled */ 144 SYSCTL_INT(_lwkt, OID_AUTO, spin_fatal, CTLFLAG_RW, 145 &lwkt_spin_fatal, 0, "LWKT scheduler spin loops till fatal panic"); 146 static int preempt_enable = 1; 147 SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW, 148 &preempt_enable, 0, "Enable preemption"); 149 static int lwkt_cache_threads = 0; 150 SYSCTL_INT(_lwkt, OID_AUTO, cache_threads, CTLFLAG_RD, 151 &lwkt_cache_threads, 0, "thread+kstack cache"); 152 153 static __cachealign int lwkt_cseq_rindex; 154 static __cachealign int lwkt_cseq_windex; 155 156 /* 157 * These helper procedures handle the runq, they can only be called from 158 * within a critical section. 159 * 160 * WARNING! Prior to SMP being brought up it is possible to enqueue and 161 * dequeue threads belonging to other cpus, so be sure to use td->td_gd 162 * instead of 'mycpu' when referencing the globaldata structure. Once 163 * SMP live enqueuing and dequeueing only occurs on the current cpu. 164 */ 165 static __inline 166 void 167 _lwkt_dequeue(thread_t td) 168 { 169 if (td->td_flags & TDF_RUNQ) { 170 struct globaldata *gd = td->td_gd; 171 172 td->td_flags &= ~TDF_RUNQ; 173 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq); 174 --gd->gd_tdrunqcount; 175 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL) 176 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING); 177 } 178 } 179 180 /* 181 * Priority enqueue. 182 * 183 * There are a limited number of lwkt threads runnable since user 184 * processes only schedule one at a time per cpu. However, there can 185 * be many user processes in kernel mode exiting from a tsleep() which 186 * become runnable. 187 * 188 * NOTE: lwkt_schedulerclock() will force a round-robin based on td_pri and 189 * will ignore user priority. This is to ensure that user threads in 190 * kernel mode get cpu at some point regardless of what the user 191 * scheduler thinks. 192 */ 193 static __inline 194 void 195 _lwkt_enqueue(thread_t td) 196 { 197 thread_t xtd; 198 199 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) { 200 struct globaldata *gd = td->td_gd; 201 202 td->td_flags |= TDF_RUNQ; 203 xtd = TAILQ_FIRST(&gd->gd_tdrunq); 204 if (xtd == NULL) { 205 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 206 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING); 207 } else { 208 /* 209 * NOTE: td_upri - higher numbers more desireable, same sense 210 * as td_pri (typically reversed from lwp_upri). 211 * 212 * In the equal priority case we want the best selection 213 * at the beginning so the less desireable selections know 214 * that they have to setrunqueue/go-to-another-cpu, even 215 * though it means switching back to the 'best' selection. 216 * This also avoids degenerate situations when many threads 217 * are runnable or waking up at the same time. 218 * 219 * If upri matches exactly place at end/round-robin. 220 */ 221 while (xtd && 222 (xtd->td_pri >= td->td_pri || 223 (xtd->td_pri == td->td_pri && 224 xtd->td_upri >= td->td_upri))) { 225 xtd = TAILQ_NEXT(xtd, td_threadq); 226 } 227 if (xtd) 228 TAILQ_INSERT_BEFORE(xtd, td, td_threadq); 229 else 230 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 231 } 232 ++gd->gd_tdrunqcount; 233 234 /* 235 * Request a LWKT reschedule if we are now at the head of the queue. 236 */ 237 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) 238 need_lwkt_resched(); 239 } 240 } 241 242 static __boolean_t 243 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags) 244 { 245 struct thread *td = (struct thread *)obj; 246 247 td->td_kstack = NULL; 248 td->td_kstack_size = 0; 249 td->td_flags = TDF_ALLOCATED_THREAD; 250 td->td_mpflags = 0; 251 return (1); 252 } 253 254 static void 255 _lwkt_thread_dtor(void *obj, void *privdata) 256 { 257 struct thread *td = (struct thread *)obj; 258 259 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD, 260 ("_lwkt_thread_dtor: not allocated from objcache")); 261 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack && 262 td->td_kstack_size > 0, 263 ("_lwkt_thread_dtor: corrupted stack")); 264 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size); 265 td->td_kstack = NULL; 266 td->td_flags = 0; 267 } 268 269 /* 270 * Initialize the lwkt s/system. 271 * 272 * Nominally cache up to 32 thread + kstack structures. Cache more on 273 * systems with a lot of cpu cores. 274 */ 275 void 276 lwkt_init(void) 277 { 278 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads); 279 if (lwkt_cache_threads == 0) { 280 lwkt_cache_threads = ncpus * 4; 281 if (lwkt_cache_threads < 32) 282 lwkt_cache_threads = 32; 283 } 284 thread_cache = objcache_create_mbacked( 285 M_THREAD, sizeof(struct thread), 286 NULL, lwkt_cache_threads, 287 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL); 288 } 289 290 /* 291 * Schedule a thread to run. As the current thread we can always safely 292 * schedule ourselves, and a shortcut procedure is provided for that 293 * function. 294 * 295 * (non-blocking, self contained on a per cpu basis) 296 */ 297 void 298 lwkt_schedule_self(thread_t td) 299 { 300 KKASSERT((td->td_flags & TDF_MIGRATING) == 0); 301 crit_enter_quick(td); 302 KASSERT(td != &td->td_gd->gd_idlethread, 303 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!")); 304 KKASSERT(td->td_lwp == NULL || 305 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 306 _lwkt_enqueue(td); 307 crit_exit_quick(td); 308 } 309 310 /* 311 * Deschedule a thread. 312 * 313 * (non-blocking, self contained on a per cpu basis) 314 */ 315 void 316 lwkt_deschedule_self(thread_t td) 317 { 318 crit_enter_quick(td); 319 _lwkt_dequeue(td); 320 crit_exit_quick(td); 321 } 322 323 /* 324 * LWKTs operate on a per-cpu basis 325 * 326 * WARNING! Called from early boot, 'mycpu' may not work yet. 327 */ 328 void 329 lwkt_gdinit(struct globaldata *gd) 330 { 331 TAILQ_INIT(&gd->gd_tdrunq); 332 TAILQ_INIT(&gd->gd_tdallq); 333 } 334 335 /* 336 * Create a new thread. The thread must be associated with a process context 337 * or LWKT start address before it can be scheduled. If the target cpu is 338 * -1 the thread will be created on the current cpu. 339 * 340 * If you intend to create a thread without a process context this function 341 * does everything except load the startup and switcher function. 342 */ 343 thread_t 344 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags) 345 { 346 static int cpu_rotator; 347 globaldata_t gd = mycpu; 348 void *stack; 349 350 /* 351 * If static thread storage is not supplied allocate a thread. Reuse 352 * a cached free thread if possible. gd_freetd is used to keep an exiting 353 * thread intact through the exit. 354 */ 355 if (td == NULL) { 356 crit_enter_gd(gd); 357 if ((td = gd->gd_freetd) != NULL) { 358 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK| 359 TDF_RUNQ)) == 0); 360 gd->gd_freetd = NULL; 361 } else { 362 td = objcache_get(thread_cache, M_WAITOK); 363 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK| 364 TDF_RUNQ)) == 0); 365 } 366 crit_exit_gd(gd); 367 KASSERT((td->td_flags & 368 (TDF_ALLOCATED_THREAD|TDF_RUNNING|TDF_PREEMPT_LOCK)) == 369 TDF_ALLOCATED_THREAD, 370 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags)); 371 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK); 372 } 373 374 /* 375 * Try to reuse cached stack. 376 */ 377 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) { 378 if (flags & TDF_ALLOCATED_STACK) { 379 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size); 380 stack = NULL; 381 } 382 } 383 if (stack == NULL) { 384 stack = (void *)kmem_alloc_stack(&kernel_map, stksize); 385 flags |= TDF_ALLOCATED_STACK; 386 } 387 if (cpu < 0) { 388 cpu = ++cpu_rotator; 389 cpu_ccfence(); 390 cpu %= ncpus; 391 } 392 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu)); 393 return(td); 394 } 395 396 /* 397 * Initialize a preexisting thread structure. This function is used by 398 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread. 399 * 400 * All threads start out in a critical section at a priority of 401 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as 402 * appropriate. This function may send an IPI message when the 403 * requested cpu is not the current cpu and consequently gd_tdallq may 404 * not be initialized synchronously from the point of view of the originating 405 * cpu. 406 * 407 * NOTE! we have to be careful in regards to creating threads for other cpus 408 * if SMP has not yet been activated. 409 */ 410 #ifdef SMP 411 412 static void 413 lwkt_init_thread_remote(void *arg) 414 { 415 thread_t td = arg; 416 417 /* 418 * Protected by critical section held by IPI dispatch 419 */ 420 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq); 421 } 422 423 #endif 424 425 /* 426 * lwkt core thread structural initialization. 427 * 428 * NOTE: All threads are initialized as mpsafe threads. 429 */ 430 void 431 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags, 432 struct globaldata *gd) 433 { 434 globaldata_t mygd = mycpu; 435 436 bzero(td, sizeof(struct thread)); 437 td->td_kstack = stack; 438 td->td_kstack_size = stksize; 439 td->td_flags = flags; 440 td->td_mpflags = 0; 441 td->td_gd = gd; 442 td->td_pri = TDPRI_KERN_DAEMON; 443 td->td_critcount = 1; 444 td->td_toks_have = NULL; 445 td->td_toks_stop = &td->td_toks_base; 446 if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT)) 447 lwkt_initport_spin(&td->td_msgport, td); 448 else 449 lwkt_initport_thread(&td->td_msgport, td); 450 pmap_init_thread(td); 451 #ifdef SMP 452 /* 453 * Normally initializing a thread for a remote cpu requires sending an 454 * IPI. However, the idlethread is setup before the other cpus are 455 * activated so we have to treat it as a special case. XXX manipulation 456 * of gd_tdallq requires the BGL. 457 */ 458 if (gd == mygd || td == &gd->gd_idlethread) { 459 crit_enter_gd(mygd); 460 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); 461 crit_exit_gd(mygd); 462 } else { 463 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td); 464 } 465 #else 466 crit_enter_gd(mygd); 467 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); 468 crit_exit_gd(mygd); 469 #endif 470 471 dsched_new_thread(td); 472 } 473 474 void 475 lwkt_set_comm(thread_t td, const char *ctl, ...) 476 { 477 __va_list va; 478 479 __va_start(va, ctl); 480 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va); 481 __va_end(va); 482 KTR_LOG(ctxsw_newtd, td, td->td_comm); 483 } 484 485 /* 486 * Prevent the thread from getting destroyed. Note that unlike PHOLD/PRELE 487 * this does not prevent the thread from migrating to another cpu so the 488 * gd_tdallq state is not protected by this. 489 */ 490 void 491 lwkt_hold(thread_t td) 492 { 493 atomic_add_int(&td->td_refs, 1); 494 } 495 496 void 497 lwkt_rele(thread_t td) 498 { 499 KKASSERT(td->td_refs > 0); 500 atomic_add_int(&td->td_refs, -1); 501 } 502 503 void 504 lwkt_free_thread(thread_t td) 505 { 506 KKASSERT(td->td_refs == 0); 507 KKASSERT((td->td_flags & (TDF_RUNNING | TDF_PREEMPT_LOCK | 508 TDF_RUNQ | TDF_TSLEEPQ)) == 0); 509 if (td->td_flags & TDF_ALLOCATED_THREAD) { 510 objcache_put(thread_cache, td); 511 } else if (td->td_flags & TDF_ALLOCATED_STACK) { 512 /* client-allocated struct with internally allocated stack */ 513 KASSERT(td->td_kstack && td->td_kstack_size > 0, 514 ("lwkt_free_thread: corrupted stack")); 515 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size); 516 td->td_kstack = NULL; 517 td->td_kstack_size = 0; 518 } 519 KTR_LOG(ctxsw_deadtd, td); 520 } 521 522 523 /* 524 * Switch to the next runnable lwkt. If no LWKTs are runnable then 525 * switch to the idlethread. Switching must occur within a critical 526 * section to avoid races with the scheduling queue. 527 * 528 * We always have full control over our cpu's run queue. Other cpus 529 * that wish to manipulate our queue must use the cpu_*msg() calls to 530 * talk to our cpu, so a critical section is all that is needed and 531 * the result is very, very fast thread switching. 532 * 533 * The LWKT scheduler uses a fixed priority model and round-robins at 534 * each priority level. User process scheduling is a totally 535 * different beast and LWKT priorities should not be confused with 536 * user process priorities. 537 * 538 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch() 539 * is not called by the current thread in the preemption case, only when 540 * the preempting thread blocks (in order to return to the original thread). 541 * 542 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread 543 * migration and tsleep deschedule the current lwkt thread and call 544 * lwkt_switch(). In particular, the target cpu of the migration fully 545 * expects the thread to become non-runnable and can deadlock against 546 * cpusync operations if we run any IPIs prior to switching the thread out. 547 * 548 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF 549 * THE CURRENT THREAD HAS BEEN DESCHEDULED! 550 */ 551 void 552 lwkt_switch(void) 553 { 554 globaldata_t gd = mycpu; 555 thread_t td = gd->gd_curthread; 556 thread_t ntd; 557 int spinning = 0; 558 559 KKASSERT(gd->gd_processing_ipiq == 0); 560 KKASSERT(td->td_flags & TDF_RUNNING); 561 562 /* 563 * Switching from within a 'fast' (non thread switched) interrupt or IPI 564 * is illegal. However, we may have to do it anyway if we hit a fatal 565 * kernel trap or we have paniced. 566 * 567 * If this case occurs save and restore the interrupt nesting level. 568 */ 569 if (gd->gd_intr_nesting_level) { 570 int savegdnest; 571 int savegdtrap; 572 573 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) { 574 panic("lwkt_switch: Attempt to switch from a " 575 "fast interrupt, ipi, or hard code section, " 576 "td %p\n", 577 td); 578 } else { 579 savegdnest = gd->gd_intr_nesting_level; 580 savegdtrap = gd->gd_trap_nesting_level; 581 gd->gd_intr_nesting_level = 0; 582 gd->gd_trap_nesting_level = 0; 583 if ((td->td_flags & TDF_PANICWARN) == 0) { 584 td->td_flags |= TDF_PANICWARN; 585 kprintf("Warning: thread switch from interrupt, IPI, " 586 "or hard code section.\n" 587 "thread %p (%s)\n", td, td->td_comm); 588 print_backtrace(-1); 589 } 590 lwkt_switch(); 591 gd->gd_intr_nesting_level = savegdnest; 592 gd->gd_trap_nesting_level = savegdtrap; 593 return; 594 } 595 } 596 597 /* 598 * Release our current user process designation if we are blocking 599 * or if a user reschedule was requested. 600 * 601 * NOTE: This function is NOT called if we are switching into or 602 * returning from a preemption. 603 * 604 * NOTE: Releasing our current user process designation may cause 605 * it to be assigned to another thread, which in turn will 606 * cause us to block in the usched acquire code when we attempt 607 * to return to userland. 608 * 609 * NOTE: On SMP systems this can be very nasty when heavy token 610 * contention is present so we want to be careful not to 611 * release the designation gratuitously. 612 */ 613 if (td->td_release && 614 (user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) { 615 td->td_release(td); 616 } 617 618 /* 619 * Release all tokens 620 */ 621 crit_enter_gd(gd); 622 if (TD_TOKS_HELD(td)) 623 lwkt_relalltokens(td); 624 625 /* 626 * We had better not be holding any spin locks, but don't get into an 627 * endless panic loop. 628 */ 629 KASSERT(gd->gd_spinlocks == 0 || panicstr != NULL, 630 ("lwkt_switch: still holding %d exclusive spinlocks!", 631 gd->gd_spinlocks)); 632 633 634 #ifdef SMP 635 #ifdef INVARIANTS 636 if (td->td_cscount) { 637 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n", 638 td); 639 if (panic_on_cscount) 640 panic("switching while mastering cpusync"); 641 } 642 #endif 643 #endif 644 645 /* 646 * If we had preempted another thread on this cpu, resume the preempted 647 * thread. This occurs transparently, whether the preempted thread 648 * was scheduled or not (it may have been preempted after descheduling 649 * itself). 650 * 651 * We have to setup the MP lock for the original thread after backing 652 * out the adjustment that was made to curthread when the original 653 * was preempted. 654 */ 655 if ((ntd = td->td_preempted) != NULL) { 656 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK); 657 ntd->td_flags |= TDF_PREEMPT_DONE; 658 659 /* 660 * The interrupt may have woken a thread up, we need to properly 661 * set the reschedule flag if the originally interrupted thread is 662 * at a lower priority. 663 * 664 * The interrupt may not have descheduled. 665 */ 666 if (TAILQ_FIRST(&gd->gd_tdrunq) != ntd) 667 need_lwkt_resched(); 668 goto havethread_preempted; 669 } 670 671 /* 672 * If we cannot obtain ownership of the tokens we cannot immediately 673 * schedule the target thread. 674 * 675 * Reminder: Again, we cannot afford to run any IPIs in this path if 676 * the current thread has been descheduled. 677 */ 678 for (;;) { 679 clear_lwkt_resched(); 680 681 /* 682 * Hotpath - pull the head of the run queue and attempt to schedule 683 * it. 684 */ 685 ntd = TAILQ_FIRST(&gd->gd_tdrunq); 686 687 if (ntd == NULL) { 688 /* 689 * Runq is empty, switch to idle to allow it to halt. 690 */ 691 ntd = &gd->gd_idlethread; 692 #ifdef SMP 693 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) 694 ASSERT_NO_TOKENS_HELD(ntd); 695 #endif 696 cpu_time.cp_msg[0] = 0; 697 cpu_time.cp_stallpc = 0; 698 goto haveidle; 699 } 700 701 /* 702 * Hotpath - schedule ntd. 703 * 704 * NOTE: For UP there is no mplock and lwkt_getalltokens() 705 * always succeeds. 706 */ 707 if (TD_TOKS_NOT_HELD(ntd) || 708 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops))) 709 { 710 goto havethread; 711 } 712 713 /* 714 * Coldpath (SMP only since tokens always succeed on UP) 715 * 716 * We had some contention on the thread we wanted to schedule. 717 * What we do now is try to find a thread that we can schedule 718 * in its stead. 719 * 720 * The coldpath scan does NOT rearrange threads in the run list. 721 * The lwkt_schedulerclock() will assert need_lwkt_resched() on 722 * the next tick whenever the current head is not the current thread. 723 */ 724 #ifdef INVARIANTS 725 ++ntd->td_contended; 726 #endif 727 ++gd->gd_cnt.v_token_colls; 728 729 if (fairq_bypass > 0) 730 goto skip; 731 732 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) { 733 #ifndef NO_LWKT_SPLIT_USERPRI 734 /* 735 * Never schedule threads returning to userland or the 736 * user thread scheduler helper thread when higher priority 737 * threads are present. The runq is sorted by priority 738 * so we can give up traversing it when we find the first 739 * low priority thread. 740 */ 741 if (ntd->td_pri < TDPRI_KERN_LPSCHED) { 742 ntd = NULL; 743 break; 744 } 745 #endif 746 747 /* 748 * Try this one. 749 */ 750 if (TD_TOKS_NOT_HELD(ntd) || 751 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops))) { 752 goto havethread; 753 } 754 #ifdef INVARIANTS 755 ++ntd->td_contended; 756 #endif 757 ++gd->gd_cnt.v_token_colls; 758 } 759 760 skip: 761 /* 762 * We exhausted the run list, meaning that all runnable threads 763 * are contested. 764 */ 765 cpu_pause(); 766 ntd = &gd->gd_idlethread; 767 #ifdef SMP 768 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) 769 ASSERT_NO_TOKENS_HELD(ntd); 770 /* contention case, do not clear contention mask */ 771 #endif 772 773 /* 774 * We are going to have to retry but if the current thread is not 775 * on the runq we instead switch through the idle thread to get away 776 * from the current thread. We have to flag for lwkt reschedule 777 * to prevent the idle thread from halting. 778 * 779 * NOTE: A non-zero spinning is passed to lwkt_getalltokens() to 780 * instruct it to deal with the potential for deadlocks by 781 * ordering the tokens by address. 782 */ 783 if ((td->td_flags & TDF_RUNQ) == 0) { 784 need_lwkt_resched(); /* prevent hlt */ 785 goto haveidle; 786 } 787 #if defined(INVARIANTS) && defined(__amd64__) 788 if ((read_rflags() & PSL_I) == 0) { 789 cpu_enable_intr(); 790 panic("lwkt_switch() called with interrupts disabled"); 791 } 792 #endif 793 794 /* 795 * Number iterations so far. After a certain point we switch to 796 * a sorted-address/monitor/mwait version of lwkt_getalltokens() 797 */ 798 if (spinning < 0x7FFFFFFF) 799 ++spinning; 800 801 #ifdef SMP 802 /* 803 * lwkt_getalltokens() failed in sorted token mode, we can use 804 * monitor/mwait in this case. 805 */ 806 if (spinning >= lwkt_spin_loops && 807 (cpu_mi_feature & CPU_MI_MONITOR) && 808 lwkt_spin_monitor) 809 { 810 cpu_mmw_pause_int(&gd->gd_reqflags, 811 (gd->gd_reqflags | RQF_SPINNING) & 812 ~RQF_IDLECHECK_WK_MASK); 813 } 814 #endif 815 816 /* 817 * We already checked that td is still scheduled so this should be 818 * safe. 819 */ 820 splz_check(); 821 822 /* 823 * This experimental resequencer is used as a fall-back to reduce 824 * hw cache line contention by placing each core's scheduler into a 825 * time-domain-multplexed slot. 826 * 827 * The resequencer is disabled by default. It's functionality has 828 * largely been superceeded by the token algorithm which limits races 829 * to a subset of cores. 830 * 831 * The resequencer algorithm tends to break down when more than 832 * 20 cores are contending. What appears to happen is that new 833 * tokens can be obtained out of address-sorted order by new cores 834 * while existing cores languish in long delays between retries and 835 * wind up being starved-out of the token acquisition. 836 */ 837 if (lwkt_spin_reseq && spinning >= lwkt_spin_reseq) { 838 int cseq = atomic_fetchadd_int(&lwkt_cseq_windex, 1); 839 int oseq; 840 841 while ((oseq = lwkt_cseq_rindex) != cseq) { 842 cpu_ccfence(); 843 #if 1 844 if (cpu_mi_feature & CPU_MI_MONITOR) { 845 cpu_mmw_pause_int(&lwkt_cseq_rindex, oseq); 846 } else { 847 #endif 848 cpu_pause(); 849 cpu_lfence(); 850 #if 1 851 } 852 #endif 853 } 854 DELAY(1); 855 atomic_add_int(&lwkt_cseq_rindex, 1); 856 } 857 /* highest level for(;;) loop */ 858 } 859 860 havethread: 861 /* 862 * Clear gd_idle_repeat when doing a normal switch to a non-idle 863 * thread. 864 */ 865 ntd->td_wmesg = NULL; 866 ++gd->gd_cnt.v_swtch; 867 gd->gd_idle_repeat = 0; 868 869 havethread_preempted: 870 /* 871 * If the new target does not need the MP lock and we are holding it, 872 * release the MP lock. If the new target requires the MP lock we have 873 * already acquired it for the target. 874 */ 875 ; 876 haveidle: 877 KASSERT(ntd->td_critcount, 878 ("priority problem in lwkt_switch %d %d", 879 td->td_critcount, ntd->td_critcount)); 880 881 if (td != ntd) { 882 /* 883 * Execute the actual thread switch operation. This function 884 * returns to the current thread and returns the previous thread 885 * (which may be different from the thread we switched to). 886 * 887 * We are responsible for marking ntd as TDF_RUNNING. 888 */ 889 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0); 890 ++switch_count; 891 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd); 892 ntd->td_flags |= TDF_RUNNING; 893 lwkt_switch_return(td->td_switch(ntd)); 894 /* ntd invalid, td_switch() can return a different thread_t */ 895 } 896 897 /* 898 * catch-all. XXX is this strictly needed? 899 */ 900 splz_check(); 901 902 /* NOTE: current cpu may have changed after switch */ 903 crit_exit_quick(td); 904 } 905 906 /* 907 * Called by assembly in the td_switch (thread restore path) for thread 908 * bootstrap cases which do not 'return' to lwkt_switch(). 909 */ 910 void 911 lwkt_switch_return(thread_t otd) 912 { 913 #ifdef SMP 914 globaldata_t rgd; 915 916 /* 917 * Check if otd was migrating. Now that we are on ntd we can finish 918 * up the migration. This is a bit messy but it is the only place 919 * where td is known to be fully descheduled. 920 * 921 * We can only activate the migration if otd was migrating but not 922 * held on the cpu due to a preemption chain. We still have to 923 * clear TDF_RUNNING on the old thread either way. 924 * 925 * We are responsible for clearing the previously running thread's 926 * TDF_RUNNING. 927 */ 928 if ((rgd = otd->td_migrate_gd) != NULL && 929 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) { 930 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) == 931 (TDF_MIGRATING | TDF_RUNNING)); 932 otd->td_migrate_gd = NULL; 933 otd->td_flags &= ~TDF_RUNNING; 934 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd); 935 } else { 936 otd->td_flags &= ~TDF_RUNNING; 937 } 938 #else 939 otd->td_flags &= ~TDF_RUNNING; 940 #endif 941 942 /* 943 * Final exit validations (see lwp_wait()). Note that otd becomes 944 * invalid the *instant* we set TDF_MP_EXITSIG. 945 */ 946 while (otd->td_flags & TDF_EXITING) { 947 u_int mpflags; 948 949 mpflags = otd->td_mpflags; 950 cpu_ccfence(); 951 952 if (mpflags & TDF_MP_EXITWAIT) { 953 if (atomic_cmpset_int(&otd->td_mpflags, mpflags, 954 mpflags | TDF_MP_EXITSIG)) { 955 wakeup(otd); 956 break; 957 } 958 } else { 959 if (atomic_cmpset_int(&otd->td_mpflags, mpflags, 960 mpflags | TDF_MP_EXITSIG)) { 961 wakeup(otd); 962 break; 963 } 964 } 965 } 966 } 967 968 /* 969 * Request that the target thread preempt the current thread. Preemption 970 * can only occur if our only critical section is the one that we were called 971 * with, the relative priority of the target thread is higher, and the target 972 * thread holds no tokens. This also only works if we are not holding any 973 * spinlocks (obviously). 974 * 975 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically 976 * this is called via lwkt_schedule() through the td_preemptable callback. 977 * critcount is the managed critical priority that we should ignore in order 978 * to determine whether preemption is possible (aka usually just the crit 979 * priority of lwkt_schedule() itself). 980 * 981 * Preemption is typically limited to interrupt threads. 982 * 983 * Operation works in a fairly straight-forward manner. The normal 984 * scheduling code is bypassed and we switch directly to the target 985 * thread. When the target thread attempts to block or switch away 986 * code at the base of lwkt_switch() will switch directly back to our 987 * thread. Our thread is able to retain whatever tokens it holds and 988 * if the target needs one of them the target will switch back to us 989 * and reschedule itself normally. 990 */ 991 void 992 lwkt_preempt(thread_t ntd, int critcount) 993 { 994 struct globaldata *gd = mycpu; 995 thread_t xtd; 996 thread_t td; 997 int save_gd_intr_nesting_level; 998 999 /* 1000 * The caller has put us in a critical section. We can only preempt 1001 * if the caller of the caller was not in a critical section (basically 1002 * a local interrupt), as determined by the 'critcount' parameter. We 1003 * also can't preempt if the caller is holding any spinlocks (even if 1004 * he isn't in a critical section). This also handles the tokens test. 1005 * 1006 * YYY The target thread must be in a critical section (else it must 1007 * inherit our critical section? I dunno yet). 1008 */ 1009 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri)); 1010 1011 td = gd->gd_curthread; 1012 if (preempt_enable == 0) { 1013 ++preempt_miss; 1014 return; 1015 } 1016 if (ntd->td_pri <= td->td_pri) { 1017 ++preempt_miss; 1018 return; 1019 } 1020 if (td->td_critcount > critcount) { 1021 ++preempt_miss; 1022 return; 1023 } 1024 #ifdef SMP 1025 if (td->td_cscount) { 1026 ++preempt_miss; 1027 return; 1028 } 1029 if (ntd->td_gd != gd) { 1030 ++preempt_miss; 1031 return; 1032 } 1033 #endif 1034 /* 1035 * We don't have to check spinlocks here as they will also bump 1036 * td_critcount. 1037 * 1038 * Do not try to preempt if the target thread is holding any tokens. 1039 * We could try to acquire the tokens but this case is so rare there 1040 * is no need to support it. 1041 */ 1042 KKASSERT(gd->gd_spinlocks == 0); 1043 1044 if (TD_TOKS_HELD(ntd)) { 1045 ++preempt_miss; 1046 return; 1047 } 1048 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) { 1049 ++preempt_weird; 1050 return; 1051 } 1052 if (ntd->td_preempted) { 1053 ++preempt_hit; 1054 return; 1055 } 1056 KKASSERT(gd->gd_processing_ipiq == 0); 1057 1058 /* 1059 * Since we are able to preempt the current thread, there is no need to 1060 * call need_lwkt_resched(). 1061 * 1062 * We must temporarily clear gd_intr_nesting_level around the switch 1063 * since switchouts from the target thread are allowed (they will just 1064 * return to our thread), and since the target thread has its own stack. 1065 * 1066 * A preemption must switch back to the original thread, assert the 1067 * case. 1068 */ 1069 ++preempt_hit; 1070 ntd->td_preempted = td; 1071 td->td_flags |= TDF_PREEMPT_LOCK; 1072 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd); 1073 save_gd_intr_nesting_level = gd->gd_intr_nesting_level; 1074 gd->gd_intr_nesting_level = 0; 1075 1076 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0); 1077 ntd->td_flags |= TDF_RUNNING; 1078 xtd = td->td_switch(ntd); 1079 KKASSERT(xtd == ntd); 1080 lwkt_switch_return(xtd); 1081 gd->gd_intr_nesting_level = save_gd_intr_nesting_level; 1082 1083 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE)); 1084 ntd->td_preempted = NULL; 1085 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE); 1086 } 1087 1088 /* 1089 * Conditionally call splz() if gd_reqflags indicates work is pending. 1090 * This will work inside a critical section but not inside a hard code 1091 * section. 1092 * 1093 * (self contained on a per cpu basis) 1094 */ 1095 void 1096 splz_check(void) 1097 { 1098 globaldata_t gd = mycpu; 1099 thread_t td = gd->gd_curthread; 1100 1101 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && 1102 gd->gd_intr_nesting_level == 0 && 1103 td->td_nest_count < 2) 1104 { 1105 splz(); 1106 } 1107 } 1108 1109 /* 1110 * This version is integrated into crit_exit, reqflags has already 1111 * been tested but td_critcount has not. 1112 * 1113 * We only want to execute the splz() on the 1->0 transition of 1114 * critcount and not in a hard code section or if too deeply nested. 1115 * 1116 * NOTE: gd->gd_spinlocks is implied to be 0 when td_critcount is 0. 1117 */ 1118 void 1119 lwkt_maybe_splz(thread_t td) 1120 { 1121 globaldata_t gd = td->td_gd; 1122 1123 if (td->td_critcount == 0 && 1124 gd->gd_intr_nesting_level == 0 && 1125 td->td_nest_count < 2) 1126 { 1127 splz(); 1128 } 1129 } 1130 1131 /* 1132 * Drivers which set up processing co-threads can call this function to 1133 * run the co-thread at a higher priority and to allow it to preempt 1134 * normal threads. 1135 */ 1136 void 1137 lwkt_set_interrupt_support_thread(void) 1138 { 1139 thread_t td = curthread; 1140 1141 lwkt_setpri_self(TDPRI_INT_SUPPORT); 1142 td->td_flags |= TDF_INTTHREAD; 1143 td->td_preemptable = lwkt_preempt; 1144 } 1145 1146 1147 /* 1148 * This function is used to negotiate a passive release of the current 1149 * process/lwp designation with the user scheduler, allowing the user 1150 * scheduler to schedule another user thread. The related kernel thread 1151 * (curthread) continues running in the released state. 1152 */ 1153 void 1154 lwkt_passive_release(struct thread *td) 1155 { 1156 struct lwp *lp = td->td_lwp; 1157 1158 #ifndef NO_LWKT_SPLIT_USERPRI 1159 td->td_release = NULL; 1160 lwkt_setpri_self(TDPRI_KERN_USER); 1161 #endif 1162 1163 lp->lwp_proc->p_usched->release_curproc(lp); 1164 } 1165 1166 1167 /* 1168 * This implements a LWKT yield, allowing a kernel thread to yield to other 1169 * kernel threads at the same or higher priority. This function can be 1170 * called in a tight loop and will typically only yield once per tick. 1171 * 1172 * Most kernel threads run at the same priority in order to allow equal 1173 * sharing. 1174 * 1175 * (self contained on a per cpu basis) 1176 */ 1177 void 1178 lwkt_yield(void) 1179 { 1180 globaldata_t gd = mycpu; 1181 thread_t td = gd->gd_curthread; 1182 1183 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1184 splz(); 1185 if (lwkt_resched_wanted()) { 1186 lwkt_schedule_self(curthread); 1187 lwkt_switch(); 1188 } 1189 } 1190 1191 /* 1192 * The quick version processes pending interrupts and higher-priority 1193 * LWKT threads but will not round-robin same-priority LWKT threads. 1194 * 1195 * When called while attempting to return to userland the only same-pri 1196 * threads are the ones which have already tried to become the current 1197 * user process. 1198 */ 1199 void 1200 lwkt_yield_quick(void) 1201 { 1202 globaldata_t gd = mycpu; 1203 thread_t td = gd->gd_curthread; 1204 1205 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1206 splz(); 1207 if (lwkt_resched_wanted()) { 1208 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) { 1209 clear_lwkt_resched(); 1210 } else { 1211 lwkt_schedule_self(curthread); 1212 lwkt_switch(); 1213 } 1214 } 1215 } 1216 1217 /* 1218 * This yield is designed for kernel threads with a user context. 1219 * 1220 * The kernel acting on behalf of the user is potentially cpu-bound, 1221 * this function will efficiently allow other threads to run and also 1222 * switch to other processes by releasing. 1223 * 1224 * The lwkt_user_yield() function is designed to have very low overhead 1225 * if no yield is determined to be needed. 1226 */ 1227 void 1228 lwkt_user_yield(void) 1229 { 1230 globaldata_t gd = mycpu; 1231 thread_t td = gd->gd_curthread; 1232 1233 /* 1234 * Always run any pending interrupts in case we are in a critical 1235 * section. 1236 */ 1237 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1238 splz(); 1239 1240 /* 1241 * Switch (which forces a release) if another kernel thread needs 1242 * the cpu, if userland wants us to resched, or if our kernel 1243 * quantum has run out. 1244 */ 1245 if (lwkt_resched_wanted() || 1246 user_resched_wanted()) 1247 { 1248 lwkt_switch(); 1249 } 1250 1251 #if 0 1252 /* 1253 * Reacquire the current process if we are released. 1254 * 1255 * XXX not implemented atm. The kernel may be holding locks and such, 1256 * so we want the thread to continue to receive cpu. 1257 */ 1258 if (td->td_release == NULL && lp) { 1259 lp->lwp_proc->p_usched->acquire_curproc(lp); 1260 td->td_release = lwkt_passive_release; 1261 lwkt_setpri_self(TDPRI_USER_NORM); 1262 } 1263 #endif 1264 } 1265 1266 /* 1267 * Generic schedule. Possibly schedule threads belonging to other cpus and 1268 * deal with threads that might be blocked on a wait queue. 1269 * 1270 * We have a little helper inline function which does additional work after 1271 * the thread has been enqueued, including dealing with preemption and 1272 * setting need_lwkt_resched() (which prevents the kernel from returning 1273 * to userland until it has processed higher priority threads). 1274 * 1275 * It is possible for this routine to be called after a failed _enqueue 1276 * (due to the target thread migrating, sleeping, or otherwise blocked). 1277 * We have to check that the thread is actually on the run queue! 1278 */ 1279 static __inline 1280 void 1281 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount) 1282 { 1283 if (ntd->td_flags & TDF_RUNQ) { 1284 if (ntd->td_preemptable) { 1285 ntd->td_preemptable(ntd, ccount); /* YYY +token */ 1286 } 1287 } 1288 } 1289 1290 static __inline 1291 void 1292 _lwkt_schedule(thread_t td) 1293 { 1294 globaldata_t mygd = mycpu; 1295 1296 KASSERT(td != &td->td_gd->gd_idlethread, 1297 ("lwkt_schedule(): scheduling gd_idlethread is illegal!")); 1298 KKASSERT((td->td_flags & TDF_MIGRATING) == 0); 1299 crit_enter_gd(mygd); 1300 KKASSERT(td->td_lwp == NULL || 1301 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 1302 1303 if (td == mygd->gd_curthread) { 1304 _lwkt_enqueue(td); 1305 } else { 1306 /* 1307 * If we own the thread, there is no race (since we are in a 1308 * critical section). If we do not own the thread there might 1309 * be a race but the target cpu will deal with it. 1310 */ 1311 #ifdef SMP 1312 if (td->td_gd == mygd) { 1313 _lwkt_enqueue(td); 1314 _lwkt_schedule_post(mygd, td, 1); 1315 } else { 1316 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0); 1317 } 1318 #else 1319 _lwkt_enqueue(td); 1320 _lwkt_schedule_post(mygd, td, 1); 1321 #endif 1322 } 1323 crit_exit_gd(mygd); 1324 } 1325 1326 void 1327 lwkt_schedule(thread_t td) 1328 { 1329 _lwkt_schedule(td); 1330 } 1331 1332 void 1333 lwkt_schedule_noresched(thread_t td) /* XXX not impl */ 1334 { 1335 _lwkt_schedule(td); 1336 } 1337 1338 #ifdef SMP 1339 1340 /* 1341 * When scheduled remotely if frame != NULL the IPIQ is being 1342 * run via doreti or an interrupt then preemption can be allowed. 1343 * 1344 * To allow preemption we have to drop the critical section so only 1345 * one is present in _lwkt_schedule_post. 1346 */ 1347 static void 1348 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame) 1349 { 1350 thread_t td = curthread; 1351 thread_t ntd = arg; 1352 1353 if (frame && ntd->td_preemptable) { 1354 crit_exit_noyield(td); 1355 _lwkt_schedule(ntd); 1356 crit_enter_quick(td); 1357 } else { 1358 _lwkt_schedule(ntd); 1359 } 1360 } 1361 1362 /* 1363 * Thread migration using a 'Pull' method. The thread may or may not be 1364 * the current thread. It MUST be descheduled and in a stable state. 1365 * lwkt_giveaway() must be called on the cpu owning the thread. 1366 * 1367 * At any point after lwkt_giveaway() is called, the target cpu may 1368 * 'pull' the thread by calling lwkt_acquire(). 1369 * 1370 * We have to make sure the thread is not sitting on a per-cpu tsleep 1371 * queue or it will blow up when it moves to another cpu. 1372 * 1373 * MPSAFE - must be called under very specific conditions. 1374 */ 1375 void 1376 lwkt_giveaway(thread_t td) 1377 { 1378 globaldata_t gd = mycpu; 1379 1380 crit_enter_gd(gd); 1381 if (td->td_flags & TDF_TSLEEPQ) 1382 tsleep_remove(td); 1383 KKASSERT(td->td_gd == gd); 1384 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); 1385 td->td_flags |= TDF_MIGRATING; 1386 crit_exit_gd(gd); 1387 } 1388 1389 void 1390 lwkt_acquire(thread_t td) 1391 { 1392 globaldata_t gd; 1393 globaldata_t mygd; 1394 int retry = 10000000; 1395 1396 KKASSERT(td->td_flags & TDF_MIGRATING); 1397 gd = td->td_gd; 1398 mygd = mycpu; 1399 if (gd != mycpu) { 1400 cpu_lfence(); 1401 KKASSERT((td->td_flags & TDF_RUNQ) == 0); 1402 crit_enter_gd(mygd); 1403 DEBUG_PUSH_INFO("lwkt_acquire"); 1404 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) { 1405 #ifdef SMP 1406 lwkt_process_ipiq(); 1407 #endif 1408 cpu_lfence(); 1409 if (--retry == 0) { 1410 kprintf("lwkt_acquire: stuck: td %p td->td_flags %08x\n", 1411 td, td->td_flags); 1412 retry = 10000000; 1413 } 1414 } 1415 DEBUG_POP_INFO(); 1416 cpu_mfence(); 1417 td->td_gd = mygd; 1418 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1419 td->td_flags &= ~TDF_MIGRATING; 1420 crit_exit_gd(mygd); 1421 } else { 1422 crit_enter_gd(mygd); 1423 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1424 td->td_flags &= ~TDF_MIGRATING; 1425 crit_exit_gd(mygd); 1426 } 1427 } 1428 1429 #endif 1430 1431 /* 1432 * Generic deschedule. Descheduling threads other then your own should be 1433 * done only in carefully controlled circumstances. Descheduling is 1434 * asynchronous. 1435 * 1436 * This function may block if the cpu has run out of messages. 1437 */ 1438 void 1439 lwkt_deschedule(thread_t td) 1440 { 1441 crit_enter(); 1442 #ifdef SMP 1443 if (td == curthread) { 1444 _lwkt_dequeue(td); 1445 } else { 1446 if (td->td_gd == mycpu) { 1447 _lwkt_dequeue(td); 1448 } else { 1449 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td); 1450 } 1451 } 1452 #else 1453 _lwkt_dequeue(td); 1454 #endif 1455 crit_exit(); 1456 } 1457 1458 /* 1459 * Set the target thread's priority. This routine does not automatically 1460 * switch to a higher priority thread, LWKT threads are not designed for 1461 * continuous priority changes. Yield if you want to switch. 1462 */ 1463 void 1464 lwkt_setpri(thread_t td, int pri) 1465 { 1466 if (td->td_pri != pri) { 1467 KKASSERT(pri >= 0); 1468 crit_enter(); 1469 if (td->td_flags & TDF_RUNQ) { 1470 KKASSERT(td->td_gd == mycpu); 1471 _lwkt_dequeue(td); 1472 td->td_pri = pri; 1473 _lwkt_enqueue(td); 1474 } else { 1475 td->td_pri = pri; 1476 } 1477 crit_exit(); 1478 } 1479 } 1480 1481 /* 1482 * Set the initial priority for a thread prior to it being scheduled for 1483 * the first time. The thread MUST NOT be scheduled before or during 1484 * this call. The thread may be assigned to a cpu other then the current 1485 * cpu. 1486 * 1487 * Typically used after a thread has been created with TDF_STOPPREQ, 1488 * and before the thread is initially scheduled. 1489 */ 1490 void 1491 lwkt_setpri_initial(thread_t td, int pri) 1492 { 1493 KKASSERT(pri >= 0); 1494 KKASSERT((td->td_flags & TDF_RUNQ) == 0); 1495 td->td_pri = pri; 1496 } 1497 1498 void 1499 lwkt_setpri_self(int pri) 1500 { 1501 thread_t td = curthread; 1502 1503 KKASSERT(pri >= 0 && pri <= TDPRI_MAX); 1504 crit_enter(); 1505 if (td->td_flags & TDF_RUNQ) { 1506 _lwkt_dequeue(td); 1507 td->td_pri = pri; 1508 _lwkt_enqueue(td); 1509 } else { 1510 td->td_pri = pri; 1511 } 1512 crit_exit(); 1513 } 1514 1515 /* 1516 * hz tick scheduler clock for LWKT threads 1517 */ 1518 void 1519 lwkt_schedulerclock(thread_t td) 1520 { 1521 globaldata_t gd = td->td_gd; 1522 thread_t xtd; 1523 1524 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) { 1525 /* 1526 * If the current thread is at the head of the runq shift it to the 1527 * end of any equal-priority threads and request a LWKT reschedule 1528 * if it moved. 1529 * 1530 * Ignore upri in this situation. There will only be one user thread 1531 * in user mode, all others will be user threads running in kernel 1532 * mode and we have to make sure they get some cpu. 1533 */ 1534 xtd = TAILQ_NEXT(td, td_threadq); 1535 if (xtd && xtd->td_pri == td->td_pri) { 1536 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq); 1537 while (xtd && xtd->td_pri == td->td_pri) 1538 xtd = TAILQ_NEXT(xtd, td_threadq); 1539 if (xtd) 1540 TAILQ_INSERT_BEFORE(xtd, td, td_threadq); 1541 else 1542 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 1543 need_lwkt_resched(); 1544 } 1545 } else { 1546 /* 1547 * If we scheduled a thread other than the one at the head of the 1548 * queue always request a reschedule every tick. 1549 */ 1550 need_lwkt_resched(); 1551 } 1552 } 1553 1554 /* 1555 * Migrate the current thread to the specified cpu. 1556 * 1557 * This is accomplished by descheduling ourselves from the current cpu 1558 * and setting td_migrate_gd. The lwkt_switch() code will detect that the 1559 * 'old' thread wants to migrate after it has been completely switched out 1560 * and will complete the migration. 1561 * 1562 * TDF_MIGRATING prevents scheduling races while the thread is being migrated. 1563 * 1564 * We must be sure to release our current process designation (if a user 1565 * process) before clearing out any tsleepq we are on because the release 1566 * code may re-add us. 1567 * 1568 * We must be sure to remove ourselves from the current cpu's tsleepq 1569 * before potentially moving to another queue. The thread can be on 1570 * a tsleepq due to a left-over tsleep_interlock(). 1571 */ 1572 1573 void 1574 lwkt_setcpu_self(globaldata_t rgd) 1575 { 1576 #ifdef SMP 1577 thread_t td = curthread; 1578 1579 if (td->td_gd != rgd) { 1580 crit_enter_quick(td); 1581 1582 if (td->td_release) 1583 td->td_release(td); 1584 if (td->td_flags & TDF_TSLEEPQ) 1585 tsleep_remove(td); 1586 1587 /* 1588 * Set TDF_MIGRATING to prevent a spurious reschedule while we are 1589 * trying to deschedule ourselves and switch away, then deschedule 1590 * ourself, remove us from tdallq, and set td_migrate_gd. Finally, 1591 * call lwkt_switch() to complete the operation. 1592 */ 1593 td->td_flags |= TDF_MIGRATING; 1594 lwkt_deschedule_self(td); 1595 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1596 td->td_migrate_gd = rgd; 1597 lwkt_switch(); 1598 1599 /* 1600 * We are now on the target cpu 1601 */ 1602 KKASSERT(rgd == mycpu); 1603 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); 1604 crit_exit_quick(td); 1605 } 1606 #endif 1607 } 1608 1609 void 1610 lwkt_migratecpu(int cpuid) 1611 { 1612 #ifdef SMP 1613 globaldata_t rgd; 1614 1615 rgd = globaldata_find(cpuid); 1616 lwkt_setcpu_self(rgd); 1617 #endif 1618 } 1619 1620 #ifdef SMP 1621 /* 1622 * Remote IPI for cpu migration (called while in a critical section so we 1623 * do not have to enter another one). 1624 * 1625 * The thread (td) has already been completely descheduled from the 1626 * originating cpu and we can simply assert the case. The thread is 1627 * assigned to the new cpu and enqueued. 1628 * 1629 * The thread will re-add itself to tdallq when it resumes execution. 1630 */ 1631 static void 1632 lwkt_setcpu_remote(void *arg) 1633 { 1634 thread_t td = arg; 1635 globaldata_t gd = mycpu; 1636 1637 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0); 1638 td->td_gd = gd; 1639 cpu_mfence(); 1640 td->td_flags &= ~TDF_MIGRATING; 1641 KKASSERT(td->td_migrate_gd == NULL); 1642 KKASSERT(td->td_lwp == NULL || 1643 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 1644 _lwkt_enqueue(td); 1645 } 1646 #endif 1647 1648 struct lwp * 1649 lwkt_preempted_proc(void) 1650 { 1651 thread_t td = curthread; 1652 while (td->td_preempted) 1653 td = td->td_preempted; 1654 return(td->td_lwp); 1655 } 1656 1657 /* 1658 * Create a kernel process/thread/whatever. It shares it's address space 1659 * with proc0 - ie: kernel only. 1660 * 1661 * If the cpu is not specified one will be selected. In the future 1662 * specifying a cpu of -1 will enable kernel thread migration between 1663 * cpus. 1664 */ 1665 int 1666 lwkt_create(void (*func)(void *), void *arg, struct thread **tdp, 1667 thread_t template, int tdflags, int cpu, const char *fmt, ...) 1668 { 1669 thread_t td; 1670 __va_list ap; 1671 1672 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu, 1673 tdflags); 1674 if (tdp) 1675 *tdp = td; 1676 cpu_set_thread_handler(td, lwkt_exit, func, arg); 1677 1678 /* 1679 * Set up arg0 for 'ps' etc 1680 */ 1681 __va_start(ap, fmt); 1682 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap); 1683 __va_end(ap); 1684 1685 /* 1686 * Schedule the thread to run 1687 */ 1688 if (td->td_flags & TDF_NOSTART) 1689 td->td_flags &= ~TDF_NOSTART; 1690 else 1691 lwkt_schedule(td); 1692 return 0; 1693 } 1694 1695 /* 1696 * Destroy an LWKT thread. Warning! This function is not called when 1697 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and 1698 * uses a different reaping mechanism. 1699 */ 1700 void 1701 lwkt_exit(void) 1702 { 1703 thread_t td = curthread; 1704 thread_t std; 1705 globaldata_t gd; 1706 1707 /* 1708 * Do any cleanup that might block here 1709 */ 1710 if (td->td_flags & TDF_VERBOSE) 1711 kprintf("kthread %p %s has exited\n", td, td->td_comm); 1712 caps_exit(td); 1713 biosched_done(td); 1714 dsched_exit_thread(td); 1715 1716 /* 1717 * Get us into a critical section to interlock gd_freetd and loop 1718 * until we can get it freed. 1719 * 1720 * We have to cache the current td in gd_freetd because objcache_put()ing 1721 * it would rip it out from under us while our thread is still active. 1722 * 1723 * We are the current thread so of course our own TDF_RUNNING bit will 1724 * be set, so unlike the lwp reap code we don't wait for it to clear. 1725 */ 1726 gd = mycpu; 1727 crit_enter_quick(td); 1728 for (;;) { 1729 if (td->td_refs) { 1730 tsleep(td, 0, "tdreap", 1); 1731 continue; 1732 } 1733 if ((std = gd->gd_freetd) != NULL) { 1734 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0); 1735 gd->gd_freetd = NULL; 1736 objcache_put(thread_cache, std); 1737 continue; 1738 } 1739 break; 1740 } 1741 1742 /* 1743 * Remove thread resources from kernel lists and deschedule us for 1744 * the last time. We cannot block after this point or we may end 1745 * up with a stale td on the tsleepq. 1746 * 1747 * None of this may block, the critical section is the only thing 1748 * protecting tdallq and the only thing preventing new lwkt_hold() 1749 * thread refs now. 1750 */ 1751 if (td->td_flags & TDF_TSLEEPQ) 1752 tsleep_remove(td); 1753 lwkt_deschedule_self(td); 1754 lwkt_remove_tdallq(td); 1755 KKASSERT(td->td_refs == 0); 1756 1757 /* 1758 * Final cleanup 1759 */ 1760 KKASSERT(gd->gd_freetd == NULL); 1761 if (td->td_flags & TDF_ALLOCATED_THREAD) 1762 gd->gd_freetd = td; 1763 cpu_thread_exit(); 1764 } 1765 1766 void 1767 lwkt_remove_tdallq(thread_t td) 1768 { 1769 KKASSERT(td->td_gd == mycpu); 1770 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1771 } 1772 1773 /* 1774 * Code reduction and branch prediction improvements. Call/return 1775 * overhead on modern cpus often degenerates into 0 cycles due to 1776 * the cpu's branch prediction hardware and return pc cache. We 1777 * can take advantage of this by not inlining medium-complexity 1778 * functions and we can also reduce the branch prediction impact 1779 * by collapsing perfectly predictable branches into a single 1780 * procedure instead of duplicating it. 1781 * 1782 * Is any of this noticeable? Probably not, so I'll take the 1783 * smaller code size. 1784 */ 1785 void 1786 crit_exit_wrapper(__DEBUG_CRIT_ARG__) 1787 { 1788 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__); 1789 } 1790 1791 void 1792 crit_panic(void) 1793 { 1794 thread_t td = curthread; 1795 int lcrit = td->td_critcount; 1796 1797 td->td_critcount = 0; 1798 panic("td_critcount is/would-go negative! %p %d", td, lcrit); 1799 /* NOT REACHED */ 1800 } 1801 1802 #ifdef SMP 1803 1804 /* 1805 * Called from debugger/panic on cpus which have been stopped. We must still 1806 * process the IPIQ while stopped, even if we were stopped while in a critical 1807 * section (XXX). 1808 * 1809 * If we are dumping also try to process any pending interrupts. This may 1810 * or may not work depending on the state of the cpu at the point it was 1811 * stopped. 1812 */ 1813 void 1814 lwkt_smp_stopped(void) 1815 { 1816 globaldata_t gd = mycpu; 1817 1818 crit_enter_gd(gd); 1819 if (dumping) { 1820 lwkt_process_ipiq(); 1821 splz(); 1822 } else { 1823 lwkt_process_ipiq(); 1824 } 1825 crit_exit_gd(gd); 1826 } 1827 1828 #endif 1829