sys/src/9/port/proc.c

   1 #include        <u.h>
   2 #include        "../port/lib.h"
   3 #include        "mem.h"
   4 #include        "dat.h"
   5 #include        "fns.h"
   6 #include        "../port/error.h"
   7 #include        "edf.h"
   8 #include        <trace.h>
   9
  10 int     schedgain = 30; /* units in seconds */
  11 int     nrdy;
  12
  13 void updatecpu(Proc*);
  14 int reprioritize(Proc*);
  15
  16 ulong   delayedscheds;  /* statistics */
  17 ulong   skipscheds;
  18 ulong   preempts;
  19 ulong   load;
  20
  21 static struct Procalloc
  22 {
  23         Lock;
  24         Proc*   ht[128];
  25         Proc*   arena;
  26         Proc*   free;
  27 } procalloc;
  28
  29 enum
  30 {
  31         Q=10,
  32         DQ=4,
  33         Scaling=2,
  34 };
  35
  36 Schedq  runq[Nrq];
  37 ulong   runvec;
  38
  39 char *statename[] =
  40 {       /* BUG: generate automatically */
  41         "Dead",
  42         "Moribund",
  43         "Ready",
  44         "Scheding",
  45         "Running",
  46         "Queueing",
  47         "QueueingR",
  48         "QueueingW",
  49         "Wakeme",
  50         "Broken",
  51         "Stopped",
  52         "Rendez",
  53         "Waitrelease",
  54 };
  55
  56 static void pidfree(Proc*);
  57 static void rebalance(void);
  58
  59 /*
  60  * Always splhi()'ed.
  61  */
  62 void
  63 schedinit(void)         /* never returns */
  64 {
  65         Edf *e;
  66
  67         setlabel(&m->sched);
  68         if(up != nil) {
  69                 if((e = up->edf) != nil && (e->flags & Admitted))
  70                         edfrecord(up);
  71                 m->proc = nil;
  72                 switch(up->state) {
  73                 case Running:
  74                         ready(up);
  75                         break;
  76                 case Moribund:
  77                         up->state = Dead;
  78                         edfstop(up);
  79                         if(up->edf != nil)
  80                                 free(up->edf);
  81                         up->edf = nil;
  82
  83                         /*
  84                          * Holding locks from pexit:
  85                          *      procalloc
  86                          *      palloc
  87                          */
  88                         mmurelease(up);
  89                         unlock(&palloc);
  90
  91                         up->mach = nil;
  92                         updatecpu(up);
  93
  94                         up->qnext = procalloc.free;
  95                         procalloc.free = up;
  96
  97                         /* proc is free now, make sure unlock() wont touch it */
  98                         up = procalloc.Lock.p = nil;
  99                         unlock(&procalloc);
 100                         sched();
 101                 }
 102                 up->mach = nil;
 103                 updatecpu(up);
 104                 up = nil;
 105         }
 106         sched();
 107 }
 108
 109 /*
 110  *  If changing this routine, look also at sleep().  It
 111  *  contains a copy of the guts of sched().
 112  */
 113 void
 114 sched(void)
 115 {
 116         Proc *p;
 117
 118         if(m->ilockdepth)
 119                 panic("cpu%d: ilockdepth %d, last lock %#p at %#p, sched called from %#p",
 120                         m->machno,
 121                         m->ilockdepth,
 122                         up != nil ? up->lastilock: nil,
 123                         (up != nil && up->lastilock != nil) ? up->lastilock->pc: 0,
 124                         getcallerpc(&p+2));
 125         if(up != nil) {
 126                 /*
 127                  * Delay the sched until the process gives up the locks
 128                  * it is holding.  This avoids dumb lock loops.
 129                  * Don't delay if the process is Moribund.
 130                  * It called sched to die.
 131                  * But do sched eventually.  This avoids a missing unlock
 132                  * from hanging the entire kernel.
 133                  * But don't reschedule procs holding palloc or procalloc.
 134                  * Those are far too important to be holding while asleep.
 135                  *
 136                  * This test is not exact.  There can still be a few instructions
 137                  * in the middle of taslock when a process holds a lock
 138                  * but Lock.p has not yet been initialized.
 139                  */
 140                 if(up->nlocks)
 141                 if(up->state != Moribund)
 142                 if(up->delaysched < 20
 143                 || palloc.Lock.p == up
 144                 || fscache.Lock.p == up
 145                 || procalloc.Lock.p == up){
 146                         up->delaysched++;
 147                         delayedscheds++;
 148                         return;
 149                 }
 150                 up->delaysched = 0;
 151
 152                 splhi();
 153
 154                 /* statistics */
 155                 m->cs++;
 156
 157                 procsave(up);
 158                 if(setlabel(&up->sched)){
 159                         procrestore(up);
 160                         spllo();
 161                         return;
 162                 }
 163                 gotolabel(&m->sched);
 164         }
 165         p = runproc();
 166         if(p->edf == nil){
 167                 updatecpu(p);
 168                 p->priority = reprioritize(p);
 169         }
 170         if(p != m->readied)
 171                 m->schedticks = m->ticks + HZ/10;
 172         m->readied = nil;
 173         up = p;
 174         up->state = Running;
 175         up->mach = MACHP(m->machno);
 176         m->proc = up;
 177         mmuswitch(up);
 178         gotolabel(&up->sched);
 179 }
 180
 181 int
 182 anyready(void)
 183 {
 184         return runvec;
 185 }
 186
 187 int
 188 anyhigher(void)
 189 {
 190         return runvec & ~((1<<(up->priority+1))-1);
 191 }
 192
 193 /*
 194  *  here once per clock tick to see if we should resched
 195  */
 196 void
 197 hzsched(void)
 198 {
 199         /* once a second, rebalance will reprioritize ready procs */
 200         if(m->machno == 0)
 201                 rebalance();
 202
 203         /* unless preempted, get to run for at least 100ms */
 204         if(anyhigher()
 205         || (!up->fixedpri && (long)(m->ticks - m->schedticks) > 0 && anyready())){
 206                 m->readied = nil;       /* avoid cooperative scheduling */
 207                 up->delaysched++;
 208         }
 209 }
 210
 211 /*
 212  *  here at the end of non-clock interrupts to see if we should preempt the
 213  *  current process.  Returns 1 if preempted, 0 otherwise.
 214  */
 215 int
 216 preempted(void)
 217 {
 218         if(up != nil && up->state == Running)
 219         if(up->preempted == 0)
 220         if(anyhigher())
 221         if(!active.exiting){
 222                 m->readied = nil;       /* avoid cooperative scheduling */
 223                 up->preempted = 1;
 224                 sched();
 225                 splhi();
 226                 up->preempted = 0;
 227                 return 1;
 228         }
 229         return 0;
 230 }
 231
 232 /*
 233  * Update the cpu time average for this particular process,
 234  * which is about to change from up -> not up or vice versa.
 235  * p->lastupdate is the last time an updatecpu happened.
 236  *
 237  * The cpu time average is a decaying average that lasts
 238  * about D clock ticks.  D is chosen to be approximately
 239  * the cpu time of a cpu-intensive "quick job".  A job has to run
 240  * for approximately D clock ticks before we home in on its
 241  * actual cpu usage.  Thus if you manage to get in and get out
 242  * quickly, you won't be penalized during your burst.  Once you
 243  * start using your share of the cpu for more than about D
 244  * clock ticks though, your p->cpu hits 1000 (1.0) and you end up
 245  * below all the other quick jobs.  Interactive tasks, because
 246  * they basically always use less than their fair share of cpu,
 247  * will be rewarded.
 248  *
 249  * If the process has not been running, then we want to
 250  * apply the filter
 251  *
 252  *      cpu = cpu * (D-1)/D
 253  *
 254  * n times, yielding
 255  *
 256  *      cpu = cpu * ((D-1)/D)^n
 257  *
 258  * but D is big enough that this is approximately
 259  *
 260  *      cpu = cpu * (D-n)/D
 261  *
 262  * so we use that instead.
 263  *
 264  * If the process has been running, we apply the filter to
 265  * 1 - cpu, yielding a similar equation.  Note that cpu is
 266  * stored in fixed point (* 1000).
 267  *
 268  * Updatecpu must be called before changing up, in order
 269  * to maintain accurate cpu usage statistics.  It can be called
 270  * at any time to bring the stats for a given proc up-to-date.
 271  */
 272 void
 273 updatecpu(Proc *p)
 274 {
 275         ulong t, ocpu, n, D;
 276
 277         if(p->edf != nil)
 278                 return;
 279
 280         t = MACHP(0)->ticks*Scaling + Scaling/2;
 281         n = t - p->lastupdate;
 282         if(n == 0)
 283                 return;
 284         p->lastupdate = t;
 285
 286         D = schedgain*HZ*Scaling;
 287         if(n > D)
 288                 n = D;
 289
 290         ocpu = p->cpu;
 291         if(p != up)
 292                 p->cpu = (ocpu*(D-n))/D;
 293         else{
 294                 t = 1000 - ocpu;
 295                 t = (t*(D-n))/D;
 296                 p->cpu = 1000 - t;
 297         }
 298 //iprint("pid %lud %s for %lud cpu %lud -> %lud\n", p->pid,p==up?"active":"inactive",n, ocpu,p->cpu);
 299 }
 300
 301 /*
 302  * On average, p has used p->cpu of a cpu recently.
 303  * Its fair share is conf.nmach/m->load of a cpu.  If it has been getting
 304  * too much, penalize it.  If it has been getting not enough, reward it.
 305  * I don't think you can get much more than your fair share that
 306  * often, so most of the queues are for using less.  Having a priority
 307  * of 3 means you're just right.  Having a higher priority (up to p->basepri)
 308  * means you're not using as much as you could.
 309  */
 310 int
 311 reprioritize(Proc *p)
 312 {
 313         int fairshare, n, load, ratio;
 314
 315         load = MACHP(0)->load;
 316         if(load == 0)
 317                 return p->basepri;
 318
 319         /*
 320          * fairshare = 1.000 * conf.nmach * 1.000/load,
 321          * except the decimal point is moved three places
 322          * on both load and fairshare.
 323          */
 324         fairshare = (conf.nmach*1000*1000)/load;
 325         n = p->cpu;
 326         if(n == 0)
 327                 n = 1;
 328         ratio = (fairshare+n/2) / n;
 329         if(ratio > p->basepri)
 330                 ratio = p->basepri;
 331         if(ratio < 0)
 332                 panic("reprioritize");
 333 //iprint("pid %lud cpu %lud load %d fair %d pri %d\n", p->pid, p->cpu, load, fairshare, ratio);
 334         return ratio;
 335 }
 336
 337 /*
 338  * add a process to a scheduling queue
 339  */
 340 void
 341 queueproc(Schedq *rq, Proc *p)
 342 {
 343         int pri;
 344
 345         pri = rq - runq;
 346         lock(runq);
 347         p->priority = pri;
 348         p->rnext = nil;
 349         if(rq->tail != nil)
 350                 rq->tail->rnext = p;
 351         else
 352                 rq->head = p;
 353         rq->tail = p;
 354         rq->n++;
 355         nrdy++;
 356         runvec |= 1<<pri;
 357         unlock(runq);
 358 }
 359
 360 /*
 361  *  try to remove a process from a scheduling queue (called splhi)
 362  */
 363 Proc*
 364 dequeueproc(Schedq *rq, Proc *tp)
 365 {
 366         Proc *l, *p;
 367
 368         if(!canlock(runq))
 369                 return nil;
 370
 371         /*
 372          *  the queue may have changed before we locked runq,
 373          *  refind the target process.
 374          */
 375         l = nil;
 376         for(p = rq->head; p != nil; p = p->rnext){
 377                 if(p == tp)
 378                         break;
 379                 l = p;
 380         }
 381
 382         /*
 383          *  p->mach==0 only when process state is saved
 384          */
 385         if(p == nil || p->mach != nil){
 386                 unlock(runq);
 387                 return nil;
 388         }
 389         if(p->rnext == nil)
 390                 rq->tail = l;
 391         if(l != nil)
 392                 l->rnext = p->rnext;
 393         else
 394                 rq->head = p->rnext;
 395         if(rq->head == nil)
 396                 runvec &= ~(1<<(rq-runq));
 397         rq->n--;
 398         nrdy--;
 399         if(p->state != Ready)
 400                 print("dequeueproc %s %lud %s\n", p->text, p->pid, statename[p->state]);
 401
 402         unlock(runq);
 403         return p;
 404 }
 405
 406 /*
 407  *  ready(p) picks a new priority for a process and sticks it in the
 408  *  runq for that priority.
 409  */
 410 void
 411 ready(Proc *p)
 412 {
 413         int s, pri;
 414         Schedq *rq;
 415         void (*pt)(Proc*, int, vlong);
 416
 417         if(p->state == Ready){
 418                 print("double ready %s %lud pc %p\n", p->text, p->pid, getcallerpc(&p));
 419                 return;
 420         }
 421
 422         s = splhi();
 423         if(edfready(p)){
 424                 splx(s);
 425                 return;
 426         }
 427
 428         if(up != p && (p->wired == nil || p->wired == MACHP(m->machno)))
 429                 m->readied = p; /* group scheduling */
 430
 431         updatecpu(p);
 432         pri = reprioritize(p);
 433         p->priority = pri;
 434         rq = &runq[pri];
 435         p->state = Ready;
 436         queueproc(rq, p);
 437         pt = proctrace;
 438         if(pt != nil)
 439                 pt(p, SReady, 0);
 440         splx(s);
 441 }
 442
 443 /*
 444  *  yield the processor and drop our priority
 445  */
 446 void
 447 yield(void)
 448 {
 449         if(anyready()){
 450                 /* pretend we just used 1/2 tick */
 451                 up->lastupdate -= Scaling/2;
 452                 sched();
 453         }
 454 }
 455
 456 /*
 457  *  recalculate priorities once a second.  We need to do this
 458  *  since priorities will otherwise only be recalculated when
 459  *  the running process blocks.
 460  */
 461 ulong balancetime;
 462
 463 static void
 464 rebalance(void)
 465 {
 466         int pri, npri, x;
 467         Schedq *rq;
 468         Proc *p;
 469         ulong t;
 470
 471         t = m->ticks;
 472         if(t - balancetime < HZ)
 473                 return;
 474         balancetime = t;
 475
 476         for(pri=0, rq=runq; pri<Npriq; pri++, rq++){
 477 another:
 478                 p = rq->head;
 479                 if(p == nil)
 480                         continue;
 481                 if(pri == p->basepri)
 482                         continue;
 483                 updatecpu(p);
 484                 npri = reprioritize(p);
 485                 if(npri != pri){
 486                         x = splhi();
 487                         p = dequeueproc(rq, p);
 488                         if(p != nil)
 489                                 queueproc(&runq[npri], p);
 490                         splx(x);
 491                         goto another;
 492                 }
 493         }
 494 }
 495
 496
 497 /*
 498  *  pick a process to run
 499  */
 500 Proc*
 501 runproc(void)
 502 {
 503         Schedq *rq;
 504         Proc *p;
 505         ulong start, now;
 506         int i;
 507         void (*pt)(Proc*, int, vlong);
 508
 509         start = perfticks();
 510
 511         /* cooperative scheduling until the clock ticks */
 512         if((p = m->readied) != nil && p->mach == nil && p->state == Ready
 513         && (p->wired == nil || p->wired == MACHP(m->machno))
 514         && runq[Nrq-1].head == nil && runq[Nrq-2].head == nil){
 515                 skipscheds++;
 516                 rq = &runq[p->priority];
 517                 goto found;
 518         }
 519
 520         preempts++;
 521
 522 loop:
 523         /*
 524          *  find a process that last ran on this processor (affinity),
 525          *  or one that hasn't moved in a while (load balancing).  Every
 526          *  time around the loop affinity goes down.
 527          */
 528         spllo();
 529         for(i = 0;; i++){
 530                 /*
 531                  *  find the highest priority target process that this
 532                  *  processor can run given affinity constraints.
 533                  *
 534                  */
 535                 for(rq = &runq[Nrq-1]; rq >= runq; rq--){
 536                         for(p = rq->head; p != nil; p = p->rnext){
 537                                 if(p->mp == nil || p->mp == MACHP(m->machno)
 538                                 || (p->wired == nil && i > 0))
 539                                         goto found;
 540                         }
 541                 }
 542
 543                 /* waste time or halt the CPU */
 544                 idlehands();
 545
 546                 /* remember how much time we're here */
 547                 now = perfticks();
 548                 m->perf.inidle += now-start;
 549                 start = now;
 550         }
 551
 552 found:
 553         splhi();
 554         p = dequeueproc(rq, p);
 555         if(p == nil)
 556                 goto loop;
 557
 558         p->state = Scheding;
 559         p->mp = MACHP(m->machno);
 560
 561         if(edflock(p)){
 562                 edfrun(p, rq == &runq[PriEdf]); /* start deadline timer and do admin */
 563                 edfunlock();
 564         }
 565         pt = proctrace;
 566         if(pt != nil)
 567                 pt(p, SRun, 0);
 568         return p;
 569 }
 570
 571 int
 572 canpage(Proc *p)
 573 {
 574         int ok = 0;
 575
 576         splhi();
 577         lock(runq);
 578         /* Only reliable way to see if we are Running */
 579         if(p->mach == nil) {
 580                 p->newtlb = 1;
 581                 ok = 1;
 582         }
 583         unlock(runq);
 584         spllo();
 585
 586         return ok;
 587 }
 588
 589 Proc*
 590 newproc(void)
 591 {
 592         char msg[64];
 593         Proc *p;
 594
 595         lock(&procalloc);
 596         for(;;) {
 597                 if((p = procalloc.free) != nil)
 598                         break;
 599
 600                 snprint(msg, sizeof msg, "no procs; %s forking",
 601                         up != nil ? up->text: "kernel");
 602                 unlock(&procalloc);
 603                 resrcwait(msg);
 604                 lock(&procalloc);
 605         }
 606         procalloc.free = p->qnext;
 607         unlock(&procalloc);
 608
 609         p->state = Scheding;
 610         p->psstate = "New";
 611         p->mach = nil;
 612         p->eql = nil;
 613         p->qnext = nil;
 614         p->nchild = 0;
 615         p->nwait = 0;
 616         p->waitq = nil;
 617         p->parent = nil;
 618         p->pgrp = nil;
 619         p->egrp = nil;
 620         p->fgrp = nil;
 621         p->rgrp = nil;
 622         p->pdbg = nil;
 623         p->fpstate = FPinit;
 624         p->kp = 0;
 625         p->procctl = 0;
 626         p->syscalltrace = nil;
 627         p->notepending = 0;
 628         p->ureg = nil;
 629         p->privatemem = 0;
 630         p->noswap = 0;
 631         p->errstr = p->errbuf0;
 632         p->syserrstr = p->errbuf1;
 633         p->errbuf0[0] = '\0';
 634         p->errbuf1[0] = '\0';
 635         p->nlocks = 0;
 636         p->delaysched = 0;
 637         p->trace = 0;
 638         kstrdup(&p->user, "*nouser");
 639         kstrdup(&p->text, "*notext");
 640         kstrdup(&p->args, "");
 641         p->nargs = 0;
 642         p->setargs = 0;
 643         memset(p->seg, 0, sizeof p->seg);
 644         p->parentpid = 0;
 645         p->noteid = pidalloc(p);
 646         if(p->kstack == nil)
 647                 p->kstack = smalloc(KSTACK);
 648
 649         /* sched params */
 650         p->mp = nil;
 651         p->wired = nil;
 652         procpriority(p, PriNormal, 0);
 653         p->cpu = 0;
 654         p->lastupdate = MACHP(0)->ticks*Scaling;
 655         p->edf = nil;
 656
 657         return p;
 658 }
 659
 660 /*
 661  * wire this proc to a machine
 662  */
 663 void
 664 procwired(Proc *p, int bm)
 665 {
 666         Proc *pp;
 667         int i;
 668         char nwired[MAXMACH];
 669         Mach *wm;
 670
 671         if(bm < 0){
 672                 /* pick a machine to wire to */
 673                 memset(nwired, 0, sizeof(nwired));
 674                 p->wired = nil;
 675                 pp = proctab(0);
 676                 for(i=0; i<conf.nproc; i++, pp++){
 677                         wm = pp->wired;
 678                         if(wm != nil && pp->pid)
 679                                 nwired[wm->machno]++;
 680                 }
 681                 bm = 0;
 682                 for(i=0; i<conf.nmach; i++)
 683                         if(nwired[i] < nwired[bm])
 684                                 bm = i;
 685         } else {
 686                 /* use the virtual machine requested */
 687                 bm = bm % conf.nmach;
 688         }
 689
 690         p->wired = MACHP(bm);
 691         p->mp = p->wired;
 692 }
 693
 694 void
 695 procpriority(Proc *p, int pri, int fixed)
 696 {
 697         if(pri >= Npriq)
 698                 pri = Npriq - 1;
 699         else if(pri < 0)
 700                 pri = 0;
 701         p->basepri = pri;
 702         p->priority = pri;
 703         if(fixed){
 704                 p->fixedpri = 1;
 705         } else {
 706                 p->fixedpri = 0;
 707         }
 708 }
 709
 710 void
 711 procinit0(void)         /* bad planning - clashes with devproc.c */
 712 {
 713         Proc *p;
 714         int i;
 715
 716         p = xalloc(conf.nproc*sizeof(Proc));
 717         if(p == nil){
 718                 xsummary();
 719                 panic("cannot allocate %lud procs (%ludMB)", conf.nproc, conf.nproc*sizeof(Proc)/(1024*1024));
 720         }
 721         procalloc.arena = p;
 722         procalloc.free = p;
 723         for(i=0; i<conf.nproc-1; i++, p++)
 724                 p->qnext = p+1;
 725         p->qnext = nil;
 726 }
 727
 728 /*
 729  *  sleep if a condition is not true.  Another process will
 730  *  awaken us after it sets the condition.  When we awaken
 731  *  the condition may no longer be true.
 732  *
 733  *  we lock both the process and the rendezvous to keep r->p
 734  *  and p->r synchronized.
 735  */
 736 void
 737 sleep(Rendez *r, int (*f)(void*), void *arg)
 738 {
 739         int s;
 740         void (*pt)(Proc*, int, vlong);
 741
 742         s = splhi();
 743
 744         if(up->nlocks)
 745                 print("process %lud sleeps with %d locks held, last lock %#p locked at pc %#p, sleep called from %#p\n",
 746                         up->pid, up->nlocks, up->lastlock, up->lastlock->pc, getcallerpc(&r));
 747         lock(r);
 748         lock(&up->rlock);
 749         if(r->p != nil){
 750                 print("double sleep called from %#p, %lud %lud\n", getcallerpc(&r), r->p->pid, up->pid);
 751                 dumpstack();
 752         }
 753
 754         /*
 755          *  Wakeup only knows there may be something to do by testing
 756          *  r->p in order to get something to lock on.
 757          *  Flush that information out to memory in case the sleep is
 758          *  committed.
 759          */
 760         r->p = up;
 761
 762         if((*f)(arg) || up->notepending){
 763                 /*
 764                  *  if condition happened or a note is pending
 765                  *  never mind
 766                  */
 767                 r->p = nil;
 768                 unlock(&up->rlock);
 769                 unlock(r);
 770         } else {
 771                 /*
 772                  *  now we are committed to
 773                  *  change state and call scheduler
 774                  */
 775                 pt = proctrace;
 776                 if(pt != nil)
 777                         pt(up, SSleep, 0);
 778                 up->state = Wakeme;
 779                 up->r = r;
 780
 781                 /* statistics */
 782                 m->cs++;
 783
 784                 procsave(up);
 785                 if(setlabel(&up->sched)) {
 786                         /*
 787                          *  here when the process is awakened
 788                          */
 789                         procrestore(up);
 790                         spllo();
 791                 } else {
 792                         /*
 793                          *  here to go to sleep (i.e. stop Running)
 794                          */
 795                         unlock(&up->rlock);
 796                         unlock(r);
 797                         gotolabel(&m->sched);
 798                 }
 799         }
 800
 801         if(up->notepending) {
 802                 up->notepending = 0;
 803                 splx(s);
 804                 interrupted();
 805         }
 806
 807         splx(s);
 808 }
 809
 810 void
 811 interrupted(void)
 812 {
 813         if(up->procctl == Proc_exitme && up->closingfgrp != nil)
 814                 forceclosefgrp();
 815         error(Eintr);
 816 }
 817
 818 static int
 819 tfn(void *arg)
 820 {
 821         return up->trend == nil || up->tfn(arg);
 822 }
 823
 824 void
 825 twakeup(Ureg*, Timer *t)
 826 {
 827         Proc *p;
 828         Rendez *trend;
 829
 830         p = t->ta;
 831         trend = p->trend;
 832         if(trend != nil){
 833                 p->trend = nil;
 834                 wakeup(trend);
 835         }
 836 }
 837
 838 void
 839 tsleep(Rendez *r, int (*fn)(void*), void *arg, ulong ms)
 840 {
 841         if(up->tt != nil){
 842                 print("%s %lud: tsleep timer active: mode %d, tf %#p, pc %#p\n",
 843                         up->text, up->pid, up->tmode, up->tf, getcallerpc(&r));
 844                 timerdel(up);
 845         }
 846         up->tns = MS2NS(ms);
 847         up->tf = twakeup;
 848         up->tmode = Trelative;
 849         up->ta = up;
 850         up->trend = r;
 851         up->tfn = fn;
 852         timeradd(up);
 853
 854         if(waserror()){
 855                 up->trend = nil;
 856                 timerdel(up);
 857                 nexterror();
 858         }
 859         sleep(r, tfn, arg);
 860         up->trend = nil;
 861         timerdel(up);
 862         poperror();
 863 }
 864
 865 /*
 866  *  Expects that only one process can call wakeup for any given Rendez.
 867  *  We hold both locks to ensure that r->p and p->r remain consistent.
 868  *  Richard Miller has a better solution that doesn't require both to
 869  *  be held simultaneously, but I'm a paranoid - presotto.
 870  */
 871 Proc*
 872 wakeup(Rendez *r)
 873 {
 874         Proc *p;
 875         int s;
 876
 877         s = splhi();
 878
 879         lock(r);
 880         p = r->p;
 881
 882         if(p != nil){
 883                 lock(&p->rlock);
 884                 if(p->state != Wakeme || p->r != r){
 885                         iprint("%p %p %d\n", p->r, r, p->state);
 886                         panic("wakeup: state");
 887                 }
 888                 r->p = nil;
 889                 p->r = nil;
 890                 ready(p);
 891                 unlock(&p->rlock);
 892         }
 893         unlock(r);
 894
 895         splx(s);
 896
 897         return p;
 898 }
 899
 900 /*
 901  *  if waking a sleeping process, this routine must hold both
 902  *  p->rlock and r->lock.  However, it can't know them in
 903  *  the same order as wakeup causing a possible lock ordering
 904  *  deadlock.  We break the deadlock by giving up the p->rlock
 905  *  lock if we can't get the r->lock and retrying.
 906  */
 907 int
 908 postnote(Proc *p, int dolock, char *n, int flag)
 909 {
 910         int s, ret;
 911         QLock *q;
 912
 913         if(p == nil)
 914                 return 0;
 915
 916         if(dolock)
 917                 qlock(&p->debug);
 918
 919         if(p->pid == 0){
 920                 if(dolock)
 921                         qunlock(&p->debug);
 922                 return 0;
 923         }
 924
 925         if(n != nil && flag != NUser && (p->notify == 0 || p->notified))
 926                 p->nnote = 0;
 927
 928         ret = 0;
 929         if(p->nnote < NNOTE && n != nil) {
 930                 kstrcpy(p->note[p->nnote].msg, n, ERRMAX);
 931                 p->note[p->nnote++].flag = flag;
 932                 ret = 1;
 933         }
 934         p->notepending = 1;
 935         if(dolock)
 936                 qunlock(&p->debug);
 937
 938         /* this loop is to avoid lock ordering problems. */
 939         for(;;){
 940                 Rendez *r;
 941
 942                 s = splhi();
 943                 lock(&p->rlock);
 944                 r = p->r;
 945
 946                 /* waiting for a wakeup? */
 947                 if(r == nil)
 948                         break;  /* no */
 949
 950                 /* try for the second lock */
 951                 if(canlock(r)){
 952                         if(p->state != Wakeme || r->p != p)
 953                                 panic("postnote: state %d %d %d", r->p != p, p->r != r, p->state);
 954                         p->r = nil;
 955                         r->p = nil;
 956                         ready(p);
 957                         unlock(r);
 958                         break;
 959                 }
 960
 961                 /* give other process time to get out of critical section and try again */
 962                 unlock(&p->rlock);
 963                 splx(s);
 964                 sched();
 965         }
 966         unlock(&p->rlock);
 967         splx(s);
 968
 969         switch(p->state){
 970         case Queueing:
 971                 /* Try and pull out of a eqlock */
 972                 if((q = p->eql) != nil){
 973                         lock(&q->use);
 974                         if(p->state == Queueing && p->eql == q){
 975                                 Proc *d, *l;
 976
 977                                 for(l = nil, d = q->head; d != nil; l = d, d = d->qnext){
 978                                         if(d == p){
 979                                                 if(l != nil)
 980                                                         l->qnext = p->qnext;
 981                                                 else
 982                                                         q->head = p->qnext;
 983                                                 if(p->qnext == nil)
 984                                                         q->tail = l;
 985                                                 p->qnext = nil;
 986                                                 p->eql = nil;   /* not taken */
 987                                                 ready(p);
 988                                                 break;
 989                                         }
 990                                 }
 991                         }
 992                         unlock(&q->use);
 993                 }
 994                 break;
 995         case Rendezvous:
 996                 /* Try and pull out of a rendezvous */
 997                 lock(p->rgrp);
 998                 if(p->state == Rendezvous) {
 999                         Proc *d, **l;
1000
1001                         l = &REND(p->rgrp, p->rendtag);
1002                         for(d = *l; d != nil; d = d->rendhash) {
1003                                 if(d == p) {
1004                                         *l = p->rendhash;
1005                                         p->rendval = ~0;
1006                                         ready(p);
1007                                         break;
1008                                 }
1009                                 l = &d->rendhash;
1010                         }
1011                 }
1012                 unlock(p->rgrp);
1013                 break;
1014         }
1015         return ret;
1016 }
1017
1018 /*
1019  * weird thing: keep at most NBROKEN around
1020  */
1021 #define NBROKEN 4
1022 struct
1023 {
1024         QLock;
1025         int     n;
1026         Proc    *p[NBROKEN];
1027 }broken;
1028
1029 void
1030 addbroken(Proc *p)
1031 {
1032         qlock(&broken);
1033         if(broken.n == NBROKEN) {
1034                 ready(broken.p[0]);
1035                 memmove(&broken.p[0], &broken.p[1], sizeof(Proc*)*(NBROKEN-1));
1036                 --broken.n;
1037         }
1038         broken.p[broken.n++] = p;
1039         qunlock(&broken);
1040
1041         edfstop(up);
1042         p->state = Broken;
1043         p->psstate = nil;
1044         sched();
1045 }
1046
1047 void
1048 unbreak(Proc *p)
1049 {
1050         int b;
1051
1052         qlock(&broken);
1053         for(b=0; b < broken.n; b++)
1054                 if(broken.p[b] == p) {
1055                         broken.n--;
1056                         memmove(&broken.p[b], &broken.p[b+1],
1057                                         sizeof(Proc*)*(NBROKEN-(b+1)));
1058                         ready(p);
1059                         break;
1060                 }
1061         qunlock(&broken);
1062 }
1063
1064 int
1065 freebroken(void)
1066 {
1067         int i, n;
1068
1069         qlock(&broken);
1070         n = broken.n;
1071         for(i=0; i<n; i++) {
1072                 ready(broken.p[i]);
1073                 broken.p[i] = nil;
1074         }
1075         broken.n = 0;
1076         qunlock(&broken);
1077         return n;
1078 }
1079
1080 void
1081 pexit(char *exitstr, int freemem)
1082 {
1083         Proc *p;
1084         Segment **s, **es;
1085         ulong utime, stime;
1086         Waitq *wq;
1087         Fgrp *fgrp;
1088         Egrp *egrp;
1089         Rgrp *rgrp;
1090         Pgrp *pgrp;
1091         Chan *dot;
1092         void (*pt)(Proc*, int, vlong);
1093
1094         up->alarm = 0;
1095         timerdel(up);
1096         pt = proctrace;
1097         if(pt != nil)
1098                 pt(up, SDead, 0);
1099
1100         /* nil out all the resources under lock (free later) */
1101         qlock(&up->debug);
1102         fgrp = up->fgrp;
1103         up->fgrp = nil;
1104         egrp = up->egrp;
1105         up->egrp = nil;
1106         rgrp = up->rgrp;
1107         up->rgrp = nil;
1108         pgrp = up->pgrp;
1109         up->pgrp = nil;
1110         dot = up->dot;
1111         up->dot = nil;
1112         qunlock(&up->debug);
1113
1114         if(fgrp != nil)
1115                 closefgrp(fgrp);
1116         if(egrp != nil)
1117                 closeegrp(egrp);
1118         if(rgrp != nil)
1119                 closergrp(rgrp);
1120         if(dot != nil)
1121                 cclose(dot);
1122         if(pgrp != nil)
1123                 closepgrp(pgrp);
1124
1125         /*
1126          * if not a kernel process and have a parent,
1127          * do some housekeeping.
1128          */
1129         if(up->kp == 0 && up->parentpid != 0) {
1130                 wq = smalloc(sizeof(Waitq));
1131                 wq->w.pid = up->pid;
1132                 utime = up->time[TUser] + up->time[TCUser];
1133                 stime = up->time[TSys] + up->time[TCSys];
1134                 wq->w.time[TUser] = tk2ms(utime);
1135                 wq->w.time[TSys] = tk2ms(stime);
1136                 wq->w.time[TReal] = tk2ms(MACHP(0)->ticks - up->time[TReal]);
1137                 if(exitstr != nil && exitstr[0])
1138                         snprint(wq->w.msg, sizeof(wq->w.msg), "%s %lud: %s", up->text, up->pid, exitstr);
1139                 else
1140                         wq->w.msg[0] = '\0';
1141
1142                 p = up->parent;
1143                 lock(&p->exl);
1144                 /*
1145                  * Check that parent is still alive.
1146                  */
1147                 if(p->pid == up->parentpid && p->state != Broken) {
1148                         p->nchild--;
1149                         p->time[TCUser] += utime;
1150                         p->time[TCSys] += stime;
1151                         /*
1152                          * If there would be more than 128 wait records
1153                          * processes for my parent, then don't leave a wait
1154                          * record behind.  This helps prevent badly written
1155                          * daemon processes from accumulating lots of wait
1156                          * records.
1157                          */
1158                         if(p->nwait < 128) {
1159                                 wq->next = p->waitq;
1160                                 p->waitq = wq;
1161                                 p->nwait++;
1162                                 wq = nil;
1163                                 wakeup(&p->waitr);
1164                         }
1165                 }
1166                 unlock(&p->exl);
1167                 if(wq != nil)
1168                         free(wq);
1169         }
1170         else if(up->kp == 0 && up->parent == nil){
1171                 if(exitstr == nil)
1172                         exitstr = "unknown";
1173                 panic("boot process died: %s", exitstr);
1174         }
1175
1176         if(!freemem)
1177                 addbroken(up);
1178
1179         qlock(&up->seglock);
1180         es = &up->seg[NSEG];
1181         for(s = up->seg; s < es; s++) {
1182                 if(*s != nil) {
1183                         putseg(*s);
1184                         *s = nil;
1185                 }
1186         }
1187         qunlock(&up->seglock);
1188
1189         lock(&up->exl);         /* Prevent my children from leaving waits */
1190         pidfree(up);
1191         up->pid = 0;
1192         wakeup(&up->waitr);
1193         unlock(&up->exl);
1194
1195         while((wq = up->waitq) != nil){
1196                 up->waitq = wq->next;
1197                 free(wq);
1198         }
1199
1200         /* release debuggers */
1201         qlock(&up->debug);
1202         if(up->pdbg != nil) {
1203                 wakeup(&up->pdbg->sleep);
1204                 up->pdbg = nil;
1205         }
1206         if(up->syscalltrace != nil) {
1207                 free(up->syscalltrace);
1208                 up->syscalltrace = nil;
1209         }
1210         if(up->watchpt != nil){
1211                 free(up->watchpt);
1212                 up->watchpt = nil;
1213         }
1214         up->nwatchpt = 0;
1215         qunlock(&up->debug);
1216
1217         /* Sched must not loop for these locks */
1218         lock(&procalloc);
1219         lock(&palloc);
1220
1221         edfstop(up);
1222         up->state = Moribund;
1223         sched();
1224         panic("pexit");
1225 }
1226
1227 static int
1228 haswaitq(void *x)
1229 {
1230         return ((Proc*)x)->waitq != nil;
1231 }
1232
1233 ulong
1234 pwait(Waitmsg *w)
1235 {
1236         ulong cpid;
1237         Waitq *wq;
1238
1239         if(!canqlock(&up->qwaitr))
1240                 error(Einuse);
1241
1242         if(waserror()) {
1243                 qunlock(&up->qwaitr);
1244                 nexterror();
1245         }
1246
1247         lock(&up->exl);
1248         while(up->waitq == nil) {
1249                 if(up->nchild == 0) {
1250                         unlock(&up->exl);
1251                         error(Enochild);
1252                 }
1253                 unlock(&up->exl);
1254                 sleep(&up->waitr, haswaitq, up);
1255                 lock(&up->exl);
1256         }
1257         wq = up->waitq;
1258         up->waitq = wq->next;
1259         up->nwait--;
1260         unlock(&up->exl);
1261
1262         qunlock(&up->qwaitr);
1263         poperror();
1264
1265         if(w != nil)
1266                 memmove(w, &wq->w, sizeof(Waitmsg));
1267         cpid = wq->w.pid;
1268         free(wq);
1269         return cpid;
1270 }
1271
1272 Proc*
1273 proctab(int i)
1274 {
1275         return &procalloc.arena[i];
1276 }
1277
1278 void
1279 dumpaproc(Proc *p)
1280 {
1281         ulong bss;
1282         char *s;
1283
1284         if(p == nil)
1285                 return;
1286
1287         bss = 0;
1288         if(p->seg[BSEG] != nil)
1289                 bss = p->seg[BSEG]->top;
1290
1291         s = p->psstate;
1292         if(s == nil)
1293                 s = statename[p->state];
1294         print("%3lud:%10s pc %#p dbgpc %#p  %8s (%s) ut %ld st %ld bss %lux qpc %#p nl %d nd %lud lpc %#p pri %lud\n",
1295                 p->pid, p->text, p->pc, dbgpc(p),  s, statename[p->state],
1296                 p->time[0], p->time[1], bss, p->qpc, p->nlocks, p->delaysched,
1297                 p->lastlock ? p->lastlock->pc : 0, p->priority);
1298 }
1299
1300 void
1301 procdump(void)
1302 {
1303         int i;
1304         Proc *p;
1305
1306         if(up != nil)
1307                 print("up %lud\n", up->pid);
1308         else
1309                 print("no current process\n");
1310         for(i=0; i<conf.nproc; i++) {
1311                 p = &procalloc.arena[i];
1312                 if(p->state != Dead)
1313                         dumpaproc(p);
1314         }
1315 }
1316
1317 /*
1318  *  wait till all matching processes have flushed their mmu
1319  */
1320 static void
1321 procflushmmu(int (*match)(Proc*, void*), void *a)
1322 {
1323         int i, nm, nwait;
1324         Proc *p;
1325
1326         /*
1327          *  tell all matching processes to flush their mmu's
1328          */
1329         nwait = 0;
1330         for(i=0; i<conf.nproc; i++) {
1331                 p = &procalloc.arena[i];
1332                 if(p->state != Dead && (*match)(p, a)){
1333                         p->newtlb = 1;
1334                         for(nm = 0; nm < conf.nmach; nm++){
1335                                 if(MACHP(nm)->proc == p){
1336                                         MACHP(nm)->flushmmu = 1;
1337                                         nwait++;
1338                                 }
1339                         }
1340                 }
1341         }
1342
1343         if(nwait == 0)
1344                 return;
1345
1346         /*
1347          *  wait for all other processors to take a clock interrupt
1348          *  and flush their mmu's
1349          */
1350         for(nm = 0; nm < conf.nmach; nm++)
1351                 while(m->machno != nm && MACHP(nm)->flushmmu)
1352                         sched();
1353 }
1354
1355 static int
1356 matchseg(Proc *p, void *a)
1357 {
1358         int ns;
1359
1360         for(ns = 0; ns < NSEG; ns++){
1361                 if(p->seg[ns] == a)
1362                         return 1;
1363         }
1364         return 0;
1365 }
1366 void
1367 procflushseg(Segment *s)
1368 {
1369         procflushmmu(matchseg, s);
1370 }
1371
1372 static int
1373 matchpseg(Proc *p, void *a)
1374 {
1375         Segment *s;
1376         int ns;
1377
1378         for(ns = 0; ns < NSEG; ns++){
1379                 s = p->seg[ns];
1380                 if(s != nil && s->pseg == a)
1381                         return 1;
1382         }
1383         return 0;
1384 }
1385 void
1386 procflushpseg(Physseg *ps)
1387 {
1388         procflushmmu(matchpseg, ps);
1389 }
1390
1391 void
1392 scheddump(void)
1393 {
1394         Proc *p;
1395         Schedq *rq;
1396
1397         for(rq = &runq[Nrq-1]; rq >= runq; rq--){
1398                 if(rq->head == nil)
1399                         continue;
1400                 print("rq%zd:", rq-runq);
1401                 for(p = rq->head; p != nil; p = p->rnext)
1402                         print(" %lud(%lud)", p->pid, m->ticks - p->readytime);
1403                 print("\n");
1404                 delay(150);
1405         }
1406         print("nrdy %d\n", nrdy);
1407 }
1408
1409 void
1410 kproc(char *name, void (*func)(void *), void *arg)
1411 {
1412         Proc *p;
1413         static Pgrp *kpgrp;
1414
1415         p = newproc();
1416         p->psstate = nil;
1417         p->procmode = 0640;
1418         p->kp = 1;
1419         p->noswap = 1;
1420
1421         p->scallnr = up->scallnr;
1422         p->s = up->s;
1423         p->nerrlab = 0;
1424         p->slash = up->slash;
1425         p->dot = up->slash;     /* unlike fork, do not inherit the dot for kprocs */
1426         if(p->dot != nil)
1427                 incref(p->dot);
1428
1429         memmove(p->note, up->note, sizeof(p->note));
1430         p->nnote = up->nnote;
1431         p->notified = 0;
1432         p->lastnote = up->lastnote;
1433         p->notify = up->notify;
1434         p->ureg = nil;
1435         p->dbgreg = nil;
1436
1437         procpriority(p, PriKproc, 0);
1438
1439         kprocchild(p, func, arg);
1440
1441         kstrdup(&p->user, eve);
1442         kstrdup(&p->text, name);
1443         if(kpgrp == nil)
1444                 kpgrp = newpgrp();
1445         p->pgrp = kpgrp;
1446         incref(kpgrp);
1447
1448         memset(p->time, 0, sizeof(p->time));
1449         p->time[TReal] = MACHP(0)->ticks;
1450         ready(p);
1451 }
1452
1453 /*
1454  *  called splhi() by notify().  See comment in notify for the
1455  *  reasoning.
1456  */
1457 void
1458 procctl(void)
1459 {
1460         char *state;
1461         ulong s;
1462
1463         switch(up->procctl) {
1464         case Proc_exitbig:
1465                 spllo();
1466                 pprint("Killed: Insufficient physical memory\n");
1467                 pexit("Killed: Insufficient physical memory", 1);
1468
1469         case Proc_exitme:
1470                 spllo();                /* pexit has locks in it */
1471                 pexit("Killed", 1);
1472
1473         case Proc_traceme:
1474                 if(up->nnote == 0)
1475                         return;
1476                 /* No break */
1477
1478         case Proc_stopme:
1479                 up->procctl = 0;
1480                 state = up->psstate;
1481                 up->psstate = "Stopped";
1482                 /* free a waiting debugger */
1483                 s = spllo();
1484                 qlock(&up->debug);
1485                 if(up->pdbg != nil) {
1486                         wakeup(&up->pdbg->sleep);
1487                         up->pdbg = nil;
1488                 }
1489                 qunlock(&up->debug);
1490                 splhi();
1491                 up->state = Stopped;
1492                 sched();
1493                 up->psstate = state;
1494                 splx(s);
1495                 return;
1496         }
1497 }
1498
1499 #include "errstr.h"
1500
1501 void
1502 error(char *err)
1503 {
1504         spllo();
1505
1506         assert(up->nerrlab < NERR);
1507         kstrcpy(up->errstr, err, ERRMAX);
1508         setlabel(&up->errlab[NERR-1]);
1509         nexterror();
1510 }
1511
1512 void
1513 nexterror(void)
1514 {
1515         assert(up->nerrlab > 0);
1516         gotolabel(&up->errlab[--up->nerrlab]);
1517 }
1518
1519 void
1520 exhausted(char *resource)
1521 {
1522         char buf[ERRMAX];
1523
1524         snprint(buf, sizeof buf, "no free %s", resource);
1525         iprint("%s\n", buf);
1526         error(buf);
1527 }
1528
1529 ulong
1530 procpagecount(Proc *p)
1531 {
1532         Segment *s;
1533         ulong pages;
1534         int i;
1535
1536         eqlock(&p->seglock);
1537         if(waserror()){
1538                 qunlock(&p->seglock);
1539                 nexterror();
1540         }
1541         pages = 0;
1542         for(i=0; i<NSEG; i++){
1543                 if((s = p->seg[i]) != nil){
1544                         eqlock(s);
1545                         pages += mcountseg(s);
1546                         qunlock(s);
1547                 }
1548         }
1549         qunlock(&p->seglock);
1550         poperror();
1551
1552         return pages;
1553 }
1554
1555 void
1556 killbig(char *why)
1557 {
1558         int i;
1559         Segment *s;
1560         ulong l, max;
1561         Proc *p, *ep, *kp;
1562
1563         max = 0;
1564         kp = nil;
1565         ep = procalloc.arena+conf.nproc;
1566         for(p = procalloc.arena; p < ep; p++) {
1567                 if(p->state == Dead || p->kp)
1568                         continue;
1569                 if((p->noswap || (p->procmode & 0222) == 0) && strcmp(eve, p->user) == 0)
1570                         continue;
1571                 l = procpagecount(p);
1572                 if(l > max){
1573                         kp = p;
1574                         max = l;
1575                 }
1576         }
1577         if(kp == nil)
1578                 return;
1579         print("%lud: %s killed: %s\n", kp->pid, kp->text, why);
1580         qlock(&kp->seglock);
1581         for(p = procalloc.arena; p < ep; p++) {
1582                 if(p->state == Dead || p->kp)
1583                         continue;
1584                 if(p != kp && p->seg[BSEG] != nil && p->seg[BSEG] == kp->seg[BSEG])
1585                         p->procctl = Proc_exitbig;
1586         }
1587         kp->procctl = Proc_exitbig;
1588         for(i = 0; i < NSEG; i++) {
1589                 s = kp->seg[i];
1590                 if(s == nil)
1591                         continue;
1592                 switch(s->type & SG_TYPE){
1593                 case SG_SHARED:
1594                 case SG_PHYSICAL:
1595                 case SG_FIXED:
1596                 case SG_STICKY:
1597                         continue;
1598                 }
1599                 qlock(s);
1600                 mfreeseg(s, s->base, (s->top - s->base)/BY2PG);
1601                 qunlock(s);
1602         }
1603         qunlock(&kp->seglock);
1604 }
1605
1606 /*
1607  *  change ownership to 'new' of all processes owned by 'old'.  Used when
1608  *  eve changes.
1609  */
1610 void
1611 renameuser(char *old, char *new)
1612 {
1613         Proc *p, *ep;
1614
1615         ep = procalloc.arena+conf.nproc;
1616         for(p = procalloc.arena; p < ep; p++)
1617                 if(p->user != nil && strcmp(old, p->user) == 0)
1618                         kstrdup(&p->user, new);
1619 }
1620
1621 /*
1622  *  time accounting called by clock() splhi'd
1623  */
1624 void
1625 accounttime(void)
1626 {
1627         Proc *p;
1628         ulong n, per;
1629         static ulong nrun;
1630
1631         p = m->proc;
1632         if(p != nil) {
1633                 nrun++;
1634                 p->time[p->insyscall]++;
1635         }
1636
1637         /* calculate decaying duty cycles */
1638         n = perfticks();
1639         per = n - m->perf.last;
1640         m->perf.last = n;
1641         per = ((uvlong)m->perf.period*(HZ-1) + per)/HZ;
1642         if(per != 0)
1643                 m->perf.period = per;
1644
1645         m->perf.avg_inidle = ((uvlong)m->perf.avg_inidle*(HZ-1)+m->perf.inidle)/HZ;
1646         m->perf.inidle = 0;
1647
1648         m->perf.avg_inintr = ((uvlong)m->perf.avg_inintr*(HZ-1)+m->perf.inintr)/HZ;
1649         m->perf.inintr = 0;
1650
1651         /* only one processor gets to compute system load averages */
1652         if(m->machno != 0)
1653                 return;
1654
1655         /*
1656          * calculate decaying load average.
1657          * if we decay by (n-1)/n then it takes
1658          * n clock ticks to go from load L to .36 L once
1659          * things quiet down.  it takes about 5 n clock
1660          * ticks to go to zero.  so using HZ means this is
1661          * approximately the load over the last second,
1662          * with a tail lasting about 5 seconds.
1663          */
1664         n = nrun;
1665         nrun = 0;
1666         n = (nrdy+n)*1000*100;
1667         load = ((uvlong)load*(HZ-1)+n)/HZ;
1668         m->load = load/100;
1669 }
1670
1671 int
1672 pidalloc(Proc *p)
1673 {
1674         static int gen, wrapped;
1675         int pid, h;
1676         Proc *x;
1677
1678         lock(&procalloc);
1679 Retry:
1680         pid = ++gen & 0x7FFFFFFF;
1681         if(pid == 0){
1682                 wrapped = 1;
1683                 goto Retry;
1684         }
1685         h = pid % nelem(procalloc.ht);
1686         if(wrapped)
1687                 for(x = procalloc.ht[h]; x != nil; x = x->pidhash)
1688                         if(x->pid == pid)
1689                                 goto Retry;
1690         if(p != nil){
1691                 p->pid = pid;
1692                 p->pidhash = procalloc.ht[h];
1693                 procalloc.ht[h] = p;
1694         }
1695         unlock(&procalloc);
1696         return pid;
1697 }
1698
1699 static void
1700 pidfree(Proc *p)
1701 {
1702         int h;
1703         Proc **l;
1704
1705         h = p->pid % nelem(procalloc.ht);
1706         lock(&procalloc);
1707         for(l = &procalloc.ht[h]; *l != nil; l = &(*l)->pidhash)
1708                 if(*l == p){
1709                         *l = p->pidhash;
1710                         break;
1711                 }
1712         unlock(&procalloc);
1713 }
1714
1715 int
1716 procindex(ulong pid)
1717 {
1718         Proc *p;
1719         int h;
1720         int s;
1721
1722         s = -1;
1723         h = pid % nelem(procalloc.ht);
1724         lock(&procalloc);
1725         for(p = procalloc.ht[h]; p != nil; p = p->pidhash)
1726                 if(p->pid == pid){
1727                         s = p - procalloc.arena;
1728                         break;
1729                 }
1730         unlock(&procalloc);
1731         return s;
1732 }