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;
75 SYSCTL_NODE(_kern, OID_AUTO, racct, CTLFLAG_RW, 0, "Resource Accounting");
76 SYSCTL_UINT(_kern_racct, OID_AUTO, pcpu_threshold, CTLFLAG_RW, &pcpu_threshold,
77 0, "Processes with higher %cpu usage than this value can be throttled.");
80 * How many seconds it takes to use the scheduler %cpu calculations. When a
81 * process starts, we compute its %cpu usage by dividing its runtime by the
82 * process wall clock time. After RACCT_PCPU_SECS pass, we use the value
83 * provided by the scheduler.
85 #define RACCT_PCPU_SECS 3
87 static struct mtx racct_lock;
88 MTX_SYSINIT(racct_lock, &racct_lock, "racct lock", MTX_DEF);
90 static uma_zone_t racct_zone;
92 static void racct_sub_racct(struct racct *dest, const struct racct *src);
93 static void racct_sub_cred_locked(struct ucred *cred, int resource,
95 static void racct_add_cred_locked(struct ucred *cred, int resource,
98 SDT_PROVIDER_DEFINE(racct);
99 SDT_PROBE_DEFINE3(racct, kernel, rusage, add, add, "struct proc *", "int",
101 SDT_PROBE_DEFINE3(racct, kernel, rusage, add_failure, add-failure,
102 "struct proc *", "int", "uint64_t");
103 SDT_PROBE_DEFINE3(racct, kernel, rusage, add_cred, add-cred, "struct ucred *",
105 SDT_PROBE_DEFINE3(racct, kernel, rusage, add_force, add-force, "struct proc *",
107 SDT_PROBE_DEFINE3(racct, kernel, rusage, set, set, "struct proc *", "int",
109 SDT_PROBE_DEFINE3(racct, kernel, rusage, set_failure, set-failure,
110 "struct proc *", "int", "uint64_t");
111 SDT_PROBE_DEFINE3(racct, kernel, rusage, sub, sub, "struct proc *", "int",
113 SDT_PROBE_DEFINE3(racct, kernel, rusage, sub_cred, sub-cred, "struct ucred *",
115 SDT_PROBE_DEFINE1(racct, kernel, racct, create, create, "struct racct *");
116 SDT_PROBE_DEFINE1(racct, kernel, racct, destroy, destroy, "struct racct *");
117 SDT_PROBE_DEFINE2(racct, kernel, racct, join, join, "struct racct *",
119 SDT_PROBE_DEFINE2(racct, kernel, racct, join_failure, join-failure,
120 "struct racct *", "struct racct *");
121 SDT_PROBE_DEFINE2(racct, kernel, racct, leave, leave, "struct racct *",
124 int racct_types[] = {
128 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
130 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
136 RACCT_RECLAIMABLE | RACCT_DENIABLE,
138 RACCT_RECLAIMABLE | RACCT_DENIABLE,
140 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
142 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
144 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
146 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
148 RACCT_RECLAIMABLE | RACCT_DENIABLE,
150 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
152 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
154 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
156 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
158 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
160 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
162 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
166 RACCT_DECAYING | RACCT_DENIABLE | RACCT_IN_MILLIONS };
168 static const fixpt_t RACCT_DECAY_FACTOR = 0.3 * FSCALE;
172 * Contains intermediate values for %cpu calculations to avoid using floating
173 * point in the kernel.
174 * ccpu_exp[k] = FSCALE * (ccpu/FSCALE)^k = FSCALE * exp(-k/20)
175 * It is needed only for the 4BSD scheduler, because in ULE, the ccpu equals to
176 * zero so the calculations are more straightforward.
178 fixpt_t ccpu_exp[] = {
180 [1] = FSCALE * 0.95122942450071400909,
181 [2] = FSCALE * 0.90483741803595957316,
182 [3] = FSCALE * 0.86070797642505780722,
183 [4] = FSCALE * 0.81873075307798185866,
184 [5] = FSCALE * 0.77880078307140486824,
185 [6] = FSCALE * 0.74081822068171786606,
186 [7] = FSCALE * 0.70468808971871343435,
187 [8] = FSCALE * 0.67032004603563930074,
188 [9] = FSCALE * 0.63762815162177329314,
189 [10] = FSCALE * 0.60653065971263342360,
190 [11] = FSCALE * 0.57694981038048669531,
191 [12] = FSCALE * 0.54881163609402643262,
192 [13] = FSCALE * 0.52204577676101604789,
193 [14] = FSCALE * 0.49658530379140951470,
194 [15] = FSCALE * 0.47236655274101470713,
195 [16] = FSCALE * 0.44932896411722159143,
196 [17] = FSCALE * 0.42741493194872666992,
197 [18] = FSCALE * 0.40656965974059911188,
198 [19] = FSCALE * 0.38674102345450120691,
199 [20] = FSCALE * 0.36787944117144232159,
200 [21] = FSCALE * 0.34993774911115535467,
201 [22] = FSCALE * 0.33287108369807955328,
202 [23] = FSCALE * 0.31663676937905321821,
203 [24] = FSCALE * 0.30119421191220209664,
204 [25] = FSCALE * 0.28650479686019010032,
205 [26] = FSCALE * 0.27253179303401260312,
206 [27] = FSCALE * 0.25924026064589150757,
207 [28] = FSCALE * 0.24659696394160647693,
208 [29] = FSCALE * 0.23457028809379765313,
209 [30] = FSCALE * 0.22313016014842982893,
210 [31] = FSCALE * 0.21224797382674305771,
211 [32] = FSCALE * 0.20189651799465540848,
212 [33] = FSCALE * 0.19204990862075411423,
213 [34] = FSCALE * 0.18268352405273465022,
214 [35] = FSCALE * 0.17377394345044512668,
215 [36] = FSCALE * 0.16529888822158653829,
216 [37] = FSCALE * 0.15723716631362761621,
217 [38] = FSCALE * 0.14956861922263505264,
218 [39] = FSCALE * 0.14227407158651357185,
219 [40] = FSCALE * 0.13533528323661269189,
220 [41] = FSCALE * 0.12873490358780421886,
221 [42] = FSCALE * 0.12245642825298191021,
222 [43] = FSCALE * 0.11648415777349695786,
223 [44] = FSCALE * 0.11080315836233388333,
224 [45] = FSCALE * 0.10539922456186433678,
225 [46] = FSCALE * 0.10025884372280373372,
226 [47] = FSCALE * 0.09536916221554961888,
227 [48] = FSCALE * 0.09071795328941250337,
228 [49] = FSCALE * 0.08629358649937051097,
229 [50] = FSCALE * 0.08208499862389879516,
230 [51] = FSCALE * 0.07808166600115315231,
231 [52] = FSCALE * 0.07427357821433388042,
232 [53] = FSCALE * 0.07065121306042958674,
233 [54] = FSCALE * 0.06720551273974976512,
234 [55] = FSCALE * 0.06392786120670757270,
235 [56] = FSCALE * 0.06081006262521796499,
236 [57] = FSCALE * 0.05784432087483846296,
237 [58] = FSCALE * 0.05502322005640722902,
238 [59] = FSCALE * 0.05233970594843239308,
239 [60] = FSCALE * 0.04978706836786394297,
240 [61] = FSCALE * 0.04735892439114092119,
241 [62] = FSCALE * 0.04504920239355780606,
242 [63] = FSCALE * 0.04285212686704017991,
243 [64] = FSCALE * 0.04076220397836621516,
244 [65] = FSCALE * 0.03877420783172200988,
245 [66] = FSCALE * 0.03688316740124000544,
246 [67] = FSCALE * 0.03508435410084502588,
247 [68] = FSCALE * 0.03337326996032607948,
248 [69] = FSCALE * 0.03174563637806794323,
249 [70] = FSCALE * 0.03019738342231850073,
250 [71] = FSCALE * 0.02872463965423942912,
251 [72] = FSCALE * 0.02732372244729256080,
252 [73] = FSCALE * 0.02599112877875534358,
253 [74] = FSCALE * 0.02472352647033939120,
254 [75] = FSCALE * 0.02351774585600910823,
255 [76] = FSCALE * 0.02237077185616559577,
256 [77] = FSCALE * 0.02127973643837716938,
257 [78] = FSCALE * 0.02024191144580438847,
258 [79] = FSCALE * 0.01925470177538692429,
259 [80] = FSCALE * 0.01831563888873418029,
260 [81] = FSCALE * 0.01742237463949351138,
261 [82] = FSCALE * 0.01657267540176124754,
262 [83] = FSCALE * 0.01576441648485449082,
263 [84] = FSCALE * 0.01499557682047770621,
264 [85] = FSCALE * 0.01426423390899925527,
265 [86] = FSCALE * 0.01356855901220093175,
266 [87] = FSCALE * 0.01290681258047986886,
267 [88] = FSCALE * 0.01227733990306844117,
268 [89] = FSCALE * 0.01167856697039544521,
269 [90] = FSCALE * 0.01110899653824230649,
270 [91] = FSCALE * 0.01056720438385265337,
271 [92] = FSCALE * 0.01005183574463358164,
272 [93] = FSCALE * 0.00956160193054350793,
273 [94] = FSCALE * 0.00909527710169581709,
274 [95] = FSCALE * 0.00865169520312063417,
275 [96] = FSCALE * 0.00822974704902002884,
276 [97] = FSCALE * 0.00782837754922577143,
277 [98] = FSCALE * 0.00744658307092434051,
278 [99] = FSCALE * 0.00708340892905212004,
279 [100] = FSCALE * 0.00673794699908546709,
280 [101] = FSCALE * 0.00640933344625638184,
281 [102] = FSCALE * 0.00609674656551563610,
282 [103] = FSCALE * 0.00579940472684214321,
283 [104] = FSCALE * 0.00551656442076077241,
284 [105] = FSCALE * 0.00524751839918138427,
285 [106] = FSCALE * 0.00499159390691021621,
286 [107] = FSCALE * 0.00474815099941147558,
287 [108] = FSCALE * 0.00451658094261266798,
288 [109] = FSCALE * 0.00429630469075234057,
289 [110] = FSCALE * 0.00408677143846406699,
293 #define CCPU_EXP_MAX 110
296 * This function is analogical to the getpcpu() function in the ps(1) command.
297 * They should both calculate in the same way so that the racct %cpu
298 * calculations are consistent with the values showed by the ps(1) tool.
299 * The calculations are more complex in the 4BSD scheduler because of the value
300 * of the ccpu variable. In ULE it is defined to be zero which saves us some
304 racct_getpcpu(struct proc *p, u_int pcpu)
308 fixpt_t pctcpu, pctcpu_next;
318 * If the process is swapped out, we count its %cpu usage as zero.
319 * This behaviour is consistent with the userland ps(1) tool.
321 if ((p->p_flag & P_INMEM) == 0)
323 swtime = (ticks - p->p_swtick) / hz;
326 * For short-lived processes, the sched_pctcpu() returns small
327 * values even for cpu intensive processes. Therefore we use
328 * our own estimate in this case.
330 if (swtime < RACCT_PCPU_SECS)
334 FOREACH_THREAD_IN_PROC(p, td) {
335 if (td == PCPU_GET(idlethread))
339 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
340 if (td == pc->pc_idlethread) {
350 pctcpu = sched_pctcpu(td);
351 /* Count also the yet unfinished second. */
352 pctcpu_next = (pctcpu * ccpu_exp[1]) >> FSHIFT;
353 pctcpu_next += sched_pctcpu_delta(td);
354 p_pctcpu += max(pctcpu, pctcpu_next);
357 * In ULE the %cpu statistics are updated on every
358 * sched_pctcpu() call. So special calculations to
359 * account for the latest (unfinished) second are
362 p_pctcpu += sched_pctcpu(td);
368 if (swtime <= CCPU_EXP_MAX)
369 return ((100 * (uint64_t)p_pctcpu * 1000000) /
370 (FSCALE - ccpu_exp[swtime]));
373 return ((100 * (uint64_t)p_pctcpu * 1000000) / FSCALE);
377 racct_add_racct(struct racct *dest, const struct racct *src)
381 mtx_assert(&racct_lock, MA_OWNED);
384 * Update resource usage in dest.
386 for (i = 0; i <= RACCT_MAX; i++) {
387 KASSERT(dest->r_resources[i] >= 0,
388 ("racct propagation meltdown: dest < 0"));
389 KASSERT(src->r_resources[i] >= 0,
390 ("racct propagation meltdown: src < 0"));
391 dest->r_resources[i] += src->r_resources[i];
396 racct_sub_racct(struct racct *dest, const struct racct *src)
400 mtx_assert(&racct_lock, MA_OWNED);
403 * Update resource usage in dest.
405 for (i = 0; i <= RACCT_MAX; i++) {
406 if (!RACCT_IS_SLOPPY(i)) {
407 KASSERT(dest->r_resources[i] >= 0,
408 ("racct propagation meltdown: dest < 0"));
409 KASSERT(src->r_resources[i] >= 0,
410 ("racct propagation meltdown: src < 0"));
411 KASSERT(src->r_resources[i] <= dest->r_resources[i],
412 ("racct propagation meltdown: src > dest"));
414 if (RACCT_CAN_DROP(i)) {
415 dest->r_resources[i] -= src->r_resources[i];
416 if (dest->r_resources[i] < 0) {
417 KASSERT(RACCT_IS_SLOPPY(i),
418 ("racct_sub_racct: usage < 0"));
419 dest->r_resources[i] = 0;
426 racct_create(struct racct **racctp)
429 SDT_PROBE(racct, kernel, racct, create, racctp, 0, 0, 0, 0);
431 KASSERT(*racctp == NULL, ("racct already allocated"));
433 *racctp = uma_zalloc(racct_zone, M_WAITOK | M_ZERO);
437 racct_destroy_locked(struct racct **racctp)
442 SDT_PROBE(racct, kernel, racct, destroy, racctp, 0, 0, 0, 0);
444 mtx_assert(&racct_lock, MA_OWNED);
445 KASSERT(racctp != NULL, ("NULL racctp"));
446 KASSERT(*racctp != NULL, ("NULL racct"));
450 for (i = 0; i <= RACCT_MAX; i++) {
451 if (RACCT_IS_SLOPPY(i))
453 if (!RACCT_IS_RECLAIMABLE(i))
455 KASSERT(racct->r_resources[i] == 0,
456 ("destroying non-empty racct: "
457 "%ju allocated for resource %d\n",
458 racct->r_resources[i], i));
460 uma_zfree(racct_zone, racct);
465 racct_destroy(struct racct **racct)
468 mtx_lock(&racct_lock);
469 racct_destroy_locked(racct);
470 mtx_unlock(&racct_lock);
474 * Increase consumption of 'resource' by 'amount' for 'racct'
475 * and all its parents. Differently from other cases, 'amount' here
476 * may be less than zero.
479 racct_alloc_resource(struct racct *racct, int resource,
483 mtx_assert(&racct_lock, MA_OWNED);
484 KASSERT(racct != NULL, ("NULL racct"));
486 racct->r_resources[resource] += amount;
487 if (racct->r_resources[resource] < 0) {
488 KASSERT(RACCT_IS_SLOPPY(resource) || RACCT_IS_DECAYING(resource),
489 ("racct_alloc_resource: usage < 0"));
490 racct->r_resources[resource] = 0;
494 * There are some cases where the racct %cpu resource would grow
496 * For example in racct_proc_exit() we add the process %cpu usage
497 * to the ucred racct containers. If too many processes terminated
498 * in a short time span, the ucred %cpu resource could grow too much.
499 * Also, the 4BSD scheduler sometimes returns for a thread more than
500 * 100% cpu usage. So we set a boundary here to 100%.
502 if ((resource == RACCT_PCTCPU) &&
503 (racct->r_resources[RACCT_PCTCPU] > 100 * 1000000))
504 racct->r_resources[RACCT_PCTCPU] = 100 * 1000000;
508 racct_add_locked(struct proc *p, int resource, uint64_t amount)
514 SDT_PROBE(racct, kernel, rusage, add, p, resource, amount, 0, 0);
517 * We need proc lock to dereference p->p_ucred.
519 PROC_LOCK_ASSERT(p, MA_OWNED);
522 error = rctl_enforce(p, resource, amount);
523 if (error && RACCT_IS_DENIABLE(resource)) {
524 SDT_PROBE(racct, kernel, rusage, add_failure, p, resource,
529 racct_alloc_resource(p->p_racct, resource, amount);
530 racct_add_cred_locked(p->p_ucred, resource, amount);
536 * Increase allocation of 'resource' by 'amount' for process 'p'.
537 * Return 0 if it's below limits, or errno, if it's not.
540 racct_add(struct proc *p, int resource, uint64_t amount)
544 mtx_lock(&racct_lock);
545 error = racct_add_locked(p, resource, amount);
546 mtx_unlock(&racct_lock);
551 racct_add_cred_locked(struct ucred *cred, int resource, uint64_t amount)
555 SDT_PROBE(racct, kernel, rusage, add_cred, cred, resource, amount,
558 racct_alloc_resource(cred->cr_ruidinfo->ui_racct, resource, amount);
559 for (pr = cred->cr_prison; pr != NULL; pr = pr->pr_parent)
560 racct_alloc_resource(pr->pr_prison_racct->prr_racct, resource,
562 racct_alloc_resource(cred->cr_loginclass->lc_racct, resource, amount);
566 * Increase allocation of 'resource' by 'amount' for credential 'cred'.
567 * Doesn't check for limits and never fails.
569 * XXX: Shouldn't this ever return an error?
572 racct_add_cred(struct ucred *cred, int resource, uint64_t amount)
575 mtx_lock(&racct_lock);
576 racct_add_cred_locked(cred, resource, amount);
577 mtx_unlock(&racct_lock);
581 * Increase allocation of 'resource' by 'amount' for process 'p'.
582 * Doesn't check for limits and never fails.
585 racct_add_force(struct proc *p, int resource, uint64_t amount)
588 SDT_PROBE(racct, kernel, rusage, add_force, p, resource, amount, 0, 0);
591 * We need proc lock to dereference p->p_ucred.
593 PROC_LOCK_ASSERT(p, MA_OWNED);
595 mtx_lock(&racct_lock);
596 racct_alloc_resource(p->p_racct, resource, amount);
597 mtx_unlock(&racct_lock);
598 racct_add_cred(p->p_ucred, resource, amount);
602 racct_set_locked(struct proc *p, int resource, uint64_t amount)
604 int64_t old_amount, decayed_amount;
605 int64_t diff_proc, diff_cred;
610 SDT_PROBE(racct, kernel, rusage, set, p, resource, amount, 0, 0);
613 * We need proc lock to dereference p->p_ucred.
615 PROC_LOCK_ASSERT(p, MA_OWNED);
617 old_amount = p->p_racct->r_resources[resource];
619 * The diffs may be negative.
621 diff_proc = amount - old_amount;
622 if (RACCT_IS_DECAYING(resource)) {
624 * Resources in per-credential racct containers may decay.
625 * If this is the case, we need to calculate the difference
626 * between the new amount and the proportional value of the
627 * old amount that has decayed in the ucred racct containers.
629 decayed_amount = old_amount * RACCT_DECAY_FACTOR / FSCALE;
630 diff_cred = amount - decayed_amount;
632 diff_cred = diff_proc;
634 KASSERT(diff_proc >= 0 || RACCT_CAN_DROP(resource),
635 ("racct_set: usage of non-droppable resource %d dropping",
640 error = rctl_enforce(p, resource, diff_proc);
641 if (error && RACCT_IS_DENIABLE(resource)) {
642 SDT_PROBE(racct, kernel, rusage, set_failure, p,
643 resource, amount, 0, 0);
648 racct_alloc_resource(p->p_racct, resource, diff_proc);
650 racct_add_cred_locked(p->p_ucred, resource, diff_cred);
651 else if (diff_cred < 0)
652 racct_sub_cred_locked(p->p_ucred, resource, -diff_cred);
658 * Set allocation of 'resource' to 'amount' for process 'p'.
659 * Return 0 if it's below limits, or errno, if it's not.
661 * Note that decreasing the allocation always returns 0,
662 * even if it's above the limit.
665 racct_set(struct proc *p, int resource, uint64_t amount)
669 mtx_lock(&racct_lock);
670 error = racct_set_locked(p, resource, amount);
671 mtx_unlock(&racct_lock);
676 racct_set_force_locked(struct proc *p, int resource, uint64_t amount)
678 int64_t old_amount, decayed_amount;
679 int64_t diff_proc, diff_cred;
681 SDT_PROBE(racct, kernel, rusage, set, p, resource, amount, 0, 0);
684 * We need proc lock to dereference p->p_ucred.
686 PROC_LOCK_ASSERT(p, MA_OWNED);
688 old_amount = p->p_racct->r_resources[resource];
690 * The diffs may be negative.
692 diff_proc = amount - old_amount;
693 if (RACCT_IS_DECAYING(resource)) {
695 * Resources in per-credential racct containers may decay.
696 * If this is the case, we need to calculate the difference
697 * between the new amount and the proportional value of the
698 * old amount that has decayed in the ucred racct containers.
700 decayed_amount = old_amount * RACCT_DECAY_FACTOR / FSCALE;
701 diff_cred = amount - decayed_amount;
703 diff_cred = diff_proc;
705 racct_alloc_resource(p->p_racct, resource, diff_proc);
707 racct_add_cred_locked(p->p_ucred, resource, diff_cred);
708 else if (diff_cred < 0)
709 racct_sub_cred_locked(p->p_ucred, resource, -diff_cred);
713 racct_set_force(struct proc *p, int resource, uint64_t amount)
715 mtx_lock(&racct_lock);
716 racct_set_force_locked(p, resource, amount);
717 mtx_unlock(&racct_lock);
721 * Returns amount of 'resource' the process 'p' can keep allocated.
722 * Allocating more than that would be denied, unless the resource
723 * is marked undeniable. Amount of already allocated resource does
727 racct_get_limit(struct proc *p, int resource)
731 return (rctl_get_limit(p, resource));
738 * Returns amount of 'resource' the process 'p' can keep allocated.
739 * Allocating more than that would be denied, unless the resource
740 * is marked undeniable. Amount of already allocated resource does
744 racct_get_available(struct proc *p, int resource)
748 return (rctl_get_available(p, resource));
755 * Returns amount of the %cpu resource that process 'p' can add to its %cpu
756 * utilization. Adding more than that would lead to the process being
760 racct_pcpu_available(struct proc *p)
764 return (rctl_pcpu_available(p));
771 * Decrease allocation of 'resource' by 'amount' for process 'p'.
774 racct_sub(struct proc *p, int resource, uint64_t amount)
777 SDT_PROBE(racct, kernel, rusage, sub, p, resource, amount, 0, 0);
780 * We need proc lock to dereference p->p_ucred.
782 PROC_LOCK_ASSERT(p, MA_OWNED);
783 KASSERT(RACCT_CAN_DROP(resource),
784 ("racct_sub: called for non-droppable resource %d", resource));
786 mtx_lock(&racct_lock);
787 KASSERT(amount <= p->p_racct->r_resources[resource],
788 ("racct_sub: freeing %ju of resource %d, which is more "
789 "than allocated %jd for %s (pid %d)", amount, resource,
790 (intmax_t)p->p_racct->r_resources[resource], p->p_comm, p->p_pid));
792 racct_alloc_resource(p->p_racct, resource, -amount);
793 racct_sub_cred_locked(p->p_ucred, resource, amount);
794 mtx_unlock(&racct_lock);
798 racct_sub_cred_locked(struct ucred *cred, int resource, uint64_t amount)
802 SDT_PROBE(racct, kernel, rusage, sub_cred, cred, resource, amount,
806 KASSERT(RACCT_CAN_DROP(resource),
807 ("racct_sub_cred: called for resource %d which can not drop",
811 racct_alloc_resource(cred->cr_ruidinfo->ui_racct, resource, -amount);
812 for (pr = cred->cr_prison; pr != NULL; pr = pr->pr_parent)
813 racct_alloc_resource(pr->pr_prison_racct->prr_racct, resource,
815 racct_alloc_resource(cred->cr_loginclass->lc_racct, resource, -amount);
819 * Decrease allocation of 'resource' by 'amount' for credential 'cred'.
822 racct_sub_cred(struct ucred *cred, int resource, uint64_t amount)
825 mtx_lock(&racct_lock);
826 racct_sub_cred_locked(cred, resource, amount);
827 mtx_unlock(&racct_lock);
831 * Inherit resource usage information from the parent process.
834 racct_proc_fork(struct proc *parent, struct proc *child)
839 * Create racct for the child process.
841 racct_create(&child->p_racct);
845 mtx_lock(&racct_lock);
848 error = rctl_proc_fork(parent, child);
853 /* Init process cpu time. */
854 child->p_prev_runtime = 0;
855 child->p_throttled = 0;
858 * Inherit resource usage.
860 for (i = 0; i <= RACCT_MAX; i++) {
861 if (parent->p_racct->r_resources[i] == 0 ||
862 !RACCT_IS_INHERITABLE(i))
865 error = racct_set_locked(child, i,
866 parent->p_racct->r_resources[i]);
871 error = racct_add_locked(child, RACCT_NPROC, 1);
872 error += racct_add_locked(child, RACCT_NTHR, 1);
875 mtx_unlock(&racct_lock);
880 racct_proc_exit(child);
886 * Called at the end of fork1(), to handle rules that require the process
887 * to be fully initialized.
890 racct_proc_fork_done(struct proc *child)
895 mtx_lock(&racct_lock);
896 rctl_enforce(child, RACCT_NPROC, 0);
897 rctl_enforce(child, RACCT_NTHR, 0);
898 mtx_unlock(&racct_lock);
904 racct_proc_exit(struct proc *p)
908 struct timeval wallclock;
909 uint64_t pct_estimate, pct;
913 * We don't need to calculate rux, proc_reap() has already done this.
915 runtime = cputick2usec(p->p_rux.rux_runtime);
917 KASSERT(runtime >= p->p_prev_runtime, ("runtime < p_prev_runtime"));
919 if (runtime < p->p_prev_runtime)
920 runtime = p->p_prev_runtime;
922 microuptime(&wallclock);
923 timevalsub(&wallclock, &p->p_stats->p_start);
924 if (wallclock.tv_sec > 0 || wallclock.tv_usec > 0) {
925 pct_estimate = (1000000 * runtime * 100) /
926 ((uint64_t)wallclock.tv_sec * 1000000 +
930 pct = racct_getpcpu(p, pct_estimate);
932 mtx_lock(&racct_lock);
933 racct_set_locked(p, RACCT_CPU, runtime);
934 racct_add_cred_locked(p->p_ucred, RACCT_PCTCPU, pct);
936 for (i = 0; i <= RACCT_MAX; i++) {
937 if (p->p_racct->r_resources[i] == 0)
939 if (!RACCT_IS_RECLAIMABLE(i))
941 racct_set_locked(p, i, 0);
944 mtx_unlock(&racct_lock);
948 rctl_racct_release(p->p_racct);
950 racct_destroy(&p->p_racct);
954 * Called after credentials change, to move resource utilisation
958 racct_proc_ucred_changed(struct proc *p, struct ucred *oldcred,
959 struct ucred *newcred)
961 struct uidinfo *olduip, *newuip;
962 struct loginclass *oldlc, *newlc;
963 struct prison *oldpr, *newpr, *pr;
965 PROC_LOCK_ASSERT(p, MA_NOTOWNED);
967 newuip = newcred->cr_ruidinfo;
968 olduip = oldcred->cr_ruidinfo;
969 newlc = newcred->cr_loginclass;
970 oldlc = oldcred->cr_loginclass;
971 newpr = newcred->cr_prison;
972 oldpr = oldcred->cr_prison;
974 mtx_lock(&racct_lock);
975 if (newuip != olduip) {
976 racct_sub_racct(olduip->ui_racct, p->p_racct);
977 racct_add_racct(newuip->ui_racct, p->p_racct);
979 if (newlc != oldlc) {
980 racct_sub_racct(oldlc->lc_racct, p->p_racct);
981 racct_add_racct(newlc->lc_racct, p->p_racct);
983 if (newpr != oldpr) {
984 for (pr = oldpr; pr != NULL; pr = pr->pr_parent)
985 racct_sub_racct(pr->pr_prison_racct->prr_racct,
987 for (pr = newpr; pr != NULL; pr = pr->pr_parent)
988 racct_add_racct(pr->pr_prison_racct->prr_racct,
991 mtx_unlock(&racct_lock);
994 rctl_proc_ucred_changed(p, newcred);
999 racct_move(struct racct *dest, struct racct *src)
1002 mtx_lock(&racct_lock);
1004 racct_add_racct(dest, src);
1005 racct_sub_racct(src, src);
1007 mtx_unlock(&racct_lock);
1011 racct_proc_throttle(struct proc *p)
1018 PROC_LOCK_ASSERT(p, MA_OWNED);
1021 * Do not block kernel processes. Also do not block processes with
1022 * low %cpu utilization to improve interactivity.
1024 if (((p->p_flag & (P_SYSTEM | P_KTHREAD)) != 0) ||
1025 (p->p_racct->r_resources[RACCT_PCTCPU] <= pcpu_threshold))
1029 FOREACH_THREAD_IN_PROC(p, td) {
1030 switch (td->td_state) {
1033 * If the thread is on the scheduler run-queue, we can
1034 * not just remove it from there. So we set the flag
1035 * TDF_NEEDRESCHED for the thread, so that once it is
1036 * running, it is taken off the cpu as soon as possible.
1039 td->td_flags |= TDF_NEEDRESCHED;
1044 * If the thread is running, we request a context
1045 * switch for it by setting the TDF_NEEDRESCHED flag.
1048 td->td_flags |= TDF_NEEDRESCHED;
1050 cpuid = td->td_oncpu;
1051 if ((cpuid != NOCPU) && (td != curthread))
1052 ipi_cpu(cpuid, IPI_AST);
1063 racct_proc_wakeup(struct proc *p)
1065 PROC_LOCK_ASSERT(p, MA_OWNED);
1067 if (p->p_throttled) {
1074 racct_decay_resource(struct racct *racct, void * res, void* dummy)
1077 int64_t r_old, r_new;
1079 resource = *(int *)res;
1080 r_old = racct->r_resources[resource];
1082 /* If there is nothing to decay, just exit. */
1086 mtx_lock(&racct_lock);
1087 r_new = r_old * RACCT_DECAY_FACTOR / FSCALE;
1088 racct->r_resources[resource] = r_new;
1089 mtx_unlock(&racct_lock);
1093 racct_decay(int resource)
1095 ui_racct_foreach(racct_decay_resource, &resource, NULL);
1096 loginclass_racct_foreach(racct_decay_resource, &resource, NULL);
1097 prison_racct_foreach(racct_decay_resource, &resource, NULL);
1105 struct timeval wallclock;
1107 uint64_t pct, pct_estimate;
1110 racct_decay(RACCT_PCTCPU);
1112 sx_slock(&allproc_lock);
1114 LIST_FOREACH(p, &zombproc, p_list) {
1116 racct_set(p, RACCT_PCTCPU, 0);
1120 FOREACH_PROC_IN_SYSTEM(p) {
1122 if (p->p_state != PRS_NORMAL) {
1127 microuptime(&wallclock);
1128 timevalsub(&wallclock, &p->p_stats->p_start);
1130 FOREACH_THREAD_IN_PROC(p, td)
1132 runtime = cputick2usec(p->p_rux.rux_runtime);
1135 KASSERT(runtime >= p->p_prev_runtime,
1136 ("runtime < p_prev_runtime"));
1138 if (runtime < p->p_prev_runtime)
1139 runtime = p->p_prev_runtime;
1141 p->p_prev_runtime = runtime;
1142 if (wallclock.tv_sec > 0 || wallclock.tv_usec > 0) {
1143 pct_estimate = (1000000 * runtime * 100) /
1144 ((uint64_t)wallclock.tv_sec * 1000000 +
1148 pct = racct_getpcpu(p, pct_estimate);
1149 mtx_lock(&racct_lock);
1150 racct_set_force_locked(p, RACCT_PCTCPU, pct);
1151 racct_set_locked(p, RACCT_CPU, runtime);
1152 racct_set_locked(p, RACCT_WALLCLOCK,
1153 (uint64_t)wallclock.tv_sec * 1000000 +
1155 mtx_unlock(&racct_lock);
1160 * To ensure that processes are throttled in a fair way, we need
1161 * to iterate over all processes again and check the limits
1162 * for %cpu resource only after ucred racct containers have been
1165 FOREACH_PROC_IN_SYSTEM(p) {
1167 if (p->p_state != PRS_NORMAL) {
1172 if (racct_pcpu_available(p) <= 0)
1173 racct_proc_throttle(p);
1174 else if (p->p_throttled)
1175 racct_proc_wakeup(p);
1178 sx_sunlock(&allproc_lock);
1183 static struct kproc_desc racctd_kp = {
1188 SYSINIT(racctd, SI_SUB_RACCTD, SI_ORDER_FIRST, kproc_start, &racctd_kp);
1194 racct_zone = uma_zcreate("racct", sizeof(struct racct),
1195 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
1197 * XXX: Move this somewhere.
1199 prison0.pr_prison_racct = prison_racct_find("0");
1201 SYSINIT(racct, SI_SUB_RACCT, SI_ORDER_FIRST, racct_init, NULL);
1206 racct_add(struct proc *p, int resource, uint64_t amount)
1213 racct_add_cred(struct ucred *cred, int resource, uint64_t amount)
1218 racct_add_force(struct proc *p, int resource, uint64_t amount)
1225 racct_set(struct proc *p, int resource, uint64_t amount)
1232 racct_set_force(struct proc *p, int resource, uint64_t amount)
1237 racct_sub(struct proc *p, int resource, uint64_t amount)
1242 racct_sub_cred(struct ucred *cred, int resource, uint64_t amount)
1247 racct_get_limit(struct proc *p, int resource)
1250 return (UINT64_MAX);
1254 racct_get_available(struct proc *p, int resource)
1257 return (UINT64_MAX);
1261 racct_create(struct racct **racctp)
1266 racct_destroy(struct racct **racctp)
1271 racct_proc_fork(struct proc *parent, struct proc *child)
1278 racct_proc_fork_done(struct proc *child)
1283 racct_proc_exit(struct proc *p)