1 /* 2 * Copyright (c) 2003-2011 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@backplane.com> 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * 3. Neither the name of The DragonFly Project nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific, prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 */ 34 35 /* 36 * Each cpu in a system has its own self-contained light weight kernel 37 * thread scheduler, which means that generally speaking we only need 38 * to use a critical section to avoid problems. Foreign thread 39 * scheduling is queued via (async) IPIs. 40 */ 41 42 #include <sys/param.h> 43 #include <sys/systm.h> 44 #include <sys/kernel.h> 45 #include <sys/proc.h> 46 #include <sys/rtprio.h> 47 #include <sys/kinfo.h> 48 #include <sys/queue.h> 49 #include <sys/sysctl.h> 50 #include <sys/kthread.h> 51 #include <machine/cpu.h> 52 #include <sys/lock.h> 53 #include <sys/spinlock.h> 54 #include <sys/ktr.h> 55 56 #include <sys/thread2.h> 57 #include <sys/spinlock2.h> 58 #include <sys/mplock2.h> 59 60 #include <sys/dsched.h> 61 62 #include <vm/vm.h> 63 #include <vm/vm_param.h> 64 #include <vm/vm_kern.h> 65 #include <vm/vm_object.h> 66 #include <vm/vm_page.h> 67 #include <vm/vm_map.h> 68 #include <vm/vm_pager.h> 69 #include <vm/vm_extern.h> 70 71 #include <machine/stdarg.h> 72 #include <machine/smp.h> 73 74 #ifdef _KERNEL_VIRTUAL 75 #include <pthread.h> 76 #endif 77 78 #if !defined(KTR_CTXSW) 79 #define KTR_CTXSW KTR_ALL 80 #endif 81 KTR_INFO_MASTER(ctxsw); 82 KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p", int cpu, struct thread *td); 83 KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p", int cpu, struct thread *td); 84 KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s", struct thread *td, char *comm); 85 KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", struct thread *td); 86 87 static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads"); 88 89 #ifdef INVARIANTS 90 static int panic_on_cscount = 0; 91 #endif 92 static int64_t switch_count = 0; 93 static int64_t preempt_hit = 0; 94 static int64_t preempt_miss = 0; 95 static int64_t preempt_weird = 0; 96 static int lwkt_use_spin_port; 97 static struct objcache *thread_cache; 98 int cpu_mwait_spin = 0; 99 100 static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame); 101 static void lwkt_setcpu_remote(void *arg); 102 103 /* 104 * We can make all thread ports use the spin backend instead of the thread 105 * backend. This should only be set to debug the spin backend. 106 */ 107 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port); 108 109 #ifdef INVARIANTS 110 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, 111 "Panic if attempting to switch lwkt's while mastering cpusync"); 112 #endif 113 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, 114 "Number of switched threads"); 115 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, 116 "Successful preemption events"); 117 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, 118 "Failed preemption events"); 119 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, 120 "Number of preempted threads."); 121 static int fairq_enable = 0; 122 SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW, 123 &fairq_enable, 0, "Turn on fairq priority accumulators"); 124 static int fairq_bypass = -1; 125 SYSCTL_INT(_lwkt, OID_AUTO, fairq_bypass, CTLFLAG_RW, 126 &fairq_bypass, 0, "Allow fairq to bypass td on token failure"); 127 extern int lwkt_sched_debug; 128 int lwkt_sched_debug = 0; 129 SYSCTL_INT(_lwkt, OID_AUTO, sched_debug, CTLFLAG_RW, 130 &lwkt_sched_debug, 0, "Scheduler debug"); 131 static int lwkt_spin_loops = 10; 132 SYSCTL_INT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW, 133 &lwkt_spin_loops, 0, "Scheduler spin loops until sorted decon"); 134 static int lwkt_spin_reseq = 0; 135 SYSCTL_INT(_lwkt, OID_AUTO, spin_reseq, CTLFLAG_RW, 136 &lwkt_spin_reseq, 0, "Scheduler resequencer enable"); 137 static int lwkt_spin_monitor = 0; 138 SYSCTL_INT(_lwkt, OID_AUTO, spin_monitor, CTLFLAG_RW, 139 &lwkt_spin_monitor, 0, "Scheduler uses monitor/mwait"); 140 static int lwkt_spin_fatal = 0; /* disabled */ 141 SYSCTL_INT(_lwkt, OID_AUTO, spin_fatal, CTLFLAG_RW, 142 &lwkt_spin_fatal, 0, "LWKT scheduler spin loops till fatal panic"); 143 static int preempt_enable = 1; 144 SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW, 145 &preempt_enable, 0, "Enable preemption"); 146 static int lwkt_cache_threads = 0; 147 SYSCTL_INT(_lwkt, OID_AUTO, cache_threads, CTLFLAG_RD, 148 &lwkt_cache_threads, 0, "thread+kstack cache"); 149 150 #ifndef _KERNEL_VIRTUAL 151 static __cachealign int lwkt_cseq_rindex; 152 static __cachealign int lwkt_cseq_windex; 153 #endif 154 155 /* 156 * These helper procedures handle the runq, they can only be called from 157 * within a critical section. 158 * 159 * WARNING! Prior to SMP being brought up it is possible to enqueue and 160 * dequeue threads belonging to other cpus, so be sure to use td->td_gd 161 * instead of 'mycpu' when referencing the globaldata structure. Once 162 * SMP live enqueuing and dequeueing only occurs on the current cpu. 163 */ 164 static __inline 165 void 166 _lwkt_dequeue(thread_t td) 167 { 168 if (td->td_flags & TDF_RUNQ) { 169 struct globaldata *gd = td->td_gd; 170 171 td->td_flags &= ~TDF_RUNQ; 172 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq); 173 --gd->gd_tdrunqcount; 174 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL) 175 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING); 176 } 177 } 178 179 /* 180 * Priority enqueue. 181 * 182 * There are a limited number of lwkt threads runnable since user 183 * processes only schedule one at a time per cpu. However, there can 184 * be many user processes in kernel mode exiting from a tsleep() which 185 * become runnable. 186 * 187 * NOTE: lwkt_schedulerclock() will force a round-robin based on td_pri and 188 * will ignore user priority. This is to ensure that user threads in 189 * kernel mode get cpu at some point regardless of what the user 190 * scheduler thinks. 191 */ 192 static __inline 193 void 194 _lwkt_enqueue(thread_t td) 195 { 196 thread_t xtd; 197 198 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) { 199 struct globaldata *gd = td->td_gd; 200 201 td->td_flags |= TDF_RUNQ; 202 xtd = TAILQ_FIRST(&gd->gd_tdrunq); 203 if (xtd == NULL) { 204 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 205 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING); 206 } else { 207 /* 208 * NOTE: td_upri - higher numbers more desireable, same sense 209 * as td_pri (typically reversed from lwp_upri). 210 * 211 * In the equal priority case we want the best selection 212 * at the beginning so the less desireable selections know 213 * that they have to setrunqueue/go-to-another-cpu, even 214 * though it means switching back to the 'best' selection. 215 * This also avoids degenerate situations when many threads 216 * are runnable or waking up at the same time. 217 * 218 * If upri matches exactly place at end/round-robin. 219 */ 220 while (xtd && 221 (xtd->td_pri >= td->td_pri || 222 (xtd->td_pri == td->td_pri && 223 xtd->td_upri >= td->td_upri))) { 224 xtd = TAILQ_NEXT(xtd, td_threadq); 225 } 226 if (xtd) 227 TAILQ_INSERT_BEFORE(xtd, td, td_threadq); 228 else 229 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 230 } 231 ++gd->gd_tdrunqcount; 232 233 /* 234 * Request a LWKT reschedule if we are now at the head of the queue. 235 */ 236 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) 237 need_lwkt_resched(); 238 } 239 } 240 241 static boolean_t 242 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags) 243 { 244 struct thread *td = (struct thread *)obj; 245 246 td->td_kstack = NULL; 247 td->td_kstack_size = 0; 248 td->td_flags = TDF_ALLOCATED_THREAD; 249 td->td_mpflags = 0; 250 return (1); 251 } 252 253 static void 254 _lwkt_thread_dtor(void *obj, void *privdata) 255 { 256 struct thread *td = (struct thread *)obj; 257 258 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD, 259 ("_lwkt_thread_dtor: not allocated from objcache")); 260 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack && 261 td->td_kstack_size > 0, 262 ("_lwkt_thread_dtor: corrupted stack")); 263 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size); 264 td->td_kstack = NULL; 265 td->td_flags = 0; 266 } 267 268 /* 269 * Initialize the lwkt s/system. 270 * 271 * Nominally cache up to 32 thread + kstack structures. Cache more on 272 * systems with a lot of cpu cores. 273 */ 274 static void 275 lwkt_init(void) 276 { 277 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads); 278 if (lwkt_cache_threads == 0) { 279 lwkt_cache_threads = ncpus * 4; 280 if (lwkt_cache_threads < 32) 281 lwkt_cache_threads = 32; 282 } 283 thread_cache = objcache_create_mbacked( 284 M_THREAD, sizeof(struct thread), 285 0, lwkt_cache_threads, 286 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL); 287 } 288 SYSINIT(lwkt_init, SI_BOOT2_LWKT_INIT, SI_ORDER_FIRST, lwkt_init, NULL); 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 (flags & TDF_FIXEDCPU) ? TRUE : FALSE); 446 } else { 447 lwkt_initport_thread(&td->td_msgport, td); 448 } 449 pmap_init_thread(td); 450 /* 451 * Normally initializing a thread for a remote cpu requires sending an 452 * IPI. However, the idlethread is setup before the other cpus are 453 * activated so we have to treat it as a special case. XXX manipulation 454 * of gd_tdallq requires the BGL. 455 */ 456 if (gd == mygd || td == &gd->gd_idlethread) { 457 crit_enter_gd(mygd); 458 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); 459 crit_exit_gd(mygd); 460 } else { 461 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td); 462 } 463 dsched_new_thread(td); 464 } 465 466 void 467 lwkt_set_comm(thread_t td, const char *ctl, ...) 468 { 469 __va_list va; 470 471 __va_start(va, ctl); 472 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va); 473 __va_end(va); 474 KTR_LOG(ctxsw_newtd, td, td->td_comm); 475 } 476 477 /* 478 * Prevent the thread from getting destroyed. Note that unlike PHOLD/PRELE 479 * this does not prevent the thread from migrating to another cpu so the 480 * gd_tdallq state is not protected by this. 481 */ 482 void 483 lwkt_hold(thread_t td) 484 { 485 atomic_add_int(&td->td_refs, 1); 486 } 487 488 void 489 lwkt_rele(thread_t td) 490 { 491 KKASSERT(td->td_refs > 0); 492 atomic_add_int(&td->td_refs, -1); 493 } 494 495 void 496 lwkt_free_thread(thread_t td) 497 { 498 KKASSERT(td->td_refs == 0); 499 KKASSERT((td->td_flags & (TDF_RUNNING | TDF_PREEMPT_LOCK | 500 TDF_RUNQ | TDF_TSLEEPQ)) == 0); 501 if (td->td_flags & TDF_ALLOCATED_THREAD) { 502 objcache_put(thread_cache, td); 503 } else if (td->td_flags & TDF_ALLOCATED_STACK) { 504 /* client-allocated struct with internally allocated stack */ 505 KASSERT(td->td_kstack && td->td_kstack_size > 0, 506 ("lwkt_free_thread: corrupted stack")); 507 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size); 508 td->td_kstack = NULL; 509 td->td_kstack_size = 0; 510 } 511 512 KTR_LOG(ctxsw_deadtd, td); 513 } 514 515 516 /* 517 * Switch to the next runnable lwkt. If no LWKTs are runnable then 518 * switch to the idlethread. Switching must occur within a critical 519 * section to avoid races with the scheduling queue. 520 * 521 * We always have full control over our cpu's run queue. Other cpus 522 * that wish to manipulate our queue must use the cpu_*msg() calls to 523 * talk to our cpu, so a critical section is all that is needed and 524 * the result is very, very fast thread switching. 525 * 526 * The LWKT scheduler uses a fixed priority model and round-robins at 527 * each priority level. User process scheduling is a totally 528 * different beast and LWKT priorities should not be confused with 529 * user process priorities. 530 * 531 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch() 532 * is not called by the current thread in the preemption case, only when 533 * the preempting thread blocks (in order to return to the original thread). 534 * 535 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread 536 * migration and tsleep deschedule the current lwkt thread and call 537 * lwkt_switch(). In particular, the target cpu of the migration fully 538 * expects the thread to become non-runnable and can deadlock against 539 * cpusync operations if we run any IPIs prior to switching the thread out. 540 * 541 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF 542 * THE CURRENT THREAD HAS BEEN DESCHEDULED! 543 */ 544 void 545 lwkt_switch(void) 546 { 547 globaldata_t gd = mycpu; 548 thread_t td = gd->gd_curthread; 549 thread_t ntd; 550 int spinning = 0; 551 552 KKASSERT(gd->gd_processing_ipiq == 0); 553 KKASSERT(td->td_flags & TDF_RUNNING); 554 555 /* 556 * Switching from within a 'fast' (non thread switched) interrupt or IPI 557 * is illegal. However, we may have to do it anyway if we hit a fatal 558 * kernel trap or we have paniced. 559 * 560 * If this case occurs save and restore the interrupt nesting level. 561 */ 562 if (gd->gd_intr_nesting_level) { 563 int savegdnest; 564 int savegdtrap; 565 566 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) { 567 panic("lwkt_switch: Attempt to switch from a " 568 "fast interrupt, ipi, or hard code section, " 569 "td %p\n", 570 td); 571 } else { 572 savegdnest = gd->gd_intr_nesting_level; 573 savegdtrap = gd->gd_trap_nesting_level; 574 gd->gd_intr_nesting_level = 0; 575 gd->gd_trap_nesting_level = 0; 576 if ((td->td_flags & TDF_PANICWARN) == 0) { 577 td->td_flags |= TDF_PANICWARN; 578 kprintf("Warning: thread switch from interrupt, IPI, " 579 "or hard code section.\n" 580 "thread %p (%s)\n", td, td->td_comm); 581 print_backtrace(-1); 582 } 583 lwkt_switch(); 584 gd->gd_intr_nesting_level = savegdnest; 585 gd->gd_trap_nesting_level = savegdtrap; 586 return; 587 } 588 } 589 590 /* 591 * Release our current user process designation if we are blocking 592 * or if a user reschedule was requested. 593 * 594 * NOTE: This function is NOT called if we are switching into or 595 * returning from a preemption. 596 * 597 * NOTE: Releasing our current user process designation may cause 598 * it to be assigned to another thread, which in turn will 599 * cause us to block in the usched acquire code when we attempt 600 * to return to userland. 601 * 602 * NOTE: On SMP systems this can be very nasty when heavy token 603 * contention is present so we want to be careful not to 604 * release the designation gratuitously. 605 */ 606 if (td->td_release && 607 (user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) { 608 td->td_release(td); 609 } 610 611 /* 612 * Release all tokens 613 */ 614 crit_enter_gd(gd); 615 if (TD_TOKS_HELD(td)) 616 lwkt_relalltokens(td); 617 618 /* 619 * We had better not be holding any spin locks, but don't get into an 620 * endless panic loop. 621 */ 622 KASSERT(gd->gd_spinlocks == 0 || panicstr != NULL, 623 ("lwkt_switch: still holding %d exclusive spinlocks!", 624 gd->gd_spinlocks)); 625 626 627 #ifdef INVARIANTS 628 if (td->td_cscount) { 629 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n", 630 td); 631 if (panic_on_cscount) 632 panic("switching while mastering cpusync"); 633 } 634 #endif 635 636 /* 637 * If we had preempted another thread on this cpu, resume the preempted 638 * thread. This occurs transparently, whether the preempted thread 639 * was scheduled or not (it may have been preempted after descheduling 640 * itself). 641 * 642 * We have to setup the MP lock for the original thread after backing 643 * out the adjustment that was made to curthread when the original 644 * was preempted. 645 */ 646 if ((ntd = td->td_preempted) != NULL) { 647 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK); 648 ntd->td_flags |= TDF_PREEMPT_DONE; 649 650 /* 651 * The interrupt may have woken a thread up, we need to properly 652 * set the reschedule flag if the originally interrupted thread is 653 * at a lower priority. 654 * 655 * The interrupt may not have descheduled. 656 */ 657 if (TAILQ_FIRST(&gd->gd_tdrunq) != ntd) 658 need_lwkt_resched(); 659 goto havethread_preempted; 660 } 661 662 /* 663 * If we cannot obtain ownership of the tokens we cannot immediately 664 * schedule the target thread. 665 * 666 * Reminder: Again, we cannot afford to run any IPIs in this path if 667 * the current thread has been descheduled. 668 */ 669 for (;;) { 670 clear_lwkt_resched(); 671 672 /* 673 * Hotpath - pull the head of the run queue and attempt to schedule 674 * it. 675 */ 676 ntd = TAILQ_FIRST(&gd->gd_tdrunq); 677 678 if (ntd == NULL) { 679 /* 680 * Runq is empty, switch to idle to allow it to halt. 681 */ 682 ntd = &gd->gd_idlethread; 683 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) 684 ASSERT_NO_TOKENS_HELD(ntd); 685 cpu_time.cp_msg[0] = 0; 686 cpu_time.cp_stallpc = 0; 687 goto haveidle; 688 } 689 690 /* 691 * Hotpath - schedule ntd. 692 * 693 * NOTE: For UP there is no mplock and lwkt_getalltokens() 694 * always succeeds. 695 */ 696 if (TD_TOKS_NOT_HELD(ntd) || 697 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops))) 698 { 699 goto havethread; 700 } 701 702 /* 703 * Coldpath (SMP only since tokens always succeed on UP) 704 * 705 * We had some contention on the thread we wanted to schedule. 706 * What we do now is try to find a thread that we can schedule 707 * in its stead. 708 * 709 * The coldpath scan does NOT rearrange threads in the run list. 710 * The lwkt_schedulerclock() will assert need_lwkt_resched() on 711 * the next tick whenever the current head is not the current thread. 712 */ 713 #ifdef INVARIANTS 714 ++ntd->td_contended; 715 #endif 716 ++gd->gd_cnt.v_lock_colls; 717 718 if (fairq_bypass > 0) 719 goto skip; 720 721 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) { 722 #ifndef NO_LWKT_SPLIT_USERPRI 723 /* 724 * Never schedule threads returning to userland or the 725 * user thread scheduler helper thread when higher priority 726 * threads are present. The runq is sorted by priority 727 * so we can give up traversing it when we find the first 728 * low priority thread. 729 */ 730 if (ntd->td_pri < TDPRI_KERN_LPSCHED) { 731 ntd = NULL; 732 break; 733 } 734 #endif 735 736 /* 737 * Try this one. 738 */ 739 if (TD_TOKS_NOT_HELD(ntd) || 740 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops))) { 741 goto havethread; 742 } 743 #ifdef INVARIANTS 744 ++ntd->td_contended; 745 #endif 746 ++gd->gd_cnt.v_lock_colls; 747 } 748 749 skip: 750 /* 751 * We exhausted the run list, meaning that all runnable threads 752 * are contested. 753 */ 754 cpu_pause(); 755 #ifdef _KERNEL_VIRTUAL 756 pthread_yield(); 757 #endif 758 ntd = &gd->gd_idlethread; 759 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) 760 ASSERT_NO_TOKENS_HELD(ntd); 761 /* contention case, do not clear contention mask */ 762 763 /* 764 * We are going to have to retry but if the current thread is not 765 * on the runq we instead switch through the idle thread to get away 766 * from the current thread. We have to flag for lwkt reschedule 767 * to prevent the idle thread from halting. 768 * 769 * NOTE: A non-zero spinning is passed to lwkt_getalltokens() to 770 * instruct it to deal with the potential for deadlocks by 771 * ordering the tokens by address. 772 */ 773 if ((td->td_flags & TDF_RUNQ) == 0) { 774 need_lwkt_resched(); /* prevent hlt */ 775 goto haveidle; 776 } 777 #if defined(INVARIANTS) && defined(__x86_64__) 778 if ((read_rflags() & PSL_I) == 0) { 779 cpu_enable_intr(); 780 panic("lwkt_switch() called with interrupts disabled"); 781 } 782 #endif 783 784 /* 785 * Number iterations so far. After a certain point we switch to 786 * a sorted-address/monitor/mwait version of lwkt_getalltokens() 787 */ 788 if (spinning < 0x7FFFFFFF) 789 ++spinning; 790 791 #ifndef _KERNEL_VIRTUAL 792 /* 793 * lwkt_getalltokens() failed in sorted token mode, we can use 794 * monitor/mwait in this case. 795 */ 796 if (spinning >= lwkt_spin_loops && 797 (cpu_mi_feature & CPU_MI_MONITOR) && 798 lwkt_spin_monitor) 799 { 800 cpu_mmw_pause_int(&gd->gd_reqflags, 801 (gd->gd_reqflags | RQF_SPINNING) & 802 ~RQF_IDLECHECK_WK_MASK, 803 cpu_mwait_spin, 0); 804 } 805 #endif 806 807 /* 808 * We already checked that td is still scheduled so this should be 809 * safe. 810 */ 811 splz_check(); 812 813 #ifndef _KERNEL_VIRTUAL 814 /* 815 * This experimental resequencer is used as a fall-back to reduce 816 * hw cache line contention by placing each core's scheduler into a 817 * time-domain-multplexed slot. 818 * 819 * The resequencer is disabled by default. It's functionality has 820 * largely been superceeded by the token algorithm which limits races 821 * to a subset of cores. 822 * 823 * The resequencer algorithm tends to break down when more than 824 * 20 cores are contending. What appears to happen is that new 825 * tokens can be obtained out of address-sorted order by new cores 826 * while existing cores languish in long delays between retries and 827 * wind up being starved-out of the token acquisition. 828 */ 829 if (lwkt_spin_reseq && spinning >= lwkt_spin_reseq) { 830 int cseq = atomic_fetchadd_int(&lwkt_cseq_windex, 1); 831 int oseq; 832 833 while ((oseq = lwkt_cseq_rindex) != cseq) { 834 cpu_ccfence(); 835 #if 1 836 if (cpu_mi_feature & CPU_MI_MONITOR) { 837 cpu_mmw_pause_int(&lwkt_cseq_rindex, oseq, 838 cpu_mwait_spin, 0); 839 } else { 840 #endif 841 cpu_pause(); 842 cpu_lfence(); 843 #if 1 844 } 845 #endif 846 } 847 DELAY(1); 848 atomic_add_int(&lwkt_cseq_rindex, 1); 849 } 850 #endif 851 /* highest level for(;;) loop */ 852 } 853 854 havethread: 855 /* 856 * Clear gd_idle_repeat when doing a normal switch to a non-idle 857 * thread. 858 */ 859 ntd->td_wmesg = NULL; 860 ++gd->gd_cnt.v_swtch; 861 gd->gd_idle_repeat = 0; 862 863 havethread_preempted: 864 /* 865 * If the new target does not need the MP lock and we are holding it, 866 * release the MP lock. If the new target requires the MP lock we have 867 * already acquired it for the target. 868 */ 869 ; 870 haveidle: 871 KASSERT(ntd->td_critcount, 872 ("priority problem in lwkt_switch %d %d", 873 td->td_critcount, ntd->td_critcount)); 874 875 if (td != ntd) { 876 /* 877 * Execute the actual thread switch operation. This function 878 * returns to the current thread and returns the previous thread 879 * (which may be different from the thread we switched to). 880 * 881 * We are responsible for marking ntd as TDF_RUNNING. 882 */ 883 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0); 884 ++switch_count; 885 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd); 886 ntd->td_flags |= TDF_RUNNING; 887 lwkt_switch_return(td->td_switch(ntd)); 888 /* ntd invalid, td_switch() can return a different thread_t */ 889 } 890 891 /* 892 * catch-all. XXX is this strictly needed? 893 */ 894 splz_check(); 895 896 /* NOTE: current cpu may have changed after switch */ 897 crit_exit_quick(td); 898 } 899 900 /* 901 * Called by assembly in the td_switch (thread restore path) for thread 902 * bootstrap cases which do not 'return' to lwkt_switch(). 903 */ 904 void 905 lwkt_switch_return(thread_t otd) 906 { 907 globaldata_t rgd; 908 909 /* 910 * Check if otd was migrating. Now that we are on ntd we can finish 911 * up the migration. This is a bit messy but it is the only place 912 * where td is known to be fully descheduled. 913 * 914 * We can only activate the migration if otd was migrating but not 915 * held on the cpu due to a preemption chain. We still have to 916 * clear TDF_RUNNING on the old thread either way. 917 * 918 * We are responsible for clearing the previously running thread's 919 * TDF_RUNNING. 920 */ 921 if ((rgd = otd->td_migrate_gd) != NULL && 922 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) { 923 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) == 924 (TDF_MIGRATING | TDF_RUNNING)); 925 otd->td_migrate_gd = NULL; 926 otd->td_flags &= ~TDF_RUNNING; 927 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd); 928 } else { 929 otd->td_flags &= ~TDF_RUNNING; 930 } 931 932 /* 933 * Final exit validations (see lwp_wait()). Note that otd becomes 934 * invalid the *instant* we set TDF_MP_EXITSIG. 935 */ 936 while (otd->td_flags & TDF_EXITING) { 937 u_int mpflags; 938 939 mpflags = otd->td_mpflags; 940 cpu_ccfence(); 941 942 if (mpflags & TDF_MP_EXITWAIT) { 943 if (atomic_cmpset_int(&otd->td_mpflags, mpflags, 944 mpflags | TDF_MP_EXITSIG)) { 945 wakeup(otd); 946 break; 947 } 948 } else { 949 if (atomic_cmpset_int(&otd->td_mpflags, mpflags, 950 mpflags | TDF_MP_EXITSIG)) { 951 wakeup(otd); 952 break; 953 } 954 } 955 } 956 } 957 958 /* 959 * Request that the target thread preempt the current thread. Preemption 960 * can only occur if our only critical section is the one that we were called 961 * with, the relative priority of the target thread is higher, and the target 962 * thread holds no tokens. This also only works if we are not holding any 963 * spinlocks (obviously). 964 * 965 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically 966 * this is called via lwkt_schedule() through the td_preemptable callback. 967 * critcount is the managed critical priority that we should ignore in order 968 * to determine whether preemption is possible (aka usually just the crit 969 * priority of lwkt_schedule() itself). 970 * 971 * Preemption is typically limited to interrupt threads. 972 * 973 * Operation works in a fairly straight-forward manner. The normal 974 * scheduling code is bypassed and we switch directly to the target 975 * thread. When the target thread attempts to block or switch away 976 * code at the base of lwkt_switch() will switch directly back to our 977 * thread. Our thread is able to retain whatever tokens it holds and 978 * if the target needs one of them the target will switch back to us 979 * and reschedule itself normally. 980 */ 981 void 982 lwkt_preempt(thread_t ntd, int critcount) 983 { 984 struct globaldata *gd = mycpu; 985 thread_t xtd; 986 thread_t td; 987 int save_gd_intr_nesting_level; 988 989 /* 990 * The caller has put us in a critical section. We can only preempt 991 * if the caller of the caller was not in a critical section (basically 992 * a local interrupt), as determined by the 'critcount' parameter. We 993 * also can't preempt if the caller is holding any spinlocks (even if 994 * he isn't in a critical section). This also handles the tokens test. 995 * 996 * YYY The target thread must be in a critical section (else it must 997 * inherit our critical section? I dunno yet). 998 */ 999 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri)); 1000 1001 td = gd->gd_curthread; 1002 if (preempt_enable == 0) { 1003 ++preempt_miss; 1004 return; 1005 } 1006 if (ntd->td_pri <= td->td_pri) { 1007 ++preempt_miss; 1008 return; 1009 } 1010 if (td->td_critcount > critcount) { 1011 ++preempt_miss; 1012 return; 1013 } 1014 if (td->td_cscount) { 1015 ++preempt_miss; 1016 return; 1017 } 1018 if (ntd->td_gd != gd) { 1019 ++preempt_miss; 1020 return; 1021 } 1022 /* 1023 * We don't have to check spinlocks here as they will also bump 1024 * td_critcount. 1025 * 1026 * Do not try to preempt if the target thread is holding any tokens. 1027 * We could try to acquire the tokens but this case is so rare there 1028 * is no need to support it. 1029 */ 1030 KKASSERT(gd->gd_spinlocks == 0); 1031 1032 if (TD_TOKS_HELD(ntd)) { 1033 ++preempt_miss; 1034 return; 1035 } 1036 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) { 1037 ++preempt_weird; 1038 return; 1039 } 1040 if (ntd->td_preempted) { 1041 ++preempt_hit; 1042 return; 1043 } 1044 KKASSERT(gd->gd_processing_ipiq == 0); 1045 1046 /* 1047 * Since we are able to preempt the current thread, there is no need to 1048 * call need_lwkt_resched(). 1049 * 1050 * We must temporarily clear gd_intr_nesting_level around the switch 1051 * since switchouts from the target thread are allowed (they will just 1052 * return to our thread), and since the target thread has its own stack. 1053 * 1054 * A preemption must switch back to the original thread, assert the 1055 * case. 1056 */ 1057 ++preempt_hit; 1058 ntd->td_preempted = td; 1059 td->td_flags |= TDF_PREEMPT_LOCK; 1060 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd); 1061 save_gd_intr_nesting_level = gd->gd_intr_nesting_level; 1062 gd->gd_intr_nesting_level = 0; 1063 1064 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0); 1065 ntd->td_flags |= TDF_RUNNING; 1066 xtd = td->td_switch(ntd); 1067 KKASSERT(xtd == ntd); 1068 lwkt_switch_return(xtd); 1069 gd->gd_intr_nesting_level = save_gd_intr_nesting_level; 1070 1071 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE)); 1072 ntd->td_preempted = NULL; 1073 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE); 1074 } 1075 1076 /* 1077 * Conditionally call splz() if gd_reqflags indicates work is pending. 1078 * This will work inside a critical section but not inside a hard code 1079 * section. 1080 * 1081 * (self contained on a per cpu basis) 1082 */ 1083 void 1084 splz_check(void) 1085 { 1086 globaldata_t gd = mycpu; 1087 thread_t td = gd->gd_curthread; 1088 1089 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && 1090 gd->gd_intr_nesting_level == 0 && 1091 td->td_nest_count < 2) 1092 { 1093 splz(); 1094 } 1095 } 1096 1097 /* 1098 * This version is integrated into crit_exit, reqflags has already 1099 * been tested but td_critcount has not. 1100 * 1101 * We only want to execute the splz() on the 1->0 transition of 1102 * critcount and not in a hard code section or if too deeply nested. 1103 * 1104 * NOTE: gd->gd_spinlocks is implied to be 0 when td_critcount is 0. 1105 */ 1106 void 1107 lwkt_maybe_splz(thread_t td) 1108 { 1109 globaldata_t gd = td->td_gd; 1110 1111 if (td->td_critcount == 0 && 1112 gd->gd_intr_nesting_level == 0 && 1113 td->td_nest_count < 2) 1114 { 1115 splz(); 1116 } 1117 } 1118 1119 /* 1120 * Drivers which set up processing co-threads can call this function to 1121 * run the co-thread at a higher priority and to allow it to preempt 1122 * normal threads. 1123 */ 1124 void 1125 lwkt_set_interrupt_support_thread(void) 1126 { 1127 thread_t td = curthread; 1128 1129 lwkt_setpri_self(TDPRI_INT_SUPPORT); 1130 td->td_flags |= TDF_INTTHREAD; 1131 td->td_preemptable = lwkt_preempt; 1132 } 1133 1134 1135 /* 1136 * This function is used to negotiate a passive release of the current 1137 * process/lwp designation with the user scheduler, allowing the user 1138 * scheduler to schedule another user thread. The related kernel thread 1139 * (curthread) continues running in the released state. 1140 */ 1141 void 1142 lwkt_passive_release(struct thread *td) 1143 { 1144 struct lwp *lp = td->td_lwp; 1145 1146 #ifndef NO_LWKT_SPLIT_USERPRI 1147 td->td_release = NULL; 1148 lwkt_setpri_self(TDPRI_KERN_USER); 1149 #endif 1150 1151 lp->lwp_proc->p_usched->release_curproc(lp); 1152 } 1153 1154 1155 /* 1156 * This implements a LWKT yield, allowing a kernel thread to yield to other 1157 * kernel threads at the same or higher priority. This function can be 1158 * called in a tight loop and will typically only yield once per tick. 1159 * 1160 * Most kernel threads run at the same priority in order to allow equal 1161 * sharing. 1162 * 1163 * (self contained on a per cpu basis) 1164 */ 1165 void 1166 lwkt_yield(void) 1167 { 1168 globaldata_t gd = mycpu; 1169 thread_t td = gd->gd_curthread; 1170 1171 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1172 splz(); 1173 if (lwkt_resched_wanted()) { 1174 lwkt_schedule_self(curthread); 1175 lwkt_switch(); 1176 } 1177 } 1178 1179 /* 1180 * The quick version processes pending interrupts and higher-priority 1181 * LWKT threads but will not round-robin same-priority LWKT threads. 1182 * 1183 * When called while attempting to return to userland the only same-pri 1184 * threads are the ones which have already tried to become the current 1185 * user process. 1186 */ 1187 void 1188 lwkt_yield_quick(void) 1189 { 1190 globaldata_t gd = mycpu; 1191 thread_t td = gd->gd_curthread; 1192 1193 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1194 splz(); 1195 if (lwkt_resched_wanted()) { 1196 crit_enter(); 1197 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) { 1198 clear_lwkt_resched(); 1199 } else { 1200 lwkt_schedule_self(curthread); 1201 lwkt_switch(); 1202 } 1203 crit_exit(); 1204 } 1205 } 1206 1207 /* 1208 * This yield is designed for kernel threads with a user context. 1209 * 1210 * The kernel acting on behalf of the user is potentially cpu-bound, 1211 * this function will efficiently allow other threads to run and also 1212 * switch to other processes by releasing. 1213 * 1214 * The lwkt_user_yield() function is designed to have very low overhead 1215 * if no yield is determined to be needed. 1216 */ 1217 void 1218 lwkt_user_yield(void) 1219 { 1220 globaldata_t gd = mycpu; 1221 thread_t td = gd->gd_curthread; 1222 1223 /* 1224 * Always run any pending interrupts in case we are in a critical 1225 * section. 1226 */ 1227 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1228 splz(); 1229 1230 /* 1231 * Switch (which forces a release) if another kernel thread needs 1232 * the cpu, if userland wants us to resched, or if our kernel 1233 * quantum has run out. 1234 */ 1235 if (lwkt_resched_wanted() || 1236 user_resched_wanted()) 1237 { 1238 lwkt_switch(); 1239 } 1240 1241 #if 0 1242 /* 1243 * Reacquire the current process if we are released. 1244 * 1245 * XXX not implemented atm. The kernel may be holding locks and such, 1246 * so we want the thread to continue to receive cpu. 1247 */ 1248 if (td->td_release == NULL && lp) { 1249 lp->lwp_proc->p_usched->acquire_curproc(lp); 1250 td->td_release = lwkt_passive_release; 1251 lwkt_setpri_self(TDPRI_USER_NORM); 1252 } 1253 #endif 1254 } 1255 1256 /* 1257 * Generic schedule. Possibly schedule threads belonging to other cpus and 1258 * deal with threads that might be blocked on a wait queue. 1259 * 1260 * We have a little helper inline function which does additional work after 1261 * the thread has been enqueued, including dealing with preemption and 1262 * setting need_lwkt_resched() (which prevents the kernel from returning 1263 * to userland until it has processed higher priority threads). 1264 * 1265 * It is possible for this routine to be called after a failed _enqueue 1266 * (due to the target thread migrating, sleeping, or otherwise blocked). 1267 * We have to check that the thread is actually on the run queue! 1268 */ 1269 static __inline 1270 void 1271 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount) 1272 { 1273 if (ntd->td_flags & TDF_RUNQ) { 1274 if (ntd->td_preemptable) { 1275 ntd->td_preemptable(ntd, ccount); /* YYY +token */ 1276 } 1277 } 1278 } 1279 1280 static __inline 1281 void 1282 _lwkt_schedule(thread_t td) 1283 { 1284 globaldata_t mygd = mycpu; 1285 1286 KASSERT(td != &td->td_gd->gd_idlethread, 1287 ("lwkt_schedule(): scheduling gd_idlethread is illegal!")); 1288 KKASSERT((td->td_flags & TDF_MIGRATING) == 0); 1289 crit_enter_gd(mygd); 1290 KKASSERT(td->td_lwp == NULL || 1291 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 1292 1293 if (td == mygd->gd_curthread) { 1294 _lwkt_enqueue(td); 1295 } else { 1296 /* 1297 * If we own the thread, there is no race (since we are in a 1298 * critical section). If we do not own the thread there might 1299 * be a race but the target cpu will deal with it. 1300 */ 1301 if (td->td_gd == mygd) { 1302 _lwkt_enqueue(td); 1303 _lwkt_schedule_post(mygd, td, 1); 1304 } else { 1305 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0); 1306 } 1307 } 1308 crit_exit_gd(mygd); 1309 } 1310 1311 void 1312 lwkt_schedule(thread_t td) 1313 { 1314 _lwkt_schedule(td); 1315 } 1316 1317 void 1318 lwkt_schedule_noresched(thread_t td) /* XXX not impl */ 1319 { 1320 _lwkt_schedule(td); 1321 } 1322 1323 /* 1324 * When scheduled remotely if frame != NULL the IPIQ is being 1325 * run via doreti or an interrupt then preemption can be allowed. 1326 * 1327 * To allow preemption we have to drop the critical section so only 1328 * one is present in _lwkt_schedule_post. 1329 */ 1330 static void 1331 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame) 1332 { 1333 thread_t td = curthread; 1334 thread_t ntd = arg; 1335 1336 if (frame && ntd->td_preemptable) { 1337 crit_exit_noyield(td); 1338 _lwkt_schedule(ntd); 1339 crit_enter_quick(td); 1340 } else { 1341 _lwkt_schedule(ntd); 1342 } 1343 } 1344 1345 /* 1346 * Thread migration using a 'Pull' method. The thread may or may not be 1347 * the current thread. It MUST be descheduled and in a stable state. 1348 * lwkt_giveaway() must be called on the cpu owning the thread. 1349 * 1350 * At any point after lwkt_giveaway() is called, the target cpu may 1351 * 'pull' the thread by calling lwkt_acquire(). 1352 * 1353 * We have to make sure the thread is not sitting on a per-cpu tsleep 1354 * queue or it will blow up when it moves to another cpu. 1355 * 1356 * MPSAFE - must be called under very specific conditions. 1357 */ 1358 void 1359 lwkt_giveaway(thread_t td) 1360 { 1361 globaldata_t gd = mycpu; 1362 1363 crit_enter_gd(gd); 1364 if (td->td_flags & TDF_TSLEEPQ) 1365 tsleep_remove(td); 1366 KKASSERT(td->td_gd == gd); 1367 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); 1368 td->td_flags |= TDF_MIGRATING; 1369 crit_exit_gd(gd); 1370 } 1371 1372 void 1373 lwkt_acquire(thread_t td) 1374 { 1375 globaldata_t gd; 1376 globaldata_t mygd; 1377 int retry = 10000000; 1378 1379 KKASSERT(td->td_flags & TDF_MIGRATING); 1380 gd = td->td_gd; 1381 mygd = mycpu; 1382 if (gd != mycpu) { 1383 cpu_lfence(); 1384 KKASSERT((td->td_flags & TDF_RUNQ) == 0); 1385 crit_enter_gd(mygd); 1386 DEBUG_PUSH_INFO("lwkt_acquire"); 1387 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) { 1388 lwkt_process_ipiq(); 1389 cpu_lfence(); 1390 if (--retry == 0) { 1391 kprintf("lwkt_acquire: stuck: td %p td->td_flags %08x\n", 1392 td, td->td_flags); 1393 retry = 10000000; 1394 } 1395 #ifdef _KERNEL_VIRTUAL 1396 pthread_yield(); 1397 #endif 1398 } 1399 DEBUG_POP_INFO(); 1400 cpu_mfence(); 1401 td->td_gd = mygd; 1402 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1403 td->td_flags &= ~TDF_MIGRATING; 1404 crit_exit_gd(mygd); 1405 } else { 1406 crit_enter_gd(mygd); 1407 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1408 td->td_flags &= ~TDF_MIGRATING; 1409 crit_exit_gd(mygd); 1410 } 1411 } 1412 1413 /* 1414 * Generic deschedule. Descheduling threads other then your own should be 1415 * done only in carefully controlled circumstances. Descheduling is 1416 * asynchronous. 1417 * 1418 * This function may block if the cpu has run out of messages. 1419 */ 1420 void 1421 lwkt_deschedule(thread_t td) 1422 { 1423 crit_enter(); 1424 if (td == curthread) { 1425 _lwkt_dequeue(td); 1426 } else { 1427 if (td->td_gd == mycpu) { 1428 _lwkt_dequeue(td); 1429 } else { 1430 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td); 1431 } 1432 } 1433 crit_exit(); 1434 } 1435 1436 /* 1437 * Set the target thread's priority. This routine does not automatically 1438 * switch to a higher priority thread, LWKT threads are not designed for 1439 * continuous priority changes. Yield if you want to switch. 1440 */ 1441 void 1442 lwkt_setpri(thread_t td, int pri) 1443 { 1444 if (td->td_pri != pri) { 1445 KKASSERT(pri >= 0); 1446 crit_enter(); 1447 if (td->td_flags & TDF_RUNQ) { 1448 KKASSERT(td->td_gd == mycpu); 1449 _lwkt_dequeue(td); 1450 td->td_pri = pri; 1451 _lwkt_enqueue(td); 1452 } else { 1453 td->td_pri = pri; 1454 } 1455 crit_exit(); 1456 } 1457 } 1458 1459 /* 1460 * Set the initial priority for a thread prior to it being scheduled for 1461 * the first time. The thread MUST NOT be scheduled before or during 1462 * this call. The thread may be assigned to a cpu other then the current 1463 * cpu. 1464 * 1465 * Typically used after a thread has been created with TDF_STOPPREQ, 1466 * and before the thread is initially scheduled. 1467 */ 1468 void 1469 lwkt_setpri_initial(thread_t td, int pri) 1470 { 1471 KKASSERT(pri >= 0); 1472 KKASSERT((td->td_flags & TDF_RUNQ) == 0); 1473 td->td_pri = pri; 1474 } 1475 1476 void 1477 lwkt_setpri_self(int pri) 1478 { 1479 thread_t td = curthread; 1480 1481 KKASSERT(pri >= 0 && pri <= TDPRI_MAX); 1482 crit_enter(); 1483 if (td->td_flags & TDF_RUNQ) { 1484 _lwkt_dequeue(td); 1485 td->td_pri = pri; 1486 _lwkt_enqueue(td); 1487 } else { 1488 td->td_pri = pri; 1489 } 1490 crit_exit(); 1491 } 1492 1493 /* 1494 * hz tick scheduler clock for LWKT threads 1495 */ 1496 void 1497 lwkt_schedulerclock(thread_t td) 1498 { 1499 globaldata_t gd = td->td_gd; 1500 thread_t xtd; 1501 1502 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) { 1503 /* 1504 * If the current thread is at the head of the runq shift it to the 1505 * end of any equal-priority threads and request a LWKT reschedule 1506 * if it moved. 1507 * 1508 * Ignore upri in this situation. There will only be one user thread 1509 * in user mode, all others will be user threads running in kernel 1510 * mode and we have to make sure they get some cpu. 1511 */ 1512 xtd = TAILQ_NEXT(td, td_threadq); 1513 if (xtd && xtd->td_pri == td->td_pri) { 1514 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq); 1515 while (xtd && xtd->td_pri == td->td_pri) 1516 xtd = TAILQ_NEXT(xtd, td_threadq); 1517 if (xtd) 1518 TAILQ_INSERT_BEFORE(xtd, td, td_threadq); 1519 else 1520 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 1521 need_lwkt_resched(); 1522 } 1523 } else { 1524 /* 1525 * If we scheduled a thread other than the one at the head of the 1526 * queue always request a reschedule every tick. 1527 */ 1528 need_lwkt_resched(); 1529 } 1530 } 1531 1532 /* 1533 * Migrate the current thread to the specified cpu. 1534 * 1535 * This is accomplished by descheduling ourselves from the current cpu 1536 * and setting td_migrate_gd. The lwkt_switch() code will detect that the 1537 * 'old' thread wants to migrate after it has been completely switched out 1538 * and will complete the migration. 1539 * 1540 * TDF_MIGRATING prevents scheduling races while the thread is being migrated. 1541 * 1542 * We must be sure to release our current process designation (if a user 1543 * process) before clearing out any tsleepq we are on because the release 1544 * code may re-add us. 1545 * 1546 * We must be sure to remove ourselves from the current cpu's tsleepq 1547 * before potentially moving to another queue. The thread can be on 1548 * a tsleepq due to a left-over tsleep_interlock(). 1549 */ 1550 1551 void 1552 lwkt_setcpu_self(globaldata_t rgd) 1553 { 1554 thread_t td = curthread; 1555 1556 if (td->td_gd != rgd) { 1557 crit_enter_quick(td); 1558 1559 if (td->td_release) 1560 td->td_release(td); 1561 if (td->td_flags & TDF_TSLEEPQ) 1562 tsleep_remove(td); 1563 1564 /* 1565 * Set TDF_MIGRATING to prevent a spurious reschedule while we are 1566 * trying to deschedule ourselves and switch away, then deschedule 1567 * ourself, remove us from tdallq, and set td_migrate_gd. Finally, 1568 * call lwkt_switch() to complete the operation. 1569 */ 1570 td->td_flags |= TDF_MIGRATING; 1571 lwkt_deschedule_self(td); 1572 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1573 td->td_migrate_gd = rgd; 1574 lwkt_switch(); 1575 1576 /* 1577 * We are now on the target cpu 1578 */ 1579 KKASSERT(rgd == mycpu); 1580 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); 1581 crit_exit_quick(td); 1582 } 1583 } 1584 1585 void 1586 lwkt_migratecpu(int cpuid) 1587 { 1588 globaldata_t rgd; 1589 1590 rgd = globaldata_find(cpuid); 1591 lwkt_setcpu_self(rgd); 1592 } 1593 1594 /* 1595 * Remote IPI for cpu migration (called while in a critical section so we 1596 * do not have to enter another one). 1597 * 1598 * The thread (td) has already been completely descheduled from the 1599 * originating cpu and we can simply assert the case. The thread is 1600 * assigned to the new cpu and enqueued. 1601 * 1602 * The thread will re-add itself to tdallq when it resumes execution. 1603 */ 1604 static void 1605 lwkt_setcpu_remote(void *arg) 1606 { 1607 thread_t td = arg; 1608 globaldata_t gd = mycpu; 1609 1610 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0); 1611 td->td_gd = gd; 1612 cpu_mfence(); 1613 td->td_flags &= ~TDF_MIGRATING; 1614 KKASSERT(td->td_migrate_gd == NULL); 1615 KKASSERT(td->td_lwp == NULL || 1616 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 1617 _lwkt_enqueue(td); 1618 } 1619 1620 struct lwp * 1621 lwkt_preempted_proc(void) 1622 { 1623 thread_t td = curthread; 1624 while (td->td_preempted) 1625 td = td->td_preempted; 1626 return(td->td_lwp); 1627 } 1628 1629 /* 1630 * Create a kernel process/thread/whatever. It shares it's address space 1631 * with proc0 - ie: kernel only. 1632 * 1633 * If the cpu is not specified one will be selected. In the future 1634 * specifying a cpu of -1 will enable kernel thread migration between 1635 * cpus. 1636 */ 1637 int 1638 lwkt_create(void (*func)(void *), void *arg, struct thread **tdp, 1639 thread_t template, int tdflags, int cpu, const char *fmt, ...) 1640 { 1641 thread_t td; 1642 __va_list ap; 1643 1644 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu, 1645 tdflags); 1646 if (tdp) 1647 *tdp = td; 1648 cpu_set_thread_handler(td, lwkt_exit, func, arg); 1649 1650 /* 1651 * Set up arg0 for 'ps' etc 1652 */ 1653 __va_start(ap, fmt); 1654 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap); 1655 __va_end(ap); 1656 1657 /* 1658 * Schedule the thread to run 1659 */ 1660 if (td->td_flags & TDF_NOSTART) 1661 td->td_flags &= ~TDF_NOSTART; 1662 else 1663 lwkt_schedule(td); 1664 return 0; 1665 } 1666 1667 /* 1668 * Destroy an LWKT thread. Warning! This function is not called when 1669 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and 1670 * uses a different reaping mechanism. 1671 */ 1672 void 1673 lwkt_exit(void) 1674 { 1675 thread_t td = curthread; 1676 thread_t std; 1677 globaldata_t gd; 1678 1679 /* 1680 * Do any cleanup that might block here 1681 */ 1682 if (td->td_flags & TDF_VERBOSE) 1683 kprintf("kthread %p %s has exited\n", td, td->td_comm); 1684 biosched_done(td); 1685 dsched_exit_thread(td); 1686 1687 /* 1688 * Get us into a critical section to interlock gd_freetd and loop 1689 * until we can get it freed. 1690 * 1691 * We have to cache the current td in gd_freetd because objcache_put()ing 1692 * it would rip it out from under us while our thread is still active. 1693 * 1694 * We are the current thread so of course our own TDF_RUNNING bit will 1695 * be set, so unlike the lwp reap code we don't wait for it to clear. 1696 */ 1697 gd = mycpu; 1698 crit_enter_quick(td); 1699 for (;;) { 1700 if (td->td_refs) { 1701 tsleep(td, 0, "tdreap", 1); 1702 continue; 1703 } 1704 if ((std = gd->gd_freetd) != NULL) { 1705 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0); 1706 gd->gd_freetd = NULL; 1707 objcache_put(thread_cache, std); 1708 continue; 1709 } 1710 break; 1711 } 1712 1713 /* 1714 * Remove thread resources from kernel lists and deschedule us for 1715 * the last time. We cannot block after this point or we may end 1716 * up with a stale td on the tsleepq. 1717 * 1718 * None of this may block, the critical section is the only thing 1719 * protecting tdallq and the only thing preventing new lwkt_hold() 1720 * thread refs now. 1721 */ 1722 if (td->td_flags & TDF_TSLEEPQ) 1723 tsleep_remove(td); 1724 lwkt_deschedule_self(td); 1725 lwkt_remove_tdallq(td); 1726 KKASSERT(td->td_refs == 0); 1727 1728 /* 1729 * Final cleanup 1730 */ 1731 KKASSERT(gd->gd_freetd == NULL); 1732 if (td->td_flags & TDF_ALLOCATED_THREAD) 1733 gd->gd_freetd = td; 1734 cpu_thread_exit(); 1735 } 1736 1737 void 1738 lwkt_remove_tdallq(thread_t td) 1739 { 1740 KKASSERT(td->td_gd == mycpu); 1741 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1742 } 1743 1744 /* 1745 * Code reduction and branch prediction improvements. Call/return 1746 * overhead on modern cpus often degenerates into 0 cycles due to 1747 * the cpu's branch prediction hardware and return pc cache. We 1748 * can take advantage of this by not inlining medium-complexity 1749 * functions and we can also reduce the branch prediction impact 1750 * by collapsing perfectly predictable branches into a single 1751 * procedure instead of duplicating it. 1752 * 1753 * Is any of this noticeable? Probably not, so I'll take the 1754 * smaller code size. 1755 */ 1756 void 1757 crit_exit_wrapper(__DEBUG_CRIT_ARG__) 1758 { 1759 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__); 1760 } 1761 1762 void 1763 crit_panic(void) 1764 { 1765 thread_t td = curthread; 1766 int lcrit = td->td_critcount; 1767 1768 td->td_critcount = 0; 1769 panic("td_critcount is/would-go negative! %p %d", td, lcrit); 1770 /* NOT REACHED */ 1771 } 1772 1773 /* 1774 * Called from debugger/panic on cpus which have been stopped. We must still 1775 * process the IPIQ while stopped. 1776 * 1777 * If we are dumping also try to process any pending interrupts. This may 1778 * or may not work depending on the state of the cpu at the point it was 1779 * stopped. 1780 */ 1781 void 1782 lwkt_smp_stopped(void) 1783 { 1784 globaldata_t gd = mycpu; 1785 1786 if (dumping) { 1787 lwkt_process_ipiq(); 1788 --gd->gd_intr_nesting_level; 1789 splz(); 1790 ++gd->gd_intr_nesting_level; 1791 } else { 1792 lwkt_process_ipiq(); 1793 } 1794 } 1795