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, , rusage, add,
108 "struct proc *", "int", "uint64_t");
109 SDT_PROBE_DEFINE3(racct, , rusage, add__failure,
110 "struct proc *", "int", "uint64_t");
111 SDT_PROBE_DEFINE3(racct, , rusage, add__cred,
112 "struct ucred *", "int", "uint64_t");
113 SDT_PROBE_DEFINE3(racct, , rusage, add__force,
114 "struct proc *", "int", "uint64_t");
115 SDT_PROBE_DEFINE3(racct, , rusage, set,
116 "struct proc *", "int", "uint64_t");
117 SDT_PROBE_DEFINE3(racct, , rusage, set__failure,
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,
125 SDT_PROBE_DEFINE1(racct, , racct, destroy,
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 *");
134 int racct_types[] = {
138 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
140 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
146 RACCT_RECLAIMABLE | RACCT_DENIABLE,
148 RACCT_RECLAIMABLE | RACCT_DENIABLE,
150 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
152 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
154 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
156 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
158 RACCT_RECLAIMABLE | RACCT_DENIABLE,
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_DENIABLE | RACCT_SLOPPY,
168 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
170 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
172 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
176 RACCT_DECAYING | RACCT_DENIABLE | RACCT_IN_MILLIONS };
178 static const fixpt_t RACCT_DECAY_FACTOR = 0.3 * FSCALE;
182 * Contains intermediate values for %cpu calculations to avoid using floating
183 * point in the kernel.
184 * ccpu_exp[k] = FSCALE * (ccpu/FSCALE)^k = FSCALE * exp(-k/20)
185 * It is needed only for the 4BSD scheduler, because in ULE, the ccpu equals to
186 * zero so the calculations are more straightforward.
188 fixpt_t ccpu_exp[] = {
190 [1] = FSCALE * 0.95122942450071400909,
191 [2] = FSCALE * 0.90483741803595957316,
192 [3] = FSCALE * 0.86070797642505780722,
193 [4] = FSCALE * 0.81873075307798185866,
194 [5] = FSCALE * 0.77880078307140486824,
195 [6] = FSCALE * 0.74081822068171786606,
196 [7] = FSCALE * 0.70468808971871343435,
197 [8] = FSCALE * 0.67032004603563930074,
198 [9] = FSCALE * 0.63762815162177329314,
199 [10] = FSCALE * 0.60653065971263342360,
200 [11] = FSCALE * 0.57694981038048669531,
201 [12] = FSCALE * 0.54881163609402643262,
202 [13] = FSCALE * 0.52204577676101604789,
203 [14] = FSCALE * 0.49658530379140951470,
204 [15] = FSCALE * 0.47236655274101470713,
205 [16] = FSCALE * 0.44932896411722159143,
206 [17] = FSCALE * 0.42741493194872666992,
207 [18] = FSCALE * 0.40656965974059911188,
208 [19] = FSCALE * 0.38674102345450120691,
209 [20] = FSCALE * 0.36787944117144232159,
210 [21] = FSCALE * 0.34993774911115535467,
211 [22] = FSCALE * 0.33287108369807955328,
212 [23] = FSCALE * 0.31663676937905321821,
213 [24] = FSCALE * 0.30119421191220209664,
214 [25] = FSCALE * 0.28650479686019010032,
215 [26] = FSCALE * 0.27253179303401260312,
216 [27] = FSCALE * 0.25924026064589150757,
217 [28] = FSCALE * 0.24659696394160647693,
218 [29] = FSCALE * 0.23457028809379765313,
219 [30] = FSCALE * 0.22313016014842982893,
220 [31] = FSCALE * 0.21224797382674305771,
221 [32] = FSCALE * 0.20189651799465540848,
222 [33] = FSCALE * 0.19204990862075411423,
223 [34] = FSCALE * 0.18268352405273465022,
224 [35] = FSCALE * 0.17377394345044512668,
225 [36] = FSCALE * 0.16529888822158653829,
226 [37] = FSCALE * 0.15723716631362761621,
227 [38] = FSCALE * 0.14956861922263505264,
228 [39] = FSCALE * 0.14227407158651357185,
229 [40] = FSCALE * 0.13533528323661269189,
230 [41] = FSCALE * 0.12873490358780421886,
231 [42] = FSCALE * 0.12245642825298191021,
232 [43] = FSCALE * 0.11648415777349695786,
233 [44] = FSCALE * 0.11080315836233388333,
234 [45] = FSCALE * 0.10539922456186433678,
235 [46] = FSCALE * 0.10025884372280373372,
236 [47] = FSCALE * 0.09536916221554961888,
237 [48] = FSCALE * 0.09071795328941250337,
238 [49] = FSCALE * 0.08629358649937051097,
239 [50] = FSCALE * 0.08208499862389879516,
240 [51] = FSCALE * 0.07808166600115315231,
241 [52] = FSCALE * 0.07427357821433388042,
242 [53] = FSCALE * 0.07065121306042958674,
243 [54] = FSCALE * 0.06720551273974976512,
244 [55] = FSCALE * 0.06392786120670757270,
245 [56] = FSCALE * 0.06081006262521796499,
246 [57] = FSCALE * 0.05784432087483846296,
247 [58] = FSCALE * 0.05502322005640722902,
248 [59] = FSCALE * 0.05233970594843239308,
249 [60] = FSCALE * 0.04978706836786394297,
250 [61] = FSCALE * 0.04735892439114092119,
251 [62] = FSCALE * 0.04504920239355780606,
252 [63] = FSCALE * 0.04285212686704017991,
253 [64] = FSCALE * 0.04076220397836621516,
254 [65] = FSCALE * 0.03877420783172200988,
255 [66] = FSCALE * 0.03688316740124000544,
256 [67] = FSCALE * 0.03508435410084502588,
257 [68] = FSCALE * 0.03337326996032607948,
258 [69] = FSCALE * 0.03174563637806794323,
259 [70] = FSCALE * 0.03019738342231850073,
260 [71] = FSCALE * 0.02872463965423942912,
261 [72] = FSCALE * 0.02732372244729256080,
262 [73] = FSCALE * 0.02599112877875534358,
263 [74] = FSCALE * 0.02472352647033939120,
264 [75] = FSCALE * 0.02351774585600910823,
265 [76] = FSCALE * 0.02237077185616559577,
266 [77] = FSCALE * 0.02127973643837716938,
267 [78] = FSCALE * 0.02024191144580438847,
268 [79] = FSCALE * 0.01925470177538692429,
269 [80] = FSCALE * 0.01831563888873418029,
270 [81] = FSCALE * 0.01742237463949351138,
271 [82] = FSCALE * 0.01657267540176124754,
272 [83] = FSCALE * 0.01576441648485449082,
273 [84] = FSCALE * 0.01499557682047770621,
274 [85] = FSCALE * 0.01426423390899925527,
275 [86] = FSCALE * 0.01356855901220093175,
276 [87] = FSCALE * 0.01290681258047986886,
277 [88] = FSCALE * 0.01227733990306844117,
278 [89] = FSCALE * 0.01167856697039544521,
279 [90] = FSCALE * 0.01110899653824230649,
280 [91] = FSCALE * 0.01056720438385265337,
281 [92] = FSCALE * 0.01005183574463358164,
282 [93] = FSCALE * 0.00956160193054350793,
283 [94] = FSCALE * 0.00909527710169581709,
284 [95] = FSCALE * 0.00865169520312063417,
285 [96] = FSCALE * 0.00822974704902002884,
286 [97] = FSCALE * 0.00782837754922577143,
287 [98] = FSCALE * 0.00744658307092434051,
288 [99] = FSCALE * 0.00708340892905212004,
289 [100] = FSCALE * 0.00673794699908546709,
290 [101] = FSCALE * 0.00640933344625638184,
291 [102] = FSCALE * 0.00609674656551563610,
292 [103] = FSCALE * 0.00579940472684214321,
293 [104] = FSCALE * 0.00551656442076077241,
294 [105] = FSCALE * 0.00524751839918138427,
295 [106] = FSCALE * 0.00499159390691021621,
296 [107] = FSCALE * 0.00474815099941147558,
297 [108] = FSCALE * 0.00451658094261266798,
298 [109] = FSCALE * 0.00429630469075234057,
299 [110] = FSCALE * 0.00408677143846406699,
303 #define CCPU_EXP_MAX 110
306 * This function is analogical to the getpcpu() function in the ps(1) command.
307 * They should both calculate in the same way so that the racct %cpu
308 * calculations are consistent with the values showed by the ps(1) tool.
309 * The calculations are more complex in the 4BSD scheduler because of the value
310 * of the ccpu variable. In ULE it is defined to be zero which saves us some
314 racct_getpcpu(struct proc *p, u_int pcpu)
318 fixpt_t pctcpu, pctcpu_next;
327 ASSERT_RACCT_ENABLED();
330 * If the process is swapped out, we count its %cpu usage as zero.
331 * This behaviour is consistent with the userland ps(1) tool.
333 if ((p->p_flag & P_INMEM) == 0)
335 swtime = (ticks - p->p_swtick) / hz;
338 * For short-lived processes, the sched_pctcpu() returns small
339 * values even for cpu intensive processes. Therefore we use
340 * our own estimate in this case.
342 if (swtime < RACCT_PCPU_SECS)
346 FOREACH_THREAD_IN_PROC(p, td) {
347 if (td == PCPU_GET(idlethread))
351 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
352 if (td == pc->pc_idlethread) {
362 pctcpu = sched_pctcpu(td);
363 /* Count also the yet unfinished second. */
364 pctcpu_next = (pctcpu * ccpu_exp[1]) >> FSHIFT;
365 pctcpu_next += sched_pctcpu_delta(td);
366 p_pctcpu += max(pctcpu, pctcpu_next);
369 * In ULE the %cpu statistics are updated on every
370 * sched_pctcpu() call. So special calculations to
371 * account for the latest (unfinished) second are
374 p_pctcpu += sched_pctcpu(td);
380 if (swtime <= CCPU_EXP_MAX)
381 return ((100 * (uint64_t)p_pctcpu * 1000000) /
382 (FSCALE - ccpu_exp[swtime]));
385 return ((100 * (uint64_t)p_pctcpu * 1000000) / FSCALE);
389 racct_add_racct(struct racct *dest, const struct racct *src)
393 ASSERT_RACCT_ENABLED();
394 mtx_assert(&racct_lock, MA_OWNED);
397 * Update resource usage in dest.
399 for (i = 0; i <= RACCT_MAX; i++) {
400 KASSERT(dest->r_resources[i] >= 0,
401 ("%s: resource %d propagation meltdown: dest < 0",
403 KASSERT(src->r_resources[i] >= 0,
404 ("%s: resource %d propagation meltdown: src < 0",
406 dest->r_resources[i] += src->r_resources[i];
411 racct_sub_racct(struct racct *dest, const struct racct *src)
415 ASSERT_RACCT_ENABLED();
416 mtx_assert(&racct_lock, MA_OWNED);
419 * Update resource usage in dest.
421 for (i = 0; i <= RACCT_MAX; i++) {
422 if (!RACCT_IS_SLOPPY(i) && !RACCT_IS_DECAYING(i)) {
423 KASSERT(dest->r_resources[i] >= 0,
424 ("%s: resource %d propagation meltdown: dest < 0",
426 KASSERT(src->r_resources[i] >= 0,
427 ("%s: resource %d propagation meltdown: src < 0",
429 KASSERT(src->r_resources[i] <= dest->r_resources[i],
430 ("%s: resource %d propagation meltdown: src > dest",
433 if (RACCT_CAN_DROP(i)) {
434 dest->r_resources[i] -= src->r_resources[i];
435 if (dest->r_resources[i] < 0) {
436 KASSERT(RACCT_IS_SLOPPY(i) ||
437 RACCT_IS_DECAYING(i),
438 ("%s: resource %d usage < 0", __func__, i));
439 dest->r_resources[i] = 0;
446 racct_create(struct racct **racctp)
452 SDT_PROBE1(racct, , racct, create, racctp);
454 KASSERT(*racctp == NULL, ("racct already allocated"));
456 *racctp = uma_zalloc(racct_zone, M_WAITOK | M_ZERO);
460 racct_destroy_locked(struct racct **racctp)
465 ASSERT_RACCT_ENABLED();
467 SDT_PROBE1(racct, , racct, destroy, racctp);
469 mtx_assert(&racct_lock, MA_OWNED);
470 KASSERT(racctp != NULL, ("NULL racctp"));
471 KASSERT(*racctp != NULL, ("NULL racct"));
475 for (i = 0; i <= RACCT_MAX; i++) {
476 if (RACCT_IS_SLOPPY(i))
478 if (!RACCT_IS_RECLAIMABLE(i))
480 KASSERT(racct->r_resources[i] == 0,
481 ("destroying non-empty racct: "
482 "%ju allocated for resource %d\n",
483 racct->r_resources[i], i));
485 uma_zfree(racct_zone, racct);
490 racct_destroy(struct racct **racct)
496 mtx_lock(&racct_lock);
497 racct_destroy_locked(racct);
498 mtx_unlock(&racct_lock);
502 * Increase consumption of 'resource' by 'amount' for 'racct'
503 * and all its parents. Differently from other cases, 'amount' here
504 * may be less than zero.
507 racct_adjust_resource(struct racct *racct, int resource,
511 ASSERT_RACCT_ENABLED();
512 mtx_assert(&racct_lock, MA_OWNED);
513 KASSERT(racct != NULL, ("NULL racct"));
515 racct->r_resources[resource] += amount;
516 if (racct->r_resources[resource] < 0) {
517 KASSERT(RACCT_IS_SLOPPY(resource) || RACCT_IS_DECAYING(resource),
518 ("%s: resource %d usage < 0", __func__, resource));
519 racct->r_resources[resource] = 0;
523 * There are some cases where the racct %cpu resource would grow
524 * beyond 100% per core. For example in racct_proc_exit() we add
525 * the process %cpu usage to the ucred racct containers. If too
526 * many processes terminated in a short time span, the ucred %cpu
527 * resource could grow too much. Also, the 4BSD scheduler sometimes
528 * returns for a thread more than 100% cpu usage. So we set a sane
529 * boundary here to 100% * the maxumum number of CPUs.
531 if ((resource == RACCT_PCTCPU) &&
532 (racct->r_resources[RACCT_PCTCPU] > 100 * 1000000 * (int64_t)MAXCPU))
533 racct->r_resources[RACCT_PCTCPU] = 100 * 1000000 * (int64_t)MAXCPU;
537 racct_add_locked(struct proc *p, int resource, uint64_t amount)
543 ASSERT_RACCT_ENABLED();
545 SDT_PROBE3(racct, , rusage, add, p, resource, amount);
548 * We need proc lock to dereference p->p_ucred.
550 PROC_LOCK_ASSERT(p, MA_OWNED);
553 error = rctl_enforce(p, resource, amount);
554 if (error && RACCT_IS_DENIABLE(resource)) {
555 SDT_PROBE3(racct, , rusage, add__failure, p, resource, amount);
559 racct_adjust_resource(p->p_racct, resource, amount);
560 racct_add_cred_locked(p->p_ucred, resource, amount);
566 * Increase allocation of 'resource' by 'amount' for process 'p'.
567 * Return 0 if it's below limits, or errno, if it's not.
570 racct_add(struct proc *p, int resource, uint64_t amount)
577 mtx_lock(&racct_lock);
578 error = racct_add_locked(p, resource, amount);
579 mtx_unlock(&racct_lock);
584 racct_add_cred_locked(struct ucred *cred, int resource, uint64_t amount)
588 ASSERT_RACCT_ENABLED();
590 SDT_PROBE3(racct, , rusage, add__cred, cred, resource, amount);
592 racct_adjust_resource(cred->cr_ruidinfo->ui_racct, resource, amount);
593 for (pr = cred->cr_prison; pr != NULL; pr = pr->pr_parent)
594 racct_adjust_resource(pr->pr_prison_racct->prr_racct, resource,
596 racct_adjust_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_PROBE3(racct, , rusage, add__force, p, resource, amount);
631 * We need proc lock to dereference p->p_ucred.
633 PROC_LOCK_ASSERT(p, MA_OWNED);
635 mtx_lock(&racct_lock);
636 racct_adjust_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_PROBE3(racct, , rusage, set, p, resource, amount);
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_PROBE3(racct, , rusage, set__failure, p, resource,
690 racct_adjust_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_PROBE3(racct, , rusage, set, p, resource, amount);
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_adjust_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_PROBE3(racct, , rusage, sub, p, resource, amount);
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_adjust_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_PROBE3(racct, , rusage, sub__cred, cred, resource, amount);
869 KASSERT(RACCT_CAN_DROP(resource),
870 ("%s: called for resource %d which can not drop", __func__,
874 racct_adjust_resource(cred->cr_ruidinfo->ui_racct, resource, -amount);
875 for (pr = cred->cr_prison; pr != NULL; pr = pr->pr_parent)
876 racct_adjust_resource(pr->pr_prison_racct->prr_racct, resource,
878 racct_adjust_resource(cred->cr_loginclass->lc_racct, resource, -amount);
882 * Decrease allocation of 'resource' by 'amount' for credential 'cred'.
885 racct_sub_cred(struct ucred *cred, int resource, uint64_t amount)
891 mtx_lock(&racct_lock);
892 racct_sub_cred_locked(cred, resource, amount);
893 mtx_unlock(&racct_lock);
897 * Inherit resource usage information from the parent process.
900 racct_proc_fork(struct proc *parent, struct proc *child)
908 * Create racct for the child process.
910 racct_create(&child->p_racct);
914 mtx_lock(&racct_lock);
917 error = rctl_proc_fork(parent, child);
922 /* Init process cpu time. */
923 child->p_prev_runtime = 0;
924 child->p_throttled = 0;
927 * Inherit resource usage.
929 for (i = 0; i <= RACCT_MAX; i++) {
930 if (parent->p_racct->r_resources[i] == 0 ||
931 !RACCT_IS_INHERITABLE(i))
934 error = racct_set_locked(child, i,
935 parent->p_racct->r_resources[i]);
940 error = racct_add_locked(child, RACCT_NPROC, 1);
941 error += racct_add_locked(child, RACCT_NTHR, 1);
944 mtx_unlock(&racct_lock);
949 racct_proc_exit(child);
955 * Called at the end of fork1(), to handle rules that require the process
956 * to be fully initialized.
959 racct_proc_fork_done(struct proc *child)
967 mtx_lock(&racct_lock);
968 rctl_enforce(child, RACCT_NPROC, 0);
969 rctl_enforce(child, RACCT_NTHR, 0);
970 mtx_unlock(&racct_lock);
976 racct_proc_exit(struct proc *p)
980 struct timeval wallclock;
981 uint64_t pct_estimate, pct;
988 * We don't need to calculate rux, proc_reap() has already done this.
990 runtime = cputick2usec(p->p_rux.rux_runtime);
992 KASSERT(runtime >= p->p_prev_runtime, ("runtime < p_prev_runtime"));
994 if (runtime < p->p_prev_runtime)
995 runtime = p->p_prev_runtime;
997 microuptime(&wallclock);
998 timevalsub(&wallclock, &p->p_stats->p_start);
999 if (wallclock.tv_sec > 0 || wallclock.tv_usec > 0) {
1000 pct_estimate = (1000000 * runtime * 100) /
1001 ((uint64_t)wallclock.tv_sec * 1000000 +
1005 pct = racct_getpcpu(p, pct_estimate);
1007 mtx_lock(&racct_lock);
1008 racct_set_locked(p, RACCT_CPU, runtime);
1009 racct_add_cred_locked(p->p_ucred, RACCT_PCTCPU, pct);
1011 KASSERT(p->p_racct->r_resources[RACCT_RSS] == 0,
1012 ("process reaped with %ju allocated for RSS\n",
1013 p->p_racct->r_resources[RACCT_RSS]));
1014 for (i = 0; i <= RACCT_MAX; i++) {
1015 if (p->p_racct->r_resources[i] == 0)
1017 if (!RACCT_IS_RECLAIMABLE(i))
1019 racct_set_locked(p, i, 0);
1022 mtx_unlock(&racct_lock);
1026 rctl_racct_release(p->p_racct);
1028 racct_destroy(&p->p_racct);
1032 * Called after credentials change, to move resource utilisation
1036 racct_proc_ucred_changed(struct proc *p, struct ucred *oldcred,
1037 struct ucred *newcred)
1039 struct uidinfo *olduip, *newuip;
1040 struct loginclass *oldlc, *newlc;
1041 struct prison *oldpr, *newpr, *pr;
1046 PROC_LOCK_ASSERT(p, MA_NOTOWNED);
1048 newuip = newcred->cr_ruidinfo;
1049 olduip = oldcred->cr_ruidinfo;
1050 newlc = newcred->cr_loginclass;
1051 oldlc = oldcred->cr_loginclass;
1052 newpr = newcred->cr_prison;
1053 oldpr = oldcred->cr_prison;
1055 mtx_lock(&racct_lock);
1056 if (newuip != olduip) {
1057 racct_sub_racct(olduip->ui_racct, p->p_racct);
1058 racct_add_racct(newuip->ui_racct, p->p_racct);
1060 if (newlc != oldlc) {
1061 racct_sub_racct(oldlc->lc_racct, p->p_racct);
1062 racct_add_racct(newlc->lc_racct, p->p_racct);
1064 if (newpr != oldpr) {
1065 for (pr = oldpr; pr != NULL; pr = pr->pr_parent)
1066 racct_sub_racct(pr->pr_prison_racct->prr_racct,
1068 for (pr = newpr; pr != NULL; pr = pr->pr_parent)
1069 racct_add_racct(pr->pr_prison_racct->prr_racct,
1072 mtx_unlock(&racct_lock);
1075 rctl_proc_ucred_changed(p, newcred);
1080 racct_move(struct racct *dest, struct racct *src)
1083 ASSERT_RACCT_ENABLED();
1085 mtx_lock(&racct_lock);
1087 racct_add_racct(dest, src);
1088 racct_sub_racct(src, src);
1090 mtx_unlock(&racct_lock);
1094 racct_proc_throttle(struct proc *p)
1101 ASSERT_RACCT_ENABLED();
1102 PROC_LOCK_ASSERT(p, MA_OWNED);
1105 * Do not block kernel processes. Also do not block processes with
1106 * low %cpu utilization to improve interactivity.
1108 if (((p->p_flag & (P_SYSTEM | P_KTHREAD)) != 0) ||
1109 (p->p_racct->r_resources[RACCT_PCTCPU] <= pcpu_threshold))
1113 FOREACH_THREAD_IN_PROC(p, td) {
1115 switch (td->td_state) {
1118 * If the thread is on the scheduler run-queue, we can
1119 * not just remove it from there. So we set the flag
1120 * TDF_NEEDRESCHED for the thread, so that once it is
1121 * running, it is taken off the cpu as soon as possible.
1123 td->td_flags |= TDF_NEEDRESCHED;
1127 * If the thread is running, we request a context
1128 * switch for it by setting the TDF_NEEDRESCHED flag.
1130 td->td_flags |= TDF_NEEDRESCHED;
1132 cpuid = td->td_oncpu;
1133 if ((cpuid != NOCPU) && (td != curthread))
1134 ipi_cpu(cpuid, IPI_AST);
1145 racct_proc_wakeup(struct proc *p)
1148 ASSERT_RACCT_ENABLED();
1150 PROC_LOCK_ASSERT(p, MA_OWNED);
1152 if (p->p_throttled) {
1159 racct_decay_resource(struct racct *racct, void * res, void* dummy)
1162 int64_t r_old, r_new;
1164 ASSERT_RACCT_ENABLED();
1166 resource = *(int *)res;
1167 r_old = racct->r_resources[resource];
1169 /* If there is nothing to decay, just exit. */
1173 mtx_lock(&racct_lock);
1174 r_new = r_old * RACCT_DECAY_FACTOR / FSCALE;
1175 racct->r_resources[resource] = r_new;
1176 mtx_unlock(&racct_lock);
1180 racct_decay(int resource)
1183 ASSERT_RACCT_ENABLED();
1185 ui_racct_foreach(racct_decay_resource, &resource, NULL);
1186 loginclass_racct_foreach(racct_decay_resource, &resource, NULL);
1187 prison_racct_foreach(racct_decay_resource, &resource, NULL);
1195 struct timeval wallclock;
1197 uint64_t pct, pct_estimate;
1199 ASSERT_RACCT_ENABLED();
1202 racct_decay(RACCT_PCTCPU);
1204 sx_slock(&allproc_lock);
1206 LIST_FOREACH(p, &zombproc, p_list) {
1208 racct_set(p, RACCT_PCTCPU, 0);
1212 FOREACH_PROC_IN_SYSTEM(p) {
1214 if (p->p_state != PRS_NORMAL) {
1219 microuptime(&wallclock);
1220 timevalsub(&wallclock, &p->p_stats->p_start);
1222 FOREACH_THREAD_IN_PROC(p, td)
1224 runtime = cputick2usec(p->p_rux.rux_runtime);
1227 KASSERT(runtime >= p->p_prev_runtime,
1228 ("runtime < p_prev_runtime"));
1230 if (runtime < p->p_prev_runtime)
1231 runtime = p->p_prev_runtime;
1233 p->p_prev_runtime = runtime;
1234 if (wallclock.tv_sec > 0 || wallclock.tv_usec > 0) {
1235 pct_estimate = (1000000 * runtime * 100) /
1236 ((uint64_t)wallclock.tv_sec * 1000000 +
1240 pct = racct_getpcpu(p, pct_estimate);
1241 mtx_lock(&racct_lock);
1242 racct_set_force_locked(p, RACCT_PCTCPU, pct);
1243 racct_set_locked(p, RACCT_CPU, runtime);
1244 racct_set_locked(p, RACCT_WALLCLOCK,
1245 (uint64_t)wallclock.tv_sec * 1000000 +
1247 mtx_unlock(&racct_lock);
1252 * To ensure that processes are throttled in a fair way, we need
1253 * to iterate over all processes again and check the limits
1254 * for %cpu resource only after ucred racct containers have been
1257 FOREACH_PROC_IN_SYSTEM(p) {
1259 if (p->p_state != PRS_NORMAL) {
1264 if (racct_pcpu_available(p) <= 0)
1265 racct_proc_throttle(p);
1266 else if (p->p_throttled)
1267 racct_proc_wakeup(p);
1270 sx_sunlock(&allproc_lock);
1275 static struct kproc_desc racctd_kp = {
1287 kproc_start(&racctd_kp);
1289 SYSINIT(racctd, SI_SUB_RACCTD, SI_ORDER_FIRST, racctd_init, NULL);
1297 racct_zone = uma_zcreate("racct", sizeof(struct racct),
1298 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
1300 * XXX: Move this somewhere.
1302 prison0.pr_prison_racct = prison_racct_find("0");
1304 SYSINIT(racct, SI_SUB_RACCT, SI_ORDER_FIRST, racct_init, NULL);
1309 racct_add(struct proc *p, int resource, uint64_t amount)
1316 racct_add_cred(struct ucred *cred, int resource, uint64_t amount)
1321 racct_add_force(struct proc *p, int resource, uint64_t amount)
1328 racct_set(struct proc *p, int resource, uint64_t amount)
1335 racct_set_force(struct proc *p, int resource, uint64_t amount)
1340 racct_sub(struct proc *p, int resource, uint64_t amount)
1345 racct_sub_cred(struct ucred *cred, int resource, uint64_t amount)
1350 racct_get_limit(struct proc *p, int resource)
1353 return (UINT64_MAX);
1357 racct_get_available(struct proc *p, int resource)
1360 return (UINT64_MAX);
1364 racct_create(struct racct **racctp)
1369 racct_destroy(struct racct **racctp)
1374 racct_proc_fork(struct proc *parent, struct proc *child)
1381 racct_proc_fork_done(struct proc *child)
1386 racct_proc_exit(struct proc *p)