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