2 * SPDX-License-Identifier: BSD-2-Clause
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
31 #include <sys/cdefs.h>
32 __FBSDID("$FreeBSD$");
34 #include "opt_sched.h"
36 #include <sys/param.h>
38 #include <sys/systm.h>
39 #include <sys/eventhandler.h>
41 #include <sys/kernel.h>
42 #include <sys/kthread.h>
44 #include <sys/loginclass.h>
45 #include <sys/malloc.h>
46 #include <sys/mutex.h>
48 #include <sys/racct.h>
49 #include <sys/resourcevar.h>
51 #include <sys/sched.h>
55 #include <sys/sysctl.h>
56 #include <sys/sysproto.h>
57 #include <sys/umtxvar.h>
58 #include <machine/smp.h>
66 FEATURE(racct, "Resource Accounting");
69 * Do not block processes that have their %cpu usage <= pcpu_threshold.
71 static int pcpu_threshold = 1;
72 #ifdef RACCT_DEFAULT_TO_DISABLED
73 bool __read_frequently racct_enable = false;
75 bool __read_frequently racct_enable = true;
78 SYSCTL_NODE(_kern, OID_AUTO, racct, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
79 "Resource Accounting");
80 SYSCTL_BOOL(_kern_racct, OID_AUTO, enable, CTLFLAG_RDTUN, &racct_enable,
81 0, "Enable RACCT/RCTL");
82 SYSCTL_UINT(_kern_racct, OID_AUTO, pcpu_threshold, CTLFLAG_RW, &pcpu_threshold,
83 0, "Processes with higher %cpu usage than this value can be throttled.");
86 * How many seconds it takes to use the scheduler %cpu calculations. When a
87 * process starts, we compute its %cpu usage by dividing its runtime by the
88 * process wall clock time. After RACCT_PCPU_SECS pass, we use the value
89 * provided by the scheduler.
91 #define RACCT_PCPU_SECS 3
93 struct mtx racct_lock;
94 MTX_SYSINIT(racct_lock, &racct_lock, "racct lock", MTX_DEF);
96 static uma_zone_t racct_zone;
98 static void racct_sub_racct(struct racct *dest, const struct racct *src);
99 static void racct_sub_cred_locked(struct ucred *cred, int resource,
101 static void racct_add_cred_locked(struct ucred *cred, int resource,
104 SDT_PROVIDER_DEFINE(racct);
105 SDT_PROBE_DEFINE3(racct, , rusage, add,
106 "struct proc *", "int", "uint64_t");
107 SDT_PROBE_DEFINE3(racct, , rusage, add__failure,
108 "struct proc *", "int", "uint64_t");
109 SDT_PROBE_DEFINE3(racct, , rusage, add__buf,
110 "struct proc *", "const struct buf *", "int");
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, set__force,
120 "struct proc *", "int", "uint64_t");
121 SDT_PROBE_DEFINE3(racct, , rusage, sub,
122 "struct proc *", "int", "uint64_t");
123 SDT_PROBE_DEFINE3(racct, , rusage, sub__cred,
124 "struct ucred *", "int", "uint64_t");
125 SDT_PROBE_DEFINE1(racct, , racct, create,
127 SDT_PROBE_DEFINE1(racct, , racct, destroy,
129 SDT_PROBE_DEFINE2(racct, , racct, join,
130 "struct racct *", "struct racct *");
131 SDT_PROBE_DEFINE2(racct, , racct, join__failure,
132 "struct racct *", "struct racct *");
133 SDT_PROBE_DEFINE2(racct, , racct, leave,
134 "struct racct *", "struct racct *");
136 int racct_types[] = {
140 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
142 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
148 RACCT_RECLAIMABLE | RACCT_DENIABLE,
150 RACCT_RECLAIMABLE | RACCT_DENIABLE,
152 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
154 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
156 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
158 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
160 RACCT_RECLAIMABLE | RACCT_DENIABLE,
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_DENIABLE | RACCT_SLOPPY,
170 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
172 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
174 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
178 RACCT_DECAYING | RACCT_DENIABLE | RACCT_IN_MILLIONS,
188 static const fixpt_t RACCT_DECAY_FACTOR = 0.3 * FSCALE;
192 * Contains intermediate values for %cpu calculations to avoid using floating
193 * point in the kernel.
194 * ccpu_exp[k] = FSCALE * (ccpu/FSCALE)^k = FSCALE * exp(-k/20)
195 * It is needed only for the 4BSD scheduler, because in ULE, the ccpu equals to
196 * zero so the calculations are more straightforward.
198 fixpt_t ccpu_exp[] = {
200 [1] = FSCALE * 0.95122942450071400909,
201 [2] = FSCALE * 0.90483741803595957316,
202 [3] = FSCALE * 0.86070797642505780722,
203 [4] = FSCALE * 0.81873075307798185866,
204 [5] = FSCALE * 0.77880078307140486824,
205 [6] = FSCALE * 0.74081822068171786606,
206 [7] = FSCALE * 0.70468808971871343435,
207 [8] = FSCALE * 0.67032004603563930074,
208 [9] = FSCALE * 0.63762815162177329314,
209 [10] = FSCALE * 0.60653065971263342360,
210 [11] = FSCALE * 0.57694981038048669531,
211 [12] = FSCALE * 0.54881163609402643262,
212 [13] = FSCALE * 0.52204577676101604789,
213 [14] = FSCALE * 0.49658530379140951470,
214 [15] = FSCALE * 0.47236655274101470713,
215 [16] = FSCALE * 0.44932896411722159143,
216 [17] = FSCALE * 0.42741493194872666992,
217 [18] = FSCALE * 0.40656965974059911188,
218 [19] = FSCALE * 0.38674102345450120691,
219 [20] = FSCALE * 0.36787944117144232159,
220 [21] = FSCALE * 0.34993774911115535467,
221 [22] = FSCALE * 0.33287108369807955328,
222 [23] = FSCALE * 0.31663676937905321821,
223 [24] = FSCALE * 0.30119421191220209664,
224 [25] = FSCALE * 0.28650479686019010032,
225 [26] = FSCALE * 0.27253179303401260312,
226 [27] = FSCALE * 0.25924026064589150757,
227 [28] = FSCALE * 0.24659696394160647693,
228 [29] = FSCALE * 0.23457028809379765313,
229 [30] = FSCALE * 0.22313016014842982893,
230 [31] = FSCALE * 0.21224797382674305771,
231 [32] = FSCALE * 0.20189651799465540848,
232 [33] = FSCALE * 0.19204990862075411423,
233 [34] = FSCALE * 0.18268352405273465022,
234 [35] = FSCALE * 0.17377394345044512668,
235 [36] = FSCALE * 0.16529888822158653829,
236 [37] = FSCALE * 0.15723716631362761621,
237 [38] = FSCALE * 0.14956861922263505264,
238 [39] = FSCALE * 0.14227407158651357185,
239 [40] = FSCALE * 0.13533528323661269189,
240 [41] = FSCALE * 0.12873490358780421886,
241 [42] = FSCALE * 0.12245642825298191021,
242 [43] = FSCALE * 0.11648415777349695786,
243 [44] = FSCALE * 0.11080315836233388333,
244 [45] = FSCALE * 0.10539922456186433678,
245 [46] = FSCALE * 0.10025884372280373372,
246 [47] = FSCALE * 0.09536916221554961888,
247 [48] = FSCALE * 0.09071795328941250337,
248 [49] = FSCALE * 0.08629358649937051097,
249 [50] = FSCALE * 0.08208499862389879516,
250 [51] = FSCALE * 0.07808166600115315231,
251 [52] = FSCALE * 0.07427357821433388042,
252 [53] = FSCALE * 0.07065121306042958674,
253 [54] = FSCALE * 0.06720551273974976512,
254 [55] = FSCALE * 0.06392786120670757270,
255 [56] = FSCALE * 0.06081006262521796499,
256 [57] = FSCALE * 0.05784432087483846296,
257 [58] = FSCALE * 0.05502322005640722902,
258 [59] = FSCALE * 0.05233970594843239308,
259 [60] = FSCALE * 0.04978706836786394297,
260 [61] = FSCALE * 0.04735892439114092119,
261 [62] = FSCALE * 0.04504920239355780606,
262 [63] = FSCALE * 0.04285212686704017991,
263 [64] = FSCALE * 0.04076220397836621516,
264 [65] = FSCALE * 0.03877420783172200988,
265 [66] = FSCALE * 0.03688316740124000544,
266 [67] = FSCALE * 0.03508435410084502588,
267 [68] = FSCALE * 0.03337326996032607948,
268 [69] = FSCALE * 0.03174563637806794323,
269 [70] = FSCALE * 0.03019738342231850073,
270 [71] = FSCALE * 0.02872463965423942912,
271 [72] = FSCALE * 0.02732372244729256080,
272 [73] = FSCALE * 0.02599112877875534358,
273 [74] = FSCALE * 0.02472352647033939120,
274 [75] = FSCALE * 0.02351774585600910823,
275 [76] = FSCALE * 0.02237077185616559577,
276 [77] = FSCALE * 0.02127973643837716938,
277 [78] = FSCALE * 0.02024191144580438847,
278 [79] = FSCALE * 0.01925470177538692429,
279 [80] = FSCALE * 0.01831563888873418029,
280 [81] = FSCALE * 0.01742237463949351138,
281 [82] = FSCALE * 0.01657267540176124754,
282 [83] = FSCALE * 0.01576441648485449082,
283 [84] = FSCALE * 0.01499557682047770621,
284 [85] = FSCALE * 0.01426423390899925527,
285 [86] = FSCALE * 0.01356855901220093175,
286 [87] = FSCALE * 0.01290681258047986886,
287 [88] = FSCALE * 0.01227733990306844117,
288 [89] = FSCALE * 0.01167856697039544521,
289 [90] = FSCALE * 0.01110899653824230649,
290 [91] = FSCALE * 0.01056720438385265337,
291 [92] = FSCALE * 0.01005183574463358164,
292 [93] = FSCALE * 0.00956160193054350793,
293 [94] = FSCALE * 0.00909527710169581709,
294 [95] = FSCALE * 0.00865169520312063417,
295 [96] = FSCALE * 0.00822974704902002884,
296 [97] = FSCALE * 0.00782837754922577143,
297 [98] = FSCALE * 0.00744658307092434051,
298 [99] = FSCALE * 0.00708340892905212004,
299 [100] = FSCALE * 0.00673794699908546709,
300 [101] = FSCALE * 0.00640933344625638184,
301 [102] = FSCALE * 0.00609674656551563610,
302 [103] = FSCALE * 0.00579940472684214321,
303 [104] = FSCALE * 0.00551656442076077241,
304 [105] = FSCALE * 0.00524751839918138427,
305 [106] = FSCALE * 0.00499159390691021621,
306 [107] = FSCALE * 0.00474815099941147558,
307 [108] = FSCALE * 0.00451658094261266798,
308 [109] = FSCALE * 0.00429630469075234057,
309 [110] = FSCALE * 0.00408677143846406699,
313 #define CCPU_EXP_MAX 110
316 * This function is analogical to the getpcpu() function in the ps(1) command.
317 * They should both calculate in the same way so that the racct %cpu
318 * calculations are consistent with the values showed by the ps(1) tool.
319 * The calculations are more complex in the 4BSD scheduler because of the value
320 * of the ccpu variable. In ULE it is defined to be zero which saves us some
324 racct_getpcpu(struct proc *p, u_int pcpu)
328 fixpt_t pctcpu, pctcpu_next;
337 ASSERT_RACCT_ENABLED();
340 * If the process is swapped out, we count its %cpu usage as zero.
341 * This behaviour is consistent with the userland ps(1) tool.
343 if ((p->p_flag & P_INMEM) == 0)
345 swtime = (ticks - p->p_swtick) / hz;
348 * For short-lived processes, the sched_pctcpu() returns small
349 * values even for cpu intensive processes. Therefore we use
350 * our own estimate in this case.
352 if (swtime < RACCT_PCPU_SECS)
356 FOREACH_THREAD_IN_PROC(p, td) {
357 if (td == PCPU_GET(idlethread))
361 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
362 if (td == pc->pc_idlethread) {
372 pctcpu = sched_pctcpu(td);
373 /* Count also the yet unfinished second. */
374 pctcpu_next = (pctcpu * ccpu_exp[1]) >> FSHIFT;
375 pctcpu_next += sched_pctcpu_delta(td);
376 p_pctcpu += max(pctcpu, pctcpu_next);
379 * In ULE the %cpu statistics are updated on every
380 * sched_pctcpu() call. So special calculations to
381 * account for the latest (unfinished) second are
384 p_pctcpu += sched_pctcpu(td);
390 if (swtime <= CCPU_EXP_MAX)
391 return ((100 * (uint64_t)p_pctcpu * 1000000) /
392 (FSCALE - ccpu_exp[swtime]));
395 return ((100 * (uint64_t)p_pctcpu * 1000000) / FSCALE);
399 racct_add_racct(struct racct *dest, const struct racct *src)
403 ASSERT_RACCT_ENABLED();
407 * Update resource usage in dest.
409 for (i = 0; i <= RACCT_MAX; i++) {
410 KASSERT(dest->r_resources[i] >= 0,
411 ("%s: resource %d propagation meltdown: dest < 0",
413 KASSERT(src->r_resources[i] >= 0,
414 ("%s: resource %d propagation meltdown: src < 0",
416 dest->r_resources[i] += src->r_resources[i];
421 racct_sub_racct(struct racct *dest, const struct racct *src)
425 ASSERT_RACCT_ENABLED();
429 * Update resource usage in dest.
431 for (i = 0; i <= RACCT_MAX; i++) {
432 if (!RACCT_IS_SLOPPY(i) && !RACCT_IS_DECAYING(i)) {
433 KASSERT(dest->r_resources[i] >= 0,
434 ("%s: resource %d propagation meltdown: dest < 0",
436 KASSERT(src->r_resources[i] >= 0,
437 ("%s: resource %d propagation meltdown: src < 0",
439 KASSERT(src->r_resources[i] <= dest->r_resources[i],
440 ("%s: resource %d propagation meltdown: src > dest",
443 if (RACCT_CAN_DROP(i)) {
444 dest->r_resources[i] -= src->r_resources[i];
445 if (dest->r_resources[i] < 0)
446 dest->r_resources[i] = 0;
452 racct_create(struct racct **racctp)
458 SDT_PROBE1(racct, , racct, create, racctp);
460 KASSERT(*racctp == NULL, ("racct already allocated"));
462 *racctp = uma_zalloc(racct_zone, M_WAITOK | M_ZERO);
466 racct_destroy_locked(struct racct **racctp)
471 ASSERT_RACCT_ENABLED();
473 SDT_PROBE1(racct, , racct, destroy, racctp);
476 KASSERT(racctp != NULL, ("NULL racctp"));
477 KASSERT(*racctp != NULL, ("NULL racct"));
481 for (i = 0; i <= RACCT_MAX; i++) {
482 if (RACCT_IS_SLOPPY(i))
484 if (!RACCT_IS_RECLAIMABLE(i))
486 KASSERT(racct->r_resources[i] == 0,
487 ("destroying non-empty racct: "
488 "%ju allocated for resource %d\n",
489 racct->r_resources[i], i));
491 uma_zfree(racct_zone, racct);
496 racct_destroy(struct racct **racct)
503 racct_destroy_locked(racct);
508 * Increase consumption of 'resource' by 'amount' for 'racct',
509 * but not its parents. Differently from other cases, 'amount' here
510 * may be less than zero.
513 racct_adjust_resource(struct racct *racct, int resource,
517 ASSERT_RACCT_ENABLED();
519 KASSERT(racct != NULL, ("NULL racct"));
521 racct->r_resources[resource] += amount;
522 if (racct->r_resources[resource] < 0) {
523 KASSERT(RACCT_IS_SLOPPY(resource) || RACCT_IS_DECAYING(resource),
524 ("%s: resource %d usage < 0", __func__, resource));
525 racct->r_resources[resource] = 0;
529 * There are some cases where the racct %cpu resource would grow
530 * beyond 100% per core. For example in racct_proc_exit() we add
531 * the process %cpu usage to the ucred racct containers. If too
532 * many processes terminated in a short time span, the ucred %cpu
533 * resource could grow too much. Also, the 4BSD scheduler sometimes
534 * returns for a thread more than 100% cpu usage. So we set a sane
535 * boundary here to 100% * the maximum number of CPUs.
537 if ((resource == RACCT_PCTCPU) &&
538 (racct->r_resources[RACCT_PCTCPU] > 100 * 1000000 * (int64_t)MAXCPU))
539 racct->r_resources[RACCT_PCTCPU] = 100 * 1000000 * (int64_t)MAXCPU;
543 racct_add_locked(struct proc *p, int resource, uint64_t amount, int force)
549 ASSERT_RACCT_ENABLED();
552 * We need proc lock to dereference p->p_ucred.
554 PROC_LOCK_ASSERT(p, MA_OWNED);
557 error = rctl_enforce(p, resource, amount);
558 if (error && !force && RACCT_IS_DENIABLE(resource)) {
559 SDT_PROBE3(racct, , rusage, add__failure, p, resource, amount);
563 racct_adjust_resource(p->p_racct, resource, amount);
564 racct_add_cred_locked(p->p_ucred, resource, amount);
570 * Increase allocation of 'resource' by 'amount' for process 'p'.
571 * Return 0 if it's below limits, or errno, if it's not.
574 racct_add(struct proc *p, int resource, uint64_t amount)
581 SDT_PROBE3(racct, , rusage, add, p, resource, amount);
584 error = racct_add_locked(p, resource, amount, 0);
590 * Increase allocation of 'resource' by 'amount' for process 'p'.
591 * Doesn't check for limits and never fails.
594 racct_add_force(struct proc *p, int resource, uint64_t amount)
600 SDT_PROBE3(racct, , rusage, add__force, p, resource, amount);
603 racct_add_locked(p, resource, amount, 1);
608 racct_add_cred_locked(struct ucred *cred, int resource, uint64_t amount)
612 ASSERT_RACCT_ENABLED();
614 racct_adjust_resource(cred->cr_ruidinfo->ui_racct, resource, amount);
615 for (pr = cred->cr_prison; pr != NULL; pr = pr->pr_parent)
616 racct_adjust_resource(pr->pr_prison_racct->prr_racct, resource,
618 racct_adjust_resource(cred->cr_loginclass->lc_racct, resource, amount);
622 * Increase allocation of 'resource' by 'amount' for credential 'cred'.
623 * Doesn't check for limits and never fails.
626 racct_add_cred(struct ucred *cred, int resource, uint64_t amount)
632 SDT_PROBE3(racct, , rusage, add__cred, cred, resource, amount);
635 racct_add_cred_locked(cred, resource, amount);
640 * Account for disk IO resource consumption. Checks for limits,
641 * but never fails, due to disk limits being undeniable.
644 racct_add_buf(struct proc *p, const struct buf *bp, int is_write)
647 ASSERT_RACCT_ENABLED();
648 PROC_LOCK_ASSERT(p, MA_OWNED);
650 SDT_PROBE3(racct, , rusage, add__buf, p, bp, is_write);
654 racct_add_locked(curproc, RACCT_WRITEBPS, bp->b_bcount, 1);
655 racct_add_locked(curproc, RACCT_WRITEIOPS, 1, 1);
657 racct_add_locked(curproc, RACCT_READBPS, bp->b_bcount, 1);
658 racct_add_locked(curproc, RACCT_READIOPS, 1, 1);
664 racct_set_locked(struct proc *p, int resource, uint64_t amount, int force)
666 int64_t old_amount, decayed_amount, diff_proc, diff_cred;
671 ASSERT_RACCT_ENABLED();
674 * We need proc lock to dereference p->p_ucred.
676 PROC_LOCK_ASSERT(p, MA_OWNED);
678 old_amount = p->p_racct->r_resources[resource];
680 * The diffs may be negative.
682 diff_proc = amount - old_amount;
683 if (resource == RACCT_PCTCPU) {
685 * Resources in per-credential racct containers may decay.
686 * If this is the case, we need to calculate the difference
687 * between the new amount and the proportional value of the
688 * old amount that has decayed in the ucred racct containers.
690 decayed_amount = old_amount * RACCT_DECAY_FACTOR / FSCALE;
691 diff_cred = amount - decayed_amount;
693 diff_cred = diff_proc;
695 KASSERT(diff_proc >= 0 || RACCT_CAN_DROP(resource),
696 ("%s: usage of non-droppable resource %d dropping", __func__,
701 error = rctl_enforce(p, resource, diff_proc);
702 if (error && !force && RACCT_IS_DENIABLE(resource)) {
703 SDT_PROBE3(racct, , rusage, set__failure, p, resource,
709 racct_adjust_resource(p->p_racct, resource, diff_proc);
711 racct_add_cred_locked(p->p_ucred, resource, diff_cred);
712 else if (diff_cred < 0)
713 racct_sub_cred_locked(p->p_ucred, resource, -diff_cred);
719 * Set allocation of 'resource' to 'amount' for process 'p'.
720 * Return 0 if it's below limits, or errno, if it's not.
722 * Note that decreasing the allocation always returns 0,
723 * even if it's above the limit.
726 racct_set_unlocked(struct proc *p, int resource, uint64_t amount)
730 ASSERT_RACCT_ENABLED();
732 error = racct_set(p, resource, amount);
738 racct_set(struct proc *p, int resource, uint64_t amount)
745 SDT_PROBE3(racct, , rusage, set__force, p, resource, amount);
748 error = racct_set_locked(p, resource, amount, 0);
754 racct_set_force(struct proc *p, int resource, uint64_t amount)
760 SDT_PROBE3(racct, , rusage, set, p, resource, amount);
763 racct_set_locked(p, resource, amount, 1);
768 * Returns amount of 'resource' the process 'p' can keep allocated.
769 * Allocating more than that would be denied, unless the resource
770 * is marked undeniable. Amount of already allocated resource does
774 racct_get_limit(struct proc *p, int resource)
783 available = rctl_get_limit(p, resource);
794 * Returns amount of 'resource' the process 'p' can keep allocated.
795 * Allocating more than that would be denied, unless the resource
796 * is marked undeniable. Amount of already allocated resource does
800 racct_get_available(struct proc *p, int resource)
809 available = rctl_get_available(p, resource);
820 * Returns amount of the %cpu resource that process 'p' can add to its %cpu
821 * utilization. Adding more than that would lead to the process being
825 racct_pcpu_available(struct proc *p)
830 ASSERT_RACCT_ENABLED();
833 available = rctl_pcpu_available(p);
844 * Decrease allocation of 'resource' by 'amount' for process 'p'.
847 racct_sub(struct proc *p, int resource, uint64_t amount)
853 SDT_PROBE3(racct, , rusage, sub, p, resource, amount);
856 * We need proc lock to dereference p->p_ucred.
858 PROC_LOCK_ASSERT(p, MA_OWNED);
859 KASSERT(RACCT_CAN_DROP(resource),
860 ("%s: called for non-droppable resource %d", __func__, resource));
863 KASSERT(amount <= p->p_racct->r_resources[resource],
864 ("%s: freeing %ju of resource %d, which is more "
865 "than allocated %jd for %s (pid %d)", __func__, amount, resource,
866 (intmax_t)p->p_racct->r_resources[resource], p->p_comm, p->p_pid));
868 racct_adjust_resource(p->p_racct, resource, -amount);
869 racct_sub_cred_locked(p->p_ucred, resource, amount);
874 racct_sub_cred_locked(struct ucred *cred, int resource, uint64_t amount)
878 ASSERT_RACCT_ENABLED();
880 racct_adjust_resource(cred->cr_ruidinfo->ui_racct, resource, -amount);
881 for (pr = cred->cr_prison; pr != NULL; pr = pr->pr_parent)
882 racct_adjust_resource(pr->pr_prison_racct->prr_racct, resource,
884 racct_adjust_resource(cred->cr_loginclass->lc_racct, resource, -amount);
888 * Decrease allocation of 'resource' by 'amount' for credential 'cred'.
891 racct_sub_cred(struct ucred *cred, int resource, uint64_t amount)
897 SDT_PROBE3(racct, , rusage, sub__cred, cred, resource, amount);
900 KASSERT(RACCT_CAN_DROP(resource),
901 ("%s: called for resource %d which can not drop", __func__,
906 racct_sub_cred_locked(cred, resource, amount);
911 * Inherit resource usage information from the parent process.
914 racct_proc_fork(struct proc *parent, struct proc *child)
922 * Create racct for the child process.
924 racct_create(&child->p_racct);
931 error = rctl_proc_fork(parent, child);
936 /* Init process cpu time. */
937 child->p_prev_runtime = 0;
938 child->p_throttled = 0;
941 * Inherit resource usage.
943 for (i = 0; i <= RACCT_MAX; i++) {
944 if (parent->p_racct->r_resources[i] == 0 ||
945 !RACCT_IS_INHERITABLE(i))
948 error = racct_set_locked(child, i,
949 parent->p_racct->r_resources[i], 0);
954 error = racct_add_locked(child, RACCT_NPROC, 1, 0);
955 error += racct_add_locked(child, RACCT_NTHR, 1, 0);
963 racct_proc_exit(child);
969 * Called at the end of fork1(), to handle rules that require the process
970 * to be fully initialized.
973 racct_proc_fork_done(struct proc *child)
982 rctl_enforce(child, RACCT_NPROC, 0);
983 rctl_enforce(child, RACCT_NTHR, 0);
990 racct_proc_exit(struct proc *p)
992 struct timeval wallclock;
993 uint64_t pct_estimate, pct, runtime;
1001 * We don't need to calculate rux, proc_reap() has already done this.
1003 runtime = cputick2usec(p->p_rux.rux_runtime);
1005 KASSERT(runtime >= p->p_prev_runtime, ("runtime < p_prev_runtime"));
1007 if (runtime < p->p_prev_runtime)
1008 runtime = p->p_prev_runtime;
1010 microuptime(&wallclock);
1011 timevalsub(&wallclock, &p->p_stats->p_start);
1012 if (wallclock.tv_sec > 0 || wallclock.tv_usec > 0) {
1013 pct_estimate = (1000000 * runtime * 100) /
1014 ((uint64_t)wallclock.tv_sec * 1000000 +
1018 pct = racct_getpcpu(p, pct_estimate);
1021 racct_set_locked(p, RACCT_CPU, runtime, 0);
1022 racct_add_cred_locked(p->p_ucred, RACCT_PCTCPU, pct);
1024 KASSERT(p->p_racct->r_resources[RACCT_RSS] == 0,
1025 ("process reaped with %ju allocated for RSS\n",
1026 p->p_racct->r_resources[RACCT_RSS]));
1027 for (i = 0; i <= RACCT_MAX; i++) {
1028 if (p->p_racct->r_resources[i] == 0)
1030 if (!RACCT_IS_RECLAIMABLE(i))
1032 racct_set_locked(p, i, 0, 0);
1036 rctl_racct_release(p->p_racct);
1038 racct_destroy_locked(&p->p_racct);
1044 * Called after credentials change, to move resource utilisation
1048 racct_proc_ucred_changed(struct proc *p, struct ucred *oldcred,
1049 struct ucred *newcred)
1051 struct uidinfo *olduip, *newuip;
1052 struct loginclass *oldlc, *newlc;
1053 struct prison *oldpr, *newpr, *pr;
1058 PROC_LOCK_ASSERT(p, MA_OWNED);
1060 newuip = newcred->cr_ruidinfo;
1061 olduip = oldcred->cr_ruidinfo;
1062 newlc = newcred->cr_loginclass;
1063 oldlc = oldcred->cr_loginclass;
1064 newpr = newcred->cr_prison;
1065 oldpr = oldcred->cr_prison;
1068 if (newuip != olduip) {
1069 racct_sub_racct(olduip->ui_racct, p->p_racct);
1070 racct_add_racct(newuip->ui_racct, p->p_racct);
1072 if (newlc != oldlc) {
1073 racct_sub_racct(oldlc->lc_racct, p->p_racct);
1074 racct_add_racct(newlc->lc_racct, p->p_racct);
1076 if (newpr != oldpr) {
1077 for (pr = oldpr; pr != NULL; pr = pr->pr_parent)
1078 racct_sub_racct(pr->pr_prison_racct->prr_racct,
1080 for (pr = newpr; pr != NULL; pr = pr->pr_parent)
1081 racct_add_racct(pr->pr_prison_racct->prr_racct,
1088 racct_move(struct racct *dest, struct racct *src)
1091 ASSERT_RACCT_ENABLED();
1094 racct_add_racct(dest, src);
1095 racct_sub_racct(src, src);
1100 ast_racct(struct thread *td, int tda __unused)
1104 ASSERT_RACCT_ENABLED();
1107 if (p->p_throttled == 0)
1111 while (p->p_throttled != 0) {
1112 msleep(p->p_racct, &p->p_mtx, 0, "racct",
1113 p->p_throttled < 0 ? 0 : p->p_throttled);
1114 if (p->p_throttled > 0)
1121 * Make the process sleep in userret() for 'timeout' ticks. Setting
1122 * timeout to -1 makes it sleep until woken up by racct_proc_wakeup().
1125 racct_proc_throttle(struct proc *p, int timeout)
1132 KASSERT(timeout != 0, ("timeout %d", timeout));
1133 ASSERT_RACCT_ENABLED();
1134 PROC_LOCK_ASSERT(p, MA_OWNED);
1137 * Do not block kernel processes. Also do not block processes with
1138 * low %cpu utilization to improve interactivity.
1140 if ((p->p_flag & (P_SYSTEM | P_KPROC)) != 0)
1143 if (p->p_throttled < 0 || (timeout > 0 && p->p_throttled > timeout))
1146 p->p_throttled = timeout;
1148 FOREACH_THREAD_IN_PROC(p, td) {
1150 ast_sched_locked(td, TDA_RACCT);
1152 switch (TD_GET_STATE(td)) {
1155 * If the thread is on the scheduler run-queue, we can
1156 * not just remove it from there. So we set the flag
1157 * TDA_SCHED for the thread, so that once it is
1158 * running, it is taken off the cpu as soon as possible.
1160 ast_sched_locked(td, TDA_SCHED);
1164 * If the thread is running, we request a context
1165 * switch for it by setting the TDA_SCHED flag.
1167 ast_sched_locked(td, TDA_SCHED);
1169 cpuid = td->td_oncpu;
1170 if ((cpuid != NOCPU) && (td != curthread))
1171 ipi_cpu(cpuid, IPI_AST);
1182 racct_proc_wakeup(struct proc *p)
1185 ASSERT_RACCT_ENABLED();
1187 PROC_LOCK_ASSERT(p, MA_OWNED);
1189 if (p->p_throttled != 0) {
1196 racct_decay_callback(struct racct *racct, void *dummy1, void *dummy2)
1198 int64_t r_old, r_new;
1200 ASSERT_RACCT_ENABLED();
1201 RACCT_LOCK_ASSERT();
1204 rctl_throttle_decay(racct, RACCT_READBPS);
1205 rctl_throttle_decay(racct, RACCT_WRITEBPS);
1206 rctl_throttle_decay(racct, RACCT_READIOPS);
1207 rctl_throttle_decay(racct, RACCT_WRITEIOPS);
1210 r_old = racct->r_resources[RACCT_PCTCPU];
1212 /* If there is nothing to decay, just exit. */
1216 r_new = r_old * RACCT_DECAY_FACTOR / FSCALE;
1217 racct->r_resources[RACCT_PCTCPU] = r_new;
1221 racct_decay_pre(void)
1228 racct_decay_post(void)
1238 ASSERT_RACCT_ENABLED();
1240 ui_racct_foreach(racct_decay_callback, racct_decay_pre,
1241 racct_decay_post, NULL, NULL);
1242 loginclass_racct_foreach(racct_decay_callback, racct_decay_pre,
1243 racct_decay_post, NULL, NULL);
1244 prison_racct_foreach(racct_decay_callback, racct_decay_pre,
1245 racct_decay_post, NULL, NULL);
1253 struct timeval wallclock;
1254 uint64_t pct, pct_estimate, runtime;
1256 ASSERT_RACCT_ENABLED();
1261 sx_slock(&allproc_lock);
1263 FOREACH_PROC_IN_SYSTEM(p) {
1265 if (p->p_state != PRS_NORMAL) {
1266 if (p->p_state == PRS_ZOMBIE)
1267 racct_set(p, RACCT_PCTCPU, 0);
1272 microuptime(&wallclock);
1273 timevalsub(&wallclock, &p->p_stats->p_start);
1275 FOREACH_THREAD_IN_PROC(p, td)
1277 runtime = cputick2usec(p->p_rux.rux_runtime);
1280 KASSERT(runtime >= p->p_prev_runtime,
1281 ("runtime < p_prev_runtime"));
1283 if (runtime < p->p_prev_runtime)
1284 runtime = p->p_prev_runtime;
1286 p->p_prev_runtime = runtime;
1287 if (wallclock.tv_sec > 0 || wallclock.tv_usec > 0) {
1288 pct_estimate = (1000000 * runtime * 100) /
1289 ((uint64_t)wallclock.tv_sec * 1000000 +
1293 pct = racct_getpcpu(p, pct_estimate);
1296 rctl_throttle_decay(p->p_racct, RACCT_READBPS);
1297 rctl_throttle_decay(p->p_racct, RACCT_WRITEBPS);
1298 rctl_throttle_decay(p->p_racct, RACCT_READIOPS);
1299 rctl_throttle_decay(p->p_racct, RACCT_WRITEIOPS);
1301 racct_set_locked(p, RACCT_PCTCPU, pct, 1);
1302 racct_set_locked(p, RACCT_CPU, runtime, 0);
1303 racct_set_locked(p, RACCT_WALLCLOCK,
1304 (uint64_t)wallclock.tv_sec * 1000000 +
1305 wallclock.tv_usec, 0);
1311 * To ensure that processes are throttled in a fair way, we need
1312 * to iterate over all processes again and check the limits
1313 * for %cpu resource only after ucred racct containers have been
1316 FOREACH_PROC_IN_SYSTEM(p) {
1318 if (p->p_state != PRS_NORMAL) {
1323 if (racct_pcpu_available(p) <= 0) {
1324 if (p->p_racct->r_resources[RACCT_PCTCPU] >
1326 racct_proc_throttle(p, -1);
1327 } else if (p->p_throttled == -1) {
1328 racct_proc_wakeup(p);
1332 sx_sunlock(&allproc_lock);
1337 static struct kproc_desc racctd_kp = {
1349 kproc_start(&racctd_kp);
1351 SYSINIT(racctd, SI_SUB_RACCTD, SI_ORDER_FIRST, racctd_init, NULL);
1359 racct_zone = uma_zcreate("racct", sizeof(struct racct),
1360 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
1361 ast_register(TDA_RACCT, ASTR_ASTF_REQUIRED, 0, ast_racct);
1364 * XXX: Move this somewhere.
1366 prison0.pr_prison_racct = prison_racct_find("0");
1368 SYSINIT(racct, SI_SUB_RACCT, SI_ORDER_FIRST, racct_init, NULL);