Commit | Line | Data |
---|---|---|
15637ed4 RG |
1 | /*- |
2 | * Copyright (c) 1982, 1986, 1991 The Regents of the University of California. | |
3 | * All rights reserved. | |
4 | * | |
5 | * Redistribution and use in source and binary forms, with or without | |
6 | * modification, are permitted provided that the following conditions | |
7 | * are met: | |
8 | * 1. Redistributions of source code must retain the above copyright | |
9 | * notice, this list of conditions and the following disclaimer. | |
10 | * 2. Redistributions in binary form must reproduce the above copyright | |
11 | * notice, this list of conditions and the following disclaimer in the | |
12 | * documentation and/or other materials provided with the distribution. | |
13 | * 3. All advertising materials mentioning features or use of this software | |
14 | * must display the following acknowledgement: | |
15 | * This product includes software developed by the University of | |
16 | * California, Berkeley and its contributors. | |
17 | * 4. Neither the name of the University nor the names of its contributors | |
18 | * may be used to endorse or promote products derived from this software | |
19 | * without specific prior written permission. | |
20 | * | |
21 | * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND | |
22 | * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE | |
23 | * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE | |
24 | * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE | |
25 | * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL | |
26 | * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS | |
27 | * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) | |
28 | * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT | |
29 | * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY | |
30 | * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF | |
31 | * SUCH DAMAGE. | |
32 | * | |
33 | * @(#)kern_clock.c 7.16 (Berkeley) 5/9/91 | |
34 | * | |
35 | * PATCHES MAGIC LEVEL PATCH THAT GOT US HERE | |
36 | * -------------------- ----- ---------------------- | |
37 | * CURRENT PATCH LEVEL: 2 00158 | |
38 | * -------------------- ----- ---------------------- | |
39 | * | |
40 | * 08 Apr 93 Poul-Henning Kamp Add support for dcfclock | |
41 | * 25 Apr 93 Bruce Evans Support new interrupt code (intr-0.1) | |
42 | */ | |
43 | ||
44 | #include "param.h" | |
45 | #include "systm.h" | |
46 | #include "dkstat.h" | |
47 | #include "callout.h" | |
48 | #include "kernel.h" | |
49 | #include "proc.h" | |
50 | #include "resourcevar.h" | |
51 | ||
52 | #include "machine/cpu.h" | |
53 | ||
54 | #ifdef GPROF | |
55 | #include "gprof.h" | |
56 | #endif | |
57 | ||
58 | /* | |
59 | * Clock handling routines. | |
60 | * | |
61 | * This code is written to operate with two timers which run | |
62 | * independently of each other. The main clock, running at hz | |
63 | * times per second, is used to do scheduling and timeout calculations. | |
64 | * The second timer does resource utilization estimation statistically | |
65 | * based on the state of the machine phz times a second. Both functions | |
66 | * can be performed by a single clock (ie hz == phz), however the | |
67 | * statistics will be much more prone to errors. Ideally a machine | |
68 | * would have separate clocks measuring time spent in user state, system | |
69 | * state, interrupt state, and idle state. These clocks would allow a non- | |
70 | * approximate measure of resource utilization. | |
71 | */ | |
72 | ||
73 | /* | |
74 | * TODO: | |
75 | * time of day, system/user timing, timeouts, profiling on separate timers | |
76 | * allocate more timeout table slots when table overflows. | |
77 | */ | |
78 | ||
79 | /* | |
80 | * Bump a timeval by a small number of usec's. | |
81 | */ | |
82 | #define BUMPTIME(t, usec) { \ | |
83 | register struct timeval *tp = (t); \ | |
84 | \ | |
85 | tp->tv_usec += (usec); \ | |
86 | if (tp->tv_usec >= 1000000) { \ | |
87 | tp->tv_usec -= 1000000; \ | |
88 | tp->tv_sec++; \ | |
89 | } \ | |
90 | } | |
91 | ||
92 | /* | |
93 | * The hz hardware interval timer. | |
94 | * We update the events relating to real time. | |
95 | * If this timer is also being used to gather statistics, | |
96 | * we run through the statistics gathering routine as well. | |
97 | */ | |
98 | hardclock(frame) | |
99 | clockframe frame; | |
100 | { | |
101 | register struct callout *p1; | |
102 | register struct proc *p = curproc; | |
103 | register struct pstats *pstats; | |
104 | register int s; | |
105 | int needsoft = 0; | |
106 | extern int tickdelta; | |
107 | extern long timedelta; | |
108 | ||
109 | /* | |
110 | * Update real-time timeout queue. | |
111 | * At front of queue are some number of events which are ``due''. | |
112 | * The time to these is <= 0 and if negative represents the | |
113 | * number of ticks which have passed since it was supposed to happen. | |
114 | * The rest of the q elements (times > 0) are events yet to happen, | |
115 | * where the time for each is given as a delta from the previous. | |
116 | * Decrementing just the first of these serves to decrement the time | |
117 | * to all events. | |
118 | */ | |
119 | p1 = calltodo.c_next; | |
120 | while (p1) { | |
121 | if (--p1->c_time > 0) | |
122 | break; | |
123 | needsoft = 1; | |
124 | if (p1->c_time == 0) | |
125 | break; | |
126 | p1 = p1->c_next; | |
127 | } | |
128 | ||
129 | /* | |
130 | * Curproc (now in p) is null if no process is running. | |
131 | * We assume that curproc is set in user mode! | |
132 | */ | |
133 | if (p) | |
134 | pstats = p->p_stats; | |
135 | /* | |
136 | * Charge the time out based on the mode the cpu is in. | |
137 | * Here again we fudge for the lack of proper interval timers | |
138 | * assuming that the current state has been around at least | |
139 | * one tick. | |
140 | */ | |
141 | if (CLKF_USERMODE(&frame)) { | |
142 | if (pstats->p_prof.pr_scale) | |
143 | needsoft = 1; | |
144 | /* | |
145 | * CPU was in user state. Increment | |
146 | * user time counter, and process process-virtual time | |
147 | * interval timer. | |
148 | */ | |
149 | BUMPTIME(&p->p_utime, tick); | |
150 | if (timerisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value) && | |
151 | itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0) | |
152 | psignal(p, SIGVTALRM); | |
153 | } else { | |
154 | /* | |
155 | * CPU was in system state. | |
156 | */ | |
157 | if (p) | |
158 | BUMPTIME(&p->p_stime, tick); | |
159 | } | |
160 | ||
161 | /* | |
162 | * If the cpu is currently scheduled to a process, then | |
163 | * charge it with resource utilization for a tick, updating | |
164 | * statistics which run in (user+system) virtual time, | |
165 | * such as the cpu time limit and profiling timers. | |
166 | * This assumes that the current process has been running | |
167 | * the entire last tick. | |
168 | */ | |
169 | if (p) { | |
170 | if ((p->p_utime.tv_sec+p->p_stime.tv_sec+1) > | |
171 | p->p_rlimit[RLIMIT_CPU].rlim_cur) { | |
172 | psignal(p, SIGXCPU); | |
173 | if (p->p_rlimit[RLIMIT_CPU].rlim_cur < | |
174 | p->p_rlimit[RLIMIT_CPU].rlim_max) | |
175 | p->p_rlimit[RLIMIT_CPU].rlim_cur += 5; | |
176 | } | |
177 | if (timerisset(&pstats->p_timer[ITIMER_PROF].it_value) && | |
178 | itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0) | |
179 | psignal(p, SIGPROF); | |
180 | ||
181 | /* | |
182 | * We adjust the priority of the current process. | |
183 | * The priority of a process gets worse as it accumulates | |
184 | * CPU time. The cpu usage estimator (p_cpu) is increased here | |
185 | * and the formula for computing priorities (in kern_synch.c) | |
186 | * will compute a different value each time the p_cpu increases | |
187 | * by 4. The cpu usage estimator ramps up quite quickly when | |
188 | * the process is running (linearly), and decays away | |
189 | * exponentially, * at a rate which is proportionally slower | |
190 | * when the system is busy. The basic principal is that the | |
191 | * system will 90% forget that a process used a lot of CPU | |
192 | * time in 5*loadav seconds. This causes the system to favor | |
193 | * processes which haven't run much recently, and to | |
194 | * round-robin among other processes. | |
195 | */ | |
196 | p->p_cpticks++; | |
197 | if (++p->p_cpu == 0) | |
198 | p->p_cpu--; | |
199 | if ((p->p_cpu&3) == 0) { | |
200 | setpri(p); | |
201 | if (p->p_pri >= PUSER) | |
202 | p->p_pri = p->p_usrpri; | |
203 | } | |
204 | } | |
205 | ||
206 | /* | |
207 | * If the alternate clock has not made itself known then | |
208 | * we must gather the statistics. | |
209 | */ | |
210 | if (phz == 0) | |
211 | gatherstats(&frame); | |
212 | ||
213 | /* | |
214 | * Increment the time-of-day, and schedule | |
215 | * processing of the callouts at a very low cpu priority, | |
216 | * so we don't keep the relatively high clock interrupt | |
217 | * priority any longer than necessary. | |
218 | */ | |
219 | if (timedelta == 0) | |
220 | BUMPTIME(&time, tick) | |
221 | else { | |
222 | register delta; | |
223 | ||
224 | if (timedelta < 0) { | |
225 | delta = tick - tickdelta; | |
226 | timedelta += tickdelta; | |
227 | } else { | |
228 | delta = tick + tickdelta; | |
229 | timedelta -= tickdelta; | |
230 | } | |
231 | BUMPTIME(&time, delta); | |
232 | } | |
233 | #ifdef DCFCLK | |
234 | /* | |
235 | * This is lousy, but until I can get the $&^%&^(!!! signal onto one | |
236 | * of the interrupt's I'll have to poll it. No, it will not work if | |
237 | * you attempt -DHZ=1000, things break. | |
238 | * But keep the NDCFCLK low, to avoid waste of cycles... | |
239 | * phk@data.fls.dk | |
240 | */ | |
241 | dcfclk_worker(); | |
242 | #endif | |
243 | if (needsoft) { | |
244 | #if 0 | |
245 | /* | |
246 | * XXX - hardclock runs at splhigh, so the splsoftclock is useless and | |
247 | * softclock runs at splhigh as well if we do this. It is not much of | |
248 | * an optimization, since the "software interrupt" is done with a call | |
249 | * from doreti, and the overhead of checking there is sometimes less | |
250 | * than checking here. Moreover, the whole %$$%$^ frame is passed by | |
251 | * value here. | |
252 | */ | |
253 | if (CLKF_BASEPRI(&frame)) { | |
254 | /* | |
255 | * Save the overhead of a software interrupt; | |
256 | * it will happen as soon as we return, so do it now. | |
257 | */ | |
258 | (void) splsoftclock(); | |
259 | softclock(frame); | |
260 | } else | |
261 | #endif | |
262 | setsoftclock(); | |
263 | } | |
264 | } | |
265 | ||
266 | int dk_ndrive = DK_NDRIVE; | |
267 | /* | |
268 | * Gather statistics on resource utilization. | |
269 | * | |
270 | * We make a gross assumption: that the system has been in the | |
271 | * state it is in (user state, kernel state, interrupt state, | |
272 | * or idle state) for the entire last time interval, and | |
273 | * update statistics accordingly. | |
274 | */ | |
275 | gatherstats(framep) | |
276 | clockframe *framep; | |
277 | { | |
278 | register int cpstate, s; | |
279 | ||
280 | /* | |
281 | * Determine what state the cpu is in. | |
282 | */ | |
283 | if (CLKF_USERMODE(framep)) { | |
284 | /* | |
285 | * CPU was in user state. | |
286 | */ | |
287 | if (curproc->p_nice > NZERO) | |
288 | cpstate = CP_NICE; | |
289 | else | |
290 | cpstate = CP_USER; | |
291 | } else { | |
292 | /* | |
293 | * CPU was in system state. If profiling kernel | |
294 | * increment a counter. If no process is running | |
295 | * then this is a system tick if we were running | |
296 | * at a non-zero IPL (in a driver). If a process is running, | |
297 | * then we charge it with system time even if we were | |
298 | * at a non-zero IPL, since the system often runs | |
299 | * this way during processing of system calls. | |
300 | * This is approximate, but the lack of true interval | |
301 | * timers makes doing anything else difficult. | |
302 | */ | |
303 | cpstate = CP_SYS; | |
304 | if (curproc == NULL && CLKF_BASEPRI(framep)) | |
305 | cpstate = CP_IDLE; | |
306 | #ifdef GPROF | |
307 | s = (u_long) CLKF_PC(framep) - (u_long) s_lowpc; | |
308 | if (profiling < 2 && s < s_textsize) | |
309 | kcount[s / (HISTFRACTION * sizeof (*kcount))]++; | |
310 | #endif | |
311 | } | |
312 | /* | |
313 | * We maintain statistics shown by user-level statistics | |
314 | * programs: the amount of time in each cpu state, and | |
315 | * the amount of time each of DK_NDRIVE ``drives'' is busy. | |
316 | */ | |
317 | cp_time[cpstate]++; | |
318 | for (s = 0; s < DK_NDRIVE; s++) | |
319 | if (dk_busy&(1<<s)) | |
320 | dk_time[s]++; | |
321 | } | |
322 | ||
323 | /* | |
324 | * Software priority level clock interrupt. | |
325 | * Run periodic events from timeout queue. | |
326 | */ | |
327 | /*ARGSUSED*/ | |
328 | softclock(frame) | |
329 | clockframe frame; | |
330 | { | |
331 | ||
332 | for (;;) { | |
333 | register struct callout *p1; | |
334 | register caddr_t arg; | |
335 | register int (*func)(); | |
336 | register int a, s; | |
337 | ||
338 | s = splhigh(); | |
339 | if ((p1 = calltodo.c_next) == 0 || p1->c_time > 0) { | |
340 | splx(s); | |
341 | break; | |
342 | } | |
343 | arg = p1->c_arg; func = p1->c_func; a = p1->c_time; | |
344 | calltodo.c_next = p1->c_next; | |
345 | p1->c_next = callfree; | |
346 | callfree = p1; | |
347 | splx(s); | |
348 | (*func)(arg, a); | |
349 | } | |
350 | ||
351 | /* | |
352 | * If no process to work with, we're finished. | |
353 | */ | |
354 | if (curproc == 0) return; | |
355 | ||
356 | /* | |
357 | * If trapped user-mode and profiling, give it | |
358 | * a profiling tick. | |
359 | */ | |
360 | if (CLKF_USERMODE(&frame)) { | |
361 | register struct proc *p = curproc; | |
362 | ||
363 | if (p->p_stats->p_prof.pr_scale) | |
364 | profile_tick(p, &frame); | |
365 | /* | |
366 | * Check to see if process has accumulated | |
367 | * more than 10 minutes of user time. If so | |
368 | * reduce priority to give others a chance. | |
369 | */ | |
370 | if (p->p_ucred->cr_uid && p->p_nice == NZERO && | |
371 | p->p_utime.tv_sec > 10 * 60) { | |
372 | p->p_nice = NZERO + 4; | |
373 | setpri(p); | |
374 | p->p_pri = p->p_usrpri; | |
375 | } | |
376 | } | |
377 | } | |
378 | ||
379 | /* | |
380 | * Arrange that (*func)(arg) is called in t/hz seconds. | |
381 | */ | |
382 | timeout(func, arg, t) | |
383 | int (*func)(); | |
384 | caddr_t arg; | |
385 | register int t; | |
386 | { | |
387 | register struct callout *p1, *p2, *pnew; | |
388 | register int s = splhigh(); | |
389 | ||
390 | if (t <= 0) | |
391 | t = 1; | |
392 | pnew = callfree; | |
393 | if (pnew == NULL) | |
394 | panic("timeout table overflow"); | |
395 | callfree = pnew->c_next; | |
396 | pnew->c_arg = arg; | |
397 | pnew->c_func = func; | |
398 | for (p1 = &calltodo; (p2 = p1->c_next) && p2->c_time < t; p1 = p2) | |
399 | if (p2->c_time > 0) | |
400 | t -= p2->c_time; | |
401 | p1->c_next = pnew; | |
402 | pnew->c_next = p2; | |
403 | pnew->c_time = t; | |
404 | if (p2) | |
405 | p2->c_time -= t; | |
406 | splx(s); | |
407 | } | |
408 | ||
409 | /* | |
410 | * untimeout is called to remove a function timeout call | |
411 | * from the callout structure. | |
412 | */ | |
413 | untimeout(func, arg) | |
414 | int (*func)(); | |
415 | caddr_t arg; | |
416 | { | |
417 | register struct callout *p1, *p2; | |
418 | register int s; | |
419 | ||
420 | s = splhigh(); | |
421 | for (p1 = &calltodo; (p2 = p1->c_next) != 0; p1 = p2) { | |
422 | if (p2->c_func == func && p2->c_arg == arg) { | |
423 | if (p2->c_next && p2->c_time > 0) | |
424 | p2->c_next->c_time += p2->c_time; | |
425 | p1->c_next = p2->c_next; | |
426 | p2->c_next = callfree; | |
427 | callfree = p2; | |
428 | break; | |
429 | } | |
430 | } | |
431 | splx(s); | |
432 | } | |
433 | ||
434 | /* | |
435 | * Compute number of hz until specified time. | |
436 | * Used to compute third argument to timeout() from an | |
437 | * absolute time. | |
438 | */ | |
439 | hzto(tv) | |
440 | struct timeval *tv; | |
441 | { | |
442 | register long ticks; | |
443 | register long sec; | |
444 | int s = splhigh(); | |
445 | ||
446 | /* | |
447 | * If number of milliseconds will fit in 32 bit arithmetic, | |
448 | * then compute number of milliseconds to time and scale to | |
449 | * ticks. Otherwise just compute number of hz in time, rounding | |
450 | * times greater than representible to maximum value. | |
451 | * | |
452 | * Delta times less than 25 days can be computed ``exactly''. | |
453 | * Maximum value for any timeout in 10ms ticks is 250 days. | |
454 | */ | |
455 | sec = tv->tv_sec - time.tv_sec; | |
456 | if (sec <= 0x7fffffff / 1000 - 1000) | |
457 | ticks = ((tv->tv_sec - time.tv_sec) * 1000 + | |
458 | (tv->tv_usec - time.tv_usec) / 1000) / (tick / 1000); | |
459 | else if (sec <= 0x7fffffff / hz) | |
460 | ticks = sec * hz; | |
461 | else | |
462 | ticks = 0x7fffffff; | |
463 | splx(s); | |
464 | return (ticks); | |
465 | } |