1 /* 2 * Copyright (c) 2003-2011 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@backplane.com> 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * 3. Neither the name of The DragonFly Project nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific, prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 */ 34 35 /* 36 * Each cpu in a system has its own self-contained light weight kernel 37 * thread scheduler, which means that generally speaking we only need 38 * to use a critical section to avoid problems. Foreign thread 39 * scheduling is queued via (async) IPIs. 40 */ 41 42 #include <sys/param.h> 43 #include <sys/systm.h> 44 #include <sys/kernel.h> 45 #include <sys/proc.h> 46 #include <sys/rtprio.h> 47 #include <sys/kinfo.h> 48 #include <sys/queue.h> 49 #include <sys/sysctl.h> 50 #include <sys/kthread.h> 51 #include <machine/cpu.h> 52 #include <sys/lock.h> 53 #include <sys/spinlock.h> 54 #include <sys/ktr.h> 55 56 #include <sys/thread2.h> 57 #include <sys/spinlock2.h> 58 #include <sys/mplock2.h> 59 60 #include <sys/dsched.h> 61 62 #include <vm/vm.h> 63 #include <vm/vm_param.h> 64 #include <vm/vm_kern.h> 65 #include <vm/vm_object.h> 66 #include <vm/vm_page.h> 67 #include <vm/vm_map.h> 68 #include <vm/vm_pager.h> 69 #include <vm/vm_extern.h> 70 71 #include <machine/stdarg.h> 72 #include <machine/smp.h> 73 74 #ifdef _KERNEL_VIRTUAL 75 #include <pthread.h> 76 #endif 77 78 #if !defined(KTR_CTXSW) 79 #define KTR_CTXSW KTR_ALL 80 #endif 81 KTR_INFO_MASTER(ctxsw); 82 KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p", int cpu, struct thread *td); 83 KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p", int cpu, struct thread *td); 84 KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s", struct thread *td, char *comm); 85 KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", struct thread *td); 86 87 static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads"); 88 89 #ifdef INVARIANTS 90 static int panic_on_cscount = 0; 91 #endif 92 static __int64_t switch_count = 0; 93 static __int64_t preempt_hit = 0; 94 static __int64_t preempt_miss = 0; 95 static __int64_t preempt_weird = 0; 96 static int lwkt_use_spin_port; 97 static struct objcache *thread_cache; 98 int cpu_mwait_spin = 0; 99 100 static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame); 101 static void lwkt_setcpu_remote(void *arg); 102 103 extern void cpu_heavy_restore(void); 104 extern void cpu_lwkt_restore(void); 105 extern void cpu_kthread_restore(void); 106 extern void cpu_idle_restore(void); 107 108 /* 109 * We can make all thread ports use the spin backend instead of the thread 110 * backend. This should only be set to debug the spin backend. 111 */ 112 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port); 113 114 #ifdef INVARIANTS 115 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, 116 "Panic if attempting to switch lwkt's while mastering cpusync"); 117 #endif 118 SYSCTL_INT(_hw, OID_AUTO, cpu_mwait_spin, CTLFLAG_RW, &cpu_mwait_spin, 0, 119 "monitor/mwait target state"); 120 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, 121 "Number of switched threads"); 122 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, 123 "Successful preemption events"); 124 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, 125 "Failed preemption events"); 126 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, 127 "Number of preempted threads."); 128 static int fairq_enable = 0; 129 SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW, 130 &fairq_enable, 0, "Turn on fairq priority accumulators"); 131 static int fairq_bypass = -1; 132 SYSCTL_INT(_lwkt, OID_AUTO, fairq_bypass, CTLFLAG_RW, 133 &fairq_bypass, 0, "Allow fairq to bypass td on token failure"); 134 extern int lwkt_sched_debug; 135 int lwkt_sched_debug = 0; 136 SYSCTL_INT(_lwkt, OID_AUTO, sched_debug, CTLFLAG_RW, 137 &lwkt_sched_debug, 0, "Scheduler debug"); 138 static int lwkt_spin_loops = 10; 139 SYSCTL_INT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW, 140 &lwkt_spin_loops, 0, "Scheduler spin loops until sorted decon"); 141 static int lwkt_spin_reseq = 0; 142 SYSCTL_INT(_lwkt, OID_AUTO, spin_reseq, CTLFLAG_RW, 143 &lwkt_spin_reseq, 0, "Scheduler resequencer enable"); 144 static int lwkt_spin_monitor = 0; 145 SYSCTL_INT(_lwkt, OID_AUTO, spin_monitor, CTLFLAG_RW, 146 &lwkt_spin_monitor, 0, "Scheduler uses monitor/mwait"); 147 static int lwkt_spin_fatal = 0; /* disabled */ 148 SYSCTL_INT(_lwkt, OID_AUTO, spin_fatal, CTLFLAG_RW, 149 &lwkt_spin_fatal, 0, "LWKT scheduler spin loops till fatal panic"); 150 static int preempt_enable = 1; 151 SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW, 152 &preempt_enable, 0, "Enable preemption"); 153 static int lwkt_cache_threads = 0; 154 SYSCTL_INT(_lwkt, OID_AUTO, cache_threads, CTLFLAG_RD, 155 &lwkt_cache_threads, 0, "thread+kstack cache"); 156 157 #ifndef _KERNEL_VIRTUAL 158 static __cachealign int lwkt_cseq_rindex; 159 static __cachealign int lwkt_cseq_windex; 160 #endif 161 162 /* 163 * These helper procedures handle the runq, they can only be called from 164 * within a critical section. 165 * 166 * WARNING! Prior to SMP being brought up it is possible to enqueue and 167 * dequeue threads belonging to other cpus, so be sure to use td->td_gd 168 * instead of 'mycpu' when referencing the globaldata structure. Once 169 * SMP live enqueuing and dequeueing only occurs on the current cpu. 170 */ 171 static __inline 172 void 173 _lwkt_dequeue(thread_t td) 174 { 175 if (td->td_flags & TDF_RUNQ) { 176 struct globaldata *gd = td->td_gd; 177 178 td->td_flags &= ~TDF_RUNQ; 179 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq); 180 --gd->gd_tdrunqcount; 181 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL) 182 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING); 183 } 184 } 185 186 /* 187 * Priority enqueue. 188 * 189 * There are a limited number of lwkt threads runnable since user 190 * processes only schedule one at a time per cpu. However, there can 191 * be many user processes in kernel mode exiting from a tsleep() which 192 * become runnable. 193 * 194 * NOTE: lwkt_schedulerclock() will force a round-robin based on td_pri and 195 * will ignore user priority. This is to ensure that user threads in 196 * kernel mode get cpu at some point regardless of what the user 197 * scheduler thinks. 198 */ 199 static __inline 200 void 201 _lwkt_enqueue(thread_t td) 202 { 203 thread_t xtd; 204 205 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) { 206 struct globaldata *gd = td->td_gd; 207 208 td->td_flags |= TDF_RUNQ; 209 xtd = TAILQ_FIRST(&gd->gd_tdrunq); 210 if (xtd == NULL) { 211 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 212 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING); 213 } else { 214 /* 215 * NOTE: td_upri - higher numbers more desireable, same sense 216 * as td_pri (typically reversed from lwp_upri). 217 * 218 * In the equal priority case we want the best selection 219 * at the beginning so the less desireable selections know 220 * that they have to setrunqueue/go-to-another-cpu, even 221 * though it means switching back to the 'best' selection. 222 * This also avoids degenerate situations when many threads 223 * are runnable or waking up at the same time. 224 * 225 * If upri matches exactly place at end/round-robin. 226 */ 227 while (xtd && 228 (xtd->td_pri >= td->td_pri || 229 (xtd->td_pri == td->td_pri && 230 xtd->td_upri >= td->td_upri))) { 231 xtd = TAILQ_NEXT(xtd, td_threadq); 232 } 233 if (xtd) 234 TAILQ_INSERT_BEFORE(xtd, td, td_threadq); 235 else 236 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 237 } 238 ++gd->gd_tdrunqcount; 239 240 /* 241 * Request a LWKT reschedule if we are now at the head of the queue. 242 */ 243 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) 244 need_lwkt_resched(); 245 } 246 } 247 248 static __boolean_t 249 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags) 250 { 251 struct thread *td = (struct thread *)obj; 252 253 td->td_kstack = NULL; 254 td->td_kstack_size = 0; 255 td->td_flags = TDF_ALLOCATED_THREAD; 256 td->td_mpflags = 0; 257 return (1); 258 } 259 260 static void 261 _lwkt_thread_dtor(void *obj, void *privdata) 262 { 263 struct thread *td = (struct thread *)obj; 264 265 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD, 266 ("_lwkt_thread_dtor: not allocated from objcache")); 267 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack && 268 td->td_kstack_size > 0, 269 ("_lwkt_thread_dtor: corrupted stack")); 270 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size); 271 td->td_kstack = NULL; 272 td->td_flags = 0; 273 } 274 275 /* 276 * Initialize the lwkt s/system. 277 * 278 * Nominally cache up to 32 thread + kstack structures. Cache more on 279 * systems with a lot of cpu cores. 280 */ 281 void 282 lwkt_init(void) 283 { 284 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads); 285 if (lwkt_cache_threads == 0) { 286 lwkt_cache_threads = ncpus * 4; 287 if (lwkt_cache_threads < 32) 288 lwkt_cache_threads = 32; 289 } 290 thread_cache = objcache_create_mbacked( 291 M_THREAD, sizeof(struct thread), 292 0, lwkt_cache_threads, 293 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL); 294 } 295 296 /* 297 * Schedule a thread to run. As the current thread we can always safely 298 * schedule ourselves, and a shortcut procedure is provided for that 299 * function. 300 * 301 * (non-blocking, self contained on a per cpu basis) 302 */ 303 void 304 lwkt_schedule_self(thread_t td) 305 { 306 KKASSERT((td->td_flags & TDF_MIGRATING) == 0); 307 crit_enter_quick(td); 308 KASSERT(td != &td->td_gd->gd_idlethread, 309 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!")); 310 KKASSERT(td->td_lwp == NULL || 311 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 312 _lwkt_enqueue(td); 313 crit_exit_quick(td); 314 } 315 316 /* 317 * Deschedule a thread. 318 * 319 * (non-blocking, self contained on a per cpu basis) 320 */ 321 void 322 lwkt_deschedule_self(thread_t td) 323 { 324 crit_enter_quick(td); 325 _lwkt_dequeue(td); 326 crit_exit_quick(td); 327 } 328 329 /* 330 * LWKTs operate on a per-cpu basis 331 * 332 * WARNING! Called from early boot, 'mycpu' may not work yet. 333 */ 334 void 335 lwkt_gdinit(struct globaldata *gd) 336 { 337 TAILQ_INIT(&gd->gd_tdrunq); 338 TAILQ_INIT(&gd->gd_tdallq); 339 } 340 341 /* 342 * Create a new thread. The thread must be associated with a process context 343 * or LWKT start address before it can be scheduled. If the target cpu is 344 * -1 the thread will be created on the current cpu. 345 * 346 * If you intend to create a thread without a process context this function 347 * does everything except load the startup and switcher function. 348 */ 349 thread_t 350 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags) 351 { 352 static int cpu_rotator; 353 globaldata_t gd = mycpu; 354 void *stack; 355 356 /* 357 * If static thread storage is not supplied allocate a thread. Reuse 358 * a cached free thread if possible. gd_freetd is used to keep an exiting 359 * thread intact through the exit. 360 */ 361 if (td == NULL) { 362 crit_enter_gd(gd); 363 if ((td = gd->gd_freetd) != NULL) { 364 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK| 365 TDF_RUNQ)) == 0); 366 gd->gd_freetd = NULL; 367 } else { 368 td = objcache_get(thread_cache, M_WAITOK); 369 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK| 370 TDF_RUNQ)) == 0); 371 } 372 crit_exit_gd(gd); 373 KASSERT((td->td_flags & 374 (TDF_ALLOCATED_THREAD|TDF_RUNNING|TDF_PREEMPT_LOCK)) == 375 TDF_ALLOCATED_THREAD, 376 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags)); 377 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK); 378 } 379 380 /* 381 * Try to reuse cached stack. 382 */ 383 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) { 384 if (flags & TDF_ALLOCATED_STACK) { 385 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size); 386 stack = NULL; 387 } 388 } 389 if (stack == NULL) { 390 stack = (void *)kmem_alloc_stack(&kernel_map, stksize); 391 flags |= TDF_ALLOCATED_STACK; 392 } 393 if (cpu < 0) { 394 cpu = ++cpu_rotator; 395 cpu_ccfence(); 396 cpu %= ncpus; 397 } 398 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu)); 399 return(td); 400 } 401 402 /* 403 * Initialize a preexisting thread structure. This function is used by 404 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread. 405 * 406 * All threads start out in a critical section at a priority of 407 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as 408 * appropriate. This function may send an IPI message when the 409 * requested cpu is not the current cpu and consequently gd_tdallq may 410 * not be initialized synchronously from the point of view of the originating 411 * cpu. 412 * 413 * NOTE! we have to be careful in regards to creating threads for other cpus 414 * if SMP has not yet been activated. 415 */ 416 static void 417 lwkt_init_thread_remote(void *arg) 418 { 419 thread_t td = arg; 420 421 /* 422 * Protected by critical section held by IPI dispatch 423 */ 424 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq); 425 } 426 427 /* 428 * lwkt core thread structural initialization. 429 * 430 * NOTE: All threads are initialized as mpsafe threads. 431 */ 432 void 433 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags, 434 struct globaldata *gd) 435 { 436 globaldata_t mygd = mycpu; 437 438 bzero(td, sizeof(struct thread)); 439 td->td_kstack = stack; 440 td->td_kstack_size = stksize; 441 td->td_flags = flags; 442 td->td_mpflags = 0; 443 td->td_type = TD_TYPE_GENERIC; 444 td->td_gd = gd; 445 td->td_pri = TDPRI_KERN_DAEMON; 446 td->td_critcount = 1; 447 td->td_toks_have = NULL; 448 td->td_toks_stop = &td->td_toks_base; 449 if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT)) { 450 lwkt_initport_spin(&td->td_msgport, td, 451 (flags & TDF_FIXEDCPU) ? TRUE : FALSE); 452 } else { 453 lwkt_initport_thread(&td->td_msgport, td); 454 } 455 pmap_init_thread(td); 456 /* 457 * Normally initializing a thread for a remote cpu requires sending an 458 * IPI. However, the idlethread is setup before the other cpus are 459 * activated so we have to treat it as a special case. XXX manipulation 460 * of gd_tdallq requires the BGL. 461 */ 462 if (gd == mygd || td == &gd->gd_idlethread) { 463 crit_enter_gd(mygd); 464 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); 465 crit_exit_gd(mygd); 466 } else { 467 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td); 468 } 469 dsched_new_thread(td); 470 } 471 472 void 473 lwkt_set_comm(thread_t td, const char *ctl, ...) 474 { 475 __va_list va; 476 477 __va_start(va, ctl); 478 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va); 479 __va_end(va); 480 KTR_LOG(ctxsw_newtd, td, td->td_comm); 481 } 482 483 /* 484 * Prevent the thread from getting destroyed. Note that unlike PHOLD/PRELE 485 * this does not prevent the thread from migrating to another cpu so the 486 * gd_tdallq state is not protected by this. 487 */ 488 void 489 lwkt_hold(thread_t td) 490 { 491 atomic_add_int(&td->td_refs, 1); 492 } 493 494 void 495 lwkt_rele(thread_t td) 496 { 497 KKASSERT(td->td_refs > 0); 498 atomic_add_int(&td->td_refs, -1); 499 } 500 501 void 502 lwkt_free_thread(thread_t td) 503 { 504 KKASSERT(td->td_refs == 0); 505 KKASSERT((td->td_flags & (TDF_RUNNING | TDF_PREEMPT_LOCK | 506 TDF_RUNQ | TDF_TSLEEPQ)) == 0); 507 if (td->td_flags & TDF_ALLOCATED_THREAD) { 508 objcache_put(thread_cache, td); 509 } else if (td->td_flags & TDF_ALLOCATED_STACK) { 510 /* client-allocated struct with internally allocated stack */ 511 KASSERT(td->td_kstack && td->td_kstack_size > 0, 512 ("lwkt_free_thread: corrupted stack")); 513 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size); 514 td->td_kstack = NULL; 515 td->td_kstack_size = 0; 516 } 517 518 KTR_LOG(ctxsw_deadtd, td); 519 } 520 521 522 /* 523 * Switch to the next runnable lwkt. If no LWKTs are runnable then 524 * switch to the idlethread. Switching must occur within a critical 525 * section to avoid races with the scheduling queue. 526 * 527 * We always have full control over our cpu's run queue. Other cpus 528 * that wish to manipulate our queue must use the cpu_*msg() calls to 529 * talk to our cpu, so a critical section is all that is needed and 530 * the result is very, very fast thread switching. 531 * 532 * The LWKT scheduler uses a fixed priority model and round-robins at 533 * each priority level. User process scheduling is a totally 534 * different beast and LWKT priorities should not be confused with 535 * user process priorities. 536 * 537 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch() 538 * is not called by the current thread in the preemption case, only when 539 * the preempting thread blocks (in order to return to the original thread). 540 * 541 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread 542 * migration and tsleep deschedule the current lwkt thread and call 543 * lwkt_switch(). In particular, the target cpu of the migration fully 544 * expects the thread to become non-runnable and can deadlock against 545 * cpusync operations if we run any IPIs prior to switching the thread out. 546 * 547 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF 548 * THE CURRENT THREAD HAS BEEN DESCHEDULED! 549 */ 550 void 551 lwkt_switch(void) 552 { 553 globaldata_t gd = mycpu; 554 thread_t td = gd->gd_curthread; 555 thread_t ntd; 556 int spinning = 0; 557 558 KKASSERT(gd->gd_processing_ipiq == 0); 559 KKASSERT(td->td_flags & TDF_RUNNING); 560 561 /* 562 * Switching from within a 'fast' (non thread switched) interrupt or IPI 563 * is illegal. However, we may have to do it anyway if we hit a fatal 564 * kernel trap or we have paniced. 565 * 566 * If this case occurs save and restore the interrupt nesting level. 567 */ 568 if (gd->gd_intr_nesting_level) { 569 int savegdnest; 570 int savegdtrap; 571 572 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) { 573 panic("lwkt_switch: Attempt to switch from a " 574 "fast interrupt, ipi, or hard code section, " 575 "td %p\n", 576 td); 577 } else { 578 savegdnest = gd->gd_intr_nesting_level; 579 savegdtrap = gd->gd_trap_nesting_level; 580 gd->gd_intr_nesting_level = 0; 581 gd->gd_trap_nesting_level = 0; 582 if ((td->td_flags & TDF_PANICWARN) == 0) { 583 td->td_flags |= TDF_PANICWARN; 584 kprintf("Warning: thread switch from interrupt, IPI, " 585 "or hard code section.\n" 586 "thread %p (%s)\n", td, td->td_comm); 587 print_backtrace(-1); 588 } 589 lwkt_switch(); 590 gd->gd_intr_nesting_level = savegdnest; 591 gd->gd_trap_nesting_level = savegdtrap; 592 return; 593 } 594 } 595 596 /* 597 * Release our current user process designation if we are blocking 598 * or if a user reschedule was requested. 599 * 600 * NOTE: This function is NOT called if we are switching into or 601 * returning from a preemption. 602 * 603 * NOTE: Releasing our current user process designation may cause 604 * it to be assigned to another thread, which in turn will 605 * cause us to block in the usched acquire code when we attempt 606 * to return to userland. 607 * 608 * NOTE: On SMP systems this can be very nasty when heavy token 609 * contention is present so we want to be careful not to 610 * release the designation gratuitously. 611 */ 612 if (td->td_release && 613 (user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) { 614 td->td_release(td); 615 } 616 617 /* 618 * Release all tokens 619 */ 620 crit_enter_gd(gd); 621 if (TD_TOKS_HELD(td)) 622 lwkt_relalltokens(td); 623 624 /* 625 * We had better not be holding any spin locks, but don't get into an 626 * endless panic loop. 627 */ 628 KASSERT(gd->gd_spinlocks == 0 || panicstr != NULL, 629 ("lwkt_switch: still holding %d exclusive spinlocks!", 630 gd->gd_spinlocks)); 631 632 633 #ifdef INVARIANTS 634 if (td->td_cscount) { 635 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n", 636 td); 637 if (panic_on_cscount) 638 panic("switching while mastering cpusync"); 639 } 640 #endif 641 642 /* 643 * If we had preempted another thread on this cpu, resume the preempted 644 * thread. This occurs transparently, whether the preempted thread 645 * was scheduled or not (it may have been preempted after descheduling 646 * itself). 647 * 648 * We have to setup the MP lock for the original thread after backing 649 * out the adjustment that was made to curthread when the original 650 * was preempted. 651 */ 652 if ((ntd = td->td_preempted) != NULL) { 653 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK); 654 ntd->td_flags |= TDF_PREEMPT_DONE; 655 656 /* 657 * The interrupt may have woken a thread up, we need to properly 658 * set the reschedule flag if the originally interrupted thread is 659 * at a lower priority. 660 * 661 * The interrupt may not have descheduled. 662 */ 663 if (TAILQ_FIRST(&gd->gd_tdrunq) != ntd) 664 need_lwkt_resched(); 665 goto havethread_preempted; 666 } 667 668 /* 669 * If we cannot obtain ownership of the tokens we cannot immediately 670 * schedule the target thread. 671 * 672 * Reminder: Again, we cannot afford to run any IPIs in this path if 673 * the current thread has been descheduled. 674 */ 675 for (;;) { 676 clear_lwkt_resched(); 677 678 /* 679 * Hotpath - pull the head of the run queue and attempt to schedule 680 * it. 681 */ 682 ntd = TAILQ_FIRST(&gd->gd_tdrunq); 683 684 if (ntd == NULL) { 685 /* 686 * Runq is empty, switch to idle to allow it to halt. 687 */ 688 ntd = &gd->gd_idlethread; 689 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) 690 ASSERT_NO_TOKENS_HELD(ntd); 691 cpu_time.cp_msg[0] = 0; 692 cpu_time.cp_stallpc = 0; 693 goto haveidle; 694 } 695 696 /* 697 * Hotpath - schedule ntd. 698 * 699 * NOTE: For UP there is no mplock and lwkt_getalltokens() 700 * always succeeds. 701 */ 702 if (TD_TOKS_NOT_HELD(ntd) || 703 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops))) 704 { 705 goto havethread; 706 } 707 708 /* 709 * Coldpath (SMP only since tokens always succeed on UP) 710 * 711 * We had some contention on the thread we wanted to schedule. 712 * What we do now is try to find a thread that we can schedule 713 * in its stead. 714 * 715 * The coldpath scan does NOT rearrange threads in the run list. 716 * The lwkt_schedulerclock() will assert need_lwkt_resched() on 717 * the next tick whenever the current head is not the current thread. 718 */ 719 #ifdef INVARIANTS 720 ++ntd->td_contended; 721 #endif 722 ++gd->gd_cnt.v_lock_colls; 723 724 if (fairq_bypass > 0) 725 goto skip; 726 727 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) { 728 #ifndef NO_LWKT_SPLIT_USERPRI 729 /* 730 * Never schedule threads returning to userland or the 731 * user thread scheduler helper thread when higher priority 732 * threads are present. The runq is sorted by priority 733 * so we can give up traversing it when we find the first 734 * low priority thread. 735 */ 736 if (ntd->td_pri < TDPRI_KERN_LPSCHED) { 737 ntd = NULL; 738 break; 739 } 740 #endif 741 742 /* 743 * Try this one. 744 */ 745 if (TD_TOKS_NOT_HELD(ntd) || 746 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops))) { 747 goto havethread; 748 } 749 #ifdef INVARIANTS 750 ++ntd->td_contended; 751 #endif 752 ++gd->gd_cnt.v_lock_colls; 753 } 754 755 skip: 756 /* 757 * We exhausted the run list, meaning that all runnable threads 758 * are contested. 759 */ 760 cpu_pause(); 761 #ifdef _KERNEL_VIRTUAL 762 pthread_yield(); 763 #endif 764 ntd = &gd->gd_idlethread; 765 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) 766 ASSERT_NO_TOKENS_HELD(ntd); 767 /* contention case, do not clear contention mask */ 768 769 /* 770 * We are going to have to retry but if the current thread is not 771 * on the runq we instead switch through the idle thread to get away 772 * from the current thread. We have to flag for lwkt reschedule 773 * to prevent the idle thread from halting. 774 * 775 * NOTE: A non-zero spinning is passed to lwkt_getalltokens() to 776 * instruct it to deal with the potential for deadlocks by 777 * ordering the tokens by address. 778 */ 779 if ((td->td_flags & TDF_RUNQ) == 0) { 780 need_lwkt_resched(); /* prevent hlt */ 781 goto haveidle; 782 } 783 #if defined(INVARIANTS) && defined(__x86_64__) 784 if ((read_rflags() & PSL_I) == 0) { 785 cpu_enable_intr(); 786 panic("lwkt_switch() called with interrupts disabled"); 787 } 788 #endif 789 790 /* 791 * Number iterations so far. After a certain point we switch to 792 * a sorted-address/monitor/mwait version of lwkt_getalltokens() 793 */ 794 if (spinning < 0x7FFFFFFF) 795 ++spinning; 796 797 #ifndef _KERNEL_VIRTUAL 798 /* 799 * lwkt_getalltokens() failed in sorted token mode, we can use 800 * monitor/mwait in this case. 801 */ 802 if (spinning >= lwkt_spin_loops && 803 (cpu_mi_feature & CPU_MI_MONITOR) && 804 lwkt_spin_monitor) 805 { 806 cpu_mmw_pause_int(&gd->gd_reqflags, 807 (gd->gd_reqflags | RQF_SPINNING) & 808 ~RQF_IDLECHECK_WK_MASK, 809 cpu_mwait_spin, 0); 810 } 811 #endif 812 813 /* 814 * We already checked that td is still scheduled so this should be 815 * safe. 816 */ 817 splz_check(); 818 819 #ifndef _KERNEL_VIRTUAL 820 /* 821 * This experimental resequencer is used as a fall-back to reduce 822 * hw cache line contention by placing each core's scheduler into a 823 * time-domain-multplexed slot. 824 * 825 * The resequencer is disabled by default. It's functionality has 826 * largely been superceeded by the token algorithm which limits races 827 * to a subset of cores. 828 * 829 * The resequencer algorithm tends to break down when more than 830 * 20 cores are contending. What appears to happen is that new 831 * tokens can be obtained out of address-sorted order by new cores 832 * while existing cores languish in long delays between retries and 833 * wind up being starved-out of the token acquisition. 834 */ 835 if (lwkt_spin_reseq && spinning >= lwkt_spin_reseq) { 836 int cseq = atomic_fetchadd_int(&lwkt_cseq_windex, 1); 837 int oseq; 838 839 while ((oseq = lwkt_cseq_rindex) != cseq) { 840 cpu_ccfence(); 841 #if 1 842 if (cpu_mi_feature & CPU_MI_MONITOR) { 843 cpu_mmw_pause_int(&lwkt_cseq_rindex, oseq, 844 cpu_mwait_spin, 0); 845 } else { 846 #endif 847 cpu_pause(); 848 cpu_lfence(); 849 #if 1 850 } 851 #endif 852 } 853 DELAY(1); 854 atomic_add_int(&lwkt_cseq_rindex, 1); 855 } 856 #endif 857 /* highest level for(;;) loop */ 858 } 859 860 havethread: 861 /* 862 * Clear gd_idle_repeat when doing a normal switch to a non-idle 863 * thread. 864 */ 865 ntd->td_wmesg = NULL; 866 ++gd->gd_cnt.v_swtch; 867 gd->gd_idle_repeat = 0; 868 869 havethread_preempted: 870 /* 871 * If the new target does not need the MP lock and we are holding it, 872 * release the MP lock. If the new target requires the MP lock we have 873 * already acquired it for the target. 874 */ 875 ; 876 haveidle: 877 KASSERT(ntd->td_critcount, 878 ("priority problem in lwkt_switch %d %d", 879 td->td_critcount, ntd->td_critcount)); 880 881 if (td != ntd) { 882 /* 883 * Execute the actual thread switch operation. This function 884 * returns to the current thread and returns the previous thread 885 * (which may be different from the thread we switched to). 886 * 887 * We are responsible for marking ntd as TDF_RUNNING. 888 */ 889 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0); 890 ++switch_count; 891 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd); 892 ntd->td_flags |= TDF_RUNNING; 893 lwkt_switch_return(td->td_switch(ntd)); 894 /* ntd invalid, td_switch() can return a different thread_t */ 895 } 896 897 /* 898 * catch-all. XXX is this strictly needed? 899 */ 900 splz_check(); 901 902 /* NOTE: current cpu may have changed after switch */ 903 crit_exit_quick(td); 904 } 905 906 /* 907 * Called by assembly in the td_switch (thread restore path) for thread 908 * bootstrap cases which do not 'return' to lwkt_switch(). 909 */ 910 void 911 lwkt_switch_return(thread_t otd) 912 { 913 globaldata_t rgd; 914 915 /* 916 * Check if otd was migrating. Now that we are on ntd we can finish 917 * up the migration. This is a bit messy but it is the only place 918 * where td is known to be fully descheduled. 919 * 920 * We can only activate the migration if otd was migrating but not 921 * held on the cpu due to a preemption chain. We still have to 922 * clear TDF_RUNNING on the old thread either way. 923 * 924 * We are responsible for clearing the previously running thread's 925 * TDF_RUNNING. 926 */ 927 if ((rgd = otd->td_migrate_gd) != NULL && 928 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) { 929 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) == 930 (TDF_MIGRATING | TDF_RUNNING)); 931 otd->td_migrate_gd = NULL; 932 otd->td_flags &= ~TDF_RUNNING; 933 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd); 934 } else { 935 otd->td_flags &= ~TDF_RUNNING; 936 } 937 938 /* 939 * Final exit validations (see lwp_wait()). Note that otd becomes 940 * invalid the *instant* we set TDF_MP_EXITSIG. 941 */ 942 while (otd->td_flags & TDF_EXITING) { 943 u_int mpflags; 944 945 mpflags = otd->td_mpflags; 946 cpu_ccfence(); 947 948 if (mpflags & TDF_MP_EXITWAIT) { 949 if (atomic_cmpset_int(&otd->td_mpflags, mpflags, 950 mpflags | TDF_MP_EXITSIG)) { 951 wakeup(otd); 952 break; 953 } 954 } else { 955 if (atomic_cmpset_int(&otd->td_mpflags, mpflags, 956 mpflags | TDF_MP_EXITSIG)) { 957 wakeup(otd); 958 break; 959 } 960 } 961 } 962 } 963 964 /* 965 * Request that the target thread preempt the current thread. Preemption 966 * can only occur if our only critical section is the one that we were called 967 * with, the relative priority of the target thread is higher, and the target 968 * thread holds no tokens. This also only works if we are not holding any 969 * spinlocks (obviously). 970 * 971 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically 972 * this is called via lwkt_schedule() through the td_preemptable callback. 973 * critcount is the managed critical priority that we should ignore in order 974 * to determine whether preemption is possible (aka usually just the crit 975 * priority of lwkt_schedule() itself). 976 * 977 * Preemption is typically limited to interrupt threads. 978 * 979 * Operation works in a fairly straight-forward manner. The normal 980 * scheduling code is bypassed and we switch directly to the target 981 * thread. When the target thread attempts to block or switch away 982 * code at the base of lwkt_switch() will switch directly back to our 983 * thread. Our thread is able to retain whatever tokens it holds and 984 * if the target needs one of them the target will switch back to us 985 * and reschedule itself normally. 986 */ 987 void 988 lwkt_preempt(thread_t ntd, int critcount) 989 { 990 struct globaldata *gd = mycpu; 991 thread_t xtd; 992 thread_t td; 993 int save_gd_intr_nesting_level; 994 995 /* 996 * The caller has put us in a critical section. We can only preempt 997 * if the caller of the caller was not in a critical section (basically 998 * a local interrupt), as determined by the 'critcount' parameter. We 999 * also can't preempt if the caller is holding any spinlocks (even if 1000 * he isn't in a critical section). This also handles the tokens test. 1001 * 1002 * YYY The target thread must be in a critical section (else it must 1003 * inherit our critical section? I dunno yet). 1004 */ 1005 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri)); 1006 1007 td = gd->gd_curthread; 1008 if (preempt_enable == 0) { 1009 ++preempt_miss; 1010 return; 1011 } 1012 if (ntd->td_pri <= td->td_pri) { 1013 ++preempt_miss; 1014 return; 1015 } 1016 if (td->td_critcount > critcount) { 1017 ++preempt_miss; 1018 return; 1019 } 1020 if (td->td_cscount) { 1021 ++preempt_miss; 1022 return; 1023 } 1024 if (ntd->td_gd != gd) { 1025 ++preempt_miss; 1026 return; 1027 } 1028 /* 1029 * We don't have to check spinlocks here as they will also bump 1030 * td_critcount. 1031 * 1032 * Do not try to preempt if the target thread is holding any tokens. 1033 * We could try to acquire the tokens but this case is so rare there 1034 * is no need to support it. 1035 */ 1036 KKASSERT(gd->gd_spinlocks == 0); 1037 1038 if (TD_TOKS_HELD(ntd)) { 1039 ++preempt_miss; 1040 return; 1041 } 1042 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) { 1043 ++preempt_weird; 1044 return; 1045 } 1046 if (ntd->td_preempted) { 1047 ++preempt_hit; 1048 return; 1049 } 1050 KKASSERT(gd->gd_processing_ipiq == 0); 1051 1052 /* 1053 * Since we are able to preempt the current thread, there is no need to 1054 * call need_lwkt_resched(). 1055 * 1056 * We must temporarily clear gd_intr_nesting_level around the switch 1057 * since switchouts from the target thread are allowed (they will just 1058 * return to our thread), and since the target thread has its own stack. 1059 * 1060 * A preemption must switch back to the original thread, assert the 1061 * case. 1062 */ 1063 ++preempt_hit; 1064 ntd->td_preempted = td; 1065 td->td_flags |= TDF_PREEMPT_LOCK; 1066 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd); 1067 save_gd_intr_nesting_level = gd->gd_intr_nesting_level; 1068 gd->gd_intr_nesting_level = 0; 1069 1070 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0); 1071 ntd->td_flags |= TDF_RUNNING; 1072 xtd = td->td_switch(ntd); 1073 KKASSERT(xtd == ntd); 1074 lwkt_switch_return(xtd); 1075 gd->gd_intr_nesting_level = save_gd_intr_nesting_level; 1076 1077 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE)); 1078 ntd->td_preempted = NULL; 1079 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE); 1080 } 1081 1082 /* 1083 * Conditionally call splz() if gd_reqflags indicates work is pending. 1084 * This will work inside a critical section but not inside a hard code 1085 * section. 1086 * 1087 * (self contained on a per cpu basis) 1088 */ 1089 void 1090 splz_check(void) 1091 { 1092 globaldata_t gd = mycpu; 1093 thread_t td = gd->gd_curthread; 1094 1095 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && 1096 gd->gd_intr_nesting_level == 0 && 1097 td->td_nest_count < 2) 1098 { 1099 splz(); 1100 } 1101 } 1102 1103 /* 1104 * This version is integrated into crit_exit, reqflags has already 1105 * been tested but td_critcount has not. 1106 * 1107 * We only want to execute the splz() on the 1->0 transition of 1108 * critcount and not in a hard code section or if too deeply nested. 1109 * 1110 * NOTE: gd->gd_spinlocks is implied to be 0 when td_critcount is 0. 1111 */ 1112 void 1113 lwkt_maybe_splz(thread_t td) 1114 { 1115 globaldata_t gd = td->td_gd; 1116 1117 if (td->td_critcount == 0 && 1118 gd->gd_intr_nesting_level == 0 && 1119 td->td_nest_count < 2) 1120 { 1121 splz(); 1122 } 1123 } 1124 1125 /* 1126 * Drivers which set up processing co-threads can call this function to 1127 * run the co-thread at a higher priority and to allow it to preempt 1128 * normal threads. 1129 */ 1130 void 1131 lwkt_set_interrupt_support_thread(void) 1132 { 1133 thread_t td = curthread; 1134 1135 lwkt_setpri_self(TDPRI_INT_SUPPORT); 1136 td->td_flags |= TDF_INTTHREAD; 1137 td->td_preemptable = lwkt_preempt; 1138 } 1139 1140 1141 /* 1142 * This function is used to negotiate a passive release of the current 1143 * process/lwp designation with the user scheduler, allowing the user 1144 * scheduler to schedule another user thread. The related kernel thread 1145 * (curthread) continues running in the released state. 1146 */ 1147 void 1148 lwkt_passive_release(struct thread *td) 1149 { 1150 struct lwp *lp = td->td_lwp; 1151 1152 #ifndef NO_LWKT_SPLIT_USERPRI 1153 td->td_release = NULL; 1154 lwkt_setpri_self(TDPRI_KERN_USER); 1155 #endif 1156 1157 lp->lwp_proc->p_usched->release_curproc(lp); 1158 } 1159 1160 1161 /* 1162 * This implements a LWKT yield, allowing a kernel thread to yield to other 1163 * kernel threads at the same or higher priority. This function can be 1164 * called in a tight loop and will typically only yield once per tick. 1165 * 1166 * Most kernel threads run at the same priority in order to allow equal 1167 * sharing. 1168 * 1169 * (self contained on a per cpu basis) 1170 */ 1171 void 1172 lwkt_yield(void) 1173 { 1174 globaldata_t gd = mycpu; 1175 thread_t td = gd->gd_curthread; 1176 1177 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1178 splz(); 1179 if (lwkt_resched_wanted()) { 1180 lwkt_schedule_self(curthread); 1181 lwkt_switch(); 1182 } 1183 } 1184 1185 /* 1186 * The quick version processes pending interrupts and higher-priority 1187 * LWKT threads but will not round-robin same-priority LWKT threads. 1188 * 1189 * When called while attempting to return to userland the only same-pri 1190 * threads are the ones which have already tried to become the current 1191 * user process. 1192 */ 1193 void 1194 lwkt_yield_quick(void) 1195 { 1196 globaldata_t gd = mycpu; 1197 thread_t td = gd->gd_curthread; 1198 1199 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1200 splz(); 1201 if (lwkt_resched_wanted()) { 1202 crit_enter(); 1203 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) { 1204 clear_lwkt_resched(); 1205 } else { 1206 lwkt_schedule_self(curthread); 1207 lwkt_switch(); 1208 } 1209 crit_exit(); 1210 } 1211 } 1212 1213 /* 1214 * This yield is designed for kernel threads with a user context. 1215 * 1216 * The kernel acting on behalf of the user is potentially cpu-bound, 1217 * this function will efficiently allow other threads to run and also 1218 * switch to other processes by releasing. 1219 * 1220 * The lwkt_user_yield() function is designed to have very low overhead 1221 * if no yield is determined to be needed. 1222 */ 1223 void 1224 lwkt_user_yield(void) 1225 { 1226 globaldata_t gd = mycpu; 1227 thread_t td = gd->gd_curthread; 1228 1229 /* 1230 * Always run any pending interrupts in case we are in a critical 1231 * section. 1232 */ 1233 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1234 splz(); 1235 1236 /* 1237 * Switch (which forces a release) if another kernel thread needs 1238 * the cpu, if userland wants us to resched, or if our kernel 1239 * quantum has run out. 1240 */ 1241 if (lwkt_resched_wanted() || 1242 user_resched_wanted()) 1243 { 1244 lwkt_switch(); 1245 } 1246 1247 #if 0 1248 /* 1249 * Reacquire the current process if we are released. 1250 * 1251 * XXX not implemented atm. The kernel may be holding locks and such, 1252 * so we want the thread to continue to receive cpu. 1253 */ 1254 if (td->td_release == NULL && lp) { 1255 lp->lwp_proc->p_usched->acquire_curproc(lp); 1256 td->td_release = lwkt_passive_release; 1257 lwkt_setpri_self(TDPRI_USER_NORM); 1258 } 1259 #endif 1260 } 1261 1262 /* 1263 * Generic schedule. Possibly schedule threads belonging to other cpus and 1264 * deal with threads that might be blocked on a wait queue. 1265 * 1266 * We have a little helper inline function which does additional work after 1267 * the thread has been enqueued, including dealing with preemption and 1268 * setting need_lwkt_resched() (which prevents the kernel from returning 1269 * to userland until it has processed higher priority threads). 1270 * 1271 * It is possible for this routine to be called after a failed _enqueue 1272 * (due to the target thread migrating, sleeping, or otherwise blocked). 1273 * We have to check that the thread is actually on the run queue! 1274 */ 1275 static __inline 1276 void 1277 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount) 1278 { 1279 if (ntd->td_flags & TDF_RUNQ) { 1280 if (ntd->td_preemptable) { 1281 ntd->td_preemptable(ntd, ccount); /* YYY +token */ 1282 } 1283 } 1284 } 1285 1286 static __inline 1287 void 1288 _lwkt_schedule(thread_t td) 1289 { 1290 globaldata_t mygd = mycpu; 1291 1292 KASSERT(td != &td->td_gd->gd_idlethread, 1293 ("lwkt_schedule(): scheduling gd_idlethread is illegal!")); 1294 KKASSERT((td->td_flags & TDF_MIGRATING) == 0); 1295 crit_enter_gd(mygd); 1296 KKASSERT(td->td_lwp == NULL || 1297 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 1298 1299 if (td == mygd->gd_curthread) { 1300 _lwkt_enqueue(td); 1301 } else { 1302 /* 1303 * If we own the thread, there is no race (since we are in a 1304 * critical section). If we do not own the thread there might 1305 * be a race but the target cpu will deal with it. 1306 */ 1307 if (td->td_gd == mygd) { 1308 _lwkt_enqueue(td); 1309 _lwkt_schedule_post(mygd, td, 1); 1310 } else { 1311 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0); 1312 } 1313 } 1314 crit_exit_gd(mygd); 1315 } 1316 1317 void 1318 lwkt_schedule(thread_t td) 1319 { 1320 _lwkt_schedule(td); 1321 } 1322 1323 void 1324 lwkt_schedule_noresched(thread_t td) /* XXX not impl */ 1325 { 1326 _lwkt_schedule(td); 1327 } 1328 1329 /* 1330 * When scheduled remotely if frame != NULL the IPIQ is being 1331 * run via doreti or an interrupt then preemption can be allowed. 1332 * 1333 * To allow preemption we have to drop the critical section so only 1334 * one is present in _lwkt_schedule_post. 1335 */ 1336 static void 1337 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame) 1338 { 1339 thread_t td = curthread; 1340 thread_t ntd = arg; 1341 1342 if (frame && ntd->td_preemptable) { 1343 crit_exit_noyield(td); 1344 _lwkt_schedule(ntd); 1345 crit_enter_quick(td); 1346 } else { 1347 _lwkt_schedule(ntd); 1348 } 1349 } 1350 1351 /* 1352 * Thread migration using a 'Pull' method. The thread may or may not be 1353 * the current thread. It MUST be descheduled and in a stable state. 1354 * lwkt_giveaway() must be called on the cpu owning the thread. 1355 * 1356 * At any point after lwkt_giveaway() is called, the target cpu may 1357 * 'pull' the thread by calling lwkt_acquire(). 1358 * 1359 * We have to make sure the thread is not sitting on a per-cpu tsleep 1360 * queue or it will blow up when it moves to another cpu. 1361 * 1362 * MPSAFE - must be called under very specific conditions. 1363 */ 1364 void 1365 lwkt_giveaway(thread_t td) 1366 { 1367 globaldata_t gd = mycpu; 1368 1369 crit_enter_gd(gd); 1370 if (td->td_flags & TDF_TSLEEPQ) 1371 tsleep_remove(td); 1372 KKASSERT(td->td_gd == gd); 1373 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); 1374 td->td_flags |= TDF_MIGRATING; 1375 crit_exit_gd(gd); 1376 } 1377 1378 void 1379 lwkt_acquire(thread_t td) 1380 { 1381 globaldata_t gd; 1382 globaldata_t mygd; 1383 int retry = 10000000; 1384 1385 KKASSERT(td->td_flags & TDF_MIGRATING); 1386 gd = td->td_gd; 1387 mygd = mycpu; 1388 if (gd != mycpu) { 1389 cpu_lfence(); 1390 KKASSERT((td->td_flags & TDF_RUNQ) == 0); 1391 crit_enter_gd(mygd); 1392 DEBUG_PUSH_INFO("lwkt_acquire"); 1393 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) { 1394 lwkt_process_ipiq(); 1395 cpu_lfence(); 1396 if (--retry == 0) { 1397 kprintf("lwkt_acquire: stuck: td %p td->td_flags %08x\n", 1398 td, td->td_flags); 1399 retry = 10000000; 1400 } 1401 #ifdef _KERNEL_VIRTUAL 1402 pthread_yield(); 1403 #endif 1404 } 1405 DEBUG_POP_INFO(); 1406 cpu_mfence(); 1407 td->td_gd = mygd; 1408 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1409 td->td_flags &= ~TDF_MIGRATING; 1410 crit_exit_gd(mygd); 1411 } else { 1412 crit_enter_gd(mygd); 1413 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1414 td->td_flags &= ~TDF_MIGRATING; 1415 crit_exit_gd(mygd); 1416 } 1417 } 1418 1419 /* 1420 * Generic deschedule. Descheduling threads other then your own should be 1421 * done only in carefully controlled circumstances. Descheduling is 1422 * asynchronous. 1423 * 1424 * This function may block if the cpu has run out of messages. 1425 */ 1426 void 1427 lwkt_deschedule(thread_t td) 1428 { 1429 crit_enter(); 1430 if (td == curthread) { 1431 _lwkt_dequeue(td); 1432 } else { 1433 if (td->td_gd == mycpu) { 1434 _lwkt_dequeue(td); 1435 } else { 1436 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td); 1437 } 1438 } 1439 crit_exit(); 1440 } 1441 1442 /* 1443 * Set the target thread's priority. This routine does not automatically 1444 * switch to a higher priority thread, LWKT threads are not designed for 1445 * continuous priority changes. Yield if you want to switch. 1446 */ 1447 void 1448 lwkt_setpri(thread_t td, int pri) 1449 { 1450 if (td->td_pri != pri) { 1451 KKASSERT(pri >= 0); 1452 crit_enter(); 1453 if (td->td_flags & TDF_RUNQ) { 1454 KKASSERT(td->td_gd == mycpu); 1455 _lwkt_dequeue(td); 1456 td->td_pri = pri; 1457 _lwkt_enqueue(td); 1458 } else { 1459 td->td_pri = pri; 1460 } 1461 crit_exit(); 1462 } 1463 } 1464 1465 /* 1466 * Set the initial priority for a thread prior to it being scheduled for 1467 * the first time. The thread MUST NOT be scheduled before or during 1468 * this call. The thread may be assigned to a cpu other then the current 1469 * cpu. 1470 * 1471 * Typically used after a thread has been created with TDF_STOPPREQ, 1472 * and before the thread is initially scheduled. 1473 */ 1474 void 1475 lwkt_setpri_initial(thread_t td, int pri) 1476 { 1477 KKASSERT(pri >= 0); 1478 KKASSERT((td->td_flags & TDF_RUNQ) == 0); 1479 td->td_pri = pri; 1480 } 1481 1482 void 1483 lwkt_setpri_self(int pri) 1484 { 1485 thread_t td = curthread; 1486 1487 KKASSERT(pri >= 0 && pri <= TDPRI_MAX); 1488 crit_enter(); 1489 if (td->td_flags & TDF_RUNQ) { 1490 _lwkt_dequeue(td); 1491 td->td_pri = pri; 1492 _lwkt_enqueue(td); 1493 } else { 1494 td->td_pri = pri; 1495 } 1496 crit_exit(); 1497 } 1498 1499 /* 1500 * hz tick scheduler clock for LWKT threads 1501 */ 1502 void 1503 lwkt_schedulerclock(thread_t td) 1504 { 1505 globaldata_t gd = td->td_gd; 1506 thread_t xtd; 1507 1508 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) { 1509 /* 1510 * If the current thread is at the head of the runq shift it to the 1511 * end of any equal-priority threads and request a LWKT reschedule 1512 * if it moved. 1513 * 1514 * Ignore upri in this situation. There will only be one user thread 1515 * in user mode, all others will be user threads running in kernel 1516 * mode and we have to make sure they get some cpu. 1517 */ 1518 xtd = TAILQ_NEXT(td, td_threadq); 1519 if (xtd && xtd->td_pri == td->td_pri) { 1520 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq); 1521 while (xtd && xtd->td_pri == td->td_pri) 1522 xtd = TAILQ_NEXT(xtd, td_threadq); 1523 if (xtd) 1524 TAILQ_INSERT_BEFORE(xtd, td, td_threadq); 1525 else 1526 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 1527 need_lwkt_resched(); 1528 } 1529 } else { 1530 /* 1531 * If we scheduled a thread other than the one at the head of the 1532 * queue always request a reschedule every tick. 1533 */ 1534 need_lwkt_resched(); 1535 } 1536 } 1537 1538 /* 1539 * Migrate the current thread to the specified cpu. 1540 * 1541 * This is accomplished by descheduling ourselves from the current cpu 1542 * and setting td_migrate_gd. The lwkt_switch() code will detect that the 1543 * 'old' thread wants to migrate after it has been completely switched out 1544 * and will complete the migration. 1545 * 1546 * TDF_MIGRATING prevents scheduling races while the thread is being migrated. 1547 * 1548 * We must be sure to release our current process designation (if a user 1549 * process) before clearing out any tsleepq we are on because the release 1550 * code may re-add us. 1551 * 1552 * We must be sure to remove ourselves from the current cpu's tsleepq 1553 * before potentially moving to another queue. The thread can be on 1554 * a tsleepq due to a left-over tsleep_interlock(). 1555 */ 1556 1557 void 1558 lwkt_setcpu_self(globaldata_t rgd) 1559 { 1560 thread_t td = curthread; 1561 1562 if (td->td_gd != rgd) { 1563 crit_enter_quick(td); 1564 1565 if (td->td_release) 1566 td->td_release(td); 1567 if (td->td_flags & TDF_TSLEEPQ) 1568 tsleep_remove(td); 1569 1570 /* 1571 * Set TDF_MIGRATING to prevent a spurious reschedule while we are 1572 * trying to deschedule ourselves and switch away, then deschedule 1573 * ourself, remove us from tdallq, and set td_migrate_gd. Finally, 1574 * call lwkt_switch() to complete the operation. 1575 */ 1576 td->td_flags |= TDF_MIGRATING; 1577 lwkt_deschedule_self(td); 1578 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1579 td->td_migrate_gd = rgd; 1580 lwkt_switch(); 1581 1582 /* 1583 * We are now on the target cpu 1584 */ 1585 KKASSERT(rgd == mycpu); 1586 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); 1587 crit_exit_quick(td); 1588 } 1589 } 1590 1591 void 1592 lwkt_migratecpu(int cpuid) 1593 { 1594 globaldata_t rgd; 1595 1596 rgd = globaldata_find(cpuid); 1597 lwkt_setcpu_self(rgd); 1598 } 1599 1600 /* 1601 * Remote IPI for cpu migration (called while in a critical section so we 1602 * do not have to enter another one). 1603 * 1604 * The thread (td) has already been completely descheduled from the 1605 * originating cpu and we can simply assert the case. The thread is 1606 * assigned to the new cpu and enqueued. 1607 * 1608 * The thread will re-add itself to tdallq when it resumes execution. 1609 */ 1610 static void 1611 lwkt_setcpu_remote(void *arg) 1612 { 1613 thread_t td = arg; 1614 globaldata_t gd = mycpu; 1615 1616 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0); 1617 td->td_gd = gd; 1618 cpu_mfence(); 1619 td->td_flags &= ~TDF_MIGRATING; 1620 KKASSERT(td->td_migrate_gd == NULL); 1621 KKASSERT(td->td_lwp == NULL || 1622 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 1623 _lwkt_enqueue(td); 1624 } 1625 1626 struct lwp * 1627 lwkt_preempted_proc(void) 1628 { 1629 thread_t td = curthread; 1630 while (td->td_preempted) 1631 td = td->td_preempted; 1632 return(td->td_lwp); 1633 } 1634 1635 /* 1636 * Create a kernel process/thread/whatever. It shares it's address space 1637 * with proc0 - ie: kernel only. 1638 * 1639 * If the cpu is not specified one will be selected. In the future 1640 * specifying a cpu of -1 will enable kernel thread migration between 1641 * cpus. 1642 */ 1643 int 1644 lwkt_create(void (*func)(void *), void *arg, struct thread **tdp, 1645 thread_t template, int tdflags, int cpu, const char *fmt, ...) 1646 { 1647 thread_t td; 1648 __va_list ap; 1649 1650 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu, 1651 tdflags); 1652 if (tdp) 1653 *tdp = td; 1654 cpu_set_thread_handler(td, lwkt_exit, func, arg); 1655 1656 /* 1657 * Set up arg0 for 'ps' etc 1658 */ 1659 __va_start(ap, fmt); 1660 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap); 1661 __va_end(ap); 1662 1663 /* 1664 * Schedule the thread to run 1665 */ 1666 if (td->td_flags & TDF_NOSTART) 1667 td->td_flags &= ~TDF_NOSTART; 1668 else 1669 lwkt_schedule(td); 1670 return 0; 1671 } 1672 1673 /* 1674 * Destroy an LWKT thread. Warning! This function is not called when 1675 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and 1676 * uses a different reaping mechanism. 1677 */ 1678 void 1679 lwkt_exit(void) 1680 { 1681 thread_t td = curthread; 1682 thread_t std; 1683 globaldata_t gd; 1684 1685 /* 1686 * Do any cleanup that might block here 1687 */ 1688 if (td->td_flags & TDF_VERBOSE) 1689 kprintf("kthread %p %s has exited\n", td, td->td_comm); 1690 biosched_done(td); 1691 dsched_exit_thread(td); 1692 1693 /* 1694 * Get us into a critical section to interlock gd_freetd and loop 1695 * until we can get it freed. 1696 * 1697 * We have to cache the current td in gd_freetd because objcache_put()ing 1698 * it would rip it out from under us while our thread is still active. 1699 * 1700 * We are the current thread so of course our own TDF_RUNNING bit will 1701 * be set, so unlike the lwp reap code we don't wait for it to clear. 1702 */ 1703 gd = mycpu; 1704 crit_enter_quick(td); 1705 for (;;) { 1706 if (td->td_refs) { 1707 tsleep(td, 0, "tdreap", 1); 1708 continue; 1709 } 1710 if ((std = gd->gd_freetd) != NULL) { 1711 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0); 1712 gd->gd_freetd = NULL; 1713 objcache_put(thread_cache, std); 1714 continue; 1715 } 1716 break; 1717 } 1718 1719 /* 1720 * Remove thread resources from kernel lists and deschedule us for 1721 * the last time. We cannot block after this point or we may end 1722 * up with a stale td on the tsleepq. 1723 * 1724 * None of this may block, the critical section is the only thing 1725 * protecting tdallq and the only thing preventing new lwkt_hold() 1726 * thread refs now. 1727 */ 1728 if (td->td_flags & TDF_TSLEEPQ) 1729 tsleep_remove(td); 1730 lwkt_deschedule_self(td); 1731 lwkt_remove_tdallq(td); 1732 KKASSERT(td->td_refs == 0); 1733 1734 /* 1735 * Final cleanup 1736 */ 1737 KKASSERT(gd->gd_freetd == NULL); 1738 if (td->td_flags & TDF_ALLOCATED_THREAD) 1739 gd->gd_freetd = td; 1740 cpu_thread_exit(); 1741 } 1742 1743 void 1744 lwkt_remove_tdallq(thread_t td) 1745 { 1746 KKASSERT(td->td_gd == mycpu); 1747 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1748 } 1749 1750 /* 1751 * Code reduction and branch prediction improvements. Call/return 1752 * overhead on modern cpus often degenerates into 0 cycles due to 1753 * the cpu's branch prediction hardware and return pc cache. We 1754 * can take advantage of this by not inlining medium-complexity 1755 * functions and we can also reduce the branch prediction impact 1756 * by collapsing perfectly predictable branches into a single 1757 * procedure instead of duplicating it. 1758 * 1759 * Is any of this noticeable? Probably not, so I'll take the 1760 * smaller code size. 1761 */ 1762 void 1763 crit_exit_wrapper(__DEBUG_CRIT_ARG__) 1764 { 1765 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__); 1766 } 1767 1768 void 1769 crit_panic(void) 1770 { 1771 thread_t td = curthread; 1772 int lcrit = td->td_critcount; 1773 1774 td->td_critcount = 0; 1775 panic("td_critcount is/would-go negative! %p %d", td, lcrit); 1776 /* NOT REACHED */ 1777 } 1778 1779 /* 1780 * Called from debugger/panic on cpus which have been stopped. We must still 1781 * process the IPIQ while stopped, even if we were stopped while in a critical 1782 * section (XXX). 1783 * 1784 * If we are dumping also try to process any pending interrupts. This may 1785 * or may not work depending on the state of the cpu at the point it was 1786 * stopped. 1787 */ 1788 void 1789 lwkt_smp_stopped(void) 1790 { 1791 globaldata_t gd = mycpu; 1792 1793 crit_enter_gd(gd); 1794 if (dumping) { 1795 lwkt_process_ipiq(); 1796 splz(); 1797 } else { 1798 lwkt_process_ipiq(); 1799 } 1800 crit_exit_gd(gd); 1801 } 1802