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/sysent.h>
59 #include <sys/sysproto.h>
61 #include <machine/smp.h>
69 FEATURE(racct, "Resource Accounting");
72 * Do not block processes that have their %cpu usage <= pcpu_threshold.
74 static int pcpu_threshold = 1;
75 #ifdef RACCT_DEFAULT_TO_DISABLED
76 bool __read_frequently racct_enable = false;
78 bool __read_frequently racct_enable = true;
81 SYSCTL_NODE(_kern, OID_AUTO, racct, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
82 "Resource Accounting");
83 SYSCTL_BOOL(_kern_racct, OID_AUTO, enable, CTLFLAG_RDTUN, &racct_enable,
84 0, "Enable RACCT/RCTL");
85 SYSCTL_UINT(_kern_racct, OID_AUTO, pcpu_threshold, CTLFLAG_RW, &pcpu_threshold,
86 0, "Processes with higher %cpu usage than this value can be throttled.");
89 * How many seconds it takes to use the scheduler %cpu calculations. When a
90 * process starts, we compute its %cpu usage by dividing its runtime by the
91 * process wall clock time. After RACCT_PCPU_SECS pass, we use the value
92 * provided by the scheduler.
94 #define RACCT_PCPU_SECS 3
96 struct mtx racct_lock;
97 MTX_SYSINIT(racct_lock, &racct_lock, "racct lock", MTX_DEF);
99 static uma_zone_t racct_zone;
101 static void racct_sub_racct(struct racct *dest, const struct racct *src);
102 static void racct_sub_cred_locked(struct ucred *cred, int resource,
104 static void racct_add_cred_locked(struct ucred *cred, int resource,
107 SDT_PROVIDER_DEFINE(racct);
108 SDT_PROBE_DEFINE3(racct, , rusage, add,
109 "struct proc *", "int", "uint64_t");
110 SDT_PROBE_DEFINE3(racct, , rusage, add__failure,
111 "struct proc *", "int", "uint64_t");
112 SDT_PROBE_DEFINE3(racct, , rusage, add__buf,
113 "struct proc *", "const struct buf *", "int");
114 SDT_PROBE_DEFINE3(racct, , rusage, add__cred,
115 "struct ucred *", "int", "uint64_t");
116 SDT_PROBE_DEFINE3(racct, , rusage, add__force,
117 "struct proc *", "int", "uint64_t");
118 SDT_PROBE_DEFINE3(racct, , rusage, set,
119 "struct proc *", "int", "uint64_t");
120 SDT_PROBE_DEFINE3(racct, , rusage, set__failure,
121 "struct proc *", "int", "uint64_t");
122 SDT_PROBE_DEFINE3(racct, , rusage, set__force,
123 "struct proc *", "int", "uint64_t");
124 SDT_PROBE_DEFINE3(racct, , rusage, sub,
125 "struct proc *", "int", "uint64_t");
126 SDT_PROBE_DEFINE3(racct, , rusage, sub__cred,
127 "struct ucred *", "int", "uint64_t");
128 SDT_PROBE_DEFINE1(racct, , racct, create,
130 SDT_PROBE_DEFINE1(racct, , racct, destroy,
132 SDT_PROBE_DEFINE2(racct, , racct, join,
133 "struct racct *", "struct racct *");
134 SDT_PROBE_DEFINE2(racct, , racct, join__failure,
135 "struct racct *", "struct racct *");
136 SDT_PROBE_DEFINE2(racct, , racct, leave,
137 "struct racct *", "struct racct *");
139 int racct_types[] = {
143 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
145 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
151 RACCT_RECLAIMABLE | RACCT_DENIABLE,
153 RACCT_RECLAIMABLE | RACCT_DENIABLE,
155 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
157 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
159 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
161 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
163 RACCT_RECLAIMABLE | RACCT_DENIABLE,
165 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
167 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
169 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
171 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
173 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
175 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
177 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
181 RACCT_DECAYING | RACCT_DENIABLE | RACCT_IN_MILLIONS,
191 static const fixpt_t RACCT_DECAY_FACTOR = 0.3 * FSCALE;
195 * Contains intermediate values for %cpu calculations to avoid using floating
196 * point in the kernel.
197 * ccpu_exp[k] = FSCALE * (ccpu/FSCALE)^k = FSCALE * exp(-k/20)
198 * It is needed only for the 4BSD scheduler, because in ULE, the ccpu equals to
199 * zero so the calculations are more straightforward.
201 fixpt_t ccpu_exp[] = {
203 [1] = FSCALE * 0.95122942450071400909,
204 [2] = FSCALE * 0.90483741803595957316,
205 [3] = FSCALE * 0.86070797642505780722,
206 [4] = FSCALE * 0.81873075307798185866,
207 [5] = FSCALE * 0.77880078307140486824,
208 [6] = FSCALE * 0.74081822068171786606,
209 [7] = FSCALE * 0.70468808971871343435,
210 [8] = FSCALE * 0.67032004603563930074,
211 [9] = FSCALE * 0.63762815162177329314,
212 [10] = FSCALE * 0.60653065971263342360,
213 [11] = FSCALE * 0.57694981038048669531,
214 [12] = FSCALE * 0.54881163609402643262,
215 [13] = FSCALE * 0.52204577676101604789,
216 [14] = FSCALE * 0.49658530379140951470,
217 [15] = FSCALE * 0.47236655274101470713,
218 [16] = FSCALE * 0.44932896411722159143,
219 [17] = FSCALE * 0.42741493194872666992,
220 [18] = FSCALE * 0.40656965974059911188,
221 [19] = FSCALE * 0.38674102345450120691,
222 [20] = FSCALE * 0.36787944117144232159,
223 [21] = FSCALE * 0.34993774911115535467,
224 [22] = FSCALE * 0.33287108369807955328,
225 [23] = FSCALE * 0.31663676937905321821,
226 [24] = FSCALE * 0.30119421191220209664,
227 [25] = FSCALE * 0.28650479686019010032,
228 [26] = FSCALE * 0.27253179303401260312,
229 [27] = FSCALE * 0.25924026064589150757,
230 [28] = FSCALE * 0.24659696394160647693,
231 [29] = FSCALE * 0.23457028809379765313,
232 [30] = FSCALE * 0.22313016014842982893,
233 [31] = FSCALE * 0.21224797382674305771,
234 [32] = FSCALE * 0.20189651799465540848,
235 [33] = FSCALE * 0.19204990862075411423,
236 [34] = FSCALE * 0.18268352405273465022,
237 [35] = FSCALE * 0.17377394345044512668,
238 [36] = FSCALE * 0.16529888822158653829,
239 [37] = FSCALE * 0.15723716631362761621,
240 [38] = FSCALE * 0.14956861922263505264,
241 [39] = FSCALE * 0.14227407158651357185,
242 [40] = FSCALE * 0.13533528323661269189,
243 [41] = FSCALE * 0.12873490358780421886,
244 [42] = FSCALE * 0.12245642825298191021,
245 [43] = FSCALE * 0.11648415777349695786,
246 [44] = FSCALE * 0.11080315836233388333,
247 [45] = FSCALE * 0.10539922456186433678,
248 [46] = FSCALE * 0.10025884372280373372,
249 [47] = FSCALE * 0.09536916221554961888,
250 [48] = FSCALE * 0.09071795328941250337,
251 [49] = FSCALE * 0.08629358649937051097,
252 [50] = FSCALE * 0.08208499862389879516,
253 [51] = FSCALE * 0.07808166600115315231,
254 [52] = FSCALE * 0.07427357821433388042,
255 [53] = FSCALE * 0.07065121306042958674,
256 [54] = FSCALE * 0.06720551273974976512,
257 [55] = FSCALE * 0.06392786120670757270,
258 [56] = FSCALE * 0.06081006262521796499,
259 [57] = FSCALE * 0.05784432087483846296,
260 [58] = FSCALE * 0.05502322005640722902,
261 [59] = FSCALE * 0.05233970594843239308,
262 [60] = FSCALE * 0.04978706836786394297,
263 [61] = FSCALE * 0.04735892439114092119,
264 [62] = FSCALE * 0.04504920239355780606,
265 [63] = FSCALE * 0.04285212686704017991,
266 [64] = FSCALE * 0.04076220397836621516,
267 [65] = FSCALE * 0.03877420783172200988,
268 [66] = FSCALE * 0.03688316740124000544,
269 [67] = FSCALE * 0.03508435410084502588,
270 [68] = FSCALE * 0.03337326996032607948,
271 [69] = FSCALE * 0.03174563637806794323,
272 [70] = FSCALE * 0.03019738342231850073,
273 [71] = FSCALE * 0.02872463965423942912,
274 [72] = FSCALE * 0.02732372244729256080,
275 [73] = FSCALE * 0.02599112877875534358,
276 [74] = FSCALE * 0.02472352647033939120,
277 [75] = FSCALE * 0.02351774585600910823,
278 [76] = FSCALE * 0.02237077185616559577,
279 [77] = FSCALE * 0.02127973643837716938,
280 [78] = FSCALE * 0.02024191144580438847,
281 [79] = FSCALE * 0.01925470177538692429,
282 [80] = FSCALE * 0.01831563888873418029,
283 [81] = FSCALE * 0.01742237463949351138,
284 [82] = FSCALE * 0.01657267540176124754,
285 [83] = FSCALE * 0.01576441648485449082,
286 [84] = FSCALE * 0.01499557682047770621,
287 [85] = FSCALE * 0.01426423390899925527,
288 [86] = FSCALE * 0.01356855901220093175,
289 [87] = FSCALE * 0.01290681258047986886,
290 [88] = FSCALE * 0.01227733990306844117,
291 [89] = FSCALE * 0.01167856697039544521,
292 [90] = FSCALE * 0.01110899653824230649,
293 [91] = FSCALE * 0.01056720438385265337,
294 [92] = FSCALE * 0.01005183574463358164,
295 [93] = FSCALE * 0.00956160193054350793,
296 [94] = FSCALE * 0.00909527710169581709,
297 [95] = FSCALE * 0.00865169520312063417,
298 [96] = FSCALE * 0.00822974704902002884,
299 [97] = FSCALE * 0.00782837754922577143,
300 [98] = FSCALE * 0.00744658307092434051,
301 [99] = FSCALE * 0.00708340892905212004,
302 [100] = FSCALE * 0.00673794699908546709,
303 [101] = FSCALE * 0.00640933344625638184,
304 [102] = FSCALE * 0.00609674656551563610,
305 [103] = FSCALE * 0.00579940472684214321,
306 [104] = FSCALE * 0.00551656442076077241,
307 [105] = FSCALE * 0.00524751839918138427,
308 [106] = FSCALE * 0.00499159390691021621,
309 [107] = FSCALE * 0.00474815099941147558,
310 [108] = FSCALE * 0.00451658094261266798,
311 [109] = FSCALE * 0.00429630469075234057,
312 [110] = FSCALE * 0.00408677143846406699,
316 #define CCPU_EXP_MAX 110
319 * This function is analogical to the getpcpu() function in the ps(1) command.
320 * They should both calculate in the same way so that the racct %cpu
321 * calculations are consistent with the values showed by the ps(1) tool.
322 * The calculations are more complex in the 4BSD scheduler because of the value
323 * of the ccpu variable. In ULE it is defined to be zero which saves us some
327 racct_getpcpu(struct proc *p, u_int pcpu)
331 fixpt_t pctcpu, pctcpu_next;
340 ASSERT_RACCT_ENABLED();
343 * If the process is swapped out, we count its %cpu usage as zero.
344 * This behaviour is consistent with the userland ps(1) tool.
346 if ((p->p_flag & P_INMEM) == 0)
348 swtime = (ticks - p->p_swtick) / hz;
351 * For short-lived processes, the sched_pctcpu() returns small
352 * values even for cpu intensive processes. Therefore we use
353 * our own estimate in this case.
355 if (swtime < RACCT_PCPU_SECS)
359 FOREACH_THREAD_IN_PROC(p, td) {
360 if (td == PCPU_GET(idlethread))
364 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
365 if (td == pc->pc_idlethread) {
375 pctcpu = sched_pctcpu(td);
376 /* Count also the yet unfinished second. */
377 pctcpu_next = (pctcpu * ccpu_exp[1]) >> FSHIFT;
378 pctcpu_next += sched_pctcpu_delta(td);
379 p_pctcpu += max(pctcpu, pctcpu_next);
382 * In ULE the %cpu statistics are updated on every
383 * sched_pctcpu() call. So special calculations to
384 * account for the latest (unfinished) second are
387 p_pctcpu += sched_pctcpu(td);
393 if (swtime <= CCPU_EXP_MAX)
394 return ((100 * (uint64_t)p_pctcpu * 1000000) /
395 (FSCALE - ccpu_exp[swtime]));
398 return ((100 * (uint64_t)p_pctcpu * 1000000) / FSCALE);
402 racct_add_racct(struct racct *dest, const struct racct *src)
406 ASSERT_RACCT_ENABLED();
410 * Update resource usage in dest.
412 for (i = 0; i <= RACCT_MAX; i++) {
413 KASSERT(dest->r_resources[i] >= 0,
414 ("%s: resource %d propagation meltdown: dest < 0",
416 KASSERT(src->r_resources[i] >= 0,
417 ("%s: resource %d propagation meltdown: src < 0",
419 dest->r_resources[i] += src->r_resources[i];
424 racct_sub_racct(struct racct *dest, const struct racct *src)
428 ASSERT_RACCT_ENABLED();
432 * Update resource usage in dest.
434 for (i = 0; i <= RACCT_MAX; i++) {
435 if (!RACCT_IS_SLOPPY(i) && !RACCT_IS_DECAYING(i)) {
436 KASSERT(dest->r_resources[i] >= 0,
437 ("%s: resource %d propagation meltdown: dest < 0",
439 KASSERT(src->r_resources[i] >= 0,
440 ("%s: resource %d propagation meltdown: src < 0",
442 KASSERT(src->r_resources[i] <= dest->r_resources[i],
443 ("%s: resource %d propagation meltdown: src > dest",
446 if (RACCT_CAN_DROP(i)) {
447 dest->r_resources[i] -= src->r_resources[i];
448 if (dest->r_resources[i] < 0)
449 dest->r_resources[i] = 0;
455 racct_create(struct racct **racctp)
461 SDT_PROBE1(racct, , racct, create, racctp);
463 KASSERT(*racctp == NULL, ("racct already allocated"));
465 *racctp = uma_zalloc(racct_zone, M_WAITOK | M_ZERO);
469 racct_destroy_locked(struct racct **racctp)
474 ASSERT_RACCT_ENABLED();
476 SDT_PROBE1(racct, , racct, destroy, racctp);
479 KASSERT(racctp != NULL, ("NULL racctp"));
480 KASSERT(*racctp != NULL, ("NULL racct"));
484 for (i = 0; i <= RACCT_MAX; i++) {
485 if (RACCT_IS_SLOPPY(i))
487 if (!RACCT_IS_RECLAIMABLE(i))
489 KASSERT(racct->r_resources[i] == 0,
490 ("destroying non-empty racct: "
491 "%ju allocated for resource %d\n",
492 racct->r_resources[i], i));
494 uma_zfree(racct_zone, racct);
499 racct_destroy(struct racct **racct)
506 racct_destroy_locked(racct);
511 * Increase consumption of 'resource' by 'amount' for 'racct',
512 * but not its parents. Differently from other cases, 'amount' here
513 * may be less than zero.
516 racct_adjust_resource(struct racct *racct, int resource,
520 ASSERT_RACCT_ENABLED();
522 KASSERT(racct != NULL, ("NULL racct"));
524 racct->r_resources[resource] += amount;
525 if (racct->r_resources[resource] < 0) {
526 KASSERT(RACCT_IS_SLOPPY(resource) || RACCT_IS_DECAYING(resource),
527 ("%s: resource %d usage < 0", __func__, resource));
528 racct->r_resources[resource] = 0;
532 * There are some cases where the racct %cpu resource would grow
533 * beyond 100% per core. For example in racct_proc_exit() we add
534 * the process %cpu usage to the ucred racct containers. If too
535 * many processes terminated in a short time span, the ucred %cpu
536 * resource could grow too much. Also, the 4BSD scheduler sometimes
537 * returns for a thread more than 100% cpu usage. So we set a sane
538 * boundary here to 100% * the maxumum number of CPUs.
540 if ((resource == RACCT_PCTCPU) &&
541 (racct->r_resources[RACCT_PCTCPU] > 100 * 1000000 * (int64_t)MAXCPU))
542 racct->r_resources[RACCT_PCTCPU] = 100 * 1000000 * (int64_t)MAXCPU;
546 racct_add_locked(struct proc *p, int resource, uint64_t amount, int force)
552 ASSERT_RACCT_ENABLED();
555 * We need proc lock to dereference p->p_ucred.
557 PROC_LOCK_ASSERT(p, MA_OWNED);
560 error = rctl_enforce(p, resource, amount);
561 if (error && !force && RACCT_IS_DENIABLE(resource)) {
562 SDT_PROBE3(racct, , rusage, add__failure, p, resource, amount);
566 racct_adjust_resource(p->p_racct, resource, amount);
567 racct_add_cred_locked(p->p_ucred, resource, amount);
573 * Increase allocation of 'resource' by 'amount' for process 'p'.
574 * Return 0 if it's below limits, or errno, if it's not.
577 racct_add(struct proc *p, int resource, uint64_t amount)
584 SDT_PROBE3(racct, , rusage, add, p, resource, amount);
587 error = racct_add_locked(p, resource, amount, 0);
593 * Increase allocation of 'resource' by 'amount' for process 'p'.
594 * Doesn't check for limits and never fails.
597 racct_add_force(struct proc *p, int resource, uint64_t amount)
603 SDT_PROBE3(racct, , rusage, add__force, p, resource, amount);
606 racct_add_locked(p, resource, amount, 1);
611 racct_add_cred_locked(struct ucred *cred, int resource, uint64_t amount)
615 ASSERT_RACCT_ENABLED();
617 racct_adjust_resource(cred->cr_ruidinfo->ui_racct, resource, amount);
618 for (pr = cred->cr_prison; pr != NULL; pr = pr->pr_parent)
619 racct_adjust_resource(pr->pr_prison_racct->prr_racct, resource,
621 racct_adjust_resource(cred->cr_loginclass->lc_racct, resource, amount);
625 * Increase allocation of 'resource' by 'amount' for credential 'cred'.
626 * Doesn't check for limits and never fails.
629 racct_add_cred(struct ucred *cred, int resource, uint64_t amount)
635 SDT_PROBE3(racct, , rusage, add__cred, cred, resource, amount);
638 racct_add_cred_locked(cred, resource, amount);
643 * Account for disk IO resource consumption. Checks for limits,
644 * but never fails, due to disk limits being undeniable.
647 racct_add_buf(struct proc *p, const struct buf *bp, int is_write)
650 ASSERT_RACCT_ENABLED();
651 PROC_LOCK_ASSERT(p, MA_OWNED);
653 SDT_PROBE3(racct, , rusage, add__buf, p, bp, is_write);
657 racct_add_locked(curproc, RACCT_WRITEBPS, bp->b_bcount, 1);
658 racct_add_locked(curproc, RACCT_WRITEIOPS, 1, 1);
660 racct_add_locked(curproc, RACCT_READBPS, bp->b_bcount, 1);
661 racct_add_locked(curproc, RACCT_READIOPS, 1, 1);
667 racct_set_locked(struct proc *p, int resource, uint64_t amount, int force)
669 int64_t old_amount, decayed_amount, diff_proc, diff_cred;
674 ASSERT_RACCT_ENABLED();
677 * We need proc lock to dereference p->p_ucred.
679 PROC_LOCK_ASSERT(p, MA_OWNED);
681 old_amount = p->p_racct->r_resources[resource];
683 * The diffs may be negative.
685 diff_proc = amount - old_amount;
686 if (resource == RACCT_PCTCPU) {
688 * Resources in per-credential racct containers may decay.
689 * If this is the case, we need to calculate the difference
690 * between the new amount and the proportional value of the
691 * old amount that has decayed in the ucred racct containers.
693 decayed_amount = old_amount * RACCT_DECAY_FACTOR / FSCALE;
694 diff_cred = amount - decayed_amount;
696 diff_cred = diff_proc;
698 KASSERT(diff_proc >= 0 || RACCT_CAN_DROP(resource),
699 ("%s: usage of non-droppable resource %d dropping", __func__,
704 error = rctl_enforce(p, resource, diff_proc);
705 if (error && !force && RACCT_IS_DENIABLE(resource)) {
706 SDT_PROBE3(racct, , rusage, set__failure, p, resource,
712 racct_adjust_resource(p->p_racct, resource, diff_proc);
714 racct_add_cred_locked(p->p_ucred, resource, diff_cred);
715 else if (diff_cred < 0)
716 racct_sub_cred_locked(p->p_ucred, resource, -diff_cred);
722 * Set allocation of 'resource' to 'amount' for process 'p'.
723 * Return 0 if it's below limits, or errno, if it's not.
725 * Note that decreasing the allocation always returns 0,
726 * even if it's above the limit.
729 racct_set_unlocked(struct proc *p, int resource, uint64_t amount)
733 ASSERT_RACCT_ENABLED();
735 error = racct_set(p, resource, amount);
741 racct_set(struct proc *p, int resource, uint64_t amount)
748 SDT_PROBE3(racct, , rusage, set__force, p, resource, amount);
751 error = racct_set_locked(p, resource, amount, 0);
757 racct_set_force(struct proc *p, int resource, uint64_t amount)
763 SDT_PROBE3(racct, , rusage, set, p, resource, amount);
766 racct_set_locked(p, resource, amount, 1);
771 * Returns amount of 'resource' the process 'p' can keep allocated.
772 * Allocating more than that would be denied, unless the resource
773 * is marked undeniable. Amount of already allocated resource does
777 racct_get_limit(struct proc *p, int resource)
786 available = rctl_get_limit(p, resource);
797 * Returns amount of 'resource' the process 'p' can keep allocated.
798 * Allocating more than that would be denied, unless the resource
799 * is marked undeniable. Amount of already allocated resource does
803 racct_get_available(struct proc *p, int resource)
812 available = rctl_get_available(p, resource);
823 * Returns amount of the %cpu resource that process 'p' can add to its %cpu
824 * utilization. Adding more than that would lead to the process being
828 racct_pcpu_available(struct proc *p)
833 ASSERT_RACCT_ENABLED();
836 available = rctl_pcpu_available(p);
847 * Decrease allocation of 'resource' by 'amount' for process 'p'.
850 racct_sub(struct proc *p, int resource, uint64_t amount)
856 SDT_PROBE3(racct, , rusage, sub, p, resource, amount);
859 * We need proc lock to dereference p->p_ucred.
861 PROC_LOCK_ASSERT(p, MA_OWNED);
862 KASSERT(RACCT_CAN_DROP(resource),
863 ("%s: called for non-droppable resource %d", __func__, resource));
866 KASSERT(amount <= p->p_racct->r_resources[resource],
867 ("%s: freeing %ju of resource %d, which is more "
868 "than allocated %jd for %s (pid %d)", __func__, amount, resource,
869 (intmax_t)p->p_racct->r_resources[resource], p->p_comm, p->p_pid));
871 racct_adjust_resource(p->p_racct, resource, -amount);
872 racct_sub_cred_locked(p->p_ucred, resource, amount);
877 racct_sub_cred_locked(struct ucred *cred, int resource, uint64_t amount)
881 ASSERT_RACCT_ENABLED();
883 racct_adjust_resource(cred->cr_ruidinfo->ui_racct, resource, -amount);
884 for (pr = cred->cr_prison; pr != NULL; pr = pr->pr_parent)
885 racct_adjust_resource(pr->pr_prison_racct->prr_racct, resource,
887 racct_adjust_resource(cred->cr_loginclass->lc_racct, resource, -amount);
891 * Decrease allocation of 'resource' by 'amount' for credential 'cred'.
894 racct_sub_cred(struct ucred *cred, int resource, uint64_t amount)
900 SDT_PROBE3(racct, , rusage, sub__cred, cred, resource, amount);
903 KASSERT(RACCT_CAN_DROP(resource),
904 ("%s: called for resource %d which can not drop", __func__,
909 racct_sub_cred_locked(cred, resource, amount);
914 * Inherit resource usage information from the parent process.
917 racct_proc_fork(struct proc *parent, struct proc *child)
925 * Create racct for the child process.
927 racct_create(&child->p_racct);
934 error = rctl_proc_fork(parent, child);
939 /* Init process cpu time. */
940 child->p_prev_runtime = 0;
941 child->p_throttled = 0;
944 * Inherit resource usage.
946 for (i = 0; i <= RACCT_MAX; i++) {
947 if (parent->p_racct->r_resources[i] == 0 ||
948 !RACCT_IS_INHERITABLE(i))
951 error = racct_set_locked(child, i,
952 parent->p_racct->r_resources[i], 0);
957 error = racct_add_locked(child, RACCT_NPROC, 1, 0);
958 error += racct_add_locked(child, RACCT_NTHR, 1, 0);
966 racct_proc_exit(child);
972 * Called at the end of fork1(), to handle rules that require the process
973 * to be fully initialized.
976 racct_proc_fork_done(struct proc *child)
985 rctl_enforce(child, RACCT_NPROC, 0);
986 rctl_enforce(child, RACCT_NTHR, 0);
993 racct_proc_exit(struct proc *p)
995 struct timeval wallclock;
996 uint64_t pct_estimate, pct, runtime;
1004 * We don't need to calculate rux, proc_reap() has already done this.
1006 runtime = cputick2usec(p->p_rux.rux_runtime);
1008 KASSERT(runtime >= p->p_prev_runtime, ("runtime < p_prev_runtime"));
1010 if (runtime < p->p_prev_runtime)
1011 runtime = p->p_prev_runtime;
1013 microuptime(&wallclock);
1014 timevalsub(&wallclock, &p->p_stats->p_start);
1015 if (wallclock.tv_sec > 0 || wallclock.tv_usec > 0) {
1016 pct_estimate = (1000000 * runtime * 100) /
1017 ((uint64_t)wallclock.tv_sec * 1000000 +
1021 pct = racct_getpcpu(p, pct_estimate);
1024 racct_set_locked(p, RACCT_CPU, runtime, 0);
1025 racct_add_cred_locked(p->p_ucred, RACCT_PCTCPU, pct);
1027 KASSERT(p->p_racct->r_resources[RACCT_RSS] == 0,
1028 ("process reaped with %ju allocated for RSS\n",
1029 p->p_racct->r_resources[RACCT_RSS]));
1030 for (i = 0; i <= RACCT_MAX; i++) {
1031 if (p->p_racct->r_resources[i] == 0)
1033 if (!RACCT_IS_RECLAIMABLE(i))
1035 racct_set_locked(p, i, 0, 0);
1039 rctl_racct_release(p->p_racct);
1041 racct_destroy_locked(&p->p_racct);
1047 * Called after credentials change, to move resource utilisation
1051 racct_proc_ucred_changed(struct proc *p, struct ucred *oldcred,
1052 struct ucred *newcred)
1054 struct uidinfo *olduip, *newuip;
1055 struct loginclass *oldlc, *newlc;
1056 struct prison *oldpr, *newpr, *pr;
1061 PROC_LOCK_ASSERT(p, MA_OWNED);
1063 newuip = newcred->cr_ruidinfo;
1064 olduip = oldcred->cr_ruidinfo;
1065 newlc = newcred->cr_loginclass;
1066 oldlc = oldcred->cr_loginclass;
1067 newpr = newcred->cr_prison;
1068 oldpr = oldcred->cr_prison;
1071 if (newuip != olduip) {
1072 racct_sub_racct(olduip->ui_racct, p->p_racct);
1073 racct_add_racct(newuip->ui_racct, p->p_racct);
1075 if (newlc != oldlc) {
1076 racct_sub_racct(oldlc->lc_racct, p->p_racct);
1077 racct_add_racct(newlc->lc_racct, p->p_racct);
1079 if (newpr != oldpr) {
1080 for (pr = oldpr; pr != NULL; pr = pr->pr_parent)
1081 racct_sub_racct(pr->pr_prison_racct->prr_racct,
1083 for (pr = newpr; pr != NULL; pr = pr->pr_parent)
1084 racct_add_racct(pr->pr_prison_racct->prr_racct,
1091 racct_move(struct racct *dest, struct racct *src)
1094 ASSERT_RACCT_ENABLED();
1097 racct_add_racct(dest, src);
1098 racct_sub_racct(src, src);
1103 racct_proc_throttled(struct proc *p)
1106 ASSERT_RACCT_ENABLED();
1109 while (p->p_throttled != 0) {
1110 msleep(p->p_racct, &p->p_mtx, 0, "racct",
1111 p->p_throttled < 0 ? 0 : p->p_throttled);
1112 if (p->p_throttled > 0)
1119 * Make the process sleep in userret() for 'timeout' ticks. Setting
1120 * timeout to -1 makes it sleep until woken up by racct_proc_wakeup().
1123 racct_proc_throttle(struct proc *p, int timeout)
1130 KASSERT(timeout != 0, ("timeout %d", timeout));
1131 ASSERT_RACCT_ENABLED();
1132 PROC_LOCK_ASSERT(p, MA_OWNED);
1135 * Do not block kernel processes. Also do not block processes with
1136 * low %cpu utilization to improve interactivity.
1138 if ((p->p_flag & (P_SYSTEM | P_KPROC)) != 0)
1141 if (p->p_throttled < 0 || (timeout > 0 && p->p_throttled > timeout))
1144 p->p_throttled = timeout;
1146 FOREACH_THREAD_IN_PROC(p, td) {
1148 td->td_flags |= TDF_ASTPENDING;
1150 switch (td->td_state) {
1153 * If the thread is on the scheduler run-queue, we can
1154 * not just remove it from there. So we set the flag
1155 * TDF_NEEDRESCHED for the thread, so that once it is
1156 * running, it is taken off the cpu as soon as possible.
1158 td->td_flags |= TDF_NEEDRESCHED;
1162 * If the thread is running, we request a context
1163 * switch for it by setting the TDF_NEEDRESCHED flag.
1165 td->td_flags |= TDF_NEEDRESCHED;
1167 cpuid = td->td_oncpu;
1168 if ((cpuid != NOCPU) && (td != curthread))
1169 ipi_cpu(cpuid, IPI_AST);
1180 racct_proc_wakeup(struct proc *p)
1183 ASSERT_RACCT_ENABLED();
1185 PROC_LOCK_ASSERT(p, MA_OWNED);
1187 if (p->p_throttled != 0) {
1194 racct_decay_callback(struct racct *racct, void *dummy1, void *dummy2)
1196 int64_t r_old, r_new;
1198 ASSERT_RACCT_ENABLED();
1199 RACCT_LOCK_ASSERT();
1202 rctl_throttle_decay(racct, RACCT_READBPS);
1203 rctl_throttle_decay(racct, RACCT_WRITEBPS);
1204 rctl_throttle_decay(racct, RACCT_READIOPS);
1205 rctl_throttle_decay(racct, RACCT_WRITEIOPS);
1208 r_old = racct->r_resources[RACCT_PCTCPU];
1210 /* If there is nothing to decay, just exit. */
1214 r_new = r_old * RACCT_DECAY_FACTOR / FSCALE;
1215 racct->r_resources[RACCT_PCTCPU] = r_new;
1219 racct_decay_pre(void)
1226 racct_decay_post(void)
1236 ASSERT_RACCT_ENABLED();
1238 ui_racct_foreach(racct_decay_callback, racct_decay_pre,
1239 racct_decay_post, NULL, NULL);
1240 loginclass_racct_foreach(racct_decay_callback, racct_decay_pre,
1241 racct_decay_post, NULL, NULL);
1242 prison_racct_foreach(racct_decay_callback, racct_decay_pre,
1243 racct_decay_post, NULL, NULL);
1251 struct timeval wallclock;
1252 uint64_t pct, pct_estimate, runtime;
1254 ASSERT_RACCT_ENABLED();
1259 sx_slock(&allproc_lock);
1261 FOREACH_PROC_IN_SYSTEM(p) {
1263 if (p->p_state != PRS_NORMAL) {
1264 if (p->p_state == PRS_ZOMBIE)
1265 racct_set(p, RACCT_PCTCPU, 0);
1270 microuptime(&wallclock);
1271 timevalsub(&wallclock, &p->p_stats->p_start);
1273 FOREACH_THREAD_IN_PROC(p, td)
1275 runtime = cputick2usec(p->p_rux.rux_runtime);
1278 KASSERT(runtime >= p->p_prev_runtime,
1279 ("runtime < p_prev_runtime"));
1281 if (runtime < p->p_prev_runtime)
1282 runtime = p->p_prev_runtime;
1284 p->p_prev_runtime = runtime;
1285 if (wallclock.tv_sec > 0 || wallclock.tv_usec > 0) {
1286 pct_estimate = (1000000 * runtime * 100) /
1287 ((uint64_t)wallclock.tv_sec * 1000000 +
1291 pct = racct_getpcpu(p, pct_estimate);
1294 rctl_throttle_decay(p->p_racct, RACCT_READBPS);
1295 rctl_throttle_decay(p->p_racct, RACCT_WRITEBPS);
1296 rctl_throttle_decay(p->p_racct, RACCT_READIOPS);
1297 rctl_throttle_decay(p->p_racct, RACCT_WRITEIOPS);
1299 racct_set_locked(p, RACCT_PCTCPU, pct, 1);
1300 racct_set_locked(p, RACCT_CPU, runtime, 0);
1301 racct_set_locked(p, RACCT_WALLCLOCK,
1302 (uint64_t)wallclock.tv_sec * 1000000 +
1303 wallclock.tv_usec, 0);
1309 * To ensure that processes are throttled in a fair way, we need
1310 * to iterate over all processes again and check the limits
1311 * for %cpu resource only after ucred racct containers have been
1314 FOREACH_PROC_IN_SYSTEM(p) {
1316 if (p->p_state != PRS_NORMAL) {
1321 if (racct_pcpu_available(p) <= 0) {
1322 if (p->p_racct->r_resources[RACCT_PCTCPU] >
1324 racct_proc_throttle(p, -1);
1325 } else if (p->p_throttled == -1) {
1326 racct_proc_wakeup(p);
1330 sx_sunlock(&allproc_lock);
1335 static struct kproc_desc racctd_kp = {
1347 kproc_start(&racctd_kp);
1349 SYSINIT(racctd, SI_SUB_RACCTD, SI_ORDER_FIRST, racctd_init, NULL);
1357 racct_zone = uma_zcreate("racct", sizeof(struct racct),
1358 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
1360 * XXX: Move this somewhere.
1362 prison0.pr_prison_racct = prison_racct_find("0");
1364 SYSINIT(racct, SI_SUB_RACCT, SI_ORDER_FIRST, racct_init, NULL);