1 /* 2 * Copyright (c) 2003,2004 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 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.107 2007/05/01 00:05:18 dillon Exp $ 35 */ 36 37 /* 38 * Each cpu in a system has its own self-contained light weight kernel 39 * thread scheduler, which means that generally speaking we only need 40 * to use a critical section to avoid problems. Foreign thread 41 * scheduling is queued via (async) IPIs. 42 */ 43 44 #ifdef _KERNEL 45 46 #include <sys/param.h> 47 #include <sys/systm.h> 48 #include <sys/kernel.h> 49 #include <sys/proc.h> 50 #include <sys/rtprio.h> 51 #include <sys/queue.h> 52 #include <sys/sysctl.h> 53 #include <sys/kthread.h> 54 #include <machine/cpu.h> 55 #include <sys/lock.h> 56 #include <sys/caps.h> 57 #include <sys/spinlock.h> 58 #include <sys/ktr.h> 59 60 #include <sys/thread2.h> 61 #include <sys/spinlock2.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 #include <vm/vm_zone.h> 72 73 #include <machine/stdarg.h> 74 #include <machine/smp.h> 75 76 #else 77 78 #include <sys/stdint.h> 79 #include <libcaps/thread.h> 80 #include <sys/thread.h> 81 #include <sys/msgport.h> 82 #include <sys/errno.h> 83 #include <libcaps/globaldata.h> 84 #include <machine/cpufunc.h> 85 #include <sys/thread2.h> 86 #include <sys/msgport2.h> 87 #include <stdio.h> 88 #include <stdlib.h> 89 #include <string.h> 90 #include <machine/lock.h> 91 #include <machine/atomic.h> 92 #include <machine/cpu.h> 93 94 #endif 95 96 static int untimely_switch = 0; 97 #ifdef INVARIANTS 98 static int panic_on_cscount = 0; 99 #endif 100 static __int64_t switch_count = 0; 101 static __int64_t preempt_hit = 0; 102 static __int64_t preempt_miss = 0; 103 static __int64_t preempt_weird = 0; 104 static __int64_t token_contention_count = 0; 105 static __int64_t mplock_contention_count = 0; 106 107 #ifdef _KERNEL 108 109 SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, ""); 110 #ifdef INVARIANTS 111 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, ""); 112 #endif 113 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, ""); 114 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, ""); 115 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, ""); 116 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, ""); 117 #ifdef INVARIANTS 118 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW, 119 &token_contention_count, 0, "spinning due to token contention"); 120 SYSCTL_QUAD(_lwkt, OID_AUTO, mplock_contention_count, CTLFLAG_RW, 121 &mplock_contention_count, 0, "spinning due to MPLOCK contention"); 122 #endif 123 #endif 124 125 /* 126 * Kernel Trace 127 */ 128 #ifdef _KERNEL 129 130 #if !defined(KTR_GIANT_CONTENTION) 131 #define KTR_GIANT_CONTENTION KTR_ALL 132 #endif 133 134 KTR_INFO_MASTER(giant); 135 KTR_INFO(KTR_GIANT_CONTENTION, giant, beg, 0, "thread=%p", sizeof(void *)); 136 KTR_INFO(KTR_GIANT_CONTENTION, giant, end, 1, "thread=%p", sizeof(void *)); 137 138 #define loggiant(name) KTR_LOG(giant_ ## name, curthread) 139 140 #endif 141 142 /* 143 * These helper procedures handle the runq, they can only be called from 144 * within a critical section. 145 * 146 * WARNING! Prior to SMP being brought up it is possible to enqueue and 147 * dequeue threads belonging to other cpus, so be sure to use td->td_gd 148 * instead of 'mycpu' when referencing the globaldata structure. Once 149 * SMP live enqueuing and dequeueing only occurs on the current cpu. 150 */ 151 static __inline 152 void 153 _lwkt_dequeue(thread_t td) 154 { 155 if (td->td_flags & TDF_RUNQ) { 156 int nq = td->td_pri & TDPRI_MASK; 157 struct globaldata *gd = td->td_gd; 158 159 td->td_flags &= ~TDF_RUNQ; 160 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq); 161 /* runqmask is passively cleaned up by the switcher */ 162 } 163 } 164 165 static __inline 166 void 167 _lwkt_enqueue(thread_t td) 168 { 169 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_TSLEEPQ|TDF_BLOCKQ)) == 0) { 170 int nq = td->td_pri & TDPRI_MASK; 171 struct globaldata *gd = td->td_gd; 172 173 td->td_flags |= TDF_RUNQ; 174 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq); 175 gd->gd_runqmask |= 1 << nq; 176 } 177 } 178 179 /* 180 * Schedule a thread to run. As the current thread we can always safely 181 * schedule ourselves, and a shortcut procedure is provided for that 182 * function. 183 * 184 * (non-blocking, self contained on a per cpu basis) 185 */ 186 void 187 lwkt_schedule_self(thread_t td) 188 { 189 crit_enter_quick(td); 190 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!")); 191 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0); 192 _lwkt_enqueue(td); 193 crit_exit_quick(td); 194 } 195 196 /* 197 * Deschedule a thread. 198 * 199 * (non-blocking, self contained on a per cpu basis) 200 */ 201 void 202 lwkt_deschedule_self(thread_t td) 203 { 204 crit_enter_quick(td); 205 _lwkt_dequeue(td); 206 crit_exit_quick(td); 207 } 208 209 #ifdef _KERNEL 210 211 /* 212 * LWKTs operate on a per-cpu basis 213 * 214 * WARNING! Called from early boot, 'mycpu' may not work yet. 215 */ 216 void 217 lwkt_gdinit(struct globaldata *gd) 218 { 219 int i; 220 221 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i) 222 TAILQ_INIT(&gd->gd_tdrunq[i]); 223 gd->gd_runqmask = 0; 224 TAILQ_INIT(&gd->gd_tdallq); 225 } 226 227 #endif /* _KERNEL */ 228 229 /* 230 * Create a new thread. The thread must be associated with a process context 231 * or LWKT start address before it can be scheduled. If the target cpu is 232 * -1 the thread will be created on the current cpu. 233 * 234 * If you intend to create a thread without a process context this function 235 * does everything except load the startup and switcher function. 236 */ 237 thread_t 238 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags) 239 { 240 void *stack; 241 globaldata_t gd = mycpu; 242 243 if (td == NULL) { 244 crit_enter_gd(gd); 245 if (gd->gd_tdfreecount > 0) { 246 --gd->gd_tdfreecount; 247 td = TAILQ_FIRST(&gd->gd_tdfreeq); 248 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0, 249 ("lwkt_alloc_thread: unexpected NULL or corrupted td")); 250 TAILQ_REMOVE(&gd->gd_tdfreeq, td, td_threadq); 251 crit_exit_gd(gd); 252 flags |= td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD); 253 } else { 254 crit_exit_gd(gd); 255 #ifdef _KERNEL 256 td = zalloc(thread_zone); 257 #else 258 td = malloc(sizeof(struct thread)); 259 #endif 260 td->td_kstack = NULL; 261 td->td_kstack_size = 0; 262 flags |= TDF_ALLOCATED_THREAD; 263 } 264 } 265 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) { 266 if (flags & TDF_ALLOCATED_STACK) { 267 #ifdef _KERNEL 268 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size); 269 #else 270 libcaps_free_stack(stack, td->td_kstack_size); 271 #endif 272 stack = NULL; 273 } 274 } 275 if (stack == NULL) { 276 #ifdef _KERNEL 277 stack = (void *)kmem_alloc(&kernel_map, stksize); 278 #else 279 stack = libcaps_alloc_stack(stksize); 280 #endif 281 flags |= TDF_ALLOCATED_STACK; 282 } 283 if (cpu < 0) 284 lwkt_init_thread(td, stack, stksize, flags, mycpu); 285 else 286 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu)); 287 return(td); 288 } 289 290 #ifdef _KERNEL 291 292 /* 293 * Initialize a preexisting thread structure. This function is used by 294 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread. 295 * 296 * All threads start out in a critical section at a priority of 297 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as 298 * appropriate. This function may send an IPI message when the 299 * requested cpu is not the current cpu and consequently gd_tdallq may 300 * not be initialized synchronously from the point of view of the originating 301 * cpu. 302 * 303 * NOTE! we have to be careful in regards to creating threads for other cpus 304 * if SMP has not yet been activated. 305 */ 306 #ifdef SMP 307 308 static void 309 lwkt_init_thread_remote(void *arg) 310 { 311 thread_t td = arg; 312 313 /* 314 * Protected by critical section held by IPI dispatch 315 */ 316 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq); 317 } 318 319 #endif 320 321 void 322 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags, 323 struct globaldata *gd) 324 { 325 globaldata_t mygd = mycpu; 326 327 bzero(td, sizeof(struct thread)); 328 td->td_kstack = stack; 329 td->td_kstack_size = stksize; 330 td->td_flags = flags; 331 td->td_gd = gd; 332 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT; 333 #ifdef SMP 334 if ((flags & TDF_MPSAFE) == 0) 335 td->td_mpcount = 1; 336 #endif 337 lwkt_initport(&td->td_msgport, td); 338 pmap_init_thread(td); 339 #ifdef SMP 340 /* 341 * Normally initializing a thread for a remote cpu requires sending an 342 * IPI. However, the idlethread is setup before the other cpus are 343 * activated so we have to treat it as a special case. XXX manipulation 344 * of gd_tdallq requires the BGL. 345 */ 346 if (gd == mygd || td == &gd->gd_idlethread) { 347 crit_enter_gd(mygd); 348 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); 349 crit_exit_gd(mygd); 350 } else { 351 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td); 352 } 353 #else 354 crit_enter_gd(mygd); 355 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); 356 crit_exit_gd(mygd); 357 #endif 358 } 359 360 #endif /* _KERNEL */ 361 362 void 363 lwkt_set_comm(thread_t td, const char *ctl, ...) 364 { 365 __va_list va; 366 367 __va_start(va, ctl); 368 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va); 369 __va_end(va); 370 } 371 372 void 373 lwkt_hold(thread_t td) 374 { 375 ++td->td_refs; 376 } 377 378 void 379 lwkt_rele(thread_t td) 380 { 381 KKASSERT(td->td_refs > 0); 382 --td->td_refs; 383 } 384 385 #ifdef _KERNEL 386 387 void 388 lwkt_wait_free(thread_t td) 389 { 390 while (td->td_refs) 391 tsleep(td, 0, "tdreap", hz); 392 } 393 394 #endif 395 396 void 397 lwkt_free_thread(thread_t td) 398 { 399 struct globaldata *gd = mycpu; 400 401 KASSERT((td->td_flags & TDF_RUNNING) == 0, 402 ("lwkt_free_thread: did not exit! %p", td)); 403 404 crit_enter_gd(gd); 405 if (gd->gd_tdfreecount < CACHE_NTHREADS && 406 (td->td_flags & TDF_ALLOCATED_THREAD) 407 ) { 408 ++gd->gd_tdfreecount; 409 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq); 410 crit_exit_gd(gd); 411 } else { 412 crit_exit_gd(gd); 413 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) { 414 #ifdef _KERNEL 415 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size); 416 #else 417 libcaps_free_stack(td->td_kstack, td->td_kstack_size); 418 #endif 419 /* gd invalid */ 420 td->td_kstack = NULL; 421 td->td_kstack_size = 0; 422 } 423 if (td->td_flags & TDF_ALLOCATED_THREAD) { 424 #ifdef _KERNEL 425 zfree(thread_zone, td); 426 #else 427 free(td); 428 #endif 429 } 430 } 431 } 432 433 434 /* 435 * Switch to the next runnable lwkt. If no LWKTs are runnable then 436 * switch to the idlethread. Switching must occur within a critical 437 * section to avoid races with the scheduling queue. 438 * 439 * We always have full control over our cpu's run queue. Other cpus 440 * that wish to manipulate our queue must use the cpu_*msg() calls to 441 * talk to our cpu, so a critical section is all that is needed and 442 * the result is very, very fast thread switching. 443 * 444 * The LWKT scheduler uses a fixed priority model and round-robins at 445 * each priority level. User process scheduling is a totally 446 * different beast and LWKT priorities should not be confused with 447 * user process priorities. 448 * 449 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch() 450 * cleans it up. Note that the td_switch() function cannot do anything that 451 * requires the MP lock since the MP lock will have already been setup for 452 * the target thread (not the current thread). It's nice to have a scheduler 453 * that does not need the MP lock to work because it allows us to do some 454 * really cool high-performance MP lock optimizations. 455 * 456 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch() 457 * is not called by the current thread in the preemption case, only when 458 * the preempting thread blocks (in order to return to the original thread). 459 */ 460 void 461 lwkt_switch(void) 462 { 463 globaldata_t gd = mycpu; 464 thread_t td = gd->gd_curthread; 465 thread_t ntd; 466 #ifdef SMP 467 int mpheld; 468 #endif 469 470 /* 471 * Switching from within a 'fast' (non thread switched) interrupt or IPI 472 * is illegal. However, we may have to do it anyway if we hit a fatal 473 * kernel trap or we have paniced. 474 * 475 * If this case occurs save and restore the interrupt nesting level. 476 */ 477 if (gd->gd_intr_nesting_level) { 478 int savegdnest; 479 int savegdtrap; 480 481 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) { 482 panic("lwkt_switch: cannot switch from within " 483 "a fast interrupt, yet, td %p\n", td); 484 } else { 485 savegdnest = gd->gd_intr_nesting_level; 486 savegdtrap = gd->gd_trap_nesting_level; 487 gd->gd_intr_nesting_level = 0; 488 gd->gd_trap_nesting_level = 0; 489 if ((td->td_flags & TDF_PANICWARN) == 0) { 490 td->td_flags |= TDF_PANICWARN; 491 kprintf("Warning: thread switch from interrupt or IPI, " 492 "thread %p (%s)\n", td, td->td_comm); 493 #ifdef DDB 494 db_print_backtrace(); 495 #endif 496 } 497 lwkt_switch(); 498 gd->gd_intr_nesting_level = savegdnest; 499 gd->gd_trap_nesting_level = savegdtrap; 500 return; 501 } 502 } 503 504 /* 505 * Passive release (used to transition from user to kernel mode 506 * when we block or switch rather then when we enter the kernel). 507 * This function is NOT called if we are switching into a preemption 508 * or returning from a preemption. Typically this causes us to lose 509 * our current process designation (if we have one) and become a true 510 * LWKT thread, and may also hand the current process designation to 511 * another process and schedule thread. 512 */ 513 if (td->td_release) 514 td->td_release(td); 515 516 crit_enter_gd(gd); 517 #ifdef SMP 518 if (td->td_toks) 519 lwkt_relalltokens(td); 520 #endif 521 522 /* 523 * We had better not be holding any spin locks, but don't get into an 524 * endless panic loop. 525 */ 526 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL, 527 ("lwkt_switch: still holding a shared spinlock %p!", 528 gd->gd_spinlock_rd)); 529 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL, 530 ("lwkt_switch: still holding %d exclusive spinlocks!", 531 gd->gd_spinlocks_wr)); 532 533 534 #ifdef SMP 535 /* 536 * td_mpcount cannot be used to determine if we currently hold the 537 * MP lock because get_mplock() will increment it prior to attempting 538 * to get the lock, and switch out if it can't. Our ownership of 539 * the actual lock will remain stable while we are in a critical section 540 * (but, of course, another cpu may own or release the lock so the 541 * actual value of mp_lock is not stable). 542 */ 543 mpheld = MP_LOCK_HELD(); 544 #ifdef INVARIANTS 545 if (td->td_cscount) { 546 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n", 547 td); 548 if (panic_on_cscount) 549 panic("switching while mastering cpusync"); 550 } 551 #endif 552 #endif 553 if ((ntd = td->td_preempted) != NULL) { 554 /* 555 * We had preempted another thread on this cpu, resume the preempted 556 * thread. This occurs transparently, whether the preempted thread 557 * was scheduled or not (it may have been preempted after descheduling 558 * itself). 559 * 560 * We have to setup the MP lock for the original thread after backing 561 * out the adjustment that was made to curthread when the original 562 * was preempted. 563 */ 564 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK); 565 #ifdef SMP 566 if (ntd->td_mpcount && mpheld == 0) { 567 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d", 568 td, ntd, td->td_mpcount, ntd->td_mpcount); 569 } 570 if (ntd->td_mpcount) { 571 td->td_mpcount -= ntd->td_mpcount; 572 KKASSERT(td->td_mpcount >= 0); 573 } 574 #endif 575 ntd->td_flags |= TDF_PREEMPT_DONE; 576 577 /* 578 * XXX. The interrupt may have woken a thread up, we need to properly 579 * set the reschedule flag if the originally interrupted thread is at 580 * a lower priority. 581 */ 582 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1) 583 need_lwkt_resched(); 584 /* YYY release mp lock on switchback if original doesn't need it */ 585 } else { 586 /* 587 * Priority queue / round-robin at each priority. Note that user 588 * processes run at a fixed, low priority and the user process 589 * scheduler deals with interactions between user processes 590 * by scheduling and descheduling them from the LWKT queue as 591 * necessary. 592 * 593 * We have to adjust the MP lock for the target thread. If we 594 * need the MP lock and cannot obtain it we try to locate a 595 * thread that does not need the MP lock. If we cannot, we spin 596 * instead of HLT. 597 * 598 * A similar issue exists for the tokens held by the target thread. 599 * If we cannot obtain ownership of the tokens we cannot immediately 600 * schedule the thread. 601 */ 602 603 /* 604 * If an LWKT reschedule was requested, well that is what we are 605 * doing now so clear it. 606 */ 607 clear_lwkt_resched(); 608 again: 609 if (gd->gd_runqmask) { 610 int nq = bsrl(gd->gd_runqmask); 611 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) { 612 gd->gd_runqmask &= ~(1 << nq); 613 goto again; 614 } 615 #ifdef SMP 616 /* 617 * THREAD SELECTION FOR AN SMP MACHINE BUILD 618 * 619 * If the target needs the MP lock and we couldn't get it, 620 * or if the target is holding tokens and we could not 621 * gain ownership of the tokens, continue looking for a 622 * thread to schedule and spin instead of HLT if we can't. 623 * 624 * NOTE: the mpheld variable invalid after this conditional, it 625 * can change due to both cpu_try_mplock() returning success 626 * AND interactions in lwkt_getalltokens() due to the fact that 627 * we are trying to check the mpcount of a thread other then 628 * the current thread. Because of this, if the current thread 629 * is not holding td_mpcount, an IPI indirectly run via 630 * lwkt_getalltokens() can obtain and release the MP lock and 631 * cause the core MP lock to be released. 632 */ 633 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) || 634 (ntd->td_toks && lwkt_getalltokens(ntd) == 0) 635 ) { 636 u_int32_t rqmask = gd->gd_runqmask; 637 638 mpheld = MP_LOCK_HELD(); 639 ntd = NULL; 640 while (rqmask) { 641 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) { 642 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) { 643 /* spinning due to MP lock being held */ 644 #ifdef INVARIANTS 645 ++mplock_contention_count; 646 #endif 647 /* mplock still not held, 'mpheld' still valid */ 648 continue; 649 } 650 651 /* 652 * mpheld state invalid after getalltokens call returns 653 * failure, but the variable is only needed for 654 * the loop. 655 */ 656 if (ntd->td_toks && !lwkt_getalltokens(ntd)) { 657 /* spinning due to token contention */ 658 #ifdef INVARIANTS 659 ++token_contention_count; 660 #endif 661 mpheld = MP_LOCK_HELD(); 662 continue; 663 } 664 break; 665 } 666 if (ntd) 667 break; 668 rqmask &= ~(1 << nq); 669 nq = bsrl(rqmask); 670 } 671 if (ntd == NULL) { 672 ntd = &gd->gd_idlethread; 673 ntd->td_flags |= TDF_IDLE_NOHLT; 674 goto using_idle_thread; 675 } else { 676 ++gd->gd_cnt.v_swtch; 677 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq); 678 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq); 679 } 680 } else { 681 ++gd->gd_cnt.v_swtch; 682 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq); 683 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq); 684 } 685 #else 686 /* 687 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to 688 * worry about tokens or the BGL. 689 */ 690 ++gd->gd_cnt.v_swtch; 691 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq); 692 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq); 693 #endif 694 } else { 695 /* 696 * We have nothing to run but only let the idle loop halt 697 * the cpu if there are no pending interrupts. 698 */ 699 ntd = &gd->gd_idlethread; 700 if (gd->gd_reqflags & RQF_IDLECHECK_MASK) 701 ntd->td_flags |= TDF_IDLE_NOHLT; 702 #ifdef SMP 703 using_idle_thread: 704 /* 705 * The idle thread should not be holding the MP lock unless we 706 * are trapping in the kernel or in a panic. Since we select the 707 * idle thread unconditionally when no other thread is available, 708 * if the MP lock is desired during a panic or kernel trap, we 709 * have to loop in the scheduler until we get it. 710 */ 711 if (ntd->td_mpcount) { 712 mpheld = MP_LOCK_HELD(); 713 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) 714 panic("Idle thread %p was holding the BGL!", ntd); 715 else if (mpheld == 0) 716 goto again; 717 } 718 #endif 719 } 720 } 721 KASSERT(ntd->td_pri >= TDPRI_CRIT, 722 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri)); 723 724 /* 725 * Do the actual switch. If the new target does not need the MP lock 726 * and we are holding it, release the MP lock. If the new target requires 727 * the MP lock we have already acquired it for the target. 728 */ 729 #ifdef SMP 730 if (ntd->td_mpcount == 0 ) { 731 if (MP_LOCK_HELD()) 732 cpu_rel_mplock(); 733 } else { 734 ASSERT_MP_LOCK_HELD(ntd); 735 } 736 #endif 737 if (td != ntd) { 738 ++switch_count; 739 td->td_switch(ntd); 740 } 741 /* NOTE: current cpu may have changed after switch */ 742 crit_exit_quick(td); 743 } 744 745 /* 746 * Request that the target thread preempt the current thread. Preemption 747 * only works under a specific set of conditions: 748 * 749 * - We are not preempting ourselves 750 * - The target thread is owned by the current cpu 751 * - We are not currently being preempted 752 * - The target is not currently being preempted 753 * - We are able to satisfy the target's MP lock requirements (if any). 754 * 755 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically 756 * this is called via lwkt_schedule() through the td_preemptable callback. 757 * critpri is the managed critical priority that we should ignore in order 758 * to determine whether preemption is possible (aka usually just the crit 759 * priority of lwkt_schedule() itself). 760 * 761 * XXX at the moment we run the target thread in a critical section during 762 * the preemption in order to prevent the target from taking interrupts 763 * that *WE* can't. Preemption is strictly limited to interrupt threads 764 * and interrupt-like threads, outside of a critical section, and the 765 * preempted source thread will be resumed the instant the target blocks 766 * whether or not the source is scheduled (i.e. preemption is supposed to 767 * be as transparent as possible). 768 * 769 * The target thread inherits our MP count (added to its own) for the 770 * duration of the preemption in order to preserve the atomicy of the 771 * MP lock during the preemption. Therefore, any preempting targets must be 772 * careful in regards to MP assertions. Note that the MP count may be 773 * out of sync with the physical mp_lock, but we do not have to preserve 774 * the original ownership of the lock if it was out of synch (that is, we 775 * can leave it synchronized on return). 776 */ 777 void 778 lwkt_preempt(thread_t ntd, int critpri) 779 { 780 struct globaldata *gd = mycpu; 781 thread_t td; 782 #ifdef SMP 783 int mpheld; 784 int savecnt; 785 #endif 786 787 /* 788 * The caller has put us in a critical section. We can only preempt 789 * if the caller of the caller was not in a critical section (basically 790 * a local interrupt), as determined by the 'critpri' parameter. We 791 * also acn't preempt if the caller is holding any spinlocks (even if 792 * he isn't in a critical section). This also handles the tokens test. 793 * 794 * YYY The target thread must be in a critical section (else it must 795 * inherit our critical section? I dunno yet). 796 * 797 * Set need_lwkt_resched() unconditionally for now YYY. 798 */ 799 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri)); 800 801 td = gd->gd_curthread; 802 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) { 803 ++preempt_miss; 804 return; 805 } 806 if ((td->td_pri & ~TDPRI_MASK) > critpri) { 807 ++preempt_miss; 808 need_lwkt_resched(); 809 return; 810 } 811 #ifdef SMP 812 if (ntd->td_gd != gd) { 813 ++preempt_miss; 814 need_lwkt_resched(); 815 return; 816 } 817 #endif 818 /* 819 * Take the easy way out and do not preempt if the target is holding 820 * any spinlocks. We could test whether the thread(s) being 821 * preempted interlock against the target thread's tokens and whether 822 * we can get all the target thread's tokens, but this situation 823 * should not occur very often so its easier to simply not preempt. 824 * Also, plain spinlocks are impossible to figure out at this point so 825 * just don't preempt. 826 */ 827 if (gd->gd_spinlock_rd || gd->gd_spinlocks_wr) { 828 ++preempt_miss; 829 need_lwkt_resched(); 830 return; 831 } 832 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) { 833 ++preempt_weird; 834 need_lwkt_resched(); 835 return; 836 } 837 if (ntd->td_preempted) { 838 ++preempt_hit; 839 need_lwkt_resched(); 840 return; 841 } 842 #ifdef SMP 843 /* 844 * note: an interrupt might have occured just as we were transitioning 845 * to or from the MP lock. In this case td_mpcount will be pre-disposed 846 * (non-zero) but not actually synchronized with the actual state of the 847 * lock. We can use it to imply an MP lock requirement for the 848 * preemption but we cannot use it to test whether we hold the MP lock 849 * or not. 850 */ 851 savecnt = td->td_mpcount; 852 mpheld = MP_LOCK_HELD(); 853 ntd->td_mpcount += td->td_mpcount; 854 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) { 855 ntd->td_mpcount -= td->td_mpcount; 856 ++preempt_miss; 857 need_lwkt_resched(); 858 return; 859 } 860 #endif 861 862 /* 863 * Since we are able to preempt the current thread, there is no need to 864 * call need_lwkt_resched(). 865 */ 866 ++preempt_hit; 867 ntd->td_preempted = td; 868 td->td_flags |= TDF_PREEMPT_LOCK; 869 td->td_switch(ntd); 870 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE)); 871 #ifdef SMP 872 KKASSERT(savecnt == td->td_mpcount); 873 mpheld = MP_LOCK_HELD(); 874 if (mpheld && td->td_mpcount == 0) 875 cpu_rel_mplock(); 876 else if (mpheld == 0 && td->td_mpcount) 877 panic("lwkt_preempt(): MP lock was not held through"); 878 #endif 879 ntd->td_preempted = NULL; 880 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE); 881 } 882 883 /* 884 * Yield our thread while higher priority threads are pending. This is 885 * typically called when we leave a critical section but it can be safely 886 * called while we are in a critical section. 887 * 888 * This function will not generally yield to equal priority threads but it 889 * can occur as a side effect. Note that lwkt_switch() is called from 890 * inside the critical section to prevent its own crit_exit() from reentering 891 * lwkt_yield_quick(). 892 * 893 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint 894 * came along but was blocked and made pending. 895 * 896 * (self contained on a per cpu basis) 897 */ 898 void 899 lwkt_yield_quick(void) 900 { 901 globaldata_t gd = mycpu; 902 thread_t td = gd->gd_curthread; 903 904 /* 905 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear 906 * it with a non-zero cpl then we might not wind up calling splz after 907 * a task switch when the critical section is exited even though the 908 * new task could accept the interrupt. 909 * 910 * XXX from crit_exit() only called after last crit section is released. 911 * If called directly will run splz() even if in a critical section. 912 * 913 * td_nest_count prevent deep nesting via splz() or doreti(). Note that 914 * except for this special case, we MUST call splz() here to handle any 915 * pending ints, particularly after we switch, or we might accidently 916 * halt the cpu with interrupts pending. 917 */ 918 if (gd->gd_reqflags && td->td_nest_count < 2) 919 splz(); 920 921 /* 922 * YYY enabling will cause wakeup() to task-switch, which really 923 * confused the old 4.x code. This is a good way to simulate 924 * preemption and MP without actually doing preemption or MP, because a 925 * lot of code assumes that wakeup() does not block. 926 */ 927 if (untimely_switch && td->td_nest_count == 0 && 928 gd->gd_intr_nesting_level == 0 929 ) { 930 crit_enter_quick(td); 931 /* 932 * YYY temporary hacks until we disassociate the userland scheduler 933 * from the LWKT scheduler. 934 */ 935 if (td->td_flags & TDF_RUNQ) { 936 lwkt_switch(); /* will not reenter yield function */ 937 } else { 938 lwkt_schedule_self(td); /* make sure we are scheduled */ 939 lwkt_switch(); /* will not reenter yield function */ 940 lwkt_deschedule_self(td); /* make sure we are descheduled */ 941 } 942 crit_exit_noyield(td); 943 } 944 } 945 946 /* 947 * This implements a normal yield which, unlike _quick, will yield to equal 948 * priority threads as well. Note that gd_reqflags tests will be handled by 949 * the crit_exit() call in lwkt_switch(). 950 * 951 * (self contained on a per cpu basis) 952 */ 953 void 954 lwkt_yield(void) 955 { 956 lwkt_schedule_self(curthread); 957 lwkt_switch(); 958 } 959 960 /* 961 * Generic schedule. Possibly schedule threads belonging to other cpus and 962 * deal with threads that might be blocked on a wait queue. 963 * 964 * We have a little helper inline function which does additional work after 965 * the thread has been enqueued, including dealing with preemption and 966 * setting need_lwkt_resched() (which prevents the kernel from returning 967 * to userland until it has processed higher priority threads). 968 * 969 * It is possible for this routine to be called after a failed _enqueue 970 * (due to the target thread migrating, sleeping, or otherwise blocked). 971 * We have to check that the thread is actually on the run queue! 972 */ 973 static __inline 974 void 975 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri) 976 { 977 if (ntd->td_flags & TDF_RUNQ) { 978 if (ntd->td_preemptable) { 979 ntd->td_preemptable(ntd, cpri); /* YYY +token */ 980 } else if ((ntd->td_flags & TDF_NORESCHED) == 0 && 981 (ntd->td_pri & TDPRI_MASK) > (gd->gd_curthread->td_pri & TDPRI_MASK) 982 ) { 983 need_lwkt_resched(); 984 } 985 } 986 } 987 988 void 989 lwkt_schedule(thread_t td) 990 { 991 globaldata_t mygd = mycpu; 992 993 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!")); 994 crit_enter_gd(mygd); 995 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0); 996 if (td == mygd->gd_curthread) { 997 _lwkt_enqueue(td); 998 } else { 999 /* 1000 * If we own the thread, there is no race (since we are in a 1001 * critical section). If we do not own the thread there might 1002 * be a race but the target cpu will deal with it. 1003 */ 1004 #ifdef SMP 1005 if (td->td_gd == mygd) { 1006 _lwkt_enqueue(td); 1007 _lwkt_schedule_post(mygd, td, TDPRI_CRIT); 1008 } else { 1009 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td); 1010 } 1011 #else 1012 _lwkt_enqueue(td); 1013 _lwkt_schedule_post(mygd, td, TDPRI_CRIT); 1014 #endif 1015 } 1016 crit_exit_gd(mygd); 1017 } 1018 1019 #ifdef SMP 1020 1021 /* 1022 * Thread migration using a 'Pull' method. The thread may or may not be 1023 * the current thread. It MUST be descheduled and in a stable state. 1024 * lwkt_giveaway() must be called on the cpu owning the thread. 1025 * 1026 * At any point after lwkt_giveaway() is called, the target cpu may 1027 * 'pull' the thread by calling lwkt_acquire(). 1028 * 1029 * MPSAFE - must be called under very specific conditions. 1030 */ 1031 void 1032 lwkt_giveaway(thread_t td) 1033 { 1034 globaldata_t gd = mycpu; 1035 1036 crit_enter_gd(gd); 1037 KKASSERT(td->td_gd == gd); 1038 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); 1039 td->td_flags |= TDF_MIGRATING; 1040 crit_exit_gd(gd); 1041 } 1042 1043 void 1044 lwkt_acquire(thread_t td) 1045 { 1046 globaldata_t gd; 1047 globaldata_t mygd; 1048 1049 KKASSERT(td->td_flags & TDF_MIGRATING); 1050 gd = td->td_gd; 1051 mygd = mycpu; 1052 if (gd != mycpu) { 1053 cpu_lfence(); 1054 KKASSERT((td->td_flags & TDF_RUNQ) == 0); 1055 crit_enter_gd(mygd); 1056 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) 1057 cpu_lfence(); 1058 td->td_gd = mygd; 1059 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1060 td->td_flags &= ~TDF_MIGRATING; 1061 crit_exit_gd(mygd); 1062 } else { 1063 crit_enter_gd(mygd); 1064 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1065 td->td_flags &= ~TDF_MIGRATING; 1066 crit_exit_gd(mygd); 1067 } 1068 } 1069 1070 #endif 1071 1072 /* 1073 * Generic deschedule. Descheduling threads other then your own should be 1074 * done only in carefully controlled circumstances. Descheduling is 1075 * asynchronous. 1076 * 1077 * This function may block if the cpu has run out of messages. 1078 */ 1079 void 1080 lwkt_deschedule(thread_t td) 1081 { 1082 crit_enter(); 1083 #ifdef SMP 1084 if (td == curthread) { 1085 _lwkt_dequeue(td); 1086 } else { 1087 if (td->td_gd == mycpu) { 1088 _lwkt_dequeue(td); 1089 } else { 1090 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td); 1091 } 1092 } 1093 #else 1094 _lwkt_dequeue(td); 1095 #endif 1096 crit_exit(); 1097 } 1098 1099 /* 1100 * Set the target thread's priority. This routine does not automatically 1101 * switch to a higher priority thread, LWKT threads are not designed for 1102 * continuous priority changes. Yield if you want to switch. 1103 * 1104 * We have to retain the critical section count which uses the high bits 1105 * of the td_pri field. The specified priority may also indicate zero or 1106 * more critical sections by adding TDPRI_CRIT*N. 1107 * 1108 * Note that we requeue the thread whether it winds up on a different runq 1109 * or not. uio_yield() depends on this and the routine is not normally 1110 * called with the same priority otherwise. 1111 */ 1112 void 1113 lwkt_setpri(thread_t td, int pri) 1114 { 1115 KKASSERT(pri >= 0); 1116 KKASSERT(td->td_gd == mycpu); 1117 crit_enter(); 1118 if (td->td_flags & TDF_RUNQ) { 1119 _lwkt_dequeue(td); 1120 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri; 1121 _lwkt_enqueue(td); 1122 } else { 1123 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri; 1124 } 1125 crit_exit(); 1126 } 1127 1128 void 1129 lwkt_setpri_self(int pri) 1130 { 1131 thread_t td = curthread; 1132 1133 KKASSERT(pri >= 0 && pri <= TDPRI_MAX); 1134 crit_enter(); 1135 if (td->td_flags & TDF_RUNQ) { 1136 _lwkt_dequeue(td); 1137 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri; 1138 _lwkt_enqueue(td); 1139 } else { 1140 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri; 1141 } 1142 crit_exit(); 1143 } 1144 1145 /* 1146 * Determine if there is a runnable thread at a higher priority then 1147 * the current thread. lwkt_setpri() does not check this automatically. 1148 * Return 1 if there is, 0 if there isn't. 1149 * 1150 * Example: if bit 31 of runqmask is set and the current thread is priority 1151 * 30, then we wind up checking the mask: 0x80000000 against 0x7fffffff. 1152 * 1153 * If nq reaches 31 the shift operation will overflow to 0 and we will wind 1154 * up comparing against 0xffffffff, a comparison that will always be false. 1155 */ 1156 int 1157 lwkt_checkpri_self(void) 1158 { 1159 globaldata_t gd = mycpu; 1160 thread_t td = gd->gd_curthread; 1161 int nq = td->td_pri & TDPRI_MASK; 1162 1163 while (gd->gd_runqmask > (__uint32_t)(2 << nq) - 1) { 1164 if (TAILQ_FIRST(&gd->gd_tdrunq[nq + 1])) 1165 return(1); 1166 ++nq; 1167 } 1168 return(0); 1169 } 1170 1171 /* 1172 * Migrate the current thread to the specified cpu. 1173 * 1174 * This is accomplished by descheduling ourselves from the current cpu, 1175 * moving our thread to the tdallq of the target cpu, IPI messaging the 1176 * target cpu, and switching out. TDF_MIGRATING prevents scheduling 1177 * races while the thread is being migrated. 1178 */ 1179 #ifdef SMP 1180 static void lwkt_setcpu_remote(void *arg); 1181 #endif 1182 1183 void 1184 lwkt_setcpu_self(globaldata_t rgd) 1185 { 1186 #ifdef SMP 1187 thread_t td = curthread; 1188 1189 if (td->td_gd != rgd) { 1190 crit_enter_quick(td); 1191 td->td_flags |= TDF_MIGRATING; 1192 lwkt_deschedule_self(td); 1193 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1194 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td); 1195 lwkt_switch(); 1196 /* we are now on the target cpu */ 1197 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); 1198 crit_exit_quick(td); 1199 } 1200 #endif 1201 } 1202 1203 void 1204 lwkt_migratecpu(int cpuid) 1205 { 1206 #ifdef SMP 1207 globaldata_t rgd; 1208 1209 rgd = globaldata_find(cpuid); 1210 lwkt_setcpu_self(rgd); 1211 #endif 1212 } 1213 1214 /* 1215 * Remote IPI for cpu migration (called while in a critical section so we 1216 * do not have to enter another one). The thread has already been moved to 1217 * our cpu's allq, but we must wait for the thread to be completely switched 1218 * out on the originating cpu before we schedule it on ours or the stack 1219 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD 1220 * change to main memory. 1221 * 1222 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races 1223 * against wakeups. It is best if this interface is used only when there 1224 * are no pending events that might try to schedule the thread. 1225 */ 1226 #ifdef SMP 1227 static void 1228 lwkt_setcpu_remote(void *arg) 1229 { 1230 thread_t td = arg; 1231 globaldata_t gd = mycpu; 1232 1233 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) 1234 cpu_lfence(); 1235 td->td_gd = gd; 1236 cpu_sfence(); 1237 td->td_flags &= ~TDF_MIGRATING; 1238 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0); 1239 _lwkt_enqueue(td); 1240 } 1241 #endif 1242 1243 struct lwp * 1244 lwkt_preempted_proc(void) 1245 { 1246 thread_t td = curthread; 1247 while (td->td_preempted) 1248 td = td->td_preempted; 1249 return(td->td_lwp); 1250 } 1251 1252 /* 1253 * Create a kernel process/thread/whatever. It shares it's address space 1254 * with proc0 - ie: kernel only. 1255 * 1256 * NOTE! By default new threads are created with the MP lock held. A 1257 * thread which does not require the MP lock should release it by calling 1258 * rel_mplock() at the start of the new thread. 1259 */ 1260 int 1261 lwkt_create(void (*func)(void *), void *arg, 1262 struct thread **tdp, thread_t template, int tdflags, int cpu, 1263 const char *fmt, ...) 1264 { 1265 thread_t td; 1266 __va_list ap; 1267 1268 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu, 1269 tdflags); 1270 if (tdp) 1271 *tdp = td; 1272 cpu_set_thread_handler(td, lwkt_exit, func, arg); 1273 1274 /* 1275 * Set up arg0 for 'ps' etc 1276 */ 1277 __va_start(ap, fmt); 1278 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap); 1279 __va_end(ap); 1280 1281 /* 1282 * Schedule the thread to run 1283 */ 1284 if ((td->td_flags & TDF_STOPREQ) == 0) 1285 lwkt_schedule(td); 1286 else 1287 td->td_flags &= ~TDF_STOPREQ; 1288 return 0; 1289 } 1290 1291 /* 1292 * kthread_* is specific to the kernel and is not needed by userland. 1293 */ 1294 #ifdef _KERNEL 1295 1296 /* 1297 * Destroy an LWKT thread. Warning! This function is not called when 1298 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and 1299 * uses a different reaping mechanism. 1300 */ 1301 void 1302 lwkt_exit(void) 1303 { 1304 thread_t td = curthread; 1305 globaldata_t gd; 1306 1307 if (td->td_flags & TDF_VERBOSE) 1308 kprintf("kthread %p %s has exited\n", td, td->td_comm); 1309 caps_exit(td); 1310 crit_enter_quick(td); 1311 lwkt_deschedule_self(td); 1312 gd = mycpu; 1313 lwkt_remove_tdallq(td); 1314 if (td->td_flags & TDF_ALLOCATED_THREAD) { 1315 ++gd->gd_tdfreecount; 1316 TAILQ_INSERT_TAIL(&gd->gd_tdfreeq, td, td_threadq); 1317 } 1318 cpu_thread_exit(); 1319 } 1320 1321 void 1322 lwkt_remove_tdallq(thread_t td) 1323 { 1324 KKASSERT(td->td_gd == mycpu); 1325 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1326 } 1327 1328 #endif /* _KERNEL */ 1329 1330 void 1331 crit_panic(void) 1332 { 1333 thread_t td = curthread; 1334 int lpri = td->td_pri; 1335 1336 td->td_pri = 0; 1337 panic("td_pri is/would-go negative! %p %d", td, lpri); 1338 } 1339 1340 #ifdef SMP 1341 1342 /* 1343 * Called from debugger/panic on cpus which have been stopped. We must still 1344 * process the IPIQ while stopped, even if we were stopped while in a critical 1345 * section (XXX). 1346 * 1347 * If we are dumping also try to process any pending interrupts. This may 1348 * or may not work depending on the state of the cpu at the point it was 1349 * stopped. 1350 */ 1351 void 1352 lwkt_smp_stopped(void) 1353 { 1354 globaldata_t gd = mycpu; 1355 1356 crit_enter_gd(gd); 1357 if (dumping) { 1358 lwkt_process_ipiq(); 1359 splz(); 1360 } else { 1361 lwkt_process_ipiq(); 1362 } 1363 crit_exit_gd(gd); 1364 } 1365 1366 /* 1367 * get_mplock() calls this routine if it is unable to obtain the MP lock. 1368 * get_mplock() has already incremented td_mpcount. We must block and 1369 * not return until giant is held. 1370 * 1371 * All we have to do is lwkt_switch() away. The LWKT scheduler will not 1372 * reschedule the thread until it can obtain the giant lock for it. 1373 */ 1374 void 1375 lwkt_mp_lock_contested(void) 1376 { 1377 #ifdef _KERNEL 1378 loggiant(beg); 1379 #endif 1380 lwkt_switch(); 1381 #ifdef _KERNEL 1382 loggiant(end); 1383 #endif 1384 } 1385 1386 #endif 1387