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_sched.h"
37 #include <sys/param.h>
39 #include <sys/systm.h>
40 #include <sys/eventhandler.h>
42 #include <sys/kernel.h>
43 #include <sys/kthread.h>
45 #include <sys/loginclass.h>
46 #include <sys/malloc.h>
47 #include <sys/mutex.h>
49 #include <sys/racct.h>
50 #include <sys/resourcevar.h>
52 #include <sys/sched.h>
56 #include <sys/sysctl.h>
57 #include <sys/sysent.h>
58 #include <sys/sysproto.h>
60 #include <machine/smp.h>
68 FEATURE(racct, "Resource Accounting");
71 * Do not block processes that have their %cpu usage <= pcpu_threshold.
73 static int pcpu_threshold = 1;
74 #ifdef RACCT_DEFAULT_TO_DISABLED
80 SYSCTL_NODE(_kern, OID_AUTO, racct, CTLFLAG_RW, 0, "Resource Accounting");
81 SYSCTL_UINT(_kern_racct, OID_AUTO, enable, CTLFLAG_RDTUN, &racct_enable,
82 0, "Enable RACCT/RCTL");
83 SYSCTL_UINT(_kern_racct, OID_AUTO, pcpu_threshold, CTLFLAG_RW, &pcpu_threshold,
84 0, "Processes with higher %cpu usage than this value can be throttled.");
87 * How many seconds it takes to use the scheduler %cpu calculations. When a
88 * process starts, we compute its %cpu usage by dividing its runtime by the
89 * process wall clock time. After RACCT_PCPU_SECS pass, we use the value
90 * provided by the scheduler.
92 #define RACCT_PCPU_SECS 3
94 struct mtx racct_lock;
95 MTX_SYSINIT(racct_lock, &racct_lock, "racct lock", MTX_DEF);
97 static uma_zone_t racct_zone;
99 static void racct_sub_racct(struct racct *dest, const struct racct *src);
100 static void racct_sub_cred_locked(struct ucred *cred, int resource,
102 static void racct_add_cred_locked(struct ucred *cred, int resource,
105 SDT_PROVIDER_DEFINE(racct);
106 SDT_PROBE_DEFINE3(racct, , rusage, add,
107 "struct proc *", "int", "uint64_t");
108 SDT_PROBE_DEFINE3(racct, , rusage, add__failure,
109 "struct proc *", "int", "uint64_t");
110 SDT_PROBE_DEFINE3(racct, , rusage, add__buf,
111 "struct proc *", "const struct buf *", "int");
112 SDT_PROBE_DEFINE3(racct, , rusage, add__cred,
113 "struct ucred *", "int", "uint64_t");
114 SDT_PROBE_DEFINE3(racct, , rusage, add__force,
115 "struct proc *", "int", "uint64_t");
116 SDT_PROBE_DEFINE3(racct, , rusage, set,
117 "struct proc *", "int", "uint64_t");
118 SDT_PROBE_DEFINE3(racct, , rusage, set__failure,
119 "struct proc *", "int", "uint64_t");
120 SDT_PROBE_DEFINE3(racct, , rusage, set__force,
121 "struct proc *", "int", "uint64_t");
122 SDT_PROBE_DEFINE3(racct, , rusage, sub,
123 "struct proc *", "int", "uint64_t");
124 SDT_PROBE_DEFINE3(racct, , rusage, sub__cred,
125 "struct ucred *", "int", "uint64_t");
126 SDT_PROBE_DEFINE1(racct, , racct, create,
128 SDT_PROBE_DEFINE1(racct, , racct, destroy,
130 SDT_PROBE_DEFINE2(racct, , racct, join,
131 "struct racct *", "struct racct *");
132 SDT_PROBE_DEFINE2(racct, , racct, join__failure,
133 "struct racct *", "struct racct *");
134 SDT_PROBE_DEFINE2(racct, , racct, leave,
135 "struct racct *", "struct racct *");
137 int racct_types[] = {
141 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
143 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
149 RACCT_RECLAIMABLE | RACCT_DENIABLE,
151 RACCT_RECLAIMABLE | RACCT_DENIABLE,
153 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
155 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
157 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
159 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
161 RACCT_RECLAIMABLE | RACCT_DENIABLE,
163 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
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_INHERITABLE | RACCT_DENIABLE,
173 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
175 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
179 RACCT_DECAYING | RACCT_DENIABLE | RACCT_IN_MILLIONS,
189 static const fixpt_t RACCT_DECAY_FACTOR = 0.3 * FSCALE;
193 * Contains intermediate values for %cpu calculations to avoid using floating
194 * point in the kernel.
195 * ccpu_exp[k] = FSCALE * (ccpu/FSCALE)^k = FSCALE * exp(-k/20)
196 * It is needed only for the 4BSD scheduler, because in ULE, the ccpu equals to
197 * zero so the calculations are more straightforward.
199 fixpt_t ccpu_exp[] = {
201 [1] = FSCALE * 0.95122942450071400909,
202 [2] = FSCALE * 0.90483741803595957316,
203 [3] = FSCALE * 0.86070797642505780722,
204 [4] = FSCALE * 0.81873075307798185866,
205 [5] = FSCALE * 0.77880078307140486824,
206 [6] = FSCALE * 0.74081822068171786606,
207 [7] = FSCALE * 0.70468808971871343435,
208 [8] = FSCALE * 0.67032004603563930074,
209 [9] = FSCALE * 0.63762815162177329314,
210 [10] = FSCALE * 0.60653065971263342360,
211 [11] = FSCALE * 0.57694981038048669531,
212 [12] = FSCALE * 0.54881163609402643262,
213 [13] = FSCALE * 0.52204577676101604789,
214 [14] = FSCALE * 0.49658530379140951470,
215 [15] = FSCALE * 0.47236655274101470713,
216 [16] = FSCALE * 0.44932896411722159143,
217 [17] = FSCALE * 0.42741493194872666992,
218 [18] = FSCALE * 0.40656965974059911188,
219 [19] = FSCALE * 0.38674102345450120691,
220 [20] = FSCALE * 0.36787944117144232159,
221 [21] = FSCALE * 0.34993774911115535467,
222 [22] = FSCALE * 0.33287108369807955328,
223 [23] = FSCALE * 0.31663676937905321821,
224 [24] = FSCALE * 0.30119421191220209664,
225 [25] = FSCALE * 0.28650479686019010032,
226 [26] = FSCALE * 0.27253179303401260312,
227 [27] = FSCALE * 0.25924026064589150757,
228 [28] = FSCALE * 0.24659696394160647693,
229 [29] = FSCALE * 0.23457028809379765313,
230 [30] = FSCALE * 0.22313016014842982893,
231 [31] = FSCALE * 0.21224797382674305771,
232 [32] = FSCALE * 0.20189651799465540848,
233 [33] = FSCALE * 0.19204990862075411423,
234 [34] = FSCALE * 0.18268352405273465022,
235 [35] = FSCALE * 0.17377394345044512668,
236 [36] = FSCALE * 0.16529888822158653829,
237 [37] = FSCALE * 0.15723716631362761621,
238 [38] = FSCALE * 0.14956861922263505264,
239 [39] = FSCALE * 0.14227407158651357185,
240 [40] = FSCALE * 0.13533528323661269189,
241 [41] = FSCALE * 0.12873490358780421886,
242 [42] = FSCALE * 0.12245642825298191021,
243 [43] = FSCALE * 0.11648415777349695786,
244 [44] = FSCALE * 0.11080315836233388333,
245 [45] = FSCALE * 0.10539922456186433678,
246 [46] = FSCALE * 0.10025884372280373372,
247 [47] = FSCALE * 0.09536916221554961888,
248 [48] = FSCALE * 0.09071795328941250337,
249 [49] = FSCALE * 0.08629358649937051097,
250 [50] = FSCALE * 0.08208499862389879516,
251 [51] = FSCALE * 0.07808166600115315231,
252 [52] = FSCALE * 0.07427357821433388042,
253 [53] = FSCALE * 0.07065121306042958674,
254 [54] = FSCALE * 0.06720551273974976512,
255 [55] = FSCALE * 0.06392786120670757270,
256 [56] = FSCALE * 0.06081006262521796499,
257 [57] = FSCALE * 0.05784432087483846296,
258 [58] = FSCALE * 0.05502322005640722902,
259 [59] = FSCALE * 0.05233970594843239308,
260 [60] = FSCALE * 0.04978706836786394297,
261 [61] = FSCALE * 0.04735892439114092119,
262 [62] = FSCALE * 0.04504920239355780606,
263 [63] = FSCALE * 0.04285212686704017991,
264 [64] = FSCALE * 0.04076220397836621516,
265 [65] = FSCALE * 0.03877420783172200988,
266 [66] = FSCALE * 0.03688316740124000544,
267 [67] = FSCALE * 0.03508435410084502588,
268 [68] = FSCALE * 0.03337326996032607948,
269 [69] = FSCALE * 0.03174563637806794323,
270 [70] = FSCALE * 0.03019738342231850073,
271 [71] = FSCALE * 0.02872463965423942912,
272 [72] = FSCALE * 0.02732372244729256080,
273 [73] = FSCALE * 0.02599112877875534358,
274 [74] = FSCALE * 0.02472352647033939120,
275 [75] = FSCALE * 0.02351774585600910823,
276 [76] = FSCALE * 0.02237077185616559577,
277 [77] = FSCALE * 0.02127973643837716938,
278 [78] = FSCALE * 0.02024191144580438847,
279 [79] = FSCALE * 0.01925470177538692429,
280 [80] = FSCALE * 0.01831563888873418029,
281 [81] = FSCALE * 0.01742237463949351138,
282 [82] = FSCALE * 0.01657267540176124754,
283 [83] = FSCALE * 0.01576441648485449082,
284 [84] = FSCALE * 0.01499557682047770621,
285 [85] = FSCALE * 0.01426423390899925527,
286 [86] = FSCALE * 0.01356855901220093175,
287 [87] = FSCALE * 0.01290681258047986886,
288 [88] = FSCALE * 0.01227733990306844117,
289 [89] = FSCALE * 0.01167856697039544521,
290 [90] = FSCALE * 0.01110899653824230649,
291 [91] = FSCALE * 0.01056720438385265337,
292 [92] = FSCALE * 0.01005183574463358164,
293 [93] = FSCALE * 0.00956160193054350793,
294 [94] = FSCALE * 0.00909527710169581709,
295 [95] = FSCALE * 0.00865169520312063417,
296 [96] = FSCALE * 0.00822974704902002884,
297 [97] = FSCALE * 0.00782837754922577143,
298 [98] = FSCALE * 0.00744658307092434051,
299 [99] = FSCALE * 0.00708340892905212004,
300 [100] = FSCALE * 0.00673794699908546709,
301 [101] = FSCALE * 0.00640933344625638184,
302 [102] = FSCALE * 0.00609674656551563610,
303 [103] = FSCALE * 0.00579940472684214321,
304 [104] = FSCALE * 0.00551656442076077241,
305 [105] = FSCALE * 0.00524751839918138427,
306 [106] = FSCALE * 0.00499159390691021621,
307 [107] = FSCALE * 0.00474815099941147558,
308 [108] = FSCALE * 0.00451658094261266798,
309 [109] = FSCALE * 0.00429630469075234057,
310 [110] = FSCALE * 0.00408677143846406699,
314 #define CCPU_EXP_MAX 110
317 * This function is analogical to the getpcpu() function in the ps(1) command.
318 * They should both calculate in the same way so that the racct %cpu
319 * calculations are consistent with the values showed by the ps(1) tool.
320 * The calculations are more complex in the 4BSD scheduler because of the value
321 * of the ccpu variable. In ULE it is defined to be zero which saves us some
325 racct_getpcpu(struct proc *p, u_int pcpu)
329 fixpt_t pctcpu, pctcpu_next;
338 ASSERT_RACCT_ENABLED();
341 * If the process is swapped out, we count its %cpu usage as zero.
342 * This behaviour is consistent with the userland ps(1) tool.
344 if ((p->p_flag & P_INMEM) == 0)
346 swtime = (ticks - p->p_swtick) / hz;
349 * For short-lived processes, the sched_pctcpu() returns small
350 * values even for cpu intensive processes. Therefore we use
351 * our own estimate in this case.
353 if (swtime < RACCT_PCPU_SECS)
357 FOREACH_THREAD_IN_PROC(p, td) {
358 if (td == PCPU_GET(idlethread))
362 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
363 if (td == pc->pc_idlethread) {
373 pctcpu = sched_pctcpu(td);
374 /* Count also the yet unfinished second. */
375 pctcpu_next = (pctcpu * ccpu_exp[1]) >> FSHIFT;
376 pctcpu_next += sched_pctcpu_delta(td);
377 p_pctcpu += max(pctcpu, pctcpu_next);
380 * In ULE the %cpu statistics are updated on every
381 * sched_pctcpu() call. So special calculations to
382 * account for the latest (unfinished) second are
385 p_pctcpu += sched_pctcpu(td);
391 if (swtime <= CCPU_EXP_MAX)
392 return ((100 * (uint64_t)p_pctcpu * 1000000) /
393 (FSCALE - ccpu_exp[swtime]));
396 return ((100 * (uint64_t)p_pctcpu * 1000000) / FSCALE);
400 racct_add_racct(struct racct *dest, const struct racct *src)
404 ASSERT_RACCT_ENABLED();
408 * Update resource usage in dest.
410 for (i = 0; i <= RACCT_MAX; i++) {
411 KASSERT(dest->r_resources[i] >= 0,
412 ("%s: resource %d propagation meltdown: dest < 0",
414 KASSERT(src->r_resources[i] >= 0,
415 ("%s: resource %d propagation meltdown: src < 0",
417 dest->r_resources[i] += src->r_resources[i];
422 racct_sub_racct(struct racct *dest, const struct racct *src)
426 ASSERT_RACCT_ENABLED();
430 * Update resource usage in dest.
432 for (i = 0; i <= RACCT_MAX; i++) {
433 if (!RACCT_IS_SLOPPY(i) && !RACCT_IS_DECAYING(i)) {
434 KASSERT(dest->r_resources[i] >= 0,
435 ("%s: resource %d propagation meltdown: dest < 0",
437 KASSERT(src->r_resources[i] >= 0,
438 ("%s: resource %d propagation meltdown: src < 0",
440 KASSERT(src->r_resources[i] <= dest->r_resources[i],
441 ("%s: resource %d propagation meltdown: src > dest",
444 if (RACCT_CAN_DROP(i)) {
445 dest->r_resources[i] -= src->r_resources[i];
446 if (dest->r_resources[i] < 0)
447 dest->r_resources[i] = 0;
453 racct_create(struct racct **racctp)
459 SDT_PROBE1(racct, , racct, create, racctp);
461 KASSERT(*racctp == NULL, ("racct already allocated"));
463 *racctp = uma_zalloc(racct_zone, M_WAITOK | M_ZERO);
467 racct_destroy_locked(struct racct **racctp)
472 ASSERT_RACCT_ENABLED();
474 SDT_PROBE1(racct, , racct, destroy, racctp);
477 KASSERT(racctp != NULL, ("NULL racctp"));
478 KASSERT(*racctp != NULL, ("NULL racct"));
482 for (i = 0; i <= RACCT_MAX; i++) {
483 if (RACCT_IS_SLOPPY(i))
485 if (!RACCT_IS_RECLAIMABLE(i))
487 KASSERT(racct->r_resources[i] == 0,
488 ("destroying non-empty racct: "
489 "%ju allocated for resource %d\n",
490 racct->r_resources[i], i));
492 uma_zfree(racct_zone, racct);
497 racct_destroy(struct racct **racct)
504 racct_destroy_locked(racct);
509 * Increase consumption of 'resource' by 'amount' for 'racct',
510 * but not its parents. Differently from other cases, 'amount' here
511 * may be less than zero.
514 racct_adjust_resource(struct racct *racct, int resource,
518 ASSERT_RACCT_ENABLED();
520 KASSERT(racct != NULL, ("NULL racct"));
522 racct->r_resources[resource] += amount;
523 if (racct->r_resources[resource] < 0) {
524 KASSERT(RACCT_IS_SLOPPY(resource) || RACCT_IS_DECAYING(resource),
525 ("%s: resource %d usage < 0", __func__, resource));
526 racct->r_resources[resource] = 0;
530 * There are some cases where the racct %cpu resource would grow
531 * beyond 100% per core. For example in racct_proc_exit() we add
532 * the process %cpu usage to the ucred racct containers. If too
533 * many processes terminated in a short time span, the ucred %cpu
534 * resource could grow too much. Also, the 4BSD scheduler sometimes
535 * returns for a thread more than 100% cpu usage. So we set a sane
536 * boundary here to 100% * the maxumum number of CPUs.
538 if ((resource == RACCT_PCTCPU) &&
539 (racct->r_resources[RACCT_PCTCPU] > 100 * 1000000 * (int64_t)MAXCPU))
540 racct->r_resources[RACCT_PCTCPU] = 100 * 1000000 * (int64_t)MAXCPU;
544 racct_add_locked(struct proc *p, int resource, uint64_t amount, int force)
550 ASSERT_RACCT_ENABLED();
553 * We need proc lock to dereference p->p_ucred.
555 PROC_LOCK_ASSERT(p, MA_OWNED);
558 error = rctl_enforce(p, resource, amount);
559 if (error && !force && RACCT_IS_DENIABLE(resource)) {
560 SDT_PROBE3(racct, , rusage, add__failure, p, resource, amount);
564 racct_adjust_resource(p->p_racct, resource, amount);
565 racct_add_cred_locked(p->p_ucred, resource, amount);
571 * Increase allocation of 'resource' by 'amount' for process 'p'.
572 * Return 0 if it's below limits, or errno, if it's not.
575 racct_add(struct proc *p, int resource, uint64_t amount)
582 SDT_PROBE3(racct, , rusage, add, p, resource, amount);
585 error = racct_add_locked(p, resource, amount, 0);
591 * Increase allocation of 'resource' by 'amount' for process 'p'.
592 * Doesn't check for limits and never fails.
595 racct_add_force(struct proc *p, int resource, uint64_t amount)
601 SDT_PROBE3(racct, , rusage, add__force, p, resource, amount);
604 racct_add_locked(p, resource, amount, 1);
609 racct_add_cred_locked(struct ucred *cred, int resource, uint64_t amount)
613 ASSERT_RACCT_ENABLED();
615 racct_adjust_resource(cred->cr_ruidinfo->ui_racct, resource, amount);
616 for (pr = cred->cr_prison; pr != NULL; pr = pr->pr_parent)
617 racct_adjust_resource(pr->pr_prison_racct->prr_racct, resource,
619 racct_adjust_resource(cred->cr_loginclass->lc_racct, resource, amount);
623 * Increase allocation of 'resource' by 'amount' for credential 'cred'.
624 * Doesn't check for limits and never fails.
627 racct_add_cred(struct ucred *cred, int resource, uint64_t amount)
633 SDT_PROBE3(racct, , rusage, add__cred, cred, resource, amount);
636 racct_add_cred_locked(cred, resource, amount);
641 * Account for disk IO resource consumption. Checks for limits,
642 * but never fails, due to disk limits being undeniable.
645 racct_add_buf(struct proc *p, const struct buf *bp, int is_write)
648 ASSERT_RACCT_ENABLED();
649 PROC_LOCK_ASSERT(p, MA_OWNED);
651 SDT_PROBE3(racct, , rusage, add__buf, p, bp, is_write);
655 racct_add_locked(curproc, RACCT_WRITEBPS, bp->b_bcount, 1);
656 racct_add_locked(curproc, RACCT_WRITEIOPS, 1, 1);
658 racct_add_locked(curproc, RACCT_READBPS, bp->b_bcount, 1);
659 racct_add_locked(curproc, RACCT_READIOPS, 1, 1);
665 racct_set_locked(struct proc *p, int resource, uint64_t amount, int force)
667 int64_t old_amount, decayed_amount, diff_proc, diff_cred;
672 ASSERT_RACCT_ENABLED();
675 * We need proc lock to dereference p->p_ucred.
677 PROC_LOCK_ASSERT(p, MA_OWNED);
679 old_amount = p->p_racct->r_resources[resource];
681 * The diffs may be negative.
683 diff_proc = amount - old_amount;
684 if (resource == RACCT_PCTCPU) {
686 * Resources in per-credential racct containers may decay.
687 * If this is the case, we need to calculate the difference
688 * between the new amount and the proportional value of the
689 * old amount that has decayed in the ucred racct containers.
691 decayed_amount = old_amount * RACCT_DECAY_FACTOR / FSCALE;
692 diff_cred = amount - decayed_amount;
694 diff_cred = diff_proc;
696 KASSERT(diff_proc >= 0 || RACCT_CAN_DROP(resource),
697 ("%s: usage of non-droppable resource %d dropping", __func__,
702 error = rctl_enforce(p, resource, diff_proc);
703 if (error && !force && RACCT_IS_DENIABLE(resource)) {
704 SDT_PROBE3(racct, , rusage, set__failure, p, resource,
710 racct_adjust_resource(p->p_racct, resource, diff_proc);
712 racct_add_cred_locked(p->p_ucred, resource, diff_cred);
713 else if (diff_cred < 0)
714 racct_sub_cred_locked(p->p_ucred, resource, -diff_cred);
720 * Set allocation of 'resource' to 'amount' for process 'p'.
721 * Return 0 if it's below limits, or errno, if it's not.
723 * Note that decreasing the allocation always returns 0,
724 * even if it's above the limit.
727 racct_set(struct proc *p, int resource, uint64_t amount)
734 SDT_PROBE3(racct, , rusage, set__force, p, resource, amount);
737 error = racct_set_locked(p, resource, amount, 0);
743 racct_set_force(struct proc *p, int resource, uint64_t amount)
749 SDT_PROBE3(racct, , rusage, set, p, resource, amount);
752 racct_set_locked(p, resource, amount, 1);
757 * Returns amount of 'resource' the process 'p' can keep allocated.
758 * Allocating more than that would be denied, unless the resource
759 * is marked undeniable. Amount of already allocated resource does
763 racct_get_limit(struct proc *p, int resource)
772 available = rctl_get_limit(p, resource);
783 * Returns amount of 'resource' the process 'p' can keep allocated.
784 * Allocating more than that would be denied, unless the resource
785 * is marked undeniable. Amount of already allocated resource does
789 racct_get_available(struct proc *p, int resource)
798 available = rctl_get_available(p, resource);
809 * Returns amount of the %cpu resource that process 'p' can add to its %cpu
810 * utilization. Adding more than that would lead to the process being
814 racct_pcpu_available(struct proc *p)
819 ASSERT_RACCT_ENABLED();
822 available = rctl_pcpu_available(p);
833 * Decrease allocation of 'resource' by 'amount' for process 'p'.
836 racct_sub(struct proc *p, int resource, uint64_t amount)
842 SDT_PROBE3(racct, , rusage, sub, p, resource, amount);
845 * We need proc lock to dereference p->p_ucred.
847 PROC_LOCK_ASSERT(p, MA_OWNED);
848 KASSERT(RACCT_CAN_DROP(resource),
849 ("%s: called for non-droppable resource %d", __func__, resource));
852 KASSERT(amount <= p->p_racct->r_resources[resource],
853 ("%s: freeing %ju of resource %d, which is more "
854 "than allocated %jd for %s (pid %d)", __func__, amount, resource,
855 (intmax_t)p->p_racct->r_resources[resource], p->p_comm, p->p_pid));
857 racct_adjust_resource(p->p_racct, resource, -amount);
858 racct_sub_cred_locked(p->p_ucred, resource, amount);
863 racct_sub_cred_locked(struct ucred *cred, int resource, uint64_t amount)
867 ASSERT_RACCT_ENABLED();
869 racct_adjust_resource(cred->cr_ruidinfo->ui_racct, resource, -amount);
870 for (pr = cred->cr_prison; pr != NULL; pr = pr->pr_parent)
871 racct_adjust_resource(pr->pr_prison_racct->prr_racct, resource,
873 racct_adjust_resource(cred->cr_loginclass->lc_racct, resource, -amount);
877 * Decrease allocation of 'resource' by 'amount' for credential 'cred'.
880 racct_sub_cred(struct ucred *cred, int resource, uint64_t amount)
886 SDT_PROBE3(racct, , rusage, sub__cred, cred, resource, amount);
889 KASSERT(RACCT_CAN_DROP(resource),
890 ("%s: called for resource %d which can not drop", __func__,
895 racct_sub_cred_locked(cred, resource, amount);
900 * Inherit resource usage information from the parent process.
903 racct_proc_fork(struct proc *parent, struct proc *child)
911 * Create racct for the child process.
913 racct_create(&child->p_racct);
920 error = rctl_proc_fork(parent, child);
925 /* Init process cpu time. */
926 child->p_prev_runtime = 0;
927 child->p_throttled = 0;
930 * Inherit resource usage.
932 for (i = 0; i <= RACCT_MAX; i++) {
933 if (parent->p_racct->r_resources[i] == 0 ||
934 !RACCT_IS_INHERITABLE(i))
937 error = racct_set_locked(child, i,
938 parent->p_racct->r_resources[i], 0);
943 error = racct_add_locked(child, RACCT_NPROC, 1, 0);
944 error += racct_add_locked(child, RACCT_NTHR, 1, 0);
952 racct_proc_exit(child);
958 * Called at the end of fork1(), to handle rules that require the process
959 * to be fully initialized.
962 racct_proc_fork_done(struct proc *child)
968 PROC_LOCK_ASSERT(child, MA_OWNED);
972 rctl_enforce(child, RACCT_NPROC, 0);
973 rctl_enforce(child, RACCT_NTHR, 0);
979 racct_proc_exit(struct proc *p)
981 struct timeval wallclock;
982 uint64_t pct_estimate, pct, runtime;
990 * We don't need to calculate rux, proc_reap() has already done this.
992 runtime = cputick2usec(p->p_rux.rux_runtime);
994 KASSERT(runtime >= p->p_prev_runtime, ("runtime < p_prev_runtime"));
996 if (runtime < p->p_prev_runtime)
997 runtime = p->p_prev_runtime;
999 microuptime(&wallclock);
1000 timevalsub(&wallclock, &p->p_stats->p_start);
1001 if (wallclock.tv_sec > 0 || wallclock.tv_usec > 0) {
1002 pct_estimate = (1000000 * runtime * 100) /
1003 ((uint64_t)wallclock.tv_sec * 1000000 +
1007 pct = racct_getpcpu(p, pct_estimate);
1010 racct_set_locked(p, RACCT_CPU, runtime, 0);
1011 racct_add_cred_locked(p->p_ucred, RACCT_PCTCPU, pct);
1013 KASSERT(p->p_racct->r_resources[RACCT_RSS] == 0,
1014 ("process reaped with %ju allocated for RSS\n",
1015 p->p_racct->r_resources[RACCT_RSS]));
1016 for (i = 0; i <= RACCT_MAX; i++) {
1017 if (p->p_racct->r_resources[i] == 0)
1019 if (!RACCT_IS_RECLAIMABLE(i))
1021 racct_set_locked(p, i, 0, 0);
1025 rctl_racct_release(p->p_racct);
1027 racct_destroy_locked(&p->p_racct);
1033 * Called after credentials change, to move resource utilisation
1037 racct_proc_ucred_changed(struct proc *p, struct ucred *oldcred,
1038 struct ucred *newcred)
1040 struct uidinfo *olduip, *newuip;
1041 struct loginclass *oldlc, *newlc;
1042 struct prison *oldpr, *newpr, *pr;
1047 PROC_LOCK_ASSERT(p, MA_NOTOWNED);
1049 newuip = newcred->cr_ruidinfo;
1050 olduip = oldcred->cr_ruidinfo;
1051 newlc = newcred->cr_loginclass;
1052 oldlc = oldcred->cr_loginclass;
1053 newpr = newcred->cr_prison;
1054 oldpr = oldcred->cr_prison;
1057 if (newuip != olduip) {
1058 racct_sub_racct(olduip->ui_racct, p->p_racct);
1059 racct_add_racct(newuip->ui_racct, p->p_racct);
1061 if (newlc != oldlc) {
1062 racct_sub_racct(oldlc->lc_racct, p->p_racct);
1063 racct_add_racct(newlc->lc_racct, p->p_racct);
1065 if (newpr != oldpr) {
1066 for (pr = oldpr; pr != NULL; pr = pr->pr_parent)
1067 racct_sub_racct(pr->pr_prison_racct->prr_racct,
1069 for (pr = newpr; pr != NULL; pr = pr->pr_parent)
1070 racct_add_racct(pr->pr_prison_racct->prr_racct,
1076 rctl_proc_ucred_changed(p, newcred);
1081 racct_move(struct racct *dest, struct racct *src)
1084 ASSERT_RACCT_ENABLED();
1087 racct_add_racct(dest, src);
1088 racct_sub_racct(src, src);
1093 * Make the process sleep in userret() for 'timeout' ticks. Setting
1094 * timeout to -1 makes it sleep until woken up by racct_proc_wakeup().
1097 racct_proc_throttle(struct proc *p, int timeout)
1104 KASSERT(timeout != 0, ("timeout %d", timeout));
1105 ASSERT_RACCT_ENABLED();
1106 PROC_LOCK_ASSERT(p, MA_OWNED);
1109 * Do not block kernel processes. Also do not block processes with
1110 * low %cpu utilization to improve interactivity.
1112 if ((p->p_flag & (P_SYSTEM | P_KPROC)) != 0)
1115 if (p->p_throttled < 0 || (timeout > 0 && p->p_throttled > timeout))
1118 p->p_throttled = timeout;
1120 FOREACH_THREAD_IN_PROC(p, td) {
1122 switch (td->td_state) {
1125 * If the thread is on the scheduler run-queue, we can
1126 * not just remove it from there. So we set the flag
1127 * TDF_NEEDRESCHED for the thread, so that once it is
1128 * running, it is taken off the cpu as soon as possible.
1130 td->td_flags |= TDF_NEEDRESCHED;
1134 * If the thread is running, we request a context
1135 * switch for it by setting the TDF_NEEDRESCHED flag.
1137 td->td_flags |= TDF_NEEDRESCHED;
1139 cpuid = td->td_oncpu;
1140 if ((cpuid != NOCPU) && (td != curthread))
1141 ipi_cpu(cpuid, IPI_AST);
1152 racct_proc_wakeup(struct proc *p)
1155 ASSERT_RACCT_ENABLED();
1157 PROC_LOCK_ASSERT(p, MA_OWNED);
1159 if (p->p_throttled != 0) {
1166 racct_decay_callback(struct racct *racct, void *dummy1, void *dummy2)
1168 int64_t r_old, r_new;
1170 ASSERT_RACCT_ENABLED();
1171 RACCT_LOCK_ASSERT();
1174 rctl_throttle_decay(racct, RACCT_READBPS);
1175 rctl_throttle_decay(racct, RACCT_WRITEBPS);
1176 rctl_throttle_decay(racct, RACCT_READIOPS);
1177 rctl_throttle_decay(racct, RACCT_WRITEIOPS);
1180 r_old = racct->r_resources[RACCT_PCTCPU];
1182 /* If there is nothing to decay, just exit. */
1186 r_new = r_old * RACCT_DECAY_FACTOR / FSCALE;
1187 racct->r_resources[RACCT_PCTCPU] = r_new;
1191 racct_decay_pre(void)
1198 racct_decay_post(void)
1208 ASSERT_RACCT_ENABLED();
1210 ui_racct_foreach(racct_decay_callback, racct_decay_pre,
1211 racct_decay_post, NULL, NULL);
1212 loginclass_racct_foreach(racct_decay_callback, racct_decay_pre,
1213 racct_decay_post, NULL, NULL);
1214 prison_racct_foreach(racct_decay_callback, racct_decay_pre,
1215 racct_decay_post, NULL, NULL);
1223 struct timeval wallclock;
1224 uint64_t pct, pct_estimate, runtime;
1226 ASSERT_RACCT_ENABLED();
1231 sx_slock(&allproc_lock);
1233 LIST_FOREACH(p, &zombproc, p_list) {
1235 racct_set(p, RACCT_PCTCPU, 0);
1239 FOREACH_PROC_IN_SYSTEM(p) {
1241 if (p->p_state != PRS_NORMAL) {
1246 microuptime(&wallclock);
1247 timevalsub(&wallclock, &p->p_stats->p_start);
1249 FOREACH_THREAD_IN_PROC(p, td)
1251 runtime = cputick2usec(p->p_rux.rux_runtime);
1254 KASSERT(runtime >= p->p_prev_runtime,
1255 ("runtime < p_prev_runtime"));
1257 if (runtime < p->p_prev_runtime)
1258 runtime = p->p_prev_runtime;
1260 p->p_prev_runtime = runtime;
1261 if (wallclock.tv_sec > 0 || wallclock.tv_usec > 0) {
1262 pct_estimate = (1000000 * runtime * 100) /
1263 ((uint64_t)wallclock.tv_sec * 1000000 +
1267 pct = racct_getpcpu(p, pct_estimate);
1270 rctl_throttle_decay(p->p_racct, RACCT_READBPS);
1271 rctl_throttle_decay(p->p_racct, RACCT_WRITEBPS);
1272 rctl_throttle_decay(p->p_racct, RACCT_READIOPS);
1273 rctl_throttle_decay(p->p_racct, RACCT_WRITEIOPS);
1275 racct_set_locked(p, RACCT_PCTCPU, pct, 1);
1276 racct_set_locked(p, RACCT_CPU, runtime, 0);
1277 racct_set_locked(p, RACCT_WALLCLOCK,
1278 (uint64_t)wallclock.tv_sec * 1000000 +
1279 wallclock.tv_usec, 0);
1285 * To ensure that processes are throttled in a fair way, we need
1286 * to iterate over all processes again and check the limits
1287 * for %cpu resource only after ucred racct containers have been
1290 FOREACH_PROC_IN_SYSTEM(p) {
1292 if (p->p_state != PRS_NORMAL) {
1297 if (racct_pcpu_available(p) <= 0) {
1298 if (p->p_racct->r_resources[RACCT_PCTCPU] >
1300 racct_proc_throttle(p, -1);
1301 } else if (p->p_throttled == -1) {
1302 racct_proc_wakeup(p);
1306 sx_sunlock(&allproc_lock);
1311 static struct kproc_desc racctd_kp = {
1323 kproc_start(&racctd_kp);
1325 SYSINIT(racctd, SI_SUB_RACCTD, SI_ORDER_FIRST, racctd_init, NULL);
1333 racct_zone = uma_zcreate("racct", sizeof(struct racct),
1334 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
1336 * XXX: Move this somewhere.
1338 prison0.pr_prison_racct = prison_racct_find("0");
1340 SYSINIT(racct, SI_SUB_RACCT, SI_ORDER_FIRST, racct_init, NULL);