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 #include <sys/indefinite.h> 56 57 #include <sys/thread2.h> 58 #include <sys/spinlock2.h> 59 #include <sys/indefinite2.h> 60 61 #include <sys/dsched.h> 62 63 #include <vm/vm.h> 64 #include <vm/vm_param.h> 65 #include <vm/vm_kern.h> 66 #include <vm/vm_object.h> 67 #include <vm/vm_page.h> 68 #include <vm/vm_map.h> 69 #include <vm/vm_pager.h> 70 #include <vm/vm_extern.h> 71 72 #include <machine/stdarg.h> 73 #include <machine/smp.h> 74 #include <machine/clock.h> 75 76 #ifdef _KERNEL_VIRTUAL 77 #include <pthread.h> 78 #endif 79 80 #define LOOPMASK 81 82 #if !defined(KTR_CTXSW) 83 #define KTR_CTXSW KTR_ALL 84 #endif 85 KTR_INFO_MASTER(ctxsw); 86 KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p", int cpu, struct thread *td); 87 KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p", int cpu, struct thread *td); 88 KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s", struct thread *td, char *comm); 89 KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", struct thread *td); 90 91 static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads"); 92 93 #ifdef INVARIANTS 94 static int panic_on_cscount = 0; 95 #endif 96 static int64_t switch_count = 0; 97 static int64_t preempt_hit = 0; 98 static int64_t preempt_miss = 0; 99 static int64_t preempt_weird = 0; 100 static int lwkt_use_spin_port; 101 static struct objcache *thread_cache; 102 int cpu_mwait_spin = 0; 103 104 static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame); 105 static void lwkt_setcpu_remote(void *arg); 106 107 /* 108 * We can make all thread ports use the spin backend instead of the thread 109 * backend. This should only be set to debug the spin backend. 110 */ 111 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port); 112 113 #ifdef INVARIANTS 114 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, 115 "Panic if attempting to switch lwkt's while mastering cpusync"); 116 #endif 117 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, 118 "Number of switched threads"); 119 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, 120 "Successful preemption events"); 121 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, 122 "Failed preemption events"); 123 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, 124 "Number of preempted threads."); 125 static int fairq_enable = 0; 126 SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW, 127 &fairq_enable, 0, "Turn on fairq priority accumulators"); 128 static int fairq_bypass = -1; 129 SYSCTL_INT(_lwkt, OID_AUTO, fairq_bypass, CTLFLAG_RW, 130 &fairq_bypass, 0, "Allow fairq to bypass td on token failure"); 131 extern int lwkt_sched_debug; 132 int lwkt_sched_debug = 0; 133 SYSCTL_INT(_lwkt, OID_AUTO, sched_debug, CTLFLAG_RW, 134 &lwkt_sched_debug, 0, "Scheduler debug"); 135 static u_int lwkt_spin_loops = 10; 136 SYSCTL_UINT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW, 137 &lwkt_spin_loops, 0, "Scheduler spin loops until sorted decon"); 138 static int preempt_enable = 1; 139 SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW, 140 &preempt_enable, 0, "Enable preemption"); 141 static int lwkt_cache_threads = 0; 142 SYSCTL_INT(_lwkt, OID_AUTO, cache_threads, CTLFLAG_RD, 143 &lwkt_cache_threads, 0, "thread+kstack cache"); 144 145 /* 146 * These helper procedures handle the runq, they can only be called from 147 * within a critical section. 148 * 149 * WARNING! Prior to SMP being brought up it is possible to enqueue and 150 * dequeue threads belonging to other cpus, so be sure to use td->td_gd 151 * instead of 'mycpu' when referencing the globaldata structure. Once 152 * SMP live enqueuing and dequeueing only occurs on the current cpu. 153 */ 154 static __inline 155 void 156 _lwkt_dequeue(thread_t td) 157 { 158 if (td->td_flags & TDF_RUNQ) { 159 struct globaldata *gd = td->td_gd; 160 161 td->td_flags &= ~TDF_RUNQ; 162 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq); 163 --gd->gd_tdrunqcount; 164 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL) 165 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING); 166 } 167 } 168 169 /* 170 * Priority enqueue. 171 * 172 * There are a limited number of lwkt threads runnable since user 173 * processes only schedule one at a time per cpu. However, there can 174 * be many user processes in kernel mode exiting from a tsleep() which 175 * become runnable. 176 * 177 * We scan the queue in both directions to help deal with degenerate 178 * situations when hundreds or thousands (or more) threads are runnable. 179 * 180 * NOTE: lwkt_schedulerclock() will force a round-robin based on td_pri and 181 * will ignore user priority. This is to ensure that user threads in 182 * kernel mode get cpu at some point regardless of what the user 183 * scheduler thinks. 184 */ 185 static __inline 186 void 187 _lwkt_enqueue(thread_t td) 188 { 189 thread_t xtd; /* forward scan */ 190 thread_t rtd; /* reverse scan */ 191 192 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) { 193 struct globaldata *gd = td->td_gd; 194 195 td->td_flags |= TDF_RUNQ; 196 xtd = TAILQ_FIRST(&gd->gd_tdrunq); 197 if (xtd == NULL) { 198 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 199 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING); 200 } else { 201 /* 202 * NOTE: td_upri - higher numbers more desireable, same sense 203 * as td_pri (typically reversed from lwp_upri). 204 * 205 * In the equal priority case we want the best selection 206 * at the beginning so the less desireable selections know 207 * that they have to setrunqueue/go-to-another-cpu, even 208 * though it means switching back to the 'best' selection. 209 * This also avoids degenerate situations when many threads 210 * are runnable or waking up at the same time. 211 * 212 * If upri matches exactly place at end/round-robin. 213 */ 214 rtd = TAILQ_LAST(&gd->gd_tdrunq, lwkt_queue); 215 216 while (xtd && 217 (xtd->td_pri > td->td_pri || 218 (xtd->td_pri == td->td_pri && 219 xtd->td_upri >= td->td_upri))) { 220 xtd = TAILQ_NEXT(xtd, td_threadq); 221 222 /* 223 * Doing a reverse scan at the same time is an optimization 224 * for the insert-closer-to-tail case that avoids having to 225 * scan the entire list. This situation can occur when 226 * thousands of threads are woken up at the same time. 227 */ 228 if (rtd->td_pri > td->td_pri || 229 (rtd->td_pri == td->td_pri && 230 rtd->td_upri >= td->td_upri)) { 231 TAILQ_INSERT_AFTER(&gd->gd_tdrunq, rtd, td, td_threadq); 232 goto skip; 233 } 234 rtd = TAILQ_PREV(rtd, lwkt_queue, td_threadq); 235 } 236 if (xtd) 237 TAILQ_INSERT_BEFORE(xtd, td, td_threadq); 238 else 239 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 240 } 241 skip: 242 ++gd->gd_tdrunqcount; 243 244 /* 245 * Request a LWKT reschedule if we are now at the head of the queue. 246 */ 247 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) 248 need_lwkt_resched(); 249 } 250 } 251 252 static boolean_t 253 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags) 254 { 255 struct thread *td = (struct thread *)obj; 256 257 td->td_kstack = NULL; 258 td->td_kstack_size = 0; 259 td->td_flags = TDF_ALLOCATED_THREAD; 260 td->td_mpflags = 0; 261 return (1); 262 } 263 264 static void 265 _lwkt_thread_dtor(void *obj, void *privdata) 266 { 267 struct thread *td = (struct thread *)obj; 268 269 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD, 270 ("_lwkt_thread_dtor: not allocated from objcache")); 271 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack && 272 td->td_kstack_size > 0, 273 ("_lwkt_thread_dtor: corrupted stack")); 274 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size); 275 td->td_kstack = NULL; 276 td->td_flags = 0; 277 } 278 279 /* 280 * Initialize the lwkt s/system. 281 * 282 * Nominally cache up to 32 thread + kstack structures. Cache more on 283 * systems with a lot of cpu cores. 284 */ 285 static void 286 lwkt_init(void) 287 { 288 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads); 289 if (lwkt_cache_threads == 0) { 290 lwkt_cache_threads = ncpus * 4; 291 if (lwkt_cache_threads < 32) 292 lwkt_cache_threads = 32; 293 } 294 thread_cache = objcache_create_mbacked( 295 M_THREAD, sizeof(struct thread), 296 0, lwkt_cache_threads, 297 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL); 298 } 299 SYSINIT(lwkt_init, SI_BOOT2_LWKT_INIT, SI_ORDER_FIRST, lwkt_init, NULL); 300 301 /* 302 * Schedule a thread to run. As the current thread we can always safely 303 * schedule ourselves, and a shortcut procedure is provided for that 304 * function. 305 * 306 * (non-blocking, self contained on a per cpu basis) 307 */ 308 void 309 lwkt_schedule_self(thread_t td) 310 { 311 KKASSERT((td->td_flags & TDF_MIGRATING) == 0); 312 crit_enter_quick(td); 313 KASSERT(td != &td->td_gd->gd_idlethread, 314 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!")); 315 KKASSERT(td->td_lwp == NULL || 316 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 317 _lwkt_enqueue(td); 318 crit_exit_quick(td); 319 } 320 321 /* 322 * Deschedule a thread. 323 * 324 * (non-blocking, self contained on a per cpu basis) 325 */ 326 void 327 lwkt_deschedule_self(thread_t td) 328 { 329 crit_enter_quick(td); 330 _lwkt_dequeue(td); 331 crit_exit_quick(td); 332 } 333 334 /* 335 * LWKTs operate on a per-cpu basis 336 * 337 * WARNING! Called from early boot, 'mycpu' may not work yet. 338 */ 339 void 340 lwkt_gdinit(struct globaldata *gd) 341 { 342 TAILQ_INIT(&gd->gd_tdrunq); 343 TAILQ_INIT(&gd->gd_tdallq); 344 lockinit(&gd->gd_sysctllock, "sysctl", 0, LK_CANRECURSE); 345 } 346 347 /* 348 * Create a new thread. The thread must be associated with a process context 349 * or LWKT start address before it can be scheduled. If the target cpu is 350 * -1 the thread will be created on the current cpu. 351 * 352 * If you intend to create a thread without a process context this function 353 * does everything except load the startup and switcher function. 354 */ 355 thread_t 356 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags) 357 { 358 static int cpu_rotator; 359 globaldata_t gd = mycpu; 360 void *stack; 361 362 /* 363 * If static thread storage is not supplied allocate a thread. Reuse 364 * a cached free thread if possible. gd_freetd is used to keep an exiting 365 * thread intact through the exit. 366 */ 367 if (td == NULL) { 368 crit_enter_gd(gd); 369 if ((td = gd->gd_freetd) != NULL) { 370 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK| 371 TDF_RUNQ)) == 0); 372 gd->gd_freetd = NULL; 373 } else { 374 td = objcache_get(thread_cache, M_WAITOK); 375 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK| 376 TDF_RUNQ)) == 0); 377 } 378 crit_exit_gd(gd); 379 KASSERT((td->td_flags & 380 (TDF_ALLOCATED_THREAD|TDF_RUNNING|TDF_PREEMPT_LOCK)) == 381 TDF_ALLOCATED_THREAD, 382 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags)); 383 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK); 384 } 385 386 /* 387 * Try to reuse cached stack. 388 */ 389 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) { 390 if (flags & TDF_ALLOCATED_STACK) { 391 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size); 392 stack = NULL; 393 } 394 } 395 if (stack == NULL) { 396 if (cpu < 0) 397 stack = (void *)kmem_alloc_stack(&kernel_map, stksize, 0); 398 else 399 stack = (void *)kmem_alloc_stack(&kernel_map, stksize, 400 KM_CPU(cpu)); 401 flags |= TDF_ALLOCATED_STACK; 402 } 403 if (cpu < 0) { 404 cpu = ++cpu_rotator; 405 cpu_ccfence(); 406 cpu = (uint32_t)cpu % (uint32_t)ncpus; 407 } 408 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu)); 409 return(td); 410 } 411 412 /* 413 * Initialize a preexisting thread structure. This function is used by 414 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread. 415 * 416 * All threads start out in a critical section at a priority of 417 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as 418 * appropriate. This function may send an IPI message when the 419 * requested cpu is not the current cpu and consequently gd_tdallq may 420 * not be initialized synchronously from the point of view of the originating 421 * cpu. 422 * 423 * NOTE! we have to be careful in regards to creating threads for other cpus 424 * if SMP has not yet been activated. 425 */ 426 static void 427 lwkt_init_thread_remote(void *arg) 428 { 429 thread_t td = arg; 430 431 /* 432 * Protected by critical section held by IPI dispatch 433 */ 434 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq); 435 } 436 437 /* 438 * lwkt core thread structural initialization. 439 * 440 * NOTE: All threads are initialized as mpsafe threads. 441 */ 442 void 443 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags, 444 struct globaldata *gd) 445 { 446 globaldata_t mygd = mycpu; 447 448 bzero(td, sizeof(struct thread)); 449 td->td_kstack = stack; 450 td->td_kstack_size = stksize; 451 td->td_flags = flags; 452 td->td_mpflags = 0; 453 td->td_type = TD_TYPE_GENERIC; 454 td->td_gd = gd; 455 td->td_pri = TDPRI_KERN_DAEMON; 456 td->td_critcount = 1; 457 td->td_toks_have = NULL; 458 td->td_toks_stop = &td->td_toks_base; 459 if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT)) { 460 lwkt_initport_spin(&td->td_msgport, td, 461 (flags & TDF_FIXEDCPU) ? TRUE : FALSE); 462 } else { 463 lwkt_initport_thread(&td->td_msgport, td); 464 } 465 pmap_init_thread(td); 466 /* 467 * Normally initializing a thread for a remote cpu requires sending an 468 * IPI. However, the idlethread is setup before the other cpus are 469 * activated so we have to treat it as a special case. XXX manipulation 470 * of gd_tdallq requires the BGL. 471 */ 472 if (gd == mygd || td == &gd->gd_idlethread) { 473 crit_enter_gd(mygd); 474 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); 475 crit_exit_gd(mygd); 476 } else { 477 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td); 478 } 479 dsched_enter_thread(td); 480 } 481 482 void 483 lwkt_set_comm(thread_t td, const char *ctl, ...) 484 { 485 __va_list va; 486 487 __va_start(va, ctl); 488 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va); 489 __va_end(va); 490 KTR_LOG(ctxsw_newtd, td, td->td_comm); 491 } 492 493 /* 494 * Prevent the thread from getting destroyed. Note that unlike PHOLD/PRELE 495 * this does not prevent the thread from migrating to another cpu so the 496 * gd_tdallq state is not protected by this. 497 */ 498 void 499 lwkt_hold(thread_t td) 500 { 501 atomic_add_int(&td->td_refs, 1); 502 } 503 504 void 505 lwkt_rele(thread_t td) 506 { 507 KKASSERT(td->td_refs > 0); 508 atomic_add_int(&td->td_refs, -1); 509 } 510 511 void 512 lwkt_free_thread(thread_t td) 513 { 514 KKASSERT(td->td_refs == 0); 515 KKASSERT((td->td_flags & (TDF_RUNNING | TDF_PREEMPT_LOCK | 516 TDF_RUNQ | TDF_TSLEEPQ)) == 0); 517 if (td->td_flags & TDF_ALLOCATED_THREAD) { 518 objcache_put(thread_cache, td); 519 } else if (td->td_flags & TDF_ALLOCATED_STACK) { 520 /* client-allocated struct with internally allocated stack */ 521 KASSERT(td->td_kstack && td->td_kstack_size > 0, 522 ("lwkt_free_thread: corrupted stack")); 523 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size); 524 td->td_kstack = NULL; 525 td->td_kstack_size = 0; 526 } 527 528 KTR_LOG(ctxsw_deadtd, td); 529 } 530 531 532 /* 533 * Switch to the next runnable lwkt. If no LWKTs are runnable then 534 * switch to the idlethread. Switching must occur within a critical 535 * section to avoid races with the scheduling queue. 536 * 537 * We always have full control over our cpu's run queue. Other cpus 538 * that wish to manipulate our queue must use the cpu_*msg() calls to 539 * talk to our cpu, so a critical section is all that is needed and 540 * the result is very, very fast thread switching. 541 * 542 * The LWKT scheduler uses a fixed priority model and round-robins at 543 * each priority level. User process scheduling is a totally 544 * different beast and LWKT priorities should not be confused with 545 * user process priorities. 546 * 547 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch() 548 * is not called by the current thread in the preemption case, only when 549 * the preempting thread blocks (in order to return to the original thread). 550 * 551 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread 552 * migration and tsleep deschedule the current lwkt thread and call 553 * lwkt_switch(). In particular, the target cpu of the migration fully 554 * expects the thread to become non-runnable and can deadlock against 555 * cpusync operations if we run any IPIs prior to switching the thread out. 556 * 557 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF 558 * THE CURRENT THREAD HAS BEEN DESCHEDULED! 559 */ 560 void 561 lwkt_switch(void) 562 { 563 globaldata_t gd = mycpu; 564 thread_t td = gd->gd_curthread; 565 thread_t ntd; 566 thread_t xtd; 567 int upri; 568 #ifdef LOOPMASK 569 uint64_t tsc_base = rdtsc(); 570 #endif 571 572 KKASSERT(gd->gd_processing_ipiq == 0); 573 KKASSERT(td->td_flags & TDF_RUNNING); 574 575 /* 576 * Switching from within a 'fast' (non thread switched) interrupt or IPI 577 * is illegal. However, we may have to do it anyway if we hit a fatal 578 * kernel trap or we have paniced. 579 * 580 * If this case occurs save and restore the interrupt nesting level. 581 */ 582 if (gd->gd_intr_nesting_level) { 583 int savegdnest; 584 int savegdtrap; 585 586 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) { 587 panic("lwkt_switch: Attempt to switch from a " 588 "fast interrupt, ipi, or hard code section, " 589 "td %p\n", 590 td); 591 } else { 592 savegdnest = gd->gd_intr_nesting_level; 593 savegdtrap = gd->gd_trap_nesting_level; 594 gd->gd_intr_nesting_level = 0; 595 gd->gd_trap_nesting_level = 0; 596 if ((td->td_flags & TDF_PANICWARN) == 0) { 597 td->td_flags |= TDF_PANICWARN; 598 kprintf("Warning: thread switch from interrupt, IPI, " 599 "or hard code section.\n" 600 "thread %p (%s)\n", td, td->td_comm); 601 print_backtrace(-1); 602 } 603 lwkt_switch(); 604 gd->gd_intr_nesting_level = savegdnest; 605 gd->gd_trap_nesting_level = savegdtrap; 606 return; 607 } 608 } 609 610 /* 611 * Release our current user process designation if we are blocking 612 * or if a user reschedule was requested. 613 * 614 * NOTE: This function is NOT called if we are switching into or 615 * returning from a preemption. 616 * 617 * NOTE: Releasing our current user process designation may cause 618 * it to be assigned to another thread, which in turn will 619 * cause us to block in the usched acquire code when we attempt 620 * to return to userland. 621 * 622 * NOTE: On SMP systems this can be very nasty when heavy token 623 * contention is present so we want to be careful not to 624 * release the designation gratuitously. 625 */ 626 if (td->td_release && 627 (user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) { 628 td->td_release(td); 629 } 630 631 /* 632 * Release all tokens. Once we do this we must remain in the critical 633 * section and cannot run IPIs or other interrupts until we switch away 634 * because they may implode if they try to get a token using our thread 635 * context. 636 */ 637 crit_enter_gd(gd); 638 if (TD_TOKS_HELD(td)) 639 lwkt_relalltokens(td); 640 641 /* 642 * We had better not be holding any spin locks, but don't get into an 643 * endless panic loop. 644 */ 645 KASSERT(gd->gd_spinlocks == 0 || panicstr != NULL, 646 ("lwkt_switch: still holding %d exclusive spinlocks!", 647 gd->gd_spinlocks)); 648 649 #ifdef INVARIANTS 650 if (td->td_cscount) { 651 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n", 652 td); 653 if (panic_on_cscount) 654 panic("switching while mastering cpusync"); 655 } 656 #endif 657 658 /* 659 * If we had preempted another thread on this cpu, resume the preempted 660 * thread. This occurs transparently, whether the preempted thread 661 * was scheduled or not (it may have been preempted after descheduling 662 * itself). 663 * 664 * We have to setup the MP lock for the original thread after backing 665 * out the adjustment that was made to curthread when the original 666 * was preempted. 667 */ 668 if ((ntd = td->td_preempted) != NULL) { 669 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK); 670 ntd->td_flags |= TDF_PREEMPT_DONE; 671 ntd->td_contended = 0; /* reset contended */ 672 673 /* 674 * The interrupt may have woken a thread up, we need to properly 675 * set the reschedule flag if the originally interrupted thread is 676 * at a lower priority. 677 * 678 * NOTE: The interrupt may not have descheduled ntd. 679 * 680 * NOTE: We do not reschedule if there are no threads on the runq. 681 * (ntd could be the idlethread). 682 */ 683 xtd = TAILQ_FIRST(&gd->gd_tdrunq); 684 if (xtd && xtd != ntd) 685 need_lwkt_resched(); 686 goto havethread_preempted; 687 } 688 689 /* 690 * Figure out switch target. If we cannot switch to our desired target 691 * look for a thread that we can switch to. 692 * 693 * NOTE! The limited spin loop and related parameters are extremely 694 * important for system performance, particularly for pipes and 695 * concurrent conflicting VM faults. 696 */ 697 clear_lwkt_resched(); 698 ntd = TAILQ_FIRST(&gd->gd_tdrunq); 699 700 if (ntd) { 701 do { 702 if (TD_TOKS_NOT_HELD(ntd) || 703 lwkt_getalltokens(ntd, (ntd->td_contended > lwkt_spin_loops))) 704 { 705 goto havethread; 706 } 707 ++ntd->td_contended; /* overflow ok */ 708 if (gd->gd_indefinite.type == 0) 709 indefinite_init(&gd->gd_indefinite, NULL, 0, 't'); 710 #ifdef LOOPMASK 711 if (tsc_frequency && rdtsc() - tsc_base > tsc_frequency) { 712 kprintf("lwkt_switch: excessive contended %d " 713 "thread %p\n", ntd->td_contended, ntd); 714 tsc_base = rdtsc(); 715 } 716 #endif 717 } while (ntd->td_contended < (lwkt_spin_loops >> 1)); 718 upri = ntd->td_upri; 719 720 /* 721 * Bleh, the thread we wanted to switch to has a contended token. 722 * See if we can switch to another thread. 723 * 724 * We generally don't want to do this because it represents a 725 * priority inversion, but contending tokens on the same cpu can 726 * cause real problems if we don't now that we have an exclusive 727 * priority mechanism over shared for tokens. 728 * 729 * The solution is to allow threads with pending tokens to compete 730 * for them (a lower priority thread will get less cpu once it 731 * returns from the kernel anyway). If a thread does not have 732 * any contending tokens, we go by td_pri and upri. 733 */ 734 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) { 735 if (TD_TOKS_NOT_HELD(ntd) && 736 ntd->td_pri < TDPRI_KERN_LPSCHED && upri > ntd->td_upri) { 737 continue; 738 } 739 if (upri < ntd->td_upri) 740 upri = ntd->td_upri; 741 742 /* 743 * Try this one. 744 */ 745 if (TD_TOKS_NOT_HELD(ntd) || 746 lwkt_getalltokens(ntd, (ntd->td_contended > lwkt_spin_loops))) { 747 goto havethread; 748 } 749 ++ntd->td_contended; /* overflow ok */ 750 } 751 752 /* 753 * Fall through, switch to idle thread to get us out of the current 754 * context. Since we were contended, prevent HLT by flagging a 755 * LWKT reschedule. 756 */ 757 need_lwkt_resched(); 758 } 759 760 /* 761 * We either contended on ntd or the runq is empty. We must switch 762 * through the idle thread to get out of the current context. 763 */ 764 ntd = &gd->gd_idlethread; 765 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) 766 ASSERT_NO_TOKENS_HELD(ntd); 767 cpu_time.cp_msg[0] = 0; 768 goto haveidle; 769 770 havethread: 771 /* 772 * Clear gd_idle_repeat when doing a normal switch to a non-idle 773 * thread. 774 */ 775 ntd->td_wmesg = NULL; 776 ntd->td_contended = 0; /* reset once scheduled */ 777 ++gd->gd_cnt.v_swtch; 778 gd->gd_idle_repeat = 0; 779 780 /* 781 * If we were busy waiting record final disposition 782 */ 783 if (gd->gd_indefinite.type) 784 indefinite_done(&gd->gd_indefinite); 785 786 havethread_preempted: 787 /* 788 * If the new target does not need the MP lock and we are holding it, 789 * release the MP lock. If the new target requires the MP lock we have 790 * already acquired it for the target. 791 */ 792 ; 793 haveidle: 794 KASSERT(ntd->td_critcount, 795 ("priority problem in lwkt_switch %d %d", 796 td->td_critcount, ntd->td_critcount)); 797 798 if (td != ntd) { 799 /* 800 * Execute the actual thread switch operation. This function 801 * returns to the current thread and returns the previous thread 802 * (which may be different from the thread we switched to). 803 * 804 * We are responsible for marking ntd as TDF_RUNNING. 805 */ 806 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0); 807 ++switch_count; 808 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd); 809 ntd->td_flags |= TDF_RUNNING; 810 lwkt_switch_return(td->td_switch(ntd)); 811 /* ntd invalid, td_switch() can return a different thread_t */ 812 } 813 814 /* 815 * catch-all. XXX is this strictly needed? 816 */ 817 splz_check(); 818 819 /* NOTE: current cpu may have changed after switch */ 820 crit_exit_quick(td); 821 } 822 823 /* 824 * Called by assembly in the td_switch (thread restore path) for thread 825 * bootstrap cases which do not 'return' to lwkt_switch(). 826 */ 827 void 828 lwkt_switch_return(thread_t otd) 829 { 830 globaldata_t rgd; 831 #ifdef LOOPMASK 832 uint64_t tsc_base = rdtsc(); 833 #endif 834 int exiting; 835 836 exiting = otd->td_flags & TDF_EXITING; 837 cpu_ccfence(); 838 839 /* 840 * Check if otd was migrating. Now that we are on ntd we can finish 841 * up the migration. This is a bit messy but it is the only place 842 * where td is known to be fully descheduled. 843 * 844 * We can only activate the migration if otd was migrating but not 845 * held on the cpu due to a preemption chain. We still have to 846 * clear TDF_RUNNING on the old thread either way. 847 * 848 * We are responsible for clearing the previously running thread's 849 * TDF_RUNNING. 850 */ 851 if ((rgd = otd->td_migrate_gd) != NULL && 852 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) { 853 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) == 854 (TDF_MIGRATING | TDF_RUNNING)); 855 otd->td_migrate_gd = NULL; 856 otd->td_flags &= ~TDF_RUNNING; 857 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd); 858 } else { 859 otd->td_flags &= ~TDF_RUNNING; 860 } 861 862 /* 863 * Final exit validations (see lwp_wait()). Note that otd becomes 864 * invalid the *instant* we set TDF_MP_EXITSIG. 865 * 866 * Use the EXITING status loaded from before we clear TDF_RUNNING, 867 * because if it is not set otd becomes invalid the instant we clear 868 * TDF_RUNNING on it (otherwise, if the system is fast enough, we 869 * might 'steal' TDF_EXITING from another switch-return!). 870 */ 871 while (exiting) { 872 u_int mpflags; 873 874 mpflags = otd->td_mpflags; 875 cpu_ccfence(); 876 877 if (mpflags & TDF_MP_EXITWAIT) { 878 if (atomic_cmpset_int(&otd->td_mpflags, mpflags, 879 mpflags | TDF_MP_EXITSIG)) { 880 wakeup(otd); 881 break; 882 } 883 } else { 884 if (atomic_cmpset_int(&otd->td_mpflags, mpflags, 885 mpflags | TDF_MP_EXITSIG)) { 886 wakeup(otd); 887 break; 888 } 889 } 890 891 #ifdef LOOPMASK 892 if (tsc_frequency && rdtsc() - tsc_base > tsc_frequency) { 893 kprintf("lwkt_switch_return: excessive TDF_EXITING " 894 "thread %p\n", otd); 895 tsc_base = rdtsc(); 896 } 897 #endif 898 } 899 } 900 901 /* 902 * Request that the target thread preempt the current thread. Preemption 903 * can only occur only: 904 * 905 * - If our critical section is the one that we were called with 906 * - The relative priority of the target thread is higher 907 * - The target is not excessively interrupt-nested via td_nest_count 908 * - The target thread holds no tokens. 909 * - The target thread is not already scheduled and belongs to the 910 * current cpu. 911 * - The current thread is not holding any spin-locks. 912 * 913 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically 914 * this is called via lwkt_schedule() through the td_preemptable callback. 915 * critcount is the managed critical priority that we should ignore in order 916 * to determine whether preemption is possible (aka usually just the crit 917 * priority of lwkt_schedule() itself). 918 * 919 * Preemption is typically limited to interrupt threads. 920 * 921 * Operation works in a fairly straight-forward manner. The normal 922 * scheduling code is bypassed and we switch directly to the target 923 * thread. When the target thread attempts to block or switch away 924 * code at the base of lwkt_switch() will switch directly back to our 925 * thread. Our thread is able to retain whatever tokens it holds and 926 * if the target needs one of them the target will switch back to us 927 * and reschedule itself normally. 928 */ 929 void 930 lwkt_preempt(thread_t ntd, int critcount) 931 { 932 struct globaldata *gd = mycpu; 933 thread_t xtd; 934 thread_t td; 935 int save_gd_intr_nesting_level; 936 937 /* 938 * The caller has put us in a critical section. We can only preempt 939 * if the caller of the caller was not in a critical section (basically 940 * a local interrupt), as determined by the 'critcount' parameter. We 941 * also can't preempt if the caller is holding any spinlocks (even if 942 * he isn't in a critical section). This also handles the tokens test. 943 * 944 * YYY The target thread must be in a critical section (else it must 945 * inherit our critical section? I dunno yet). 946 */ 947 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri)); 948 949 td = gd->gd_curthread; 950 if (preempt_enable == 0) { 951 ++preempt_miss; 952 return; 953 } 954 if (ntd->td_pri <= td->td_pri) { 955 ++preempt_miss; 956 return; 957 } 958 if (td->td_critcount > critcount) { 959 ++preempt_miss; 960 return; 961 } 962 if (td->td_nest_count >= 2) { 963 ++preempt_miss; 964 return; 965 } 966 if (td->td_cscount) { 967 ++preempt_miss; 968 return; 969 } 970 if (ntd->td_gd != gd) { 971 ++preempt_miss; 972 return; 973 } 974 975 /* 976 * We don't have to check spinlocks here as they will also bump 977 * td_critcount. 978 * 979 * Do not try to preempt if the target thread is holding any tokens. 980 * We could try to acquire the tokens but this case is so rare there 981 * is no need to support it. 982 */ 983 KKASSERT(gd->gd_spinlocks == 0); 984 985 if (TD_TOKS_HELD(ntd)) { 986 ++preempt_miss; 987 return; 988 } 989 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) { 990 ++preempt_weird; 991 return; 992 } 993 if (ntd->td_preempted) { 994 ++preempt_hit; 995 return; 996 } 997 KKASSERT(gd->gd_processing_ipiq == 0); 998 999 /* 1000 * Since we are able to preempt the current thread, there is no need to 1001 * call need_lwkt_resched(). 1002 * 1003 * We must temporarily clear gd_intr_nesting_level around the switch 1004 * since switchouts from the target thread are allowed (they will just 1005 * return to our thread), and since the target thread has its own stack. 1006 * 1007 * A preemption must switch back to the original thread, assert the 1008 * case. 1009 */ 1010 ++preempt_hit; 1011 ntd->td_preempted = td; 1012 td->td_flags |= TDF_PREEMPT_LOCK; 1013 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd); 1014 save_gd_intr_nesting_level = gd->gd_intr_nesting_level; 1015 gd->gd_intr_nesting_level = 0; 1016 1017 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0); 1018 ntd->td_flags |= TDF_RUNNING; 1019 xtd = td->td_switch(ntd); 1020 KKASSERT(xtd == ntd); 1021 lwkt_switch_return(xtd); 1022 gd->gd_intr_nesting_level = save_gd_intr_nesting_level; 1023 1024 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE)); 1025 ntd->td_preempted = NULL; 1026 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE); 1027 } 1028 1029 /* 1030 * Conditionally call splz() if gd_reqflags indicates work is pending. 1031 * This will work inside a critical section but not inside a hard code 1032 * section. 1033 * 1034 * (self contained on a per cpu basis) 1035 */ 1036 void 1037 splz_check(void) 1038 { 1039 globaldata_t gd = mycpu; 1040 thread_t td = gd->gd_curthread; 1041 1042 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && 1043 gd->gd_intr_nesting_level == 0 && 1044 td->td_nest_count < 2) 1045 { 1046 splz(); 1047 } 1048 } 1049 1050 /* 1051 * This version is integrated into crit_exit, reqflags has already 1052 * been tested but td_critcount has not. 1053 * 1054 * We only want to execute the splz() on the 1->0 transition of 1055 * critcount and not in a hard code section or if too deeply nested. 1056 * 1057 * NOTE: gd->gd_spinlocks is implied to be 0 when td_critcount is 0. 1058 */ 1059 void 1060 lwkt_maybe_splz(thread_t td) 1061 { 1062 globaldata_t gd = td->td_gd; 1063 1064 if (td->td_critcount == 0 && 1065 gd->gd_intr_nesting_level == 0 && 1066 td->td_nest_count < 2) 1067 { 1068 splz(); 1069 } 1070 } 1071 1072 /* 1073 * Drivers which set up processing co-threads can call this function to 1074 * run the co-thread at a higher priority and to allow it to preempt 1075 * normal threads. 1076 */ 1077 void 1078 lwkt_set_interrupt_support_thread(void) 1079 { 1080 thread_t td = curthread; 1081 1082 lwkt_setpri_self(TDPRI_INT_SUPPORT); 1083 td->td_flags |= TDF_INTTHREAD; 1084 td->td_preemptable = lwkt_preempt; 1085 } 1086 1087 1088 /* 1089 * This function is used to negotiate a passive release of the current 1090 * process/lwp designation with the user scheduler, allowing the user 1091 * scheduler to schedule another user thread. The related kernel thread 1092 * (curthread) continues running in the released state. 1093 */ 1094 void 1095 lwkt_passive_release(struct thread *td) 1096 { 1097 struct lwp *lp = td->td_lwp; 1098 1099 td->td_release = NULL; 1100 lwkt_setpri_self(TDPRI_KERN_USER); 1101 1102 lp->lwp_proc->p_usched->release_curproc(lp); 1103 } 1104 1105 1106 /* 1107 * This implements a LWKT yield, allowing a kernel thread to yield to other 1108 * kernel threads at the same or higher priority. This function can be 1109 * called in a tight loop and will typically only yield once per tick. 1110 * 1111 * Most kernel threads run at the same priority in order to allow equal 1112 * sharing. 1113 * 1114 * (self contained on a per cpu basis) 1115 */ 1116 void 1117 lwkt_yield(void) 1118 { 1119 globaldata_t gd = mycpu; 1120 thread_t td = gd->gd_curthread; 1121 1122 /* 1123 * Should never be called with spinlocks held but there is a path 1124 * via ACPI where it might happen. 1125 */ 1126 if (gd->gd_spinlocks) 1127 return; 1128 1129 /* 1130 * Safe to call splz if we are not too-heavily nested. 1131 */ 1132 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1133 splz(); 1134 1135 /* 1136 * Caller allows switching 1137 */ 1138 if (lwkt_resched_wanted()) { 1139 lwkt_schedule_self(curthread); 1140 lwkt_switch(); 1141 } 1142 } 1143 1144 /* 1145 * The quick version processes pending interrupts and higher-priority 1146 * LWKT threads but will not round-robin same-priority LWKT threads. 1147 * 1148 * When called while attempting to return to userland the only same-pri 1149 * threads are the ones which have already tried to become the current 1150 * user process. 1151 */ 1152 void 1153 lwkt_yield_quick(void) 1154 { 1155 globaldata_t gd = mycpu; 1156 thread_t td = gd->gd_curthread; 1157 1158 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1159 splz(); 1160 if (lwkt_resched_wanted()) { 1161 crit_enter(); 1162 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) { 1163 clear_lwkt_resched(); 1164 } else { 1165 lwkt_schedule_self(curthread); 1166 lwkt_switch(); 1167 } 1168 crit_exit(); 1169 } 1170 } 1171 1172 /* 1173 * This yield is designed for kernel threads with a user context. 1174 * 1175 * The kernel acting on behalf of the user is potentially cpu-bound, 1176 * this function will efficiently allow other threads to run and also 1177 * switch to other processes by releasing. 1178 * 1179 * The lwkt_user_yield() function is designed to have very low overhead 1180 * if no yield is determined to be needed. 1181 */ 1182 void 1183 lwkt_user_yield(void) 1184 { 1185 globaldata_t gd = mycpu; 1186 thread_t td = gd->gd_curthread; 1187 1188 /* 1189 * Should never be called with spinlocks held but there is a path 1190 * via ACPI where it might happen. 1191 */ 1192 if (gd->gd_spinlocks) 1193 return; 1194 1195 /* 1196 * Always run any pending interrupts in case we are in a critical 1197 * section. 1198 */ 1199 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1200 splz(); 1201 1202 /* 1203 * Switch (which forces a release) if another kernel thread needs 1204 * the cpu, if userland wants us to resched, or if our kernel 1205 * quantum has run out. 1206 */ 1207 if (lwkt_resched_wanted() || 1208 user_resched_wanted()) 1209 { 1210 lwkt_switch(); 1211 } 1212 1213 #if 0 1214 /* 1215 * Reacquire the current process if we are released. 1216 * 1217 * XXX not implemented atm. The kernel may be holding locks and such, 1218 * so we want the thread to continue to receive cpu. 1219 */ 1220 if (td->td_release == NULL && lp) { 1221 lp->lwp_proc->p_usched->acquire_curproc(lp); 1222 td->td_release = lwkt_passive_release; 1223 lwkt_setpri_self(TDPRI_USER_NORM); 1224 } 1225 #endif 1226 } 1227 1228 /* 1229 * Generic schedule. Possibly schedule threads belonging to other cpus and 1230 * deal with threads that might be blocked on a wait queue. 1231 * 1232 * We have a little helper inline function which does additional work after 1233 * the thread has been enqueued, including dealing with preemption and 1234 * setting need_lwkt_resched() (which prevents the kernel from returning 1235 * to userland until it has processed higher priority threads). 1236 * 1237 * It is possible for this routine to be called after a failed _enqueue 1238 * (due to the target thread migrating, sleeping, or otherwise blocked). 1239 * We have to check that the thread is actually on the run queue! 1240 */ 1241 static __inline 1242 void 1243 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount) 1244 { 1245 if (ntd->td_flags & TDF_RUNQ) { 1246 if (ntd->td_preemptable) { 1247 ntd->td_preemptable(ntd, ccount); /* YYY +token */ 1248 } 1249 } 1250 } 1251 1252 static __inline 1253 void 1254 _lwkt_schedule(thread_t td) 1255 { 1256 globaldata_t mygd = mycpu; 1257 1258 KASSERT(td != &td->td_gd->gd_idlethread, 1259 ("lwkt_schedule(): scheduling gd_idlethread is illegal!")); 1260 KKASSERT((td->td_flags & TDF_MIGRATING) == 0); 1261 crit_enter_gd(mygd); 1262 KKASSERT(td->td_lwp == NULL || 1263 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 1264 1265 if (td == mygd->gd_curthread) { 1266 _lwkt_enqueue(td); 1267 } else { 1268 /* 1269 * If we own the thread, there is no race (since we are in a 1270 * critical section). If we do not own the thread there might 1271 * be a race but the target cpu will deal with it. 1272 */ 1273 if (td->td_gd == mygd) { 1274 _lwkt_enqueue(td); 1275 _lwkt_schedule_post(mygd, td, 1); 1276 } else { 1277 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0); 1278 } 1279 } 1280 crit_exit_gd(mygd); 1281 } 1282 1283 void 1284 lwkt_schedule(thread_t td) 1285 { 1286 _lwkt_schedule(td); 1287 } 1288 1289 void 1290 lwkt_schedule_noresched(thread_t td) /* XXX not impl */ 1291 { 1292 _lwkt_schedule(td); 1293 } 1294 1295 /* 1296 * When scheduled remotely if frame != NULL the IPIQ is being 1297 * run via doreti or an interrupt then preemption can be allowed. 1298 * 1299 * To allow preemption we have to drop the critical section so only 1300 * one is present in _lwkt_schedule_post. 1301 */ 1302 static void 1303 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame) 1304 { 1305 thread_t td = curthread; 1306 thread_t ntd = arg; 1307 1308 if (frame && ntd->td_preemptable) { 1309 crit_exit_noyield(td); 1310 _lwkt_schedule(ntd); 1311 crit_enter_quick(td); 1312 } else { 1313 _lwkt_schedule(ntd); 1314 } 1315 } 1316 1317 /* 1318 * Thread migration using a 'Pull' method. The thread may or may not be 1319 * the current thread. It MUST be descheduled and in a stable state. 1320 * lwkt_giveaway() must be called on the cpu owning the thread. 1321 * 1322 * At any point after lwkt_giveaway() is called, the target cpu may 1323 * 'pull' the thread by calling lwkt_acquire(). 1324 * 1325 * We have to make sure the thread is not sitting on a per-cpu tsleep 1326 * queue or it will blow up when it moves to another cpu. 1327 * 1328 * MPSAFE - must be called under very specific conditions. 1329 */ 1330 void 1331 lwkt_giveaway(thread_t td) 1332 { 1333 globaldata_t gd = mycpu; 1334 1335 crit_enter_gd(gd); 1336 if (td->td_flags & TDF_TSLEEPQ) 1337 tsleep_remove(td); 1338 KKASSERT(td->td_gd == gd); 1339 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); 1340 td->td_flags |= TDF_MIGRATING; 1341 crit_exit_gd(gd); 1342 } 1343 1344 void 1345 lwkt_acquire(thread_t td) 1346 { 1347 globaldata_t gd; 1348 globaldata_t mygd; 1349 1350 KKASSERT(td->td_flags & TDF_MIGRATING); 1351 gd = td->td_gd; 1352 mygd = mycpu; 1353 if (gd != mycpu) { 1354 #ifdef LOOPMASK 1355 uint64_t tsc_base = rdtsc(); 1356 #endif 1357 cpu_lfence(); 1358 KKASSERT((td->td_flags & TDF_RUNQ) == 0); 1359 crit_enter_gd(mygd); 1360 DEBUG_PUSH_INFO("lwkt_acquire"); 1361 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) { 1362 lwkt_process_ipiq(); 1363 cpu_lfence(); 1364 #ifdef _KERNEL_VIRTUAL 1365 pthread_yield(); 1366 #endif 1367 #ifdef LOOPMASK 1368 if (tsc_frequency && rdtsc() - tsc_base > tsc_frequency) { 1369 kprintf("lwkt_acquire: stuck td %p td->td_flags %08x\n", 1370 td, td->td_flags); 1371 tsc_base = rdtsc(); 1372 } 1373 #endif 1374 } 1375 DEBUG_POP_INFO(); 1376 cpu_mfence(); 1377 td->td_gd = mygd; 1378 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1379 td->td_flags &= ~TDF_MIGRATING; 1380 crit_exit_gd(mygd); 1381 } else { 1382 crit_enter_gd(mygd); 1383 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1384 td->td_flags &= ~TDF_MIGRATING; 1385 crit_exit_gd(mygd); 1386 } 1387 } 1388 1389 /* 1390 * Generic deschedule. Descheduling threads other then your own should be 1391 * done only in carefully controlled circumstances. Descheduling is 1392 * asynchronous. 1393 * 1394 * This function may block if the cpu has run out of messages. 1395 */ 1396 void 1397 lwkt_deschedule(thread_t td) 1398 { 1399 crit_enter(); 1400 if (td == curthread) { 1401 _lwkt_dequeue(td); 1402 } else { 1403 if (td->td_gd == mycpu) { 1404 _lwkt_dequeue(td); 1405 } else { 1406 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td); 1407 } 1408 } 1409 crit_exit(); 1410 } 1411 1412 /* 1413 * Set the target thread's priority. This routine does not automatically 1414 * switch to a higher priority thread, LWKT threads are not designed for 1415 * continuous priority changes. Yield if you want to switch. 1416 */ 1417 void 1418 lwkt_setpri(thread_t td, int pri) 1419 { 1420 if (td->td_pri != pri) { 1421 KKASSERT(pri >= 0); 1422 crit_enter(); 1423 if (td->td_flags & TDF_RUNQ) { 1424 KKASSERT(td->td_gd == mycpu); 1425 _lwkt_dequeue(td); 1426 td->td_pri = pri; 1427 _lwkt_enqueue(td); 1428 } else { 1429 td->td_pri = pri; 1430 } 1431 crit_exit(); 1432 } 1433 } 1434 1435 /* 1436 * Set the initial priority for a thread prior to it being scheduled for 1437 * the first time. The thread MUST NOT be scheduled before or during 1438 * this call. The thread may be assigned to a cpu other then the current 1439 * cpu. 1440 * 1441 * Typically used after a thread has been created with TDF_STOPPREQ, 1442 * and before the thread is initially scheduled. 1443 */ 1444 void 1445 lwkt_setpri_initial(thread_t td, int pri) 1446 { 1447 KKASSERT(pri >= 0); 1448 KKASSERT((td->td_flags & TDF_RUNQ) == 0); 1449 td->td_pri = pri; 1450 } 1451 1452 void 1453 lwkt_setpri_self(int pri) 1454 { 1455 thread_t td = curthread; 1456 1457 KKASSERT(pri >= 0 && pri <= TDPRI_MAX); 1458 crit_enter(); 1459 if (td->td_flags & TDF_RUNQ) { 1460 _lwkt_dequeue(td); 1461 td->td_pri = pri; 1462 _lwkt_enqueue(td); 1463 } else { 1464 td->td_pri = pri; 1465 } 1466 crit_exit(); 1467 } 1468 1469 /* 1470 * hz tick scheduler clock for LWKT threads 1471 */ 1472 void 1473 lwkt_schedulerclock(thread_t td) 1474 { 1475 globaldata_t gd = td->td_gd; 1476 thread_t xtd; 1477 1478 xtd = TAILQ_FIRST(&gd->gd_tdrunq); 1479 if (xtd == td) { 1480 /* 1481 * If the current thread is at the head of the runq shift it to the 1482 * end of any equal-priority threads and request a LWKT reschedule 1483 * if it moved. 1484 * 1485 * Ignore upri in this situation. There will only be one user thread 1486 * in user mode, all others will be user threads running in kernel 1487 * mode and we have to make sure they get some cpu. 1488 */ 1489 xtd = TAILQ_NEXT(td, td_threadq); 1490 if (xtd && xtd->td_pri == td->td_pri) { 1491 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq); 1492 while (xtd && xtd->td_pri == td->td_pri) 1493 xtd = TAILQ_NEXT(xtd, td_threadq); 1494 if (xtd) 1495 TAILQ_INSERT_BEFORE(xtd, td, td_threadq); 1496 else 1497 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 1498 need_lwkt_resched(); 1499 } 1500 } else if (xtd) { 1501 /* 1502 * If we scheduled a thread other than the one at the head of the 1503 * queue always request a reschedule every tick. 1504 */ 1505 need_lwkt_resched(); 1506 } 1507 /* else curthread probably the idle thread, no need to reschedule */ 1508 } 1509 1510 /* 1511 * Migrate the current thread to the specified cpu. 1512 * 1513 * This is accomplished by descheduling ourselves from the current cpu 1514 * and setting td_migrate_gd. The lwkt_switch() code will detect that the 1515 * 'old' thread wants to migrate after it has been completely switched out 1516 * and will complete the migration. 1517 * 1518 * TDF_MIGRATING prevents scheduling races while the thread is being migrated. 1519 * 1520 * We must be sure to release our current process designation (if a user 1521 * process) before clearing out any tsleepq we are on because the release 1522 * code may re-add us. 1523 * 1524 * We must be sure to remove ourselves from the current cpu's tsleepq 1525 * before potentially moving to another queue. The thread can be on 1526 * a tsleepq due to a left-over tsleep_interlock(). 1527 */ 1528 1529 void 1530 lwkt_setcpu_self(globaldata_t rgd) 1531 { 1532 thread_t td = curthread; 1533 1534 if (td->td_gd != rgd) { 1535 crit_enter_quick(td); 1536 1537 if (td->td_release) 1538 td->td_release(td); 1539 if (td->td_flags & TDF_TSLEEPQ) 1540 tsleep_remove(td); 1541 1542 /* 1543 * Set TDF_MIGRATING to prevent a spurious reschedule while we are 1544 * trying to deschedule ourselves and switch away, then deschedule 1545 * ourself, remove us from tdallq, and set td_migrate_gd. Finally, 1546 * call lwkt_switch() to complete the operation. 1547 */ 1548 td->td_flags |= TDF_MIGRATING; 1549 lwkt_deschedule_self(td); 1550 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1551 td->td_migrate_gd = rgd; 1552 lwkt_switch(); 1553 1554 /* 1555 * We are now on the target cpu 1556 */ 1557 KKASSERT(rgd == mycpu); 1558 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); 1559 crit_exit_quick(td); 1560 } 1561 } 1562 1563 void 1564 lwkt_migratecpu(int cpuid) 1565 { 1566 globaldata_t rgd; 1567 1568 rgd = globaldata_find(cpuid); 1569 lwkt_setcpu_self(rgd); 1570 } 1571 1572 /* 1573 * Remote IPI for cpu migration (called while in a critical section so we 1574 * do not have to enter another one). 1575 * 1576 * The thread (td) has already been completely descheduled from the 1577 * originating cpu and we can simply assert the case. The thread is 1578 * assigned to the new cpu and enqueued. 1579 * 1580 * The thread will re-add itself to tdallq when it resumes execution. 1581 */ 1582 static void 1583 lwkt_setcpu_remote(void *arg) 1584 { 1585 thread_t td = arg; 1586 globaldata_t gd = mycpu; 1587 1588 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0); 1589 td->td_gd = gd; 1590 cpu_mfence(); 1591 td->td_flags &= ~TDF_MIGRATING; 1592 KKASSERT(td->td_migrate_gd == NULL); 1593 KKASSERT(td->td_lwp == NULL || 1594 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 1595 _lwkt_enqueue(td); 1596 } 1597 1598 struct lwp * 1599 lwkt_preempted_proc(void) 1600 { 1601 thread_t td = curthread; 1602 while (td->td_preempted) 1603 td = td->td_preempted; 1604 return(td->td_lwp); 1605 } 1606 1607 /* 1608 * Create a kernel process/thread/whatever. It shares it's address space 1609 * with proc0 - ie: kernel only. 1610 * 1611 * If the cpu is not specified one will be selected. In the future 1612 * specifying a cpu of -1 will enable kernel thread migration between 1613 * cpus. 1614 */ 1615 int 1616 lwkt_create(void (*func)(void *), void *arg, struct thread **tdp, 1617 thread_t template, int tdflags, int cpu, const char *fmt, ...) 1618 { 1619 thread_t td; 1620 __va_list ap; 1621 1622 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu, 1623 tdflags); 1624 if (tdp) 1625 *tdp = td; 1626 cpu_set_thread_handler(td, lwkt_exit, func, arg); 1627 1628 /* 1629 * Set up arg0 for 'ps' etc 1630 */ 1631 __va_start(ap, fmt); 1632 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap); 1633 __va_end(ap); 1634 1635 /* 1636 * Schedule the thread to run 1637 */ 1638 if (td->td_flags & TDF_NOSTART) 1639 td->td_flags &= ~TDF_NOSTART; 1640 else 1641 lwkt_schedule(td); 1642 return 0; 1643 } 1644 1645 /* 1646 * Destroy an LWKT thread. Warning! This function is not called when 1647 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and 1648 * uses a different reaping mechanism. 1649 */ 1650 void 1651 lwkt_exit(void) 1652 { 1653 thread_t td = curthread; 1654 thread_t std; 1655 globaldata_t gd; 1656 1657 /* 1658 * Do any cleanup that might block here 1659 */ 1660 if (td->td_flags & TDF_VERBOSE) 1661 kprintf("kthread %p %s has exited\n", td, td->td_comm); 1662 biosched_done(td); 1663 dsched_exit_thread(td); 1664 1665 /* 1666 * Get us into a critical section to interlock gd_freetd and loop 1667 * until we can get it freed. 1668 * 1669 * We have to cache the current td in gd_freetd because objcache_put()ing 1670 * it would rip it out from under us while our thread is still active. 1671 * 1672 * We are the current thread so of course our own TDF_RUNNING bit will 1673 * be set, so unlike the lwp reap code we don't wait for it to clear. 1674 */ 1675 gd = mycpu; 1676 crit_enter_quick(td); 1677 for (;;) { 1678 if (td->td_refs) { 1679 tsleep(td, 0, "tdreap", 1); 1680 continue; 1681 } 1682 if ((std = gd->gd_freetd) != NULL) { 1683 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0); 1684 gd->gd_freetd = NULL; 1685 objcache_put(thread_cache, std); 1686 continue; 1687 } 1688 break; 1689 } 1690 1691 /* 1692 * Remove thread resources from kernel lists and deschedule us for 1693 * the last time. We cannot block after this point or we may end 1694 * up with a stale td on the tsleepq. 1695 * 1696 * None of this may block, the critical section is the only thing 1697 * protecting tdallq and the only thing preventing new lwkt_hold() 1698 * thread refs now. 1699 */ 1700 if (td->td_flags & TDF_TSLEEPQ) 1701 tsleep_remove(td); 1702 lwkt_deschedule_self(td); 1703 lwkt_remove_tdallq(td); 1704 KKASSERT(td->td_refs == 0); 1705 1706 /* 1707 * Final cleanup 1708 */ 1709 KKASSERT(gd->gd_freetd == NULL); 1710 if (td->td_flags & TDF_ALLOCATED_THREAD) 1711 gd->gd_freetd = td; 1712 cpu_thread_exit(); 1713 } 1714 1715 void 1716 lwkt_remove_tdallq(thread_t td) 1717 { 1718 KKASSERT(td->td_gd == mycpu); 1719 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1720 } 1721 1722 /* 1723 * Code reduction and branch prediction improvements. Call/return 1724 * overhead on modern cpus often degenerates into 0 cycles due to 1725 * the cpu's branch prediction hardware and return pc cache. We 1726 * can take advantage of this by not inlining medium-complexity 1727 * functions and we can also reduce the branch prediction impact 1728 * by collapsing perfectly predictable branches into a single 1729 * procedure instead of duplicating it. 1730 * 1731 * Is any of this noticeable? Probably not, so I'll take the 1732 * smaller code size. 1733 */ 1734 void 1735 crit_exit_wrapper(__DEBUG_CRIT_ARG__) 1736 { 1737 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__); 1738 } 1739 1740 void 1741 crit_panic(void) 1742 { 1743 thread_t td = curthread; 1744 int lcrit = td->td_critcount; 1745 1746 td->td_critcount = 0; 1747 cpu_ccfence(); 1748 panic("td_critcount is/would-go negative! %p %d", td, lcrit); 1749 /* NOT REACHED */ 1750 } 1751 1752 /* 1753 * Called from debugger/panic on cpus which have been stopped. We must still 1754 * process the IPIQ while stopped. 1755 * 1756 * If we are dumping also try to process any pending interrupts. This may 1757 * or may not work depending on the state of the cpu at the point it was 1758 * stopped. 1759 */ 1760 void 1761 lwkt_smp_stopped(void) 1762 { 1763 globaldata_t gd = mycpu; 1764 1765 if (dumping) { 1766 lwkt_process_ipiq(); 1767 --gd->gd_intr_nesting_level; 1768 splz(); 1769 ++gd->gd_intr_nesting_level; 1770 } else { 1771 lwkt_process_ipiq(); 1772 } 1773 cpu_smp_stopped(); 1774 } 1775