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 #ifdef DEBUG_LWKT_THREAD 97 static int64_t switch_count = 0; 98 static int64_t preempt_hit = 0; 99 static int64_t preempt_miss = 0; 100 static int64_t preempt_weird = 0; 101 #endif 102 static int lwkt_use_spin_port; 103 __read_mostly static struct objcache *thread_cache; 104 int cpu_mwait_spin = 0; 105 106 static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame); 107 static void lwkt_setcpu_remote(void *arg); 108 109 /* 110 * We can make all thread ports use the spin backend instead of the thread 111 * backend. This should only be set to debug the spin backend. 112 */ 113 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port); 114 115 #ifdef INVARIANTS 116 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, 117 "Panic if attempting to switch lwkt's while mastering cpusync"); 118 #endif 119 #ifdef DEBUG_LWKT_THREAD 120 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, 121 "Number of switched threads"); 122 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, 123 "Successful preemption events"); 124 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, 125 "Failed preemption events"); 126 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, 127 "Number of preempted threads."); 128 #endif 129 extern int lwkt_sched_debug; 130 int lwkt_sched_debug = 0; 131 SYSCTL_INT(_lwkt, OID_AUTO, sched_debug, CTLFLAG_RW, 132 &lwkt_sched_debug, 0, "Scheduler debug"); 133 __read_mostly static u_int lwkt_spin_loops = 10; 134 SYSCTL_UINT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW, 135 &lwkt_spin_loops, 0, "Scheduler spin loops until sorted decon"); 136 __read_mostly static int preempt_enable = 1; 137 SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW, 138 &preempt_enable, 0, "Enable preemption"); 139 static int lwkt_cache_threads = 0; 140 SYSCTL_INT(_lwkt, OID_AUTO, cache_threads, CTLFLAG_RD, 141 &lwkt_cache_threads, 0, "thread+kstack cache"); 142 143 /* 144 * These helper procedures handle the runq, they can only be called from 145 * within a critical section. 146 * 147 * WARNING! Prior to SMP being brought up it is possible to enqueue and 148 * dequeue threads belonging to other cpus, so be sure to use td->td_gd 149 * instead of 'mycpu' when referencing the globaldata structure. Once 150 * SMP live enqueuing and dequeueing only occurs on the current cpu. 151 */ 152 static __inline 153 void 154 _lwkt_dequeue(thread_t td) 155 { 156 if (td->td_flags & TDF_RUNQ) { 157 struct globaldata *gd = td->td_gd; 158 159 td->td_flags &= ~TDF_RUNQ; 160 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq); 161 --gd->gd_tdrunqcount; 162 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL) 163 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING); 164 } 165 } 166 167 /* 168 * Priority enqueue. 169 * 170 * There are a limited number of lwkt threads runnable since user 171 * processes only schedule one at a time per cpu. However, there can 172 * be many user processes in kernel mode exiting from a tsleep() which 173 * become runnable. 174 * 175 * We scan the queue in both directions to help deal with degenerate 176 * situations when hundreds or thousands (or more) threads are runnable. 177 * 178 * NOTE: lwkt_schedulerclock() will force a round-robin based on td_pri and 179 * will ignore user priority. This is to ensure that user threads in 180 * kernel mode get cpu at some point regardless of what the user 181 * scheduler thinks. 182 */ 183 static __inline 184 void 185 _lwkt_enqueue(thread_t td) 186 { 187 thread_t xtd; /* forward scan */ 188 thread_t rtd; /* reverse scan */ 189 190 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) { 191 struct globaldata *gd = td->td_gd; 192 193 td->td_flags |= TDF_RUNQ; 194 xtd = TAILQ_FIRST(&gd->gd_tdrunq); 195 if (xtd == NULL) { 196 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 197 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING); 198 } else { 199 /* 200 * NOTE: td_upri - higher numbers more desireable, same sense 201 * as td_pri (typically reversed from lwp_upri). 202 * 203 * In the equal priority case we want the best selection 204 * at the beginning so the less desireable selections know 205 * that they have to setrunqueue/go-to-another-cpu, even 206 * though it means switching back to the 'best' selection. 207 * This also avoids degenerate situations when many threads 208 * are runnable or waking up at the same time. 209 * 210 * If upri matches exactly place at end/round-robin. 211 */ 212 rtd = TAILQ_LAST(&gd->gd_tdrunq, lwkt_queue); 213 214 while (xtd && 215 (xtd->td_pri > td->td_pri || 216 (xtd->td_pri == td->td_pri && 217 xtd->td_upri >= td->td_upri))) { 218 xtd = TAILQ_NEXT(xtd, td_threadq); 219 220 /* 221 * Doing a reverse scan at the same time is an optimization 222 * for the insert-closer-to-tail case that avoids having to 223 * scan the entire list. This situation can occur when 224 * thousands of threads are woken up at the same time. 225 */ 226 if (rtd->td_pri > td->td_pri || 227 (rtd->td_pri == td->td_pri && 228 rtd->td_upri >= td->td_upri)) { 229 TAILQ_INSERT_AFTER(&gd->gd_tdrunq, rtd, td, td_threadq); 230 goto skip; 231 } 232 rtd = TAILQ_PREV(rtd, lwkt_queue, td_threadq); 233 } 234 if (xtd) 235 TAILQ_INSERT_BEFORE(xtd, td, td_threadq); 236 else 237 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 238 } 239 skip: 240 ++gd->gd_tdrunqcount; 241 242 /* 243 * Request a LWKT reschedule if we are now at the head of the queue. 244 */ 245 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) 246 need_lwkt_resched(); 247 } 248 } 249 250 static boolean_t 251 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags) 252 { 253 struct thread *td = (struct thread *)obj; 254 255 td->td_kstack = NULL; 256 td->td_kstack_size = 0; 257 td->td_flags = TDF_ALLOCATED_THREAD; 258 td->td_mpflags = 0; 259 return (1); 260 } 261 262 static void 263 _lwkt_thread_dtor(void *obj, void *privdata) 264 { 265 struct thread *td = (struct thread *)obj; 266 267 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD, 268 ("_lwkt_thread_dtor: not allocated from objcache")); 269 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack && 270 td->td_kstack_size > 0, 271 ("_lwkt_thread_dtor: corrupted stack")); 272 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size); 273 td->td_kstack = NULL; 274 td->td_flags = 0; 275 } 276 277 /* 278 * Initialize the lwkt s/system. 279 * 280 * Nominally cache up to 32 thread + kstack structures. Cache more on 281 * systems with a lot of cpu cores. 282 */ 283 static void 284 lwkt_init(void) 285 { 286 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads); 287 if (lwkt_cache_threads == 0) { 288 lwkt_cache_threads = ncpus * 4; 289 if (lwkt_cache_threads < 32) 290 lwkt_cache_threads = 32; 291 } 292 thread_cache = objcache_create_mbacked( 293 M_THREAD, sizeof(struct thread), 294 0, lwkt_cache_threads, 295 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL); 296 } 297 SYSINIT(lwkt_init, SI_BOOT2_LWKT_INIT, SI_ORDER_FIRST, lwkt_init, NULL); 298 299 /* 300 * Schedule a thread to run. As the current thread we can always safely 301 * schedule ourselves, and a shortcut procedure is provided for that 302 * function. 303 * 304 * (non-blocking, self contained on a per cpu basis) 305 */ 306 void 307 lwkt_schedule_self(thread_t td) 308 { 309 KKASSERT((td->td_flags & TDF_MIGRATING) == 0); 310 crit_enter_quick(td); 311 KASSERT(td != &td->td_gd->gd_idlethread, 312 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!")); 313 KKASSERT(td->td_lwp == NULL || 314 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 315 _lwkt_enqueue(td); 316 crit_exit_quick(td); 317 } 318 319 /* 320 * Deschedule a thread. 321 * 322 * (non-blocking, self contained on a per cpu basis) 323 */ 324 void 325 lwkt_deschedule_self(thread_t td) 326 { 327 crit_enter_quick(td); 328 _lwkt_dequeue(td); 329 crit_exit_quick(td); 330 } 331 332 /* 333 * LWKTs operate on a per-cpu basis 334 * 335 * WARNING! Called from early boot, 'mycpu' may not work yet. 336 */ 337 void 338 lwkt_gdinit(struct globaldata *gd) 339 { 340 TAILQ_INIT(&gd->gd_tdrunq); 341 TAILQ_INIT(&gd->gd_tdallq); 342 lockinit(&gd->gd_sysctllock, "sysctl", 0, LK_CANRECURSE); 343 } 344 345 /* 346 * Create a new thread. The thread must be associated with a process context 347 * or LWKT start address before it can be scheduled. If the target cpu is 348 * -1 the thread will be created on the current cpu. 349 * 350 * If you intend to create a thread without a process context this function 351 * does everything except load the startup and switcher function. 352 */ 353 thread_t 354 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags) 355 { 356 static int cpu_rotator; 357 globaldata_t gd = mycpu; 358 void *stack; 359 360 /* 361 * If static thread storage is not supplied allocate a thread. Reuse 362 * a cached free thread if possible. gd_freetd is used to keep an exiting 363 * thread intact through the exit. 364 */ 365 if (td == NULL) { 366 crit_enter_gd(gd); 367 if ((td = gd->gd_freetd) != NULL) { 368 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK| 369 TDF_RUNQ)) == 0); 370 gd->gd_freetd = NULL; 371 } else { 372 td = objcache_get(thread_cache, M_WAITOK); 373 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK| 374 TDF_RUNQ)) == 0); 375 } 376 crit_exit_gd(gd); 377 KASSERT((td->td_flags & 378 (TDF_ALLOCATED_THREAD|TDF_RUNNING|TDF_PREEMPT_LOCK)) == 379 TDF_ALLOCATED_THREAD, 380 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags)); 381 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK); 382 } 383 384 /* 385 * Try to reuse cached stack. 386 */ 387 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) { 388 if (flags & TDF_ALLOCATED_STACK) { 389 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size); 390 stack = NULL; 391 } 392 } 393 if (stack == NULL) { 394 if (cpu < 0) 395 stack = (void *)kmem_alloc_stack(&kernel_map, stksize, 0); 396 else 397 stack = (void *)kmem_alloc_stack(&kernel_map, stksize, 398 KM_CPU(cpu)); 399 flags |= TDF_ALLOCATED_STACK; 400 } 401 if (cpu < 0) { 402 cpu = ++cpu_rotator; 403 cpu_ccfence(); 404 cpu = (uint32_t)cpu % (uint32_t)ncpus; 405 } 406 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu)); 407 return(td); 408 } 409 410 /* 411 * Initialize a preexisting thread structure. This function is used by 412 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread. 413 * 414 * All threads start out in a critical section at a priority of 415 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as 416 * appropriate. This function may send an IPI message when the 417 * requested cpu is not the current cpu and consequently gd_tdallq may 418 * not be initialized synchronously from the point of view of the originating 419 * cpu. 420 * 421 * NOTE! we have to be careful in regards to creating threads for other cpus 422 * if SMP has not yet been activated. 423 */ 424 static void 425 lwkt_init_thread_remote(void *arg) 426 { 427 thread_t td = arg; 428 429 /* 430 * Protected by critical section held by IPI dispatch 431 */ 432 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq); 433 } 434 435 /* 436 * lwkt core thread structural initialization. 437 * 438 * NOTE: All threads are initialized as mpsafe threads. 439 */ 440 void 441 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags, 442 struct globaldata *gd) 443 { 444 globaldata_t mygd = mycpu; 445 446 bzero(td, sizeof(struct thread)); 447 td->td_kstack = stack; 448 td->td_kstack_size = stksize; 449 td->td_flags = flags; 450 td->td_mpflags = 0; 451 td->td_type = TD_TYPE_GENERIC; 452 td->td_gd = gd; 453 td->td_pri = TDPRI_KERN_DAEMON; 454 td->td_critcount = 1; 455 td->td_toks_have = NULL; 456 td->td_toks_stop = &td->td_toks_base; 457 if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT)) { 458 lwkt_initport_spin(&td->td_msgport, td, 459 (flags & TDF_FIXEDCPU) ? TRUE : FALSE); 460 } else { 461 lwkt_initport_thread(&td->td_msgport, td); 462 } 463 pmap_init_thread(td); 464 /* 465 * Normally initializing a thread for a remote cpu requires sending an 466 * IPI. However, the idlethread is setup before the other cpus are 467 * activated so we have to treat it as a special case. XXX manipulation 468 * of gd_tdallq requires the BGL. 469 */ 470 if (gd == mygd || td == &gd->gd_idlethread) { 471 crit_enter_gd(mygd); 472 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); 473 crit_exit_gd(mygd); 474 } else { 475 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td); 476 } 477 dsched_enter_thread(td); 478 } 479 480 void 481 lwkt_set_comm(thread_t td, const char *ctl, ...) 482 { 483 __va_list va; 484 485 __va_start(va, ctl); 486 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va); 487 __va_end(va); 488 KTR_LOG(ctxsw_newtd, td, td->td_comm); 489 } 490 491 /* 492 * Prevent the thread from getting destroyed. Note that unlike PHOLD/PRELE 493 * this does not prevent the thread from migrating to another cpu so the 494 * gd_tdallq state is not protected by this. 495 */ 496 void 497 lwkt_hold(thread_t td) 498 { 499 atomic_add_int(&td->td_refs, 1); 500 } 501 502 void 503 lwkt_rele(thread_t td) 504 { 505 KKASSERT(td->td_refs > 0); 506 atomic_add_int(&td->td_refs, -1); 507 } 508 509 void 510 lwkt_free_thread(thread_t td) 511 { 512 KKASSERT(td->td_refs == 0); 513 KKASSERT((td->td_flags & (TDF_RUNNING | TDF_PREEMPT_LOCK | 514 TDF_RUNQ | TDF_TSLEEPQ)) == 0); 515 if (td->td_flags & TDF_ALLOCATED_THREAD) { 516 objcache_put(thread_cache, td); 517 } else if (td->td_flags & TDF_ALLOCATED_STACK) { 518 /* client-allocated struct with internally allocated stack */ 519 KASSERT(td->td_kstack && td->td_kstack_size > 0, 520 ("lwkt_free_thread: corrupted stack")); 521 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size); 522 td->td_kstack = NULL; 523 td->td_kstack_size = 0; 524 } 525 526 KTR_LOG(ctxsw_deadtd, td); 527 } 528 529 530 /* 531 * Switch to the next runnable lwkt. If no LWKTs are runnable then 532 * switch to the idlethread. Switching must occur within a critical 533 * section to avoid races with the scheduling queue. 534 * 535 * We always have full control over our cpu's run queue. Other cpus 536 * that wish to manipulate our queue must use the cpu_*msg() calls to 537 * talk to our cpu, so a critical section is all that is needed and 538 * the result is very, very fast thread switching. 539 * 540 * The LWKT scheduler uses a fixed priority model and round-robins at 541 * each priority level. User process scheduling is a totally 542 * different beast and LWKT priorities should not be confused with 543 * user process priorities. 544 * 545 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch() 546 * is not called by the current thread in the preemption case, only when 547 * the preempting thread blocks (in order to return to the original thread). 548 * 549 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread 550 * migration and tsleep deschedule the current lwkt thread and call 551 * lwkt_switch(). In particular, the target cpu of the migration fully 552 * expects the thread to become non-runnable and can deadlock against 553 * cpusync operations if we run any IPIs prior to switching the thread out. 554 * 555 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF 556 * THE CURRENT THREAD HAS BEEN DESCHEDULED! 557 */ 558 void 559 lwkt_switch(void) 560 { 561 globaldata_t gd = mycpu; 562 thread_t td = gd->gd_curthread; 563 thread_t ntd; 564 thread_t xtd; 565 int upri; 566 #ifdef LOOPMASK 567 uint64_t tsc_base = rdtsc(); 568 #endif 569 570 KKASSERT(gd->gd_processing_ipiq == 0); 571 KKASSERT(td->td_flags & TDF_RUNNING); 572 573 /* 574 * Switching from within a 'fast' (non thread switched) interrupt or IPI 575 * is illegal. However, we may have to do it anyway if we hit a fatal 576 * kernel trap or we have paniced. 577 * 578 * If this case occurs save and restore the interrupt nesting level. 579 */ 580 if (gd->gd_intr_nesting_level) { 581 int savegdnest; 582 int savegdtrap; 583 584 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) { 585 panic("lwkt_switch: Attempt to switch from a " 586 "fast interrupt, ipi, or hard code section, " 587 "td %p\n", 588 td); 589 } else { 590 savegdnest = gd->gd_intr_nesting_level; 591 savegdtrap = gd->gd_trap_nesting_level; 592 gd->gd_intr_nesting_level = 0; 593 gd->gd_trap_nesting_level = 0; 594 if ((td->td_flags & TDF_PANICWARN) == 0) { 595 td->td_flags |= TDF_PANICWARN; 596 kprintf("Warning: thread switch from interrupt, IPI, " 597 "or hard code section.\n" 598 "thread %p (%s)\n", td, td->td_comm); 599 print_backtrace(-1); 600 } 601 lwkt_switch(); 602 gd->gd_intr_nesting_level = savegdnest; 603 gd->gd_trap_nesting_level = savegdtrap; 604 return; 605 } 606 } 607 608 /* 609 * Release our current user process designation if we are blocking 610 * or if a user reschedule was requested. 611 * 612 * NOTE: This function is NOT called if we are switching into or 613 * returning from a preemption. 614 * 615 * NOTE: Releasing our current user process designation may cause 616 * it to be assigned to another thread, which in turn will 617 * cause us to block in the usched acquire code when we attempt 618 * to return to userland. 619 * 620 * NOTE: On SMP systems this can be very nasty when heavy token 621 * contention is present so we want to be careful not to 622 * release the designation gratuitously. 623 */ 624 if (td->td_release && 625 (user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) { 626 td->td_release(td); 627 } 628 629 /* 630 * Release all tokens. Once we do this we must remain in the critical 631 * section and cannot run IPIs or other interrupts until we switch away 632 * because they may implode if they try to get a token using our thread 633 * context. 634 */ 635 crit_enter_gd(gd); 636 if (TD_TOKS_HELD(td)) 637 lwkt_relalltokens(td); 638 639 /* 640 * We had better not be holding any spin locks, but don't get into an 641 * endless panic loop. 642 */ 643 KASSERT(gd->gd_spinlocks == 0 || panicstr != NULL, 644 ("lwkt_switch: still holding %d exclusive spinlocks!", 645 gd->gd_spinlocks)); 646 647 #ifdef INVARIANTS 648 if (td->td_cscount) { 649 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n", 650 td); 651 if (panic_on_cscount) 652 panic("switching while mastering cpusync"); 653 } 654 #endif 655 656 /* 657 * If we had preempted another thread on this cpu, resume the preempted 658 * thread. This occurs transparently, whether the preempted thread 659 * was scheduled or not (it may have been preempted after descheduling 660 * itself). 661 * 662 * We have to setup the MP lock for the original thread after backing 663 * out the adjustment that was made to curthread when the original 664 * was preempted. 665 */ 666 if ((ntd = td->td_preempted) != NULL) { 667 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK); 668 ntd->td_flags |= TDF_PREEMPT_DONE; 669 ntd->td_contended = 0; /* reset contended */ 670 671 /* 672 * The interrupt may have woken a thread up, we need to properly 673 * set the reschedule flag if the originally interrupted thread is 674 * at a lower priority. 675 * 676 * NOTE: The interrupt may not have descheduled ntd. 677 * 678 * NOTE: We do not reschedule if there are no threads on the runq. 679 * (ntd could be the idlethread). 680 */ 681 xtd = TAILQ_FIRST(&gd->gd_tdrunq); 682 if (xtd && xtd != ntd) 683 need_lwkt_resched(); 684 goto havethread_preempted; 685 } 686 687 /* 688 * Figure out switch target. If we cannot switch to our desired target 689 * look for a thread that we can switch to. 690 * 691 * NOTE! The limited spin loop and related parameters are extremely 692 * important for system performance, particularly for pipes and 693 * concurrent conflicting VM faults. 694 */ 695 clear_lwkt_resched(); 696 ntd = TAILQ_FIRST(&gd->gd_tdrunq); 697 698 if (ntd) { 699 do { 700 if (TD_TOKS_NOT_HELD(ntd) || 701 lwkt_getalltokens(ntd, (ntd->td_contended > lwkt_spin_loops))) 702 { 703 goto havethread; 704 } 705 ++ntd->td_contended; /* overflow ok */ 706 if (gd->gd_indefinite.type == 0) 707 indefinite_init(&gd->gd_indefinite, NULL, 0, 't'); 708 #ifdef LOOPMASK 709 if (tsc_frequency && rdtsc() - tsc_base > tsc_frequency) { 710 kprintf("lwkt_switch: excessive contended %d " 711 "thread %p\n", ntd->td_contended, ntd); 712 tsc_base = rdtsc(); 713 } 714 #endif 715 } while (ntd->td_contended < (lwkt_spin_loops >> 1)); 716 upri = ntd->td_upri; 717 718 /* 719 * Bleh, the thread we wanted to switch to has a contended token. 720 * See if we can switch to another thread. 721 * 722 * We generally don't want to do this because it represents a 723 * priority inversion, but contending tokens on the same cpu can 724 * cause real problems if we don't now that we have an exclusive 725 * priority mechanism over shared for tokens. 726 * 727 * The solution is to allow threads with pending tokens to compete 728 * for them (a lower priority thread will get less cpu once it 729 * returns from the kernel anyway). If a thread does not have 730 * any contending tokens, we go by td_pri and upri. 731 */ 732 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) { 733 if (TD_TOKS_NOT_HELD(ntd) && 734 ntd->td_pri < TDPRI_KERN_LPSCHED && upri > ntd->td_upri) { 735 continue; 736 } 737 if (upri < ntd->td_upri) 738 upri = ntd->td_upri; 739 740 /* 741 * Try this one. 742 */ 743 if (TD_TOKS_NOT_HELD(ntd) || 744 lwkt_getalltokens(ntd, (ntd->td_contended > lwkt_spin_loops))) { 745 goto havethread; 746 } 747 ++ntd->td_contended; /* overflow ok */ 748 } 749 750 /* 751 * Fall through, switch to idle thread to get us out of the current 752 * context. Since we were contended, prevent HLT by flagging a 753 * LWKT reschedule. 754 */ 755 need_lwkt_resched(); 756 } 757 758 /* 759 * We either contended on ntd or the runq is empty. We must switch 760 * through the idle thread to get out of the current context. 761 */ 762 ntd = &gd->gd_idlethread; 763 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) 764 ASSERT_NO_TOKENS_HELD(ntd); 765 cpu_time.cp_msg[0] = 0; 766 goto haveidle; 767 768 havethread: 769 /* 770 * Clear gd_idle_repeat when doing a normal switch to a non-idle 771 * thread. 772 */ 773 ntd->td_wmesg = NULL; 774 ntd->td_contended = 0; /* reset once scheduled */ 775 ++gd->gd_cnt.v_swtch; 776 gd->gd_idle_repeat = 0; 777 778 /* 779 * If we were busy waiting record final disposition 780 */ 781 if (gd->gd_indefinite.type) 782 indefinite_done(&gd->gd_indefinite); 783 784 havethread_preempted: 785 /* 786 * If the new target does not need the MP lock and we are holding it, 787 * release the MP lock. If the new target requires the MP lock we have 788 * already acquired it for the target. 789 */ 790 ; 791 haveidle: 792 KASSERT(ntd->td_critcount, 793 ("priority problem in lwkt_switch %d %d", 794 td->td_critcount, ntd->td_critcount)); 795 796 if (td != ntd) { 797 /* 798 * Execute the actual thread switch operation. This function 799 * returns to the current thread and returns the previous thread 800 * (which may be different from the thread we switched to). 801 * 802 * We are responsible for marking ntd as TDF_RUNNING. 803 */ 804 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0); 805 #ifdef DEBUG_LWKT_THREAD 806 ++switch_count; 807 #endif 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 #ifdef DEBUG_LWKT_THREAD 952 ++preempt_miss; 953 #endif 954 return; 955 } 956 if (ntd->td_pri <= td->td_pri) { 957 #ifdef DEBUG_LWKT_THREAD 958 ++preempt_miss; 959 #endif 960 return; 961 } 962 if (td->td_critcount > critcount) { 963 #ifdef DEBUG_LWKT_THREAD 964 ++preempt_miss; 965 #endif 966 return; 967 } 968 if (td->td_nest_count >= 2) { 969 #ifdef DEBUG_LWKT_THREAD 970 ++preempt_miss; 971 #endif 972 return; 973 } 974 if (td->td_cscount) { 975 #ifdef DEBUG_LWKT_THREAD 976 ++preempt_miss; 977 #endif 978 return; 979 } 980 if (ntd->td_gd != gd) { 981 #ifdef DEBUG_LWKT_THREAD 982 ++preempt_miss; 983 #endif 984 return; 985 } 986 987 /* 988 * We don't have to check spinlocks here as they will also bump 989 * td_critcount. 990 * 991 * Do not try to preempt if the target thread is holding any tokens. 992 * We could try to acquire the tokens but this case is so rare there 993 * is no need to support it. 994 */ 995 KKASSERT(gd->gd_spinlocks == 0); 996 997 if (TD_TOKS_HELD(ntd)) { 998 #ifdef DEBUG_LWKT_THREAD 999 ++preempt_miss; 1000 #endif 1001 return; 1002 } 1003 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) { 1004 #ifdef DEBUG_LWKT_THREAD 1005 ++preempt_weird; 1006 #endif 1007 return; 1008 } 1009 if (ntd->td_preempted) { 1010 #ifdef DEBUG_LWKT_THREAD 1011 ++preempt_hit; 1012 #endif 1013 return; 1014 } 1015 KKASSERT(gd->gd_processing_ipiq == 0); 1016 1017 /* 1018 * Since we are able to preempt the current thread, there is no need to 1019 * call need_lwkt_resched(). 1020 * 1021 * We must temporarily clear gd_intr_nesting_level around the switch 1022 * since switchouts from the target thread are allowed (they will just 1023 * return to our thread), and since the target thread has its own stack. 1024 * 1025 * A preemption must switch back to the original thread, assert the 1026 * case. 1027 */ 1028 #ifdef DEBUG_LWKT_THREAD 1029 ++preempt_hit; 1030 #endif 1031 ntd->td_preempted = td; 1032 td->td_flags |= TDF_PREEMPT_LOCK; 1033 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd); 1034 save_gd_intr_nesting_level = gd->gd_intr_nesting_level; 1035 gd->gd_intr_nesting_level = 0; 1036 1037 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0); 1038 ntd->td_flags |= TDF_RUNNING; 1039 xtd = td->td_switch(ntd); 1040 KKASSERT(xtd == ntd); 1041 lwkt_switch_return(xtd); 1042 gd->gd_intr_nesting_level = save_gd_intr_nesting_level; 1043 1044 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE)); 1045 ntd->td_preempted = NULL; 1046 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE); 1047 } 1048 1049 /* 1050 * Conditionally call splz() if gd_reqflags indicates work is pending. 1051 * This will work inside a critical section but not inside a hard code 1052 * section. 1053 * 1054 * (self contained on a per cpu basis) 1055 */ 1056 void 1057 splz_check(void) 1058 { 1059 globaldata_t gd = mycpu; 1060 thread_t td = gd->gd_curthread; 1061 1062 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && 1063 gd->gd_intr_nesting_level == 0 && 1064 td->td_nest_count < 2) 1065 { 1066 splz(); 1067 } 1068 } 1069 1070 /* 1071 * This version is integrated into crit_exit, reqflags has already 1072 * been tested but td_critcount has not. 1073 * 1074 * We only want to execute the splz() on the 1->0 transition of 1075 * critcount and not in a hard code section or if too deeply nested. 1076 * 1077 * NOTE: gd->gd_spinlocks is implied to be 0 when td_critcount is 0. 1078 */ 1079 void 1080 lwkt_maybe_splz(thread_t td) 1081 { 1082 globaldata_t gd = td->td_gd; 1083 1084 if (td->td_critcount == 0 && 1085 gd->gd_intr_nesting_level == 0 && 1086 td->td_nest_count < 2) 1087 { 1088 splz(); 1089 } 1090 } 1091 1092 /* 1093 * Drivers which set up processing co-threads can call this function to 1094 * run the co-thread at a higher priority and to allow it to preempt 1095 * normal threads. 1096 */ 1097 void 1098 lwkt_set_interrupt_support_thread(void) 1099 { 1100 thread_t td = curthread; 1101 1102 lwkt_setpri_self(TDPRI_INT_SUPPORT); 1103 td->td_flags |= TDF_INTTHREAD; 1104 td->td_preemptable = lwkt_preempt; 1105 } 1106 1107 1108 /* 1109 * This function is used to negotiate a passive release of the current 1110 * process/lwp designation with the user scheduler, allowing the user 1111 * scheduler to schedule another user thread. The related kernel thread 1112 * (curthread) continues running in the released state. 1113 */ 1114 void 1115 lwkt_passive_release(struct thread *td) 1116 { 1117 struct lwp *lp = td->td_lwp; 1118 1119 td->td_release = NULL; 1120 lwkt_setpri_self(TDPRI_KERN_USER); 1121 1122 lp->lwp_proc->p_usched->release_curproc(lp); 1123 } 1124 1125 1126 /* 1127 * This implements a LWKT yield, allowing a kernel thread to yield to other 1128 * kernel threads at the same or higher priority. This function can be 1129 * called in a tight loop and will typically only yield once per tick. 1130 * 1131 * Most kernel threads run at the same priority in order to allow equal 1132 * sharing. 1133 * 1134 * (self contained on a per cpu basis) 1135 */ 1136 void 1137 lwkt_yield(void) 1138 { 1139 globaldata_t gd = mycpu; 1140 thread_t td = gd->gd_curthread; 1141 1142 /* 1143 * Should never be called with spinlocks held but there is a path 1144 * via ACPI where it might happen. 1145 */ 1146 if (gd->gd_spinlocks) 1147 return; 1148 1149 /* 1150 * Safe to call splz if we are not too-heavily nested. 1151 */ 1152 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1153 splz(); 1154 1155 /* 1156 * Caller allows switching 1157 */ 1158 if (lwkt_resched_wanted()) { 1159 atomic_set_int(&td->td_mpflags, TDF_MP_DIDYIELD); 1160 lwkt_schedule_self(td); 1161 lwkt_switch(); 1162 } 1163 } 1164 1165 /* 1166 * The quick version processes pending interrupts and higher-priority 1167 * LWKT threads but will not round-robin same-priority LWKT threads. 1168 * 1169 * When called while attempting to return to userland the only same-pri 1170 * threads are the ones which have already tried to become the current 1171 * user process. 1172 */ 1173 void 1174 lwkt_yield_quick(void) 1175 { 1176 globaldata_t gd = mycpu; 1177 thread_t td = gd->gd_curthread; 1178 1179 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1180 splz(); 1181 if (lwkt_resched_wanted()) { 1182 crit_enter(); 1183 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) { 1184 clear_lwkt_resched(); 1185 } else { 1186 atomic_set_int(&td->td_mpflags, TDF_MP_DIDYIELD); 1187 lwkt_schedule_self(curthread); 1188 lwkt_switch(); 1189 } 1190 crit_exit(); 1191 } 1192 } 1193 1194 /* 1195 * This yield is designed for kernel threads with a user context. 1196 * 1197 * The kernel acting on behalf of the user is potentially cpu-bound, 1198 * this function will efficiently allow other threads to run and also 1199 * switch to other processes by releasing. 1200 * 1201 * The lwkt_user_yield() function is designed to have very low overhead 1202 * if no yield is determined to be needed. 1203 */ 1204 void 1205 lwkt_user_yield(void) 1206 { 1207 globaldata_t gd = mycpu; 1208 thread_t td = gd->gd_curthread; 1209 1210 /* 1211 * Should never be called with spinlocks held but there is a path 1212 * via ACPI where it might happen. 1213 */ 1214 if (gd->gd_spinlocks) 1215 return; 1216 1217 /* 1218 * Always run any pending interrupts in case we are in a critical 1219 * section. 1220 */ 1221 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1222 splz(); 1223 1224 /* 1225 * Switch (which forces a release) if another kernel thread needs 1226 * the cpu, if userland wants us to resched, or if our kernel 1227 * quantum has run out. 1228 */ 1229 if (lwkt_resched_wanted() || 1230 user_resched_wanted()) 1231 { 1232 lwkt_switch(); 1233 } 1234 1235 #if 0 1236 /* 1237 * Reacquire the current process if we are released. 1238 * 1239 * XXX not implemented atm. The kernel may be holding locks and such, 1240 * so we want the thread to continue to receive cpu. 1241 */ 1242 if (td->td_release == NULL && lp) { 1243 lp->lwp_proc->p_usched->acquire_curproc(lp); 1244 td->td_release = lwkt_passive_release; 1245 lwkt_setpri_self(TDPRI_USER_NORM); 1246 } 1247 #endif 1248 } 1249 1250 /* 1251 * Generic schedule. Possibly schedule threads belonging to other cpus and 1252 * deal with threads that might be blocked on a wait queue. 1253 * 1254 * We have a little helper inline function which does additional work after 1255 * the thread has been enqueued, including dealing with preemption and 1256 * setting need_lwkt_resched() (which prevents the kernel from returning 1257 * to userland until it has processed higher priority threads). 1258 * 1259 * It is possible for this routine to be called after a failed _enqueue 1260 * (due to the target thread migrating, sleeping, or otherwise blocked). 1261 * We have to check that the thread is actually on the run queue! 1262 */ 1263 static __inline 1264 void 1265 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount) 1266 { 1267 if (ntd->td_flags & TDF_RUNQ) { 1268 if (ntd->td_preemptable) { 1269 ntd->td_preemptable(ntd, ccount); /* YYY +token */ 1270 } 1271 } 1272 } 1273 1274 static __inline 1275 void 1276 _lwkt_schedule(thread_t td) 1277 { 1278 globaldata_t mygd = mycpu; 1279 1280 KASSERT(td != &td->td_gd->gd_idlethread, 1281 ("lwkt_schedule(): scheduling gd_idlethread is illegal!")); 1282 KKASSERT((td->td_flags & TDF_MIGRATING) == 0); 1283 crit_enter_gd(mygd); 1284 KKASSERT(td->td_lwp == NULL || 1285 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 1286 1287 if (td == mygd->gd_curthread) { 1288 _lwkt_enqueue(td); 1289 } else { 1290 /* 1291 * If we own the thread, there is no race (since we are in a 1292 * critical section). If we do not own the thread there might 1293 * be a race but the target cpu will deal with it. 1294 */ 1295 if (td->td_gd == mygd) { 1296 _lwkt_enqueue(td); 1297 _lwkt_schedule_post(mygd, td, 1); 1298 } else { 1299 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0); 1300 } 1301 } 1302 crit_exit_gd(mygd); 1303 } 1304 1305 void 1306 lwkt_schedule(thread_t td) 1307 { 1308 _lwkt_schedule(td); 1309 } 1310 1311 void 1312 lwkt_schedule_noresched(thread_t td) /* XXX not impl */ 1313 { 1314 _lwkt_schedule(td); 1315 } 1316 1317 /* 1318 * When scheduled remotely if frame != NULL the IPIQ is being 1319 * run via doreti or an interrupt then preemption can be allowed. 1320 * 1321 * To allow preemption we have to drop the critical section so only 1322 * one is present in _lwkt_schedule_post. 1323 */ 1324 static void 1325 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame) 1326 { 1327 thread_t td = curthread; 1328 thread_t ntd = arg; 1329 1330 if (frame && ntd->td_preemptable) { 1331 crit_exit_noyield(td); 1332 _lwkt_schedule(ntd); 1333 crit_enter_quick(td); 1334 } else { 1335 _lwkt_schedule(ntd); 1336 } 1337 } 1338 1339 /* 1340 * Thread migration using a 'Pull' method. The thread may or may not be 1341 * the current thread. It MUST be descheduled and in a stable state. 1342 * lwkt_giveaway() must be called on the cpu owning the thread. 1343 * 1344 * At any point after lwkt_giveaway() is called, the target cpu may 1345 * 'pull' the thread by calling lwkt_acquire(). 1346 * 1347 * We have to make sure the thread is not sitting on a per-cpu tsleep 1348 * queue or it will blow up when it moves to another cpu. 1349 * 1350 * MPSAFE - must be called under very specific conditions. 1351 */ 1352 void 1353 lwkt_giveaway(thread_t td) 1354 { 1355 globaldata_t gd = mycpu; 1356 1357 crit_enter_gd(gd); 1358 if (td->td_flags & TDF_TSLEEPQ) 1359 tsleep_remove(td); 1360 KKASSERT(td->td_gd == gd); 1361 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); 1362 td->td_flags |= TDF_MIGRATING; 1363 crit_exit_gd(gd); 1364 } 1365 1366 void 1367 lwkt_acquire(thread_t td) 1368 { 1369 globaldata_t gd; 1370 globaldata_t mygd; 1371 1372 KKASSERT(td->td_flags & TDF_MIGRATING); 1373 gd = td->td_gd; 1374 mygd = mycpu; 1375 if (gd != mycpu) { 1376 #ifdef LOOPMASK 1377 uint64_t tsc_base = rdtsc(); 1378 #endif 1379 cpu_lfence(); 1380 KKASSERT((td->td_flags & TDF_RUNQ) == 0); 1381 crit_enter_gd(mygd); 1382 DEBUG_PUSH_INFO("lwkt_acquire"); 1383 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) { 1384 lwkt_process_ipiq(); 1385 cpu_lfence(); 1386 #ifdef _KERNEL_VIRTUAL 1387 pthread_yield(); 1388 #endif 1389 #ifdef LOOPMASK 1390 if (tsc_frequency && rdtsc() - tsc_base > tsc_frequency) { 1391 kprintf("lwkt_acquire: stuck td %p td->td_flags %08x\n", 1392 td, td->td_flags); 1393 tsc_base = rdtsc(); 1394 } 1395 #endif 1396 } 1397 DEBUG_POP_INFO(); 1398 cpu_mfence(); 1399 td->td_gd = mygd; 1400 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1401 td->td_flags &= ~TDF_MIGRATING; 1402 crit_exit_gd(mygd); 1403 } else { 1404 crit_enter_gd(mygd); 1405 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1406 td->td_flags &= ~TDF_MIGRATING; 1407 crit_exit_gd(mygd); 1408 } 1409 } 1410 1411 /* 1412 * Generic deschedule. Descheduling threads other then your own should be 1413 * done only in carefully controlled circumstances. Descheduling is 1414 * asynchronous. 1415 * 1416 * This function may block if the cpu has run out of messages. 1417 */ 1418 void 1419 lwkt_deschedule(thread_t td) 1420 { 1421 crit_enter(); 1422 if (td == curthread) { 1423 _lwkt_dequeue(td); 1424 } else { 1425 if (td->td_gd == mycpu) { 1426 _lwkt_dequeue(td); 1427 } else { 1428 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td); 1429 } 1430 } 1431 crit_exit(); 1432 } 1433 1434 /* 1435 * Set the target thread's priority. This routine does not automatically 1436 * switch to a higher priority thread, LWKT threads are not designed for 1437 * continuous priority changes. Yield if you want to switch. 1438 */ 1439 void 1440 lwkt_setpri(thread_t td, int pri) 1441 { 1442 if (td->td_pri != pri) { 1443 KKASSERT(pri >= 0); 1444 crit_enter(); 1445 if (td->td_flags & TDF_RUNQ) { 1446 KKASSERT(td->td_gd == mycpu); 1447 _lwkt_dequeue(td); 1448 td->td_pri = pri; 1449 _lwkt_enqueue(td); 1450 } else { 1451 td->td_pri = pri; 1452 } 1453 crit_exit(); 1454 } 1455 } 1456 1457 /* 1458 * Set the initial priority for a thread prior to it being scheduled for 1459 * the first time. The thread MUST NOT be scheduled before or during 1460 * this call. The thread may be assigned to a cpu other then the current 1461 * cpu. 1462 * 1463 * Typically used after a thread has been created with TDF_STOPPREQ, 1464 * and before the thread is initially scheduled. 1465 */ 1466 void 1467 lwkt_setpri_initial(thread_t td, int pri) 1468 { 1469 KKASSERT(pri >= 0); 1470 KKASSERT((td->td_flags & TDF_RUNQ) == 0); 1471 td->td_pri = pri; 1472 } 1473 1474 void 1475 lwkt_setpri_self(int pri) 1476 { 1477 thread_t td = curthread; 1478 1479 KKASSERT(pri >= 0 && pri <= TDPRI_MAX); 1480 crit_enter(); 1481 if (td->td_flags & TDF_RUNQ) { 1482 _lwkt_dequeue(td); 1483 td->td_pri = pri; 1484 _lwkt_enqueue(td); 1485 } else { 1486 td->td_pri = pri; 1487 } 1488 crit_exit(); 1489 } 1490 1491 /* 1492 * hz tick scheduler clock for LWKT threads 1493 */ 1494 void 1495 lwkt_schedulerclock(thread_t td) 1496 { 1497 globaldata_t gd = td->td_gd; 1498 thread_t xtd; 1499 1500 xtd = TAILQ_FIRST(&gd->gd_tdrunq); 1501 if (xtd == td) { 1502 /* 1503 * If the current thread is at the head of the runq shift it to the 1504 * end of any equal-priority threads and request a LWKT reschedule 1505 * if it moved. 1506 * 1507 * Ignore upri in this situation. There will only be one user thread 1508 * in user mode, all others will be user threads running in kernel 1509 * mode and we have to make sure they get some cpu. 1510 */ 1511 xtd = TAILQ_NEXT(td, td_threadq); 1512 if (xtd && xtd->td_pri == td->td_pri) { 1513 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq); 1514 while (xtd && xtd->td_pri == td->td_pri) 1515 xtd = TAILQ_NEXT(xtd, td_threadq); 1516 if (xtd) 1517 TAILQ_INSERT_BEFORE(xtd, td, td_threadq); 1518 else 1519 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 1520 need_lwkt_resched(); 1521 } 1522 } else if (xtd) { 1523 /* 1524 * If we scheduled a thread other than the one at the head of the 1525 * queue always request a reschedule every tick. 1526 */ 1527 need_lwkt_resched(); 1528 } 1529 /* else curthread probably the idle thread, no need to reschedule */ 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 biosched_done(td); 1683 dsched_exit_thread(td); 1684 1685 /* 1686 * Get us into a critical section to interlock gd_freetd and loop 1687 * until we can get it freed. 1688 * 1689 * We have to cache the current td in gd_freetd because objcache_put()ing 1690 * it would rip it out from under us while our thread is still active. 1691 * 1692 * We are the current thread so of course our own TDF_RUNNING bit will 1693 * be set, so unlike the lwp reap code we don't wait for it to clear. 1694 */ 1695 gd = mycpu; 1696 crit_enter_quick(td); 1697 for (;;) { 1698 if (td->td_refs) { 1699 tsleep(td, 0, "tdreap", 1); 1700 continue; 1701 } 1702 if ((std = gd->gd_freetd) != NULL) { 1703 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0); 1704 gd->gd_freetd = NULL; 1705 objcache_put(thread_cache, std); 1706 continue; 1707 } 1708 break; 1709 } 1710 1711 /* 1712 * Remove thread resources from kernel lists and deschedule us for 1713 * the last time. We cannot block after this point or we may end 1714 * up with a stale td on the tsleepq. 1715 * 1716 * None of this may block, the critical section is the only thing 1717 * protecting tdallq and the only thing preventing new lwkt_hold() 1718 * thread refs now. 1719 */ 1720 if (td->td_flags & TDF_TSLEEPQ) 1721 tsleep_remove(td); 1722 lwkt_deschedule_self(td); 1723 lwkt_remove_tdallq(td); 1724 KKASSERT(td->td_refs == 0); 1725 1726 /* 1727 * Final cleanup 1728 */ 1729 KKASSERT(gd->gd_freetd == NULL); 1730 if (td->td_flags & TDF_ALLOCATED_THREAD) 1731 gd->gd_freetd = td; 1732 cpu_thread_exit(); 1733 } 1734 1735 void 1736 lwkt_remove_tdallq(thread_t td) 1737 { 1738 KKASSERT(td->td_gd == mycpu); 1739 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1740 } 1741 1742 /* 1743 * Code reduction and branch prediction improvements. Call/return 1744 * overhead on modern cpus often degenerates into 0 cycles due to 1745 * the cpu's branch prediction hardware and return pc cache. We 1746 * can take advantage of this by not inlining medium-complexity 1747 * functions and we can also reduce the branch prediction impact 1748 * by collapsing perfectly predictable branches into a single 1749 * procedure instead of duplicating it. 1750 * 1751 * Is any of this noticeable? Probably not, so I'll take the 1752 * smaller code size. 1753 */ 1754 void 1755 crit_exit_wrapper(__DEBUG_CRIT_ARG__) 1756 { 1757 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__); 1758 } 1759 1760 void 1761 crit_panic(void) 1762 { 1763 thread_t td = curthread; 1764 int lcrit = td->td_critcount; 1765 1766 td->td_critcount = 0; 1767 cpu_ccfence(); 1768 panic("td_critcount is/would-go negative! %p %d", td, lcrit); 1769 /* NOT REACHED */ 1770 } 1771 1772 /* 1773 * Called from debugger/panic on cpus which have been stopped. We must still 1774 * process the IPIQ while stopped. 1775 * 1776 * If we are dumping also try to process any pending interrupts. This may 1777 * or may not work depending on the state of the cpu at the point it was 1778 * stopped. 1779 */ 1780 void 1781 lwkt_smp_stopped(void) 1782 { 1783 globaldata_t gd = mycpu; 1784 1785 if (dumping) { 1786 lwkt_process_ipiq(); 1787 --gd->gd_intr_nesting_level; 1788 splz(); 1789 ++gd->gd_intr_nesting_level; 1790 } else { 1791 lwkt_process_ipiq(); 1792 } 1793 cpu_smp_stopped(); 1794 } 1795