FreeBSD/Linux Kernel Cross Reference
sys/kern/kern_racct.c
1 /*-
2 * Copyright (c) 2010 The FreeBSD Foundation
3 * All rights reserved.
4 *
5 * This software was developed by Edward Tomasz Napierala under sponsorship
6 * from the FreeBSD Foundation.
7 *
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
10 * are met:
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.
16 *
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
27 * SUCH DAMAGE.
28 *
29 * $FreeBSD$
30 */
31
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
34
35 #include "opt_kdtrace.h"
36 #include "opt_sched.h"
37
38 #include <sys/param.h>
39 #include <sys/systm.h>
40 #include <sys/eventhandler.h>
41 #include <sys/jail.h>
42 #include <sys/kernel.h>
43 #include <sys/kthread.h>
44 #include <sys/lock.h>
45 #include <sys/loginclass.h>
46 #include <sys/malloc.h>
47 #include <sys/mutex.h>
48 #include <sys/proc.h>
49 #include <sys/racct.h>
50 #include <sys/resourcevar.h>
51 #include <sys/sbuf.h>
52 #include <sys/sched.h>
53 #include <sys/sdt.h>
54 #include <sys/smp.h>
55 #include <sys/sx.h>
56 #include <sys/sysctl.h>
57 #include <sys/sysent.h>
58 #include <sys/sysproto.h>
59 #include <sys/umtx.h>
60 #include <machine/smp.h>
61
62 #ifdef RCTL
63 #include <sys/rctl.h>
64 #endif
65
66 #ifdef RACCT
67
68 FEATURE(racct, "Resource Accounting");
69
70 /*
71 * Do not block processes that have their %cpu usage <= pcpu_threshold.
72 */
73 static int pcpu_threshold = 1;
74
75 SYSCTL_NODE(_kern, OID_AUTO, racct, CTLFLAG_RW, 0, "Resource Accounting");
76 SYSCTL_UINT(_kern_racct, OID_AUTO, pcpu_threshold, CTLFLAG_RW, &pcpu_threshold,
77 0, "Processes with higher %cpu usage than this value can be throttled.");
78
79 /*
80 * How many seconds it takes to use the scheduler %cpu calculations. When a
81 * process starts, we compute its %cpu usage by dividing its runtime by the
82 * process wall clock time. After RACCT_PCPU_SECS pass, we use the value
83 * provided by the scheduler.
84 */
85 #define RACCT_PCPU_SECS 3
86
87 static struct mtx racct_lock;
88 MTX_SYSINIT(racct_lock, &racct_lock, "racct lock", MTX_DEF);
89
90 static uma_zone_t racct_zone;
91
92 static void racct_sub_racct(struct racct *dest, const struct racct *src);
93 static void racct_sub_cred_locked(struct ucred *cred, int resource,
94 uint64_t amount);
95 static void racct_add_cred_locked(struct ucred *cred, int resource,
96 uint64_t amount);
97
98 SDT_PROVIDER_DEFINE(racct);
99 SDT_PROBE_DEFINE3(racct, kernel, rusage, add, "struct proc *", "int",
100 "uint64_t");
101 SDT_PROBE_DEFINE3(racct, kernel, rusage, add__failure,
102 "struct proc *", "int", "uint64_t");
103 SDT_PROBE_DEFINE3(racct, kernel, rusage, add__cred, "struct ucred *",
104 "int", "uint64_t");
105 SDT_PROBE_DEFINE3(racct, kernel, rusage, add__force, "struct proc *",
106 "int", "uint64_t");
107 SDT_PROBE_DEFINE3(racct, kernel, rusage, set, "struct proc *", "int",
108 "uint64_t");
109 SDT_PROBE_DEFINE3(racct, kernel, rusage, set__failure,
110 "struct proc *", "int", "uint64_t");
111 SDT_PROBE_DEFINE3(racct, kernel, rusage, sub, "struct proc *", "int",
112 "uint64_t");
113 SDT_PROBE_DEFINE3(racct, kernel, rusage, sub__cred, "struct ucred *",
114 "int", "uint64_t");
115 SDT_PROBE_DEFINE1(racct, kernel, racct, create, "struct racct *");
116 SDT_PROBE_DEFINE1(racct, kernel, racct, destroy, "struct racct *");
117 SDT_PROBE_DEFINE2(racct, kernel, racct, join, "struct racct *",
118 "struct racct *");
119 SDT_PROBE_DEFINE2(racct, kernel, racct, join__failure,
120 "struct racct *", "struct racct *");
121 SDT_PROBE_DEFINE2(racct, kernel, racct, leave, "struct racct *",
122 "struct racct *");
123
124 int racct_types[] = {
125 [RACCT_CPU] =
126 RACCT_IN_MILLIONS,
127 [RACCT_DATA] =
128 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
129 [RACCT_STACK] =
130 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
131 [RACCT_CORE] =
132 RACCT_DENIABLE,
133 [RACCT_RSS] =
134 RACCT_RECLAIMABLE,
135 [RACCT_MEMLOCK] =
136 RACCT_RECLAIMABLE | RACCT_DENIABLE,
137 [RACCT_NPROC] =
138 RACCT_RECLAIMABLE | RACCT_DENIABLE,
139 [RACCT_NOFILE] =
140 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
141 [RACCT_VMEM] =
142 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
143 [RACCT_NPTS] =
144 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
145 [RACCT_SWAP] =
146 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
147 [RACCT_NTHR] =
148 RACCT_RECLAIMABLE | RACCT_DENIABLE,
149 [RACCT_MSGQQUEUED] =
150 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
151 [RACCT_MSGQSIZE] =
152 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
153 [RACCT_NMSGQ] =
154 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
155 [RACCT_NSEM] =
156 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
157 [RACCT_NSEMOP] =
158 RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
159 [RACCT_NSHM] =
160 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
161 [RACCT_SHMSIZE] =
162 RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
163 [RACCT_WALLCLOCK] =
164 RACCT_IN_MILLIONS,
165 [RACCT_PCTCPU] =
166 RACCT_DECAYING | RACCT_DENIABLE | RACCT_IN_MILLIONS };
167
168 static const fixpt_t RACCT_DECAY_FACTOR = 0.3 * FSCALE;
169
170 #ifdef SCHED_4BSD
171 /*
172 * Contains intermediate values for %cpu calculations to avoid using floating
173 * point in the kernel.
174 * ccpu_exp[k] = FSCALE * (ccpu/FSCALE)^k = FSCALE * exp(-k/20)
175 * It is needed only for the 4BSD scheduler, because in ULE, the ccpu equals to
176 * zero so the calculations are more straightforward.
177 */
178 fixpt_t ccpu_exp[] = {
179 [0] = FSCALE * 1,
180 [1] = FSCALE * 0.95122942450071400909,
181 [2] = FSCALE * 0.90483741803595957316,
182 [3] = FSCALE * 0.86070797642505780722,
183 [4] = FSCALE * 0.81873075307798185866,
184 [5] = FSCALE * 0.77880078307140486824,
185 [6] = FSCALE * 0.74081822068171786606,
186 [7] = FSCALE * 0.70468808971871343435,
187 [8] = FSCALE * 0.67032004603563930074,
188 [9] = FSCALE * 0.63762815162177329314,
189 [10] = FSCALE * 0.60653065971263342360,
190 [11] = FSCALE * 0.57694981038048669531,
191 [12] = FSCALE * 0.54881163609402643262,
192 [13] = FSCALE * 0.52204577676101604789,
193 [14] = FSCALE * 0.49658530379140951470,
194 [15] = FSCALE * 0.47236655274101470713,
195 [16] = FSCALE * 0.44932896411722159143,
196 [17] = FSCALE * 0.42741493194872666992,
197 [18] = FSCALE * 0.40656965974059911188,
198 [19] = FSCALE * 0.38674102345450120691,
199 [20] = FSCALE * 0.36787944117144232159,
200 [21] = FSCALE * 0.34993774911115535467,
201 [22] = FSCALE * 0.33287108369807955328,
202 [23] = FSCALE * 0.31663676937905321821,
203 [24] = FSCALE * 0.30119421191220209664,
204 [25] = FSCALE * 0.28650479686019010032,
205 [26] = FSCALE * 0.27253179303401260312,
206 [27] = FSCALE * 0.25924026064589150757,
207 [28] = FSCALE * 0.24659696394160647693,
208 [29] = FSCALE * 0.23457028809379765313,
209 [30] = FSCALE * 0.22313016014842982893,
210 [31] = FSCALE * 0.21224797382674305771,
211 [32] = FSCALE * 0.20189651799465540848,
212 [33] = FSCALE * 0.19204990862075411423,
213 [34] = FSCALE * 0.18268352405273465022,
214 [35] = FSCALE * 0.17377394345044512668,
215 [36] = FSCALE * 0.16529888822158653829,
216 [37] = FSCALE * 0.15723716631362761621,
217 [38] = FSCALE * 0.14956861922263505264,
218 [39] = FSCALE * 0.14227407158651357185,
219 [40] = FSCALE * 0.13533528323661269189,
220 [41] = FSCALE * 0.12873490358780421886,
221 [42] = FSCALE * 0.12245642825298191021,
222 [43] = FSCALE * 0.11648415777349695786,
223 [44] = FSCALE * 0.11080315836233388333,
224 [45] = FSCALE * 0.10539922456186433678,
225 [46] = FSCALE * 0.10025884372280373372,
226 [47] = FSCALE * 0.09536916221554961888,
227 [48] = FSCALE * 0.09071795328941250337,
228 [49] = FSCALE * 0.08629358649937051097,
229 [50] = FSCALE * 0.08208499862389879516,
230 [51] = FSCALE * 0.07808166600115315231,
231 [52] = FSCALE * 0.07427357821433388042,
232 [53] = FSCALE * 0.07065121306042958674,
233 [54] = FSCALE * 0.06720551273974976512,
234 [55] = FSCALE * 0.06392786120670757270,
235 [56] = FSCALE * 0.06081006262521796499,
236 [57] = FSCALE * 0.05784432087483846296,
237 [58] = FSCALE * 0.05502322005640722902,
238 [59] = FSCALE * 0.05233970594843239308,
239 [60] = FSCALE * 0.04978706836786394297,
240 [61] = FSCALE * 0.04735892439114092119,
241 [62] = FSCALE * 0.04504920239355780606,
242 [63] = FSCALE * 0.04285212686704017991,
243 [64] = FSCALE * 0.04076220397836621516,
244 [65] = FSCALE * 0.03877420783172200988,
245 [66] = FSCALE * 0.03688316740124000544,
246 [67] = FSCALE * 0.03508435410084502588,
247 [68] = FSCALE * 0.03337326996032607948,
248 [69] = FSCALE * 0.03174563637806794323,
249 [70] = FSCALE * 0.03019738342231850073,
250 [71] = FSCALE * 0.02872463965423942912,
251 [72] = FSCALE * 0.02732372244729256080,
252 [73] = FSCALE * 0.02599112877875534358,
253 [74] = FSCALE * 0.02472352647033939120,
254 [75] = FSCALE * 0.02351774585600910823,
255 [76] = FSCALE * 0.02237077185616559577,
256 [77] = FSCALE * 0.02127973643837716938,
257 [78] = FSCALE * 0.02024191144580438847,
258 [79] = FSCALE * 0.01925470177538692429,
259 [80] = FSCALE * 0.01831563888873418029,
260 [81] = FSCALE * 0.01742237463949351138,
261 [82] = FSCALE * 0.01657267540176124754,
262 [83] = FSCALE * 0.01576441648485449082,
263 [84] = FSCALE * 0.01499557682047770621,
264 [85] = FSCALE * 0.01426423390899925527,
265 [86] = FSCALE * 0.01356855901220093175,
266 [87] = FSCALE * 0.01290681258047986886,
267 [88] = FSCALE * 0.01227733990306844117,
268 [89] = FSCALE * 0.01167856697039544521,
269 [90] = FSCALE * 0.01110899653824230649,
270 [91] = FSCALE * 0.01056720438385265337,
271 [92] = FSCALE * 0.01005183574463358164,
272 [93] = FSCALE * 0.00956160193054350793,
273 [94] = FSCALE * 0.00909527710169581709,
274 [95] = FSCALE * 0.00865169520312063417,
275 [96] = FSCALE * 0.00822974704902002884,
276 [97] = FSCALE * 0.00782837754922577143,
277 [98] = FSCALE * 0.00744658307092434051,
278 [99] = FSCALE * 0.00708340892905212004,
279 [100] = FSCALE * 0.00673794699908546709,
280 [101] = FSCALE * 0.00640933344625638184,
281 [102] = FSCALE * 0.00609674656551563610,
282 [103] = FSCALE * 0.00579940472684214321,
283 [104] = FSCALE * 0.00551656442076077241,
284 [105] = FSCALE * 0.00524751839918138427,
285 [106] = FSCALE * 0.00499159390691021621,
286 [107] = FSCALE * 0.00474815099941147558,
287 [108] = FSCALE * 0.00451658094261266798,
288 [109] = FSCALE * 0.00429630469075234057,
289 [110] = FSCALE * 0.00408677143846406699,
290 };
291 #endif
292
293 #define CCPU_EXP_MAX 110
294
295 /*
296 * This function is analogical to the getpcpu() function in the ps(1) command.
297 * They should both calculate in the same way so that the racct %cpu
298 * calculations are consistent with the values showed by the ps(1) tool.
299 * The calculations are more complex in the 4BSD scheduler because of the value
300 * of the ccpu variable. In ULE it is defined to be zero which saves us some
301 * work.
302 */
303 static uint64_t
304 racct_getpcpu(struct proc *p, u_int pcpu)
305 {
306 u_int swtime;
307 #ifdef SCHED_4BSD
308 fixpt_t pctcpu, pctcpu_next;
309 #endif
310 #ifdef SMP
311 struct pcpu *pc;
312 int found;
313 #endif
314 fixpt_t p_pctcpu;
315 struct thread *td;
316
317 /*
318 * If the process is swapped out, we count its %cpu usage as zero.
319 * This behaviour is consistent with the userland ps(1) tool.
320 */
321 if ((p->p_flag & P_INMEM) == 0)
322 return (0);
323 swtime = (ticks - p->p_swtick) / hz;
324
325 /*
326 * For short-lived processes, the sched_pctcpu() returns small
327 * values even for cpu intensive processes. Therefore we use
328 * our own estimate in this case.
329 */
330 if (swtime < RACCT_PCPU_SECS)
331 return (pcpu);
332
333 p_pctcpu = 0;
334 FOREACH_THREAD_IN_PROC(p, td) {
335 if (td == PCPU_GET(idlethread))
336 continue;
337 #ifdef SMP
338 found = 0;
339 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
340 if (td == pc->pc_idlethread) {
341 found = 1;
342 break;
343 }
344 }
345 if (found)
346 continue;
347 #endif
348 thread_lock(td);
349 #ifdef SCHED_4BSD
350 pctcpu = sched_pctcpu(td);
351 /* Count also the yet unfinished second. */
352 pctcpu_next = (pctcpu * ccpu_exp[1]) >> FSHIFT;
353 pctcpu_next += sched_pctcpu_delta(td);
354 p_pctcpu += max(pctcpu, pctcpu_next);
355 #else
356 /*
357 * In ULE the %cpu statistics are updated on every
358 * sched_pctcpu() call. So special calculations to
359 * account for the latest (unfinished) second are
360 * not needed.
361 */
362 p_pctcpu += sched_pctcpu(td);
363 #endif
364 thread_unlock(td);
365 }
366
367 #ifdef SCHED_4BSD
368 if (swtime <= CCPU_EXP_MAX)
369 return ((100 * (uint64_t)p_pctcpu * 1000000) /
370 (FSCALE - ccpu_exp[swtime]));
371 #endif
372
373 return ((100 * (uint64_t)p_pctcpu * 1000000) / FSCALE);
374 }
375
376 static void
377 racct_add_racct(struct racct *dest, const struct racct *src)
378 {
379 int i;
380
381 mtx_assert(&racct_lock, MA_OWNED);
382
383 /*
384 * Update resource usage in dest.
385 */
386 for (i = 0; i <= RACCT_MAX; i++) {
387 KASSERT(dest->r_resources[i] >= 0,
388 ("%s: resource %d propagation meltdown: dest < 0",
389 __func__, i));
390 KASSERT(src->r_resources[i] >= 0,
391 ("%s: resource %d propagation meltdown: src < 0",
392 __func__, i));
393 dest->r_resources[i] += src->r_resources[i];
394 }
395 }
396
397 static void
398 racct_sub_racct(struct racct *dest, const struct racct *src)
399 {
400 int i;
401
402 mtx_assert(&racct_lock, MA_OWNED);
403
404 /*
405 * Update resource usage in dest.
406 */
407 for (i = 0; i <= RACCT_MAX; i++) {
408 if (!RACCT_IS_SLOPPY(i) && !RACCT_IS_DECAYING(i)) {
409 KASSERT(dest->r_resources[i] >= 0,
410 ("%s: resource %d propagation meltdown: dest < 0",
411 __func__, i));
412 KASSERT(src->r_resources[i] >= 0,
413 ("%s: resource %d propagation meltdown: src < 0",
414 __func__, i));
415 KASSERT(src->r_resources[i] <= dest->r_resources[i],
416 ("%s: resource %d propagation meltdown: src > dest",
417 __func__, i));
418 }
419 if (RACCT_CAN_DROP(i)) {
420 dest->r_resources[i] -= src->r_resources[i];
421 if (dest->r_resources[i] < 0) {
422 KASSERT(RACCT_IS_SLOPPY(i) ||
423 RACCT_IS_DECAYING(i),
424 ("%s: resource %d usage < 0", __func__, i));
425 dest->r_resources[i] = 0;
426 }
427 }
428 }
429 }
430
431 void
432 racct_create(struct racct **racctp)
433 {
434
435 SDT_PROBE1(racct, kernel, racct, create, racctp);
436
437 KASSERT(*racctp == NULL, ("racct already allocated"));
438
439 *racctp = uma_zalloc(racct_zone, M_WAITOK | M_ZERO);
440 }
441
442 static void
443 racct_destroy_locked(struct racct **racctp)
444 {
445 int i;
446 struct racct *racct;
447
448 SDT_PROBE1(racct, kernel, racct, destroy, racctp);
449
450 mtx_assert(&racct_lock, MA_OWNED);
451 KASSERT(racctp != NULL, ("NULL racctp"));
452 KASSERT(*racctp != NULL, ("NULL racct"));
453
454 racct = *racctp;
455
456 for (i = 0; i <= RACCT_MAX; i++) {
457 if (RACCT_IS_SLOPPY(i))
458 continue;
459 if (!RACCT_IS_RECLAIMABLE(i))
460 continue;
461 KASSERT(racct->r_resources[i] == 0,
462 ("destroying non-empty racct: "
463 "%ju allocated for resource %d\n",
464 racct->r_resources[i], i));
465 }
466 uma_zfree(racct_zone, racct);
467 *racctp = NULL;
468 }
469
470 void
471 racct_destroy(struct racct **racct)
472 {
473
474 mtx_lock(&racct_lock);
475 racct_destroy_locked(racct);
476 mtx_unlock(&racct_lock);
477 }
478
479 /*
480 * Increase consumption of 'resource' by 'amount' for 'racct'
481 * and all its parents. Differently from other cases, 'amount' here
482 * may be less than zero.
483 */
484 static void
485 racct_alloc_resource(struct racct *racct, int resource,
486 uint64_t amount)
487 {
488
489 mtx_assert(&racct_lock, MA_OWNED);
490 KASSERT(racct != NULL, ("NULL racct"));
491
492 racct->r_resources[resource] += amount;
493 if (racct->r_resources[resource] < 0) {
494 KASSERT(RACCT_IS_SLOPPY(resource) || RACCT_IS_DECAYING(resource),
495 ("%s: resource %d usage < 0", __func__, resource));
496 racct->r_resources[resource] = 0;
497 }
498
499 /*
500 * There are some cases where the racct %cpu resource would grow
501 * beyond 100%.
502 * For example in racct_proc_exit() we add the process %cpu usage
503 * to the ucred racct containers. If too many processes terminated
504 * in a short time span, the ucred %cpu resource could grow too much.
505 * Also, the 4BSD scheduler sometimes returns for a thread more than
506 * 100% cpu usage. So we set a boundary here to 100%.
507 */
508 if ((resource == RACCT_PCTCPU) &&
509 (racct->r_resources[RACCT_PCTCPU] > 100 * 1000000))
510 racct->r_resources[RACCT_PCTCPU] = 100 * 1000000;
511 }
512
513 static int
514 racct_add_locked(struct proc *p, int resource, uint64_t amount)
515 {
516 #ifdef RCTL
517 int error;
518 #endif
519
520 SDT_PROBE3(racct, kernel, rusage, add, p, resource, amount);
521
522 /*
523 * We need proc lock to dereference p->p_ucred.
524 */
525 PROC_LOCK_ASSERT(p, MA_OWNED);
526
527 #ifdef RCTL
528 error = rctl_enforce(p, resource, amount);
529 if (error && RACCT_IS_DENIABLE(resource)) {
530 SDT_PROBE3(racct, kernel, rusage, add__failure, p, resource,
531 amount);
532 return (error);
533 }
534 #endif
535 racct_alloc_resource(p->p_racct, resource, amount);
536 racct_add_cred_locked(p->p_ucred, resource, amount);
537
538 return (0);
539 }
540
541 /*
542 * Increase allocation of 'resource' by 'amount' for process 'p'.
543 * Return 0 if it's below limits, or errno, if it's not.
544 */
545 int
546 racct_add(struct proc *p, int resource, uint64_t amount)
547 {
548 int error;
549
550 mtx_lock(&racct_lock);
551 error = racct_add_locked(p, resource, amount);
552 mtx_unlock(&racct_lock);
553 return (error);
554 }
555
556 static void
557 racct_add_cred_locked(struct ucred *cred, int resource, uint64_t amount)
558 {
559 struct prison *pr;
560
561 SDT_PROBE3(racct, kernel, rusage, add__cred, cred, resource, amount);
562
563 racct_alloc_resource(cred->cr_ruidinfo->ui_racct, resource, amount);
564 for (pr = cred->cr_prison; pr != NULL; pr = pr->pr_parent)
565 racct_alloc_resource(pr->pr_prison_racct->prr_racct, resource,
566 amount);
567 racct_alloc_resource(cred->cr_loginclass->lc_racct, resource, amount);
568 }
569
570 /*
571 * Increase allocation of 'resource' by 'amount' for credential 'cred'.
572 * Doesn't check for limits and never fails.
573 *
574 * XXX: Shouldn't this ever return an error?
575 */
576 void
577 racct_add_cred(struct ucred *cred, int resource, uint64_t amount)
578 {
579
580 mtx_lock(&racct_lock);
581 racct_add_cred_locked(cred, resource, amount);
582 mtx_unlock(&racct_lock);
583 }
584
585 /*
586 * Increase allocation of 'resource' by 'amount' for process 'p'.
587 * Doesn't check for limits and never fails.
588 */
589 void
590 racct_add_force(struct proc *p, int resource, uint64_t amount)
591 {
592
593 SDT_PROBE3(racct, kernel, rusage, add__force, p, resource, amount);
594
595 /*
596 * We need proc lock to dereference p->p_ucred.
597 */
598 PROC_LOCK_ASSERT(p, MA_OWNED);
599
600 mtx_lock(&racct_lock);
601 racct_alloc_resource(p->p_racct, resource, amount);
602 mtx_unlock(&racct_lock);
603 racct_add_cred(p->p_ucred, resource, amount);
604 }
605
606 static int
607 racct_set_locked(struct proc *p, int resource, uint64_t amount)
608 {
609 int64_t old_amount, decayed_amount;
610 int64_t diff_proc, diff_cred;
611 #ifdef RCTL
612 int error;
613 #endif
614
615 SDT_PROBE3(racct, kernel, rusage, set, p, resource, amount);
616
617 /*
618 * We need proc lock to dereference p->p_ucred.
619 */
620 PROC_LOCK_ASSERT(p, MA_OWNED);
621
622 old_amount = p->p_racct->r_resources[resource];
623 /*
624 * The diffs may be negative.
625 */
626 diff_proc = amount - old_amount;
627 if (RACCT_IS_DECAYING(resource)) {
628 /*
629 * Resources in per-credential racct containers may decay.
630 * If this is the case, we need to calculate the difference
631 * between the new amount and the proportional value of the
632 * old amount that has decayed in the ucred racct containers.
633 */
634 decayed_amount = old_amount * RACCT_DECAY_FACTOR / FSCALE;
635 diff_cred = amount - decayed_amount;
636 } else
637 diff_cred = diff_proc;
638 #ifdef notyet
639 KASSERT(diff_proc >= 0 || RACCT_CAN_DROP(resource),
640 ("%s: usage of non-droppable resource %d dropping", __func__,
641 resource));
642 #endif
643 #ifdef RCTL
644 if (diff_proc > 0) {
645 error = rctl_enforce(p, resource, diff_proc);
646 if (error && RACCT_IS_DENIABLE(resource)) {
647 SDT_PROBE3(racct, kernel, rusage, set__failure, p,
648 resource, amount);
649 return (error);
650 }
651 }
652 #endif
653 racct_alloc_resource(p->p_racct, resource, diff_proc);
654 if (diff_cred > 0)
655 racct_add_cred_locked(p->p_ucred, resource, diff_cred);
656 else if (diff_cred < 0)
657 racct_sub_cred_locked(p->p_ucred, resource, -diff_cred);
658
659 return (0);
660 }
661
662 /*
663 * Set allocation of 'resource' to 'amount' for process 'p'.
664 * Return 0 if it's below limits, or errno, if it's not.
665 *
666 * Note that decreasing the allocation always returns 0,
667 * even if it's above the limit.
668 */
669 int
670 racct_set(struct proc *p, int resource, uint64_t amount)
671 {
672 int error;
673
674 mtx_lock(&racct_lock);
675 error = racct_set_locked(p, resource, amount);
676 mtx_unlock(&racct_lock);
677 return (error);
678 }
679
680 static void
681 racct_set_force_locked(struct proc *p, int resource, uint64_t amount)
682 {
683 int64_t old_amount, decayed_amount;
684 int64_t diff_proc, diff_cred;
685
686 SDT_PROBE3(racct, kernel, rusage, set, p, resource, amount);
687
688 /*
689 * We need proc lock to dereference p->p_ucred.
690 */
691 PROC_LOCK_ASSERT(p, MA_OWNED);
692
693 old_amount = p->p_racct->r_resources[resource];
694 /*
695 * The diffs may be negative.
696 */
697 diff_proc = amount - old_amount;
698 if (RACCT_IS_DECAYING(resource)) {
699 /*
700 * Resources in per-credential racct containers may decay.
701 * If this is the case, we need to calculate the difference
702 * between the new amount and the proportional value of the
703 * old amount that has decayed in the ucred racct containers.
704 */
705 decayed_amount = old_amount * RACCT_DECAY_FACTOR / FSCALE;
706 diff_cred = amount - decayed_amount;
707 } else
708 diff_cred = diff_proc;
709
710 racct_alloc_resource(p->p_racct, resource, diff_proc);
711 if (diff_cred > 0)
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);
715 }
716
717 void
718 racct_set_force(struct proc *p, int resource, uint64_t amount)
719 {
720 mtx_lock(&racct_lock);
721 racct_set_force_locked(p, resource, amount);
722 mtx_unlock(&racct_lock);
723 }
724
725 /*
726 * Returns amount of 'resource' the process 'p' can keep allocated.
727 * Allocating more than that would be denied, unless the resource
728 * is marked undeniable. Amount of already allocated resource does
729 * not matter.
730 */
731 uint64_t
732 racct_get_limit(struct proc *p, int resource)
733 {
734
735 #ifdef RCTL
736 return (rctl_get_limit(p, resource));
737 #else
738 return (UINT64_MAX);
739 #endif
740 }
741
742 /*
743 * Returns amount of 'resource' the process 'p' can keep allocated.
744 * Allocating more than that would be denied, unless the resource
745 * is marked undeniable. Amount of already allocated resource does
746 * matter.
747 */
748 uint64_t
749 racct_get_available(struct proc *p, int resource)
750 {
751
752 #ifdef RCTL
753 return (rctl_get_available(p, resource));
754 #else
755 return (UINT64_MAX);
756 #endif
757 }
758
759 /*
760 * Returns amount of the %cpu resource that process 'p' can add to its %cpu
761 * utilization. Adding more than that would lead to the process being
762 * throttled.
763 */
764 static int64_t
765 racct_pcpu_available(struct proc *p)
766 {
767
768 #ifdef RCTL
769 return (rctl_pcpu_available(p));
770 #else
771 return (INT64_MAX);
772 #endif
773 }
774
775 /*
776 * Decrease allocation of 'resource' by 'amount' for process 'p'.
777 */
778 void
779 racct_sub(struct proc *p, int resource, uint64_t amount)
780 {
781
782 SDT_PROBE3(racct, kernel, rusage, sub, p, resource, amount);
783
784 /*
785 * We need proc lock to dereference p->p_ucred.
786 */
787 PROC_LOCK_ASSERT(p, MA_OWNED);
788 KASSERT(RACCT_CAN_DROP(resource),
789 ("%s: called for non-droppable resource %d", __func__, resource));
790
791 mtx_lock(&racct_lock);
792 KASSERT(amount <= p->p_racct->r_resources[resource],
793 ("%s: freeing %ju of resource %d, which is more "
794 "than allocated %jd for %s (pid %d)", __func__, amount, resource,
795 (intmax_t)p->p_racct->r_resources[resource], p->p_comm, p->p_pid));
796
797 racct_alloc_resource(p->p_racct, resource, -amount);
798 racct_sub_cred_locked(p->p_ucred, resource, amount);
799 mtx_unlock(&racct_lock);
800 }
801
802 static void
803 racct_sub_cred_locked(struct ucred *cred, int resource, uint64_t amount)
804 {
805 struct prison *pr;
806
807 SDT_PROBE3(racct, kernel, rusage, sub__cred, cred, resource, amount);
808
809 #ifdef notyet
810 KASSERT(RACCT_CAN_DROP(resource),
811 ("%s: called for resource %d which can not drop", __func__,
812 resource));
813 #endif
814
815 racct_alloc_resource(cred->cr_ruidinfo->ui_racct, resource, -amount);
816 for (pr = cred->cr_prison; pr != NULL; pr = pr->pr_parent)
817 racct_alloc_resource(pr->pr_prison_racct->prr_racct, resource,
818 -amount);
819 racct_alloc_resource(cred->cr_loginclass->lc_racct, resource, -amount);
820 }
821
822 /*
823 * Decrease allocation of 'resource' by 'amount' for credential 'cred'.
824 */
825 void
826 racct_sub_cred(struct ucred *cred, int resource, uint64_t amount)
827 {
828
829 mtx_lock(&racct_lock);
830 racct_sub_cred_locked(cred, resource, amount);
831 mtx_unlock(&racct_lock);
832 }
833
834 /*
835 * Inherit resource usage information from the parent process.
836 */
837 int
838 racct_proc_fork(struct proc *parent, struct proc *child)
839 {
840 int i, error = 0;
841
842 /*
843 * Create racct for the child process.
844 */
845 racct_create(&child->p_racct);
846
847 PROC_LOCK(parent);
848 PROC_LOCK(child);
849 mtx_lock(&racct_lock);
850
851 #ifdef RCTL
852 error = rctl_proc_fork(parent, child);
853 if (error != 0)
854 goto out;
855 #endif
856
857 /* Init process cpu time. */
858 child->p_prev_runtime = 0;
859 child->p_throttled = 0;
860
861 /*
862 * Inherit resource usage.
863 */
864 for (i = 0; i <= RACCT_MAX; i++) {
865 if (parent->p_racct->r_resources[i] == 0 ||
866 !RACCT_IS_INHERITABLE(i))
867 continue;
868
869 error = racct_set_locked(child, i,
870 parent->p_racct->r_resources[i]);
871 if (error != 0)
872 goto out;
873 }
874
875 error = racct_add_locked(child, RACCT_NPROC, 1);
876 error += racct_add_locked(child, RACCT_NTHR, 1);
877
878 out:
879 mtx_unlock(&racct_lock);
880 PROC_UNLOCK(child);
881 PROC_UNLOCK(parent);
882
883 if (error != 0)
884 racct_proc_exit(child);
885
886 return (error);
887 }
888
889 /*
890 * Called at the end of fork1(), to handle rules that require the process
891 * to be fully initialized.
892 */
893 void
894 racct_proc_fork_done(struct proc *child)
895 {
896
897 #ifdef RCTL
898 PROC_LOCK(child);
899 mtx_lock(&racct_lock);
900 rctl_enforce(child, RACCT_NPROC, 0);
901 rctl_enforce(child, RACCT_NTHR, 0);
902 mtx_unlock(&racct_lock);
903 PROC_UNLOCK(child);
904 #endif
905 }
906
907 void
908 racct_proc_exit(struct proc *p)
909 {
910 int i;
911 uint64_t runtime;
912 struct timeval wallclock;
913 uint64_t pct_estimate, pct;
914
915 PROC_LOCK(p);
916 /*
917 * We don't need to calculate rux, proc_reap() has already done this.
918 */
919 runtime = cputick2usec(p->p_rux.rux_runtime);
920 #ifdef notyet
921 KASSERT(runtime >= p->p_prev_runtime, ("runtime < p_prev_runtime"));
922 #else
923 if (runtime < p->p_prev_runtime)
924 runtime = p->p_prev_runtime;
925 #endif
926 microuptime(&wallclock);
927 timevalsub(&wallclock, &p->p_stats->p_start);
928 if (wallclock.tv_sec > 0 || wallclock.tv_usec > 0) {
929 pct_estimate = (1000000 * runtime * 100) /
930 ((uint64_t)wallclock.tv_sec * 1000000 +
931 wallclock.tv_usec);
932 } else
933 pct_estimate = 0;
934 pct = racct_getpcpu(p, pct_estimate);
935
936 mtx_lock(&racct_lock);
937 racct_set_locked(p, RACCT_CPU, runtime);
938 racct_add_cred_locked(p->p_ucred, RACCT_PCTCPU, pct);
939
940 for (i = 0; i <= RACCT_MAX; i++) {
941 if (p->p_racct->r_resources[i] == 0)
942 continue;
943 if (!RACCT_IS_RECLAIMABLE(i))
944 continue;
945 racct_set_locked(p, i, 0);
946 }
947
948 mtx_unlock(&racct_lock);
949 PROC_UNLOCK(p);
950
951 #ifdef RCTL
952 rctl_racct_release(p->p_racct);
953 #endif
954 racct_destroy(&p->p_racct);
955 }
956
957 /*
958 * Called after credentials change, to move resource utilisation
959 * between raccts.
960 */
961 void
962 racct_proc_ucred_changed(struct proc *p, struct ucred *oldcred,
963 struct ucred *newcred)
964 {
965 struct uidinfo *olduip, *newuip;
966 struct loginclass *oldlc, *newlc;
967 struct prison *oldpr, *newpr, *pr;
968
969 PROC_LOCK_ASSERT(p, MA_NOTOWNED);
970
971 newuip = newcred->cr_ruidinfo;
972 olduip = oldcred->cr_ruidinfo;
973 newlc = newcred->cr_loginclass;
974 oldlc = oldcred->cr_loginclass;
975 newpr = newcred->cr_prison;
976 oldpr = oldcred->cr_prison;
977
978 mtx_lock(&racct_lock);
979 if (newuip != olduip) {
980 racct_sub_racct(olduip->ui_racct, p->p_racct);
981 racct_add_racct(newuip->ui_racct, p->p_racct);
982 }
983 if (newlc != oldlc) {
984 racct_sub_racct(oldlc->lc_racct, p->p_racct);
985 racct_add_racct(newlc->lc_racct, p->p_racct);
986 }
987 if (newpr != oldpr) {
988 for (pr = oldpr; pr != NULL; pr = pr->pr_parent)
989 racct_sub_racct(pr->pr_prison_racct->prr_racct,
990 p->p_racct);
991 for (pr = newpr; pr != NULL; pr = pr->pr_parent)
992 racct_add_racct(pr->pr_prison_racct->prr_racct,
993 p->p_racct);
994 }
995 mtx_unlock(&racct_lock);
996
997 #ifdef RCTL
998 rctl_proc_ucred_changed(p, newcred);
999 #endif
1000 }
1001
1002 void
1003 racct_move(struct racct *dest, struct racct *src)
1004 {
1005
1006 mtx_lock(&racct_lock);
1007
1008 racct_add_racct(dest, src);
1009 racct_sub_racct(src, src);
1010
1011 mtx_unlock(&racct_lock);
1012 }
1013
1014 static void
1015 racct_proc_throttle(struct proc *p)
1016 {
1017 struct thread *td;
1018 #ifdef SMP
1019 int cpuid;
1020 #endif
1021
1022 PROC_LOCK_ASSERT(p, MA_OWNED);
1023
1024 /*
1025 * Do not block kernel processes. Also do not block processes with
1026 * low %cpu utilization to improve interactivity.
1027 */
1028 if (((p->p_flag & (P_SYSTEM | P_KTHREAD)) != 0) ||
1029 (p->p_racct->r_resources[RACCT_PCTCPU] <= pcpu_threshold))
1030 return;
1031 p->p_throttled = 1;
1032
1033 FOREACH_THREAD_IN_PROC(p, td) {
1034 thread_lock(td);
1035 switch (td->td_state) {
1036 case TDS_RUNQ:
1037 /*
1038 * If the thread is on the scheduler run-queue, we can
1039 * not just remove it from there. So we set the flag
1040 * TDF_NEEDRESCHED for the thread, so that once it is
1041 * running, it is taken off the cpu as soon as possible.
1042 */
1043 td->td_flags |= TDF_NEEDRESCHED;
1044 break;
1045 case TDS_RUNNING:
1046 /*
1047 * If the thread is running, we request a context
1048 * switch for it by setting the TDF_NEEDRESCHED flag.
1049 */
1050 td->td_flags |= TDF_NEEDRESCHED;
1051 #ifdef SMP
1052 cpuid = td->td_oncpu;
1053 if ((cpuid != NOCPU) && (td != curthread))
1054 ipi_cpu(cpuid, IPI_AST);
1055 #endif
1056 break;
1057 default:
1058 break;
1059 }
1060 thread_unlock(td);
1061 }
1062 }
1063
1064 static void
1065 racct_proc_wakeup(struct proc *p)
1066 {
1067 PROC_LOCK_ASSERT(p, MA_OWNED);
1068
1069 if (p->p_throttled) {
1070 p->p_throttled = 0;
1071 wakeup(p->p_racct);
1072 }
1073 }
1074
1075 static void
1076 racct_decay_resource(struct racct *racct, void * res, void* dummy)
1077 {
1078 int resource;
1079 int64_t r_old, r_new;
1080
1081 resource = *(int *)res;
1082 r_old = racct->r_resources[resource];
1083
1084 /* If there is nothing to decay, just exit. */
1085 if (r_old <= 0)
1086 return;
1087
1088 mtx_lock(&racct_lock);
1089 r_new = r_old * RACCT_DECAY_FACTOR / FSCALE;
1090 racct->r_resources[resource] = r_new;
1091 mtx_unlock(&racct_lock);
1092 }
1093
1094 static void
1095 racct_decay(int resource)
1096 {
1097 ui_racct_foreach(racct_decay_resource, &resource, NULL);
1098 loginclass_racct_foreach(racct_decay_resource, &resource, NULL);
1099 prison_racct_foreach(racct_decay_resource, &resource, NULL);
1100 }
1101
1102 static void
1103 racctd(void)
1104 {
1105 struct thread *td;
1106 struct proc *p;
1107 struct timeval wallclock;
1108 uint64_t runtime;
1109 uint64_t pct, pct_estimate;
1110
1111 for (;;) {
1112 racct_decay(RACCT_PCTCPU);
1113
1114 sx_slock(&allproc_lock);
1115
1116 LIST_FOREACH(p, &zombproc, p_list) {
1117 PROC_LOCK(p);
1118 racct_set(p, RACCT_PCTCPU, 0);
1119 PROC_UNLOCK(p);
1120 }
1121
1122 FOREACH_PROC_IN_SYSTEM(p) {
1123 PROC_LOCK(p);
1124 if (p->p_state != PRS_NORMAL) {
1125 PROC_UNLOCK(p);
1126 continue;
1127 }
1128
1129 microuptime(&wallclock);
1130 timevalsub(&wallclock, &p->p_stats->p_start);
1131 PROC_SLOCK(p);
1132 FOREACH_THREAD_IN_PROC(p, td)
1133 ruxagg(p, td);
1134 runtime = cputick2usec(p->p_rux.rux_runtime);
1135 PROC_SUNLOCK(p);
1136 #ifdef notyet
1137 KASSERT(runtime >= p->p_prev_runtime,
1138 ("runtime < p_prev_runtime"));
1139 #else
1140 if (runtime < p->p_prev_runtime)
1141 runtime = p->p_prev_runtime;
1142 #endif
1143 p->p_prev_runtime = runtime;
1144 if (wallclock.tv_sec > 0 || wallclock.tv_usec > 0) {
1145 pct_estimate = (1000000 * runtime * 100) /
1146 ((uint64_t)wallclock.tv_sec * 1000000 +
1147 wallclock.tv_usec);
1148 } else
1149 pct_estimate = 0;
1150 pct = racct_getpcpu(p, pct_estimate);
1151 mtx_lock(&racct_lock);
1152 racct_set_force_locked(p, RACCT_PCTCPU, pct);
1153 racct_set_locked(p, RACCT_CPU, runtime);
1154 racct_set_locked(p, RACCT_WALLCLOCK,
1155 (uint64_t)wallclock.tv_sec * 1000000 +
1156 wallclock.tv_usec);
1157 mtx_unlock(&racct_lock);
1158 PROC_UNLOCK(p);
1159 }
1160
1161 /*
1162 * To ensure that processes are throttled in a fair way, we need
1163 * to iterate over all processes again and check the limits
1164 * for %cpu resource only after ucred racct containers have been
1165 * properly filled.
1166 */
1167 FOREACH_PROC_IN_SYSTEM(p) {
1168 PROC_LOCK(p);
1169 if (p->p_state != PRS_NORMAL) {
1170 PROC_UNLOCK(p);
1171 continue;
1172 }
1173
1174 if (racct_pcpu_available(p) <= 0)
1175 racct_proc_throttle(p);
1176 else if (p->p_throttled)
1177 racct_proc_wakeup(p);
1178 PROC_UNLOCK(p);
1179 }
1180 sx_sunlock(&allproc_lock);
1181 pause("-", hz);
1182 }
1183 }
1184
1185 static struct kproc_desc racctd_kp = {
1186 "racctd",
1187 racctd,
1188 NULL
1189 };
1190 SYSINIT(racctd, SI_SUB_RACCTD, SI_ORDER_FIRST, kproc_start, &racctd_kp);
1191
1192 static void
1193 racct_init(void)
1194 {
1195
1196 racct_zone = uma_zcreate("racct", sizeof(struct racct),
1197 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
1198 /*
1199 * XXX: Move this somewhere.
1200 */
1201 prison0.pr_prison_racct = prison_racct_find("");
1202 }
1203 SYSINIT(racct, SI_SUB_RACCT, SI_ORDER_FIRST, racct_init, NULL);
1204
1205 #else /* !RACCT */
1206
1207 int
1208 racct_add(struct proc *p, int resource, uint64_t amount)
1209 {
1210
1211 return (0);
1212 }
1213
1214 void
1215 racct_add_cred(struct ucred *cred, int resource, uint64_t amount)
1216 {
1217 }
1218
1219 void
1220 racct_add_force(struct proc *p, int resource, uint64_t amount)
1221 {
1222
1223 return;
1224 }
1225
1226 int
1227 racct_set(struct proc *p, int resource, uint64_t amount)
1228 {
1229
1230 return (0);
1231 }
1232
1233 void
1234 racct_set_force(struct proc *p, int resource, uint64_t amount)
1235 {
1236 }
1237
1238 void
1239 racct_sub(struct proc *p, int resource, uint64_t amount)
1240 {
1241 }
1242
1243 void
1244 racct_sub_cred(struct ucred *cred, int resource, uint64_t amount)
1245 {
1246 }
1247
1248 uint64_t
1249 racct_get_limit(struct proc *p, int resource)
1250 {
1251
1252 return (UINT64_MAX);
1253 }
1254
1255 uint64_t
1256 racct_get_available(struct proc *p, int resource)
1257 {
1258
1259 return (UINT64_MAX);
1260 }
1261
1262 void
1263 racct_create(struct racct **racctp)
1264 {
1265 }
1266
1267 void
1268 racct_destroy(struct racct **racctp)
1269 {
1270 }
1271
1272 int
1273 racct_proc_fork(struct proc *parent, struct proc *child)
1274 {
1275
1276 return (0);
1277 }
1278
1279 void
1280 racct_proc_fork_done(struct proc *child)
1281 {
1282 }
1283
1284 void
1285 racct_proc_exit(struct proc *p)
1286 {
1287 }
1288
1289 #endif /* !RACCT */
Cache object: 634c99f8ed083b365f522e70ecec5d63
|