xref: /freebsd/sys/kern/kern_racct.c (revision e9ac41698b2f322d55ccf9da50a3596edb2c1800)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2010 The FreeBSD Foundation
5  *
6  * This software was developed by Edward Tomasz Napierala under sponsorship
7  * from the FreeBSD Foundation.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  */
30 
31 #include <sys/cdefs.h>
32 #include "opt_sched.h"
33 
34 #include <sys/param.h>
35 #include <sys/buf.h>
36 #include <sys/systm.h>
37 #include <sys/eventhandler.h>
38 #include <sys/jail.h>
39 #include <sys/kernel.h>
40 #include <sys/kthread.h>
41 #include <sys/lock.h>
42 #include <sys/loginclass.h>
43 #include <sys/malloc.h>
44 #include <sys/mutex.h>
45 #include <sys/proc.h>
46 #include <sys/racct.h>
47 #include <sys/resourcevar.h>
48 #include <sys/sbuf.h>
49 #include <sys/sched.h>
50 #include <sys/sdt.h>
51 #include <sys/smp.h>
52 #include <sys/sx.h>
53 #include <sys/sysctl.h>
54 #include <sys/sysproto.h>
55 #include <sys/umtxvar.h>
56 #include <machine/smp.h>
57 
58 #ifdef RCTL
59 #include <sys/rctl.h>
60 #endif
61 
62 FEATURE(racct, "Resource Accounting");
63 
64 /*
65  * Do not block processes that have their %cpu usage <= pcpu_threshold.
66  */
67 static int pcpu_threshold = 1;
68 #ifdef RACCT_DEFAULT_TO_DISABLED
69 bool __read_frequently racct_enable = false;
70 #else
71 bool __read_frequently racct_enable = true;
72 #endif
73 
74 SYSCTL_NODE(_kern, OID_AUTO, racct, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
75     "Resource Accounting");
76 SYSCTL_BOOL(_kern_racct, OID_AUTO, enable, CTLFLAG_RDTUN, &racct_enable,
77     0, "Enable RACCT/RCTL");
78 SYSCTL_UINT(_kern_racct, OID_AUTO, pcpu_threshold, CTLFLAG_RW, &pcpu_threshold,
79     0, "Processes with higher %cpu usage than this value can be throttled.");
80 
81 /*
82  * How many seconds it takes to use the scheduler %cpu calculations.  When a
83  * process starts, we compute its %cpu usage by dividing its runtime by the
84  * process wall clock time.  After RACCT_PCPU_SECS pass, we use the value
85  * provided by the scheduler.
86  */
87 #define RACCT_PCPU_SECS		3
88 
89 struct mtx racct_lock;
90 MTX_SYSINIT(racct_lock, &racct_lock, "racct lock", MTX_DEF);
91 
92 static uma_zone_t racct_zone;
93 
94 static void racct_sub_racct(struct racct *dest, const struct racct *src);
95 static void racct_sub_cred_locked(struct ucred *cred, int resource,
96 		uint64_t amount);
97 static void racct_add_cred_locked(struct ucred *cred, int resource,
98 		uint64_t amount);
99 
100 SDT_PROVIDER_DEFINE(racct);
101 SDT_PROBE_DEFINE3(racct, , rusage, add,
102     "struct proc *", "int", "uint64_t");
103 SDT_PROBE_DEFINE3(racct, , rusage, add__failure,
104     "struct proc *", "int", "uint64_t");
105 SDT_PROBE_DEFINE3(racct, , rusage, add__buf,
106     "struct proc *", "const struct buf *", "int");
107 SDT_PROBE_DEFINE3(racct, , rusage, add__cred,
108     "struct ucred *", "int", "uint64_t");
109 SDT_PROBE_DEFINE3(racct, , rusage, add__force,
110     "struct proc *", "int", "uint64_t");
111 SDT_PROBE_DEFINE3(racct, , rusage, set,
112     "struct proc *", "int", "uint64_t");
113 SDT_PROBE_DEFINE3(racct, , rusage, set__failure,
114     "struct proc *", "int", "uint64_t");
115 SDT_PROBE_DEFINE3(racct, , rusage, set__force,
116     "struct proc *", "int", "uint64_t");
117 SDT_PROBE_DEFINE3(racct, , rusage, sub,
118     "struct proc *", "int", "uint64_t");
119 SDT_PROBE_DEFINE3(racct, , rusage, sub__cred,
120     "struct ucred *", "int", "uint64_t");
121 SDT_PROBE_DEFINE1(racct, , racct, create,
122     "struct racct *");
123 SDT_PROBE_DEFINE1(racct, , racct, destroy,
124     "struct racct *");
125 SDT_PROBE_DEFINE2(racct, , racct, join,
126     "struct racct *", "struct racct *");
127 SDT_PROBE_DEFINE2(racct, , racct, join__failure,
128     "struct racct *", "struct racct *");
129 SDT_PROBE_DEFINE2(racct, , racct, leave,
130     "struct racct *", "struct racct *");
131 
132 int racct_types[] = {
133 	[RACCT_CPU] =
134 		RACCT_IN_MILLIONS,
135 	[RACCT_DATA] =
136 		RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
137 	[RACCT_STACK] =
138 		RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
139 	[RACCT_CORE] =
140 		RACCT_DENIABLE,
141 	[RACCT_RSS] =
142 		RACCT_RECLAIMABLE,
143 	[RACCT_MEMLOCK] =
144 		RACCT_RECLAIMABLE | RACCT_DENIABLE,
145 	[RACCT_NPROC] =
146 		RACCT_RECLAIMABLE | RACCT_DENIABLE,
147 	[RACCT_NOFILE] =
148 		RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
149 	[RACCT_VMEM] =
150 		RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
151 	[RACCT_NPTS] =
152 		RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
153 	[RACCT_SWAP] =
154 		RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
155 	[RACCT_NTHR] =
156 		RACCT_RECLAIMABLE | RACCT_DENIABLE,
157 	[RACCT_MSGQQUEUED] =
158 		RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
159 	[RACCT_MSGQSIZE] =
160 		RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
161 	[RACCT_NMSGQ] =
162 		RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
163 	[RACCT_NSEM] =
164 		RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
165 	[RACCT_NSEMOP] =
166 		RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
167 	[RACCT_NSHM] =
168 		RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
169 	[RACCT_SHMSIZE] =
170 		RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
171 	[RACCT_WALLCLOCK] =
172 		RACCT_IN_MILLIONS,
173 	[RACCT_PCTCPU] =
174 		RACCT_DECAYING | RACCT_DENIABLE | RACCT_IN_MILLIONS,
175 	[RACCT_READBPS] =
176 		RACCT_DECAYING,
177 	[RACCT_WRITEBPS] =
178 		RACCT_DECAYING,
179 	[RACCT_READIOPS] =
180 		RACCT_DECAYING,
181 	[RACCT_WRITEIOPS] =
182 		RACCT_DECAYING };
183 
184 static const fixpt_t RACCT_DECAY_FACTOR = 0.3 * FSCALE;
185 
186 #ifdef SCHED_4BSD
187 /*
188  * Contains intermediate values for %cpu calculations to avoid using floating
189  * point in the kernel.
190  * ccpu_exp[k] = FSCALE * (ccpu/FSCALE)^k = FSCALE * exp(-k/20)
191  * It is needed only for the 4BSD scheduler, because in ULE, the ccpu equals to
192  * zero so the calculations are more straightforward.
193  */
194 fixpt_t ccpu_exp[] = {
195 	[0] = FSCALE * 1,
196 	[1] = FSCALE * 0.95122942450071400909,
197 	[2] = FSCALE * 0.90483741803595957316,
198 	[3] = FSCALE * 0.86070797642505780722,
199 	[4] = FSCALE * 0.81873075307798185866,
200 	[5] = FSCALE * 0.77880078307140486824,
201 	[6] = FSCALE * 0.74081822068171786606,
202 	[7] = FSCALE * 0.70468808971871343435,
203 	[8] = FSCALE * 0.67032004603563930074,
204 	[9] = FSCALE * 0.63762815162177329314,
205 	[10] = FSCALE * 0.60653065971263342360,
206 	[11] = FSCALE * 0.57694981038048669531,
207 	[12] = FSCALE * 0.54881163609402643262,
208 	[13] = FSCALE * 0.52204577676101604789,
209 	[14] = FSCALE * 0.49658530379140951470,
210 	[15] = FSCALE * 0.47236655274101470713,
211 	[16] = FSCALE * 0.44932896411722159143,
212 	[17] = FSCALE * 0.42741493194872666992,
213 	[18] = FSCALE * 0.40656965974059911188,
214 	[19] = FSCALE * 0.38674102345450120691,
215 	[20] = FSCALE * 0.36787944117144232159,
216 	[21] = FSCALE * 0.34993774911115535467,
217 	[22] = FSCALE * 0.33287108369807955328,
218 	[23] = FSCALE * 0.31663676937905321821,
219 	[24] = FSCALE * 0.30119421191220209664,
220 	[25] = FSCALE * 0.28650479686019010032,
221 	[26] = FSCALE * 0.27253179303401260312,
222 	[27] = FSCALE * 0.25924026064589150757,
223 	[28] = FSCALE * 0.24659696394160647693,
224 	[29] = FSCALE * 0.23457028809379765313,
225 	[30] = FSCALE * 0.22313016014842982893,
226 	[31] = FSCALE * 0.21224797382674305771,
227 	[32] = FSCALE * 0.20189651799465540848,
228 	[33] = FSCALE * 0.19204990862075411423,
229 	[34] = FSCALE * 0.18268352405273465022,
230 	[35] = FSCALE * 0.17377394345044512668,
231 	[36] = FSCALE * 0.16529888822158653829,
232 	[37] = FSCALE * 0.15723716631362761621,
233 	[38] = FSCALE * 0.14956861922263505264,
234 	[39] = FSCALE * 0.14227407158651357185,
235 	[40] = FSCALE * 0.13533528323661269189,
236 	[41] = FSCALE * 0.12873490358780421886,
237 	[42] = FSCALE * 0.12245642825298191021,
238 	[43] = FSCALE * 0.11648415777349695786,
239 	[44] = FSCALE * 0.11080315836233388333,
240 	[45] = FSCALE * 0.10539922456186433678,
241 	[46] = FSCALE * 0.10025884372280373372,
242 	[47] = FSCALE * 0.09536916221554961888,
243 	[48] = FSCALE * 0.09071795328941250337,
244 	[49] = FSCALE * 0.08629358649937051097,
245 	[50] = FSCALE * 0.08208499862389879516,
246 	[51] = FSCALE * 0.07808166600115315231,
247 	[52] = FSCALE * 0.07427357821433388042,
248 	[53] = FSCALE * 0.07065121306042958674,
249 	[54] = FSCALE * 0.06720551273974976512,
250 	[55] = FSCALE * 0.06392786120670757270,
251 	[56] = FSCALE * 0.06081006262521796499,
252 	[57] = FSCALE * 0.05784432087483846296,
253 	[58] = FSCALE * 0.05502322005640722902,
254 	[59] = FSCALE * 0.05233970594843239308,
255 	[60] = FSCALE * 0.04978706836786394297,
256 	[61] = FSCALE * 0.04735892439114092119,
257 	[62] = FSCALE * 0.04504920239355780606,
258 	[63] = FSCALE * 0.04285212686704017991,
259 	[64] = FSCALE * 0.04076220397836621516,
260 	[65] = FSCALE * 0.03877420783172200988,
261 	[66] = FSCALE * 0.03688316740124000544,
262 	[67] = FSCALE * 0.03508435410084502588,
263 	[68] = FSCALE * 0.03337326996032607948,
264 	[69] = FSCALE * 0.03174563637806794323,
265 	[70] = FSCALE * 0.03019738342231850073,
266 	[71] = FSCALE * 0.02872463965423942912,
267 	[72] = FSCALE * 0.02732372244729256080,
268 	[73] = FSCALE * 0.02599112877875534358,
269 	[74] = FSCALE * 0.02472352647033939120,
270 	[75] = FSCALE * 0.02351774585600910823,
271 	[76] = FSCALE * 0.02237077185616559577,
272 	[77] = FSCALE * 0.02127973643837716938,
273 	[78] = FSCALE * 0.02024191144580438847,
274 	[79] = FSCALE * 0.01925470177538692429,
275 	[80] = FSCALE * 0.01831563888873418029,
276 	[81] = FSCALE * 0.01742237463949351138,
277 	[82] = FSCALE * 0.01657267540176124754,
278 	[83] = FSCALE * 0.01576441648485449082,
279 	[84] = FSCALE * 0.01499557682047770621,
280 	[85] = FSCALE * 0.01426423390899925527,
281 	[86] = FSCALE * 0.01356855901220093175,
282 	[87] = FSCALE * 0.01290681258047986886,
283 	[88] = FSCALE * 0.01227733990306844117,
284 	[89] = FSCALE * 0.01167856697039544521,
285 	[90] = FSCALE * 0.01110899653824230649,
286 	[91] = FSCALE * 0.01056720438385265337,
287 	[92] = FSCALE * 0.01005183574463358164,
288 	[93] = FSCALE * 0.00956160193054350793,
289 	[94] = FSCALE * 0.00909527710169581709,
290 	[95] = FSCALE * 0.00865169520312063417,
291 	[96] = FSCALE * 0.00822974704902002884,
292 	[97] = FSCALE * 0.00782837754922577143,
293 	[98] = FSCALE * 0.00744658307092434051,
294 	[99] = FSCALE * 0.00708340892905212004,
295 	[100] = FSCALE * 0.00673794699908546709,
296 	[101] = FSCALE * 0.00640933344625638184,
297 	[102] = FSCALE * 0.00609674656551563610,
298 	[103] = FSCALE * 0.00579940472684214321,
299 	[104] = FSCALE * 0.00551656442076077241,
300 	[105] = FSCALE * 0.00524751839918138427,
301 	[106] = FSCALE * 0.00499159390691021621,
302 	[107] = FSCALE * 0.00474815099941147558,
303 	[108] = FSCALE * 0.00451658094261266798,
304 	[109] = FSCALE * 0.00429630469075234057,
305 	[110] = FSCALE * 0.00408677143846406699,
306 };
307 #endif
308 
309 #define	CCPU_EXP_MAX	110
310 
311 /*
312  * This function is analogical to the getpcpu() function in the ps(1) command.
313  * They should both calculate in the same way so that the racct %cpu
314  * calculations are consistent with the values showed by the ps(1) tool.
315  * The calculations are more complex in the 4BSD scheduler because of the value
316  * of the ccpu variable.  In ULE it is defined to be zero which saves us some
317  * work.
318  */
319 static uint64_t
320 racct_getpcpu(struct proc *p, u_int pcpu)
321 {
322 	u_int swtime;
323 #ifdef SCHED_4BSD
324 	fixpt_t pctcpu, pctcpu_next;
325 #endif
326 #ifdef SMP
327 	struct pcpu *pc;
328 	int found;
329 #endif
330 	fixpt_t p_pctcpu;
331 	struct thread *td;
332 
333 	ASSERT_RACCT_ENABLED();
334 
335 	/*
336 	 * If the process is swapped out, we count its %cpu usage as zero.
337 	 * This behaviour is consistent with the userland ps(1) tool.
338 	 */
339 	if ((p->p_flag & P_INMEM) == 0)
340 		return (0);
341 	swtime = (ticks - p->p_swtick) / hz;
342 
343 	/*
344 	 * For short-lived processes, the sched_pctcpu() returns small
345 	 * values even for cpu intensive processes.  Therefore we use
346 	 * our own estimate in this case.
347 	 */
348 	if (swtime < RACCT_PCPU_SECS)
349 		return (pcpu);
350 
351 	p_pctcpu = 0;
352 	FOREACH_THREAD_IN_PROC(p, td) {
353 		if (td == PCPU_GET(idlethread))
354 			continue;
355 #ifdef SMP
356 		found = 0;
357 		STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
358 			if (td == pc->pc_idlethread) {
359 				found = 1;
360 				break;
361 			}
362 		}
363 		if (found)
364 			continue;
365 #endif
366 		thread_lock(td);
367 #ifdef SCHED_4BSD
368 		pctcpu = sched_pctcpu(td);
369 		/* Count also the yet unfinished second. */
370 		pctcpu_next = (pctcpu * ccpu_exp[1]) >> FSHIFT;
371 		pctcpu_next += sched_pctcpu_delta(td);
372 		p_pctcpu += max(pctcpu, pctcpu_next);
373 #else
374 		/*
375 		 * In ULE the %cpu statistics are updated on every
376 		 * sched_pctcpu() call.  So special calculations to
377 		 * account for the latest (unfinished) second are
378 		 * not needed.
379 		 */
380 		p_pctcpu += sched_pctcpu(td);
381 #endif
382 		thread_unlock(td);
383 	}
384 
385 #ifdef SCHED_4BSD
386 	if (swtime <= CCPU_EXP_MAX)
387 		return ((100 * (uint64_t)p_pctcpu * 1000000) /
388 		    (FSCALE - ccpu_exp[swtime]));
389 #endif
390 
391 	return ((100 * (uint64_t)p_pctcpu * 1000000) / FSCALE);
392 }
393 
394 static void
395 racct_add_racct(struct racct *dest, const struct racct *src)
396 {
397 	int i;
398 
399 	ASSERT_RACCT_ENABLED();
400 	RACCT_LOCK_ASSERT();
401 
402 	/*
403 	 * Update resource usage in dest.
404 	 */
405 	for (i = 0; i <= RACCT_MAX; i++) {
406 		KASSERT(dest->r_resources[i] >= 0,
407 		    ("%s: resource %d propagation meltdown: dest < 0",
408 		    __func__, i));
409 		KASSERT(src->r_resources[i] >= 0,
410 		    ("%s: resource %d propagation meltdown: src < 0",
411 		    __func__, i));
412 		dest->r_resources[i] += src->r_resources[i];
413 	}
414 }
415 
416 static void
417 racct_sub_racct(struct racct *dest, const struct racct *src)
418 {
419 	int i;
420 
421 	ASSERT_RACCT_ENABLED();
422 	RACCT_LOCK_ASSERT();
423 
424 	/*
425 	 * Update resource usage in dest.
426 	 */
427 	for (i = 0; i <= RACCT_MAX; i++) {
428 		if (!RACCT_IS_SLOPPY(i) && !RACCT_IS_DECAYING(i)) {
429 			KASSERT(dest->r_resources[i] >= 0,
430 			    ("%s: resource %d propagation meltdown: dest < 0",
431 			    __func__, i));
432 			KASSERT(src->r_resources[i] >= 0,
433 			    ("%s: resource %d propagation meltdown: src < 0",
434 			    __func__, i));
435 			KASSERT(src->r_resources[i] <= dest->r_resources[i],
436 			    ("%s: resource %d propagation meltdown: src > dest",
437 			    __func__, i));
438 		}
439 		if (RACCT_CAN_DROP(i)) {
440 			dest->r_resources[i] -= src->r_resources[i];
441 			if (dest->r_resources[i] < 0)
442 				dest->r_resources[i] = 0;
443 		}
444 	}
445 }
446 
447 void
448 racct_create(struct racct **racctp)
449 {
450 
451 	if (!racct_enable)
452 		return;
453 
454 	SDT_PROBE1(racct, , racct, create, racctp);
455 
456 	KASSERT(*racctp == NULL, ("racct already allocated"));
457 
458 	*racctp = uma_zalloc(racct_zone, M_WAITOK | M_ZERO);
459 }
460 
461 static void
462 racct_destroy_locked(struct racct **racctp)
463 {
464 	struct racct *racct;
465 	int i;
466 
467 	ASSERT_RACCT_ENABLED();
468 
469 	SDT_PROBE1(racct, , racct, destroy, racctp);
470 
471 	RACCT_LOCK_ASSERT();
472 	KASSERT(racctp != NULL, ("NULL racctp"));
473 	KASSERT(*racctp != NULL, ("NULL racct"));
474 
475 	racct = *racctp;
476 
477 	for (i = 0; i <= RACCT_MAX; i++) {
478 		if (RACCT_IS_SLOPPY(i))
479 			continue;
480 		if (!RACCT_IS_RECLAIMABLE(i))
481 			continue;
482 		KASSERT(racct->r_resources[i] == 0,
483 		    ("destroying non-empty racct: "
484 		    "%ju allocated for resource %d\n",
485 		    racct->r_resources[i], i));
486 	}
487 	uma_zfree(racct_zone, racct);
488 	*racctp = NULL;
489 }
490 
491 void
492 racct_destroy(struct racct **racct)
493 {
494 
495 	if (!racct_enable)
496 		return;
497 
498 	RACCT_LOCK();
499 	racct_destroy_locked(racct);
500 	RACCT_UNLOCK();
501 }
502 
503 /*
504  * Increase consumption of 'resource' by 'amount' for 'racct',
505  * but not its parents.  Differently from other cases, 'amount' here
506  * may be less than zero.
507  */
508 static void
509 racct_adjust_resource(struct racct *racct, int resource,
510     int64_t amount)
511 {
512 
513 	ASSERT_RACCT_ENABLED();
514 	RACCT_LOCK_ASSERT();
515 	KASSERT(racct != NULL, ("NULL racct"));
516 
517 	racct->r_resources[resource] += amount;
518 	if (racct->r_resources[resource] < 0) {
519 		KASSERT(RACCT_IS_SLOPPY(resource) || RACCT_IS_DECAYING(resource),
520 		    ("%s: resource %d usage < 0", __func__, resource));
521 		racct->r_resources[resource] = 0;
522 	}
523 
524 	/*
525 	 * There are some cases where the racct %cpu resource would grow
526 	 * beyond 100% per core.  For example in racct_proc_exit() we add
527 	 * the process %cpu usage to the ucred racct containers.  If too
528 	 * many processes terminated in a short time span, the ucred %cpu
529 	 * resource could grow too much.  Also, the 4BSD scheduler sometimes
530 	 * returns for a thread more than 100% cpu usage. So we set a sane
531 	 * boundary here to 100% * the maximum number of CPUs.
532 	 */
533 	if ((resource == RACCT_PCTCPU) &&
534 	    (racct->r_resources[RACCT_PCTCPU] > 100 * 1000000 * (int64_t)MAXCPU))
535 		racct->r_resources[RACCT_PCTCPU] = 100 * 1000000 * (int64_t)MAXCPU;
536 }
537 
538 static int
539 racct_add_locked(struct proc *p, int resource, uint64_t amount, int force)
540 {
541 #ifdef RCTL
542 	int error;
543 #endif
544 
545 	ASSERT_RACCT_ENABLED();
546 
547 	/*
548 	 * We need proc lock to dereference p->p_ucred.
549 	 */
550 	PROC_LOCK_ASSERT(p, MA_OWNED);
551 
552 #ifdef RCTL
553 	error = rctl_enforce(p, resource, amount);
554 	if (error && !force && RACCT_IS_DENIABLE(resource)) {
555 		SDT_PROBE3(racct, , rusage, add__failure, p, resource, amount);
556 		return (error);
557 	}
558 #endif
559 	racct_adjust_resource(p->p_racct, resource, amount);
560 	racct_add_cred_locked(p->p_ucred, resource, amount);
561 
562 	return (0);
563 }
564 
565 /*
566  * Increase allocation of 'resource' by 'amount' for process 'p'.
567  * Return 0 if it's below limits, or errno, if it's not.
568  */
569 int
570 racct_add(struct proc *p, int resource, uint64_t amount)
571 {
572 	int error;
573 
574 	if (!racct_enable)
575 		return (0);
576 
577 	SDT_PROBE3(racct, , rusage, add, p, resource, amount);
578 
579 	RACCT_LOCK();
580 	error = racct_add_locked(p, resource, amount, 0);
581 	RACCT_UNLOCK();
582 	return (error);
583 }
584 
585 /*
586  * Increase allocation of 'resource' by 'amount' for process 'p'.
587  * Doesn't check for limits and never fails.
588  */
589 void
590 racct_add_force(struct proc *p, int resource, uint64_t amount)
591 {
592 
593 	if (!racct_enable)
594 		return;
595 
596 	SDT_PROBE3(racct, , rusage, add__force, p, resource, amount);
597 
598 	RACCT_LOCK();
599 	racct_add_locked(p, resource, amount, 1);
600 	RACCT_UNLOCK();
601 }
602 
603 static void
604 racct_add_cred_locked(struct ucred *cred, int resource, uint64_t amount)
605 {
606 	struct prison *pr;
607 
608 	ASSERT_RACCT_ENABLED();
609 
610 	racct_adjust_resource(cred->cr_ruidinfo->ui_racct, resource, amount);
611 	for (pr = cred->cr_prison; pr != NULL; pr = pr->pr_parent)
612 		racct_adjust_resource(pr->pr_prison_racct->prr_racct, resource,
613 		    amount);
614 	racct_adjust_resource(cred->cr_loginclass->lc_racct, resource, amount);
615 }
616 
617 /*
618  * Increase allocation of 'resource' by 'amount' for credential 'cred'.
619  * Doesn't check for limits and never fails.
620  */
621 void
622 racct_add_cred(struct ucred *cred, int resource, uint64_t amount)
623 {
624 
625 	if (!racct_enable)
626 		return;
627 
628 	SDT_PROBE3(racct, , rusage, add__cred, cred, resource, amount);
629 
630 	RACCT_LOCK();
631 	racct_add_cred_locked(cred, resource, amount);
632 	RACCT_UNLOCK();
633 }
634 
635 /*
636  * Account for disk IO resource consumption.  Checks for limits,
637  * but never fails, due to disk limits being undeniable.
638  */
639 void
640 racct_add_buf(struct proc *p, const struct buf *bp, int is_write)
641 {
642 
643 	ASSERT_RACCT_ENABLED();
644 	PROC_LOCK_ASSERT(p, MA_OWNED);
645 
646 	SDT_PROBE3(racct, , rusage, add__buf, p, bp, is_write);
647 
648 	RACCT_LOCK();
649 	if (is_write) {
650 		racct_add_locked(curproc, RACCT_WRITEBPS, bp->b_bcount, 1);
651 		racct_add_locked(curproc, RACCT_WRITEIOPS, 1, 1);
652 	} else {
653 		racct_add_locked(curproc, RACCT_READBPS, bp->b_bcount, 1);
654 		racct_add_locked(curproc, RACCT_READIOPS, 1, 1);
655 	}
656 	RACCT_UNLOCK();
657 }
658 
659 static int
660 racct_set_locked(struct proc *p, int resource, uint64_t amount, int force)
661 {
662 	int64_t old_amount, decayed_amount, diff_proc, diff_cred;
663 #ifdef RCTL
664 	int error;
665 #endif
666 
667 	ASSERT_RACCT_ENABLED();
668 
669 	/*
670 	 * We need proc lock to dereference p->p_ucred.
671 	 */
672 	PROC_LOCK_ASSERT(p, MA_OWNED);
673 
674 	old_amount = p->p_racct->r_resources[resource];
675 	/*
676 	 * The diffs may be negative.
677 	 */
678 	diff_proc = amount - old_amount;
679 	if (resource == RACCT_PCTCPU) {
680 		/*
681 		 * Resources in per-credential racct containers may decay.
682 		 * If this is the case, we need to calculate the difference
683 		 * between the new amount and the proportional value of the
684 		 * old amount that has decayed in the ucred racct containers.
685 		 */
686 		decayed_amount = old_amount * RACCT_DECAY_FACTOR / FSCALE;
687 		diff_cred = amount - decayed_amount;
688 	} else
689 		diff_cred = diff_proc;
690 #ifdef notyet
691 	KASSERT(diff_proc >= 0 || RACCT_CAN_DROP(resource),
692 	    ("%s: usage of non-droppable resource %d dropping", __func__,
693 	     resource));
694 #endif
695 #ifdef RCTL
696 	if (diff_proc > 0) {
697 		error = rctl_enforce(p, resource, diff_proc);
698 		if (error && !force && RACCT_IS_DENIABLE(resource)) {
699 			SDT_PROBE3(racct, , rusage, set__failure, p, resource,
700 			    amount);
701 			return (error);
702 		}
703 	}
704 #endif
705 	racct_adjust_resource(p->p_racct, resource, diff_proc);
706 	if (diff_cred > 0)
707 		racct_add_cred_locked(p->p_ucred, resource, diff_cred);
708 	else if (diff_cred < 0)
709 		racct_sub_cred_locked(p->p_ucred, resource, -diff_cred);
710 
711 	return (0);
712 }
713 
714 /*
715  * Set allocation of 'resource' to 'amount' for process 'p'.
716  * Return 0 if it's below limits, or errno, if it's not.
717  *
718  * Note that decreasing the allocation always returns 0,
719  * even if it's above the limit.
720  */
721 int
722 racct_set_unlocked(struct proc *p, int resource, uint64_t amount)
723 {
724 	int error;
725 
726 	ASSERT_RACCT_ENABLED();
727 	PROC_LOCK(p);
728 	error = racct_set(p, resource, amount);
729 	PROC_UNLOCK(p);
730 	return (error);
731 }
732 
733 int
734 racct_set(struct proc *p, int resource, uint64_t amount)
735 {
736 	int error;
737 
738 	if (!racct_enable)
739 		return (0);
740 
741 	SDT_PROBE3(racct, , rusage, set__force, p, resource, amount);
742 
743 	RACCT_LOCK();
744 	error = racct_set_locked(p, resource, amount, 0);
745 	RACCT_UNLOCK();
746 	return (error);
747 }
748 
749 void
750 racct_set_force(struct proc *p, int resource, uint64_t amount)
751 {
752 
753 	if (!racct_enable)
754 		return;
755 
756 	SDT_PROBE3(racct, , rusage, set, p, resource, amount);
757 
758 	RACCT_LOCK();
759 	racct_set_locked(p, resource, amount, 1);
760 	RACCT_UNLOCK();
761 }
762 
763 /*
764  * Returns amount of 'resource' the process 'p' can keep allocated.
765  * Allocating more than that would be denied, unless the resource
766  * is marked undeniable.  Amount of already allocated resource does
767  * not matter.
768  */
769 uint64_t
770 racct_get_limit(struct proc *p, int resource)
771 {
772 #ifdef RCTL
773 	uint64_t available;
774 
775 	if (!racct_enable)
776 		return (UINT64_MAX);
777 
778 	RACCT_LOCK();
779 	available = rctl_get_limit(p, resource);
780 	RACCT_UNLOCK();
781 
782 	return (available);
783 #else
784 
785 	return (UINT64_MAX);
786 #endif
787 }
788 
789 /*
790  * Returns amount of 'resource' the process 'p' can keep allocated.
791  * Allocating more than that would be denied, unless the resource
792  * is marked undeniable.  Amount of already allocated resource does
793  * matter.
794  */
795 uint64_t
796 racct_get_available(struct proc *p, int resource)
797 {
798 #ifdef RCTL
799 	uint64_t available;
800 
801 	if (!racct_enable)
802 		return (UINT64_MAX);
803 
804 	RACCT_LOCK();
805 	available = rctl_get_available(p, resource);
806 	RACCT_UNLOCK();
807 
808 	return (available);
809 #else
810 
811 	return (UINT64_MAX);
812 #endif
813 }
814 
815 /*
816  * Returns amount of the %cpu resource that process 'p' can add to its %cpu
817  * utilization.  Adding more than that would lead to the process being
818  * throttled.
819  */
820 static int64_t
821 racct_pcpu_available(struct proc *p)
822 {
823 #ifdef RCTL
824 	uint64_t available;
825 
826 	ASSERT_RACCT_ENABLED();
827 
828 	RACCT_LOCK();
829 	available = rctl_pcpu_available(p);
830 	RACCT_UNLOCK();
831 
832 	return (available);
833 #else
834 
835 	return (INT64_MAX);
836 #endif
837 }
838 
839 /*
840  * Decrease allocation of 'resource' by 'amount' for process 'p'.
841  */
842 void
843 racct_sub(struct proc *p, int resource, uint64_t amount)
844 {
845 
846 	if (!racct_enable)
847 		return;
848 
849 	SDT_PROBE3(racct, , rusage, sub, p, resource, amount);
850 
851 	/*
852 	 * We need proc lock to dereference p->p_ucred.
853 	 */
854 	PROC_LOCK_ASSERT(p, MA_OWNED);
855 	KASSERT(RACCT_CAN_DROP(resource),
856 	    ("%s: called for non-droppable resource %d", __func__, resource));
857 
858 	RACCT_LOCK();
859 	KASSERT(amount <= p->p_racct->r_resources[resource],
860 	    ("%s: freeing %ju of resource %d, which is more "
861 	     "than allocated %jd for %s (pid %d)", __func__, amount, resource,
862 	    (intmax_t)p->p_racct->r_resources[resource], p->p_comm, p->p_pid));
863 
864 	racct_adjust_resource(p->p_racct, resource, -amount);
865 	racct_sub_cred_locked(p->p_ucred, resource, amount);
866 	RACCT_UNLOCK();
867 }
868 
869 static void
870 racct_sub_cred_locked(struct ucred *cred, int resource, uint64_t amount)
871 {
872 	struct prison *pr;
873 
874 	ASSERT_RACCT_ENABLED();
875 
876 	racct_adjust_resource(cred->cr_ruidinfo->ui_racct, resource, -amount);
877 	for (pr = cred->cr_prison; pr != NULL; pr = pr->pr_parent)
878 		racct_adjust_resource(pr->pr_prison_racct->prr_racct, resource,
879 		    -amount);
880 	racct_adjust_resource(cred->cr_loginclass->lc_racct, resource, -amount);
881 }
882 
883 /*
884  * Decrease allocation of 'resource' by 'amount' for credential 'cred'.
885  */
886 void
887 racct_sub_cred(struct ucred *cred, int resource, uint64_t amount)
888 {
889 
890 	if (!racct_enable)
891 		return;
892 
893 	SDT_PROBE3(racct, , rusage, sub__cred, cred, resource, amount);
894 
895 #ifdef notyet
896 	KASSERT(RACCT_CAN_DROP(resource),
897 	    ("%s: called for resource %d which can not drop", __func__,
898 	     resource));
899 #endif
900 
901 	RACCT_LOCK();
902 	racct_sub_cred_locked(cred, resource, amount);
903 	RACCT_UNLOCK();
904 }
905 
906 /*
907  * Inherit resource usage information from the parent process.
908  */
909 int
910 racct_proc_fork(struct proc *parent, struct proc *child)
911 {
912 	int i, error = 0;
913 
914 	if (!racct_enable)
915 		return (0);
916 
917 	/*
918 	 * Create racct for the child process.
919 	 */
920 	racct_create(&child->p_racct);
921 
922 	PROC_LOCK(parent);
923 	PROC_LOCK(child);
924 	RACCT_LOCK();
925 
926 #ifdef RCTL
927 	error = rctl_proc_fork(parent, child);
928 	if (error != 0)
929 		goto out;
930 #endif
931 
932 	/* Init process cpu time. */
933 	child->p_prev_runtime = 0;
934 	child->p_throttled = 0;
935 
936 	/*
937 	 * Inherit resource usage.
938 	 */
939 	for (i = 0; i <= RACCT_MAX; i++) {
940 		if (parent->p_racct->r_resources[i] == 0 ||
941 		    !RACCT_IS_INHERITABLE(i))
942 			continue;
943 
944 		error = racct_set_locked(child, i,
945 		    parent->p_racct->r_resources[i], 0);
946 		if (error != 0)
947 			goto out;
948 	}
949 
950 	error = racct_add_locked(child, RACCT_NPROC, 1, 0);
951 	error += racct_add_locked(child, RACCT_NTHR, 1, 0);
952 
953 out:
954 	RACCT_UNLOCK();
955 	PROC_UNLOCK(child);
956 	PROC_UNLOCK(parent);
957 
958 	if (error != 0)
959 		racct_proc_exit(child);
960 
961 	return (error);
962 }
963 
964 /*
965  * Called at the end of fork1(), to handle rules that require the process
966  * to be fully initialized.
967  */
968 void
969 racct_proc_fork_done(struct proc *child)
970 {
971 
972 	if (!racct_enable)
973 		return;
974 
975 #ifdef RCTL
976 	PROC_LOCK(child);
977 	RACCT_LOCK();
978 	rctl_enforce(child, RACCT_NPROC, 0);
979 	rctl_enforce(child, RACCT_NTHR, 0);
980 	RACCT_UNLOCK();
981 	PROC_UNLOCK(child);
982 #endif
983 }
984 
985 void
986 racct_proc_exit(struct proc *p)
987 {
988 	struct timeval wallclock;
989 	uint64_t pct_estimate, pct, runtime;
990 	int i;
991 
992 	if (!racct_enable)
993 		return;
994 
995 	PROC_LOCK(p);
996 	/*
997 	 * We don't need to calculate rux, proc_reap() has already done this.
998 	 */
999 	runtime = cputick2usec(p->p_rux.rux_runtime);
1000 #ifdef notyet
1001 	KASSERT(runtime >= p->p_prev_runtime, ("runtime < p_prev_runtime"));
1002 #else
1003 	if (runtime < p->p_prev_runtime)
1004 		runtime = p->p_prev_runtime;
1005 #endif
1006 	microuptime(&wallclock);
1007 	timevalsub(&wallclock, &p->p_stats->p_start);
1008 	if (wallclock.tv_sec > 0 || wallclock.tv_usec > 0) {
1009 		pct_estimate = (1000000 * runtime * 100) /
1010 		    ((uint64_t)wallclock.tv_sec * 1000000 +
1011 		    wallclock.tv_usec);
1012 	} else
1013 		pct_estimate = 0;
1014 	pct = racct_getpcpu(p, pct_estimate);
1015 
1016 	RACCT_LOCK();
1017 	racct_set_locked(p, RACCT_CPU, runtime, 0);
1018 	racct_add_cred_locked(p->p_ucred, RACCT_PCTCPU, pct);
1019 
1020 	KASSERT(p->p_racct->r_resources[RACCT_RSS] == 0,
1021 	    ("process reaped with %ju allocated for RSS\n",
1022 	    p->p_racct->r_resources[RACCT_RSS]));
1023 	for (i = 0; i <= RACCT_MAX; i++) {
1024 		if (p->p_racct->r_resources[i] == 0)
1025 			continue;
1026 		if (!RACCT_IS_RECLAIMABLE(i))
1027 			continue;
1028 		racct_set_locked(p, i, 0, 0);
1029 	}
1030 
1031 #ifdef RCTL
1032 	rctl_racct_release(p->p_racct);
1033 #endif
1034 	racct_destroy_locked(&p->p_racct);
1035 	RACCT_UNLOCK();
1036 	PROC_UNLOCK(p);
1037 }
1038 
1039 /*
1040  * Called after credentials change, to move resource utilisation
1041  * between raccts.
1042  */
1043 void
1044 racct_proc_ucred_changed(struct proc *p, struct ucred *oldcred,
1045     struct ucred *newcred)
1046 {
1047 	struct uidinfo *olduip, *newuip;
1048 	struct loginclass *oldlc, *newlc;
1049 	struct prison *oldpr, *newpr, *pr;
1050 
1051 	if (!racct_enable)
1052 		return;
1053 
1054 	PROC_LOCK_ASSERT(p, MA_OWNED);
1055 
1056 	newuip = newcred->cr_ruidinfo;
1057 	olduip = oldcred->cr_ruidinfo;
1058 	newlc = newcred->cr_loginclass;
1059 	oldlc = oldcred->cr_loginclass;
1060 	newpr = newcred->cr_prison;
1061 	oldpr = oldcred->cr_prison;
1062 
1063 	RACCT_LOCK();
1064 	if (newuip != olduip) {
1065 		racct_sub_racct(olduip->ui_racct, p->p_racct);
1066 		racct_add_racct(newuip->ui_racct, p->p_racct);
1067 	}
1068 	if (newlc != oldlc) {
1069 		racct_sub_racct(oldlc->lc_racct, p->p_racct);
1070 		racct_add_racct(newlc->lc_racct, p->p_racct);
1071 	}
1072 	if (newpr != oldpr) {
1073 		for (pr = oldpr; pr != NULL; pr = pr->pr_parent)
1074 			racct_sub_racct(pr->pr_prison_racct->prr_racct,
1075 			    p->p_racct);
1076 		for (pr = newpr; pr != NULL; pr = pr->pr_parent)
1077 			racct_add_racct(pr->pr_prison_racct->prr_racct,
1078 			    p->p_racct);
1079 	}
1080 	RACCT_UNLOCK();
1081 }
1082 
1083 void
1084 racct_move(struct racct *dest, struct racct *src)
1085 {
1086 
1087 	ASSERT_RACCT_ENABLED();
1088 
1089 	RACCT_LOCK();
1090 	racct_add_racct(dest, src);
1091 	racct_sub_racct(src, src);
1092 	RACCT_UNLOCK();
1093 }
1094 
1095 static void
1096 ast_racct(struct thread *td, int tda __unused)
1097 {
1098 	struct proc *p;
1099 
1100 	ASSERT_RACCT_ENABLED();
1101 
1102 	p = td->td_proc;
1103 	if (p->p_throttled == 0)
1104 		return;
1105 
1106 	PROC_LOCK(p);
1107 	while (p->p_throttled != 0) {
1108 		msleep(p->p_racct, &p->p_mtx, 0, "racct",
1109 		    p->p_throttled < 0 ? 0 : p->p_throttled);
1110 		if (p->p_throttled > 0)
1111 			p->p_throttled = 0;
1112 	}
1113 	PROC_UNLOCK(p);
1114 }
1115 
1116 /*
1117  * Make the process sleep in userret() for 'timeout' ticks.  Setting
1118  * timeout to -1 makes it sleep until woken up by racct_proc_wakeup().
1119  */
1120 void
1121 racct_proc_throttle(struct proc *p, int timeout)
1122 {
1123 	struct thread *td;
1124 #ifdef SMP
1125 	int cpuid;
1126 #endif
1127 
1128 	KASSERT(timeout != 0, ("timeout %d", timeout));
1129 	ASSERT_RACCT_ENABLED();
1130 	PROC_LOCK_ASSERT(p, MA_OWNED);
1131 
1132 	/*
1133 	 * Do not block kernel processes.  Also do not block processes with
1134 	 * low %cpu utilization to improve interactivity.
1135 	 */
1136 	if ((p->p_flag & (P_SYSTEM | P_KPROC)) != 0)
1137 		return;
1138 
1139 	if (p->p_throttled < 0 || (timeout > 0 && p->p_throttled > timeout))
1140 		return;
1141 
1142 	p->p_throttled = timeout;
1143 
1144 	FOREACH_THREAD_IN_PROC(p, td) {
1145 		thread_lock(td);
1146 		ast_sched_locked(td, TDA_RACCT);
1147 
1148 		switch (TD_GET_STATE(td)) {
1149 		case TDS_RUNQ:
1150 			/*
1151 			 * If the thread is on the scheduler run-queue, we can
1152 			 * not just remove it from there.  So we set the flag
1153 			 * TDA_SCHED for the thread, so that once it is
1154 			 * running, it is taken off the cpu as soon as possible.
1155 			 */
1156 			ast_sched_locked(td, TDA_SCHED);
1157 			break;
1158 		case TDS_RUNNING:
1159 			/*
1160 			 * If the thread is running, we request a context
1161 			 * switch for it by setting the TDA_SCHED flag.
1162 			 */
1163 			ast_sched_locked(td, TDA_SCHED);
1164 #ifdef SMP
1165 			cpuid = td->td_oncpu;
1166 			if ((cpuid != NOCPU) && (td != curthread))
1167 				ipi_cpu(cpuid, IPI_AST);
1168 #endif
1169 			break;
1170 		default:
1171 			break;
1172 		}
1173 		thread_unlock(td);
1174 	}
1175 }
1176 
1177 static void
1178 racct_proc_wakeup(struct proc *p)
1179 {
1180 
1181 	ASSERT_RACCT_ENABLED();
1182 
1183 	PROC_LOCK_ASSERT(p, MA_OWNED);
1184 
1185 	if (p->p_throttled != 0) {
1186 		p->p_throttled = 0;
1187 		wakeup(p->p_racct);
1188 	}
1189 }
1190 
1191 static void
1192 racct_decay_callback(struct racct *racct, void *dummy1, void *dummy2)
1193 {
1194 	int64_t r_old, r_new;
1195 
1196 	ASSERT_RACCT_ENABLED();
1197 	RACCT_LOCK_ASSERT();
1198 
1199 #ifdef RCTL
1200 	rctl_throttle_decay(racct, RACCT_READBPS);
1201 	rctl_throttle_decay(racct, RACCT_WRITEBPS);
1202 	rctl_throttle_decay(racct, RACCT_READIOPS);
1203 	rctl_throttle_decay(racct, RACCT_WRITEIOPS);
1204 #endif
1205 
1206 	r_old = racct->r_resources[RACCT_PCTCPU];
1207 
1208 	/* If there is nothing to decay, just exit. */
1209 	if (r_old <= 0)
1210 		return;
1211 
1212 	r_new = r_old * RACCT_DECAY_FACTOR / FSCALE;
1213 	racct->r_resources[RACCT_PCTCPU] = r_new;
1214 }
1215 
1216 static void
1217 racct_decay_pre(void)
1218 {
1219 
1220 	RACCT_LOCK();
1221 }
1222 
1223 static void
1224 racct_decay_post(void)
1225 {
1226 
1227 	RACCT_UNLOCK();
1228 }
1229 
1230 static void
1231 racct_decay(void)
1232 {
1233 
1234 	ASSERT_RACCT_ENABLED();
1235 
1236 	ui_racct_foreach(racct_decay_callback, racct_decay_pre,
1237 	    racct_decay_post, NULL, NULL);
1238 	loginclass_racct_foreach(racct_decay_callback, racct_decay_pre,
1239 	    racct_decay_post, NULL, NULL);
1240 	prison_racct_foreach(racct_decay_callback, racct_decay_pre,
1241 	    racct_decay_post, NULL, NULL);
1242 }
1243 
1244 static void
1245 racctd(void)
1246 {
1247 	struct thread *td;
1248 	struct proc *p;
1249 	struct timeval wallclock;
1250 	uint64_t pct, pct_estimate, runtime;
1251 
1252 	ASSERT_RACCT_ENABLED();
1253 
1254 	for (;;) {
1255 		racct_decay();
1256 
1257 		sx_slock(&allproc_lock);
1258 
1259 		FOREACH_PROC_IN_SYSTEM(p) {
1260 			PROC_LOCK(p);
1261 			if (p->p_state != PRS_NORMAL) {
1262 				if (p->p_state == PRS_ZOMBIE)
1263 					racct_set(p, RACCT_PCTCPU, 0);
1264 				PROC_UNLOCK(p);
1265 				continue;
1266 			}
1267 
1268 			microuptime(&wallclock);
1269 			timevalsub(&wallclock, &p->p_stats->p_start);
1270 			PROC_STATLOCK(p);
1271 			FOREACH_THREAD_IN_PROC(p, td)
1272 				ruxagg(p, td);
1273 			runtime = cputick2usec(p->p_rux.rux_runtime);
1274 			PROC_STATUNLOCK(p);
1275 #ifdef notyet
1276 			KASSERT(runtime >= p->p_prev_runtime,
1277 			    ("runtime < p_prev_runtime"));
1278 #else
1279 			if (runtime < p->p_prev_runtime)
1280 				runtime = p->p_prev_runtime;
1281 #endif
1282 			p->p_prev_runtime = runtime;
1283 			if (wallclock.tv_sec > 0 || wallclock.tv_usec > 0) {
1284 				pct_estimate = (1000000 * runtime * 100) /
1285 				    ((uint64_t)wallclock.tv_sec * 1000000 +
1286 				    wallclock.tv_usec);
1287 			} else
1288 				pct_estimate = 0;
1289 			pct = racct_getpcpu(p, pct_estimate);
1290 			RACCT_LOCK();
1291 #ifdef RCTL
1292 			rctl_throttle_decay(p->p_racct, RACCT_READBPS);
1293 			rctl_throttle_decay(p->p_racct, RACCT_WRITEBPS);
1294 			rctl_throttle_decay(p->p_racct, RACCT_READIOPS);
1295 			rctl_throttle_decay(p->p_racct, RACCT_WRITEIOPS);
1296 #endif
1297 			racct_set_locked(p, RACCT_PCTCPU, pct, 1);
1298 			racct_set_locked(p, RACCT_CPU, runtime, 0);
1299 			racct_set_locked(p, RACCT_WALLCLOCK,
1300 			    (uint64_t)wallclock.tv_sec * 1000000 +
1301 			    wallclock.tv_usec, 0);
1302 			RACCT_UNLOCK();
1303 			PROC_UNLOCK(p);
1304 		}
1305 
1306 		/*
1307 		 * To ensure that processes are throttled in a fair way, we need
1308 		 * to iterate over all processes again and check the limits
1309 		 * for %cpu resource only after ucred racct containers have been
1310 		 * properly filled.
1311 		 */
1312 		FOREACH_PROC_IN_SYSTEM(p) {
1313 			PROC_LOCK(p);
1314 			if (p->p_state != PRS_NORMAL) {
1315 				PROC_UNLOCK(p);
1316 				continue;
1317 			}
1318 
1319 			if (racct_pcpu_available(p) <= 0) {
1320 				if (p->p_racct->r_resources[RACCT_PCTCPU] >
1321 				    pcpu_threshold)
1322 					racct_proc_throttle(p, -1);
1323 			} else if (p->p_throttled == -1) {
1324 				racct_proc_wakeup(p);
1325 			}
1326 			PROC_UNLOCK(p);
1327 		}
1328 		sx_sunlock(&allproc_lock);
1329 		pause("-", hz);
1330 	}
1331 }
1332 
1333 static struct kproc_desc racctd_kp = {
1334 	"racctd",
1335 	racctd,
1336 	NULL
1337 };
1338 
1339 static void
1340 racctd_init(void)
1341 {
1342 	if (!racct_enable)
1343 		return;
1344 
1345 	kproc_start(&racctd_kp);
1346 }
1347 SYSINIT(racctd, SI_SUB_RACCTD, SI_ORDER_FIRST, racctd_init, NULL);
1348 
1349 static void
1350 racct_init(void)
1351 {
1352 	if (!racct_enable)
1353 		return;
1354 
1355 	racct_zone = uma_zcreate("racct", sizeof(struct racct),
1356 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
1357 	ast_register(TDA_RACCT, ASTR_ASTF_REQUIRED, 0, ast_racct);
1358 
1359 	/*
1360 	 * XXX: Move this somewhere.
1361 	 */
1362 	prison0.pr_prison_racct = prison_racct_find("0");
1363 }
1364 SYSINIT(racct, SI_SUB_RACCT, SI_ORDER_FIRST, racct_init, NULL);
1365