xref: /freebsd/sys/kern/kern_racct.c (revision 5f4c09dd85bff675e0ca63c55ea3c517e0fddfcc)
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 #ifdef RACCT
63 
64 FEATURE(racct, "Resource Accounting");
65 
66 /*
67  * Do not block processes that have their %cpu usage <= pcpu_threshold.
68  */
69 static int pcpu_threshold = 1;
70 #ifdef RACCT_DEFAULT_TO_DISABLED
71 bool __read_frequently racct_enable = false;
72 #else
73 bool __read_frequently racct_enable = true;
74 #endif
75 
76 SYSCTL_NODE(_kern, OID_AUTO, racct, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
77     "Resource Accounting");
78 SYSCTL_BOOL(_kern_racct, OID_AUTO, enable, CTLFLAG_RDTUN, &racct_enable,
79     0, "Enable RACCT/RCTL");
80 SYSCTL_UINT(_kern_racct, OID_AUTO, pcpu_threshold, CTLFLAG_RW, &pcpu_threshold,
81     0, "Processes with higher %cpu usage than this value can be throttled.");
82 
83 /*
84  * How many seconds it takes to use the scheduler %cpu calculations.  When a
85  * process starts, we compute its %cpu usage by dividing its runtime by the
86  * process wall clock time.  After RACCT_PCPU_SECS pass, we use the value
87  * provided by the scheduler.
88  */
89 #define RACCT_PCPU_SECS		3
90 
91 struct mtx racct_lock;
92 MTX_SYSINIT(racct_lock, &racct_lock, "racct lock", MTX_DEF);
93 
94 static uma_zone_t racct_zone;
95 
96 static void racct_sub_racct(struct racct *dest, const struct racct *src);
97 static void racct_sub_cred_locked(struct ucred *cred, int resource,
98 		uint64_t amount);
99 static void racct_add_cred_locked(struct ucred *cred, int resource,
100 		uint64_t amount);
101 
102 SDT_PROVIDER_DEFINE(racct);
103 SDT_PROBE_DEFINE3(racct, , rusage, add,
104     "struct proc *", "int", "uint64_t");
105 SDT_PROBE_DEFINE3(racct, , rusage, add__failure,
106     "struct proc *", "int", "uint64_t");
107 SDT_PROBE_DEFINE3(racct, , rusage, add__buf,
108     "struct proc *", "const struct buf *", "int");
109 SDT_PROBE_DEFINE3(racct, , rusage, add__cred,
110     "struct ucred *", "int", "uint64_t");
111 SDT_PROBE_DEFINE3(racct, , rusage, add__force,
112     "struct proc *", "int", "uint64_t");
113 SDT_PROBE_DEFINE3(racct, , rusage, set,
114     "struct proc *", "int", "uint64_t");
115 SDT_PROBE_DEFINE3(racct, , rusage, set__failure,
116     "struct proc *", "int", "uint64_t");
117 SDT_PROBE_DEFINE3(racct, , rusage, set__force,
118     "struct proc *", "int", "uint64_t");
119 SDT_PROBE_DEFINE3(racct, , rusage, sub,
120     "struct proc *", "int", "uint64_t");
121 SDT_PROBE_DEFINE3(racct, , rusage, sub__cred,
122     "struct ucred *", "int", "uint64_t");
123 SDT_PROBE_DEFINE1(racct, , racct, create,
124     "struct racct *");
125 SDT_PROBE_DEFINE1(racct, , racct, destroy,
126     "struct racct *");
127 SDT_PROBE_DEFINE2(racct, , racct, join,
128     "struct racct *", "struct racct *");
129 SDT_PROBE_DEFINE2(racct, , racct, join__failure,
130     "struct racct *", "struct racct *");
131 SDT_PROBE_DEFINE2(racct, , racct, leave,
132     "struct racct *", "struct racct *");
133 
134 int racct_types[] = {
135 	[RACCT_CPU] =
136 		RACCT_IN_MILLIONS,
137 	[RACCT_DATA] =
138 		RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
139 	[RACCT_STACK] =
140 		RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
141 	[RACCT_CORE] =
142 		RACCT_DENIABLE,
143 	[RACCT_RSS] =
144 		RACCT_RECLAIMABLE,
145 	[RACCT_MEMLOCK] =
146 		RACCT_RECLAIMABLE | RACCT_DENIABLE,
147 	[RACCT_NPROC] =
148 		RACCT_RECLAIMABLE | RACCT_DENIABLE,
149 	[RACCT_NOFILE] =
150 		RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
151 	[RACCT_VMEM] =
152 		RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
153 	[RACCT_NPTS] =
154 		RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
155 	[RACCT_SWAP] =
156 		RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
157 	[RACCT_NTHR] =
158 		RACCT_RECLAIMABLE | RACCT_DENIABLE,
159 	[RACCT_MSGQQUEUED] =
160 		RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
161 	[RACCT_MSGQSIZE] =
162 		RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
163 	[RACCT_NMSGQ] =
164 		RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
165 	[RACCT_NSEM] =
166 		RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
167 	[RACCT_NSEMOP] =
168 		RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
169 	[RACCT_NSHM] =
170 		RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
171 	[RACCT_SHMSIZE] =
172 		RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
173 	[RACCT_WALLCLOCK] =
174 		RACCT_IN_MILLIONS,
175 	[RACCT_PCTCPU] =
176 		RACCT_DECAYING | RACCT_DENIABLE | RACCT_IN_MILLIONS,
177 	[RACCT_READBPS] =
178 		RACCT_DECAYING,
179 	[RACCT_WRITEBPS] =
180 		RACCT_DECAYING,
181 	[RACCT_READIOPS] =
182 		RACCT_DECAYING,
183 	[RACCT_WRITEIOPS] =
184 		RACCT_DECAYING };
185 
186 static const fixpt_t RACCT_DECAY_FACTOR = 0.3 * FSCALE;
187 
188 #ifdef SCHED_4BSD
189 /*
190  * Contains intermediate values for %cpu calculations to avoid using floating
191  * point in the kernel.
192  * ccpu_exp[k] = FSCALE * (ccpu/FSCALE)^k = FSCALE * exp(-k/20)
193  * It is needed only for the 4BSD scheduler, because in ULE, the ccpu equals to
194  * zero so the calculations are more straightforward.
195  */
196 fixpt_t ccpu_exp[] = {
197 	[0] = FSCALE * 1,
198 	[1] = FSCALE * 0.95122942450071400909,
199 	[2] = FSCALE * 0.90483741803595957316,
200 	[3] = FSCALE * 0.86070797642505780722,
201 	[4] = FSCALE * 0.81873075307798185866,
202 	[5] = FSCALE * 0.77880078307140486824,
203 	[6] = FSCALE * 0.74081822068171786606,
204 	[7] = FSCALE * 0.70468808971871343435,
205 	[8] = FSCALE * 0.67032004603563930074,
206 	[9] = FSCALE * 0.63762815162177329314,
207 	[10] = FSCALE * 0.60653065971263342360,
208 	[11] = FSCALE * 0.57694981038048669531,
209 	[12] = FSCALE * 0.54881163609402643262,
210 	[13] = FSCALE * 0.52204577676101604789,
211 	[14] = FSCALE * 0.49658530379140951470,
212 	[15] = FSCALE * 0.47236655274101470713,
213 	[16] = FSCALE * 0.44932896411722159143,
214 	[17] = FSCALE * 0.42741493194872666992,
215 	[18] = FSCALE * 0.40656965974059911188,
216 	[19] = FSCALE * 0.38674102345450120691,
217 	[20] = FSCALE * 0.36787944117144232159,
218 	[21] = FSCALE * 0.34993774911115535467,
219 	[22] = FSCALE * 0.33287108369807955328,
220 	[23] = FSCALE * 0.31663676937905321821,
221 	[24] = FSCALE * 0.30119421191220209664,
222 	[25] = FSCALE * 0.28650479686019010032,
223 	[26] = FSCALE * 0.27253179303401260312,
224 	[27] = FSCALE * 0.25924026064589150757,
225 	[28] = FSCALE * 0.24659696394160647693,
226 	[29] = FSCALE * 0.23457028809379765313,
227 	[30] = FSCALE * 0.22313016014842982893,
228 	[31] = FSCALE * 0.21224797382674305771,
229 	[32] = FSCALE * 0.20189651799465540848,
230 	[33] = FSCALE * 0.19204990862075411423,
231 	[34] = FSCALE * 0.18268352405273465022,
232 	[35] = FSCALE * 0.17377394345044512668,
233 	[36] = FSCALE * 0.16529888822158653829,
234 	[37] = FSCALE * 0.15723716631362761621,
235 	[38] = FSCALE * 0.14956861922263505264,
236 	[39] = FSCALE * 0.14227407158651357185,
237 	[40] = FSCALE * 0.13533528323661269189,
238 	[41] = FSCALE * 0.12873490358780421886,
239 	[42] = FSCALE * 0.12245642825298191021,
240 	[43] = FSCALE * 0.11648415777349695786,
241 	[44] = FSCALE * 0.11080315836233388333,
242 	[45] = FSCALE * 0.10539922456186433678,
243 	[46] = FSCALE * 0.10025884372280373372,
244 	[47] = FSCALE * 0.09536916221554961888,
245 	[48] = FSCALE * 0.09071795328941250337,
246 	[49] = FSCALE * 0.08629358649937051097,
247 	[50] = FSCALE * 0.08208499862389879516,
248 	[51] = FSCALE * 0.07808166600115315231,
249 	[52] = FSCALE * 0.07427357821433388042,
250 	[53] = FSCALE * 0.07065121306042958674,
251 	[54] = FSCALE * 0.06720551273974976512,
252 	[55] = FSCALE * 0.06392786120670757270,
253 	[56] = FSCALE * 0.06081006262521796499,
254 	[57] = FSCALE * 0.05784432087483846296,
255 	[58] = FSCALE * 0.05502322005640722902,
256 	[59] = FSCALE * 0.05233970594843239308,
257 	[60] = FSCALE * 0.04978706836786394297,
258 	[61] = FSCALE * 0.04735892439114092119,
259 	[62] = FSCALE * 0.04504920239355780606,
260 	[63] = FSCALE * 0.04285212686704017991,
261 	[64] = FSCALE * 0.04076220397836621516,
262 	[65] = FSCALE * 0.03877420783172200988,
263 	[66] = FSCALE * 0.03688316740124000544,
264 	[67] = FSCALE * 0.03508435410084502588,
265 	[68] = FSCALE * 0.03337326996032607948,
266 	[69] = FSCALE * 0.03174563637806794323,
267 	[70] = FSCALE * 0.03019738342231850073,
268 	[71] = FSCALE * 0.02872463965423942912,
269 	[72] = FSCALE * 0.02732372244729256080,
270 	[73] = FSCALE * 0.02599112877875534358,
271 	[74] = FSCALE * 0.02472352647033939120,
272 	[75] = FSCALE * 0.02351774585600910823,
273 	[76] = FSCALE * 0.02237077185616559577,
274 	[77] = FSCALE * 0.02127973643837716938,
275 	[78] = FSCALE * 0.02024191144580438847,
276 	[79] = FSCALE * 0.01925470177538692429,
277 	[80] = FSCALE * 0.01831563888873418029,
278 	[81] = FSCALE * 0.01742237463949351138,
279 	[82] = FSCALE * 0.01657267540176124754,
280 	[83] = FSCALE * 0.01576441648485449082,
281 	[84] = FSCALE * 0.01499557682047770621,
282 	[85] = FSCALE * 0.01426423390899925527,
283 	[86] = FSCALE * 0.01356855901220093175,
284 	[87] = FSCALE * 0.01290681258047986886,
285 	[88] = FSCALE * 0.01227733990306844117,
286 	[89] = FSCALE * 0.01167856697039544521,
287 	[90] = FSCALE * 0.01110899653824230649,
288 	[91] = FSCALE * 0.01056720438385265337,
289 	[92] = FSCALE * 0.01005183574463358164,
290 	[93] = FSCALE * 0.00956160193054350793,
291 	[94] = FSCALE * 0.00909527710169581709,
292 	[95] = FSCALE * 0.00865169520312063417,
293 	[96] = FSCALE * 0.00822974704902002884,
294 	[97] = FSCALE * 0.00782837754922577143,
295 	[98] = FSCALE * 0.00744658307092434051,
296 	[99] = FSCALE * 0.00708340892905212004,
297 	[100] = FSCALE * 0.00673794699908546709,
298 	[101] = FSCALE * 0.00640933344625638184,
299 	[102] = FSCALE * 0.00609674656551563610,
300 	[103] = FSCALE * 0.00579940472684214321,
301 	[104] = FSCALE * 0.00551656442076077241,
302 	[105] = FSCALE * 0.00524751839918138427,
303 	[106] = FSCALE * 0.00499159390691021621,
304 	[107] = FSCALE * 0.00474815099941147558,
305 	[108] = FSCALE * 0.00451658094261266798,
306 	[109] = FSCALE * 0.00429630469075234057,
307 	[110] = FSCALE * 0.00408677143846406699,
308 };
309 #endif
310 
311 #define	CCPU_EXP_MAX	110
312 
313 /*
314  * This function is analogical to the getpcpu() function in the ps(1) command.
315  * They should both calculate in the same way so that the racct %cpu
316  * calculations are consistent with the values showed by the ps(1) tool.
317  * The calculations are more complex in the 4BSD scheduler because of the value
318  * of the ccpu variable.  In ULE it is defined to be zero which saves us some
319  * work.
320  */
321 static uint64_t
322 racct_getpcpu(struct proc *p, u_int pcpu)
323 {
324 	u_int swtime;
325 #ifdef SCHED_4BSD
326 	fixpt_t pctcpu, pctcpu_next;
327 #endif
328 #ifdef SMP
329 	struct pcpu *pc;
330 	int found;
331 #endif
332 	fixpt_t p_pctcpu;
333 	struct thread *td;
334 
335 	ASSERT_RACCT_ENABLED();
336 
337 	/*
338 	 * If the process is swapped out, we count its %cpu usage as zero.
339 	 * This behaviour is consistent with the userland ps(1) tool.
340 	 */
341 	if ((p->p_flag & P_INMEM) == 0)
342 		return (0);
343 	swtime = (ticks - p->p_swtick) / hz;
344 
345 	/*
346 	 * For short-lived processes, the sched_pctcpu() returns small
347 	 * values even for cpu intensive processes.  Therefore we use
348 	 * our own estimate in this case.
349 	 */
350 	if (swtime < RACCT_PCPU_SECS)
351 		return (pcpu);
352 
353 	p_pctcpu = 0;
354 	FOREACH_THREAD_IN_PROC(p, td) {
355 		if (td == PCPU_GET(idlethread))
356 			continue;
357 #ifdef SMP
358 		found = 0;
359 		STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
360 			if (td == pc->pc_idlethread) {
361 				found = 1;
362 				break;
363 			}
364 		}
365 		if (found)
366 			continue;
367 #endif
368 		thread_lock(td);
369 #ifdef SCHED_4BSD
370 		pctcpu = sched_pctcpu(td);
371 		/* Count also the yet unfinished second. */
372 		pctcpu_next = (pctcpu * ccpu_exp[1]) >> FSHIFT;
373 		pctcpu_next += sched_pctcpu_delta(td);
374 		p_pctcpu += max(pctcpu, pctcpu_next);
375 #else
376 		/*
377 		 * In ULE the %cpu statistics are updated on every
378 		 * sched_pctcpu() call.  So special calculations to
379 		 * account for the latest (unfinished) second are
380 		 * not needed.
381 		 */
382 		p_pctcpu += sched_pctcpu(td);
383 #endif
384 		thread_unlock(td);
385 	}
386 
387 #ifdef SCHED_4BSD
388 	if (swtime <= CCPU_EXP_MAX)
389 		return ((100 * (uint64_t)p_pctcpu * 1000000) /
390 		    (FSCALE - ccpu_exp[swtime]));
391 #endif
392 
393 	return ((100 * (uint64_t)p_pctcpu * 1000000) / FSCALE);
394 }
395 
396 static void
397 racct_add_racct(struct racct *dest, const struct racct *src)
398 {
399 	int i;
400 
401 	ASSERT_RACCT_ENABLED();
402 	RACCT_LOCK_ASSERT();
403 
404 	/*
405 	 * Update resource usage in dest.
406 	 */
407 	for (i = 0; i <= RACCT_MAX; i++) {
408 		KASSERT(dest->r_resources[i] >= 0,
409 		    ("%s: resource %d propagation meltdown: dest < 0",
410 		    __func__, i));
411 		KASSERT(src->r_resources[i] >= 0,
412 		    ("%s: resource %d propagation meltdown: src < 0",
413 		    __func__, i));
414 		dest->r_resources[i] += src->r_resources[i];
415 	}
416 }
417 
418 static void
419 racct_sub_racct(struct racct *dest, const struct racct *src)
420 {
421 	int i;
422 
423 	ASSERT_RACCT_ENABLED();
424 	RACCT_LOCK_ASSERT();
425 
426 	/*
427 	 * Update resource usage in dest.
428 	 */
429 	for (i = 0; i <= RACCT_MAX; i++) {
430 		if (!RACCT_IS_SLOPPY(i) && !RACCT_IS_DECAYING(i)) {
431 			KASSERT(dest->r_resources[i] >= 0,
432 			    ("%s: resource %d propagation meltdown: dest < 0",
433 			    __func__, i));
434 			KASSERT(src->r_resources[i] >= 0,
435 			    ("%s: resource %d propagation meltdown: src < 0",
436 			    __func__, i));
437 			KASSERT(src->r_resources[i] <= dest->r_resources[i],
438 			    ("%s: resource %d propagation meltdown: src > dest",
439 			    __func__, i));
440 		}
441 		if (RACCT_CAN_DROP(i)) {
442 			dest->r_resources[i] -= src->r_resources[i];
443 			if (dest->r_resources[i] < 0)
444 				dest->r_resources[i] = 0;
445 		}
446 	}
447 }
448 
449 void
450 racct_create(struct racct **racctp)
451 {
452 
453 	if (!racct_enable)
454 		return;
455 
456 	SDT_PROBE1(racct, , racct, create, racctp);
457 
458 	KASSERT(*racctp == NULL, ("racct already allocated"));
459 
460 	*racctp = uma_zalloc(racct_zone, M_WAITOK | M_ZERO);
461 }
462 
463 static void
464 racct_destroy_locked(struct racct **racctp)
465 {
466 	struct racct *racct;
467 	int i;
468 
469 	ASSERT_RACCT_ENABLED();
470 
471 	SDT_PROBE1(racct, , racct, destroy, racctp);
472 
473 	RACCT_LOCK_ASSERT();
474 	KASSERT(racctp != NULL, ("NULL racctp"));
475 	KASSERT(*racctp != NULL, ("NULL racct"));
476 
477 	racct = *racctp;
478 
479 	for (i = 0; i <= RACCT_MAX; i++) {
480 		if (RACCT_IS_SLOPPY(i))
481 			continue;
482 		if (!RACCT_IS_RECLAIMABLE(i))
483 			continue;
484 		KASSERT(racct->r_resources[i] == 0,
485 		    ("destroying non-empty racct: "
486 		    "%ju allocated for resource %d\n",
487 		    racct->r_resources[i], i));
488 	}
489 	uma_zfree(racct_zone, racct);
490 	*racctp = NULL;
491 }
492 
493 void
494 racct_destroy(struct racct **racct)
495 {
496 
497 	if (!racct_enable)
498 		return;
499 
500 	RACCT_LOCK();
501 	racct_destroy_locked(racct);
502 	RACCT_UNLOCK();
503 }
504 
505 /*
506  * Increase consumption of 'resource' by 'amount' for 'racct',
507  * but not its parents.  Differently from other cases, 'amount' here
508  * may be less than zero.
509  */
510 static void
511 racct_adjust_resource(struct racct *racct, int resource,
512     int64_t amount)
513 {
514 
515 	ASSERT_RACCT_ENABLED();
516 	RACCT_LOCK_ASSERT();
517 	KASSERT(racct != NULL, ("NULL racct"));
518 
519 	racct->r_resources[resource] += amount;
520 	if (racct->r_resources[resource] < 0) {
521 		KASSERT(RACCT_IS_SLOPPY(resource) || RACCT_IS_DECAYING(resource),
522 		    ("%s: resource %d usage < 0", __func__, resource));
523 		racct->r_resources[resource] = 0;
524 	}
525 
526 	/*
527 	 * There are some cases where the racct %cpu resource would grow
528 	 * beyond 100% per core.  For example in racct_proc_exit() we add
529 	 * the process %cpu usage to the ucred racct containers.  If too
530 	 * many processes terminated in a short time span, the ucred %cpu
531 	 * resource could grow too much.  Also, the 4BSD scheduler sometimes
532 	 * returns for a thread more than 100% cpu usage. So we set a sane
533 	 * boundary here to 100% * the maximum number of CPUs.
534 	 */
535 	if ((resource == RACCT_PCTCPU) &&
536 	    (racct->r_resources[RACCT_PCTCPU] > 100 * 1000000 * (int64_t)MAXCPU))
537 		racct->r_resources[RACCT_PCTCPU] = 100 * 1000000 * (int64_t)MAXCPU;
538 }
539 
540 static int
541 racct_add_locked(struct proc *p, int resource, uint64_t amount, int force)
542 {
543 #ifdef RCTL
544 	int error;
545 #endif
546 
547 	ASSERT_RACCT_ENABLED();
548 
549 	/*
550 	 * We need proc lock to dereference p->p_ucred.
551 	 */
552 	PROC_LOCK_ASSERT(p, MA_OWNED);
553 
554 #ifdef RCTL
555 	error = rctl_enforce(p, resource, amount);
556 	if (error && !force && RACCT_IS_DENIABLE(resource)) {
557 		SDT_PROBE3(racct, , rusage, add__failure, p, resource, amount);
558 		return (error);
559 	}
560 #endif
561 	racct_adjust_resource(p->p_racct, resource, amount);
562 	racct_add_cred_locked(p->p_ucred, resource, amount);
563 
564 	return (0);
565 }
566 
567 /*
568  * Increase allocation of 'resource' by 'amount' for process 'p'.
569  * Return 0 if it's below limits, or errno, if it's not.
570  */
571 int
572 racct_add(struct proc *p, int resource, uint64_t amount)
573 {
574 	int error;
575 
576 	if (!racct_enable)
577 		return (0);
578 
579 	SDT_PROBE3(racct, , rusage, add, p, resource, amount);
580 
581 	RACCT_LOCK();
582 	error = racct_add_locked(p, resource, amount, 0);
583 	RACCT_UNLOCK();
584 	return (error);
585 }
586 
587 /*
588  * Increase allocation of 'resource' by 'amount' for process 'p'.
589  * Doesn't check for limits and never fails.
590  */
591 void
592 racct_add_force(struct proc *p, int resource, uint64_t amount)
593 {
594 
595 	if (!racct_enable)
596 		return;
597 
598 	SDT_PROBE3(racct, , rusage, add__force, p, resource, amount);
599 
600 	RACCT_LOCK();
601 	racct_add_locked(p, resource, amount, 1);
602 	RACCT_UNLOCK();
603 }
604 
605 static void
606 racct_add_cred_locked(struct ucred *cred, int resource, uint64_t amount)
607 {
608 	struct prison *pr;
609 
610 	ASSERT_RACCT_ENABLED();
611 
612 	racct_adjust_resource(cred->cr_ruidinfo->ui_racct, resource, amount);
613 	for (pr = cred->cr_prison; pr != NULL; pr = pr->pr_parent)
614 		racct_adjust_resource(pr->pr_prison_racct->prr_racct, resource,
615 		    amount);
616 	racct_adjust_resource(cred->cr_loginclass->lc_racct, resource, amount);
617 }
618 
619 /*
620  * Increase allocation of 'resource' by 'amount' for credential 'cred'.
621  * Doesn't check for limits and never fails.
622  */
623 void
624 racct_add_cred(struct ucred *cred, int resource, uint64_t amount)
625 {
626 
627 	if (!racct_enable)
628 		return;
629 
630 	SDT_PROBE3(racct, , rusage, add__cred, cred, resource, amount);
631 
632 	RACCT_LOCK();
633 	racct_add_cred_locked(cred, resource, amount);
634 	RACCT_UNLOCK();
635 }
636 
637 /*
638  * Account for disk IO resource consumption.  Checks for limits,
639  * but never fails, due to disk limits being undeniable.
640  */
641 void
642 racct_add_buf(struct proc *p, const struct buf *bp, int is_write)
643 {
644 
645 	ASSERT_RACCT_ENABLED();
646 	PROC_LOCK_ASSERT(p, MA_OWNED);
647 
648 	SDT_PROBE3(racct, , rusage, add__buf, p, bp, is_write);
649 
650 	RACCT_LOCK();
651 	if (is_write) {
652 		racct_add_locked(curproc, RACCT_WRITEBPS, bp->b_bcount, 1);
653 		racct_add_locked(curproc, RACCT_WRITEIOPS, 1, 1);
654 	} else {
655 		racct_add_locked(curproc, RACCT_READBPS, bp->b_bcount, 1);
656 		racct_add_locked(curproc, RACCT_READIOPS, 1, 1);
657 	}
658 	RACCT_UNLOCK();
659 }
660 
661 static int
662 racct_set_locked(struct proc *p, int resource, uint64_t amount, int force)
663 {
664 	int64_t old_amount, decayed_amount, diff_proc, diff_cred;
665 #ifdef RCTL
666 	int error;
667 #endif
668 
669 	ASSERT_RACCT_ENABLED();
670 
671 	/*
672 	 * We need proc lock to dereference p->p_ucred.
673 	 */
674 	PROC_LOCK_ASSERT(p, MA_OWNED);
675 
676 	old_amount = p->p_racct->r_resources[resource];
677 	/*
678 	 * The diffs may be negative.
679 	 */
680 	diff_proc = amount - old_amount;
681 	if (resource == RACCT_PCTCPU) {
682 		/*
683 		 * Resources in per-credential racct containers may decay.
684 		 * If this is the case, we need to calculate the difference
685 		 * between the new amount and the proportional value of the
686 		 * old amount that has decayed in the ucred racct containers.
687 		 */
688 		decayed_amount = old_amount * RACCT_DECAY_FACTOR / FSCALE;
689 		diff_cred = amount - decayed_amount;
690 	} else
691 		diff_cred = diff_proc;
692 #ifdef notyet
693 	KASSERT(diff_proc >= 0 || RACCT_CAN_DROP(resource),
694 	    ("%s: usage of non-droppable resource %d dropping", __func__,
695 	     resource));
696 #endif
697 #ifdef RCTL
698 	if (diff_proc > 0) {
699 		error = rctl_enforce(p, resource, diff_proc);
700 		if (error && !force && RACCT_IS_DENIABLE(resource)) {
701 			SDT_PROBE3(racct, , rusage, set__failure, p, resource,
702 			    amount);
703 			return (error);
704 		}
705 	}
706 #endif
707 	racct_adjust_resource(p->p_racct, resource, diff_proc);
708 	if (diff_cred > 0)
709 		racct_add_cred_locked(p->p_ucred, resource, diff_cred);
710 	else if (diff_cred < 0)
711 		racct_sub_cred_locked(p->p_ucred, resource, -diff_cred);
712 
713 	return (0);
714 }
715 
716 /*
717  * Set allocation of 'resource' to 'amount' for process 'p'.
718  * Return 0 if it's below limits, or errno, if it's not.
719  *
720  * Note that decreasing the allocation always returns 0,
721  * even if it's above the limit.
722  */
723 int
724 racct_set_unlocked(struct proc *p, int resource, uint64_t amount)
725 {
726 	int error;
727 
728 	ASSERT_RACCT_ENABLED();
729 	PROC_LOCK(p);
730 	error = racct_set(p, resource, amount);
731 	PROC_UNLOCK(p);
732 	return (error);
733 }
734 
735 int
736 racct_set(struct proc *p, int resource, uint64_t amount)
737 {
738 	int error;
739 
740 	if (!racct_enable)
741 		return (0);
742 
743 	SDT_PROBE3(racct, , rusage, set__force, p, resource, amount);
744 
745 	RACCT_LOCK();
746 	error = racct_set_locked(p, resource, amount, 0);
747 	RACCT_UNLOCK();
748 	return (error);
749 }
750 
751 void
752 racct_set_force(struct proc *p, int resource, uint64_t amount)
753 {
754 
755 	if (!racct_enable)
756 		return;
757 
758 	SDT_PROBE3(racct, , rusage, set, p, resource, amount);
759 
760 	RACCT_LOCK();
761 	racct_set_locked(p, resource, amount, 1);
762 	RACCT_UNLOCK();
763 }
764 
765 /*
766  * Returns amount of 'resource' the process 'p' can keep allocated.
767  * Allocating more than that would be denied, unless the resource
768  * is marked undeniable.  Amount of already allocated resource does
769  * not matter.
770  */
771 uint64_t
772 racct_get_limit(struct proc *p, int resource)
773 {
774 #ifdef RCTL
775 	uint64_t available;
776 
777 	if (!racct_enable)
778 		return (UINT64_MAX);
779 
780 	RACCT_LOCK();
781 	available = rctl_get_limit(p, resource);
782 	RACCT_UNLOCK();
783 
784 	return (available);
785 #else
786 
787 	return (UINT64_MAX);
788 #endif
789 }
790 
791 /*
792  * Returns amount of 'resource' the process 'p' can keep allocated.
793  * Allocating more than that would be denied, unless the resource
794  * is marked undeniable.  Amount of already allocated resource does
795  * matter.
796  */
797 uint64_t
798 racct_get_available(struct proc *p, int resource)
799 {
800 #ifdef RCTL
801 	uint64_t available;
802 
803 	if (!racct_enable)
804 		return (UINT64_MAX);
805 
806 	RACCT_LOCK();
807 	available = rctl_get_available(p, resource);
808 	RACCT_UNLOCK();
809 
810 	return (available);
811 #else
812 
813 	return (UINT64_MAX);
814 #endif
815 }
816 
817 /*
818  * Returns amount of the %cpu resource that process 'p' can add to its %cpu
819  * utilization.  Adding more than that would lead to the process being
820  * throttled.
821  */
822 static int64_t
823 racct_pcpu_available(struct proc *p)
824 {
825 #ifdef RCTL
826 	uint64_t available;
827 
828 	ASSERT_RACCT_ENABLED();
829 
830 	RACCT_LOCK();
831 	available = rctl_pcpu_available(p);
832 	RACCT_UNLOCK();
833 
834 	return (available);
835 #else
836 
837 	return (INT64_MAX);
838 #endif
839 }
840 
841 /*
842  * Decrease allocation of 'resource' by 'amount' for process 'p'.
843  */
844 void
845 racct_sub(struct proc *p, int resource, uint64_t amount)
846 {
847 
848 	if (!racct_enable)
849 		return;
850 
851 	SDT_PROBE3(racct, , rusage, sub, p, resource, amount);
852 
853 	/*
854 	 * We need proc lock to dereference p->p_ucred.
855 	 */
856 	PROC_LOCK_ASSERT(p, MA_OWNED);
857 	KASSERT(RACCT_CAN_DROP(resource),
858 	    ("%s: called for non-droppable resource %d", __func__, resource));
859 
860 	RACCT_LOCK();
861 	KASSERT(amount <= p->p_racct->r_resources[resource],
862 	    ("%s: freeing %ju of resource %d, which is more "
863 	     "than allocated %jd for %s (pid %d)", __func__, amount, resource,
864 	    (intmax_t)p->p_racct->r_resources[resource], p->p_comm, p->p_pid));
865 
866 	racct_adjust_resource(p->p_racct, resource, -amount);
867 	racct_sub_cred_locked(p->p_ucred, resource, amount);
868 	RACCT_UNLOCK();
869 }
870 
871 static void
872 racct_sub_cred_locked(struct ucred *cred, int resource, uint64_t amount)
873 {
874 	struct prison *pr;
875 
876 	ASSERT_RACCT_ENABLED();
877 
878 	racct_adjust_resource(cred->cr_ruidinfo->ui_racct, resource, -amount);
879 	for (pr = cred->cr_prison; pr != NULL; pr = pr->pr_parent)
880 		racct_adjust_resource(pr->pr_prison_racct->prr_racct, resource,
881 		    -amount);
882 	racct_adjust_resource(cred->cr_loginclass->lc_racct, resource, -amount);
883 }
884 
885 /*
886  * Decrease allocation of 'resource' by 'amount' for credential 'cred'.
887  */
888 void
889 racct_sub_cred(struct ucred *cred, int resource, uint64_t amount)
890 {
891 
892 	if (!racct_enable)
893 		return;
894 
895 	SDT_PROBE3(racct, , rusage, sub__cred, cred, resource, amount);
896 
897 #ifdef notyet
898 	KASSERT(RACCT_CAN_DROP(resource),
899 	    ("%s: called for resource %d which can not drop", __func__,
900 	     resource));
901 #endif
902 
903 	RACCT_LOCK();
904 	racct_sub_cred_locked(cred, resource, amount);
905 	RACCT_UNLOCK();
906 }
907 
908 /*
909  * Inherit resource usage information from the parent process.
910  */
911 int
912 racct_proc_fork(struct proc *parent, struct proc *child)
913 {
914 	int i, error = 0;
915 
916 	if (!racct_enable)
917 		return (0);
918 
919 	/*
920 	 * Create racct for the child process.
921 	 */
922 	racct_create(&child->p_racct);
923 
924 	PROC_LOCK(parent);
925 	PROC_LOCK(child);
926 	RACCT_LOCK();
927 
928 #ifdef RCTL
929 	error = rctl_proc_fork(parent, child);
930 	if (error != 0)
931 		goto out;
932 #endif
933 
934 	/* Init process cpu time. */
935 	child->p_prev_runtime = 0;
936 	child->p_throttled = 0;
937 
938 	/*
939 	 * Inherit resource usage.
940 	 */
941 	for (i = 0; i <= RACCT_MAX; i++) {
942 		if (parent->p_racct->r_resources[i] == 0 ||
943 		    !RACCT_IS_INHERITABLE(i))
944 			continue;
945 
946 		error = racct_set_locked(child, i,
947 		    parent->p_racct->r_resources[i], 0);
948 		if (error != 0)
949 			goto out;
950 	}
951 
952 	error = racct_add_locked(child, RACCT_NPROC, 1, 0);
953 	error += racct_add_locked(child, RACCT_NTHR, 1, 0);
954 
955 out:
956 	RACCT_UNLOCK();
957 	PROC_UNLOCK(child);
958 	PROC_UNLOCK(parent);
959 
960 	if (error != 0)
961 		racct_proc_exit(child);
962 
963 	return (error);
964 }
965 
966 /*
967  * Called at the end of fork1(), to handle rules that require the process
968  * to be fully initialized.
969  */
970 void
971 racct_proc_fork_done(struct proc *child)
972 {
973 
974 	if (!racct_enable)
975 		return;
976 
977 #ifdef RCTL
978 	PROC_LOCK(child);
979 	RACCT_LOCK();
980 	rctl_enforce(child, RACCT_NPROC, 0);
981 	rctl_enforce(child, RACCT_NTHR, 0);
982 	RACCT_UNLOCK();
983 	PROC_UNLOCK(child);
984 #endif
985 }
986 
987 void
988 racct_proc_exit(struct proc *p)
989 {
990 	struct timeval wallclock;
991 	uint64_t pct_estimate, pct, runtime;
992 	int i;
993 
994 	if (!racct_enable)
995 		return;
996 
997 	PROC_LOCK(p);
998 	/*
999 	 * We don't need to calculate rux, proc_reap() has already done this.
1000 	 */
1001 	runtime = cputick2usec(p->p_rux.rux_runtime);
1002 #ifdef notyet
1003 	KASSERT(runtime >= p->p_prev_runtime, ("runtime < p_prev_runtime"));
1004 #else
1005 	if (runtime < p->p_prev_runtime)
1006 		runtime = p->p_prev_runtime;
1007 #endif
1008 	microuptime(&wallclock);
1009 	timevalsub(&wallclock, &p->p_stats->p_start);
1010 	if (wallclock.tv_sec > 0 || wallclock.tv_usec > 0) {
1011 		pct_estimate = (1000000 * runtime * 100) /
1012 		    ((uint64_t)wallclock.tv_sec * 1000000 +
1013 		    wallclock.tv_usec);
1014 	} else
1015 		pct_estimate = 0;
1016 	pct = racct_getpcpu(p, pct_estimate);
1017 
1018 	RACCT_LOCK();
1019 	racct_set_locked(p, RACCT_CPU, runtime, 0);
1020 	racct_add_cred_locked(p->p_ucred, RACCT_PCTCPU, pct);
1021 
1022 	KASSERT(p->p_racct->r_resources[RACCT_RSS] == 0,
1023 	    ("process reaped with %ju allocated for RSS\n",
1024 	    p->p_racct->r_resources[RACCT_RSS]));
1025 	for (i = 0; i <= RACCT_MAX; i++) {
1026 		if (p->p_racct->r_resources[i] == 0)
1027 			continue;
1028 		if (!RACCT_IS_RECLAIMABLE(i))
1029 			continue;
1030 		racct_set_locked(p, i, 0, 0);
1031 	}
1032 
1033 #ifdef RCTL
1034 	rctl_racct_release(p->p_racct);
1035 #endif
1036 	racct_destroy_locked(&p->p_racct);
1037 	RACCT_UNLOCK();
1038 	PROC_UNLOCK(p);
1039 }
1040 
1041 /*
1042  * Called after credentials change, to move resource utilisation
1043  * between raccts.
1044  */
1045 void
1046 racct_proc_ucred_changed(struct proc *p, struct ucred *oldcred,
1047     struct ucred *newcred)
1048 {
1049 	struct uidinfo *olduip, *newuip;
1050 	struct loginclass *oldlc, *newlc;
1051 	struct prison *oldpr, *newpr, *pr;
1052 
1053 	if (!racct_enable)
1054 		return;
1055 
1056 	PROC_LOCK_ASSERT(p, MA_OWNED);
1057 
1058 	newuip = newcred->cr_ruidinfo;
1059 	olduip = oldcred->cr_ruidinfo;
1060 	newlc = newcred->cr_loginclass;
1061 	oldlc = oldcred->cr_loginclass;
1062 	newpr = newcred->cr_prison;
1063 	oldpr = oldcred->cr_prison;
1064 
1065 	RACCT_LOCK();
1066 	if (newuip != olduip) {
1067 		racct_sub_racct(olduip->ui_racct, p->p_racct);
1068 		racct_add_racct(newuip->ui_racct, p->p_racct);
1069 	}
1070 	if (newlc != oldlc) {
1071 		racct_sub_racct(oldlc->lc_racct, p->p_racct);
1072 		racct_add_racct(newlc->lc_racct, p->p_racct);
1073 	}
1074 	if (newpr != oldpr) {
1075 		for (pr = oldpr; pr != NULL; pr = pr->pr_parent)
1076 			racct_sub_racct(pr->pr_prison_racct->prr_racct,
1077 			    p->p_racct);
1078 		for (pr = newpr; pr != NULL; pr = pr->pr_parent)
1079 			racct_add_racct(pr->pr_prison_racct->prr_racct,
1080 			    p->p_racct);
1081 	}
1082 	RACCT_UNLOCK();
1083 }
1084 
1085 void
1086 racct_move(struct racct *dest, struct racct *src)
1087 {
1088 
1089 	ASSERT_RACCT_ENABLED();
1090 
1091 	RACCT_LOCK();
1092 	racct_add_racct(dest, src);
1093 	racct_sub_racct(src, src);
1094 	RACCT_UNLOCK();
1095 }
1096 
1097 static void
1098 ast_racct(struct thread *td, int tda __unused)
1099 {
1100 	struct proc *p;
1101 
1102 	ASSERT_RACCT_ENABLED();
1103 
1104 	p = td->td_proc;
1105 	if (p->p_throttled == 0)
1106 		return;
1107 
1108 	PROC_LOCK(p);
1109 	while (p->p_throttled != 0) {
1110 		msleep(p->p_racct, &p->p_mtx, 0, "racct",
1111 		    p->p_throttled < 0 ? 0 : p->p_throttled);
1112 		if (p->p_throttled > 0)
1113 			p->p_throttled = 0;
1114 	}
1115 	PROC_UNLOCK(p);
1116 }
1117 
1118 /*
1119  * Make the process sleep in userret() for 'timeout' ticks.  Setting
1120  * timeout to -1 makes it sleep until woken up by racct_proc_wakeup().
1121  */
1122 void
1123 racct_proc_throttle(struct proc *p, int timeout)
1124 {
1125 	struct thread *td;
1126 #ifdef SMP
1127 	int cpuid;
1128 #endif
1129 
1130 	KASSERT(timeout != 0, ("timeout %d", timeout));
1131 	ASSERT_RACCT_ENABLED();
1132 	PROC_LOCK_ASSERT(p, MA_OWNED);
1133 
1134 	/*
1135 	 * Do not block kernel processes.  Also do not block processes with
1136 	 * low %cpu utilization to improve interactivity.
1137 	 */
1138 	if ((p->p_flag & (P_SYSTEM | P_KPROC)) != 0)
1139 		return;
1140 
1141 	if (p->p_throttled < 0 || (timeout > 0 && p->p_throttled > timeout))
1142 		return;
1143 
1144 	p->p_throttled = timeout;
1145 
1146 	FOREACH_THREAD_IN_PROC(p, td) {
1147 		thread_lock(td);
1148 		ast_sched_locked(td, TDA_RACCT);
1149 
1150 		switch (TD_GET_STATE(td)) {
1151 		case TDS_RUNQ:
1152 			/*
1153 			 * If the thread is on the scheduler run-queue, we can
1154 			 * not just remove it from there.  So we set the flag
1155 			 * TDA_SCHED for the thread, so that once it is
1156 			 * running, it is taken off the cpu as soon as possible.
1157 			 */
1158 			ast_sched_locked(td, TDA_SCHED);
1159 			break;
1160 		case TDS_RUNNING:
1161 			/*
1162 			 * If the thread is running, we request a context
1163 			 * switch for it by setting the TDA_SCHED flag.
1164 			 */
1165 			ast_sched_locked(td, TDA_SCHED);
1166 #ifdef SMP
1167 			cpuid = td->td_oncpu;
1168 			if ((cpuid != NOCPU) && (td != curthread))
1169 				ipi_cpu(cpuid, IPI_AST);
1170 #endif
1171 			break;
1172 		default:
1173 			break;
1174 		}
1175 		thread_unlock(td);
1176 	}
1177 }
1178 
1179 static void
1180 racct_proc_wakeup(struct proc *p)
1181 {
1182 
1183 	ASSERT_RACCT_ENABLED();
1184 
1185 	PROC_LOCK_ASSERT(p, MA_OWNED);
1186 
1187 	if (p->p_throttled != 0) {
1188 		p->p_throttled = 0;
1189 		wakeup(p->p_racct);
1190 	}
1191 }
1192 
1193 static void
1194 racct_decay_callback(struct racct *racct, void *dummy1, void *dummy2)
1195 {
1196 	int64_t r_old, r_new;
1197 
1198 	ASSERT_RACCT_ENABLED();
1199 	RACCT_LOCK_ASSERT();
1200 
1201 #ifdef RCTL
1202 	rctl_throttle_decay(racct, RACCT_READBPS);
1203 	rctl_throttle_decay(racct, RACCT_WRITEBPS);
1204 	rctl_throttle_decay(racct, RACCT_READIOPS);
1205 	rctl_throttle_decay(racct, RACCT_WRITEIOPS);
1206 #endif
1207 
1208 	r_old = racct->r_resources[RACCT_PCTCPU];
1209 
1210 	/* If there is nothing to decay, just exit. */
1211 	if (r_old <= 0)
1212 		return;
1213 
1214 	r_new = r_old * RACCT_DECAY_FACTOR / FSCALE;
1215 	racct->r_resources[RACCT_PCTCPU] = r_new;
1216 }
1217 
1218 static void
1219 racct_decay_pre(void)
1220 {
1221 
1222 	RACCT_LOCK();
1223 }
1224 
1225 static void
1226 racct_decay_post(void)
1227 {
1228 
1229 	RACCT_UNLOCK();
1230 }
1231 
1232 static void
1233 racct_decay(void)
1234 {
1235 
1236 	ASSERT_RACCT_ENABLED();
1237 
1238 	ui_racct_foreach(racct_decay_callback, racct_decay_pre,
1239 	    racct_decay_post, NULL, NULL);
1240 	loginclass_racct_foreach(racct_decay_callback, racct_decay_pre,
1241 	    racct_decay_post, NULL, NULL);
1242 	prison_racct_foreach(racct_decay_callback, racct_decay_pre,
1243 	    racct_decay_post, NULL, NULL);
1244 }
1245 
1246 static void
1247 racctd(void)
1248 {
1249 	struct thread *td;
1250 	struct proc *p;
1251 	struct timeval wallclock;
1252 	uint64_t pct, pct_estimate, runtime;
1253 
1254 	ASSERT_RACCT_ENABLED();
1255 
1256 	for (;;) {
1257 		racct_decay();
1258 
1259 		sx_slock(&allproc_lock);
1260 
1261 		FOREACH_PROC_IN_SYSTEM(p) {
1262 			PROC_LOCK(p);
1263 			if (p->p_state != PRS_NORMAL) {
1264 				if (p->p_state == PRS_ZOMBIE)
1265 					racct_set(p, RACCT_PCTCPU, 0);
1266 				PROC_UNLOCK(p);
1267 				continue;
1268 			}
1269 
1270 			microuptime(&wallclock);
1271 			timevalsub(&wallclock, &p->p_stats->p_start);
1272 			PROC_STATLOCK(p);
1273 			FOREACH_THREAD_IN_PROC(p, td)
1274 				ruxagg(p, td);
1275 			runtime = cputick2usec(p->p_rux.rux_runtime);
1276 			PROC_STATUNLOCK(p);
1277 #ifdef notyet
1278 			KASSERT(runtime >= p->p_prev_runtime,
1279 			    ("runtime < p_prev_runtime"));
1280 #else
1281 			if (runtime < p->p_prev_runtime)
1282 				runtime = p->p_prev_runtime;
1283 #endif
1284 			p->p_prev_runtime = runtime;
1285 			if (wallclock.tv_sec > 0 || wallclock.tv_usec > 0) {
1286 				pct_estimate = (1000000 * runtime * 100) /
1287 				    ((uint64_t)wallclock.tv_sec * 1000000 +
1288 				    wallclock.tv_usec);
1289 			} else
1290 				pct_estimate = 0;
1291 			pct = racct_getpcpu(p, pct_estimate);
1292 			RACCT_LOCK();
1293 #ifdef RCTL
1294 			rctl_throttle_decay(p->p_racct, RACCT_READBPS);
1295 			rctl_throttle_decay(p->p_racct, RACCT_WRITEBPS);
1296 			rctl_throttle_decay(p->p_racct, RACCT_READIOPS);
1297 			rctl_throttle_decay(p->p_racct, RACCT_WRITEIOPS);
1298 #endif
1299 			racct_set_locked(p, RACCT_PCTCPU, pct, 1);
1300 			racct_set_locked(p, RACCT_CPU, runtime, 0);
1301 			racct_set_locked(p, RACCT_WALLCLOCK,
1302 			    (uint64_t)wallclock.tv_sec * 1000000 +
1303 			    wallclock.tv_usec, 0);
1304 			RACCT_UNLOCK();
1305 			PROC_UNLOCK(p);
1306 		}
1307 
1308 		/*
1309 		 * To ensure that processes are throttled in a fair way, we need
1310 		 * to iterate over all processes again and check the limits
1311 		 * for %cpu resource only after ucred racct containers have been
1312 		 * properly filled.
1313 		 */
1314 		FOREACH_PROC_IN_SYSTEM(p) {
1315 			PROC_LOCK(p);
1316 			if (p->p_state != PRS_NORMAL) {
1317 				PROC_UNLOCK(p);
1318 				continue;
1319 			}
1320 
1321 			if (racct_pcpu_available(p) <= 0) {
1322 				if (p->p_racct->r_resources[RACCT_PCTCPU] >
1323 				    pcpu_threshold)
1324 					racct_proc_throttle(p, -1);
1325 			} else if (p->p_throttled == -1) {
1326 				racct_proc_wakeup(p);
1327 			}
1328 			PROC_UNLOCK(p);
1329 		}
1330 		sx_sunlock(&allproc_lock);
1331 		pause("-", hz);
1332 	}
1333 }
1334 
1335 static struct kproc_desc racctd_kp = {
1336 	"racctd",
1337 	racctd,
1338 	NULL
1339 };
1340 
1341 static void
1342 racctd_init(void)
1343 {
1344 	if (!racct_enable)
1345 		return;
1346 
1347 	kproc_start(&racctd_kp);
1348 }
1349 SYSINIT(racctd, SI_SUB_RACCTD, SI_ORDER_FIRST, racctd_init, NULL);
1350 
1351 static void
1352 racct_init(void)
1353 {
1354 	if (!racct_enable)
1355 		return;
1356 
1357 	racct_zone = uma_zcreate("racct", sizeof(struct racct),
1358 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
1359 	ast_register(TDA_RACCT, ASTR_ASTF_REQUIRED, 0, ast_racct);
1360 
1361 	/*
1362 	 * XXX: Move this somewhere.
1363 	 */
1364 	prison0.pr_prison_racct = prison_racct_find("0");
1365 }
1366 SYSINIT(racct, SI_SUB_RACCT, SI_ORDER_FIRST, racct_init, NULL);
1367 
1368 #endif /* !RACCT */
1369