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, &fairq_enable, 0, 134 "Turn on fairq priority accumulators"); 135 static int user_pri_sched = 0; 136 SYSCTL_INT(_lwkt, OID_AUTO, user_pri_sched, CTLFLAG_RW, &user_pri_sched, 0, 137 ""); 138 static int preempt_enable = 1; 139 SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW, &preempt_enable, 0, 140 "Enable preemption"); 141 142 143 /* 144 * These helper procedures handle the runq, they can only be called from 145 * within a critical section. 146 * 147 * WARNING! Prior to SMP being brought up it is possible to enqueue and 148 * dequeue threads belonging to other cpus, so be sure to use td->td_gd 149 * instead of 'mycpu' when referencing the globaldata structure. Once 150 * SMP live enqueuing and dequeueing only occurs on the current cpu. 151 */ 152 static __inline 153 void 154 _lwkt_dequeue(thread_t td) 155 { 156 if (td->td_flags & TDF_RUNQ) { 157 struct globaldata *gd = td->td_gd; 158 159 td->td_flags &= ~TDF_RUNQ; 160 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq); 161 gd->gd_fairq_total_pri -= td->td_pri; 162 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL) 163 atomic_clear_int_nonlocked(&gd->gd_reqflags, RQF_RUNNING); 164 } 165 } 166 167 /* 168 * Priority enqueue. 169 * 170 * NOTE: There are a limited number of lwkt threads runnable since user 171 * processes only schedule one at a time per cpu. 172 */ 173 static __inline 174 void 175 _lwkt_enqueue(thread_t td) 176 { 177 thread_t xtd; 178 179 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) { 180 struct globaldata *gd = td->td_gd; 181 182 td->td_flags |= TDF_RUNQ; 183 xtd = TAILQ_FIRST(&gd->gd_tdrunq); 184 if (xtd == NULL) { 185 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 186 atomic_set_int_nonlocked(&gd->gd_reqflags, RQF_RUNNING); 187 } else { 188 while (xtd && xtd->td_pri > td->td_pri) 189 xtd = TAILQ_NEXT(xtd, td_threadq); 190 if (xtd) 191 TAILQ_INSERT_BEFORE(xtd, td, td_threadq); 192 else 193 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 194 } 195 gd->gd_fairq_total_pri += td->td_pri; 196 } 197 } 198 199 static __boolean_t 200 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags) 201 { 202 struct thread *td = (struct thread *)obj; 203 204 td->td_kstack = NULL; 205 td->td_kstack_size = 0; 206 td->td_flags = TDF_ALLOCATED_THREAD; 207 return (1); 208 } 209 210 static void 211 _lwkt_thread_dtor(void *obj, void *privdata) 212 { 213 struct thread *td = (struct thread *)obj; 214 215 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD, 216 ("_lwkt_thread_dtor: not allocated from objcache")); 217 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack && 218 td->td_kstack_size > 0, 219 ("_lwkt_thread_dtor: corrupted stack")); 220 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size); 221 } 222 223 /* 224 * Initialize the lwkt s/system. 225 */ 226 void 227 lwkt_init(void) 228 { 229 /* An objcache has 2 magazines per CPU so divide cache size by 2. */ 230 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread), 231 NULL, CACHE_NTHREADS/2, 232 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL); 233 } 234 235 /* 236 * Schedule a thread to run. As the current thread we can always safely 237 * schedule ourselves, and a shortcut procedure is provided for that 238 * function. 239 * 240 * (non-blocking, self contained on a per cpu basis) 241 */ 242 void 243 lwkt_schedule_self(thread_t td) 244 { 245 crit_enter_quick(td); 246 KASSERT(td != &td->td_gd->gd_idlethread, 247 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!")); 248 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0); 249 _lwkt_enqueue(td); 250 crit_exit_quick(td); 251 } 252 253 /* 254 * Deschedule a thread. 255 * 256 * (non-blocking, self contained on a per cpu basis) 257 */ 258 void 259 lwkt_deschedule_self(thread_t td) 260 { 261 crit_enter_quick(td); 262 _lwkt_dequeue(td); 263 crit_exit_quick(td); 264 } 265 266 /* 267 * LWKTs operate on a per-cpu basis 268 * 269 * WARNING! Called from early boot, 'mycpu' may not work yet. 270 */ 271 void 272 lwkt_gdinit(struct globaldata *gd) 273 { 274 TAILQ_INIT(&gd->gd_tdrunq); 275 TAILQ_INIT(&gd->gd_tdallq); 276 } 277 278 /* 279 * Create a new thread. The thread must be associated with a process context 280 * or LWKT start address before it can be scheduled. If the target cpu is 281 * -1 the thread will be created on the current cpu. 282 * 283 * If you intend to create a thread without a process context this function 284 * does everything except load the startup and switcher function. 285 */ 286 thread_t 287 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags) 288 { 289 globaldata_t gd = mycpu; 290 void *stack; 291 292 /* 293 * If static thread storage is not supplied allocate a thread. Reuse 294 * a cached free thread if possible. gd_freetd is used to keep an exiting 295 * thread intact through the exit. 296 */ 297 if (td == NULL) { 298 crit_enter_gd(gd); 299 if ((td = gd->gd_freetd) != NULL) { 300 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK| 301 TDF_RUNQ)) == 0); 302 gd->gd_freetd = NULL; 303 } else { 304 td = objcache_get(thread_cache, M_WAITOK); 305 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK| 306 TDF_RUNQ)) == 0); 307 } 308 crit_exit_gd(gd); 309 KASSERT((td->td_flags & 310 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD, 311 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags)); 312 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK); 313 } 314 315 /* 316 * Try to reuse cached stack. 317 */ 318 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) { 319 if (flags & TDF_ALLOCATED_STACK) { 320 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size); 321 stack = NULL; 322 } 323 } 324 if (stack == NULL) { 325 stack = (void *)kmem_alloc_stack(&kernel_map, stksize); 326 flags |= TDF_ALLOCATED_STACK; 327 } 328 if (cpu < 0) 329 lwkt_init_thread(td, stack, stksize, flags, gd); 330 else 331 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu)); 332 return(td); 333 } 334 335 /* 336 * Initialize a preexisting thread structure. This function is used by 337 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread. 338 * 339 * All threads start out in a critical section at a priority of 340 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as 341 * appropriate. This function may send an IPI message when the 342 * requested cpu is not the current cpu and consequently gd_tdallq may 343 * not be initialized synchronously from the point of view of the originating 344 * cpu. 345 * 346 * NOTE! we have to be careful in regards to creating threads for other cpus 347 * if SMP has not yet been activated. 348 */ 349 #ifdef SMP 350 351 static void 352 lwkt_init_thread_remote(void *arg) 353 { 354 thread_t td = arg; 355 356 /* 357 * Protected by critical section held by IPI dispatch 358 */ 359 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq); 360 } 361 362 #endif 363 364 /* 365 * lwkt core thread structural initialization. 366 * 367 * NOTE: All threads are initialized as mpsafe threads. 368 */ 369 void 370 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags, 371 struct globaldata *gd) 372 { 373 globaldata_t mygd = mycpu; 374 375 bzero(td, sizeof(struct thread)); 376 td->td_kstack = stack; 377 td->td_kstack_size = stksize; 378 td->td_flags = flags; 379 td->td_gd = gd; 380 td->td_pri = TDPRI_KERN_DAEMON; 381 td->td_critcount = 1; 382 td->td_toks_stop = &td->td_toks_base; 383 if (lwkt_use_spin_port) 384 lwkt_initport_spin(&td->td_msgport); 385 else 386 lwkt_initport_thread(&td->td_msgport, td); 387 pmap_init_thread(td); 388 #ifdef SMP 389 /* 390 * Normally initializing a thread for a remote cpu requires sending an 391 * IPI. However, the idlethread is setup before the other cpus are 392 * activated so we have to treat it as a special case. XXX manipulation 393 * of gd_tdallq requires the BGL. 394 */ 395 if (gd == mygd || td == &gd->gd_idlethread) { 396 crit_enter_gd(mygd); 397 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); 398 crit_exit_gd(mygd); 399 } else { 400 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td); 401 } 402 #else 403 crit_enter_gd(mygd); 404 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); 405 crit_exit_gd(mygd); 406 #endif 407 408 dsched_new_thread(td); 409 } 410 411 void 412 lwkt_set_comm(thread_t td, const char *ctl, ...) 413 { 414 __va_list va; 415 416 __va_start(va, ctl); 417 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va); 418 __va_end(va); 419 KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]); 420 } 421 422 void 423 lwkt_hold(thread_t td) 424 { 425 ++td->td_refs; 426 } 427 428 void 429 lwkt_rele(thread_t td) 430 { 431 KKASSERT(td->td_refs > 0); 432 --td->td_refs; 433 } 434 435 void 436 lwkt_wait_free(thread_t td) 437 { 438 while (td->td_refs) 439 tsleep(td, 0, "tdreap", hz); 440 } 441 442 void 443 lwkt_free_thread(thread_t td) 444 { 445 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|TDF_RUNQ)) == 0); 446 if (td->td_flags & TDF_ALLOCATED_THREAD) { 447 objcache_put(thread_cache, td); 448 } else if (td->td_flags & TDF_ALLOCATED_STACK) { 449 /* client-allocated struct with internally allocated stack */ 450 KASSERT(td->td_kstack && td->td_kstack_size > 0, 451 ("lwkt_free_thread: corrupted stack")); 452 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size); 453 td->td_kstack = NULL; 454 td->td_kstack_size = 0; 455 } 456 KTR_LOG(ctxsw_deadtd, td); 457 } 458 459 460 /* 461 * Switch to the next runnable lwkt. If no LWKTs are runnable then 462 * switch to the idlethread. Switching must occur within a critical 463 * section to avoid races with the scheduling queue. 464 * 465 * We always have full control over our cpu's run queue. Other cpus 466 * that wish to manipulate our queue must use the cpu_*msg() calls to 467 * talk to our cpu, so a critical section is all that is needed and 468 * the result is very, very fast thread switching. 469 * 470 * The LWKT scheduler uses a fixed priority model and round-robins at 471 * each priority level. User process scheduling is a totally 472 * different beast and LWKT priorities should not be confused with 473 * user process priorities. 474 * 475 * The MP lock may be out of sync with the thread's td_mpcount + td_xpcount. 476 * lwkt_switch() cleans it up. 477 * 478 * Note that the td_switch() function cannot do anything that requires 479 * the MP lock since the MP lock will have already been setup for 480 * the target thread (not the current thread). It's nice to have a scheduler 481 * that does not need the MP lock to work because it allows us to do some 482 * really cool high-performance MP lock optimizations. 483 * 484 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch() 485 * is not called by the current thread in the preemption case, only when 486 * the preempting thread blocks (in order to return to the original thread). 487 */ 488 void 489 lwkt_switch(void) 490 { 491 globaldata_t gd = mycpu; 492 thread_t td = gd->gd_curthread; 493 thread_t ntd; 494 thread_t xtd; 495 thread_t nlast; 496 int nquserok; 497 #ifdef SMP 498 int mpheld; 499 #endif 500 int didaccumulate; 501 const char *lmsg; /* diagnostic - 'systat -pv 1' */ 502 const void *laddr; 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 /* 566 * td_mpcount + td_xpcount cannot be used to determine if we currently 567 * hold the MP lock because get_mplock() will increment it prior to 568 * attempting to get the lock, and switch out if it can't. Our 569 * ownership of the actual lock will remain stable while we are 570 * in a critical section, and once we actually acquire the underlying 571 * lock as long as the count is greater than 0. 572 */ 573 mpheld = MP_LOCK_HELD(gd); 574 #ifdef INVARIANTS 575 if (td->td_cscount) { 576 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n", 577 td); 578 if (panic_on_cscount) 579 panic("switching while mastering cpusync"); 580 } 581 #endif 582 #endif 583 584 /* 585 * If we had preempted another thread on this cpu, resume the preempted 586 * thread. This occurs transparently, whether the preempted thread 587 * was scheduled or not (it may have been preempted after descheduling 588 * itself). 589 * 590 * We have to setup the MP lock for the original thread after backing 591 * out the adjustment that was made to curthread when the original 592 * was preempted. 593 */ 594 if ((ntd = td->td_preempted) != NULL) { 595 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK); 596 #ifdef SMP 597 if (ntd->td_mpcount + ntd->td_xpcount && mpheld == 0) { 598 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d", 599 td, ntd, td->td_mpcount, ntd->td_mpcount + ntd->td_xpcount); 600 } 601 td->td_xpcount = 0; 602 #endif 603 ntd->td_flags |= TDF_PREEMPT_DONE; 604 605 /* 606 * The interrupt may have woken a thread up, we need to properly 607 * set the reschedule flag if the originally interrupted thread is 608 * at a lower priority. 609 */ 610 if (TAILQ_FIRST(&gd->gd_tdrunq) && 611 TAILQ_FIRST(&gd->gd_tdrunq)->td_pri > ntd->td_pri) { 612 need_lwkt_resched(); 613 } 614 /* YYY release mp lock on switchback if original doesn't need it */ 615 goto havethread_preempted; 616 } 617 618 /* 619 * Implement round-robin fairq with priority insertion. The priority 620 * insertion is handled by _lwkt_enqueue() 621 * 622 * We have to adjust the MP lock for the target thread. If we 623 * need the MP lock and cannot obtain it we try to locate a 624 * thread that does not need the MP lock. If we cannot, we spin 625 * instead of HLT. 626 * 627 * A similar issue exists for the tokens held by the target thread. 628 * If we cannot obtain ownership of the tokens we cannot immediately 629 * schedule the thread. 630 */ 631 for (;;) { 632 clear_lwkt_resched(); 633 didaccumulate = 0; 634 ntd = TAILQ_FIRST(&gd->gd_tdrunq); 635 636 /* 637 * Hotpath if we can get all necessary resources. 638 * 639 * If nothing is runnable switch to the idle thread 640 */ 641 if (ntd == NULL) { 642 ntd = &gd->gd_idlethread; 643 if (gd->gd_reqflags & RQF_IDLECHECK_MASK) 644 ntd->td_flags |= TDF_IDLE_NOHLT; 645 #ifdef SMP 646 KKASSERT(ntd->td_xpcount == 0); 647 if (ntd->td_mpcount) { 648 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) 649 panic("Idle thread %p was holding the BGL!", ntd); 650 if (mpheld == 0) { 651 set_cpu_contention_mask(gd); 652 handle_cpu_contention_mask(); 653 cpu_try_mplock(); 654 mpheld = MP_LOCK_HELD(gd); 655 cpu_pause(); 656 continue; 657 } 658 } 659 clr_cpu_contention_mask(gd); 660 #endif 661 cpu_time.cp_msg[0] = 0; 662 cpu_time.cp_stallpc = 0; 663 goto haveidle; 664 } 665 666 /* 667 * Hotpath schedule 668 * 669 * NOTE: For UP there is no mplock and lwkt_getalltokens() 670 * always succeeds. 671 */ 672 if (ntd->td_fairq_accum >= 0 && 673 #ifdef SMP 674 (ntd->td_mpcount + ntd->td_xpcount == 0 || 675 mpheld || cpu_try_mplock()) && 676 #endif 677 (!TD_TOKS_HELD(ntd) || lwkt_getalltokens(ntd, &lmsg, &laddr)) 678 ) { 679 #ifdef SMP 680 clr_cpu_contention_mask(gd); 681 #endif 682 goto havethread; 683 } 684 685 lmsg = NULL; 686 laddr = NULL; 687 688 #ifdef SMP 689 if (ntd->td_fairq_accum >= 0) 690 set_cpu_contention_mask(gd); 691 /* Reload mpheld (it become stale after mplock/token ops) */ 692 mpheld = MP_LOCK_HELD(gd); 693 if (ntd->td_mpcount + ntd->td_xpcount && mpheld == 0) { 694 lmsg = "mplock"; 695 laddr = ntd->td_mplock_stallpc; 696 } 697 #endif 698 699 /* 700 * Coldpath - unable to schedule ntd, continue looking for threads 701 * to schedule. This is only allowed of the (presumably) kernel 702 * thread exhausted its fair share. A kernel thread stuck on 703 * resources does not currently allow a user thread to get in 704 * front of it. 705 */ 706 #ifdef SMP 707 nquserok = ((ntd->td_pri < TDPRI_KERN_LPSCHED) || 708 (ntd->td_fairq_accum < 0)); 709 #else 710 nquserok = 1; 711 #endif 712 nlast = NULL; 713 714 for (;;) { 715 /* 716 * If the fair-share scheduler ran out ntd gets moved to the 717 * end and its accumulator will be bumped, if it didn't we 718 * maintain the same queue position. 719 * 720 * nlast keeps track of the last element prior to any moves. 721 */ 722 if (ntd->td_fairq_accum < 0) { 723 lwkt_fairq_accumulate(gd, ntd); 724 didaccumulate = 1; 725 726 /* 727 * Move to end 728 */ 729 xtd = TAILQ_NEXT(ntd, td_threadq); 730 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq); 731 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, ntd, td_threadq); 732 733 /* 734 * Set terminal element (nlast) 735 */ 736 if (nlast == NULL) { 737 nlast = ntd; 738 if (xtd == NULL) 739 xtd = ntd; 740 } 741 ntd = xtd; 742 } else { 743 ntd = TAILQ_NEXT(ntd, td_threadq); 744 } 745 746 /* 747 * If we exhausted the run list switch to the idle thread. 748 * Since one or more threads had resource acquisition issues 749 * we do not allow the idle thread to halt. 750 * 751 * NOTE: nlast can be NULL. 752 */ 753 if (ntd == nlast) { 754 cpu_pause(); 755 ntd = &gd->gd_idlethread; 756 ntd->td_flags |= TDF_IDLE_NOHLT; 757 #ifdef SMP 758 KKASSERT(ntd->td_xpcount == 0); 759 if (ntd->td_mpcount) { 760 mpheld = MP_LOCK_HELD(gd); 761 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) 762 panic("Idle thread %p was holding the BGL!", ntd); 763 if (mpheld == 0) { 764 set_cpu_contention_mask(gd); 765 handle_cpu_contention_mask(); 766 cpu_try_mplock(); 767 mpheld = MP_LOCK_HELD(gd); 768 cpu_pause(); 769 break; /* try again from the top, almost */ 770 } 771 } 772 #endif 773 774 /* 775 * If fairq accumulations occured we do not schedule the 776 * idle thread. This will cause us to try again from 777 * the (almost) top. 778 */ 779 if (didaccumulate) 780 break; /* try again from the top, almost */ 781 if (lmsg) 782 strlcpy(cpu_time.cp_msg, lmsg, sizeof(cpu_time.cp_msg)); 783 cpu_time.cp_stallpc = (uintptr_t)laddr; 784 goto haveidle; 785 } 786 787 /* 788 * Try to switch to this thread. 789 * 790 * NOTE: For UP there is no mplock and lwkt_getalltokens() 791 * always succeeds. 792 */ 793 if ((ntd->td_pri >= TDPRI_KERN_LPSCHED || nquserok || 794 user_pri_sched) && ntd->td_fairq_accum >= 0 && 795 #ifdef SMP 796 (ntd->td_mpcount + ntd->td_xpcount == 0 || 797 mpheld || cpu_try_mplock()) && 798 #endif 799 (!TD_TOKS_HELD(ntd) || lwkt_getalltokens(ntd, &lmsg, &laddr)) 800 ) { 801 #ifdef SMP 802 clr_cpu_contention_mask(gd); 803 #endif 804 goto havethread; 805 } 806 #ifdef SMP 807 if (ntd->td_fairq_accum >= 0) 808 set_cpu_contention_mask(gd); 809 /* 810 * Reload mpheld (it become stale after mplock/token ops). 811 */ 812 mpheld = MP_LOCK_HELD(gd); 813 if (ntd->td_mpcount + ntd->td_xpcount && mpheld == 0) { 814 lmsg = "mplock"; 815 laddr = ntd->td_mplock_stallpc; 816 } 817 if (ntd->td_pri >= TDPRI_KERN_LPSCHED && ntd->td_fairq_accum >= 0) 818 nquserok = 0; 819 #endif 820 } 821 822 /* 823 * All threads exhausted but we can loop due to a negative 824 * accumulator. 825 * 826 * While we are looping in the scheduler be sure to service 827 * any interrupts which were made pending due to our critical 828 * section, otherwise we could livelock (e.g.) IPIs. 829 * 830 * NOTE: splz can enter and exit the mplock so mpheld is 831 * stale after this call. 832 */ 833 splz_check(); 834 835 #ifdef SMP 836 /* 837 * Our mplock can be cached and cause other cpus to livelock 838 * if we loop due to e.g. not being able to acquire tokens. 839 */ 840 if (MP_LOCK_HELD(gd)) 841 cpu_rel_mplock(gd->gd_cpuid); 842 mpheld = 0; 843 #endif 844 } 845 846 /* 847 * Do the actual switch. WARNING: mpheld is stale here. 848 * 849 * We must always decrement td_fairq_accum on non-idle threads just 850 * in case a thread never gets a tick due to being in a continuous 851 * critical section. The page-zeroing code does that. 852 * 853 * If the thread we came up with is a higher or equal priority verses 854 * the thread at the head of the queue we move our thread to the 855 * front. This way we can always check the front of the queue. 856 */ 857 havethread: 858 ++gd->gd_cnt.v_swtch; 859 --ntd->td_fairq_accum; 860 xtd = TAILQ_FIRST(&gd->gd_tdrunq); 861 if (ntd != xtd && ntd->td_pri >= xtd->td_pri) { 862 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq); 863 TAILQ_INSERT_HEAD(&gd->gd_tdrunq, ntd, td_threadq); 864 } 865 havethread_preempted: 866 ; 867 /* 868 * If the new target does not need the MP lock and we are holding it, 869 * release the MP lock. If the new target requires the MP lock we have 870 * already acquired it for the target. 871 * 872 * WARNING: mpheld is stale here. 873 */ 874 haveidle: 875 KASSERT(ntd->td_critcount, 876 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri)); 877 #ifdef SMP 878 if (ntd->td_mpcount + ntd->td_xpcount == 0 ) { 879 if (MP_LOCK_HELD(gd)) 880 cpu_rel_mplock(gd->gd_cpuid); 881 } else { 882 ASSERT_MP_LOCK_HELD(ntd); 883 } 884 #endif 885 if (td != ntd) { 886 ++switch_count; 887 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd); 888 td->td_switch(ntd); 889 } 890 /* NOTE: current cpu may have changed after switch */ 891 crit_exit_quick(td); 892 } 893 894 /* 895 * Request that the target thread preempt the current thread. Preemption 896 * only works under a specific set of conditions: 897 * 898 * - We are not preempting ourselves 899 * - The target thread is owned by the current cpu 900 * - We are not currently being preempted 901 * - The target is not currently being preempted 902 * - We are not holding any spin locks 903 * - The target thread is not holding any tokens 904 * - We are able to satisfy the target's MP lock requirements (if any). 905 * 906 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically 907 * this is called via lwkt_schedule() through the td_preemptable callback. 908 * critcount is the managed critical priority that we should ignore in order 909 * to determine whether preemption is possible (aka usually just the crit 910 * priority of lwkt_schedule() itself). 911 * 912 * XXX at the moment we run the target thread in a critical section during 913 * the preemption in order to prevent the target from taking interrupts 914 * that *WE* can't. Preemption is strictly limited to interrupt threads 915 * and interrupt-like threads, outside of a critical section, and the 916 * preempted source thread will be resumed the instant the target blocks 917 * whether or not the source is scheduled (i.e. preemption is supposed to 918 * be as transparent as possible). 919 * 920 * The target thread inherits our MP count (added to its own) for the 921 * duration of the preemption in order to preserve the atomicy of the 922 * MP lock during the preemption. Therefore, any preempting targets must be 923 * careful in regards to MP assertions. Note that the MP count may be 924 * out of sync with the physical mp_lock, but we do not have to preserve 925 * the original ownership of the lock if it was out of synch (that is, we 926 * can leave it synchronized on return). 927 */ 928 void 929 lwkt_preempt(thread_t ntd, int critcount) 930 { 931 struct globaldata *gd = mycpu; 932 thread_t td; 933 #ifdef SMP 934 int mpheld; 935 int savecnt; 936 #endif 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 #ifdef SMP 1001 /* 1002 * NOTE: An interrupt might have occured just as we were transitioning 1003 * to or from the MP lock. In this case td_mpcount will be pre-disposed 1004 * (non-zero) but not actually synchronized with the mp_lock itself. 1005 * We can use it to imply an MP lock requirement for the preemption but 1006 * we cannot use it to test whether we hold the MP lock or not. 1007 */ 1008 savecnt = td->td_mpcount; 1009 mpheld = MP_LOCK_HELD(gd); 1010 ntd->td_xpcount = td->td_mpcount + td->td_xpcount; 1011 if (mpheld == 0 && ntd->td_mpcount + ntd->td_xpcount && !cpu_try_mplock()) { 1012 ntd->td_xpcount = 0; 1013 ++preempt_miss; 1014 need_lwkt_resched(); 1015 return; 1016 } 1017 #endif 1018 1019 /* 1020 * Since we are able to preempt the current thread, there is no need to 1021 * call need_lwkt_resched(). 1022 * 1023 * We must temporarily clear gd_intr_nesting_level around the switch 1024 * since switchouts from the target thread are allowed (they will just 1025 * return to our thread), and since the target thread has its own stack. 1026 */ 1027 ++preempt_hit; 1028 ntd->td_preempted = td; 1029 td->td_flags |= TDF_PREEMPT_LOCK; 1030 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd); 1031 save_gd_intr_nesting_level = gd->gd_intr_nesting_level; 1032 gd->gd_intr_nesting_level = 0; 1033 td->td_switch(ntd); 1034 gd->gd_intr_nesting_level = save_gd_intr_nesting_level; 1035 1036 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE)); 1037 #ifdef SMP 1038 KKASSERT(savecnt == td->td_mpcount); 1039 mpheld = MP_LOCK_HELD(gd); 1040 if (mpheld && td->td_mpcount == 0) 1041 cpu_rel_mplock(gd->gd_cpuid); 1042 else if (mpheld == 0 && td->td_mpcount + td->td_xpcount) 1043 panic("lwkt_preempt(): MP lock was not held through"); 1044 #endif 1045 ntd->td_preempted = NULL; 1046 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE); 1047 } 1048 1049 /* 1050 * Conditionally call splz() if gd_reqflags indicates work is pending. 1051 * This will work inside a critical section but not inside a hard code 1052 * section. 1053 * 1054 * (self contained on a per cpu basis) 1055 */ 1056 void 1057 splz_check(void) 1058 { 1059 globaldata_t gd = mycpu; 1060 thread_t td = gd->gd_curthread; 1061 1062 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && 1063 gd->gd_intr_nesting_level == 0 && 1064 td->td_nest_count < 2) 1065 { 1066 splz(); 1067 } 1068 } 1069 1070 /* 1071 * This version is integrated into crit_exit, reqflags has already 1072 * been tested but td_critcount has not. 1073 * 1074 * We only want to execute the splz() on the 1->0 transition of 1075 * critcount and not in a hard code section or if too deeply nested. 1076 */ 1077 void 1078 lwkt_maybe_splz(thread_t td) 1079 { 1080 globaldata_t gd = td->td_gd; 1081 1082 if (td->td_critcount == 0 && 1083 gd->gd_intr_nesting_level == 0 && 1084 td->td_nest_count < 2) 1085 { 1086 splz(); 1087 } 1088 } 1089 1090 /* 1091 * This function is used to negotiate a passive release of the current 1092 * process/lwp designation with the user scheduler, allowing the user 1093 * scheduler to schedule another user thread. The related kernel thread 1094 * (curthread) continues running in the released state. 1095 */ 1096 void 1097 lwkt_passive_release(struct thread *td) 1098 { 1099 struct lwp *lp = td->td_lwp; 1100 1101 td->td_release = NULL; 1102 lwkt_setpri_self(TDPRI_KERN_USER); 1103 lp->lwp_proc->p_usched->release_curproc(lp); 1104 } 1105 1106 1107 /* 1108 * This implements a normal yield. This routine is virtually a nop if 1109 * there is nothing to yield to but it will always run any pending interrupts 1110 * if called from a critical section. 1111 * 1112 * This yield is designed for kernel threads without a user context. 1113 * 1114 * (self contained on a per cpu basis) 1115 */ 1116 void 1117 lwkt_yield(void) 1118 { 1119 globaldata_t gd = mycpu; 1120 thread_t td = gd->gd_curthread; 1121 thread_t xtd; 1122 1123 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1124 splz(); 1125 if (td->td_fairq_accum < 0) { 1126 lwkt_schedule_self(curthread); 1127 lwkt_switch(); 1128 } else { 1129 xtd = TAILQ_FIRST(&gd->gd_tdrunq); 1130 if (xtd && xtd->td_pri > td->td_pri) { 1131 lwkt_schedule_self(curthread); 1132 lwkt_switch(); 1133 } 1134 } 1135 } 1136 1137 /* 1138 * This yield is designed for kernel threads with a user context. 1139 * 1140 * The kernel acting on behalf of the user is potentially cpu-bound, 1141 * this function will efficiently allow other threads to run and also 1142 * switch to other processes by releasing. 1143 * 1144 * The lwkt_user_yield() function is designed to have very low overhead 1145 * if no yield is determined to be needed. 1146 */ 1147 void 1148 lwkt_user_yield(void) 1149 { 1150 globaldata_t gd = mycpu; 1151 thread_t td = gd->gd_curthread; 1152 1153 /* 1154 * Always run any pending interrupts in case we are in a critical 1155 * section. 1156 */ 1157 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1158 splz(); 1159 1160 #ifdef SMP 1161 /* 1162 * XXX SEVERE TEMPORARY HACK. A cpu-bound operation running in the 1163 * kernel can prevent other cpus from servicing interrupt threads 1164 * which still require the MP lock (which is a lot of them). This 1165 * has a chaining effect since if the interrupt is blocked, so is 1166 * the event, so normal scheduling will not pick up on the problem. 1167 */ 1168 if (cpu_contention_mask && td->td_mpcount + td->td_xpcount) { 1169 yield_mplock(td); 1170 } 1171 #endif 1172 1173 /* 1174 * Switch (which forces a release) if another kernel thread needs 1175 * the cpu, if userland wants us to resched, or if our kernel 1176 * quantum has run out. 1177 */ 1178 if (lwkt_resched_wanted() || 1179 user_resched_wanted() || 1180 td->td_fairq_accum < 0) 1181 { 1182 lwkt_switch(); 1183 } 1184 1185 #if 0 1186 /* 1187 * Reacquire the current process if we are released. 1188 * 1189 * XXX not implemented atm. The kernel may be holding locks and such, 1190 * so we want the thread to continue to receive cpu. 1191 */ 1192 if (td->td_release == NULL && lp) { 1193 lp->lwp_proc->p_usched->acquire_curproc(lp); 1194 td->td_release = lwkt_passive_release; 1195 lwkt_setpri_self(TDPRI_USER_NORM); 1196 } 1197 #endif 1198 } 1199 1200 /* 1201 * Generic schedule. Possibly schedule threads belonging to other cpus and 1202 * deal with threads that might be blocked on a wait queue. 1203 * 1204 * We have a little helper inline function which does additional work after 1205 * the thread has been enqueued, including dealing with preemption and 1206 * setting need_lwkt_resched() (which prevents the kernel from returning 1207 * to userland until it has processed higher priority threads). 1208 * 1209 * It is possible for this routine to be called after a failed _enqueue 1210 * (due to the target thread migrating, sleeping, or otherwise blocked). 1211 * We have to check that the thread is actually on the run queue! 1212 * 1213 * reschedok is an optimized constant propagated from lwkt_schedule() or 1214 * lwkt_schedule_noresched(). By default it is non-zero, causing a 1215 * reschedule to be requested if the target thread has a higher priority. 1216 * The port messaging code will set MSG_NORESCHED and cause reschedok to 1217 * be 0, prevented undesired reschedules. 1218 */ 1219 static __inline 1220 void 1221 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount, int reschedok) 1222 { 1223 thread_t otd; 1224 1225 if (ntd->td_flags & TDF_RUNQ) { 1226 if (ntd->td_preemptable && reschedok) { 1227 ntd->td_preemptable(ntd, ccount); /* YYY +token */ 1228 } else if (reschedok) { 1229 otd = curthread; 1230 if (ntd->td_pri > otd->td_pri) 1231 need_lwkt_resched(); 1232 } 1233 1234 /* 1235 * Give the thread a little fair share scheduler bump if it 1236 * has been asleep for a while. This is primarily to avoid 1237 * a degenerate case for interrupt threads where accumulator 1238 * crosses into negative territory unnecessarily. 1239 */ 1240 if (ntd->td_fairq_lticks != ticks) { 1241 ntd->td_fairq_lticks = ticks; 1242 ntd->td_fairq_accum += gd->gd_fairq_total_pri; 1243 if (ntd->td_fairq_accum > TDFAIRQ_MAX(gd)) 1244 ntd->td_fairq_accum = TDFAIRQ_MAX(gd); 1245 } 1246 } 1247 } 1248 1249 static __inline 1250 void 1251 _lwkt_schedule(thread_t td, int reschedok) 1252 { 1253 globaldata_t mygd = mycpu; 1254 1255 KASSERT(td != &td->td_gd->gd_idlethread, 1256 ("lwkt_schedule(): scheduling gd_idlethread is illegal!")); 1257 crit_enter_gd(mygd); 1258 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0); 1259 if (td == mygd->gd_curthread) { 1260 _lwkt_enqueue(td); 1261 } else { 1262 /* 1263 * If we own the thread, there is no race (since we are in a 1264 * critical section). If we do not own the thread there might 1265 * be a race but the target cpu will deal with it. 1266 */ 1267 #ifdef SMP 1268 if (td->td_gd == mygd) { 1269 _lwkt_enqueue(td); 1270 _lwkt_schedule_post(mygd, td, 1, reschedok); 1271 } else { 1272 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0); 1273 } 1274 #else 1275 _lwkt_enqueue(td); 1276 _lwkt_schedule_post(mygd, td, 1, reschedok); 1277 #endif 1278 } 1279 crit_exit_gd(mygd); 1280 } 1281 1282 void 1283 lwkt_schedule(thread_t td) 1284 { 1285 _lwkt_schedule(td, 1); 1286 } 1287 1288 void 1289 lwkt_schedule_noresched(thread_t td) 1290 { 1291 _lwkt_schedule(td, 0); 1292 } 1293 1294 #ifdef SMP 1295 1296 /* 1297 * When scheduled remotely if frame != NULL the IPIQ is being 1298 * run via doreti or an interrupt then preemption can be allowed. 1299 * 1300 * To allow preemption we have to drop the critical section so only 1301 * one is present in _lwkt_schedule_post. 1302 */ 1303 static void 1304 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame) 1305 { 1306 thread_t td = curthread; 1307 thread_t ntd = arg; 1308 1309 if (frame && ntd->td_preemptable) { 1310 crit_exit_noyield(td); 1311 _lwkt_schedule(ntd, 1); 1312 crit_enter_quick(td); 1313 } else { 1314 _lwkt_schedule(ntd, 1); 1315 } 1316 } 1317 1318 /* 1319 * Thread migration using a 'Pull' method. The thread may or may not be 1320 * the current thread. It MUST be descheduled and in a stable state. 1321 * lwkt_giveaway() must be called on the cpu owning the thread. 1322 * 1323 * At any point after lwkt_giveaway() is called, the target cpu may 1324 * 'pull' the thread by calling lwkt_acquire(). 1325 * 1326 * We have to make sure the thread is not sitting on a per-cpu tsleep 1327 * queue or it will blow up when it moves to another cpu. 1328 * 1329 * MPSAFE - must be called under very specific conditions. 1330 */ 1331 void 1332 lwkt_giveaway(thread_t td) 1333 { 1334 globaldata_t gd = mycpu; 1335 1336 crit_enter_gd(gd); 1337 if (td->td_flags & TDF_TSLEEPQ) 1338 tsleep_remove(td); 1339 KKASSERT(td->td_gd == gd); 1340 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); 1341 td->td_flags |= TDF_MIGRATING; 1342 crit_exit_gd(gd); 1343 } 1344 1345 void 1346 lwkt_acquire(thread_t td) 1347 { 1348 globaldata_t gd; 1349 globaldata_t mygd; 1350 1351 KKASSERT(td->td_flags & TDF_MIGRATING); 1352 gd = td->td_gd; 1353 mygd = mycpu; 1354 if (gd != mycpu) { 1355 cpu_lfence(); 1356 KKASSERT((td->td_flags & TDF_RUNQ) == 0); 1357 crit_enter_gd(mygd); 1358 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) { 1359 #ifdef SMP 1360 lwkt_process_ipiq(); 1361 #endif 1362 cpu_lfence(); 1363 } 1364 cpu_mfence(); 1365 td->td_gd = mygd; 1366 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1367 td->td_flags &= ~TDF_MIGRATING; 1368 crit_exit_gd(mygd); 1369 } else { 1370 crit_enter_gd(mygd); 1371 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1372 td->td_flags &= ~TDF_MIGRATING; 1373 crit_exit_gd(mygd); 1374 } 1375 } 1376 1377 #endif 1378 1379 /* 1380 * Generic deschedule. Descheduling threads other then your own should be 1381 * done only in carefully controlled circumstances. Descheduling is 1382 * asynchronous. 1383 * 1384 * This function may block if the cpu has run out of messages. 1385 */ 1386 void 1387 lwkt_deschedule(thread_t td) 1388 { 1389 crit_enter(); 1390 #ifdef SMP 1391 if (td == curthread) { 1392 _lwkt_dequeue(td); 1393 } else { 1394 if (td->td_gd == mycpu) { 1395 _lwkt_dequeue(td); 1396 } else { 1397 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td); 1398 } 1399 } 1400 #else 1401 _lwkt_dequeue(td); 1402 #endif 1403 crit_exit(); 1404 } 1405 1406 /* 1407 * Set the target thread's priority. This routine does not automatically 1408 * switch to a higher priority thread, LWKT threads are not designed for 1409 * continuous priority changes. Yield if you want to switch. 1410 */ 1411 void 1412 lwkt_setpri(thread_t td, int pri) 1413 { 1414 KKASSERT(td->td_gd == mycpu); 1415 if (td->td_pri != pri) { 1416 KKASSERT(pri >= 0); 1417 crit_enter(); 1418 if (td->td_flags & TDF_RUNQ) { 1419 _lwkt_dequeue(td); 1420 td->td_pri = pri; 1421 _lwkt_enqueue(td); 1422 } else { 1423 td->td_pri = pri; 1424 } 1425 crit_exit(); 1426 } 1427 } 1428 1429 /* 1430 * Set the initial priority for a thread prior to it being scheduled for 1431 * the first time. The thread MUST NOT be scheduled before or during 1432 * this call. The thread may be assigned to a cpu other then the current 1433 * cpu. 1434 * 1435 * Typically used after a thread has been created with TDF_STOPPREQ, 1436 * and before the thread is initially scheduled. 1437 */ 1438 void 1439 lwkt_setpri_initial(thread_t td, int pri) 1440 { 1441 KKASSERT(pri >= 0); 1442 KKASSERT((td->td_flags & TDF_RUNQ) == 0); 1443 td->td_pri = pri; 1444 } 1445 1446 void 1447 lwkt_setpri_self(int pri) 1448 { 1449 thread_t td = curthread; 1450 1451 KKASSERT(pri >= 0 && pri <= TDPRI_MAX); 1452 crit_enter(); 1453 if (td->td_flags & TDF_RUNQ) { 1454 _lwkt_dequeue(td); 1455 td->td_pri = pri; 1456 _lwkt_enqueue(td); 1457 } else { 1458 td->td_pri = pri; 1459 } 1460 crit_exit(); 1461 } 1462 1463 /* 1464 * 1/hz tick (typically 10ms) x TDFAIRQ_SCALE (typ 8) = 80ms full cycle. 1465 * 1466 * Example: two competing threads, same priority N. decrement by (2*N) 1467 * increment by N*8, each thread will get 4 ticks. 1468 */ 1469 void 1470 lwkt_fairq_schedulerclock(thread_t td) 1471 { 1472 if (fairq_enable) { 1473 while (td) { 1474 if (td != &td->td_gd->gd_idlethread) { 1475 td->td_fairq_accum -= td->td_gd->gd_fairq_total_pri; 1476 if (td->td_fairq_accum < -TDFAIRQ_MAX(td->td_gd)) 1477 td->td_fairq_accum = -TDFAIRQ_MAX(td->td_gd); 1478 if (td->td_fairq_accum < 0) 1479 need_lwkt_resched(); 1480 td->td_fairq_lticks = ticks; 1481 } 1482 td = td->td_preempted; 1483 } 1484 } 1485 } 1486 1487 static void 1488 lwkt_fairq_accumulate(globaldata_t gd, thread_t td) 1489 { 1490 td->td_fairq_accum += td->td_pri * TDFAIRQ_SCALE; 1491 if (td->td_fairq_accum > TDFAIRQ_MAX(td->td_gd)) 1492 td->td_fairq_accum = TDFAIRQ_MAX(td->td_gd); 1493 } 1494 1495 /* 1496 * Migrate the current thread to the specified cpu. 1497 * 1498 * This is accomplished by descheduling ourselves from the current cpu, 1499 * moving our thread to the tdallq of the target cpu, IPI messaging the 1500 * target cpu, and switching out. TDF_MIGRATING prevents scheduling 1501 * races while the thread is being migrated. 1502 * 1503 * We must be sure to remove ourselves from the current cpu's tsleepq 1504 * before potentially moving to another queue. The thread can be on 1505 * a tsleepq due to a left-over tsleep_interlock(). 1506 */ 1507 #ifdef SMP 1508 static void lwkt_setcpu_remote(void *arg); 1509 #endif 1510 1511 void 1512 lwkt_setcpu_self(globaldata_t rgd) 1513 { 1514 #ifdef SMP 1515 thread_t td = curthread; 1516 1517 if (td->td_gd != rgd) { 1518 crit_enter_quick(td); 1519 if (td->td_flags & TDF_TSLEEPQ) 1520 tsleep_remove(td); 1521 td->td_flags |= TDF_MIGRATING; 1522 lwkt_deschedule_self(td); 1523 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1524 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td); 1525 lwkt_switch(); 1526 /* we are now on the target cpu */ 1527 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); 1528 crit_exit_quick(td); 1529 } 1530 #endif 1531 } 1532 1533 void 1534 lwkt_migratecpu(int cpuid) 1535 { 1536 #ifdef SMP 1537 globaldata_t rgd; 1538 1539 rgd = globaldata_find(cpuid); 1540 lwkt_setcpu_self(rgd); 1541 #endif 1542 } 1543 1544 /* 1545 * Remote IPI for cpu migration (called while in a critical section so we 1546 * do not have to enter another one). The thread has already been moved to 1547 * our cpu's allq, but we must wait for the thread to be completely switched 1548 * out on the originating cpu before we schedule it on ours or the stack 1549 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD 1550 * change to main memory. 1551 * 1552 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races 1553 * against wakeups. It is best if this interface is used only when there 1554 * are no pending events that might try to schedule the thread. 1555 */ 1556 #ifdef SMP 1557 static void 1558 lwkt_setcpu_remote(void *arg) 1559 { 1560 thread_t td = arg; 1561 globaldata_t gd = mycpu; 1562 1563 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) { 1564 #ifdef SMP 1565 lwkt_process_ipiq(); 1566 #endif 1567 cpu_lfence(); 1568 cpu_pause(); 1569 } 1570 td->td_gd = gd; 1571 cpu_mfence(); 1572 td->td_flags &= ~TDF_MIGRATING; 1573 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0); 1574 _lwkt_enqueue(td); 1575 } 1576 #endif 1577 1578 struct lwp * 1579 lwkt_preempted_proc(void) 1580 { 1581 thread_t td = curthread; 1582 while (td->td_preempted) 1583 td = td->td_preempted; 1584 return(td->td_lwp); 1585 } 1586 1587 /* 1588 * Create a kernel process/thread/whatever. It shares it's address space 1589 * with proc0 - ie: kernel only. 1590 * 1591 * NOTE! By default new threads are created with the MP lock held. A 1592 * thread which does not require the MP lock should release it by calling 1593 * rel_mplock() at the start of the new thread. 1594 */ 1595 int 1596 lwkt_create(void (*func)(void *), void *arg, struct thread **tdp, 1597 thread_t template, int tdflags, int cpu, const char *fmt, ...) 1598 { 1599 thread_t td; 1600 __va_list ap; 1601 1602 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu, 1603 tdflags); 1604 if (tdp) 1605 *tdp = td; 1606 cpu_set_thread_handler(td, lwkt_exit, func, arg); 1607 1608 /* 1609 * Set up arg0 for 'ps' etc 1610 */ 1611 __va_start(ap, fmt); 1612 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap); 1613 __va_end(ap); 1614 1615 /* 1616 * Schedule the thread to run 1617 */ 1618 if ((td->td_flags & TDF_STOPREQ) == 0) 1619 lwkt_schedule(td); 1620 else 1621 td->td_flags &= ~TDF_STOPREQ; 1622 return 0; 1623 } 1624 1625 /* 1626 * Destroy an LWKT thread. Warning! This function is not called when 1627 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and 1628 * uses a different reaping mechanism. 1629 */ 1630 void 1631 lwkt_exit(void) 1632 { 1633 thread_t td = curthread; 1634 thread_t std; 1635 globaldata_t gd; 1636 1637 /* 1638 * Do any cleanup that might block here 1639 */ 1640 if (td->td_flags & TDF_VERBOSE) 1641 kprintf("kthread %p %s has exited\n", td, td->td_comm); 1642 caps_exit(td); 1643 biosched_done(td); 1644 dsched_exit_thread(td); 1645 1646 /* 1647 * Get us into a critical section to interlock gd_freetd and loop 1648 * until we can get it freed. 1649 * 1650 * We have to cache the current td in gd_freetd because objcache_put()ing 1651 * it would rip it out from under us while our thread is still active. 1652 */ 1653 gd = mycpu; 1654 crit_enter_quick(td); 1655 while ((std = gd->gd_freetd) != NULL) { 1656 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0); 1657 gd->gd_freetd = NULL; 1658 objcache_put(thread_cache, std); 1659 } 1660 1661 /* 1662 * Remove thread resources from kernel lists and deschedule us for 1663 * the last time. We cannot block after this point or we may end 1664 * up with a stale td on the tsleepq. 1665 */ 1666 if (td->td_flags & TDF_TSLEEPQ) 1667 tsleep_remove(td); 1668 lwkt_deschedule_self(td); 1669 lwkt_remove_tdallq(td); 1670 1671 /* 1672 * Final cleanup 1673 */ 1674 KKASSERT(gd->gd_freetd == NULL); 1675 if (td->td_flags & TDF_ALLOCATED_THREAD) 1676 gd->gd_freetd = td; 1677 cpu_thread_exit(); 1678 } 1679 1680 void 1681 lwkt_remove_tdallq(thread_t td) 1682 { 1683 KKASSERT(td->td_gd == mycpu); 1684 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1685 } 1686 1687 /* 1688 * Code reduction and branch prediction improvements. Call/return 1689 * overhead on modern cpus often degenerates into 0 cycles due to 1690 * the cpu's branch prediction hardware and return pc cache. We 1691 * can take advantage of this by not inlining medium-complexity 1692 * functions and we can also reduce the branch prediction impact 1693 * by collapsing perfectly predictable branches into a single 1694 * procedure instead of duplicating it. 1695 * 1696 * Is any of this noticeable? Probably not, so I'll take the 1697 * smaller code size. 1698 */ 1699 void 1700 crit_exit_wrapper(__DEBUG_CRIT_ARG__) 1701 { 1702 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__); 1703 } 1704 1705 void 1706 crit_panic(void) 1707 { 1708 thread_t td = curthread; 1709 int lcrit = td->td_critcount; 1710 1711 td->td_critcount = 0; 1712 panic("td_critcount is/would-go negative! %p %d", td, lcrit); 1713 /* NOT REACHED */ 1714 } 1715 1716 #ifdef SMP 1717 1718 /* 1719 * Called from debugger/panic on cpus which have been stopped. We must still 1720 * process the IPIQ while stopped, even if we were stopped while in a critical 1721 * section (XXX). 1722 * 1723 * If we are dumping also try to process any pending interrupts. This may 1724 * or may not work depending on the state of the cpu at the point it was 1725 * stopped. 1726 */ 1727 void 1728 lwkt_smp_stopped(void) 1729 { 1730 globaldata_t gd = mycpu; 1731 1732 crit_enter_gd(gd); 1733 if (dumping) { 1734 lwkt_process_ipiq(); 1735 splz(); 1736 } else { 1737 lwkt_process_ipiq(); 1738 } 1739 crit_exit_gd(gd); 1740 } 1741 1742 #endif 1743