Commit | Line | Data |
---|---|---|
da7c5cc6 | 1 | /* |
0157085f | 2 | * Copyright (c) 1982, 1986, 1991 Regents of the University of California. |
da7c5cc6 KM |
3 | * All rights reserved. The Berkeley software License Agreement |
4 | * specifies the terms and conditions for redistribution. | |
5 | * | |
0157085f | 6 | * @(#)kern_clock.c 7.12 (Berkeley) %G% |
da7c5cc6 | 7 | */ |
961945a8 | 8 | |
94368568 JB |
9 | #include "param.h" |
10 | #include "systm.h" | |
fb1db32c | 11 | #include "dkstat.h" |
94368568 | 12 | #include "callout.h" |
d9c2f47f | 13 | #include "user.h" |
94368568 JB |
14 | #include "kernel.h" |
15 | #include "proc.h" | |
83be5fac | 16 | |
0157085f | 17 | #include "machine/cpu.h" |
961945a8 | 18 | |
8487304f | 19 | #ifdef GPROF |
94368568 | 20 | #include "gprof.h" |
8487304f KM |
21 | #endif |
22 | ||
45e9acec MK |
23 | #define ADJTIME /* For now... */ |
24 | #define ADJ_TICK 1000 | |
25 | int adjtimedelta; | |
26 | ||
76b2a182 BJ |
27 | /* |
28 | * Clock handling routines. | |
29 | * | |
53a32545 SL |
30 | * This code is written to operate with two timers which run |
31 | * independently of each other. The main clock, running at hz | |
32 | * times per second, is used to do scheduling and timeout calculations. | |
33 | * The second timer does resource utilization estimation statistically | |
34 | * based on the state of the machine phz times a second. Both functions | |
35 | * can be performed by a single clock (ie hz == phz), however the | |
36 | * statistics will be much more prone to errors. Ideally a machine | |
37 | * would have separate clocks measuring time spent in user state, system | |
38 | * state, interrupt state, and idle state. These clocks would allow a non- | |
39 | * approximate measure of resource utilization. | |
76b2a182 | 40 | */ |
6602c75b | 41 | |
76b2a182 BJ |
42 | /* |
43 | * TODO: | |
88a7a62a SL |
44 | * time of day, system/user timing, timeouts, profiling on separate timers |
45 | * allocate more timeout table slots when table overflows. | |
76b2a182 | 46 | */ |
9c5cfb8b | 47 | |
ad8023d1 KM |
48 | /* |
49 | * Bump a timeval by a small number of usec's. | |
50 | */ | |
ad8023d1 KM |
51 | #define BUMPTIME(t, usec) { \ |
52 | register struct timeval *tp = (t); \ | |
53 | \ | |
54 | tp->tv_usec += (usec); \ | |
55 | if (tp->tv_usec >= 1000000) { \ | |
56 | tp->tv_usec -= 1000000; \ | |
57 | tp->tv_sec++; \ | |
58 | } \ | |
59 | } | |
83be5fac | 60 | |
76b2a182 | 61 | /* |
53a32545 SL |
62 | * The hz hardware interval timer. |
63 | * We update the events relating to real time. | |
64 | * If this timer is also being used to gather statistics, | |
65 | * we run through the statistics gathering routine as well. | |
76b2a182 | 66 | */ |
d293217c | 67 | hardclock(frame) |
0157085f | 68 | clockframe frame; |
83be5fac | 69 | { |
0a34b6fd | 70 | register struct callout *p1; |
0157085f MK |
71 | register struct proc *p = curproc; |
72 | register struct pstats *pstats = p->p_stats; | |
0b355a6e | 73 | register int s; |
83be5fac | 74 | |
76b2a182 BJ |
75 | /* |
76 | * Update real-time timeout queue. | |
77 | * At front of queue are some number of events which are ``due''. | |
78 | * The time to these is <= 0 and if negative represents the | |
79 | * number of ticks which have passed since it was supposed to happen. | |
80 | * The rest of the q elements (times > 0) are events yet to happen, | |
81 | * where the time for each is given as a delta from the previous. | |
82 | * Decrementing just the first of these serves to decrement the time | |
83 | * to all events. | |
84 | */ | |
88a7a62a SL |
85 | p1 = calltodo.c_next; |
86 | while (p1) { | |
87 | if (--p1->c_time > 0) | |
88 | break; | |
88a7a62a SL |
89 | if (p1->c_time == 0) |
90 | break; | |
91 | p1 = p1->c_next; | |
92 | } | |
5da67d35 | 93 | |
76b2a182 BJ |
94 | /* |
95 | * Charge the time out based on the mode the cpu is in. | |
96 | * Here again we fudge for the lack of proper interval timers | |
97 | * assuming that the current state has been around at least | |
98 | * one tick. | |
99 | */ | |
76b2a182 BJ |
100 | /* |
101 | * CPU was in user state. Increment | |
102 | * user time counter, and process process-virtual time | |
877ef342 | 103 | * interval timer. |
76b2a182 | 104 | */ |
53fbb3b3 | 105 | BUMPTIME(&p->p_utime, tick); |
0157085f MK |
106 | if (timerisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value) && |
107 | itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0) | |
53fbb3b3 | 108 | psignal(p, SIGVTALRM); |
83be5fac | 109 | } else { |
76b2a182 | 110 | /* |
0b355a6e | 111 | * CPU was in system state. |
76b2a182 | 112 | */ |
9c5cfb8b | 113 | if (!noproc) |
53fbb3b3 | 114 | BUMPTIME(&p->p_stime, tick); |
83be5fac | 115 | } |
27b91f59 | 116 | |
9fb1a8d0 SL |
117 | /* |
118 | * If the cpu is currently scheduled to a process, then | |
119 | * charge it with resource utilization for a tick, updating | |
120 | * statistics which run in (user+system) virtual time, | |
121 | * such as the cpu time limit and profiling timers. | |
122 | * This assumes that the current process has been running | |
123 | * the entire last tick. | |
124 | */ | |
405de916 | 125 | if (noproc == 0) { |
53fbb3b3 | 126 | if ((p->p_utime.tv_sec+p->p_stime.tv_sec+1) > |
0157085f | 127 | p->p_rlimit[RLIMIT_CPU].rlim_cur) { |
53fbb3b3 | 128 | psignal(p, SIGXCPU); |
0157085f MK |
129 | if (p->p_rlimit[RLIMIT_CPU].rlim_cur < |
130 | p->p_rlimit[RLIMIT_CPU].rlim_max) | |
131 | p->p_rlimit[RLIMIT_CPU].rlim_cur += 5; | |
9fb1a8d0 | 132 | } |
0157085f MK |
133 | if (timerisset(&pstats->p_timer[ITIMER_PROF].it_value) && |
134 | itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0) | |
53fbb3b3 | 135 | psignal(p, SIGPROF); |
9fb1a8d0 | 136 | |
0157085f MK |
137 | /* |
138 | * We adjust the priority of the current process. | |
139 | * The priority of a process gets worse as it accumulates | |
140 | * CPU time. The cpu usage estimator (p_cpu) is increased here | |
141 | * and the formula for computing priorities (in kern_synch.c) | |
142 | * will compute a different value each time the p_cpu increases | |
143 | * by 4. The cpu usage estimator ramps up quite quickly when | |
144 | * the process is running (linearly), and decays away | |
145 | * exponentially, * at a rate which is proportionally slower | |
146 | * when the system is busy. The basic principal is that the | |
147 | * system will 90% forget that a process used a lot of CPU | |
148 | * time in 5*loadav seconds. This causes the system to favor | |
149 | * processes which haven't run much recently, and to | |
150 | * round-robin among other processes. | |
151 | */ | |
27b91f59 BJ |
152 | p->p_cpticks++; |
153 | if (++p->p_cpu == 0) | |
154 | p->p_cpu--; | |
76b2a182 | 155 | if ((p->p_cpu&3) == 0) { |
0157085f | 156 | setpri(p); |
27b91f59 BJ |
157 | if (p->p_pri >= PUSER) |
158 | p->p_pri = p->p_usrpri; | |
83be5fac BJ |
159 | } |
160 | } | |
76b2a182 | 161 | |
53a32545 SL |
162 | /* |
163 | * If the alternate clock has not made itself known then | |
164 | * we must gather the statistics. | |
165 | */ | |
166 | if (phz == 0) | |
0157085f | 167 | gatherstats(&frame); |
53a32545 | 168 | |
76b2a182 BJ |
169 | /* |
170 | * Increment the time-of-day, and schedule | |
171 | * processing of the callouts at a very low cpu priority, | |
172 | * so we don't keep the relatively high clock interrupt | |
173 | * priority any longer than necessary. | |
174 | */ | |
45e9acec MK |
175 | #ifdef ADJTIME |
176 | if (adjtimedelta == 0) | |
177 | bumptime(&time, tick); | |
178 | else { | |
179 | if (adjtimedelta < 0) { | |
180 | bumptime(&time, tick-ADJ_TICK); | |
181 | adjtimedelta++; | |
182 | } else { | |
183 | bumptime(&time, tick+ADJ_TICK); | |
184 | adjtimedelta--; | |
185 | } | |
186 | } | |
187 | #else | |
4ca0d0d6 | 188 | if (timedelta == 0) |
99e47f6b MK |
189 | BUMPTIME(&time, tick) |
190 | else { | |
191 | register delta; | |
192 | ||
4ca0d0d6 MK |
193 | if (timedelta < 0) { |
194 | delta = tick - tickdelta; | |
195 | timedelta += tickdelta; | |
99e47f6b | 196 | } else { |
4ca0d0d6 MK |
197 | delta = tick + tickdelta; |
198 | timedelta -= tickdelta; | |
99e47f6b MK |
199 | } |
200 | BUMPTIME(&time, delta); | |
201 | } | |
45e9acec | 202 | #endif |
ca6b57a4 | 203 | setsoftclock(); |
f403d99f BJ |
204 | } |
205 | ||
d976d466 | 206 | int dk_ndrive = DK_NDRIVE; |
53a32545 SL |
207 | /* |
208 | * Gather statistics on resource utilization. | |
209 | * | |
210 | * We make a gross assumption: that the system has been in the | |
211 | * state it is in (user state, kernel state, interrupt state, | |
212 | * or idle state) for the entire last time interval, and | |
213 | * update statistics accordingly. | |
214 | */ | |
0157085f MK |
215 | gatherstats(framep) |
216 | clockframe *framep; | |
53a32545 | 217 | { |
9c5cfb8b | 218 | register int cpstate, s; |
53a32545 SL |
219 | |
220 | /* | |
221 | * Determine what state the cpu is in. | |
222 | */ | |
0157085f | 223 | if (CLKF_USERMODE(framep)) { |
53a32545 SL |
224 | /* |
225 | * CPU was in user state. | |
226 | */ | |
0157085f | 227 | if (curproc->p_nice > NZERO) |
53a32545 SL |
228 | cpstate = CP_NICE; |
229 | else | |
230 | cpstate = CP_USER; | |
231 | } else { | |
232 | /* | |
233 | * CPU was in system state. If profiling kernel | |
0b355a6e JB |
234 | * increment a counter. If no process is running |
235 | * then this is a system tick if we were running | |
236 | * at a non-zero IPL (in a driver). If a process is running, | |
237 | * then we charge it with system time even if we were | |
238 | * at a non-zero IPL, since the system often runs | |
239 | * this way during processing of system calls. | |
240 | * This is approximate, but the lack of true interval | |
241 | * timers makes doing anything else difficult. | |
53a32545 SL |
242 | */ |
243 | cpstate = CP_SYS; | |
0157085f | 244 | if (noproc && CLKF_BASEPRI(framep)) |
53a32545 SL |
245 | cpstate = CP_IDLE; |
246 | #ifdef GPROF | |
0157085f | 247 | s = CLKF_PC(framep) - s_lowpc; |
53a32545 SL |
248 | if (profiling < 2 && s < s_textsize) |
249 | kcount[s / (HISTFRACTION * sizeof (*kcount))]++; | |
250 | #endif | |
251 | } | |
252 | /* | |
253 | * We maintain statistics shown by user-level statistics | |
254 | * programs: the amount of time in each cpu state, and | |
255 | * the amount of time each of DK_NDRIVE ``drives'' is busy. | |
256 | */ | |
257 | cp_time[cpstate]++; | |
258 | for (s = 0; s < DK_NDRIVE; s++) | |
fb1db32c | 259 | if (dk_busy&(1<<s)) |
53a32545 SL |
260 | dk_time[s]++; |
261 | } | |
262 | ||
76b2a182 BJ |
263 | /* |
264 | * Software priority level clock interrupt. | |
265 | * Run periodic events from timeout queue. | |
266 | */ | |
260ea681 | 267 | /*ARGSUSED*/ |
d293217c | 268 | softclock(frame) |
0157085f | 269 | clockframe frame; |
f403d99f | 270 | { |
f403d99f | 271 | |
27b91f59 | 272 | for (;;) { |
76b2a182 BJ |
273 | register struct callout *p1; |
274 | register caddr_t arg; | |
275 | register int (*func)(); | |
276 | register int a, s; | |
277 | ||
9c5cfb8b | 278 | s = splhigh(); |
27b91f59 BJ |
279 | if ((p1 = calltodo.c_next) == 0 || p1->c_time > 0) { |
280 | splx(s); | |
281 | break; | |
f403d99f | 282 | } |
76b2a182 | 283 | arg = p1->c_arg; func = p1->c_func; a = p1->c_time; |
27b91f59 | 284 | calltodo.c_next = p1->c_next; |
27b91f59 BJ |
285 | p1->c_next = callfree; |
286 | callfree = p1; | |
4f083fd7 | 287 | splx(s); |
d01b68d6 | 288 | (*func)(arg, a); |
f403d99f | 289 | } |
877ef342 | 290 | /* |
db1f1262 SL |
291 | * If trapped user-mode and profiling, give it |
292 | * a profiling tick. | |
877ef342 | 293 | */ |
0157085f MK |
294 | if (CLKF_USERMODE(&frame)) { |
295 | register struct proc *p = curproc; | |
db1f1262 | 296 | |
0157085f MK |
297 | if (p->p_stats->p_prof.pr_scale) |
298 | profile_tick(p, &frame); | |
db1f1262 SL |
299 | /* |
300 | * Check to see if process has accumulated | |
301 | * more than 10 minutes of user time. If so | |
302 | * reduce priority to give others a chance. | |
303 | */ | |
0157085f | 304 | if (p->p_ucred->cr_uid && p->p_nice == NZERO && |
53fbb3b3 | 305 | p->p_utime.tv_sec > 10 * 60) { |
0157085f MK |
306 | p->p_nice = NZERO + 4; |
307 | setpri(p); | |
db1f1262 SL |
308 | p->p_pri = p->p_usrpri; |
309 | } | |
877ef342 | 310 | } |
83be5fac BJ |
311 | } |
312 | ||
88a7a62a | 313 | /* |
0157085f | 314 | * Arrange that (*func)(arg) is called in t/hz seconds. |
83be5fac | 315 | */ |
0157085f MK |
316 | timeout(func, arg, t) |
317 | int (*func)(); | |
4512b9a4 | 318 | caddr_t arg; |
88a7a62a | 319 | register int t; |
83be5fac | 320 | { |
c4710996 | 321 | register struct callout *p1, *p2, *pnew; |
9c5cfb8b | 322 | register int s = splhigh(); |
83be5fac | 323 | |
ba96129b | 324 | if (t <= 0) |
88a7a62a | 325 | t = 1; |
c4710996 BJ |
326 | pnew = callfree; |
327 | if (pnew == NULL) | |
328 | panic("timeout table overflow"); | |
329 | callfree = pnew->c_next; | |
330 | pnew->c_arg = arg; | |
0157085f | 331 | pnew->c_func = func; |
c4710996 | 332 | for (p1 = &calltodo; (p2 = p1->c_next) && p2->c_time < t; p1 = p2) |
d45b61eb SL |
333 | if (p2->c_time > 0) |
334 | t -= p2->c_time; | |
c4710996 BJ |
335 | p1->c_next = pnew; |
336 | pnew->c_next = p2; | |
337 | pnew->c_time = t; | |
338 | if (p2) | |
339 | p2->c_time -= t; | |
83be5fac BJ |
340 | splx(s); |
341 | } | |
1fa9ff62 SL |
342 | |
343 | /* | |
344 | * untimeout is called to remove a function timeout call | |
345 | * from the callout structure. | |
346 | */ | |
0157085f MK |
347 | untimeout(func, arg) |
348 | int (*func)(); | |
1fa9ff62 SL |
349 | caddr_t arg; |
350 | { | |
1fa9ff62 SL |
351 | register struct callout *p1, *p2; |
352 | register int s; | |
353 | ||
9c5cfb8b | 354 | s = splhigh(); |
1fa9ff62 | 355 | for (p1 = &calltodo; (p2 = p1->c_next) != 0; p1 = p2) { |
0157085f | 356 | if (p2->c_func == func && p2->c_arg == arg) { |
d01b68d6 | 357 | if (p2->c_next && p2->c_time > 0) |
1fa9ff62 SL |
358 | p2->c_next->c_time += p2->c_time; |
359 | p1->c_next = p2->c_next; | |
360 | p2->c_next = callfree; | |
361 | callfree = p2; | |
362 | break; | |
363 | } | |
364 | } | |
365 | splx(s); | |
366 | } | |
d01b68d6 | 367 | |
76b2a182 BJ |
368 | /* |
369 | * Compute number of hz until specified time. | |
370 | * Used to compute third argument to timeout() from an | |
371 | * absolute time. | |
372 | */ | |
d01b68d6 BJ |
373 | hzto(tv) |
374 | struct timeval *tv; | |
375 | { | |
76b2a182 BJ |
376 | register long ticks; |
377 | register long sec; | |
9c5cfb8b | 378 | int s = splhigh(); |
d01b68d6 | 379 | |
76b2a182 BJ |
380 | /* |
381 | * If number of milliseconds will fit in 32 bit arithmetic, | |
382 | * then compute number of milliseconds to time and scale to | |
383 | * ticks. Otherwise just compute number of hz in time, rounding | |
384 | * times greater than representible to maximum value. | |
385 | * | |
386 | * Delta times less than 25 days can be computed ``exactly''. | |
387 | * Maximum value for any timeout in 10ms ticks is 250 days. | |
388 | */ | |
389 | sec = tv->tv_sec - time.tv_sec; | |
390 | if (sec <= 0x7fffffff / 1000 - 1000) | |
391 | ticks = ((tv->tv_sec - time.tv_sec) * 1000 + | |
392 | (tv->tv_usec - time.tv_usec) / 1000) / (tick / 1000); | |
393 | else if (sec <= 0x7fffffff / hz) | |
394 | ticks = sec * hz; | |
395 | else | |
396 | ticks = 0x7fffffff; | |
d01b68d6 BJ |
397 | splx(s); |
398 | return (ticks); | |
399 | } | |
88a7a62a | 400 | |
330a6ff9 KM |
401 | /* ARGSUSED */ |
402 | profil(p, uap, retval) | |
403 | struct proc *p; | |
404 | register struct args { | |
88a7a62a SL |
405 | short *bufbase; |
406 | unsigned bufsize; | |
407 | unsigned pcoffset; | |
408 | unsigned pcscale; | |
330a6ff9 KM |
409 | } *uap; |
410 | int *retval; | |
411 | { | |
0157085f | 412 | register struct uprof *upp = &p->p_stats->p_prof; |
88a7a62a SL |
413 | |
414 | upp->pr_base = uap->bufbase; | |
415 | upp->pr_size = uap->bufsize; | |
416 | upp->pr_off = uap->pcoffset; | |
417 | upp->pr_scale = uap->pcscale; | |
d9c2f47f | 418 | return (0); |
88a7a62a | 419 | } |