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