1 /* 2 * Copyright (c) 2003,2004 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@backplane.com> 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * 3. Neither the name of The DragonFly Project nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific, prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 */ 34 35 /* 36 * Each cpu in a system has its own self-contained light weight kernel 37 * thread scheduler, which means that generally speaking we only need 38 * to use a critical section to avoid problems. Foreign thread 39 * scheduling is queued via (async) IPIs. 40 */ 41 42 #include <sys/param.h> 43 #include <sys/systm.h> 44 #include <sys/kernel.h> 45 #include <sys/proc.h> 46 #include <sys/rtprio.h> 47 #include <sys/queue.h> 48 #include <sys/sysctl.h> 49 #include <sys/kthread.h> 50 #include <machine/cpu.h> 51 #include <sys/lock.h> 52 #include <sys/caps.h> 53 #include <sys/spinlock.h> 54 #include <sys/ktr.h> 55 56 #include <sys/thread2.h> 57 #include <sys/spinlock2.h> 58 59 #include <vm/vm.h> 60 #include <vm/vm_param.h> 61 #include <vm/vm_kern.h> 62 #include <vm/vm_object.h> 63 #include <vm/vm_page.h> 64 #include <vm/vm_map.h> 65 #include <vm/vm_pager.h> 66 #include <vm/vm_extern.h> 67 68 #include <machine/stdarg.h> 69 #include <machine/smp.h> 70 71 72 static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads"); 73 74 #ifdef SMP 75 static int mplock_countx = 0; 76 #endif 77 #ifdef INVARIANTS 78 static int panic_on_cscount = 0; 79 #endif 80 static __int64_t switch_count = 0; 81 static __int64_t preempt_hit = 0; 82 static __int64_t preempt_miss = 0; 83 static __int64_t preempt_weird = 0; 84 static __int64_t token_contention_count = 0; 85 static __int64_t mplock_contention_count = 0; 86 static int lwkt_use_spin_port; 87 #ifdef SMP 88 static int chain_mplock = 0; 89 #endif 90 static struct objcache *thread_cache; 91 92 volatile cpumask_t mp_lock_contention_mask; 93 94 extern void cpu_heavy_restore(void); 95 extern void cpu_lwkt_restore(void); 96 extern void cpu_kthread_restore(void); 97 extern void cpu_idle_restore(void); 98 99 #ifdef __amd64__ 100 101 static int 102 jg_tos_ok(struct thread *td) 103 { 104 void *tos; 105 int tos_ok; 106 107 if (td == NULL) { 108 return 1; 109 } 110 KKASSERT(td->td_sp != NULL); 111 tos = ((void **)td->td_sp)[0]; 112 tos_ok = 0; 113 if ((tos == cpu_heavy_restore) || (tos == cpu_lwkt_restore) || 114 (tos == cpu_kthread_restore) || (tos == cpu_idle_restore)) { 115 tos_ok = 1; 116 } 117 return tos_ok; 118 } 119 120 #endif 121 122 /* 123 * We can make all thread ports use the spin backend instead of the thread 124 * backend. This should only be set to debug the spin backend. 125 */ 126 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port); 127 128 #ifdef INVARIANTS 129 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, ""); 130 #endif 131 #ifdef SMP 132 SYSCTL_INT(_lwkt, OID_AUTO, chain_mplock, CTLFLAG_RW, &chain_mplock, 0, ""); 133 #endif 134 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, ""); 135 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, ""); 136 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, ""); 137 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, ""); 138 #ifdef INVARIANTS 139 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW, 140 &token_contention_count, 0, "spinning due to token contention"); 141 SYSCTL_QUAD(_lwkt, OID_AUTO, mplock_contention_count, CTLFLAG_RW, 142 &mplock_contention_count, 0, "spinning due to MPLOCK contention"); 143 #endif 144 145 /* 146 * Kernel Trace 147 */ 148 #if !defined(KTR_GIANT_CONTENTION) 149 #define KTR_GIANT_CONTENTION KTR_ALL 150 #endif 151 152 KTR_INFO_MASTER(giant); 153 KTR_INFO(KTR_GIANT_CONTENTION, giant, beg, 0, "thread=%p", sizeof(void *)); 154 KTR_INFO(KTR_GIANT_CONTENTION, giant, end, 1, "thread=%p", sizeof(void *)); 155 156 #define loggiant(name) KTR_LOG(giant_ ## name, curthread) 157 158 /* 159 * These helper procedures handle the runq, they can only be called from 160 * within a critical section. 161 * 162 * WARNING! Prior to SMP being brought up it is possible to enqueue and 163 * dequeue threads belonging to other cpus, so be sure to use td->td_gd 164 * instead of 'mycpu' when referencing the globaldata structure. Once 165 * SMP live enqueuing and dequeueing only occurs on the current cpu. 166 */ 167 static __inline 168 void 169 _lwkt_dequeue(thread_t td) 170 { 171 if (td->td_flags & TDF_RUNQ) { 172 int nq = td->td_pri & TDPRI_MASK; 173 struct globaldata *gd = td->td_gd; 174 175 td->td_flags &= ~TDF_RUNQ; 176 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq); 177 /* runqmask is passively cleaned up by the switcher */ 178 } 179 } 180 181 static __inline 182 void 183 _lwkt_enqueue(thread_t td) 184 { 185 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) { 186 int nq = td->td_pri & TDPRI_MASK; 187 struct globaldata *gd = td->td_gd; 188 189 td->td_flags |= TDF_RUNQ; 190 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq); 191 gd->gd_runqmask |= 1 << nq; 192 } 193 } 194 195 static __boolean_t 196 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags) 197 { 198 struct thread *td = (struct thread *)obj; 199 200 td->td_kstack = NULL; 201 td->td_kstack_size = 0; 202 td->td_flags = TDF_ALLOCATED_THREAD; 203 return (1); 204 } 205 206 static void 207 _lwkt_thread_dtor(void *obj, void *privdata) 208 { 209 struct thread *td = (struct thread *)obj; 210 211 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD, 212 ("_lwkt_thread_dtor: not allocated from objcache")); 213 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack && 214 td->td_kstack_size > 0, 215 ("_lwkt_thread_dtor: corrupted stack")); 216 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size); 217 } 218 219 /* 220 * Initialize the lwkt s/system. 221 */ 222 void 223 lwkt_init(void) 224 { 225 /* An objcache has 2 magazines per CPU so divide cache size by 2. */ 226 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread), 227 NULL, CACHE_NTHREADS/2, 228 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL); 229 } 230 231 /* 232 * Schedule a thread to run. As the current thread we can always safely 233 * schedule ourselves, and a shortcut procedure is provided for that 234 * function. 235 * 236 * (non-blocking, self contained on a per cpu basis) 237 */ 238 void 239 lwkt_schedule_self(thread_t td) 240 { 241 crit_enter_quick(td); 242 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!")); 243 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0); 244 _lwkt_enqueue(td); 245 crit_exit_quick(td); 246 } 247 248 /* 249 * Deschedule a thread. 250 * 251 * (non-blocking, self contained on a per cpu basis) 252 */ 253 void 254 lwkt_deschedule_self(thread_t td) 255 { 256 crit_enter_quick(td); 257 _lwkt_dequeue(td); 258 crit_exit_quick(td); 259 } 260 261 /* 262 * LWKTs operate on a per-cpu basis 263 * 264 * WARNING! Called from early boot, 'mycpu' may not work yet. 265 */ 266 void 267 lwkt_gdinit(struct globaldata *gd) 268 { 269 int i; 270 271 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i) 272 TAILQ_INIT(&gd->gd_tdrunq[i]); 273 gd->gd_runqmask = 0; 274 TAILQ_INIT(&gd->gd_tdallq); 275 } 276 277 /* 278 * Create a new thread. The thread must be associated with a process context 279 * or LWKT start address before it can be scheduled. If the target cpu is 280 * -1 the thread will be created on the current cpu. 281 * 282 * If you intend to create a thread without a process context this function 283 * does everything except load the startup and switcher function. 284 */ 285 thread_t 286 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags) 287 { 288 globaldata_t gd = mycpu; 289 void *stack; 290 291 /* 292 * If static thread storage is not supplied allocate a thread. Reuse 293 * a cached free thread if possible. gd_freetd is used to keep an exiting 294 * thread intact through the exit. 295 */ 296 if (td == NULL) { 297 if ((td = gd->gd_freetd) != NULL) 298 gd->gd_freetd = NULL; 299 else 300 td = objcache_get(thread_cache, M_WAITOK); 301 KASSERT((td->td_flags & 302 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD, 303 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags)); 304 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK); 305 } 306 307 /* 308 * Try to reuse cached stack. 309 */ 310 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) { 311 if (flags & TDF_ALLOCATED_STACK) { 312 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size); 313 stack = NULL; 314 } 315 } 316 if (stack == NULL) { 317 stack = (void *)kmem_alloc(&kernel_map, stksize); 318 flags |= TDF_ALLOCATED_STACK; 319 } 320 if (cpu < 0) 321 lwkt_init_thread(td, stack, stksize, flags, gd); 322 else 323 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu)); 324 return(td); 325 } 326 327 /* 328 * Initialize a preexisting thread structure. This function is used by 329 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread. 330 * 331 * All threads start out in a critical section at a priority of 332 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as 333 * appropriate. This function may send an IPI message when the 334 * requested cpu is not the current cpu and consequently gd_tdallq may 335 * not be initialized synchronously from the point of view of the originating 336 * cpu. 337 * 338 * NOTE! we have to be careful in regards to creating threads for other cpus 339 * if SMP has not yet been activated. 340 */ 341 #ifdef SMP 342 343 static void 344 lwkt_init_thread_remote(void *arg) 345 { 346 thread_t td = arg; 347 348 /* 349 * Protected by critical section held by IPI dispatch 350 */ 351 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq); 352 } 353 354 #endif 355 356 void 357 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags, 358 struct globaldata *gd) 359 { 360 globaldata_t mygd = mycpu; 361 362 bzero(td, sizeof(struct thread)); 363 td->td_kstack = stack; 364 td->td_kstack_size = stksize; 365 td->td_flags = flags; 366 td->td_gd = gd; 367 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT; 368 #ifdef SMP 369 if ((flags & TDF_MPSAFE) == 0) 370 td->td_mpcount = 1; 371 #endif 372 if (lwkt_use_spin_port) 373 lwkt_initport_spin(&td->td_msgport); 374 else 375 lwkt_initport_thread(&td->td_msgport, td); 376 pmap_init_thread(td); 377 #ifdef SMP 378 /* 379 * Normally initializing a thread for a remote cpu requires sending an 380 * IPI. However, the idlethread is setup before the other cpus are 381 * activated so we have to treat it as a special case. XXX manipulation 382 * of gd_tdallq requires the BGL. 383 */ 384 if (gd == mygd || td == &gd->gd_idlethread) { 385 crit_enter_gd(mygd); 386 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); 387 crit_exit_gd(mygd); 388 } else { 389 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td); 390 } 391 #else 392 crit_enter_gd(mygd); 393 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); 394 crit_exit_gd(mygd); 395 #endif 396 } 397 398 void 399 lwkt_set_comm(thread_t td, const char *ctl, ...) 400 { 401 __va_list va; 402 403 __va_start(va, ctl); 404 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va); 405 __va_end(va); 406 } 407 408 void 409 lwkt_hold(thread_t td) 410 { 411 ++td->td_refs; 412 } 413 414 void 415 lwkt_rele(thread_t td) 416 { 417 KKASSERT(td->td_refs > 0); 418 --td->td_refs; 419 } 420 421 void 422 lwkt_wait_free(thread_t td) 423 { 424 while (td->td_refs) 425 tsleep(td, 0, "tdreap", hz); 426 } 427 428 void 429 lwkt_free_thread(thread_t td) 430 { 431 KASSERT((td->td_flags & TDF_RUNNING) == 0, 432 ("lwkt_free_thread: did not exit! %p", td)); 433 434 if (td->td_flags & TDF_ALLOCATED_THREAD) { 435 objcache_put(thread_cache, td); 436 } else if (td->td_flags & TDF_ALLOCATED_STACK) { 437 /* client-allocated struct with internally allocated stack */ 438 KASSERT(td->td_kstack && td->td_kstack_size > 0, 439 ("lwkt_free_thread: corrupted stack")); 440 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size); 441 td->td_kstack = NULL; 442 td->td_kstack_size = 0; 443 } 444 } 445 446 447 /* 448 * Switch to the next runnable lwkt. If no LWKTs are runnable then 449 * switch to the idlethread. Switching must occur within a critical 450 * section to avoid races with the scheduling queue. 451 * 452 * We always have full control over our cpu's run queue. Other cpus 453 * that wish to manipulate our queue must use the cpu_*msg() calls to 454 * talk to our cpu, so a critical section is all that is needed and 455 * the result is very, very fast thread switching. 456 * 457 * The LWKT scheduler uses a fixed priority model and round-robins at 458 * each priority level. User process scheduling is a totally 459 * different beast and LWKT priorities should not be confused with 460 * user process priorities. 461 * 462 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch() 463 * cleans it up. Note that the td_switch() function cannot do anything that 464 * requires the MP lock since the MP lock will have already been setup for 465 * the target thread (not the current thread). It's nice to have a scheduler 466 * that does not need the MP lock to work because it allows us to do some 467 * really cool high-performance MP lock optimizations. 468 * 469 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch() 470 * is not called by the current thread in the preemption case, only when 471 * the preempting thread blocks (in order to return to the original thread). 472 */ 473 void 474 lwkt_switch(void) 475 { 476 globaldata_t gd = mycpu; 477 thread_t td = gd->gd_curthread; 478 thread_t ntd; 479 #ifdef SMP 480 int mpheld; 481 #endif 482 483 /* 484 * Switching from within a 'fast' (non thread switched) interrupt or IPI 485 * is illegal. However, we may have to do it anyway if we hit a fatal 486 * kernel trap or we have paniced. 487 * 488 * If this case occurs save and restore the interrupt nesting level. 489 */ 490 if (gd->gd_intr_nesting_level) { 491 int savegdnest; 492 int savegdtrap; 493 494 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) { 495 panic("lwkt_switch: cannot switch from within " 496 "a fast interrupt, yet, td %p\n", td); 497 } else { 498 savegdnest = gd->gd_intr_nesting_level; 499 savegdtrap = gd->gd_trap_nesting_level; 500 gd->gd_intr_nesting_level = 0; 501 gd->gd_trap_nesting_level = 0; 502 if ((td->td_flags & TDF_PANICWARN) == 0) { 503 td->td_flags |= TDF_PANICWARN; 504 kprintf("Warning: thread switch from interrupt or IPI, " 505 "thread %p (%s)\n", td, td->td_comm); 506 print_backtrace(); 507 } 508 lwkt_switch(); 509 gd->gd_intr_nesting_level = savegdnest; 510 gd->gd_trap_nesting_level = savegdtrap; 511 return; 512 } 513 } 514 515 /* 516 * Passive release (used to transition from user to kernel mode 517 * when we block or switch rather then when we enter the kernel). 518 * This function is NOT called if we are switching into a preemption 519 * or returning from a preemption. Typically this causes us to lose 520 * our current process designation (if we have one) and become a true 521 * LWKT thread, and may also hand the current process designation to 522 * another process and schedule thread. 523 */ 524 if (td->td_release) 525 td->td_release(td); 526 527 crit_enter_gd(gd); 528 if (td->td_toks) 529 lwkt_relalltokens(td); 530 531 /* 532 * We had better not be holding any spin locks, but don't get into an 533 * endless panic loop. 534 */ 535 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL, 536 ("lwkt_switch: still holding a shared spinlock %p!", 537 gd->gd_spinlock_rd)); 538 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL, 539 ("lwkt_switch: still holding %d exclusive spinlocks!", 540 gd->gd_spinlocks_wr)); 541 542 543 #ifdef SMP 544 /* 545 * td_mpcount cannot be used to determine if we currently hold the 546 * MP lock because get_mplock() will increment it prior to attempting 547 * to get the lock, and switch out if it can't. Our ownership of 548 * the actual lock will remain stable while we are in a critical section 549 * (but, of course, another cpu may own or release the lock so the 550 * actual value of mp_lock is not stable). 551 */ 552 mpheld = MP_LOCK_HELD(); 553 #ifdef INVARIANTS 554 if (td->td_cscount) { 555 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n", 556 td); 557 if (panic_on_cscount) 558 panic("switching while mastering cpusync"); 559 } 560 #endif 561 #endif 562 if ((ntd = td->td_preempted) != NULL) { 563 /* 564 * We had preempted another thread on this cpu, resume the preempted 565 * thread. This occurs transparently, whether the preempted thread 566 * was scheduled or not (it may have been preempted after descheduling 567 * itself). 568 * 569 * We have to setup the MP lock for the original thread after backing 570 * out the adjustment that was made to curthread when the original 571 * was preempted. 572 */ 573 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK); 574 #ifdef SMP 575 if (ntd->td_mpcount && mpheld == 0) { 576 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d", 577 td, ntd, td->td_mpcount, ntd->td_mpcount); 578 } 579 if (ntd->td_mpcount) { 580 td->td_mpcount -= ntd->td_mpcount; 581 KKASSERT(td->td_mpcount >= 0); 582 } 583 #endif 584 ntd->td_flags |= TDF_PREEMPT_DONE; 585 586 /* 587 * The interrupt may have woken a thread up, we need to properly 588 * set the reschedule flag if the originally interrupted thread is 589 * at a lower priority. 590 */ 591 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1) 592 need_lwkt_resched(); 593 /* YYY release mp lock on switchback if original doesn't need it */ 594 } else { 595 /* 596 * Priority queue / round-robin at each priority. Note that user 597 * processes run at a fixed, low priority and the user process 598 * scheduler deals with interactions between user processes 599 * by scheduling and descheduling them from the LWKT queue as 600 * necessary. 601 * 602 * We have to adjust the MP lock for the target thread. If we 603 * need the MP lock and cannot obtain it we try to locate a 604 * thread that does not need the MP lock. If we cannot, we spin 605 * instead of HLT. 606 * 607 * A similar issue exists for the tokens held by the target thread. 608 * If we cannot obtain ownership of the tokens we cannot immediately 609 * schedule the thread. 610 */ 611 612 /* 613 * If an LWKT reschedule was requested, well that is what we are 614 * doing now so clear it. 615 */ 616 clear_lwkt_resched(); 617 again: 618 if (gd->gd_runqmask) { 619 int nq = bsrl(gd->gd_runqmask); 620 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) { 621 gd->gd_runqmask &= ~(1 << nq); 622 goto again; 623 } 624 #ifdef SMP 625 /* 626 * THREAD SELECTION FOR AN SMP MACHINE BUILD 627 * 628 * If the target needs the MP lock and we couldn't get it, 629 * or if the target is holding tokens and we could not 630 * gain ownership of the tokens, continue looking for a 631 * thread to schedule and spin instead of HLT if we can't. 632 * 633 * NOTE: the mpheld variable invalid after this conditional, it 634 * can change due to both cpu_try_mplock() returning success 635 * AND interactions in lwkt_getalltokens() due to the fact that 636 * we are trying to check the mpcount of a thread other then 637 * the current thread. Because of this, if the current thread 638 * is not holding td_mpcount, an IPI indirectly run via 639 * lwkt_getalltokens() can obtain and release the MP lock and 640 * cause the core MP lock to be released. 641 */ 642 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) || 643 (ntd->td_toks && lwkt_getalltokens(ntd) == 0) 644 ) { 645 u_int32_t rqmask = gd->gd_runqmask; 646 647 mpheld = MP_LOCK_HELD(); 648 ntd = NULL; 649 while (rqmask) { 650 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) { 651 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) { 652 /* spinning due to MP lock being held */ 653 #ifdef INVARIANTS 654 ++mplock_contention_count; 655 #endif 656 /* mplock still not held, 'mpheld' still valid */ 657 continue; 658 } 659 660 /* 661 * mpheld state invalid after getalltokens call returns 662 * failure, but the variable is only needed for 663 * the loop. 664 */ 665 if (ntd->td_toks && !lwkt_getalltokens(ntd)) { 666 /* spinning due to token contention */ 667 #ifdef INVARIANTS 668 ++token_contention_count; 669 #endif 670 mpheld = MP_LOCK_HELD(); 671 continue; 672 } 673 break; 674 } 675 if (ntd) 676 break; 677 rqmask &= ~(1 << nq); 678 nq = bsrl(rqmask); 679 680 /* 681 * We have two choices. We can either refuse to run a 682 * user thread when a kernel thread needs the MP lock 683 * but could not get it, or we can allow it to run but 684 * then expect an IPI (hopefully) later on to force a 685 * reschedule when the MP lock might become available. 686 */ 687 if (nq < TDPRI_KERN_LPSCHED) { 688 if (chain_mplock == 0) 689 break; 690 atomic_set_int(&mp_lock_contention_mask, 691 gd->gd_cpumask); 692 /* continue loop, allow user threads to be scheduled */ 693 } 694 } 695 if (ntd == NULL) { 696 cpu_mplock_contested(); 697 ntd = &gd->gd_idlethread; 698 ntd->td_flags |= TDF_IDLE_NOHLT; 699 goto using_idle_thread; 700 } else { 701 ++gd->gd_cnt.v_swtch; 702 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq); 703 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq); 704 } 705 } else { 706 if (ntd->td_mpcount) 707 ++mplock_countx; 708 ++gd->gd_cnt.v_swtch; 709 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq); 710 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq); 711 } 712 #else 713 /* 714 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to 715 * worry about tokens or the BGL. However, we still have 716 * to call lwkt_getalltokens() in order to properly detect 717 * stale tokens. This call cannot fail for a UP build! 718 */ 719 lwkt_getalltokens(ntd); 720 ++gd->gd_cnt.v_swtch; 721 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq); 722 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq); 723 #endif 724 } else { 725 /* 726 * We have nothing to run but only let the idle loop halt 727 * the cpu if there are no pending interrupts. 728 */ 729 ntd = &gd->gd_idlethread; 730 if (gd->gd_reqflags & RQF_IDLECHECK_MASK) 731 ntd->td_flags |= TDF_IDLE_NOHLT; 732 #ifdef SMP 733 using_idle_thread: 734 /* 735 * The idle thread should not be holding the MP lock unless we 736 * are trapping in the kernel or in a panic. Since we select the 737 * idle thread unconditionally when no other thread is available, 738 * if the MP lock is desired during a panic or kernel trap, we 739 * have to loop in the scheduler until we get it. 740 */ 741 if (ntd->td_mpcount) { 742 mpheld = MP_LOCK_HELD(); 743 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) { 744 panic("Idle thread %p was holding the BGL!", ntd); 745 } else if (mpheld == 0) { 746 cpu_mplock_contested(); 747 goto again; 748 } 749 } 750 #endif 751 } 752 } 753 KASSERT(ntd->td_pri >= TDPRI_CRIT, 754 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri)); 755 756 /* 757 * Do the actual switch. If the new target does not need the MP lock 758 * and we are holding it, release the MP lock. If the new target requires 759 * the MP lock we have already acquired it for the target. 760 */ 761 #ifdef SMP 762 if (ntd->td_mpcount == 0 ) { 763 if (MP_LOCK_HELD()) 764 cpu_rel_mplock(); 765 } else { 766 ASSERT_MP_LOCK_HELD(ntd); 767 } 768 #endif 769 if (td != ntd) { 770 ++switch_count; 771 #ifdef __amd64__ 772 KKASSERT(jg_tos_ok(ntd)); 773 #endif 774 td->td_switch(ntd); 775 } 776 /* NOTE: current cpu may have changed after switch */ 777 crit_exit_quick(td); 778 } 779 780 /* 781 * Request that the target thread preempt the current thread. Preemption 782 * only works under a specific set of conditions: 783 * 784 * - We are not preempting ourselves 785 * - The target thread is owned by the current cpu 786 * - We are not currently being preempted 787 * - The target is not currently being preempted 788 * - We are not holding any spin locks 789 * - The target thread is not holding any tokens 790 * - We are able to satisfy the target's MP lock requirements (if any). 791 * 792 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically 793 * this is called via lwkt_schedule() through the td_preemptable callback. 794 * critpri is the managed critical priority that we should ignore in order 795 * to determine whether preemption is possible (aka usually just the crit 796 * priority of lwkt_schedule() itself). 797 * 798 * XXX at the moment we run the target thread in a critical section during 799 * the preemption in order to prevent the target from taking interrupts 800 * that *WE* can't. Preemption is strictly limited to interrupt threads 801 * and interrupt-like threads, outside of a critical section, and the 802 * preempted source thread will be resumed the instant the target blocks 803 * whether or not the source is scheduled (i.e. preemption is supposed to 804 * be as transparent as possible). 805 * 806 * The target thread inherits our MP count (added to its own) for the 807 * duration of the preemption in order to preserve the atomicy of the 808 * MP lock during the preemption. Therefore, any preempting targets must be 809 * careful in regards to MP assertions. Note that the MP count may be 810 * out of sync with the physical mp_lock, but we do not have to preserve 811 * the original ownership of the lock if it was out of synch (that is, we 812 * can leave it synchronized on return). 813 */ 814 void 815 lwkt_preempt(thread_t ntd, int critpri) 816 { 817 struct globaldata *gd = mycpu; 818 thread_t td; 819 #ifdef SMP 820 int mpheld; 821 int savecnt; 822 #endif 823 824 /* 825 * The caller has put us in a critical section. We can only preempt 826 * if the caller of the caller was not in a critical section (basically 827 * a local interrupt), as determined by the 'critpri' parameter. We 828 * also can't preempt if the caller is holding any spinlocks (even if 829 * he isn't in a critical section). This also handles the tokens test. 830 * 831 * YYY The target thread must be in a critical section (else it must 832 * inherit our critical section? I dunno yet). 833 * 834 * Set need_lwkt_resched() unconditionally for now YYY. 835 */ 836 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri)); 837 838 td = gd->gd_curthread; 839 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) { 840 ++preempt_miss; 841 return; 842 } 843 if ((td->td_pri & ~TDPRI_MASK) > critpri) { 844 ++preempt_miss; 845 need_lwkt_resched(); 846 return; 847 } 848 #ifdef SMP 849 if (ntd->td_gd != gd) { 850 ++preempt_miss; 851 need_lwkt_resched(); 852 return; 853 } 854 #endif 855 /* 856 * Take the easy way out and do not preempt if we are holding 857 * any spinlocks. We could test whether the thread(s) being 858 * preempted interlock against the target thread's tokens and whether 859 * we can get all the target thread's tokens, but this situation 860 * should not occur very often so its easier to simply not preempt. 861 * Also, plain spinlocks are impossible to figure out at this point so 862 * just don't preempt. 863 * 864 * Do not try to preempt if the target thread is holding any tokens. 865 * We could try to acquire the tokens but this case is so rare there 866 * is no need to support it. 867 */ 868 if (gd->gd_spinlock_rd || gd->gd_spinlocks_wr) { 869 ++preempt_miss; 870 need_lwkt_resched(); 871 return; 872 } 873 if (ntd->td_toks) { 874 ++preempt_miss; 875 need_lwkt_resched(); 876 return; 877 } 878 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) { 879 ++preempt_weird; 880 need_lwkt_resched(); 881 return; 882 } 883 if (ntd->td_preempted) { 884 ++preempt_hit; 885 need_lwkt_resched(); 886 return; 887 } 888 #ifdef SMP 889 /* 890 * note: an interrupt might have occured just as we were transitioning 891 * to or from the MP lock. In this case td_mpcount will be pre-disposed 892 * (non-zero) but not actually synchronized with the actual state of the 893 * lock. We can use it to imply an MP lock requirement for the 894 * preemption but we cannot use it to test whether we hold the MP lock 895 * or not. 896 */ 897 savecnt = td->td_mpcount; 898 mpheld = MP_LOCK_HELD(); 899 ntd->td_mpcount += td->td_mpcount; 900 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) { 901 ntd->td_mpcount -= td->td_mpcount; 902 ++preempt_miss; 903 need_lwkt_resched(); 904 return; 905 } 906 #endif 907 908 /* 909 * Since we are able to preempt the current thread, there is no need to 910 * call need_lwkt_resched(). 911 */ 912 ++preempt_hit; 913 ntd->td_preempted = td; 914 td->td_flags |= TDF_PREEMPT_LOCK; 915 td->td_switch(ntd); 916 917 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE)); 918 #ifdef SMP 919 KKASSERT(savecnt == td->td_mpcount); 920 mpheld = MP_LOCK_HELD(); 921 if (mpheld && td->td_mpcount == 0) 922 cpu_rel_mplock(); 923 else if (mpheld == 0 && td->td_mpcount) 924 panic("lwkt_preempt(): MP lock was not held through"); 925 #endif 926 ntd->td_preempted = NULL; 927 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE); 928 } 929 930 /* 931 * Conditionally call splz() if gd_reqflags indicates work is pending. 932 * 933 * td_nest_count prevents deep nesting via splz() or doreti() which 934 * might otherwise blow out the kernel stack. Note that except for 935 * this special case, we MUST call splz() here to handle any 936 * pending ints, particularly after we switch, or we might accidently 937 * halt the cpu with interrupts pending. 938 * 939 * (self contained on a per cpu basis) 940 */ 941 void 942 splz_check(void) 943 { 944 globaldata_t gd = mycpu; 945 thread_t td = gd->gd_curthread; 946 947 if (gd->gd_reqflags && td->td_nest_count < 2) 948 splz(); 949 } 950 951 /* 952 * This implements a normal yield which will yield to equal priority 953 * threads as well as higher priority threads. Note that gd_reqflags 954 * tests will be handled by the crit_exit() call in lwkt_switch(). 955 * 956 * (self contained on a per cpu basis) 957 */ 958 void 959 lwkt_yield(void) 960 { 961 lwkt_schedule_self(curthread); 962 lwkt_switch(); 963 } 964 965 /* 966 * This function is used along with the lwkt_passive_recover() inline 967 * by the trap code to negotiate a passive release of the current 968 * process/lwp designation with the user scheduler. 969 */ 970 void 971 lwkt_passive_release(struct thread *td) 972 { 973 struct lwp *lp = td->td_lwp; 974 975 td->td_release = NULL; 976 lwkt_setpri_self(TDPRI_KERN_USER); 977 lp->lwp_proc->p_usched->release_curproc(lp); 978 } 979 980 /* 981 * Make a kernel thread act as if it were in user mode with regards 982 * to scheduling, to avoid becoming cpu-bound in the kernel. Kernel 983 * loops which may be potentially cpu-bound can call lwkt_user_yield(). 984 * 985 * The lwkt_user_yield() function is designed to have very low overhead 986 * if no yield is determined to be needed. 987 */ 988 void 989 lwkt_user_yield(void) 990 { 991 thread_t td = curthread; 992 struct lwp *lp = td->td_lwp; 993 994 #ifdef SMP 995 /* 996 * XXX SEVERE TEMPORARY HACK. A cpu-bound operation running in the 997 * kernel can prevent other cpus from servicing interrupt threads 998 * which still require the MP lock (which is a lot of them). This 999 * has a chaining effect since if the interrupt is blocked, so is 1000 * the event, so normal scheduling will not pick up on the problem. 1001 */ 1002 if (mplock_countx && td->td_mpcount) { 1003 int savecnt = td->td_mpcount; 1004 1005 td->td_mpcount = 1; 1006 rel_mplock(); 1007 DELAY(10); 1008 get_mplock(); 1009 td->td_mpcount = savecnt; 1010 mplock_countx = 0; 1011 } 1012 #endif 1013 1014 /* 1015 * Another kernel thread wants the cpu 1016 */ 1017 if (lwkt_resched_wanted()) 1018 lwkt_switch(); 1019 1020 /* 1021 * If the user scheduler has asynchronously determined that the current 1022 * process (when running in user mode) needs to lose the cpu then make 1023 * sure we are released. 1024 */ 1025 if (user_resched_wanted()) { 1026 if (td->td_release) 1027 td->td_release(td); 1028 } 1029 1030 /* 1031 * If we are released reduce our priority 1032 */ 1033 if (td->td_release == NULL) { 1034 if (lwkt_check_resched(td) > 0) 1035 lwkt_switch(); 1036 lp->lwp_proc->p_usched->acquire_curproc(lp); 1037 td->td_release = lwkt_passive_release; 1038 lwkt_setpri_self(TDPRI_USER_NORM); 1039 } 1040 } 1041 1042 /* 1043 * Return 0 if no runnable threads are pending at the same or higher 1044 * priority as the passed thread. 1045 * 1046 * Return 1 if runnable threads are pending at the same priority. 1047 * 1048 * Return 2 if runnable threads are pending at a higher priority. 1049 */ 1050 int 1051 lwkt_check_resched(thread_t td) 1052 { 1053 int pri = td->td_pri & TDPRI_MASK; 1054 1055 if (td->td_gd->gd_runqmask > (2 << pri) - 1) 1056 return(2); 1057 if (TAILQ_NEXT(td, td_threadq)) 1058 return(1); 1059 return(0); 1060 } 1061 1062 /* 1063 * Generic schedule. Possibly schedule threads belonging to other cpus and 1064 * deal with threads that might be blocked on a wait queue. 1065 * 1066 * We have a little helper inline function which does additional work after 1067 * the thread has been enqueued, including dealing with preemption and 1068 * setting need_lwkt_resched() (which prevents the kernel from returning 1069 * to userland until it has processed higher priority threads). 1070 * 1071 * It is possible for this routine to be called after a failed _enqueue 1072 * (due to the target thread migrating, sleeping, or otherwise blocked). 1073 * We have to check that the thread is actually on the run queue! 1074 * 1075 * reschedok is an optimized constant propagated from lwkt_schedule() or 1076 * lwkt_schedule_noresched(). By default it is non-zero, causing a 1077 * reschedule to be requested if the target thread has a higher priority. 1078 * The port messaging code will set MSG_NORESCHED and cause reschedok to 1079 * be 0, prevented undesired reschedules. 1080 */ 1081 static __inline 1082 void 1083 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri, int reschedok) 1084 { 1085 thread_t otd; 1086 1087 if (ntd->td_flags & TDF_RUNQ) { 1088 if (ntd->td_preemptable && reschedok) { 1089 ntd->td_preemptable(ntd, cpri); /* YYY +token */ 1090 } else if (reschedok) { 1091 otd = curthread; 1092 if ((ntd->td_pri & TDPRI_MASK) > (otd->td_pri & TDPRI_MASK)) 1093 need_lwkt_resched(); 1094 } 1095 } 1096 } 1097 1098 static __inline 1099 void 1100 _lwkt_schedule(thread_t td, int reschedok) 1101 { 1102 globaldata_t mygd = mycpu; 1103 1104 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!")); 1105 crit_enter_gd(mygd); 1106 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0); 1107 if (td == mygd->gd_curthread) { 1108 _lwkt_enqueue(td); 1109 } else { 1110 /* 1111 * If we own the thread, there is no race (since we are in a 1112 * critical section). If we do not own the thread there might 1113 * be a race but the target cpu will deal with it. 1114 */ 1115 #ifdef SMP 1116 if (td->td_gd == mygd) { 1117 _lwkt_enqueue(td); 1118 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok); 1119 } else { 1120 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td); 1121 } 1122 #else 1123 _lwkt_enqueue(td); 1124 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok); 1125 #endif 1126 } 1127 crit_exit_gd(mygd); 1128 } 1129 1130 void 1131 lwkt_schedule(thread_t td) 1132 { 1133 _lwkt_schedule(td, 1); 1134 } 1135 1136 void 1137 lwkt_schedule_noresched(thread_t td) 1138 { 1139 _lwkt_schedule(td, 0); 1140 } 1141 1142 #ifdef SMP 1143 1144 /* 1145 * Thread migration using a 'Pull' method. The thread may or may not be 1146 * the current thread. It MUST be descheduled and in a stable state. 1147 * lwkt_giveaway() must be called on the cpu owning the thread. 1148 * 1149 * At any point after lwkt_giveaway() is called, the target cpu may 1150 * 'pull' the thread by calling lwkt_acquire(). 1151 * 1152 * We have to make sure the thread is not sitting on a per-cpu tsleep 1153 * queue or it will blow up when it moves to another cpu. 1154 * 1155 * MPSAFE - must be called under very specific conditions. 1156 */ 1157 void 1158 lwkt_giveaway(thread_t td) 1159 { 1160 globaldata_t gd = mycpu; 1161 1162 crit_enter_gd(gd); 1163 if (td->td_flags & TDF_TSLEEPQ) 1164 tsleep_remove(td); 1165 KKASSERT(td->td_gd == gd); 1166 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); 1167 td->td_flags |= TDF_MIGRATING; 1168 crit_exit_gd(gd); 1169 } 1170 1171 void 1172 lwkt_acquire(thread_t td) 1173 { 1174 globaldata_t gd; 1175 globaldata_t mygd; 1176 1177 KKASSERT(td->td_flags & TDF_MIGRATING); 1178 gd = td->td_gd; 1179 mygd = mycpu; 1180 if (gd != mycpu) { 1181 cpu_lfence(); 1182 KKASSERT((td->td_flags & TDF_RUNQ) == 0); 1183 crit_enter_gd(mygd); 1184 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) { 1185 #ifdef SMP 1186 lwkt_process_ipiq(); 1187 #endif 1188 cpu_lfence(); 1189 } 1190 td->td_gd = mygd; 1191 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1192 td->td_flags &= ~TDF_MIGRATING; 1193 crit_exit_gd(mygd); 1194 } else { 1195 crit_enter_gd(mygd); 1196 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); 1197 td->td_flags &= ~TDF_MIGRATING; 1198 crit_exit_gd(mygd); 1199 } 1200 } 1201 1202 #endif 1203 1204 /* 1205 * Generic deschedule. Descheduling threads other then your own should be 1206 * done only in carefully controlled circumstances. Descheduling is 1207 * asynchronous. 1208 * 1209 * This function may block if the cpu has run out of messages. 1210 */ 1211 void 1212 lwkt_deschedule(thread_t td) 1213 { 1214 crit_enter(); 1215 #ifdef SMP 1216 if (td == curthread) { 1217 _lwkt_dequeue(td); 1218 } else { 1219 if (td->td_gd == mycpu) { 1220 _lwkt_dequeue(td); 1221 } else { 1222 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td); 1223 } 1224 } 1225 #else 1226 _lwkt_dequeue(td); 1227 #endif 1228 crit_exit(); 1229 } 1230 1231 /* 1232 * Set the target thread's priority. This routine does not automatically 1233 * switch to a higher priority thread, LWKT threads are not designed for 1234 * continuous priority changes. Yield if you want to switch. 1235 * 1236 * We have to retain the critical section count which uses the high bits 1237 * of the td_pri field. The specified priority may also indicate zero or 1238 * more critical sections by adding TDPRI_CRIT*N. 1239 * 1240 * Note that we requeue the thread whether it winds up on a different runq 1241 * or not. uio_yield() depends on this and the routine is not normally 1242 * called with the same priority otherwise. 1243 */ 1244 void 1245 lwkt_setpri(thread_t td, int pri) 1246 { 1247 KKASSERT(pri >= 0); 1248 KKASSERT(td->td_gd == mycpu); 1249 crit_enter(); 1250 if (td->td_flags & TDF_RUNQ) { 1251 _lwkt_dequeue(td); 1252 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri; 1253 _lwkt_enqueue(td); 1254 } else { 1255 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri; 1256 } 1257 crit_exit(); 1258 } 1259 1260 void 1261 lwkt_setpri_self(int pri) 1262 { 1263 thread_t td = curthread; 1264 1265 KKASSERT(pri >= 0 && pri <= TDPRI_MAX); 1266 crit_enter(); 1267 if (td->td_flags & TDF_RUNQ) { 1268 _lwkt_dequeue(td); 1269 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri; 1270 _lwkt_enqueue(td); 1271 } else { 1272 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri; 1273 } 1274 crit_exit(); 1275 } 1276 1277 /* 1278 * Migrate the current thread to the specified cpu. 1279 * 1280 * This is accomplished by descheduling ourselves from the current cpu, 1281 * moving our thread to the tdallq of the target cpu, IPI messaging the 1282 * target cpu, and switching out. TDF_MIGRATING prevents scheduling 1283 * races while the thread is being migrated. 1284 * 1285 * We must be sure to remove ourselves from the current cpu's tsleepq 1286 * before potentially moving to another queue. The thread can be on 1287 * a tsleepq due to a left-over tsleep_interlock(). 1288 */ 1289 #ifdef SMP 1290 static void lwkt_setcpu_remote(void *arg); 1291 #endif 1292 1293 void 1294 lwkt_setcpu_self(globaldata_t rgd) 1295 { 1296 #ifdef SMP 1297 thread_t td = curthread; 1298 1299 if (td->td_gd != rgd) { 1300 crit_enter_quick(td); 1301 if (td->td_flags & TDF_TSLEEPQ) 1302 tsleep_remove(td); 1303 td->td_flags |= TDF_MIGRATING; 1304 lwkt_deschedule_self(td); 1305 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1306 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td); 1307 lwkt_switch(); 1308 /* we are now on the target cpu */ 1309 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); 1310 crit_exit_quick(td); 1311 } 1312 #endif 1313 } 1314 1315 void 1316 lwkt_migratecpu(int cpuid) 1317 { 1318 #ifdef SMP 1319 globaldata_t rgd; 1320 1321 rgd = globaldata_find(cpuid); 1322 lwkt_setcpu_self(rgd); 1323 #endif 1324 } 1325 1326 /* 1327 * Remote IPI for cpu migration (called while in a critical section so we 1328 * do not have to enter another one). The thread has already been moved to 1329 * our cpu's allq, but we must wait for the thread to be completely switched 1330 * out on the originating cpu before we schedule it on ours or the stack 1331 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD 1332 * change to main memory. 1333 * 1334 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races 1335 * against wakeups. It is best if this interface is used only when there 1336 * are no pending events that might try to schedule the thread. 1337 */ 1338 #ifdef SMP 1339 static void 1340 lwkt_setcpu_remote(void *arg) 1341 { 1342 thread_t td = arg; 1343 globaldata_t gd = mycpu; 1344 1345 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) { 1346 #ifdef SMP 1347 lwkt_process_ipiq(); 1348 #endif 1349 cpu_lfence(); 1350 } 1351 td->td_gd = gd; 1352 cpu_sfence(); 1353 td->td_flags &= ~TDF_MIGRATING; 1354 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0); 1355 _lwkt_enqueue(td); 1356 } 1357 #endif 1358 1359 struct lwp * 1360 lwkt_preempted_proc(void) 1361 { 1362 thread_t td = curthread; 1363 while (td->td_preempted) 1364 td = td->td_preempted; 1365 return(td->td_lwp); 1366 } 1367 1368 /* 1369 * Create a kernel process/thread/whatever. It shares it's address space 1370 * with proc0 - ie: kernel only. 1371 * 1372 * NOTE! By default new threads are created with the MP lock held. A 1373 * thread which does not require the MP lock should release it by calling 1374 * rel_mplock() at the start of the new thread. 1375 */ 1376 int 1377 lwkt_create(void (*func)(void *), void *arg, 1378 struct thread **tdp, thread_t template, int tdflags, int cpu, 1379 const char *fmt, ...) 1380 { 1381 thread_t td; 1382 __va_list ap; 1383 1384 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu, 1385 tdflags); 1386 if (tdp) 1387 *tdp = td; 1388 cpu_set_thread_handler(td, lwkt_exit, func, arg); 1389 1390 /* 1391 * Set up arg0 for 'ps' etc 1392 */ 1393 __va_start(ap, fmt); 1394 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap); 1395 __va_end(ap); 1396 1397 /* 1398 * Schedule the thread to run 1399 */ 1400 if ((td->td_flags & TDF_STOPREQ) == 0) 1401 lwkt_schedule(td); 1402 else 1403 td->td_flags &= ~TDF_STOPREQ; 1404 return 0; 1405 } 1406 1407 /* 1408 * Destroy an LWKT thread. Warning! This function is not called when 1409 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and 1410 * uses a different reaping mechanism. 1411 */ 1412 void 1413 lwkt_exit(void) 1414 { 1415 thread_t td = curthread; 1416 thread_t std; 1417 globaldata_t gd; 1418 1419 if (td->td_flags & TDF_VERBOSE) 1420 kprintf("kthread %p %s has exited\n", td, td->td_comm); 1421 caps_exit(td); 1422 1423 /* 1424 * Get us into a critical section to interlock gd_freetd and loop 1425 * until we can get it freed. 1426 * 1427 * We have to cache the current td in gd_freetd because objcache_put()ing 1428 * it would rip it out from under us while our thread is still active. 1429 */ 1430 gd = mycpu; 1431 crit_enter_quick(td); 1432 while ((std = gd->gd_freetd) != NULL) { 1433 gd->gd_freetd = NULL; 1434 objcache_put(thread_cache, std); 1435 } 1436 1437 /* 1438 * Remove thread resources from kernel lists and deschedule us for 1439 * the last time. 1440 */ 1441 if (td->td_flags & TDF_TSLEEPQ) 1442 tsleep_remove(td); 1443 lwkt_deschedule_self(td); 1444 lwkt_remove_tdallq(td); 1445 if (td->td_flags & TDF_ALLOCATED_THREAD) 1446 gd->gd_freetd = td; 1447 cpu_thread_exit(); 1448 } 1449 1450 void 1451 lwkt_remove_tdallq(thread_t td) 1452 { 1453 KKASSERT(td->td_gd == mycpu); 1454 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); 1455 } 1456 1457 void 1458 crit_panic(void) 1459 { 1460 thread_t td = curthread; 1461 int lpri = td->td_pri; 1462 1463 td->td_pri = 0; 1464 panic("td_pri is/would-go negative! %p %d", td, lpri); 1465 } 1466 1467 #ifdef SMP 1468 1469 /* 1470 * Called from debugger/panic on cpus which have been stopped. We must still 1471 * process the IPIQ while stopped, even if we were stopped while in a critical 1472 * section (XXX). 1473 * 1474 * If we are dumping also try to process any pending interrupts. This may 1475 * or may not work depending on the state of the cpu at the point it was 1476 * stopped. 1477 */ 1478 void 1479 lwkt_smp_stopped(void) 1480 { 1481 globaldata_t gd = mycpu; 1482 1483 crit_enter_gd(gd); 1484 if (dumping) { 1485 lwkt_process_ipiq(); 1486 splz(); 1487 } else { 1488 lwkt_process_ipiq(); 1489 } 1490 crit_exit_gd(gd); 1491 } 1492 1493 /* 1494 * get_mplock() calls this routine if it is unable to obtain the MP lock. 1495 * get_mplock() has already incremented td_mpcount. We must block and 1496 * not return until giant is held. 1497 * 1498 * All we have to do is lwkt_switch() away. The LWKT scheduler will not 1499 * reschedule the thread until it can obtain the giant lock for it. 1500 */ 1501 void 1502 lwkt_mp_lock_contested(void) 1503 { 1504 ++mplock_countx; 1505 loggiant(beg); 1506 lwkt_switch(); 1507 loggiant(end); 1508 } 1509 1510 /* 1511 * The rel_mplock() code will call this function after releasing the 1512 * last reference on the MP lock if mp_lock_contention_mask is non-zero. 1513 * 1514 * We then chain an IPI to a single other cpu potentially needing the 1515 * lock. This is a bit heuristical and we can wind up with IPIs flying 1516 * all over the place. 1517 */ 1518 static void lwkt_mp_lock_uncontested_remote(void *arg __unused); 1519 1520 void 1521 lwkt_mp_lock_uncontested(void) 1522 { 1523 globaldata_t gd; 1524 globaldata_t dgd; 1525 cpumask_t mask; 1526 cpumask_t tmpmask; 1527 int cpuid; 1528 1529 if (chain_mplock) { 1530 gd = mycpu; 1531 atomic_clear_int(&mp_lock_contention_mask, gd->gd_cpumask); 1532 mask = mp_lock_contention_mask; 1533 tmpmask = ~((1 << gd->gd_cpuid) - 1); 1534 1535 if (mask) { 1536 if (mask & tmpmask) 1537 cpuid = bsfl(mask & tmpmask); 1538 else 1539 cpuid = bsfl(mask); 1540 atomic_clear_int(&mp_lock_contention_mask, 1 << cpuid); 1541 dgd = globaldata_find(cpuid); 1542 lwkt_send_ipiq(dgd, lwkt_mp_lock_uncontested_remote, NULL); 1543 } 1544 } 1545 } 1546 1547 /* 1548 * The idea is for this IPI to interrupt a potentially lower priority 1549 * thread, such as a user thread, to allow the scheduler to reschedule 1550 * a higher priority kernel thread that needs the MP lock. 1551 * 1552 * For now we set the LWKT reschedule flag which generates an AST in 1553 * doreti, though theoretically it is also possible to possibly preempt 1554 * here if the underlying thread was operating in user mode. Nah. 1555 */ 1556 static void 1557 lwkt_mp_lock_uncontested_remote(void *arg __unused) 1558 { 1559 need_lwkt_resched(); 1560 } 1561 1562 #endif 1563