xref: /freebsd/sys/kern/kern_racct.c (revision 5ffd83dbcc34f10e07f6d3e968ae6365869615f4)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2010 The FreeBSD Foundation
5  * All rights reserved.
6  *
7  * This software was developed by Edward Tomasz Napierala under sponsorship
8  * from the FreeBSD Foundation.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  *
31  * $FreeBSD$
32  */
33 
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
36 
37 #include "opt_sched.h"
38 
39 #include <sys/param.h>
40 #include <sys/buf.h>
41 #include <sys/systm.h>
42 #include <sys/eventhandler.h>
43 #include <sys/jail.h>
44 #include <sys/kernel.h>
45 #include <sys/kthread.h>
46 #include <sys/lock.h>
47 #include <sys/loginclass.h>
48 #include <sys/malloc.h>
49 #include <sys/mutex.h>
50 #include <sys/proc.h>
51 #include <sys/racct.h>
52 #include <sys/resourcevar.h>
53 #include <sys/sbuf.h>
54 #include <sys/sched.h>
55 #include <sys/sdt.h>
56 #include <sys/smp.h>
57 #include <sys/sx.h>
58 #include <sys/sysctl.h>
59 #include <sys/sysent.h>
60 #include <sys/sysproto.h>
61 #include <sys/umtx.h>
62 #include <machine/smp.h>
63 
64 #ifdef RCTL
65 #include <sys/rctl.h>
66 #endif
67 
68 #ifdef RACCT
69 
70 FEATURE(racct, "Resource Accounting");
71 
72 /*
73  * Do not block processes that have their %cpu usage <= pcpu_threshold.
74  */
75 static int pcpu_threshold = 1;
76 #ifdef RACCT_DEFAULT_TO_DISABLED
77 bool __read_frequently racct_enable = false;
78 #else
79 bool __read_frequently racct_enable = true;
80 #endif
81 
82 SYSCTL_NODE(_kern, OID_AUTO, racct, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
83     "Resource Accounting");
84 SYSCTL_BOOL(_kern_racct, OID_AUTO, enable, CTLFLAG_RDTUN, &racct_enable,
85     0, "Enable RACCT/RCTL");
86 SYSCTL_UINT(_kern_racct, OID_AUTO, pcpu_threshold, CTLFLAG_RW, &pcpu_threshold,
87     0, "Processes with higher %cpu usage than this value can be throttled.");
88 
89 /*
90  * How many seconds it takes to use the scheduler %cpu calculations.  When a
91  * process starts, we compute its %cpu usage by dividing its runtime by the
92  * process wall clock time.  After RACCT_PCPU_SECS pass, we use the value
93  * provided by the scheduler.
94  */
95 #define RACCT_PCPU_SECS		3
96 
97 struct mtx racct_lock;
98 MTX_SYSINIT(racct_lock, &racct_lock, "racct lock", MTX_DEF);
99 
100 static uma_zone_t racct_zone;
101 
102 static void racct_sub_racct(struct racct *dest, const struct racct *src);
103 static void racct_sub_cred_locked(struct ucred *cred, int resource,
104 		uint64_t amount);
105 static void racct_add_cred_locked(struct ucred *cred, int resource,
106 		uint64_t amount);
107 
108 SDT_PROVIDER_DEFINE(racct);
109 SDT_PROBE_DEFINE3(racct, , rusage, add,
110     "struct proc *", "int", "uint64_t");
111 SDT_PROBE_DEFINE3(racct, , rusage, add__failure,
112     "struct proc *", "int", "uint64_t");
113 SDT_PROBE_DEFINE3(racct, , rusage, add__buf,
114     "struct proc *", "const struct buf *", "int");
115 SDT_PROBE_DEFINE3(racct, , rusage, add__cred,
116     "struct ucred *", "int", "uint64_t");
117 SDT_PROBE_DEFINE3(racct, , rusage, add__force,
118     "struct proc *", "int", "uint64_t");
119 SDT_PROBE_DEFINE3(racct, , rusage, set,
120     "struct proc *", "int", "uint64_t");
121 SDT_PROBE_DEFINE3(racct, , rusage, set__failure,
122     "struct proc *", "int", "uint64_t");
123 SDT_PROBE_DEFINE3(racct, , rusage, set__force,
124     "struct proc *", "int", "uint64_t");
125 SDT_PROBE_DEFINE3(racct, , rusage, sub,
126     "struct proc *", "int", "uint64_t");
127 SDT_PROBE_DEFINE3(racct, , rusage, sub__cred,
128     "struct ucred *", "int", "uint64_t");
129 SDT_PROBE_DEFINE1(racct, , racct, create,
130     "struct racct *");
131 SDT_PROBE_DEFINE1(racct, , racct, destroy,
132     "struct racct *");
133 SDT_PROBE_DEFINE2(racct, , racct, join,
134     "struct racct *", "struct racct *");
135 SDT_PROBE_DEFINE2(racct, , racct, join__failure,
136     "struct racct *", "struct racct *");
137 SDT_PROBE_DEFINE2(racct, , racct, leave,
138     "struct racct *", "struct racct *");
139 
140 int racct_types[] = {
141 	[RACCT_CPU] =
142 		RACCT_IN_MILLIONS,
143 	[RACCT_DATA] =
144 		RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
145 	[RACCT_STACK] =
146 		RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
147 	[RACCT_CORE] =
148 		RACCT_DENIABLE,
149 	[RACCT_RSS] =
150 		RACCT_RECLAIMABLE,
151 	[RACCT_MEMLOCK] =
152 		RACCT_RECLAIMABLE | RACCT_DENIABLE,
153 	[RACCT_NPROC] =
154 		RACCT_RECLAIMABLE | RACCT_DENIABLE,
155 	[RACCT_NOFILE] =
156 		RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
157 	[RACCT_VMEM] =
158 		RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
159 	[RACCT_NPTS] =
160 		RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
161 	[RACCT_SWAP] =
162 		RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
163 	[RACCT_NTHR] =
164 		RACCT_RECLAIMABLE | RACCT_DENIABLE,
165 	[RACCT_MSGQQUEUED] =
166 		RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
167 	[RACCT_MSGQSIZE] =
168 		RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
169 	[RACCT_NMSGQ] =
170 		RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
171 	[RACCT_NSEM] =
172 		RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
173 	[RACCT_NSEMOP] =
174 		RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
175 	[RACCT_NSHM] =
176 		RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
177 	[RACCT_SHMSIZE] =
178 		RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
179 	[RACCT_WALLCLOCK] =
180 		RACCT_IN_MILLIONS,
181 	[RACCT_PCTCPU] =
182 		RACCT_DECAYING | RACCT_DENIABLE | RACCT_IN_MILLIONS,
183 	[RACCT_READBPS] =
184 		RACCT_DECAYING,
185 	[RACCT_WRITEBPS] =
186 		RACCT_DECAYING,
187 	[RACCT_READIOPS] =
188 		RACCT_DECAYING,
189 	[RACCT_WRITEIOPS] =
190 		RACCT_DECAYING };
191 
192 static const fixpt_t RACCT_DECAY_FACTOR = 0.3 * FSCALE;
193 
194 #ifdef SCHED_4BSD
195 /*
196  * Contains intermediate values for %cpu calculations to avoid using floating
197  * point in the kernel.
198  * ccpu_exp[k] = FSCALE * (ccpu/FSCALE)^k = FSCALE * exp(-k/20)
199  * It is needed only for the 4BSD scheduler, because in ULE, the ccpu equals to
200  * zero so the calculations are more straightforward.
201  */
202 fixpt_t ccpu_exp[] = {
203 	[0] = FSCALE * 1,
204 	[1] = FSCALE * 0.95122942450071400909,
205 	[2] = FSCALE * 0.90483741803595957316,
206 	[3] = FSCALE * 0.86070797642505780722,
207 	[4] = FSCALE * 0.81873075307798185866,
208 	[5] = FSCALE * 0.77880078307140486824,
209 	[6] = FSCALE * 0.74081822068171786606,
210 	[7] = FSCALE * 0.70468808971871343435,
211 	[8] = FSCALE * 0.67032004603563930074,
212 	[9] = FSCALE * 0.63762815162177329314,
213 	[10] = FSCALE * 0.60653065971263342360,
214 	[11] = FSCALE * 0.57694981038048669531,
215 	[12] = FSCALE * 0.54881163609402643262,
216 	[13] = FSCALE * 0.52204577676101604789,
217 	[14] = FSCALE * 0.49658530379140951470,
218 	[15] = FSCALE * 0.47236655274101470713,
219 	[16] = FSCALE * 0.44932896411722159143,
220 	[17] = FSCALE * 0.42741493194872666992,
221 	[18] = FSCALE * 0.40656965974059911188,
222 	[19] = FSCALE * 0.38674102345450120691,
223 	[20] = FSCALE * 0.36787944117144232159,
224 	[21] = FSCALE * 0.34993774911115535467,
225 	[22] = FSCALE * 0.33287108369807955328,
226 	[23] = FSCALE * 0.31663676937905321821,
227 	[24] = FSCALE * 0.30119421191220209664,
228 	[25] = FSCALE * 0.28650479686019010032,
229 	[26] = FSCALE * 0.27253179303401260312,
230 	[27] = FSCALE * 0.25924026064589150757,
231 	[28] = FSCALE * 0.24659696394160647693,
232 	[29] = FSCALE * 0.23457028809379765313,
233 	[30] = FSCALE * 0.22313016014842982893,
234 	[31] = FSCALE * 0.21224797382674305771,
235 	[32] = FSCALE * 0.20189651799465540848,
236 	[33] = FSCALE * 0.19204990862075411423,
237 	[34] = FSCALE * 0.18268352405273465022,
238 	[35] = FSCALE * 0.17377394345044512668,
239 	[36] = FSCALE * 0.16529888822158653829,
240 	[37] = FSCALE * 0.15723716631362761621,
241 	[38] = FSCALE * 0.14956861922263505264,
242 	[39] = FSCALE * 0.14227407158651357185,
243 	[40] = FSCALE * 0.13533528323661269189,
244 	[41] = FSCALE * 0.12873490358780421886,
245 	[42] = FSCALE * 0.12245642825298191021,
246 	[43] = FSCALE * 0.11648415777349695786,
247 	[44] = FSCALE * 0.11080315836233388333,
248 	[45] = FSCALE * 0.10539922456186433678,
249 	[46] = FSCALE * 0.10025884372280373372,
250 	[47] = FSCALE * 0.09536916221554961888,
251 	[48] = FSCALE * 0.09071795328941250337,
252 	[49] = FSCALE * 0.08629358649937051097,
253 	[50] = FSCALE * 0.08208499862389879516,
254 	[51] = FSCALE * 0.07808166600115315231,
255 	[52] = FSCALE * 0.07427357821433388042,
256 	[53] = FSCALE * 0.07065121306042958674,
257 	[54] = FSCALE * 0.06720551273974976512,
258 	[55] = FSCALE * 0.06392786120670757270,
259 	[56] = FSCALE * 0.06081006262521796499,
260 	[57] = FSCALE * 0.05784432087483846296,
261 	[58] = FSCALE * 0.05502322005640722902,
262 	[59] = FSCALE * 0.05233970594843239308,
263 	[60] = FSCALE * 0.04978706836786394297,
264 	[61] = FSCALE * 0.04735892439114092119,
265 	[62] = FSCALE * 0.04504920239355780606,
266 	[63] = FSCALE * 0.04285212686704017991,
267 	[64] = FSCALE * 0.04076220397836621516,
268 	[65] = FSCALE * 0.03877420783172200988,
269 	[66] = FSCALE * 0.03688316740124000544,
270 	[67] = FSCALE * 0.03508435410084502588,
271 	[68] = FSCALE * 0.03337326996032607948,
272 	[69] = FSCALE * 0.03174563637806794323,
273 	[70] = FSCALE * 0.03019738342231850073,
274 	[71] = FSCALE * 0.02872463965423942912,
275 	[72] = FSCALE * 0.02732372244729256080,
276 	[73] = FSCALE * 0.02599112877875534358,
277 	[74] = FSCALE * 0.02472352647033939120,
278 	[75] = FSCALE * 0.02351774585600910823,
279 	[76] = FSCALE * 0.02237077185616559577,
280 	[77] = FSCALE * 0.02127973643837716938,
281 	[78] = FSCALE * 0.02024191144580438847,
282 	[79] = FSCALE * 0.01925470177538692429,
283 	[80] = FSCALE * 0.01831563888873418029,
284 	[81] = FSCALE * 0.01742237463949351138,
285 	[82] = FSCALE * 0.01657267540176124754,
286 	[83] = FSCALE * 0.01576441648485449082,
287 	[84] = FSCALE * 0.01499557682047770621,
288 	[85] = FSCALE * 0.01426423390899925527,
289 	[86] = FSCALE * 0.01356855901220093175,
290 	[87] = FSCALE * 0.01290681258047986886,
291 	[88] = FSCALE * 0.01227733990306844117,
292 	[89] = FSCALE * 0.01167856697039544521,
293 	[90] = FSCALE * 0.01110899653824230649,
294 	[91] = FSCALE * 0.01056720438385265337,
295 	[92] = FSCALE * 0.01005183574463358164,
296 	[93] = FSCALE * 0.00956160193054350793,
297 	[94] = FSCALE * 0.00909527710169581709,
298 	[95] = FSCALE * 0.00865169520312063417,
299 	[96] = FSCALE * 0.00822974704902002884,
300 	[97] = FSCALE * 0.00782837754922577143,
301 	[98] = FSCALE * 0.00744658307092434051,
302 	[99] = FSCALE * 0.00708340892905212004,
303 	[100] = FSCALE * 0.00673794699908546709,
304 	[101] = FSCALE * 0.00640933344625638184,
305 	[102] = FSCALE * 0.00609674656551563610,
306 	[103] = FSCALE * 0.00579940472684214321,
307 	[104] = FSCALE * 0.00551656442076077241,
308 	[105] = FSCALE * 0.00524751839918138427,
309 	[106] = FSCALE * 0.00499159390691021621,
310 	[107] = FSCALE * 0.00474815099941147558,
311 	[108] = FSCALE * 0.00451658094261266798,
312 	[109] = FSCALE * 0.00429630469075234057,
313 	[110] = FSCALE * 0.00408677143846406699,
314 };
315 #endif
316 
317 #define	CCPU_EXP_MAX	110
318 
319 /*
320  * This function is analogical to the getpcpu() function in the ps(1) command.
321  * They should both calculate in the same way so that the racct %cpu
322  * calculations are consistent with the values showed by the ps(1) tool.
323  * The calculations are more complex in the 4BSD scheduler because of the value
324  * of the ccpu variable.  In ULE it is defined to be zero which saves us some
325  * work.
326  */
327 static uint64_t
328 racct_getpcpu(struct proc *p, u_int pcpu)
329 {
330 	u_int swtime;
331 #ifdef SCHED_4BSD
332 	fixpt_t pctcpu, pctcpu_next;
333 #endif
334 #ifdef SMP
335 	struct pcpu *pc;
336 	int found;
337 #endif
338 	fixpt_t p_pctcpu;
339 	struct thread *td;
340 
341 	ASSERT_RACCT_ENABLED();
342 
343 	/*
344 	 * If the process is swapped out, we count its %cpu usage as zero.
345 	 * This behaviour is consistent with the userland ps(1) tool.
346 	 */
347 	if ((p->p_flag & P_INMEM) == 0)
348 		return (0);
349 	swtime = (ticks - p->p_swtick) / hz;
350 
351 	/*
352 	 * For short-lived processes, the sched_pctcpu() returns small
353 	 * values even for cpu intensive processes.  Therefore we use
354 	 * our own estimate in this case.
355 	 */
356 	if (swtime < RACCT_PCPU_SECS)
357 		return (pcpu);
358 
359 	p_pctcpu = 0;
360 	FOREACH_THREAD_IN_PROC(p, td) {
361 		if (td == PCPU_GET(idlethread))
362 			continue;
363 #ifdef SMP
364 		found = 0;
365 		STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
366 			if (td == pc->pc_idlethread) {
367 				found = 1;
368 				break;
369 			}
370 		}
371 		if (found)
372 			continue;
373 #endif
374 		thread_lock(td);
375 #ifdef SCHED_4BSD
376 		pctcpu = sched_pctcpu(td);
377 		/* Count also the yet unfinished second. */
378 		pctcpu_next = (pctcpu * ccpu_exp[1]) >> FSHIFT;
379 		pctcpu_next += sched_pctcpu_delta(td);
380 		p_pctcpu += max(pctcpu, pctcpu_next);
381 #else
382 		/*
383 		 * In ULE the %cpu statistics are updated on every
384 		 * sched_pctcpu() call.  So special calculations to
385 		 * account for the latest (unfinished) second are
386 		 * not needed.
387 		 */
388 		p_pctcpu += sched_pctcpu(td);
389 #endif
390 		thread_unlock(td);
391 	}
392 
393 #ifdef SCHED_4BSD
394 	if (swtime <= CCPU_EXP_MAX)
395 		return ((100 * (uint64_t)p_pctcpu * 1000000) /
396 		    (FSCALE - ccpu_exp[swtime]));
397 #endif
398 
399 	return ((100 * (uint64_t)p_pctcpu * 1000000) / FSCALE);
400 }
401 
402 static void
403 racct_add_racct(struct racct *dest, const struct racct *src)
404 {
405 	int i;
406 
407 	ASSERT_RACCT_ENABLED();
408 	RACCT_LOCK_ASSERT();
409 
410 	/*
411 	 * Update resource usage in dest.
412 	 */
413 	for (i = 0; i <= RACCT_MAX; i++) {
414 		KASSERT(dest->r_resources[i] >= 0,
415 		    ("%s: resource %d propagation meltdown: dest < 0",
416 		    __func__, i));
417 		KASSERT(src->r_resources[i] >= 0,
418 		    ("%s: resource %d propagation meltdown: src < 0",
419 		    __func__, i));
420 		dest->r_resources[i] += src->r_resources[i];
421 	}
422 }
423 
424 static void
425 racct_sub_racct(struct racct *dest, const struct racct *src)
426 {
427 	int i;
428 
429 	ASSERT_RACCT_ENABLED();
430 	RACCT_LOCK_ASSERT();
431 
432 	/*
433 	 * Update resource usage in dest.
434 	 */
435 	for (i = 0; i <= RACCT_MAX; i++) {
436 		if (!RACCT_IS_SLOPPY(i) && !RACCT_IS_DECAYING(i)) {
437 			KASSERT(dest->r_resources[i] >= 0,
438 			    ("%s: resource %d propagation meltdown: dest < 0",
439 			    __func__, i));
440 			KASSERT(src->r_resources[i] >= 0,
441 			    ("%s: resource %d propagation meltdown: src < 0",
442 			    __func__, i));
443 			KASSERT(src->r_resources[i] <= dest->r_resources[i],
444 			    ("%s: resource %d propagation meltdown: src > dest",
445 			    __func__, i));
446 		}
447 		if (RACCT_CAN_DROP(i)) {
448 			dest->r_resources[i] -= src->r_resources[i];
449 			if (dest->r_resources[i] < 0)
450 				dest->r_resources[i] = 0;
451 		}
452 	}
453 }
454 
455 void
456 racct_create(struct racct **racctp)
457 {
458 
459 	if (!racct_enable)
460 		return;
461 
462 	SDT_PROBE1(racct, , racct, create, racctp);
463 
464 	KASSERT(*racctp == NULL, ("racct already allocated"));
465 
466 	*racctp = uma_zalloc(racct_zone, M_WAITOK | M_ZERO);
467 }
468 
469 static void
470 racct_destroy_locked(struct racct **racctp)
471 {
472 	struct racct *racct;
473 	int i;
474 
475 	ASSERT_RACCT_ENABLED();
476 
477 	SDT_PROBE1(racct, , racct, destroy, racctp);
478 
479 	RACCT_LOCK_ASSERT();
480 	KASSERT(racctp != NULL, ("NULL racctp"));
481 	KASSERT(*racctp != NULL, ("NULL racct"));
482 
483 	racct = *racctp;
484 
485 	for (i = 0; i <= RACCT_MAX; i++) {
486 		if (RACCT_IS_SLOPPY(i))
487 			continue;
488 		if (!RACCT_IS_RECLAIMABLE(i))
489 			continue;
490 		KASSERT(racct->r_resources[i] == 0,
491 		    ("destroying non-empty racct: "
492 		    "%ju allocated for resource %d\n",
493 		    racct->r_resources[i], i));
494 	}
495 	uma_zfree(racct_zone, racct);
496 	*racctp = NULL;
497 }
498 
499 void
500 racct_destroy(struct racct **racct)
501 {
502 
503 	if (!racct_enable)
504 		return;
505 
506 	RACCT_LOCK();
507 	racct_destroy_locked(racct);
508 	RACCT_UNLOCK();
509 }
510 
511 /*
512  * Increase consumption of 'resource' by 'amount' for 'racct',
513  * but not its parents.  Differently from other cases, 'amount' here
514  * may be less than zero.
515  */
516 static void
517 racct_adjust_resource(struct racct *racct, int resource,
518     int64_t amount)
519 {
520 
521 	ASSERT_RACCT_ENABLED();
522 	RACCT_LOCK_ASSERT();
523 	KASSERT(racct != NULL, ("NULL racct"));
524 
525 	racct->r_resources[resource] += amount;
526 	if (racct->r_resources[resource] < 0) {
527 		KASSERT(RACCT_IS_SLOPPY(resource) || RACCT_IS_DECAYING(resource),
528 		    ("%s: resource %d usage < 0", __func__, resource));
529 		racct->r_resources[resource] = 0;
530 	}
531 
532 	/*
533 	 * There are some cases where the racct %cpu resource would grow
534 	 * beyond 100% per core.  For example in racct_proc_exit() we add
535 	 * the process %cpu usage to the ucred racct containers.  If too
536 	 * many processes terminated in a short time span, the ucred %cpu
537 	 * resource could grow too much.  Also, the 4BSD scheduler sometimes
538 	 * returns for a thread more than 100% cpu usage. So we set a sane
539 	 * boundary here to 100% * the maxumum number of CPUs.
540 	 */
541 	if ((resource == RACCT_PCTCPU) &&
542 	    (racct->r_resources[RACCT_PCTCPU] > 100 * 1000000 * (int64_t)MAXCPU))
543 		racct->r_resources[RACCT_PCTCPU] = 100 * 1000000 * (int64_t)MAXCPU;
544 }
545 
546 static int
547 racct_add_locked(struct proc *p, int resource, uint64_t amount, int force)
548 {
549 #ifdef RCTL
550 	int error;
551 #endif
552 
553 	ASSERT_RACCT_ENABLED();
554 
555 	/*
556 	 * We need proc lock to dereference p->p_ucred.
557 	 */
558 	PROC_LOCK_ASSERT(p, MA_OWNED);
559 
560 #ifdef RCTL
561 	error = rctl_enforce(p, resource, amount);
562 	if (error && !force && RACCT_IS_DENIABLE(resource)) {
563 		SDT_PROBE3(racct, , rusage, add__failure, p, resource, amount);
564 		return (error);
565 	}
566 #endif
567 	racct_adjust_resource(p->p_racct, resource, amount);
568 	racct_add_cred_locked(p->p_ucred, resource, amount);
569 
570 	return (0);
571 }
572 
573 /*
574  * Increase allocation of 'resource' by 'amount' for process 'p'.
575  * Return 0 if it's below limits, or errno, if it's not.
576  */
577 int
578 racct_add(struct proc *p, int resource, uint64_t amount)
579 {
580 	int error;
581 
582 	if (!racct_enable)
583 		return (0);
584 
585 	SDT_PROBE3(racct, , rusage, add, p, resource, amount);
586 
587 	RACCT_LOCK();
588 	error = racct_add_locked(p, resource, amount, 0);
589 	RACCT_UNLOCK();
590 	return (error);
591 }
592 
593 /*
594  * Increase allocation of 'resource' by 'amount' for process 'p'.
595  * Doesn't check for limits and never fails.
596  */
597 void
598 racct_add_force(struct proc *p, int resource, uint64_t amount)
599 {
600 
601 	if (!racct_enable)
602 		return;
603 
604 	SDT_PROBE3(racct, , rusage, add__force, p, resource, amount);
605 
606 	RACCT_LOCK();
607 	racct_add_locked(p, resource, amount, 1);
608 	RACCT_UNLOCK();
609 }
610 
611 static void
612 racct_add_cred_locked(struct ucred *cred, int resource, uint64_t amount)
613 {
614 	struct prison *pr;
615 
616 	ASSERT_RACCT_ENABLED();
617 
618 	racct_adjust_resource(cred->cr_ruidinfo->ui_racct, resource, amount);
619 	for (pr = cred->cr_prison; pr != NULL; pr = pr->pr_parent)
620 		racct_adjust_resource(pr->pr_prison_racct->prr_racct, resource,
621 		    amount);
622 	racct_adjust_resource(cred->cr_loginclass->lc_racct, resource, amount);
623 }
624 
625 /*
626  * Increase allocation of 'resource' by 'amount' for credential 'cred'.
627  * Doesn't check for limits and never fails.
628  */
629 void
630 racct_add_cred(struct ucred *cred, int resource, uint64_t amount)
631 {
632 
633 	if (!racct_enable)
634 		return;
635 
636 	SDT_PROBE3(racct, , rusage, add__cred, cred, resource, amount);
637 
638 	RACCT_LOCK();
639 	racct_add_cred_locked(cred, resource, amount);
640 	RACCT_UNLOCK();
641 }
642 
643 /*
644  * Account for disk IO resource consumption.  Checks for limits,
645  * but never fails, due to disk limits being undeniable.
646  */
647 void
648 racct_add_buf(struct proc *p, const struct buf *bp, int is_write)
649 {
650 
651 	ASSERT_RACCT_ENABLED();
652 	PROC_LOCK_ASSERT(p, MA_OWNED);
653 
654 	SDT_PROBE3(racct, , rusage, add__buf, p, bp, is_write);
655 
656 	RACCT_LOCK();
657 	if (is_write) {
658 		racct_add_locked(curproc, RACCT_WRITEBPS, bp->b_bcount, 1);
659 		racct_add_locked(curproc, RACCT_WRITEIOPS, 1, 1);
660 	} else {
661 		racct_add_locked(curproc, RACCT_READBPS, bp->b_bcount, 1);
662 		racct_add_locked(curproc, RACCT_READIOPS, 1, 1);
663 	}
664 	RACCT_UNLOCK();
665 }
666 
667 static int
668 racct_set_locked(struct proc *p, int resource, uint64_t amount, int force)
669 {
670 	int64_t old_amount, decayed_amount, diff_proc, diff_cred;
671 #ifdef RCTL
672 	int error;
673 #endif
674 
675 	ASSERT_RACCT_ENABLED();
676 
677 	/*
678 	 * We need proc lock to dereference p->p_ucred.
679 	 */
680 	PROC_LOCK_ASSERT(p, MA_OWNED);
681 
682 	old_amount = p->p_racct->r_resources[resource];
683 	/*
684 	 * The diffs may be negative.
685 	 */
686 	diff_proc = amount - old_amount;
687 	if (resource == RACCT_PCTCPU) {
688 		/*
689 		 * Resources in per-credential racct containers may decay.
690 		 * If this is the case, we need to calculate the difference
691 		 * between the new amount and the proportional value of the
692 		 * old amount that has decayed in the ucred racct containers.
693 		 */
694 		decayed_amount = old_amount * RACCT_DECAY_FACTOR / FSCALE;
695 		diff_cred = amount - decayed_amount;
696 	} else
697 		diff_cred = diff_proc;
698 #ifdef notyet
699 	KASSERT(diff_proc >= 0 || RACCT_CAN_DROP(resource),
700 	    ("%s: usage of non-droppable resource %d dropping", __func__,
701 	     resource));
702 #endif
703 #ifdef RCTL
704 	if (diff_proc > 0) {
705 		error = rctl_enforce(p, resource, diff_proc);
706 		if (error && !force && RACCT_IS_DENIABLE(resource)) {
707 			SDT_PROBE3(racct, , rusage, set__failure, p, resource,
708 			    amount);
709 			return (error);
710 		}
711 	}
712 #endif
713 	racct_adjust_resource(p->p_racct, resource, diff_proc);
714 	if (diff_cred > 0)
715 		racct_add_cred_locked(p->p_ucred, resource, diff_cred);
716 	else if (diff_cred < 0)
717 		racct_sub_cred_locked(p->p_ucred, resource, -diff_cred);
718 
719 	return (0);
720 }
721 
722 /*
723  * Set allocation of 'resource' to 'amount' for process 'p'.
724  * Return 0 if it's below limits, or errno, if it's not.
725  *
726  * Note that decreasing the allocation always returns 0,
727  * even if it's above the limit.
728  */
729 int
730 racct_set_unlocked(struct proc *p, int resource, uint64_t amount)
731 {
732 	int error;
733 
734 	ASSERT_RACCT_ENABLED();
735 	PROC_LOCK(p);
736 	error = racct_set(p, resource, amount);
737 	PROC_UNLOCK(p);
738 	return (error);
739 }
740 
741 int
742 racct_set(struct proc *p, int resource, uint64_t amount)
743 {
744 	int error;
745 
746 	if (!racct_enable)
747 		return (0);
748 
749 	SDT_PROBE3(racct, , rusage, set__force, p, resource, amount);
750 
751 	RACCT_LOCK();
752 	error = racct_set_locked(p, resource, amount, 0);
753 	RACCT_UNLOCK();
754 	return (error);
755 }
756 
757 void
758 racct_set_force(struct proc *p, int resource, uint64_t amount)
759 {
760 
761 	if (!racct_enable)
762 		return;
763 
764 	SDT_PROBE3(racct, , rusage, set, p, resource, amount);
765 
766 	RACCT_LOCK();
767 	racct_set_locked(p, resource, amount, 1);
768 	RACCT_UNLOCK();
769 }
770 
771 /*
772  * Returns amount of 'resource' the process 'p' can keep allocated.
773  * Allocating more than that would be denied, unless the resource
774  * is marked undeniable.  Amount of already allocated resource does
775  * not matter.
776  */
777 uint64_t
778 racct_get_limit(struct proc *p, int resource)
779 {
780 #ifdef RCTL
781 	uint64_t available;
782 
783 	if (!racct_enable)
784 		return (UINT64_MAX);
785 
786 	RACCT_LOCK();
787 	available = rctl_get_limit(p, resource);
788 	RACCT_UNLOCK();
789 
790 	return (available);
791 #else
792 
793 	return (UINT64_MAX);
794 #endif
795 }
796 
797 /*
798  * Returns amount of 'resource' the process 'p' can keep allocated.
799  * Allocating more than that would be denied, unless the resource
800  * is marked undeniable.  Amount of already allocated resource does
801  * matter.
802  */
803 uint64_t
804 racct_get_available(struct proc *p, int resource)
805 {
806 #ifdef RCTL
807 	uint64_t available;
808 
809 	if (!racct_enable)
810 		return (UINT64_MAX);
811 
812 	RACCT_LOCK();
813 	available = rctl_get_available(p, resource);
814 	RACCT_UNLOCK();
815 
816 	return (available);
817 #else
818 
819 	return (UINT64_MAX);
820 #endif
821 }
822 
823 /*
824  * Returns amount of the %cpu resource that process 'p' can add to its %cpu
825  * utilization.  Adding more than that would lead to the process being
826  * throttled.
827  */
828 static int64_t
829 racct_pcpu_available(struct proc *p)
830 {
831 #ifdef RCTL
832 	uint64_t available;
833 
834 	ASSERT_RACCT_ENABLED();
835 
836 	RACCT_LOCK();
837 	available = rctl_pcpu_available(p);
838 	RACCT_UNLOCK();
839 
840 	return (available);
841 #else
842 
843 	return (INT64_MAX);
844 #endif
845 }
846 
847 /*
848  * Decrease allocation of 'resource' by 'amount' for process 'p'.
849  */
850 void
851 racct_sub(struct proc *p, int resource, uint64_t amount)
852 {
853 
854 	if (!racct_enable)
855 		return;
856 
857 	SDT_PROBE3(racct, , rusage, sub, p, resource, amount);
858 
859 	/*
860 	 * We need proc lock to dereference p->p_ucred.
861 	 */
862 	PROC_LOCK_ASSERT(p, MA_OWNED);
863 	KASSERT(RACCT_CAN_DROP(resource),
864 	    ("%s: called for non-droppable resource %d", __func__, resource));
865 
866 	RACCT_LOCK();
867 	KASSERT(amount <= p->p_racct->r_resources[resource],
868 	    ("%s: freeing %ju of resource %d, which is more "
869 	     "than allocated %jd for %s (pid %d)", __func__, amount, resource,
870 	    (intmax_t)p->p_racct->r_resources[resource], p->p_comm, p->p_pid));
871 
872 	racct_adjust_resource(p->p_racct, resource, -amount);
873 	racct_sub_cred_locked(p->p_ucred, resource, amount);
874 	RACCT_UNLOCK();
875 }
876 
877 static void
878 racct_sub_cred_locked(struct ucred *cred, int resource, uint64_t amount)
879 {
880 	struct prison *pr;
881 
882 	ASSERT_RACCT_ENABLED();
883 
884 	racct_adjust_resource(cred->cr_ruidinfo->ui_racct, resource, -amount);
885 	for (pr = cred->cr_prison; pr != NULL; pr = pr->pr_parent)
886 		racct_adjust_resource(pr->pr_prison_racct->prr_racct, resource,
887 		    -amount);
888 	racct_adjust_resource(cred->cr_loginclass->lc_racct, resource, -amount);
889 }
890 
891 /*
892  * Decrease allocation of 'resource' by 'amount' for credential 'cred'.
893  */
894 void
895 racct_sub_cred(struct ucred *cred, int resource, uint64_t amount)
896 {
897 
898 	if (!racct_enable)
899 		return;
900 
901 	SDT_PROBE3(racct, , rusage, sub__cred, cred, resource, amount);
902 
903 #ifdef notyet
904 	KASSERT(RACCT_CAN_DROP(resource),
905 	    ("%s: called for resource %d which can not drop", __func__,
906 	     resource));
907 #endif
908 
909 	RACCT_LOCK();
910 	racct_sub_cred_locked(cred, resource, amount);
911 	RACCT_UNLOCK();
912 }
913 
914 /*
915  * Inherit resource usage information from the parent process.
916  */
917 int
918 racct_proc_fork(struct proc *parent, struct proc *child)
919 {
920 	int i, error = 0;
921 
922 	if (!racct_enable)
923 		return (0);
924 
925 	/*
926 	 * Create racct for the child process.
927 	 */
928 	racct_create(&child->p_racct);
929 
930 	PROC_LOCK(parent);
931 	PROC_LOCK(child);
932 	RACCT_LOCK();
933 
934 #ifdef RCTL
935 	error = rctl_proc_fork(parent, child);
936 	if (error != 0)
937 		goto out;
938 #endif
939 
940 	/* Init process cpu time. */
941 	child->p_prev_runtime = 0;
942 	child->p_throttled = 0;
943 
944 	/*
945 	 * Inherit resource usage.
946 	 */
947 	for (i = 0; i <= RACCT_MAX; i++) {
948 		if (parent->p_racct->r_resources[i] == 0 ||
949 		    !RACCT_IS_INHERITABLE(i))
950 			continue;
951 
952 		error = racct_set_locked(child, i,
953 		    parent->p_racct->r_resources[i], 0);
954 		if (error != 0)
955 			goto out;
956 	}
957 
958 	error = racct_add_locked(child, RACCT_NPROC, 1, 0);
959 	error += racct_add_locked(child, RACCT_NTHR, 1, 0);
960 
961 out:
962 	RACCT_UNLOCK();
963 	PROC_UNLOCK(child);
964 	PROC_UNLOCK(parent);
965 
966 	if (error != 0)
967 		racct_proc_exit(child);
968 
969 	return (error);
970 }
971 
972 /*
973  * Called at the end of fork1(), to handle rules that require the process
974  * to be fully initialized.
975  */
976 void
977 racct_proc_fork_done(struct proc *child)
978 {
979 
980 	if (!racct_enable)
981 		return;
982 
983 #ifdef RCTL
984 	PROC_LOCK(child);
985 	RACCT_LOCK();
986 	rctl_enforce(child, RACCT_NPROC, 0);
987 	rctl_enforce(child, RACCT_NTHR, 0);
988 	RACCT_UNLOCK();
989 	PROC_UNLOCK(child);
990 #endif
991 }
992 
993 void
994 racct_proc_exit(struct proc *p)
995 {
996 	struct timeval wallclock;
997 	uint64_t pct_estimate, pct, runtime;
998 	int i;
999 
1000 	if (!racct_enable)
1001 		return;
1002 
1003 	PROC_LOCK(p);
1004 	/*
1005 	 * We don't need to calculate rux, proc_reap() has already done this.
1006 	 */
1007 	runtime = cputick2usec(p->p_rux.rux_runtime);
1008 #ifdef notyet
1009 	KASSERT(runtime >= p->p_prev_runtime, ("runtime < p_prev_runtime"));
1010 #else
1011 	if (runtime < p->p_prev_runtime)
1012 		runtime = p->p_prev_runtime;
1013 #endif
1014 	microuptime(&wallclock);
1015 	timevalsub(&wallclock, &p->p_stats->p_start);
1016 	if (wallclock.tv_sec > 0 || wallclock.tv_usec > 0) {
1017 		pct_estimate = (1000000 * runtime * 100) /
1018 		    ((uint64_t)wallclock.tv_sec * 1000000 +
1019 		    wallclock.tv_usec);
1020 	} else
1021 		pct_estimate = 0;
1022 	pct = racct_getpcpu(p, pct_estimate);
1023 
1024 	RACCT_LOCK();
1025 	racct_set_locked(p, RACCT_CPU, runtime, 0);
1026 	racct_add_cred_locked(p->p_ucred, RACCT_PCTCPU, pct);
1027 
1028 	KASSERT(p->p_racct->r_resources[RACCT_RSS] == 0,
1029 	    ("process reaped with %ju allocated for RSS\n",
1030 	    p->p_racct->r_resources[RACCT_RSS]));
1031 	for (i = 0; i <= RACCT_MAX; i++) {
1032 		if (p->p_racct->r_resources[i] == 0)
1033 			continue;
1034 		if (!RACCT_IS_RECLAIMABLE(i))
1035 			continue;
1036 		racct_set_locked(p, i, 0, 0);
1037 	}
1038 
1039 #ifdef RCTL
1040 	rctl_racct_release(p->p_racct);
1041 #endif
1042 	racct_destroy_locked(&p->p_racct);
1043 	RACCT_UNLOCK();
1044 	PROC_UNLOCK(p);
1045 }
1046 
1047 /*
1048  * Called after credentials change, to move resource utilisation
1049  * between raccts.
1050  */
1051 void
1052 racct_proc_ucred_changed(struct proc *p, struct ucred *oldcred,
1053     struct ucred *newcred)
1054 {
1055 	struct uidinfo *olduip, *newuip;
1056 	struct loginclass *oldlc, *newlc;
1057 	struct prison *oldpr, *newpr, *pr;
1058 
1059 	if (!racct_enable)
1060 		return;
1061 
1062 	PROC_LOCK_ASSERT(p, MA_OWNED);
1063 
1064 	newuip = newcred->cr_ruidinfo;
1065 	olduip = oldcred->cr_ruidinfo;
1066 	newlc = newcred->cr_loginclass;
1067 	oldlc = oldcred->cr_loginclass;
1068 	newpr = newcred->cr_prison;
1069 	oldpr = oldcred->cr_prison;
1070 
1071 	RACCT_LOCK();
1072 	if (newuip != olduip) {
1073 		racct_sub_racct(olduip->ui_racct, p->p_racct);
1074 		racct_add_racct(newuip->ui_racct, p->p_racct);
1075 	}
1076 	if (newlc != oldlc) {
1077 		racct_sub_racct(oldlc->lc_racct, p->p_racct);
1078 		racct_add_racct(newlc->lc_racct, p->p_racct);
1079 	}
1080 	if (newpr != oldpr) {
1081 		for (pr = oldpr; pr != NULL; pr = pr->pr_parent)
1082 			racct_sub_racct(pr->pr_prison_racct->prr_racct,
1083 			    p->p_racct);
1084 		for (pr = newpr; pr != NULL; pr = pr->pr_parent)
1085 			racct_add_racct(pr->pr_prison_racct->prr_racct,
1086 			    p->p_racct);
1087 	}
1088 	RACCT_UNLOCK();
1089 }
1090 
1091 void
1092 racct_move(struct racct *dest, struct racct *src)
1093 {
1094 
1095 	ASSERT_RACCT_ENABLED();
1096 
1097 	RACCT_LOCK();
1098 	racct_add_racct(dest, src);
1099 	racct_sub_racct(src, src);
1100 	RACCT_UNLOCK();
1101 }
1102 
1103 void
1104 racct_proc_throttled(struct proc *p)
1105 {
1106 
1107 	ASSERT_RACCT_ENABLED();
1108 
1109 	PROC_LOCK(p);
1110 	while (p->p_throttled != 0) {
1111 		msleep(p->p_racct, &p->p_mtx, 0, "racct",
1112 		    p->p_throttled < 0 ? 0 : p->p_throttled);
1113 		if (p->p_throttled > 0)
1114 			p->p_throttled = 0;
1115 	}
1116 	PROC_UNLOCK(p);
1117 }
1118 
1119 /*
1120  * Make the process sleep in userret() for 'timeout' ticks.  Setting
1121  * timeout to -1 makes it sleep until woken up by racct_proc_wakeup().
1122  */
1123 void
1124 racct_proc_throttle(struct proc *p, int timeout)
1125 {
1126 	struct thread *td;
1127 #ifdef SMP
1128 	int cpuid;
1129 #endif
1130 
1131 	KASSERT(timeout != 0, ("timeout %d", timeout));
1132 	ASSERT_RACCT_ENABLED();
1133 	PROC_LOCK_ASSERT(p, MA_OWNED);
1134 
1135 	/*
1136 	 * Do not block kernel processes.  Also do not block processes with
1137 	 * low %cpu utilization to improve interactivity.
1138 	 */
1139 	if ((p->p_flag & (P_SYSTEM | P_KPROC)) != 0)
1140 		return;
1141 
1142 	if (p->p_throttled < 0 || (timeout > 0 && p->p_throttled > timeout))
1143 		return;
1144 
1145 	p->p_throttled = timeout;
1146 
1147 	FOREACH_THREAD_IN_PROC(p, td) {
1148 		thread_lock(td);
1149 		switch (td->td_state) {
1150 		case TDS_RUNQ:
1151 			/*
1152 			 * If the thread is on the scheduler run-queue, we can
1153 			 * not just remove it from there.  So we set the flag
1154 			 * TDF_NEEDRESCHED for the thread, so that once it is
1155 			 * running, it is taken off the cpu as soon as possible.
1156 			 */
1157 			td->td_flags |= TDF_NEEDRESCHED;
1158 			break;
1159 		case TDS_RUNNING:
1160 			/*
1161 			 * If the thread is running, we request a context
1162 			 * switch for it by setting the TDF_NEEDRESCHED flag.
1163 			 */
1164 			td->td_flags |= TDF_NEEDRESCHED;
1165 #ifdef SMP
1166 			cpuid = td->td_oncpu;
1167 			if ((cpuid != NOCPU) && (td != curthread))
1168 				ipi_cpu(cpuid, IPI_AST);
1169 #endif
1170 			break;
1171 		default:
1172 			break;
1173 		}
1174 		thread_unlock(td);
1175 	}
1176 }
1177 
1178 static void
1179 racct_proc_wakeup(struct proc *p)
1180 {
1181 
1182 	ASSERT_RACCT_ENABLED();
1183 
1184 	PROC_LOCK_ASSERT(p, MA_OWNED);
1185 
1186 	if (p->p_throttled != 0) {
1187 		p->p_throttled = 0;
1188 		wakeup(p->p_racct);
1189 	}
1190 }
1191 
1192 static void
1193 racct_decay_callback(struct racct *racct, void *dummy1, void *dummy2)
1194 {
1195 	int64_t r_old, r_new;
1196 
1197 	ASSERT_RACCT_ENABLED();
1198 	RACCT_LOCK_ASSERT();
1199 
1200 #ifdef RCTL
1201 	rctl_throttle_decay(racct, RACCT_READBPS);
1202 	rctl_throttle_decay(racct, RACCT_WRITEBPS);
1203 	rctl_throttle_decay(racct, RACCT_READIOPS);
1204 	rctl_throttle_decay(racct, RACCT_WRITEIOPS);
1205 #endif
1206 
1207 	r_old = racct->r_resources[RACCT_PCTCPU];
1208 
1209 	/* If there is nothing to decay, just exit. */
1210 	if (r_old <= 0)
1211 		return;
1212 
1213 	r_new = r_old * RACCT_DECAY_FACTOR / FSCALE;
1214 	racct->r_resources[RACCT_PCTCPU] = r_new;
1215 }
1216 
1217 static void
1218 racct_decay_pre(void)
1219 {
1220 
1221 	RACCT_LOCK();
1222 }
1223 
1224 static void
1225 racct_decay_post(void)
1226 {
1227 
1228 	RACCT_UNLOCK();
1229 }
1230 
1231 static void
1232 racct_decay(void)
1233 {
1234 
1235 	ASSERT_RACCT_ENABLED();
1236 
1237 	ui_racct_foreach(racct_decay_callback, racct_decay_pre,
1238 	    racct_decay_post, NULL, NULL);
1239 	loginclass_racct_foreach(racct_decay_callback, racct_decay_pre,
1240 	    racct_decay_post, NULL, NULL);
1241 	prison_racct_foreach(racct_decay_callback, racct_decay_pre,
1242 	    racct_decay_post, NULL, NULL);
1243 }
1244 
1245 static void
1246 racctd(void)
1247 {
1248 	struct thread *td;
1249 	struct proc *p;
1250 	struct timeval wallclock;
1251 	uint64_t pct, pct_estimate, runtime;
1252 
1253 	ASSERT_RACCT_ENABLED();
1254 
1255 	for (;;) {
1256 		racct_decay();
1257 
1258 		sx_slock(&allproc_lock);
1259 
1260 		FOREACH_PROC_IN_SYSTEM(p) {
1261 			PROC_LOCK(p);
1262 			if (p->p_state != PRS_NORMAL) {
1263 				if (p->p_state == PRS_ZOMBIE)
1264 					racct_set(p, RACCT_PCTCPU, 0);
1265 				PROC_UNLOCK(p);
1266 				continue;
1267 			}
1268 
1269 			microuptime(&wallclock);
1270 			timevalsub(&wallclock, &p->p_stats->p_start);
1271 			PROC_STATLOCK(p);
1272 			FOREACH_THREAD_IN_PROC(p, td)
1273 				ruxagg(p, td);
1274 			runtime = cputick2usec(p->p_rux.rux_runtime);
1275 			PROC_STATUNLOCK(p);
1276 #ifdef notyet
1277 			KASSERT(runtime >= p->p_prev_runtime,
1278 			    ("runtime < p_prev_runtime"));
1279 #else
1280 			if (runtime < p->p_prev_runtime)
1281 				runtime = p->p_prev_runtime;
1282 #endif
1283 			p->p_prev_runtime = runtime;
1284 			if (wallclock.tv_sec > 0 || wallclock.tv_usec > 0) {
1285 				pct_estimate = (1000000 * runtime * 100) /
1286 				    ((uint64_t)wallclock.tv_sec * 1000000 +
1287 				    wallclock.tv_usec);
1288 			} else
1289 				pct_estimate = 0;
1290 			pct = racct_getpcpu(p, pct_estimate);
1291 			RACCT_LOCK();
1292 #ifdef RCTL
1293 			rctl_throttle_decay(p->p_racct, RACCT_READBPS);
1294 			rctl_throttle_decay(p->p_racct, RACCT_WRITEBPS);
1295 			rctl_throttle_decay(p->p_racct, RACCT_READIOPS);
1296 			rctl_throttle_decay(p->p_racct, RACCT_WRITEIOPS);
1297 #endif
1298 			racct_set_locked(p, RACCT_PCTCPU, pct, 1);
1299 			racct_set_locked(p, RACCT_CPU, runtime, 0);
1300 			racct_set_locked(p, RACCT_WALLCLOCK,
1301 			    (uint64_t)wallclock.tv_sec * 1000000 +
1302 			    wallclock.tv_usec, 0);
1303 			RACCT_UNLOCK();
1304 			PROC_UNLOCK(p);
1305 		}
1306 
1307 		/*
1308 		 * To ensure that processes are throttled in a fair way, we need
1309 		 * to iterate over all processes again and check the limits
1310 		 * for %cpu resource only after ucred racct containers have been
1311 		 * properly filled.
1312 		 */
1313 		FOREACH_PROC_IN_SYSTEM(p) {
1314 			PROC_LOCK(p);
1315 			if (p->p_state != PRS_NORMAL) {
1316 				PROC_UNLOCK(p);
1317 				continue;
1318 			}
1319 
1320 			if (racct_pcpu_available(p) <= 0) {
1321 				if (p->p_racct->r_resources[RACCT_PCTCPU] >
1322 				    pcpu_threshold)
1323 					racct_proc_throttle(p, -1);
1324 			} else if (p->p_throttled == -1) {
1325 				racct_proc_wakeup(p);
1326 			}
1327 			PROC_UNLOCK(p);
1328 		}
1329 		sx_sunlock(&allproc_lock);
1330 		pause("-", hz);
1331 	}
1332 }
1333 
1334 static struct kproc_desc racctd_kp = {
1335 	"racctd",
1336 	racctd,
1337 	NULL
1338 };
1339 
1340 static void
1341 racctd_init(void)
1342 {
1343 	if (!racct_enable)
1344 		return;
1345 
1346 	kproc_start(&racctd_kp);
1347 }
1348 SYSINIT(racctd, SI_SUB_RACCTD, SI_ORDER_FIRST, racctd_init, NULL);
1349 
1350 static void
1351 racct_init(void)
1352 {
1353 	if (!racct_enable)
1354 		return;
1355 
1356 	racct_zone = uma_zcreate("racct", sizeof(struct racct),
1357 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
1358 	/*
1359 	 * XXX: Move this somewhere.
1360 	 */
1361 	prison0.pr_prison_racct = prison_racct_find("0");
1362 }
1363 SYSINIT(racct, SI_SUB_RACCT, SI_ORDER_FIRST, racct_init, NULL);
1364 
1365 #endif /* !RACCT */
1366