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 1.112 2008/03/01 06:21:28 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 static int lwkt_use_spin_port; 107 108 #ifdef _KERNEL 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 SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, ""); 117 #ifdef INVARIANTS 118 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, ""); 119 #endif 120 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, ""); 121 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, ""); 122 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, ""); 123 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, ""); 124 #ifdef INVARIANTS 125 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW, 126 &token_contention_count, 0, "spinning due to token contention"); 127 SYSCTL_QUAD(_lwkt, OID_AUTO, mplock_contention_count, CTLFLAG_RW, 128 &mplock_contention_count, 0, "spinning due to MPLOCK contention"); 129 #endif 130 #endif 131 132 /* 133 * Kernel Trace 134 */ 135 #ifdef _KERNEL 136 137 #if !defined(KTR_GIANT_CONTENTION) 138 #define KTR_GIANT_CONTENTION KTR_ALL 139 #endif 140 141 KTR_INFO_MASTER(giant); 142 KTR_INFO(KTR_GIANT_CONTENTION, giant, beg, 0, "thread=%p", sizeof(void *)); 143 KTR_INFO(KTR_GIANT_CONTENTION, giant, end, 1, "thread=%p", sizeof(void *)); 144 145 #define loggiant(name) KTR_LOG(giant_ ## name, curthread) 146 147 #endif 148 149 /* 150 * These helper procedures handle the runq, they can only be called from 151 * within a critical section. 152 * 153 * WARNING! Prior to SMP being brought up it is possible to enqueue and 154 * dequeue threads belonging to other cpus, so be sure to use td->td_gd 155 * instead of 'mycpu' when referencing the globaldata structure. Once 156 * SMP live enqueuing and dequeueing only occurs on the current cpu. 157 */ 158 static __inline 159 void 160 _lwkt_dequeue(thread_t td) 161 { 162 if (td->td_flags & TDF_RUNQ) { 163 int nq = td->td_pri & TDPRI_MASK; 164 struct globaldata *gd = td->td_gd; 165 166 td->td_flags &= ~TDF_RUNQ; 167 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq); 168 /* runqmask is passively cleaned up by the switcher */ 169 } 170 } 171 172 static __inline 173 void 174 _lwkt_enqueue(thread_t td) 175 { 176 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_TSLEEPQ|TDF_BLOCKQ)) == 0) { 177 int nq = td->td_pri & TDPRI_MASK; 178 struct globaldata *gd = td->td_gd; 179 180 td->td_flags |= TDF_RUNQ; 181 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq); 182 gd->gd_runqmask |= 1 << nq; 183 } 184 } 185 186 /* 187 * Schedule a thread to run. As the current thread we can always safely 188 * schedule ourselves, and a shortcut procedure is provided for that 189 * function. 190 * 191 * (non-blocking, self contained on a per cpu basis) 192 */ 193 void 194 lwkt_schedule_self(thread_t td) 195 { 196 crit_enter_quick(td); 197 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!")); 198 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0); 199 _lwkt_enqueue(td); 200 crit_exit_quick(td); 201 } 202 203 /* 204 * Deschedule a thread. 205 * 206 * (non-blocking, self contained on a per cpu basis) 207 */ 208 void 209 lwkt_deschedule_self(thread_t td) 210 { 211 crit_enter_quick(td); 212 _lwkt_dequeue(td); 213 crit_exit_quick(td); 214 } 215 216 #ifdef _KERNEL 217 218 /* 219 * LWKTs operate on a per-cpu basis 220 * 221 * WARNING! Called from early boot, 'mycpu' may not work yet. 222 */ 223 void 224 lwkt_gdinit(struct globaldata *gd) 225 { 226 int i; 227 228 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i) 229 TAILQ_INIT(&gd->gd_tdrunq[i]); 230 gd->gd_runqmask = 0; 231 TAILQ_INIT(&gd->gd_tdallq); 232 } 233 234 #endif /* _KERNEL */ 235 236 /* 237 * Create a new thread. The thread must be associated with a process context 238 * or LWKT start address before it can be scheduled. If the target cpu is 239 * -1 the thread will be created on the current cpu. 240 * 241 * If you intend to create a thread without a process context this function 242 * does everything except load the startup and switcher function. 243 */ 244 thread_t 245 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags) 246 { 247 void *stack; 248 globaldata_t gd = mycpu; 249 250 if (td == NULL) { 251 crit_enter_gd(gd); 252 if (gd->gd_tdfreecount > 0) { 253 --gd->gd_tdfreecount; 254 td = TAILQ_FIRST(&gd->gd_tdfreeq); 255 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0, 256 ("lwkt_alloc_thread: unexpected NULL or corrupted td")); 257 TAILQ_REMOVE(&gd->gd_tdfreeq, td, td_threadq); 258 crit_exit_gd(gd); 259 flags |= td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD); 260 } else { 261 crit_exit_gd(gd); 262 #ifdef _KERNEL 263 td = zalloc(thread_zone); 264 #else 265 td = malloc(sizeof(struct thread)); 266 #endif 267 td->td_kstack = NULL; 268 td->td_kstack_size = 0; 269 flags |= TDF_ALLOCATED_THREAD; 270 } 271 } 272 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) { 273 if (flags & TDF_ALLOCATED_STACK) { 274 #ifdef _KERNEL 275 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size); 276 #else 277 libcaps_free_stack(stack, td->td_kstack_size); 278 #endif 279 stack = NULL; 280 } 281 } 282 if (stack == NULL) { 283 #ifdef _KERNEL 284 stack = (void *)kmem_alloc(&kernel_map, stksize); 285 #else 286 stack = libcaps_alloc_stack(stksize); 287 #endif 288 flags |= TDF_ALLOCATED_STACK; 289 } 290 if (cpu < 0) 291 lwkt_init_thread(td, stack, stksize, flags, mycpu); 292 else 293 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu)); 294 return(td); 295 } 296 297 #ifdef _KERNEL 298 299 /* 300 * Initialize a preexisting thread structure. This function is used by 301 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread. 302 * 303 * All threads start out in a critical section at a priority of 304 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as 305 * appropriate. This function may send an IPI message when the 306 * requested cpu is not the current cpu and consequently gd_tdallq may 307 * not be initialized synchronously from the point of view of the originating 308 * cpu. 309 * 310 * NOTE! we have to be careful in regards to creating threads for other cpus 311 * if SMP has not yet been activated. 312 */ 313 #ifdef SMP 314 315 static void 316 lwkt_init_thread_remote(void *arg) 317 { 318 thread_t td = arg; 319 320 /* 321 * Protected by critical section held by IPI dispatch 322 */ 323 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq); 324 } 325 326 #endif 327 328 void 329 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags, 330 struct globaldata *gd) 331 { 332 globaldata_t mygd = mycpu; 333 334 bzero(td, sizeof(struct thread)); 335 td->td_kstack = stack; 336 td->td_kstack_size = stksize; 337 td->td_flags = flags; 338 td->td_gd = gd; 339 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT; 340 #ifdef SMP 341 if ((flags & TDF_MPSAFE) == 0) 342 td->td_mpcount = 1; 343 #endif 344 if (lwkt_use_spin_port) 345 lwkt_initport_spin(&td->td_msgport); 346 else 347 lwkt_initport_thread(&td->td_msgport, td); 348 pmap_init_thread(td); 349 #ifdef SMP 350 /* 351 * Normally initializing a thread for a remote cpu requires sending an 352 * IPI. However, the idlethread is setup before the other cpus are 353 * activated so we have to treat it as a special case. XXX manipulation 354 * of gd_tdallq requires the BGL. 355 */ 356 if (gd == mygd || td == &gd->gd_idlethread) { 357 crit_enter_gd(mygd); 358 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); 359 crit_exit_gd(mygd); 360 } else { 361 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td); 362 } 363 #else 364 crit_enter_gd(mygd); 365 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); 366 crit_exit_gd(mygd); 367 #endif 368 } 369 370 #endif /* _KERNEL */ 371 372 void 373 lwkt_set_comm(thread_t td, const char *ctl, ...) 374 { 375 __va_list va; 376 377 __va_start(va, ctl); 378 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va); 379 __va_end(va); 380 } 381 382 void 383 lwkt_hold(thread_t td) 384 { 385 ++td->td_refs; 386 } 387 388 void 389 lwkt_rele(thread_t td) 390 { 391 KKASSERT(td->td_refs > 0); 392 --td->td_refs; 393 } 394 395 #ifdef _KERNEL 396 397 void 398 lwkt_wait_free(thread_t td) 399 { 400 while (td->td_refs) 401 tsleep(td, 0, "tdreap", hz); 402 } 403 404 #endif 405 406 void 407 lwkt_free_thread(thread_t td) 408 { 409 struct globaldata *gd = mycpu; 410 411 KASSERT((td->td_flags & TDF_RUNNING) == 0, 412 ("lwkt_free_thread: did not exit! %p", td)); 413 414 crit_enter_gd(gd); 415 if (gd->gd_tdfreecount < CACHE_NTHREADS && 416 (td->td_flags & TDF_ALLOCATED_THREAD) 417 ) { 418 ++gd->gd_tdfreecount; 419 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq); 420 crit_exit_gd(gd); 421 } else { 422 crit_exit_gd(gd); 423 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) { 424 #ifdef _KERNEL 425 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size); 426 #else 427 libcaps_free_stack(td->td_kstack, td->td_kstack_size); 428 #endif 429 /* gd invalid */ 430 td->td_kstack = NULL; 431 td->td_kstack_size = 0; 432 } 433 if (td->td_flags & TDF_ALLOCATED_THREAD) { 434 #ifdef _KERNEL 435 zfree(thread_zone, td); 436 #else 437 free(td); 438 #endif 439 } 440 } 441 } 442 443 444 /* 445 * Switch to the next runnable lwkt. If no LWKTs are runnable then 446 * switch to the idlethread. Switching must occur within a critical 447 * section to avoid races with the scheduling queue. 448 * 449 * We always have full control over our cpu's run queue. Other cpus 450 * that wish to manipulate our queue must use the cpu_*msg() calls to 451 * talk to our cpu, so a critical section is all that is needed and 452 * the result is very, very fast thread switching. 453 * 454 * The LWKT scheduler uses a fixed priority model and round-robins at 455 * each priority level. User process scheduling is a totally 456 * different beast and LWKT priorities should not be confused with 457 * user process priorities. 458 * 459 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch() 460 * cleans it up. Note that the td_switch() function cannot do anything that 461 * requires the MP lock since the MP lock will have already been setup for 462 * the target thread (not the current thread). It's nice to have a scheduler 463 * that does not need the MP lock to work because it allows us to do some 464 * really cool high-performance MP lock optimizations. 465 * 466 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch() 467 * is not called by the current thread in the preemption case, only when 468 * the preempting thread blocks (in order to return to the original thread). 469 */ 470 void 471 lwkt_switch(void) 472 { 473 globaldata_t gd = mycpu; 474 thread_t td = gd->gd_curthread; 475 thread_t ntd; 476 #ifdef SMP 477 int mpheld; 478 #endif 479 480 /* 481 * Switching from within a 'fast' (non thread switched) interrupt or IPI 482 * is illegal. However, we may have to do it anyway if we hit a fatal 483 * kernel trap or we have paniced. 484 * 485 * If this case occurs save and restore the interrupt nesting level. 486 */ 487 if (gd->gd_intr_nesting_level) { 488 int savegdnest; 489 int savegdtrap; 490 491 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) { 492 panic("lwkt_switch: cannot switch from within " 493 "a fast interrupt, yet, td %p\n", td); 494 } else { 495 savegdnest = gd->gd_intr_nesting_level; 496 savegdtrap = gd->gd_trap_nesting_level; 497 gd->gd_intr_nesting_level = 0; 498 gd->gd_trap_nesting_level = 0; 499 if ((td->td_flags & TDF_PANICWARN) == 0) { 500 td->td_flags |= TDF_PANICWARN; 501 kprintf("Warning: thread switch from interrupt or IPI, " 502 "thread %p (%s)\n", td, td->td_comm); 503 #ifdef DDB 504 db_print_backtrace(); 505 #endif 506 } 507 lwkt_switch(); 508 gd->gd_intr_nesting_level = savegdnest; 509 gd->gd_trap_nesting_level = savegdtrap; 510 return; 511 } 512 } 513 514 /* 515 * Passive release (used to transition from user to kernel mode 516 * when we block or switch rather then when we enter the kernel). 517 * This function is NOT called if we are switching into a preemption 518 * or returning from a preemption. Typically this causes us to lose 519 * our current process designation (if we have one) and become a true 520 * LWKT thread, and may also hand the current process designation to 521 * another process and schedule thread. 522 */ 523 if (td->td_release) 524 td->td_release(td); 525 526 crit_enter_gd(gd); 527 if (td->td_toks) 528 lwkt_relalltokens(td); 529 530 /* 531 * We had better not be holding any spin locks, but don't get into an 532 * endless panic loop. 533 */ 534 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL, 535 ("lwkt_switch: still holding a shared spinlock %p!", 536 gd->gd_spinlock_rd)); 537 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL, 538 ("lwkt_switch: still holding %d exclusive spinlocks!", 539 gd->gd_spinlocks_wr)); 540 541 542 #ifdef SMP 543 /* 544 * td_mpcount cannot be used to determine if we currently hold the 545 * MP lock because get_mplock() will increment it prior to attempting 546 * to get the lock, and switch out if it can't. Our ownership of 547 * the actual lock will remain stable while we are in a critical section 548 * (but, of course, another cpu may own or release the lock so the 549 * actual value of mp_lock is not stable). 550 */ 551 mpheld = MP_LOCK_HELD(); 552 #ifdef INVARIANTS 553 if (td->td_cscount) { 554 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n", 555 td); 556 if (panic_on_cscount) 557 panic("switching while mastering cpusync"); 558 } 559 #endif 560 #endif 561 if ((ntd = td->td_preempted) != NULL) { 562 /* 563 * We had preempted another thread on this cpu, resume the preempted 564 * thread. This occurs transparently, whether the preempted thread 565 * was scheduled or not (it may have been preempted after descheduling 566 * itself). 567 * 568 * We have to setup the MP lock for the original thread after backing 569 * out the adjustment that was made to curthread when the original 570 * was preempted. 571 */ 572 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK); 573 #ifdef SMP 574 if (ntd->td_mpcount && mpheld == 0) { 575 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d", 576 td, ntd, td->td_mpcount, ntd->td_mpcount); 577 } 578 if (ntd->td_mpcount) { 579 td->td_mpcount -= ntd->td_mpcount; 580 KKASSERT(td->td_mpcount >= 0); 581 } 582 #endif 583 ntd->td_flags |= TDF_PREEMPT_DONE; 584 585 /* 586 * XXX. The interrupt may have woken a thread up, we need to properly 587 * set the reschedule flag if the originally interrupted thread is at 588 * a lower priority. 589 */ 590 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1) 591 need_lwkt_resched(); 592 /* YYY release mp lock on switchback if original doesn't need it */ 593 } else { 594 /* 595 * Priority queue / round-robin at each priority. Note that user 596 * processes run at a fixed, low priority and the user process 597 * scheduler deals with interactions between user processes 598 * by scheduling and descheduling them from the LWKT queue as 599 * necessary. 600 * 601 * We have to adjust the MP lock for the target thread. If we 602 * need the MP lock and cannot obtain it we try to locate a 603 * thread that does not need the MP lock. If we cannot, we spin 604 * instead of HLT. 605 * 606 * A similar issue exists for the tokens held by the target thread. 607 * If we cannot obtain ownership of the tokens we cannot immediately 608 * schedule the thread. 609 */ 610 611 /* 612 * If an LWKT reschedule was requested, well that is what we are 613 * doing now so clear it. 614 */ 615 clear_lwkt_resched(); 616 again: 617 if (gd->gd_runqmask) { 618 int nq = bsrl(gd->gd_runqmask); 619 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) { 620 gd->gd_runqmask &= ~(1 << nq); 621 goto again; 622 } 623 #ifdef SMP 624 /* 625 * THREAD SELECTION FOR AN SMP MACHINE BUILD 626 * 627 * If the target needs the MP lock and we couldn't get it, 628 * or if the target is holding tokens and we could not 629 * gain ownership of the tokens, continue looking for a 630 * thread to schedule and spin instead of HLT if we can't. 631 * 632 * NOTE: the mpheld variable invalid after this conditional, it 633 * can change due to both cpu_try_mplock() returning success 634 * AND interactions in lwkt_getalltokens() due to the fact that 635 * we are trying to check the mpcount of a thread other then 636 * the current thread. Because of this, if the current thread 637 * is not holding td_mpcount, an IPI indirectly run via 638 * lwkt_getalltokens() can obtain and release the MP lock and 639 * cause the core MP lock to be released. 640 */ 641 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) || 642 (ntd->td_toks && lwkt_getalltokens(ntd) == 0) 643 ) { 644 u_int32_t rqmask = gd->gd_runqmask; 645 646 mpheld = MP_LOCK_HELD(); 647 ntd = NULL; 648 while (rqmask) { 649 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) { 650 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) { 651 /* spinning due to MP lock being held */ 652 #ifdef INVARIANTS 653 ++mplock_contention_count; 654 #endif 655 /* mplock still not held, 'mpheld' still valid */ 656 continue; 657 } 658 659 /* 660 * mpheld state invalid after getalltokens call returns 661 * failure, but the variable is only needed for 662 * the loop. 663 */ 664 if (ntd->td_toks && !lwkt_getalltokens(ntd)) { 665 /* spinning due to token contention */ 666 #ifdef INVARIANTS 667 ++token_contention_count; 668 #endif 669 mpheld = MP_LOCK_HELD(); 670 continue; 671 } 672 break; 673 } 674 if (ntd) 675 break; 676 rqmask &= ~(1 << nq); 677 nq = bsrl(rqmask); 678 } 679 if (ntd == NULL) { 680 cpu_mplock_contested(); 681 ntd = &gd->gd_idlethread; 682 ntd->td_flags |= TDF_IDLE_NOHLT; 683 goto using_idle_thread; 684 } else { 685 ++gd->gd_cnt.v_swtch; 686 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq); 687 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq); 688 } 689 } else { 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 } 694 #else 695 /* 696 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to 697 * worry about tokens or the BGL. However, we still have 698 * to call lwkt_getalltokens() in order to properly detect 699 * stale tokens. This call cannot fail for a UP build! 700 */ 701 lwkt_getalltokens(ntd); 702 ++gd->gd_cnt.v_swtch; 703 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq); 704 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq); 705 #endif 706 } else { 707 /* 708 * We have nothing to run but only let the idle loop halt 709 * the cpu if there are no pending interrupts. 710 */ 711 ntd = &gd->gd_idlethread; 712 if (gd->gd_reqflags & RQF_IDLECHECK_MASK) 713 ntd->td_flags |= TDF_IDLE_NOHLT; 714 #ifdef SMP 715 using_idle_thread: 716 /* 717 * The idle thread should not be holding the MP lock unless we 718 * are trapping in the kernel or in a panic. Since we select the 719 * idle thread unconditionally when no other thread is available, 720 * if the MP lock is desired during a panic or kernel trap, we 721 * have to loop in the scheduler until we get it. 722 */ 723 if (ntd->td_mpcount) { 724 mpheld = MP_LOCK_HELD(); 725 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) { 726 panic("Idle thread %p was holding the BGL!", ntd); 727 } else if (mpheld == 0) { 728 cpu_mplock_contested(); 729 goto again; 730 } 731 } 732 #endif 733 } 734 } 735 KASSERT(ntd->td_pri >= TDPRI_CRIT, 736 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri)); 737 738 /* 739 * Do the actual switch. If the new target does not need the MP lock 740 * and we are holding it, release the MP lock. If the new target requires 741 * the MP lock we have already acquired it for the target. 742 */ 743 #ifdef SMP 744 if (ntd->td_mpcount == 0 ) { 745 if (MP_LOCK_HELD()) 746 cpu_rel_mplock(); 747 } else { 748 ASSERT_MP_LOCK_HELD(ntd); 749 } 750 #endif 751 if (td != ntd) { 752 ++switch_count; 753 td->td_switch(ntd); 754 } 755 /* NOTE: current cpu may have changed after switch */ 756 crit_exit_quick(td); 757 } 758 759 /* 760 * Request that the target thread preempt the current thread. Preemption 761 * only works under a specific set of conditions: 762 * 763 * - We are not preempting ourselves 764 * - The target thread is owned by the current cpu 765 * - We are not currently being preempted 766 * - The target is not currently being preempted 767 * - We are not holding any spin locks 768 * - The target thread is not holding any tokens 769 * - We are able to satisfy the target's MP lock requirements (if any). 770 * 771 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically 772 * this is called via lwkt_schedule() through the td_preemptable callback. 773 * critpri is the managed critical priority that we should ignore in order 774 * to determine whether preemption is possible (aka usually just the crit 775 * priority of lwkt_schedule() itself). 776 * 777 * XXX at the moment we run the target thread in a critical section during 778 * the preemption in order to prevent the target from taking interrupts 779 * that *WE* can't. Preemption is strictly limited to interrupt threads 780 * and interrupt-like threads, outside of a critical section, and the 781 * preempted source thread will be resumed the instant the target blocks 782 * whether or not the source is scheduled (i.e. preemption is supposed to 783 * be as transparent as possible). 784 * 785 * The target thread inherits our MP count (added to its own) for the 786 * duration of the preemption in order to preserve the atomicy of the 787 * MP lock during the preemption. Therefore, any preempting targets must be 788 * careful in regards to MP assertions. Note that the MP count may be 789 * out of sync with the physical mp_lock, but we do not have to preserve 790 * the original ownership of the lock if it was out of synch (that is, we 791 * can leave it synchronized on return). 792 */ 793 void 794 lwkt_preempt(thread_t ntd, int critpri) 795 { 796 struct globaldata *gd = mycpu; 797 thread_t td; 798 #ifdef SMP 799 int mpheld; 800 int savecnt; 801 #endif 802 803 /* 804 * The caller has put us in a critical section. We can only preempt 805 * if the caller of the caller was not in a critical section (basically 806 * a local interrupt), as determined by the 'critpri' parameter. We 807 * also can't preempt if the caller is holding any spinlocks (even if 808 * he isn't in a critical section). This also handles the tokens test. 809 * 810 * YYY The target thread must be in a critical section (else it must 811 * inherit our critical section? I dunno yet). 812 * 813 * Set need_lwkt_resched() unconditionally for now YYY. 814 */ 815 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri)); 816 817 td = gd->gd_curthread; 818 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) { 819 ++preempt_miss; 820 return; 821 } 822 if ((td->td_pri & ~TDPRI_MASK) > critpri) { 823 ++preempt_miss; 824 need_lwkt_resched(); 825 return; 826 } 827 #ifdef SMP 828 if (ntd->td_gd != gd) { 829 ++preempt_miss; 830 need_lwkt_resched(); 831 return; 832 } 833 #endif 834 /* 835 * Take the easy way out and do not preempt if we are holding 836 * any spinlocks. We could test whether the thread(s) being 837 * preempted interlock against the target thread's tokens and whether 838 * we can get all the target thread's tokens, but this situation 839 * should not occur very often so its easier to simply not preempt. 840 * Also, plain spinlocks are impossible to figure out at this point so 841 * just don't preempt. 842 * 843 * Do not try to preempt if the target thread is holding any tokens. 844 * We could try to acquire the tokens but this case is so rare there 845 * is no need to support it. 846 */ 847 if (gd->gd_spinlock_rd || gd->gd_spinlocks_wr) { 848 ++preempt_miss; 849 need_lwkt_resched(); 850 return; 851 } 852 if (ntd->td_toks) { 853 ++preempt_miss; 854 need_lwkt_resched(); 855 return; 856 } 857 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) { 858 ++preempt_weird; 859 need_lwkt_resched(); 860 return; 861 } 862 if (ntd->td_preempted) { 863 ++preempt_hit; 864 need_lwkt_resched(); 865 return; 866 } 867 #ifdef SMP 868 /* 869 * note: an interrupt might have occured just as we were transitioning 870 * to or from the MP lock. In this case td_mpcount will be pre-disposed 871 * (non-zero) but not actually synchronized with the actual state of the 872 * lock. We can use it to imply an MP lock requirement for the 873 * preemption but we cannot use it to test whether we hold the MP lock 874 * or not. 875 */ 876 savecnt = td->td_mpcount; 877 mpheld = MP_LOCK_HELD(); 878 ntd->td_mpcount += td->td_mpcount; 879 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) { 880 ntd->td_mpcount -= td->td_mpcount; 881 ++preempt_miss; 882 need_lwkt_resched(); 883 return; 884 } 885 #endif 886 887 /* 888 * Since we are able to preempt the current thread, there is no need to 889 * call need_lwkt_resched(). 890 */ 891 ++preempt_hit; 892 ntd->td_preempted = td; 893 td->td_flags |= TDF_PREEMPT_LOCK; 894 td->td_switch(ntd); 895 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE)); 896 #ifdef SMP 897 KKASSERT(savecnt == td->td_mpcount); 898 mpheld = MP_LOCK_HELD(); 899 if (mpheld && td->td_mpcount == 0) 900 cpu_rel_mplock(); 901 else if (mpheld == 0 && td->td_mpcount) 902 panic("lwkt_preempt(): MP lock was not held through"); 903 #endif 904 ntd->td_preempted = NULL; 905 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE); 906 } 907 908 /* 909 * Yield our thread while higher priority threads are pending. This is 910 * typically called when we leave a critical section but it can be safely 911 * called while we are in a critical section. 912 * 913 * This function will not generally yield to equal priority threads but it 914 * can occur as a side effect. Note that lwkt_switch() is called from 915 * inside the critical section to prevent its own crit_exit() from reentering 916 * lwkt_yield_quick(). 917 * 918 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint 919 * came along but was blocked and made pending. 920 * 921 * (self contained on a per cpu basis) 922 */ 923 void 924 lwkt_yield_quick(void) 925 { 926 globaldata_t gd = mycpu; 927 thread_t td = gd->gd_curthread; 928 929 /* 930 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear 931 * it with a non-zero cpl then we might not wind up calling splz after 932 * a task switch when the critical section is exited even though the 933 * new task could accept the interrupt. 934 * 935 * XXX from crit_exit() only called after last crit section is released. 936 * If called directly will run splz() even if in a critical section. 937 * 938 * td_nest_count prevent deep nesting via splz() or doreti(). Note that 939 * except for this special case, we MUST call splz() here to handle any 940 * pending ints, particularly after we switch, or we might accidently 941 * halt the cpu with interrupts pending. 942 */ 943 if (gd->gd_reqflags && td->td_nest_count < 2) 944 splz(); 945 946 /* 947 * YYY enabling will cause wakeup() to task-switch, which really 948 * confused the old 4.x code. This is a good way to simulate 949 * preemption and MP without actually doing preemption or MP, because a 950 * lot of code assumes that wakeup() does not block. 951 */ 952 if (untimely_switch && td->td_nest_count == 0 && 953 gd->gd_intr_nesting_level == 0 954 ) { 955 crit_enter_quick(td); 956 /* 957 * YYY temporary hacks until we disassociate the userland scheduler 958 * from the LWKT scheduler. 959 */ 960 if (td->td_flags & TDF_RUNQ) { 961 lwkt_switch(); /* will not reenter yield function */ 962 } else { 963 lwkt_schedule_self(td); /* make sure we are scheduled */ 964 lwkt_switch(); /* will not reenter yield function */ 965 lwkt_deschedule_self(td); /* make sure we are descheduled */ 966 } 967 crit_exit_noyield(td); 968 } 969 } 970 971 /* 972 * This implements a normal yield which, unlike _quick, will yield to equal 973 * priority threads as well. Note that gd_reqflags tests will be handled by 974 * the crit_exit() call in lwkt_switch(). 975 * 976 * (self contained on a per cpu basis) 977 */ 978 void 979 lwkt_yield(void) 980 { 981 lwkt_schedule_self(curthread); 982 lwkt_switch(); 983 } 984 985 /* 986 * Generic schedule. Possibly schedule threads belonging to other cpus and 987 * deal with threads that might be blocked on a wait queue. 988 * 989 * We have a little helper inline function which does additional work after 990 * the thread has been enqueued, including dealing with preemption and 991 * setting need_lwkt_resched() (which prevents the kernel from returning 992 * to userland until it has processed higher priority threads). 993 * 994 * It is possible for this routine to be called after a failed _enqueue 995 * (due to the target thread migrating, sleeping, or otherwise blocked). 996 * We have to check that the thread is actually on the run queue! 997 */ 998 static __inline 999 void 1000 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri) 1001 { 1002 if (ntd->td_flags & TDF_RUNQ) { 1003 if (ntd->td_preemptable) { 1004 ntd->td_preemptable(ntd, cpri); /* YYY +token */ 1005 } else if ((ntd->td_flags & TDF_NORESCHED) == 0 && 1006 (ntd->td_pri & TDPRI_MASK) > (gd->gd_curthread->td_pri & TDPRI_MASK) 1007 ) { 1008 need_lwkt_resched(); 1009 } 1010 } 1011 } 1012 1013 void 1014 lwkt_schedule(thread_t td) 1015 { 1016 globaldata_t mygd = mycpu; 1017 1018 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!")); 1019 crit_enter_gd(mygd); 1020 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0); 1021 if (td == mygd->gd_curthread) { 1022 _lwkt_enqueue(td); 1023 } else { 1024 /* 1025 * If we own the thread, there is no race (since we are in a 1026 * critical section). If we do not own the thread there might 1027 * be a race but the target cpu will deal with it. 1028 */ 1029 #ifdef SMP 1030 if (td->td_gd == mygd) { 1031 _lwkt_enqueue(td); 1032 _lwkt_schedule_post(mygd, td, TDPRI_CRIT); 1033 } else { 1034 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td); 1035 } 1036 #else 1037 _lwkt_enqueue(td); 1038 _lwkt_schedule_post(mygd, td, TDPRI_CRIT); 1039 #endif 1040 } 1041 crit_exit_gd(mygd); 1042 } 1043 1044 #ifdef SMP 1045 1046 /* 1047 * Thread migration using a 'Pull' method. The thread may or may not be 1048 * the current thread. It MUST be descheduled and in a stable state. 1049 * lwkt_giveaway() must be called on the cpu owning the thread. 1050 * 1051 * At any point after lwkt_giveaway() is called, the target cpu may 1052 * 'pull' the thread by calling lwkt_acquire(). 1053 * 1054 * MPSAFE - must be called under very specific conditions. 1055 */ 1056 void 1057 lwkt_giveaway(thread_t td) 1058 { 1059 globaldata_t gd = mycpu; 1060 1061 crit_enter_gd(gd); 1062 KKASSERT(td->td_gd == gd); 1063 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); 1064 td->td_flags |= TDF_MIGRATING; 1065 crit_exit_gd(gd); 1066 } 1067 1068 void 1069 lwkt_acquire(thread_t td) 1070 { 1071 globaldata_t gd; 1072 globaldata_t mygd; 1073 1074 KKASSERT(td->td_flags & TDF_MIGRATING); 1075 gd = td->td_gd; 1076 mygd = mycpu; 1077 if (gd != mycpu) { 1078 cpu_lfence(); 1079 KKASSERT((td->td_flags & TDF_RUNQ) == 0); 1080 crit_enter_gd(mygd); 1081 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) 1082 cpu_lfence(); 1083 td->td_gd = mygd; 1084 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1085 td->td_flags &= ~TDF_MIGRATING; 1086 crit_exit_gd(mygd); 1087 } else { 1088 crit_enter_gd(mygd); 1089 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1090 td->td_flags &= ~TDF_MIGRATING; 1091 crit_exit_gd(mygd); 1092 } 1093 } 1094 1095 #endif 1096 1097 /* 1098 * Generic deschedule. Descheduling threads other then your own should be 1099 * done only in carefully controlled circumstances. Descheduling is 1100 * asynchronous. 1101 * 1102 * This function may block if the cpu has run out of messages. 1103 */ 1104 void 1105 lwkt_deschedule(thread_t td) 1106 { 1107 crit_enter(); 1108 #ifdef SMP 1109 if (td == curthread) { 1110 _lwkt_dequeue(td); 1111 } else { 1112 if (td->td_gd == mycpu) { 1113 _lwkt_dequeue(td); 1114 } else { 1115 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td); 1116 } 1117 } 1118 #else 1119 _lwkt_dequeue(td); 1120 #endif 1121 crit_exit(); 1122 } 1123 1124 /* 1125 * Set the target thread's priority. This routine does not automatically 1126 * switch to a higher priority thread, LWKT threads are not designed for 1127 * continuous priority changes. Yield if you want to switch. 1128 * 1129 * We have to retain the critical section count which uses the high bits 1130 * of the td_pri field. The specified priority may also indicate zero or 1131 * more critical sections by adding TDPRI_CRIT*N. 1132 * 1133 * Note that we requeue the thread whether it winds up on a different runq 1134 * or not. uio_yield() depends on this and the routine is not normally 1135 * called with the same priority otherwise. 1136 */ 1137 void 1138 lwkt_setpri(thread_t td, int pri) 1139 { 1140 KKASSERT(pri >= 0); 1141 KKASSERT(td->td_gd == mycpu); 1142 crit_enter(); 1143 if (td->td_flags & TDF_RUNQ) { 1144 _lwkt_dequeue(td); 1145 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri; 1146 _lwkt_enqueue(td); 1147 } else { 1148 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri; 1149 } 1150 crit_exit(); 1151 } 1152 1153 void 1154 lwkt_setpri_self(int pri) 1155 { 1156 thread_t td = curthread; 1157 1158 KKASSERT(pri >= 0 && pri <= TDPRI_MAX); 1159 crit_enter(); 1160 if (td->td_flags & TDF_RUNQ) { 1161 _lwkt_dequeue(td); 1162 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri; 1163 _lwkt_enqueue(td); 1164 } else { 1165 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri; 1166 } 1167 crit_exit(); 1168 } 1169 1170 /* 1171 * Determine if there is a runnable thread at a higher priority then 1172 * the current thread. lwkt_setpri() does not check this automatically. 1173 * Return 1 if there is, 0 if there isn't. 1174 * 1175 * Example: if bit 31 of runqmask is set and the current thread is priority 1176 * 30, then we wind up checking the mask: 0x80000000 against 0x7fffffff. 1177 * 1178 * If nq reaches 31 the shift operation will overflow to 0 and we will wind 1179 * up comparing against 0xffffffff, a comparison that will always be false. 1180 */ 1181 int 1182 lwkt_checkpri_self(void) 1183 { 1184 globaldata_t gd = mycpu; 1185 thread_t td = gd->gd_curthread; 1186 int nq = td->td_pri & TDPRI_MASK; 1187 1188 while (gd->gd_runqmask > (__uint32_t)(2 << nq) - 1) { 1189 if (TAILQ_FIRST(&gd->gd_tdrunq[nq + 1])) 1190 return(1); 1191 ++nq; 1192 } 1193 return(0); 1194 } 1195 1196 /* 1197 * Migrate the current thread to the specified cpu. 1198 * 1199 * This is accomplished by descheduling ourselves from the current cpu, 1200 * moving our thread to the tdallq of the target cpu, IPI messaging the 1201 * target cpu, and switching out. TDF_MIGRATING prevents scheduling 1202 * races while the thread is being migrated. 1203 */ 1204 #ifdef SMP 1205 static void lwkt_setcpu_remote(void *arg); 1206 #endif 1207 1208 void 1209 lwkt_setcpu_self(globaldata_t rgd) 1210 { 1211 #ifdef SMP 1212 thread_t td = curthread; 1213 1214 if (td->td_gd != rgd) { 1215 crit_enter_quick(td); 1216 td->td_flags |= TDF_MIGRATING; 1217 lwkt_deschedule_self(td); 1218 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1219 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td); 1220 lwkt_switch(); 1221 /* we are now on the target cpu */ 1222 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); 1223 crit_exit_quick(td); 1224 } 1225 #endif 1226 } 1227 1228 void 1229 lwkt_migratecpu(int cpuid) 1230 { 1231 #ifdef SMP 1232 globaldata_t rgd; 1233 1234 rgd = globaldata_find(cpuid); 1235 lwkt_setcpu_self(rgd); 1236 #endif 1237 } 1238 1239 /* 1240 * Remote IPI for cpu migration (called while in a critical section so we 1241 * do not have to enter another one). The thread has already been moved to 1242 * our cpu's allq, but we must wait for the thread to be completely switched 1243 * out on the originating cpu before we schedule it on ours or the stack 1244 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD 1245 * change to main memory. 1246 * 1247 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races 1248 * against wakeups. It is best if this interface is used only when there 1249 * are no pending events that might try to schedule the thread. 1250 */ 1251 #ifdef SMP 1252 static void 1253 lwkt_setcpu_remote(void *arg) 1254 { 1255 thread_t td = arg; 1256 globaldata_t gd = mycpu; 1257 1258 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) 1259 cpu_lfence(); 1260 td->td_gd = gd; 1261 cpu_sfence(); 1262 td->td_flags &= ~TDF_MIGRATING; 1263 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0); 1264 _lwkt_enqueue(td); 1265 } 1266 #endif 1267 1268 struct lwp * 1269 lwkt_preempted_proc(void) 1270 { 1271 thread_t td = curthread; 1272 while (td->td_preempted) 1273 td = td->td_preempted; 1274 return(td->td_lwp); 1275 } 1276 1277 /* 1278 * Create a kernel process/thread/whatever. It shares it's address space 1279 * with proc0 - ie: kernel only. 1280 * 1281 * NOTE! By default new threads are created with the MP lock held. A 1282 * thread which does not require the MP lock should release it by calling 1283 * rel_mplock() at the start of the new thread. 1284 */ 1285 int 1286 lwkt_create(void (*func)(void *), void *arg, 1287 struct thread **tdp, thread_t template, int tdflags, int cpu, 1288 const char *fmt, ...) 1289 { 1290 thread_t td; 1291 __va_list ap; 1292 1293 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu, 1294 tdflags); 1295 if (tdp) 1296 *tdp = td; 1297 cpu_set_thread_handler(td, lwkt_exit, func, arg); 1298 1299 /* 1300 * Set up arg0 for 'ps' etc 1301 */ 1302 __va_start(ap, fmt); 1303 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap); 1304 __va_end(ap); 1305 1306 /* 1307 * Schedule the thread to run 1308 */ 1309 if ((td->td_flags & TDF_STOPREQ) == 0) 1310 lwkt_schedule(td); 1311 else 1312 td->td_flags &= ~TDF_STOPREQ; 1313 return 0; 1314 } 1315 1316 /* 1317 * kthread_* is specific to the kernel and is not needed by userland. 1318 */ 1319 #ifdef _KERNEL 1320 1321 /* 1322 * Destroy an LWKT thread. Warning! This function is not called when 1323 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and 1324 * uses a different reaping mechanism. 1325 */ 1326 void 1327 lwkt_exit(void) 1328 { 1329 thread_t td = curthread; 1330 globaldata_t gd; 1331 1332 if (td->td_flags & TDF_VERBOSE) 1333 kprintf("kthread %p %s has exited\n", td, td->td_comm); 1334 caps_exit(td); 1335 crit_enter_quick(td); 1336 lwkt_deschedule_self(td); 1337 gd = mycpu; 1338 lwkt_remove_tdallq(td); 1339 if (td->td_flags & TDF_ALLOCATED_THREAD) { 1340 ++gd->gd_tdfreecount; 1341 TAILQ_INSERT_TAIL(&gd->gd_tdfreeq, td, td_threadq); 1342 } 1343 cpu_thread_exit(); 1344 } 1345 1346 void 1347 lwkt_remove_tdallq(thread_t td) 1348 { 1349 KKASSERT(td->td_gd == mycpu); 1350 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1351 } 1352 1353 #endif /* _KERNEL */ 1354 1355 void 1356 crit_panic(void) 1357 { 1358 thread_t td = curthread; 1359 int lpri = td->td_pri; 1360 1361 td->td_pri = 0; 1362 panic("td_pri is/would-go negative! %p %d", td, lpri); 1363 } 1364 1365 #ifdef SMP 1366 1367 /* 1368 * Called from debugger/panic on cpus which have been stopped. We must still 1369 * process the IPIQ while stopped, even if we were stopped while in a critical 1370 * section (XXX). 1371 * 1372 * If we are dumping also try to process any pending interrupts. This may 1373 * or may not work depending on the state of the cpu at the point it was 1374 * stopped. 1375 */ 1376 void 1377 lwkt_smp_stopped(void) 1378 { 1379 globaldata_t gd = mycpu; 1380 1381 crit_enter_gd(gd); 1382 if (dumping) { 1383 lwkt_process_ipiq(); 1384 splz(); 1385 } else { 1386 lwkt_process_ipiq(); 1387 } 1388 crit_exit_gd(gd); 1389 } 1390 1391 /* 1392 * get_mplock() calls this routine if it is unable to obtain the MP lock. 1393 * get_mplock() has already incremented td_mpcount. We must block and 1394 * not return until giant is held. 1395 * 1396 * All we have to do is lwkt_switch() away. The LWKT scheduler will not 1397 * reschedule the thread until it can obtain the giant lock for it. 1398 */ 1399 void 1400 lwkt_mp_lock_contested(void) 1401 { 1402 #ifdef _KERNEL 1403 loggiant(beg); 1404 #endif 1405 lwkt_switch(); 1406 #ifdef _KERNEL 1407 loggiant(end); 1408 #endif 1409 } 1410 1411 #endif 1412