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 KASSERT(RACCT_IS_SLOPPY(i
) ||
448 RACCT_IS_DECAYING(i
),
449 ("%s: resource %d usage < 0", __func__
, i
));
450 dest
->r_resources
[i
] = 0;
457 racct_create(struct racct
**racctp
)
463 SDT_PROBE1(racct
, , racct
, create
, racctp
);
465 KASSERT(*racctp
== NULL
, ("racct already allocated"));
467 *racctp
= uma_zalloc(racct_zone
, M_WAITOK
| M_ZERO
);
471 racct_destroy_locked(struct racct
**racctp
)
476 ASSERT_RACCT_ENABLED();
478 SDT_PROBE1(racct
, , racct
, destroy
, racctp
);
481 KASSERT(racctp
!= NULL
, ("NULL racctp"));
482 KASSERT(*racctp
!= NULL
, ("NULL racct"));
486 for (i
= 0; i
<= RACCT_MAX
; i
++) {
487 if (RACCT_IS_SLOPPY(i
))
489 if (!RACCT_IS_RECLAIMABLE(i
))
491 KASSERT(racct
->r_resources
[i
] == 0,
492 ("destroying non-empty racct: "
493 "%ju allocated for resource %d\n",
494 racct
->r_resources
[i
], i
));
496 uma_zfree(racct_zone
, racct
);
501 racct_destroy(struct racct
**racct
)
508 racct_destroy_locked(racct
);
513 * Increase consumption of 'resource' by 'amount' for 'racct',
514 * but not its parents. Differently from other cases, 'amount' here
515 * may be less than zero.
518 racct_adjust_resource(struct racct
*racct
, int resource
,
522 ASSERT_RACCT_ENABLED();
524 KASSERT(racct
!= NULL
, ("NULL racct"));
526 racct
->r_resources
[resource
] += amount
;
527 if (racct
->r_resources
[resource
] < 0) {
528 KASSERT(RACCT_IS_SLOPPY(resource
) || RACCT_IS_DECAYING(resource
),
529 ("%s: resource %d usage < 0", __func__
, resource
));
530 racct
->r_resources
[resource
] = 0;
534 * There are some cases where the racct %cpu resource would grow
535 * beyond 100% per core. For example in racct_proc_exit() we add
536 * the process %cpu usage to the ucred racct containers. If too
537 * many processes terminated in a short time span, the ucred %cpu
538 * resource could grow too much. Also, the 4BSD scheduler sometimes
539 * returns for a thread more than 100% cpu usage. So we set a sane
540 * boundary here to 100% * the maxumum number of CPUs.
542 if ((resource
== RACCT_PCTCPU
) &&
543 (racct
->r_resources
[RACCT_PCTCPU
] > 100 * 1000000 * (int64_t)MAXCPU
))
544 racct
->r_resources
[RACCT_PCTCPU
] = 100 * 1000000 * (int64_t)MAXCPU
;
548 racct_add_locked(struct proc
*p
, int resource
, uint64_t amount
, int force
)
554 ASSERT_RACCT_ENABLED();
557 * We need proc lock to dereference p->p_ucred.
559 PROC_LOCK_ASSERT(p
, MA_OWNED
);
562 error
= rctl_enforce(p
, resource
, amount
);
563 if (error
&& !force
&& RACCT_IS_DENIABLE(resource
)) {
564 SDT_PROBE3(racct
, , rusage
, add__failure
, p
, resource
, amount
);
568 racct_adjust_resource(p
->p_racct
, resource
, amount
);
569 racct_add_cred_locked(p
->p_ucred
, resource
, amount
);
575 * Increase allocation of 'resource' by 'amount' for process 'p'.
576 * Return 0 if it's below limits, or errno, if it's not.
579 racct_add(struct proc
*p
, int resource
, uint64_t amount
)
586 SDT_PROBE3(racct
, , rusage
, add
, p
, resource
, amount
);
589 error
= racct_add_locked(p
, resource
, amount
, 0);
595 * Increase allocation of 'resource' by 'amount' for process 'p'.
596 * Doesn't check for limits and never fails.
599 racct_add_force(struct proc
*p
, int resource
, uint64_t amount
)
605 SDT_PROBE3(racct
, , rusage
, add__force
, p
, resource
, amount
);
608 racct_add_locked(p
, resource
, amount
, 1);
613 racct_add_cred_locked(struct ucred
*cred
, int resource
, uint64_t amount
)
617 ASSERT_RACCT_ENABLED();
619 racct_adjust_resource(cred
->cr_ruidinfo
->ui_racct
, resource
, amount
);
620 for (pr
= cred
->cr_prison
; pr
!= NULL
; pr
= pr
->pr_parent
)
621 racct_adjust_resource(pr
->pr_prison_racct
->prr_racct
, resource
,
623 racct_adjust_resource(cred
->cr_loginclass
->lc_racct
, resource
, amount
);
627 * Increase allocation of 'resource' by 'amount' for credential 'cred'.
628 * Doesn't check for limits and never fails.
631 racct_add_cred(struct ucred
*cred
, int resource
, uint64_t amount
)
637 SDT_PROBE3(racct
, , rusage
, add__cred
, cred
, resource
, amount
);
640 racct_add_cred_locked(cred
, resource
, amount
);
645 * Account for disk IO resource consumption. Checks for limits,
646 * but never fails, due to disk limits being undeniable.
649 racct_add_buf(struct proc
*p
, const struct buf
*bp
, int is_write
)
652 ASSERT_RACCT_ENABLED();
653 PROC_LOCK_ASSERT(p
, MA_OWNED
);
655 SDT_PROBE3(racct
, , rusage
, add__buf
, p
, bp
, is_write
);
659 racct_add_locked(curproc
, RACCT_WRITEBPS
, bp
->b_bcount
, 1);
660 racct_add_locked(curproc
, RACCT_WRITEIOPS
, 1, 1);
662 racct_add_locked(curproc
, RACCT_READBPS
, bp
->b_bcount
, 1);
663 racct_add_locked(curproc
, RACCT_READIOPS
, 1, 1);
669 racct_set_locked(struct proc
*p
, int resource
, uint64_t amount
, int force
)
671 int64_t old_amount
, decayed_amount
, diff_proc
, diff_cred
;
676 ASSERT_RACCT_ENABLED();
679 * We need proc lock to dereference p->p_ucred.
681 PROC_LOCK_ASSERT(p
, MA_OWNED
);
683 old_amount
= p
->p_racct
->r_resources
[resource
];
685 * The diffs may be negative.
687 diff_proc
= amount
- old_amount
;
688 if (resource
== RACCT_PCTCPU
) {
690 * Resources in per-credential racct containers may decay.
691 * If this is the case, we need to calculate the difference
692 * between the new amount and the proportional value of the
693 * old amount that has decayed in the ucred racct containers.
695 decayed_amount
= old_amount
* RACCT_DECAY_FACTOR
/ FSCALE
;
696 diff_cred
= amount
- decayed_amount
;
698 diff_cred
= diff_proc
;
700 KASSERT(diff_proc
>= 0 || RACCT_CAN_DROP(resource
),
701 ("%s: usage of non-droppable resource %d dropping", __func__
,
706 error
= rctl_enforce(p
, resource
, diff_proc
);
707 if (error
&& !force
&& RACCT_IS_DENIABLE(resource
)) {
708 SDT_PROBE3(racct
, , rusage
, set__failure
, p
, resource
,
714 racct_adjust_resource(p
->p_racct
, resource
, diff_proc
);
716 racct_add_cred_locked(p
->p_ucred
, resource
, diff_cred
);
717 else if (diff_cred
< 0)
718 racct_sub_cred_locked(p
->p_ucred
, resource
, -diff_cred
);
724 * Set allocation of 'resource' to 'amount' for process 'p'.
725 * Return 0 if it's below limits, or errno, if it's not.
727 * Note that decreasing the allocation always returns 0,
728 * even if it's above the limit.
731 racct_set(struct proc
*p
, int resource
, uint64_t amount
)
738 SDT_PROBE3(racct
, , rusage
, set__force
, p
, resource
, amount
);
741 error
= racct_set_locked(p
, resource
, amount
, 0);
747 racct_set_force(struct proc
*p
, int resource
, uint64_t amount
)
753 SDT_PROBE3(racct
, , rusage
, set
, p
, resource
, amount
);
756 racct_set_locked(p
, resource
, amount
, 1);
761 * Returns amount of 'resource' the process 'p' can keep allocated.
762 * Allocating more than that would be denied, unless the resource
763 * is marked undeniable. Amount of already allocated resource does
767 racct_get_limit(struct proc
*p
, int resource
)
776 available
= rctl_get_limit(p
, resource
);
787 * Returns amount of 'resource' the process 'p' can keep allocated.
788 * Allocating more than that would be denied, unless the resource
789 * is marked undeniable. Amount of already allocated resource does
793 racct_get_available(struct proc
*p
, int resource
)
802 available
= rctl_get_available(p
, resource
);
813 * Returns amount of the %cpu resource that process 'p' can add to its %cpu
814 * utilization. Adding more than that would lead to the process being
818 racct_pcpu_available(struct proc
*p
)
823 ASSERT_RACCT_ENABLED();
826 available
= rctl_pcpu_available(p
);
837 * Decrease allocation of 'resource' by 'amount' for process 'p'.
840 racct_sub(struct proc
*p
, int resource
, uint64_t amount
)
846 SDT_PROBE3(racct
, , rusage
, sub
, p
, resource
, amount
);
849 * We need proc lock to dereference p->p_ucred.
851 PROC_LOCK_ASSERT(p
, MA_OWNED
);
852 KASSERT(RACCT_CAN_DROP(resource
),
853 ("%s: called for non-droppable resource %d", __func__
, resource
));
856 KASSERT(amount
<= p
->p_racct
->r_resources
[resource
],
857 ("%s: freeing %ju of resource %d, which is more "
858 "than allocated %jd for %s (pid %d)", __func__
, amount
, resource
,
859 (intmax_t)p
->p_racct
->r_resources
[resource
], p
->p_comm
, p
->p_pid
));
861 racct_adjust_resource(p
->p_racct
, resource
, -amount
);
862 racct_sub_cred_locked(p
->p_ucred
, resource
, amount
);
867 racct_sub_cred_locked(struct ucred
*cred
, int resource
, uint64_t amount
)
871 ASSERT_RACCT_ENABLED();
873 racct_adjust_resource(cred
->cr_ruidinfo
->ui_racct
, resource
, -amount
);
874 for (pr
= cred
->cr_prison
; pr
!= NULL
; pr
= pr
->pr_parent
)
875 racct_adjust_resource(pr
->pr_prison_racct
->prr_racct
, resource
,
877 racct_adjust_resource(cred
->cr_loginclass
->lc_racct
, resource
, -amount
);
881 * Decrease allocation of 'resource' by 'amount' for credential 'cred'.
884 racct_sub_cred(struct ucred
*cred
, int resource
, uint64_t amount
)
890 SDT_PROBE3(racct
, , rusage
, sub__cred
, cred
, resource
, amount
);
893 KASSERT(RACCT_CAN_DROP(resource
),
894 ("%s: called for resource %d which can not drop", __func__
,
899 racct_sub_cred_locked(cred
, resource
, amount
);
904 * Inherit resource usage information from the parent process.
907 racct_proc_fork(struct proc
*parent
, struct proc
*child
)
915 * Create racct for the child process.
917 racct_create(&child
->p_racct
);
924 error
= rctl_proc_fork(parent
, child
);
929 /* Init process cpu time. */
930 child
->p_prev_runtime
= 0;
931 child
->p_throttled
= 0;
934 * Inherit resource usage.
936 for (i
= 0; i
<= RACCT_MAX
; i
++) {
937 if (parent
->p_racct
->r_resources
[i
] == 0 ||
938 !RACCT_IS_INHERITABLE(i
))
941 error
= racct_set_locked(child
, i
,
942 parent
->p_racct
->r_resources
[i
], 0);
947 error
= racct_add_locked(child
, RACCT_NPROC
, 1, 0);
948 error
+= racct_add_locked(child
, RACCT_NTHR
, 1, 0);
956 racct_proc_exit(child
);
962 * Called at the end of fork1(), to handle rules that require the process
963 * to be fully initialized.
966 racct_proc_fork_done(struct proc
*child
)
972 PROC_LOCK_ASSERT(child
, MA_OWNED
);
976 rctl_enforce(child
, RACCT_NPROC
, 0);
977 rctl_enforce(child
, RACCT_NTHR
, 0);
983 racct_proc_exit(struct proc
*p
)
985 struct timeval wallclock
;
986 uint64_t pct_estimate
, pct
, runtime
;
994 * We don't need to calculate rux, proc_reap() has already done this.
996 runtime
= cputick2usec(p
->p_rux
.rux_runtime
);
998 KASSERT(runtime
>= p
->p_prev_runtime
, ("runtime < p_prev_runtime"));
1000 if (runtime
< p
->p_prev_runtime
)
1001 runtime
= p
->p_prev_runtime
;
1003 microuptime(&wallclock
);
1004 timevalsub(&wallclock
, &p
->p_stats
->p_start
);
1005 if (wallclock
.tv_sec
> 0 || wallclock
.tv_usec
> 0) {
1006 pct_estimate
= (1000000 * runtime
* 100) /
1007 ((uint64_t)wallclock
.tv_sec
* 1000000 +
1011 pct
= racct_getpcpu(p
, pct_estimate
);
1014 racct_set_locked(p
, RACCT_CPU
, runtime
, 0);
1015 racct_add_cred_locked(p
->p_ucred
, RACCT_PCTCPU
, pct
);
1017 for (i
= 0; i
<= RACCT_MAX
; i
++) {
1018 if (p
->p_racct
->r_resources
[i
] == 0)
1020 if (!RACCT_IS_RECLAIMABLE(i
))
1022 racct_set_locked(p
, i
, 0, 0);
1026 rctl_racct_release(p
->p_racct
);
1028 racct_destroy_locked(&p
->p_racct
);
1034 * Called after credentials change, to move resource utilisation
1038 racct_proc_ucred_changed(struct proc
*p
, struct ucred
*oldcred
,
1039 struct ucred
*newcred
)
1041 struct uidinfo
*olduip
, *newuip
;
1042 struct loginclass
*oldlc
, *newlc
;
1043 struct prison
*oldpr
, *newpr
, *pr
;
1048 PROC_LOCK_ASSERT(p
, MA_NOTOWNED
);
1050 newuip
= newcred
->cr_ruidinfo
;
1051 olduip
= oldcred
->cr_ruidinfo
;
1052 newlc
= newcred
->cr_loginclass
;
1053 oldlc
= oldcred
->cr_loginclass
;
1054 newpr
= newcred
->cr_prison
;
1055 oldpr
= oldcred
->cr_prison
;
1058 if (newuip
!= olduip
) {
1059 racct_sub_racct(olduip
->ui_racct
, p
->p_racct
);
1060 racct_add_racct(newuip
->ui_racct
, p
->p_racct
);
1062 if (newlc
!= oldlc
) {
1063 racct_sub_racct(oldlc
->lc_racct
, p
->p_racct
);
1064 racct_add_racct(newlc
->lc_racct
, p
->p_racct
);
1066 if (newpr
!= oldpr
) {
1067 for (pr
= oldpr
; pr
!= NULL
; pr
= pr
->pr_parent
)
1068 racct_sub_racct(pr
->pr_prison_racct
->prr_racct
,
1070 for (pr
= newpr
; pr
!= NULL
; pr
= pr
->pr_parent
)
1071 racct_add_racct(pr
->pr_prison_racct
->prr_racct
,
1077 rctl_proc_ucred_changed(p
, newcred
);
1082 racct_move(struct racct
*dest
, struct racct
*src
)
1085 ASSERT_RACCT_ENABLED();
1088 racct_add_racct(dest
, src
);
1089 racct_sub_racct(src
, src
);
1094 * Make the process sleep in userret() for 'timeout' ticks. Setting
1095 * timeout to -1 makes it sleep until woken up by racct_proc_wakeup().
1098 racct_proc_throttle(struct proc
*p
, int timeout
)
1105 KASSERT(timeout
!= 0, ("timeout %d", timeout
));
1106 ASSERT_RACCT_ENABLED();
1107 PROC_LOCK_ASSERT(p
, MA_OWNED
);
1110 * Do not block kernel processes. Also do not block processes with
1111 * low %cpu utilization to improve interactivity.
1113 if ((p
->p_flag
& (P_SYSTEM
| P_KPROC
)) != 0)
1116 if (p
->p_throttled
< 0 || (timeout
> 0 && p
->p_throttled
> timeout
))
1119 p
->p_throttled
= timeout
;
1121 FOREACH_THREAD_IN_PROC(p
, td
) {
1123 switch (td
->td_state
) {
1126 * If the thread is on the scheduler run-queue, we can
1127 * not just remove it from there. So we set the flag
1128 * TDF_NEEDRESCHED for the thread, so that once it is
1129 * running, it is taken off the cpu as soon as possible.
1131 td
->td_flags
|= TDF_NEEDRESCHED
;
1135 * If the thread is running, we request a context
1136 * switch for it by setting the TDF_NEEDRESCHED flag.
1138 td
->td_flags
|= TDF_NEEDRESCHED
;
1140 cpuid
= td
->td_oncpu
;
1141 if ((cpuid
!= NOCPU
) && (td
!= curthread
))
1142 ipi_cpu(cpuid
, IPI_AST
);
1153 racct_proc_wakeup(struct proc
*p
)
1156 ASSERT_RACCT_ENABLED();
1158 PROC_LOCK_ASSERT(p
, MA_OWNED
);
1160 if (p
->p_throttled
!= 0) {
1167 racct_decay_callback(struct racct
*racct
, void *dummy1
, void *dummy2
)
1169 int64_t r_old
, r_new
;
1171 ASSERT_RACCT_ENABLED();
1172 RACCT_LOCK_ASSERT();
1175 rctl_throttle_decay(racct
, RACCT_READBPS
);
1176 rctl_throttle_decay(racct
, RACCT_WRITEBPS
);
1177 rctl_throttle_decay(racct
, RACCT_READIOPS
);
1178 rctl_throttle_decay(racct
, RACCT_WRITEIOPS
);
1181 r_old
= racct
->r_resources
[RACCT_PCTCPU
];
1183 /* If there is nothing to decay, just exit. */
1187 r_new
= r_old
* RACCT_DECAY_FACTOR
/ FSCALE
;
1188 racct
->r_resources
[RACCT_PCTCPU
] = r_new
;
1192 racct_decay_pre(void)
1199 racct_decay_post(void)
1209 ASSERT_RACCT_ENABLED();
1211 ui_racct_foreach(racct_decay_callback
, racct_decay_pre
,
1212 racct_decay_post
, NULL
, NULL
);
1213 loginclass_racct_foreach(racct_decay_callback
, racct_decay_pre
,
1214 racct_decay_post
, NULL
, NULL
);
1215 prison_racct_foreach(racct_decay_callback
, racct_decay_pre
,
1216 racct_decay_post
, NULL
, NULL
);
1224 struct timeval wallclock
;
1225 uint64_t pct
, pct_estimate
, runtime
;
1227 ASSERT_RACCT_ENABLED();
1232 sx_slock(&allproc_lock
);
1234 LIST_FOREACH(p
, &zombproc
, p_list
) {
1236 racct_set(p
, RACCT_PCTCPU
, 0);
1240 FOREACH_PROC_IN_SYSTEM(p
) {
1242 if (p
->p_state
!= PRS_NORMAL
) {
1247 microuptime(&wallclock
);
1248 timevalsub(&wallclock
, &p
->p_stats
->p_start
);
1250 FOREACH_THREAD_IN_PROC(p
, td
)
1252 runtime
= cputick2usec(p
->p_rux
.rux_runtime
);
1255 KASSERT(runtime
>= p
->p_prev_runtime
,
1256 ("runtime < p_prev_runtime"));
1258 if (runtime
< p
->p_prev_runtime
)
1259 runtime
= p
->p_prev_runtime
;
1261 p
->p_prev_runtime
= runtime
;
1262 if (wallclock
.tv_sec
> 0 || wallclock
.tv_usec
> 0) {
1263 pct_estimate
= (1000000 * runtime
* 100) /
1264 ((uint64_t)wallclock
.tv_sec
* 1000000 +
1268 pct
= racct_getpcpu(p
, pct_estimate
);
1271 rctl_throttle_decay(p
->p_racct
, RACCT_READBPS
);
1272 rctl_throttle_decay(p
->p_racct
, RACCT_WRITEBPS
);
1273 rctl_throttle_decay(p
->p_racct
, RACCT_READIOPS
);
1274 rctl_throttle_decay(p
->p_racct
, RACCT_WRITEIOPS
);
1276 racct_set_locked(p
, RACCT_PCTCPU
, pct
, 1);
1277 racct_set_locked(p
, RACCT_CPU
, runtime
, 0);
1278 racct_set_locked(p
, RACCT_WALLCLOCK
,
1279 (uint64_t)wallclock
.tv_sec
* 1000000 +
1280 wallclock
.tv_usec
, 0);
1286 * To ensure that processes are throttled in a fair way, we need
1287 * to iterate over all processes again and check the limits
1288 * for %cpu resource only after ucred racct containers have been
1291 FOREACH_PROC_IN_SYSTEM(p
) {
1293 if (p
->p_state
!= PRS_NORMAL
) {
1298 if (racct_pcpu_available(p
) <= 0) {
1299 if (p
->p_racct
->r_resources
[RACCT_PCTCPU
] >
1301 racct_proc_throttle(p
, -1);
1302 } else if (p
->p_throttled
== -1) {
1303 racct_proc_wakeup(p
);
1307 sx_sunlock(&allproc_lock
);
1312 static struct kproc_desc racctd_kp
= {
1324 kproc_start(&racctd_kp
);
1326 SYSINIT(racctd
, SI_SUB_RACCTD
, SI_ORDER_FIRST
, racctd_init
, NULL
);
1334 racct_zone
= uma_zcreate("racct", sizeof(struct racct
),
1335 NULL
, NULL
, NULL
, NULL
, UMA_ALIGN_PTR
, 0);
1337 * XXX: Move this somewhere.
1339 prison0
.pr_prison_racct
= prison_racct_find("0");
1341 SYSINIT(racct
, SI_SUB_RACCT
, SI_ORDER_FIRST
, racct_init
, NULL
);