1 /* 2 * Copyright (c) 2003-2011 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@backplane.com> 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * 3. Neither the name of The DragonFly Project nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific, prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 */ 34 35 /* 36 * Each cpu in a system has its own self-contained light weight kernel 37 * thread scheduler, which means that generally speaking we only need 38 * to use a critical section to avoid problems. Foreign thread 39 * scheduling is queued via (async) IPIs. 40 */ 41 42 #include <sys/param.h> 43 #include <sys/systm.h> 44 #include <sys/kernel.h> 45 #include <sys/proc.h> 46 #include <sys/rtprio.h> 47 #include <sys/kinfo.h> 48 #include <sys/queue.h> 49 #include <sys/sysctl.h> 50 #include <sys/kthread.h> 51 #include <machine/cpu.h> 52 #include <sys/lock.h> 53 #include <sys/spinlock.h> 54 #include <sys/ktr.h> 55 56 #include <sys/thread2.h> 57 #include <sys/spinlock2.h> 58 #include <sys/mplock2.h> 59 60 #include <sys/dsched.h> 61 62 #include <vm/vm.h> 63 #include <vm/vm_param.h> 64 #include <vm/vm_kern.h> 65 #include <vm/vm_object.h> 66 #include <vm/vm_page.h> 67 #include <vm/vm_map.h> 68 #include <vm/vm_pager.h> 69 #include <vm/vm_extern.h> 70 71 #include <machine/stdarg.h> 72 #include <machine/smp.h> 73 74 #if !defined(KTR_CTXSW) 75 #define KTR_CTXSW KTR_ALL 76 #endif 77 KTR_INFO_MASTER(ctxsw); 78 KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p", int cpu, struct thread *td); 79 KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p", int cpu, struct thread *td); 80 KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s", struct thread *td, char *comm); 81 KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", struct thread *td); 82 83 static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads"); 84 85 #ifdef INVARIANTS 86 static int panic_on_cscount = 0; 87 #endif 88 static __int64_t switch_count = 0; 89 static __int64_t preempt_hit = 0; 90 static __int64_t preempt_miss = 0; 91 static __int64_t preempt_weird = 0; 92 static int lwkt_use_spin_port; 93 static struct objcache *thread_cache; 94 int cpu_mwait_spin = 0; 95 96 static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame); 97 static void lwkt_setcpu_remote(void *arg); 98 99 extern void cpu_heavy_restore(void); 100 extern void cpu_lwkt_restore(void); 101 extern void cpu_kthread_restore(void); 102 extern void cpu_idle_restore(void); 103 104 /* 105 * We can make all thread ports use the spin backend instead of the thread 106 * backend. This should only be set to debug the spin backend. 107 */ 108 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port); 109 110 #ifdef INVARIANTS 111 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, 112 "Panic if attempting to switch lwkt's while mastering cpusync"); 113 #endif 114 SYSCTL_INT(_hw, OID_AUTO, cpu_mwait_spin, CTLFLAG_RW, &cpu_mwait_spin, 0, 115 "monitor/mwait target state"); 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 0, 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 static void 411 lwkt_init_thread_remote(void *arg) 412 { 413 thread_t td = arg; 414 415 /* 416 * Protected by critical section held by IPI dispatch 417 */ 418 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq); 419 } 420 421 /* 422 * lwkt core thread structural initialization. 423 * 424 * NOTE: All threads are initialized as mpsafe threads. 425 */ 426 void 427 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags, 428 struct globaldata *gd) 429 { 430 globaldata_t mygd = mycpu; 431 432 bzero(td, sizeof(struct thread)); 433 td->td_kstack = stack; 434 td->td_kstack_size = stksize; 435 td->td_flags = flags; 436 td->td_mpflags = 0; 437 td->td_type = TD_TYPE_GENERIC; 438 td->td_gd = gd; 439 td->td_pri = TDPRI_KERN_DAEMON; 440 td->td_critcount = 1; 441 td->td_toks_have = NULL; 442 td->td_toks_stop = &td->td_toks_base; 443 if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT)) 444 lwkt_initport_spin(&td->td_msgport, td); 445 else 446 lwkt_initport_thread(&td->td_msgport, td); 447 pmap_init_thread(td); 448 /* 449 * Normally initializing a thread for a remote cpu requires sending an 450 * IPI. However, the idlethread is setup before the other cpus are 451 * activated so we have to treat it as a special case. XXX manipulation 452 * of gd_tdallq requires the BGL. 453 */ 454 if (gd == mygd || td == &gd->gd_idlethread) { 455 crit_enter_gd(mygd); 456 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); 457 crit_exit_gd(mygd); 458 } else { 459 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td); 460 } 461 dsched_new_thread(td); 462 } 463 464 void 465 lwkt_set_comm(thread_t td, const char *ctl, ...) 466 { 467 __va_list va; 468 469 __va_start(va, ctl); 470 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va); 471 __va_end(va); 472 KTR_LOG(ctxsw_newtd, td, td->td_comm); 473 } 474 475 /* 476 * Prevent the thread from getting destroyed. Note that unlike PHOLD/PRELE 477 * this does not prevent the thread from migrating to another cpu so the 478 * gd_tdallq state is not protected by this. 479 */ 480 void 481 lwkt_hold(thread_t td) 482 { 483 atomic_add_int(&td->td_refs, 1); 484 } 485 486 void 487 lwkt_rele(thread_t td) 488 { 489 KKASSERT(td->td_refs > 0); 490 atomic_add_int(&td->td_refs, -1); 491 } 492 493 void 494 lwkt_free_thread(thread_t td) 495 { 496 KKASSERT(td->td_refs == 0); 497 KKASSERT((td->td_flags & (TDF_RUNNING | TDF_PREEMPT_LOCK | 498 TDF_RUNQ | TDF_TSLEEPQ)) == 0); 499 if (td->td_flags & TDF_ALLOCATED_THREAD) { 500 objcache_put(thread_cache, td); 501 } else if (td->td_flags & TDF_ALLOCATED_STACK) { 502 /* client-allocated struct with internally allocated stack */ 503 KASSERT(td->td_kstack && td->td_kstack_size > 0, 504 ("lwkt_free_thread: corrupted stack")); 505 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size); 506 td->td_kstack = NULL; 507 td->td_kstack_size = 0; 508 } 509 KTR_LOG(ctxsw_deadtd, td); 510 } 511 512 513 /* 514 * Switch to the next runnable lwkt. If no LWKTs are runnable then 515 * switch to the idlethread. Switching must occur within a critical 516 * section to avoid races with the scheduling queue. 517 * 518 * We always have full control over our cpu's run queue. Other cpus 519 * that wish to manipulate our queue must use the cpu_*msg() calls to 520 * talk to our cpu, so a critical section is all that is needed and 521 * the result is very, very fast thread switching. 522 * 523 * The LWKT scheduler uses a fixed priority model and round-robins at 524 * each priority level. User process scheduling is a totally 525 * different beast and LWKT priorities should not be confused with 526 * user process priorities. 527 * 528 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch() 529 * is not called by the current thread in the preemption case, only when 530 * the preempting thread blocks (in order to return to the original thread). 531 * 532 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread 533 * migration and tsleep deschedule the current lwkt thread and call 534 * lwkt_switch(). In particular, the target cpu of the migration fully 535 * expects the thread to become non-runnable and can deadlock against 536 * cpusync operations if we run any IPIs prior to switching the thread out. 537 * 538 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF 539 * THE CURRENT THREAD HAS BEEN DESCHEDULED! 540 */ 541 void 542 lwkt_switch(void) 543 { 544 globaldata_t gd = mycpu; 545 thread_t td = gd->gd_curthread; 546 thread_t ntd; 547 int spinning = 0; 548 549 KKASSERT(gd->gd_processing_ipiq == 0); 550 KKASSERT(td->td_flags & TDF_RUNNING); 551 552 /* 553 * Switching from within a 'fast' (non thread switched) interrupt or IPI 554 * is illegal. However, we may have to do it anyway if we hit a fatal 555 * kernel trap or we have paniced. 556 * 557 * If this case occurs save and restore the interrupt nesting level. 558 */ 559 if (gd->gd_intr_nesting_level) { 560 int savegdnest; 561 int savegdtrap; 562 563 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) { 564 panic("lwkt_switch: Attempt to switch from a " 565 "fast interrupt, ipi, or hard code section, " 566 "td %p\n", 567 td); 568 } else { 569 savegdnest = gd->gd_intr_nesting_level; 570 savegdtrap = gd->gd_trap_nesting_level; 571 gd->gd_intr_nesting_level = 0; 572 gd->gd_trap_nesting_level = 0; 573 if ((td->td_flags & TDF_PANICWARN) == 0) { 574 td->td_flags |= TDF_PANICWARN; 575 kprintf("Warning: thread switch from interrupt, IPI, " 576 "or hard code section.\n" 577 "thread %p (%s)\n", td, td->td_comm); 578 print_backtrace(-1); 579 } 580 lwkt_switch(); 581 gd->gd_intr_nesting_level = savegdnest; 582 gd->gd_trap_nesting_level = savegdtrap; 583 return; 584 } 585 } 586 587 /* 588 * Release our current user process designation if we are blocking 589 * or if a user reschedule was requested. 590 * 591 * NOTE: This function is NOT called if we are switching into or 592 * returning from a preemption. 593 * 594 * NOTE: Releasing our current user process designation may cause 595 * it to be assigned to another thread, which in turn will 596 * cause us to block in the usched acquire code when we attempt 597 * to return to userland. 598 * 599 * NOTE: On SMP systems this can be very nasty when heavy token 600 * contention is present so we want to be careful not to 601 * release the designation gratuitously. 602 */ 603 if (td->td_release && 604 (user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) { 605 td->td_release(td); 606 } 607 608 /* 609 * Release all tokens 610 */ 611 crit_enter_gd(gd); 612 if (TD_TOKS_HELD(td)) 613 lwkt_relalltokens(td); 614 615 /* 616 * We had better not be holding any spin locks, but don't get into an 617 * endless panic loop. 618 */ 619 KASSERT(gd->gd_spinlocks == 0 || panicstr != NULL, 620 ("lwkt_switch: still holding %d exclusive spinlocks!", 621 gd->gd_spinlocks)); 622 623 624 #ifdef INVARIANTS 625 if (td->td_cscount) { 626 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n", 627 td); 628 if (panic_on_cscount) 629 panic("switching while mastering cpusync"); 630 } 631 #endif 632 633 /* 634 * If we had preempted another thread on this cpu, resume the preempted 635 * thread. This occurs transparently, whether the preempted thread 636 * was scheduled or not (it may have been preempted after descheduling 637 * itself). 638 * 639 * We have to setup the MP lock for the original thread after backing 640 * out the adjustment that was made to curthread when the original 641 * was preempted. 642 */ 643 if ((ntd = td->td_preempted) != NULL) { 644 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK); 645 ntd->td_flags |= TDF_PREEMPT_DONE; 646 647 /* 648 * The interrupt may have woken a thread up, we need to properly 649 * set the reschedule flag if the originally interrupted thread is 650 * at a lower priority. 651 * 652 * The interrupt may not have descheduled. 653 */ 654 if (TAILQ_FIRST(&gd->gd_tdrunq) != ntd) 655 need_lwkt_resched(); 656 goto havethread_preempted; 657 } 658 659 /* 660 * If we cannot obtain ownership of the tokens we cannot immediately 661 * schedule the target thread. 662 * 663 * Reminder: Again, we cannot afford to run any IPIs in this path if 664 * the current thread has been descheduled. 665 */ 666 for (;;) { 667 clear_lwkt_resched(); 668 669 /* 670 * Hotpath - pull the head of the run queue and attempt to schedule 671 * it. 672 */ 673 ntd = TAILQ_FIRST(&gd->gd_tdrunq); 674 675 if (ntd == NULL) { 676 /* 677 * Runq is empty, switch to idle to allow it to halt. 678 */ 679 ntd = &gd->gd_idlethread; 680 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) 681 ASSERT_NO_TOKENS_HELD(ntd); 682 cpu_time.cp_msg[0] = 0; 683 cpu_time.cp_stallpc = 0; 684 goto haveidle; 685 } 686 687 /* 688 * Hotpath - schedule ntd. 689 * 690 * NOTE: For UP there is no mplock and lwkt_getalltokens() 691 * always succeeds. 692 */ 693 if (TD_TOKS_NOT_HELD(ntd) || 694 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops))) 695 { 696 goto havethread; 697 } 698 699 /* 700 * Coldpath (SMP only since tokens always succeed on UP) 701 * 702 * We had some contention on the thread we wanted to schedule. 703 * What we do now is try to find a thread that we can schedule 704 * in its stead. 705 * 706 * The coldpath scan does NOT rearrange threads in the run list. 707 * The lwkt_schedulerclock() will assert need_lwkt_resched() on 708 * the next tick whenever the current head is not the current thread. 709 */ 710 #ifdef INVARIANTS 711 ++ntd->td_contended; 712 #endif 713 ++gd->gd_cnt.v_token_colls; 714 715 if (fairq_bypass > 0) 716 goto skip; 717 718 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) { 719 #ifndef NO_LWKT_SPLIT_USERPRI 720 /* 721 * Never schedule threads returning to userland or the 722 * user thread scheduler helper thread when higher priority 723 * threads are present. The runq is sorted by priority 724 * so we can give up traversing it when we find the first 725 * low priority thread. 726 */ 727 if (ntd->td_pri < TDPRI_KERN_LPSCHED) { 728 ntd = NULL; 729 break; 730 } 731 #endif 732 733 /* 734 * Try this one. 735 */ 736 if (TD_TOKS_NOT_HELD(ntd) || 737 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops))) { 738 goto havethread; 739 } 740 #ifdef INVARIANTS 741 ++ntd->td_contended; 742 #endif 743 ++gd->gd_cnt.v_token_colls; 744 } 745 746 skip: 747 /* 748 * We exhausted the run list, meaning that all runnable threads 749 * are contested. 750 */ 751 cpu_pause(); 752 ntd = &gd->gd_idlethread; 753 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) 754 ASSERT_NO_TOKENS_HELD(ntd); 755 /* contention case, do not clear contention mask */ 756 757 /* 758 * We are going to have to retry but if the current thread is not 759 * on the runq we instead switch through the idle thread to get away 760 * from the current thread. We have to flag for lwkt reschedule 761 * to prevent the idle thread from halting. 762 * 763 * NOTE: A non-zero spinning is passed to lwkt_getalltokens() to 764 * instruct it to deal with the potential for deadlocks by 765 * ordering the tokens by address. 766 */ 767 if ((td->td_flags & TDF_RUNQ) == 0) { 768 need_lwkt_resched(); /* prevent hlt */ 769 goto haveidle; 770 } 771 #if defined(INVARIANTS) && defined(__x86_64__) 772 if ((read_rflags() & PSL_I) == 0) { 773 cpu_enable_intr(); 774 panic("lwkt_switch() called with interrupts disabled"); 775 } 776 #endif 777 778 /* 779 * Number iterations so far. After a certain point we switch to 780 * a sorted-address/monitor/mwait version of lwkt_getalltokens() 781 */ 782 if (spinning < 0x7FFFFFFF) 783 ++spinning; 784 785 /* 786 * lwkt_getalltokens() failed in sorted token mode, we can use 787 * monitor/mwait in this case. 788 */ 789 if (spinning >= lwkt_spin_loops && 790 (cpu_mi_feature & CPU_MI_MONITOR) && 791 lwkt_spin_monitor) 792 { 793 cpu_mmw_pause_int(&gd->gd_reqflags, 794 (gd->gd_reqflags | RQF_SPINNING) & 795 ~RQF_IDLECHECK_WK_MASK, 796 cpu_mwait_spin); 797 } 798 799 /* 800 * We already checked that td is still scheduled so this should be 801 * safe. 802 */ 803 splz_check(); 804 805 /* 806 * This experimental resequencer is used as a fall-back to reduce 807 * hw cache line contention by placing each core's scheduler into a 808 * time-domain-multplexed slot. 809 * 810 * The resequencer is disabled by default. It's functionality has 811 * largely been superceeded by the token algorithm which limits races 812 * to a subset of cores. 813 * 814 * The resequencer algorithm tends to break down when more than 815 * 20 cores are contending. What appears to happen is that new 816 * tokens can be obtained out of address-sorted order by new cores 817 * while existing cores languish in long delays between retries and 818 * wind up being starved-out of the token acquisition. 819 */ 820 if (lwkt_spin_reseq && spinning >= lwkt_spin_reseq) { 821 int cseq = atomic_fetchadd_int(&lwkt_cseq_windex, 1); 822 int oseq; 823 824 while ((oseq = lwkt_cseq_rindex) != cseq) { 825 cpu_ccfence(); 826 #if 1 827 if (cpu_mi_feature & CPU_MI_MONITOR) { 828 cpu_mmw_pause_int(&lwkt_cseq_rindex, oseq, cpu_mwait_spin); 829 } else { 830 #endif 831 cpu_pause(); 832 cpu_lfence(); 833 #if 1 834 } 835 #endif 836 } 837 DELAY(1); 838 atomic_add_int(&lwkt_cseq_rindex, 1); 839 } 840 /* highest level for(;;) loop */ 841 } 842 843 havethread: 844 /* 845 * Clear gd_idle_repeat when doing a normal switch to a non-idle 846 * thread. 847 */ 848 ntd->td_wmesg = NULL; 849 ++gd->gd_cnt.v_swtch; 850 gd->gd_idle_repeat = 0; 851 852 havethread_preempted: 853 /* 854 * If the new target does not need the MP lock and we are holding it, 855 * release the MP lock. If the new target requires the MP lock we have 856 * already acquired it for the target. 857 */ 858 ; 859 haveidle: 860 KASSERT(ntd->td_critcount, 861 ("priority problem in lwkt_switch %d %d", 862 td->td_critcount, ntd->td_critcount)); 863 864 if (td != ntd) { 865 /* 866 * Execute the actual thread switch operation. This function 867 * returns to the current thread and returns the previous thread 868 * (which may be different from the thread we switched to). 869 * 870 * We are responsible for marking ntd as TDF_RUNNING. 871 */ 872 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0); 873 ++switch_count; 874 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd); 875 ntd->td_flags |= TDF_RUNNING; 876 lwkt_switch_return(td->td_switch(ntd)); 877 /* ntd invalid, td_switch() can return a different thread_t */ 878 } 879 880 /* 881 * catch-all. XXX is this strictly needed? 882 */ 883 splz_check(); 884 885 /* NOTE: current cpu may have changed after switch */ 886 crit_exit_quick(td); 887 } 888 889 /* 890 * Called by assembly in the td_switch (thread restore path) for thread 891 * bootstrap cases which do not 'return' to lwkt_switch(). 892 */ 893 void 894 lwkt_switch_return(thread_t otd) 895 { 896 globaldata_t rgd; 897 898 /* 899 * Check if otd was migrating. Now that we are on ntd we can finish 900 * up the migration. This is a bit messy but it is the only place 901 * where td is known to be fully descheduled. 902 * 903 * We can only activate the migration if otd was migrating but not 904 * held on the cpu due to a preemption chain. We still have to 905 * clear TDF_RUNNING on the old thread either way. 906 * 907 * We are responsible for clearing the previously running thread's 908 * TDF_RUNNING. 909 */ 910 if ((rgd = otd->td_migrate_gd) != NULL && 911 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) { 912 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) == 913 (TDF_MIGRATING | TDF_RUNNING)); 914 otd->td_migrate_gd = NULL; 915 otd->td_flags &= ~TDF_RUNNING; 916 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd); 917 } else { 918 otd->td_flags &= ~TDF_RUNNING; 919 } 920 921 /* 922 * Final exit validations (see lwp_wait()). Note that otd becomes 923 * invalid the *instant* we set TDF_MP_EXITSIG. 924 */ 925 while (otd->td_flags & TDF_EXITING) { 926 u_int mpflags; 927 928 mpflags = otd->td_mpflags; 929 cpu_ccfence(); 930 931 if (mpflags & TDF_MP_EXITWAIT) { 932 if (atomic_cmpset_int(&otd->td_mpflags, mpflags, 933 mpflags | TDF_MP_EXITSIG)) { 934 wakeup(otd); 935 break; 936 } 937 } else { 938 if (atomic_cmpset_int(&otd->td_mpflags, mpflags, 939 mpflags | TDF_MP_EXITSIG)) { 940 wakeup(otd); 941 break; 942 } 943 } 944 } 945 } 946 947 /* 948 * Request that the target thread preempt the current thread. Preemption 949 * can only occur if our only critical section is the one that we were called 950 * with, the relative priority of the target thread is higher, and the target 951 * thread holds no tokens. This also only works if we are not holding any 952 * spinlocks (obviously). 953 * 954 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically 955 * this is called via lwkt_schedule() through the td_preemptable callback. 956 * critcount is the managed critical priority that we should ignore in order 957 * to determine whether preemption is possible (aka usually just the crit 958 * priority of lwkt_schedule() itself). 959 * 960 * Preemption is typically limited to interrupt threads. 961 * 962 * Operation works in a fairly straight-forward manner. The normal 963 * scheduling code is bypassed and we switch directly to the target 964 * thread. When the target thread attempts to block or switch away 965 * code at the base of lwkt_switch() will switch directly back to our 966 * thread. Our thread is able to retain whatever tokens it holds and 967 * if the target needs one of them the target will switch back to us 968 * and reschedule itself normally. 969 */ 970 void 971 lwkt_preempt(thread_t ntd, int critcount) 972 { 973 struct globaldata *gd = mycpu; 974 thread_t xtd; 975 thread_t td; 976 int save_gd_intr_nesting_level; 977 978 /* 979 * The caller has put us in a critical section. We can only preempt 980 * if the caller of the caller was not in a critical section (basically 981 * a local interrupt), as determined by the 'critcount' parameter. We 982 * also can't preempt if the caller is holding any spinlocks (even if 983 * he isn't in a critical section). This also handles the tokens test. 984 * 985 * YYY The target thread must be in a critical section (else it must 986 * inherit our critical section? I dunno yet). 987 */ 988 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri)); 989 990 td = gd->gd_curthread; 991 if (preempt_enable == 0) { 992 ++preempt_miss; 993 return; 994 } 995 if (ntd->td_pri <= td->td_pri) { 996 ++preempt_miss; 997 return; 998 } 999 if (td->td_critcount > critcount) { 1000 ++preempt_miss; 1001 return; 1002 } 1003 if (td->td_cscount) { 1004 ++preempt_miss; 1005 return; 1006 } 1007 if (ntd->td_gd != gd) { 1008 ++preempt_miss; 1009 return; 1010 } 1011 /* 1012 * We don't have to check spinlocks here as they will also bump 1013 * td_critcount. 1014 * 1015 * Do not try to preempt if the target thread is holding any tokens. 1016 * We could try to acquire the tokens but this case is so rare there 1017 * is no need to support it. 1018 */ 1019 KKASSERT(gd->gd_spinlocks == 0); 1020 1021 if (TD_TOKS_HELD(ntd)) { 1022 ++preempt_miss; 1023 return; 1024 } 1025 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) { 1026 ++preempt_weird; 1027 return; 1028 } 1029 if (ntd->td_preempted) { 1030 ++preempt_hit; 1031 return; 1032 } 1033 KKASSERT(gd->gd_processing_ipiq == 0); 1034 1035 /* 1036 * Since we are able to preempt the current thread, there is no need to 1037 * call need_lwkt_resched(). 1038 * 1039 * We must temporarily clear gd_intr_nesting_level around the switch 1040 * since switchouts from the target thread are allowed (they will just 1041 * return to our thread), and since the target thread has its own stack. 1042 * 1043 * A preemption must switch back to the original thread, assert the 1044 * case. 1045 */ 1046 ++preempt_hit; 1047 ntd->td_preempted = td; 1048 td->td_flags |= TDF_PREEMPT_LOCK; 1049 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd); 1050 save_gd_intr_nesting_level = gd->gd_intr_nesting_level; 1051 gd->gd_intr_nesting_level = 0; 1052 1053 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0); 1054 ntd->td_flags |= TDF_RUNNING; 1055 xtd = td->td_switch(ntd); 1056 KKASSERT(xtd == ntd); 1057 lwkt_switch_return(xtd); 1058 gd->gd_intr_nesting_level = save_gd_intr_nesting_level; 1059 1060 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE)); 1061 ntd->td_preempted = NULL; 1062 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE); 1063 } 1064 1065 /* 1066 * Conditionally call splz() if gd_reqflags indicates work is pending. 1067 * This will work inside a critical section but not inside a hard code 1068 * section. 1069 * 1070 * (self contained on a per cpu basis) 1071 */ 1072 void 1073 splz_check(void) 1074 { 1075 globaldata_t gd = mycpu; 1076 thread_t td = gd->gd_curthread; 1077 1078 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && 1079 gd->gd_intr_nesting_level == 0 && 1080 td->td_nest_count < 2) 1081 { 1082 splz(); 1083 } 1084 } 1085 1086 /* 1087 * This version is integrated into crit_exit, reqflags has already 1088 * been tested but td_critcount has not. 1089 * 1090 * We only want to execute the splz() on the 1->0 transition of 1091 * critcount and not in a hard code section or if too deeply nested. 1092 * 1093 * NOTE: gd->gd_spinlocks is implied to be 0 when td_critcount is 0. 1094 */ 1095 void 1096 lwkt_maybe_splz(thread_t td) 1097 { 1098 globaldata_t gd = td->td_gd; 1099 1100 if (td->td_critcount == 0 && 1101 gd->gd_intr_nesting_level == 0 && 1102 td->td_nest_count < 2) 1103 { 1104 splz(); 1105 } 1106 } 1107 1108 /* 1109 * Drivers which set up processing co-threads can call this function to 1110 * run the co-thread at a higher priority and to allow it to preempt 1111 * normal threads. 1112 */ 1113 void 1114 lwkt_set_interrupt_support_thread(void) 1115 { 1116 thread_t td = curthread; 1117 1118 lwkt_setpri_self(TDPRI_INT_SUPPORT); 1119 td->td_flags |= TDF_INTTHREAD; 1120 td->td_preemptable = lwkt_preempt; 1121 } 1122 1123 1124 /* 1125 * This function is used to negotiate a passive release of the current 1126 * process/lwp designation with the user scheduler, allowing the user 1127 * scheduler to schedule another user thread. The related kernel thread 1128 * (curthread) continues running in the released state. 1129 */ 1130 void 1131 lwkt_passive_release(struct thread *td) 1132 { 1133 struct lwp *lp = td->td_lwp; 1134 1135 #ifndef NO_LWKT_SPLIT_USERPRI 1136 td->td_release = NULL; 1137 lwkt_setpri_self(TDPRI_KERN_USER); 1138 #endif 1139 1140 lp->lwp_proc->p_usched->release_curproc(lp); 1141 } 1142 1143 1144 /* 1145 * This implements a LWKT yield, allowing a kernel thread to yield to other 1146 * kernel threads at the same or higher priority. This function can be 1147 * called in a tight loop and will typically only yield once per tick. 1148 * 1149 * Most kernel threads run at the same priority in order to allow equal 1150 * sharing. 1151 * 1152 * (self contained on a per cpu basis) 1153 */ 1154 void 1155 lwkt_yield(void) 1156 { 1157 globaldata_t gd = mycpu; 1158 thread_t td = gd->gd_curthread; 1159 1160 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1161 splz(); 1162 if (lwkt_resched_wanted()) { 1163 lwkt_schedule_self(curthread); 1164 lwkt_switch(); 1165 } 1166 } 1167 1168 /* 1169 * The quick version processes pending interrupts and higher-priority 1170 * LWKT threads but will not round-robin same-priority LWKT threads. 1171 * 1172 * When called while attempting to return to userland the only same-pri 1173 * threads are the ones which have already tried to become the current 1174 * user process. 1175 */ 1176 void 1177 lwkt_yield_quick(void) 1178 { 1179 globaldata_t gd = mycpu; 1180 thread_t td = gd->gd_curthread; 1181 1182 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1183 splz(); 1184 if (lwkt_resched_wanted()) { 1185 crit_enter(); 1186 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) { 1187 clear_lwkt_resched(); 1188 } else { 1189 lwkt_schedule_self(curthread); 1190 lwkt_switch(); 1191 } 1192 crit_exit(); 1193 } 1194 } 1195 1196 /* 1197 * This yield is designed for kernel threads with a user context. 1198 * 1199 * The kernel acting on behalf of the user is potentially cpu-bound, 1200 * this function will efficiently allow other threads to run and also 1201 * switch to other processes by releasing. 1202 * 1203 * The lwkt_user_yield() function is designed to have very low overhead 1204 * if no yield is determined to be needed. 1205 */ 1206 void 1207 lwkt_user_yield(void) 1208 { 1209 globaldata_t gd = mycpu; 1210 thread_t td = gd->gd_curthread; 1211 1212 /* 1213 * Always run any pending interrupts in case we are in a critical 1214 * section. 1215 */ 1216 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1217 splz(); 1218 1219 /* 1220 * Switch (which forces a release) if another kernel thread needs 1221 * the cpu, if userland wants us to resched, or if our kernel 1222 * quantum has run out. 1223 */ 1224 if (lwkt_resched_wanted() || 1225 user_resched_wanted()) 1226 { 1227 lwkt_switch(); 1228 } 1229 1230 #if 0 1231 /* 1232 * Reacquire the current process if we are released. 1233 * 1234 * XXX not implemented atm. The kernel may be holding locks and such, 1235 * so we want the thread to continue to receive cpu. 1236 */ 1237 if (td->td_release == NULL && lp) { 1238 lp->lwp_proc->p_usched->acquire_curproc(lp); 1239 td->td_release = lwkt_passive_release; 1240 lwkt_setpri_self(TDPRI_USER_NORM); 1241 } 1242 #endif 1243 } 1244 1245 /* 1246 * Generic schedule. Possibly schedule threads belonging to other cpus and 1247 * deal with threads that might be blocked on a wait queue. 1248 * 1249 * We have a little helper inline function which does additional work after 1250 * the thread has been enqueued, including dealing with preemption and 1251 * setting need_lwkt_resched() (which prevents the kernel from returning 1252 * to userland until it has processed higher priority threads). 1253 * 1254 * It is possible for this routine to be called after a failed _enqueue 1255 * (due to the target thread migrating, sleeping, or otherwise blocked). 1256 * We have to check that the thread is actually on the run queue! 1257 */ 1258 static __inline 1259 void 1260 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount) 1261 { 1262 if (ntd->td_flags & TDF_RUNQ) { 1263 if (ntd->td_preemptable) { 1264 ntd->td_preemptable(ntd, ccount); /* YYY +token */ 1265 } 1266 } 1267 } 1268 1269 static __inline 1270 void 1271 _lwkt_schedule(thread_t td) 1272 { 1273 globaldata_t mygd = mycpu; 1274 1275 KASSERT(td != &td->td_gd->gd_idlethread, 1276 ("lwkt_schedule(): scheduling gd_idlethread is illegal!")); 1277 KKASSERT((td->td_flags & TDF_MIGRATING) == 0); 1278 crit_enter_gd(mygd); 1279 KKASSERT(td->td_lwp == NULL || 1280 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 1281 1282 if (td == mygd->gd_curthread) { 1283 _lwkt_enqueue(td); 1284 } else { 1285 /* 1286 * If we own the thread, there is no race (since we are in a 1287 * critical section). If we do not own the thread there might 1288 * be a race but the target cpu will deal with it. 1289 */ 1290 if (td->td_gd == mygd) { 1291 _lwkt_enqueue(td); 1292 _lwkt_schedule_post(mygd, td, 1); 1293 } else { 1294 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0); 1295 } 1296 } 1297 crit_exit_gd(mygd); 1298 } 1299 1300 void 1301 lwkt_schedule(thread_t td) 1302 { 1303 _lwkt_schedule(td); 1304 } 1305 1306 void 1307 lwkt_schedule_noresched(thread_t td) /* XXX not impl */ 1308 { 1309 _lwkt_schedule(td); 1310 } 1311 1312 /* 1313 * When scheduled remotely if frame != NULL the IPIQ is being 1314 * run via doreti or an interrupt then preemption can be allowed. 1315 * 1316 * To allow preemption we have to drop the critical section so only 1317 * one is present in _lwkt_schedule_post. 1318 */ 1319 static void 1320 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame) 1321 { 1322 thread_t td = curthread; 1323 thread_t ntd = arg; 1324 1325 if (frame && ntd->td_preemptable) { 1326 crit_exit_noyield(td); 1327 _lwkt_schedule(ntd); 1328 crit_enter_quick(td); 1329 } else { 1330 _lwkt_schedule(ntd); 1331 } 1332 } 1333 1334 /* 1335 * Thread migration using a 'Pull' method. The thread may or may not be 1336 * the current thread. It MUST be descheduled and in a stable state. 1337 * lwkt_giveaway() must be called on the cpu owning the thread. 1338 * 1339 * At any point after lwkt_giveaway() is called, the target cpu may 1340 * 'pull' the thread by calling lwkt_acquire(). 1341 * 1342 * We have to make sure the thread is not sitting on a per-cpu tsleep 1343 * queue or it will blow up when it moves to another cpu. 1344 * 1345 * MPSAFE - must be called under very specific conditions. 1346 */ 1347 void 1348 lwkt_giveaway(thread_t td) 1349 { 1350 globaldata_t gd = mycpu; 1351 1352 crit_enter_gd(gd); 1353 if (td->td_flags & TDF_TSLEEPQ) 1354 tsleep_remove(td); 1355 KKASSERT(td->td_gd == gd); 1356 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); 1357 td->td_flags |= TDF_MIGRATING; 1358 crit_exit_gd(gd); 1359 } 1360 1361 void 1362 lwkt_acquire(thread_t td) 1363 { 1364 globaldata_t gd; 1365 globaldata_t mygd; 1366 int retry = 10000000; 1367 1368 KKASSERT(td->td_flags & TDF_MIGRATING); 1369 gd = td->td_gd; 1370 mygd = mycpu; 1371 if (gd != mycpu) { 1372 cpu_lfence(); 1373 KKASSERT((td->td_flags & TDF_RUNQ) == 0); 1374 crit_enter_gd(mygd); 1375 DEBUG_PUSH_INFO("lwkt_acquire"); 1376 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) { 1377 lwkt_process_ipiq(); 1378 cpu_lfence(); 1379 if (--retry == 0) { 1380 kprintf("lwkt_acquire: stuck: td %p td->td_flags %08x\n", 1381 td, td->td_flags); 1382 retry = 10000000; 1383 } 1384 } 1385 DEBUG_POP_INFO(); 1386 cpu_mfence(); 1387 td->td_gd = mygd; 1388 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1389 td->td_flags &= ~TDF_MIGRATING; 1390 crit_exit_gd(mygd); 1391 } else { 1392 crit_enter_gd(mygd); 1393 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1394 td->td_flags &= ~TDF_MIGRATING; 1395 crit_exit_gd(mygd); 1396 } 1397 } 1398 1399 /* 1400 * Generic deschedule. Descheduling threads other then your own should be 1401 * done only in carefully controlled circumstances. Descheduling is 1402 * asynchronous. 1403 * 1404 * This function may block if the cpu has run out of messages. 1405 */ 1406 void 1407 lwkt_deschedule(thread_t td) 1408 { 1409 crit_enter(); 1410 if (td == curthread) { 1411 _lwkt_dequeue(td); 1412 } else { 1413 if (td->td_gd == mycpu) { 1414 _lwkt_dequeue(td); 1415 } else { 1416 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td); 1417 } 1418 } 1419 crit_exit(); 1420 } 1421 1422 /* 1423 * Set the target thread's priority. This routine does not automatically 1424 * switch to a higher priority thread, LWKT threads are not designed for 1425 * continuous priority changes. Yield if you want to switch. 1426 */ 1427 void 1428 lwkt_setpri(thread_t td, int pri) 1429 { 1430 if (td->td_pri != pri) { 1431 KKASSERT(pri >= 0); 1432 crit_enter(); 1433 if (td->td_flags & TDF_RUNQ) { 1434 KKASSERT(td->td_gd == mycpu); 1435 _lwkt_dequeue(td); 1436 td->td_pri = pri; 1437 _lwkt_enqueue(td); 1438 } else { 1439 td->td_pri = pri; 1440 } 1441 crit_exit(); 1442 } 1443 } 1444 1445 /* 1446 * Set the initial priority for a thread prior to it being scheduled for 1447 * the first time. The thread MUST NOT be scheduled before or during 1448 * this call. The thread may be assigned to a cpu other then the current 1449 * cpu. 1450 * 1451 * Typically used after a thread has been created with TDF_STOPPREQ, 1452 * and before the thread is initially scheduled. 1453 */ 1454 void 1455 lwkt_setpri_initial(thread_t td, int pri) 1456 { 1457 KKASSERT(pri >= 0); 1458 KKASSERT((td->td_flags & TDF_RUNQ) == 0); 1459 td->td_pri = pri; 1460 } 1461 1462 void 1463 lwkt_setpri_self(int pri) 1464 { 1465 thread_t td = curthread; 1466 1467 KKASSERT(pri >= 0 && pri <= TDPRI_MAX); 1468 crit_enter(); 1469 if (td->td_flags & TDF_RUNQ) { 1470 _lwkt_dequeue(td); 1471 td->td_pri = pri; 1472 _lwkt_enqueue(td); 1473 } else { 1474 td->td_pri = pri; 1475 } 1476 crit_exit(); 1477 } 1478 1479 /* 1480 * hz tick scheduler clock for LWKT threads 1481 */ 1482 void 1483 lwkt_schedulerclock(thread_t td) 1484 { 1485 globaldata_t gd = td->td_gd; 1486 thread_t xtd; 1487 1488 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) { 1489 /* 1490 * If the current thread is at the head of the runq shift it to the 1491 * end of any equal-priority threads and request a LWKT reschedule 1492 * if it moved. 1493 * 1494 * Ignore upri in this situation. There will only be one user thread 1495 * in user mode, all others will be user threads running in kernel 1496 * mode and we have to make sure they get some cpu. 1497 */ 1498 xtd = TAILQ_NEXT(td, td_threadq); 1499 if (xtd && xtd->td_pri == td->td_pri) { 1500 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq); 1501 while (xtd && xtd->td_pri == td->td_pri) 1502 xtd = TAILQ_NEXT(xtd, td_threadq); 1503 if (xtd) 1504 TAILQ_INSERT_BEFORE(xtd, td, td_threadq); 1505 else 1506 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 1507 need_lwkt_resched(); 1508 } 1509 } else { 1510 /* 1511 * If we scheduled a thread other than the one at the head of the 1512 * queue always request a reschedule every tick. 1513 */ 1514 need_lwkt_resched(); 1515 } 1516 } 1517 1518 /* 1519 * Migrate the current thread to the specified cpu. 1520 * 1521 * This is accomplished by descheduling ourselves from the current cpu 1522 * and setting td_migrate_gd. The lwkt_switch() code will detect that the 1523 * 'old' thread wants to migrate after it has been completely switched out 1524 * and will complete the migration. 1525 * 1526 * TDF_MIGRATING prevents scheduling races while the thread is being migrated. 1527 * 1528 * We must be sure to release our current process designation (if a user 1529 * process) before clearing out any tsleepq we are on because the release 1530 * code may re-add us. 1531 * 1532 * We must be sure to remove ourselves from the current cpu's tsleepq 1533 * before potentially moving to another queue. The thread can be on 1534 * a tsleepq due to a left-over tsleep_interlock(). 1535 */ 1536 1537 void 1538 lwkt_setcpu_self(globaldata_t rgd) 1539 { 1540 thread_t td = curthread; 1541 1542 if (td->td_gd != rgd) { 1543 crit_enter_quick(td); 1544 1545 if (td->td_release) 1546 td->td_release(td); 1547 if (td->td_flags & TDF_TSLEEPQ) 1548 tsleep_remove(td); 1549 1550 /* 1551 * Set TDF_MIGRATING to prevent a spurious reschedule while we are 1552 * trying to deschedule ourselves and switch away, then deschedule 1553 * ourself, remove us from tdallq, and set td_migrate_gd. Finally, 1554 * call lwkt_switch() to complete the operation. 1555 */ 1556 td->td_flags |= TDF_MIGRATING; 1557 lwkt_deschedule_self(td); 1558 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1559 td->td_migrate_gd = rgd; 1560 lwkt_switch(); 1561 1562 /* 1563 * We are now on the target cpu 1564 */ 1565 KKASSERT(rgd == mycpu); 1566 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); 1567 crit_exit_quick(td); 1568 } 1569 } 1570 1571 void 1572 lwkt_migratecpu(int cpuid) 1573 { 1574 globaldata_t rgd; 1575 1576 rgd = globaldata_find(cpuid); 1577 lwkt_setcpu_self(rgd); 1578 } 1579 1580 /* 1581 * Remote IPI for cpu migration (called while in a critical section so we 1582 * do not have to enter another one). 1583 * 1584 * The thread (td) has already been completely descheduled from the 1585 * originating cpu and we can simply assert the case. The thread is 1586 * assigned to the new cpu and enqueued. 1587 * 1588 * The thread will re-add itself to tdallq when it resumes execution. 1589 */ 1590 static void 1591 lwkt_setcpu_remote(void *arg) 1592 { 1593 thread_t td = arg; 1594 globaldata_t gd = mycpu; 1595 1596 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0); 1597 td->td_gd = gd; 1598 cpu_mfence(); 1599 td->td_flags &= ~TDF_MIGRATING; 1600 KKASSERT(td->td_migrate_gd == NULL); 1601 KKASSERT(td->td_lwp == NULL || 1602 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 1603 _lwkt_enqueue(td); 1604 } 1605 1606 struct lwp * 1607 lwkt_preempted_proc(void) 1608 { 1609 thread_t td = curthread; 1610 while (td->td_preempted) 1611 td = td->td_preempted; 1612 return(td->td_lwp); 1613 } 1614 1615 /* 1616 * Create a kernel process/thread/whatever. It shares it's address space 1617 * with proc0 - ie: kernel only. 1618 * 1619 * If the cpu is not specified one will be selected. In the future 1620 * specifying a cpu of -1 will enable kernel thread migration between 1621 * cpus. 1622 */ 1623 int 1624 lwkt_create(void (*func)(void *), void *arg, struct thread **tdp, 1625 thread_t template, int tdflags, int cpu, const char *fmt, ...) 1626 { 1627 thread_t td; 1628 __va_list ap; 1629 1630 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu, 1631 tdflags); 1632 if (tdp) 1633 *tdp = td; 1634 cpu_set_thread_handler(td, lwkt_exit, func, arg); 1635 1636 /* 1637 * Set up arg0 for 'ps' etc 1638 */ 1639 __va_start(ap, fmt); 1640 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap); 1641 __va_end(ap); 1642 1643 /* 1644 * Schedule the thread to run 1645 */ 1646 if (td->td_flags & TDF_NOSTART) 1647 td->td_flags &= ~TDF_NOSTART; 1648 else 1649 lwkt_schedule(td); 1650 return 0; 1651 } 1652 1653 /* 1654 * Destroy an LWKT thread. Warning! This function is not called when 1655 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and 1656 * uses a different reaping mechanism. 1657 */ 1658 void 1659 lwkt_exit(void) 1660 { 1661 thread_t td = curthread; 1662 thread_t std; 1663 globaldata_t gd; 1664 1665 /* 1666 * Do any cleanup that might block here 1667 */ 1668 if (td->td_flags & TDF_VERBOSE) 1669 kprintf("kthread %p %s has exited\n", td, td->td_comm); 1670 biosched_done(td); 1671 dsched_exit_thread(td); 1672 1673 /* 1674 * Get us into a critical section to interlock gd_freetd and loop 1675 * until we can get it freed. 1676 * 1677 * We have to cache the current td in gd_freetd because objcache_put()ing 1678 * it would rip it out from under us while our thread is still active. 1679 * 1680 * We are the current thread so of course our own TDF_RUNNING bit will 1681 * be set, so unlike the lwp reap code we don't wait for it to clear. 1682 */ 1683 gd = mycpu; 1684 crit_enter_quick(td); 1685 for (;;) { 1686 if (td->td_refs) { 1687 tsleep(td, 0, "tdreap", 1); 1688 continue; 1689 } 1690 if ((std = gd->gd_freetd) != NULL) { 1691 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0); 1692 gd->gd_freetd = NULL; 1693 objcache_put(thread_cache, std); 1694 continue; 1695 } 1696 break; 1697 } 1698 1699 /* 1700 * Remove thread resources from kernel lists and deschedule us for 1701 * the last time. We cannot block after this point or we may end 1702 * up with a stale td on the tsleepq. 1703 * 1704 * None of this may block, the critical section is the only thing 1705 * protecting tdallq and the only thing preventing new lwkt_hold() 1706 * thread refs now. 1707 */ 1708 if (td->td_flags & TDF_TSLEEPQ) 1709 tsleep_remove(td); 1710 lwkt_deschedule_self(td); 1711 lwkt_remove_tdallq(td); 1712 KKASSERT(td->td_refs == 0); 1713 1714 /* 1715 * Final cleanup 1716 */ 1717 KKASSERT(gd->gd_freetd == NULL); 1718 if (td->td_flags & TDF_ALLOCATED_THREAD) 1719 gd->gd_freetd = td; 1720 cpu_thread_exit(); 1721 } 1722 1723 void 1724 lwkt_remove_tdallq(thread_t td) 1725 { 1726 KKASSERT(td->td_gd == mycpu); 1727 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1728 } 1729 1730 /* 1731 * Code reduction and branch prediction improvements. Call/return 1732 * overhead on modern cpus often degenerates into 0 cycles due to 1733 * the cpu's branch prediction hardware and return pc cache. We 1734 * can take advantage of this by not inlining medium-complexity 1735 * functions and we can also reduce the branch prediction impact 1736 * by collapsing perfectly predictable branches into a single 1737 * procedure instead of duplicating it. 1738 * 1739 * Is any of this noticeable? Probably not, so I'll take the 1740 * smaller code size. 1741 */ 1742 void 1743 crit_exit_wrapper(__DEBUG_CRIT_ARG__) 1744 { 1745 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__); 1746 } 1747 1748 void 1749 crit_panic(void) 1750 { 1751 thread_t td = curthread; 1752 int lcrit = td->td_critcount; 1753 1754 td->td_critcount = 0; 1755 panic("td_critcount is/would-go negative! %p %d", td, lcrit); 1756 /* NOT REACHED */ 1757 } 1758 1759 /* 1760 * Called from debugger/panic on cpus which have been stopped. We must still 1761 * process the IPIQ while stopped, even if we were stopped while in a critical 1762 * section (XXX). 1763 * 1764 * If we are dumping also try to process any pending interrupts. This may 1765 * or may not work depending on the state of the cpu at the point it was 1766 * stopped. 1767 */ 1768 void 1769 lwkt_smp_stopped(void) 1770 { 1771 globaldata_t gd = mycpu; 1772 1773 crit_enter_gd(gd); 1774 if (dumping) { 1775 lwkt_process_ipiq(); 1776 splz(); 1777 } else { 1778 lwkt_process_ipiq(); 1779 } 1780 crit_exit_gd(gd); 1781 } 1782