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 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.65 2004/07/16 05:51:10 dillon Exp $ 35 */ 36 37 /* 38 * Each cpu in a system has its own self-contained light weight kernel 39 * thread scheduler, which means that generally speaking we only need 40 * to use a critical section to avoid problems. Foreign thread 41 * scheduling is queued via (async) IPIs. 42 */ 43 44 #ifdef _KERNEL 45 46 #include <sys/param.h> 47 #include <sys/systm.h> 48 #include <sys/kernel.h> 49 #include <sys/proc.h> 50 #include <sys/rtprio.h> 51 #include <sys/queue.h> 52 #include <sys/thread2.h> 53 #include <sys/sysctl.h> 54 #include <sys/kthread.h> 55 #include <machine/cpu.h> 56 #include <sys/lock.h> 57 #include <sys/caps.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 #include <vm/vm_zone.h> 68 69 #include <machine/stdarg.h> 70 #include <machine/ipl.h> 71 #include <machine/smp.h> 72 73 #define THREAD_STACK (UPAGES * PAGE_SIZE) 74 75 #else 76 77 #include <sys/stdint.h> 78 #include <libcaps/thread.h> 79 #include <sys/thread.h> 80 #include <sys/msgport.h> 81 #include <sys/errno.h> 82 #include <libcaps/globaldata.h> 83 #include <machine/cpufunc.h> 84 #include <sys/thread2.h> 85 #include <sys/msgport2.h> 86 #include <stdio.h> 87 #include <stdlib.h> 88 #include <string.h> 89 #include <machine/lock.h> 90 91 #endif 92 93 static int untimely_switch = 0; 94 #ifdef INVARIANTS 95 static int panic_on_cscount = 0; 96 #endif 97 static __int64_t switch_count = 0; 98 static __int64_t preempt_hit = 0; 99 static __int64_t preempt_miss = 0; 100 static __int64_t preempt_weird = 0; 101 102 #ifdef _KERNEL 103 104 SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, ""); 105 #ifdef INVARIANTS 106 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, ""); 107 #endif 108 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, ""); 109 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, ""); 110 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, ""); 111 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, ""); 112 113 #endif 114 115 /* 116 * These helper procedures handle the runq, they can only be called from 117 * within a critical section. 118 * 119 * WARNING! Prior to SMP being brought up it is possible to enqueue and 120 * dequeue threads belonging to other cpus, so be sure to use td->td_gd 121 * instead of 'mycpu' when referencing the globaldata structure. Once 122 * SMP live enqueuing and dequeueing only occurs on the current cpu. 123 */ 124 static __inline 125 void 126 _lwkt_dequeue(thread_t td) 127 { 128 if (td->td_flags & TDF_RUNQ) { 129 int nq = td->td_pri & TDPRI_MASK; 130 struct globaldata *gd = td->td_gd; 131 132 td->td_flags &= ~TDF_RUNQ; 133 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq); 134 /* runqmask is passively cleaned up by the switcher */ 135 } 136 } 137 138 static __inline 139 void 140 _lwkt_enqueue(thread_t td) 141 { 142 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING)) == 0) { 143 int nq = td->td_pri & TDPRI_MASK; 144 struct globaldata *gd = td->td_gd; 145 146 td->td_flags |= TDF_RUNQ; 147 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq); 148 gd->gd_runqmask |= 1 << nq; 149 } 150 } 151 152 /* 153 * Schedule a thread to run. As the current thread we can always safely 154 * schedule ourselves, and a shortcut procedure is provided for that 155 * function. 156 * 157 * (non-blocking, self contained on a per cpu basis) 158 */ 159 void 160 lwkt_schedule_self(thread_t td) 161 { 162 crit_enter_quick(td); 163 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!")); 164 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!")); 165 _lwkt_enqueue(td); 166 #ifdef _KERNEL 167 if (td->td_proc && td->td_proc->p_stat == SSLEEP) 168 panic("SCHED SELF PANIC"); 169 #endif 170 crit_exit_quick(td); 171 } 172 173 /* 174 * Deschedule a thread. 175 * 176 * (non-blocking, self contained on a per cpu basis) 177 */ 178 void 179 lwkt_deschedule_self(thread_t td) 180 { 181 crit_enter_quick(td); 182 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!")); 183 _lwkt_dequeue(td); 184 crit_exit_quick(td); 185 } 186 187 #ifdef _KERNEL 188 189 /* 190 * LWKTs operate on a per-cpu basis 191 * 192 * WARNING! Called from early boot, 'mycpu' may not work yet. 193 */ 194 void 195 lwkt_gdinit(struct globaldata *gd) 196 { 197 int i; 198 199 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i) 200 TAILQ_INIT(&gd->gd_tdrunq[i]); 201 gd->gd_runqmask = 0; 202 TAILQ_INIT(&gd->gd_tdallq); 203 } 204 205 #endif /* _KERNEL */ 206 207 /* 208 * Initialize a thread wait structure prior to first use. 209 * 210 * NOTE! called from low level boot code, we cannot do anything fancy! 211 */ 212 void 213 lwkt_wait_init(lwkt_wait_t w) 214 { 215 lwkt_token_init(&w->wa_token); 216 TAILQ_INIT(&w->wa_waitq); 217 w->wa_gen = 0; 218 w->wa_count = 0; 219 } 220 221 /* 222 * Create a new thread. The thread must be associated with a process context 223 * or LWKT start address before it can be scheduled. If the target cpu is 224 * -1 the thread will be created on the current cpu. 225 * 226 * If you intend to create a thread without a process context this function 227 * does everything except load the startup and switcher function. 228 */ 229 thread_t 230 lwkt_alloc_thread(struct thread *td, int cpu) 231 { 232 void *stack; 233 int flags = 0; 234 globaldata_t gd = mycpu; 235 236 if (td == NULL) { 237 crit_enter_gd(gd); 238 if (gd->gd_tdfreecount > 0) { 239 --gd->gd_tdfreecount; 240 td = TAILQ_FIRST(&gd->gd_tdfreeq); 241 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0, 242 ("lwkt_alloc_thread: unexpected NULL or corrupted td")); 243 TAILQ_REMOVE(&gd->gd_tdfreeq, td, td_threadq); 244 crit_exit_gd(gd); 245 stack = td->td_kstack; 246 flags = td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD); 247 } else { 248 crit_exit_gd(gd); 249 #ifdef _KERNEL 250 td = zalloc(thread_zone); 251 #else 252 td = malloc(sizeof(struct thread)); 253 #endif 254 td->td_kstack = NULL; 255 flags |= TDF_ALLOCATED_THREAD; 256 } 257 } 258 if ((stack = td->td_kstack) == NULL) { 259 #ifdef _KERNEL 260 stack = (void *)kmem_alloc(kernel_map, THREAD_STACK); 261 #else 262 stack = libcaps_alloc_stack(THREAD_STACK); 263 #endif 264 flags |= TDF_ALLOCATED_STACK; 265 } 266 if (cpu < 0) 267 lwkt_init_thread(td, stack, flags, mycpu); 268 else 269 lwkt_init_thread(td, stack, flags, globaldata_find(cpu)); 270 return(td); 271 } 272 273 #ifdef _KERNEL 274 275 /* 276 * Initialize a preexisting thread structure. This function is used by 277 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread. 278 * 279 * All threads start out in a critical section at a priority of 280 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as 281 * appropriate. This function may send an IPI message when the 282 * requested cpu is not the current cpu and consequently gd_tdallq may 283 * not be initialized synchronously from the point of view of the originating 284 * cpu. 285 * 286 * NOTE! we have to be careful in regards to creating threads for other cpus 287 * if SMP has not yet been activated. 288 */ 289 #ifdef SMP 290 291 static void 292 lwkt_init_thread_remote(void *arg) 293 { 294 thread_t td = arg; 295 296 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq); 297 } 298 299 #endif 300 301 void 302 lwkt_init_thread(thread_t td, void *stack, int flags, struct globaldata *gd) 303 { 304 globaldata_t mygd = mycpu; 305 306 bzero(td, sizeof(struct thread)); 307 td->td_kstack = stack; 308 td->td_flags |= flags; 309 td->td_gd = gd; 310 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT; 311 lwkt_initport(&td->td_msgport, td); 312 pmap_init_thread(td); 313 #ifdef SMP 314 /* 315 * Normally initializing a thread for a remote cpu requires sending an 316 * IPI. However, the idlethread is setup before the other cpus are 317 * activated so we have to treat it as a special case. XXX manipulation 318 * of gd_tdallq requires the BGL. 319 */ 320 if (gd == mygd || td == &gd->gd_idlethread) { 321 crit_enter_gd(mygd); 322 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); 323 crit_exit_gd(mygd); 324 } else { 325 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td); 326 } 327 #else 328 crit_enter_gd(mygd); 329 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); 330 crit_exit_gd(mygd); 331 #endif 332 } 333 334 #endif /* _KERNEL */ 335 336 void 337 lwkt_set_comm(thread_t td, const char *ctl, ...) 338 { 339 __va_list va; 340 341 __va_start(va, ctl); 342 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va); 343 __va_end(va); 344 } 345 346 void 347 lwkt_hold(thread_t td) 348 { 349 ++td->td_refs; 350 } 351 352 void 353 lwkt_rele(thread_t td) 354 { 355 KKASSERT(td->td_refs > 0); 356 --td->td_refs; 357 } 358 359 #ifdef _KERNEL 360 361 void 362 lwkt_wait_free(thread_t td) 363 { 364 while (td->td_refs) 365 tsleep(td, 0, "tdreap", hz); 366 } 367 368 #endif 369 370 void 371 lwkt_free_thread(thread_t td) 372 { 373 struct globaldata *gd = mycpu; 374 375 KASSERT((td->td_flags & TDF_RUNNING) == 0, 376 ("lwkt_free_thread: did not exit! %p", td)); 377 378 crit_enter_gd(gd); 379 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); 380 if (gd->gd_tdfreecount < CACHE_NTHREADS && 381 (td->td_flags & TDF_ALLOCATED_THREAD) 382 ) { 383 ++gd->gd_tdfreecount; 384 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq); 385 crit_exit_gd(gd); 386 } else { 387 crit_exit_gd(gd); 388 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) { 389 #ifdef _KERNEL 390 kmem_free(kernel_map, (vm_offset_t)td->td_kstack, THREAD_STACK); 391 #else 392 libcaps_free_stack(td->td_kstack, THREAD_STACK); 393 #endif 394 /* gd invalid */ 395 td->td_kstack = NULL; 396 } 397 if (td->td_flags & TDF_ALLOCATED_THREAD) { 398 #ifdef _KERNEL 399 zfree(thread_zone, td); 400 #else 401 free(td); 402 #endif 403 } 404 } 405 } 406 407 408 /* 409 * Switch to the next runnable lwkt. If no LWKTs are runnable then 410 * switch to the idlethread. Switching must occur within a critical 411 * section to avoid races with the scheduling queue. 412 * 413 * We always have full control over our cpu's run queue. Other cpus 414 * that wish to manipulate our queue must use the cpu_*msg() calls to 415 * talk to our cpu, so a critical section is all that is needed and 416 * the result is very, very fast thread switching. 417 * 418 * The LWKT scheduler uses a fixed priority model and round-robins at 419 * each priority level. User process scheduling is a totally 420 * different beast and LWKT priorities should not be confused with 421 * user process priorities. 422 * 423 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch() 424 * cleans it up. Note that the td_switch() function cannot do anything that 425 * requires the MP lock since the MP lock will have already been setup for 426 * the target thread (not the current thread). It's nice to have a scheduler 427 * that does not need the MP lock to work because it allows us to do some 428 * really cool high-performance MP lock optimizations. 429 */ 430 431 void 432 lwkt_switch(void) 433 { 434 globaldata_t gd = mycpu; 435 thread_t td = gd->gd_curthread; 436 thread_t ntd; 437 #ifdef SMP 438 int mpheld; 439 #endif 440 441 /* 442 * Switching from within a 'fast' (non thread switched) interrupt is 443 * illegal. 444 */ 445 if (gd->gd_intr_nesting_level && panicstr == NULL) { 446 panic("lwkt_switch: cannot switch from within a fast interrupt, yet"); 447 } 448 449 /* 450 * Passive release (used to transition from user to kernel mode 451 * when we block or switch rather then when we enter the kernel). 452 * This function is NOT called if we are switching into a preemption 453 * or returning from a preemption. Typically this causes us to lose 454 * our current process designation (if we have one) and become a true 455 * LWKT thread, and may also hand the current process designation to 456 * another process and schedule thread. 457 */ 458 if (td->td_release) 459 td->td_release(td); 460 461 crit_enter_gd(gd); 462 ++switch_count; 463 464 #ifdef SMP 465 /* 466 * td_mpcount cannot be used to determine if we currently hold the 467 * MP lock because get_mplock() will increment it prior to attempting 468 * to get the lock, and switch out if it can't. Our ownership of 469 * the actual lock will remain stable while we are in a critical section 470 * (but, of course, another cpu may own or release the lock so the 471 * actual value of mp_lock is not stable). 472 */ 473 mpheld = MP_LOCK_HELD(); 474 #ifdef INVARIANTS 475 if (td->td_cscount) { 476 printf("Diagnostic: attempt to switch while mastering cpusync: %p\n", 477 td); 478 if (panic_on_cscount) 479 panic("switching while mastering cpusync"); 480 } 481 #endif 482 #endif 483 if ((ntd = td->td_preempted) != NULL) { 484 /* 485 * We had preempted another thread on this cpu, resume the preempted 486 * thread. This occurs transparently, whether the preempted thread 487 * was scheduled or not (it may have been preempted after descheduling 488 * itself). 489 * 490 * We have to setup the MP lock for the original thread after backing 491 * out the adjustment that was made to curthread when the original 492 * was preempted. 493 */ 494 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK); 495 #ifdef SMP 496 if (ntd->td_mpcount && mpheld == 0) { 497 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d", 498 td, ntd, td->td_mpcount, ntd->td_mpcount); 499 } 500 if (ntd->td_mpcount) { 501 td->td_mpcount -= ntd->td_mpcount; 502 KKASSERT(td->td_mpcount >= 0); 503 } 504 #endif 505 ntd->td_flags |= TDF_PREEMPT_DONE; 506 /* YYY release mp lock on switchback if original doesn't need it */ 507 } else { 508 /* 509 * Priority queue / round-robin at each priority. Note that user 510 * processes run at a fixed, low priority and the user process 511 * scheduler deals with interactions between user processes 512 * by scheduling and descheduling them from the LWKT queue as 513 * necessary. 514 * 515 * We have to adjust the MP lock for the target thread. If we 516 * need the MP lock and cannot obtain it we try to locate a 517 * thread that does not need the MP lock. If we cannot, we spin 518 * instead of HLT. 519 * 520 * A similar issue exists for the tokens held by the target thread. 521 * If we cannot obtain ownership of the tokens we cannot immediately 522 * schedule the thread. 523 */ 524 525 /* 526 * We are switching threads. If there are any pending requests for 527 * tokens we can satisfy all of them here. 528 */ 529 #ifdef SMP 530 if (gd->gd_tokreqbase) 531 lwkt_drain_token_requests(); 532 #endif 533 534 again: 535 if (gd->gd_runqmask) { 536 int nq = bsrl(gd->gd_runqmask); 537 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) { 538 gd->gd_runqmask &= ~(1 << nq); 539 goto again; 540 } 541 #ifdef SMP 542 /* 543 * If the target needs the MP lock and we couldn't get it, 544 * or if the target is holding tokens and we could not 545 * gain ownership of the tokens, continue looking for a 546 * thread to schedule and spin instead of HLT if we can't. 547 */ 548 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) || 549 (ntd->td_toks && lwkt_chktokens(ntd) == 0) 550 ) { 551 u_int32_t rqmask = gd->gd_runqmask; 552 while (rqmask) { 553 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) { 554 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) 555 continue; 556 mpheld = MP_LOCK_HELD(); 557 if (ntd->td_toks && !lwkt_chktokens(ntd)) 558 continue; 559 break; 560 } 561 if (ntd) 562 break; 563 rqmask &= ~(1 << nq); 564 nq = bsrl(rqmask); 565 } 566 if (ntd == NULL) { 567 ntd = &gd->gd_idlethread; 568 ntd->td_flags |= TDF_IDLE_NOHLT; 569 } else { 570 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq); 571 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq); 572 } 573 } else { 574 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq); 575 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq); 576 } 577 #else 578 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq); 579 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq); 580 #endif 581 } else { 582 /* 583 * We have nothing to run but only let the idle loop halt 584 * the cpu if there are no pending interrupts. 585 */ 586 ntd = &gd->gd_idlethread; 587 if (gd->gd_reqflags & RQF_IDLECHECK_MASK) 588 ntd->td_flags |= TDF_IDLE_NOHLT; 589 } 590 } 591 KASSERT(ntd->td_pri >= TDPRI_CRIT, 592 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri)); 593 594 /* 595 * Do the actual switch. If the new target does not need the MP lock 596 * and we are holding it, release the MP lock. If the new target requires 597 * the MP lock we have already acquired it for the target. 598 */ 599 #ifdef SMP 600 if (ntd->td_mpcount == 0 ) { 601 if (MP_LOCK_HELD()) 602 cpu_rel_mplock(); 603 } else { 604 ASSERT_MP_LOCK_HELD(); 605 } 606 #endif 607 if (td != ntd) 608 td->td_switch(ntd); 609 /* NOTE: current cpu may have changed after switch */ 610 crit_exit_quick(td); 611 } 612 613 /* 614 * Request that the target thread preempt the current thread. Preemption 615 * only works under a specific set of conditions: 616 * 617 * - We are not preempting ourselves 618 * - The target thread is owned by the current cpu 619 * - We are not currently being preempted 620 * - The target is not currently being preempted 621 * - We are able to satisfy the target's MP lock requirements (if any). 622 * 623 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically 624 * this is called via lwkt_schedule() through the td_preemptable callback. 625 * critpri is the managed critical priority that we should ignore in order 626 * to determine whether preemption is possible (aka usually just the crit 627 * priority of lwkt_schedule() itself). 628 * 629 * XXX at the moment we run the target thread in a critical section during 630 * the preemption in order to prevent the target from taking interrupts 631 * that *WE* can't. Preemption is strictly limited to interrupt threads 632 * and interrupt-like threads, outside of a critical section, and the 633 * preempted source thread will be resumed the instant the target blocks 634 * whether or not the source is scheduled (i.e. preemption is supposed to 635 * be as transparent as possible). 636 * 637 * The target thread inherits our MP count (added to its own) for the 638 * duration of the preemption in order to preserve the atomicy of the 639 * MP lock during the preemption. Therefore, any preempting targets must be 640 * careful in regards to MP assertions. Note that the MP count may be 641 * out of sync with the physical mp_lock, but we do not have to preserve 642 * the original ownership of the lock if it was out of synch (that is, we 643 * can leave it synchronized on return). 644 */ 645 void 646 lwkt_preempt(thread_t ntd, int critpri) 647 { 648 struct globaldata *gd = mycpu; 649 thread_t td; 650 #ifdef SMP 651 int mpheld; 652 int savecnt; 653 #endif 654 655 /* 656 * The caller has put us in a critical section. We can only preempt 657 * if the caller of the caller was not in a critical section (basically 658 * a local interrupt), as determined by the 'critpri' parameter. 659 * 660 * YYY The target thread must be in a critical section (else it must 661 * inherit our critical section? I dunno yet). 662 * 663 * Any tokens held by the target may not be held by thread(s) being 664 * preempted. We take the easy way out and do not preempt if 665 * the target is holding tokens. 666 * 667 * Set need_lwkt_resched() unconditionally for now YYY. 668 */ 669 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri)); 670 671 td = gd->gd_curthread; 672 need_lwkt_resched(); 673 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) { 674 ++preempt_miss; 675 return; 676 } 677 if ((td->td_pri & ~TDPRI_MASK) > critpri) { 678 ++preempt_miss; 679 return; 680 } 681 #ifdef SMP 682 if (ntd->td_gd != gd) { 683 ++preempt_miss; 684 return; 685 } 686 #endif 687 /* 688 * Take the easy way out and do not preempt if the target is holding 689 * one or more tokens. We could test whether the thread(s) being 690 * preempted interlock against the target thread's tokens and whether 691 * we can get all the target thread's tokens, but this situation 692 * should not occur very often so its easier to simply not preempt. 693 */ 694 if (ntd->td_toks != NULL) { 695 ++preempt_miss; 696 return; 697 } 698 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) { 699 ++preempt_weird; 700 return; 701 } 702 if (ntd->td_preempted) { 703 ++preempt_hit; 704 return; 705 } 706 #ifdef SMP 707 /* 708 * note: an interrupt might have occured just as we were transitioning 709 * to or from the MP lock. In this case td_mpcount will be pre-disposed 710 * (non-zero) but not actually synchronized with the actual state of the 711 * lock. We can use it to imply an MP lock requirement for the 712 * preemption but we cannot use it to test whether we hold the MP lock 713 * or not. 714 */ 715 savecnt = td->td_mpcount; 716 mpheld = MP_LOCK_HELD(); 717 ntd->td_mpcount += td->td_mpcount; 718 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) { 719 ntd->td_mpcount -= td->td_mpcount; 720 ++preempt_miss; 721 return; 722 } 723 #endif 724 725 ++preempt_hit; 726 ntd->td_preempted = td; 727 td->td_flags |= TDF_PREEMPT_LOCK; 728 td->td_switch(ntd); 729 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE)); 730 #ifdef SMP 731 KKASSERT(savecnt == td->td_mpcount); 732 mpheld = MP_LOCK_HELD(); 733 if (mpheld && td->td_mpcount == 0) 734 cpu_rel_mplock(); 735 else if (mpheld == 0 && td->td_mpcount) 736 panic("lwkt_preempt(): MP lock was not held through"); 737 #endif 738 ntd->td_preempted = NULL; 739 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE); 740 } 741 742 /* 743 * Yield our thread while higher priority threads are pending. This is 744 * typically called when we leave a critical section but it can be safely 745 * called while we are in a critical section. 746 * 747 * This function will not generally yield to equal priority threads but it 748 * can occur as a side effect. Note that lwkt_switch() is called from 749 * inside the critical section to prevent its own crit_exit() from reentering 750 * lwkt_yield_quick(). 751 * 752 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint 753 * came along but was blocked and made pending. 754 * 755 * (self contained on a per cpu basis) 756 */ 757 void 758 lwkt_yield_quick(void) 759 { 760 globaldata_t gd = mycpu; 761 thread_t td = gd->gd_curthread; 762 763 /* 764 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear 765 * it with a non-zero cpl then we might not wind up calling splz after 766 * a task switch when the critical section is exited even though the 767 * new task could accept the interrupt. 768 * 769 * XXX from crit_exit() only called after last crit section is released. 770 * If called directly will run splz() even if in a critical section. 771 * 772 * td_nest_count prevent deep nesting via splz() or doreti(). Note that 773 * except for this special case, we MUST call splz() here to handle any 774 * pending ints, particularly after we switch, or we might accidently 775 * halt the cpu with interrupts pending. 776 */ 777 if (gd->gd_reqflags && td->td_nest_count < 2) 778 splz(); 779 780 /* 781 * YYY enabling will cause wakeup() to task-switch, which really 782 * confused the old 4.x code. This is a good way to simulate 783 * preemption and MP without actually doing preemption or MP, because a 784 * lot of code assumes that wakeup() does not block. 785 */ 786 if (untimely_switch && td->td_nest_count == 0 && 787 gd->gd_intr_nesting_level == 0 788 ) { 789 crit_enter_quick(td); 790 /* 791 * YYY temporary hacks until we disassociate the userland scheduler 792 * from the LWKT scheduler. 793 */ 794 if (td->td_flags & TDF_RUNQ) { 795 lwkt_switch(); /* will not reenter yield function */ 796 } else { 797 lwkt_schedule_self(td); /* make sure we are scheduled */ 798 lwkt_switch(); /* will not reenter yield function */ 799 lwkt_deschedule_self(td); /* make sure we are descheduled */ 800 } 801 crit_exit_noyield(td); 802 } 803 } 804 805 /* 806 * This implements a normal yield which, unlike _quick, will yield to equal 807 * priority threads as well. Note that gd_reqflags tests will be handled by 808 * the crit_exit() call in lwkt_switch(). 809 * 810 * (self contained on a per cpu basis) 811 */ 812 void 813 lwkt_yield(void) 814 { 815 lwkt_schedule_self(curthread); 816 lwkt_switch(); 817 } 818 819 /* 820 * Generic schedule. Possibly schedule threads belonging to other cpus and 821 * deal with threads that might be blocked on a wait queue. 822 * 823 * We have a little helper inline function which does additional work after 824 * the thread has been enqueued, including dealing with preemption and 825 * setting need_lwkt_resched() (which prevents the kernel from returning 826 * to userland until it has processed higher priority threads). 827 */ 828 static __inline 829 void 830 _lwkt_schedule_post(thread_t ntd, int cpri) 831 { 832 if (ntd->td_preemptable) { 833 ntd->td_preemptable(ntd, cpri); /* YYY +token */ 834 } else { 835 if ((ntd->td_flags & TDF_NORESCHED) == 0) { 836 if ((ntd->td_pri & TDPRI_MASK) >= TDPRI_KERN_USER) 837 need_lwkt_resched(); 838 } 839 } 840 } 841 842 void 843 lwkt_schedule(thread_t td) 844 { 845 globaldata_t mygd = mycpu; 846 847 #ifdef INVARIANTS 848 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!")); 849 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc 850 && td->td_proc->p_stat == SSLEEP 851 ) { 852 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n", 853 curthread, 854 curthread->td_proc ? curthread->td_proc->p_pid : -1, 855 curthread->td_proc ? curthread->td_proc->p_stat : -1, 856 td, 857 td->td_proc ? curthread->td_proc->p_pid : -1, 858 td->td_proc ? curthread->td_proc->p_stat : -1 859 ); 860 panic("SCHED PANIC"); 861 } 862 #endif 863 crit_enter_gd(mygd); 864 if (td == mygd->gd_curthread) { 865 _lwkt_enqueue(td); 866 } else { 867 lwkt_wait_t w; 868 869 /* 870 * If the thread is on a wait list we have to send our scheduling 871 * request to the owner of the wait structure. Otherwise we send 872 * the scheduling request to the cpu owning the thread. Races 873 * are ok, the target will forward the message as necessary (the 874 * message may chase the thread around before it finally gets 875 * acted upon). 876 * 877 * (remember, wait structures use stable storage) 878 * 879 * NOTE: tokens no longer enter a critical section, so we only need 880 * to account for the crit_enter() above when calling 881 * _lwkt_schedule_post(). 882 */ 883 if ((w = td->td_wait) != NULL) { 884 lwkt_tokref wref; 885 886 if (lwkt_trytoken(&wref, &w->wa_token)) { 887 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq); 888 --w->wa_count; 889 td->td_wait = NULL; 890 #ifdef SMP 891 if (td->td_gd == mycpu) { 892 _lwkt_enqueue(td); 893 _lwkt_schedule_post(td, TDPRI_CRIT); 894 } else { 895 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td); 896 } 897 #else 898 _lwkt_enqueue(td); 899 _lwkt_schedule_post(td, TDPRI_CRIT); 900 #endif 901 lwkt_reltoken(&wref); 902 } else { 903 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td); 904 } 905 } else { 906 /* 907 * If the wait structure is NULL and we own the thread, there 908 * is no race (since we are in a critical section). If we 909 * do not own the thread there might be a race but the 910 * target cpu will deal with it. 911 */ 912 #ifdef SMP 913 if (td->td_gd == mygd) { 914 _lwkt_enqueue(td); 915 _lwkt_schedule_post(td, TDPRI_CRIT); 916 } else { 917 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td); 918 } 919 #else 920 _lwkt_enqueue(td); 921 _lwkt_schedule_post(td, TDPRI_CRIT); 922 #endif 923 } 924 } 925 crit_exit_gd(mygd); 926 } 927 928 /* 929 * Managed acquisition. This code assumes that the MP lock is held for 930 * the tdallq operation and that the thread has been descheduled from its 931 * original cpu. We also have to wait for the thread to be entirely switched 932 * out on its original cpu (this is usually fast enough that we never loop) 933 * since the LWKT system does not have to hold the MP lock while switching 934 * and the target may have released it before switching. 935 */ 936 void 937 lwkt_acquire(thread_t td) 938 { 939 globaldata_t gd; 940 globaldata_t mygd; 941 942 gd = td->td_gd; 943 mygd = mycpu; 944 KKASSERT((td->td_flags & TDF_RUNQ) == 0); 945 while (td->td_flags & TDF_RUNNING) /* XXX spin */ 946 cpu_mb1(); 947 if (gd != mygd) { 948 crit_enter_gd(mygd); 949 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */ 950 td->td_gd = mygd; 951 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); /* protected by BGL */ 952 crit_exit_gd(mygd); 953 } 954 } 955 956 /* 957 * Generic deschedule. Descheduling threads other then your own should be 958 * done only in carefully controlled circumstances. Descheduling is 959 * asynchronous. 960 * 961 * This function may block if the cpu has run out of messages. 962 */ 963 void 964 lwkt_deschedule(thread_t td) 965 { 966 crit_enter(); 967 if (td == curthread) { 968 _lwkt_dequeue(td); 969 } else { 970 if (td->td_gd == mycpu) { 971 _lwkt_dequeue(td); 972 } else { 973 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_deschedule, td); 974 } 975 } 976 crit_exit(); 977 } 978 979 /* 980 * Set the target thread's priority. This routine does not automatically 981 * switch to a higher priority thread, LWKT threads are not designed for 982 * continuous priority changes. Yield if you want to switch. 983 * 984 * We have to retain the critical section count which uses the high bits 985 * of the td_pri field. The specified priority may also indicate zero or 986 * more critical sections by adding TDPRI_CRIT*N. 987 * 988 * Note that we requeue the thread whether it winds up on a different runq 989 * or not. uio_yield() depends on this and the routine is not normally 990 * called with the same priority otherwise. 991 */ 992 void 993 lwkt_setpri(thread_t td, int pri) 994 { 995 KKASSERT(pri >= 0); 996 KKASSERT(td->td_gd == mycpu); 997 crit_enter(); 998 if (td->td_flags & TDF_RUNQ) { 999 _lwkt_dequeue(td); 1000 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri; 1001 _lwkt_enqueue(td); 1002 } else { 1003 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri; 1004 } 1005 crit_exit(); 1006 } 1007 1008 void 1009 lwkt_setpri_self(int pri) 1010 { 1011 thread_t td = curthread; 1012 1013 KKASSERT(pri >= 0 && pri <= TDPRI_MAX); 1014 crit_enter(); 1015 if (td->td_flags & TDF_RUNQ) { 1016 _lwkt_dequeue(td); 1017 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri; 1018 _lwkt_enqueue(td); 1019 } else { 1020 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri; 1021 } 1022 crit_exit(); 1023 } 1024 1025 /* 1026 * Migrate the current thread to the specified cpu. The BGL must be held 1027 * (for the gd_tdallq manipulation XXX). This is accomplished by 1028 * descheduling ourselves from the current cpu, moving our thread to the 1029 * tdallq of the target cpu, IPI messaging the target cpu, and switching out. 1030 * TDF_MIGRATING prevents scheduling races while the thread is being migrated. 1031 */ 1032 #ifdef SMP 1033 static void lwkt_setcpu_remote(void *arg); 1034 #endif 1035 1036 void 1037 lwkt_setcpu_self(globaldata_t rgd) 1038 { 1039 #ifdef SMP 1040 thread_t td = curthread; 1041 1042 if (td->td_gd != rgd) { 1043 crit_enter_quick(td); 1044 td->td_flags |= TDF_MIGRATING; 1045 lwkt_deschedule_self(td); 1046 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); /* protected by BGL */ 1047 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); /* protected by BGL */ 1048 lwkt_send_ipiq(rgd, (ipifunc_t)lwkt_setcpu_remote, td); 1049 lwkt_switch(); 1050 /* we are now on the target cpu */ 1051 crit_exit_quick(td); 1052 } 1053 #endif 1054 } 1055 1056 /* 1057 * Remote IPI for cpu migration (called while in a critical section so we 1058 * do not have to enter another one). The thread has already been moved to 1059 * our cpu's allq, but we must wait for the thread to be completely switched 1060 * out on the originating cpu before we schedule it on ours or the stack 1061 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD 1062 * change to main memory. 1063 * 1064 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races 1065 * against wakeups. It is best if this interface is used only when there 1066 * are no pending events that might try to schedule the thread. 1067 */ 1068 #ifdef SMP 1069 static void 1070 lwkt_setcpu_remote(void *arg) 1071 { 1072 thread_t td = arg; 1073 globaldata_t gd = mycpu; 1074 1075 while (td->td_flags & TDF_RUNNING) 1076 cpu_mb1(); 1077 td->td_gd = gd; 1078 cpu_mb2(); 1079 td->td_flags &= ~TDF_MIGRATING; 1080 _lwkt_enqueue(td); 1081 } 1082 #endif 1083 1084 struct proc * 1085 lwkt_preempted_proc(void) 1086 { 1087 thread_t td = curthread; 1088 while (td->td_preempted) 1089 td = td->td_preempted; 1090 return(td->td_proc); 1091 } 1092 1093 /* 1094 * Block on the specified wait queue until signaled. A generation number 1095 * must be supplied to interlock the wait queue. The function will 1096 * return immediately if the generation number does not match the wait 1097 * structure's generation number. 1098 */ 1099 void 1100 lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen) 1101 { 1102 thread_t td = curthread; 1103 lwkt_tokref ilock; 1104 1105 lwkt_gettoken(&ilock, &w->wa_token); 1106 crit_enter(); 1107 if (w->wa_gen == *gen) { 1108 _lwkt_dequeue(td); 1109 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq); 1110 ++w->wa_count; 1111 td->td_wait = w; 1112 td->td_wmesg = wmesg; 1113 again: 1114 lwkt_switch(); 1115 if (td->td_wmesg != NULL) { 1116 _lwkt_dequeue(td); 1117 goto again; 1118 } 1119 } 1120 crit_exit(); 1121 *gen = w->wa_gen; 1122 lwkt_reltoken(&ilock); 1123 } 1124 1125 /* 1126 * Signal a wait queue. We gain ownership of the wait queue in order to 1127 * signal it. Once a thread is removed from the wait queue we have to 1128 * deal with the cpu owning the thread. 1129 * 1130 * Note: alternatively we could message the target cpu owning the wait 1131 * queue. YYY implement as sysctl. 1132 */ 1133 void 1134 lwkt_signal(lwkt_wait_t w, int count) 1135 { 1136 thread_t td; 1137 lwkt_tokref ilock; 1138 1139 lwkt_gettoken(&ilock, &w->wa_token); 1140 ++w->wa_gen; 1141 crit_enter(); 1142 if (count < 0) 1143 count = w->wa_count; 1144 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) { 1145 --count; 1146 --w->wa_count; 1147 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq); 1148 td->td_wait = NULL; 1149 td->td_wmesg = NULL; 1150 if (td->td_gd == mycpu) { 1151 _lwkt_enqueue(td); 1152 } else { 1153 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td); 1154 } 1155 } 1156 crit_exit(); 1157 lwkt_reltoken(&ilock); 1158 } 1159 1160 /* 1161 * Create a kernel process/thread/whatever. It shares it's address space 1162 * with proc0 - ie: kernel only. 1163 * 1164 * NOTE! By default new threads are created with the MP lock held. A 1165 * thread which does not require the MP lock should release it by calling 1166 * rel_mplock() at the start of the new thread. 1167 */ 1168 int 1169 lwkt_create(void (*func)(void *), void *arg, 1170 struct thread **tdp, thread_t template, int tdflags, int cpu, 1171 const char *fmt, ...) 1172 { 1173 thread_t td; 1174 __va_list ap; 1175 1176 td = lwkt_alloc_thread(template, cpu); 1177 if (tdp) 1178 *tdp = td; 1179 cpu_set_thread_handler(td, lwkt_exit, func, arg); 1180 td->td_flags |= TDF_VERBOSE | tdflags; 1181 #ifdef SMP 1182 td->td_mpcount = 1; 1183 #endif 1184 1185 /* 1186 * Set up arg0 for 'ps' etc 1187 */ 1188 __va_start(ap, fmt); 1189 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap); 1190 __va_end(ap); 1191 1192 /* 1193 * Schedule the thread to run 1194 */ 1195 if ((td->td_flags & TDF_STOPREQ) == 0) 1196 lwkt_schedule(td); 1197 else 1198 td->td_flags &= ~TDF_STOPREQ; 1199 return 0; 1200 } 1201 1202 /* 1203 * kthread_* is specific to the kernel and is not needed by userland. 1204 */ 1205 #ifdef _KERNEL 1206 1207 /* 1208 * Destroy an LWKT thread. Warning! This function is not called when 1209 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and 1210 * uses a different reaping mechanism. 1211 */ 1212 void 1213 lwkt_exit(void) 1214 { 1215 thread_t td = curthread; 1216 globaldata_t gd; 1217 1218 if (td->td_flags & TDF_VERBOSE) 1219 printf("kthread %p %s has exited\n", td, td->td_comm); 1220 caps_exit(td); 1221 crit_enter_quick(td); 1222 lwkt_deschedule_self(td); 1223 gd = mycpu; 1224 KKASSERT(gd == td->td_gd); 1225 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); 1226 if (td->td_flags & TDF_ALLOCATED_THREAD) { 1227 ++gd->gd_tdfreecount; 1228 TAILQ_INSERT_TAIL(&gd->gd_tdfreeq, td, td_threadq); 1229 } 1230 cpu_thread_exit(); 1231 } 1232 1233 /* 1234 * Create a kernel process/thread/whatever. It shares it's address space 1235 * with proc0 - ie: kernel only. 5.x compatible. 1236 * 1237 * NOTE! By default kthreads are created with the MP lock held. A 1238 * thread which does not require the MP lock should release it by calling 1239 * rel_mplock() at the start of the new thread. 1240 */ 1241 int 1242 kthread_create(void (*func)(void *), void *arg, 1243 struct thread **tdp, const char *fmt, ...) 1244 { 1245 thread_t td; 1246 __va_list ap; 1247 1248 td = lwkt_alloc_thread(NULL, -1); 1249 if (tdp) 1250 *tdp = td; 1251 cpu_set_thread_handler(td, kthread_exit, func, arg); 1252 td->td_flags |= TDF_VERBOSE; 1253 #ifdef SMP 1254 td->td_mpcount = 1; 1255 #endif 1256 1257 /* 1258 * Set up arg0 for 'ps' etc 1259 */ 1260 __va_start(ap, fmt); 1261 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap); 1262 __va_end(ap); 1263 1264 /* 1265 * Schedule the thread to run 1266 */ 1267 lwkt_schedule(td); 1268 return 0; 1269 } 1270 1271 /* 1272 * Destroy an LWKT thread. Warning! This function is not called when 1273 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and 1274 * uses a different reaping mechanism. 1275 * 1276 * XXX duplicates lwkt_exit() 1277 */ 1278 void 1279 kthread_exit(void) 1280 { 1281 lwkt_exit(); 1282 } 1283 1284 #endif /* _KERNEL */ 1285 1286 void 1287 crit_panic(void) 1288 { 1289 thread_t td = curthread; 1290 int lpri = td->td_pri; 1291 1292 td->td_pri = 0; 1293 panic("td_pri is/would-go negative! %p %d", td, lpri); 1294 } 1295 1296