| 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 1.90 2005/12/10 18:50:36 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 | #else |
| 74 | |
| 75 | #include <sys/stdint.h> |
| 76 | #include <libcaps/thread.h> |
| 77 | #include <sys/thread.h> |
| 78 | #include <sys/msgport.h> |
| 79 | #include <sys/errno.h> |
| 80 | #include <libcaps/globaldata.h> |
| 81 | #include <machine/cpufunc.h> |
| 82 | #include <sys/thread2.h> |
| 83 | #include <sys/msgport2.h> |
| 84 | #include <stdio.h> |
| 85 | #include <stdlib.h> |
| 86 | #include <string.h> |
| 87 | #include <machine/lock.h> |
| 88 | #include <machine/atomic.h> |
| 89 | #include <machine/cpu.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 | static __int64_t token_contention_count = 0; |
| 102 | static __int64_t mplock_contention_count = 0; |
| 103 | |
| 104 | #ifdef _KERNEL |
| 105 | |
| 106 | SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, ""); |
| 107 | #ifdef INVARIANTS |
| 108 | SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, ""); |
| 109 | #endif |
| 110 | SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, ""); |
| 111 | SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, ""); |
| 112 | SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, ""); |
| 113 | SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, ""); |
| 114 | #ifdef INVARIANTS |
| 115 | SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW, |
| 116 | &token_contention_count, 0, "spinning due to token contention"); |
| 117 | SYSCTL_QUAD(_lwkt, OID_AUTO, mplock_contention_count, CTLFLAG_RW, |
| 118 | &mplock_contention_count, 0, "spinning due to MPLOCK contention"); |
| 119 | #endif |
| 120 | #endif |
| 121 | |
| 122 | /* |
| 123 | * These helper procedures handle the runq, they can only be called from |
| 124 | * within a critical section. |
| 125 | * |
| 126 | * WARNING! Prior to SMP being brought up it is possible to enqueue and |
| 127 | * dequeue threads belonging to other cpus, so be sure to use td->td_gd |
| 128 | * instead of 'mycpu' when referencing the globaldata structure. Once |
| 129 | * SMP live enqueuing and dequeueing only occurs on the current cpu. |
| 130 | */ |
| 131 | static __inline |
| 132 | void |
| 133 | _lwkt_dequeue(thread_t td) |
| 134 | { |
| 135 | if (td->td_flags & TDF_RUNQ) { |
| 136 | int nq = td->td_pri & TDPRI_MASK; |
| 137 | struct globaldata *gd = td->td_gd; |
| 138 | |
| 139 | td->td_flags &= ~TDF_RUNQ; |
| 140 | TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq); |
| 141 | /* runqmask is passively cleaned up by the switcher */ |
| 142 | } |
| 143 | } |
| 144 | |
| 145 | static __inline |
| 146 | void |
| 147 | _lwkt_enqueue(thread_t td) |
| 148 | { |
| 149 | if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_TSLEEPQ|TDF_BLOCKQ)) == 0) { |
| 150 | int nq = td->td_pri & TDPRI_MASK; |
| 151 | struct globaldata *gd = td->td_gd; |
| 152 | |
| 153 | td->td_flags |= TDF_RUNQ; |
| 154 | TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq); |
| 155 | gd->gd_runqmask |= 1 << nq; |
| 156 | } |
| 157 | } |
| 158 | |
| 159 | /* |
| 160 | * Schedule a thread to run. As the current thread we can always safely |
| 161 | * schedule ourselves, and a shortcut procedure is provided for that |
| 162 | * function. |
| 163 | * |
| 164 | * (non-blocking, self contained on a per cpu basis) |
| 165 | */ |
| 166 | void |
| 167 | lwkt_schedule_self(thread_t td) |
| 168 | { |
| 169 | crit_enter_quick(td); |
| 170 | KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!")); |
| 171 | KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!")); |
| 172 | KKASSERT(td->td_proc == NULL || (td->td_proc->p_flag & P_ONRUNQ) == 0); |
| 173 | _lwkt_enqueue(td); |
| 174 | crit_exit_quick(td); |
| 175 | } |
| 176 | |
| 177 | /* |
| 178 | * Deschedule a thread. |
| 179 | * |
| 180 | * (non-blocking, self contained on a per cpu basis) |
| 181 | */ |
| 182 | void |
| 183 | lwkt_deschedule_self(thread_t td) |
| 184 | { |
| 185 | crit_enter_quick(td); |
| 186 | KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!")); |
| 187 | _lwkt_dequeue(td); |
| 188 | crit_exit_quick(td); |
| 189 | } |
| 190 | |
| 191 | #ifdef _KERNEL |
| 192 | |
| 193 | /* |
| 194 | * LWKTs operate on a per-cpu basis |
| 195 | * |
| 196 | * WARNING! Called from early boot, 'mycpu' may not work yet. |
| 197 | */ |
| 198 | void |
| 199 | lwkt_gdinit(struct globaldata *gd) |
| 200 | { |
| 201 | int i; |
| 202 | |
| 203 | for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i) |
| 204 | TAILQ_INIT(&gd->gd_tdrunq[i]); |
| 205 | gd->gd_runqmask = 0; |
| 206 | TAILQ_INIT(&gd->gd_tdallq); |
| 207 | } |
| 208 | |
| 209 | #endif /* _KERNEL */ |
| 210 | |
| 211 | /* |
| 212 | * Initialize a thread wait structure prior to first use. |
| 213 | * |
| 214 | * NOTE! called from low level boot code, we cannot do anything fancy! |
| 215 | */ |
| 216 | void |
| 217 | lwkt_wait_init(lwkt_wait_t w) |
| 218 | { |
| 219 | lwkt_token_init(&w->wa_token); |
| 220 | TAILQ_INIT(&w->wa_waitq); |
| 221 | w->wa_gen = 0; |
| 222 | w->wa_count = 0; |
| 223 | } |
| 224 | |
| 225 | /* |
| 226 | * Create a new thread. The thread must be associated with a process context |
| 227 | * or LWKT start address before it can be scheduled. If the target cpu is |
| 228 | * -1 the thread will be created on the current cpu. |
| 229 | * |
| 230 | * If you intend to create a thread without a process context this function |
| 231 | * does everything except load the startup and switcher function. |
| 232 | */ |
| 233 | thread_t |
| 234 | lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags) |
| 235 | { |
| 236 | void *stack; |
| 237 | globaldata_t gd = mycpu; |
| 238 | |
| 239 | if (td == NULL) { |
| 240 | crit_enter_gd(gd); |
| 241 | if (gd->gd_tdfreecount > 0) { |
| 242 | --gd->gd_tdfreecount; |
| 243 | td = TAILQ_FIRST(&gd->gd_tdfreeq); |
| 244 | KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0, |
| 245 | ("lwkt_alloc_thread: unexpected NULL or corrupted td")); |
| 246 | TAILQ_REMOVE(&gd->gd_tdfreeq, td, td_threadq); |
| 247 | crit_exit_gd(gd); |
| 248 | flags |= td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD); |
| 249 | } else { |
| 250 | crit_exit_gd(gd); |
| 251 | #ifdef _KERNEL |
| 252 | td = zalloc(thread_zone); |
| 253 | #else |
| 254 | td = malloc(sizeof(struct thread)); |
| 255 | #endif |
| 256 | td->td_kstack = NULL; |
| 257 | td->td_kstack_size = 0; |
| 258 | flags |= TDF_ALLOCATED_THREAD; |
| 259 | } |
| 260 | } |
| 261 | if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) { |
| 262 | if (flags & TDF_ALLOCATED_STACK) { |
| 263 | #ifdef _KERNEL |
| 264 | kmem_free(kernel_map, (vm_offset_t)stack, td->td_kstack_size); |
| 265 | #else |
| 266 | libcaps_free_stack(stack, td->td_kstack_size); |
| 267 | #endif |
| 268 | stack = NULL; |
| 269 | } |
| 270 | } |
| 271 | if (stack == NULL) { |
| 272 | #ifdef _KERNEL |
| 273 | stack = (void *)kmem_alloc(kernel_map, stksize); |
| 274 | #else |
| 275 | stack = libcaps_alloc_stack(stksize); |
| 276 | #endif |
| 277 | flags |= TDF_ALLOCATED_STACK; |
| 278 | } |
| 279 | if (cpu < 0) |
| 280 | lwkt_init_thread(td, stack, stksize, flags, mycpu); |
| 281 | else |
| 282 | lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu)); |
| 283 | return(td); |
| 284 | } |
| 285 | |
| 286 | #ifdef _KERNEL |
| 287 | |
| 288 | /* |
| 289 | * Initialize a preexisting thread structure. This function is used by |
| 290 | * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread. |
| 291 | * |
| 292 | * All threads start out in a critical section at a priority of |
| 293 | * TDPRI_KERN_DAEMON. Higher level code will modify the priority as |
| 294 | * appropriate. This function may send an IPI message when the |
| 295 | * requested cpu is not the current cpu and consequently gd_tdallq may |
| 296 | * not be initialized synchronously from the point of view of the originating |
| 297 | * cpu. |
| 298 | * |
| 299 | * NOTE! we have to be careful in regards to creating threads for other cpus |
| 300 | * if SMP has not yet been activated. |
| 301 | */ |
| 302 | #ifdef SMP |
| 303 | |
| 304 | static void |
| 305 | lwkt_init_thread_remote(void *arg) |
| 306 | { |
| 307 | thread_t td = arg; |
| 308 | |
| 309 | TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq); |
| 310 | } |
| 311 | |
| 312 | #endif |
| 313 | |
| 314 | void |
| 315 | lwkt_init_thread(thread_t td, void *stack, int stksize, int flags, |
| 316 | struct globaldata *gd) |
| 317 | { |
| 318 | globaldata_t mygd = mycpu; |
| 319 | |
| 320 | bzero(td, sizeof(struct thread)); |
| 321 | td->td_kstack = stack; |
| 322 | td->td_kstack_size = stksize; |
| 323 | td->td_flags = flags; |
| 324 | td->td_gd = gd; |
| 325 | td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT; |
| 326 | #ifdef SMP |
| 327 | if ((flags & TDF_MPSAFE) == 0) |
| 328 | td->td_mpcount = 1; |
| 329 | #endif |
| 330 | lwkt_initport(&td->td_msgport, td); |
| 331 | pmap_init_thread(td); |
| 332 | #ifdef SMP |
| 333 | /* |
| 334 | * Normally initializing a thread for a remote cpu requires sending an |
| 335 | * IPI. However, the idlethread is setup before the other cpus are |
| 336 | * activated so we have to treat it as a special case. XXX manipulation |
| 337 | * of gd_tdallq requires the BGL. |
| 338 | */ |
| 339 | if (gd == mygd || td == &gd->gd_idlethread) { |
| 340 | crit_enter_gd(mygd); |
| 341 | TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); |
| 342 | crit_exit_gd(mygd); |
| 343 | } else { |
| 344 | lwkt_send_ipiq(gd, lwkt_init_thread_remote, td); |
| 345 | } |
| 346 | #else |
| 347 | crit_enter_gd(mygd); |
| 348 | TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); |
| 349 | crit_exit_gd(mygd); |
| 350 | #endif |
| 351 | } |
| 352 | |
| 353 | #endif /* _KERNEL */ |
| 354 | |
| 355 | void |
| 356 | lwkt_set_comm(thread_t td, const char *ctl, ...) |
| 357 | { |
| 358 | __va_list va; |
| 359 | |
| 360 | __va_start(va, ctl); |
| 361 | vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va); |
| 362 | __va_end(va); |
| 363 | } |
| 364 | |
| 365 | void |
| 366 | lwkt_hold(thread_t td) |
| 367 | { |
| 368 | ++td->td_refs; |
| 369 | } |
| 370 | |
| 371 | void |
| 372 | lwkt_rele(thread_t td) |
| 373 | { |
| 374 | KKASSERT(td->td_refs > 0); |
| 375 | --td->td_refs; |
| 376 | } |
| 377 | |
| 378 | #ifdef _KERNEL |
| 379 | |
| 380 | void |
| 381 | lwkt_wait_free(thread_t td) |
| 382 | { |
| 383 | while (td->td_refs) |
| 384 | tsleep(td, 0, "tdreap", hz); |
| 385 | } |
| 386 | |
| 387 | #endif |
| 388 | |
| 389 | void |
| 390 | lwkt_free_thread(thread_t td) |
| 391 | { |
| 392 | struct globaldata *gd = mycpu; |
| 393 | |
| 394 | KASSERT((td->td_flags & TDF_RUNNING) == 0, |
| 395 | ("lwkt_free_thread: did not exit! %p", td)); |
| 396 | |
| 397 | crit_enter_gd(gd); |
| 398 | TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); |
| 399 | if (gd->gd_tdfreecount < CACHE_NTHREADS && |
| 400 | (td->td_flags & TDF_ALLOCATED_THREAD) |
| 401 | ) { |
| 402 | ++gd->gd_tdfreecount; |
| 403 | TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq); |
| 404 | crit_exit_gd(gd); |
| 405 | } else { |
| 406 | crit_exit_gd(gd); |
| 407 | if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) { |
| 408 | #ifdef _KERNEL |
| 409 | kmem_free(kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size); |
| 410 | #else |
| 411 | libcaps_free_stack(td->td_kstack, td->td_kstack_size); |
| 412 | #endif |
| 413 | /* gd invalid */ |
| 414 | td->td_kstack = NULL; |
| 415 | td->td_kstack_size = 0; |
| 416 | } |
| 417 | if (td->td_flags & TDF_ALLOCATED_THREAD) { |
| 418 | #ifdef _KERNEL |
| 419 | zfree(thread_zone, td); |
| 420 | #else |
| 421 | free(td); |
| 422 | #endif |
| 423 | } |
| 424 | } |
| 425 | } |
| 426 | |
| 427 | |
| 428 | /* |
| 429 | * Switch to the next runnable lwkt. If no LWKTs are runnable then |
| 430 | * switch to the idlethread. Switching must occur within a critical |
| 431 | * section to avoid races with the scheduling queue. |
| 432 | * |
| 433 | * We always have full control over our cpu's run queue. Other cpus |
| 434 | * that wish to manipulate our queue must use the cpu_*msg() calls to |
| 435 | * talk to our cpu, so a critical section is all that is needed and |
| 436 | * the result is very, very fast thread switching. |
| 437 | * |
| 438 | * The LWKT scheduler uses a fixed priority model and round-robins at |
| 439 | * each priority level. User process scheduling is a totally |
| 440 | * different beast and LWKT priorities should not be confused with |
| 441 | * user process priorities. |
| 442 | * |
| 443 | * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch() |
| 444 | * cleans it up. Note that the td_switch() function cannot do anything that |
| 445 | * requires the MP lock since the MP lock will have already been setup for |
| 446 | * the target thread (not the current thread). It's nice to have a scheduler |
| 447 | * that does not need the MP lock to work because it allows us to do some |
| 448 | * really cool high-performance MP lock optimizations. |
| 449 | * |
| 450 | * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch() |
| 451 | * is not called by the current thread in the preemption case, only when |
| 452 | * the preempting thread blocks (in order to return to the original thread). |
| 453 | */ |
| 454 | void |
| 455 | lwkt_switch(void) |
| 456 | { |
| 457 | globaldata_t gd = mycpu; |
| 458 | thread_t td = gd->gd_curthread; |
| 459 | thread_t ntd; |
| 460 | #ifdef SMP |
| 461 | int mpheld; |
| 462 | #endif |
| 463 | |
| 464 | /* |
| 465 | * We had better not be holding any spin locks. |
| 466 | */ |
| 467 | KKASSERT(td->td_spinlocks == 0); |
| 468 | |
| 469 | /* |
| 470 | * Switching from within a 'fast' (non thread switched) interrupt or IPI |
| 471 | * is illegal. However, we may have to do it anyway if we hit a fatal |
| 472 | * kernel trap or we have paniced. |
| 473 | * |
| 474 | * If this case occurs save and restore the interrupt nesting level. |
| 475 | */ |
| 476 | if (gd->gd_intr_nesting_level) { |
| 477 | int savegdnest; |
| 478 | int savegdtrap; |
| 479 | |
| 480 | if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) { |
| 481 | panic("lwkt_switch: cannot switch from within " |
| 482 | "a fast interrupt, yet, td %p\n", td); |
| 483 | } else { |
| 484 | savegdnest = gd->gd_intr_nesting_level; |
| 485 | savegdtrap = gd->gd_trap_nesting_level; |
| 486 | gd->gd_intr_nesting_level = 0; |
| 487 | gd->gd_trap_nesting_level = 0; |
| 488 | if ((td->td_flags & TDF_PANICWARN) == 0) { |
| 489 | td->td_flags |= TDF_PANICWARN; |
| 490 | printf("Warning: thread switch from interrupt or IPI, " |
| 491 | "thread %p (%s)\n", td, td->td_comm); |
| 492 | #ifdef DDB |
| 493 | db_print_backtrace(); |
| 494 | #endif |
| 495 | } |
| 496 | lwkt_switch(); |
| 497 | gd->gd_intr_nesting_level = savegdnest; |
| 498 | gd->gd_trap_nesting_level = savegdtrap; |
| 499 | return; |
| 500 | } |
| 501 | } |
| 502 | |
| 503 | /* |
| 504 | * Passive release (used to transition from user to kernel mode |
| 505 | * when we block or switch rather then when we enter the kernel). |
| 506 | * This function is NOT called if we are switching into a preemption |
| 507 | * or returning from a preemption. Typically this causes us to lose |
| 508 | * our current process designation (if we have one) and become a true |
| 509 | * LWKT thread, and may also hand the current process designation to |
| 510 | * another process and schedule thread. |
| 511 | */ |
| 512 | if (td->td_release) |
| 513 | td->td_release(td); |
| 514 | |
| 515 | crit_enter_gd(gd); |
| 516 | |
| 517 | #ifdef SMP |
| 518 | /* |
| 519 | * td_mpcount cannot be used to determine if we currently hold the |
| 520 | * MP lock because get_mplock() will increment it prior to attempting |
| 521 | * to get the lock, and switch out if it can't. Our ownership of |
| 522 | * the actual lock will remain stable while we are in a critical section |
| 523 | * (but, of course, another cpu may own or release the lock so the |
| 524 | * actual value of mp_lock is not stable). |
| 525 | */ |
| 526 | mpheld = MP_LOCK_HELD(); |
| 527 | #ifdef INVARIANTS |
| 528 | if (td->td_cscount) { |
| 529 | printf("Diagnostic: attempt to switch while mastering cpusync: %p\n", |
| 530 | td); |
| 531 | if (panic_on_cscount) |
| 532 | panic("switching while mastering cpusync"); |
| 533 | } |
| 534 | #endif |
| 535 | #endif |
| 536 | if ((ntd = td->td_preempted) != NULL) { |
| 537 | /* |
| 538 | * We had preempted another thread on this cpu, resume the preempted |
| 539 | * thread. This occurs transparently, whether the preempted thread |
| 540 | * was scheduled or not (it may have been preempted after descheduling |
| 541 | * itself). |
| 542 | * |
| 543 | * We have to setup the MP lock for the original thread after backing |
| 544 | * out the adjustment that was made to curthread when the original |
| 545 | * was preempted. |
| 546 | */ |
| 547 | KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK); |
| 548 | #ifdef SMP |
| 549 | if (ntd->td_mpcount && mpheld == 0) { |
| 550 | panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d", |
| 551 | td, ntd, td->td_mpcount, ntd->td_mpcount); |
| 552 | } |
| 553 | if (ntd->td_mpcount) { |
| 554 | td->td_mpcount -= ntd->td_mpcount; |
| 555 | KKASSERT(td->td_mpcount >= 0); |
| 556 | } |
| 557 | #endif |
| 558 | ntd->td_flags |= TDF_PREEMPT_DONE; |
| 559 | |
| 560 | /* |
| 561 | * XXX. The interrupt may have woken a thread up, we need to properly |
| 562 | * set the reschedule flag if the originally interrupted thread is at |
| 563 | * a lower priority. |
| 564 | */ |
| 565 | if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1) |
| 566 | need_lwkt_resched(); |
| 567 | /* YYY release mp lock on switchback if original doesn't need it */ |
| 568 | } else { |
| 569 | /* |
| 570 | * Priority queue / round-robin at each priority. Note that user |
| 571 | * processes run at a fixed, low priority and the user process |
| 572 | * scheduler deals with interactions between user processes |
| 573 | * by scheduling and descheduling them from the LWKT queue as |
| 574 | * necessary. |
| 575 | * |
| 576 | * We have to adjust the MP lock for the target thread. If we |
| 577 | * need the MP lock and cannot obtain it we try to locate a |
| 578 | * thread that does not need the MP lock. If we cannot, we spin |
| 579 | * instead of HLT. |
| 580 | * |
| 581 | * A similar issue exists for the tokens held by the target thread. |
| 582 | * If we cannot obtain ownership of the tokens we cannot immediately |
| 583 | * schedule the thread. |
| 584 | */ |
| 585 | |
| 586 | /* |
| 587 | * We are switching threads. If there are any pending requests for |
| 588 | * tokens we can satisfy all of them here. |
| 589 | */ |
| 590 | #ifdef SMP |
| 591 | if (gd->gd_tokreqbase) |
| 592 | lwkt_drain_token_requests(); |
| 593 | #endif |
| 594 | |
| 595 | /* |
| 596 | * If an LWKT reschedule was requested, well that is what we are |
| 597 | * doing now so clear it. |
| 598 | */ |
| 599 | clear_lwkt_resched(); |
| 600 | again: |
| 601 | if (gd->gd_runqmask) { |
| 602 | int nq = bsrl(gd->gd_runqmask); |
| 603 | if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) { |
| 604 | gd->gd_runqmask &= ~(1 << nq); |
| 605 | goto again; |
| 606 | } |
| 607 | #ifdef SMP |
| 608 | /* |
| 609 | * THREAD SELECTION FOR AN SMP MACHINE BUILD |
| 610 | * |
| 611 | * If the target needs the MP lock and we couldn't get it, |
| 612 | * or if the target is holding tokens and we could not |
| 613 | * gain ownership of the tokens, continue looking for a |
| 614 | * thread to schedule and spin instead of HLT if we can't. |
| 615 | * |
| 616 | * NOTE: the mpheld variable invalid after this conditional, it |
| 617 | * can change due to both cpu_try_mplock() returning success |
| 618 | * AND interactions in lwkt_chktokens() due to the fact that |
| 619 | * we are trying to check the mpcount of a thread other then |
| 620 | * the current thread. Because of this, if the current thread |
| 621 | * is not holding td_mpcount, an IPI indirectly run via |
| 622 | * lwkt_chktokens() can obtain and release the MP lock and |
| 623 | * cause the core MP lock to be released. |
| 624 | */ |
| 625 | if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) || |
| 626 | (ntd->td_toks && lwkt_chktokens(ntd) == 0) |
| 627 | ) { |
| 628 | u_int32_t rqmask = gd->gd_runqmask; |
| 629 | |
| 630 | mpheld = MP_LOCK_HELD(); |
| 631 | ntd = NULL; |
| 632 | while (rqmask) { |
| 633 | TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) { |
| 634 | if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) { |
| 635 | /* spinning due to MP lock being held */ |
| 636 | #ifdef INVARIANTS |
| 637 | ++mplock_contention_count; |
| 638 | #endif |
| 639 | /* mplock still not held, 'mpheld' still valid */ |
| 640 | continue; |
| 641 | } |
| 642 | |
| 643 | /* |
| 644 | * mpheld state invalid after chktokens call returns |
| 645 | * failure, but the variable is only needed for |
| 646 | * the loop. |
| 647 | */ |
| 648 | if (ntd->td_toks && !lwkt_chktokens(ntd)) { |
| 649 | /* spinning due to token contention */ |
| 650 | #ifdef INVARIANTS |
| 651 | ++token_contention_count; |
| 652 | #endif |
| 653 | mpheld = MP_LOCK_HELD(); |
| 654 | continue; |
| 655 | } |
| 656 | break; |
| 657 | } |
| 658 | if (ntd) |
| 659 | break; |
| 660 | rqmask &= ~(1 << nq); |
| 661 | nq = bsrl(rqmask); |
| 662 | } |
| 663 | if (ntd == NULL) { |
| 664 | ntd = &gd->gd_idlethread; |
| 665 | ntd->td_flags |= TDF_IDLE_NOHLT; |
| 666 | goto using_idle_thread; |
| 667 | } else { |
| 668 | ++gd->gd_cnt.v_swtch; |
| 669 | TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq); |
| 670 | TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq); |
| 671 | } |
| 672 | } else { |
| 673 | ++gd->gd_cnt.v_swtch; |
| 674 | TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq); |
| 675 | TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq); |
| 676 | } |
| 677 | #else |
| 678 | /* |
| 679 | * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to |
| 680 | * worry about tokens or the BGL. |
| 681 | */ |
| 682 | ++gd->gd_cnt.v_swtch; |
| 683 | TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq); |
| 684 | TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq); |
| 685 | #endif |
| 686 | } else { |
| 687 | /* |
| 688 | * We have nothing to run but only let the idle loop halt |
| 689 | * the cpu if there are no pending interrupts. |
| 690 | */ |
| 691 | ntd = &gd->gd_idlethread; |
| 692 | if (gd->gd_reqflags & RQF_IDLECHECK_MASK) |
| 693 | ntd->td_flags |= TDF_IDLE_NOHLT; |
| 694 | #ifdef SMP |
| 695 | using_idle_thread: |
| 696 | /* |
| 697 | * The idle thread should not be holding the MP lock unless we |
| 698 | * are trapping in the kernel or in a panic. Since we select the |
| 699 | * idle thread unconditionally when no other thread is available, |
| 700 | * if the MP lock is desired during a panic or kernel trap, we |
| 701 | * have to loop in the scheduler until we get it. |
| 702 | */ |
| 703 | if (ntd->td_mpcount) { |
| 704 | mpheld = MP_LOCK_HELD(); |
| 705 | if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) |
| 706 | panic("Idle thread %p was holding the BGL!", ntd); |
| 707 | else if (mpheld == 0) |
| 708 | goto again; |
| 709 | } |
| 710 | #endif |
| 711 | } |
| 712 | } |
| 713 | KASSERT(ntd->td_pri >= TDPRI_CRIT, |
| 714 | ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri)); |
| 715 | |
| 716 | /* |
| 717 | * Do the actual switch. If the new target does not need the MP lock |
| 718 | * and we are holding it, release the MP lock. If the new target requires |
| 719 | * the MP lock we have already acquired it for the target. |
| 720 | */ |
| 721 | #ifdef SMP |
| 722 | if (ntd->td_mpcount == 0 ) { |
| 723 | if (MP_LOCK_HELD()) |
| 724 | cpu_rel_mplock(); |
| 725 | } else { |
| 726 | ASSERT_MP_LOCK_HELD(ntd); |
| 727 | } |
| 728 | #endif |
| 729 | if (td != ntd) { |
| 730 | ++switch_count; |
| 731 | td->td_switch(ntd); |
| 732 | } |
| 733 | /* NOTE: current cpu may have changed after switch */ |
| 734 | crit_exit_quick(td); |
| 735 | } |
| 736 | |
| 737 | /* |
| 738 | * Request that the target thread preempt the current thread. Preemption |
| 739 | * only works under a specific set of conditions: |
| 740 | * |
| 741 | * - We are not preempting ourselves |
| 742 | * - The target thread is owned by the current cpu |
| 743 | * - We are not currently being preempted |
| 744 | * - The target is not currently being preempted |
| 745 | * - We are able to satisfy the target's MP lock requirements (if any). |
| 746 | * |
| 747 | * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically |
| 748 | * this is called via lwkt_schedule() through the td_preemptable callback. |
| 749 | * critpri is the managed critical priority that we should ignore in order |
| 750 | * to determine whether preemption is possible (aka usually just the crit |
| 751 | * priority of lwkt_schedule() itself). |
| 752 | * |
| 753 | * XXX at the moment we run the target thread in a critical section during |
| 754 | * the preemption in order to prevent the target from taking interrupts |
| 755 | * that *WE* can't. Preemption is strictly limited to interrupt threads |
| 756 | * and interrupt-like threads, outside of a critical section, and the |
| 757 | * preempted source thread will be resumed the instant the target blocks |
| 758 | * whether or not the source is scheduled (i.e. preemption is supposed to |
| 759 | * be as transparent as possible). |
| 760 | * |
| 761 | * The target thread inherits our MP count (added to its own) for the |
| 762 | * duration of the preemption in order to preserve the atomicy of the |
| 763 | * MP lock during the preemption. Therefore, any preempting targets must be |
| 764 | * careful in regards to MP assertions. Note that the MP count may be |
| 765 | * out of sync with the physical mp_lock, but we do not have to preserve |
| 766 | * the original ownership of the lock if it was out of synch (that is, we |
| 767 | * can leave it synchronized on return). |
| 768 | */ |
| 769 | void |
| 770 | lwkt_preempt(thread_t ntd, int critpri) |
| 771 | { |
| 772 | struct globaldata *gd = mycpu; |
| 773 | thread_t td; |
| 774 | #ifdef SMP |
| 775 | int mpheld; |
| 776 | int savecnt; |
| 777 | #endif |
| 778 | |
| 779 | /* |
| 780 | * The caller has put us in a critical section. We can only preempt |
| 781 | * if the caller of the caller was not in a critical section (basically |
| 782 | * a local interrupt), as determined by the 'critpri' parameter. |
| 783 | * |
| 784 | * YYY The target thread must be in a critical section (else it must |
| 785 | * inherit our critical section? I dunno yet). |
| 786 | * |
| 787 | * Any tokens held by the target may not be held by thread(s) being |
| 788 | * preempted. We take the easy way out and do not preempt if |
| 789 | * the target is holding tokens. |
| 790 | * |
| 791 | * Set need_lwkt_resched() unconditionally for now YYY. |
| 792 | */ |
| 793 | KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri)); |
| 794 | |
| 795 | td = gd->gd_curthread; |
| 796 | if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) { |
| 797 | ++preempt_miss; |
| 798 | return; |
| 799 | } |
| 800 | if ((td->td_pri & ~TDPRI_MASK) > critpri) { |
| 801 | ++preempt_miss; |
| 802 | need_lwkt_resched(); |
| 803 | return; |
| 804 | } |
| 805 | #ifdef SMP |
| 806 | if (ntd->td_gd != gd) { |
| 807 | ++preempt_miss; |
| 808 | need_lwkt_resched(); |
| 809 | return; |
| 810 | } |
| 811 | #endif |
| 812 | /* |
| 813 | * Take the easy way out and do not preempt if the target is holding |
| 814 | * one or more tokens. We could test whether the thread(s) being |
| 815 | * preempted interlock against the target thread's tokens and whether |
| 816 | * we can get all the target thread's tokens, but this situation |
| 817 | * should not occur very often so its easier to simply not preempt. |
| 818 | */ |
| 819 | if (ntd->td_toks != NULL) { |
| 820 | ++preempt_miss; |
| 821 | need_lwkt_resched(); |
| 822 | return; |
| 823 | } |
| 824 | if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) { |
| 825 | ++preempt_weird; |
| 826 | need_lwkt_resched(); |
| 827 | return; |
| 828 | } |
| 829 | if (ntd->td_preempted) { |
| 830 | ++preempt_hit; |
| 831 | need_lwkt_resched(); |
| 832 | return; |
| 833 | } |
| 834 | #ifdef SMP |
| 835 | /* |
| 836 | * note: an interrupt might have occured just as we were transitioning |
| 837 | * to or from the MP lock. In this case td_mpcount will be pre-disposed |
| 838 | * (non-zero) but not actually synchronized with the actual state of the |
| 839 | * lock. We can use it to imply an MP lock requirement for the |
| 840 | * preemption but we cannot use it to test whether we hold the MP lock |
| 841 | * or not. |
| 842 | */ |
| 843 | savecnt = td->td_mpcount; |
| 844 | mpheld = MP_LOCK_HELD(); |
| 845 | ntd->td_mpcount += td->td_mpcount; |
| 846 | if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) { |
| 847 | ntd->td_mpcount -= td->td_mpcount; |
| 848 | ++preempt_miss; |
| 849 | need_lwkt_resched(); |
| 850 | return; |
| 851 | } |
| 852 | #endif |
| 853 | |
| 854 | /* |
| 855 | * Since we are able to preempt the current thread, there is no need to |
| 856 | * call need_lwkt_resched(). |
| 857 | */ |
| 858 | ++preempt_hit; |
| 859 | ntd->td_preempted = td; |
| 860 | td->td_flags |= TDF_PREEMPT_LOCK; |
| 861 | td->td_switch(ntd); |
| 862 | KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE)); |
| 863 | #ifdef SMP |
| 864 | KKASSERT(savecnt == td->td_mpcount); |
| 865 | mpheld = MP_LOCK_HELD(); |
| 866 | if (mpheld && td->td_mpcount == 0) |
| 867 | cpu_rel_mplock(); |
| 868 | else if (mpheld == 0 && td->td_mpcount) |
| 869 | panic("lwkt_preempt(): MP lock was not held through"); |
| 870 | #endif |
| 871 | ntd->td_preempted = NULL; |
| 872 | td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE); |
| 873 | } |
| 874 | |
| 875 | /* |
| 876 | * Yield our thread while higher priority threads are pending. This is |
| 877 | * typically called when we leave a critical section but it can be safely |
| 878 | * called while we are in a critical section. |
| 879 | * |
| 880 | * This function will not generally yield to equal priority threads but it |
| 881 | * can occur as a side effect. Note that lwkt_switch() is called from |
| 882 | * inside the critical section to prevent its own crit_exit() from reentering |
| 883 | * lwkt_yield_quick(). |
| 884 | * |
| 885 | * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint |
| 886 | * came along but was blocked and made pending. |
| 887 | * |
| 888 | * (self contained on a per cpu basis) |
| 889 | */ |
| 890 | void |
| 891 | lwkt_yield_quick(void) |
| 892 | { |
| 893 | globaldata_t gd = mycpu; |
| 894 | thread_t td = gd->gd_curthread; |
| 895 | |
| 896 | /* |
| 897 | * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear |
| 898 | * it with a non-zero cpl then we might not wind up calling splz after |
| 899 | * a task switch when the critical section is exited even though the |
| 900 | * new task could accept the interrupt. |
| 901 | * |
| 902 | * XXX from crit_exit() only called after last crit section is released. |
| 903 | * If called directly will run splz() even if in a critical section. |
| 904 | * |
| 905 | * td_nest_count prevent deep nesting via splz() or doreti(). Note that |
| 906 | * except for this special case, we MUST call splz() here to handle any |
| 907 | * pending ints, particularly after we switch, or we might accidently |
| 908 | * halt the cpu with interrupts pending. |
| 909 | */ |
| 910 | if (gd->gd_reqflags && td->td_nest_count < 2) |
| 911 | splz(); |
| 912 | |
| 913 | /* |
| 914 | * YYY enabling will cause wakeup() to task-switch, which really |
| 915 | * confused the old 4.x code. This is a good way to simulate |
| 916 | * preemption and MP without actually doing preemption or MP, because a |
| 917 | * lot of code assumes that wakeup() does not block. |
| 918 | */ |
| 919 | if (untimely_switch && td->td_nest_count == 0 && |
| 920 | gd->gd_intr_nesting_level == 0 |
| 921 | ) { |
| 922 | crit_enter_quick(td); |
| 923 | /* |
| 924 | * YYY temporary hacks until we disassociate the userland scheduler |
| 925 | * from the LWKT scheduler. |
| 926 | */ |
| 927 | if (td->td_flags & TDF_RUNQ) { |
| 928 | lwkt_switch(); /* will not reenter yield function */ |
| 929 | } else { |
| 930 | lwkt_schedule_self(td); /* make sure we are scheduled */ |
| 931 | lwkt_switch(); /* will not reenter yield function */ |
| 932 | lwkt_deschedule_self(td); /* make sure we are descheduled */ |
| 933 | } |
| 934 | crit_exit_noyield(td); |
| 935 | } |
| 936 | } |
| 937 | |
| 938 | /* |
| 939 | * This implements a normal yield which, unlike _quick, will yield to equal |
| 940 | * priority threads as well. Note that gd_reqflags tests will be handled by |
| 941 | * the crit_exit() call in lwkt_switch(). |
| 942 | * |
| 943 | * (self contained on a per cpu basis) |
| 944 | */ |
| 945 | void |
| 946 | lwkt_yield(void) |
| 947 | { |
| 948 | lwkt_schedule_self(curthread); |
| 949 | lwkt_switch(); |
| 950 | } |
| 951 | |
| 952 | /* |
| 953 | * Generic schedule. Possibly schedule threads belonging to other cpus and |
| 954 | * deal with threads that might be blocked on a wait queue. |
| 955 | * |
| 956 | * We have a little helper inline function which does additional work after |
| 957 | * the thread has been enqueued, including dealing with preemption and |
| 958 | * setting need_lwkt_resched() (which prevents the kernel from returning |
| 959 | * to userland until it has processed higher priority threads). |
| 960 | * |
| 961 | * It is possible for this routine to be called after a failed _enqueue |
| 962 | * (due to the target thread migrating, sleeping, or otherwise blocked). |
| 963 | * We have to check that the thread is actually on the run queue! |
| 964 | */ |
| 965 | static __inline |
| 966 | void |
| 967 | _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri) |
| 968 | { |
| 969 | if (ntd->td_flags & TDF_RUNQ) { |
| 970 | if (ntd->td_preemptable) { |
| 971 | ntd->td_preemptable(ntd, cpri); /* YYY +token */ |
| 972 | } else if ((ntd->td_flags & TDF_NORESCHED) == 0 && |
| 973 | (ntd->td_pri & TDPRI_MASK) > (gd->gd_curthread->td_pri & TDPRI_MASK) |
| 974 | ) { |
| 975 | need_lwkt_resched(); |
| 976 | } |
| 977 | } |
| 978 | } |
| 979 | |
| 980 | void |
| 981 | lwkt_schedule(thread_t td) |
| 982 | { |
| 983 | globaldata_t mygd = mycpu; |
| 984 | |
| 985 | KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!")); |
| 986 | crit_enter_gd(mygd); |
| 987 | KKASSERT(td->td_proc == NULL || (td->td_proc->p_flag & P_ONRUNQ) == 0); |
| 988 | if (td == mygd->gd_curthread) { |
| 989 | _lwkt_enqueue(td); |
| 990 | } else { |
| 991 | lwkt_wait_t w; |
| 992 | |
| 993 | /* |
| 994 | * If the thread is on a wait list we have to send our scheduling |
| 995 | * request to the owner of the wait structure. Otherwise we send |
| 996 | * the scheduling request to the cpu owning the thread. Races |
| 997 | * are ok, the target will forward the message as necessary (the |
| 998 | * message may chase the thread around before it finally gets |
| 999 | * acted upon). |
| 1000 | * |
| 1001 | * (remember, wait structures use stable storage) |
| 1002 | * |
| 1003 | * NOTE: we have to account for the number of critical sections |
| 1004 | * under our control when calling _lwkt_schedule_post() so it |
| 1005 | * can figure out whether preemption is allowed. |
| 1006 | * |
| 1007 | * NOTE: The wait structure algorithms are a mess and need to be |
| 1008 | * rewritten. |
| 1009 | * |
| 1010 | * NOTE: We cannot safely acquire or release a token, even |
| 1011 | * non-blocking, because this routine may be called in the context |
| 1012 | * of a thread already holding the token and thus not provide any |
| 1013 | * interlock protection. We cannot safely manipulate the td_toks |
| 1014 | * list for the same reason. Instead we depend on our critical |
| 1015 | * section if the token is owned by our cpu. |
| 1016 | */ |
| 1017 | if ((w = td->td_wait) != NULL) { |
| 1018 | if (w->wa_token.t_cpu == mygd) { |
| 1019 | TAILQ_REMOVE(&w->wa_waitq, td, td_threadq); |
| 1020 | --w->wa_count; |
| 1021 | td->td_wait = NULL; |
| 1022 | #ifdef SMP |
| 1023 | if (td->td_gd == mygd) { |
| 1024 | _lwkt_enqueue(td); |
| 1025 | _lwkt_schedule_post(mygd, td, TDPRI_CRIT); |
| 1026 | } else { |
| 1027 | lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td); |
| 1028 | } |
| 1029 | #else |
| 1030 | _lwkt_enqueue(td); |
| 1031 | _lwkt_schedule_post(mygd, td, TDPRI_CRIT); |
| 1032 | #endif |
| 1033 | } else { |
| 1034 | #ifdef SMP |
| 1035 | lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc1_t)lwkt_schedule, td); |
| 1036 | #else |
| 1037 | panic("bad token %p", &w->wa_token); |
| 1038 | #endif |
| 1039 | } |
| 1040 | } else { |
| 1041 | /* |
| 1042 | * If the wait structure is NULL and we own the thread, there |
| 1043 | * is no race (since we are in a critical section). If we |
| 1044 | * do not own the thread there might be a race but the |
| 1045 | * target cpu will deal with it. |
| 1046 | */ |
| 1047 | #ifdef SMP |
| 1048 | if (td->td_gd == mygd) { |
| 1049 | _lwkt_enqueue(td); |
| 1050 | _lwkt_schedule_post(mygd, td, TDPRI_CRIT); |
| 1051 | } else { |
| 1052 | lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td); |
| 1053 | } |
| 1054 | #else |
| 1055 | _lwkt_enqueue(td); |
| 1056 | _lwkt_schedule_post(mygd, td, TDPRI_CRIT); |
| 1057 | #endif |
| 1058 | } |
| 1059 | } |
| 1060 | crit_exit_gd(mygd); |
| 1061 | } |
| 1062 | |
| 1063 | /* |
| 1064 | * Managed acquisition. This code assumes that the MP lock is held for |
| 1065 | * the tdallq operation and that the thread has been descheduled from its |
| 1066 | * original cpu. We also have to wait for the thread to be entirely switched |
| 1067 | * out on its original cpu (this is usually fast enough that we never loop) |
| 1068 | * since the LWKT system does not have to hold the MP lock while switching |
| 1069 | * and the target may have released it before switching. |
| 1070 | */ |
| 1071 | void |
| 1072 | lwkt_acquire(thread_t td) |
| 1073 | { |
| 1074 | globaldata_t gd; |
| 1075 | globaldata_t mygd; |
| 1076 | |
| 1077 | gd = td->td_gd; |
| 1078 | mygd = mycpu; |
| 1079 | cpu_lfence(); |
| 1080 | KKASSERT((td->td_flags & TDF_RUNQ) == 0); |
| 1081 | while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) /* XXX spin */ |
| 1082 | cpu_lfence(); |
| 1083 | if (gd != mygd) { |
| 1084 | crit_enter_gd(mygd); |
| 1085 | TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */ |
| 1086 | td->td_gd = mygd; |
| 1087 | TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); /* protected by BGL */ |
| 1088 | crit_exit_gd(mygd); |
| 1089 | } |
| 1090 | } |
| 1091 | |
| 1092 | /* |
| 1093 | * Generic deschedule. Descheduling threads other then your own should be |
| 1094 | * done only in carefully controlled circumstances. Descheduling is |
| 1095 | * asynchronous. |
| 1096 | * |
| 1097 | * This function may block if the cpu has run out of messages. |
| 1098 | */ |
| 1099 | void |
| 1100 | lwkt_deschedule(thread_t td) |
| 1101 | { |
| 1102 | crit_enter(); |
| 1103 | #ifdef SMP |
| 1104 | if (td == curthread) { |
| 1105 | _lwkt_dequeue(td); |
| 1106 | } else { |
| 1107 | if (td->td_gd == mycpu) { |
| 1108 | _lwkt_dequeue(td); |
| 1109 | } else { |
| 1110 | lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td); |
| 1111 | } |
| 1112 | } |
| 1113 | #else |
| 1114 | _lwkt_dequeue(td); |
| 1115 | #endif |
| 1116 | crit_exit(); |
| 1117 | } |
| 1118 | |
| 1119 | /* |
| 1120 | * Set the target thread's priority. This routine does not automatically |
| 1121 | * switch to a higher priority thread, LWKT threads are not designed for |
| 1122 | * continuous priority changes. Yield if you want to switch. |
| 1123 | * |
| 1124 | * We have to retain the critical section count which uses the high bits |
| 1125 | * of the td_pri field. The specified priority may also indicate zero or |
| 1126 | * more critical sections by adding TDPRI_CRIT*N. |
| 1127 | * |
| 1128 | * Note that we requeue the thread whether it winds up on a different runq |
| 1129 | * or not. uio_yield() depends on this and the routine is not normally |
| 1130 | * called with the same priority otherwise. |
| 1131 | */ |
| 1132 | void |
| 1133 | lwkt_setpri(thread_t td, int pri) |
| 1134 | { |
| 1135 | KKASSERT(pri >= 0); |
| 1136 | KKASSERT(td->td_gd == mycpu); |
| 1137 | crit_enter(); |
| 1138 | if (td->td_flags & TDF_RUNQ) { |
| 1139 | _lwkt_dequeue(td); |
| 1140 | td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri; |
| 1141 | _lwkt_enqueue(td); |
| 1142 | } else { |
| 1143 | td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri; |
| 1144 | } |
| 1145 | crit_exit(); |
| 1146 | } |
| 1147 | |
| 1148 | void |
| 1149 | lwkt_setpri_self(int pri) |
| 1150 | { |
| 1151 | thread_t td = curthread; |
| 1152 | |
| 1153 | KKASSERT(pri >= 0 && pri <= TDPRI_MAX); |
| 1154 | crit_enter(); |
| 1155 | if (td->td_flags & TDF_RUNQ) { |
| 1156 | _lwkt_dequeue(td); |
| 1157 | td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri; |
| 1158 | _lwkt_enqueue(td); |
| 1159 | } else { |
| 1160 | td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri; |
| 1161 | } |
| 1162 | crit_exit(); |
| 1163 | } |
| 1164 | |
| 1165 | /* |
| 1166 | * Determine if there is a runnable thread at a higher priority then |
| 1167 | * the current thread. lwkt_setpri() does not check this automatically. |
| 1168 | * Return 1 if there is, 0 if there isn't. |
| 1169 | * |
| 1170 | * Example: if bit 31 of runqmask is set and the current thread is priority |
| 1171 | * 30, then we wind up checking the mask: 0x80000000 against 0x7fffffff. |
| 1172 | * |
| 1173 | * If nq reaches 31 the shift operation will overflow to 0 and we will wind |
| 1174 | * up comparing against 0xffffffff, a comparison that will always be false. |
| 1175 | */ |
| 1176 | int |
| 1177 | lwkt_checkpri_self(void) |
| 1178 | { |
| 1179 | globaldata_t gd = mycpu; |
| 1180 | thread_t td = gd->gd_curthread; |
| 1181 | int nq = td->td_pri & TDPRI_MASK; |
| 1182 | |
| 1183 | while (gd->gd_runqmask > (__uint32_t)(2 << nq) - 1) { |
| 1184 | if (TAILQ_FIRST(&gd->gd_tdrunq[nq + 1])) |
| 1185 | return(1); |
| 1186 | ++nq; |
| 1187 | } |
| 1188 | return(0); |
| 1189 | } |
| 1190 | |
| 1191 | /* |
| 1192 | * Migrate the current thread to the specified cpu. The BGL must be held |
| 1193 | * (for the gd_tdallq manipulation XXX). This is accomplished by |
| 1194 | * descheduling ourselves from the current cpu, moving our thread to the |
| 1195 | * tdallq of the target cpu, IPI messaging the target cpu, and switching out. |
| 1196 | * TDF_MIGRATING prevents scheduling races while the thread is being migrated. |
| 1197 | */ |
| 1198 | #ifdef SMP |
| 1199 | static void lwkt_setcpu_remote(void *arg); |
| 1200 | #endif |
| 1201 | |
| 1202 | void |
| 1203 | lwkt_setcpu_self(globaldata_t rgd) |
| 1204 | { |
| 1205 | #ifdef SMP |
| 1206 | thread_t td = curthread; |
| 1207 | |
| 1208 | if (td->td_gd != rgd) { |
| 1209 | crit_enter_quick(td); |
| 1210 | td->td_flags |= TDF_MIGRATING; |
| 1211 | lwkt_deschedule_self(td); |
| 1212 | TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); /* protected by BGL */ |
| 1213 | TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); /* protected by BGL */ |
| 1214 | lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td); |
| 1215 | lwkt_switch(); |
| 1216 | /* we are now on the target cpu */ |
| 1217 | crit_exit_quick(td); |
| 1218 | } |
| 1219 | #endif |
| 1220 | } |
| 1221 | |
| 1222 | /* |
| 1223 | * Remote IPI for cpu migration (called while in a critical section so we |
| 1224 | * do not have to enter another one). The thread has already been moved to |
| 1225 | * our cpu's allq, but we must wait for the thread to be completely switched |
| 1226 | * out on the originating cpu before we schedule it on ours or the stack |
| 1227 | * state may be corrupt. We clear TDF_MIGRATING after flushing the GD |
| 1228 | * change to main memory. |
| 1229 | * |
| 1230 | * XXX The use of TDF_MIGRATING might not be sufficient to avoid races |
| 1231 | * against wakeups. It is best if this interface is used only when there |
| 1232 | * are no pending events that might try to schedule the thread. |
| 1233 | */ |
| 1234 | #ifdef SMP |
| 1235 | static void |
| 1236 | lwkt_setcpu_remote(void *arg) |
| 1237 | { |
| 1238 | thread_t td = arg; |
| 1239 | globaldata_t gd = mycpu; |
| 1240 | |
| 1241 | while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) |
| 1242 | cpu_lfence(); |
| 1243 | td->td_gd = gd; |
| 1244 | cpu_sfence(); |
| 1245 | td->td_flags &= ~TDF_MIGRATING; |
| 1246 | KKASSERT(td->td_proc == NULL || (td->td_proc->p_flag & P_ONRUNQ) == 0); |
| 1247 | _lwkt_enqueue(td); |
| 1248 | } |
| 1249 | #endif |
| 1250 | |
| 1251 | struct lwp * |
| 1252 | lwkt_preempted_proc(void) |
| 1253 | { |
| 1254 | thread_t td = curthread; |
| 1255 | while (td->td_preempted) |
| 1256 | td = td->td_preempted; |
| 1257 | return(td->td_lwp); |
| 1258 | } |
| 1259 | |
| 1260 | /* |
| 1261 | * Block on the specified wait queue until signaled. A generation number |
| 1262 | * must be supplied to interlock the wait queue. The function will |
| 1263 | * return immediately if the generation number does not match the wait |
| 1264 | * structure's generation number. |
| 1265 | */ |
| 1266 | void |
| 1267 | lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen) |
| 1268 | { |
| 1269 | thread_t td = curthread; |
| 1270 | lwkt_tokref ilock; |
| 1271 | |
| 1272 | lwkt_gettoken(&ilock, &w->wa_token); |
| 1273 | crit_enter(); |
| 1274 | if (w->wa_gen == *gen) { |
| 1275 | _lwkt_dequeue(td); |
| 1276 | td->td_flags |= TDF_BLOCKQ; |
| 1277 | TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq); |
| 1278 | ++w->wa_count; |
| 1279 | td->td_wait = w; |
| 1280 | td->td_wmesg = wmesg; |
| 1281 | lwkt_switch(); |
| 1282 | KKASSERT((td->td_flags & TDF_BLOCKQ) == 0); |
| 1283 | td->td_wmesg = NULL; |
| 1284 | } |
| 1285 | crit_exit(); |
| 1286 | *gen = w->wa_gen; |
| 1287 | lwkt_reltoken(&ilock); |
| 1288 | } |
| 1289 | |
| 1290 | /* |
| 1291 | * Signal a wait queue. We gain ownership of the wait queue in order to |
| 1292 | * signal it. Once a thread is removed from the wait queue we have to |
| 1293 | * deal with the cpu owning the thread. |
| 1294 | * |
| 1295 | * Note: alternatively we could message the target cpu owning the wait |
| 1296 | * queue. YYY implement as sysctl. |
| 1297 | */ |
| 1298 | void |
| 1299 | lwkt_signal(lwkt_wait_t w, int count) |
| 1300 | { |
| 1301 | thread_t td; |
| 1302 | lwkt_tokref ilock; |
| 1303 | |
| 1304 | lwkt_gettoken(&ilock, &w->wa_token); |
| 1305 | ++w->wa_gen; |
| 1306 | crit_enter(); |
| 1307 | if (count < 0) |
| 1308 | count = w->wa_count; |
| 1309 | while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) { |
| 1310 | --count; |
| 1311 | --w->wa_count; |
| 1312 | KKASSERT(td->td_flags & TDF_BLOCKQ); |
| 1313 | TAILQ_REMOVE(&w->wa_waitq, td, td_threadq); |
| 1314 | td->td_flags &= ~TDF_BLOCKQ; |
| 1315 | td->td_wait = NULL; |
| 1316 | KKASSERT(td->td_proc == NULL || (td->td_proc->p_flag & P_ONRUNQ) == 0); |
| 1317 | #ifdef SMP |
| 1318 | if (td->td_gd == mycpu) { |
| 1319 | _lwkt_enqueue(td); |
| 1320 | } else { |
| 1321 | lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td); |
| 1322 | } |
| 1323 | #else |
| 1324 | _lwkt_enqueue(td); |
| 1325 | #endif |
| 1326 | } |
| 1327 | crit_exit(); |
| 1328 | lwkt_reltoken(&ilock); |
| 1329 | } |
| 1330 | |
| 1331 | /* |
| 1332 | * Create a kernel process/thread/whatever. It shares it's address space |
| 1333 | * with proc0 - ie: kernel only. |
| 1334 | * |
| 1335 | * NOTE! By default new threads are created with the MP lock held. A |
| 1336 | * thread which does not require the MP lock should release it by calling |
| 1337 | * rel_mplock() at the start of the new thread. |
| 1338 | */ |
| 1339 | int |
| 1340 | lwkt_create(void (*func)(void *), void *arg, |
| 1341 | struct thread **tdp, thread_t template, int tdflags, int cpu, |
| 1342 | const char *fmt, ...) |
| 1343 | { |
| 1344 | thread_t td; |
| 1345 | __va_list ap; |
| 1346 | |
| 1347 | td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu, |
| 1348 | tdflags | TDF_VERBOSE); |
| 1349 | if (tdp) |
| 1350 | *tdp = td; |
| 1351 | cpu_set_thread_handler(td, lwkt_exit, func, arg); |
| 1352 | |
| 1353 | /* |
| 1354 | * Set up arg0 for 'ps' etc |
| 1355 | */ |
| 1356 | __va_start(ap, fmt); |
| 1357 | vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap); |
| 1358 | __va_end(ap); |
| 1359 | |
| 1360 | /* |
| 1361 | * Schedule the thread to run |
| 1362 | */ |
| 1363 | if ((td->td_flags & TDF_STOPREQ) == 0) |
| 1364 | lwkt_schedule(td); |
| 1365 | else |
| 1366 | td->td_flags &= ~TDF_STOPREQ; |
| 1367 | return 0; |
| 1368 | } |
| 1369 | |
| 1370 | /* |
| 1371 | * kthread_* is specific to the kernel and is not needed by userland. |
| 1372 | */ |
| 1373 | #ifdef _KERNEL |
| 1374 | |
| 1375 | /* |
| 1376 | * Destroy an LWKT thread. Warning! This function is not called when |
| 1377 | * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and |
| 1378 | * uses a different reaping mechanism. |
| 1379 | */ |
| 1380 | void |
| 1381 | lwkt_exit(void) |
| 1382 | { |
| 1383 | thread_t td = curthread; |
| 1384 | globaldata_t gd; |
| 1385 | |
| 1386 | if (td->td_flags & TDF_VERBOSE) |
| 1387 | printf("kthread %p %s has exited\n", td, td->td_comm); |
| 1388 | caps_exit(td); |
| 1389 | crit_enter_quick(td); |
| 1390 | lwkt_deschedule_self(td); |
| 1391 | gd = mycpu; |
| 1392 | KKASSERT(gd == td->td_gd); |
| 1393 | TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); |
| 1394 | if (td->td_flags & TDF_ALLOCATED_THREAD) { |
| 1395 | ++gd->gd_tdfreecount; |
| 1396 | TAILQ_INSERT_TAIL(&gd->gd_tdfreeq, td, td_threadq); |
| 1397 | } |
| 1398 | cpu_thread_exit(); |
| 1399 | } |
| 1400 | |
| 1401 | #endif /* _KERNEL */ |
| 1402 | |
| 1403 | void |
| 1404 | crit_panic(void) |
| 1405 | { |
| 1406 | thread_t td = curthread; |
| 1407 | int lpri = td->td_pri; |
| 1408 | |
| 1409 | td->td_pri = 0; |
| 1410 | panic("td_pri is/would-go negative! %p %d", td, lpri); |
| 1411 | } |
| 1412 | |
| 1413 | #ifdef SMP |
| 1414 | |
| 1415 | /* |
| 1416 | * Called from debugger/panic on cpus which have been stopped. We must still |
| 1417 | * process the IPIQ while stopped, even if we were stopped while in a critical |
| 1418 | * section (XXX). |
| 1419 | * |
| 1420 | * If we are dumping also try to process any pending interrupts. This may |
| 1421 | * or may not work depending on the state of the cpu at the point it was |
| 1422 | * stopped. |
| 1423 | */ |
| 1424 | void |
| 1425 | lwkt_smp_stopped(void) |
| 1426 | { |
| 1427 | globaldata_t gd = mycpu; |
| 1428 | |
| 1429 | crit_enter_gd(gd); |
| 1430 | if (dumping) { |
| 1431 | lwkt_process_ipiq(); |
| 1432 | splz(); |
| 1433 | } else { |
| 1434 | lwkt_process_ipiq(); |
| 1435 | } |
| 1436 | crit_exit_gd(gd); |
| 1437 | } |
| 1438 | |
| 1439 | #endif |