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", int cpu, struct thread *td); 80 KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p", int cpu, struct thread *td); 81 KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s", struct thread *td, char *comm); 82 KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", struct thread *td); 83 84 static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads"); 85 86 #ifdef INVARIANTS 87 static int panic_on_cscount = 0; 88 #endif 89 static __int64_t switch_count = 0; 90 static __int64_t preempt_hit = 0; 91 static __int64_t preempt_miss = 0; 92 static __int64_t preempt_weird = 0; 93 static __int64_t token_contention_count[TDPRI_MAX+1] __debugvar; 94 static int lwkt_use_spin_port; 95 static struct objcache *thread_cache; 96 97 #ifdef SMP 98 static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame); 99 static void lwkt_setcpu_remote(void *arg); 100 #endif 101 102 extern void cpu_heavy_restore(void); 103 extern void cpu_lwkt_restore(void); 104 extern void cpu_kthread_restore(void); 105 extern void cpu_idle_restore(void); 106 107 /* 108 * We can make all thread ports use the spin backend instead of the thread 109 * backend. This should only be set to debug the spin backend. 110 */ 111 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port); 112 113 #ifdef INVARIANTS 114 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, 115 "Panic if attempting to switch lwkt's while mastering cpusync"); 116 #endif 117 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, 118 "Number of switched threads"); 119 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, 120 "Successful preemption events"); 121 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, 122 "Failed preemption events"); 123 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, 124 "Number of preempted threads."); 125 #ifdef INVARIANTS 126 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_00, CTLFLAG_RW, 127 &token_contention_count[0], 0, "spinning due to token contention"); 128 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_01, CTLFLAG_RW, 129 &token_contention_count[1], 0, "spinning due to token contention"); 130 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_02, CTLFLAG_RW, 131 &token_contention_count[2], 0, "spinning due to token contention"); 132 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_03, CTLFLAG_RW, 133 &token_contention_count[3], 0, "spinning due to token contention"); 134 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_04, CTLFLAG_RW, 135 &token_contention_count[4], 0, "spinning due to token contention"); 136 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_05, CTLFLAG_RW, 137 &token_contention_count[5], 0, "spinning due to token contention"); 138 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_06, CTLFLAG_RW, 139 &token_contention_count[6], 0, "spinning due to token contention"); 140 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_07, CTLFLAG_RW, 141 &token_contention_count[7], 0, "spinning due to token contention"); 142 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_08, CTLFLAG_RW, 143 &token_contention_count[8], 0, "spinning due to token contention"); 144 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_09, CTLFLAG_RW, 145 &token_contention_count[9], 0, "spinning due to token contention"); 146 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_10, CTLFLAG_RW, 147 &token_contention_count[10], 0, "spinning due to token contention"); 148 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_11, CTLFLAG_RW, 149 &token_contention_count[11], 0, "spinning due to token contention"); 150 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_12, CTLFLAG_RW, 151 &token_contention_count[12], 0, "spinning due to token contention"); 152 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_13, CTLFLAG_RW, 153 &token_contention_count[13], 0, "spinning due to token contention"); 154 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_14, CTLFLAG_RW, 155 &token_contention_count[14], 0, "spinning due to token contention"); 156 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_15, CTLFLAG_RW, 157 &token_contention_count[15], 0, "spinning due to token contention"); 158 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_16, CTLFLAG_RW, 159 &token_contention_count[16], 0, "spinning due to token contention"); 160 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_17, CTLFLAG_RW, 161 &token_contention_count[17], 0, "spinning due to token contention"); 162 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_18, CTLFLAG_RW, 163 &token_contention_count[18], 0, "spinning due to token contention"); 164 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_19, CTLFLAG_RW, 165 &token_contention_count[19], 0, "spinning due to token contention"); 166 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_20, CTLFLAG_RW, 167 &token_contention_count[20], 0, "spinning due to token contention"); 168 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_21, CTLFLAG_RW, 169 &token_contention_count[21], 0, "spinning due to token contention"); 170 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_22, CTLFLAG_RW, 171 &token_contention_count[22], 0, "spinning due to token contention"); 172 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_23, CTLFLAG_RW, 173 &token_contention_count[23], 0, "spinning due to token contention"); 174 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_24, CTLFLAG_RW, 175 &token_contention_count[24], 0, "spinning due to token contention"); 176 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_25, CTLFLAG_RW, 177 &token_contention_count[25], 0, "spinning due to token contention"); 178 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_26, CTLFLAG_RW, 179 &token_contention_count[26], 0, "spinning due to token contention"); 180 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_27, CTLFLAG_RW, 181 &token_contention_count[27], 0, "spinning due to token contention"); 182 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_28, CTLFLAG_RW, 183 &token_contention_count[28], 0, "spinning due to token contention"); 184 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_29, CTLFLAG_RW, 185 &token_contention_count[29], 0, "spinning due to token contention"); 186 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_30, CTLFLAG_RW, 187 &token_contention_count[30], 0, "spinning due to token contention"); 188 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_31, CTLFLAG_RW, 189 &token_contention_count[31], 0, "spinning due to token contention"); 190 #endif 191 static int fairq_enable = 0; 192 SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW, 193 &fairq_enable, 0, "Turn on fairq priority accumulators"); 194 static int fairq_bypass = -1; 195 SYSCTL_INT(_lwkt, OID_AUTO, fairq_bypass, CTLFLAG_RW, 196 &fairq_bypass, 0, "Allow fairq to bypass td on token failure"); 197 extern int lwkt_sched_debug; 198 int lwkt_sched_debug = 0; 199 SYSCTL_INT(_lwkt, OID_AUTO, sched_debug, CTLFLAG_RW, 200 &lwkt_sched_debug, 0, "Scheduler debug"); 201 static int lwkt_spin_loops = 10; 202 SYSCTL_INT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW, 203 &lwkt_spin_loops, 0, "Scheduler spin loops until sorted decon"); 204 static int lwkt_spin_reseq = 0; 205 SYSCTL_INT(_lwkt, OID_AUTO, spin_reseq, CTLFLAG_RW, 206 &lwkt_spin_reseq, 0, "Scheduler resequencer enable"); 207 static int lwkt_spin_monitor = 0; 208 SYSCTL_INT(_lwkt, OID_AUTO, spin_monitor, CTLFLAG_RW, 209 &lwkt_spin_monitor, 0, "Scheduler uses monitor/mwait"); 210 static int lwkt_spin_fatal = 0; /* disabled */ 211 SYSCTL_INT(_lwkt, OID_AUTO, spin_fatal, CTLFLAG_RW, 212 &lwkt_spin_fatal, 0, "LWKT scheduler spin loops till fatal panic"); 213 static int preempt_enable = 1; 214 SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW, 215 &preempt_enable, 0, "Enable preemption"); 216 static int lwkt_cache_threads = 0; 217 SYSCTL_INT(_lwkt, OID_AUTO, cache_threads, CTLFLAG_RD, 218 &lwkt_cache_threads, 0, "thread+kstack cache"); 219 220 static __cachealign int lwkt_cseq_rindex; 221 static __cachealign int lwkt_cseq_windex; 222 223 /* 224 * These helper procedures handle the runq, they can only be called from 225 * within a critical section. 226 * 227 * WARNING! Prior to SMP being brought up it is possible to enqueue and 228 * dequeue threads belonging to other cpus, so be sure to use td->td_gd 229 * instead of 'mycpu' when referencing the globaldata structure. Once 230 * SMP live enqueuing and dequeueing only occurs on the current cpu. 231 */ 232 static __inline 233 void 234 _lwkt_dequeue(thread_t td) 235 { 236 if (td->td_flags & TDF_RUNQ) { 237 struct globaldata *gd = td->td_gd; 238 239 td->td_flags &= ~TDF_RUNQ; 240 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq); 241 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL) 242 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING); 243 } 244 } 245 246 /* 247 * Priority enqueue. 248 * 249 * NOTE: There are a limited number of lwkt threads runnable since user 250 * processes only schedule one at a time per cpu. 251 */ 252 static __inline 253 void 254 _lwkt_enqueue(thread_t td) 255 { 256 thread_t xtd; 257 258 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) { 259 struct globaldata *gd = td->td_gd; 260 261 td->td_flags |= TDF_RUNQ; 262 xtd = TAILQ_FIRST(&gd->gd_tdrunq); 263 if (xtd == NULL) { 264 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 265 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING); 266 } else { 267 while (xtd && xtd->td_pri >= td->td_pri) 268 xtd = TAILQ_NEXT(xtd, td_threadq); 269 if (xtd) 270 TAILQ_INSERT_BEFORE(xtd, td, td_threadq); 271 else 272 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 273 } 274 275 /* 276 * Request a LWKT reschedule if we are now at the head of the queue. 277 */ 278 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) 279 need_lwkt_resched(); 280 } 281 } 282 283 static __boolean_t 284 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags) 285 { 286 struct thread *td = (struct thread *)obj; 287 288 td->td_kstack = NULL; 289 td->td_kstack_size = 0; 290 td->td_flags = TDF_ALLOCATED_THREAD; 291 td->td_mpflags = 0; 292 return (1); 293 } 294 295 static void 296 _lwkt_thread_dtor(void *obj, void *privdata) 297 { 298 struct thread *td = (struct thread *)obj; 299 300 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD, 301 ("_lwkt_thread_dtor: not allocated from objcache")); 302 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack && 303 td->td_kstack_size > 0, 304 ("_lwkt_thread_dtor: corrupted stack")); 305 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size); 306 td->td_kstack = NULL; 307 td->td_flags = 0; 308 } 309 310 /* 311 * Initialize the lwkt s/system. 312 * 313 * Nominally cache up to 32 thread + kstack structures. Cache more on 314 * systems with a lot of cpu cores. 315 */ 316 void 317 lwkt_init(void) 318 { 319 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads); 320 if (lwkt_cache_threads == 0) { 321 lwkt_cache_threads = ncpus * 4; 322 if (lwkt_cache_threads < 32) 323 lwkt_cache_threads = 32; 324 } 325 thread_cache = objcache_create_mbacked( 326 M_THREAD, sizeof(struct thread), 327 NULL, lwkt_cache_threads, 328 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL); 329 } 330 331 /* 332 * Schedule a thread to run. As the current thread we can always safely 333 * schedule ourselves, and a shortcut procedure is provided for that 334 * function. 335 * 336 * (non-blocking, self contained on a per cpu basis) 337 */ 338 void 339 lwkt_schedule_self(thread_t td) 340 { 341 KKASSERT((td->td_flags & TDF_MIGRATING) == 0); 342 crit_enter_quick(td); 343 KASSERT(td != &td->td_gd->gd_idlethread, 344 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!")); 345 KKASSERT(td->td_lwp == NULL || 346 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 347 _lwkt_enqueue(td); 348 crit_exit_quick(td); 349 } 350 351 /* 352 * Deschedule a thread. 353 * 354 * (non-blocking, self contained on a per cpu basis) 355 */ 356 void 357 lwkt_deschedule_self(thread_t td) 358 { 359 crit_enter_quick(td); 360 _lwkt_dequeue(td); 361 crit_exit_quick(td); 362 } 363 364 /* 365 * LWKTs operate on a per-cpu basis 366 * 367 * WARNING! Called from early boot, 'mycpu' may not work yet. 368 */ 369 void 370 lwkt_gdinit(struct globaldata *gd) 371 { 372 TAILQ_INIT(&gd->gd_tdrunq); 373 TAILQ_INIT(&gd->gd_tdallq); 374 } 375 376 /* 377 * Create a new thread. The thread must be associated with a process context 378 * or LWKT start address before it can be scheduled. If the target cpu is 379 * -1 the thread will be created on the current cpu. 380 * 381 * If you intend to create a thread without a process context this function 382 * does everything except load the startup and switcher function. 383 */ 384 thread_t 385 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags) 386 { 387 static int cpu_rotator; 388 globaldata_t gd = mycpu; 389 void *stack; 390 391 /* 392 * If static thread storage is not supplied allocate a thread. Reuse 393 * a cached free thread if possible. gd_freetd is used to keep an exiting 394 * thread intact through the exit. 395 */ 396 if (td == NULL) { 397 crit_enter_gd(gd); 398 if ((td = gd->gd_freetd) != NULL) { 399 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK| 400 TDF_RUNQ)) == 0); 401 gd->gd_freetd = NULL; 402 } else { 403 td = objcache_get(thread_cache, M_WAITOK); 404 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK| 405 TDF_RUNQ)) == 0); 406 } 407 crit_exit_gd(gd); 408 KASSERT((td->td_flags & 409 (TDF_ALLOCATED_THREAD|TDF_RUNNING|TDF_PREEMPT_LOCK)) == 410 TDF_ALLOCATED_THREAD, 411 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags)); 412 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK); 413 } 414 415 /* 416 * Try to reuse cached stack. 417 */ 418 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) { 419 if (flags & TDF_ALLOCATED_STACK) { 420 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size); 421 stack = NULL; 422 } 423 } 424 if (stack == NULL) { 425 stack = (void *)kmem_alloc_stack(&kernel_map, stksize); 426 flags |= TDF_ALLOCATED_STACK; 427 } 428 if (cpu < 0) { 429 cpu = ++cpu_rotator; 430 cpu_ccfence(); 431 cpu %= ncpus; 432 } 433 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu)); 434 return(td); 435 } 436 437 /* 438 * Initialize a preexisting thread structure. This function is used by 439 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread. 440 * 441 * All threads start out in a critical section at a priority of 442 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as 443 * appropriate. This function may send an IPI message when the 444 * requested cpu is not the current cpu and consequently gd_tdallq may 445 * not be initialized synchronously from the point of view of the originating 446 * cpu. 447 * 448 * NOTE! we have to be careful in regards to creating threads for other cpus 449 * if SMP has not yet been activated. 450 */ 451 #ifdef SMP 452 453 static void 454 lwkt_init_thread_remote(void *arg) 455 { 456 thread_t td = arg; 457 458 /* 459 * Protected by critical section held by IPI dispatch 460 */ 461 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq); 462 } 463 464 #endif 465 466 /* 467 * lwkt core thread structural initialization. 468 * 469 * NOTE: All threads are initialized as mpsafe threads. 470 */ 471 void 472 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags, 473 struct globaldata *gd) 474 { 475 globaldata_t mygd = mycpu; 476 477 bzero(td, sizeof(struct thread)); 478 td->td_kstack = stack; 479 td->td_kstack_size = stksize; 480 td->td_flags = flags; 481 td->td_mpflags = 0; 482 td->td_gd = gd; 483 td->td_pri = TDPRI_KERN_DAEMON; 484 td->td_critcount = 1; 485 td->td_toks_have = NULL; 486 td->td_toks_stop = &td->td_toks_base; 487 if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT)) 488 lwkt_initport_spin(&td->td_msgport); 489 else 490 lwkt_initport_thread(&td->td_msgport, td); 491 pmap_init_thread(td); 492 #ifdef SMP 493 /* 494 * Normally initializing a thread for a remote cpu requires sending an 495 * IPI. However, the idlethread is setup before the other cpus are 496 * activated so we have to treat it as a special case. XXX manipulation 497 * of gd_tdallq requires the BGL. 498 */ 499 if (gd == mygd || td == &gd->gd_idlethread) { 500 crit_enter_gd(mygd); 501 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); 502 crit_exit_gd(mygd); 503 } else { 504 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td); 505 } 506 #else 507 crit_enter_gd(mygd); 508 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); 509 crit_exit_gd(mygd); 510 #endif 511 512 dsched_new_thread(td); 513 } 514 515 void 516 lwkt_set_comm(thread_t td, const char *ctl, ...) 517 { 518 __va_list va; 519 520 __va_start(va, ctl); 521 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va); 522 __va_end(va); 523 KTR_LOG(ctxsw_newtd, td, td->td_comm); 524 } 525 526 /* 527 * Prevent the thread from getting destroyed. Note that unlike PHOLD/PRELE 528 * this does not prevent the thread from migrating to another cpu so the 529 * gd_tdallq state is not protected by this. 530 */ 531 void 532 lwkt_hold(thread_t td) 533 { 534 atomic_add_int(&td->td_refs, 1); 535 } 536 537 void 538 lwkt_rele(thread_t td) 539 { 540 KKASSERT(td->td_refs > 0); 541 atomic_add_int(&td->td_refs, -1); 542 } 543 544 void 545 lwkt_free_thread(thread_t td) 546 { 547 KKASSERT(td->td_refs == 0); 548 KKASSERT((td->td_flags & (TDF_RUNNING | TDF_PREEMPT_LOCK | 549 TDF_RUNQ | TDF_TSLEEPQ)) == 0); 550 if (td->td_flags & TDF_ALLOCATED_THREAD) { 551 objcache_put(thread_cache, td); 552 } else if (td->td_flags & TDF_ALLOCATED_STACK) { 553 /* client-allocated struct with internally allocated stack */ 554 KASSERT(td->td_kstack && td->td_kstack_size > 0, 555 ("lwkt_free_thread: corrupted stack")); 556 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size); 557 td->td_kstack = NULL; 558 td->td_kstack_size = 0; 559 } 560 KTR_LOG(ctxsw_deadtd, td); 561 } 562 563 564 /* 565 * Switch to the next runnable lwkt. If no LWKTs are runnable then 566 * switch to the idlethread. Switching must occur within a critical 567 * section to avoid races with the scheduling queue. 568 * 569 * We always have full control over our cpu's run queue. Other cpus 570 * that wish to manipulate our queue must use the cpu_*msg() calls to 571 * talk to our cpu, so a critical section is all that is needed and 572 * the result is very, very fast thread switching. 573 * 574 * The LWKT scheduler uses a fixed priority model and round-robins at 575 * each priority level. User process scheduling is a totally 576 * different beast and LWKT priorities should not be confused with 577 * user process priorities. 578 * 579 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch() 580 * is not called by the current thread in the preemption case, only when 581 * the preempting thread blocks (in order to return to the original thread). 582 * 583 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread 584 * migration and tsleep deschedule the current lwkt thread and call 585 * lwkt_switch(). In particular, the target cpu of the migration fully 586 * expects the thread to become non-runnable and can deadlock against 587 * cpusync operations if we run any IPIs prior to switching the thread out. 588 * 589 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF 590 * THE CURRENT THREAD HAS BEEN DESCHEDULED! 591 */ 592 void 593 lwkt_switch(void) 594 { 595 globaldata_t gd = mycpu; 596 thread_t td = gd->gd_curthread; 597 thread_t ntd; 598 int spinning = 0; 599 600 KKASSERT(gd->gd_processing_ipiq == 0); 601 KKASSERT(td->td_flags & TDF_RUNNING); 602 603 /* 604 * Switching from within a 'fast' (non thread switched) interrupt or IPI 605 * is illegal. However, we may have to do it anyway if we hit a fatal 606 * kernel trap or we have paniced. 607 * 608 * If this case occurs save and restore the interrupt nesting level. 609 */ 610 if (gd->gd_intr_nesting_level) { 611 int savegdnest; 612 int savegdtrap; 613 614 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) { 615 panic("lwkt_switch: Attempt to switch from a " 616 "a fast interrupt, ipi, or hard code section, " 617 "td %p\n", 618 td); 619 } else { 620 savegdnest = gd->gd_intr_nesting_level; 621 savegdtrap = gd->gd_trap_nesting_level; 622 gd->gd_intr_nesting_level = 0; 623 gd->gd_trap_nesting_level = 0; 624 if ((td->td_flags & TDF_PANICWARN) == 0) { 625 td->td_flags |= TDF_PANICWARN; 626 kprintf("Warning: thread switch from interrupt, IPI, " 627 "or hard code section.\n" 628 "thread %p (%s)\n", td, td->td_comm); 629 print_backtrace(-1); 630 } 631 lwkt_switch(); 632 gd->gd_intr_nesting_level = savegdnest; 633 gd->gd_trap_nesting_level = savegdtrap; 634 return; 635 } 636 } 637 638 /* 639 * Release our current user process designation if we are blocking 640 * or if a user reschedule was requested. 641 * 642 * NOTE: This function is NOT called if we are switching into or 643 * returning from a preemption. 644 * 645 * NOTE: Releasing our current user process designation may cause 646 * it to be assigned to another thread, which in turn will 647 * cause us to block in the usched acquire code when we attempt 648 * to return to userland. 649 * 650 * NOTE: On SMP systems this can be very nasty when heavy token 651 * contention is present so we want to be careful not to 652 * release the designation gratuitously. 653 */ 654 if (td->td_release && 655 (user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) { 656 td->td_release(td); 657 } 658 659 /* 660 * Release all tokens 661 */ 662 crit_enter_gd(gd); 663 if (TD_TOKS_HELD(td)) 664 lwkt_relalltokens(td); 665 666 /* 667 * We had better not be holding any spin locks, but don't get into an 668 * endless panic loop. 669 */ 670 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL, 671 ("lwkt_switch: still holding %d exclusive spinlocks!", 672 gd->gd_spinlocks_wr)); 673 674 675 #ifdef SMP 676 #ifdef INVARIANTS 677 if (td->td_cscount) { 678 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n", 679 td); 680 if (panic_on_cscount) 681 panic("switching while mastering cpusync"); 682 } 683 #endif 684 #endif 685 686 /* 687 * If we had preempted another thread on this cpu, resume the preempted 688 * thread. This occurs transparently, whether the preempted thread 689 * was scheduled or not (it may have been preempted after descheduling 690 * itself). 691 * 692 * We have to setup the MP lock for the original thread after backing 693 * out the adjustment that was made to curthread when the original 694 * was preempted. 695 */ 696 if ((ntd = td->td_preempted) != NULL) { 697 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK); 698 ntd->td_flags |= TDF_PREEMPT_DONE; 699 700 /* 701 * The interrupt may have woken a thread up, we need to properly 702 * set the reschedule flag if the originally interrupted thread is 703 * at a lower priority. 704 * 705 * The interrupt may not have descheduled. 706 */ 707 if (TAILQ_FIRST(&gd->gd_tdrunq) != ntd) 708 need_lwkt_resched(); 709 goto havethread_preempted; 710 } 711 712 /* 713 * If we cannot obtain ownership of the tokens we cannot immediately 714 * schedule the target thread. 715 * 716 * Reminder: Again, we cannot afford to run any IPIs in this path if 717 * the current thread has been descheduled. 718 */ 719 for (;;) { 720 clear_lwkt_resched(); 721 722 /* 723 * Hotpath - pull the head of the run queue and attempt to schedule 724 * it. 725 */ 726 for (;;) { 727 ntd = TAILQ_FIRST(&gd->gd_tdrunq); 728 729 if (ntd == NULL) { 730 /* 731 * Runq is empty, switch to idle to allow it to halt. 732 */ 733 ntd = &gd->gd_idlethread; 734 #ifdef SMP 735 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) 736 ASSERT_NO_TOKENS_HELD(ntd); 737 #endif 738 cpu_time.cp_msg[0] = 0; 739 cpu_time.cp_stallpc = 0; 740 goto haveidle; 741 } 742 break; 743 } 744 745 /* 746 * Hotpath - schedule ntd. 747 * 748 * NOTE: For UP there is no mplock and lwkt_getalltokens() 749 * always succeeds. 750 */ 751 if (TD_TOKS_NOT_HELD(ntd) || 752 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops))) 753 { 754 goto havethread; 755 } 756 757 /* 758 * Coldpath (SMP only since tokens always succeed on UP) 759 * 760 * We had some contention on the thread we wanted to schedule. 761 * What we do now is try to find a thread that we can schedule 762 * in its stead. 763 * 764 * The coldpath scan does NOT rearrange threads in the run list. 765 * The lwkt_schedulerclock() will assert need_lwkt_resched() on 766 * the next tick whenever the current head is not the current thread. 767 */ 768 #ifdef INVARIANTS 769 ++token_contention_count[ntd->td_pri]; 770 ++ntd->td_contended; 771 #endif 772 773 if (fairq_bypass > 0) 774 goto skip; 775 776 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) { 777 /* 778 * Never schedule threads returning to userland or the 779 * user thread scheduler helper thread when higher priority 780 * threads are present. 781 */ 782 if (ntd->td_pri < TDPRI_KERN_LPSCHED) { 783 ntd = NULL; 784 break; 785 } 786 787 /* 788 * Try this one. 789 */ 790 if (TD_TOKS_NOT_HELD(ntd) || 791 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops))) { 792 goto havethread; 793 } 794 #ifdef INVARIANTS 795 ++token_contention_count[ntd->td_pri]; 796 ++ntd->td_contended; 797 #endif 798 } 799 800 skip: 801 /* 802 * We exhausted the run list, meaning that all runnable threads 803 * are contested. 804 */ 805 cpu_pause(); 806 ntd = &gd->gd_idlethread; 807 #ifdef SMP 808 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) 809 ASSERT_NO_TOKENS_HELD(ntd); 810 /* contention case, do not clear contention mask */ 811 #endif 812 813 /* 814 * We are going to have to retry but if the current thread is not 815 * on the runq we instead switch through the idle thread to get away 816 * from the current thread. We have to flag for lwkt reschedule 817 * to prevent the idle thread from halting. 818 * 819 * NOTE: A non-zero spinning is passed to lwkt_getalltokens() to 820 * instruct it to deal with the potential for deadlocks by 821 * ordering the tokens by address. 822 */ 823 if ((td->td_flags & TDF_RUNQ) == 0) { 824 need_lwkt_resched(); /* prevent hlt */ 825 goto haveidle; 826 } 827 #if defined(INVARIANTS) && defined(__amd64__) 828 if ((read_rflags() & PSL_I) == 0) { 829 cpu_enable_intr(); 830 panic("lwkt_switch() called with interrupts disabled"); 831 } 832 #endif 833 834 /* 835 * Number iterations so far. After a certain point we switch to 836 * a sorted-address/monitor/mwait version of lwkt_getalltokens() 837 */ 838 if (spinning < 0x7FFFFFFF) 839 ++spinning; 840 841 #ifdef SMP 842 /* 843 * lwkt_getalltokens() failed in sorted token mode, we can use 844 * monitor/mwait in this case. 845 */ 846 if (spinning >= lwkt_spin_loops && 847 (cpu_mi_feature & CPU_MI_MONITOR) && 848 lwkt_spin_monitor) 849 { 850 cpu_mmw_pause_int(&gd->gd_reqflags, 851 (gd->gd_reqflags | RQF_SPINNING) & 852 ~RQF_IDLECHECK_WK_MASK); 853 } 854 #endif 855 856 /* 857 * We already checked that td is still scheduled so this should be 858 * safe. 859 */ 860 splz_check(); 861 862 /* 863 * This experimental resequencer is used as a fall-back to reduce 864 * hw cache line contention by placing each core's scheduler into a 865 * time-domain-multplexed slot. 866 * 867 * The resequencer is disabled by default. It's functionality has 868 * largely been superceeded by the token algorithm which limits races 869 * to a subset of cores. 870 * 871 * The resequencer algorithm tends to break down when more than 872 * 20 cores are contending. What appears to happen is that new 873 * tokens can be obtained out of address-sorted order by new cores 874 * while existing cores languish in long delays between retries and 875 * wind up being starved-out of the token acquisition. 876 */ 877 if (lwkt_spin_reseq && spinning >= lwkt_spin_reseq) { 878 int cseq = atomic_fetchadd_int(&lwkt_cseq_windex, 1); 879 int oseq; 880 881 while ((oseq = lwkt_cseq_rindex) != cseq) { 882 cpu_ccfence(); 883 #if 1 884 if (cpu_mi_feature & CPU_MI_MONITOR) { 885 cpu_mmw_pause_int(&lwkt_cseq_rindex, oseq); 886 } else { 887 #endif 888 cpu_pause(); 889 cpu_lfence(); 890 #if 1 891 } 892 #endif 893 } 894 DELAY(1); 895 atomic_add_int(&lwkt_cseq_rindex, 1); 896 } 897 /* highest level for(;;) loop */ 898 } 899 900 havethread: 901 /* 902 * Clear gd_idle_repeat when doing a normal switch to a non-idle 903 * thread. 904 */ 905 ntd->td_wmesg = NULL; 906 ++gd->gd_cnt.v_swtch; 907 gd->gd_idle_repeat = 0; 908 909 havethread_preempted: 910 /* 911 * If the new target does not need the MP lock and we are holding it, 912 * release the MP lock. If the new target requires the MP lock we have 913 * already acquired it for the target. 914 */ 915 ; 916 haveidle: 917 KASSERT(ntd->td_critcount, 918 ("priority problem in lwkt_switch %d %d", 919 td->td_critcount, ntd->td_critcount)); 920 921 if (td != ntd) { 922 /* 923 * Execute the actual thread switch operation. This function 924 * returns to the current thread and returns the previous thread 925 * (which may be different from the thread we switched to). 926 * 927 * We are responsible for marking ntd as TDF_RUNNING. 928 */ 929 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0); 930 ++switch_count; 931 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd); 932 ntd->td_flags |= TDF_RUNNING; 933 lwkt_switch_return(td->td_switch(ntd)); 934 /* ntd invalid, td_switch() can return a different thread_t */ 935 } 936 937 /* 938 * catch-all. XXX is this strictly needed? 939 */ 940 splz_check(); 941 942 /* NOTE: current cpu may have changed after switch */ 943 crit_exit_quick(td); 944 } 945 946 /* 947 * Called by assembly in the td_switch (thread restore path) for thread 948 * bootstrap cases which do not 'return' to lwkt_switch(). 949 */ 950 void 951 lwkt_switch_return(thread_t otd) 952 { 953 #ifdef SMP 954 globaldata_t rgd; 955 956 /* 957 * Check if otd was migrating. Now that we are on ntd we can finish 958 * up the migration. This is a bit messy but it is the only place 959 * where td is known to be fully descheduled. 960 * 961 * We can only activate the migration if otd was migrating but not 962 * held on the cpu due to a preemption chain. We still have to 963 * clear TDF_RUNNING on the old thread either way. 964 * 965 * We are responsible for clearing the previously running thread's 966 * TDF_RUNNING. 967 */ 968 if ((rgd = otd->td_migrate_gd) != NULL && 969 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) { 970 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) == 971 (TDF_MIGRATING | TDF_RUNNING)); 972 otd->td_migrate_gd = NULL; 973 otd->td_flags &= ~TDF_RUNNING; 974 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd); 975 } else { 976 otd->td_flags &= ~TDF_RUNNING; 977 } 978 #else 979 otd->td_flags &= ~TDF_RUNNING; 980 #endif 981 } 982 983 /* 984 * Request that the target thread preempt the current thread. Preemption 985 * can only occur if our only critical section is the one that we were called 986 * with, the relative priority of the target thread is higher, and the target 987 * thread holds no tokens. This also only works if we are not holding any 988 * spinlocks (obviously). 989 * 990 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically 991 * this is called via lwkt_schedule() through the td_preemptable callback. 992 * critcount is the managed critical priority that we should ignore in order 993 * to determine whether preemption is possible (aka usually just the crit 994 * priority of lwkt_schedule() itself). 995 * 996 * Preemption is typically limited to interrupt threads. 997 * 998 * Operation works in a fairly straight-forward manner. The normal 999 * scheduling code is bypassed and we switch directly to the target 1000 * thread. When the target thread attempts to block or switch away 1001 * code at the base of lwkt_switch() will switch directly back to our 1002 * thread. Our thread is able to retain whatever tokens it holds and 1003 * if the target needs one of them the target will switch back to us 1004 * and reschedule itself normally. 1005 */ 1006 void 1007 lwkt_preempt(thread_t ntd, int critcount) 1008 { 1009 struct globaldata *gd = mycpu; 1010 thread_t xtd; 1011 thread_t td; 1012 int save_gd_intr_nesting_level; 1013 1014 /* 1015 * The caller has put us in a critical section. We can only preempt 1016 * if the caller of the caller was not in a critical section (basically 1017 * a local interrupt), as determined by the 'critcount' parameter. We 1018 * also can't preempt if the caller is holding any spinlocks (even if 1019 * he isn't in a critical section). This also handles the tokens test. 1020 * 1021 * YYY The target thread must be in a critical section (else it must 1022 * inherit our critical section? I dunno yet). 1023 */ 1024 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri)); 1025 1026 td = gd->gd_curthread; 1027 if (preempt_enable == 0) { 1028 ++preempt_miss; 1029 return; 1030 } 1031 if (ntd->td_pri <= td->td_pri) { 1032 ++preempt_miss; 1033 return; 1034 } 1035 if (td->td_critcount > critcount) { 1036 ++preempt_miss; 1037 return; 1038 } 1039 #ifdef SMP 1040 if (td->td_cscount) { 1041 ++preempt_miss; 1042 return; 1043 } 1044 if (ntd->td_gd != gd) { 1045 ++preempt_miss; 1046 return; 1047 } 1048 #endif 1049 /* 1050 * We don't have to check spinlocks here as they will also bump 1051 * td_critcount. 1052 * 1053 * Do not try to preempt if the target thread is holding any tokens. 1054 * We could try to acquire the tokens but this case is so rare there 1055 * is no need to support it. 1056 */ 1057 KKASSERT(gd->gd_spinlocks_wr == 0); 1058 1059 if (TD_TOKS_HELD(ntd)) { 1060 ++preempt_miss; 1061 return; 1062 } 1063 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) { 1064 ++preempt_weird; 1065 return; 1066 } 1067 if (ntd->td_preempted) { 1068 ++preempt_hit; 1069 return; 1070 } 1071 KKASSERT(gd->gd_processing_ipiq == 0); 1072 1073 /* 1074 * Since we are able to preempt the current thread, there is no need to 1075 * call need_lwkt_resched(). 1076 * 1077 * We must temporarily clear gd_intr_nesting_level around the switch 1078 * since switchouts from the target thread are allowed (they will just 1079 * return to our thread), and since the target thread has its own stack. 1080 * 1081 * A preemption must switch back to the original thread, assert the 1082 * case. 1083 */ 1084 ++preempt_hit; 1085 ntd->td_preempted = td; 1086 td->td_flags |= TDF_PREEMPT_LOCK; 1087 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd); 1088 save_gd_intr_nesting_level = gd->gd_intr_nesting_level; 1089 gd->gd_intr_nesting_level = 0; 1090 1091 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0); 1092 ntd->td_flags |= TDF_RUNNING; 1093 xtd = td->td_switch(ntd); 1094 KKASSERT(xtd == ntd); 1095 lwkt_switch_return(xtd); 1096 gd->gd_intr_nesting_level = save_gd_intr_nesting_level; 1097 1098 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE)); 1099 ntd->td_preempted = NULL; 1100 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE); 1101 } 1102 1103 /* 1104 * Conditionally call splz() if gd_reqflags indicates work is pending. 1105 * This will work inside a critical section but not inside a hard code 1106 * section. 1107 * 1108 * (self contained on a per cpu basis) 1109 */ 1110 void 1111 splz_check(void) 1112 { 1113 globaldata_t gd = mycpu; 1114 thread_t td = gd->gd_curthread; 1115 1116 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && 1117 gd->gd_intr_nesting_level == 0 && 1118 td->td_nest_count < 2) 1119 { 1120 splz(); 1121 } 1122 } 1123 1124 /* 1125 * This version is integrated into crit_exit, reqflags has already 1126 * been tested but td_critcount has not. 1127 * 1128 * We only want to execute the splz() on the 1->0 transition of 1129 * critcount and not in a hard code section or if too deeply nested. 1130 * 1131 * NOTE: gd->gd_spinlocks_wr is implied to be 0 when td_critcount is 0. 1132 */ 1133 void 1134 lwkt_maybe_splz(thread_t td) 1135 { 1136 globaldata_t gd = td->td_gd; 1137 1138 if (td->td_critcount == 0 && 1139 gd->gd_intr_nesting_level == 0 && 1140 td->td_nest_count < 2) 1141 { 1142 splz(); 1143 } 1144 } 1145 1146 /* 1147 * Drivers which set up processing co-threads can call this function to 1148 * run the co-thread at a higher priority and to allow it to preempt 1149 * normal threads. 1150 */ 1151 void 1152 lwkt_set_interrupt_support_thread(void) 1153 { 1154 thread_t td = curthread; 1155 1156 lwkt_setpri_self(TDPRI_INT_SUPPORT); 1157 td->td_flags |= TDF_INTTHREAD; 1158 td->td_preemptable = lwkt_preempt; 1159 } 1160 1161 1162 /* 1163 * This function is used to negotiate a passive release of the current 1164 * process/lwp designation with the user scheduler, allowing the user 1165 * scheduler to schedule another user thread. The related kernel thread 1166 * (curthread) continues running in the released state. 1167 */ 1168 void 1169 lwkt_passive_release(struct thread *td) 1170 { 1171 struct lwp *lp = td->td_lwp; 1172 1173 td->td_release = NULL; 1174 lwkt_setpri_self(TDPRI_KERN_USER); 1175 lp->lwp_proc->p_usched->release_curproc(lp); 1176 } 1177 1178 1179 /* 1180 * This implements a LWKT yield, allowing a kernel thread to yield to other 1181 * kernel threads at the same or higher priority. This function can be 1182 * called in a tight loop and will typically only yield once per tick. 1183 * 1184 * Most kernel threads run at the same priority in order to allow equal 1185 * sharing. 1186 * 1187 * (self contained on a per cpu basis) 1188 */ 1189 void 1190 lwkt_yield(void) 1191 { 1192 globaldata_t gd = mycpu; 1193 thread_t td = gd->gd_curthread; 1194 1195 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1196 splz(); 1197 if (lwkt_resched_wanted()) { 1198 lwkt_schedule_self(curthread); 1199 lwkt_switch(); 1200 } 1201 } 1202 1203 /* 1204 * This yield is designed for kernel threads with a user context. 1205 * 1206 * The kernel acting on behalf of the user is potentially cpu-bound, 1207 * this function will efficiently allow other threads to run and also 1208 * switch to other processes by releasing. 1209 * 1210 * The lwkt_user_yield() function is designed to have very low overhead 1211 * if no yield is determined to be needed. 1212 */ 1213 void 1214 lwkt_user_yield(void) 1215 { 1216 globaldata_t gd = mycpu; 1217 thread_t td = gd->gd_curthread; 1218 1219 /* 1220 * Always run any pending interrupts in case we are in a critical 1221 * section. 1222 */ 1223 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2) 1224 splz(); 1225 1226 /* 1227 * Switch (which forces a release) if another kernel thread needs 1228 * the cpu, if userland wants us to resched, or if our kernel 1229 * quantum has run out. 1230 */ 1231 if (lwkt_resched_wanted() || 1232 user_resched_wanted()) 1233 { 1234 lwkt_switch(); 1235 } 1236 1237 #if 0 1238 /* 1239 * Reacquire the current process if we are released. 1240 * 1241 * XXX not implemented atm. The kernel may be holding locks and such, 1242 * so we want the thread to continue to receive cpu. 1243 */ 1244 if (td->td_release == NULL && lp) { 1245 lp->lwp_proc->p_usched->acquire_curproc(lp); 1246 td->td_release = lwkt_passive_release; 1247 lwkt_setpri_self(TDPRI_USER_NORM); 1248 } 1249 #endif 1250 } 1251 1252 /* 1253 * Generic schedule. Possibly schedule threads belonging to other cpus and 1254 * deal with threads that might be blocked on a wait queue. 1255 * 1256 * We have a little helper inline function which does additional work after 1257 * the thread has been enqueued, including dealing with preemption and 1258 * setting need_lwkt_resched() (which prevents the kernel from returning 1259 * to userland until it has processed higher priority threads). 1260 * 1261 * It is possible for this routine to be called after a failed _enqueue 1262 * (due to the target thread migrating, sleeping, or otherwise blocked). 1263 * We have to check that the thread is actually on the run queue! 1264 */ 1265 static __inline 1266 void 1267 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount) 1268 { 1269 if (ntd->td_flags & TDF_RUNQ) { 1270 if (ntd->td_preemptable) { 1271 ntd->td_preemptable(ntd, ccount); /* YYY +token */ 1272 } 1273 } 1274 } 1275 1276 static __inline 1277 void 1278 _lwkt_schedule(thread_t td) 1279 { 1280 globaldata_t mygd = mycpu; 1281 1282 KASSERT(td != &td->td_gd->gd_idlethread, 1283 ("lwkt_schedule(): scheduling gd_idlethread is illegal!")); 1284 KKASSERT((td->td_flags & TDF_MIGRATING) == 0); 1285 crit_enter_gd(mygd); 1286 KKASSERT(td->td_lwp == NULL || 1287 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 1288 1289 if (td == mygd->gd_curthread) { 1290 _lwkt_enqueue(td); 1291 } else { 1292 /* 1293 * If we own the thread, there is no race (since we are in a 1294 * critical section). If we do not own the thread there might 1295 * be a race but the target cpu will deal with it. 1296 */ 1297 #ifdef SMP 1298 if (td->td_gd == mygd) { 1299 _lwkt_enqueue(td); 1300 _lwkt_schedule_post(mygd, td, 1); 1301 } else { 1302 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0); 1303 } 1304 #else 1305 _lwkt_enqueue(td); 1306 _lwkt_schedule_post(mygd, td, 1); 1307 #endif 1308 } 1309 crit_exit_gd(mygd); 1310 } 1311 1312 void 1313 lwkt_schedule(thread_t td) 1314 { 1315 _lwkt_schedule(td); 1316 } 1317 1318 void 1319 lwkt_schedule_noresched(thread_t td) /* XXX not impl */ 1320 { 1321 _lwkt_schedule(td); 1322 } 1323 1324 #ifdef SMP 1325 1326 /* 1327 * When scheduled remotely if frame != NULL the IPIQ is being 1328 * run via doreti or an interrupt then preemption can be allowed. 1329 * 1330 * To allow preemption we have to drop the critical section so only 1331 * one is present in _lwkt_schedule_post. 1332 */ 1333 static void 1334 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame) 1335 { 1336 thread_t td = curthread; 1337 thread_t ntd = arg; 1338 1339 if (frame && ntd->td_preemptable) { 1340 crit_exit_noyield(td); 1341 _lwkt_schedule(ntd); 1342 crit_enter_quick(td); 1343 } else { 1344 _lwkt_schedule(ntd); 1345 } 1346 } 1347 1348 /* 1349 * Thread migration using a 'Pull' method. The thread may or may not be 1350 * the current thread. It MUST be descheduled and in a stable state. 1351 * lwkt_giveaway() must be called on the cpu owning the thread. 1352 * 1353 * At any point after lwkt_giveaway() is called, the target cpu may 1354 * 'pull' the thread by calling lwkt_acquire(). 1355 * 1356 * We have to make sure the thread is not sitting on a per-cpu tsleep 1357 * queue or it will blow up when it moves to another cpu. 1358 * 1359 * MPSAFE - must be called under very specific conditions. 1360 */ 1361 void 1362 lwkt_giveaway(thread_t td) 1363 { 1364 globaldata_t gd = mycpu; 1365 1366 crit_enter_gd(gd); 1367 if (td->td_flags & TDF_TSLEEPQ) 1368 tsleep_remove(td); 1369 KKASSERT(td->td_gd == gd); 1370 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); 1371 td->td_flags |= TDF_MIGRATING; 1372 crit_exit_gd(gd); 1373 } 1374 1375 void 1376 lwkt_acquire(thread_t td) 1377 { 1378 globaldata_t gd; 1379 globaldata_t mygd; 1380 int retry = 10000000; 1381 1382 KKASSERT(td->td_flags & TDF_MIGRATING); 1383 gd = td->td_gd; 1384 mygd = mycpu; 1385 if (gd != mycpu) { 1386 cpu_lfence(); 1387 KKASSERT((td->td_flags & TDF_RUNQ) == 0); 1388 crit_enter_gd(mygd); 1389 DEBUG_PUSH_INFO("lwkt_acquire"); 1390 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) { 1391 #ifdef SMP 1392 lwkt_process_ipiq(); 1393 #endif 1394 cpu_lfence(); 1395 if (--retry == 0) { 1396 kprintf("lwkt_acquire: stuck: td %p td->td_flags %08x\n", 1397 td, td->td_flags); 1398 retry = 10000000; 1399 } 1400 } 1401 DEBUG_POP_INFO(); 1402 cpu_mfence(); 1403 td->td_gd = mygd; 1404 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1405 td->td_flags &= ~TDF_MIGRATING; 1406 crit_exit_gd(mygd); 1407 } else { 1408 crit_enter_gd(mygd); 1409 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1410 td->td_flags &= ~TDF_MIGRATING; 1411 crit_exit_gd(mygd); 1412 } 1413 } 1414 1415 #endif 1416 1417 /* 1418 * Generic deschedule. Descheduling threads other then your own should be 1419 * done only in carefully controlled circumstances. Descheduling is 1420 * asynchronous. 1421 * 1422 * This function may block if the cpu has run out of messages. 1423 */ 1424 void 1425 lwkt_deschedule(thread_t td) 1426 { 1427 crit_enter(); 1428 #ifdef SMP 1429 if (td == curthread) { 1430 _lwkt_dequeue(td); 1431 } else { 1432 if (td->td_gd == mycpu) { 1433 _lwkt_dequeue(td); 1434 } else { 1435 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td); 1436 } 1437 } 1438 #else 1439 _lwkt_dequeue(td); 1440 #endif 1441 crit_exit(); 1442 } 1443 1444 /* 1445 * Set the target thread's priority. This routine does not automatically 1446 * switch to a higher priority thread, LWKT threads are not designed for 1447 * continuous priority changes. Yield if you want to switch. 1448 */ 1449 void 1450 lwkt_setpri(thread_t td, int pri) 1451 { 1452 if (td->td_pri != pri) { 1453 KKASSERT(pri >= 0); 1454 crit_enter(); 1455 if (td->td_flags & TDF_RUNQ) { 1456 KKASSERT(td->td_gd == mycpu); 1457 _lwkt_dequeue(td); 1458 td->td_pri = pri; 1459 _lwkt_enqueue(td); 1460 } else { 1461 td->td_pri = pri; 1462 } 1463 crit_exit(); 1464 } 1465 } 1466 1467 /* 1468 * Set the initial priority for a thread prior to it being scheduled for 1469 * the first time. The thread MUST NOT be scheduled before or during 1470 * this call. The thread may be assigned to a cpu other then the current 1471 * cpu. 1472 * 1473 * Typically used after a thread has been created with TDF_STOPPREQ, 1474 * and before the thread is initially scheduled. 1475 */ 1476 void 1477 lwkt_setpri_initial(thread_t td, int pri) 1478 { 1479 KKASSERT(pri >= 0); 1480 KKASSERT((td->td_flags & TDF_RUNQ) == 0); 1481 td->td_pri = pri; 1482 } 1483 1484 void 1485 lwkt_setpri_self(int pri) 1486 { 1487 thread_t td = curthread; 1488 1489 KKASSERT(pri >= 0 && pri <= TDPRI_MAX); 1490 crit_enter(); 1491 if (td->td_flags & TDF_RUNQ) { 1492 _lwkt_dequeue(td); 1493 td->td_pri = pri; 1494 _lwkt_enqueue(td); 1495 } else { 1496 td->td_pri = pri; 1497 } 1498 crit_exit(); 1499 } 1500 1501 /* 1502 * hz tick scheduler clock for LWKT threads 1503 */ 1504 void 1505 lwkt_schedulerclock(thread_t td) 1506 { 1507 globaldata_t gd = td->td_gd; 1508 thread_t xtd; 1509 1510 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) { 1511 /* 1512 * If the current thread is at the head of the runq shift it to the 1513 * end of any equal-priority threads and request a LWKT reschedule 1514 * if it moved. 1515 */ 1516 xtd = TAILQ_NEXT(td, td_threadq); 1517 if (xtd && xtd->td_pri == td->td_pri) { 1518 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq); 1519 while (xtd && xtd->td_pri == td->td_pri) 1520 xtd = TAILQ_NEXT(xtd, td_threadq); 1521 if (xtd) 1522 TAILQ_INSERT_BEFORE(xtd, td, td_threadq); 1523 else 1524 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq); 1525 need_lwkt_resched(); 1526 } 1527 } else { 1528 /* 1529 * If we scheduled a thread other than the one at the head of the 1530 * queue always request a reschedule every tick. 1531 */ 1532 need_lwkt_resched(); 1533 } 1534 } 1535 1536 /* 1537 * Migrate the current thread to the specified cpu. 1538 * 1539 * This is accomplished by descheduling ourselves from the current cpu 1540 * and setting td_migrate_gd. The lwkt_switch() code will detect that the 1541 * 'old' thread wants to migrate after it has been completely switched out 1542 * and will complete the migration. 1543 * 1544 * TDF_MIGRATING prevents scheduling races while the thread is being migrated. 1545 * 1546 * We must be sure to release our current process designation (if a user 1547 * process) before clearing out any tsleepq we are on because the release 1548 * code may re-add us. 1549 * 1550 * We must be sure to remove ourselves from the current cpu's tsleepq 1551 * before potentially moving to another queue. The thread can be on 1552 * a tsleepq due to a left-over tsleep_interlock(). 1553 */ 1554 1555 void 1556 lwkt_setcpu_self(globaldata_t rgd) 1557 { 1558 #ifdef SMP 1559 thread_t td = curthread; 1560 1561 if (td->td_gd != rgd) { 1562 crit_enter_quick(td); 1563 1564 if (td->td_release) 1565 td->td_release(td); 1566 if (td->td_flags & TDF_TSLEEPQ) 1567 tsleep_remove(td); 1568 1569 /* 1570 * Set TDF_MIGRATING to prevent a spurious reschedule while we are 1571 * trying to deschedule ourselves and switch away, then deschedule 1572 * ourself, remove us from tdallq, and set td_migrate_gd. Finally, 1573 * call lwkt_switch() to complete the operation. 1574 */ 1575 td->td_flags |= TDF_MIGRATING; 1576 lwkt_deschedule_self(td); 1577 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1578 td->td_migrate_gd = rgd; 1579 lwkt_switch(); 1580 1581 /* 1582 * We are now on the target cpu 1583 */ 1584 KKASSERT(rgd == mycpu); 1585 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); 1586 crit_exit_quick(td); 1587 } 1588 #endif 1589 } 1590 1591 void 1592 lwkt_migratecpu(int cpuid) 1593 { 1594 #ifdef SMP 1595 globaldata_t rgd; 1596 1597 rgd = globaldata_find(cpuid); 1598 lwkt_setcpu_self(rgd); 1599 #endif 1600 } 1601 1602 #ifdef SMP 1603 /* 1604 * Remote IPI for cpu migration (called while in a critical section so we 1605 * do not have to enter another one). 1606 * 1607 * The thread (td) has already been completely descheduled from the 1608 * originating cpu and we can simply assert the case. The thread is 1609 * assigned to the new cpu and enqueued. 1610 * 1611 * The thread will re-add itself to tdallq when it resumes execution. 1612 */ 1613 static void 1614 lwkt_setcpu_remote(void *arg) 1615 { 1616 thread_t td = arg; 1617 globaldata_t gd = mycpu; 1618 1619 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0); 1620 td->td_gd = gd; 1621 cpu_mfence(); 1622 td->td_flags &= ~TDF_MIGRATING; 1623 KKASSERT(td->td_migrate_gd == NULL); 1624 KKASSERT(td->td_lwp == NULL || 1625 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0); 1626 _lwkt_enqueue(td); 1627 } 1628 #endif 1629 1630 struct lwp * 1631 lwkt_preempted_proc(void) 1632 { 1633 thread_t td = curthread; 1634 while (td->td_preempted) 1635 td = td->td_preempted; 1636 return(td->td_lwp); 1637 } 1638 1639 /* 1640 * Create a kernel process/thread/whatever. It shares it's address space 1641 * with proc0 - ie: kernel only. 1642 * 1643 * If the cpu is not specified one will be selected. In the future 1644 * specifying a cpu of -1 will enable kernel thread migration between 1645 * cpus. 1646 */ 1647 int 1648 lwkt_create(void (*func)(void *), void *arg, struct thread **tdp, 1649 thread_t template, int tdflags, int cpu, const char *fmt, ...) 1650 { 1651 thread_t td; 1652 __va_list ap; 1653 1654 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu, 1655 tdflags); 1656 if (tdp) 1657 *tdp = td; 1658 cpu_set_thread_handler(td, lwkt_exit, func, arg); 1659 1660 /* 1661 * Set up arg0 for 'ps' etc 1662 */ 1663 __va_start(ap, fmt); 1664 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap); 1665 __va_end(ap); 1666 1667 /* 1668 * Schedule the thread to run 1669 */ 1670 if (td->td_flags & TDF_NOSTART) 1671 td->td_flags &= ~TDF_NOSTART; 1672 else 1673 lwkt_schedule(td); 1674 return 0; 1675 } 1676 1677 /* 1678 * Destroy an LWKT thread. Warning! This function is not called when 1679 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and 1680 * uses a different reaping mechanism. 1681 */ 1682 void 1683 lwkt_exit(void) 1684 { 1685 thread_t td = curthread; 1686 thread_t std; 1687 globaldata_t gd; 1688 1689 /* 1690 * Do any cleanup that might block here 1691 */ 1692 if (td->td_flags & TDF_VERBOSE) 1693 kprintf("kthread %p %s has exited\n", td, td->td_comm); 1694 caps_exit(td); 1695 biosched_done(td); 1696 dsched_exit_thread(td); 1697 1698 /* 1699 * Get us into a critical section to interlock gd_freetd and loop 1700 * until we can get it freed. 1701 * 1702 * We have to cache the current td in gd_freetd because objcache_put()ing 1703 * it would rip it out from under us while our thread is still active. 1704 * 1705 * We are the current thread so of course our own TDF_RUNNING bit will 1706 * be set, so unlike the lwp reap code we don't wait for it to clear. 1707 */ 1708 gd = mycpu; 1709 crit_enter_quick(td); 1710 for (;;) { 1711 if (td->td_refs) { 1712 tsleep(td, 0, "tdreap", 1); 1713 continue; 1714 } 1715 if ((std = gd->gd_freetd) != NULL) { 1716 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0); 1717 gd->gd_freetd = NULL; 1718 objcache_put(thread_cache, std); 1719 continue; 1720 } 1721 break; 1722 } 1723 1724 /* 1725 * Remove thread resources from kernel lists and deschedule us for 1726 * the last time. We cannot block after this point or we may end 1727 * up with a stale td on the tsleepq. 1728 * 1729 * None of this may block, the critical section is the only thing 1730 * protecting tdallq and the only thing preventing new lwkt_hold() 1731 * thread refs now. 1732 */ 1733 if (td->td_flags & TDF_TSLEEPQ) 1734 tsleep_remove(td); 1735 lwkt_deschedule_self(td); 1736 lwkt_remove_tdallq(td); 1737 KKASSERT(td->td_refs == 0); 1738 1739 /* 1740 * Final cleanup 1741 */ 1742 KKASSERT(gd->gd_freetd == NULL); 1743 if (td->td_flags & TDF_ALLOCATED_THREAD) 1744 gd->gd_freetd = td; 1745 cpu_thread_exit(); 1746 } 1747 1748 void 1749 lwkt_remove_tdallq(thread_t td) 1750 { 1751 KKASSERT(td->td_gd == mycpu); 1752 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1753 } 1754 1755 /* 1756 * Code reduction and branch prediction improvements. Call/return 1757 * overhead on modern cpus often degenerates into 0 cycles due to 1758 * the cpu's branch prediction hardware and return pc cache. We 1759 * can take advantage of this by not inlining medium-complexity 1760 * functions and we can also reduce the branch prediction impact 1761 * by collapsing perfectly predictable branches into a single 1762 * procedure instead of duplicating it. 1763 * 1764 * Is any of this noticeable? Probably not, so I'll take the 1765 * smaller code size. 1766 */ 1767 void 1768 crit_exit_wrapper(__DEBUG_CRIT_ARG__) 1769 { 1770 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__); 1771 } 1772 1773 void 1774 crit_panic(void) 1775 { 1776 thread_t td = curthread; 1777 int lcrit = td->td_critcount; 1778 1779 td->td_critcount = 0; 1780 panic("td_critcount is/would-go negative! %p %d", td, lcrit); 1781 /* NOT REACHED */ 1782 } 1783 1784 #ifdef SMP 1785 1786 /* 1787 * Called from debugger/panic on cpus which have been stopped. We must still 1788 * process the IPIQ while stopped, even if we were stopped while in a critical 1789 * section (XXX). 1790 * 1791 * If we are dumping also try to process any pending interrupts. This may 1792 * or may not work depending on the state of the cpu at the point it was 1793 * stopped. 1794 */ 1795 void 1796 lwkt_smp_stopped(void) 1797 { 1798 globaldata_t gd = mycpu; 1799 1800 crit_enter_gd(gd); 1801 if (dumping) { 1802 lwkt_process_ipiq(); 1803 splz(); 1804 } else { 1805 lwkt_process_ipiq(); 1806 } 1807 crit_exit_gd(gd); 1808 } 1809 1810 #endif 1811