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/malloc.h> 49 #include <sys/queue.h> 50 #include <sys/sysctl.h> 51 #include <sys/kthread.h> 52 #include <machine/cpu.h> 53 #include <sys/lock.h> 54 #include <sys/spinlock.h> 55 #include <sys/ktr.h> 56 #include <sys/indefinite.h> 57 58 #include <sys/thread2.h> 59 #include <sys/spinlock2.h> 60 #include <sys/indefinite2.h> 61 62 #include <sys/dsched.h> 63 64 #include <vm/vm.h> 65 #include <vm/vm_param.h> 66 #include <vm/vm_kern.h> 67 #include <vm/vm_object.h> 68 #include <vm/vm_page.h> 69 #include <vm/vm_map.h> 70 #include <vm/vm_pager.h> 71 #include <vm/vm_extern.h> 72 73 #include <machine/stdarg.h> 74 #include <machine/smp.h> 75 #include <machine/clock.h> 76 77 #define LOOPMASK 78 79 #if !defined(KTR_CTXSW) 80 #define KTR_CTXSW KTR_ALL 81 #endif 82 KTR_INFO_MASTER(ctxsw); 83 KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p", int cpu, struct thread *td); 84 KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p", int cpu, struct thread *td); 85 KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s", struct thread *td, char *comm); 86 KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", struct thread *td); 87 88 static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads"); 89 MALLOC_DEFINE(M_FPUCTX, "fpuctx", "kernel FPU contexts"); 90 91 #ifdef INVARIANTS 92 static int panic_on_cscount = 0; 93 #endif 94 #ifdef DEBUG_LWKT_THREAD 95 static int64_t switch_count = 0; 96 static int64_t preempt_hit = 0; 97 static int64_t preempt_miss = 0; 98 static int64_t preempt_weird = 0; 99 #endif 100 static int lwkt_use_spin_port; 101 __read_mostly 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 #ifdef DEBUG_LWKT_THREAD 118 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, 119 "Number of switched threads"); 120 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, 121 "Successful preemption events"); 122 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, 123 "Failed preemption events"); 124 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, 125 "Number of preempted threads."); 126 #endif 127 extern int lwkt_sched_debug; 128 int lwkt_sched_debug = 0; 129 SYSCTL_INT(_lwkt, OID_AUTO, sched_debug, CTLFLAG_RW, 130 &lwkt_sched_debug, 0, "Scheduler debug"); 131 __read_mostly static u_int lwkt_spin_loops = 10; 132 SYSCTL_UINT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW, 133 &lwkt_spin_loops, 0, "Scheduler spin loops until sorted decon"); 134 __read_mostly static int preempt_enable = 1; 135 SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW, 136 &preempt_enable, 0, "Enable preemption"); 137 static int lwkt_cache_threads = 0; 138 SYSCTL_INT(_lwkt, OID_AUTO, cache_threads, CTLFLAG_RD, 139 &lwkt_cache_threads, 0, "thread+kstack cache"); 140 141 /* 142 * These helper procedures handle the runq, they can only be called from 143 * within a critical section. 144 * 145 * WARNING! Prior to SMP being brought up it is possible to enqueue and 146 * dequeue threads belonging to other cpus, so be sure to use td->td_gd 147 * instead of 'mycpu' when referencing the globaldata structure. Once 148 * SMP live enqueuing and dequeueing only occurs on the current cpu. 149 */ 150 static __inline 151 void 152 _lwkt_dequeue(thread_t td) 153 { 154 if (td->td_flags & TDF_RUNQ) { 155 struct globaldata *gd = td->td_gd; 156 157 td->td_flags &= ~TDF_RUNQ; 158 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq); 159 --gd->gd_tdrunqcount; 160 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL) 161 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING); 162 } 163 } 164 165 /* 166 * Priority enqueue. 167 * 168 * There are a limited number of lwkt threads runnable since user 169 * processes only schedule one at a time per cpu. However, there can 170 * be many user processes in kernel mode exiting from a tsleep() which 171 * become runnable. 172 * 173 * We scan the queue in both directions to help deal with degenerate 174 * situations when hundreds or thousands (or more) threads are runnable. 175 * 176 * NOTE: lwkt_schedulerclock() will force a round-robin based on td_pri and 177 * will ignore user priority. This is to ensure that user threads in 178 * kernel mode get cpu at some point regardless of what the user 179 * scheduler thinks. 180 */ 181 static __inline 182 void 183 _lwkt_enqueue(thread_t td) 184 { 185 thread_t xtd; /* forward scan */ 186 thread_t rtd; /* reverse scan */ 187 188 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) { 189 struct globaldata *gd = td->td_gd; 190 191 td->td_flags |= TDF_RUNQ; 192 xtd = TAILQ_FIRST(&gd->gd_tdrunq); 193 if (xtd == NULL) { 194 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 195 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING); 196 } else { 197 /* 198 * NOTE: td_upri - higher numbers more desireable, same sense 199 * as td_pri (typically reversed from lwp_upri). 200 * 201 * In the equal priority case we want the best selection 202 * at the beginning so the less desireable selections know 203 * that they have to setrunqueue/go-to-another-cpu, even 204 * though it means switching back to the 'best' selection. 205 * This also avoids degenerate situations when many threads 206 * are runnable or waking up at the same time. 207 * 208 * If upri matches exactly place at end/round-robin. 209 */ 210 rtd = TAILQ_LAST(&gd->gd_tdrunq, lwkt_queue); 211 212 while (xtd && 213 (xtd->td_pri > td->td_pri || 214 (xtd->td_pri == td->td_pri && 215 xtd->td_upri >= td->td_upri))) { 216 xtd = TAILQ_NEXT(xtd, td_threadq); 217 218 /* 219 * Doing a reverse scan at the same time is an optimization 220 * for the insert-closer-to-tail case that avoids having to 221 * scan the entire list. This situation can occur when 222 * thousands of threads are woken up at the same time. 223 */ 224 if (rtd->td_pri > td->td_pri || 225 (rtd->td_pri == td->td_pri && 226 rtd->td_upri >= td->td_upri)) { 227 TAILQ_INSERT_AFTER(&gd->gd_tdrunq, rtd, td, td_threadq); 228 goto skip; 229 } 230 rtd = TAILQ_PREV(rtd, lwkt_queue, td_threadq); 231 } 232 if (xtd) 233 TAILQ_INSERT_BEFORE(xtd, td, td_threadq); 234 else 235 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 236 } 237 skip: 238 ++gd->gd_tdrunqcount; 239 240 /* 241 * Request a LWKT reschedule if we are now at the head of the queue. 242 */ 243 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) 244 need_lwkt_resched(); 245 } 246 } 247 248 static boolean_t 249 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags) 250 { 251 struct thread *td = (struct thread *)obj; 252 253 td->td_kstack = NULL; 254 td->td_kstack_size = 0; 255 td->td_flags = TDF_ALLOCATED_THREAD; 256 td->td_mpflags = 0; 257 return (1); 258 } 259 260 static void 261 _lwkt_thread_dtor(void *obj, void *privdata) 262 { 263 struct thread *td = (struct thread *)obj; 264 265 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD, 266 ("_lwkt_thread_dtor: not allocated from objcache")); 267 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack && 268 td->td_kstack_size > 0, 269 ("_lwkt_thread_dtor: corrupted stack")); 270 kmem_free(kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size); 271 td->td_kstack = NULL; 272 td->td_flags = 0; 273 } 274 275 /* 276 * Initialize the lwkt s/system. 277 * 278 * Nominally cache up to 32 thread + kstack structures. Cache more on 279 * systems with a lot of cpu cores. 280 */ 281 static void 282 lwkt_init(void) 283 { 284 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads); 285 if (lwkt_cache_threads == 0) { 286 lwkt_cache_threads = ncpus * 4; 287 if (lwkt_cache_threads < 32) 288 lwkt_cache_threads = 32; 289 } 290 thread_cache = objcache_create_mbacked( 291 M_THREAD, sizeof(struct thread), 292 0, lwkt_cache_threads, 293 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL); 294 } 295 SYSINIT(lwkt_init, SI_BOOT2_LWKT_INIT, SI_ORDER_FIRST, lwkt_init, NULL); 296 297 /* 298 * Schedule a thread to run. As the current thread we can always safely 299 * schedule ourselves, and a shortcut procedure is provided for that 300 * function. 301 * 302 * (non-blocking, self contained on a per cpu basis) 303 */ 304 void 305 lwkt_schedule_self(thread_t td) 306 { 307 KKASSERT((td->td_flags & TDF_MIGRATING) == 0); 308 crit_enter_quick(td); 309 KASSERT(td != &td->td_gd->gd_idlethread, 310 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!")); 311 KKASSERT(td->td_lwp == NULL || 312 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 313 _lwkt_enqueue(td); 314 crit_exit_quick(td); 315 } 316 317 /* 318 * Deschedule a thread. 319 * 320 * (non-blocking, self contained on a per cpu basis) 321 */ 322 void 323 lwkt_deschedule_self(thread_t td) 324 { 325 crit_enter_quick(td); 326 _lwkt_dequeue(td); 327 crit_exit_quick(td); 328 } 329 330 /* 331 * LWKTs operate on a per-cpu basis 332 * 333 * WARNING! Called from early boot, 'mycpu' may not work yet. 334 */ 335 void 336 lwkt_gdinit(struct globaldata *gd) 337 { 338 TAILQ_INIT(&gd->gd_tdrunq); 339 TAILQ_INIT(&gd->gd_tdallq); 340 lockinit(&gd->gd_sysctllock, "sysctl", 0, LK_CANRECURSE); 341 } 342 343 /* 344 * Create a new thread. The thread must be associated with a process context 345 * or LWKT start address before it can be scheduled. If the target cpu is 346 * -1 the thread will be created on the current cpu. 347 * 348 * If you intend to create a thread without a process context this function 349 * does everything except load the startup and switcher function. 350 */ 351 thread_t 352 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags) 353 { 354 static int cpu_rotator; 355 globaldata_t gd = mycpu; 356 void *stack; 357 358 /* 359 * If static thread storage is not supplied allocate a thread. Reuse 360 * a cached free thread if possible. gd_freetd is used to keep an exiting 361 * thread intact through the exit. 362 */ 363 if (td == NULL) { 364 crit_enter_gd(gd); 365 if ((td = gd->gd_freetd) != NULL) { 366 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK| 367 TDF_RUNQ)) == 0); 368 gd->gd_freetd = NULL; 369 } else { 370 td = objcache_get(thread_cache, M_WAITOK); 371 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK| 372 TDF_RUNQ)) == 0); 373 } 374 crit_exit_gd(gd); 375 KASSERT((td->td_flags & 376 (TDF_ALLOCATED_THREAD|TDF_RUNNING|TDF_PREEMPT_LOCK)) == 377 TDF_ALLOCATED_THREAD, 378 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags)); 379 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK); 380 } 381 382 /* 383 * Try to reuse cached stack. 384 */ 385 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) { 386 if (flags & TDF_ALLOCATED_STACK) { 387 kmem_free(kernel_map, (vm_offset_t)stack, td->td_kstack_size); 388 stack = NULL; 389 } 390 } 391 if (stack == NULL) { 392 if (cpu < 0) { 393 stack = (void *)kmem_alloc_stack(kernel_map, stksize, 0); 394 } else { 395 stack = (void *)kmem_alloc_stack(kernel_map, stksize, 396 KM_CPU(cpu)); 397 } 398 flags |= TDF_ALLOCATED_STACK; 399 } 400 if (cpu < 0) { 401 cpu = ++cpu_rotator; 402 cpu_ccfence(); 403 cpu = (uint32_t)cpu % (uint32_t)ncpus; 404 } 405 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu)); 406 return(td); 407 } 408 409 /* 410 * Initialize a preexisting thread structure. This function is used by 411 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread. 412 * 413 * All threads start out in a critical section at a priority of 414 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as 415 * appropriate. This function may send an IPI message when the 416 * requested cpu is not the current cpu and consequently gd_tdallq may 417 * not be initialized synchronously from the point of view of the originating 418 * cpu. 419 * 420 * NOTE! we have to be careful in regards to creating threads for other cpus 421 * if SMP has not yet been activated. 422 */ 423 static void 424 lwkt_init_thread_remote(void *arg) 425 { 426 thread_t td = arg; 427 428 /* 429 * Protected by critical section held by IPI dispatch 430 */ 431 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq); 432 } 433 434 /* 435 * lwkt core thread structural initialization. 436 * 437 * NOTE: All threads are initialized as mpsafe threads. 438 */ 439 void 440 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags, 441 struct globaldata *gd) 442 { 443 globaldata_t mygd = mycpu; 444 445 bzero(td, sizeof(struct thread)); 446 td->td_kstack = stack; 447 td->td_kstack_size = stksize; 448 td->td_flags = flags; 449 td->td_mpflags = 0; 450 td->td_type = TD_TYPE_GENERIC; 451 td->td_gd = gd; 452 td->td_pri = TDPRI_KERN_DAEMON; 453 td->td_critcount = 1; 454 td->td_toks_have = NULL; 455 td->td_toks_stop = &td->td_toks_base; 456 if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT)) { 457 lwkt_initport_spin(&td->td_msgport, td, 458 (flags & TDF_FIXEDCPU) ? TRUE : FALSE); 459 } else { 460 lwkt_initport_thread(&td->td_msgport, td); 461 } 462 pmap_init_thread(td); 463 /* 464 * Normally initializing a thread for a remote cpu requires sending an 465 * IPI. However, the idlethread is setup before the other cpus are 466 * activated so we have to treat it as a special case. XXX manipulation 467 * of gd_tdallq requires the BGL. 468 */ 469 if (gd == mygd || td == &gd->gd_idlethread) { 470 crit_enter_gd(mygd); 471 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); 472 crit_exit_gd(mygd); 473 } else { 474 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td); 475 } 476 dsched_enter_thread(td); 477 } 478 479 void 480 lwkt_set_comm(thread_t td, const char *ctl, ...) 481 { 482 __va_list va; 483 484 __va_start(va, ctl); 485 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va); 486 __va_end(va); 487 KTR_LOG(ctxsw_newtd, td, td->td_comm); 488 } 489 490 /* 491 * Prevent the thread from getting destroyed. Note that unlike PHOLD/PRELE 492 * this does not prevent the thread from migrating to another cpu so the 493 * gd_tdallq state is not protected by this. 494 */ 495 void 496 lwkt_hold(thread_t td) 497 { 498 atomic_add_int(&td->td_refs, 1); 499 } 500 501 void 502 lwkt_rele(thread_t td) 503 { 504 KKASSERT(td->td_refs > 0); 505 atomic_add_int(&td->td_refs, -1); 506 } 507 508 void 509 lwkt_free_thread(thread_t td) 510 { 511 KKASSERT(td->td_refs == 0); 512 KKASSERT((td->td_flags & (TDF_RUNNING | TDF_PREEMPT_LOCK | 513 TDF_RUNQ | TDF_TSLEEPQ | TDF_KERNELFP)) == 0); 514 515 if (td->td_kfpuctx) { 516 kfree(td->td_kfpuctx, M_FPUCTX); 517 td->td_kfpuctx = NULL; 518 } 519 520 if (td->td_flags & TDF_ALLOCATED_THREAD) { 521 objcache_put(thread_cache, td); 522 } else if (td->td_flags & TDF_ALLOCATED_STACK) { 523 /* client-allocated struct with internally allocated stack */ 524 KASSERT(td->td_kstack && td->td_kstack_size > 0, 525 ("lwkt_free_thread: corrupted stack")); 526 kmem_free(kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size); 527 td->td_kstack = NULL; 528 td->td_kstack_size = 0; 529 } 530 531 KTR_LOG(ctxsw_deadtd, td); 532 } 533 534 535 /* 536 * Switch to the next runnable lwkt. If no LWKTs are runnable then 537 * switch to the idlethread. Switching must occur within a critical 538 * section to avoid races with the scheduling queue. 539 * 540 * We always have full control over our cpu's run queue. Other cpus 541 * that wish to manipulate our queue must use the cpu_*msg() calls to 542 * talk to our cpu, so a critical section is all that is needed and 543 * the result is very, very fast thread switching. 544 * 545 * The LWKT scheduler uses a fixed priority model and round-robins at 546 * each priority level. User process scheduling is a totally 547 * different beast and LWKT priorities should not be confused with 548 * user process priorities. 549 * 550 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch() 551 * is not called by the current thread in the preemption case, only when 552 * the preempting thread blocks (in order to return to the original thread). 553 * 554 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread 555 * migration and tsleep deschedule the current lwkt thread and call 556 * lwkt_switch(). In particular, the target cpu of the migration fully 557 * expects the thread to become non-runnable and can deadlock against 558 * cpusync operations if we run any IPIs prior to switching the thread out. 559 * 560 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF 561 * THE CURRENT THREAD HAS BEEN DESCHEDULED! 562 */ 563 void 564 lwkt_switch(void) 565 { 566 globaldata_t gd = mycpu; 567 thread_t td = gd->gd_curthread; 568 thread_t ntd; 569 thread_t xtd; 570 int upri; 571 #ifdef LOOPMASK 572 uint64_t tsc_base = rdtsc(); 573 #endif 574 575 KKASSERT(gd->gd_processing_ipiq == 0); 576 KKASSERT(td->td_flags & TDF_RUNNING); 577 578 /* 579 * Switching from within a 'fast' (non thread switched) interrupt or IPI 580 * is illegal. However, we may have to do it anyway if we hit a fatal 581 * kernel trap or we have paniced. 582 * 583 * If this case occurs save and restore the interrupt nesting level. 584 */ 585 if (gd->gd_intr_nesting_level) { 586 int savegdnest; 587 int savegdtrap; 588 589 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) { 590 panic("lwkt_switch: Attempt to switch from a " 591 "fast interrupt, ipi, or hard code section, " 592 "td %p\n", 593 td); 594 } else { 595 savegdnest = gd->gd_intr_nesting_level; 596 savegdtrap = gd->gd_trap_nesting_level; 597 gd->gd_intr_nesting_level = 0; 598 gd->gd_trap_nesting_level = 0; 599 if ((td->td_flags & TDF_PANICWARN) == 0) { 600 td->td_flags |= TDF_PANICWARN; 601 kprintf("Warning: thread switch from interrupt, IPI, " 602 "or hard code section.\n" 603 "thread %p (%s)\n", td, td->td_comm); 604 print_backtrace(-1); 605 } 606 lwkt_switch(); 607 gd->gd_intr_nesting_level = savegdnest; 608 gd->gd_trap_nesting_level = savegdtrap; 609 return; 610 } 611 } 612 613 /* 614 * Release our current user process designation if we are blocking 615 * or if a user reschedule was requested. 616 * 617 * NOTE: This function is NOT called if we are switching into or 618 * returning from a preemption. 619 * 620 * NOTE: Releasing our current user process designation may cause 621 * it to be assigned to another thread, which in turn will 622 * cause us to block in the usched acquire code when we attempt 623 * to return to userland. 624 * 625 * NOTE: On SMP systems this can be very nasty when heavy token 626 * contention is present so we want to be careful not to 627 * release the designation gratuitously. 628 */ 629 if (td->td_release && 630 (user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) { 631 td->td_release(td); 632 } 633 634 /* 635 * Release all tokens. Once we do this we must remain in the critical 636 * section and cannot run IPIs or other interrupts until we switch away 637 * because they may implode if they try to get a token using our thread 638 * context. 639 */ 640 crit_enter_gd(gd); 641 if (TD_TOKS_HELD(td)) 642 lwkt_relalltokens(td); 643 644 /* 645 * We had better not be holding any spin locks, but don't get into an 646 * endless panic loop. 647 */ 648 KASSERT(gd->gd_spinlocks == 0 || panicstr != NULL, 649 ("lwkt_switch: still holding %d exclusive spinlocks!", 650 gd->gd_spinlocks)); 651 652 #ifdef INVARIANTS 653 if (td->td_cscount) { 654 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n", 655 td); 656 if (panic_on_cscount) 657 panic("switching while mastering cpusync"); 658 } 659 #endif 660 661 /* 662 * If we had preempted another thread on this cpu, resume the preempted 663 * thread. This occurs transparently, whether the preempted thread 664 * was scheduled or not (it may have been preempted after descheduling 665 * itself). 666 * 667 * We have to setup the MP lock for the original thread after backing 668 * out the adjustment that was made to curthread when the original 669 * was preempted. 670 */ 671 if ((ntd = td->td_preempted) != NULL) { 672 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK); 673 ntd->td_flags |= TDF_PREEMPT_DONE; 674 ntd->td_contended = 0; /* reset contended */ 675 676 /* 677 * The interrupt may have woken a thread up, we need to properly 678 * set the reschedule flag if the originally interrupted thread is 679 * at a lower priority. 680 * 681 * NOTE: The interrupt may not have descheduled ntd. 682 * 683 * NOTE: We do not reschedule if there are no threads on the runq. 684 * (ntd could be the idlethread). 685 */ 686 xtd = TAILQ_FIRST(&gd->gd_tdrunq); 687 if (xtd && xtd != ntd) 688 need_lwkt_resched(); 689 goto havethread_preempted; 690 } 691 692 /* 693 * Figure out switch target. If we cannot switch to our desired target 694 * look for a thread that we can switch to. 695 * 696 * NOTE! The limited spin loop and related parameters are extremely 697 * important for system performance, particularly for pipes and 698 * concurrent conflicting VM faults. 699 */ 700 clear_lwkt_resched(); 701 ntd = TAILQ_FIRST(&gd->gd_tdrunq); 702 703 if (ntd) { 704 do { 705 if (TD_TOKS_NOT_HELD(ntd) || 706 lwkt_getalltokens(ntd, (ntd->td_contended > lwkt_spin_loops))) 707 { 708 goto havethread; 709 } 710 ++ntd->td_contended; /* overflow ok */ 711 if (gd->gd_indefinite.type == 0) 712 indefinite_init(&gd->gd_indefinite, NULL, 0, 't'); 713 #ifdef LOOPMASK 714 if (tsc_frequency && rdtsc() - tsc_base > tsc_frequency) { 715 kprintf("lwkt_switch: WARNING, excessive token contention " 716 "cpu %d, %d sec, " 717 "td %p (%s)\n", 718 gd->gd_cpuid, 719 ntd->td_contended, 720 ntd, 721 ntd->td_comm); 722 tsc_base = rdtsc(); 723 } 724 #endif 725 } while (ntd->td_contended < (lwkt_spin_loops >> 1)); 726 upri = ntd->td_upri; 727 728 /* 729 * Bleh, the thread we wanted to switch to has a contended token. 730 * See if we can switch to another thread. 731 * 732 * We generally don't want to do this because it represents a 733 * priority inversion, but contending tokens on the same cpu can 734 * cause real problems if we don't now that we have an exclusive 735 * priority mechanism over shared for tokens. 736 * 737 * The solution is to allow threads with pending tokens to compete 738 * for them (a lower priority thread will get less cpu once it 739 * returns from the kernel anyway). If a thread does not have 740 * any contending tokens, we go by td_pri and upri. 741 */ 742 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) { 743 if (TD_TOKS_NOT_HELD(ntd) && 744 ntd->td_pri < TDPRI_KERN_LPSCHED && upri > ntd->td_upri) { 745 continue; 746 } 747 if (upri < ntd->td_upri) 748 upri = ntd->td_upri; 749 750 /* 751 * Try this one. 752 */ 753 if (TD_TOKS_NOT_HELD(ntd) || 754 lwkt_getalltokens(ntd, (ntd->td_contended > lwkt_spin_loops))) { 755 goto havethread; 756 } 757 ++ntd->td_contended; /* overflow ok */ 758 } 759 760 /* 761 * Fall through, switch to idle thread to get us out of the current 762 * context. Since we were contended, prevent HLT by flagging a 763 * LWKT reschedule. 764 */ 765 need_lwkt_resched(); 766 } 767 768 /* 769 * We either contended on ntd or the runq is empty. We must switch 770 * through the idle thread to get out of the current context. 771 */ 772 ntd = &gd->gd_idlethread; 773 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) 774 ASSERT_NO_TOKENS_HELD(ntd); 775 cpu_time.cp_msg[0] = 0; 776 goto haveidle; 777 778 havethread: 779 /* 780 * Clear gd_idle_repeat when doing a normal switch to a non-idle 781 * thread. 782 */ 783 ntd->td_wmesg = NULL; 784 ntd->td_contended = 0; /* reset once scheduled */ 785 ++gd->gd_cnt.v_swtch; 786 gd->gd_idle_repeat = 0; 787 788 /* 789 * If we were busy waiting record final disposition 790 */ 791 if (gd->gd_indefinite.type) 792 indefinite_done(&gd->gd_indefinite); 793 794 havethread_preempted: 795 /* 796 * If the new target does not need the MP lock and we are holding it, 797 * release the MP lock. If the new target requires the MP lock we have 798 * already acquired it for the target. 799 */ 800 ; 801 haveidle: 802 KASSERT(ntd->td_critcount, 803 ("priority problem in lwkt_switch %d %d", 804 td->td_critcount, ntd->td_critcount)); 805 806 if (td != ntd) { 807 /* 808 * Execute the actual thread switch operation. This function 809 * returns to the current thread and returns the previous thread 810 * (which may be different from the thread we switched to). 811 * 812 * We are responsible for marking ntd as TDF_RUNNING. 813 */ 814 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0); 815 #ifdef DEBUG_LWKT_THREAD 816 ++switch_count; 817 #endif 818 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd); 819 ntd->td_flags |= TDF_RUNNING; 820 lwkt_switch_return(td->td_switch(ntd)); 821 /* ntd invalid, td_switch() can return a different thread_t */ 822 } 823 824 /* 825 * catch-all. XXX is this strictly needed? 826 */ 827 splz_check(); 828 829 /* NOTE: current cpu may have changed after switch */ 830 crit_exit_quick(td); 831 } 832 833 /* 834 * Called by assembly in the td_switch (thread restore path) for thread 835 * bootstrap cases which do not 'return' to lwkt_switch(). 836 */ 837 void 838 lwkt_switch_return(thread_t otd) 839 { 840 globaldata_t rgd; 841 #ifdef LOOPMASK 842 uint64_t tsc_base = rdtsc(); 843 #endif 844 int exiting; 845 846 exiting = otd->td_flags & TDF_EXITING; 847 cpu_ccfence(); 848 849 /* 850 * Check if otd was migrating. Now that we are on ntd we can finish 851 * up the migration. This is a bit messy but it is the only place 852 * where td is known to be fully descheduled. 853 * 854 * We can only activate the migration if otd was migrating but not 855 * held on the cpu due to a preemption chain. We still have to 856 * clear TDF_RUNNING on the old thread either way. 857 * 858 * We are responsible for clearing the previously running thread's 859 * TDF_RUNNING. 860 */ 861 if ((rgd = otd->td_migrate_gd) != NULL && 862 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) { 863 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) == 864 (TDF_MIGRATING | TDF_RUNNING)); 865 otd->td_migrate_gd = NULL; 866 otd->td_flags &= ~TDF_RUNNING; 867 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd); 868 } else { 869 otd->td_flags &= ~TDF_RUNNING; 870 } 871 872 /* 873 * Final exit validations (see lwp_wait()). Note that otd becomes 874 * invalid the *instant* we set TDF_MP_EXITSIG. 875 * 876 * Use the EXITING status loaded from before we clear TDF_RUNNING, 877 * because if it is not set otd becomes invalid the instant we clear 878 * TDF_RUNNING on it (otherwise, if the system is fast enough, we 879 * might 'steal' TDF_EXITING from another switch-return!). 880 */ 881 while (exiting) { 882 u_int mpflags; 883 884 mpflags = otd->td_mpflags; 885 cpu_ccfence(); 886 887 if (mpflags & TDF_MP_EXITWAIT) { 888 if (atomic_cmpset_int(&otd->td_mpflags, mpflags, 889 mpflags | TDF_MP_EXITSIG)) { 890 wakeup(otd); 891 break; 892 } 893 } else { 894 if (atomic_cmpset_int(&otd->td_mpflags, mpflags, 895 mpflags | TDF_MP_EXITSIG)) { 896 wakeup(otd); 897 break; 898 } 899 } 900 901 #ifdef LOOPMASK 902 if (tsc_frequency && rdtsc() - tsc_base > tsc_frequency) { 903 kprintf("lwkt_switch_return: excessive TDF_EXITING " 904 "thread %p\n", otd); 905 tsc_base = rdtsc(); 906 } 907 #endif 908 } 909 } 910 911 /* 912 * Request that the target thread preempt the current thread. Preemption 913 * can only occur only: 914 * 915 * - If our critical section is the one that we were called with 916 * - The relative priority of the target thread is higher 917 * - The target is not excessively interrupt-nested via td_nest_count 918 * - The target thread holds no tokens. 919 * - The target thread is not already scheduled and belongs to the 920 * current cpu. 921 * - The current thread is not holding any spin-locks. 922 * 923 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically 924 * this is called via lwkt_schedule() through the td_preemptable callback. 925 * critcount is the managed critical priority that we should ignore in order 926 * to determine whether preemption is possible (aka usually just the crit 927 * priority of lwkt_schedule() itself). 928 * 929 * Preemption is typically limited to interrupt threads. 930 * 931 * Operation works in a fairly straight-forward manner. The normal 932 * scheduling code is bypassed and we switch directly to the target 933 * thread. When the target thread attempts to block or switch away 934 * code at the base of lwkt_switch() will switch directly back to our 935 * thread. Our thread is able to retain whatever tokens it holds and 936 * if the target needs one of them the target will switch back to us 937 * and reschedule itself normally. 938 */ 939 void 940 lwkt_preempt(thread_t ntd, int critcount) 941 { 942 struct globaldata *gd = mycpu; 943 thread_t xtd; 944 thread_t td; 945 int save_gd_intr_nesting_level; 946 947 /* 948 * The caller has put us in a critical section. We can only preempt 949 * if the caller of the caller was not in a critical section (basically 950 * a local interrupt), as determined by the 'critcount' parameter. We 951 * also can't preempt if the caller is holding any spinlocks (even if 952 * he isn't in a critical section). This also handles the tokens test. 953 * 954 * YYY The target thread must be in a critical section (else it must 955 * inherit our critical section? I dunno yet). 956 */ 957 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri)); 958 959 td = gd->gd_curthread; 960 if (preempt_enable == 0) { 961 #ifdef DEBUG_LWKT_THREAD 962 ++preempt_miss; 963 #endif 964 return; 965 } 966 if (ntd->td_pri <= td->td_pri) { 967 #ifdef DEBUG_LWKT_THREAD 968 ++preempt_miss; 969 #endif 970 return; 971 } 972 if (td->td_critcount > critcount) { 973 #ifdef DEBUG_LWKT_THREAD 974 ++preempt_miss; 975 #endif 976 return; 977 } 978 if (td->td_nest_count >= 2) { 979 #ifdef DEBUG_LWKT_THREAD 980 ++preempt_miss; 981 #endif 982 return; 983 } 984 if (td->td_cscount) { 985 #ifdef DEBUG_LWKT_THREAD 986 ++preempt_miss; 987 #endif 988 return; 989 } 990 if (ntd->td_gd != gd) { 991 #ifdef DEBUG_LWKT_THREAD 992 ++preempt_miss; 993 #endif 994 return; 995 } 996 997 /* 998 * We don't have to check spinlocks here as they will also bump 999 * td_critcount. 1000 * 1001 * Do not try to preempt if the target thread is holding any tokens. 1002 * We could try to acquire the tokens but this case is so rare there 1003 * is no need to support it. 1004 */ 1005 KKASSERT(gd->gd_spinlocks == 0); 1006 1007 if (TD_TOKS_HELD(ntd)) { 1008 #ifdef DEBUG_LWKT_THREAD 1009 ++preempt_miss; 1010 #endif 1011 return; 1012 } 1013 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) { 1014 #ifdef DEBUG_LWKT_THREAD 1015 ++preempt_weird; 1016 #endif 1017 return; 1018 } 1019 if (ntd->td_preempted) { 1020 #ifdef DEBUG_LWKT_THREAD 1021 ++preempt_hit; 1022 #endif 1023 return; 1024 } 1025 KKASSERT(gd->gd_processing_ipiq == 0); 1026 1027 /* 1028 * Since we are able to preempt the current thread, there is no need to 1029 * call need_lwkt_resched(). 1030 * 1031 * We must temporarily clear gd_intr_nesting_level around the switch 1032 * since switchouts from the target thread are allowed (they will just 1033 * return to our thread), and since the target thread has its own stack. 1034 * 1035 * A preemption must switch back to the original thread, assert the 1036 * case. 1037 */ 1038 #ifdef DEBUG_LWKT_THREAD 1039 ++preempt_hit; 1040 #endif 1041 ntd->td_preempted = td; 1042 td->td_flags |= TDF_PREEMPT_LOCK; 1043 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd); 1044 save_gd_intr_nesting_level = gd->gd_intr_nesting_level; 1045 gd->gd_intr_nesting_level = 0; 1046 1047 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0); 1048 ntd->td_flags |= TDF_RUNNING; 1049 xtd = td->td_switch(ntd); 1050 KKASSERT(xtd == ntd); 1051 lwkt_switch_return(xtd); 1052 gd->gd_intr_nesting_level = save_gd_intr_nesting_level; 1053 1054 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE)); 1055 ntd->td_preempted = NULL; 1056 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE); 1057 } 1058 1059 /* 1060 * Conditionally call splz() if gd_reqflags indicates work is pending. 1061 * This will work inside a critical section but not inside a hard code 1062 * section. 1063 * 1064 * (self contained on a per cpu basis) 1065 */ 1066 void 1067 splz_check(void) 1068 { 1069 globaldata_t gd = mycpu; 1070 thread_t td = gd->gd_curthread; 1071 1072 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && 1073 gd->gd_intr_nesting_level == 0 && 1074 td->td_nest_count < 2) 1075 { 1076 splz(); 1077 } 1078 } 1079 1080 /* 1081 * This version is integrated into crit_exit, reqflags has already 1082 * been tested but td_critcount has not. 1083 * 1084 * We only want to execute the splz() on the 1->0 transition of 1085 * critcount and not in a hard code section or if too deeply nested. 1086 * 1087 * NOTE: gd->gd_spinlocks is implied to be 0 when td_critcount is 0. 1088 */ 1089 void 1090 lwkt_maybe_splz(thread_t td) 1091 { 1092 globaldata_t gd = td->td_gd; 1093 1094 if (td->td_critcount == 0 && 1095 gd->gd_intr_nesting_level == 0 && 1096 td->td_nest_count < 2) 1097 { 1098 splz(); 1099 } 1100 } 1101 1102 /* 1103 * Drivers which set up processing co-threads can call this function to 1104 * run the co-thread at a higher priority and to allow it to preempt 1105 * normal threads. 1106 */ 1107 void 1108 lwkt_set_interrupt_support_thread(void) 1109 { 1110 thread_t td = curthread; 1111 1112 lwkt_setpri_self(TDPRI_INT_SUPPORT); 1113 td->td_flags |= TDF_INTTHREAD; 1114 td->td_preemptable = lwkt_preempt; 1115 } 1116 1117 1118 /* 1119 * This function is used to negotiate a passive release of the current 1120 * process/lwp designation with the user scheduler, allowing the user 1121 * scheduler to schedule another user thread. The related kernel thread 1122 * (curthread) continues running in the released state. 1123 */ 1124 void 1125 lwkt_passive_release(struct thread *td) 1126 { 1127 struct lwp *lp = td->td_lwp; 1128 1129 td->td_release = NULL; 1130 lwkt_setpri_self(TDPRI_KERN_USER); 1131 1132 lp->lwp_proc->p_usched->release_curproc(lp); 1133 } 1134 1135 1136 /* 1137 * This implements a LWKT yield, allowing a kernel thread to yield to other 1138 * kernel threads at the same or higher priority. This function can be 1139 * called in a tight loop and will typically only yield once per tick. 1140 * 1141 * Most kernel threads run at the same priority in order to allow equal 1142 * sharing. 1143 * 1144 * (self contained on a per cpu basis) 1145 */ 1146 void 1147 lwkt_yield(void) 1148 { 1149 globaldata_t gd = mycpu; 1150 thread_t td = gd->gd_curthread; 1151 1152 /* 1153 * Should never be called with spinlocks held but there is a path 1154 * via ACPI where it might happen. 1155 */ 1156 if (gd->gd_spinlocks) 1157 return; 1158 1159 /* 1160 * Safe to call splz if we are not too-heavily nested. 1161 */ 1162 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1163 splz(); 1164 1165 /* 1166 * Caller allows switching 1167 */ 1168 if (lwkt_resched_wanted()) { 1169 atomic_set_int(&td->td_mpflags, TDF_MP_DIDYIELD); 1170 lwkt_schedule_self(td); 1171 lwkt_switch(); 1172 } 1173 } 1174 1175 /* 1176 * The quick version processes pending interrupts and higher-priority 1177 * LWKT threads but will not round-robin same-priority LWKT threads. 1178 * 1179 * When called while attempting to return to userland the only same-pri 1180 * threads are the ones which have already tried to become the current 1181 * user process. 1182 */ 1183 void 1184 lwkt_yield_quick(void) 1185 { 1186 globaldata_t gd = mycpu; 1187 thread_t td = gd->gd_curthread; 1188 1189 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1190 splz(); 1191 if (lwkt_resched_wanted()) { 1192 crit_enter(); 1193 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) { 1194 clear_lwkt_resched(); 1195 } else { 1196 atomic_set_int(&td->td_mpflags, TDF_MP_DIDYIELD); 1197 lwkt_schedule_self(curthread); 1198 lwkt_switch(); 1199 } 1200 crit_exit(); 1201 } 1202 } 1203 1204 /* 1205 * This yield is designed for kernel threads with a user context. 1206 * 1207 * The kernel acting on behalf of the user is potentially cpu-bound, 1208 * this function will efficiently allow other threads to run and also 1209 * switch to other processes by releasing. 1210 * 1211 * The lwkt_user_yield() function is designed to have very low overhead 1212 * if no yield is determined to be needed. 1213 */ 1214 void 1215 lwkt_user_yield(void) 1216 { 1217 globaldata_t gd = mycpu; 1218 thread_t td = gd->gd_curthread; 1219 1220 /* 1221 * Should never be called with spinlocks held but there is a path 1222 * via ACPI where it might happen. 1223 */ 1224 if (gd->gd_spinlocks) 1225 return; 1226 1227 /* 1228 * Always run any pending interrupts in case we are in a critical 1229 * section. 1230 */ 1231 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1232 splz(); 1233 1234 /* 1235 * Switch (which forces a release) if another kernel thread needs 1236 * the cpu, if userland wants us to resched, or if our kernel 1237 * quantum has run out. 1238 */ 1239 if (lwkt_resched_wanted() || 1240 user_resched_wanted()) 1241 { 1242 lwkt_switch(); 1243 } 1244 1245 #if 0 1246 /* 1247 * Reacquire the current process if we are released. 1248 * 1249 * XXX not implemented atm. The kernel may be holding locks and such, 1250 * so we want the thread to continue to receive cpu. 1251 */ 1252 if (td->td_release == NULL && lp) { 1253 lp->lwp_proc->p_usched->acquire_curproc(lp); 1254 td->td_release = lwkt_passive_release; 1255 lwkt_setpri_self(TDPRI_USER_NORM); 1256 } 1257 #endif 1258 } 1259 1260 /* 1261 * Generic schedule. Possibly schedule threads belonging to other cpus and 1262 * deal with threads that might be blocked on a wait queue. 1263 * 1264 * We have a little helper inline function which does additional work after 1265 * the thread has been enqueued, including dealing with preemption and 1266 * setting need_lwkt_resched() (which prevents the kernel from returning 1267 * to userland until it has processed higher priority threads). 1268 * 1269 * It is possible for this routine to be called after a failed _enqueue 1270 * (due to the target thread migrating, sleeping, or otherwise blocked). 1271 * We have to check that the thread is actually on the run queue! 1272 */ 1273 static __inline 1274 void 1275 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount) 1276 { 1277 if (ntd->td_flags & TDF_RUNQ) { 1278 if (ntd->td_preemptable) { 1279 ntd->td_preemptable(ntd, ccount); /* YYY +token */ 1280 } 1281 } 1282 } 1283 1284 static __inline 1285 void 1286 _lwkt_schedule(thread_t td) 1287 { 1288 globaldata_t mygd = mycpu; 1289 1290 KASSERT(td != &td->td_gd->gd_idlethread, 1291 ("lwkt_schedule(): scheduling gd_idlethread is illegal!")); 1292 KKASSERT((td->td_flags & TDF_MIGRATING) == 0); 1293 crit_enter_gd(mygd); 1294 KKASSERT(td->td_lwp == NULL || 1295 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 1296 1297 if (td == mygd->gd_curthread) { 1298 _lwkt_enqueue(td); 1299 } else { 1300 /* 1301 * If we own the thread, there is no race (since we are in a 1302 * critical section). If we do not own the thread there might 1303 * be a race but the target cpu will deal with it. 1304 */ 1305 if (td->td_gd == mygd) { 1306 _lwkt_enqueue(td); 1307 _lwkt_schedule_post(mygd, td, 1); 1308 } else { 1309 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0); 1310 } 1311 } 1312 crit_exit_gd(mygd); 1313 } 1314 1315 void 1316 lwkt_schedule(thread_t td) 1317 { 1318 _lwkt_schedule(td); 1319 } 1320 1321 void 1322 lwkt_schedule_noresched(thread_t td) /* XXX not impl */ 1323 { 1324 _lwkt_schedule(td); 1325 } 1326 1327 /* 1328 * When scheduled remotely if frame != NULL the IPIQ is being 1329 * run via doreti or an interrupt then preemption can be allowed. 1330 * 1331 * To allow preemption we have to drop the critical section so only 1332 * one is present in _lwkt_schedule_post. 1333 */ 1334 static void 1335 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame) 1336 { 1337 thread_t td = curthread; 1338 thread_t ntd = arg; 1339 1340 if (frame && ntd->td_preemptable) { 1341 crit_exit_noyield(td); 1342 _lwkt_schedule(ntd); 1343 crit_enter_quick(td); 1344 } else { 1345 _lwkt_schedule(ntd); 1346 } 1347 } 1348 1349 /* 1350 * Thread migration using a 'Pull' method. The thread may or may not be 1351 * the current thread. It MUST be descheduled and in a stable state. 1352 * lwkt_giveaway() must be called on the cpu owning the thread. 1353 * 1354 * At any point after lwkt_giveaway() is called, the target cpu may 1355 * 'pull' the thread by calling lwkt_acquire(). 1356 * 1357 * We have to make sure the thread is not sitting on a per-cpu tsleep 1358 * queue or it will blow up when it moves to another cpu. 1359 * 1360 * MPSAFE - must be called under very specific conditions. 1361 */ 1362 void 1363 lwkt_giveaway(thread_t td) 1364 { 1365 globaldata_t gd = mycpu; 1366 1367 crit_enter_gd(gd); 1368 if (td->td_flags & TDF_TSLEEPQ) 1369 tsleep_remove(td); 1370 KKASSERT(td->td_gd == gd); 1371 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); 1372 td->td_flags |= TDF_MIGRATING; 1373 crit_exit_gd(gd); 1374 } 1375 1376 void 1377 lwkt_acquire(thread_t td) 1378 { 1379 globaldata_t gd; 1380 globaldata_t mygd; 1381 1382 KKASSERT(td->td_flags & TDF_MIGRATING); 1383 gd = td->td_gd; 1384 mygd = mycpu; 1385 if (gd != mycpu) { 1386 #ifdef LOOPMASK 1387 uint64_t tsc_base = rdtsc(); 1388 #endif 1389 cpu_lfence(); 1390 KKASSERT((td->td_flags & TDF_RUNQ) == 0); 1391 crit_enter_gd(mygd); 1392 DEBUG_PUSH_INFO("lwkt_acquire"); 1393 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) { 1394 lwkt_process_ipiq(); 1395 cpu_lfence(); 1396 #ifdef _KERNEL_VIRTUAL 1397 vkernel_yield(); 1398 #endif 1399 #ifdef LOOPMASK 1400 if (tsc_frequency && rdtsc() - tsc_base > tsc_frequency) { 1401 kprintf("lwkt_acquire: stuck td %p td->td_flags %08x\n", 1402 td, td->td_flags); 1403 tsc_base = rdtsc(); 1404 } 1405 #endif 1406 } 1407 DEBUG_POP_INFO(); 1408 cpu_mfence(); 1409 td->td_gd = mygd; 1410 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1411 td->td_flags &= ~TDF_MIGRATING; 1412 crit_exit_gd(mygd); 1413 } else { 1414 crit_enter_gd(mygd); 1415 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1416 td->td_flags &= ~TDF_MIGRATING; 1417 crit_exit_gd(mygd); 1418 } 1419 } 1420 1421 /* 1422 * Generic deschedule. Descheduling threads other then your own should be 1423 * done only in carefully controlled circumstances. Descheduling is 1424 * asynchronous. 1425 * 1426 * This function may block if the cpu has run out of messages. 1427 */ 1428 void 1429 lwkt_deschedule(thread_t td) 1430 { 1431 crit_enter(); 1432 if (td == curthread) { 1433 _lwkt_dequeue(td); 1434 } else { 1435 if (td->td_gd == mycpu) { 1436 _lwkt_dequeue(td); 1437 } else { 1438 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td); 1439 } 1440 } 1441 crit_exit(); 1442 } 1443 1444 /* 1445 * Set the target thread's priority. This routine does not automatically 1446 * switch to a higher priority thread, LWKT threads are not designed for 1447 * continuous priority changes. Yield if you want to switch. 1448 */ 1449 void 1450 lwkt_setpri(thread_t td, int pri) 1451 { 1452 if (td->td_pri != pri) { 1453 KKASSERT(pri >= 0); 1454 crit_enter(); 1455 if (td->td_flags & TDF_RUNQ) { 1456 KKASSERT(td->td_gd == mycpu); 1457 _lwkt_dequeue(td); 1458 td->td_pri = pri; 1459 _lwkt_enqueue(td); 1460 } else { 1461 td->td_pri = pri; 1462 } 1463 crit_exit(); 1464 } 1465 } 1466 1467 /* 1468 * Set the initial priority for a thread prior to it being scheduled for 1469 * the first time. The thread MUST NOT be scheduled before or during 1470 * this call. The thread may be assigned to a cpu other then the current 1471 * cpu. 1472 * 1473 * Typically used after a thread has been created with TDF_STOPPREQ, 1474 * and before the thread is initially scheduled. 1475 */ 1476 void 1477 lwkt_setpri_initial(thread_t td, int pri) 1478 { 1479 KKASSERT(pri >= 0); 1480 KKASSERT((td->td_flags & TDF_RUNQ) == 0); 1481 td->td_pri = pri; 1482 } 1483 1484 void 1485 lwkt_setpri_self(int pri) 1486 { 1487 thread_t td = curthread; 1488 1489 KKASSERT(pri >= 0 && pri <= TDPRI_MAX); 1490 crit_enter(); 1491 if (td->td_flags & TDF_RUNQ) { 1492 _lwkt_dequeue(td); 1493 td->td_pri = pri; 1494 _lwkt_enqueue(td); 1495 } else { 1496 td->td_pri = pri; 1497 } 1498 crit_exit(); 1499 } 1500 1501 /* 1502 * hz tick scheduler clock for LWKT threads 1503 */ 1504 void 1505 lwkt_schedulerclock(thread_t td) 1506 { 1507 globaldata_t gd = td->td_gd; 1508 thread_t xtd; 1509 1510 xtd = TAILQ_FIRST(&gd->gd_tdrunq); 1511 if (xtd == td) { 1512 /* 1513 * If the current thread is at the head of the runq shift it to the 1514 * end of any equal-priority threads and request a LWKT reschedule 1515 * if it moved. 1516 * 1517 * Ignore upri in this situation. There will only be one user thread 1518 * in user mode, all others will be user threads running in kernel 1519 * mode and we have to make sure they get some cpu. 1520 */ 1521 xtd = TAILQ_NEXT(td, td_threadq); 1522 if (xtd && xtd->td_pri == td->td_pri) { 1523 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq); 1524 while (xtd && xtd->td_pri == td->td_pri) 1525 xtd = TAILQ_NEXT(xtd, td_threadq); 1526 if (xtd) 1527 TAILQ_INSERT_BEFORE(xtd, td, td_threadq); 1528 else 1529 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 1530 need_lwkt_resched(); 1531 } 1532 } else if (xtd) { 1533 /* 1534 * If we scheduled a thread other than the one at the head of the 1535 * queue always request a reschedule every tick. 1536 */ 1537 need_lwkt_resched(); 1538 } 1539 /* else curthread probably the idle thread, no need to reschedule */ 1540 } 1541 1542 /* 1543 * Migrate the current thread to the specified cpu. 1544 * 1545 * This is accomplished by descheduling ourselves from the current cpu 1546 * and setting td_migrate_gd. The lwkt_switch() code will detect that the 1547 * 'old' thread wants to migrate after it has been completely switched out 1548 * and will complete the migration. 1549 * 1550 * TDF_MIGRATING prevents scheduling races while the thread is being migrated. 1551 * 1552 * We must be sure to release our current process designation (if a user 1553 * process) before clearing out any tsleepq we are on because the release 1554 * code may re-add us. 1555 * 1556 * We must be sure to remove ourselves from the current cpu's tsleepq 1557 * before potentially moving to another queue. The thread can be on 1558 * a tsleepq due to a left-over tsleep_interlock(). 1559 */ 1560 1561 void 1562 lwkt_setcpu_self(globaldata_t rgd) 1563 { 1564 thread_t td = curthread; 1565 1566 if (td->td_gd != rgd) { 1567 crit_enter_quick(td); 1568 1569 if (td->td_release) 1570 td->td_release(td); 1571 if (td->td_flags & TDF_TSLEEPQ) 1572 tsleep_remove(td); 1573 1574 /* 1575 * Set TDF_MIGRATING to prevent a spurious reschedule while we are 1576 * trying to deschedule ourselves and switch away, then deschedule 1577 * ourself, remove us from tdallq, and set td_migrate_gd. Finally, 1578 * call lwkt_switch() to complete the operation. 1579 */ 1580 td->td_flags |= TDF_MIGRATING; 1581 lwkt_deschedule_self(td); 1582 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1583 td->td_migrate_gd = rgd; 1584 lwkt_switch(); 1585 1586 /* 1587 * We are now on the target cpu 1588 */ 1589 KKASSERT(rgd == mycpu); 1590 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); 1591 crit_exit_quick(td); 1592 } 1593 } 1594 1595 void 1596 lwkt_migratecpu(int cpuid) 1597 { 1598 globaldata_t rgd; 1599 1600 rgd = globaldata_find(cpuid); 1601 lwkt_setcpu_self(rgd); 1602 } 1603 1604 /* 1605 * Remote IPI for cpu migration (called while in a critical section so we 1606 * do not have to enter another one). 1607 * 1608 * The thread (td) has already been completely descheduled from the 1609 * originating cpu and we can simply assert the case. The thread is 1610 * assigned to the new cpu and enqueued. 1611 * 1612 * The thread will re-add itself to tdallq when it resumes execution. 1613 */ 1614 static void 1615 lwkt_setcpu_remote(void *arg) 1616 { 1617 thread_t td = arg; 1618 globaldata_t gd = mycpu; 1619 1620 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0); 1621 td->td_gd = gd; 1622 cpu_mfence(); 1623 td->td_flags &= ~TDF_MIGRATING; 1624 KKASSERT(td->td_migrate_gd == NULL); 1625 KKASSERT(td->td_lwp == NULL || 1626 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 1627 _lwkt_enqueue(td); 1628 } 1629 1630 struct lwp * 1631 lwkt_preempted_proc(void) 1632 { 1633 thread_t td = curthread; 1634 while (td->td_preempted) 1635 td = td->td_preempted; 1636 return(td->td_lwp); 1637 } 1638 1639 /* 1640 * Create a kernel process/thread/whatever. It shares it's address space 1641 * with proc0 - ie: kernel only. 1642 * 1643 * If the cpu is not specified one will be selected. In the future 1644 * specifying a cpu of -1 will enable kernel thread migration between 1645 * cpus. 1646 */ 1647 int 1648 lwkt_create(void (*func)(void *), void *arg, struct thread **tdp, 1649 thread_t template, int tdflags, int cpu, const char *fmt, ...) 1650 { 1651 thread_t td; 1652 __va_list ap; 1653 1654 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu, 1655 tdflags); 1656 if (tdp) 1657 *tdp = td; 1658 cpu_set_thread_handler(td, lwkt_exit, func, arg); 1659 1660 /* 1661 * Set up arg0 for 'ps' etc 1662 */ 1663 __va_start(ap, fmt); 1664 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap); 1665 __va_end(ap); 1666 1667 /* 1668 * Schedule the thread to run 1669 */ 1670 if (td->td_flags & TDF_NOSTART) 1671 td->td_flags &= ~TDF_NOSTART; 1672 else 1673 lwkt_schedule(td); 1674 return 0; 1675 } 1676 1677 /* 1678 * Destroy an LWKT thread. Warning! This function is not called when 1679 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and 1680 * uses a different reaping mechanism. 1681 */ 1682 void 1683 lwkt_exit(void) 1684 { 1685 thread_t td = curthread; 1686 thread_t std; 1687 globaldata_t gd; 1688 1689 /* 1690 * Do any cleanup that might block here 1691 */ 1692 biosched_done(td); 1693 dsched_exit_thread(td); 1694 1695 /* 1696 * Get us into a critical section to interlock gd_freetd and loop 1697 * until we can get it freed. 1698 * 1699 * We have to cache the current td in gd_freetd because objcache_put()ing 1700 * it would rip it out from under us while our thread is still active. 1701 * 1702 * We are the current thread so of course our own TDF_RUNNING bit will 1703 * be set, so unlike the lwp reap code we don't wait for it to clear. 1704 */ 1705 gd = mycpu; 1706 crit_enter_quick(td); 1707 for (;;) { 1708 if (td->td_refs) { 1709 tsleep(td, 0, "tdreap", 1); 1710 continue; 1711 } 1712 if ((std = gd->gd_freetd) != NULL) { 1713 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0); 1714 gd->gd_freetd = NULL; 1715 objcache_put(thread_cache, std); 1716 continue; 1717 } 1718 break; 1719 } 1720 1721 /* 1722 * Remove thread resources from kernel lists and deschedule us for 1723 * the last time. We cannot block after this point or we may end 1724 * up with a stale td on the tsleepq. 1725 * 1726 * None of this may block, the critical section is the only thing 1727 * protecting tdallq and the only thing preventing new lwkt_hold() 1728 * thread refs now. 1729 */ 1730 if (td->td_flags & TDF_TSLEEPQ) 1731 tsleep_remove(td); 1732 lwkt_deschedule_self(td); 1733 lwkt_remove_tdallq(td); 1734 KKASSERT(td->td_refs == 0); 1735 1736 /* 1737 * Final cleanup 1738 */ 1739 KKASSERT(gd->gd_freetd == NULL); 1740 if (td->td_flags & TDF_ALLOCATED_THREAD) 1741 gd->gd_freetd = td; 1742 cpu_thread_exit(); 1743 } 1744 1745 void 1746 lwkt_remove_tdallq(thread_t td) 1747 { 1748 KKASSERT(td->td_gd == mycpu); 1749 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1750 } 1751 1752 /* 1753 * Code reduction and branch prediction improvements. Call/return 1754 * overhead on modern cpus often degenerates into 0 cycles due to 1755 * the cpu's branch prediction hardware and return pc cache. We 1756 * can take advantage of this by not inlining medium-complexity 1757 * functions and we can also reduce the branch prediction impact 1758 * by collapsing perfectly predictable branches into a single 1759 * procedure instead of duplicating it. 1760 * 1761 * Is any of this noticeable? Probably not, so I'll take the 1762 * smaller code size. 1763 */ 1764 void 1765 crit_exit_wrapper(__DEBUG_CRIT_ARG__) 1766 { 1767 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__); 1768 } 1769 1770 void 1771 crit_panic(void) 1772 { 1773 thread_t td = curthread; 1774 int lcrit = td->td_critcount; 1775 1776 td->td_critcount = 0; 1777 cpu_ccfence(); 1778 panic("td_critcount is/would-go negative! %p %d", td, lcrit); 1779 /* NOT REACHED */ 1780 } 1781 1782 /* 1783 * Called from debugger/panic on cpus which have been stopped. We must still 1784 * process the IPIQ while stopped. 1785 * 1786 * If we are dumping also try to process any pending interrupts. This may 1787 * or may not work depending on the state of the cpu at the point it was 1788 * stopped. 1789 */ 1790 void 1791 lwkt_smp_stopped(void) 1792 { 1793 globaldata_t gd = mycpu; 1794 1795 if (dumping) { 1796 lwkt_process_ipiq(); 1797 --gd->gd_intr_nesting_level; 1798 splz(); 1799 ++gd->gd_intr_nesting_level; 1800 } else { 1801 lwkt_process_ipiq(); 1802 } 1803 cpu_smp_stopped(); 1804 } 1805