2 * Copyright (c) 2010 The FreeBSD Foundation
5 * This software was developed by Edward Tomasz Napierala under sponsorship
6 * from the FreeBSD Foundation.
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
35 #include "opt_sched.h"
37 #include <sys/param.h>
38 #include <sys/systm.h>
39 #include <sys/eventhandler.h>
41 #include <sys/kernel.h>
42 #include <sys/kthread.h>
44 #include <sys/loginclass.h>
45 #include <sys/malloc.h>
46 #include <sys/mutex.h>
48 #include <sys/racct.h>
49 #include <sys/resourcevar.h>
51 #include <sys/sched.h>
55 #include <sys/sysctl.h>
56 #include <sys/sysent.h>
57 #include <sys/sysproto.h>
59 #include <machine/smp.h>
67 FEATURE(racct, "Resource Accounting");
70 * Do not block processes that have their %cpu usage <= pcpu_threshold.
72 static int pcpu_threshold = 1;
73 #ifdef RACCT_DEFAULT_TO_DISABLED
79 SYSCTL_NODE(_kern, OID_AUTO, racct, CTLFLAG_RW, 0, "Resource Accounting");
80 SYSCTL_UINT(_kern_racct, OID_AUTO, enable, CTLFLAG_RDTUN, &racct_enable,
81 0, "Enable RACCT/RCTL");
82 SYSCTL_UINT(_kern_racct, OID_AUTO, pcpu_threshold, CTLFLAG_RW, &pcpu_threshold,
83 0, "Processes with higher %cpu usage than this value can be throttled.");
86 * How many seconds it takes to use the scheduler %cpu calculations. When a
87 * process starts, we compute its %cpu usage by dividing its runtime by the
88 * process wall clock time. After RACCT_PCPU_SECS pass, we use the value
89 * provided by the scheduler.
91 #define RACCT_PCPU_SECS 3
93 static struct mtx racct_lock;
94 MTX_SYSINIT(racct_lock, &racct_lock, "racct lock", MTX_DEF);
96 static uma_zone_t racct_zone;
98 static void racct_sub_racct(struct racct *dest, const struct racct *src);
99 static void racct_sub_cred_locked(struct ucred *cred, int resource,
101 static void racct_add_cred_locked(struct ucred *cred, int resource,
104 SDT_PROVIDER_DEFINE(racct);
105 SDT_PROBE_DEFINE3(racct, kernel, rusage, add, "struct proc *", "int",
107 SDT_PROBE_DEFINE3(racct, kernel, rusage, add__failure,
108 "struct proc *", "int", "uint64_t");
109 SDT_PROBE_DEFINE3(racct, kernel, rusage, add__cred, "struct ucred *",
111 SDT_PROBE_DEFINE3(racct, kernel, rusage, add__force, "struct proc *",
113 SDT_PROBE_DEFINE3(racct, kernel, rusage, set, "struct proc *", "int",
115 SDT_PROBE_DEFINE3(racct, kernel, rusage, set__failure,
116 "struct proc *", "int", "uint64_t");
117 SDT_PROBE_DEFINE3(racct, kernel, rusage, sub, "struct proc *", "int",
119 SDT_PROBE_DEFINE3(racct, kernel, rusage, sub__cred, "struct ucred *",
121 SDT_PROBE_DEFINE1(racct, kernel, racct, create, "struct racct *");
122 SDT_PROBE_DEFINE1(racct, kernel, racct, destroy, "struct racct *");
123 SDT_PROBE_DEFINE2(racct, kernel, racct, join, "struct racct *",
125 SDT_PROBE_DEFINE2(racct, kernel, racct, join__failure,
126 "struct racct *", "struct racct *");
127 SDT_PROBE_DEFINE2(racct, kernel, racct, leave, "struct racct *",
130 int racct_types[] = {
134 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
136 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
142 RACCT_RECLAIMABLE | RACCT_DENIABLE,
144 RACCT_RECLAIMABLE | RACCT_DENIABLE,
146 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
148 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
150 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
152 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
154 RACCT_RECLAIMABLE | RACCT_DENIABLE,
156 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
158 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
160 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
162 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
164 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
166 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
168 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
172 RACCT_DECAYING | RACCT_DENIABLE | RACCT_IN_MILLIONS };
174 static const fixpt_t RACCT_DECAY_FACTOR = 0.3 * FSCALE;
178 * Contains intermediate values for %cpu calculations to avoid using floating
179 * point in the kernel.
180 * ccpu_exp[k] = FSCALE * (ccpu/FSCALE)^k = FSCALE * exp(-k/20)
181 * It is needed only for the 4BSD scheduler, because in ULE, the ccpu equals to
182 * zero so the calculations are more straightforward.
184 fixpt_t ccpu_exp[] = {
186 [1] = FSCALE * 0.95122942450071400909,
187 [2] = FSCALE * 0.90483741803595957316,
188 [3] = FSCALE * 0.86070797642505780722,
189 [4] = FSCALE * 0.81873075307798185866,
190 [5] = FSCALE * 0.77880078307140486824,
191 [6] = FSCALE * 0.74081822068171786606,
192 [7] = FSCALE * 0.70468808971871343435,
193 [8] = FSCALE * 0.67032004603563930074,
194 [9] = FSCALE * 0.63762815162177329314,
195 [10] = FSCALE * 0.60653065971263342360,
196 [11] = FSCALE * 0.57694981038048669531,
197 [12] = FSCALE * 0.54881163609402643262,
198 [13] = FSCALE * 0.52204577676101604789,
199 [14] = FSCALE * 0.49658530379140951470,
200 [15] = FSCALE * 0.47236655274101470713,
201 [16] = FSCALE * 0.44932896411722159143,
202 [17] = FSCALE * 0.42741493194872666992,
203 [18] = FSCALE * 0.40656965974059911188,
204 [19] = FSCALE * 0.38674102345450120691,
205 [20] = FSCALE * 0.36787944117144232159,
206 [21] = FSCALE * 0.34993774911115535467,
207 [22] = FSCALE * 0.33287108369807955328,
208 [23] = FSCALE * 0.31663676937905321821,
209 [24] = FSCALE * 0.30119421191220209664,
210 [25] = FSCALE * 0.28650479686019010032,
211 [26] = FSCALE * 0.27253179303401260312,
212 [27] = FSCALE * 0.25924026064589150757,
213 [28] = FSCALE * 0.24659696394160647693,
214 [29] = FSCALE * 0.23457028809379765313,
215 [30] = FSCALE * 0.22313016014842982893,
216 [31] = FSCALE * 0.21224797382674305771,
217 [32] = FSCALE * 0.20189651799465540848,
218 [33] = FSCALE * 0.19204990862075411423,
219 [34] = FSCALE * 0.18268352405273465022,
220 [35] = FSCALE * 0.17377394345044512668,
221 [36] = FSCALE * 0.16529888822158653829,
222 [37] = FSCALE * 0.15723716631362761621,
223 [38] = FSCALE * 0.14956861922263505264,
224 [39] = FSCALE * 0.14227407158651357185,
225 [40] = FSCALE * 0.13533528323661269189,
226 [41] = FSCALE * 0.12873490358780421886,
227 [42] = FSCALE * 0.12245642825298191021,
228 [43] = FSCALE * 0.11648415777349695786,
229 [44] = FSCALE * 0.11080315836233388333,
230 [45] = FSCALE * 0.10539922456186433678,
231 [46] = FSCALE * 0.10025884372280373372,
232 [47] = FSCALE * 0.09536916221554961888,
233 [48] = FSCALE * 0.09071795328941250337,
234 [49] = FSCALE * 0.08629358649937051097,
235 [50] = FSCALE * 0.08208499862389879516,
236 [51] = FSCALE * 0.07808166600115315231,
237 [52] = FSCALE * 0.07427357821433388042,
238 [53] = FSCALE * 0.07065121306042958674,
239 [54] = FSCALE * 0.06720551273974976512,
240 [55] = FSCALE * 0.06392786120670757270,
241 [56] = FSCALE * 0.06081006262521796499,
242 [57] = FSCALE * 0.05784432087483846296,
243 [58] = FSCALE * 0.05502322005640722902,
244 [59] = FSCALE * 0.05233970594843239308,
245 [60] = FSCALE * 0.04978706836786394297,
246 [61] = FSCALE * 0.04735892439114092119,
247 [62] = FSCALE * 0.04504920239355780606,
248 [63] = FSCALE * 0.04285212686704017991,
249 [64] = FSCALE * 0.04076220397836621516,
250 [65] = FSCALE * 0.03877420783172200988,
251 [66] = FSCALE * 0.03688316740124000544,
252 [67] = FSCALE * 0.03508435410084502588,
253 [68] = FSCALE * 0.03337326996032607948,
254 [69] = FSCALE * 0.03174563637806794323,
255 [70] = FSCALE * 0.03019738342231850073,
256 [71] = FSCALE * 0.02872463965423942912,
257 [72] = FSCALE * 0.02732372244729256080,
258 [73] = FSCALE * 0.02599112877875534358,
259 [74] = FSCALE * 0.02472352647033939120,
260 [75] = FSCALE * 0.02351774585600910823,
261 [76] = FSCALE * 0.02237077185616559577,
262 [77] = FSCALE * 0.02127973643837716938,
263 [78] = FSCALE * 0.02024191144580438847,
264 [79] = FSCALE * 0.01925470177538692429,
265 [80] = FSCALE * 0.01831563888873418029,
266 [81] = FSCALE * 0.01742237463949351138,
267 [82] = FSCALE * 0.01657267540176124754,
268 [83] = FSCALE * 0.01576441648485449082,
269 [84] = FSCALE * 0.01499557682047770621,
270 [85] = FSCALE * 0.01426423390899925527,
271 [86] = FSCALE * 0.01356855901220093175,
272 [87] = FSCALE * 0.01290681258047986886,
273 [88] = FSCALE * 0.01227733990306844117,
274 [89] = FSCALE * 0.01167856697039544521,
275 [90] = FSCALE * 0.01110899653824230649,
276 [91] = FSCALE * 0.01056720438385265337,
277 [92] = FSCALE * 0.01005183574463358164,
278 [93] = FSCALE * 0.00956160193054350793,
279 [94] = FSCALE * 0.00909527710169581709,
280 [95] = FSCALE * 0.00865169520312063417,
281 [96] = FSCALE * 0.00822974704902002884,
282 [97] = FSCALE * 0.00782837754922577143,
283 [98] = FSCALE * 0.00744658307092434051,
284 [99] = FSCALE * 0.00708340892905212004,
285 [100] = FSCALE * 0.00673794699908546709,
286 [101] = FSCALE * 0.00640933344625638184,
287 [102] = FSCALE * 0.00609674656551563610,
288 [103] = FSCALE * 0.00579940472684214321,
289 [104] = FSCALE * 0.00551656442076077241,
290 [105] = FSCALE * 0.00524751839918138427,
291 [106] = FSCALE * 0.00499159390691021621,
292 [107] = FSCALE * 0.00474815099941147558,
293 [108] = FSCALE * 0.00451658094261266798,
294 [109] = FSCALE * 0.00429630469075234057,
295 [110] = FSCALE * 0.00408677143846406699,
299 #define CCPU_EXP_MAX 110
302 * This function is analogical to the getpcpu() function in the ps(1) command.
303 * They should both calculate in the same way so that the racct %cpu
304 * calculations are consistent with the values showed by the ps(1) tool.
305 * The calculations are more complex in the 4BSD scheduler because of the value
306 * of the ccpu variable. In ULE it is defined to be zero which saves us some
310 racct_getpcpu(struct proc *p, u_int pcpu)
314 fixpt_t pctcpu, pctcpu_next;
323 ASSERT_RACCT_ENABLED();
326 * If the process is swapped out, we count its %cpu usage as zero.
327 * This behaviour is consistent with the userland ps(1) tool.
329 if ((p->p_flag & P_INMEM) == 0)
331 swtime = (ticks - p->p_swtick) / hz;
334 * For short-lived processes, the sched_pctcpu() returns small
335 * values even for cpu intensive processes. Therefore we use
336 * our own estimate in this case.
338 if (swtime < RACCT_PCPU_SECS)
342 FOREACH_THREAD_IN_PROC(p, td) {
343 if (td == PCPU_GET(idlethread))
347 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
348 if (td == pc->pc_idlethread) {
358 pctcpu = sched_pctcpu(td);
359 /* Count also the yet unfinished second. */
360 pctcpu_next = (pctcpu * ccpu_exp[1]) >> FSHIFT;
361 pctcpu_next += sched_pctcpu_delta(td);
362 p_pctcpu += max(pctcpu, pctcpu_next);
365 * In ULE the %cpu statistics are updated on every
366 * sched_pctcpu() call. So special calculations to
367 * account for the latest (unfinished) second are
370 p_pctcpu += sched_pctcpu(td);
376 if (swtime <= CCPU_EXP_MAX)
377 return ((100 * (uint64_t)p_pctcpu * 1000000) /
378 (FSCALE - ccpu_exp[swtime]));
381 return ((100 * (uint64_t)p_pctcpu * 1000000) / FSCALE);
385 racct_add_racct(struct racct *dest, const struct racct *src)
389 ASSERT_RACCT_ENABLED();
390 mtx_assert(&racct_lock, MA_OWNED);
393 * Update resource usage in dest.
395 for (i = 0; i <= RACCT_MAX; i++) {
396 KASSERT(dest->r_resources[i] >= 0,
397 ("%s: resource %d propagation meltdown: dest < 0",
399 KASSERT(src->r_resources[i] >= 0,
400 ("%s: resource %d propagation meltdown: src < 0",
402 dest->r_resources[i] += src->r_resources[i];
407 racct_sub_racct(struct racct *dest, const struct racct *src)
411 ASSERT_RACCT_ENABLED();
412 mtx_assert(&racct_lock, MA_OWNED);
415 * Update resource usage in dest.
417 for (i = 0; i <= RACCT_MAX; i++) {
418 if (!RACCT_IS_SLOPPY(i) && !RACCT_IS_DECAYING(i)) {
419 KASSERT(dest->r_resources[i] >= 0,
420 ("%s: resource %d propagation meltdown: dest < 0",
422 KASSERT(src->r_resources[i] >= 0,
423 ("%s: resource %d propagation meltdown: src < 0",
425 KASSERT(src->r_resources[i] <= dest->r_resources[i],
426 ("%s: resource %d propagation meltdown: src > dest",
429 if (RACCT_CAN_DROP(i)) {
430 dest->r_resources[i] -= src->r_resources[i];
431 if (dest->r_resources[i] < 0) {
432 KASSERT(RACCT_IS_SLOPPY(i) ||
433 RACCT_IS_DECAYING(i),
434 ("%s: resource %d usage < 0", __func__, i));
435 dest->r_resources[i] = 0;
442 racct_create(struct racct **racctp)
448 SDT_PROBE(racct, kernel, racct, create, racctp, 0, 0, 0, 0);
450 KASSERT(*racctp == NULL, ("racct already allocated"));
452 *racctp = uma_zalloc(racct_zone, M_WAITOK | M_ZERO);
456 racct_destroy_locked(struct racct **racctp)
461 ASSERT_RACCT_ENABLED();
463 SDT_PROBE(racct, kernel, racct, destroy, racctp, 0, 0, 0, 0);
465 mtx_assert(&racct_lock, MA_OWNED);
466 KASSERT(racctp != NULL, ("NULL racctp"));
467 KASSERT(*racctp != NULL, ("NULL racct"));
471 for (i = 0; i <= RACCT_MAX; i++) {
472 if (RACCT_IS_SLOPPY(i))
474 if (!RACCT_IS_RECLAIMABLE(i))
476 KASSERT(racct->r_resources[i] == 0,
477 ("destroying non-empty racct: "
478 "%ju allocated for resource %d\n",
479 racct->r_resources[i], i));
481 uma_zfree(racct_zone, racct);
486 racct_destroy(struct racct **racct)
492 mtx_lock(&racct_lock);
493 racct_destroy_locked(racct);
494 mtx_unlock(&racct_lock);
498 * Increase consumption of 'resource' by 'amount' for 'racct'
499 * and all its parents. Differently from other cases, 'amount' here
500 * may be less than zero.
503 racct_adjust_resource(struct racct *racct, int resource,
507 ASSERT_RACCT_ENABLED();
508 mtx_assert(&racct_lock, MA_OWNED);
509 KASSERT(racct != NULL, ("NULL racct"));
511 racct->r_resources[resource] += amount;
512 if (racct->r_resources[resource] < 0) {
513 KASSERT(RACCT_IS_SLOPPY(resource) || RACCT_IS_DECAYING(resource),
514 ("%s: resource %d usage < 0", __func__, resource));
515 racct->r_resources[resource] = 0;
519 * There are some cases where the racct %cpu resource would grow
521 * For example in racct_proc_exit() we add the process %cpu usage
522 * to the ucred racct containers. If too many processes terminated
523 * in a short time span, the ucred %cpu resource could grow too much.
524 * Also, the 4BSD scheduler sometimes returns for a thread more than
525 * 100% cpu usage. So we set a boundary here to 100%.
527 if ((resource == RACCT_PCTCPU) &&
528 (racct->r_resources[RACCT_PCTCPU] > 100 * 1000000))
529 racct->r_resources[RACCT_PCTCPU] = 100 * 1000000;
533 racct_add_locked(struct proc *p, int resource, uint64_t amount)
539 ASSERT_RACCT_ENABLED();
541 SDT_PROBE(racct, kernel, rusage, add, p, resource, amount, 0, 0);
544 * We need proc lock to dereference p->p_ucred.
546 PROC_LOCK_ASSERT(p, MA_OWNED);
549 error = rctl_enforce(p, resource, amount);
550 if (error && RACCT_IS_DENIABLE(resource)) {
551 SDT_PROBE(racct, kernel, rusage, add__failure, p, resource,
556 racct_adjust_resource(p->p_racct, resource, amount);
557 racct_add_cred_locked(p->p_ucred, resource, amount);
563 * Increase allocation of 'resource' by 'amount' for process 'p'.
564 * Return 0 if it's below limits, or errno, if it's not.
567 racct_add(struct proc *p, int resource, uint64_t amount)
574 mtx_lock(&racct_lock);
575 error = racct_add_locked(p, resource, amount);
576 mtx_unlock(&racct_lock);
581 racct_add_cred_locked(struct ucred *cred, int resource, uint64_t amount)
585 ASSERT_RACCT_ENABLED();
587 SDT_PROBE(racct, kernel, rusage, add__cred, cred, resource, amount,
590 racct_adjust_resource(cred->cr_ruidinfo->ui_racct, resource, amount);
591 for (pr = cred->cr_prison; pr != NULL; pr = pr->pr_parent)
592 racct_adjust_resource(pr->pr_prison_racct->prr_racct, resource,
594 racct_adjust_resource(cred->cr_loginclass->lc_racct, resource, amount);
598 * Increase allocation of 'resource' by 'amount' for credential 'cred'.
599 * Doesn't check for limits and never fails.
601 * XXX: Shouldn't this ever return an error?
604 racct_add_cred(struct ucred *cred, int resource, uint64_t amount)
610 mtx_lock(&racct_lock);
611 racct_add_cred_locked(cred, resource, amount);
612 mtx_unlock(&racct_lock);
616 * Increase allocation of 'resource' by 'amount' for process 'p'.
617 * Doesn't check for limits and never fails.
620 racct_add_force(struct proc *p, int resource, uint64_t amount)
626 SDT_PROBE(racct, kernel, rusage, add__force, p, resource, amount, 0, 0);
629 * We need proc lock to dereference p->p_ucred.
631 PROC_LOCK_ASSERT(p, MA_OWNED);
633 mtx_lock(&racct_lock);
634 racct_adjust_resource(p->p_racct, resource, amount);
635 mtx_unlock(&racct_lock);
636 racct_add_cred(p->p_ucred, resource, amount);
640 racct_set_locked(struct proc *p, int resource, uint64_t amount)
642 int64_t old_amount, decayed_amount;
643 int64_t diff_proc, diff_cred;
648 ASSERT_RACCT_ENABLED();
650 SDT_PROBE(racct, kernel, rusage, set, p, resource, amount, 0, 0);
653 * We need proc lock to dereference p->p_ucred.
655 PROC_LOCK_ASSERT(p, MA_OWNED);
657 old_amount = p->p_racct->r_resources[resource];
659 * The diffs may be negative.
661 diff_proc = amount - old_amount;
662 if (RACCT_IS_DECAYING(resource)) {
664 * Resources in per-credential racct containers may decay.
665 * If this is the case, we need to calculate the difference
666 * between the new amount and the proportional value of the
667 * old amount that has decayed in the ucred racct containers.
669 decayed_amount = old_amount * RACCT_DECAY_FACTOR / FSCALE;
670 diff_cred = amount - decayed_amount;
672 diff_cred = diff_proc;
674 KASSERT(diff_proc >= 0 || RACCT_CAN_DROP(resource),
675 ("%s: usage of non-droppable resource %d dropping", __func__,
680 error = rctl_enforce(p, resource, diff_proc);
681 if (error && RACCT_IS_DENIABLE(resource)) {
682 SDT_PROBE(racct, kernel, rusage, set__failure, p,
683 resource, amount, 0, 0);
688 racct_adjust_resource(p->p_racct, resource, diff_proc);
690 racct_add_cred_locked(p->p_ucred, resource, diff_cred);
691 else if (diff_cred < 0)
692 racct_sub_cred_locked(p->p_ucred, resource, -diff_cred);
698 * Set allocation of 'resource' to 'amount' for process 'p'.
699 * Return 0 if it's below limits, or errno, if it's not.
701 * Note that decreasing the allocation always returns 0,
702 * even if it's above the limit.
705 racct_set(struct proc *p, int resource, uint64_t amount)
712 mtx_lock(&racct_lock);
713 error = racct_set_locked(p, resource, amount);
714 mtx_unlock(&racct_lock);
719 racct_set_force_locked(struct proc *p, int resource, uint64_t amount)
721 int64_t old_amount, decayed_amount;
722 int64_t diff_proc, diff_cred;
724 ASSERT_RACCT_ENABLED();
726 SDT_PROBE(racct, kernel, rusage, set, p, resource, amount, 0, 0);
729 * We need proc lock to dereference p->p_ucred.
731 PROC_LOCK_ASSERT(p, MA_OWNED);
733 old_amount = p->p_racct->r_resources[resource];
735 * The diffs may be negative.
737 diff_proc = amount - old_amount;
738 if (RACCT_IS_DECAYING(resource)) {
740 * Resources in per-credential racct containers may decay.
741 * If this is the case, we need to calculate the difference
742 * between the new amount and the proportional value of the
743 * old amount that has decayed in the ucred racct containers.
745 decayed_amount = old_amount * RACCT_DECAY_FACTOR / FSCALE;
746 diff_cred = amount - decayed_amount;
748 diff_cred = diff_proc;
750 racct_adjust_resource(p->p_racct, resource, diff_proc);
752 racct_add_cred_locked(p->p_ucred, resource, diff_cred);
753 else if (diff_cred < 0)
754 racct_sub_cred_locked(p->p_ucred, resource, -diff_cred);
758 racct_set_force(struct proc *p, int resource, uint64_t amount)
764 mtx_lock(&racct_lock);
765 racct_set_force_locked(p, resource, amount);
766 mtx_unlock(&racct_lock);
770 * Returns amount of 'resource' the process 'p' can keep allocated.
771 * Allocating more than that would be denied, unless the resource
772 * is marked undeniable. Amount of already allocated resource does
776 racct_get_limit(struct proc *p, int resource)
783 return (rctl_get_limit(p, resource));
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
796 racct_get_available(struct proc *p, int resource)
803 return (rctl_get_available(p, resource));
810 * Returns amount of the %cpu resource that process 'p' can add to its %cpu
811 * utilization. Adding more than that would lead to the process being
815 racct_pcpu_available(struct proc *p)
818 ASSERT_RACCT_ENABLED();
821 return (rctl_pcpu_available(p));
828 * Decrease allocation of 'resource' by 'amount' for process 'p'.
831 racct_sub(struct proc *p, int resource, uint64_t amount)
837 SDT_PROBE(racct, kernel, rusage, sub, p, resource, amount, 0, 0);
840 * We need proc lock to dereference p->p_ucred.
842 PROC_LOCK_ASSERT(p, MA_OWNED);
843 KASSERT(RACCT_CAN_DROP(resource),
844 ("%s: called for non-droppable resource %d", __func__, resource));
846 mtx_lock(&racct_lock);
847 KASSERT(amount <= p->p_racct->r_resources[resource],
848 ("%s: freeing %ju of resource %d, which is more "
849 "than allocated %jd for %s (pid %d)", __func__, amount, resource,
850 (intmax_t)p->p_racct->r_resources[resource], p->p_comm, p->p_pid));
852 racct_adjust_resource(p->p_racct, resource, -amount);
853 racct_sub_cred_locked(p->p_ucred, resource, amount);
854 mtx_unlock(&racct_lock);
858 racct_sub_cred_locked(struct ucred *cred, int resource, uint64_t amount)
862 ASSERT_RACCT_ENABLED();
864 SDT_PROBE(racct, kernel, rusage, sub__cred, cred, resource, amount,
868 KASSERT(RACCT_CAN_DROP(resource),
869 ("%s: called for resource %d which can not drop", __func__,
873 racct_adjust_resource(cred->cr_ruidinfo->ui_racct, resource, -amount);
874 for (pr = cred->cr_prison; pr != NULL; pr = pr->pr_parent)
875 racct_adjust_resource(pr->pr_prison_racct->prr_racct, resource,
877 racct_adjust_resource(cred->cr_loginclass->lc_racct, resource, -amount);
881 * Decrease allocation of 'resource' by 'amount' for credential 'cred'.
884 racct_sub_cred(struct ucred *cred, int resource, uint64_t amount)
890 mtx_lock(&racct_lock);
891 racct_sub_cred_locked(cred, resource, amount);
892 mtx_unlock(&racct_lock);
896 * Inherit resource usage information from the parent process.
899 racct_proc_fork(struct proc *parent, struct proc *child)
907 * Create racct for the child process.
909 racct_create(&child->p_racct);
913 mtx_lock(&racct_lock);
916 error = rctl_proc_fork(parent, child);
921 /* Init process cpu time. */
922 child->p_prev_runtime = 0;
923 child->p_throttled = 0;
926 * Inherit resource usage.
928 for (i = 0; i <= RACCT_MAX; i++) {
929 if (parent->p_racct->r_resources[i] == 0 ||
930 !RACCT_IS_INHERITABLE(i))
933 error = racct_set_locked(child, i,
934 parent->p_racct->r_resources[i]);
939 error = racct_add_locked(child, RACCT_NPROC, 1);
940 error += racct_add_locked(child, RACCT_NTHR, 1);
943 mtx_unlock(&racct_lock);
948 racct_proc_exit(child);
954 * Called at the end of fork1(), to handle rules that require the process
955 * to be fully initialized.
958 racct_proc_fork_done(struct proc *child)
966 mtx_lock(&racct_lock);
967 rctl_enforce(child, RACCT_NPROC, 0);
968 rctl_enforce(child, RACCT_NTHR, 0);
969 mtx_unlock(&racct_lock);
975 racct_proc_exit(struct proc *p)
979 struct timeval wallclock;
980 uint64_t pct_estimate, pct;
987 * We don't need to calculate rux, proc_reap() has already done this.
989 runtime = cputick2usec(p->p_rux.rux_runtime);
991 KASSERT(runtime >= p->p_prev_runtime, ("runtime < p_prev_runtime"));
993 if (runtime < p->p_prev_runtime)
994 runtime = p->p_prev_runtime;
996 microuptime(&wallclock);
997 timevalsub(&wallclock, &p->p_stats->p_start);
998 if (wallclock.tv_sec > 0 || wallclock.tv_usec > 0) {
999 pct_estimate = (1000000 * runtime * 100) /
1000 ((uint64_t)wallclock.tv_sec * 1000000 +
1004 pct = racct_getpcpu(p, pct_estimate);
1006 mtx_lock(&racct_lock);
1007 racct_set_locked(p, RACCT_CPU, runtime);
1008 racct_add_cred_locked(p->p_ucred, RACCT_PCTCPU, pct);
1010 for (i = 0; i <= RACCT_MAX; i++) {
1011 if (p->p_racct->r_resources[i] == 0)
1013 if (!RACCT_IS_RECLAIMABLE(i))
1015 racct_set_locked(p, i, 0);
1018 mtx_unlock(&racct_lock);
1022 rctl_racct_release(p->p_racct);
1024 racct_destroy(&p->p_racct);
1028 * Called after credentials change, to move resource utilisation
1032 racct_proc_ucred_changed(struct proc *p, struct ucred *oldcred,
1033 struct ucred *newcred)
1035 struct uidinfo *olduip, *newuip;
1036 struct loginclass *oldlc, *newlc;
1037 struct prison *oldpr, *newpr, *pr;
1042 PROC_LOCK_ASSERT(p, MA_NOTOWNED);
1044 newuip = newcred->cr_ruidinfo;
1045 olduip = oldcred->cr_ruidinfo;
1046 newlc = newcred->cr_loginclass;
1047 oldlc = oldcred->cr_loginclass;
1048 newpr = newcred->cr_prison;
1049 oldpr = oldcred->cr_prison;
1051 mtx_lock(&racct_lock);
1052 if (newuip != olduip) {
1053 racct_sub_racct(olduip->ui_racct, p->p_racct);
1054 racct_add_racct(newuip->ui_racct, p->p_racct);
1056 if (newlc != oldlc) {
1057 racct_sub_racct(oldlc->lc_racct, p->p_racct);
1058 racct_add_racct(newlc->lc_racct, p->p_racct);
1060 if (newpr != oldpr) {
1061 for (pr = oldpr; pr != NULL; pr = pr->pr_parent)
1062 racct_sub_racct(pr->pr_prison_racct->prr_racct,
1064 for (pr = newpr; pr != NULL; pr = pr->pr_parent)
1065 racct_add_racct(pr->pr_prison_racct->prr_racct,
1068 mtx_unlock(&racct_lock);
1071 rctl_proc_ucred_changed(p, newcred);
1076 racct_move(struct racct *dest, struct racct *src)
1079 ASSERT_RACCT_ENABLED();
1081 mtx_lock(&racct_lock);
1083 racct_add_racct(dest, src);
1084 racct_sub_racct(src, src);
1086 mtx_unlock(&racct_lock);
1090 racct_proc_throttle(struct proc *p)
1097 ASSERT_RACCT_ENABLED();
1098 PROC_LOCK_ASSERT(p, MA_OWNED);
1101 * Do not block kernel processes. Also do not block processes with
1102 * low %cpu utilization to improve interactivity.
1104 if (((p->p_flag & (P_SYSTEM | P_KTHREAD)) != 0) ||
1105 (p->p_racct->r_resources[RACCT_PCTCPU] <= pcpu_threshold))
1109 FOREACH_THREAD_IN_PROC(p, td) {
1111 switch (td->td_state) {
1114 * If the thread is on the scheduler run-queue, we can
1115 * not just remove it from there. So we set the flag
1116 * TDF_NEEDRESCHED for the thread, so that once it is
1117 * running, it is taken off the cpu as soon as possible.
1119 td->td_flags |= TDF_NEEDRESCHED;
1123 * If the thread is running, we request a context
1124 * switch for it by setting the TDF_NEEDRESCHED flag.
1126 td->td_flags |= TDF_NEEDRESCHED;
1128 cpuid = td->td_oncpu;
1129 if ((cpuid != NOCPU) && (td != curthread))
1130 ipi_cpu(cpuid, IPI_AST);
1141 racct_proc_wakeup(struct proc *p)
1144 ASSERT_RACCT_ENABLED();
1146 PROC_LOCK_ASSERT(p, MA_OWNED);
1148 if (p->p_throttled) {
1155 racct_decay_resource(struct racct *racct, void * res, void* dummy)
1158 int64_t r_old, r_new;
1160 ASSERT_RACCT_ENABLED();
1162 resource = *(int *)res;
1163 r_old = racct->r_resources[resource];
1165 /* If there is nothing to decay, just exit. */
1169 mtx_lock(&racct_lock);
1170 r_new = r_old * RACCT_DECAY_FACTOR / FSCALE;
1171 racct->r_resources[resource] = r_new;
1172 mtx_unlock(&racct_lock);
1176 racct_decay(int resource)
1179 ASSERT_RACCT_ENABLED();
1181 ui_racct_foreach(racct_decay_resource, &resource, NULL);
1182 loginclass_racct_foreach(racct_decay_resource, &resource, NULL);
1183 prison_racct_foreach(racct_decay_resource, &resource, NULL);
1191 struct timeval wallclock;
1193 uint64_t pct, pct_estimate;
1195 ASSERT_RACCT_ENABLED();
1198 racct_decay(RACCT_PCTCPU);
1200 sx_slock(&allproc_lock);
1202 LIST_FOREACH(p, &zombproc, p_list) {
1204 racct_set(p, RACCT_PCTCPU, 0);
1208 FOREACH_PROC_IN_SYSTEM(p) {
1210 if (p->p_state != PRS_NORMAL) {
1215 microuptime(&wallclock);
1216 timevalsub(&wallclock, &p->p_stats->p_start);
1218 FOREACH_THREAD_IN_PROC(p, td)
1220 runtime = cputick2usec(p->p_rux.rux_runtime);
1223 KASSERT(runtime >= p->p_prev_runtime,
1224 ("runtime < p_prev_runtime"));
1226 if (runtime < p->p_prev_runtime)
1227 runtime = p->p_prev_runtime;
1229 p->p_prev_runtime = runtime;
1230 if (wallclock.tv_sec > 0 || wallclock.tv_usec > 0) {
1231 pct_estimate = (1000000 * runtime * 100) /
1232 ((uint64_t)wallclock.tv_sec * 1000000 +
1236 pct = racct_getpcpu(p, pct_estimate);
1237 mtx_lock(&racct_lock);
1238 racct_set_force_locked(p, RACCT_PCTCPU, pct);
1239 racct_set_locked(p, RACCT_CPU, runtime);
1240 racct_set_locked(p, RACCT_WALLCLOCK,
1241 (uint64_t)wallclock.tv_sec * 1000000 +
1243 mtx_unlock(&racct_lock);
1248 * To ensure that processes are throttled in a fair way, we need
1249 * to iterate over all processes again and check the limits
1250 * for %cpu resource only after ucred racct containers have been
1253 FOREACH_PROC_IN_SYSTEM(p) {
1255 if (p->p_state != PRS_NORMAL) {
1260 if (racct_pcpu_available(p) <= 0)
1261 racct_proc_throttle(p);
1262 else if (p->p_throttled)
1263 racct_proc_wakeup(p);
1266 sx_sunlock(&allproc_lock);
1271 static struct kproc_desc racctd_kp = {
1283 kproc_start(&racctd_kp);
1285 SYSINIT(racctd, SI_SUB_RACCTD, SI_ORDER_FIRST, racctd_init, NULL);
1293 racct_zone = uma_zcreate("racct", sizeof(struct racct),
1294 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
1296 * XXX: Move this somewhere.
1298 prison0.pr_prison_racct = prison_racct_find("0");
1300 SYSINIT(racct, SI_SUB_RACCT, SI_ORDER_FIRST, racct_init, NULL);