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_kdtrace.h"
36 #include "opt_sched.h"
38 #include <sys/param.h>
39 #include <sys/systm.h>
40 #include <sys/eventhandler.h>
42 #include <sys/kernel.h>
43 #include <sys/kthread.h>
45 #include <sys/loginclass.h>
46 #include <sys/malloc.h>
47 #include <sys/mutex.h>
49 #include <sys/racct.h>
50 #include <sys/resourcevar.h>
52 #include <sys/sched.h>
56 #include <sys/sysctl.h>
57 #include <sys/sysent.h>
58 #include <sys/sysproto.h>
60 #include <machine/smp.h>
68 FEATURE(racct, "Resource Accounting");
71 * Do not block processes that have their %cpu usage <= pcpu_threshold.
73 static int pcpu_threshold = 1;
74 #ifdef RACCT_DEFAULT_TO_DISABLED
80 SYSCTL_NODE(_kern, OID_AUTO, racct, CTLFLAG_RW, 0, "Resource Accounting");
81 TUNABLE_INT("kern.racct.enable", &racct_enable);
82 SYSCTL_UINT(_kern_racct, OID_AUTO, enable, CTLFLAG_RDTUN, &racct_enable,
83 0, "Enable RACCT/RCTL");
84 SYSCTL_UINT(_kern_racct, OID_AUTO, pcpu_threshold, CTLFLAG_RW, &pcpu_threshold,
85 0, "Processes with higher %cpu usage than this value can be throttled.");
88 * How many seconds it takes to use the scheduler %cpu calculations. When a
89 * process starts, we compute its %cpu usage by dividing its runtime by the
90 * process wall clock time. After RACCT_PCPU_SECS pass, we use the value
91 * provided by the scheduler.
93 #define RACCT_PCPU_SECS 3
95 static struct mtx racct_lock;
96 MTX_SYSINIT(racct_lock, &racct_lock, "racct lock", MTX_DEF);
98 static uma_zone_t racct_zone;
100 static void racct_sub_racct(struct racct *dest, const struct racct *src);
101 static void racct_sub_cred_locked(struct ucred *cred, int resource,
103 static void racct_add_cred_locked(struct ucred *cred, int resource,
106 SDT_PROVIDER_DEFINE(racct);
107 SDT_PROBE_DEFINE3(racct, kernel, rusage, add, "struct proc *", "int",
109 SDT_PROBE_DEFINE3(racct, kernel, rusage, add__failure,
110 "struct proc *", "int", "uint64_t");
111 SDT_PROBE_DEFINE3(racct, kernel, rusage, add__cred, "struct ucred *",
113 SDT_PROBE_DEFINE3(racct, kernel, rusage, add__force, "struct proc *",
115 SDT_PROBE_DEFINE3(racct, kernel, rusage, set, "struct proc *", "int",
117 SDT_PROBE_DEFINE3(racct, kernel, rusage, set__failure,
118 "struct proc *", "int", "uint64_t");
119 SDT_PROBE_DEFINE3(racct, kernel, rusage, sub, "struct proc *", "int",
121 SDT_PROBE_DEFINE3(racct, kernel, rusage, sub__cred, "struct ucred *",
123 SDT_PROBE_DEFINE1(racct, kernel, racct, create, "struct racct *");
124 SDT_PROBE_DEFINE1(racct, kernel, racct, destroy, "struct racct *");
125 SDT_PROBE_DEFINE2(racct, kernel, racct, join, "struct racct *",
127 SDT_PROBE_DEFINE2(racct, kernel, racct, join__failure,
128 "struct racct *", "struct racct *");
129 SDT_PROBE_DEFINE2(racct, kernel, racct, leave, "struct racct *",
132 int racct_types[] = {
136 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
138 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
144 RACCT_RECLAIMABLE | RACCT_DENIABLE,
146 RACCT_RECLAIMABLE | RACCT_DENIABLE,
148 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
150 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
152 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
154 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
156 RACCT_RECLAIMABLE | RACCT_DENIABLE,
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_DENIABLE | RACCT_SLOPPY,
166 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
168 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
170 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
174 RACCT_DECAYING | RACCT_DENIABLE | RACCT_IN_MILLIONS };
176 static const fixpt_t RACCT_DECAY_FACTOR = 0.3 * FSCALE;
180 * Contains intermediate values for %cpu calculations to avoid using floating
181 * point in the kernel.
182 * ccpu_exp[k] = FSCALE * (ccpu/FSCALE)^k = FSCALE * exp(-k/20)
183 * It is needed only for the 4BSD scheduler, because in ULE, the ccpu equals to
184 * zero so the calculations are more straightforward.
186 fixpt_t ccpu_exp[] = {
188 [1] = FSCALE * 0.95122942450071400909,
189 [2] = FSCALE * 0.90483741803595957316,
190 [3] = FSCALE * 0.86070797642505780722,
191 [4] = FSCALE * 0.81873075307798185866,
192 [5] = FSCALE * 0.77880078307140486824,
193 [6] = FSCALE * 0.74081822068171786606,
194 [7] = FSCALE * 0.70468808971871343435,
195 [8] = FSCALE * 0.67032004603563930074,
196 [9] = FSCALE * 0.63762815162177329314,
197 [10] = FSCALE * 0.60653065971263342360,
198 [11] = FSCALE * 0.57694981038048669531,
199 [12] = FSCALE * 0.54881163609402643262,
200 [13] = FSCALE * 0.52204577676101604789,
201 [14] = FSCALE * 0.49658530379140951470,
202 [15] = FSCALE * 0.47236655274101470713,
203 [16] = FSCALE * 0.44932896411722159143,
204 [17] = FSCALE * 0.42741493194872666992,
205 [18] = FSCALE * 0.40656965974059911188,
206 [19] = FSCALE * 0.38674102345450120691,
207 [20] = FSCALE * 0.36787944117144232159,
208 [21] = FSCALE * 0.34993774911115535467,
209 [22] = FSCALE * 0.33287108369807955328,
210 [23] = FSCALE * 0.31663676937905321821,
211 [24] = FSCALE * 0.30119421191220209664,
212 [25] = FSCALE * 0.28650479686019010032,
213 [26] = FSCALE * 0.27253179303401260312,
214 [27] = FSCALE * 0.25924026064589150757,
215 [28] = FSCALE * 0.24659696394160647693,
216 [29] = FSCALE * 0.23457028809379765313,
217 [30] = FSCALE * 0.22313016014842982893,
218 [31] = FSCALE * 0.21224797382674305771,
219 [32] = FSCALE * 0.20189651799465540848,
220 [33] = FSCALE * 0.19204990862075411423,
221 [34] = FSCALE * 0.18268352405273465022,
222 [35] = FSCALE * 0.17377394345044512668,
223 [36] = FSCALE * 0.16529888822158653829,
224 [37] = FSCALE * 0.15723716631362761621,
225 [38] = FSCALE * 0.14956861922263505264,
226 [39] = FSCALE * 0.14227407158651357185,
227 [40] = FSCALE * 0.13533528323661269189,
228 [41] = FSCALE * 0.12873490358780421886,
229 [42] = FSCALE * 0.12245642825298191021,
230 [43] = FSCALE * 0.11648415777349695786,
231 [44] = FSCALE * 0.11080315836233388333,
232 [45] = FSCALE * 0.10539922456186433678,
233 [46] = FSCALE * 0.10025884372280373372,
234 [47] = FSCALE * 0.09536916221554961888,
235 [48] = FSCALE * 0.09071795328941250337,
236 [49] = FSCALE * 0.08629358649937051097,
237 [50] = FSCALE * 0.08208499862389879516,
238 [51] = FSCALE * 0.07808166600115315231,
239 [52] = FSCALE * 0.07427357821433388042,
240 [53] = FSCALE * 0.07065121306042958674,
241 [54] = FSCALE * 0.06720551273974976512,
242 [55] = FSCALE * 0.06392786120670757270,
243 [56] = FSCALE * 0.06081006262521796499,
244 [57] = FSCALE * 0.05784432087483846296,
245 [58] = FSCALE * 0.05502322005640722902,
246 [59] = FSCALE * 0.05233970594843239308,
247 [60] = FSCALE * 0.04978706836786394297,
248 [61] = FSCALE * 0.04735892439114092119,
249 [62] = FSCALE * 0.04504920239355780606,
250 [63] = FSCALE * 0.04285212686704017991,
251 [64] = FSCALE * 0.04076220397836621516,
252 [65] = FSCALE * 0.03877420783172200988,
253 [66] = FSCALE * 0.03688316740124000544,
254 [67] = FSCALE * 0.03508435410084502588,
255 [68] = FSCALE * 0.03337326996032607948,
256 [69] = FSCALE * 0.03174563637806794323,
257 [70] = FSCALE * 0.03019738342231850073,
258 [71] = FSCALE * 0.02872463965423942912,
259 [72] = FSCALE * 0.02732372244729256080,
260 [73] = FSCALE * 0.02599112877875534358,
261 [74] = FSCALE * 0.02472352647033939120,
262 [75] = FSCALE * 0.02351774585600910823,
263 [76] = FSCALE * 0.02237077185616559577,
264 [77] = FSCALE * 0.02127973643837716938,
265 [78] = FSCALE * 0.02024191144580438847,
266 [79] = FSCALE * 0.01925470177538692429,
267 [80] = FSCALE * 0.01831563888873418029,
268 [81] = FSCALE * 0.01742237463949351138,
269 [82] = FSCALE * 0.01657267540176124754,
270 [83] = FSCALE * 0.01576441648485449082,
271 [84] = FSCALE * 0.01499557682047770621,
272 [85] = FSCALE * 0.01426423390899925527,
273 [86] = FSCALE * 0.01356855901220093175,
274 [87] = FSCALE * 0.01290681258047986886,
275 [88] = FSCALE * 0.01227733990306844117,
276 [89] = FSCALE * 0.01167856697039544521,
277 [90] = FSCALE * 0.01110899653824230649,
278 [91] = FSCALE * 0.01056720438385265337,
279 [92] = FSCALE * 0.01005183574463358164,
280 [93] = FSCALE * 0.00956160193054350793,
281 [94] = FSCALE * 0.00909527710169581709,
282 [95] = FSCALE * 0.00865169520312063417,
283 [96] = FSCALE * 0.00822974704902002884,
284 [97] = FSCALE * 0.00782837754922577143,
285 [98] = FSCALE * 0.00744658307092434051,
286 [99] = FSCALE * 0.00708340892905212004,
287 [100] = FSCALE * 0.00673794699908546709,
288 [101] = FSCALE * 0.00640933344625638184,
289 [102] = FSCALE * 0.00609674656551563610,
290 [103] = FSCALE * 0.00579940472684214321,
291 [104] = FSCALE * 0.00551656442076077241,
292 [105] = FSCALE * 0.00524751839918138427,
293 [106] = FSCALE * 0.00499159390691021621,
294 [107] = FSCALE * 0.00474815099941147558,
295 [108] = FSCALE * 0.00451658094261266798,
296 [109] = FSCALE * 0.00429630469075234057,
297 [110] = FSCALE * 0.00408677143846406699,
301 #define CCPU_EXP_MAX 110
304 * This function is analogical to the getpcpu() function in the ps(1) command.
305 * They should both calculate in the same way so that the racct %cpu
306 * calculations are consistent with the values showed by the ps(1) tool.
307 * The calculations are more complex in the 4BSD scheduler because of the value
308 * of the ccpu variable. In ULE it is defined to be zero which saves us some
312 racct_getpcpu(struct proc *p, u_int pcpu)
316 fixpt_t pctcpu, pctcpu_next;
325 ASSERT_RACCT_ENABLED();
328 * If the process is swapped out, we count its %cpu usage as zero.
329 * This behaviour is consistent with the userland ps(1) tool.
331 if ((p->p_flag & P_INMEM) == 0)
333 swtime = (ticks - p->p_swtick) / hz;
336 * For short-lived processes, the sched_pctcpu() returns small
337 * values even for cpu intensive processes. Therefore we use
338 * our own estimate in this case.
340 if (swtime < RACCT_PCPU_SECS)
344 FOREACH_THREAD_IN_PROC(p, td) {
345 if (td == PCPU_GET(idlethread))
349 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
350 if (td == pc->pc_idlethread) {
360 pctcpu = sched_pctcpu(td);
361 /* Count also the yet unfinished second. */
362 pctcpu_next = (pctcpu * ccpu_exp[1]) >> FSHIFT;
363 pctcpu_next += sched_pctcpu_delta(td);
364 p_pctcpu += max(pctcpu, pctcpu_next);
367 * In ULE the %cpu statistics are updated on every
368 * sched_pctcpu() call. So special calculations to
369 * account for the latest (unfinished) second are
372 p_pctcpu += sched_pctcpu(td);
378 if (swtime <= CCPU_EXP_MAX)
379 return ((100 * (uint64_t)p_pctcpu * 1000000) /
380 (FSCALE - ccpu_exp[swtime]));
383 return ((100 * (uint64_t)p_pctcpu * 1000000) / FSCALE);
387 racct_add_racct(struct racct *dest, const struct racct *src)
391 ASSERT_RACCT_ENABLED();
392 mtx_assert(&racct_lock, MA_OWNED);
395 * Update resource usage in dest.
397 for (i = 0; i <= RACCT_MAX; i++) {
398 KASSERT(dest->r_resources[i] >= 0,
399 ("%s: resource %d propagation meltdown: dest < 0",
401 KASSERT(src->r_resources[i] >= 0,
402 ("%s: resource %d propagation meltdown: src < 0",
404 dest->r_resources[i] += src->r_resources[i];
409 racct_sub_racct(struct racct *dest, const struct racct *src)
413 ASSERT_RACCT_ENABLED();
414 mtx_assert(&racct_lock, MA_OWNED);
417 * Update resource usage in dest.
419 for (i = 0; i <= RACCT_MAX; i++) {
420 if (!RACCT_IS_SLOPPY(i) && !RACCT_IS_DECAYING(i)) {
421 KASSERT(dest->r_resources[i] >= 0,
422 ("%s: resource %d propagation meltdown: dest < 0",
424 KASSERT(src->r_resources[i] >= 0,
425 ("%s: resource %d propagation meltdown: src < 0",
427 KASSERT(src->r_resources[i] <= dest->r_resources[i],
428 ("%s: resource %d propagation meltdown: src > dest",
431 if (RACCT_CAN_DROP(i)) {
432 dest->r_resources[i] -= src->r_resources[i];
433 if (dest->r_resources[i] < 0) {
434 KASSERT(RACCT_IS_SLOPPY(i) ||
435 RACCT_IS_DECAYING(i),
436 ("%s: resource %d usage < 0", __func__, i));
437 dest->r_resources[i] = 0;
444 racct_create(struct racct **racctp)
450 SDT_PROBE(racct, kernel, racct, create, racctp, 0, 0, 0, 0);
452 KASSERT(*racctp == NULL, ("racct already allocated"));
454 *racctp = uma_zalloc(racct_zone, M_WAITOK | M_ZERO);
458 racct_destroy_locked(struct racct **racctp)
463 ASSERT_RACCT_ENABLED();
465 SDT_PROBE(racct, kernel, racct, destroy, racctp, 0, 0, 0, 0);
467 mtx_assert(&racct_lock, MA_OWNED);
468 KASSERT(racctp != NULL, ("NULL racctp"));
469 KASSERT(*racctp != NULL, ("NULL racct"));
473 for (i = 0; i <= RACCT_MAX; i++) {
474 if (RACCT_IS_SLOPPY(i))
476 if (!RACCT_IS_RECLAIMABLE(i))
478 KASSERT(racct->r_resources[i] == 0,
479 ("destroying non-empty racct: "
480 "%ju allocated for resource %d\n",
481 racct->r_resources[i], i));
483 uma_zfree(racct_zone, racct);
488 racct_destroy(struct racct **racct)
494 mtx_lock(&racct_lock);
495 racct_destroy_locked(racct);
496 mtx_unlock(&racct_lock);
500 * Increase consumption of 'resource' by 'amount' for 'racct'
501 * and all its parents. Differently from other cases, 'amount' here
502 * may be less than zero.
505 racct_alloc_resource(struct racct *racct, int resource,
509 ASSERT_RACCT_ENABLED();
510 mtx_assert(&racct_lock, MA_OWNED);
511 KASSERT(racct != NULL, ("NULL racct"));
513 racct->r_resources[resource] += amount;
514 if (racct->r_resources[resource] < 0) {
515 KASSERT(RACCT_IS_SLOPPY(resource) || RACCT_IS_DECAYING(resource),
516 ("%s: resource %d usage < 0", __func__, resource));
517 racct->r_resources[resource] = 0;
521 * There are some cases where the racct %cpu resource would grow
523 * For example in racct_proc_exit() we add the process %cpu usage
524 * to the ucred racct containers. If too many processes terminated
525 * in a short time span, the ucred %cpu resource could grow too much.
526 * Also, the 4BSD scheduler sometimes returns for a thread more than
527 * 100% cpu usage. So we set a boundary here to 100%.
529 if ((resource == RACCT_PCTCPU) &&
530 (racct->r_resources[RACCT_PCTCPU] > 100 * 1000000))
531 racct->r_resources[RACCT_PCTCPU] = 100 * 1000000;
535 racct_add_locked(struct proc *p, int resource, uint64_t amount)
541 ASSERT_RACCT_ENABLED();
543 SDT_PROBE(racct, kernel, rusage, add, p, resource, amount, 0, 0);
546 * We need proc lock to dereference p->p_ucred.
548 PROC_LOCK_ASSERT(p, MA_OWNED);
551 error = rctl_enforce(p, resource, amount);
552 if (error && RACCT_IS_DENIABLE(resource)) {
553 SDT_PROBE(racct, kernel, rusage, add__failure, p, resource,
558 racct_alloc_resource(p->p_racct, resource, amount);
559 racct_add_cred_locked(p->p_ucred, resource, amount);
565 * Increase allocation of 'resource' by 'amount' for process 'p'.
566 * Return 0 if it's below limits, or errno, if it's not.
569 racct_add(struct proc *p, int resource, uint64_t amount)
576 mtx_lock(&racct_lock);
577 error = racct_add_locked(p, resource, amount);
578 mtx_unlock(&racct_lock);
583 racct_add_cred_locked(struct ucred *cred, int resource, uint64_t amount)
587 ASSERT_RACCT_ENABLED();
589 SDT_PROBE(racct, kernel, rusage, add__cred, cred, resource, amount,
592 racct_alloc_resource(cred->cr_ruidinfo->ui_racct, resource, amount);
593 for (pr = cred->cr_prison; pr != NULL; pr = pr->pr_parent)
594 racct_alloc_resource(pr->pr_prison_racct->prr_racct, resource,
596 racct_alloc_resource(cred->cr_loginclass->lc_racct, resource, amount);
600 * Increase allocation of 'resource' by 'amount' for credential 'cred'.
601 * Doesn't check for limits and never fails.
603 * XXX: Shouldn't this ever return an error?
606 racct_add_cred(struct ucred *cred, int resource, uint64_t amount)
612 mtx_lock(&racct_lock);
613 racct_add_cred_locked(cred, resource, amount);
614 mtx_unlock(&racct_lock);
618 * Increase allocation of 'resource' by 'amount' for process 'p'.
619 * Doesn't check for limits and never fails.
622 racct_add_force(struct proc *p, int resource, uint64_t amount)
628 SDT_PROBE(racct, kernel, rusage, add__force, p, resource, amount, 0, 0);
631 * We need proc lock to dereference p->p_ucred.
633 PROC_LOCK_ASSERT(p, MA_OWNED);
635 mtx_lock(&racct_lock);
636 racct_alloc_resource(p->p_racct, resource, amount);
637 mtx_unlock(&racct_lock);
638 racct_add_cred(p->p_ucred, resource, amount);
642 racct_set_locked(struct proc *p, int resource, uint64_t amount)
644 int64_t old_amount, decayed_amount;
645 int64_t diff_proc, diff_cred;
650 ASSERT_RACCT_ENABLED();
652 SDT_PROBE(racct, kernel, rusage, set, p, resource, amount, 0, 0);
655 * We need proc lock to dereference p->p_ucred.
657 PROC_LOCK_ASSERT(p, MA_OWNED);
659 old_amount = p->p_racct->r_resources[resource];
661 * The diffs may be negative.
663 diff_proc = amount - old_amount;
664 if (RACCT_IS_DECAYING(resource)) {
666 * Resources in per-credential racct containers may decay.
667 * If this is the case, we need to calculate the difference
668 * between the new amount and the proportional value of the
669 * old amount that has decayed in the ucred racct containers.
671 decayed_amount = old_amount * RACCT_DECAY_FACTOR / FSCALE;
672 diff_cred = amount - decayed_amount;
674 diff_cred = diff_proc;
676 KASSERT(diff_proc >= 0 || RACCT_CAN_DROP(resource),
677 ("%s: usage of non-droppable resource %d dropping", __func__,
682 error = rctl_enforce(p, resource, diff_proc);
683 if (error && RACCT_IS_DENIABLE(resource)) {
684 SDT_PROBE(racct, kernel, rusage, set__failure, p,
685 resource, amount, 0, 0);
690 racct_alloc_resource(p->p_racct, resource, diff_proc);
692 racct_add_cred_locked(p->p_ucred, resource, diff_cred);
693 else if (diff_cred < 0)
694 racct_sub_cred_locked(p->p_ucred, resource, -diff_cred);
700 * Set allocation of 'resource' to 'amount' for process 'p'.
701 * Return 0 if it's below limits, or errno, if it's not.
703 * Note that decreasing the allocation always returns 0,
704 * even if it's above the limit.
707 racct_set(struct proc *p, int resource, uint64_t amount)
714 mtx_lock(&racct_lock);
715 error = racct_set_locked(p, resource, amount);
716 mtx_unlock(&racct_lock);
721 racct_set_force_locked(struct proc *p, int resource, uint64_t amount)
723 int64_t old_amount, decayed_amount;
724 int64_t diff_proc, diff_cred;
726 ASSERT_RACCT_ENABLED();
728 SDT_PROBE(racct, kernel, rusage, set, p, resource, amount, 0, 0);
731 * We need proc lock to dereference p->p_ucred.
733 PROC_LOCK_ASSERT(p, MA_OWNED);
735 old_amount = p->p_racct->r_resources[resource];
737 * The diffs may be negative.
739 diff_proc = amount - old_amount;
740 if (RACCT_IS_DECAYING(resource)) {
742 * Resources in per-credential racct containers may decay.
743 * If this is the case, we need to calculate the difference
744 * between the new amount and the proportional value of the
745 * old amount that has decayed in the ucred racct containers.
747 decayed_amount = old_amount * RACCT_DECAY_FACTOR / FSCALE;
748 diff_cred = amount - decayed_amount;
750 diff_cred = diff_proc;
752 racct_alloc_resource(p->p_racct, resource, diff_proc);
754 racct_add_cred_locked(p->p_ucred, resource, diff_cred);
755 else if (diff_cred < 0)
756 racct_sub_cred_locked(p->p_ucred, resource, -diff_cred);
760 racct_set_force(struct proc *p, int resource, uint64_t amount)
766 mtx_lock(&racct_lock);
767 racct_set_force_locked(p, resource, amount);
768 mtx_unlock(&racct_lock);
772 * Returns amount of 'resource' the process 'p' can keep allocated.
773 * Allocating more than that would be denied, unless the resource
774 * is marked undeniable. Amount of already allocated resource does
778 racct_get_limit(struct proc *p, int resource)
785 return (rctl_get_limit(p, resource));
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
798 racct_get_available(struct proc *p, int resource)
805 return (rctl_get_available(p, resource));
812 * Returns amount of the %cpu resource that process 'p' can add to its %cpu
813 * utilization. Adding more than that would lead to the process being
817 racct_pcpu_available(struct proc *p)
820 ASSERT_RACCT_ENABLED();
823 return (rctl_pcpu_available(p));
830 * Decrease allocation of 'resource' by 'amount' for process 'p'.
833 racct_sub(struct proc *p, int resource, uint64_t amount)
839 SDT_PROBE(racct, kernel, rusage, sub, p, resource, amount, 0, 0);
842 * We need proc lock to dereference p->p_ucred.
844 PROC_LOCK_ASSERT(p, MA_OWNED);
845 KASSERT(RACCT_CAN_DROP(resource),
846 ("%s: called for non-droppable resource %d", __func__, resource));
848 mtx_lock(&racct_lock);
849 KASSERT(amount <= p->p_racct->r_resources[resource],
850 ("%s: freeing %ju of resource %d, which is more "
851 "than allocated %jd for %s (pid %d)", __func__, amount, resource,
852 (intmax_t)p->p_racct->r_resources[resource], p->p_comm, p->p_pid));
854 racct_alloc_resource(p->p_racct, resource, -amount);
855 racct_sub_cred_locked(p->p_ucred, resource, amount);
856 mtx_unlock(&racct_lock);
860 racct_sub_cred_locked(struct ucred *cred, int resource, uint64_t amount)
864 ASSERT_RACCT_ENABLED();
866 SDT_PROBE(racct, kernel, rusage, sub__cred, cred, resource, amount,
870 KASSERT(RACCT_CAN_DROP(resource),
871 ("%s: called for resource %d which can not drop", __func__,
875 racct_alloc_resource(cred->cr_ruidinfo->ui_racct, resource, -amount);
876 for (pr = cred->cr_prison; pr != NULL; pr = pr->pr_parent)
877 racct_alloc_resource(pr->pr_prison_racct->prr_racct, resource,
879 racct_alloc_resource(cred->cr_loginclass->lc_racct, resource, -amount);
883 * Decrease allocation of 'resource' by 'amount' for credential 'cred'.
886 racct_sub_cred(struct ucred *cred, int resource, uint64_t amount)
892 mtx_lock(&racct_lock);
893 racct_sub_cred_locked(cred, resource, amount);
894 mtx_unlock(&racct_lock);
898 * Inherit resource usage information from the parent process.
901 racct_proc_fork(struct proc *parent, struct proc *child)
909 * Create racct for the child process.
911 racct_create(&child->p_racct);
915 mtx_lock(&racct_lock);
918 error = rctl_proc_fork(parent, child);
923 /* Init process cpu time. */
924 child->p_prev_runtime = 0;
925 child->p_throttled = 0;
928 * Inherit resource usage.
930 for (i = 0; i <= RACCT_MAX; i++) {
931 if (parent->p_racct->r_resources[i] == 0 ||
932 !RACCT_IS_INHERITABLE(i))
935 error = racct_set_locked(child, i,
936 parent->p_racct->r_resources[i]);
941 error = racct_add_locked(child, RACCT_NPROC, 1);
942 error += racct_add_locked(child, RACCT_NTHR, 1);
945 mtx_unlock(&racct_lock);
950 racct_proc_exit(child);
956 * Called at the end of fork1(), to handle rules that require the process
957 * to be fully initialized.
960 racct_proc_fork_done(struct proc *child)
968 mtx_lock(&racct_lock);
969 rctl_enforce(child, RACCT_NPROC, 0);
970 rctl_enforce(child, RACCT_NTHR, 0);
971 mtx_unlock(&racct_lock);
977 racct_proc_exit(struct proc *p)
981 struct timeval wallclock;
982 uint64_t pct_estimate, pct;
989 * We don't need to calculate rux, proc_reap() has already done this.
991 runtime = cputick2usec(p->p_rux.rux_runtime);
993 KASSERT(runtime >= p->p_prev_runtime, ("runtime < p_prev_runtime"));
995 if (runtime < p->p_prev_runtime)
996 runtime = p->p_prev_runtime;
998 microuptime(&wallclock);
999 timevalsub(&wallclock, &p->p_stats->p_start);
1000 if (wallclock.tv_sec > 0 || wallclock.tv_usec > 0) {
1001 pct_estimate = (1000000 * runtime * 100) /
1002 ((uint64_t)wallclock.tv_sec * 1000000 +
1006 pct = racct_getpcpu(p, pct_estimate);
1008 mtx_lock(&racct_lock);
1009 racct_set_locked(p, RACCT_CPU, runtime);
1010 racct_add_cred_locked(p->p_ucred, RACCT_PCTCPU, pct);
1012 for (i = 0; i <= RACCT_MAX; i++) {
1013 if (p->p_racct->r_resources[i] == 0)
1015 if (!RACCT_IS_RECLAIMABLE(i))
1017 racct_set_locked(p, i, 0);
1020 mtx_unlock(&racct_lock);
1024 rctl_racct_release(p->p_racct);
1026 racct_destroy(&p->p_racct);
1030 * Called after credentials change, to move resource utilisation
1034 racct_proc_ucred_changed(struct proc *p, struct ucred *oldcred,
1035 struct ucred *newcred)
1037 struct uidinfo *olduip, *newuip;
1038 struct loginclass *oldlc, *newlc;
1039 struct prison *oldpr, *newpr, *pr;
1044 PROC_LOCK_ASSERT(p, MA_NOTOWNED);
1046 newuip = newcred->cr_ruidinfo;
1047 olduip = oldcred->cr_ruidinfo;
1048 newlc = newcred->cr_loginclass;
1049 oldlc = oldcred->cr_loginclass;
1050 newpr = newcred->cr_prison;
1051 oldpr = oldcred->cr_prison;
1053 mtx_lock(&racct_lock);
1054 if (newuip != olduip) {
1055 racct_sub_racct(olduip->ui_racct, p->p_racct);
1056 racct_add_racct(newuip->ui_racct, p->p_racct);
1058 if (newlc != oldlc) {
1059 racct_sub_racct(oldlc->lc_racct, p->p_racct);
1060 racct_add_racct(newlc->lc_racct, p->p_racct);
1062 if (newpr != oldpr) {
1063 for (pr = oldpr; pr != NULL; pr = pr->pr_parent)
1064 racct_sub_racct(pr->pr_prison_racct->prr_racct,
1066 for (pr = newpr; pr != NULL; pr = pr->pr_parent)
1067 racct_add_racct(pr->pr_prison_racct->prr_racct,
1070 mtx_unlock(&racct_lock);
1073 rctl_proc_ucred_changed(p, newcred);
1078 racct_move(struct racct *dest, struct racct *src)
1081 ASSERT_RACCT_ENABLED();
1083 mtx_lock(&racct_lock);
1085 racct_add_racct(dest, src);
1086 racct_sub_racct(src, src);
1088 mtx_unlock(&racct_lock);
1092 racct_proc_throttle(struct proc *p)
1099 ASSERT_RACCT_ENABLED();
1100 PROC_LOCK_ASSERT(p, MA_OWNED);
1103 * Do not block kernel processes. Also do not block processes with
1104 * low %cpu utilization to improve interactivity.
1106 if (((p->p_flag & (P_SYSTEM | P_KTHREAD)) != 0) ||
1107 (p->p_racct->r_resources[RACCT_PCTCPU] <= pcpu_threshold))
1111 FOREACH_THREAD_IN_PROC(p, td) {
1113 switch (td->td_state) {
1116 * If the thread is on the scheduler run-queue, we can
1117 * not just remove it from there. So we set the flag
1118 * TDF_NEEDRESCHED for the thread, so that once it is
1119 * running, it is taken off the cpu as soon as possible.
1121 td->td_flags |= TDF_NEEDRESCHED;
1125 * If the thread is running, we request a context
1126 * switch for it by setting the TDF_NEEDRESCHED flag.
1128 td->td_flags |= TDF_NEEDRESCHED;
1130 cpuid = td->td_oncpu;
1131 if ((cpuid != NOCPU) && (td != curthread))
1132 ipi_cpu(cpuid, IPI_AST);
1143 racct_proc_wakeup(struct proc *p)
1146 ASSERT_RACCT_ENABLED();
1148 PROC_LOCK_ASSERT(p, MA_OWNED);
1150 if (p->p_throttled) {
1157 racct_decay_resource(struct racct *racct, void * res, void* dummy)
1160 int64_t r_old, r_new;
1162 ASSERT_RACCT_ENABLED();
1164 resource = *(int *)res;
1165 r_old = racct->r_resources[resource];
1167 /* If there is nothing to decay, just exit. */
1171 mtx_lock(&racct_lock);
1172 r_new = r_old * RACCT_DECAY_FACTOR / FSCALE;
1173 racct->r_resources[resource] = r_new;
1174 mtx_unlock(&racct_lock);
1178 racct_decay(int resource)
1181 ASSERT_RACCT_ENABLED();
1183 ui_racct_foreach(racct_decay_resource, &resource, NULL);
1184 loginclass_racct_foreach(racct_decay_resource, &resource, NULL);
1185 prison_racct_foreach(racct_decay_resource, &resource, NULL);
1193 struct timeval wallclock;
1195 uint64_t pct, pct_estimate;
1197 ASSERT_RACCT_ENABLED();
1200 racct_decay(RACCT_PCTCPU);
1202 sx_slock(&allproc_lock);
1204 LIST_FOREACH(p, &zombproc, p_list) {
1206 racct_set(p, RACCT_PCTCPU, 0);
1210 FOREACH_PROC_IN_SYSTEM(p) {
1212 if (p->p_state != PRS_NORMAL) {
1217 microuptime(&wallclock);
1218 timevalsub(&wallclock, &p->p_stats->p_start);
1220 FOREACH_THREAD_IN_PROC(p, td)
1222 runtime = cputick2usec(p->p_rux.rux_runtime);
1225 KASSERT(runtime >= p->p_prev_runtime,
1226 ("runtime < p_prev_runtime"));
1228 if (runtime < p->p_prev_runtime)
1229 runtime = p->p_prev_runtime;
1231 p->p_prev_runtime = runtime;
1232 if (wallclock.tv_sec > 0 || wallclock.tv_usec > 0) {
1233 pct_estimate = (1000000 * runtime * 100) /
1234 ((uint64_t)wallclock.tv_sec * 1000000 +
1238 pct = racct_getpcpu(p, pct_estimate);
1239 mtx_lock(&racct_lock);
1240 racct_set_force_locked(p, RACCT_PCTCPU, pct);
1241 racct_set_locked(p, RACCT_CPU, runtime);
1242 racct_set_locked(p, RACCT_WALLCLOCK,
1243 (uint64_t)wallclock.tv_sec * 1000000 +
1245 mtx_unlock(&racct_lock);
1250 * To ensure that processes are throttled in a fair way, we need
1251 * to iterate over all processes again and check the limits
1252 * for %cpu resource only after ucred racct containers have been
1255 FOREACH_PROC_IN_SYSTEM(p) {
1257 if (p->p_state != PRS_NORMAL) {
1262 if (racct_pcpu_available(p) <= 0)
1263 racct_proc_throttle(p);
1264 else if (p->p_throttled)
1265 racct_proc_wakeup(p);
1268 sx_sunlock(&allproc_lock);
1273 static struct kproc_desc racctd_kp = {
1285 kproc_start(&racctd_kp);
1287 SYSINIT(racctd, SI_SUB_RACCTD, SI_ORDER_FIRST, racctd_init, NULL);
1295 racct_zone = uma_zcreate("racct", sizeof(struct racct),
1296 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
1298 * XXX: Move this somewhere.
1300 prison0.pr_prison_racct = prison_racct_find("0");
1302 SYSINIT(racct, SI_SUB_RACCT, SI_ORDER_FIRST, racct_init, NULL);
1307 racct_add(struct proc *p, int resource, uint64_t amount)
1314 racct_add_cred(struct ucred *cred, int resource, uint64_t amount)
1319 racct_add_force(struct proc *p, int resource, uint64_t amount)
1326 racct_set(struct proc *p, int resource, uint64_t amount)
1333 racct_set_force(struct proc *p, int resource, uint64_t amount)
1338 racct_sub(struct proc *p, int resource, uint64_t amount)
1343 racct_sub_cred(struct ucred *cred, int resource, uint64_t amount)
1348 racct_get_limit(struct proc *p, int resource)
1351 return (UINT64_MAX);
1355 racct_get_available(struct proc *p, int resource)
1358 return (UINT64_MAX);
1362 racct_create(struct racct **racctp)
1367 racct_destroy(struct racct **racctp)
1372 racct_proc_fork(struct proc *parent, struct proc *child)
1379 racct_proc_fork_done(struct proc *child)
1384 racct_proc_exit(struct proc *p)