| 1 | /* |
| 2 | * Copyright (c) 2003 Matthew Dillon <dillon@backplane.com> |
| 3 | * All rights reserved. |
| 4 | * |
| 5 | * Redistribution and use in source and binary forms, with or without |
| 6 | * modification, are permitted provided that the following conditions |
| 7 | * are met: |
| 8 | * 1. Redistributions of source code must retain the above copyright |
| 9 | * notice, this list of conditions and the following disclaimer. |
| 10 | * 2. Redistributions in binary form must reproduce the above copyright |
| 11 | * notice, this list of conditions and the following disclaimer in the |
| 12 | * documentation and/or other materials provided with the distribution. |
| 13 | * |
| 14 | * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND |
| 15 | * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE |
| 16 | * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE |
| 17 | * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE |
| 18 | * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL |
| 19 | * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS |
| 20 | * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) |
| 21 | * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT |
| 22 | * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY |
| 23 | * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF |
| 24 | * SUCH DAMAGE. |
| 25 | * |
| 26 | * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.51 2004/02/09 21:13:18 dillon Exp $ |
| 27 | */ |
| 28 | |
| 29 | /* |
| 30 | * Each cpu in a system has its own self-contained light weight kernel |
| 31 | * thread scheduler, which means that generally speaking we only need |
| 32 | * to use a critical section to avoid problems. Foreign thread |
| 33 | * scheduling is queued via (async) IPIs. |
| 34 | * |
| 35 | * NOTE: on UP machines smp_active is defined to be 0. On SMP machines |
| 36 | * smp_active is 0 prior to SMP activation, then it is 1. The LWKT module |
| 37 | * uses smp_active to optimize UP builds and to avoid sending IPIs during |
| 38 | * early boot (primarily interrupt and network thread initialization). |
| 39 | */ |
| 40 | |
| 41 | #ifdef _KERNEL |
| 42 | |
| 43 | #include <sys/param.h> |
| 44 | #include <sys/systm.h> |
| 45 | #include <sys/kernel.h> |
| 46 | #include <sys/proc.h> |
| 47 | #include <sys/rtprio.h> |
| 48 | #include <sys/queue.h> |
| 49 | #include <sys/thread2.h> |
| 50 | #include <sys/sysctl.h> |
| 51 | #include <sys/kthread.h> |
| 52 | #include <machine/cpu.h> |
| 53 | #include <sys/lock.h> |
| 54 | #include <sys/caps.h> |
| 55 | |
| 56 | #include <vm/vm.h> |
| 57 | #include <vm/vm_param.h> |
| 58 | #include <vm/vm_kern.h> |
| 59 | #include <vm/vm_object.h> |
| 60 | #include <vm/vm_page.h> |
| 61 | #include <vm/vm_map.h> |
| 62 | #include <vm/vm_pager.h> |
| 63 | #include <vm/vm_extern.h> |
| 64 | #include <vm/vm_zone.h> |
| 65 | |
| 66 | #include <machine/stdarg.h> |
| 67 | #include <machine/ipl.h> |
| 68 | #include <machine/smp.h> |
| 69 | |
| 70 | #define THREAD_STACK (UPAGES * PAGE_SIZE) |
| 71 | |
| 72 | #else |
| 73 | |
| 74 | #include <sys/stdint.h> |
| 75 | #include <libcaps/thread.h> |
| 76 | #include <sys/thread.h> |
| 77 | #include <sys/msgport.h> |
| 78 | #include <sys/errno.h> |
| 79 | #include <libcaps/globaldata.h> |
| 80 | #include <sys/thread2.h> |
| 81 | #include <sys/msgport2.h> |
| 82 | #include <stdio.h> |
| 83 | #include <stdlib.h> |
| 84 | #include <string.h> |
| 85 | #include <machine/cpufunc.h> |
| 86 | #include <machine/lock.h> |
| 87 | |
| 88 | #endif |
| 89 | |
| 90 | static int untimely_switch = 0; |
| 91 | static __int64_t switch_count = 0; |
| 92 | static __int64_t preempt_hit = 0; |
| 93 | static __int64_t preempt_miss = 0; |
| 94 | static __int64_t preempt_weird = 0; |
| 95 | #ifdef SMP |
| 96 | static __int64_t ipiq_count = 0; |
| 97 | static __int64_t ipiq_fifofull = 0; |
| 98 | #endif |
| 99 | |
| 100 | #ifdef _KERNEL |
| 101 | |
| 102 | SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, ""); |
| 103 | SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, ""); |
| 104 | SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, ""); |
| 105 | SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, ""); |
| 106 | SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, ""); |
| 107 | #ifdef SMP |
| 108 | SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_count, CTLFLAG_RW, &ipiq_count, 0, ""); |
| 109 | SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_fifofull, CTLFLAG_RW, &ipiq_fifofull, 0, ""); |
| 110 | #endif |
| 111 | |
| 112 | #endif |
| 113 | |
| 114 | /* |
| 115 | * These helper procedures handle the runq, they can only be called from |
| 116 | * within a critical section. |
| 117 | * |
| 118 | * WARNING! Prior to SMP being brought up it is possible to enqueue and |
| 119 | * dequeue threads belonging to other cpus, so be sure to use td->td_gd |
| 120 | * instead of 'mycpu' when referencing the globaldata structure. Once |
| 121 | * SMP live enqueuing and dequeueing only occurs on the current cpu. |
| 122 | */ |
| 123 | static __inline |
| 124 | void |
| 125 | _lwkt_dequeue(thread_t td) |
| 126 | { |
| 127 | if (td->td_flags & TDF_RUNQ) { |
| 128 | int nq = td->td_pri & TDPRI_MASK; |
| 129 | struct globaldata *gd = td->td_gd; |
| 130 | |
| 131 | td->td_flags &= ~TDF_RUNQ; |
| 132 | TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq); |
| 133 | /* runqmask is passively cleaned up by the switcher */ |
| 134 | } |
| 135 | } |
| 136 | |
| 137 | static __inline |
| 138 | void |
| 139 | _lwkt_enqueue(thread_t td) |
| 140 | { |
| 141 | if ((td->td_flags & TDF_RUNQ) == 0) { |
| 142 | int nq = td->td_pri & TDPRI_MASK; |
| 143 | struct globaldata *gd = td->td_gd; |
| 144 | |
| 145 | td->td_flags |= TDF_RUNQ; |
| 146 | TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq); |
| 147 | gd->gd_runqmask |= 1 << nq; |
| 148 | } |
| 149 | } |
| 150 | |
| 151 | static __inline |
| 152 | int |
| 153 | _lwkt_wantresched(thread_t ntd, thread_t cur) |
| 154 | { |
| 155 | return((ntd->td_pri & TDPRI_MASK) > (cur->td_pri & TDPRI_MASK)); |
| 156 | } |
| 157 | |
| 158 | #ifdef _KERNEL |
| 159 | |
| 160 | /* |
| 161 | * LWKTs operate on a per-cpu basis |
| 162 | * |
| 163 | * WARNING! Called from early boot, 'mycpu' may not work yet. |
| 164 | */ |
| 165 | void |
| 166 | lwkt_gdinit(struct globaldata *gd) |
| 167 | { |
| 168 | int i; |
| 169 | |
| 170 | for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i) |
| 171 | TAILQ_INIT(&gd->gd_tdrunq[i]); |
| 172 | gd->gd_runqmask = 0; |
| 173 | TAILQ_INIT(&gd->gd_tdallq); |
| 174 | } |
| 175 | |
| 176 | #endif /* _KERNEL */ |
| 177 | |
| 178 | /* |
| 179 | * Initialize a thread wait structure prior to first use. |
| 180 | * |
| 181 | * NOTE! called from low level boot code, we cannot do anything fancy! |
| 182 | */ |
| 183 | void |
| 184 | lwkt_init_wait(lwkt_wait_t w) |
| 185 | { |
| 186 | TAILQ_INIT(&w->wa_waitq); |
| 187 | } |
| 188 | |
| 189 | /* |
| 190 | * Create a new thread. The thread must be associated with a process context |
| 191 | * or LWKT start address before it can be scheduled. If the target cpu is |
| 192 | * -1 the thread will be created on the current cpu. |
| 193 | * |
| 194 | * If you intend to create a thread without a process context this function |
| 195 | * does everything except load the startup and switcher function. |
| 196 | */ |
| 197 | thread_t |
| 198 | lwkt_alloc_thread(struct thread *td, int cpu) |
| 199 | { |
| 200 | void *stack; |
| 201 | int flags = 0; |
| 202 | |
| 203 | if (td == NULL) { |
| 204 | crit_enter(); |
| 205 | if (mycpu->gd_tdfreecount > 0) { |
| 206 | --mycpu->gd_tdfreecount; |
| 207 | td = TAILQ_FIRST(&mycpu->gd_tdfreeq); |
| 208 | KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0, |
| 209 | ("lwkt_alloc_thread: unexpected NULL or corrupted td")); |
| 210 | TAILQ_REMOVE(&mycpu->gd_tdfreeq, td, td_threadq); |
| 211 | crit_exit(); |
| 212 | stack = td->td_kstack; |
| 213 | flags = td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD); |
| 214 | } else { |
| 215 | crit_exit(); |
| 216 | #ifdef _KERNEL |
| 217 | td = zalloc(thread_zone); |
| 218 | #else |
| 219 | td = malloc(sizeof(struct thread)); |
| 220 | #endif |
| 221 | td->td_kstack = NULL; |
| 222 | flags |= TDF_ALLOCATED_THREAD; |
| 223 | } |
| 224 | } |
| 225 | if ((stack = td->td_kstack) == NULL) { |
| 226 | #ifdef _KERNEL |
| 227 | stack = (void *)kmem_alloc(kernel_map, THREAD_STACK); |
| 228 | #else |
| 229 | stack = libcaps_alloc_stack(THREAD_STACK); |
| 230 | #endif |
| 231 | flags |= TDF_ALLOCATED_STACK; |
| 232 | } |
| 233 | if (cpu < 0) |
| 234 | lwkt_init_thread(td, stack, flags, mycpu); |
| 235 | else |
| 236 | lwkt_init_thread(td, stack, flags, globaldata_find(cpu)); |
| 237 | return(td); |
| 238 | } |
| 239 | |
| 240 | #ifdef _KERNEL |
| 241 | |
| 242 | /* |
| 243 | * Initialize a preexisting thread structure. This function is used by |
| 244 | * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread. |
| 245 | * |
| 246 | * All threads start out in a critical section at a priority of |
| 247 | * TDPRI_KERN_DAEMON. Higher level code will modify the priority as |
| 248 | * appropriate. This function may send an IPI message when the |
| 249 | * requested cpu is not the current cpu and consequently gd_tdallq may |
| 250 | * not be initialized synchronously from the point of view of the originating |
| 251 | * cpu. |
| 252 | * |
| 253 | * NOTE! we have to be careful in regards to creating threads for other cpus |
| 254 | * if SMP has not yet been activated. |
| 255 | */ |
| 256 | static void |
| 257 | lwkt_init_thread_remote(void *arg) |
| 258 | { |
| 259 | thread_t td = arg; |
| 260 | |
| 261 | TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq); |
| 262 | } |
| 263 | |
| 264 | void |
| 265 | lwkt_init_thread(thread_t td, void *stack, int flags, struct globaldata *gd) |
| 266 | { |
| 267 | bzero(td, sizeof(struct thread)); |
| 268 | td->td_kstack = stack; |
| 269 | td->td_flags |= flags; |
| 270 | td->td_gd = gd; |
| 271 | td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT; |
| 272 | lwkt_initport(&td->td_msgport, td); |
| 273 | pmap_init_thread(td); |
| 274 | if (smp_active == 0 || gd == mycpu) { |
| 275 | crit_enter(); |
| 276 | TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); |
| 277 | crit_exit(); |
| 278 | } else { |
| 279 | lwkt_send_ipiq(gd->gd_cpuid, lwkt_init_thread_remote, td); |
| 280 | } |
| 281 | } |
| 282 | |
| 283 | #endif /* _KERNEL */ |
| 284 | |
| 285 | void |
| 286 | lwkt_set_comm(thread_t td, const char *ctl, ...) |
| 287 | { |
| 288 | __va_list va; |
| 289 | |
| 290 | __va_start(va, ctl); |
| 291 | vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va); |
| 292 | __va_end(va); |
| 293 | } |
| 294 | |
| 295 | void |
| 296 | lwkt_hold(thread_t td) |
| 297 | { |
| 298 | ++td->td_refs; |
| 299 | } |
| 300 | |
| 301 | void |
| 302 | lwkt_rele(thread_t td) |
| 303 | { |
| 304 | KKASSERT(td->td_refs > 0); |
| 305 | --td->td_refs; |
| 306 | } |
| 307 | |
| 308 | #ifdef _KERNEL |
| 309 | |
| 310 | void |
| 311 | lwkt_wait_free(thread_t td) |
| 312 | { |
| 313 | while (td->td_refs) |
| 314 | tsleep(td, 0, "tdreap", hz); |
| 315 | } |
| 316 | |
| 317 | #endif |
| 318 | |
| 319 | void |
| 320 | lwkt_free_thread(thread_t td) |
| 321 | { |
| 322 | struct globaldata *gd = mycpu; |
| 323 | |
| 324 | KASSERT((td->td_flags & TDF_RUNNING) == 0, |
| 325 | ("lwkt_free_thread: did not exit! %p", td)); |
| 326 | |
| 327 | crit_enter(); |
| 328 | TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); |
| 329 | if (gd->gd_tdfreecount < CACHE_NTHREADS && |
| 330 | (td->td_flags & TDF_ALLOCATED_THREAD) |
| 331 | ) { |
| 332 | ++gd->gd_tdfreecount; |
| 333 | TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq); |
| 334 | crit_exit(); |
| 335 | } else { |
| 336 | crit_exit(); |
| 337 | if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) { |
| 338 | #ifdef _KERNEL |
| 339 | kmem_free(kernel_map, (vm_offset_t)td->td_kstack, THREAD_STACK); |
| 340 | #else |
| 341 | libcaps_free_stack(td->td_kstack, THREAD_STACK); |
| 342 | #endif |
| 343 | /* gd invalid */ |
| 344 | td->td_kstack = NULL; |
| 345 | } |
| 346 | if (td->td_flags & TDF_ALLOCATED_THREAD) { |
| 347 | #ifdef _KERNEL |
| 348 | zfree(thread_zone, td); |
| 349 | #else |
| 350 | free(td); |
| 351 | #endif |
| 352 | } |
| 353 | } |
| 354 | } |
| 355 | |
| 356 | |
| 357 | /* |
| 358 | * Switch to the next runnable lwkt. If no LWKTs are runnable then |
| 359 | * switch to the idlethread. Switching must occur within a critical |
| 360 | * section to avoid races with the scheduling queue. |
| 361 | * |
| 362 | * We always have full control over our cpu's run queue. Other cpus |
| 363 | * that wish to manipulate our queue must use the cpu_*msg() calls to |
| 364 | * talk to our cpu, so a critical section is all that is needed and |
| 365 | * the result is very, very fast thread switching. |
| 366 | * |
| 367 | * The LWKT scheduler uses a fixed priority model and round-robins at |
| 368 | * each priority level. User process scheduling is a totally |
| 369 | * different beast and LWKT priorities should not be confused with |
| 370 | * user process priorities. |
| 371 | * |
| 372 | * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch() |
| 373 | * cleans it up. Note that the td_switch() function cannot do anything that |
| 374 | * requires the MP lock since the MP lock will have already been setup for |
| 375 | * the target thread (not the current thread). It's nice to have a scheduler |
| 376 | * that does not need the MP lock to work because it allows us to do some |
| 377 | * really cool high-performance MP lock optimizations. |
| 378 | */ |
| 379 | |
| 380 | void |
| 381 | lwkt_switch(void) |
| 382 | { |
| 383 | struct globaldata *gd; |
| 384 | thread_t td = curthread; |
| 385 | thread_t ntd; |
| 386 | #ifdef SMP |
| 387 | int mpheld; |
| 388 | #endif |
| 389 | |
| 390 | /* |
| 391 | * Switching from within a 'fast' (non thread switched) interrupt is |
| 392 | * illegal. |
| 393 | */ |
| 394 | if (mycpu->gd_intr_nesting_level && panicstr == NULL) { |
| 395 | panic("lwkt_switch: cannot switch from within a fast interrupt, yet\n"); |
| 396 | } |
| 397 | |
| 398 | /* |
| 399 | * Passive release (used to transition from user to kernel mode |
| 400 | * when we block or switch rather then when we enter the kernel). |
| 401 | * This function is NOT called if we are switching into a preemption |
| 402 | * or returning from a preemption. Typically this causes us to lose |
| 403 | * our P_CURPROC designation (if we have one) and become a true LWKT |
| 404 | * thread, and may also hand P_CURPROC to another process and schedule |
| 405 | * its thread. |
| 406 | */ |
| 407 | if (td->td_release) |
| 408 | td->td_release(td); |
| 409 | |
| 410 | crit_enter(); |
| 411 | ++switch_count; |
| 412 | |
| 413 | #ifdef SMP |
| 414 | /* |
| 415 | * td_mpcount cannot be used to determine if we currently hold the |
| 416 | * MP lock because get_mplock() will increment it prior to attempting |
| 417 | * to get the lock, and switch out if it can't. Our ownership of |
| 418 | * the actual lock will remain stable while we are in a critical section |
| 419 | * (but, of course, another cpu may own or release the lock so the |
| 420 | * actual value of mp_lock is not stable). |
| 421 | */ |
| 422 | mpheld = MP_LOCK_HELD(); |
| 423 | #endif |
| 424 | if ((ntd = td->td_preempted) != NULL) { |
| 425 | /* |
| 426 | * We had preempted another thread on this cpu, resume the preempted |
| 427 | * thread. This occurs transparently, whether the preempted thread |
| 428 | * was scheduled or not (it may have been preempted after descheduling |
| 429 | * itself). |
| 430 | * |
| 431 | * We have to setup the MP lock for the original thread after backing |
| 432 | * out the adjustment that was made to curthread when the original |
| 433 | * was preempted. |
| 434 | */ |
| 435 | KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK); |
| 436 | #ifdef SMP |
| 437 | if (ntd->td_mpcount && mpheld == 0) { |
| 438 | panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d\n", |
| 439 | td, ntd, td->td_mpcount, ntd->td_mpcount); |
| 440 | } |
| 441 | if (ntd->td_mpcount) { |
| 442 | td->td_mpcount -= ntd->td_mpcount; |
| 443 | KKASSERT(td->td_mpcount >= 0); |
| 444 | } |
| 445 | #endif |
| 446 | ntd->td_flags |= TDF_PREEMPT_DONE; |
| 447 | /* YYY release mp lock on switchback if original doesn't need it */ |
| 448 | } else { |
| 449 | /* |
| 450 | * Priority queue / round-robin at each priority. Note that user |
| 451 | * processes run at a fixed, low priority and the user process |
| 452 | * scheduler deals with interactions between user processes |
| 453 | * by scheduling and descheduling them from the LWKT queue as |
| 454 | * necessary. |
| 455 | * |
| 456 | * We have to adjust the MP lock for the target thread. If we |
| 457 | * need the MP lock and cannot obtain it we try to locate a |
| 458 | * thread that does not need the MP lock. |
| 459 | */ |
| 460 | gd = mycpu; |
| 461 | again: |
| 462 | if (gd->gd_runqmask) { |
| 463 | int nq = bsrl(gd->gd_runqmask); |
| 464 | if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) { |
| 465 | gd->gd_runqmask &= ~(1 << nq); |
| 466 | goto again; |
| 467 | } |
| 468 | #ifdef SMP |
| 469 | if (ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) { |
| 470 | /* |
| 471 | * Target needs MP lock and we couldn't get it, try |
| 472 | * to locate a thread which does not need the MP lock |
| 473 | * to run. If we cannot locate a thread spin in idle. |
| 474 | */ |
| 475 | u_int32_t rqmask = gd->gd_runqmask; |
| 476 | while (rqmask) { |
| 477 | TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) { |
| 478 | if (ntd->td_mpcount == 0) |
| 479 | break; |
| 480 | } |
| 481 | if (ntd) |
| 482 | break; |
| 483 | rqmask &= ~(1 << nq); |
| 484 | nq = bsrl(rqmask); |
| 485 | } |
| 486 | if (ntd == NULL) { |
| 487 | ntd = &gd->gd_idlethread; |
| 488 | ntd->td_flags |= TDF_IDLE_NOHLT; |
| 489 | } else { |
| 490 | TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq); |
| 491 | TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq); |
| 492 | } |
| 493 | } else { |
| 494 | TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq); |
| 495 | TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq); |
| 496 | } |
| 497 | #else |
| 498 | TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq); |
| 499 | TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq); |
| 500 | #endif |
| 501 | } else { |
| 502 | /* |
| 503 | * We have nothing to run but only let the idle loop halt |
| 504 | * the cpu if there are no pending interrupts. |
| 505 | */ |
| 506 | ntd = &gd->gd_idlethread; |
| 507 | if (gd->gd_reqflags & RQF_IDLECHECK_MASK) |
| 508 | ntd->td_flags |= TDF_IDLE_NOHLT; |
| 509 | } |
| 510 | } |
| 511 | KASSERT(ntd->td_pri >= TDPRI_CRIT, |
| 512 | ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri)); |
| 513 | |
| 514 | /* |
| 515 | * Do the actual switch. If the new target does not need the MP lock |
| 516 | * and we are holding it, release the MP lock. If the new target requires |
| 517 | * the MP lock we have already acquired it for the target. |
| 518 | */ |
| 519 | #ifdef SMP |
| 520 | if (ntd->td_mpcount == 0 ) { |
| 521 | if (MP_LOCK_HELD()) |
| 522 | cpu_rel_mplock(); |
| 523 | } else { |
| 524 | ASSERT_MP_LOCK_HELD(); |
| 525 | } |
| 526 | #endif |
| 527 | if (td != ntd) { |
| 528 | td->td_switch(ntd); |
| 529 | } |
| 530 | |
| 531 | crit_exit(); |
| 532 | } |
| 533 | |
| 534 | /* |
| 535 | * Switch if another thread has a higher priority. Do not switch to other |
| 536 | * threads at the same priority. |
| 537 | */ |
| 538 | void |
| 539 | lwkt_maybe_switch() |
| 540 | { |
| 541 | struct globaldata *gd = mycpu; |
| 542 | struct thread *td = gd->gd_curthread; |
| 543 | |
| 544 | if ((td->td_pri & TDPRI_MASK) < bsrl(gd->gd_runqmask)) { |
| 545 | lwkt_switch(); |
| 546 | } |
| 547 | } |
| 548 | |
| 549 | /* |
| 550 | * Request that the target thread preempt the current thread. Preemption |
| 551 | * only works under a specific set of conditions: |
| 552 | * |
| 553 | * - We are not preempting ourselves |
| 554 | * - The target thread is owned by the current cpu |
| 555 | * - We are not currently being preempted |
| 556 | * - The target is not currently being preempted |
| 557 | * - We are able to satisfy the target's MP lock requirements (if any). |
| 558 | * |
| 559 | * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically |
| 560 | * this is called via lwkt_schedule() through the td_preemptable callback. |
| 561 | * critpri is the managed critical priority that we should ignore in order |
| 562 | * to determine whether preemption is possible (aka usually just the crit |
| 563 | * priority of lwkt_schedule() itself). |
| 564 | * |
| 565 | * XXX at the moment we run the target thread in a critical section during |
| 566 | * the preemption in order to prevent the target from taking interrupts |
| 567 | * that *WE* can't. Preemption is strictly limited to interrupt threads |
| 568 | * and interrupt-like threads, outside of a critical section, and the |
| 569 | * preempted source thread will be resumed the instant the target blocks |
| 570 | * whether or not the source is scheduled (i.e. preemption is supposed to |
| 571 | * be as transparent as possible). |
| 572 | * |
| 573 | * The target thread inherits our MP count (added to its own) for the |
| 574 | * duration of the preemption in order to preserve the atomicy of the |
| 575 | * MP lock during the preemption. Therefore, any preempting targets must be |
| 576 | * careful in regards to MP assertions. Note that the MP count may be |
| 577 | * out of sync with the physical mp_lock, but we do not have to preserve |
| 578 | * the original ownership of the lock if it was out of synch (that is, we |
| 579 | * can leave it synchronized on return). |
| 580 | */ |
| 581 | void |
| 582 | lwkt_preempt(thread_t ntd, int critpri) |
| 583 | { |
| 584 | struct globaldata *gd = mycpu; |
| 585 | thread_t td = gd->gd_curthread; |
| 586 | #ifdef SMP |
| 587 | int mpheld; |
| 588 | int savecnt; |
| 589 | #endif |
| 590 | |
| 591 | /* |
| 592 | * The caller has put us in a critical section. We can only preempt |
| 593 | * if the caller of the caller was not in a critical section (basically |
| 594 | * a local interrupt), as determined by the 'critpri' parameter. If |
| 595 | * we are unable to preempt |
| 596 | * |
| 597 | * YYY The target thread must be in a critical section (else it must |
| 598 | * inherit our critical section? I dunno yet). |
| 599 | */ |
| 600 | KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri)); |
| 601 | |
| 602 | need_resched(); |
| 603 | if (!_lwkt_wantresched(ntd, td)) { |
| 604 | ++preempt_miss; |
| 605 | return; |
| 606 | } |
| 607 | if ((td->td_pri & ~TDPRI_MASK) > critpri) { |
| 608 | ++preempt_miss; |
| 609 | return; |
| 610 | } |
| 611 | #ifdef SMP |
| 612 | if (ntd->td_gd != gd) { |
| 613 | ++preempt_miss; |
| 614 | return; |
| 615 | } |
| 616 | #endif |
| 617 | if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) { |
| 618 | ++preempt_weird; |
| 619 | return; |
| 620 | } |
| 621 | if (ntd->td_preempted) { |
| 622 | ++preempt_hit; |
| 623 | return; |
| 624 | } |
| 625 | #ifdef SMP |
| 626 | /* |
| 627 | * note: an interrupt might have occured just as we were transitioning |
| 628 | * to or from the MP lock. In this case td_mpcount will be pre-disposed |
| 629 | * (non-zero) but not actually synchronized with the actual state of the |
| 630 | * lock. We can use it to imply an MP lock requirement for the |
| 631 | * preemption but we cannot use it to test whether we hold the MP lock |
| 632 | * or not. |
| 633 | */ |
| 634 | savecnt = td->td_mpcount; |
| 635 | mpheld = MP_LOCK_HELD(); |
| 636 | ntd->td_mpcount += td->td_mpcount; |
| 637 | if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) { |
| 638 | ntd->td_mpcount -= td->td_mpcount; |
| 639 | ++preempt_miss; |
| 640 | return; |
| 641 | } |
| 642 | #endif |
| 643 | |
| 644 | ++preempt_hit; |
| 645 | ntd->td_preempted = td; |
| 646 | td->td_flags |= TDF_PREEMPT_LOCK; |
| 647 | td->td_switch(ntd); |
| 648 | KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE)); |
| 649 | #ifdef SMP |
| 650 | KKASSERT(savecnt == td->td_mpcount); |
| 651 | mpheld = MP_LOCK_HELD(); |
| 652 | if (mpheld && td->td_mpcount == 0) |
| 653 | cpu_rel_mplock(); |
| 654 | else if (mpheld == 0 && td->td_mpcount) |
| 655 | panic("lwkt_preempt(): MP lock was not held through"); |
| 656 | #endif |
| 657 | ntd->td_preempted = NULL; |
| 658 | td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE); |
| 659 | } |
| 660 | |
| 661 | /* |
| 662 | * Yield our thread while higher priority threads are pending. This is |
| 663 | * typically called when we leave a critical section but it can be safely |
| 664 | * called while we are in a critical section. |
| 665 | * |
| 666 | * This function will not generally yield to equal priority threads but it |
| 667 | * can occur as a side effect. Note that lwkt_switch() is called from |
| 668 | * inside the critical section to prevent its own crit_exit() from reentering |
| 669 | * lwkt_yield_quick(). |
| 670 | * |
| 671 | * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint |
| 672 | * came along but was blocked and made pending. |
| 673 | * |
| 674 | * (self contained on a per cpu basis) |
| 675 | */ |
| 676 | void |
| 677 | lwkt_yield_quick(void) |
| 678 | { |
| 679 | globaldata_t gd = mycpu; |
| 680 | thread_t td = gd->gd_curthread; |
| 681 | |
| 682 | /* |
| 683 | * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear |
| 684 | * it with a non-zero cpl then we might not wind up calling splz after |
| 685 | * a task switch when the critical section is exited even though the |
| 686 | * new task could accept the interrupt. |
| 687 | * |
| 688 | * XXX from crit_exit() only called after last crit section is released. |
| 689 | * If called directly will run splz() even if in a critical section. |
| 690 | * |
| 691 | * td_nest_count prevent deep nesting via splz() or doreti(). Note that |
| 692 | * except for this special case, we MUST call splz() here to handle any |
| 693 | * pending ints, particularly after we switch, or we might accidently |
| 694 | * halt the cpu with interrupts pending. |
| 695 | */ |
| 696 | if (gd->gd_reqflags && td->td_nest_count < 2) |
| 697 | splz(); |
| 698 | |
| 699 | /* |
| 700 | * YYY enabling will cause wakeup() to task-switch, which really |
| 701 | * confused the old 4.x code. This is a good way to simulate |
| 702 | * preemption and MP without actually doing preemption or MP, because a |
| 703 | * lot of code assumes that wakeup() does not block. |
| 704 | */ |
| 705 | if (untimely_switch && td->td_nest_count == 0 && |
| 706 | gd->gd_intr_nesting_level == 0 |
| 707 | ) { |
| 708 | crit_enter(); |
| 709 | /* |
| 710 | * YYY temporary hacks until we disassociate the userland scheduler |
| 711 | * from the LWKT scheduler. |
| 712 | */ |
| 713 | if (td->td_flags & TDF_RUNQ) { |
| 714 | lwkt_switch(); /* will not reenter yield function */ |
| 715 | } else { |
| 716 | lwkt_schedule_self(); /* make sure we are scheduled */ |
| 717 | lwkt_switch(); /* will not reenter yield function */ |
| 718 | lwkt_deschedule_self(); /* make sure we are descheduled */ |
| 719 | } |
| 720 | crit_exit_noyield(td); |
| 721 | } |
| 722 | } |
| 723 | |
| 724 | /* |
| 725 | * This implements a normal yield which, unlike _quick, will yield to equal |
| 726 | * priority threads as well. Note that gd_reqflags tests will be handled by |
| 727 | * the crit_exit() call in lwkt_switch(). |
| 728 | * |
| 729 | * (self contained on a per cpu basis) |
| 730 | */ |
| 731 | void |
| 732 | lwkt_yield(void) |
| 733 | { |
| 734 | lwkt_schedule_self(); |
| 735 | lwkt_switch(); |
| 736 | } |
| 737 | |
| 738 | /* |
| 739 | * Schedule a thread to run. As the current thread we can always safely |
| 740 | * schedule ourselves, and a shortcut procedure is provided for that |
| 741 | * function. |
| 742 | * |
| 743 | * (non-blocking, self contained on a per cpu basis) |
| 744 | */ |
| 745 | void |
| 746 | lwkt_schedule_self(void) |
| 747 | { |
| 748 | thread_t td = curthread; |
| 749 | |
| 750 | crit_enter(); |
| 751 | KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!")); |
| 752 | _lwkt_enqueue(td); |
| 753 | #ifdef _KERNEL |
| 754 | if (td->td_proc && td->td_proc->p_stat == SSLEEP) |
| 755 | panic("SCHED SELF PANIC"); |
| 756 | #endif |
| 757 | crit_exit(); |
| 758 | } |
| 759 | |
| 760 | /* |
| 761 | * Generic schedule. Possibly schedule threads belonging to other cpus and |
| 762 | * deal with threads that might be blocked on a wait queue. |
| 763 | * |
| 764 | * YYY this is one of the best places to implement load balancing code. |
| 765 | * Load balancing can be accomplished by requesting other sorts of actions |
| 766 | * for the thread in question. |
| 767 | */ |
| 768 | void |
| 769 | lwkt_schedule(thread_t td) |
| 770 | { |
| 771 | #ifdef INVARIANTS |
| 772 | if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc |
| 773 | && td->td_proc->p_stat == SSLEEP |
| 774 | ) { |
| 775 | printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n", |
| 776 | curthread, |
| 777 | curthread->td_proc ? curthread->td_proc->p_pid : -1, |
| 778 | curthread->td_proc ? curthread->td_proc->p_stat : -1, |
| 779 | td, |
| 780 | td->td_proc ? curthread->td_proc->p_pid : -1, |
| 781 | td->td_proc ? curthread->td_proc->p_stat : -1 |
| 782 | ); |
| 783 | panic("SCHED PANIC"); |
| 784 | } |
| 785 | #endif |
| 786 | crit_enter(); |
| 787 | if (td == curthread) { |
| 788 | _lwkt_enqueue(td); |
| 789 | } else { |
| 790 | lwkt_wait_t w; |
| 791 | |
| 792 | /* |
| 793 | * If the thread is on a wait list we have to send our scheduling |
| 794 | * request to the owner of the wait structure. Otherwise we send |
| 795 | * the scheduling request to the cpu owning the thread. Races |
| 796 | * are ok, the target will forward the message as necessary (the |
| 797 | * message may chase the thread around before it finally gets |
| 798 | * acted upon). |
| 799 | * |
| 800 | * (remember, wait structures use stable storage) |
| 801 | */ |
| 802 | if ((w = td->td_wait) != NULL) { |
| 803 | if (lwkt_trytoken(&w->wa_token)) { |
| 804 | TAILQ_REMOVE(&w->wa_waitq, td, td_threadq); |
| 805 | --w->wa_count; |
| 806 | td->td_wait = NULL; |
| 807 | if (smp_active == 0 || td->td_gd == mycpu) { |
| 808 | _lwkt_enqueue(td); |
| 809 | if (td->td_preemptable) { |
| 810 | td->td_preemptable(td, TDPRI_CRIT*2); /* YYY +token */ |
| 811 | } else if (_lwkt_wantresched(td, curthread)) { |
| 812 | need_resched(); |
| 813 | } |
| 814 | } else { |
| 815 | lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td); |
| 816 | } |
| 817 | lwkt_reltoken(&w->wa_token); |
| 818 | } else { |
| 819 | lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td); |
| 820 | } |
| 821 | } else { |
| 822 | /* |
| 823 | * If the wait structure is NULL and we own the thread, there |
| 824 | * is no race (since we are in a critical section). If we |
| 825 | * do not own the thread there might be a race but the |
| 826 | * target cpu will deal with it. |
| 827 | */ |
| 828 | if (smp_active == 0 || td->td_gd == mycpu) { |
| 829 | _lwkt_enqueue(td); |
| 830 | if (td->td_preemptable) { |
| 831 | td->td_preemptable(td, TDPRI_CRIT); |
| 832 | } else if (_lwkt_wantresched(td, curthread)) { |
| 833 | need_resched(); |
| 834 | } |
| 835 | } else { |
| 836 | lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td); |
| 837 | } |
| 838 | } |
| 839 | } |
| 840 | crit_exit(); |
| 841 | } |
| 842 | |
| 843 | /* |
| 844 | * Managed acquisition. This code assumes that the MP lock is held for |
| 845 | * the tdallq operation and that the thread has been descheduled from its |
| 846 | * original cpu. We also have to wait for the thread to be entirely switched |
| 847 | * out on its original cpu (this is usually fast enough that we never loop) |
| 848 | * since the LWKT system does not have to hold the MP lock while switching |
| 849 | * and the target may have released it before switching. |
| 850 | */ |
| 851 | void |
| 852 | lwkt_acquire(thread_t td) |
| 853 | { |
| 854 | struct globaldata *gd; |
| 855 | |
| 856 | gd = td->td_gd; |
| 857 | KKASSERT((td->td_flags & TDF_RUNQ) == 0); |
| 858 | while (td->td_flags & TDF_RUNNING) /* XXX spin */ |
| 859 | ; |
| 860 | if (gd != mycpu) { |
| 861 | crit_enter(); |
| 862 | TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */ |
| 863 | gd = mycpu; |
| 864 | td->td_gd = gd; |
| 865 | TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); /* protected by BGL */ |
| 866 | crit_exit(); |
| 867 | } |
| 868 | } |
| 869 | |
| 870 | /* |
| 871 | * Deschedule a thread. |
| 872 | * |
| 873 | * (non-blocking, self contained on a per cpu basis) |
| 874 | */ |
| 875 | void |
| 876 | lwkt_deschedule_self(void) |
| 877 | { |
| 878 | thread_t td = curthread; |
| 879 | |
| 880 | crit_enter(); |
| 881 | KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!")); |
| 882 | _lwkt_dequeue(td); |
| 883 | crit_exit(); |
| 884 | } |
| 885 | |
| 886 | /* |
| 887 | * Generic deschedule. Descheduling threads other then your own should be |
| 888 | * done only in carefully controlled circumstances. Descheduling is |
| 889 | * asynchronous. |
| 890 | * |
| 891 | * This function may block if the cpu has run out of messages. |
| 892 | */ |
| 893 | void |
| 894 | lwkt_deschedule(thread_t td) |
| 895 | { |
| 896 | crit_enter(); |
| 897 | if (td == curthread) { |
| 898 | _lwkt_dequeue(td); |
| 899 | } else { |
| 900 | if (td->td_gd == mycpu) { |
| 901 | _lwkt_dequeue(td); |
| 902 | } else { |
| 903 | lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_deschedule, td); |
| 904 | } |
| 905 | } |
| 906 | crit_exit(); |
| 907 | } |
| 908 | |
| 909 | /* |
| 910 | * Set the target thread's priority. This routine does not automatically |
| 911 | * switch to a higher priority thread, LWKT threads are not designed for |
| 912 | * continuous priority changes. Yield if you want to switch. |
| 913 | * |
| 914 | * We have to retain the critical section count which uses the high bits |
| 915 | * of the td_pri field. The specified priority may also indicate zero or |
| 916 | * more critical sections by adding TDPRI_CRIT*N. |
| 917 | */ |
| 918 | void |
| 919 | lwkt_setpri(thread_t td, int pri) |
| 920 | { |
| 921 | KKASSERT(pri >= 0); |
| 922 | KKASSERT(td->td_gd == mycpu); |
| 923 | crit_enter(); |
| 924 | if (td->td_flags & TDF_RUNQ) { |
| 925 | _lwkt_dequeue(td); |
| 926 | td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri; |
| 927 | _lwkt_enqueue(td); |
| 928 | } else { |
| 929 | td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri; |
| 930 | } |
| 931 | crit_exit(); |
| 932 | } |
| 933 | |
| 934 | void |
| 935 | lwkt_setpri_self(int pri) |
| 936 | { |
| 937 | thread_t td = curthread; |
| 938 | |
| 939 | KKASSERT(pri >= 0 && pri <= TDPRI_MAX); |
| 940 | crit_enter(); |
| 941 | if (td->td_flags & TDF_RUNQ) { |
| 942 | _lwkt_dequeue(td); |
| 943 | td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri; |
| 944 | _lwkt_enqueue(td); |
| 945 | } else { |
| 946 | td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri; |
| 947 | } |
| 948 | crit_exit(); |
| 949 | } |
| 950 | |
| 951 | struct proc * |
| 952 | lwkt_preempted_proc(void) |
| 953 | { |
| 954 | thread_t td = curthread; |
| 955 | while (td->td_preempted) |
| 956 | td = td->td_preempted; |
| 957 | return(td->td_proc); |
| 958 | } |
| 959 | |
| 960 | #if 0 |
| 961 | |
| 962 | /* |
| 963 | * This function deschedules the current thread and blocks on the specified |
| 964 | * wait queue. We obtain ownership of the wait queue in order to block |
| 965 | * on it. A generation number is used to interlock the wait queue in case |
| 966 | * it gets signalled while we are blocked waiting on the token. |
| 967 | * |
| 968 | * Note: alternatively we could dequeue our thread and then message the |
| 969 | * target cpu owning the wait queue. YYY implement as sysctl. |
| 970 | * |
| 971 | * Note: wait queue signals normally ping-pong the cpu as an optimization. |
| 972 | */ |
| 973 | |
| 974 | void |
| 975 | lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen) |
| 976 | { |
| 977 | thread_t td = curthread; |
| 978 | |
| 979 | lwkt_gettoken(&w->wa_token); |
| 980 | if (w->wa_gen == *gen) { |
| 981 | _lwkt_dequeue(td); |
| 982 | TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq); |
| 983 | ++w->wa_count; |
| 984 | td->td_wait = w; |
| 985 | td->td_wmesg = wmesg; |
| 986 | again: |
| 987 | lwkt_switch(); |
| 988 | lwkt_regettoken(&w->wa_token); |
| 989 | if (td->td_wmesg != NULL) { |
| 990 | _lwkt_dequeue(td); |
| 991 | goto again; |
| 992 | } |
| 993 | } |
| 994 | /* token might be lost, doesn't matter for gen update */ |
| 995 | *gen = w->wa_gen; |
| 996 | lwkt_reltoken(&w->wa_token); |
| 997 | } |
| 998 | |
| 999 | /* |
| 1000 | * Signal a wait queue. We gain ownership of the wait queue in order to |
| 1001 | * signal it. Once a thread is removed from the wait queue we have to |
| 1002 | * deal with the cpu owning the thread. |
| 1003 | * |
| 1004 | * Note: alternatively we could message the target cpu owning the wait |
| 1005 | * queue. YYY implement as sysctl. |
| 1006 | */ |
| 1007 | void |
| 1008 | lwkt_signal(lwkt_wait_t w, int count) |
| 1009 | { |
| 1010 | thread_t td; |
| 1011 | int count; |
| 1012 | |
| 1013 | lwkt_gettoken(&w->wa_token); |
| 1014 | ++w->wa_gen; |
| 1015 | if (count < 0) |
| 1016 | count = w->wa_count; |
| 1017 | while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) { |
| 1018 | --count; |
| 1019 | --w->wa_count; |
| 1020 | TAILQ_REMOVE(&w->wa_waitq, td, td_threadq); |
| 1021 | td->td_wait = NULL; |
| 1022 | td->td_wmesg = NULL; |
| 1023 | if (td->td_gd == mycpu) { |
| 1024 | _lwkt_enqueue(td); |
| 1025 | } else { |
| 1026 | lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td); |
| 1027 | } |
| 1028 | lwkt_regettoken(&w->wa_token); |
| 1029 | } |
| 1030 | lwkt_reltoken(&w->wa_token); |
| 1031 | } |
| 1032 | |
| 1033 | #endif |
| 1034 | |
| 1035 | /* |
| 1036 | * Create a kernel process/thread/whatever. It shares it's address space |
| 1037 | * with proc0 - ie: kernel only. |
| 1038 | * |
| 1039 | * NOTE! By default new threads are created with the MP lock held. A |
| 1040 | * thread which does not require the MP lock should release it by calling |
| 1041 | * rel_mplock() at the start of the new thread. |
| 1042 | */ |
| 1043 | int |
| 1044 | lwkt_create(void (*func)(void *), void *arg, |
| 1045 | struct thread **tdp, thread_t template, int tdflags, int cpu, |
| 1046 | const char *fmt, ...) |
| 1047 | { |
| 1048 | thread_t td; |
| 1049 | __va_list ap; |
| 1050 | |
| 1051 | td = lwkt_alloc_thread(template, cpu); |
| 1052 | if (tdp) |
| 1053 | *tdp = td; |
| 1054 | cpu_set_thread_handler(td, lwkt_exit, func, arg); |
| 1055 | td->td_flags |= TDF_VERBOSE | tdflags; |
| 1056 | #ifdef SMP |
| 1057 | td->td_mpcount = 1; |
| 1058 | #endif |
| 1059 | |
| 1060 | /* |
| 1061 | * Set up arg0 for 'ps' etc |
| 1062 | */ |
| 1063 | __va_start(ap, fmt); |
| 1064 | vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap); |
| 1065 | __va_end(ap); |
| 1066 | |
| 1067 | /* |
| 1068 | * Schedule the thread to run |
| 1069 | */ |
| 1070 | if ((td->td_flags & TDF_STOPREQ) == 0) |
| 1071 | lwkt_schedule(td); |
| 1072 | else |
| 1073 | td->td_flags &= ~TDF_STOPREQ; |
| 1074 | return 0; |
| 1075 | } |
| 1076 | |
| 1077 | /* |
| 1078 | * kthread_* is specific to the kernel and is not needed by userland. |
| 1079 | */ |
| 1080 | #ifdef _KERNEL |
| 1081 | |
| 1082 | /* |
| 1083 | * Destroy an LWKT thread. Warning! This function is not called when |
| 1084 | * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and |
| 1085 | * uses a different reaping mechanism. |
| 1086 | */ |
| 1087 | void |
| 1088 | lwkt_exit(void) |
| 1089 | { |
| 1090 | thread_t td = curthread; |
| 1091 | |
| 1092 | if (td->td_flags & TDF_VERBOSE) |
| 1093 | printf("kthread %p %s has exited\n", td, td->td_comm); |
| 1094 | caps_exit(td); |
| 1095 | crit_enter(); |
| 1096 | lwkt_deschedule_self(); |
| 1097 | ++mycpu->gd_tdfreecount; |
| 1098 | TAILQ_INSERT_TAIL(&mycpu->gd_tdfreeq, td, td_threadq); |
| 1099 | cpu_thread_exit(); |
| 1100 | } |
| 1101 | |
| 1102 | /* |
| 1103 | * Create a kernel process/thread/whatever. It shares it's address space |
| 1104 | * with proc0 - ie: kernel only. 5.x compatible. |
| 1105 | * |
| 1106 | * NOTE! By default kthreads are created with the MP lock held. A |
| 1107 | * thread which does not require the MP lock should release it by calling |
| 1108 | * rel_mplock() at the start of the new thread. |
| 1109 | */ |
| 1110 | int |
| 1111 | kthread_create(void (*func)(void *), void *arg, |
| 1112 | struct thread **tdp, const char *fmt, ...) |
| 1113 | { |
| 1114 | thread_t td; |
| 1115 | __va_list ap; |
| 1116 | |
| 1117 | td = lwkt_alloc_thread(NULL, -1); |
| 1118 | if (tdp) |
| 1119 | *tdp = td; |
| 1120 | cpu_set_thread_handler(td, kthread_exit, func, arg); |
| 1121 | td->td_flags |= TDF_VERBOSE; |
| 1122 | #ifdef SMP |
| 1123 | td->td_mpcount = 1; |
| 1124 | #endif |
| 1125 | |
| 1126 | /* |
| 1127 | * Set up arg0 for 'ps' etc |
| 1128 | */ |
| 1129 | __va_start(ap, fmt); |
| 1130 | vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap); |
| 1131 | __va_end(ap); |
| 1132 | |
| 1133 | /* |
| 1134 | * Schedule the thread to run |
| 1135 | */ |
| 1136 | lwkt_schedule(td); |
| 1137 | return 0; |
| 1138 | } |
| 1139 | |
| 1140 | /* |
| 1141 | * Destroy an LWKT thread. Warning! This function is not called when |
| 1142 | * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and |
| 1143 | * uses a different reaping mechanism. |
| 1144 | * |
| 1145 | * XXX duplicates lwkt_exit() |
| 1146 | */ |
| 1147 | void |
| 1148 | kthread_exit(void) |
| 1149 | { |
| 1150 | lwkt_exit(); |
| 1151 | } |
| 1152 | |
| 1153 | #endif /* _KERNEL */ |
| 1154 | |
| 1155 | void |
| 1156 | crit_panic(void) |
| 1157 | { |
| 1158 | thread_t td = curthread; |
| 1159 | int lpri = td->td_pri; |
| 1160 | |
| 1161 | td->td_pri = 0; |
| 1162 | panic("td_pri is/would-go negative! %p %d", td, lpri); |
| 1163 | } |
| 1164 | |
| 1165 | #ifdef SMP |
| 1166 | |
| 1167 | /* |
| 1168 | * Send a function execution request to another cpu. The request is queued |
| 1169 | * on the cpu<->cpu ipiq matrix. Each cpu owns a unique ipiq FIFO for every |
| 1170 | * possible target cpu. The FIFO can be written. |
| 1171 | * |
| 1172 | * YYY If the FIFO fills up we have to enable interrupts and process the |
| 1173 | * IPIQ while waiting for it to empty or we may deadlock with another cpu. |
| 1174 | * Create a CPU_*() function to do this! |
| 1175 | * |
| 1176 | * We can safely bump gd_intr_nesting_level because our crit_exit() at the |
| 1177 | * end will take care of any pending interrupts. |
| 1178 | * |
| 1179 | * Must be called from a critical section. |
| 1180 | */ |
| 1181 | int |
| 1182 | lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg) |
| 1183 | { |
| 1184 | lwkt_ipiq_t ip; |
| 1185 | int windex; |
| 1186 | struct globaldata *gd = mycpu; |
| 1187 | |
| 1188 | if (dcpu == gd->gd_cpuid) { |
| 1189 | func(arg); |
| 1190 | return(0); |
| 1191 | } |
| 1192 | crit_enter(); |
| 1193 | ++gd->gd_intr_nesting_level; |
| 1194 | #ifdef INVARIANTS |
| 1195 | if (gd->gd_intr_nesting_level > 20) |
| 1196 | panic("lwkt_send_ipiq: TOO HEAVILY NESTED!"); |
| 1197 | #endif |
| 1198 | KKASSERT(curthread->td_pri >= TDPRI_CRIT); |
| 1199 | KKASSERT(dcpu >= 0 && dcpu < ncpus); |
| 1200 | ++ipiq_count; |
| 1201 | ip = &gd->gd_ipiq[dcpu]; |
| 1202 | |
| 1203 | /* |
| 1204 | * We always drain before the FIFO becomes full so it should never |
| 1205 | * become full. We need to leave enough entries to deal with |
| 1206 | * reentrancy. |
| 1207 | */ |
| 1208 | KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO); |
| 1209 | windex = ip->ip_windex & MAXCPUFIFO_MASK; |
| 1210 | ip->ip_func[windex] = func; |
| 1211 | ip->ip_arg[windex] = arg; |
| 1212 | /* YYY memory barrier */ |
| 1213 | ++ip->ip_windex; |
| 1214 | if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) { |
| 1215 | unsigned int eflags = read_eflags(); |
| 1216 | cpu_enable_intr(); |
| 1217 | ++ipiq_fifofull; |
| 1218 | while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) { |
| 1219 | KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1); |
| 1220 | lwkt_process_ipiq(); |
| 1221 | } |
| 1222 | write_eflags(eflags); |
| 1223 | } |
| 1224 | --gd->gd_intr_nesting_level; |
| 1225 | cpu_send_ipiq(dcpu); /* issues memory barrier if appropriate */ |
| 1226 | crit_exit(); |
| 1227 | return(ip->ip_windex); |
| 1228 | } |
| 1229 | |
| 1230 | /* |
| 1231 | * Send a message to several target cpus. Typically used for scheduling. |
| 1232 | * The message will not be sent to stopped cpus. |
| 1233 | */ |
| 1234 | void |
| 1235 | lwkt_send_ipiq_mask(u_int32_t mask, ipifunc_t func, void *arg) |
| 1236 | { |
| 1237 | int cpuid; |
| 1238 | |
| 1239 | mask &= ~stopped_cpus; |
| 1240 | while (mask) { |
| 1241 | cpuid = bsfl(mask); |
| 1242 | lwkt_send_ipiq(cpuid, func, arg); |
| 1243 | mask &= ~(1 << cpuid); |
| 1244 | } |
| 1245 | } |
| 1246 | |
| 1247 | /* |
| 1248 | * Wait for the remote cpu to finish processing a function. |
| 1249 | * |
| 1250 | * YYY we have to enable interrupts and process the IPIQ while waiting |
| 1251 | * for it to empty or we may deadlock with another cpu. Create a CPU_*() |
| 1252 | * function to do this! YYY we really should 'block' here. |
| 1253 | * |
| 1254 | * Must be called from a critical section. Thsi routine may be called |
| 1255 | * from an interrupt (for example, if an interrupt wakes a foreign thread |
| 1256 | * up). |
| 1257 | */ |
| 1258 | void |
| 1259 | lwkt_wait_ipiq(int dcpu, int seq) |
| 1260 | { |
| 1261 | lwkt_ipiq_t ip; |
| 1262 | int maxc = 100000000; |
| 1263 | |
| 1264 | if (dcpu != mycpu->gd_cpuid) { |
| 1265 | KKASSERT(dcpu >= 0 && dcpu < ncpus); |
| 1266 | ip = &mycpu->gd_ipiq[dcpu]; |
| 1267 | if ((int)(ip->ip_xindex - seq) < 0) { |
| 1268 | unsigned int eflags = read_eflags(); |
| 1269 | cpu_enable_intr(); |
| 1270 | while ((int)(ip->ip_xindex - seq) < 0) { |
| 1271 | lwkt_process_ipiq(); |
| 1272 | if (--maxc == 0) |
| 1273 | printf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n", mycpu->gd_cpuid, dcpu, ip->ip_xindex - seq); |
| 1274 | if (maxc < -1000000) |
| 1275 | panic("LWKT_WAIT_IPIQ"); |
| 1276 | } |
| 1277 | write_eflags(eflags); |
| 1278 | } |
| 1279 | } |
| 1280 | } |
| 1281 | |
| 1282 | /* |
| 1283 | * Called from IPI interrupt (like a fast interrupt), which has placed |
| 1284 | * us in a critical section. The MP lock may or may not be held. |
| 1285 | * May also be called from doreti or splz, or be reentrantly called |
| 1286 | * indirectly through the ip_func[] we run. |
| 1287 | * |
| 1288 | * There are two versions, one where no interrupt frame is available (when |
| 1289 | * called from the send code and from splz, and one where an interrupt |
| 1290 | * frame is available. |
| 1291 | */ |
| 1292 | void |
| 1293 | lwkt_process_ipiq(void) |
| 1294 | { |
| 1295 | int n; |
| 1296 | int cpuid = mycpu->gd_cpuid; |
| 1297 | |
| 1298 | for (n = 0; n < ncpus; ++n) { |
| 1299 | lwkt_ipiq_t ip; |
| 1300 | int ri; |
| 1301 | |
| 1302 | if (n == cpuid) |
| 1303 | continue; |
| 1304 | ip = globaldata_find(n)->gd_ipiq; |
| 1305 | if (ip == NULL) |
| 1306 | continue; |
| 1307 | ip = &ip[cpuid]; |
| 1308 | |
| 1309 | /* |
| 1310 | * Note: xindex is only updated after we are sure the function has |
| 1311 | * finished execution. Beware lwkt_process_ipiq() reentrancy! The |
| 1312 | * function may send an IPI which may block/drain. |
| 1313 | */ |
| 1314 | while (ip->ip_rindex != ip->ip_windex) { |
| 1315 | ri = ip->ip_rindex & MAXCPUFIFO_MASK; |
| 1316 | ++ip->ip_rindex; |
| 1317 | ip->ip_func[ri](ip->ip_arg[ri], NULL); |
| 1318 | /* YYY memory barrier */ |
| 1319 | ip->ip_xindex = ip->ip_rindex; |
| 1320 | } |
| 1321 | } |
| 1322 | } |
| 1323 | |
| 1324 | #ifdef _KERNEL |
| 1325 | void |
| 1326 | lwkt_process_ipiq_frame(struct intrframe frame) |
| 1327 | { |
| 1328 | int n; |
| 1329 | int cpuid = mycpu->gd_cpuid; |
| 1330 | |
| 1331 | for (n = 0; n < ncpus; ++n) { |
| 1332 | lwkt_ipiq_t ip; |
| 1333 | int ri; |
| 1334 | |
| 1335 | if (n == cpuid) |
| 1336 | continue; |
| 1337 | ip = globaldata_find(n)->gd_ipiq; |
| 1338 | if (ip == NULL) |
| 1339 | continue; |
| 1340 | ip = &ip[cpuid]; |
| 1341 | |
| 1342 | /* |
| 1343 | * Note: xindex is only updated after we are sure the function has |
| 1344 | * finished execution. Beware lwkt_process_ipiq() reentrancy! The |
| 1345 | * function may send an IPI which may block/drain. |
| 1346 | */ |
| 1347 | while (ip->ip_rindex != ip->ip_windex) { |
| 1348 | ri = ip->ip_rindex & MAXCPUFIFO_MASK; |
| 1349 | ++ip->ip_rindex; |
| 1350 | ip->ip_func[ri](ip->ip_arg[ri], &frame); |
| 1351 | /* YYY memory barrier */ |
| 1352 | ip->ip_xindex = ip->ip_rindex; |
| 1353 | } |
| 1354 | } |
| 1355 | } |
| 1356 | #endif |
| 1357 | |
| 1358 | #else |
| 1359 | |
| 1360 | int |
| 1361 | lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg) |
| 1362 | { |
| 1363 | panic("lwkt_send_ipiq: UP box! (%d,%p,%p)", dcpu, func, arg); |
| 1364 | return(0); /* NOT REACHED */ |
| 1365 | } |
| 1366 | |
| 1367 | void |
| 1368 | lwkt_wait_ipiq(int dcpu, int seq) |
| 1369 | { |
| 1370 | panic("lwkt_wait_ipiq: UP box! (%d,%d)", dcpu, seq); |
| 1371 | } |
| 1372 | |
| 1373 | #endif |