sys/kern/usched_dummy.c

   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
  54 #define MAXPRI                  128
  55 #define PRIBASE_REALTIME        0
  56 #define PRIBASE_NORMAL          MAXPRI
  57 #define PRIBASE_IDLE            (MAXPRI * 2)
  58 #define PRIBASE_THREAD          (MAXPRI * 3)
  59 #define PRIBASE_NULL            (MAXPRI * 4)
  60
  61 #define lwp_priority    lwp_usdata.bsd4.priority
  62 #define lwp_estcpu      lwp_usdata.bsd4.estcpu
  63
  64 static void dummy_acquire_curproc(struct lwp *lp);
  65 static void dummy_release_curproc(struct lwp *lp);
  66 static void dummy_select_curproc(globaldata_t gd);
  67 static void dummy_setrunqueue(struct lwp *lp);
  68 static void dummy_schedulerclock(struct lwp *lp, sysclock_t period,
  69                                 sysclock_t cpstamp);
  70 static void dummy_recalculate_estcpu(struct lwp *lp);
  71 static void dummy_resetpriority(struct lwp *lp);
  72 static void dummy_forking(struct lwp *plp, struct lwp *lp);
  73 static void dummy_exiting(struct lwp *plp, struct lwp *lp);
  74 static void dummy_yield(struct lwp *lp);
  75
  76 struct usched usched_dummy = {
  77         { NULL },
  78         "dummy", "Dummy DragonFly Scheduler",
  79         NULL,                   /* default registration */
  80         NULL,                   /* default deregistration */
  81         dummy_acquire_curproc,
  82         dummy_release_curproc,
  83         dummy_setrunqueue,
  84         dummy_schedulerclock,
  85         dummy_recalculate_estcpu,
  86         dummy_resetpriority,
  87         dummy_forking,
  88         dummy_exiting,
  89         NULL,                   /* setcpumask not supported */
  90         dummy_yield
  91 };
  92
  93 struct usched_dummy_pcpu {
  94         int     rrcount;
  95         struct thread helper_thread;
  96         struct lwp *uschedcp;
  97 };
  98
  99 typedef struct usched_dummy_pcpu *dummy_pcpu_t;
 100
 101 static struct usched_dummy_pcpu dummy_pcpu[MAXCPU];
 102 static cpumask_t dummy_curprocmask = -1;
 103 static cpumask_t dummy_rdyprocmask;
 104 static struct spinlock dummy_spin;
 105 static TAILQ_HEAD(rq, lwp) dummy_runq;
 106 static int dummy_runqcount;
 107
 108 static int usched_dummy_rrinterval = (ESTCPUFREQ + 9) / 10;
 109 SYSCTL_INT(_kern, OID_AUTO, usched_dummy_rrinterval, CTLFLAG_RW,
 110         &usched_dummy_rrinterval, 0, "");
 111
 112 /*
 113  * Initialize the run queues at boot time, clear cpu 0 in curprocmask
 114  * to allow dummy scheduling on cpu 0.
 115  */
 116 static void
 117 dummyinit(void *dummy)
 118 {
 119         TAILQ_INIT(&dummy_runq);
 120         spin_init(&dummy_spin);
 121         atomic_clear_int(&dummy_curprocmask, 1);
 122 }
 123 SYSINIT(runqueue, SI_BOOT2_USCHED, SI_ORDER_FIRST, dummyinit, NULL)
 124
 125 /*
 126  * DUMMY_ACQUIRE_CURPROC
 127  *
 128  * This function is called when the kernel intends to return to userland.
 129  * It is responsible for making the thread the current designated userland
 130  * thread for this cpu, blocking if necessary.
 131  *
 132  * We are expected to handle userland reschedule requests here too.
 133  *
 134  * WARNING! THIS FUNCTION IS ALLOWED TO CAUSE THE CURRENT THREAD TO MIGRATE
 135  * TO ANOTHER CPU!  Because most of the kernel assumes that no migration will
 136  * occur, this function is called only under very controlled circumstances.
 137  *
 138  * MPSAFE
 139  */
 140 static void
 141 dummy_acquire_curproc(struct lwp *lp)
 142 {
 143         globaldata_t gd = mycpu;
 144         dummy_pcpu_t dd = &dummy_pcpu[gd->gd_cpuid];
 145         thread_t td = lp->lwp_thread;
 146
 147         /*
 148          * Possibly select another thread
 149          */
 150         if (user_resched_wanted())
 151                 dummy_select_curproc(gd);
 152
 153         /*
 154          * If this cpu has no current thread, select ourself
 155          */
 156         if (dd->uschedcp == lp ||
 157             (dd->uschedcp == NULL && TAILQ_EMPTY(&dummy_runq))) {
 158                 atomic_set_int(&dummy_curprocmask, gd->gd_cpumask);
 159                 dd->uschedcp = lp;
 160                 return;
 161         }
 162
 163         /*
 164          * If this cpu's current user process thread is not our thread,
 165          * deschedule ourselves and place us on the run queue, then
 166          * switch away.
 167          *
 168          * We loop until we become the current process.  Its a good idea
 169          * to run any passive release(s) before we mess with the scheduler
 170          * so our thread is in the expected state.
 171          */
 172         KKASSERT(dd->uschedcp != lp);
 173         if (td->td_release)
 174                 td->td_release(lp->lwp_thread);
 175         do {
 176                 crit_enter();
 177                 lwkt_deschedule_self(td);
 178                 dummy_setrunqueue(lp);
 179                 if ((td->td_flags & TDF_RUNQ) == 0)
 180                         ++lp->lwp_ru.ru_nivcsw;
 181                 lwkt_switch();          /* WE MAY MIGRATE TO ANOTHER CPU */
 182                 crit_exit();
 183                 gd = mycpu;
 184                 dd = &dummy_pcpu[gd->gd_cpuid];
 185                 KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0);
 186         } while (dd->uschedcp != lp);
 187 }
 188
 189 /*
 190  * DUMMY_RELEASE_CURPROC
 191  *
 192  * This routine detaches the current thread from the userland scheduler,
 193  * usually because the thread needs to run in the kernel (at kernel priority)
 194  * for a while.
 195  *
 196  * This routine is also responsible for selecting a new thread to
 197  * make the current thread.
 198  *
 199  * WARNING!  The MP lock may be in an unsynchronized state due to the
 200  * way get_mplock() works and the fact that this function may be called
 201  * from a passive release during a lwkt_switch().   try_mplock() will deal
 202  * with this for us but you should be aware that td_mpcount may not be
 203  * useable.
 204  *
 205  * MPSAFE
 206  */
 207 static void
 208 dummy_release_curproc(struct lwp *lp)
 209 {
 210         globaldata_t gd = mycpu;
 211         dummy_pcpu_t dd = &dummy_pcpu[gd->gd_cpuid];
 212
 213         KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0);
 214         if (dd->uschedcp == lp) {
 215                 dummy_select_curproc(gd);
 216         }
 217 }
 218
 219 /*
 220  * DUMMY_SELECT_CURPROC
 221  *
 222  * Select a new current process for this cpu.  This satisfies a user
 223  * scheduler reschedule request so clear that too.
 224  *
 225  * This routine is also responsible for equal-priority round-robining,
 226  * typically triggered from dummy_schedulerclock().  In our dummy example
 227  * all the 'user' threads are LWKT scheduled all at once and we just
 228  * call lwkt_switch().
 229  *
 230  * MPSAFE
 231  */
 232 static
 233 void
 234 dummy_select_curproc(globaldata_t gd)
 235 {
 236         dummy_pcpu_t dd = &dummy_pcpu[gd->gd_cpuid];
 237         struct lwp *lp;
 238
 239         clear_user_resched();
 240         spin_lock_wr(&dummy_spin);
 241         if ((lp = TAILQ_FIRST(&dummy_runq)) == NULL) {
 242                 dd->uschedcp = NULL;
 243                 atomic_clear_int(&dummy_curprocmask, gd->gd_cpumask);
 244                 spin_unlock_wr(&dummy_spin);
 245         } else {
 246                 --dummy_runqcount;
 247                 TAILQ_REMOVE(&dummy_runq, lp, lwp_procq);
 248                 lp->lwp_flag &= ~LWP_ONRUNQ;
 249                 dd->uschedcp = lp;
 250                 atomic_set_int(&dummy_curprocmask, gd->gd_cpumask);
 251                 spin_unlock_wr(&dummy_spin);
 252 #ifdef SMP
 253                 lwkt_acquire(lp->lwp_thread);
 254 #endif
 255                 lwkt_schedule(lp->lwp_thread);
 256         }
 257 }
 258
 259 /*
 260  * DUMMY_SETRUNQUEUE
 261  *
 262  * This routine is called to schedule a new user process after a fork.
 263  * The scheduler module itself might also call this routine to place
 264  * the current process on the userland scheduler's run queue prior
 265  * to calling dummy_select_curproc().
 266  *
 267  * The caller may set P_PASSIVE_ACQ in p_flag to indicate that we should
 268  * attempt to leave the thread on the current cpu.
 269  *
 270  * MPSAFE
 271  */
 272 static void
 273 dummy_setrunqueue(struct lwp *lp)
 274 {
 275         globaldata_t gd = mycpu;
 276         dummy_pcpu_t dd = &dummy_pcpu[gd->gd_cpuid];
 277         cpumask_t mask;
 278         int cpuid;
 279
 280         if (dd->uschedcp == NULL) {
 281                 dd->uschedcp = lp;
 282                 atomic_set_int(&dummy_curprocmask, gd->gd_cpumask);
 283                 lwkt_schedule(lp->lwp_thread);
 284         } else {
 285                 /*
 286                  * Add to our global runq
 287                  */
 288                 KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0);
 289                 spin_lock_wr(&dummy_spin);
 290                 ++dummy_runqcount;
 291                 TAILQ_INSERT_TAIL(&dummy_runq, lp, lwp_procq);
 292                 lp->lwp_flag |= LWP_ONRUNQ;
 293 #ifdef SMP
 294                 lwkt_giveaway(lp->lwp_thread);
 295 #endif
 296
 297                 /* lp = TAILQ_FIRST(&dummy_runq); */
 298
 299                 /*
 300                  * Notify the next available cpu.  P.S. some
 301                  * cpu affinity could be done here.
 302                  *
 303                  * The rdyprocmask bit placeholds the knowledge that there
 304                  * is a process on the runq that needs service.  If the
 305                  * helper thread cannot find a home for it it will forward
 306                  * the request to another available cpu.
 307                  */
 308                 mask = ~dummy_curprocmask & dummy_rdyprocmask &
 309                        gd->gd_other_cpus;
 310                 if (mask) {
 311                         cpuid = bsfl(mask);
 312                         atomic_clear_int(&dummy_rdyprocmask, 1 << cpuid);
 313                         spin_unlock_wr(&dummy_spin);
 314                         lwkt_schedule(&dummy_pcpu[cpuid].helper_thread);
 315                 } else {
 316                         spin_unlock_wr(&dummy_spin);
 317                 }
 318         }
 319 }
 320
 321 /*
 322  * This routine is called from a systimer IPI.  Thus it is called with
 323  * a critical section held.  Any spinlocks we get here that are also
 324  * obtained in other procedures must be proected by a critical section
 325  * in those other procedures to avoid a deadlock.
 326  *
 327  * The MP lock may or may not be held on entry and cannot be obtained
 328  * by this routine (because it is called from a systimer IPI).  Additionally,
 329  * because this is equivalent to a FAST interrupt, spinlocks cannot be used
 330  * (or at least, you have to check that gd_spin* counts are 0 before you
 331  * can).
 332  *
 333  * This routine is called at ESTCPUFREQ on each cpu independantly.
 334  *
 335  * This routine typically queues a reschedule request, which will cause
 336  * the scheduler's BLAH_select_curproc() to be called as soon as possible.
 337  *
 338  * MPSAFE
 339  */
 340 static
 341 void
 342 dummy_schedulerclock(struct lwp *lp, sysclock_t period, sysclock_t cpstamp)
 343 {
 344         globaldata_t gd = mycpu;
 345         dummy_pcpu_t dd = &dummy_pcpu[gd->gd_cpuid];
 346
 347         if (++dd->rrcount >= usched_dummy_rrinterval) {
 348                 dd->rrcount = 0;
 349                 need_user_resched();
 350         }
 351 }
 352
 353 /*
 354  * DUMMY_RECALCULATE_ESTCPU
 355  *
 356  * Called once a second for any process that is running or has slept
 357  * for less then 2 seconds.
 358  *
 359  * MPSAFE
 360  */
 361 static
 362 void
 363 dummy_recalculate_estcpu(struct lwp *lp)
 364 {
 365 }
 366
 367 static
 368 void
 369 dummy_yield(struct lwp *lp)
 370 {
 371         need_user_resched();
 372 }
 373
 374 /*
 375  * DUMMY_RESETPRIORITY
 376  *
 377  * This routine is called after the kernel has potentially modified
 378  * the lwp_rtprio structure.  The target process may be running or sleeping
 379  * or scheduled but not yet running or owned by another cpu.  Basically,
 380  * it can be in virtually any state.
 381  *
 382  * This routine is called by fork1() for initial setup with the process
 383  * of the run queue, and also may be called normally with the process on or
 384  * off the run queue.
 385  *
 386  * MPSAFE
 387  */
 388 static void
 389 dummy_resetpriority(struct lwp *lp)
 390 {
 391         /* XXX spinlock usually needed */
 392         /*
 393          * Set p_priority for general process comparisons
 394          */
 395         switch(lp->lwp_rtprio.type) {
 396         case RTP_PRIO_REALTIME:
 397                 lp->lwp_priority = PRIBASE_REALTIME + lp->lwp_rtprio.prio;
 398                 return;
 399         case RTP_PRIO_NORMAL:
 400                 lp->lwp_priority = PRIBASE_NORMAL + lp->lwp_rtprio.prio;
 401                 break;
 402         case RTP_PRIO_IDLE:
 403                 lp->lwp_priority = PRIBASE_IDLE + lp->lwp_rtprio.prio;
 404                 return;
 405         case RTP_PRIO_THREAD:
 406                 lp->lwp_priority = PRIBASE_THREAD + lp->lwp_rtprio.prio;
 407                 return;
 408         }
 409         /* XXX spinlock usually needed */
 410 }
 411
 412
 413 /*
 414  * DUMMY_FORKING
 415  *
 416  * Called from fork1() when a new child process is being created.  Allows
 417  * the scheduler to predispose the child process before it gets scheduled.
 418  *
 419  * MPSAFE
 420  */
 421 static void
 422 dummy_forking(struct lwp *plp, struct lwp *lp)
 423 {
 424         lp->lwp_estcpu = plp->lwp_estcpu;
 425 #if 0
 426         ++plp->lwp_estcpu;
 427 #endif
 428 }
 429
 430 /*
 431  * DUMMY_EXITING
 432  *
 433  * Called when the parent reaps a child.   Typically used to propogate cpu
 434  * use by the child back to the parent as part of a batch detection
 435  * heuristic.
 436  *
 437  * NOTE: cpu use is not normally back-propogated to PID 1.
 438  *
 439  * MPSAFE
 440  */
 441 static void
 442 dummy_exiting(struct lwp *plp, struct lwp *lp)
 443 {
 444 }
 445
 446 /*
 447  * SMP systems may need a scheduler helper thread.  This is how one can be
 448  * setup.
 449  *
 450  * We use a neat LWKT scheduling trick to interlock the helper thread.  It
 451  * is possible to deschedule an LWKT thread and then do some work before
 452  * switching away.  The thread can be rescheduled at any time, even before
 453  * we switch away.
 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 = 1 << cpuid;
 472
 473     /*
 474      * Our Scheduler helper thread does not need to hold the MP lock
 475      */
 476     rel_mplock();
 477
 478     for (;;) {
 479         lwkt_deschedule_self(gd->gd_curthread);         /* interlock */
 480         atomic_set_int(&dummy_rdyprocmask, cpumask);
 481         spin_lock_wr(&dummy_spin);
 482         if (dd->uschedcp) {
 483                 /*
 484                  * We raced another cpu trying to schedule a thread onto us.
 485                  * If the runq isn't empty hit another free cpu.
 486                  */
 487                 tmpmask = ~dummy_curprocmask & dummy_rdyprocmask &
 488                           gd->gd_other_cpus;
 489                 if (tmpmask && dummy_runqcount) {
 490                         tmpid = bsfl(tmpmask);
 491                         KKASSERT(tmpid != cpuid);
 492                         atomic_clear_int(&dummy_rdyprocmask, 1 << tmpid);
 493                         spin_unlock_wr(&dummy_spin);
 494                         lwkt_schedule(&dummy_pcpu[tmpid].helper_thread);
 495                 } else {
 496                         spin_unlock_wr(&dummy_spin);
 497                 }
 498         } else if ((lp = TAILQ_FIRST(&dummy_runq)) != NULL) {
 499                 --dummy_runqcount;
 500                 TAILQ_REMOVE(&dummy_runq, lp, lwp_procq);
 501                 lp->lwp_flag &= ~LWP_ONRUNQ;
 502                 dd->uschedcp = lp;
 503                 atomic_set_int(&dummy_curprocmask, cpumask);
 504                 spin_unlock_wr(&dummy_spin);
 505 #ifdef SMP
 506                 lwkt_acquire(lp->lwp_thread);
 507 #endif
 508                 lwkt_schedule(lp->lwp_thread);
 509         } else {
 510                 spin_unlock_wr(&dummy_spin);
 511         }
 512         lwkt_switch();
 513     }
 514 }
 515
 516 /*
 517  * Setup our scheduler helpers.  Note that curprocmask bit 0 has already
 518  * been cleared by rqinit() and we should not mess with it further.
 519  */
 520 static void
 521 dummy_sched_thread_cpu_init(void)
 522 {
 523     int i;
 524
 525     if (bootverbose)
 526         kprintf("start dummy scheduler helpers on cpus:");
 527
 528     for (i = 0; i < ncpus; ++i) {
 529         dummy_pcpu_t dd = &dummy_pcpu[i];
 530         cpumask_t mask = 1 << i;
 531
 532         if ((mask & smp_active_mask) == 0)
 533             continue;
 534
 535         if (bootverbose)
 536             kprintf(" %d", i);
 537
 538         lwkt_create(dummy_sched_thread, NULL, NULL, &dd->helper_thread,
 539                     TDF_STOPREQ, i, "dsched %d", i);
 540
 541         /*
 542          * Allow user scheduling on the target cpu.  cpu #0 has already
 543          * been enabled in rqinit().
 544          */
 545         if (i)
 546             atomic_clear_int(&dummy_curprocmask, mask);
 547         atomic_set_int(&dummy_rdyprocmask, mask);
 548     }
 549     if (bootverbose)
 550         kprintf("\n");
 551 }
 552 SYSINIT(uschedtd, SI_BOOT2_USCHED, SI_ORDER_SECOND,
 553         dummy_sched_thread_cpu_init, NULL)
 554
 555 #endif
 556