1 /* 2 * Copyright (c) 2003-2011 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@backplane.com> 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * 3. Neither the name of The DragonFly Project nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific, prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 */ 34 35 /* 36 * Each cpu in a system has its own self-contained light weight kernel 37 * thread scheduler, which means that generally speaking we only need 38 * to use a critical section to avoid problems. Foreign thread 39 * scheduling is queued via (async) IPIs. 40 */ 41 42 #include <sys/param.h> 43 #include <sys/systm.h> 44 #include <sys/kernel.h> 45 #include <sys/proc.h> 46 #include <sys/rtprio.h> 47 #include <sys/kinfo.h> 48 #include <sys/queue.h> 49 #include <sys/sysctl.h> 50 #include <sys/kthread.h> 51 #include <machine/cpu.h> 52 #include <sys/lock.h> 53 #include <sys/caps.h> 54 #include <sys/spinlock.h> 55 #include <sys/ktr.h> 56 57 #include <sys/thread2.h> 58 #include <sys/spinlock2.h> 59 #include <sys/mplock2.h> 60 61 #include <sys/dsched.h> 62 63 #include <vm/vm.h> 64 #include <vm/vm_param.h> 65 #include <vm/vm_kern.h> 66 #include <vm/vm_object.h> 67 #include <vm/vm_page.h> 68 #include <vm/vm_map.h> 69 #include <vm/vm_pager.h> 70 #include <vm/vm_extern.h> 71 72 #include <machine/stdarg.h> 73 #include <machine/smp.h> 74 75 #if !defined(KTR_CTXSW) 76 #define KTR_CTXSW KTR_ALL 77 #endif 78 KTR_INFO_MASTER(ctxsw); 79 KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p", 80 sizeof(int) + sizeof(struct thread *)); 81 KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p", 82 sizeof(int) + sizeof(struct thread *)); 83 KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s", 84 sizeof (struct thread *) + sizeof(char *)); 85 KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", sizeof (struct thread *)); 86 87 static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads"); 88 89 #ifdef INVARIANTS 90 static int panic_on_cscount = 0; 91 #endif 92 static __int64_t switch_count = 0; 93 static __int64_t preempt_hit = 0; 94 static __int64_t preempt_miss = 0; 95 static __int64_t preempt_weird = 0; 96 static __int64_t token_contention_count[TDPRI_MAX+1] __debugvar; 97 static int lwkt_use_spin_port; 98 static struct objcache *thread_cache; 99 100 #ifdef SMP 101 static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame); 102 static void lwkt_setcpu_remote(void *arg); 103 #endif 104 105 extern void cpu_heavy_restore(void); 106 extern void cpu_lwkt_restore(void); 107 extern void cpu_kthread_restore(void); 108 extern void cpu_idle_restore(void); 109 110 /* 111 * We can make all thread ports use the spin backend instead of the thread 112 * backend. This should only be set to debug the spin backend. 113 */ 114 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port); 115 116 #ifdef INVARIANTS 117 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, 118 "Panic if attempting to switch lwkt's while mastering cpusync"); 119 #endif 120 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, 121 "Number of switched threads"); 122 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, 123 "Successful preemption events"); 124 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, 125 "Failed preemption events"); 126 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, 127 "Number of preempted threads."); 128 #ifdef INVARIANTS 129 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_00, CTLFLAG_RW, 130 &token_contention_count[0], 0, "spinning due to token contention"); 131 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_01, CTLFLAG_RW, 132 &token_contention_count[1], 0, "spinning due to token contention"); 133 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_02, CTLFLAG_RW, 134 &token_contention_count[2], 0, "spinning due to token contention"); 135 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_03, CTLFLAG_RW, 136 &token_contention_count[3], 0, "spinning due to token contention"); 137 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_04, CTLFLAG_RW, 138 &token_contention_count[4], 0, "spinning due to token contention"); 139 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_05, CTLFLAG_RW, 140 &token_contention_count[5], 0, "spinning due to token contention"); 141 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_06, CTLFLAG_RW, 142 &token_contention_count[6], 0, "spinning due to token contention"); 143 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_07, CTLFLAG_RW, 144 &token_contention_count[7], 0, "spinning due to token contention"); 145 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_08, CTLFLAG_RW, 146 &token_contention_count[8], 0, "spinning due to token contention"); 147 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_09, CTLFLAG_RW, 148 &token_contention_count[9], 0, "spinning due to token contention"); 149 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_10, CTLFLAG_RW, 150 &token_contention_count[10], 0, "spinning due to token contention"); 151 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_11, CTLFLAG_RW, 152 &token_contention_count[11], 0, "spinning due to token contention"); 153 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_12, CTLFLAG_RW, 154 &token_contention_count[12], 0, "spinning due to token contention"); 155 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_13, CTLFLAG_RW, 156 &token_contention_count[13], 0, "spinning due to token contention"); 157 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_14, CTLFLAG_RW, 158 &token_contention_count[14], 0, "spinning due to token contention"); 159 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_15, CTLFLAG_RW, 160 &token_contention_count[15], 0, "spinning due to token contention"); 161 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_16, CTLFLAG_RW, 162 &token_contention_count[16], 0, "spinning due to token contention"); 163 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_17, CTLFLAG_RW, 164 &token_contention_count[17], 0, "spinning due to token contention"); 165 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_18, CTLFLAG_RW, 166 &token_contention_count[18], 0, "spinning due to token contention"); 167 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_19, CTLFLAG_RW, 168 &token_contention_count[19], 0, "spinning due to token contention"); 169 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_20, CTLFLAG_RW, 170 &token_contention_count[20], 0, "spinning due to token contention"); 171 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_21, CTLFLAG_RW, 172 &token_contention_count[21], 0, "spinning due to token contention"); 173 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_22, CTLFLAG_RW, 174 &token_contention_count[22], 0, "spinning due to token contention"); 175 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_23, CTLFLAG_RW, 176 &token_contention_count[23], 0, "spinning due to token contention"); 177 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_24, CTLFLAG_RW, 178 &token_contention_count[24], 0, "spinning due to token contention"); 179 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_25, CTLFLAG_RW, 180 &token_contention_count[25], 0, "spinning due to token contention"); 181 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_26, CTLFLAG_RW, 182 &token_contention_count[26], 0, "spinning due to token contention"); 183 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_27, CTLFLAG_RW, 184 &token_contention_count[27], 0, "spinning due to token contention"); 185 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_28, CTLFLAG_RW, 186 &token_contention_count[28], 0, "spinning due to token contention"); 187 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_29, CTLFLAG_RW, 188 &token_contention_count[29], 0, "spinning due to token contention"); 189 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_30, CTLFLAG_RW, 190 &token_contention_count[30], 0, "spinning due to token contention"); 191 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_31, CTLFLAG_RW, 192 &token_contention_count[31], 0, "spinning due to token contention"); 193 #endif 194 static int fairq_enable = 0; 195 SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW, 196 &fairq_enable, 0, "Turn on fairq priority accumulators"); 197 static int fairq_bypass = -1; 198 SYSCTL_INT(_lwkt, OID_AUTO, fairq_bypass, CTLFLAG_RW, 199 &fairq_bypass, 0, "Allow fairq to bypass td on token failure"); 200 extern int lwkt_sched_debug; 201 int lwkt_sched_debug = 0; 202 SYSCTL_INT(_lwkt, OID_AUTO, sched_debug, CTLFLAG_RW, 203 &lwkt_sched_debug, 0, "Scheduler debug"); 204 static int lwkt_spin_loops = 10; 205 SYSCTL_INT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW, 206 &lwkt_spin_loops, 0, "Scheduler spin loops until sorted decon"); 207 static int lwkt_spin_reseq = 0; 208 SYSCTL_INT(_lwkt, OID_AUTO, spin_reseq, CTLFLAG_RW, 209 &lwkt_spin_reseq, 0, "Scheduler resequencer enable"); 210 static int lwkt_spin_monitor = 0; 211 SYSCTL_INT(_lwkt, OID_AUTO, spin_monitor, CTLFLAG_RW, 212 &lwkt_spin_monitor, 0, "Scheduler uses monitor/mwait"); 213 static int lwkt_spin_fatal = 0; /* disabled */ 214 SYSCTL_INT(_lwkt, OID_AUTO, spin_fatal, CTLFLAG_RW, 215 &lwkt_spin_fatal, 0, "LWKT scheduler spin loops till fatal panic"); 216 static int preempt_enable = 1; 217 SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW, 218 &preempt_enable, 0, "Enable preemption"); 219 static int lwkt_cache_threads = 32; 220 SYSCTL_INT(_lwkt, OID_AUTO, cache_threads, CTLFLAG_RD, 221 &lwkt_cache_threads, 0, "thread+kstack cache"); 222 223 static __cachealign int lwkt_cseq_rindex; 224 static __cachealign int lwkt_cseq_windex; 225 226 /* 227 * These helper procedures handle the runq, they can only be called from 228 * within a critical section. 229 * 230 * WARNING! Prior to SMP being brought up it is possible to enqueue and 231 * dequeue threads belonging to other cpus, so be sure to use td->td_gd 232 * instead of 'mycpu' when referencing the globaldata structure. Once 233 * SMP live enqueuing and dequeueing only occurs on the current cpu. 234 */ 235 static __inline 236 void 237 _lwkt_dequeue(thread_t td) 238 { 239 if (td->td_flags & TDF_RUNQ) { 240 struct globaldata *gd = td->td_gd; 241 242 td->td_flags &= ~TDF_RUNQ; 243 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq); 244 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL) 245 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING); 246 } 247 } 248 249 /* 250 * Priority enqueue. 251 * 252 * NOTE: There are a limited number of lwkt threads runnable since user 253 * processes only schedule one at a time per cpu. 254 */ 255 static __inline 256 void 257 _lwkt_enqueue(thread_t td) 258 { 259 thread_t xtd; 260 261 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) { 262 struct globaldata *gd = td->td_gd; 263 264 td->td_flags |= TDF_RUNQ; 265 xtd = TAILQ_FIRST(&gd->gd_tdrunq); 266 if (xtd == NULL) { 267 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 268 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING); 269 } else { 270 while (xtd && xtd->td_pri >= td->td_pri) 271 xtd = TAILQ_NEXT(xtd, td_threadq); 272 if (xtd) 273 TAILQ_INSERT_BEFORE(xtd, td, td_threadq); 274 else 275 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 276 } 277 278 /* 279 * Request a LWKT reschedule if we are now at the head of the queue. 280 */ 281 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) 282 need_lwkt_resched(); 283 } 284 } 285 286 static __boolean_t 287 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags) 288 { 289 struct thread *td = (struct thread *)obj; 290 291 td->td_kstack = NULL; 292 td->td_kstack_size = 0; 293 td->td_flags = TDF_ALLOCATED_THREAD; 294 td->td_mpflags = 0; 295 return (1); 296 } 297 298 static void 299 _lwkt_thread_dtor(void *obj, void *privdata) 300 { 301 struct thread *td = (struct thread *)obj; 302 303 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD, 304 ("_lwkt_thread_dtor: not allocated from objcache")); 305 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack && 306 td->td_kstack_size > 0, 307 ("_lwkt_thread_dtor: corrupted stack")); 308 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size); 309 } 310 311 /* 312 * Initialize the lwkt s/system. 313 * 314 * Nominally cache up to 32 thread + kstack structures. 315 */ 316 void 317 lwkt_init(void) 318 { 319 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads); 320 thread_cache = objcache_create_mbacked( 321 M_THREAD, sizeof(struct thread), 322 NULL, lwkt_cache_threads, 323 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL); 324 } 325 326 /* 327 * Schedule a thread to run. As the current thread we can always safely 328 * schedule ourselves, and a shortcut procedure is provided for that 329 * function. 330 * 331 * (non-blocking, self contained on a per cpu basis) 332 */ 333 void 334 lwkt_schedule_self(thread_t td) 335 { 336 KKASSERT((td->td_flags & TDF_MIGRATING) == 0); 337 crit_enter_quick(td); 338 KASSERT(td != &td->td_gd->gd_idlethread, 339 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!")); 340 KKASSERT(td->td_lwp == NULL || 341 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 342 _lwkt_enqueue(td); 343 crit_exit_quick(td); 344 } 345 346 /* 347 * Deschedule a thread. 348 * 349 * (non-blocking, self contained on a per cpu basis) 350 */ 351 void 352 lwkt_deschedule_self(thread_t td) 353 { 354 crit_enter_quick(td); 355 _lwkt_dequeue(td); 356 crit_exit_quick(td); 357 } 358 359 /* 360 * LWKTs operate on a per-cpu basis 361 * 362 * WARNING! Called from early boot, 'mycpu' may not work yet. 363 */ 364 void 365 lwkt_gdinit(struct globaldata *gd) 366 { 367 TAILQ_INIT(&gd->gd_tdrunq); 368 TAILQ_INIT(&gd->gd_tdallq); 369 } 370 371 /* 372 * Create a new thread. The thread must be associated with a process context 373 * or LWKT start address before it can be scheduled. If the target cpu is 374 * -1 the thread will be created on the current cpu. 375 * 376 * If you intend to create a thread without a process context this function 377 * does everything except load the startup and switcher function. 378 */ 379 thread_t 380 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags) 381 { 382 static int cpu_rotator; 383 globaldata_t gd = mycpu; 384 void *stack; 385 386 /* 387 * If static thread storage is not supplied allocate a thread. Reuse 388 * a cached free thread if possible. gd_freetd is used to keep an exiting 389 * thread intact through the exit. 390 */ 391 if (td == NULL) { 392 crit_enter_gd(gd); 393 if ((td = gd->gd_freetd) != NULL) { 394 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK| 395 TDF_RUNQ)) == 0); 396 gd->gd_freetd = NULL; 397 } else { 398 td = objcache_get(thread_cache, M_WAITOK); 399 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK| 400 TDF_RUNQ)) == 0); 401 } 402 crit_exit_gd(gd); 403 KASSERT((td->td_flags & 404 (TDF_ALLOCATED_THREAD|TDF_RUNNING|TDF_PREEMPT_LOCK)) == 405 TDF_ALLOCATED_THREAD, 406 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags)); 407 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK); 408 } 409 410 /* 411 * Try to reuse cached stack. 412 */ 413 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) { 414 if (flags & TDF_ALLOCATED_STACK) { 415 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size); 416 stack = NULL; 417 } 418 } 419 if (stack == NULL) { 420 stack = (void *)kmem_alloc_stack(&kernel_map, stksize); 421 flags |= TDF_ALLOCATED_STACK; 422 } 423 if (cpu < 0) { 424 cpu = ++cpu_rotator; 425 cpu_ccfence(); 426 cpu %= ncpus; 427 } 428 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu)); 429 return(td); 430 } 431 432 /* 433 * Initialize a preexisting thread structure. This function is used by 434 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread. 435 * 436 * All threads start out in a critical section at a priority of 437 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as 438 * appropriate. This function may send an IPI message when the 439 * requested cpu is not the current cpu and consequently gd_tdallq may 440 * not be initialized synchronously from the point of view of the originating 441 * cpu. 442 * 443 * NOTE! we have to be careful in regards to creating threads for other cpus 444 * if SMP has not yet been activated. 445 */ 446 #ifdef SMP 447 448 static void 449 lwkt_init_thread_remote(void *arg) 450 { 451 thread_t td = arg; 452 453 /* 454 * Protected by critical section held by IPI dispatch 455 */ 456 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq); 457 } 458 459 #endif 460 461 /* 462 * lwkt core thread structural initialization. 463 * 464 * NOTE: All threads are initialized as mpsafe threads. 465 */ 466 void 467 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags, 468 struct globaldata *gd) 469 { 470 globaldata_t mygd = mycpu; 471 472 bzero(td, sizeof(struct thread)); 473 td->td_kstack = stack; 474 td->td_kstack_size = stksize; 475 td->td_flags = flags; 476 td->td_mpflags = 0; 477 td->td_gd = gd; 478 td->td_pri = TDPRI_KERN_DAEMON; 479 td->td_critcount = 1; 480 td->td_toks_have = NULL; 481 td->td_toks_stop = &td->td_toks_base; 482 if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT)) 483 lwkt_initport_spin(&td->td_msgport); 484 else 485 lwkt_initport_thread(&td->td_msgport, td); 486 pmap_init_thread(td); 487 #ifdef SMP 488 /* 489 * Normally initializing a thread for a remote cpu requires sending an 490 * IPI. However, the idlethread is setup before the other cpus are 491 * activated so we have to treat it as a special case. XXX manipulation 492 * of gd_tdallq requires the BGL. 493 */ 494 if (gd == mygd || td == &gd->gd_idlethread) { 495 crit_enter_gd(mygd); 496 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); 497 crit_exit_gd(mygd); 498 } else { 499 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td); 500 } 501 #else 502 crit_enter_gd(mygd); 503 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); 504 crit_exit_gd(mygd); 505 #endif 506 507 dsched_new_thread(td); 508 } 509 510 void 511 lwkt_set_comm(thread_t td, const char *ctl, ...) 512 { 513 __va_list va; 514 515 __va_start(va, ctl); 516 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va); 517 __va_end(va); 518 KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]); 519 } 520 521 /* 522 * Prevent the thread from getting destroyed. Note that unlike PHOLD/PRELE 523 * this does not prevent the thread from migrating to another cpu so the 524 * gd_tdallq state is not protected by this. 525 */ 526 void 527 lwkt_hold(thread_t td) 528 { 529 atomic_add_int(&td->td_refs, 1); 530 } 531 532 void 533 lwkt_rele(thread_t td) 534 { 535 KKASSERT(td->td_refs > 0); 536 atomic_add_int(&td->td_refs, -1); 537 } 538 539 void 540 lwkt_free_thread(thread_t td) 541 { 542 KKASSERT(td->td_refs == 0); 543 KKASSERT((td->td_flags & (TDF_RUNNING | TDF_PREEMPT_LOCK | 544 TDF_RUNQ | TDF_TSLEEPQ)) == 0); 545 if (td->td_flags & TDF_ALLOCATED_THREAD) { 546 objcache_put(thread_cache, td); 547 } else if (td->td_flags & TDF_ALLOCATED_STACK) { 548 /* client-allocated struct with internally allocated stack */ 549 KASSERT(td->td_kstack && td->td_kstack_size > 0, 550 ("lwkt_free_thread: corrupted stack")); 551 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size); 552 td->td_kstack = NULL; 553 td->td_kstack_size = 0; 554 } 555 KTR_LOG(ctxsw_deadtd, td); 556 } 557 558 559 /* 560 * Switch to the next runnable lwkt. If no LWKTs are runnable then 561 * switch to the idlethread. Switching must occur within a critical 562 * section to avoid races with the scheduling queue. 563 * 564 * We always have full control over our cpu's run queue. Other cpus 565 * that wish to manipulate our queue must use the cpu_*msg() calls to 566 * talk to our cpu, so a critical section is all that is needed and 567 * the result is very, very fast thread switching. 568 * 569 * The LWKT scheduler uses a fixed priority model and round-robins at 570 * each priority level. User process scheduling is a totally 571 * different beast and LWKT priorities should not be confused with 572 * user process priorities. 573 * 574 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch() 575 * is not called by the current thread in the preemption case, only when 576 * the preempting thread blocks (in order to return to the original thread). 577 * 578 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread 579 * migration and tsleep deschedule the current lwkt thread and call 580 * lwkt_switch(). In particular, the target cpu of the migration fully 581 * expects the thread to become non-runnable and can deadlock against 582 * cpusync operations if we run any IPIs prior to switching the thread out. 583 * 584 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF 585 * THE CURRENT THREAD HAS BEEN DESCHEDULED! 586 */ 587 void 588 lwkt_switch(void) 589 { 590 globaldata_t gd = mycpu; 591 thread_t td = gd->gd_curthread; 592 thread_t ntd; 593 thread_t xtd; 594 int spinning = 0; 595 596 KKASSERT(gd->gd_processing_ipiq == 0); 597 KKASSERT(td->td_flags & TDF_RUNNING); 598 599 /* 600 * Switching from within a 'fast' (non thread switched) interrupt or IPI 601 * is illegal. However, we may have to do it anyway if we hit a fatal 602 * kernel trap or we have paniced. 603 * 604 * If this case occurs save and restore the interrupt nesting level. 605 */ 606 if (gd->gd_intr_nesting_level) { 607 int savegdnest; 608 int savegdtrap; 609 610 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) { 611 panic("lwkt_switch: Attempt to switch from a " 612 "a fast interrupt, ipi, or hard code section, " 613 "td %p\n", 614 td); 615 } else { 616 savegdnest = gd->gd_intr_nesting_level; 617 savegdtrap = gd->gd_trap_nesting_level; 618 gd->gd_intr_nesting_level = 0; 619 gd->gd_trap_nesting_level = 0; 620 if ((td->td_flags & TDF_PANICWARN) == 0) { 621 td->td_flags |= TDF_PANICWARN; 622 kprintf("Warning: thread switch from interrupt, IPI, " 623 "or hard code section.\n" 624 "thread %p (%s)\n", td, td->td_comm); 625 print_backtrace(-1); 626 } 627 lwkt_switch(); 628 gd->gd_intr_nesting_level = savegdnest; 629 gd->gd_trap_nesting_level = savegdtrap; 630 return; 631 } 632 } 633 634 /* 635 * Release our current user process designation if we are blocking 636 * or if a user reschedule was requested. 637 * 638 * NOTE: This function is NOT called if we are switching into or 639 * returning from a preemption. 640 * 641 * NOTE: Releasing our current user process designation may cause 642 * it to be assigned to another thread, which in turn will 643 * cause us to block in the usched acquire code when we attempt 644 * to return to userland. 645 * 646 * NOTE: On SMP systems this can be very nasty when heavy token 647 * contention is present so we want to be careful not to 648 * release the designation gratuitously. 649 */ 650 if (td->td_release && 651 (user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) { 652 td->td_release(td); 653 } 654 655 /* 656 * Release all tokens 657 */ 658 crit_enter_gd(gd); 659 if (TD_TOKS_HELD(td)) 660 lwkt_relalltokens(td); 661 662 /* 663 * We had better not be holding any spin locks, but don't get into an 664 * endless panic loop. 665 */ 666 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL, 667 ("lwkt_switch: still holding %d exclusive spinlocks!", 668 gd->gd_spinlocks_wr)); 669 670 671 #ifdef SMP 672 #ifdef INVARIANTS 673 if (td->td_cscount) { 674 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n", 675 td); 676 if (panic_on_cscount) 677 panic("switching while mastering cpusync"); 678 } 679 #endif 680 #endif 681 682 /* 683 * If we had preempted another thread on this cpu, resume the preempted 684 * thread. This occurs transparently, whether the preempted thread 685 * was scheduled or not (it may have been preempted after descheduling 686 * itself). 687 * 688 * We have to setup the MP lock for the original thread after backing 689 * out the adjustment that was made to curthread when the original 690 * was preempted. 691 */ 692 if ((ntd = td->td_preempted) != NULL) { 693 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK); 694 ntd->td_flags |= TDF_PREEMPT_DONE; 695 696 /* 697 * The interrupt may have woken a thread up, we need to properly 698 * set the reschedule flag if the originally interrupted thread is 699 * at a lower priority. 700 * 701 * The interrupt may not have descheduled. 702 */ 703 if (TAILQ_FIRST(&gd->gd_tdrunq) != ntd) 704 need_lwkt_resched(); 705 goto havethread_preempted; 706 } 707 708 /* 709 * If we cannot obtain ownership of the tokens we cannot immediately 710 * schedule the target thread. 711 * 712 * Reminder: Again, we cannot afford to run any IPIs in this path if 713 * the current thread has been descheduled. 714 */ 715 for (;;) { 716 clear_lwkt_resched(); 717 718 /* 719 * Hotpath - pull the head of the run queue and attempt to schedule 720 * it. 721 */ 722 for (;;) { 723 ntd = TAILQ_FIRST(&gd->gd_tdrunq); 724 725 if (ntd == NULL) { 726 /* 727 * Runq is empty, switch to idle to allow it to halt. 728 */ 729 ntd = &gd->gd_idlethread; 730 #ifdef SMP 731 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) 732 ASSERT_NO_TOKENS_HELD(ntd); 733 #endif 734 cpu_time.cp_msg[0] = 0; 735 cpu_time.cp_stallpc = 0; 736 goto haveidle; 737 } 738 break; 739 } 740 741 /* 742 * Hotpath - schedule ntd. 743 * 744 * NOTE: For UP there is no mplock and lwkt_getalltokens() 745 * always succeeds. 746 */ 747 if (TD_TOKS_NOT_HELD(ntd) || 748 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops))) 749 { 750 goto havethread; 751 } 752 753 /* 754 * Coldpath (SMP only since tokens always succeed on UP) 755 * 756 * We had some contention on the thread we wanted to schedule. 757 * What we do now is try to find a thread that we can schedule 758 * in its stead. 759 * 760 * The coldpath scan does NOT rearrange threads in the run list. 761 * The lwkt_schedulerclock() will assert need_lwkt_resched() on 762 * the next tick whenever the current head is not the current thread. 763 */ 764 #ifdef INVARIANTS 765 ++token_contention_count[ntd->td_pri]; 766 ++ntd->td_contended; 767 #endif 768 769 if (fairq_bypass > 0) 770 goto skip; 771 772 xtd = NULL; 773 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) { 774 /* 775 * Never schedule threads returning to userland or the 776 * user thread scheduler helper thread when higher priority 777 * threads are present. 778 */ 779 if (ntd->td_pri < TDPRI_KERN_LPSCHED) { 780 ntd = NULL; 781 break; 782 } 783 784 /* 785 * Try this one. 786 */ 787 if (TD_TOKS_NOT_HELD(ntd) || 788 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops))) { 789 goto havethread; 790 } 791 #ifdef INVARIANTS 792 ++token_contention_count[ntd->td_pri]; 793 ++ntd->td_contended; 794 #endif 795 } 796 797 skip: 798 /* 799 * We exhausted the run list, meaning that all runnable threads 800 * are contested. 801 */ 802 cpu_pause(); 803 ntd = &gd->gd_idlethread; 804 #ifdef SMP 805 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) 806 ASSERT_NO_TOKENS_HELD(ntd); 807 /* contention case, do not clear contention mask */ 808 #endif 809 810 /* 811 * We are going to have to retry but if the current thread is not 812 * on the runq we instead switch through the idle thread to get away 813 * from the current thread. We have to flag for lwkt reschedule 814 * to prevent the idle thread from halting. 815 * 816 * NOTE: A non-zero spinning is passed to lwkt_getalltokens() to 817 * instruct it to deal with the potential for deadlocks by 818 * ordering the tokens by address. 819 */ 820 if ((td->td_flags & TDF_RUNQ) == 0) { 821 need_lwkt_resched(); /* prevent hlt */ 822 goto haveidle; 823 } 824 #if defined(INVARIANTS) && defined(__amd64__) 825 if ((read_rflags() & PSL_I) == 0) { 826 cpu_enable_intr(); 827 panic("lwkt_switch() called with interrupts disabled"); 828 } 829 #endif 830 831 /* 832 * Number iterations so far. After a certain point we switch to 833 * a sorted-address/monitor/mwait version of lwkt_getalltokens() 834 */ 835 if (spinning < 0x7FFFFFFF) 836 ++spinning; 837 838 #ifdef SMP 839 /* 840 * lwkt_getalltokens() failed in sorted token mode, we can use 841 * monitor/mwait in this case. 842 */ 843 if (spinning >= lwkt_spin_loops && 844 (cpu_mi_feature & CPU_MI_MONITOR) && 845 lwkt_spin_monitor) 846 { 847 cpu_mmw_pause_int(&gd->gd_reqflags, 848 (gd->gd_reqflags | RQF_SPINNING) & 849 ~RQF_IDLECHECK_WK_MASK); 850 } 851 #endif 852 853 /* 854 * We already checked that td is still scheduled so this should be 855 * safe. 856 */ 857 splz_check(); 858 859 /* 860 * This experimental resequencer is used as a fall-back to reduce 861 * hw cache line contention by placing each core's scheduler into a 862 * time-domain-multplexed slot. 863 * 864 * The resequencer is disabled by default. It's functionality has 865 * largely been superceeded by the token algorithm which limits races 866 * to a subset of cores. 867 * 868 * The resequencer algorithm tends to break down when more than 869 * 20 cores are contending. What appears to happen is that new 870 * tokens can be obtained out of address-sorted order by new cores 871 * while existing cores languish in long delays between retries and 872 * wind up being starved-out of the token acquisition. 873 */ 874 if (lwkt_spin_reseq && spinning >= lwkt_spin_reseq) { 875 int cseq = atomic_fetchadd_int(&lwkt_cseq_windex, 1); 876 int oseq; 877 878 while ((oseq = lwkt_cseq_rindex) != cseq) { 879 cpu_ccfence(); 880 #if 1 881 if (cpu_mi_feature & CPU_MI_MONITOR) { 882 cpu_mmw_pause_int(&lwkt_cseq_rindex, oseq); 883 } else { 884 #endif 885 cpu_pause(); 886 cpu_lfence(); 887 #if 1 888 } 889 #endif 890 } 891 DELAY(1); 892 atomic_add_int(&lwkt_cseq_rindex, 1); 893 } 894 /* highest level for(;;) loop */ 895 } 896 897 havethread: 898 /* 899 * Clear gd_idle_repeat when doing a normal switch to a non-idle 900 * thread. 901 */ 902 ntd->td_wmesg = NULL; 903 ++gd->gd_cnt.v_swtch; 904 gd->gd_idle_repeat = 0; 905 906 havethread_preempted: 907 /* 908 * If the new target does not need the MP lock and we are holding it, 909 * release the MP lock. If the new target requires the MP lock we have 910 * already acquired it for the target. 911 */ 912 ; 913 haveidle: 914 KASSERT(ntd->td_critcount, 915 ("priority problem in lwkt_switch %d %d", 916 td->td_critcount, ntd->td_critcount)); 917 918 if (td != ntd) { 919 /* 920 * Execute the actual thread switch operation. This function 921 * returns to the current thread and returns the previous thread 922 * (which may be different from the thread we switched to). 923 * 924 * We are responsible for marking ntd as TDF_RUNNING. 925 */ 926 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0); 927 ++switch_count; 928 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd); 929 ntd->td_flags |= TDF_RUNNING; 930 lwkt_switch_return(td->td_switch(ntd)); 931 /* ntd invalid, td_switch() can return a different thread_t */ 932 } 933 934 /* 935 * catch-all. XXX is this strictly needed? 936 */ 937 splz_check(); 938 939 /* NOTE: current cpu may have changed after switch */ 940 crit_exit_quick(td); 941 } 942 943 /* 944 * Called by assembly in the td_switch (thread restore path) for thread 945 * bootstrap cases which do not 'return' to lwkt_switch(). 946 */ 947 void 948 lwkt_switch_return(thread_t otd) 949 { 950 #ifdef SMP 951 globaldata_t rgd; 952 953 /* 954 * Check if otd was migrating. Now that we are on ntd we can finish 955 * up the migration. This is a bit messy but it is the only place 956 * where td is known to be fully descheduled. 957 * 958 * We can only activate the migration if otd was migrating but not 959 * held on the cpu due to a preemption chain. We still have to 960 * clear TDF_RUNNING on the old thread either way. 961 * 962 * We are responsible for clearing the previously running thread's 963 * TDF_RUNNING. 964 */ 965 if ((rgd = otd->td_migrate_gd) != NULL && 966 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) { 967 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) == 968 (TDF_MIGRATING | TDF_RUNNING)); 969 otd->td_migrate_gd = NULL; 970 otd->td_flags &= ~TDF_RUNNING; 971 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd); 972 } else { 973 otd->td_flags &= ~TDF_RUNNING; 974 } 975 #else 976 otd->td_flags &= ~TDF_RUNNING; 977 #endif 978 } 979 980 /* 981 * Request that the target thread preempt the current thread. Preemption 982 * can only occur if our only critical section is the one that we were called 983 * with, the relative priority of the target thread is higher, and the target 984 * thread holds no tokens. This also only works if we are not holding any 985 * spinlocks (obviously). 986 * 987 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically 988 * this is called via lwkt_schedule() through the td_preemptable callback. 989 * critcount is the managed critical priority that we should ignore in order 990 * to determine whether preemption is possible (aka usually just the crit 991 * priority of lwkt_schedule() itself). 992 * 993 * Preemption is typically limited to interrupt threads. 994 * 995 * Operation works in a fairly straight-forward manner. The normal 996 * scheduling code is bypassed and we switch directly to the target 997 * thread. When the target thread attempts to block or switch away 998 * code at the base of lwkt_switch() will switch directly back to our 999 * thread. Our thread is able to retain whatever tokens it holds and 1000 * if the target needs one of them the target will switch back to us 1001 * and reschedule itself normally. 1002 */ 1003 void 1004 lwkt_preempt(thread_t ntd, int critcount) 1005 { 1006 struct globaldata *gd = mycpu; 1007 thread_t xtd; 1008 thread_t td; 1009 int save_gd_intr_nesting_level; 1010 1011 /* 1012 * The caller has put us in a critical section. We can only preempt 1013 * if the caller of the caller was not in a critical section (basically 1014 * a local interrupt), as determined by the 'critcount' parameter. We 1015 * also can't preempt if the caller is holding any spinlocks (even if 1016 * he isn't in a critical section). This also handles the tokens test. 1017 * 1018 * YYY The target thread must be in a critical section (else it must 1019 * inherit our critical section? I dunno yet). 1020 */ 1021 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri)); 1022 1023 td = gd->gd_curthread; 1024 if (preempt_enable == 0) { 1025 ++preempt_miss; 1026 return; 1027 } 1028 if (ntd->td_pri <= td->td_pri) { 1029 ++preempt_miss; 1030 return; 1031 } 1032 if (td->td_critcount > critcount) { 1033 ++preempt_miss; 1034 return; 1035 } 1036 #ifdef SMP 1037 if (td->td_cscount) { 1038 ++preempt_miss; 1039 return; 1040 } 1041 if (ntd->td_gd != gd) { 1042 ++preempt_miss; 1043 return; 1044 } 1045 #endif 1046 /* 1047 * We don't have to check spinlocks here as they will also bump 1048 * td_critcount. 1049 * 1050 * Do not try to preempt if the target thread is holding any tokens. 1051 * We could try to acquire the tokens but this case is so rare there 1052 * is no need to support it. 1053 */ 1054 KKASSERT(gd->gd_spinlocks_wr == 0); 1055 1056 if (TD_TOKS_HELD(ntd)) { 1057 ++preempt_miss; 1058 return; 1059 } 1060 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) { 1061 ++preempt_weird; 1062 return; 1063 } 1064 if (ntd->td_preempted) { 1065 ++preempt_hit; 1066 return; 1067 } 1068 KKASSERT(gd->gd_processing_ipiq == 0); 1069 1070 /* 1071 * Since we are able to preempt the current thread, there is no need to 1072 * call need_lwkt_resched(). 1073 * 1074 * We must temporarily clear gd_intr_nesting_level around the switch 1075 * since switchouts from the target thread are allowed (they will just 1076 * return to our thread), and since the target thread has its own stack. 1077 * 1078 * A preemption must switch back to the original thread, assert the 1079 * case. 1080 */ 1081 ++preempt_hit; 1082 ntd->td_preempted = td; 1083 td->td_flags |= TDF_PREEMPT_LOCK; 1084 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd); 1085 save_gd_intr_nesting_level = gd->gd_intr_nesting_level; 1086 gd->gd_intr_nesting_level = 0; 1087 1088 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0); 1089 ntd->td_flags |= TDF_RUNNING; 1090 xtd = td->td_switch(ntd); 1091 KKASSERT(xtd == ntd); 1092 lwkt_switch_return(xtd); 1093 gd->gd_intr_nesting_level = save_gd_intr_nesting_level; 1094 1095 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE)); 1096 ntd->td_preempted = NULL; 1097 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE); 1098 } 1099 1100 /* 1101 * Conditionally call splz() if gd_reqflags indicates work is pending. 1102 * This will work inside a critical section but not inside a hard code 1103 * section. 1104 * 1105 * (self contained on a per cpu basis) 1106 */ 1107 void 1108 splz_check(void) 1109 { 1110 globaldata_t gd = mycpu; 1111 thread_t td = gd->gd_curthread; 1112 1113 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && 1114 gd->gd_intr_nesting_level == 0 && 1115 td->td_nest_count < 2) 1116 { 1117 splz(); 1118 } 1119 } 1120 1121 /* 1122 * This version is integrated into crit_exit, reqflags has already 1123 * been tested but td_critcount has not. 1124 * 1125 * We only want to execute the splz() on the 1->0 transition of 1126 * critcount and not in a hard code section or if too deeply nested. 1127 */ 1128 void 1129 lwkt_maybe_splz(thread_t td) 1130 { 1131 globaldata_t gd = td->td_gd; 1132 1133 if (td->td_critcount == 0 && 1134 gd->gd_intr_nesting_level == 0 && 1135 td->td_nest_count < 2) 1136 { 1137 splz(); 1138 } 1139 } 1140 1141 /* 1142 * Drivers which set up processing co-threads can call this function to 1143 * run the co-thread at a higher priority and to allow it to preempt 1144 * normal threads. 1145 */ 1146 void 1147 lwkt_set_interrupt_support_thread(void) 1148 { 1149 thread_t td = curthread; 1150 1151 lwkt_setpri_self(TDPRI_INT_SUPPORT); 1152 td->td_flags |= TDF_INTTHREAD; 1153 td->td_preemptable = lwkt_preempt; 1154 } 1155 1156 1157 /* 1158 * This function is used to negotiate a passive release of the current 1159 * process/lwp designation with the user scheduler, allowing the user 1160 * scheduler to schedule another user thread. The related kernel thread 1161 * (curthread) continues running in the released state. 1162 */ 1163 void 1164 lwkt_passive_release(struct thread *td) 1165 { 1166 struct lwp *lp = td->td_lwp; 1167 1168 td->td_release = NULL; 1169 lwkt_setpri_self(TDPRI_KERN_USER); 1170 lp->lwp_proc->p_usched->release_curproc(lp); 1171 } 1172 1173 1174 /* 1175 * This implements a LWKT yield, allowing a kernel thread to yield to other 1176 * kernel threads at the same or higher priority. This function can be 1177 * called in a tight loop and will typically only yield once per tick. 1178 * 1179 * Most kernel threads run at the same priority in order to allow equal 1180 * sharing. 1181 * 1182 * (self contained on a per cpu basis) 1183 */ 1184 void 1185 lwkt_yield(void) 1186 { 1187 globaldata_t gd = mycpu; 1188 thread_t td = gd->gd_curthread; 1189 1190 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1191 splz(); 1192 if (lwkt_resched_wanted()) { 1193 lwkt_schedule_self(curthread); 1194 lwkt_switch(); 1195 } 1196 } 1197 1198 /* 1199 * This yield is designed for kernel threads with a user context. 1200 * 1201 * The kernel acting on behalf of the user is potentially cpu-bound, 1202 * this function will efficiently allow other threads to run and also 1203 * switch to other processes by releasing. 1204 * 1205 * The lwkt_user_yield() function is designed to have very low overhead 1206 * if no yield is determined to be needed. 1207 */ 1208 void 1209 lwkt_user_yield(void) 1210 { 1211 globaldata_t gd = mycpu; 1212 thread_t td = gd->gd_curthread; 1213 1214 /* 1215 * Always run any pending interrupts in case we are in a critical 1216 * section. 1217 */ 1218 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1219 splz(); 1220 1221 /* 1222 * Switch (which forces a release) if another kernel thread needs 1223 * the cpu, if userland wants us to resched, or if our kernel 1224 * quantum has run out. 1225 */ 1226 if (lwkt_resched_wanted() || 1227 user_resched_wanted()) 1228 { 1229 lwkt_switch(); 1230 } 1231 1232 #if 0 1233 /* 1234 * Reacquire the current process if we are released. 1235 * 1236 * XXX not implemented atm. The kernel may be holding locks and such, 1237 * so we want the thread to continue to receive cpu. 1238 */ 1239 if (td->td_release == NULL && lp) { 1240 lp->lwp_proc->p_usched->acquire_curproc(lp); 1241 td->td_release = lwkt_passive_release; 1242 lwkt_setpri_self(TDPRI_USER_NORM); 1243 } 1244 #endif 1245 } 1246 1247 /* 1248 * Generic schedule. Possibly schedule threads belonging to other cpus and 1249 * deal with threads that might be blocked on a wait queue. 1250 * 1251 * We have a little helper inline function which does additional work after 1252 * the thread has been enqueued, including dealing with preemption and 1253 * setting need_lwkt_resched() (which prevents the kernel from returning 1254 * to userland until it has processed higher priority threads). 1255 * 1256 * It is possible for this routine to be called after a failed _enqueue 1257 * (due to the target thread migrating, sleeping, or otherwise blocked). 1258 * We have to check that the thread is actually on the run queue! 1259 */ 1260 static __inline 1261 void 1262 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount) 1263 { 1264 if (ntd->td_flags & TDF_RUNQ) { 1265 if (ntd->td_preemptable) { 1266 ntd->td_preemptable(ntd, ccount); /* YYY +token */ 1267 } 1268 } 1269 } 1270 1271 static __inline 1272 void 1273 _lwkt_schedule(thread_t td) 1274 { 1275 globaldata_t mygd = mycpu; 1276 1277 KASSERT(td != &td->td_gd->gd_idlethread, 1278 ("lwkt_schedule(): scheduling gd_idlethread is illegal!")); 1279 KKASSERT((td->td_flags & TDF_MIGRATING) == 0); 1280 crit_enter_gd(mygd); 1281 KKASSERT(td->td_lwp == NULL || 1282 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 1283 1284 if (td == mygd->gd_curthread) { 1285 _lwkt_enqueue(td); 1286 } else { 1287 /* 1288 * If we own the thread, there is no race (since we are in a 1289 * critical section). If we do not own the thread there might 1290 * be a race but the target cpu will deal with it. 1291 */ 1292 #ifdef SMP 1293 if (td->td_gd == mygd) { 1294 _lwkt_enqueue(td); 1295 _lwkt_schedule_post(mygd, td, 1); 1296 } else { 1297 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0); 1298 } 1299 #else 1300 _lwkt_enqueue(td); 1301 _lwkt_schedule_post(mygd, td, 1); 1302 #endif 1303 } 1304 crit_exit_gd(mygd); 1305 } 1306 1307 void 1308 lwkt_schedule(thread_t td) 1309 { 1310 _lwkt_schedule(td); 1311 } 1312 1313 void 1314 lwkt_schedule_noresched(thread_t td) /* XXX not impl */ 1315 { 1316 _lwkt_schedule(td); 1317 } 1318 1319 #ifdef SMP 1320 1321 /* 1322 * When scheduled remotely if frame != NULL the IPIQ is being 1323 * run via doreti or an interrupt then preemption can be allowed. 1324 * 1325 * To allow preemption we have to drop the critical section so only 1326 * one is present in _lwkt_schedule_post. 1327 */ 1328 static void 1329 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame) 1330 { 1331 thread_t td = curthread; 1332 thread_t ntd = arg; 1333 1334 if (frame && ntd->td_preemptable) { 1335 crit_exit_noyield(td); 1336 _lwkt_schedule(ntd); 1337 crit_enter_quick(td); 1338 } else { 1339 _lwkt_schedule(ntd); 1340 } 1341 } 1342 1343 /* 1344 * Thread migration using a 'Pull' method. The thread may or may not be 1345 * the current thread. It MUST be descheduled and in a stable state. 1346 * lwkt_giveaway() must be called on the cpu owning the thread. 1347 * 1348 * At any point after lwkt_giveaway() is called, the target cpu may 1349 * 'pull' the thread by calling lwkt_acquire(). 1350 * 1351 * We have to make sure the thread is not sitting on a per-cpu tsleep 1352 * queue or it will blow up when it moves to another cpu. 1353 * 1354 * MPSAFE - must be called under very specific conditions. 1355 */ 1356 void 1357 lwkt_giveaway(thread_t td) 1358 { 1359 globaldata_t gd = mycpu; 1360 1361 crit_enter_gd(gd); 1362 if (td->td_flags & TDF_TSLEEPQ) 1363 tsleep_remove(td); 1364 KKASSERT(td->td_gd == gd); 1365 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); 1366 td->td_flags |= TDF_MIGRATING; 1367 crit_exit_gd(gd); 1368 } 1369 1370 void 1371 lwkt_acquire(thread_t td) 1372 { 1373 globaldata_t gd; 1374 globaldata_t mygd; 1375 int retry = 10000000; 1376 1377 KKASSERT(td->td_flags & TDF_MIGRATING); 1378 gd = td->td_gd; 1379 mygd = mycpu; 1380 if (gd != mycpu) { 1381 cpu_lfence(); 1382 KKASSERT((td->td_flags & TDF_RUNQ) == 0); 1383 crit_enter_gd(mygd); 1384 DEBUG_PUSH_INFO("lwkt_acquire"); 1385 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) { 1386 #ifdef SMP 1387 lwkt_process_ipiq(); 1388 #endif 1389 cpu_lfence(); 1390 if (--retry == 0) { 1391 kprintf("lwkt_acquire: stuck: td %p td->td_flags %08x\n", 1392 td, td->td_flags); 1393 retry = 10000000; 1394 } 1395 } 1396 DEBUG_POP_INFO(); 1397 cpu_mfence(); 1398 td->td_gd = mygd; 1399 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1400 td->td_flags &= ~TDF_MIGRATING; 1401 crit_exit_gd(mygd); 1402 } else { 1403 crit_enter_gd(mygd); 1404 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1405 td->td_flags &= ~TDF_MIGRATING; 1406 crit_exit_gd(mygd); 1407 } 1408 } 1409 1410 #endif 1411 1412 /* 1413 * Generic deschedule. Descheduling threads other then your own should be 1414 * done only in carefully controlled circumstances. Descheduling is 1415 * asynchronous. 1416 * 1417 * This function may block if the cpu has run out of messages. 1418 */ 1419 void 1420 lwkt_deschedule(thread_t td) 1421 { 1422 crit_enter(); 1423 #ifdef SMP 1424 if (td == curthread) { 1425 _lwkt_dequeue(td); 1426 } else { 1427 if (td->td_gd == mycpu) { 1428 _lwkt_dequeue(td); 1429 } else { 1430 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td); 1431 } 1432 } 1433 #else 1434 _lwkt_dequeue(td); 1435 #endif 1436 crit_exit(); 1437 } 1438 1439 /* 1440 * Set the target thread's priority. This routine does not automatically 1441 * switch to a higher priority thread, LWKT threads are not designed for 1442 * continuous priority changes. Yield if you want to switch. 1443 */ 1444 void 1445 lwkt_setpri(thread_t td, int pri) 1446 { 1447 if (td->td_pri != pri) { 1448 KKASSERT(pri >= 0); 1449 crit_enter(); 1450 if (td->td_flags & TDF_RUNQ) { 1451 KKASSERT(td->td_gd == mycpu); 1452 _lwkt_dequeue(td); 1453 td->td_pri = pri; 1454 _lwkt_enqueue(td); 1455 } else { 1456 td->td_pri = pri; 1457 } 1458 crit_exit(); 1459 } 1460 } 1461 1462 /* 1463 * Set the initial priority for a thread prior to it being scheduled for 1464 * the first time. The thread MUST NOT be scheduled before or during 1465 * this call. The thread may be assigned to a cpu other then the current 1466 * cpu. 1467 * 1468 * Typically used after a thread has been created with TDF_STOPPREQ, 1469 * and before the thread is initially scheduled. 1470 */ 1471 void 1472 lwkt_setpri_initial(thread_t td, int pri) 1473 { 1474 KKASSERT(pri >= 0); 1475 KKASSERT((td->td_flags & TDF_RUNQ) == 0); 1476 td->td_pri = pri; 1477 } 1478 1479 void 1480 lwkt_setpri_self(int pri) 1481 { 1482 thread_t td = curthread; 1483 1484 KKASSERT(pri >= 0 && pri <= TDPRI_MAX); 1485 crit_enter(); 1486 if (td->td_flags & TDF_RUNQ) { 1487 _lwkt_dequeue(td); 1488 td->td_pri = pri; 1489 _lwkt_enqueue(td); 1490 } else { 1491 td->td_pri = pri; 1492 } 1493 crit_exit(); 1494 } 1495 1496 /* 1497 * hz tick scheduler clock for LWKT threads 1498 */ 1499 void 1500 lwkt_schedulerclock(thread_t td) 1501 { 1502 globaldata_t gd = td->td_gd; 1503 thread_t xtd; 1504 1505 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) { 1506 /* 1507 * If the current thread is at the head of the runq shift it to the 1508 * end of any equal-priority threads and request a LWKT reschedule 1509 * if it moved. 1510 */ 1511 xtd = TAILQ_NEXT(td, td_threadq); 1512 if (xtd && xtd->td_pri == td->td_pri) { 1513 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq); 1514 while (xtd && xtd->td_pri == td->td_pri) 1515 xtd = TAILQ_NEXT(xtd, td_threadq); 1516 if (xtd) 1517 TAILQ_INSERT_BEFORE(xtd, td, td_threadq); 1518 else 1519 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 1520 need_lwkt_resched(); 1521 } 1522 } else { 1523 /* 1524 * If we scheduled a thread other than the one at the head of the 1525 * queue always request a reschedule every tick. 1526 */ 1527 need_lwkt_resched(); 1528 } 1529 } 1530 1531 /* 1532 * Migrate the current thread to the specified cpu. 1533 * 1534 * This is accomplished by descheduling ourselves from the current cpu 1535 * and setting td_migrate_gd. The lwkt_switch() code will detect that the 1536 * 'old' thread wants to migrate after it has been completely switched out 1537 * and will complete the migration. 1538 * 1539 * TDF_MIGRATING prevents scheduling races while the thread is being migrated. 1540 * 1541 * We must be sure to release our current process designation (if a user 1542 * process) before clearing out any tsleepq we are on because the release 1543 * code may re-add us. 1544 * 1545 * We must be sure to remove ourselves from the current cpu's tsleepq 1546 * before potentially moving to another queue. The thread can be on 1547 * a tsleepq due to a left-over tsleep_interlock(). 1548 */ 1549 1550 void 1551 lwkt_setcpu_self(globaldata_t rgd) 1552 { 1553 #ifdef SMP 1554 thread_t td = curthread; 1555 1556 if (td->td_gd != rgd) { 1557 crit_enter_quick(td); 1558 1559 if (td->td_release) 1560 td->td_release(td); 1561 if (td->td_flags & TDF_TSLEEPQ) 1562 tsleep_remove(td); 1563 1564 /* 1565 * Set TDF_MIGRATING to prevent a spurious reschedule while we are 1566 * trying to deschedule ourselves and switch away, then deschedule 1567 * ourself, remove us from tdallq, and set td_migrate_gd. Finally, 1568 * call lwkt_switch() to complete the operation. 1569 */ 1570 td->td_flags |= TDF_MIGRATING; 1571 lwkt_deschedule_self(td); 1572 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1573 td->td_migrate_gd = rgd; 1574 lwkt_switch(); 1575 1576 /* 1577 * We are now on the target cpu 1578 */ 1579 KKASSERT(rgd == mycpu); 1580 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); 1581 crit_exit_quick(td); 1582 } 1583 #endif 1584 } 1585 1586 void 1587 lwkt_migratecpu(int cpuid) 1588 { 1589 #ifdef SMP 1590 globaldata_t rgd; 1591 1592 rgd = globaldata_find(cpuid); 1593 lwkt_setcpu_self(rgd); 1594 #endif 1595 } 1596 1597 #ifdef SMP 1598 /* 1599 * Remote IPI for cpu migration (called while in a critical section so we 1600 * do not have to enter another one). 1601 * 1602 * The thread (td) has already been completely descheduled from the 1603 * originating cpu and we can simply assert the case. The thread is 1604 * assigned to the new cpu and enqueued. 1605 * 1606 * The thread will re-add itself to tdallq when it resumes execution. 1607 */ 1608 static void 1609 lwkt_setcpu_remote(void *arg) 1610 { 1611 thread_t td = arg; 1612 globaldata_t gd = mycpu; 1613 1614 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0); 1615 td->td_gd = gd; 1616 cpu_mfence(); 1617 td->td_flags &= ~TDF_MIGRATING; 1618 KKASSERT(td->td_migrate_gd == NULL); 1619 KKASSERT(td->td_lwp == NULL || 1620 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 1621 _lwkt_enqueue(td); 1622 } 1623 #endif 1624 1625 struct lwp * 1626 lwkt_preempted_proc(void) 1627 { 1628 thread_t td = curthread; 1629 while (td->td_preempted) 1630 td = td->td_preempted; 1631 return(td->td_lwp); 1632 } 1633 1634 /* 1635 * Create a kernel process/thread/whatever. It shares it's address space 1636 * with proc0 - ie: kernel only. 1637 * 1638 * If the cpu is not specified one will be selected. In the future 1639 * specifying a cpu of -1 will enable kernel thread migration between 1640 * cpus. 1641 */ 1642 int 1643 lwkt_create(void (*func)(void *), void *arg, struct thread **tdp, 1644 thread_t template, int tdflags, int cpu, const char *fmt, ...) 1645 { 1646 thread_t td; 1647 __va_list ap; 1648 1649 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu, 1650 tdflags); 1651 if (tdp) 1652 *tdp = td; 1653 cpu_set_thread_handler(td, lwkt_exit, func, arg); 1654 1655 /* 1656 * Set up arg0 for 'ps' etc 1657 */ 1658 __va_start(ap, fmt); 1659 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap); 1660 __va_end(ap); 1661 1662 /* 1663 * Schedule the thread to run 1664 */ 1665 if (td->td_flags & TDF_NOSTART) 1666 td->td_flags &= ~TDF_NOSTART; 1667 else 1668 lwkt_schedule(td); 1669 return 0; 1670 } 1671 1672 /* 1673 * Destroy an LWKT thread. Warning! This function is not called when 1674 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and 1675 * uses a different reaping mechanism. 1676 */ 1677 void 1678 lwkt_exit(void) 1679 { 1680 thread_t td = curthread; 1681 thread_t std; 1682 globaldata_t gd; 1683 1684 /* 1685 * Do any cleanup that might block here 1686 */ 1687 if (td->td_flags & TDF_VERBOSE) 1688 kprintf("kthread %p %s has exited\n", td, td->td_comm); 1689 caps_exit(td); 1690 biosched_done(td); 1691 dsched_exit_thread(td); 1692 1693 /* 1694 * Get us into a critical section to interlock gd_freetd and loop 1695 * until we can get it freed. 1696 * 1697 * We have to cache the current td in gd_freetd because objcache_put()ing 1698 * it would rip it out from under us while our thread is still active. 1699 * 1700 * We are the current thread so of course our own TDF_RUNNING bit will 1701 * be set, so unlike the lwp reap code we don't wait for it to clear. 1702 */ 1703 gd = mycpu; 1704 crit_enter_quick(td); 1705 for (;;) { 1706 if (td->td_refs) { 1707 tsleep(td, 0, "tdreap", 1); 1708 continue; 1709 } 1710 if ((std = gd->gd_freetd) != NULL) { 1711 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0); 1712 gd->gd_freetd = NULL; 1713 objcache_put(thread_cache, std); 1714 continue; 1715 } 1716 break; 1717 } 1718 1719 /* 1720 * Remove thread resources from kernel lists and deschedule us for 1721 * the last time. We cannot block after this point or we may end 1722 * up with a stale td on the tsleepq. 1723 * 1724 * None of this may block, the critical section is the only thing 1725 * protecting tdallq and the only thing preventing new lwkt_hold() 1726 * thread refs now. 1727 */ 1728 if (td->td_flags & TDF_TSLEEPQ) 1729 tsleep_remove(td); 1730 lwkt_deschedule_self(td); 1731 lwkt_remove_tdallq(td); 1732 KKASSERT(td->td_refs == 0); 1733 1734 /* 1735 * Final cleanup 1736 */ 1737 KKASSERT(gd->gd_freetd == NULL); 1738 if (td->td_flags & TDF_ALLOCATED_THREAD) 1739 gd->gd_freetd = td; 1740 cpu_thread_exit(); 1741 } 1742 1743 void 1744 lwkt_remove_tdallq(thread_t td) 1745 { 1746 KKASSERT(td->td_gd == mycpu); 1747 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1748 } 1749 1750 /* 1751 * Code reduction and branch prediction improvements. Call/return 1752 * overhead on modern cpus often degenerates into 0 cycles due to 1753 * the cpu's branch prediction hardware and return pc cache. We 1754 * can take advantage of this by not inlining medium-complexity 1755 * functions and we can also reduce the branch prediction impact 1756 * by collapsing perfectly predictable branches into a single 1757 * procedure instead of duplicating it. 1758 * 1759 * Is any of this noticeable? Probably not, so I'll take the 1760 * smaller code size. 1761 */ 1762 void 1763 crit_exit_wrapper(__DEBUG_CRIT_ARG__) 1764 { 1765 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__); 1766 } 1767 1768 void 1769 crit_panic(void) 1770 { 1771 thread_t td = curthread; 1772 int lcrit = td->td_critcount; 1773 1774 td->td_critcount = 0; 1775 panic("td_critcount is/would-go negative! %p %d", td, lcrit); 1776 /* NOT REACHED */ 1777 } 1778 1779 #ifdef SMP 1780 1781 /* 1782 * Called from debugger/panic on cpus which have been stopped. We must still 1783 * process the IPIQ while stopped, even if we were stopped while in a critical 1784 * section (XXX). 1785 * 1786 * If we are dumping also try to process any pending interrupts. This may 1787 * or may not work depending on the state of the cpu at the point it was 1788 * stopped. 1789 */ 1790 void 1791 lwkt_smp_stopped(void) 1792 { 1793 globaldata_t gd = mycpu; 1794 1795 crit_enter_gd(gd); 1796 if (dumping) { 1797 lwkt_process_ipiq(); 1798 splz(); 1799 } else { 1800 lwkt_process_ipiq(); 1801 } 1802 crit_exit_gd(gd); 1803 } 1804 1805 #endif 1806