1 /* 2 * Copyright (c) 2003-2010 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/caps.h> 54 #include <sys/spinlock.h> 55 #include <sys/ktr.h> 56 57 #include <sys/thread2.h> 58 #include <sys/spinlock2.h> 59 #include <sys/mplock2.h> 60 61 #include <sys/dsched.h> 62 63 #include <vm/vm.h> 64 #include <vm/vm_param.h> 65 #include <vm/vm_kern.h> 66 #include <vm/vm_object.h> 67 #include <vm/vm_page.h> 68 #include <vm/vm_map.h> 69 #include <vm/vm_pager.h> 70 #include <vm/vm_extern.h> 71 72 #include <machine/stdarg.h> 73 #include <machine/smp.h> 74 75 #if !defined(KTR_CTXSW) 76 #define KTR_CTXSW KTR_ALL 77 #endif 78 KTR_INFO_MASTER(ctxsw); 79 KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p", 80 sizeof(int) + sizeof(struct thread *)); 81 KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p", 82 sizeof(int) + sizeof(struct thread *)); 83 KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s", 84 sizeof (struct thread *) + sizeof(char *)); 85 KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", sizeof (struct thread *)); 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 __int64_t token_contention_count __debugvar = 0; 97 static int lwkt_use_spin_port; 98 static struct objcache *thread_cache; 99 100 #ifdef SMP 101 static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame); 102 #endif 103 static void lwkt_fairq_accumulate(globaldata_t gd, thread_t td); 104 105 extern void cpu_heavy_restore(void); 106 extern void cpu_lwkt_restore(void); 107 extern void cpu_kthread_restore(void); 108 extern void cpu_idle_restore(void); 109 110 /* 111 * We can make all thread ports use the spin backend instead of the thread 112 * backend. This should only be set to debug the spin backend. 113 */ 114 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port); 115 116 #ifdef INVARIANTS 117 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, 118 "Panic if attempting to switch lwkt's while mastering cpusync"); 119 #endif 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 #ifdef INVARIANTS 129 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW, 130 &token_contention_count, 0, "spinning due to token contention"); 131 #endif 132 static int fairq_enable = 1; 133 SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW, 134 &fairq_enable, 0, "Turn on fairq priority accumulators"); 135 static int lwkt_spin_loops = 10; 136 SYSCTL_INT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW, 137 &lwkt_spin_loops, 0, ""); 138 static int lwkt_spin_delay = 1; 139 SYSCTL_INT(_lwkt, OID_AUTO, spin_delay, CTLFLAG_RW, 140 &lwkt_spin_delay, 0, "Scheduler spin delay in microseconds 0=auto"); 141 static int lwkt_spin_method = 1; 142 SYSCTL_INT(_lwkt, OID_AUTO, spin_method, CTLFLAG_RW, 143 &lwkt_spin_method, 0, "LWKT scheduler behavior when contended"); 144 static int preempt_enable = 1; 145 SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW, 146 &preempt_enable, 0, "Enable preemption"); 147 148 static __cachealign int lwkt_cseq_rindex; 149 static __cachealign int lwkt_cseq_windex; 150 151 /* 152 * These helper procedures handle the runq, they can only be called from 153 * within a critical section. 154 * 155 * WARNING! Prior to SMP being brought up it is possible to enqueue and 156 * dequeue threads belonging to other cpus, so be sure to use td->td_gd 157 * instead of 'mycpu' when referencing the globaldata structure. Once 158 * SMP live enqueuing and dequeueing only occurs on the current cpu. 159 */ 160 static __inline 161 void 162 _lwkt_dequeue(thread_t td) 163 { 164 if (td->td_flags & TDF_RUNQ) { 165 struct globaldata *gd = td->td_gd; 166 167 td->td_flags &= ~TDF_RUNQ; 168 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq); 169 gd->gd_fairq_total_pri -= td->td_pri; 170 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL) 171 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING); 172 } 173 } 174 175 /* 176 * Priority enqueue. 177 * 178 * NOTE: There are a limited number of lwkt threads runnable since user 179 * processes only schedule one at a time per cpu. 180 */ 181 static __inline 182 void 183 _lwkt_enqueue(thread_t td) 184 { 185 thread_t xtd; 186 187 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) { 188 struct globaldata *gd = td->td_gd; 189 190 td->td_flags |= TDF_RUNQ; 191 xtd = TAILQ_FIRST(&gd->gd_tdrunq); 192 if (xtd == NULL) { 193 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 194 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING); 195 } else { 196 while (xtd && xtd->td_pri > td->td_pri) 197 xtd = TAILQ_NEXT(xtd, td_threadq); 198 if (xtd) 199 TAILQ_INSERT_BEFORE(xtd, td, td_threadq); 200 else 201 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 202 } 203 gd->gd_fairq_total_pri += td->td_pri; 204 } 205 } 206 207 static __boolean_t 208 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags) 209 { 210 struct thread *td = (struct thread *)obj; 211 212 td->td_kstack = NULL; 213 td->td_kstack_size = 0; 214 td->td_flags = TDF_ALLOCATED_THREAD; 215 return (1); 216 } 217 218 static void 219 _lwkt_thread_dtor(void *obj, void *privdata) 220 { 221 struct thread *td = (struct thread *)obj; 222 223 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD, 224 ("_lwkt_thread_dtor: not allocated from objcache")); 225 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack && 226 td->td_kstack_size > 0, 227 ("_lwkt_thread_dtor: corrupted stack")); 228 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size); 229 } 230 231 /* 232 * Initialize the lwkt s/system. 233 */ 234 void 235 lwkt_init(void) 236 { 237 /* An objcache has 2 magazines per CPU so divide cache size by 2. */ 238 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread), 239 NULL, CACHE_NTHREADS/2, 240 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL); 241 } 242 243 /* 244 * Schedule a thread to run. As the current thread we can always safely 245 * schedule ourselves, and a shortcut procedure is provided for that 246 * function. 247 * 248 * (non-blocking, self contained on a per cpu basis) 249 */ 250 void 251 lwkt_schedule_self(thread_t td) 252 { 253 crit_enter_quick(td); 254 KASSERT(td != &td->td_gd->gd_idlethread, 255 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!")); 256 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0); 257 _lwkt_enqueue(td); 258 crit_exit_quick(td); 259 } 260 261 /* 262 * Deschedule a thread. 263 * 264 * (non-blocking, self contained on a per cpu basis) 265 */ 266 void 267 lwkt_deschedule_self(thread_t td) 268 { 269 crit_enter_quick(td); 270 _lwkt_dequeue(td); 271 crit_exit_quick(td); 272 } 273 274 /* 275 * LWKTs operate on a per-cpu basis 276 * 277 * WARNING! Called from early boot, 'mycpu' may not work yet. 278 */ 279 void 280 lwkt_gdinit(struct globaldata *gd) 281 { 282 TAILQ_INIT(&gd->gd_tdrunq); 283 TAILQ_INIT(&gd->gd_tdallq); 284 } 285 286 /* 287 * Create a new thread. The thread must be associated with a process context 288 * or LWKT start address before it can be scheduled. If the target cpu is 289 * -1 the thread will be created on the current cpu. 290 * 291 * If you intend to create a thread without a process context this function 292 * does everything except load the startup and switcher function. 293 */ 294 thread_t 295 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags) 296 { 297 globaldata_t gd = mycpu; 298 void *stack; 299 300 /* 301 * If static thread storage is not supplied allocate a thread. Reuse 302 * a cached free thread if possible. gd_freetd is used to keep an exiting 303 * thread intact through the exit. 304 */ 305 if (td == NULL) { 306 crit_enter_gd(gd); 307 if ((td = gd->gd_freetd) != NULL) { 308 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK| 309 TDF_RUNQ)) == 0); 310 gd->gd_freetd = NULL; 311 } else { 312 td = objcache_get(thread_cache, M_WAITOK); 313 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK| 314 TDF_RUNQ)) == 0); 315 } 316 crit_exit_gd(gd); 317 KASSERT((td->td_flags & 318 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD, 319 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags)); 320 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK); 321 } 322 323 /* 324 * Try to reuse cached stack. 325 */ 326 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) { 327 if (flags & TDF_ALLOCATED_STACK) { 328 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size); 329 stack = NULL; 330 } 331 } 332 if (stack == NULL) { 333 stack = (void *)kmem_alloc_stack(&kernel_map, stksize); 334 flags |= TDF_ALLOCATED_STACK; 335 } 336 if (cpu < 0) 337 lwkt_init_thread(td, stack, stksize, flags, gd); 338 else 339 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu)); 340 return(td); 341 } 342 343 /* 344 * Initialize a preexisting thread structure. This function is used by 345 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread. 346 * 347 * All threads start out in a critical section at a priority of 348 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as 349 * appropriate. This function may send an IPI message when the 350 * requested cpu is not the current cpu and consequently gd_tdallq may 351 * not be initialized synchronously from the point of view of the originating 352 * cpu. 353 * 354 * NOTE! we have to be careful in regards to creating threads for other cpus 355 * if SMP has not yet been activated. 356 */ 357 #ifdef SMP 358 359 static void 360 lwkt_init_thread_remote(void *arg) 361 { 362 thread_t td = arg; 363 364 /* 365 * Protected by critical section held by IPI dispatch 366 */ 367 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq); 368 } 369 370 #endif 371 372 /* 373 * lwkt core thread structural initialization. 374 * 375 * NOTE: All threads are initialized as mpsafe threads. 376 */ 377 void 378 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags, 379 struct globaldata *gd) 380 { 381 globaldata_t mygd = mycpu; 382 383 bzero(td, sizeof(struct thread)); 384 td->td_kstack = stack; 385 td->td_kstack_size = stksize; 386 td->td_flags = flags; 387 td->td_gd = gd; 388 td->td_pri = TDPRI_KERN_DAEMON; 389 td->td_critcount = 1; 390 td->td_toks_stop = &td->td_toks_base; 391 if (lwkt_use_spin_port) 392 lwkt_initport_spin(&td->td_msgport); 393 else 394 lwkt_initport_thread(&td->td_msgport, td); 395 pmap_init_thread(td); 396 #ifdef SMP 397 /* 398 * Normally initializing a thread for a remote cpu requires sending an 399 * IPI. However, the idlethread is setup before the other cpus are 400 * activated so we have to treat it as a special case. XXX manipulation 401 * of gd_tdallq requires the BGL. 402 */ 403 if (gd == mygd || td == &gd->gd_idlethread) { 404 crit_enter_gd(mygd); 405 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); 406 crit_exit_gd(mygd); 407 } else { 408 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td); 409 } 410 #else 411 crit_enter_gd(mygd); 412 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); 413 crit_exit_gd(mygd); 414 #endif 415 416 dsched_new_thread(td); 417 } 418 419 void 420 lwkt_set_comm(thread_t td, const char *ctl, ...) 421 { 422 __va_list va; 423 424 __va_start(va, ctl); 425 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va); 426 __va_end(va); 427 KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]); 428 } 429 430 void 431 lwkt_hold(thread_t td) 432 { 433 ++td->td_refs; 434 } 435 436 void 437 lwkt_rele(thread_t td) 438 { 439 KKASSERT(td->td_refs > 0); 440 --td->td_refs; 441 } 442 443 void 444 lwkt_wait_free(thread_t td) 445 { 446 while (td->td_refs) 447 tsleep(td, 0, "tdreap", hz); 448 } 449 450 void 451 lwkt_free_thread(thread_t td) 452 { 453 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|TDF_RUNQ)) == 0); 454 if (td->td_flags & TDF_ALLOCATED_THREAD) { 455 objcache_put(thread_cache, td); 456 } else if (td->td_flags & TDF_ALLOCATED_STACK) { 457 /* client-allocated struct with internally allocated stack */ 458 KASSERT(td->td_kstack && td->td_kstack_size > 0, 459 ("lwkt_free_thread: corrupted stack")); 460 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size); 461 td->td_kstack = NULL; 462 td->td_kstack_size = 0; 463 } 464 KTR_LOG(ctxsw_deadtd, td); 465 } 466 467 468 /* 469 * Switch to the next runnable lwkt. If no LWKTs are runnable then 470 * switch to the idlethread. Switching must occur within a critical 471 * section to avoid races with the scheduling queue. 472 * 473 * We always have full control over our cpu's run queue. Other cpus 474 * that wish to manipulate our queue must use the cpu_*msg() calls to 475 * talk to our cpu, so a critical section is all that is needed and 476 * the result is very, very fast thread switching. 477 * 478 * The LWKT scheduler uses a fixed priority model and round-robins at 479 * each priority level. User process scheduling is a totally 480 * different beast and LWKT priorities should not be confused with 481 * user process priorities. 482 * 483 * Note that the td_switch() function cannot do anything that requires 484 * the MP lock since the MP lock will have already been setup for 485 * the target thread (not the current thread). It's nice to have a scheduler 486 * that does not need the MP lock to work because it allows us to do some 487 * really cool high-performance MP lock optimizations. 488 * 489 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch() 490 * is not called by the current thread in the preemption case, only when 491 * the preempting thread blocks (in order to return to the original thread). 492 */ 493 void 494 lwkt_switch(void) 495 { 496 globaldata_t gd = mycpu; 497 thread_t td = gd->gd_curthread; 498 thread_t ntd; 499 thread_t xtd; 500 int spinning = lwkt_spin_loops; /* loops before HLTing */ 501 int reqflags; 502 int cseq; 503 504 /* 505 * Switching from within a 'fast' (non thread switched) interrupt or IPI 506 * is illegal. However, we may have to do it anyway if we hit a fatal 507 * kernel trap or we have paniced. 508 * 509 * If this case occurs save and restore the interrupt nesting level. 510 */ 511 if (gd->gd_intr_nesting_level) { 512 int savegdnest; 513 int savegdtrap; 514 515 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) { 516 panic("lwkt_switch: Attempt to switch from a " 517 "a fast interrupt, ipi, or hard code section, " 518 "td %p\n", 519 td); 520 } else { 521 savegdnest = gd->gd_intr_nesting_level; 522 savegdtrap = gd->gd_trap_nesting_level; 523 gd->gd_intr_nesting_level = 0; 524 gd->gd_trap_nesting_level = 0; 525 if ((td->td_flags & TDF_PANICWARN) == 0) { 526 td->td_flags |= TDF_PANICWARN; 527 kprintf("Warning: thread switch from interrupt, IPI, " 528 "or hard code section.\n" 529 "thread %p (%s)\n", td, td->td_comm); 530 print_backtrace(-1); 531 } 532 lwkt_switch(); 533 gd->gd_intr_nesting_level = savegdnest; 534 gd->gd_trap_nesting_level = savegdtrap; 535 return; 536 } 537 } 538 539 /* 540 * Passive release (used to transition from user to kernel mode 541 * when we block or switch rather then when we enter the kernel). 542 * This function is NOT called if we are switching into a preemption 543 * or returning from a preemption. Typically this causes us to lose 544 * our current process designation (if we have one) and become a true 545 * LWKT thread, and may also hand the current process designation to 546 * another process and schedule thread. 547 */ 548 if (td->td_release) 549 td->td_release(td); 550 551 crit_enter_gd(gd); 552 if (TD_TOKS_HELD(td)) 553 lwkt_relalltokens(td); 554 555 /* 556 * We had better not be holding any spin locks, but don't get into an 557 * endless panic loop. 558 */ 559 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL, 560 ("lwkt_switch: still holding %d exclusive spinlocks!", 561 gd->gd_spinlocks_wr)); 562 563 564 #ifdef SMP 565 #ifdef INVARIANTS 566 if (td->td_cscount) { 567 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n", 568 td); 569 if (panic_on_cscount) 570 panic("switching while mastering cpusync"); 571 } 572 #endif 573 #endif 574 575 /* 576 * If we had preempted another thread on this cpu, resume the preempted 577 * thread. This occurs transparently, whether the preempted thread 578 * was scheduled or not (it may have been preempted after descheduling 579 * itself). 580 * 581 * We have to setup the MP lock for the original thread after backing 582 * out the adjustment that was made to curthread when the original 583 * was preempted. 584 */ 585 if ((ntd = td->td_preempted) != NULL) { 586 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK); 587 ntd->td_flags |= TDF_PREEMPT_DONE; 588 589 /* 590 * The interrupt may have woken a thread up, we need to properly 591 * set the reschedule flag if the originally interrupted thread is 592 * at a lower priority. 593 */ 594 if (TAILQ_FIRST(&gd->gd_tdrunq) && 595 TAILQ_FIRST(&gd->gd_tdrunq)->td_pri > ntd->td_pri) { 596 need_lwkt_resched(); 597 } 598 /* YYY release mp lock on switchback if original doesn't need it */ 599 goto havethread_preempted; 600 } 601 602 /* 603 * Implement round-robin fairq with priority insertion. The priority 604 * insertion is handled by _lwkt_enqueue() 605 * 606 * We have to adjust the MP lock for the target thread. If we 607 * need the MP lock and cannot obtain it we try to locate a 608 * thread that does not need the MP lock. If we cannot, we spin 609 * instead of HLT. 610 * 611 * A similar issue exists for the tokens held by the target thread. 612 * If we cannot obtain ownership of the tokens we cannot immediately 613 * schedule the thread. 614 */ 615 for (;;) { 616 /* 617 * Clear RQF_AST_LWKT_RESCHED (we handle the reschedule request) 618 * and set RQF_WAKEUP (prevent unnecessary IPIs from being 619 * received). 620 */ 621 for (;;) { 622 reqflags = gd->gd_reqflags; 623 if (atomic_cmpset_int(&gd->gd_reqflags, reqflags, 624 (reqflags & ~RQF_AST_LWKT_RESCHED) | 625 RQF_WAKEUP)) { 626 break; 627 } 628 } 629 630 /* 631 * Hotpath - pull the head of the run queue and attempt to schedule 632 * it. Fairq exhaustion moves the task to the end of the list. If 633 * no threads are runnable we switch to the idle thread. 634 */ 635 for (;;) { 636 ntd = TAILQ_FIRST(&gd->gd_tdrunq); 637 638 if (ntd == NULL) { 639 /* 640 * Runq is empty, switch to idle and clear RQF_WAKEUP 641 * to allow it to halt. 642 */ 643 ntd = &gd->gd_idlethread; 644 #ifdef SMP 645 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) 646 ASSERT_NO_TOKENS_HELD(ntd); 647 #endif 648 cpu_time.cp_msg[0] = 0; 649 cpu_time.cp_stallpc = 0; 650 atomic_clear_int(&gd->gd_reqflags, RQF_WAKEUP); 651 goto haveidle; 652 } 653 654 if (ntd->td_fairq_accum >= 0) 655 break; 656 657 splz_check(); 658 lwkt_fairq_accumulate(gd, ntd); 659 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq); 660 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, ntd, td_threadq); 661 } 662 663 /* 664 * Hotpath - schedule ntd. Leaves RQF_WAKEUP set to prevent 665 * unwanted decontention IPIs. 666 * 667 * NOTE: For UP there is no mplock and lwkt_getalltokens() 668 * always succeeds. 669 */ 670 if (TD_TOKS_NOT_HELD(ntd) || lwkt_getalltokens(ntd)) 671 goto havethread; 672 673 /* 674 * Coldpath (SMP only since tokens always succeed on UP) 675 * 676 * We had some contention on the thread we wanted to schedule. 677 * What we do now is try to find a thread that we can schedule 678 * in its stead until decontention reschedules on our cpu. 679 * 680 * The coldpath scan does NOT rearrange threads in the run list 681 * and it also ignores the accumulator. 682 * 683 * We do not immediately schedule a user priority thread, instead 684 * we record it in xtd and continue looking for kernel threads. 685 * A cpu can only have one user priority thread (normally) so just 686 * record the first one. 687 * 688 * NOTE: This scan will also include threads whos fairq's were 689 * accumulated in the first loop. 690 */ 691 ++token_contention_count; 692 xtd = NULL; 693 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) { 694 /* 695 * Try to switch to this thread. If the thread is running at 696 * user priority we clear WAKEUP to allow decontention IPIs 697 * (since this thread is simply running until the one we wanted 698 * decontends), and we make sure that LWKT_RESCHED is not set. 699 * 700 * Otherwise for kernel threads we leave WAKEUP set to avoid 701 * unnecessary decontention IPIs. 702 */ 703 if (ntd->td_pri < TDPRI_KERN_LPSCHED) { 704 if (xtd == NULL) 705 xtd = ntd; 706 continue; 707 } 708 709 /* 710 * Do not let the fairq get too negative. Even though we are 711 * ignoring it atm once the scheduler decontends a very negative 712 * thread will get moved to the end of the queue. 713 */ 714 if (TD_TOKS_NOT_HELD(ntd) || lwkt_getalltokens(ntd)) { 715 if (ntd->td_fairq_accum < -TDFAIRQ_MAX(gd)) 716 ntd->td_fairq_accum = -TDFAIRQ_MAX(gd); 717 goto havethread; 718 } 719 720 /* 721 * Well fubar, this thread is contended as well, loop 722 */ 723 /* */ 724 } 725 726 /* 727 * We exhausted the run list but we may have recorded a user 728 * thread to try. We have three choices based on 729 * lwkt.decontention_method. 730 * 731 * (0) Atomically clear RQF_WAKEUP in order to receive decontention 732 * IPIs (to interrupt the user process) and test 733 * RQF_AST_LWKT_RESCHED at the same time. 734 * 735 * This results in significant decontention IPI traffic but may 736 * be more responsive. 737 * 738 * (1) Leave RQF_WAKEUP set so we do not receive a decontention IPI. 739 * An automatic LWKT reschedule will occur on the next hardclock 740 * (typically 100hz). 741 * 742 * This results in no decontention IPI traffic but may be less 743 * responsive. This is the default. 744 * 745 * (2) Refuse to schedule the user process at this time. 746 * 747 * This is highly experimental and should not be used under 748 * normal circumstances. This can cause a user process to 749 * get starved out in situations where kernel threads are 750 * fighting each other for tokens. 751 */ 752 if (xtd) { 753 ntd = xtd; 754 755 switch(lwkt_spin_method) { 756 case 0: 757 for (;;) { 758 reqflags = gd->gd_reqflags; 759 if (atomic_cmpset_int(&gd->gd_reqflags, 760 reqflags, 761 reqflags & ~RQF_WAKEUP)) { 762 break; 763 } 764 } 765 break; 766 case 1: 767 reqflags = gd->gd_reqflags; 768 break; 769 default: 770 goto skip; 771 break; 772 } 773 if ((reqflags & RQF_AST_LWKT_RESCHED) == 0 && 774 (TD_TOKS_NOT_HELD(ntd) || lwkt_getalltokens(ntd)) 775 ) { 776 if (ntd->td_fairq_accum < -TDFAIRQ_MAX(gd)) 777 ntd->td_fairq_accum = -TDFAIRQ_MAX(gd); 778 goto havethread; 779 } 780 781 skip: 782 /* 783 * Make sure RQF_WAKEUP is set if we failed to schedule the 784 * user thread to prevent the idle thread from halting. 785 */ 786 atomic_set_int(&gd->gd_reqflags, RQF_WAKEUP); 787 } 788 789 /* 790 * We exhausted the run list, meaning that all runnable threads 791 * are contended. 792 */ 793 cpu_pause(); 794 ntd = &gd->gd_idlethread; 795 #ifdef SMP 796 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) 797 ASSERT_NO_TOKENS_HELD(ntd); 798 /* contention case, do not clear contention mask */ 799 #endif 800 801 /* 802 * Ok, we might want to spin a few times as some tokens are held for 803 * very short periods of time and IPI overhead is 1uS or worse 804 * (meaning it is usually better to spin). Regardless we have to 805 * call splz_check() to be sure to service any interrupts blocked 806 * by our critical section, otherwise we could livelock e.g. IPIs. 807 * 808 * The IPI mechanic is really a last resort. In nearly all other 809 * cases RQF_WAKEUP is left set to prevent decontention IPIs. 810 * 811 * When we decide not to spin we clear RQF_WAKEUP and switch to 812 * the idle thread. Clearing RQF_WEAKEUP allows the idle thread 813 * to halt and decontended tokens will issue an IPI to us. The 814 * idle thread will check for pending reschedules already set 815 * (RQF_AST_LWKT_RESCHED) before actually halting so we don't have 816 * to here. 817 */ 818 if (spinning <= 0) { 819 atomic_clear_int(&gd->gd_reqflags, RQF_WAKEUP); 820 goto haveidle; 821 } 822 --spinning; 823 824 /* 825 * When spinning a delay is required both to avoid livelocks from 826 * token order reversals (a thread may be trying to acquire multiple 827 * tokens), and also to reduce cpu cache management traffic. 828 * 829 * In order to scale to a large number of CPUs we use a time slot 830 * resequencer to force contending cpus into non-contending 831 * time-slots. The scheduler may still contend with the lock holder 832 * but will not (generally) contend with all the other cpus trying 833 * trying to get the same token. 834 * 835 * The resequencer uses a FIFO counter mechanic. The owner of the 836 * rindex at the head of the FIFO is allowed to pull itself off 837 * the FIFO and fetchadd is used to enter into the FIFO. This bit 838 * of code is VERY cache friendly and forces all spinning schedulers 839 * into their own time slots. 840 * 841 * This code has been tested to 48-cpus and caps the cache 842 * contention load at ~1uS intervals regardless of the number of 843 * cpus. Scaling beyond 64 cpus might require additional smarts 844 * (such as separate FIFOs for specific token cases). 845 * 846 * WARNING! We can't call splz_check() or anything else here as 847 * it could cause a deadlock. 848 */ 849 cseq = atomic_fetchadd_int(&lwkt_cseq_windex, 1); 850 while (lwkt_cseq_rindex != cseq) { 851 DELAY(1); 852 cpu_lfence(); 853 } 854 cseq = lwkt_spin_delay; /* don't trust the system operator */ 855 cpu_ccfence(); 856 if (cseq < 1) 857 cseq = 1; 858 if (cseq > 1000) 859 cseq = 1000; 860 DELAY(cseq); 861 atomic_add_int(&lwkt_cseq_rindex, 1); 862 splz_check(); 863 /* highest level for(;;) loop */ 864 } 865 866 havethread: 867 /* 868 * We must always decrement td_fairq_accum on non-idle threads just 869 * in case a thread never gets a tick due to being in a continuous 870 * critical section. The page-zeroing code does this, for example. 871 * 872 * If the thread we came up with is a higher or equal priority verses 873 * the thread at the head of the queue we move our thread to the 874 * front. This way we can always check the front of the queue. 875 */ 876 ++gd->gd_cnt.v_swtch; 877 --ntd->td_fairq_accum; 878 ntd->td_wmesg = NULL; 879 xtd = TAILQ_FIRST(&gd->gd_tdrunq); 880 if (ntd != xtd && ntd->td_pri >= xtd->td_pri) { 881 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq); 882 TAILQ_INSERT_HEAD(&gd->gd_tdrunq, ntd, td_threadq); 883 } 884 885 havethread_preempted: 886 /* 887 * If the new target does not need the MP lock and we are holding it, 888 * release the MP lock. If the new target requires the MP lock we have 889 * already acquired it for the target. 890 */ 891 ; 892 haveidle: 893 KASSERT(ntd->td_critcount, 894 ("priority problem in lwkt_switch %d %d", 895 td->td_critcount, ntd->td_critcount)); 896 897 if (td != ntd) { 898 ++switch_count; 899 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd); 900 td->td_switch(ntd); 901 } 902 /* NOTE: current cpu may have changed after switch */ 903 crit_exit_quick(td); 904 } 905 906 /* 907 * Request that the target thread preempt the current thread. Preemption 908 * only works under a specific set of conditions: 909 * 910 * - We are not preempting ourselves 911 * - The target thread is owned by the current cpu 912 * - We are not currently being preempted 913 * - The target is not currently being preempted 914 * - We are not holding any spin locks 915 * - The target thread is not holding any tokens 916 * - We are able to satisfy the target's MP lock requirements (if any). 917 * 918 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically 919 * this is called via lwkt_schedule() through the td_preemptable callback. 920 * critcount is the managed critical priority that we should ignore in order 921 * to determine whether preemption is possible (aka usually just the crit 922 * priority of lwkt_schedule() itself). 923 * 924 * XXX at the moment we run the target thread in a critical section during 925 * the preemption in order to prevent the target from taking interrupts 926 * that *WE* can't. Preemption is strictly limited to interrupt threads 927 * and interrupt-like threads, outside of a critical section, and the 928 * preempted source thread will be resumed the instant the target blocks 929 * whether or not the source is scheduled (i.e. preemption is supposed to 930 * be as transparent as possible). 931 */ 932 void 933 lwkt_preempt(thread_t ntd, int critcount) 934 { 935 struct globaldata *gd = mycpu; 936 thread_t td; 937 int save_gd_intr_nesting_level; 938 939 /* 940 * The caller has put us in a critical section. We can only preempt 941 * if the caller of the caller was not in a critical section (basically 942 * a local interrupt), as determined by the 'critcount' parameter. We 943 * also can't preempt if the caller is holding any spinlocks (even if 944 * he isn't in a critical section). This also handles the tokens test. 945 * 946 * YYY The target thread must be in a critical section (else it must 947 * inherit our critical section? I dunno yet). 948 * 949 * Set need_lwkt_resched() unconditionally for now YYY. 950 */ 951 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri)); 952 953 if (preempt_enable == 0) { 954 ++preempt_miss; 955 return; 956 } 957 958 td = gd->gd_curthread; 959 if (ntd->td_pri <= td->td_pri) { 960 ++preempt_miss; 961 return; 962 } 963 if (td->td_critcount > critcount) { 964 ++preempt_miss; 965 need_lwkt_resched(); 966 return; 967 } 968 #ifdef SMP 969 if (ntd->td_gd != gd) { 970 ++preempt_miss; 971 need_lwkt_resched(); 972 return; 973 } 974 #endif 975 /* 976 * We don't have to check spinlocks here as they will also bump 977 * td_critcount. 978 * 979 * Do not try to preempt if the target thread is holding any tokens. 980 * We could try to acquire the tokens but this case is so rare there 981 * is no need to support it. 982 */ 983 KKASSERT(gd->gd_spinlocks_wr == 0); 984 985 if (TD_TOKS_HELD(ntd)) { 986 ++preempt_miss; 987 need_lwkt_resched(); 988 return; 989 } 990 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) { 991 ++preempt_weird; 992 need_lwkt_resched(); 993 return; 994 } 995 if (ntd->td_preempted) { 996 ++preempt_hit; 997 need_lwkt_resched(); 998 return; 999 } 1000 1001 /* 1002 * Since we are able to preempt the current thread, there is no need to 1003 * call need_lwkt_resched(). 1004 * 1005 * We must temporarily clear gd_intr_nesting_level around the switch 1006 * since switchouts from the target thread are allowed (they will just 1007 * return to our thread), and since the target thread has its own stack. 1008 */ 1009 ++preempt_hit; 1010 ntd->td_preempted = td; 1011 td->td_flags |= TDF_PREEMPT_LOCK; 1012 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd); 1013 save_gd_intr_nesting_level = gd->gd_intr_nesting_level; 1014 gd->gd_intr_nesting_level = 0; 1015 td->td_switch(ntd); 1016 gd->gd_intr_nesting_level = save_gd_intr_nesting_level; 1017 1018 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE)); 1019 ntd->td_preempted = NULL; 1020 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE); 1021 } 1022 1023 /* 1024 * Conditionally call splz() if gd_reqflags indicates work is pending. 1025 * This will work inside a critical section but not inside a hard code 1026 * section. 1027 * 1028 * (self contained on a per cpu basis) 1029 */ 1030 void 1031 splz_check(void) 1032 { 1033 globaldata_t gd = mycpu; 1034 thread_t td = gd->gd_curthread; 1035 1036 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && 1037 gd->gd_intr_nesting_level == 0 && 1038 td->td_nest_count < 2) 1039 { 1040 splz(); 1041 } 1042 } 1043 1044 /* 1045 * This version is integrated into crit_exit, reqflags has already 1046 * been tested but td_critcount has not. 1047 * 1048 * We only want to execute the splz() on the 1->0 transition of 1049 * critcount and not in a hard code section or if too deeply nested. 1050 */ 1051 void 1052 lwkt_maybe_splz(thread_t td) 1053 { 1054 globaldata_t gd = td->td_gd; 1055 1056 if (td->td_critcount == 0 && 1057 gd->gd_intr_nesting_level == 0 && 1058 td->td_nest_count < 2) 1059 { 1060 splz(); 1061 } 1062 } 1063 1064 /* 1065 * This function is used to negotiate a passive release of the current 1066 * process/lwp designation with the user scheduler, allowing the user 1067 * scheduler to schedule another user thread. The related kernel thread 1068 * (curthread) continues running in the released state. 1069 */ 1070 void 1071 lwkt_passive_release(struct thread *td) 1072 { 1073 struct lwp *lp = td->td_lwp; 1074 1075 td->td_release = NULL; 1076 lwkt_setpri_self(TDPRI_KERN_USER); 1077 lp->lwp_proc->p_usched->release_curproc(lp); 1078 } 1079 1080 1081 /* 1082 * This implements a normal yield. This routine is virtually a nop if 1083 * there is nothing to yield to but it will always run any pending interrupts 1084 * if called from a critical section. 1085 * 1086 * This yield is designed for kernel threads without a user context. 1087 * 1088 * (self contained on a per cpu basis) 1089 */ 1090 void 1091 lwkt_yield(void) 1092 { 1093 globaldata_t gd = mycpu; 1094 thread_t td = gd->gd_curthread; 1095 thread_t xtd; 1096 1097 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1098 splz(); 1099 if (td->td_fairq_accum < 0) { 1100 lwkt_schedule_self(curthread); 1101 lwkt_switch(); 1102 } else { 1103 xtd = TAILQ_FIRST(&gd->gd_tdrunq); 1104 if (xtd && xtd->td_pri > td->td_pri) { 1105 lwkt_schedule_self(curthread); 1106 lwkt_switch(); 1107 } 1108 } 1109 } 1110 1111 /* 1112 * This yield is designed for kernel threads with a user context. 1113 * 1114 * The kernel acting on behalf of the user is potentially cpu-bound, 1115 * this function will efficiently allow other threads to run and also 1116 * switch to other processes by releasing. 1117 * 1118 * The lwkt_user_yield() function is designed to have very low overhead 1119 * if no yield is determined to be needed. 1120 */ 1121 void 1122 lwkt_user_yield(void) 1123 { 1124 globaldata_t gd = mycpu; 1125 thread_t td = gd->gd_curthread; 1126 1127 /* 1128 * Always run any pending interrupts in case we are in a critical 1129 * section. 1130 */ 1131 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1132 splz(); 1133 1134 /* 1135 * Switch (which forces a release) if another kernel thread needs 1136 * the cpu, if userland wants us to resched, or if our kernel 1137 * quantum has run out. 1138 */ 1139 if (lwkt_resched_wanted() || 1140 user_resched_wanted() || 1141 td->td_fairq_accum < 0) 1142 { 1143 lwkt_switch(); 1144 } 1145 1146 #if 0 1147 /* 1148 * Reacquire the current process if we are released. 1149 * 1150 * XXX not implemented atm. The kernel may be holding locks and such, 1151 * so we want the thread to continue to receive cpu. 1152 */ 1153 if (td->td_release == NULL && lp) { 1154 lp->lwp_proc->p_usched->acquire_curproc(lp); 1155 td->td_release = lwkt_passive_release; 1156 lwkt_setpri_self(TDPRI_USER_NORM); 1157 } 1158 #endif 1159 } 1160 1161 /* 1162 * Generic schedule. Possibly schedule threads belonging to other cpus and 1163 * deal with threads that might be blocked on a wait queue. 1164 * 1165 * We have a little helper inline function which does additional work after 1166 * the thread has been enqueued, including dealing with preemption and 1167 * setting need_lwkt_resched() (which prevents the kernel from returning 1168 * to userland until it has processed higher priority threads). 1169 * 1170 * It is possible for this routine to be called after a failed _enqueue 1171 * (due to the target thread migrating, sleeping, or otherwise blocked). 1172 * We have to check that the thread is actually on the run queue! 1173 * 1174 * reschedok is an optimized constant propagated from lwkt_schedule() or 1175 * lwkt_schedule_noresched(). By default it is non-zero, causing a 1176 * reschedule to be requested if the target thread has a higher priority. 1177 * The port messaging code will set MSG_NORESCHED and cause reschedok to 1178 * be 0, prevented undesired reschedules. 1179 */ 1180 static __inline 1181 void 1182 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount, int reschedok) 1183 { 1184 thread_t otd; 1185 1186 if (ntd->td_flags & TDF_RUNQ) { 1187 if (ntd->td_preemptable && reschedok) { 1188 ntd->td_preemptable(ntd, ccount); /* YYY +token */ 1189 } else if (reschedok) { 1190 otd = curthread; 1191 if (ntd->td_pri > otd->td_pri) 1192 need_lwkt_resched(); 1193 } 1194 1195 /* 1196 * Give the thread a little fair share scheduler bump if it 1197 * has been asleep for a while. This is primarily to avoid 1198 * a degenerate case for interrupt threads where accumulator 1199 * crosses into negative territory unnecessarily. 1200 */ 1201 if (ntd->td_fairq_lticks != ticks) { 1202 ntd->td_fairq_lticks = ticks; 1203 ntd->td_fairq_accum += gd->gd_fairq_total_pri; 1204 if (ntd->td_fairq_accum > TDFAIRQ_MAX(gd)) 1205 ntd->td_fairq_accum = TDFAIRQ_MAX(gd); 1206 } 1207 } 1208 } 1209 1210 static __inline 1211 void 1212 _lwkt_schedule(thread_t td, int reschedok) 1213 { 1214 globaldata_t mygd = mycpu; 1215 1216 KASSERT(td != &td->td_gd->gd_idlethread, 1217 ("lwkt_schedule(): scheduling gd_idlethread is illegal!")); 1218 crit_enter_gd(mygd); 1219 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0); 1220 if (td == mygd->gd_curthread) { 1221 _lwkt_enqueue(td); 1222 } else { 1223 /* 1224 * If we own the thread, there is no race (since we are in a 1225 * critical section). If we do not own the thread there might 1226 * be a race but the target cpu will deal with it. 1227 */ 1228 #ifdef SMP 1229 if (td->td_gd == mygd) { 1230 _lwkt_enqueue(td); 1231 _lwkt_schedule_post(mygd, td, 1, reschedok); 1232 } else { 1233 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0); 1234 } 1235 #else 1236 _lwkt_enqueue(td); 1237 _lwkt_schedule_post(mygd, td, 1, reschedok); 1238 #endif 1239 } 1240 crit_exit_gd(mygd); 1241 } 1242 1243 void 1244 lwkt_schedule(thread_t td) 1245 { 1246 _lwkt_schedule(td, 1); 1247 } 1248 1249 void 1250 lwkt_schedule_noresched(thread_t td) 1251 { 1252 _lwkt_schedule(td, 0); 1253 } 1254 1255 #ifdef SMP 1256 1257 /* 1258 * When scheduled remotely if frame != NULL the IPIQ is being 1259 * run via doreti or an interrupt then preemption can be allowed. 1260 * 1261 * To allow preemption we have to drop the critical section so only 1262 * one is present in _lwkt_schedule_post. 1263 */ 1264 static void 1265 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame) 1266 { 1267 thread_t td = curthread; 1268 thread_t ntd = arg; 1269 1270 if (frame && ntd->td_preemptable) { 1271 crit_exit_noyield(td); 1272 _lwkt_schedule(ntd, 1); 1273 crit_enter_quick(td); 1274 } else { 1275 _lwkt_schedule(ntd, 1); 1276 } 1277 } 1278 1279 /* 1280 * Thread migration using a 'Pull' method. The thread may or may not be 1281 * the current thread. It MUST be descheduled and in a stable state. 1282 * lwkt_giveaway() must be called on the cpu owning the thread. 1283 * 1284 * At any point after lwkt_giveaway() is called, the target cpu may 1285 * 'pull' the thread by calling lwkt_acquire(). 1286 * 1287 * We have to make sure the thread is not sitting on a per-cpu tsleep 1288 * queue or it will blow up when it moves to another cpu. 1289 * 1290 * MPSAFE - must be called under very specific conditions. 1291 */ 1292 void 1293 lwkt_giveaway(thread_t td) 1294 { 1295 globaldata_t gd = mycpu; 1296 1297 crit_enter_gd(gd); 1298 if (td->td_flags & TDF_TSLEEPQ) 1299 tsleep_remove(td); 1300 KKASSERT(td->td_gd == gd); 1301 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); 1302 td->td_flags |= TDF_MIGRATING; 1303 crit_exit_gd(gd); 1304 } 1305 1306 void 1307 lwkt_acquire(thread_t td) 1308 { 1309 globaldata_t gd; 1310 globaldata_t mygd; 1311 1312 KKASSERT(td->td_flags & TDF_MIGRATING); 1313 gd = td->td_gd; 1314 mygd = mycpu; 1315 if (gd != mycpu) { 1316 cpu_lfence(); 1317 KKASSERT((td->td_flags & TDF_RUNQ) == 0); 1318 crit_enter_gd(mygd); 1319 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) { 1320 #ifdef SMP 1321 lwkt_process_ipiq(); 1322 #endif 1323 cpu_lfence(); 1324 } 1325 cpu_mfence(); 1326 td->td_gd = mygd; 1327 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1328 td->td_flags &= ~TDF_MIGRATING; 1329 crit_exit_gd(mygd); 1330 } else { 1331 crit_enter_gd(mygd); 1332 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1333 td->td_flags &= ~TDF_MIGRATING; 1334 crit_exit_gd(mygd); 1335 } 1336 } 1337 1338 #endif 1339 1340 /* 1341 * Generic deschedule. Descheduling threads other then your own should be 1342 * done only in carefully controlled circumstances. Descheduling is 1343 * asynchronous. 1344 * 1345 * This function may block if the cpu has run out of messages. 1346 */ 1347 void 1348 lwkt_deschedule(thread_t td) 1349 { 1350 crit_enter(); 1351 #ifdef SMP 1352 if (td == curthread) { 1353 _lwkt_dequeue(td); 1354 } else { 1355 if (td->td_gd == mycpu) { 1356 _lwkt_dequeue(td); 1357 } else { 1358 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td); 1359 } 1360 } 1361 #else 1362 _lwkt_dequeue(td); 1363 #endif 1364 crit_exit(); 1365 } 1366 1367 /* 1368 * Set the target thread's priority. This routine does not automatically 1369 * switch to a higher priority thread, LWKT threads are not designed for 1370 * continuous priority changes. Yield if you want to switch. 1371 */ 1372 void 1373 lwkt_setpri(thread_t td, int pri) 1374 { 1375 KKASSERT(td->td_gd == mycpu); 1376 if (td->td_pri != pri) { 1377 KKASSERT(pri >= 0); 1378 crit_enter(); 1379 if (td->td_flags & TDF_RUNQ) { 1380 _lwkt_dequeue(td); 1381 td->td_pri = pri; 1382 _lwkt_enqueue(td); 1383 } else { 1384 td->td_pri = pri; 1385 } 1386 crit_exit(); 1387 } 1388 } 1389 1390 /* 1391 * Set the initial priority for a thread prior to it being scheduled for 1392 * the first time. The thread MUST NOT be scheduled before or during 1393 * this call. The thread may be assigned to a cpu other then the current 1394 * cpu. 1395 * 1396 * Typically used after a thread has been created with TDF_STOPPREQ, 1397 * and before the thread is initially scheduled. 1398 */ 1399 void 1400 lwkt_setpri_initial(thread_t td, int pri) 1401 { 1402 KKASSERT(pri >= 0); 1403 KKASSERT((td->td_flags & TDF_RUNQ) == 0); 1404 td->td_pri = pri; 1405 } 1406 1407 void 1408 lwkt_setpri_self(int pri) 1409 { 1410 thread_t td = curthread; 1411 1412 KKASSERT(pri >= 0 && pri <= TDPRI_MAX); 1413 crit_enter(); 1414 if (td->td_flags & TDF_RUNQ) { 1415 _lwkt_dequeue(td); 1416 td->td_pri = pri; 1417 _lwkt_enqueue(td); 1418 } else { 1419 td->td_pri = pri; 1420 } 1421 crit_exit(); 1422 } 1423 1424 /* 1425 * 1/hz tick (typically 10ms) x TDFAIRQ_SCALE (typ 8) = 80ms full cycle. 1426 * 1427 * Example: two competing threads, same priority N. decrement by (2*N) 1428 * increment by N*8, each thread will get 4 ticks. 1429 */ 1430 void 1431 lwkt_fairq_schedulerclock(thread_t td) 1432 { 1433 globaldata_t gd; 1434 1435 if (fairq_enable) { 1436 while (td) { 1437 gd = td->td_gd; 1438 if (td != &gd->gd_idlethread) { 1439 td->td_fairq_accum -= gd->gd_fairq_total_pri; 1440 if (td->td_fairq_accum < -TDFAIRQ_MAX(gd)) 1441 td->td_fairq_accum = -TDFAIRQ_MAX(gd); 1442 if (td->td_fairq_accum < 0) 1443 need_lwkt_resched(); 1444 td->td_fairq_lticks = ticks; 1445 } 1446 td = td->td_preempted; 1447 } 1448 } 1449 } 1450 1451 static void 1452 lwkt_fairq_accumulate(globaldata_t gd, thread_t td) 1453 { 1454 td->td_fairq_accum += td->td_pri * TDFAIRQ_SCALE; 1455 if (td->td_fairq_accum > TDFAIRQ_MAX(td->td_gd)) 1456 td->td_fairq_accum = TDFAIRQ_MAX(td->td_gd); 1457 } 1458 1459 /* 1460 * Migrate the current thread to the specified cpu. 1461 * 1462 * This is accomplished by descheduling ourselves from the current cpu, 1463 * moving our thread to the tdallq of the target cpu, IPI messaging the 1464 * target cpu, and switching out. TDF_MIGRATING prevents scheduling 1465 * races while the thread is being migrated. 1466 * 1467 * We must be sure to remove ourselves from the current cpu's tsleepq 1468 * before potentially moving to another queue. The thread can be on 1469 * a tsleepq due to a left-over tsleep_interlock(). 1470 */ 1471 #ifdef SMP 1472 static void lwkt_setcpu_remote(void *arg); 1473 #endif 1474 1475 void 1476 lwkt_setcpu_self(globaldata_t rgd) 1477 { 1478 #ifdef SMP 1479 thread_t td = curthread; 1480 1481 if (td->td_gd != rgd) { 1482 crit_enter_quick(td); 1483 if (td->td_flags & TDF_TSLEEPQ) 1484 tsleep_remove(td); 1485 td->td_flags |= TDF_MIGRATING; 1486 lwkt_deschedule_self(td); 1487 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1488 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td); 1489 lwkt_switch(); 1490 /* we are now on the target cpu */ 1491 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); 1492 crit_exit_quick(td); 1493 } 1494 #endif 1495 } 1496 1497 void 1498 lwkt_migratecpu(int cpuid) 1499 { 1500 #ifdef SMP 1501 globaldata_t rgd; 1502 1503 rgd = globaldata_find(cpuid); 1504 lwkt_setcpu_self(rgd); 1505 #endif 1506 } 1507 1508 /* 1509 * Remote IPI for cpu migration (called while in a critical section so we 1510 * do not have to enter another one). The thread has already been moved to 1511 * our cpu's allq, but we must wait for the thread to be completely switched 1512 * out on the originating cpu before we schedule it on ours or the stack 1513 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD 1514 * change to main memory. 1515 * 1516 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races 1517 * against wakeups. It is best if this interface is used only when there 1518 * are no pending events that might try to schedule the thread. 1519 */ 1520 #ifdef SMP 1521 static void 1522 lwkt_setcpu_remote(void *arg) 1523 { 1524 thread_t td = arg; 1525 globaldata_t gd = mycpu; 1526 1527 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) { 1528 #ifdef SMP 1529 lwkt_process_ipiq(); 1530 #endif 1531 cpu_lfence(); 1532 cpu_pause(); 1533 } 1534 td->td_gd = gd; 1535 cpu_mfence(); 1536 td->td_flags &= ~TDF_MIGRATING; 1537 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0); 1538 _lwkt_enqueue(td); 1539 } 1540 #endif 1541 1542 struct lwp * 1543 lwkt_preempted_proc(void) 1544 { 1545 thread_t td = curthread; 1546 while (td->td_preempted) 1547 td = td->td_preempted; 1548 return(td->td_lwp); 1549 } 1550 1551 /* 1552 * Create a kernel process/thread/whatever. It shares it's address space 1553 * with proc0 - ie: kernel only. 1554 * 1555 * NOTE! By default new threads are created with the MP lock held. A 1556 * thread which does not require the MP lock should release it by calling 1557 * rel_mplock() at the start of the new thread. 1558 */ 1559 int 1560 lwkt_create(void (*func)(void *), void *arg, struct thread **tdp, 1561 thread_t template, int tdflags, int cpu, const char *fmt, ...) 1562 { 1563 thread_t td; 1564 __va_list ap; 1565 1566 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu, 1567 tdflags); 1568 if (tdp) 1569 *tdp = td; 1570 cpu_set_thread_handler(td, lwkt_exit, func, arg); 1571 1572 /* 1573 * Set up arg0 for 'ps' etc 1574 */ 1575 __va_start(ap, fmt); 1576 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap); 1577 __va_end(ap); 1578 1579 /* 1580 * Schedule the thread to run 1581 */ 1582 if ((td->td_flags & TDF_STOPREQ) == 0) 1583 lwkt_schedule(td); 1584 else 1585 td->td_flags &= ~TDF_STOPREQ; 1586 return 0; 1587 } 1588 1589 /* 1590 * Destroy an LWKT thread. Warning! This function is not called when 1591 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and 1592 * uses a different reaping mechanism. 1593 */ 1594 void 1595 lwkt_exit(void) 1596 { 1597 thread_t td = curthread; 1598 thread_t std; 1599 globaldata_t gd; 1600 1601 /* 1602 * Do any cleanup that might block here 1603 */ 1604 if (td->td_flags & TDF_VERBOSE) 1605 kprintf("kthread %p %s has exited\n", td, td->td_comm); 1606 caps_exit(td); 1607 biosched_done(td); 1608 dsched_exit_thread(td); 1609 1610 /* 1611 * Get us into a critical section to interlock gd_freetd and loop 1612 * until we can get it freed. 1613 * 1614 * We have to cache the current td in gd_freetd because objcache_put()ing 1615 * it would rip it out from under us while our thread is still active. 1616 */ 1617 gd = mycpu; 1618 crit_enter_quick(td); 1619 while ((std = gd->gd_freetd) != NULL) { 1620 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0); 1621 gd->gd_freetd = NULL; 1622 objcache_put(thread_cache, std); 1623 } 1624 1625 /* 1626 * Remove thread resources from kernel lists and deschedule us for 1627 * the last time. We cannot block after this point or we may end 1628 * up with a stale td on the tsleepq. 1629 */ 1630 if (td->td_flags & TDF_TSLEEPQ) 1631 tsleep_remove(td); 1632 lwkt_deschedule_self(td); 1633 lwkt_remove_tdallq(td); 1634 1635 /* 1636 * Final cleanup 1637 */ 1638 KKASSERT(gd->gd_freetd == NULL); 1639 if (td->td_flags & TDF_ALLOCATED_THREAD) 1640 gd->gd_freetd = td; 1641 cpu_thread_exit(); 1642 } 1643 1644 void 1645 lwkt_remove_tdallq(thread_t td) 1646 { 1647 KKASSERT(td->td_gd == mycpu); 1648 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1649 } 1650 1651 /* 1652 * Code reduction and branch prediction improvements. Call/return 1653 * overhead on modern cpus often degenerates into 0 cycles due to 1654 * the cpu's branch prediction hardware and return pc cache. We 1655 * can take advantage of this by not inlining medium-complexity 1656 * functions and we can also reduce the branch prediction impact 1657 * by collapsing perfectly predictable branches into a single 1658 * procedure instead of duplicating it. 1659 * 1660 * Is any of this noticeable? Probably not, so I'll take the 1661 * smaller code size. 1662 */ 1663 void 1664 crit_exit_wrapper(__DEBUG_CRIT_ARG__) 1665 { 1666 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__); 1667 } 1668 1669 void 1670 crit_panic(void) 1671 { 1672 thread_t td = curthread; 1673 int lcrit = td->td_critcount; 1674 1675 td->td_critcount = 0; 1676 panic("td_critcount is/would-go negative! %p %d", td, lcrit); 1677 /* NOT REACHED */ 1678 } 1679 1680 #ifdef SMP 1681 1682 /* 1683 * Called from debugger/panic on cpus which have been stopped. We must still 1684 * process the IPIQ while stopped, even if we were stopped while in a critical 1685 * section (XXX). 1686 * 1687 * If we are dumping also try to process any pending interrupts. This may 1688 * or may not work depending on the state of the cpu at the point it was 1689 * stopped. 1690 */ 1691 void 1692 lwkt_smp_stopped(void) 1693 { 1694 globaldata_t gd = mycpu; 1695 1696 crit_enter_gd(gd); 1697 if (dumping) { 1698 lwkt_process_ipiq(); 1699 splz(); 1700 } else { 1701 lwkt_process_ipiq(); 1702 } 1703 crit_exit_gd(gd); 1704 } 1705 1706 #endif 1707