2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2010 The FreeBSD Foundation
6 * This software was developed by Edward Tomasz Napierala under sponsorship
7 * from the FreeBSD Foundation.
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in the
16 * documentation and/or other materials provided with the distribution.
18 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
19 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
22 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD$");
36 #include "opt_sched.h"
38 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/eventhandler.h>
43 #include <sys/kernel.h>
44 #include <sys/kthread.h>
46 #include <sys/loginclass.h>
47 #include <sys/malloc.h>
48 #include <sys/mutex.h>
50 #include <sys/racct.h>
51 #include <sys/resourcevar.h>
53 #include <sys/sched.h>
57 #include <sys/sysctl.h>
58 #include <sys/sysproto.h>
59 #include <sys/umtxvar.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
75 bool __read_frequently racct_enable = false;
77 bool __read_frequently racct_enable = true;
80 SYSCTL_NODE(_kern, OID_AUTO, racct, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
81 "Resource Accounting");
82 SYSCTL_BOOL(_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 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__buf,
112 "struct proc *", "const struct buf *", "int");
113 SDT_PROBE_DEFINE3(racct, , rusage, add__cred,
114 "struct ucred *", "int", "uint64_t");
115 SDT_PROBE_DEFINE3(racct, , rusage, add__force,
116 "struct proc *", "int", "uint64_t");
117 SDT_PROBE_DEFINE3(racct, , rusage, set,
118 "struct proc *", "int", "uint64_t");
119 SDT_PROBE_DEFINE3(racct, , rusage, set__failure,
120 "struct proc *", "int", "uint64_t");
121 SDT_PROBE_DEFINE3(racct, , rusage, set__force,
122 "struct proc *", "int", "uint64_t");
123 SDT_PROBE_DEFINE3(racct, , rusage, sub,
124 "struct proc *", "int", "uint64_t");
125 SDT_PROBE_DEFINE3(racct, , rusage, sub__cred,
126 "struct ucred *", "int", "uint64_t");
127 SDT_PROBE_DEFINE1(racct, , racct, create,
129 SDT_PROBE_DEFINE1(racct, , racct, destroy,
131 SDT_PROBE_DEFINE2(racct, , racct, join,
132 "struct racct *", "struct racct *");
133 SDT_PROBE_DEFINE2(racct, , racct, join__failure,
134 "struct racct *", "struct racct *");
135 SDT_PROBE_DEFINE2(racct, , racct, leave,
136 "struct racct *", "struct racct *");
138 int racct_types[] = {
142 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
144 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
150 RACCT_RECLAIMABLE | RACCT_DENIABLE,
152 RACCT_RECLAIMABLE | RACCT_DENIABLE,
154 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
156 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
158 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
160 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
162 RACCT_RECLAIMABLE | RACCT_DENIABLE,
164 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
166 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
168 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
170 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
172 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
174 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
176 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
180 RACCT_DECAYING | RACCT_DENIABLE | RACCT_IN_MILLIONS,
190 static const fixpt_t RACCT_DECAY_FACTOR = 0.3 * FSCALE;
194 * Contains intermediate values for %cpu calculations to avoid using floating
195 * point in the kernel.
196 * ccpu_exp[k] = FSCALE * (ccpu/FSCALE)^k = FSCALE * exp(-k/20)
197 * It is needed only for the 4BSD scheduler, because in ULE, the ccpu equals to
198 * zero so the calculations are more straightforward.
200 fixpt_t ccpu_exp[] = {
202 [1] = FSCALE * 0.95122942450071400909,
203 [2] = FSCALE * 0.90483741803595957316,
204 [3] = FSCALE * 0.86070797642505780722,
205 [4] = FSCALE * 0.81873075307798185866,
206 [5] = FSCALE * 0.77880078307140486824,
207 [6] = FSCALE * 0.74081822068171786606,
208 [7] = FSCALE * 0.70468808971871343435,
209 [8] = FSCALE * 0.67032004603563930074,
210 [9] = FSCALE * 0.63762815162177329314,
211 [10] = FSCALE * 0.60653065971263342360,
212 [11] = FSCALE * 0.57694981038048669531,
213 [12] = FSCALE * 0.54881163609402643262,
214 [13] = FSCALE * 0.52204577676101604789,
215 [14] = FSCALE * 0.49658530379140951470,
216 [15] = FSCALE * 0.47236655274101470713,
217 [16] = FSCALE * 0.44932896411722159143,
218 [17] = FSCALE * 0.42741493194872666992,
219 [18] = FSCALE * 0.40656965974059911188,
220 [19] = FSCALE * 0.38674102345450120691,
221 [20] = FSCALE * 0.36787944117144232159,
222 [21] = FSCALE * 0.34993774911115535467,
223 [22] = FSCALE * 0.33287108369807955328,
224 [23] = FSCALE * 0.31663676937905321821,
225 [24] = FSCALE * 0.30119421191220209664,
226 [25] = FSCALE * 0.28650479686019010032,
227 [26] = FSCALE * 0.27253179303401260312,
228 [27] = FSCALE * 0.25924026064589150757,
229 [28] = FSCALE * 0.24659696394160647693,
230 [29] = FSCALE * 0.23457028809379765313,
231 [30] = FSCALE * 0.22313016014842982893,
232 [31] = FSCALE * 0.21224797382674305771,
233 [32] = FSCALE * 0.20189651799465540848,
234 [33] = FSCALE * 0.19204990862075411423,
235 [34] = FSCALE * 0.18268352405273465022,
236 [35] = FSCALE * 0.17377394345044512668,
237 [36] = FSCALE * 0.16529888822158653829,
238 [37] = FSCALE * 0.15723716631362761621,
239 [38] = FSCALE * 0.14956861922263505264,
240 [39] = FSCALE * 0.14227407158651357185,
241 [40] = FSCALE * 0.13533528323661269189,
242 [41] = FSCALE * 0.12873490358780421886,
243 [42] = FSCALE * 0.12245642825298191021,
244 [43] = FSCALE * 0.11648415777349695786,
245 [44] = FSCALE * 0.11080315836233388333,
246 [45] = FSCALE * 0.10539922456186433678,
247 [46] = FSCALE * 0.10025884372280373372,
248 [47] = FSCALE * 0.09536916221554961888,
249 [48] = FSCALE * 0.09071795328941250337,
250 [49] = FSCALE * 0.08629358649937051097,
251 [50] = FSCALE * 0.08208499862389879516,
252 [51] = FSCALE * 0.07808166600115315231,
253 [52] = FSCALE * 0.07427357821433388042,
254 [53] = FSCALE * 0.07065121306042958674,
255 [54] = FSCALE * 0.06720551273974976512,
256 [55] = FSCALE * 0.06392786120670757270,
257 [56] = FSCALE * 0.06081006262521796499,
258 [57] = FSCALE * 0.05784432087483846296,
259 [58] = FSCALE * 0.05502322005640722902,
260 [59] = FSCALE * 0.05233970594843239308,
261 [60] = FSCALE * 0.04978706836786394297,
262 [61] = FSCALE * 0.04735892439114092119,
263 [62] = FSCALE * 0.04504920239355780606,
264 [63] = FSCALE * 0.04285212686704017991,
265 [64] = FSCALE * 0.04076220397836621516,
266 [65] = FSCALE * 0.03877420783172200988,
267 [66] = FSCALE * 0.03688316740124000544,
268 [67] = FSCALE * 0.03508435410084502588,
269 [68] = FSCALE * 0.03337326996032607948,
270 [69] = FSCALE * 0.03174563637806794323,
271 [70] = FSCALE * 0.03019738342231850073,
272 [71] = FSCALE * 0.02872463965423942912,
273 [72] = FSCALE * 0.02732372244729256080,
274 [73] = FSCALE * 0.02599112877875534358,
275 [74] = FSCALE * 0.02472352647033939120,
276 [75] = FSCALE * 0.02351774585600910823,
277 [76] = FSCALE * 0.02237077185616559577,
278 [77] = FSCALE * 0.02127973643837716938,
279 [78] = FSCALE * 0.02024191144580438847,
280 [79] = FSCALE * 0.01925470177538692429,
281 [80] = FSCALE * 0.01831563888873418029,
282 [81] = FSCALE * 0.01742237463949351138,
283 [82] = FSCALE * 0.01657267540176124754,
284 [83] = FSCALE * 0.01576441648485449082,
285 [84] = FSCALE * 0.01499557682047770621,
286 [85] = FSCALE * 0.01426423390899925527,
287 [86] = FSCALE * 0.01356855901220093175,
288 [87] = FSCALE * 0.01290681258047986886,
289 [88] = FSCALE * 0.01227733990306844117,
290 [89] = FSCALE * 0.01167856697039544521,
291 [90] = FSCALE * 0.01110899653824230649,
292 [91] = FSCALE * 0.01056720438385265337,
293 [92] = FSCALE * 0.01005183574463358164,
294 [93] = FSCALE * 0.00956160193054350793,
295 [94] = FSCALE * 0.00909527710169581709,
296 [95] = FSCALE * 0.00865169520312063417,
297 [96] = FSCALE * 0.00822974704902002884,
298 [97] = FSCALE * 0.00782837754922577143,
299 [98] = FSCALE * 0.00744658307092434051,
300 [99] = FSCALE * 0.00708340892905212004,
301 [100] = FSCALE * 0.00673794699908546709,
302 [101] = FSCALE * 0.00640933344625638184,
303 [102] = FSCALE * 0.00609674656551563610,
304 [103] = FSCALE * 0.00579940472684214321,
305 [104] = FSCALE * 0.00551656442076077241,
306 [105] = FSCALE * 0.00524751839918138427,
307 [106] = FSCALE * 0.00499159390691021621,
308 [107] = FSCALE * 0.00474815099941147558,
309 [108] = FSCALE * 0.00451658094261266798,
310 [109] = FSCALE * 0.00429630469075234057,
311 [110] = FSCALE * 0.00408677143846406699,
315 #define CCPU_EXP_MAX 110
318 * This function is analogical to the getpcpu() function in the ps(1) command.
319 * They should both calculate in the same way so that the racct %cpu
320 * calculations are consistent with the values showed by the ps(1) tool.
321 * The calculations are more complex in the 4BSD scheduler because of the value
322 * of the ccpu variable. In ULE it is defined to be zero which saves us some
326 racct_getpcpu(struct proc *p, u_int pcpu)
330 fixpt_t pctcpu, pctcpu_next;
339 ASSERT_RACCT_ENABLED();
342 * If the process is swapped out, we count its %cpu usage as zero.
343 * This behaviour is consistent with the userland ps(1) tool.
345 if ((p->p_flag & P_INMEM) == 0)
347 swtime = (ticks - p->p_swtick) / hz;
350 * For short-lived processes, the sched_pctcpu() returns small
351 * values even for cpu intensive processes. Therefore we use
352 * our own estimate in this case.
354 if (swtime < RACCT_PCPU_SECS)
358 FOREACH_THREAD_IN_PROC(p, td) {
359 if (td == PCPU_GET(idlethread))
363 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
364 if (td == pc->pc_idlethread) {
374 pctcpu = sched_pctcpu(td);
375 /* Count also the yet unfinished second. */
376 pctcpu_next = (pctcpu * ccpu_exp[1]) >> FSHIFT;
377 pctcpu_next += sched_pctcpu_delta(td);
378 p_pctcpu += max(pctcpu, pctcpu_next);
381 * In ULE the %cpu statistics are updated on every
382 * sched_pctcpu() call. So special calculations to
383 * account for the latest (unfinished) second are
386 p_pctcpu += sched_pctcpu(td);
392 if (swtime <= CCPU_EXP_MAX)
393 return ((100 * (uint64_t)p_pctcpu * 1000000) /
394 (FSCALE - ccpu_exp[swtime]));
397 return ((100 * (uint64_t)p_pctcpu * 1000000) / FSCALE);
401 racct_add_racct(struct racct *dest, const struct racct *src)
405 ASSERT_RACCT_ENABLED();
409 * Update resource usage in dest.
411 for (i = 0; i <= RACCT_MAX; i++) {
412 KASSERT(dest->r_resources[i] >= 0,
413 ("%s: resource %d propagation meltdown: dest < 0",
415 KASSERT(src->r_resources[i] >= 0,
416 ("%s: resource %d propagation meltdown: src < 0",
418 dest->r_resources[i] += src->r_resources[i];
423 racct_sub_racct(struct racct *dest, const struct racct *src)
427 ASSERT_RACCT_ENABLED();
431 * Update resource usage in dest.
433 for (i = 0; i <= RACCT_MAX; i++) {
434 if (!RACCT_IS_SLOPPY(i) && !RACCT_IS_DECAYING(i)) {
435 KASSERT(dest->r_resources[i] >= 0,
436 ("%s: resource %d propagation meltdown: dest < 0",
438 KASSERT(src->r_resources[i] >= 0,
439 ("%s: resource %d propagation meltdown: src < 0",
441 KASSERT(src->r_resources[i] <= dest->r_resources[i],
442 ("%s: resource %d propagation meltdown: src > dest",
445 if (RACCT_CAN_DROP(i)) {
446 dest->r_resources[i] -= src->r_resources[i];
447 if (dest->r_resources[i] < 0)
448 dest->r_resources[i] = 0;
454 racct_create(struct racct **racctp)
460 SDT_PROBE1(racct, , racct, create, racctp);
462 KASSERT(*racctp == NULL, ("racct already allocated"));
464 *racctp = uma_zalloc(racct_zone, M_WAITOK | M_ZERO);
468 racct_destroy_locked(struct racct **racctp)
473 ASSERT_RACCT_ENABLED();
475 SDT_PROBE1(racct, , racct, destroy, racctp);
478 KASSERT(racctp != NULL, ("NULL racctp"));
479 KASSERT(*racctp != NULL, ("NULL racct"));
483 for (i = 0; i <= RACCT_MAX; i++) {
484 if (RACCT_IS_SLOPPY(i))
486 if (!RACCT_IS_RECLAIMABLE(i))
488 KASSERT(racct->r_resources[i] == 0,
489 ("destroying non-empty racct: "
490 "%ju allocated for resource %d\n",
491 racct->r_resources[i], i));
493 uma_zfree(racct_zone, racct);
498 racct_destroy(struct racct **racct)
505 racct_destroy_locked(racct);
510 * Increase consumption of 'resource' by 'amount' for 'racct',
511 * but not its parents. Differently from other cases, 'amount' here
512 * may be less than zero.
515 racct_adjust_resource(struct racct *racct, int resource,
519 ASSERT_RACCT_ENABLED();
521 KASSERT(racct != NULL, ("NULL racct"));
523 racct->r_resources[resource] += amount;
524 if (racct->r_resources[resource] < 0) {
525 KASSERT(RACCT_IS_SLOPPY(resource) || RACCT_IS_DECAYING(resource),
526 ("%s: resource %d usage < 0", __func__, resource));
527 racct->r_resources[resource] = 0;
531 * There are some cases where the racct %cpu resource would grow
532 * beyond 100% per core. For example in racct_proc_exit() we add
533 * the process %cpu usage to the ucred racct containers. If too
534 * many processes terminated in a short time span, the ucred %cpu
535 * resource could grow too much. Also, the 4BSD scheduler sometimes
536 * returns for a thread more than 100% cpu usage. So we set a sane
537 * boundary here to 100% * the maximum number of CPUs.
539 if ((resource == RACCT_PCTCPU) &&
540 (racct->r_resources[RACCT_PCTCPU] > 100 * 1000000 * (int64_t)MAXCPU))
541 racct->r_resources[RACCT_PCTCPU] = 100 * 1000000 * (int64_t)MAXCPU;
545 racct_add_locked(struct proc *p, int resource, uint64_t amount, int force)
551 ASSERT_RACCT_ENABLED();
554 * We need proc lock to dereference p->p_ucred.
556 PROC_LOCK_ASSERT(p, MA_OWNED);
559 error = rctl_enforce(p, resource, amount);
560 if (error && !force && RACCT_IS_DENIABLE(resource)) {
561 SDT_PROBE3(racct, , rusage, add__failure, p, resource, amount);
565 racct_adjust_resource(p->p_racct, resource, amount);
566 racct_add_cred_locked(p->p_ucred, resource, amount);
572 * Increase allocation of 'resource' by 'amount' for process 'p'.
573 * Return 0 if it's below limits, or errno, if it's not.
576 racct_add(struct proc *p, int resource, uint64_t amount)
583 SDT_PROBE3(racct, , rusage, add, p, resource, amount);
586 error = racct_add_locked(p, resource, amount, 0);
592 * Increase allocation of 'resource' by 'amount' for process 'p'.
593 * Doesn't check for limits and never fails.
596 racct_add_force(struct proc *p, int resource, uint64_t amount)
602 SDT_PROBE3(racct, , rusage, add__force, p, resource, amount);
605 racct_add_locked(p, resource, amount, 1);
610 racct_add_cred_locked(struct ucred *cred, int resource, uint64_t amount)
614 ASSERT_RACCT_ENABLED();
616 racct_adjust_resource(cred->cr_ruidinfo->ui_racct, resource, amount);
617 for (pr = cred->cr_prison; pr != NULL; pr = pr->pr_parent)
618 racct_adjust_resource(pr->pr_prison_racct->prr_racct, resource,
620 racct_adjust_resource(cred->cr_loginclass->lc_racct, resource, amount);
624 * Increase allocation of 'resource' by 'amount' for credential 'cred'.
625 * Doesn't check for limits and never fails.
628 racct_add_cred(struct ucred *cred, int resource, uint64_t amount)
634 SDT_PROBE3(racct, , rusage, add__cred, cred, resource, amount);
637 racct_add_cred_locked(cred, resource, amount);
642 * Account for disk IO resource consumption. Checks for limits,
643 * but never fails, due to disk limits being undeniable.
646 racct_add_buf(struct proc *p, const struct buf *bp, int is_write)
649 ASSERT_RACCT_ENABLED();
650 PROC_LOCK_ASSERT(p, MA_OWNED);
652 SDT_PROBE3(racct, , rusage, add__buf, p, bp, is_write);
656 racct_add_locked(curproc, RACCT_WRITEBPS, bp->b_bcount, 1);
657 racct_add_locked(curproc, RACCT_WRITEIOPS, 1, 1);
659 racct_add_locked(curproc, RACCT_READBPS, bp->b_bcount, 1);
660 racct_add_locked(curproc, RACCT_READIOPS, 1, 1);
666 racct_set_locked(struct proc *p, int resource, uint64_t amount, int force)
668 int64_t old_amount, decayed_amount, diff_proc, diff_cred;
673 ASSERT_RACCT_ENABLED();
676 * We need proc lock to dereference p->p_ucred.
678 PROC_LOCK_ASSERT(p, MA_OWNED);
680 old_amount = p->p_racct->r_resources[resource];
682 * The diffs may be negative.
684 diff_proc = amount - old_amount;
685 if (resource == RACCT_PCTCPU) {
687 * Resources in per-credential racct containers may decay.
688 * If this is the case, we need to calculate the difference
689 * between the new amount and the proportional value of the
690 * old amount that has decayed in the ucred racct containers.
692 decayed_amount = old_amount * RACCT_DECAY_FACTOR / FSCALE;
693 diff_cred = amount - decayed_amount;
695 diff_cred = diff_proc;
697 KASSERT(diff_proc >= 0 || RACCT_CAN_DROP(resource),
698 ("%s: usage of non-droppable resource %d dropping", __func__,
703 error = rctl_enforce(p, resource, diff_proc);
704 if (error && !force && RACCT_IS_DENIABLE(resource)) {
705 SDT_PROBE3(racct, , rusage, set__failure, p, resource,
711 racct_adjust_resource(p->p_racct, resource, diff_proc);
713 racct_add_cred_locked(p->p_ucred, resource, diff_cred);
714 else if (diff_cred < 0)
715 racct_sub_cred_locked(p->p_ucred, resource, -diff_cred);
721 * Set allocation of 'resource' to 'amount' for process 'p'.
722 * Return 0 if it's below limits, or errno, if it's not.
724 * Note that decreasing the allocation always returns 0,
725 * even if it's above the limit.
728 racct_set_unlocked(struct proc *p, int resource, uint64_t amount)
732 ASSERT_RACCT_ENABLED();
734 error = racct_set(p, resource, amount);
740 racct_set(struct proc *p, int resource, uint64_t amount)
747 SDT_PROBE3(racct, , rusage, set__force, p, resource, amount);
750 error = racct_set_locked(p, resource, amount, 0);
756 racct_set_force(struct proc *p, int resource, uint64_t amount)
762 SDT_PROBE3(racct, , rusage, set, p, resource, amount);
765 racct_set_locked(p, resource, amount, 1);
770 * Returns amount of 'resource' the process 'p' can keep allocated.
771 * Allocating more than that would be denied, unless the resource
772 * is marked undeniable. Amount of already allocated resource does
776 racct_get_limit(struct proc *p, int resource)
785 available = rctl_get_limit(p, resource);
796 * Returns amount of 'resource' the process 'p' can keep allocated.
797 * Allocating more than that would be denied, unless the resource
798 * is marked undeniable. Amount of already allocated resource does
802 racct_get_available(struct proc *p, int resource)
811 available = rctl_get_available(p, resource);
822 * Returns amount of the %cpu resource that process 'p' can add to its %cpu
823 * utilization. Adding more than that would lead to the process being
827 racct_pcpu_available(struct proc *p)
832 ASSERT_RACCT_ENABLED();
835 available = rctl_pcpu_available(p);
846 * Decrease allocation of 'resource' by 'amount' for process 'p'.
849 racct_sub(struct proc *p, int resource, uint64_t amount)
855 SDT_PROBE3(racct, , rusage, sub, p, resource, amount);
858 * We need proc lock to dereference p->p_ucred.
860 PROC_LOCK_ASSERT(p, MA_OWNED);
861 KASSERT(RACCT_CAN_DROP(resource),
862 ("%s: called for non-droppable resource %d", __func__, resource));
865 KASSERT(amount <= p->p_racct->r_resources[resource],
866 ("%s: freeing %ju of resource %d, which is more "
867 "than allocated %jd for %s (pid %d)", __func__, amount, resource,
868 (intmax_t)p->p_racct->r_resources[resource], p->p_comm, p->p_pid));
870 racct_adjust_resource(p->p_racct, resource, -amount);
871 racct_sub_cred_locked(p->p_ucred, resource, amount);
876 racct_sub_cred_locked(struct ucred *cred, int resource, uint64_t amount)
880 ASSERT_RACCT_ENABLED();
882 racct_adjust_resource(cred->cr_ruidinfo->ui_racct, resource, -amount);
883 for (pr = cred->cr_prison; pr != NULL; pr = pr->pr_parent)
884 racct_adjust_resource(pr->pr_prison_racct->prr_racct, resource,
886 racct_adjust_resource(cred->cr_loginclass->lc_racct, resource, -amount);
890 * Decrease allocation of 'resource' by 'amount' for credential 'cred'.
893 racct_sub_cred(struct ucred *cred, int resource, uint64_t amount)
899 SDT_PROBE3(racct, , rusage, sub__cred, cred, resource, amount);
902 KASSERT(RACCT_CAN_DROP(resource),
903 ("%s: called for resource %d which can not drop", __func__,
908 racct_sub_cred_locked(cred, resource, amount);
913 * Inherit resource usage information from the parent process.
916 racct_proc_fork(struct proc *parent, struct proc *child)
924 * Create racct for the child process.
926 racct_create(&child->p_racct);
933 error = rctl_proc_fork(parent, child);
938 /* Init process cpu time. */
939 child->p_prev_runtime = 0;
940 child->p_throttled = 0;
943 * Inherit resource usage.
945 for (i = 0; i <= RACCT_MAX; i++) {
946 if (parent->p_racct->r_resources[i] == 0 ||
947 !RACCT_IS_INHERITABLE(i))
950 error = racct_set_locked(child, i,
951 parent->p_racct->r_resources[i], 0);
956 error = racct_add_locked(child, RACCT_NPROC, 1, 0);
957 error += racct_add_locked(child, RACCT_NTHR, 1, 0);
965 racct_proc_exit(child);
971 * Called at the end of fork1(), to handle rules that require the process
972 * to be fully initialized.
975 racct_proc_fork_done(struct proc *child)
984 rctl_enforce(child, RACCT_NPROC, 0);
985 rctl_enforce(child, RACCT_NTHR, 0);
992 racct_proc_exit(struct proc *p)
994 struct timeval wallclock;
995 uint64_t pct_estimate, pct, runtime;
1003 * We don't need to calculate rux, proc_reap() has already done this.
1005 runtime = cputick2usec(p->p_rux.rux_runtime);
1007 KASSERT(runtime >= p->p_prev_runtime, ("runtime < p_prev_runtime"));
1009 if (runtime < p->p_prev_runtime)
1010 runtime = p->p_prev_runtime;
1012 microuptime(&wallclock);
1013 timevalsub(&wallclock, &p->p_stats->p_start);
1014 if (wallclock.tv_sec > 0 || wallclock.tv_usec > 0) {
1015 pct_estimate = (1000000 * runtime * 100) /
1016 ((uint64_t)wallclock.tv_sec * 1000000 +
1020 pct = racct_getpcpu(p, pct_estimate);
1023 racct_set_locked(p, RACCT_CPU, runtime, 0);
1024 racct_add_cred_locked(p->p_ucred, RACCT_PCTCPU, pct);
1026 KASSERT(p->p_racct->r_resources[RACCT_RSS] == 0,
1027 ("process reaped with %ju allocated for RSS\n",
1028 p->p_racct->r_resources[RACCT_RSS]));
1029 for (i = 0; i <= RACCT_MAX; i++) {
1030 if (p->p_racct->r_resources[i] == 0)
1032 if (!RACCT_IS_RECLAIMABLE(i))
1034 racct_set_locked(p, i, 0, 0);
1038 rctl_racct_release(p->p_racct);
1040 racct_destroy_locked(&p->p_racct);
1046 * Called after credentials change, to move resource utilisation
1050 racct_proc_ucred_changed(struct proc *p, struct ucred *oldcred,
1051 struct ucred *newcred)
1053 struct uidinfo *olduip, *newuip;
1054 struct loginclass *oldlc, *newlc;
1055 struct prison *oldpr, *newpr, *pr;
1060 PROC_LOCK_ASSERT(p, MA_OWNED);
1062 newuip = newcred->cr_ruidinfo;
1063 olduip = oldcred->cr_ruidinfo;
1064 newlc = newcred->cr_loginclass;
1065 oldlc = oldcred->cr_loginclass;
1066 newpr = newcred->cr_prison;
1067 oldpr = oldcred->cr_prison;
1070 if (newuip != olduip) {
1071 racct_sub_racct(olduip->ui_racct, p->p_racct);
1072 racct_add_racct(newuip->ui_racct, p->p_racct);
1074 if (newlc != oldlc) {
1075 racct_sub_racct(oldlc->lc_racct, p->p_racct);
1076 racct_add_racct(newlc->lc_racct, p->p_racct);
1078 if (newpr != oldpr) {
1079 for (pr = oldpr; pr != NULL; pr = pr->pr_parent)
1080 racct_sub_racct(pr->pr_prison_racct->prr_racct,
1082 for (pr = newpr; pr != NULL; pr = pr->pr_parent)
1083 racct_add_racct(pr->pr_prison_racct->prr_racct,
1090 racct_move(struct racct *dest, struct racct *src)
1093 ASSERT_RACCT_ENABLED();
1096 racct_add_racct(dest, src);
1097 racct_sub_racct(src, src);
1102 ast_racct(struct thread *td, int tda __unused)
1106 ASSERT_RACCT_ENABLED();
1109 if (p->p_throttled == 0)
1113 while (p->p_throttled != 0) {
1114 msleep(p->p_racct, &p->p_mtx, 0, "racct",
1115 p->p_throttled < 0 ? 0 : p->p_throttled);
1116 if (p->p_throttled > 0)
1123 * Make the process sleep in userret() for 'timeout' ticks. Setting
1124 * timeout to -1 makes it sleep until woken up by racct_proc_wakeup().
1127 racct_proc_throttle(struct proc *p, int timeout)
1134 KASSERT(timeout != 0, ("timeout %d", timeout));
1135 ASSERT_RACCT_ENABLED();
1136 PROC_LOCK_ASSERT(p, MA_OWNED);
1139 * Do not block kernel processes. Also do not block processes with
1140 * low %cpu utilization to improve interactivity.
1142 if ((p->p_flag & (P_SYSTEM | P_KPROC)) != 0)
1145 if (p->p_throttled < 0 || (timeout > 0 && p->p_throttled > timeout))
1148 p->p_throttled = timeout;
1150 FOREACH_THREAD_IN_PROC(p, td) {
1152 ast_sched_locked(td, TDA_RACCT);
1154 switch (TD_GET_STATE(td)) {
1157 * If the thread is on the scheduler run-queue, we can
1158 * not just remove it from there. So we set the flag
1159 * TDA_SCHED for the thread, so that once it is
1160 * running, it is taken off the cpu as soon as possible.
1162 ast_sched_locked(td, TDA_SCHED);
1166 * If the thread is running, we request a context
1167 * switch for it by setting the TDA_SCHED flag.
1169 ast_sched_locked(td, TDA_SCHED);
1171 cpuid = td->td_oncpu;
1172 if ((cpuid != NOCPU) && (td != curthread))
1173 ipi_cpu(cpuid, IPI_AST);
1184 racct_proc_wakeup(struct proc *p)
1187 ASSERT_RACCT_ENABLED();
1189 PROC_LOCK_ASSERT(p, MA_OWNED);
1191 if (p->p_throttled != 0) {
1198 racct_decay_callback(struct racct *racct, void *dummy1, void *dummy2)
1200 int64_t r_old, r_new;
1202 ASSERT_RACCT_ENABLED();
1203 RACCT_LOCK_ASSERT();
1206 rctl_throttle_decay(racct, RACCT_READBPS);
1207 rctl_throttle_decay(racct, RACCT_WRITEBPS);
1208 rctl_throttle_decay(racct, RACCT_READIOPS);
1209 rctl_throttle_decay(racct, RACCT_WRITEIOPS);
1212 r_old = racct->r_resources[RACCT_PCTCPU];
1214 /* If there is nothing to decay, just exit. */
1218 r_new = r_old * RACCT_DECAY_FACTOR / FSCALE;
1219 racct->r_resources[RACCT_PCTCPU] = r_new;
1223 racct_decay_pre(void)
1230 racct_decay_post(void)
1240 ASSERT_RACCT_ENABLED();
1242 ui_racct_foreach(racct_decay_callback, racct_decay_pre,
1243 racct_decay_post, NULL, NULL);
1244 loginclass_racct_foreach(racct_decay_callback, racct_decay_pre,
1245 racct_decay_post, NULL, NULL);
1246 prison_racct_foreach(racct_decay_callback, racct_decay_pre,
1247 racct_decay_post, NULL, NULL);
1255 struct timeval wallclock;
1256 uint64_t pct, pct_estimate, runtime;
1258 ASSERT_RACCT_ENABLED();
1263 sx_slock(&allproc_lock);
1265 FOREACH_PROC_IN_SYSTEM(p) {
1267 if (p->p_state != PRS_NORMAL) {
1268 if (p->p_state == PRS_ZOMBIE)
1269 racct_set(p, RACCT_PCTCPU, 0);
1274 microuptime(&wallclock);
1275 timevalsub(&wallclock, &p->p_stats->p_start);
1277 FOREACH_THREAD_IN_PROC(p, td)
1279 runtime = cputick2usec(p->p_rux.rux_runtime);
1282 KASSERT(runtime >= p->p_prev_runtime,
1283 ("runtime < p_prev_runtime"));
1285 if (runtime < p->p_prev_runtime)
1286 runtime = p->p_prev_runtime;
1288 p->p_prev_runtime = runtime;
1289 if (wallclock.tv_sec > 0 || wallclock.tv_usec > 0) {
1290 pct_estimate = (1000000 * runtime * 100) /
1291 ((uint64_t)wallclock.tv_sec * 1000000 +
1295 pct = racct_getpcpu(p, pct_estimate);
1298 rctl_throttle_decay(p->p_racct, RACCT_READBPS);
1299 rctl_throttle_decay(p->p_racct, RACCT_WRITEBPS);
1300 rctl_throttle_decay(p->p_racct, RACCT_READIOPS);
1301 rctl_throttle_decay(p->p_racct, RACCT_WRITEIOPS);
1303 racct_set_locked(p, RACCT_PCTCPU, pct, 1);
1304 racct_set_locked(p, RACCT_CPU, runtime, 0);
1305 racct_set_locked(p, RACCT_WALLCLOCK,
1306 (uint64_t)wallclock.tv_sec * 1000000 +
1307 wallclock.tv_usec, 0);
1313 * To ensure that processes are throttled in a fair way, we need
1314 * to iterate over all processes again and check the limits
1315 * for %cpu resource only after ucred racct containers have been
1318 FOREACH_PROC_IN_SYSTEM(p) {
1320 if (p->p_state != PRS_NORMAL) {
1325 if (racct_pcpu_available(p) <= 0) {
1326 if (p->p_racct->r_resources[RACCT_PCTCPU] >
1328 racct_proc_throttle(p, -1);
1329 } else if (p->p_throttled == -1) {
1330 racct_proc_wakeup(p);
1334 sx_sunlock(&allproc_lock);
1339 static struct kproc_desc racctd_kp = {
1351 kproc_start(&racctd_kp);
1353 SYSINIT(racctd, SI_SUB_RACCTD, SI_ORDER_FIRST, racctd_init, NULL);
1361 racct_zone = uma_zcreate("racct", sizeof(struct racct),
1362 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
1363 ast_register(TDA_RACCT, ASTR_ASTF_REQUIRED, 0, ast_racct);
1366 * XXX: Move this somewhere.
1368 prison0.pr_prison_racct = prison_racct_find("0");
1370 SYSINIT(racct, SI_SUB_RACCT, SI_ORDER_FIRST, racct_init, NULL);