1 /* 2 * Copyright (c) 2006 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/usched_dummy.c,v 1.9 2008/04/21 15:24:46 dillon Exp $ 35 */ 36 37 #include <sys/param.h> 38 #include <sys/systm.h> 39 #include <sys/kernel.h> 40 #include <sys/lock.h> 41 #include <sys/queue.h> 42 #include <sys/proc.h> 43 #include <sys/rtprio.h> 44 #include <sys/uio.h> 45 #include <sys/sysctl.h> 46 #include <sys/resourcevar.h> 47 #include <sys/spinlock.h> 48 #include <machine/cpu.h> 49 #include <machine/smp.h> 50 51 #include <sys/thread2.h> 52 #include <sys/spinlock2.h> 53 #include <sys/mplock2.h> 54 55 #define MAXPRI 128 56 #define PRIBASE_REALTIME 0 57 #define PRIBASE_NORMAL MAXPRI 58 #define PRIBASE_IDLE (MAXPRI * 2) 59 #define PRIBASE_THREAD (MAXPRI * 3) 60 #define PRIBASE_NULL (MAXPRI * 4) 61 62 #define lwp_priority lwp_usdata.bsd4.priority 63 #define lwp_estcpu lwp_usdata.bsd4.estcpu 64 65 static void dummy_acquire_curproc(struct lwp *lp); 66 static void dummy_release_curproc(struct lwp *lp); 67 static void dummy_select_curproc(globaldata_t gd); 68 static void dummy_setrunqueue(struct lwp *lp); 69 static void dummy_schedulerclock(struct lwp *lp, sysclock_t period, 70 sysclock_t cpstamp); 71 static void dummy_recalculate_estcpu(struct lwp *lp); 72 static void dummy_resetpriority(struct lwp *lp); 73 static void dummy_forking(struct lwp *plp, struct lwp *lp); 74 static void dummy_exiting(struct lwp *plp, struct proc *child); 75 static void dummy_yield(struct lwp *lp); 76 77 struct usched usched_dummy = { 78 { NULL }, 79 "dummy", "Dummy DragonFly Scheduler", 80 NULL, /* default registration */ 81 NULL, /* default deregistration */ 82 dummy_acquire_curproc, 83 dummy_release_curproc, 84 dummy_setrunqueue, 85 dummy_schedulerclock, 86 dummy_recalculate_estcpu, 87 dummy_resetpriority, 88 dummy_forking, 89 dummy_exiting, 90 NULL, /* setcpumask not supported */ 91 dummy_yield 92 }; 93 94 struct usched_dummy_pcpu { 95 int rrcount; 96 struct thread helper_thread; 97 struct lwp *uschedcp; 98 }; 99 100 typedef struct usched_dummy_pcpu *dummy_pcpu_t; 101 102 static struct usched_dummy_pcpu dummy_pcpu[MAXCPU]; 103 static cpumask_t dummy_curprocmask = -1; 104 static cpumask_t dummy_rdyprocmask; 105 static struct spinlock dummy_spin; 106 static TAILQ_HEAD(rq, lwp) dummy_runq; 107 static int dummy_runqcount; 108 109 static int usched_dummy_rrinterval = (ESTCPUFREQ + 9) / 10; 110 SYSCTL_INT(_kern, OID_AUTO, usched_dummy_rrinterval, CTLFLAG_RW, 111 &usched_dummy_rrinterval, 0, ""); 112 113 /* 114 * Initialize the run queues at boot time, clear cpu 0 in curprocmask 115 * to allow dummy scheduling on cpu 0. 116 */ 117 static void 118 dummyinit(void *dummy) 119 { 120 TAILQ_INIT(&dummy_runq); 121 spin_init(&dummy_spin); 122 atomic_clear_cpumask(&dummy_curprocmask, 1); 123 } 124 SYSINIT(runqueue, SI_BOOT2_USCHED, SI_ORDER_FIRST, dummyinit, NULL) 125 126 /* 127 * DUMMY_ACQUIRE_CURPROC 128 * 129 * This function is called when the kernel intends to return to userland. 130 * It is responsible for making the thread the current designated userland 131 * thread for this cpu, blocking if necessary. 132 * 133 * We are expected to handle userland reschedule requests here too. 134 * 135 * WARNING! THIS FUNCTION IS ALLOWED TO CAUSE THE CURRENT THREAD TO MIGRATE 136 * TO ANOTHER CPU! Because most of the kernel assumes that no migration will 137 * occur, this function is called only under very controlled circumstances. 138 * 139 * MPSAFE 140 */ 141 static void 142 dummy_acquire_curproc(struct lwp *lp) 143 { 144 globaldata_t gd = mycpu; 145 dummy_pcpu_t dd = &dummy_pcpu[gd->gd_cpuid]; 146 thread_t td = lp->lwp_thread; 147 148 /* 149 * Possibly select another thread 150 */ 151 if (user_resched_wanted()) 152 dummy_select_curproc(gd); 153 154 /* 155 * If this cpu has no current thread, select ourself 156 */ 157 if (dd->uschedcp == lp || 158 (dd->uschedcp == NULL && TAILQ_EMPTY(&dummy_runq))) { 159 atomic_set_cpumask(&dummy_curprocmask, gd->gd_cpumask); 160 dd->uschedcp = lp; 161 return; 162 } 163 164 /* 165 * If this cpu's current user process thread is not our thread, 166 * deschedule ourselves and place us on the run queue, then 167 * switch away. 168 * 169 * We loop until we become the current process. Its a good idea 170 * to run any passive release(s) before we mess with the scheduler 171 * so our thread is in the expected state. 172 */ 173 KKASSERT(dd->uschedcp != lp); 174 if (td->td_release) 175 td->td_release(lp->lwp_thread); 176 do { 177 crit_enter(); 178 lwkt_deschedule_self(td); 179 dummy_setrunqueue(lp); 180 if ((td->td_flags & TDF_RUNQ) == 0) 181 ++lp->lwp_ru.ru_nivcsw; 182 lwkt_switch(); /* WE MAY MIGRATE TO ANOTHER CPU */ 183 crit_exit(); 184 gd = mycpu; 185 dd = &dummy_pcpu[gd->gd_cpuid]; 186 KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0); 187 } while (dd->uschedcp != lp); 188 } 189 190 /* 191 * DUMMY_RELEASE_CURPROC 192 * 193 * This routine detaches the current thread from the userland scheduler, 194 * usually because the thread needs to run in the kernel (at kernel priority) 195 * for a while. 196 * 197 * This routine is also responsible for selecting a new thread to 198 * make the current thread. 199 * 200 * MPSAFE 201 */ 202 static void 203 dummy_release_curproc(struct lwp *lp) 204 { 205 globaldata_t gd = mycpu; 206 dummy_pcpu_t dd = &dummy_pcpu[gd->gd_cpuid]; 207 208 KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0); 209 if (dd->uschedcp == lp) { 210 dummy_select_curproc(gd); 211 } 212 } 213 214 /* 215 * DUMMY_SELECT_CURPROC 216 * 217 * Select a new current process for this cpu. This satisfies a user 218 * scheduler reschedule request so clear that too. 219 * 220 * This routine is also responsible for equal-priority round-robining, 221 * typically triggered from dummy_schedulerclock(). In our dummy example 222 * all the 'user' threads are LWKT scheduled all at once and we just 223 * call lwkt_switch(). 224 * 225 * MPSAFE 226 */ 227 static 228 void 229 dummy_select_curproc(globaldata_t gd) 230 { 231 dummy_pcpu_t dd = &dummy_pcpu[gd->gd_cpuid]; 232 struct lwp *lp; 233 234 clear_user_resched(); 235 spin_lock(&dummy_spin); 236 if ((lp = TAILQ_FIRST(&dummy_runq)) == NULL) { 237 dd->uschedcp = NULL; 238 atomic_clear_cpumask(&dummy_curprocmask, gd->gd_cpumask); 239 spin_unlock(&dummy_spin); 240 } else { 241 --dummy_runqcount; 242 TAILQ_REMOVE(&dummy_runq, lp, lwp_procq); 243 lp->lwp_flag &= ~LWP_ONRUNQ; 244 dd->uschedcp = lp; 245 atomic_set_cpumask(&dummy_curprocmask, gd->gd_cpumask); 246 spin_unlock(&dummy_spin); 247 #ifdef SMP 248 lwkt_acquire(lp->lwp_thread); 249 #endif 250 lwkt_schedule(lp->lwp_thread); 251 } 252 } 253 254 /* 255 * DUMMY_SETRUNQUEUE 256 * 257 * This routine is called to schedule a new user process after a fork. 258 * The scheduler module itself might also call this routine to place 259 * the current process on the userland scheduler's run queue prior 260 * to calling dummy_select_curproc(). 261 * 262 * The caller may set P_PASSIVE_ACQ in p_flag to indicate that we should 263 * attempt to leave the thread on the current cpu. 264 * 265 * MPSAFE 266 */ 267 static void 268 dummy_setrunqueue(struct lwp *lp) 269 { 270 globaldata_t gd = mycpu; 271 dummy_pcpu_t dd = &dummy_pcpu[gd->gd_cpuid]; 272 cpumask_t mask; 273 int cpuid; 274 275 if (dd->uschedcp == NULL) { 276 dd->uschedcp = lp; 277 atomic_set_cpumask(&dummy_curprocmask, gd->gd_cpumask); 278 lwkt_schedule(lp->lwp_thread); 279 } else { 280 /* 281 * Add to our global runq 282 */ 283 KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0); 284 spin_lock(&dummy_spin); 285 ++dummy_runqcount; 286 TAILQ_INSERT_TAIL(&dummy_runq, lp, lwp_procq); 287 lp->lwp_flag |= LWP_ONRUNQ; 288 #ifdef SMP 289 lwkt_giveaway(lp->lwp_thread); 290 #endif 291 292 /* lp = TAILQ_FIRST(&dummy_runq); */ 293 294 /* 295 * Notify the next available cpu. P.S. some 296 * cpu affinity could be done here. 297 * 298 * The rdyprocmask bit placeholds the knowledge that there 299 * is a process on the runq that needs service. If the 300 * helper thread cannot find a home for it it will forward 301 * the request to another available cpu. 302 */ 303 mask = ~dummy_curprocmask & dummy_rdyprocmask & 304 gd->gd_other_cpus; 305 if (mask) { 306 cpuid = BSFCPUMASK(mask); 307 atomic_clear_cpumask(&dummy_rdyprocmask, CPUMASK(cpuid)); 308 spin_unlock(&dummy_spin); 309 lwkt_schedule(&dummy_pcpu[cpuid].helper_thread); 310 } else { 311 spin_unlock(&dummy_spin); 312 } 313 } 314 } 315 316 /* 317 * This routine is called from a systimer IPI. Thus it is called with 318 * a critical section held. Any spinlocks we get here that are also 319 * obtained in other procedures must be proected by a critical section 320 * in those other procedures to avoid a deadlock. 321 * 322 * The MP lock may or may not be held on entry and cannot be obtained 323 * by this routine (because it is called from a systimer IPI). Additionally, 324 * because this is equivalent to a FAST interrupt, spinlocks cannot be used 325 * (or at least, you have to check that gd_spin* counts are 0 before you 326 * can). 327 * 328 * This routine is called at ESTCPUFREQ on each cpu independantly. 329 * 330 * This routine typically queues a reschedule request, which will cause 331 * the scheduler's BLAH_select_curproc() to be called as soon as possible. 332 * 333 * MPSAFE 334 */ 335 static 336 void 337 dummy_schedulerclock(struct lwp *lp, sysclock_t period, sysclock_t cpstamp) 338 { 339 globaldata_t gd = mycpu; 340 dummy_pcpu_t dd = &dummy_pcpu[gd->gd_cpuid]; 341 342 if (++dd->rrcount >= usched_dummy_rrinterval) { 343 dd->rrcount = 0; 344 need_user_resched(); 345 } 346 } 347 348 /* 349 * DUMMY_RECALCULATE_ESTCPU 350 * 351 * Called once a second for any process that is running or has slept 352 * for less then 2 seconds. 353 * 354 * MPSAFE 355 */ 356 static 357 void 358 dummy_recalculate_estcpu(struct lwp *lp) 359 { 360 } 361 362 /* 363 * MPSAFE 364 */ 365 static 366 void 367 dummy_yield(struct lwp *lp) 368 { 369 need_user_resched(); 370 } 371 372 /* 373 * DUMMY_RESETPRIORITY 374 * 375 * This routine is called after the kernel has potentially modified 376 * the lwp_rtprio structure. The target process may be running or sleeping 377 * or scheduled but not yet running or owned by another cpu. Basically, 378 * it can be in virtually any state. 379 * 380 * This routine is called by fork1() for initial setup with the process 381 * of the run queue, and also may be called normally with the process on or 382 * off the run queue. 383 * 384 * MPSAFE 385 */ 386 static void 387 dummy_resetpriority(struct lwp *lp) 388 { 389 /* XXX spinlock usually needed */ 390 /* 391 * Set p_priority for general process comparisons 392 */ 393 switch(lp->lwp_rtprio.type) { 394 case RTP_PRIO_REALTIME: 395 lp->lwp_priority = PRIBASE_REALTIME + lp->lwp_rtprio.prio; 396 return; 397 case RTP_PRIO_NORMAL: 398 lp->lwp_priority = PRIBASE_NORMAL + lp->lwp_rtprio.prio; 399 break; 400 case RTP_PRIO_IDLE: 401 lp->lwp_priority = PRIBASE_IDLE + lp->lwp_rtprio.prio; 402 return; 403 case RTP_PRIO_THREAD: 404 lp->lwp_priority = PRIBASE_THREAD + lp->lwp_rtprio.prio; 405 return; 406 } 407 /* XXX spinlock usually needed */ 408 } 409 410 411 /* 412 * DUMMY_FORKING 413 * 414 * Called from fork1() when a new child process is being created. Allows 415 * the scheduler to predispose the child process before it gets scheduled. 416 * 417 * MPSAFE 418 */ 419 static void 420 dummy_forking(struct lwp *plp, struct lwp *lp) 421 { 422 lp->lwp_estcpu = plp->lwp_estcpu; 423 #if 0 424 ++plp->lwp_estcpu; 425 #endif 426 } 427 428 /* 429 * DUMMY_EXITING 430 * 431 * Called when the parent reaps a child. Typically used to propogate cpu 432 * use by the child back to the parent as part of a batch detection 433 * heuristic. 434 * 435 * NOTE: cpu use is not normally back-propogated to PID 1. 436 * 437 * MPSAFE 438 */ 439 static void 440 dummy_exiting(struct lwp *plp, struct proc *child) 441 { 442 } 443 444 /* 445 * SMP systems may need a scheduler helper thread. This is how one can be 446 * setup. 447 * 448 * We use a neat LWKT scheduling trick to interlock the helper thread. It 449 * is possible to deschedule an LWKT thread and then do some work before 450 * switching away. The thread can be rescheduled at any time, even before 451 * we switch away. 452 * 453 * MPSAFE 454 */ 455 #ifdef SMP 456 457 static void 458 dummy_sched_thread(void *dummy) 459 { 460 globaldata_t gd; 461 dummy_pcpu_t dd; 462 struct lwp *lp; 463 cpumask_t cpumask; 464 cpumask_t tmpmask; 465 int cpuid; 466 int tmpid; 467 468 gd = mycpu; 469 cpuid = gd->gd_cpuid; 470 dd = &dummy_pcpu[cpuid]; 471 cpumask = CPUMASK(cpuid); 472 473 for (;;) { 474 lwkt_deschedule_self(gd->gd_curthread); /* interlock */ 475 atomic_set_cpumask(&dummy_rdyprocmask, cpumask); 476 spin_lock(&dummy_spin); 477 if (dd->uschedcp) { 478 /* 479 * We raced another cpu trying to schedule a thread onto us. 480 * If the runq isn't empty hit another free cpu. 481 */ 482 tmpmask = ~dummy_curprocmask & dummy_rdyprocmask & 483 gd->gd_other_cpus; 484 if (tmpmask && dummy_runqcount) { 485 tmpid = BSFCPUMASK(tmpmask); 486 KKASSERT(tmpid != cpuid); 487 atomic_clear_cpumask(&dummy_rdyprocmask, CPUMASK(tmpid)); 488 spin_unlock(&dummy_spin); 489 lwkt_schedule(&dummy_pcpu[tmpid].helper_thread); 490 } else { 491 spin_unlock(&dummy_spin); 492 } 493 } else if ((lp = TAILQ_FIRST(&dummy_runq)) != NULL) { 494 --dummy_runqcount; 495 TAILQ_REMOVE(&dummy_runq, lp, lwp_procq); 496 lp->lwp_flag &= ~LWP_ONRUNQ; 497 dd->uschedcp = lp; 498 atomic_set_cpumask(&dummy_curprocmask, cpumask); 499 spin_unlock(&dummy_spin); 500 #ifdef SMP 501 lwkt_acquire(lp->lwp_thread); 502 #endif 503 lwkt_schedule(lp->lwp_thread); 504 } else { 505 spin_unlock(&dummy_spin); 506 } 507 lwkt_switch(); 508 } 509 } 510 511 /* 512 * Setup our scheduler helpers. Note that curprocmask bit 0 has already 513 * been cleared by rqinit() and we should not mess with it further. 514 */ 515 static void 516 dummy_sched_thread_cpu_init(void) 517 { 518 int i; 519 520 if (bootverbose) 521 kprintf("start dummy scheduler helpers on cpus:"); 522 523 for (i = 0; i < ncpus; ++i) { 524 dummy_pcpu_t dd = &dummy_pcpu[i]; 525 cpumask_t mask = CPUMASK(i); 526 527 if ((mask & smp_active_mask) == 0) 528 continue; 529 530 if (bootverbose) 531 kprintf(" %d", i); 532 533 lwkt_create(dummy_sched_thread, NULL, NULL, &dd->helper_thread, 534 TDF_STOPREQ, i, "dsched %d", i); 535 536 /* 537 * Allow user scheduling on the target cpu. cpu #0 has already 538 * been enabled in rqinit(). 539 */ 540 if (i) 541 atomic_clear_cpumask(&dummy_curprocmask, mask); 542 atomic_set_cpumask(&dummy_rdyprocmask, mask); 543 } 544 if (bootverbose) 545 kprintf("\n"); 546 } 547 SYSINIT(uschedtd, SI_BOOT2_USCHED, SI_ORDER_SECOND, 548 dummy_sched_thread_cpu_init, NULL) 549 550 #endif 551 552