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