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