1 /* kern_clock.c 4.22 81/06/11 */ 2 3 #include "../h/param.h" 4 #include "../h/systm.h" 5 #include "../h/dk.h" 6 #include "../h/callout.h" 7 #include "../h/seg.h" 8 #include "../h/dir.h" 9 #include "../h/user.h" 10 #include "../h/proc.h" 11 #include "../h/reg.h" 12 #include "../h/psl.h" 13 #include "../h/vm.h" 14 #include "../h/buf.h" 15 #include "../h/text.h" 16 #include "../h/vlimit.h" 17 #include "../h/mtpr.h" 18 #include "../h/clock.h" 19 #include "../h/cpu.h" 20 21 #include "bk.h" 22 #include "dh.h" 23 #include "dz.h" 24 25 /* 26 * Hardclock is called straight from 27 * the real time clock interrupt. 28 * We limit the work we do at real clock interrupt time to: 29 * reloading clock 30 * decrementing time to callouts 31 * recording cpu time usage 32 * modifying priority of current process 33 * arrange for soft clock interrupt 34 * kernel pc profiling 35 * 36 * At software (softclock) interrupt time we: 37 * implement callouts 38 * maintain date 39 * lightning bolt wakeup (every second) 40 * alarm clock signals 41 * jab the scheduler 42 * 43 * On the vax softclock interrupts are implemented by 44 * software interrupts. Note that we may have multiple softclock 45 * interrupts compressed into one (due to excessive interrupt load), 46 * but that hardclock interrupts should never be lost. 47 */ 48 49 /*ARGSUSED*/ 50 hardclock(pc, ps) 51 caddr_t pc; 52 { 53 register struct callout *p1; 54 register struct proc *pp; 55 register int s, cpstate; 56 57 /* 58 * reprime clock 59 */ 60 clkreld(); 61 62 /* 63 * update callout times 64 */ 65 for (p1 = calltodo.c_next; p1 && p1->c_time <= 0; p1 = p1->c_next) 66 ; 67 if (p1) 68 p1->c_time--; 69 70 /* 71 * Maintain iostat and per-process cpu statistics 72 */ 73 if (!noproc) { 74 s = u.u_procp->p_rssize; 75 u.u_vm.vm_idsrss += s; 76 if (u.u_procp->p_textp) { 77 register int xrss = u.u_procp->p_textp->x_rssize; 78 79 s += xrss; 80 u.u_vm.vm_ixrss += xrss; 81 } 82 if (s > u.u_vm.vm_maxrss) 83 u.u_vm.vm_maxrss = s; 84 if ((u.u_vm.vm_utime+u.u_vm.vm_stime+1)/hz > u.u_limit[LIM_CPU]) { 85 psignal(u.u_procp, SIGXCPU); 86 if (u.u_limit[LIM_CPU] < INFINITY - 5) 87 u.u_limit[LIM_CPU] += 5; 88 } 89 } 90 /* 91 * Update iostat information. 92 */ 93 if (USERMODE(ps)) { 94 u.u_vm.vm_utime++; 95 if(u.u_procp->p_nice > NZERO) 96 cpstate = CP_NICE; 97 else 98 cpstate = CP_USER; 99 } else { 100 cpstate = CP_SYS; 101 if (noproc) 102 cpstate = CP_IDLE; 103 else 104 u.u_vm.vm_stime++; 105 } 106 cp_time[cpstate]++; 107 for (s = 0; s < DK_NDRIVE; s++) 108 if (dk_busy&(1<<s)) 109 dk_time[s]++; 110 /* 111 * Adjust priority of current process. 112 */ 113 if (!noproc) { 114 pp = u.u_procp; 115 pp->p_cpticks++; 116 if(++pp->p_cpu == 0) 117 pp->p_cpu--; 118 if(pp->p_cpu % 4 == 0) { 119 (void) setpri(pp); 120 if (pp->p_pri >= PUSER) 121 pp->p_pri = pp->p_usrpri; 122 } 123 } 124 /* 125 * Time moves on. 126 */ 127 ++lbolt; 128 #if VAX780 129 /* 130 * On 780's, impelement a fast UBA watcher, 131 * to make sure uba's don't get stuck. 132 */ 133 if (cpu == VAX_780 && panicstr == 0 && !BASEPRI(ps)) 134 unhang(); 135 #endif 136 /* 137 * Schedule a software interrupt for the rest 138 * of clock activities. 139 */ 140 setsoftclock(); 141 } 142 143 /* 144 * The digital decay cpu usage priority assignment is scaled to run in 145 * time as expanded by the 1 minute load average. Each second we 146 * multiply the the previous cpu usage estimate by 147 * nrscale*avenrun[0] 148 * The following relates the load average to the period over which 149 * cpu usage is 90% forgotten: 150 * loadav 1 5 seconds 151 * loadav 5 24 seconds 152 * loadav 10 47 seconds 153 * loadav 20 93 seconds 154 * This is a great improvement on the previous algorithm which 155 * decayed the priorities by a constant, and decayed away all knowledge 156 * of previous activity in about 20 seconds. Under heavy load, 157 * the previous algorithm degenerated to round-robin with poor response 158 * time when there was a high load average. 159 */ 160 #define ave(a,b) ((int)(((int)(a*b))/(b+1))) 161 int nrscale = 2; 162 double avenrun[]; 163 164 /* 165 * Constant for decay filter for cpu usage field 166 * in process table (used by ps au). 167 */ 168 double ccpu = 0.95122942450071400909; /* exp(-1/20) */ 169 170 /* 171 * Software clock interrupt. 172 * This routine runs at lower priority than device interrupts. 173 */ 174 /*ARGSUSED*/ 175 softclock(pc, ps) 176 caddr_t pc; 177 { 178 register struct callout *p1; 179 register struct proc *pp; 180 register int a, s; 181 caddr_t arg; 182 int (*func)(); 183 184 /* 185 * Perform callouts (but not after panic's!) 186 */ 187 if (panicstr == 0) { 188 for (;;) { 189 s = spl7(); 190 if ((p1 = calltodo.c_next) == 0 || p1->c_time > 0) 191 break; 192 calltodo.c_next = p1->c_next; 193 arg = p1->c_arg; 194 func = p1->c_func; 195 p1->c_next = callfree; 196 callfree = p1; 197 (void) splx(s); 198 (*func)(arg); 199 } 200 } 201 202 /* 203 * Drain silos. 204 */ 205 #if NBK > 0 206 #if NDH > 0 207 s = spl5(); dhtimer(); splx(s); 208 #endif 209 #if NDZ > 0 210 s = spl5(); dztimer(); splx(s); 211 #endif 212 #endif 213 214 /* 215 * If idling and processes are waiting to swap in, 216 * check on them. 217 */ 218 if (noproc && runin) { 219 runin = 0; 220 wakeup((caddr_t)&runin); 221 } 222 223 /* 224 * Run paging daemon every 1/4 sec. 225 */ 226 if (lbolt % (hz/4) == 0) { 227 vmpago(); 228 } 229 230 /* 231 * Reschedule every 1/10 sec. 232 */ 233 if (lbolt % (hz/10) == 0) { 234 runrun++; 235 aston(); 236 } 237 238 /* 239 * Lightning bolt every second: 240 * sleep timeouts 241 * process priority recomputation 242 * process %cpu averaging 243 * virtual memory metering 244 * kick swapper if processes want in 245 */ 246 if (lbolt >= hz) { 247 /* 248 * This doesn't mean much on VAX since we run at 249 * software interrupt time... if hardclock() 250 * calls softclock() directly, it prevents 251 * this code from running when the priority 252 * was raised when the clock interrupt occurred. 253 */ 254 if (BASEPRI(ps)) 255 return; 256 257 /* 258 * If we didn't run a few times because of 259 * long blockage at high ipl, we don't 260 * really want to run this code several times, 261 * so squish out all multiples of hz here. 262 */ 263 time += lbolt / hz; 264 lbolt %= hz; 265 266 /* 267 * Wakeup lightning bolt sleepers. 268 * Processes sleep on lbolt to wait 269 * for short amounts of time (e.g. 1 second). 270 */ 271 wakeup((caddr_t)&lbolt); 272 273 /* 274 * Recompute process priority and process 275 * sleep() system calls as well as internal 276 * sleeps with timeouts (tsleep() kernel routine). 277 */ 278 for (pp = proc; pp < procNPROC; pp++) 279 if (pp->p_stat && pp->p_stat!=SZOMB) { 280 /* 281 * Increase resident time, to max of 127 seconds 282 * (it is kept in a character.) For 283 * loaded processes this is time in core; for 284 * swapped processes, this is time on drum. 285 */ 286 if (pp->p_time != 127) 287 pp->p_time++; 288 /* 289 * If process has clock counting down, and it 290 * expires, set it running (if this is a tsleep()), 291 * or give it an SIGALRM (if the user process 292 * is using alarm signals. 293 */ 294 if (pp->p_clktim && --pp->p_clktim == 0) 295 if (pp->p_flag & STIMO) { 296 s = spl6(); 297 switch (pp->p_stat) { 298 299 case SSLEEP: 300 setrun(pp); 301 break; 302 303 case SSTOP: 304 unsleep(pp); 305 break; 306 } 307 pp->p_flag &= ~STIMO; 308 splx(s); 309 } else 310 psignal(pp, SIGALRM); 311 /* 312 * If process is blocked, increment computed 313 * time blocked. This is used in swap scheduling. 314 */ 315 if (pp->p_stat==SSLEEP || pp->p_stat==SSTOP) 316 if (pp->p_slptime != 127) 317 pp->p_slptime++; 318 /* 319 * Update digital filter estimation of process 320 * cpu utilization for loaded processes. 321 */ 322 if (pp->p_flag&SLOAD) 323 pp->p_pctcpu = ccpu * pp->p_pctcpu + 324 (1.0 - ccpu) * (pp->p_cpticks/(float)hz); 325 /* 326 * Recompute process priority. The number p_cpu 327 * is a weighted estimate of cpu time consumed. 328 * A process which consumes cpu time has this 329 * increase regularly. We here decrease it by 330 * a fraction based on load average giving a digital 331 * decay filter which damps out in about 5 seconds 332 * when seconds are measured in time expanded by the 333 * load average. 334 * 335 * If a process is niced, then the nice directly 336 * affects the new priority. The final priority 337 * is in the range 0 to 255, to fit in a character. 338 */ 339 pp->p_cpticks = 0; 340 a = ave((pp->p_cpu & 0377), avenrun[0]*nrscale) + 341 pp->p_nice - NZERO; 342 if (a < 0) 343 a = 0; 344 if (a > 255) 345 a = 255; 346 pp->p_cpu = a; 347 (void) setpri(pp); 348 /* 349 * Now have computed new process priority 350 * in p->p_usrpri. Carefully change p->p_pri. 351 * A process is on a run queue associated with 352 * this priority, so we must block out process 353 * state changes during the transition. 354 */ 355 s = spl6(); 356 if (pp->p_pri >= PUSER) { 357 if ((pp != u.u_procp || noproc) && 358 pp->p_stat == SRUN && 359 (pp->p_flag & SLOAD) && 360 pp->p_pri != pp->p_usrpri) { 361 remrq(pp); 362 pp->p_pri = pp->p_usrpri; 363 setrq(pp); 364 } else 365 pp->p_pri = pp->p_usrpri; 366 } 367 splx(s); 368 } 369 370 /* 371 * Perform virtual memory metering. 372 */ 373 vmmeter(); 374 375 /* 376 * If the swap process is trying to bring 377 * a process in, have it look again to see 378 * if it is possible now. 379 */ 380 if (runin!=0) { 381 runin = 0; 382 wakeup((caddr_t)&runin); 383 } 384 385 /* 386 * If there are pages that have been cleaned, 387 * jolt the pageout daemon to process them. 388 * We do this here so that these pages will be 389 * freed if there is an abundance of memory and the 390 * daemon would not be awakened otherwise. 391 */ 392 if (bclnlist != NULL) 393 wakeup((caddr_t)&proc[2]); 394 395 /* 396 * If the trap occurred from usermode, 397 * then check to see if it has now been 398 * running more than 10 minutes of user time 399 * and should thus run with reduced priority 400 * to give other processes a chance. 401 */ 402 if (USERMODE(ps)) { 403 pp = u.u_procp; 404 if (pp->p_uid && pp->p_nice == NZERO && 405 u.u_vm.vm_utime > 600 * hz) 406 pp->p_nice = NZERO+4; 407 (void) setpri(pp); 408 pp->p_pri = pp->p_usrpri; 409 } 410 } 411 /* 412 * If trapped user-mode, give it a profiling tick. 413 */ 414 if (USERMODE(ps) && u.u_prof.pr_scale) { 415 u.u_procp->p_flag |= SOWEUPC; 416 aston(); 417 } 418 } 419 420 /* 421 * Timeout is called to arrange that 422 * fun(arg) is called in tim/hz seconds. 423 * An entry is linked into the callout 424 * structure. The time in each structure 425 * entry is the number of hz's more 426 * than the previous entry. 427 * In this way, decrementing the 428 * first entry has the effect of 429 * updating all entries. 430 * 431 * The panic is there because there is nothing 432 * intelligent to be done if an entry won't fit. 433 */ 434 timeout(fun, arg, tim) 435 int (*fun)(); 436 caddr_t arg; 437 { 438 register struct callout *p1, *p2, *pnew; 439 register int t; 440 int s; 441 442 /* DEBUGGING CODE */ 443 int ttrstrt(); 444 445 if (fun == ttrstrt && arg == 0) 446 panic("timeout ttrstr arg"); 447 /* END DEBUGGING CODE */ 448 t = tim; 449 s = spl7(); 450 pnew = callfree; 451 if (pnew == NULL) 452 panic("timeout table overflow"); 453 callfree = pnew->c_next; 454 pnew->c_arg = arg; 455 pnew->c_func = fun; 456 for (p1 = &calltodo; (p2 = p1->c_next) && p2->c_time < t; p1 = p2) 457 t -= p2->c_time; 458 p1->c_next = pnew; 459 pnew->c_next = p2; 460 pnew->c_time = t; 461 if (p2) 462 p2->c_time -= t; 463 splx(s); 464 } 465