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 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/queue.h> 48 #include <sys/sysctl.h> 49 #include <sys/kthread.h> 50 #include <machine/cpu.h> 51 #include <sys/lock.h> 52 #include <sys/caps.h> 53 #include <sys/spinlock.h> 54 #include <sys/ktr.h> 55 56 #include <sys/thread2.h> 57 #include <sys/spinlock2.h> 58 #include <sys/mplock2.h> 59 60 #include <sys/dsched.h> 61 62 #include <vm/vm.h> 63 #include <vm/vm_param.h> 64 #include <vm/vm_kern.h> 65 #include <vm/vm_object.h> 66 #include <vm/vm_page.h> 67 #include <vm/vm_map.h> 68 #include <vm/vm_pager.h> 69 #include <vm/vm_extern.h> 70 71 #include <machine/stdarg.h> 72 #include <machine/smp.h> 73 74 #if !defined(KTR_CTXSW) 75 #define KTR_CTXSW KTR_ALL 76 #endif 77 KTR_INFO_MASTER(ctxsw); 78 KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p", 79 sizeof(int) + sizeof(struct thread *)); 80 KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p", 81 sizeof(int) + sizeof(struct thread *)); 82 KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s", 83 sizeof (struct thread *) + sizeof(char *)); 84 KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", sizeof (struct thread *)); 85 86 static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads"); 87 88 #ifdef INVARIANTS 89 static int panic_on_cscount = 0; 90 #endif 91 static __int64_t switch_count = 0; 92 static __int64_t preempt_hit = 0; 93 static __int64_t preempt_miss = 0; 94 static __int64_t preempt_weird = 0; 95 static __int64_t token_contention_count __debugvar = 0; 96 static int lwkt_use_spin_port; 97 static struct objcache *thread_cache; 98 99 #ifdef SMP 100 static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame); 101 #endif 102 103 extern void cpu_heavy_restore(void); 104 extern void cpu_lwkt_restore(void); 105 extern void cpu_kthread_restore(void); 106 extern void cpu_idle_restore(void); 107 108 #ifdef __x86_64__ 109 110 static int 111 jg_tos_ok(struct thread *td) 112 { 113 void *tos; 114 int tos_ok; 115 116 if (td == NULL) { 117 return 1; 118 } 119 KKASSERT(td->td_sp != NULL); 120 tos = ((void **)td->td_sp)[0]; 121 tos_ok = 0; 122 if ((tos == cpu_heavy_restore) || (tos == cpu_lwkt_restore) || 123 (tos == cpu_kthread_restore) || (tos == cpu_idle_restore)) { 124 tos_ok = 1; 125 } 126 return tos_ok; 127 } 128 129 #endif 130 131 /* 132 * We can make all thread ports use the spin backend instead of the thread 133 * backend. This should only be set to debug the spin backend. 134 */ 135 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port); 136 137 #ifdef INVARIANTS 138 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, ""); 139 #endif 140 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, ""); 141 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, ""); 142 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, ""); 143 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, ""); 144 #ifdef INVARIANTS 145 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW, 146 &token_contention_count, 0, "spinning due to token contention"); 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_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 static __boolean_t 187 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags) 188 { 189 struct thread *td = (struct thread *)obj; 190 191 td->td_kstack = NULL; 192 td->td_kstack_size = 0; 193 td->td_flags = TDF_ALLOCATED_THREAD; 194 return (1); 195 } 196 197 static void 198 _lwkt_thread_dtor(void *obj, void *privdata) 199 { 200 struct thread *td = (struct thread *)obj; 201 202 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD, 203 ("_lwkt_thread_dtor: not allocated from objcache")); 204 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack && 205 td->td_kstack_size > 0, 206 ("_lwkt_thread_dtor: corrupted stack")); 207 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size); 208 } 209 210 /* 211 * Initialize the lwkt s/system. 212 */ 213 void 214 lwkt_init(void) 215 { 216 /* An objcache has 2 magazines per CPU so divide cache size by 2. */ 217 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread), 218 NULL, CACHE_NTHREADS/2, 219 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL); 220 } 221 222 /* 223 * Schedule a thread to run. As the current thread we can always safely 224 * schedule ourselves, and a shortcut procedure is provided for that 225 * function. 226 * 227 * (non-blocking, self contained on a per cpu basis) 228 */ 229 void 230 lwkt_schedule_self(thread_t td) 231 { 232 crit_enter_quick(td); 233 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!")); 234 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0); 235 _lwkt_enqueue(td); 236 crit_exit_quick(td); 237 } 238 239 /* 240 * Deschedule a thread. 241 * 242 * (non-blocking, self contained on a per cpu basis) 243 */ 244 void 245 lwkt_deschedule_self(thread_t td) 246 { 247 crit_enter_quick(td); 248 _lwkt_dequeue(td); 249 crit_exit_quick(td); 250 } 251 252 /* 253 * LWKTs operate on a per-cpu basis 254 * 255 * WARNING! Called from early boot, 'mycpu' may not work yet. 256 */ 257 void 258 lwkt_gdinit(struct globaldata *gd) 259 { 260 int i; 261 262 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i) 263 TAILQ_INIT(&gd->gd_tdrunq[i]); 264 gd->gd_runqmask = 0; 265 TAILQ_INIT(&gd->gd_tdallq); 266 } 267 268 /* 269 * Create a new thread. The thread must be associated with a process context 270 * or LWKT start address before it can be scheduled. If the target cpu is 271 * -1 the thread will be created on the current cpu. 272 * 273 * If you intend to create a thread without a process context this function 274 * does everything except load the startup and switcher function. 275 */ 276 thread_t 277 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags) 278 { 279 globaldata_t gd = mycpu; 280 void *stack; 281 282 /* 283 * If static thread storage is not supplied allocate a thread. Reuse 284 * a cached free thread if possible. gd_freetd is used to keep an exiting 285 * thread intact through the exit. 286 */ 287 if (td == NULL) { 288 if ((td = gd->gd_freetd) != NULL) 289 gd->gd_freetd = NULL; 290 else 291 td = objcache_get(thread_cache, M_WAITOK); 292 KASSERT((td->td_flags & 293 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD, 294 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags)); 295 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK); 296 } 297 298 /* 299 * Try to reuse cached stack. 300 */ 301 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) { 302 if (flags & TDF_ALLOCATED_STACK) { 303 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size); 304 stack = NULL; 305 } 306 } 307 if (stack == NULL) { 308 stack = (void *)kmem_alloc(&kernel_map, stksize); 309 flags |= TDF_ALLOCATED_STACK; 310 } 311 if (cpu < 0) 312 lwkt_init_thread(td, stack, stksize, flags, gd); 313 else 314 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu)); 315 return(td); 316 } 317 318 /* 319 * Initialize a preexisting thread structure. This function is used by 320 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread. 321 * 322 * All threads start out in a critical section at a priority of 323 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as 324 * appropriate. This function may send an IPI message when the 325 * requested cpu is not the current cpu and consequently gd_tdallq may 326 * not be initialized synchronously from the point of view of the originating 327 * cpu. 328 * 329 * NOTE! we have to be careful in regards to creating threads for other cpus 330 * if SMP has not yet been activated. 331 */ 332 #ifdef SMP 333 334 static void 335 lwkt_init_thread_remote(void *arg) 336 { 337 thread_t td = arg; 338 339 /* 340 * Protected by critical section held by IPI dispatch 341 */ 342 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq); 343 } 344 345 #endif 346 347 void 348 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags, 349 struct globaldata *gd) 350 { 351 globaldata_t mygd = mycpu; 352 353 bzero(td, sizeof(struct thread)); 354 td->td_kstack = stack; 355 td->td_kstack_size = stksize; 356 td->td_flags = flags; 357 td->td_gd = gd; 358 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT; 359 #ifdef SMP 360 if ((flags & TDF_MPSAFE) == 0) 361 td->td_mpcount = 1; 362 #endif 363 if (lwkt_use_spin_port) 364 lwkt_initport_spin(&td->td_msgport); 365 else 366 lwkt_initport_thread(&td->td_msgport, td); 367 pmap_init_thread(td); 368 #ifdef SMP 369 /* 370 * Normally initializing a thread for a remote cpu requires sending an 371 * IPI. However, the idlethread is setup before the other cpus are 372 * activated so we have to treat it as a special case. XXX manipulation 373 * of gd_tdallq requires the BGL. 374 */ 375 if (gd == mygd || td == &gd->gd_idlethread) { 376 crit_enter_gd(mygd); 377 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); 378 crit_exit_gd(mygd); 379 } else { 380 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td); 381 } 382 #else 383 crit_enter_gd(mygd); 384 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); 385 crit_exit_gd(mygd); 386 #endif 387 388 dsched_new_thread(td); 389 } 390 391 void 392 lwkt_set_comm(thread_t td, const char *ctl, ...) 393 { 394 __va_list va; 395 396 __va_start(va, ctl); 397 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va); 398 __va_end(va); 399 KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]); 400 } 401 402 void 403 lwkt_hold(thread_t td) 404 { 405 ++td->td_refs; 406 } 407 408 void 409 lwkt_rele(thread_t td) 410 { 411 KKASSERT(td->td_refs > 0); 412 --td->td_refs; 413 } 414 415 void 416 lwkt_wait_free(thread_t td) 417 { 418 while (td->td_refs) 419 tsleep(td, 0, "tdreap", hz); 420 } 421 422 void 423 lwkt_free_thread(thread_t td) 424 { 425 KASSERT((td->td_flags & TDF_RUNNING) == 0, 426 ("lwkt_free_thread: did not exit! %p", td)); 427 428 if (td->td_flags & TDF_ALLOCATED_THREAD) { 429 objcache_put(thread_cache, td); 430 } else if (td->td_flags & TDF_ALLOCATED_STACK) { 431 /* client-allocated struct with internally allocated stack */ 432 KASSERT(td->td_kstack && td->td_kstack_size > 0, 433 ("lwkt_free_thread: corrupted stack")); 434 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size); 435 td->td_kstack = NULL; 436 td->td_kstack_size = 0; 437 } 438 KTR_LOG(ctxsw_deadtd, td); 439 } 440 441 442 /* 443 * Switch to the next runnable lwkt. If no LWKTs are runnable then 444 * switch to the idlethread. Switching must occur within a critical 445 * section to avoid races with the scheduling queue. 446 * 447 * We always have full control over our cpu's run queue. Other cpus 448 * that wish to manipulate our queue must use the cpu_*msg() calls to 449 * talk to our cpu, so a critical section is all that is needed and 450 * the result is very, very fast thread switching. 451 * 452 * The LWKT scheduler uses a fixed priority model and round-robins at 453 * each priority level. User process scheduling is a totally 454 * different beast and LWKT priorities should not be confused with 455 * user process priorities. 456 * 457 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch() 458 * cleans it up. Note that the td_switch() function cannot do anything that 459 * requires the MP lock since the MP lock will have already been setup for 460 * the target thread (not the current thread). It's nice to have a scheduler 461 * that does not need the MP lock to work because it allows us to do some 462 * really cool high-performance MP lock optimizations. 463 * 464 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch() 465 * is not called by the current thread in the preemption case, only when 466 * the preempting thread blocks (in order to return to the original thread). 467 */ 468 void 469 lwkt_switch(void) 470 { 471 globaldata_t gd = mycpu; 472 thread_t td = gd->gd_curthread; 473 thread_t ntd; 474 #ifdef SMP 475 int mpheld; 476 #endif 477 478 /* 479 * Switching from within a 'fast' (non thread switched) interrupt or IPI 480 * is illegal. However, we may have to do it anyway if we hit a fatal 481 * kernel trap or we have paniced. 482 * 483 * If this case occurs save and restore the interrupt nesting level. 484 */ 485 if (gd->gd_intr_nesting_level) { 486 int savegdnest; 487 int savegdtrap; 488 489 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) { 490 panic("lwkt_switch: cannot switch from within " 491 "a fast interrupt, yet, td %p\n", td); 492 } else { 493 savegdnest = gd->gd_intr_nesting_level; 494 savegdtrap = gd->gd_trap_nesting_level; 495 gd->gd_intr_nesting_level = 0; 496 gd->gd_trap_nesting_level = 0; 497 if ((td->td_flags & TDF_PANICWARN) == 0) { 498 td->td_flags |= TDF_PANICWARN; 499 kprintf("Warning: thread switch from interrupt or IPI, " 500 "thread %p (%s)\n", td, td->td_comm); 501 print_backtrace(-1); 502 } 503 lwkt_switch(); 504 gd->gd_intr_nesting_level = savegdnest; 505 gd->gd_trap_nesting_level = savegdtrap; 506 return; 507 } 508 } 509 510 /* 511 * Passive release (used to transition from user to kernel mode 512 * when we block or switch rather then when we enter the kernel). 513 * This function is NOT called if we are switching into a preemption 514 * or returning from a preemption. Typically this causes us to lose 515 * our current process designation (if we have one) and become a true 516 * LWKT thread, and may also hand the current process designation to 517 * another process and schedule thread. 518 */ 519 if (td->td_release) 520 td->td_release(td); 521 522 crit_enter_gd(gd); 523 if (td->td_toks) 524 lwkt_relalltokens(td); 525 526 /* 527 * We had better not be holding any spin locks, but don't get into an 528 * endless panic loop. 529 */ 530 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL, 531 ("lwkt_switch: still holding a shared spinlock %p!", 532 gd->gd_spinlock_rd)); 533 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL, 534 ("lwkt_switch: still holding %d exclusive spinlocks!", 535 gd->gd_spinlocks_wr)); 536 537 538 #ifdef SMP 539 /* 540 * td_mpcount cannot be used to determine if we currently hold the 541 * MP lock because get_mplock() will increment it prior to attempting 542 * to get the lock, and switch out if it can't. Our ownership of 543 * the actual lock will remain stable while we are in a critical section 544 * (but, of course, another cpu may own or release the lock so the 545 * actual value of mp_lock is not stable). 546 */ 547 mpheld = MP_LOCK_HELD(); 548 #ifdef INVARIANTS 549 if (td->td_cscount) { 550 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n", 551 td); 552 if (panic_on_cscount) 553 panic("switching while mastering cpusync"); 554 } 555 #endif 556 #endif 557 if ((ntd = td->td_preempted) != NULL) { 558 /* 559 * We had preempted another thread on this cpu, resume the preempted 560 * thread. This occurs transparently, whether the preempted thread 561 * was scheduled or not (it may have been preempted after descheduling 562 * itself). 563 * 564 * We have to setup the MP lock for the original thread after backing 565 * out the adjustment that was made to curthread when the original 566 * was preempted. 567 */ 568 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK); 569 #ifdef SMP 570 if (ntd->td_mpcount && mpheld == 0) { 571 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d", 572 td, ntd, td->td_mpcount, ntd->td_mpcount); 573 } 574 if (ntd->td_mpcount) { 575 td->td_mpcount -= ntd->td_mpcount; 576 KKASSERT(td->td_mpcount >= 0); 577 } 578 #endif 579 ntd->td_flags |= TDF_PREEMPT_DONE; 580 581 /* 582 * The interrupt may have woken a thread up, we need to properly 583 * set the reschedule flag if the originally interrupted thread is 584 * at a lower priority. 585 */ 586 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1) 587 need_lwkt_resched(); 588 /* YYY release mp lock on switchback if original doesn't need it */ 589 } else { 590 /* 591 * Priority queue / round-robin at each priority. Note that user 592 * processes run at a fixed, low priority and the user process 593 * scheduler deals with interactions between user processes 594 * by scheduling and descheduling them from the LWKT queue as 595 * necessary. 596 * 597 * We have to adjust the MP lock for the target thread. If we 598 * need the MP lock and cannot obtain it we try to locate a 599 * thread that does not need the MP lock. If we cannot, we spin 600 * instead of HLT. 601 * 602 * A similar issue exists for the tokens held by the target thread. 603 * If we cannot obtain ownership of the tokens we cannot immediately 604 * schedule the thread. 605 */ 606 607 /* 608 * If an LWKT reschedule was requested, well that is what we are 609 * doing now so clear it. 610 */ 611 clear_lwkt_resched(); 612 again: 613 if (gd->gd_runqmask) { 614 int nq = bsrl(gd->gd_runqmask); 615 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) { 616 gd->gd_runqmask &= ~(1 << nq); 617 goto again; 618 } 619 #ifdef SMP 620 /* 621 * THREAD SELECTION FOR AN SMP MACHINE BUILD 622 * 623 * If the target needs the MP lock and we couldn't get it, 624 * or if the target is holding tokens and we could not 625 * gain ownership of the tokens, continue looking for a 626 * thread to schedule and spin instead of HLT if we can't. 627 * 628 * NOTE: the mpheld variable invalid after this conditional, it 629 * can change due to both cpu_try_mplock() returning success 630 * AND interactions in lwkt_getalltokens() due to the fact that 631 * we are trying to check the mpcount of a thread other then 632 * the current thread. Because of this, if the current thread 633 * is not holding td_mpcount, an IPI indirectly run via 634 * lwkt_getalltokens() can obtain and release the MP lock and 635 * cause the core MP lock to be released. 636 */ 637 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) || 638 (ntd->td_toks && lwkt_getalltokens(ntd) == 0) 639 ) { 640 u_int32_t rqmask = gd->gd_runqmask; 641 642 mpheld = MP_LOCK_HELD(); 643 ntd = NULL; 644 while (rqmask) { 645 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) { 646 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) { 647 /* spinning due to MP lock being held */ 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 /* 672 * We have two choices. We can either refuse to run a 673 * user thread when a kernel thread needs the MP lock 674 * but could not get it, or we can allow it to run but 675 * then expect an IPI (hopefully) later on to force a 676 * reschedule when the MP lock might become available. 677 */ 678 if (nq < TDPRI_KERN_LPSCHED) { 679 break; /* for now refuse to run */ 680 #if 0 681 if (chain_mplock == 0) 682 break; 683 /* continue loop, allow user threads to be scheduled */ 684 #endif 685 } 686 } 687 688 /* 689 * Case where a (kernel) thread needed the MP lock and could 690 * not get one, and we may or may not have found another 691 * thread which does not need the MP lock to run while 692 * we wait (ntd). 693 */ 694 if (ntd == NULL) { 695 ntd = &gd->gd_idlethread; 696 ntd->td_flags |= TDF_IDLE_NOHLT; 697 set_mplock_contention_mask(gd); 698 cpu_mplock_contested(); 699 goto using_idle_thread; 700 } else { 701 clr_mplock_contention_mask(gd); 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 } 706 } else { 707 clr_mplock_contention_mask(gd); 708 ++gd->gd_cnt.v_swtch; 709 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq); 710 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq); 711 } 712 #else 713 /* 714 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to 715 * worry about tokens or the BGL. However, we still have 716 * to call lwkt_getalltokens() in order to properly detect 717 * stale tokens. This call cannot fail for a UP build! 718 */ 719 lwkt_getalltokens(ntd); 720 ++gd->gd_cnt.v_swtch; 721 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq); 722 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq); 723 #endif 724 } else { 725 /* 726 * We have nothing to run but only let the idle loop halt 727 * the cpu if there are no pending interrupts. 728 */ 729 ntd = &gd->gd_idlethread; 730 if (gd->gd_reqflags & RQF_IDLECHECK_MASK) 731 ntd->td_flags |= TDF_IDLE_NOHLT; 732 #ifdef SMP 733 using_idle_thread: 734 /* 735 * The idle thread should not be holding the MP lock unless we 736 * are trapping in the kernel or in a panic. Since we select the 737 * idle thread unconditionally when no other thread is available, 738 * if the MP lock is desired during a panic or kernel trap, we 739 * have to loop in the scheduler until we get it. 740 */ 741 if (ntd->td_mpcount) { 742 mpheld = MP_LOCK_HELD(); 743 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) 744 panic("Idle thread %p was holding the BGL!", ntd); 745 if (mpheld == 0) 746 goto again; 747 } 748 #endif 749 } 750 } 751 KASSERT(ntd->td_pri >= TDPRI_CRIT, 752 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri)); 753 754 /* 755 * Do the actual switch. If the new target does not need the MP lock 756 * and we are holding it, release the MP lock. If the new target requires 757 * the MP lock we have already acquired it for the target. 758 */ 759 #ifdef SMP 760 if (ntd->td_mpcount == 0 ) { 761 if (MP_LOCK_HELD()) 762 cpu_rel_mplock(); 763 } else { 764 ASSERT_MP_LOCK_HELD(ntd); 765 } 766 #endif 767 if (td != ntd) { 768 ++switch_count; 769 #ifdef __x86_64__ 770 { 771 int tos_ok __debugvar = jg_tos_ok(ntd); 772 KKASSERT(tos_ok); 773 } 774 #endif 775 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd); 776 td->td_switch(ntd); 777 } 778 /* NOTE: current cpu may have changed after switch */ 779 crit_exit_quick(td); 780 } 781 782 /* 783 * Request that the target thread preempt the current thread. Preemption 784 * only works under a specific set of conditions: 785 * 786 * - We are not preempting ourselves 787 * - The target thread is owned by the current cpu 788 * - We are not currently being preempted 789 * - The target is not currently being preempted 790 * - We are not holding any spin locks 791 * - The target thread is not holding any tokens 792 * - We are able to satisfy the target's MP lock requirements (if any). 793 * 794 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically 795 * this is called via lwkt_schedule() through the td_preemptable callback. 796 * critpri is the managed critical priority that we should ignore in order 797 * to determine whether preemption is possible (aka usually just the crit 798 * priority of lwkt_schedule() itself). 799 * 800 * XXX at the moment we run the target thread in a critical section during 801 * the preemption in order to prevent the target from taking interrupts 802 * that *WE* can't. Preemption is strictly limited to interrupt threads 803 * and interrupt-like threads, outside of a critical section, and the 804 * preempted source thread will be resumed the instant the target blocks 805 * whether or not the source is scheduled (i.e. preemption is supposed to 806 * be as transparent as possible). 807 * 808 * The target thread inherits our MP count (added to its own) for the 809 * duration of the preemption in order to preserve the atomicy of the 810 * MP lock during the preemption. Therefore, any preempting targets must be 811 * careful in regards to MP assertions. Note that the MP count may be 812 * out of sync with the physical mp_lock, but we do not have to preserve 813 * the original ownership of the lock if it was out of synch (that is, we 814 * can leave it synchronized on return). 815 */ 816 void 817 lwkt_preempt(thread_t ntd, int critpri) 818 { 819 struct globaldata *gd = mycpu; 820 thread_t td; 821 #ifdef SMP 822 int mpheld; 823 int savecnt; 824 #endif 825 826 /* 827 * The caller has put us in a critical section. We can only preempt 828 * if the caller of the caller was not in a critical section (basically 829 * a local interrupt), as determined by the 'critpri' parameter. We 830 * also can't preempt if the caller is holding any spinlocks (even if 831 * he isn't in a critical section). This also handles the tokens test. 832 * 833 * YYY The target thread must be in a critical section (else it must 834 * inherit our critical section? I dunno yet). 835 * 836 * Set need_lwkt_resched() unconditionally for now YYY. 837 */ 838 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri)); 839 840 td = gd->gd_curthread; 841 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) { 842 ++preempt_miss; 843 return; 844 } 845 if ((td->td_pri & ~TDPRI_MASK) > critpri) { 846 ++preempt_miss; 847 need_lwkt_resched(); 848 return; 849 } 850 #ifdef SMP 851 if (ntd->td_gd != gd) { 852 ++preempt_miss; 853 need_lwkt_resched(); 854 return; 855 } 856 #endif 857 /* 858 * Take the easy way out and do not preempt if we are holding 859 * any spinlocks. We could test whether the thread(s) being 860 * preempted interlock against the target thread's tokens and whether 861 * we can get all the target thread's tokens, but this situation 862 * should not occur very often so its easier to simply not preempt. 863 * Also, plain spinlocks are impossible to figure out at this point so 864 * just don't preempt. 865 * 866 * Do not try to preempt if the target thread is holding any tokens. 867 * We could try to acquire the tokens but this case is so rare there 868 * is no need to support it. 869 */ 870 if (gd->gd_spinlock_rd || gd->gd_spinlocks_wr) { 871 ++preempt_miss; 872 need_lwkt_resched(); 873 return; 874 } 875 if (ntd->td_toks) { 876 ++preempt_miss; 877 need_lwkt_resched(); 878 return; 879 } 880 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) { 881 ++preempt_weird; 882 need_lwkt_resched(); 883 return; 884 } 885 if (ntd->td_preempted) { 886 ++preempt_hit; 887 need_lwkt_resched(); 888 return; 889 } 890 #ifdef SMP 891 /* 892 * note: an interrupt might have occured just as we were transitioning 893 * to or from the MP lock. In this case td_mpcount will be pre-disposed 894 * (non-zero) but not actually synchronized with the actual state of the 895 * lock. We can use it to imply an MP lock requirement for the 896 * preemption but we cannot use it to test whether we hold the MP lock 897 * or not. 898 */ 899 savecnt = td->td_mpcount; 900 mpheld = MP_LOCK_HELD(); 901 ntd->td_mpcount += td->td_mpcount; 902 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) { 903 ntd->td_mpcount -= td->td_mpcount; 904 ++preempt_miss; 905 need_lwkt_resched(); 906 return; 907 } 908 #endif 909 910 /* 911 * Since we are able to preempt the current thread, there is no need to 912 * call need_lwkt_resched(). 913 */ 914 ++preempt_hit; 915 ntd->td_preempted = td; 916 td->td_flags |= TDF_PREEMPT_LOCK; 917 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd); 918 td->td_switch(ntd); 919 920 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE)); 921 #ifdef SMP 922 KKASSERT(savecnt == td->td_mpcount); 923 mpheld = MP_LOCK_HELD(); 924 if (mpheld && td->td_mpcount == 0) 925 cpu_rel_mplock(); 926 else if (mpheld == 0 && td->td_mpcount) 927 panic("lwkt_preempt(): MP lock was not held through"); 928 #endif 929 ntd->td_preempted = NULL; 930 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE); 931 } 932 933 /* 934 * Conditionally call splz() if gd_reqflags indicates work is pending. 935 * 936 * td_nest_count prevents deep nesting via splz() or doreti() which 937 * might otherwise blow out the kernel stack. Note that except for 938 * this special case, we MUST call splz() here to handle any 939 * pending ints, particularly after we switch, or we might accidently 940 * halt the cpu with interrupts pending. 941 * 942 * (self contained on a per cpu basis) 943 */ 944 void 945 splz_check(void) 946 { 947 globaldata_t gd = mycpu; 948 thread_t td = gd->gd_curthread; 949 950 if (gd->gd_reqflags && td->td_nest_count < 2) 951 splz(); 952 } 953 954 /* 955 * This implements a normal yield which will yield to equal priority 956 * threads as well as higher priority threads. Note that gd_reqflags 957 * tests will be handled by the crit_exit() call in lwkt_switch(). 958 * 959 * (self contained on a per cpu basis) 960 */ 961 void 962 lwkt_yield(void) 963 { 964 lwkt_schedule_self(curthread); 965 lwkt_switch(); 966 } 967 968 /* 969 * This function is used along with the lwkt_passive_recover() inline 970 * by the trap code to negotiate a passive release of the current 971 * process/lwp designation with the user scheduler. 972 */ 973 void 974 lwkt_passive_release(struct thread *td) 975 { 976 struct lwp *lp = td->td_lwp; 977 978 td->td_release = NULL; 979 lwkt_setpri_self(TDPRI_KERN_USER); 980 lp->lwp_proc->p_usched->release_curproc(lp); 981 } 982 983 /* 984 * Make a kernel thread act as if it were in user mode with regards 985 * to scheduling, to avoid becoming cpu-bound in the kernel. Kernel 986 * loops which may be potentially cpu-bound can call lwkt_user_yield(). 987 * 988 * The lwkt_user_yield() function is designed to have very low overhead 989 * if no yield is determined to be needed. 990 */ 991 void 992 lwkt_user_yield(void) 993 { 994 thread_t td = curthread; 995 struct lwp *lp = td->td_lwp; 996 997 #ifdef SMP 998 /* 999 * XXX SEVERE TEMPORARY HACK. A cpu-bound operation running in the 1000 * kernel can prevent other cpus from servicing interrupt threads 1001 * which still require the MP lock (which is a lot of them). This 1002 * has a chaining effect since if the interrupt is blocked, so is 1003 * the event, so normal scheduling will not pick up on the problem. 1004 */ 1005 if (mp_lock_contention_mask && td->td_mpcount) { 1006 yield_mplock(td); 1007 } 1008 #endif 1009 1010 /* 1011 * Another kernel thread wants the cpu 1012 */ 1013 if (lwkt_resched_wanted()) 1014 lwkt_switch(); 1015 1016 /* 1017 * If the user scheduler has asynchronously determined that the current 1018 * process (when running in user mode) needs to lose the cpu then make 1019 * sure we are released. 1020 */ 1021 if (user_resched_wanted()) { 1022 if (td->td_release) 1023 td->td_release(td); 1024 } 1025 1026 /* 1027 * If we are released reduce our priority 1028 */ 1029 if (td->td_release == NULL) { 1030 if (lwkt_check_resched(td) > 0) 1031 lwkt_switch(); 1032 if (lp) { 1033 lp->lwp_proc->p_usched->acquire_curproc(lp); 1034 td->td_release = lwkt_passive_release; 1035 lwkt_setpri_self(TDPRI_USER_NORM); 1036 } 1037 } 1038 } 1039 1040 /* 1041 * Return 0 if no runnable threads are pending at the same or higher 1042 * priority as the passed thread. 1043 * 1044 * Return 1 if runnable threads are pending at the same priority. 1045 * 1046 * Return 2 if runnable threads are pending at a higher priority. 1047 */ 1048 int 1049 lwkt_check_resched(thread_t td) 1050 { 1051 int pri = td->td_pri & TDPRI_MASK; 1052 1053 if (td->td_gd->gd_runqmask > (2 << pri) - 1) 1054 return(2); 1055 if (TAILQ_NEXT(td, td_threadq)) 1056 return(1); 1057 return(0); 1058 } 1059 1060 /* 1061 * Generic schedule. Possibly schedule threads belonging to other cpus and 1062 * deal with threads that might be blocked on a wait queue. 1063 * 1064 * We have a little helper inline function which does additional work after 1065 * the thread has been enqueued, including dealing with preemption and 1066 * setting need_lwkt_resched() (which prevents the kernel from returning 1067 * to userland until it has processed higher priority threads). 1068 * 1069 * It is possible for this routine to be called after a failed _enqueue 1070 * (due to the target thread migrating, sleeping, or otherwise blocked). 1071 * We have to check that the thread is actually on the run queue! 1072 * 1073 * reschedok is an optimized constant propagated from lwkt_schedule() or 1074 * lwkt_schedule_noresched(). By default it is non-zero, causing a 1075 * reschedule to be requested if the target thread has a higher priority. 1076 * The port messaging code will set MSG_NORESCHED and cause reschedok to 1077 * be 0, prevented undesired reschedules. 1078 */ 1079 static __inline 1080 void 1081 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri, int reschedok) 1082 { 1083 thread_t otd; 1084 1085 if (ntd->td_flags & TDF_RUNQ) { 1086 if (ntd->td_preemptable && reschedok) { 1087 ntd->td_preemptable(ntd, cpri); /* YYY +token */ 1088 } else if (reschedok) { 1089 otd = curthread; 1090 if ((ntd->td_pri & TDPRI_MASK) > (otd->td_pri & TDPRI_MASK)) 1091 need_lwkt_resched(); 1092 } 1093 } 1094 } 1095 1096 static __inline 1097 void 1098 _lwkt_schedule(thread_t td, int reschedok) 1099 { 1100 globaldata_t mygd = mycpu; 1101 1102 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!")); 1103 crit_enter_gd(mygd); 1104 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0); 1105 if (td == mygd->gd_curthread) { 1106 _lwkt_enqueue(td); 1107 } else { 1108 /* 1109 * If we own the thread, there is no race (since we are in a 1110 * critical section). If we do not own the thread there might 1111 * be a race but the target cpu will deal with it. 1112 */ 1113 #ifdef SMP 1114 if (td->td_gd == mygd) { 1115 _lwkt_enqueue(td); 1116 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok); 1117 } else { 1118 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0); 1119 } 1120 #else 1121 _lwkt_enqueue(td); 1122 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok); 1123 #endif 1124 } 1125 crit_exit_gd(mygd); 1126 } 1127 1128 void 1129 lwkt_schedule(thread_t td) 1130 { 1131 _lwkt_schedule(td, 1); 1132 } 1133 1134 void 1135 lwkt_schedule_noresched(thread_t td) 1136 { 1137 _lwkt_schedule(td, 0); 1138 } 1139 1140 #ifdef SMP 1141 1142 /* 1143 * When scheduled remotely if frame != NULL the IPIQ is being 1144 * run via doreti or an interrupt then preemption can be allowed. 1145 * 1146 * To allow preemption we have to drop the critical section so only 1147 * one is present in _lwkt_schedule_post. 1148 */ 1149 static void 1150 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame) 1151 { 1152 thread_t td = curthread; 1153 thread_t ntd = arg; 1154 1155 if (frame && ntd->td_preemptable) { 1156 crit_exit_noyield(td); 1157 _lwkt_schedule(ntd, 1); 1158 crit_enter_quick(td); 1159 } else { 1160 _lwkt_schedule(ntd, 1); 1161 } 1162 } 1163 1164 /* 1165 * Thread migration using a 'Pull' method. The thread may or may not be 1166 * the current thread. It MUST be descheduled and in a stable state. 1167 * lwkt_giveaway() must be called on the cpu owning the thread. 1168 * 1169 * At any point after lwkt_giveaway() is called, the target cpu may 1170 * 'pull' the thread by calling lwkt_acquire(). 1171 * 1172 * We have to make sure the thread is not sitting on a per-cpu tsleep 1173 * queue or it will blow up when it moves to another cpu. 1174 * 1175 * MPSAFE - must be called under very specific conditions. 1176 */ 1177 void 1178 lwkt_giveaway(thread_t td) 1179 { 1180 globaldata_t gd = mycpu; 1181 1182 crit_enter_gd(gd); 1183 if (td->td_flags & TDF_TSLEEPQ) 1184 tsleep_remove(td); 1185 KKASSERT(td->td_gd == gd); 1186 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); 1187 td->td_flags |= TDF_MIGRATING; 1188 crit_exit_gd(gd); 1189 } 1190 1191 void 1192 lwkt_acquire(thread_t td) 1193 { 1194 globaldata_t gd; 1195 globaldata_t mygd; 1196 1197 KKASSERT(td->td_flags & TDF_MIGRATING); 1198 gd = td->td_gd; 1199 mygd = mycpu; 1200 if (gd != mycpu) { 1201 cpu_lfence(); 1202 KKASSERT((td->td_flags & TDF_RUNQ) == 0); 1203 crit_enter_gd(mygd); 1204 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) { 1205 #ifdef SMP 1206 lwkt_process_ipiq(); 1207 #endif 1208 cpu_lfence(); 1209 } 1210 td->td_gd = mygd; 1211 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1212 td->td_flags &= ~TDF_MIGRATING; 1213 crit_exit_gd(mygd); 1214 } else { 1215 crit_enter_gd(mygd); 1216 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1217 td->td_flags &= ~TDF_MIGRATING; 1218 crit_exit_gd(mygd); 1219 } 1220 } 1221 1222 #endif 1223 1224 /* 1225 * Generic deschedule. Descheduling threads other then your own should be 1226 * done only in carefully controlled circumstances. Descheduling is 1227 * asynchronous. 1228 * 1229 * This function may block if the cpu has run out of messages. 1230 */ 1231 void 1232 lwkt_deschedule(thread_t td) 1233 { 1234 crit_enter(); 1235 #ifdef SMP 1236 if (td == curthread) { 1237 _lwkt_dequeue(td); 1238 } else { 1239 if (td->td_gd == mycpu) { 1240 _lwkt_dequeue(td); 1241 } else { 1242 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td); 1243 } 1244 } 1245 #else 1246 _lwkt_dequeue(td); 1247 #endif 1248 crit_exit(); 1249 } 1250 1251 /* 1252 * Set the target thread's priority. This routine does not automatically 1253 * switch to a higher priority thread, LWKT threads are not designed for 1254 * continuous priority changes. Yield if you want to switch. 1255 * 1256 * We have to retain the critical section count which uses the high bits 1257 * of the td_pri field. The specified priority may also indicate zero or 1258 * more critical sections by adding TDPRI_CRIT*N. 1259 * 1260 * Note that we requeue the thread whether it winds up on a different runq 1261 * or not. uio_yield() depends on this and the routine is not normally 1262 * called with the same priority otherwise. 1263 */ 1264 void 1265 lwkt_setpri(thread_t td, int pri) 1266 { 1267 KKASSERT(pri >= 0); 1268 KKASSERT(td->td_gd == mycpu); 1269 crit_enter(); 1270 if (td->td_flags & TDF_RUNQ) { 1271 _lwkt_dequeue(td); 1272 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri; 1273 _lwkt_enqueue(td); 1274 } else { 1275 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri; 1276 } 1277 crit_exit(); 1278 } 1279 1280 /* 1281 * Set the initial priority for a thread prior to it being scheduled for 1282 * the first time. The thread MUST NOT be scheduled before or during 1283 * this call. The thread may be assigned to a cpu other then the current 1284 * cpu. 1285 * 1286 * Typically used after a thread has been created with TDF_STOPPREQ, 1287 * and before the thread is initially scheduled. 1288 */ 1289 void 1290 lwkt_setpri_initial(thread_t td, int pri) 1291 { 1292 KKASSERT(pri >= 0); 1293 KKASSERT((td->td_flags & TDF_RUNQ) == 0); 1294 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri; 1295 } 1296 1297 void 1298 lwkt_setpri_self(int pri) 1299 { 1300 thread_t td = curthread; 1301 1302 KKASSERT(pri >= 0 && pri <= TDPRI_MAX); 1303 crit_enter(); 1304 if (td->td_flags & TDF_RUNQ) { 1305 _lwkt_dequeue(td); 1306 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri; 1307 _lwkt_enqueue(td); 1308 } else { 1309 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri; 1310 } 1311 crit_exit(); 1312 } 1313 1314 /* 1315 * Migrate the current thread to the specified cpu. 1316 * 1317 * This is accomplished by descheduling ourselves from the current cpu, 1318 * moving our thread to the tdallq of the target cpu, IPI messaging the 1319 * target cpu, and switching out. TDF_MIGRATING prevents scheduling 1320 * races while the thread is being migrated. 1321 * 1322 * We must be sure to remove ourselves from the current cpu's tsleepq 1323 * before potentially moving to another queue. The thread can be on 1324 * a tsleepq due to a left-over tsleep_interlock(). 1325 */ 1326 #ifdef SMP 1327 static void lwkt_setcpu_remote(void *arg); 1328 #endif 1329 1330 void 1331 lwkt_setcpu_self(globaldata_t rgd) 1332 { 1333 #ifdef SMP 1334 thread_t td = curthread; 1335 1336 if (td->td_gd != rgd) { 1337 crit_enter_quick(td); 1338 if (td->td_flags & TDF_TSLEEPQ) 1339 tsleep_remove(td); 1340 td->td_flags |= TDF_MIGRATING; 1341 lwkt_deschedule_self(td); 1342 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1343 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td); 1344 lwkt_switch(); 1345 /* we are now on the target cpu */ 1346 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); 1347 crit_exit_quick(td); 1348 } 1349 #endif 1350 } 1351 1352 void 1353 lwkt_migratecpu(int cpuid) 1354 { 1355 #ifdef SMP 1356 globaldata_t rgd; 1357 1358 rgd = globaldata_find(cpuid); 1359 lwkt_setcpu_self(rgd); 1360 #endif 1361 } 1362 1363 /* 1364 * Remote IPI for cpu migration (called while in a critical section so we 1365 * do not have to enter another one). The thread has already been moved to 1366 * our cpu's allq, but we must wait for the thread to be completely switched 1367 * out on the originating cpu before we schedule it on ours or the stack 1368 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD 1369 * change to main memory. 1370 * 1371 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races 1372 * against wakeups. It is best if this interface is used only when there 1373 * are no pending events that might try to schedule the thread. 1374 */ 1375 #ifdef SMP 1376 static void 1377 lwkt_setcpu_remote(void *arg) 1378 { 1379 thread_t td = arg; 1380 globaldata_t gd = mycpu; 1381 1382 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) { 1383 #ifdef SMP 1384 lwkt_process_ipiq(); 1385 #endif 1386 cpu_lfence(); 1387 } 1388 td->td_gd = gd; 1389 cpu_sfence(); 1390 td->td_flags &= ~TDF_MIGRATING; 1391 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0); 1392 _lwkt_enqueue(td); 1393 } 1394 #endif 1395 1396 struct lwp * 1397 lwkt_preempted_proc(void) 1398 { 1399 thread_t td = curthread; 1400 while (td->td_preempted) 1401 td = td->td_preempted; 1402 return(td->td_lwp); 1403 } 1404 1405 /* 1406 * Create a kernel process/thread/whatever. It shares it's address space 1407 * with proc0 - ie: kernel only. 1408 * 1409 * NOTE! By default new threads are created with the MP lock held. A 1410 * thread which does not require the MP lock should release it by calling 1411 * rel_mplock() at the start of the new thread. 1412 */ 1413 int 1414 lwkt_create(void (*func)(void *), void *arg, 1415 struct thread **tdp, thread_t template, int tdflags, int cpu, 1416 const char *fmt, ...) 1417 { 1418 thread_t td; 1419 __va_list ap; 1420 1421 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu, 1422 tdflags); 1423 if (tdp) 1424 *tdp = td; 1425 cpu_set_thread_handler(td, lwkt_exit, func, arg); 1426 1427 /* 1428 * Set up arg0 for 'ps' etc 1429 */ 1430 __va_start(ap, fmt); 1431 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap); 1432 __va_end(ap); 1433 1434 /* 1435 * Schedule the thread to run 1436 */ 1437 if ((td->td_flags & TDF_STOPREQ) == 0) 1438 lwkt_schedule(td); 1439 else 1440 td->td_flags &= ~TDF_STOPREQ; 1441 return 0; 1442 } 1443 1444 /* 1445 * Destroy an LWKT thread. Warning! This function is not called when 1446 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and 1447 * uses a different reaping mechanism. 1448 */ 1449 void 1450 lwkt_exit(void) 1451 { 1452 thread_t td = curthread; 1453 thread_t std; 1454 globaldata_t gd; 1455 1456 if (td->td_flags & TDF_VERBOSE) 1457 kprintf("kthread %p %s has exited\n", td, td->td_comm); 1458 caps_exit(td); 1459 1460 /* 1461 * Get us into a critical section to interlock gd_freetd and loop 1462 * until we can get it freed. 1463 * 1464 * We have to cache the current td in gd_freetd because objcache_put()ing 1465 * it would rip it out from under us while our thread is still active. 1466 */ 1467 gd = mycpu; 1468 crit_enter_quick(td); 1469 while ((std = gd->gd_freetd) != NULL) { 1470 gd->gd_freetd = NULL; 1471 objcache_put(thread_cache, std); 1472 } 1473 1474 /* 1475 * Remove thread resources from kernel lists and deschedule us for 1476 * the last time. 1477 */ 1478 if (td->td_flags & TDF_TSLEEPQ) 1479 tsleep_remove(td); 1480 biosched_done(td); 1481 dsched_exit_thread(td); 1482 lwkt_deschedule_self(td); 1483 lwkt_remove_tdallq(td); 1484 if (td->td_flags & TDF_ALLOCATED_THREAD) 1485 gd->gd_freetd = td; 1486 cpu_thread_exit(); 1487 } 1488 1489 void 1490 lwkt_remove_tdallq(thread_t td) 1491 { 1492 KKASSERT(td->td_gd == mycpu); 1493 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1494 } 1495 1496 void 1497 crit_panic(void) 1498 { 1499 thread_t td = curthread; 1500 int lpri = td->td_pri; 1501 1502 td->td_pri = 0; 1503 panic("td_pri is/would-go negative! %p %d", td, lpri); 1504 } 1505 1506 #ifdef SMP 1507 1508 /* 1509 * Called from debugger/panic on cpus which have been stopped. We must still 1510 * process the IPIQ while stopped, even if we were stopped while in a critical 1511 * section (XXX). 1512 * 1513 * If we are dumping also try to process any pending interrupts. This may 1514 * or may not work depending on the state of the cpu at the point it was 1515 * stopped. 1516 */ 1517 void 1518 lwkt_smp_stopped(void) 1519 { 1520 globaldata_t gd = mycpu; 1521 1522 crit_enter_gd(gd); 1523 if (dumping) { 1524 lwkt_process_ipiq(); 1525 splz(); 1526 } else { 1527 lwkt_process_ipiq(); 1528 } 1529 crit_exit_gd(gd); 1530 } 1531 1532 #endif 1533