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