1 /* 2 * Copyright (c) 2003-2011 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@backplane.com> 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * 3. Neither the name of The DragonFly Project nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific, prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 */ 34 35 /* 36 * Each cpu in a system has its own self-contained light weight kernel 37 * thread scheduler, which means that generally speaking we only need 38 * to use a critical section to avoid problems. Foreign thread 39 * scheduling is queued via (async) IPIs. 40 */ 41 42 #include <sys/param.h> 43 #include <sys/systm.h> 44 #include <sys/kernel.h> 45 #include <sys/proc.h> 46 #include <sys/rtprio.h> 47 #include <sys/kinfo.h> 48 #include <sys/queue.h> 49 #include <sys/sysctl.h> 50 #include <sys/kthread.h> 51 #include <machine/cpu.h> 52 #include <sys/lock.h> 53 #include <sys/spinlock.h> 54 #include <sys/ktr.h> 55 56 #include <sys/thread2.h> 57 #include <sys/spinlock2.h> 58 #include <sys/mplock2.h> 59 60 #include <sys/dsched.h> 61 62 #include <vm/vm.h> 63 #include <vm/vm_param.h> 64 #include <vm/vm_kern.h> 65 #include <vm/vm_object.h> 66 #include <vm/vm_page.h> 67 #include <vm/vm_map.h> 68 #include <vm/vm_pager.h> 69 #include <vm/vm_extern.h> 70 71 #include <machine/stdarg.h> 72 #include <machine/smp.h> 73 #include <machine/clock.h> 74 75 #ifdef _KERNEL_VIRTUAL 76 #include <pthread.h> 77 #endif 78 79 #define LOOPMASK 80 81 #if !defined(KTR_CTXSW) 82 #define KTR_CTXSW KTR_ALL 83 #endif 84 KTR_INFO_MASTER(ctxsw); 85 KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p", int cpu, struct thread *td); 86 KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p", int cpu, struct thread *td); 87 KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s", struct thread *td, char *comm); 88 KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", struct thread *td); 89 90 static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads"); 91 92 #ifdef INVARIANTS 93 static int panic_on_cscount = 0; 94 #endif 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 static int lwkt_use_spin_port; 100 static struct objcache *thread_cache; 101 int cpu_mwait_spin = 0; 102 103 static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame); 104 static void lwkt_setcpu_remote(void *arg); 105 106 /* 107 * We can make all thread ports use the spin backend instead of the thread 108 * backend. This should only be set to debug the spin backend. 109 */ 110 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port); 111 112 #ifdef INVARIANTS 113 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, 114 "Panic if attempting to switch lwkt's while mastering cpusync"); 115 #endif 116 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, 117 "Number of switched threads"); 118 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, 119 "Successful preemption events"); 120 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, 121 "Failed preemption events"); 122 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, 123 "Number of preempted threads."); 124 static int fairq_enable = 0; 125 SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW, 126 &fairq_enable, 0, "Turn on fairq priority accumulators"); 127 static int fairq_bypass = -1; 128 SYSCTL_INT(_lwkt, OID_AUTO, fairq_bypass, CTLFLAG_RW, 129 &fairq_bypass, 0, "Allow fairq to bypass td on token failure"); 130 extern int lwkt_sched_debug; 131 int lwkt_sched_debug = 0; 132 SYSCTL_INT(_lwkt, OID_AUTO, sched_debug, CTLFLAG_RW, 133 &lwkt_sched_debug, 0, "Scheduler debug"); 134 static u_int lwkt_spin_loops = 10; 135 SYSCTL_UINT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW, 136 &lwkt_spin_loops, 0, "Scheduler spin loops until sorted decon"); 137 static int preempt_enable = 1; 138 SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW, 139 &preempt_enable, 0, "Enable preemption"); 140 static int lwkt_cache_threads = 0; 141 SYSCTL_INT(_lwkt, OID_AUTO, cache_threads, CTLFLAG_RD, 142 &lwkt_cache_threads, 0, "thread+kstack cache"); 143 144 /* 145 * These helper procedures handle the runq, they can only be called from 146 * within a critical section. 147 * 148 * WARNING! Prior to SMP being brought up it is possible to enqueue and 149 * dequeue threads belonging to other cpus, so be sure to use td->td_gd 150 * instead of 'mycpu' when referencing the globaldata structure. Once 151 * SMP live enqueuing and dequeueing only occurs on the current cpu. 152 */ 153 static __inline 154 void 155 _lwkt_dequeue(thread_t td) 156 { 157 if (td->td_flags & TDF_RUNQ) { 158 struct globaldata *gd = td->td_gd; 159 160 td->td_flags &= ~TDF_RUNQ; 161 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq); 162 --gd->gd_tdrunqcount; 163 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL) 164 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING); 165 } 166 } 167 168 /* 169 * Priority enqueue. 170 * 171 * There are a limited number of lwkt threads runnable since user 172 * processes only schedule one at a time per cpu. However, there can 173 * be many user processes in kernel mode exiting from a tsleep() which 174 * become 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; 186 187 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) { 188 struct globaldata *gd = td->td_gd; 189 190 td->td_flags |= TDF_RUNQ; 191 xtd = TAILQ_FIRST(&gd->gd_tdrunq); 192 if (xtd == NULL) { 193 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 194 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING); 195 } else { 196 /* 197 * NOTE: td_upri - higher numbers more desireable, same sense 198 * as td_pri (typically reversed from lwp_upri). 199 * 200 * In the equal priority case we want the best selection 201 * at the beginning so the less desireable selections know 202 * that they have to setrunqueue/go-to-another-cpu, even 203 * though it means switching back to the 'best' selection. 204 * This also avoids degenerate situations when many threads 205 * are runnable or waking up at the same time. 206 * 207 * If upri matches exactly place at end/round-robin. 208 */ 209 while (xtd && 210 (xtd->td_pri >= td->td_pri || 211 (xtd->td_pri == td->td_pri && 212 xtd->td_upri >= td->td_upri))) { 213 xtd = TAILQ_NEXT(xtd, td_threadq); 214 } 215 if (xtd) 216 TAILQ_INSERT_BEFORE(xtd, td, td_threadq); 217 else 218 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 219 } 220 ++gd->gd_tdrunqcount; 221 222 /* 223 * Request a LWKT reschedule if we are now at the head of the queue. 224 */ 225 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) 226 need_lwkt_resched(); 227 } 228 } 229 230 static boolean_t 231 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags) 232 { 233 struct thread *td = (struct thread *)obj; 234 235 td->td_kstack = NULL; 236 td->td_kstack_size = 0; 237 td->td_flags = TDF_ALLOCATED_THREAD; 238 td->td_mpflags = 0; 239 return (1); 240 } 241 242 static void 243 _lwkt_thread_dtor(void *obj, void *privdata) 244 { 245 struct thread *td = (struct thread *)obj; 246 247 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD, 248 ("_lwkt_thread_dtor: not allocated from objcache")); 249 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack && 250 td->td_kstack_size > 0, 251 ("_lwkt_thread_dtor: corrupted stack")); 252 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size); 253 td->td_kstack = NULL; 254 td->td_flags = 0; 255 } 256 257 /* 258 * Initialize the lwkt s/system. 259 * 260 * Nominally cache up to 32 thread + kstack structures. Cache more on 261 * systems with a lot of cpu cores. 262 */ 263 static void 264 lwkt_init(void) 265 { 266 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads); 267 if (lwkt_cache_threads == 0) { 268 lwkt_cache_threads = ncpus * 4; 269 if (lwkt_cache_threads < 32) 270 lwkt_cache_threads = 32; 271 } 272 thread_cache = objcache_create_mbacked( 273 M_THREAD, sizeof(struct thread), 274 0, lwkt_cache_threads, 275 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL); 276 } 277 SYSINIT(lwkt_init, SI_BOOT2_LWKT_INIT, SI_ORDER_FIRST, lwkt_init, NULL); 278 279 /* 280 * Schedule a thread to run. As the current thread we can always safely 281 * schedule ourselves, and a shortcut procedure is provided for that 282 * function. 283 * 284 * (non-blocking, self contained on a per cpu basis) 285 */ 286 void 287 lwkt_schedule_self(thread_t td) 288 { 289 KKASSERT((td->td_flags & TDF_MIGRATING) == 0); 290 crit_enter_quick(td); 291 KASSERT(td != &td->td_gd->gd_idlethread, 292 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!")); 293 KKASSERT(td->td_lwp == NULL || 294 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 295 _lwkt_enqueue(td); 296 crit_exit_quick(td); 297 } 298 299 /* 300 * Deschedule a thread. 301 * 302 * (non-blocking, self contained on a per cpu basis) 303 */ 304 void 305 lwkt_deschedule_self(thread_t td) 306 { 307 crit_enter_quick(td); 308 _lwkt_dequeue(td); 309 crit_exit_quick(td); 310 } 311 312 /* 313 * LWKTs operate on a per-cpu basis 314 * 315 * WARNING! Called from early boot, 'mycpu' may not work yet. 316 */ 317 void 318 lwkt_gdinit(struct globaldata *gd) 319 { 320 TAILQ_INIT(&gd->gd_tdrunq); 321 TAILQ_INIT(&gd->gd_tdallq); 322 } 323 324 /* 325 * Create a new thread. The thread must be associated with a process context 326 * or LWKT start address before it can be scheduled. If the target cpu is 327 * -1 the thread will be created on the current cpu. 328 * 329 * If you intend to create a thread without a process context this function 330 * does everything except load the startup and switcher function. 331 */ 332 thread_t 333 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags) 334 { 335 static int cpu_rotator; 336 globaldata_t gd = mycpu; 337 void *stack; 338 339 /* 340 * If static thread storage is not supplied allocate a thread. Reuse 341 * a cached free thread if possible. gd_freetd is used to keep an exiting 342 * thread intact through the exit. 343 */ 344 if (td == NULL) { 345 crit_enter_gd(gd); 346 if ((td = gd->gd_freetd) != NULL) { 347 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK| 348 TDF_RUNQ)) == 0); 349 gd->gd_freetd = NULL; 350 } else { 351 td = objcache_get(thread_cache, M_WAITOK); 352 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK| 353 TDF_RUNQ)) == 0); 354 } 355 crit_exit_gd(gd); 356 KASSERT((td->td_flags & 357 (TDF_ALLOCATED_THREAD|TDF_RUNNING|TDF_PREEMPT_LOCK)) == 358 TDF_ALLOCATED_THREAD, 359 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags)); 360 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK); 361 } 362 363 /* 364 * Try to reuse cached stack. 365 */ 366 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) { 367 if (flags & TDF_ALLOCATED_STACK) { 368 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size); 369 stack = NULL; 370 } 371 } 372 if (stack == NULL) { 373 stack = (void *)kmem_alloc_stack(&kernel_map, stksize); 374 flags |= TDF_ALLOCATED_STACK; 375 } 376 if (cpu < 0) { 377 cpu = ++cpu_rotator; 378 cpu_ccfence(); 379 cpu %= ncpus; 380 } 381 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu)); 382 return(td); 383 } 384 385 /* 386 * Initialize a preexisting thread structure. This function is used by 387 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread. 388 * 389 * All threads start out in a critical section at a priority of 390 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as 391 * appropriate. This function may send an IPI message when the 392 * requested cpu is not the current cpu and consequently gd_tdallq may 393 * not be initialized synchronously from the point of view of the originating 394 * cpu. 395 * 396 * NOTE! we have to be careful in regards to creating threads for other cpus 397 * if SMP has not yet been activated. 398 */ 399 static void 400 lwkt_init_thread_remote(void *arg) 401 { 402 thread_t td = arg; 403 404 /* 405 * Protected by critical section held by IPI dispatch 406 */ 407 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq); 408 } 409 410 /* 411 * lwkt core thread structural initialization. 412 * 413 * NOTE: All threads are initialized as mpsafe threads. 414 */ 415 void 416 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags, 417 struct globaldata *gd) 418 { 419 globaldata_t mygd = mycpu; 420 421 bzero(td, sizeof(struct thread)); 422 td->td_kstack = stack; 423 td->td_kstack_size = stksize; 424 td->td_flags = flags; 425 td->td_mpflags = 0; 426 td->td_type = TD_TYPE_GENERIC; 427 td->td_gd = gd; 428 td->td_pri = TDPRI_KERN_DAEMON; 429 td->td_critcount = 1; 430 td->td_toks_have = NULL; 431 td->td_toks_stop = &td->td_toks_base; 432 if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT)) { 433 lwkt_initport_spin(&td->td_msgport, td, 434 (flags & TDF_FIXEDCPU) ? TRUE : FALSE); 435 } else { 436 lwkt_initport_thread(&td->td_msgport, td); 437 } 438 pmap_init_thread(td); 439 /* 440 * Normally initializing a thread for a remote cpu requires sending an 441 * IPI. However, the idlethread is setup before the other cpus are 442 * activated so we have to treat it as a special case. XXX manipulation 443 * of gd_tdallq requires the BGL. 444 */ 445 if (gd == mygd || td == &gd->gd_idlethread) { 446 crit_enter_gd(mygd); 447 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); 448 crit_exit_gd(mygd); 449 } else { 450 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td); 451 } 452 dsched_enter_thread(td); 453 } 454 455 void 456 lwkt_set_comm(thread_t td, const char *ctl, ...) 457 { 458 __va_list va; 459 460 __va_start(va, ctl); 461 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va); 462 __va_end(va); 463 KTR_LOG(ctxsw_newtd, td, td->td_comm); 464 } 465 466 /* 467 * Prevent the thread from getting destroyed. Note that unlike PHOLD/PRELE 468 * this does not prevent the thread from migrating to another cpu so the 469 * gd_tdallq state is not protected by this. 470 */ 471 void 472 lwkt_hold(thread_t td) 473 { 474 atomic_add_int(&td->td_refs, 1); 475 } 476 477 void 478 lwkt_rele(thread_t td) 479 { 480 KKASSERT(td->td_refs > 0); 481 atomic_add_int(&td->td_refs, -1); 482 } 483 484 void 485 lwkt_free_thread(thread_t td) 486 { 487 KKASSERT(td->td_refs == 0); 488 KKASSERT((td->td_flags & (TDF_RUNNING | TDF_PREEMPT_LOCK | 489 TDF_RUNQ | TDF_TSLEEPQ)) == 0); 490 if (td->td_flags & TDF_ALLOCATED_THREAD) { 491 objcache_put(thread_cache, td); 492 } else if (td->td_flags & TDF_ALLOCATED_STACK) { 493 /* client-allocated struct with internally allocated stack */ 494 KASSERT(td->td_kstack && td->td_kstack_size > 0, 495 ("lwkt_free_thread: corrupted stack")); 496 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size); 497 td->td_kstack = NULL; 498 td->td_kstack_size = 0; 499 } 500 501 KTR_LOG(ctxsw_deadtd, td); 502 } 503 504 505 /* 506 * Switch to the next runnable lwkt. If no LWKTs are runnable then 507 * switch to the idlethread. Switching must occur within a critical 508 * section to avoid races with the scheduling queue. 509 * 510 * We always have full control over our cpu's run queue. Other cpus 511 * that wish to manipulate our queue must use the cpu_*msg() calls to 512 * talk to our cpu, so a critical section is all that is needed and 513 * the result is very, very fast thread switching. 514 * 515 * The LWKT scheduler uses a fixed priority model and round-robins at 516 * each priority level. User process scheduling is a totally 517 * different beast and LWKT priorities should not be confused with 518 * user process priorities. 519 * 520 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch() 521 * is not called by the current thread in the preemption case, only when 522 * the preempting thread blocks (in order to return to the original thread). 523 * 524 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread 525 * migration and tsleep deschedule the current lwkt thread and call 526 * lwkt_switch(). In particular, the target cpu of the migration fully 527 * expects the thread to become non-runnable and can deadlock against 528 * cpusync operations if we run any IPIs prior to switching the thread out. 529 * 530 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF 531 * THE CURRENT THREAD HAS BEEN DESCHEDULED! 532 */ 533 void 534 lwkt_switch(void) 535 { 536 globaldata_t gd = mycpu; 537 thread_t td = gd->gd_curthread; 538 thread_t ntd; 539 int upri; 540 #ifdef LOOPMASK 541 uint64_t tsc_base = rdtsc(); 542 #endif 543 544 KKASSERT(gd->gd_processing_ipiq == 0); 545 KKASSERT(td->td_flags & TDF_RUNNING); 546 547 /* 548 * Switching from within a 'fast' (non thread switched) interrupt or IPI 549 * is illegal. However, we may have to do it anyway if we hit a fatal 550 * kernel trap or we have paniced. 551 * 552 * If this case occurs save and restore the interrupt nesting level. 553 */ 554 if (gd->gd_intr_nesting_level) { 555 int savegdnest; 556 int savegdtrap; 557 558 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) { 559 panic("lwkt_switch: Attempt to switch from a " 560 "fast interrupt, ipi, or hard code section, " 561 "td %p\n", 562 td); 563 } else { 564 savegdnest = gd->gd_intr_nesting_level; 565 savegdtrap = gd->gd_trap_nesting_level; 566 gd->gd_intr_nesting_level = 0; 567 gd->gd_trap_nesting_level = 0; 568 if ((td->td_flags & TDF_PANICWARN) == 0) { 569 td->td_flags |= TDF_PANICWARN; 570 kprintf("Warning: thread switch from interrupt, IPI, " 571 "or hard code section.\n" 572 "thread %p (%s)\n", td, td->td_comm); 573 print_backtrace(-1); 574 } 575 lwkt_switch(); 576 gd->gd_intr_nesting_level = savegdnest; 577 gd->gd_trap_nesting_level = savegdtrap; 578 return; 579 } 580 } 581 582 /* 583 * Release our current user process designation if we are blocking 584 * or if a user reschedule was requested. 585 * 586 * NOTE: This function is NOT called if we are switching into or 587 * returning from a preemption. 588 * 589 * NOTE: Releasing our current user process designation may cause 590 * it to be assigned to another thread, which in turn will 591 * cause us to block in the usched acquire code when we attempt 592 * to return to userland. 593 * 594 * NOTE: On SMP systems this can be very nasty when heavy token 595 * contention is present so we want to be careful not to 596 * release the designation gratuitously. 597 */ 598 if (td->td_release && 599 (user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) { 600 td->td_release(td); 601 } 602 603 /* 604 * Release all tokens. Once we do this we must remain in the critical 605 * section and cannot run IPIs or other interrupts until we switch away 606 * because they may implode if they try to get a token using our thread 607 * context. 608 */ 609 crit_enter_gd(gd); 610 if (TD_TOKS_HELD(td)) 611 lwkt_relalltokens(td); 612 613 /* 614 * We had better not be holding any spin locks, but don't get into an 615 * endless panic loop. 616 */ 617 KASSERT(gd->gd_spinlocks == 0 || panicstr != NULL, 618 ("lwkt_switch: still holding %d exclusive spinlocks!", 619 gd->gd_spinlocks)); 620 621 #ifdef INVARIANTS 622 if (td->td_cscount) { 623 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n", 624 td); 625 if (panic_on_cscount) 626 panic("switching while mastering cpusync"); 627 } 628 #endif 629 630 /* 631 * If we had preempted another thread on this cpu, resume the preempted 632 * thread. This occurs transparently, whether the preempted thread 633 * was scheduled or not (it may have been preempted after descheduling 634 * itself). 635 * 636 * We have to setup the MP lock for the original thread after backing 637 * out the adjustment that was made to curthread when the original 638 * was preempted. 639 */ 640 if ((ntd = td->td_preempted) != NULL) { 641 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK); 642 ntd->td_flags |= TDF_PREEMPT_DONE; 643 ntd->td_contended = 0; /* reset contended */ 644 645 /* 646 * The interrupt may have woken a thread up, we need to properly 647 * set the reschedule flag if the originally interrupted thread is 648 * at a lower priority. 649 * 650 * The interrupt may not have descheduled. 651 */ 652 if (TAILQ_FIRST(&gd->gd_tdrunq) != ntd) 653 need_lwkt_resched(); 654 goto havethread_preempted; 655 } 656 657 /* 658 * Figure out switch target. If we cannot switch to our desired target 659 * look for a thread that we can switch to. 660 * 661 * NOTE! The limited spin loop and related parameters are extremely 662 * important for system performance, particularly for pipes and 663 * concurrent conflicting VM faults. 664 */ 665 clear_lwkt_resched(); 666 ntd = TAILQ_FIRST(&gd->gd_tdrunq); 667 668 if (ntd) { 669 do { 670 if (TD_TOKS_NOT_HELD(ntd) || 671 lwkt_getalltokens(ntd, (ntd->td_contended > lwkt_spin_loops))) 672 { 673 goto havethread; 674 } 675 ++gd->gd_cnt.v_lock_colls; 676 ++ntd->td_contended; /* overflow ok */ 677 #ifdef LOOPMASK 678 if (tsc_frequency && rdtsc() - tsc_base > tsc_frequency) { 679 kprintf("lwkt_switch: excessive contended %d " 680 "thread %p\n", ntd->td_contended, ntd); 681 tsc_base = rdtsc(); 682 } 683 #endif 684 } while (ntd->td_contended < (lwkt_spin_loops >> 1)); 685 upri = ntd->td_upri; 686 687 /* 688 * Bleh, the thread we wanted to switch to has a contended token. 689 * See if we can switch to another thread. 690 * 691 * We generally don't want to do this because it represents a 692 * priority inversion. Do not allow the case if the thread 693 * is returning to userland (not a kernel thread) AND the thread 694 * has a lower upri. 695 */ 696 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) { 697 if (ntd->td_pri < TDPRI_KERN_LPSCHED && upri > ntd->td_upri) 698 break; 699 upri = ntd->td_upri; 700 701 /* 702 * Try this one. 703 */ 704 if (TD_TOKS_NOT_HELD(ntd) || 705 lwkt_getalltokens(ntd, (ntd->td_contended > lwkt_spin_loops))) { 706 goto havethread; 707 } 708 ++ntd->td_contended; /* overflow ok */ 709 ++gd->gd_cnt.v_lock_colls; 710 } 711 712 /* 713 * Fall through, switch to idle thread to get us out of the current 714 * context. Since we were contended, prevent HLT by flagging a 715 * LWKT reschedule. 716 */ 717 need_lwkt_resched(); 718 } 719 720 /* 721 * We either contended on ntd or the runq is empty. We must switch 722 * through the idle thread to get out of the current context. 723 */ 724 ntd = &gd->gd_idlethread; 725 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) 726 ASSERT_NO_TOKENS_HELD(ntd); 727 cpu_time.cp_msg[0] = 0; 728 goto haveidle; 729 730 havethread: 731 /* 732 * Clear gd_idle_repeat when doing a normal switch to a non-idle 733 * thread. 734 */ 735 ntd->td_wmesg = NULL; 736 ntd->td_contended = 0; /* reset once scheduled */ 737 ++gd->gd_cnt.v_swtch; 738 gd->gd_idle_repeat = 0; 739 740 havethread_preempted: 741 /* 742 * If the new target does not need the MP lock and we are holding it, 743 * release the MP lock. If the new target requires the MP lock we have 744 * already acquired it for the target. 745 */ 746 ; 747 haveidle: 748 KASSERT(ntd->td_critcount, 749 ("priority problem in lwkt_switch %d %d", 750 td->td_critcount, ntd->td_critcount)); 751 752 if (td != ntd) { 753 /* 754 * Execute the actual thread switch operation. This function 755 * returns to the current thread and returns the previous thread 756 * (which may be different from the thread we switched to). 757 * 758 * We are responsible for marking ntd as TDF_RUNNING. 759 */ 760 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0); 761 ++switch_count; 762 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd); 763 ntd->td_flags |= TDF_RUNNING; 764 lwkt_switch_return(td->td_switch(ntd)); 765 /* ntd invalid, td_switch() can return a different thread_t */ 766 } 767 768 /* 769 * catch-all. XXX is this strictly needed? 770 */ 771 splz_check(); 772 773 /* NOTE: current cpu may have changed after switch */ 774 crit_exit_quick(td); 775 } 776 777 /* 778 * Called by assembly in the td_switch (thread restore path) for thread 779 * bootstrap cases which do not 'return' to lwkt_switch(). 780 */ 781 void 782 lwkt_switch_return(thread_t otd) 783 { 784 globaldata_t rgd; 785 #ifdef LOOPMASK 786 uint64_t tsc_base = rdtsc(); 787 #endif 788 int exiting; 789 790 exiting = otd->td_flags & TDF_EXITING; 791 cpu_ccfence(); 792 793 /* 794 * Check if otd was migrating. Now that we are on ntd we can finish 795 * up the migration. This is a bit messy but it is the only place 796 * where td is known to be fully descheduled. 797 * 798 * We can only activate the migration if otd was migrating but not 799 * held on the cpu due to a preemption chain. We still have to 800 * clear TDF_RUNNING on the old thread either way. 801 * 802 * We are responsible for clearing the previously running thread's 803 * TDF_RUNNING. 804 */ 805 if ((rgd = otd->td_migrate_gd) != NULL && 806 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) { 807 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) == 808 (TDF_MIGRATING | TDF_RUNNING)); 809 otd->td_migrate_gd = NULL; 810 otd->td_flags &= ~TDF_RUNNING; 811 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd); 812 } else { 813 otd->td_flags &= ~TDF_RUNNING; 814 } 815 816 /* 817 * Final exit validations (see lwp_wait()). Note that otd becomes 818 * invalid the *instant* we set TDF_MP_EXITSIG. 819 * 820 * Use the EXITING status loaded from before we clear TDF_RUNNING, 821 * because if it is not set otd becomes invalid the instant we clear 822 * TDF_RUNNING on it (otherwise, if the system is fast enough, we 823 * might 'steal' TDF_EXITING from another switch-return!). 824 */ 825 while (exiting) { 826 u_int mpflags; 827 828 mpflags = otd->td_mpflags; 829 cpu_ccfence(); 830 831 if (mpflags & TDF_MP_EXITWAIT) { 832 if (atomic_cmpset_int(&otd->td_mpflags, mpflags, 833 mpflags | TDF_MP_EXITSIG)) { 834 wakeup(otd); 835 break; 836 } 837 } else { 838 if (atomic_cmpset_int(&otd->td_mpflags, mpflags, 839 mpflags | TDF_MP_EXITSIG)) { 840 wakeup(otd); 841 break; 842 } 843 } 844 845 #ifdef LOOPMASK 846 if (tsc_frequency && rdtsc() - tsc_base > tsc_frequency) { 847 kprintf("lwkt_switch_return: excessive TDF_EXITING " 848 "thread %p\n", otd); 849 tsc_base = rdtsc(); 850 } 851 #endif 852 } 853 } 854 855 /* 856 * Request that the target thread preempt the current thread. Preemption 857 * can only occur if our only critical section is the one that we were called 858 * with, the relative priority of the target thread is higher, and the target 859 * thread holds no tokens. This also only works if we are not holding any 860 * spinlocks (obviously). 861 * 862 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically 863 * this is called via lwkt_schedule() through the td_preemptable callback. 864 * critcount is the managed critical priority that we should ignore in order 865 * to determine whether preemption is possible (aka usually just the crit 866 * priority of lwkt_schedule() itself). 867 * 868 * Preemption is typically limited to interrupt threads. 869 * 870 * Operation works in a fairly straight-forward manner. The normal 871 * scheduling code is bypassed and we switch directly to the target 872 * thread. When the target thread attempts to block or switch away 873 * code at the base of lwkt_switch() will switch directly back to our 874 * thread. Our thread is able to retain whatever tokens it holds and 875 * if the target needs one of them the target will switch back to us 876 * and reschedule itself normally. 877 */ 878 void 879 lwkt_preempt(thread_t ntd, int critcount) 880 { 881 struct globaldata *gd = mycpu; 882 thread_t xtd; 883 thread_t td; 884 int save_gd_intr_nesting_level; 885 886 /* 887 * The caller has put us in a critical section. We can only preempt 888 * if the caller of the caller was not in a critical section (basically 889 * a local interrupt), as determined by the 'critcount' parameter. We 890 * also can't preempt if the caller is holding any spinlocks (even if 891 * he isn't in a critical section). This also handles the tokens test. 892 * 893 * YYY The target thread must be in a critical section (else it must 894 * inherit our critical section? I dunno yet). 895 */ 896 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri)); 897 898 td = gd->gd_curthread; 899 if (preempt_enable == 0) { 900 ++preempt_miss; 901 return; 902 } 903 if (ntd->td_pri <= td->td_pri) { 904 ++preempt_miss; 905 return; 906 } 907 if (td->td_critcount > critcount) { 908 ++preempt_miss; 909 return; 910 } 911 if (td->td_cscount) { 912 ++preempt_miss; 913 return; 914 } 915 if (ntd->td_gd != gd) { 916 ++preempt_miss; 917 return; 918 } 919 920 /* 921 * We don't have to check spinlocks here as they will also bump 922 * td_critcount. 923 * 924 * Do not try to preempt if the target thread is holding any tokens. 925 * We could try to acquire the tokens but this case is so rare there 926 * is no need to support it. 927 */ 928 KKASSERT(gd->gd_spinlocks == 0); 929 930 if (TD_TOKS_HELD(ntd)) { 931 ++preempt_miss; 932 return; 933 } 934 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) { 935 ++preempt_weird; 936 return; 937 } 938 if (ntd->td_preempted) { 939 ++preempt_hit; 940 return; 941 } 942 KKASSERT(gd->gd_processing_ipiq == 0); 943 944 /* 945 * Since we are able to preempt the current thread, there is no need to 946 * call need_lwkt_resched(). 947 * 948 * We must temporarily clear gd_intr_nesting_level around the switch 949 * since switchouts from the target thread are allowed (they will just 950 * return to our thread), and since the target thread has its own stack. 951 * 952 * A preemption must switch back to the original thread, assert the 953 * case. 954 */ 955 ++preempt_hit; 956 ntd->td_preempted = td; 957 td->td_flags |= TDF_PREEMPT_LOCK; 958 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd); 959 save_gd_intr_nesting_level = gd->gd_intr_nesting_level; 960 gd->gd_intr_nesting_level = 0; 961 962 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0); 963 ntd->td_flags |= TDF_RUNNING; 964 xtd = td->td_switch(ntd); 965 KKASSERT(xtd == ntd); 966 lwkt_switch_return(xtd); 967 gd->gd_intr_nesting_level = save_gd_intr_nesting_level; 968 969 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE)); 970 ntd->td_preempted = NULL; 971 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE); 972 } 973 974 /* 975 * Conditionally call splz() if gd_reqflags indicates work is pending. 976 * This will work inside a critical section but not inside a hard code 977 * section. 978 * 979 * (self contained on a per cpu basis) 980 */ 981 void 982 splz_check(void) 983 { 984 globaldata_t gd = mycpu; 985 thread_t td = gd->gd_curthread; 986 987 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && 988 gd->gd_intr_nesting_level == 0 && 989 td->td_nest_count < 2) 990 { 991 splz(); 992 } 993 } 994 995 /* 996 * This version is integrated into crit_exit, reqflags has already 997 * been tested but td_critcount has not. 998 * 999 * We only want to execute the splz() on the 1->0 transition of 1000 * critcount and not in a hard code section or if too deeply nested. 1001 * 1002 * NOTE: gd->gd_spinlocks is implied to be 0 when td_critcount is 0. 1003 */ 1004 void 1005 lwkt_maybe_splz(thread_t td) 1006 { 1007 globaldata_t gd = td->td_gd; 1008 1009 if (td->td_critcount == 0 && 1010 gd->gd_intr_nesting_level == 0 && 1011 td->td_nest_count < 2) 1012 { 1013 splz(); 1014 } 1015 } 1016 1017 /* 1018 * Drivers which set up processing co-threads can call this function to 1019 * run the co-thread at a higher priority and to allow it to preempt 1020 * normal threads. 1021 */ 1022 void 1023 lwkt_set_interrupt_support_thread(void) 1024 { 1025 thread_t td = curthread; 1026 1027 lwkt_setpri_self(TDPRI_INT_SUPPORT); 1028 td->td_flags |= TDF_INTTHREAD; 1029 td->td_preemptable = lwkt_preempt; 1030 } 1031 1032 1033 /* 1034 * This function is used to negotiate a passive release of the current 1035 * process/lwp designation with the user scheduler, allowing the user 1036 * scheduler to schedule another user thread. The related kernel thread 1037 * (curthread) continues running in the released state. 1038 */ 1039 void 1040 lwkt_passive_release(struct thread *td) 1041 { 1042 struct lwp *lp = td->td_lwp; 1043 1044 td->td_release = NULL; 1045 lwkt_setpri_self(TDPRI_KERN_USER); 1046 1047 lp->lwp_proc->p_usched->release_curproc(lp); 1048 } 1049 1050 1051 /* 1052 * This implements a LWKT yield, allowing a kernel thread to yield to other 1053 * kernel threads at the same or higher priority. This function can be 1054 * called in a tight loop and will typically only yield once per tick. 1055 * 1056 * Most kernel threads run at the same priority in order to allow equal 1057 * sharing. 1058 * 1059 * (self contained on a per cpu basis) 1060 */ 1061 void 1062 lwkt_yield(void) 1063 { 1064 globaldata_t gd = mycpu; 1065 thread_t td = gd->gd_curthread; 1066 1067 /* 1068 * Should never be called with spinlocks held but there is a path 1069 * via ACPI where it might happen. 1070 */ 1071 if (gd->gd_spinlocks) 1072 return; 1073 1074 /* 1075 * Safe to call splz if we are not too-heavily nested. 1076 */ 1077 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1078 splz(); 1079 1080 /* 1081 * Caller allows switching 1082 */ 1083 if (lwkt_resched_wanted()) { 1084 lwkt_schedule_self(curthread); 1085 lwkt_switch(); 1086 } 1087 } 1088 1089 /* 1090 * The quick version processes pending interrupts and higher-priority 1091 * LWKT threads but will not round-robin same-priority LWKT threads. 1092 * 1093 * When called while attempting to return to userland the only same-pri 1094 * threads are the ones which have already tried to become the current 1095 * user process. 1096 */ 1097 void 1098 lwkt_yield_quick(void) 1099 { 1100 globaldata_t gd = mycpu; 1101 thread_t td = gd->gd_curthread; 1102 1103 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1104 splz(); 1105 if (lwkt_resched_wanted()) { 1106 crit_enter(); 1107 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) { 1108 clear_lwkt_resched(); 1109 } else { 1110 lwkt_schedule_self(curthread); 1111 lwkt_switch(); 1112 } 1113 crit_exit(); 1114 } 1115 } 1116 1117 /* 1118 * This yield is designed for kernel threads with a user context. 1119 * 1120 * The kernel acting on behalf of the user is potentially cpu-bound, 1121 * this function will efficiently allow other threads to run and also 1122 * switch to other processes by releasing. 1123 * 1124 * The lwkt_user_yield() function is designed to have very low overhead 1125 * if no yield is determined to be needed. 1126 */ 1127 void 1128 lwkt_user_yield(void) 1129 { 1130 globaldata_t gd = mycpu; 1131 thread_t td = gd->gd_curthread; 1132 1133 /* 1134 * Should never be called with spinlocks held but there is a path 1135 * via ACPI where it might happen. 1136 */ 1137 if (gd->gd_spinlocks) 1138 return; 1139 1140 /* 1141 * Always run any pending interrupts in case we are in a critical 1142 * section. 1143 */ 1144 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1145 splz(); 1146 1147 /* 1148 * Switch (which forces a release) if another kernel thread needs 1149 * the cpu, if userland wants us to resched, or if our kernel 1150 * quantum has run out. 1151 */ 1152 if (lwkt_resched_wanted() || 1153 user_resched_wanted()) 1154 { 1155 lwkt_switch(); 1156 } 1157 1158 #if 0 1159 /* 1160 * Reacquire the current process if we are released. 1161 * 1162 * XXX not implemented atm. The kernel may be holding locks and such, 1163 * so we want the thread to continue to receive cpu. 1164 */ 1165 if (td->td_release == NULL && lp) { 1166 lp->lwp_proc->p_usched->acquire_curproc(lp); 1167 td->td_release = lwkt_passive_release; 1168 lwkt_setpri_self(TDPRI_USER_NORM); 1169 } 1170 #endif 1171 } 1172 1173 /* 1174 * Generic schedule. Possibly schedule threads belonging to other cpus and 1175 * deal with threads that might be blocked on a wait queue. 1176 * 1177 * We have a little helper inline function which does additional work after 1178 * the thread has been enqueued, including dealing with preemption and 1179 * setting need_lwkt_resched() (which prevents the kernel from returning 1180 * to userland until it has processed higher priority threads). 1181 * 1182 * It is possible for this routine to be called after a failed _enqueue 1183 * (due to the target thread migrating, sleeping, or otherwise blocked). 1184 * We have to check that the thread is actually on the run queue! 1185 */ 1186 static __inline 1187 void 1188 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount) 1189 { 1190 if (ntd->td_flags & TDF_RUNQ) { 1191 if (ntd->td_preemptable) { 1192 ntd->td_preemptable(ntd, ccount); /* YYY +token */ 1193 } 1194 } 1195 } 1196 1197 static __inline 1198 void 1199 _lwkt_schedule(thread_t td) 1200 { 1201 globaldata_t mygd = mycpu; 1202 1203 KASSERT(td != &td->td_gd->gd_idlethread, 1204 ("lwkt_schedule(): scheduling gd_idlethread is illegal!")); 1205 KKASSERT((td->td_flags & TDF_MIGRATING) == 0); 1206 crit_enter_gd(mygd); 1207 KKASSERT(td->td_lwp == NULL || 1208 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 1209 1210 if (td == mygd->gd_curthread) { 1211 _lwkt_enqueue(td); 1212 } else { 1213 /* 1214 * If we own the thread, there is no race (since we are in a 1215 * critical section). If we do not own the thread there might 1216 * be a race but the target cpu will deal with it. 1217 */ 1218 if (td->td_gd == mygd) { 1219 _lwkt_enqueue(td); 1220 _lwkt_schedule_post(mygd, td, 1); 1221 } else { 1222 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0); 1223 } 1224 } 1225 crit_exit_gd(mygd); 1226 } 1227 1228 void 1229 lwkt_schedule(thread_t td) 1230 { 1231 _lwkt_schedule(td); 1232 } 1233 1234 void 1235 lwkt_schedule_noresched(thread_t td) /* XXX not impl */ 1236 { 1237 _lwkt_schedule(td); 1238 } 1239 1240 /* 1241 * When scheduled remotely if frame != NULL the IPIQ is being 1242 * run via doreti or an interrupt then preemption can be allowed. 1243 * 1244 * To allow preemption we have to drop the critical section so only 1245 * one is present in _lwkt_schedule_post. 1246 */ 1247 static void 1248 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame) 1249 { 1250 thread_t td = curthread; 1251 thread_t ntd = arg; 1252 1253 if (frame && ntd->td_preemptable) { 1254 crit_exit_noyield(td); 1255 _lwkt_schedule(ntd); 1256 crit_enter_quick(td); 1257 } else { 1258 _lwkt_schedule(ntd); 1259 } 1260 } 1261 1262 /* 1263 * Thread migration using a 'Pull' method. The thread may or may not be 1264 * the current thread. It MUST be descheduled and in a stable state. 1265 * lwkt_giveaway() must be called on the cpu owning the thread. 1266 * 1267 * At any point after lwkt_giveaway() is called, the target cpu may 1268 * 'pull' the thread by calling lwkt_acquire(). 1269 * 1270 * We have to make sure the thread is not sitting on a per-cpu tsleep 1271 * queue or it will blow up when it moves to another cpu. 1272 * 1273 * MPSAFE - must be called under very specific conditions. 1274 */ 1275 void 1276 lwkt_giveaway(thread_t td) 1277 { 1278 globaldata_t gd = mycpu; 1279 1280 crit_enter_gd(gd); 1281 if (td->td_flags & TDF_TSLEEPQ) 1282 tsleep_remove(td); 1283 KKASSERT(td->td_gd == gd); 1284 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); 1285 td->td_flags |= TDF_MIGRATING; 1286 crit_exit_gd(gd); 1287 } 1288 1289 void 1290 lwkt_acquire(thread_t td) 1291 { 1292 globaldata_t gd; 1293 globaldata_t mygd; 1294 1295 KKASSERT(td->td_flags & TDF_MIGRATING); 1296 gd = td->td_gd; 1297 mygd = mycpu; 1298 if (gd != mycpu) { 1299 #ifdef LOOPMASK 1300 uint64_t tsc_base = rdtsc(); 1301 #endif 1302 cpu_lfence(); 1303 KKASSERT((td->td_flags & TDF_RUNQ) == 0); 1304 crit_enter_gd(mygd); 1305 DEBUG_PUSH_INFO("lwkt_acquire"); 1306 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) { 1307 lwkt_process_ipiq(); 1308 cpu_lfence(); 1309 #ifdef _KERNEL_VIRTUAL 1310 pthread_yield(); 1311 #endif 1312 #ifdef LOOPMASK 1313 if (tsc_frequency && rdtsc() - tsc_base > tsc_frequency) { 1314 kprintf("lwkt_acquire: stuck td %p td->td_flags %08x\n", 1315 td, td->td_flags); 1316 tsc_base = rdtsc(); 1317 } 1318 #endif 1319 } 1320 DEBUG_POP_INFO(); 1321 cpu_mfence(); 1322 td->td_gd = mygd; 1323 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1324 td->td_flags &= ~TDF_MIGRATING; 1325 crit_exit_gd(mygd); 1326 } else { 1327 crit_enter_gd(mygd); 1328 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1329 td->td_flags &= ~TDF_MIGRATING; 1330 crit_exit_gd(mygd); 1331 } 1332 } 1333 1334 /* 1335 * Generic deschedule. Descheduling threads other then your own should be 1336 * done only in carefully controlled circumstances. Descheduling is 1337 * asynchronous. 1338 * 1339 * This function may block if the cpu has run out of messages. 1340 */ 1341 void 1342 lwkt_deschedule(thread_t td) 1343 { 1344 crit_enter(); 1345 if (td == curthread) { 1346 _lwkt_dequeue(td); 1347 } else { 1348 if (td->td_gd == mycpu) { 1349 _lwkt_dequeue(td); 1350 } else { 1351 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td); 1352 } 1353 } 1354 crit_exit(); 1355 } 1356 1357 /* 1358 * Set the target thread's priority. This routine does not automatically 1359 * switch to a higher priority thread, LWKT threads are not designed for 1360 * continuous priority changes. Yield if you want to switch. 1361 */ 1362 void 1363 lwkt_setpri(thread_t td, int pri) 1364 { 1365 if (td->td_pri != pri) { 1366 KKASSERT(pri >= 0); 1367 crit_enter(); 1368 if (td->td_flags & TDF_RUNQ) { 1369 KKASSERT(td->td_gd == mycpu); 1370 _lwkt_dequeue(td); 1371 td->td_pri = pri; 1372 _lwkt_enqueue(td); 1373 } else { 1374 td->td_pri = pri; 1375 } 1376 crit_exit(); 1377 } 1378 } 1379 1380 /* 1381 * Set the initial priority for a thread prior to it being scheduled for 1382 * the first time. The thread MUST NOT be scheduled before or during 1383 * this call. The thread may be assigned to a cpu other then the current 1384 * cpu. 1385 * 1386 * Typically used after a thread has been created with TDF_STOPPREQ, 1387 * and before the thread is initially scheduled. 1388 */ 1389 void 1390 lwkt_setpri_initial(thread_t td, int pri) 1391 { 1392 KKASSERT(pri >= 0); 1393 KKASSERT((td->td_flags & TDF_RUNQ) == 0); 1394 td->td_pri = pri; 1395 } 1396 1397 void 1398 lwkt_setpri_self(int pri) 1399 { 1400 thread_t td = curthread; 1401 1402 KKASSERT(pri >= 0 && pri <= TDPRI_MAX); 1403 crit_enter(); 1404 if (td->td_flags & TDF_RUNQ) { 1405 _lwkt_dequeue(td); 1406 td->td_pri = pri; 1407 _lwkt_enqueue(td); 1408 } else { 1409 td->td_pri = pri; 1410 } 1411 crit_exit(); 1412 } 1413 1414 /* 1415 * hz tick scheduler clock for LWKT threads 1416 */ 1417 void 1418 lwkt_schedulerclock(thread_t td) 1419 { 1420 globaldata_t gd = td->td_gd; 1421 thread_t xtd; 1422 1423 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) { 1424 /* 1425 * If the current thread is at the head of the runq shift it to the 1426 * end of any equal-priority threads and request a LWKT reschedule 1427 * if it moved. 1428 * 1429 * Ignore upri in this situation. There will only be one user thread 1430 * in user mode, all others will be user threads running in kernel 1431 * mode and we have to make sure they get some cpu. 1432 */ 1433 xtd = TAILQ_NEXT(td, td_threadq); 1434 if (xtd && xtd->td_pri == td->td_pri) { 1435 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq); 1436 while (xtd && xtd->td_pri == td->td_pri) 1437 xtd = TAILQ_NEXT(xtd, td_threadq); 1438 if (xtd) 1439 TAILQ_INSERT_BEFORE(xtd, td, td_threadq); 1440 else 1441 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 1442 need_lwkt_resched(); 1443 } 1444 } else { 1445 /* 1446 * If we scheduled a thread other than the one at the head of the 1447 * queue always request a reschedule every tick. 1448 */ 1449 need_lwkt_resched(); 1450 } 1451 } 1452 1453 /* 1454 * Migrate the current thread to the specified cpu. 1455 * 1456 * This is accomplished by descheduling ourselves from the current cpu 1457 * and setting td_migrate_gd. The lwkt_switch() code will detect that the 1458 * 'old' thread wants to migrate after it has been completely switched out 1459 * and will complete the migration. 1460 * 1461 * TDF_MIGRATING prevents scheduling races while the thread is being migrated. 1462 * 1463 * We must be sure to release our current process designation (if a user 1464 * process) before clearing out any tsleepq we are on because the release 1465 * code may re-add us. 1466 * 1467 * We must be sure to remove ourselves from the current cpu's tsleepq 1468 * before potentially moving to another queue. The thread can be on 1469 * a tsleepq due to a left-over tsleep_interlock(). 1470 */ 1471 1472 void 1473 lwkt_setcpu_self(globaldata_t rgd) 1474 { 1475 thread_t td = curthread; 1476 1477 if (td->td_gd != rgd) { 1478 crit_enter_quick(td); 1479 1480 if (td->td_release) 1481 td->td_release(td); 1482 if (td->td_flags & TDF_TSLEEPQ) 1483 tsleep_remove(td); 1484 1485 /* 1486 * Set TDF_MIGRATING to prevent a spurious reschedule while we are 1487 * trying to deschedule ourselves and switch away, then deschedule 1488 * ourself, remove us from tdallq, and set td_migrate_gd. Finally, 1489 * call lwkt_switch() to complete the operation. 1490 */ 1491 td->td_flags |= TDF_MIGRATING; 1492 lwkt_deschedule_self(td); 1493 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1494 td->td_migrate_gd = rgd; 1495 lwkt_switch(); 1496 1497 /* 1498 * We are now on the target cpu 1499 */ 1500 KKASSERT(rgd == mycpu); 1501 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); 1502 crit_exit_quick(td); 1503 } 1504 } 1505 1506 void 1507 lwkt_migratecpu(int cpuid) 1508 { 1509 globaldata_t rgd; 1510 1511 rgd = globaldata_find(cpuid); 1512 lwkt_setcpu_self(rgd); 1513 } 1514 1515 /* 1516 * Remote IPI for cpu migration (called while in a critical section so we 1517 * do not have to enter another one). 1518 * 1519 * The thread (td) has already been completely descheduled from the 1520 * originating cpu and we can simply assert the case. The thread is 1521 * assigned to the new cpu and enqueued. 1522 * 1523 * The thread will re-add itself to tdallq when it resumes execution. 1524 */ 1525 static void 1526 lwkt_setcpu_remote(void *arg) 1527 { 1528 thread_t td = arg; 1529 globaldata_t gd = mycpu; 1530 1531 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0); 1532 td->td_gd = gd; 1533 cpu_mfence(); 1534 td->td_flags &= ~TDF_MIGRATING; 1535 KKASSERT(td->td_migrate_gd == NULL); 1536 KKASSERT(td->td_lwp == NULL || 1537 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 1538 _lwkt_enqueue(td); 1539 } 1540 1541 struct lwp * 1542 lwkt_preempted_proc(void) 1543 { 1544 thread_t td = curthread; 1545 while (td->td_preempted) 1546 td = td->td_preempted; 1547 return(td->td_lwp); 1548 } 1549 1550 /* 1551 * Create a kernel process/thread/whatever. It shares it's address space 1552 * with proc0 - ie: kernel only. 1553 * 1554 * If the cpu is not specified one will be selected. In the future 1555 * specifying a cpu of -1 will enable kernel thread migration between 1556 * cpus. 1557 */ 1558 int 1559 lwkt_create(void (*func)(void *), void *arg, struct thread **tdp, 1560 thread_t template, int tdflags, int cpu, const char *fmt, ...) 1561 { 1562 thread_t td; 1563 __va_list ap; 1564 1565 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu, 1566 tdflags); 1567 if (tdp) 1568 *tdp = td; 1569 cpu_set_thread_handler(td, lwkt_exit, func, arg); 1570 1571 /* 1572 * Set up arg0 for 'ps' etc 1573 */ 1574 __va_start(ap, fmt); 1575 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap); 1576 __va_end(ap); 1577 1578 /* 1579 * Schedule the thread to run 1580 */ 1581 if (td->td_flags & TDF_NOSTART) 1582 td->td_flags &= ~TDF_NOSTART; 1583 else 1584 lwkt_schedule(td); 1585 return 0; 1586 } 1587 1588 /* 1589 * Destroy an LWKT thread. Warning! This function is not called when 1590 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and 1591 * uses a different reaping mechanism. 1592 */ 1593 void 1594 lwkt_exit(void) 1595 { 1596 thread_t td = curthread; 1597 thread_t std; 1598 globaldata_t gd; 1599 1600 /* 1601 * Do any cleanup that might block here 1602 */ 1603 if (td->td_flags & TDF_VERBOSE) 1604 kprintf("kthread %p %s has exited\n", td, td->td_comm); 1605 biosched_done(td); 1606 dsched_exit_thread(td); 1607 1608 /* 1609 * Get us into a critical section to interlock gd_freetd and loop 1610 * until we can get it freed. 1611 * 1612 * We have to cache the current td in gd_freetd because objcache_put()ing 1613 * it would rip it out from under us while our thread is still active. 1614 * 1615 * We are the current thread so of course our own TDF_RUNNING bit will 1616 * be set, so unlike the lwp reap code we don't wait for it to clear. 1617 */ 1618 gd = mycpu; 1619 crit_enter_quick(td); 1620 for (;;) { 1621 if (td->td_refs) { 1622 tsleep(td, 0, "tdreap", 1); 1623 continue; 1624 } 1625 if ((std = gd->gd_freetd) != NULL) { 1626 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0); 1627 gd->gd_freetd = NULL; 1628 objcache_put(thread_cache, std); 1629 continue; 1630 } 1631 break; 1632 } 1633 1634 /* 1635 * Remove thread resources from kernel lists and deschedule us for 1636 * the last time. We cannot block after this point or we may end 1637 * up with a stale td on the tsleepq. 1638 * 1639 * None of this may block, the critical section is the only thing 1640 * protecting tdallq and the only thing preventing new lwkt_hold() 1641 * thread refs now. 1642 */ 1643 if (td->td_flags & TDF_TSLEEPQ) 1644 tsleep_remove(td); 1645 lwkt_deschedule_self(td); 1646 lwkt_remove_tdallq(td); 1647 KKASSERT(td->td_refs == 0); 1648 1649 /* 1650 * Final cleanup 1651 */ 1652 KKASSERT(gd->gd_freetd == NULL); 1653 if (td->td_flags & TDF_ALLOCATED_THREAD) 1654 gd->gd_freetd = td; 1655 cpu_thread_exit(); 1656 } 1657 1658 void 1659 lwkt_remove_tdallq(thread_t td) 1660 { 1661 KKASSERT(td->td_gd == mycpu); 1662 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1663 } 1664 1665 /* 1666 * Code reduction and branch prediction improvements. Call/return 1667 * overhead on modern cpus often degenerates into 0 cycles due to 1668 * the cpu's branch prediction hardware and return pc cache. We 1669 * can take advantage of this by not inlining medium-complexity 1670 * functions and we can also reduce the branch prediction impact 1671 * by collapsing perfectly predictable branches into a single 1672 * procedure instead of duplicating it. 1673 * 1674 * Is any of this noticeable? Probably not, so I'll take the 1675 * smaller code size. 1676 */ 1677 void 1678 crit_exit_wrapper(__DEBUG_CRIT_ARG__) 1679 { 1680 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__); 1681 } 1682 1683 void 1684 crit_panic(void) 1685 { 1686 thread_t td = curthread; 1687 int lcrit = td->td_critcount; 1688 1689 td->td_critcount = 0; 1690 panic("td_critcount is/would-go negative! %p %d", td, lcrit); 1691 /* NOT REACHED */ 1692 } 1693 1694 /* 1695 * Called from debugger/panic on cpus which have been stopped. We must still 1696 * process the IPIQ while stopped. 1697 * 1698 * If we are dumping also try to process any pending interrupts. This may 1699 * or may not work depending on the state of the cpu at the point it was 1700 * stopped. 1701 */ 1702 void 1703 lwkt_smp_stopped(void) 1704 { 1705 globaldata_t gd = mycpu; 1706 1707 if (dumping) { 1708 lwkt_process_ipiq(); 1709 --gd->gd_intr_nesting_level; 1710 splz(); 1711 ++gd->gd_intr_nesting_level; 1712 } else { 1713 lwkt_process_ipiq(); 1714 } 1715 cpu_smp_stopped(); 1716 } 1717