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