1097055e2SEdward Tomasz Napierala /*- 2*4d846d26SWarner Losh * SPDX-License-Identifier: BSD-2-Clause 38a36da99SPedro F. Giffuni * 4097055e2SEdward Tomasz Napierala * Copyright (c) 2010 The FreeBSD Foundation 5097055e2SEdward Tomasz Napierala * 6097055e2SEdward Tomasz Napierala * This software was developed by Edward Tomasz Napierala under sponsorship 7097055e2SEdward Tomasz Napierala * from the FreeBSD Foundation. 8097055e2SEdward Tomasz Napierala * 9097055e2SEdward Tomasz Napierala * Redistribution and use in source and binary forms, with or without 10097055e2SEdward Tomasz Napierala * modification, are permitted provided that the following conditions 11097055e2SEdward Tomasz Napierala * are met: 12097055e2SEdward Tomasz Napierala * 1. Redistributions of source code must retain the above copyright 13097055e2SEdward Tomasz Napierala * notice, this list of conditions and the following disclaimer. 14097055e2SEdward Tomasz Napierala * 2. Redistributions in binary form must reproduce the above copyright 15097055e2SEdward Tomasz Napierala * notice, this list of conditions and the following disclaimer in the 16097055e2SEdward Tomasz Napierala * documentation and/or other materials provided with the distribution. 17097055e2SEdward Tomasz Napierala * 18097055e2SEdward Tomasz Napierala * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 19097055e2SEdward Tomasz Napierala * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 20097055e2SEdward Tomasz Napierala * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 21097055e2SEdward Tomasz Napierala * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 22097055e2SEdward Tomasz Napierala * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 23097055e2SEdward Tomasz Napierala * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 24097055e2SEdward Tomasz Napierala * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 25097055e2SEdward Tomasz Napierala * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 26097055e2SEdward Tomasz Napierala * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 27097055e2SEdward Tomasz Napierala * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 28097055e2SEdward Tomasz Napierala * SUCH DAMAGE. 29097055e2SEdward Tomasz Napierala */ 30097055e2SEdward Tomasz Napierala 31097055e2SEdward Tomasz Napierala #include <sys/cdefs.h> 3236af9869SEdward Tomasz Napierala #include "opt_sched.h" 33097055e2SEdward Tomasz Napierala 34097055e2SEdward Tomasz Napierala #include <sys/param.h> 35ae34b6ffSEdward Tomasz Napierala #include <sys/buf.h> 360e225211SAndriy Gapon #include <sys/systm.h> 37097055e2SEdward Tomasz Napierala #include <sys/eventhandler.h> 38097055e2SEdward Tomasz Napierala #include <sys/jail.h> 39097055e2SEdward Tomasz Napierala #include <sys/kernel.h> 40097055e2SEdward Tomasz Napierala #include <sys/kthread.h> 41097055e2SEdward Tomasz Napierala #include <sys/lock.h> 42097055e2SEdward Tomasz Napierala #include <sys/loginclass.h> 43097055e2SEdward Tomasz Napierala #include <sys/malloc.h> 44097055e2SEdward Tomasz Napierala #include <sys/mutex.h> 45097055e2SEdward Tomasz Napierala #include <sys/proc.h> 46097055e2SEdward Tomasz Napierala #include <sys/racct.h> 47097055e2SEdward Tomasz Napierala #include <sys/resourcevar.h> 48097055e2SEdward Tomasz Napierala #include <sys/sbuf.h> 49097055e2SEdward Tomasz Napierala #include <sys/sched.h> 50097055e2SEdward Tomasz Napierala #include <sys/sdt.h> 5136af9869SEdward Tomasz Napierala #include <sys/smp.h> 52097055e2SEdward Tomasz Napierala #include <sys/sx.h> 5336af9869SEdward Tomasz Napierala #include <sys/sysctl.h> 54097055e2SEdward Tomasz Napierala #include <sys/sysproto.h> 55af29f399SDmitry Chagin #include <sys/umtxvar.h> 5636af9869SEdward Tomasz Napierala #include <machine/smp.h> 57097055e2SEdward Tomasz Napierala 58097055e2SEdward Tomasz Napierala #ifdef RCTL 59097055e2SEdward Tomasz Napierala #include <sys/rctl.h> 60097055e2SEdward Tomasz Napierala #endif 61097055e2SEdward Tomasz Napierala 62097055e2SEdward Tomasz Napierala FEATURE(racct, "Resource Accounting"); 63097055e2SEdward Tomasz Napierala 6436af9869SEdward Tomasz Napierala /* 6536af9869SEdward Tomasz Napierala * Do not block processes that have their %cpu usage <= pcpu_threshold. 6636af9869SEdward Tomasz Napierala */ 6736af9869SEdward Tomasz Napierala static int pcpu_threshold = 1; 68ba8f0eb8SEdward Tomasz Napierala #ifdef RACCT_DEFAULT_TO_DISABLED 69eec8d0a3SMateusz Guzik bool __read_frequently racct_enable = false; 704b5c9cf6SEdward Tomasz Napierala #else 71eec8d0a3SMateusz Guzik bool __read_frequently racct_enable = true; 724b5c9cf6SEdward Tomasz Napierala #endif 7336af9869SEdward Tomasz Napierala 747029da5cSPawel Biernacki SYSCTL_NODE(_kern, OID_AUTO, racct, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 757029da5cSPawel Biernacki "Resource Accounting"); 76eec8d0a3SMateusz Guzik SYSCTL_BOOL(_kern_racct, OID_AUTO, enable, CTLFLAG_RDTUN, &racct_enable, 774b5c9cf6SEdward Tomasz Napierala 0, "Enable RACCT/RCTL"); 7836af9869SEdward Tomasz Napierala SYSCTL_UINT(_kern_racct, OID_AUTO, pcpu_threshold, CTLFLAG_RW, &pcpu_threshold, 7936af9869SEdward Tomasz Napierala 0, "Processes with higher %cpu usage than this value can be throttled."); 8036af9869SEdward Tomasz Napierala 8136af9869SEdward Tomasz Napierala /* 8236af9869SEdward Tomasz Napierala * How many seconds it takes to use the scheduler %cpu calculations. When a 8336af9869SEdward Tomasz Napierala * process starts, we compute its %cpu usage by dividing its runtime by the 8436af9869SEdward Tomasz Napierala * process wall clock time. After RACCT_PCPU_SECS pass, we use the value 8536af9869SEdward Tomasz Napierala * provided by the scheduler. 8636af9869SEdward Tomasz Napierala */ 8736af9869SEdward Tomasz Napierala #define RACCT_PCPU_SECS 3 8836af9869SEdward Tomasz Napierala 89bbe4eb6dSEdward Tomasz Napierala struct mtx racct_lock; 90097055e2SEdward Tomasz Napierala MTX_SYSINIT(racct_lock, &racct_lock, "racct lock", MTX_DEF); 91097055e2SEdward Tomasz Napierala 92097055e2SEdward Tomasz Napierala static uma_zone_t racct_zone; 93097055e2SEdward Tomasz Napierala 94097055e2SEdward Tomasz Napierala static void racct_sub_racct(struct racct *dest, const struct racct *src); 95097055e2SEdward Tomasz Napierala static void racct_sub_cred_locked(struct ucred *cred, int resource, 96097055e2SEdward Tomasz Napierala uint64_t amount); 97097055e2SEdward Tomasz Napierala static void racct_add_cred_locked(struct ucred *cred, int resource, 98097055e2SEdward Tomasz Napierala uint64_t amount); 99097055e2SEdward Tomasz Napierala 100097055e2SEdward Tomasz Napierala SDT_PROVIDER_DEFINE(racct); 10136160958SMark Johnston SDT_PROBE_DEFINE3(racct, , rusage, add, 102097055e2SEdward Tomasz Napierala "struct proc *", "int", "uint64_t"); 10336160958SMark Johnston SDT_PROBE_DEFINE3(racct, , rusage, add__failure, 104097055e2SEdward Tomasz Napierala "struct proc *", "int", "uint64_t"); 105bbe4eb6dSEdward Tomasz Napierala SDT_PROBE_DEFINE3(racct, , rusage, add__buf, 106bbe4eb6dSEdward Tomasz Napierala "struct proc *", "const struct buf *", "int"); 10736160958SMark Johnston SDT_PROBE_DEFINE3(racct, , rusage, add__cred, 10836160958SMark Johnston "struct ucred *", "int", "uint64_t"); 10936160958SMark Johnston SDT_PROBE_DEFINE3(racct, , rusage, add__force, 11036160958SMark Johnston "struct proc *", "int", "uint64_t"); 11136160958SMark Johnston SDT_PROBE_DEFINE3(racct, , rusage, set, 11236160958SMark Johnston "struct proc *", "int", "uint64_t"); 11336160958SMark Johnston SDT_PROBE_DEFINE3(racct, , rusage, set__failure, 11436160958SMark Johnston "struct proc *", "int", "uint64_t"); 11510287198SEdward Tomasz Napierala SDT_PROBE_DEFINE3(racct, , rusage, set__force, 11610287198SEdward Tomasz Napierala "struct proc *", "int", "uint64_t"); 11736160958SMark Johnston SDT_PROBE_DEFINE3(racct, , rusage, sub, 11836160958SMark Johnston "struct proc *", "int", "uint64_t"); 11936160958SMark Johnston SDT_PROBE_DEFINE3(racct, , rusage, sub__cred, 12036160958SMark Johnston "struct ucred *", "int", "uint64_t"); 12136160958SMark Johnston SDT_PROBE_DEFINE1(racct, , racct, create, 122097055e2SEdward Tomasz Napierala "struct racct *"); 12336160958SMark Johnston SDT_PROBE_DEFINE1(racct, , racct, destroy, 12436160958SMark Johnston "struct racct *"); 12536160958SMark Johnston SDT_PROBE_DEFINE2(racct, , racct, join, 126097055e2SEdward Tomasz Napierala "struct racct *", "struct racct *"); 12736160958SMark Johnston SDT_PROBE_DEFINE2(racct, , racct, join__failure, 12836160958SMark Johnston "struct racct *", "struct racct *"); 12936160958SMark Johnston SDT_PROBE_DEFINE2(racct, , racct, leave, 13036160958SMark Johnston "struct racct *", "struct racct *"); 131097055e2SEdward Tomasz Napierala 132097055e2SEdward Tomasz Napierala int racct_types[] = { 133097055e2SEdward Tomasz Napierala [RACCT_CPU] = 13485a2f1b4SEdward Tomasz Napierala RACCT_IN_MILLIONS, 135097055e2SEdward Tomasz Napierala [RACCT_DATA] = 136097055e2SEdward Tomasz Napierala RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE, 137097055e2SEdward Tomasz Napierala [RACCT_STACK] = 138097055e2SEdward Tomasz Napierala RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE, 139097055e2SEdward Tomasz Napierala [RACCT_CORE] = 140097055e2SEdward Tomasz Napierala RACCT_DENIABLE, 141097055e2SEdward Tomasz Napierala [RACCT_RSS] = 142097055e2SEdward Tomasz Napierala RACCT_RECLAIMABLE, 143097055e2SEdward Tomasz Napierala [RACCT_MEMLOCK] = 144097055e2SEdward Tomasz Napierala RACCT_RECLAIMABLE | RACCT_DENIABLE, 145097055e2SEdward Tomasz Napierala [RACCT_NPROC] = 146097055e2SEdward Tomasz Napierala RACCT_RECLAIMABLE | RACCT_DENIABLE, 147097055e2SEdward Tomasz Napierala [RACCT_NOFILE] = 148097055e2SEdward Tomasz Napierala RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE, 149097055e2SEdward Tomasz Napierala [RACCT_VMEM] = 150097055e2SEdward Tomasz Napierala RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE, 151097055e2SEdward Tomasz Napierala [RACCT_NPTS] = 152097055e2SEdward Tomasz Napierala RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY, 153097055e2SEdward Tomasz Napierala [RACCT_SWAP] = 154097055e2SEdward Tomasz Napierala RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY, 155097055e2SEdward Tomasz Napierala [RACCT_NTHR] = 156097055e2SEdward Tomasz Napierala RACCT_RECLAIMABLE | RACCT_DENIABLE, 157097055e2SEdward Tomasz Napierala [RACCT_MSGQQUEUED] = 158097055e2SEdward Tomasz Napierala RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY, 159097055e2SEdward Tomasz Napierala [RACCT_MSGQSIZE] = 160097055e2SEdward Tomasz Napierala RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY, 161097055e2SEdward Tomasz Napierala [RACCT_NMSGQ] = 162097055e2SEdward Tomasz Napierala RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY, 163097055e2SEdward Tomasz Napierala [RACCT_NSEM] = 164097055e2SEdward Tomasz Napierala RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY, 165097055e2SEdward Tomasz Napierala [RACCT_NSEMOP] = 166097055e2SEdward Tomasz Napierala RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE, 167097055e2SEdward Tomasz Napierala [RACCT_NSHM] = 168097055e2SEdward Tomasz Napierala RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY, 169097055e2SEdward Tomasz Napierala [RACCT_SHMSIZE] = 170097055e2SEdward Tomasz Napierala RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY, 171097055e2SEdward Tomasz Napierala [RACCT_WALLCLOCK] = 17236af9869SEdward Tomasz Napierala RACCT_IN_MILLIONS, 17336af9869SEdward Tomasz Napierala [RACCT_PCTCPU] = 174ae34b6ffSEdward Tomasz Napierala RACCT_DECAYING | RACCT_DENIABLE | RACCT_IN_MILLIONS, 175ae34b6ffSEdward Tomasz Napierala [RACCT_READBPS] = 176ae34b6ffSEdward Tomasz Napierala RACCT_DECAYING, 177ae34b6ffSEdward Tomasz Napierala [RACCT_WRITEBPS] = 178ae34b6ffSEdward Tomasz Napierala RACCT_DECAYING, 179ae34b6ffSEdward Tomasz Napierala [RACCT_READIOPS] = 180ae34b6ffSEdward Tomasz Napierala RACCT_DECAYING, 181ae34b6ffSEdward Tomasz Napierala [RACCT_WRITEIOPS] = 182ae34b6ffSEdward Tomasz Napierala RACCT_DECAYING }; 18336af9869SEdward Tomasz Napierala 18436af9869SEdward Tomasz Napierala static const fixpt_t RACCT_DECAY_FACTOR = 0.3 * FSCALE; 18536af9869SEdward Tomasz Napierala 18636af9869SEdward Tomasz Napierala #ifdef SCHED_4BSD 18736af9869SEdward Tomasz Napierala /* 18836af9869SEdward Tomasz Napierala * Contains intermediate values for %cpu calculations to avoid using floating 18936af9869SEdward Tomasz Napierala * point in the kernel. 19036af9869SEdward Tomasz Napierala * ccpu_exp[k] = FSCALE * (ccpu/FSCALE)^k = FSCALE * exp(-k/20) 19136af9869SEdward Tomasz Napierala * It is needed only for the 4BSD scheduler, because in ULE, the ccpu equals to 19236af9869SEdward Tomasz Napierala * zero so the calculations are more straightforward. 19336af9869SEdward Tomasz Napierala */ 19436af9869SEdward Tomasz Napierala fixpt_t ccpu_exp[] = { 19536af9869SEdward Tomasz Napierala [0] = FSCALE * 1, 19636af9869SEdward Tomasz Napierala [1] = FSCALE * 0.95122942450071400909, 19736af9869SEdward Tomasz Napierala [2] = FSCALE * 0.90483741803595957316, 19836af9869SEdward Tomasz Napierala [3] = FSCALE * 0.86070797642505780722, 19936af9869SEdward Tomasz Napierala [4] = FSCALE * 0.81873075307798185866, 20036af9869SEdward Tomasz Napierala [5] = FSCALE * 0.77880078307140486824, 20136af9869SEdward Tomasz Napierala [6] = FSCALE * 0.74081822068171786606, 20236af9869SEdward Tomasz Napierala [7] = FSCALE * 0.70468808971871343435, 20336af9869SEdward Tomasz Napierala [8] = FSCALE * 0.67032004603563930074, 20436af9869SEdward Tomasz Napierala [9] = FSCALE * 0.63762815162177329314, 20536af9869SEdward Tomasz Napierala [10] = FSCALE * 0.60653065971263342360, 20636af9869SEdward Tomasz Napierala [11] = FSCALE * 0.57694981038048669531, 20736af9869SEdward Tomasz Napierala [12] = FSCALE * 0.54881163609402643262, 20836af9869SEdward Tomasz Napierala [13] = FSCALE * 0.52204577676101604789, 20936af9869SEdward Tomasz Napierala [14] = FSCALE * 0.49658530379140951470, 21036af9869SEdward Tomasz Napierala [15] = FSCALE * 0.47236655274101470713, 21136af9869SEdward Tomasz Napierala [16] = FSCALE * 0.44932896411722159143, 21236af9869SEdward Tomasz Napierala [17] = FSCALE * 0.42741493194872666992, 21336af9869SEdward Tomasz Napierala [18] = FSCALE * 0.40656965974059911188, 21436af9869SEdward Tomasz Napierala [19] = FSCALE * 0.38674102345450120691, 21536af9869SEdward Tomasz Napierala [20] = FSCALE * 0.36787944117144232159, 21636af9869SEdward Tomasz Napierala [21] = FSCALE * 0.34993774911115535467, 21736af9869SEdward Tomasz Napierala [22] = FSCALE * 0.33287108369807955328, 21836af9869SEdward Tomasz Napierala [23] = FSCALE * 0.31663676937905321821, 21936af9869SEdward Tomasz Napierala [24] = FSCALE * 0.30119421191220209664, 22036af9869SEdward Tomasz Napierala [25] = FSCALE * 0.28650479686019010032, 22136af9869SEdward Tomasz Napierala [26] = FSCALE * 0.27253179303401260312, 22236af9869SEdward Tomasz Napierala [27] = FSCALE * 0.25924026064589150757, 22336af9869SEdward Tomasz Napierala [28] = FSCALE * 0.24659696394160647693, 22436af9869SEdward Tomasz Napierala [29] = FSCALE * 0.23457028809379765313, 22536af9869SEdward Tomasz Napierala [30] = FSCALE * 0.22313016014842982893, 22636af9869SEdward Tomasz Napierala [31] = FSCALE * 0.21224797382674305771, 22736af9869SEdward Tomasz Napierala [32] = FSCALE * 0.20189651799465540848, 22836af9869SEdward Tomasz Napierala [33] = FSCALE * 0.19204990862075411423, 22936af9869SEdward Tomasz Napierala [34] = FSCALE * 0.18268352405273465022, 23036af9869SEdward Tomasz Napierala [35] = FSCALE * 0.17377394345044512668, 23136af9869SEdward Tomasz Napierala [36] = FSCALE * 0.16529888822158653829, 23236af9869SEdward Tomasz Napierala [37] = FSCALE * 0.15723716631362761621, 23336af9869SEdward Tomasz Napierala [38] = FSCALE * 0.14956861922263505264, 23436af9869SEdward Tomasz Napierala [39] = FSCALE * 0.14227407158651357185, 23536af9869SEdward Tomasz Napierala [40] = FSCALE * 0.13533528323661269189, 23636af9869SEdward Tomasz Napierala [41] = FSCALE * 0.12873490358780421886, 23736af9869SEdward Tomasz Napierala [42] = FSCALE * 0.12245642825298191021, 23836af9869SEdward Tomasz Napierala [43] = FSCALE * 0.11648415777349695786, 23936af9869SEdward Tomasz Napierala [44] = FSCALE * 0.11080315836233388333, 24036af9869SEdward Tomasz Napierala [45] = FSCALE * 0.10539922456186433678, 24136af9869SEdward Tomasz Napierala [46] = FSCALE * 0.10025884372280373372, 24236af9869SEdward Tomasz Napierala [47] = FSCALE * 0.09536916221554961888, 24336af9869SEdward Tomasz Napierala [48] = FSCALE * 0.09071795328941250337, 24436af9869SEdward Tomasz Napierala [49] = FSCALE * 0.08629358649937051097, 24536af9869SEdward Tomasz Napierala [50] = FSCALE * 0.08208499862389879516, 24636af9869SEdward Tomasz Napierala [51] = FSCALE * 0.07808166600115315231, 24736af9869SEdward Tomasz Napierala [52] = FSCALE * 0.07427357821433388042, 24836af9869SEdward Tomasz Napierala [53] = FSCALE * 0.07065121306042958674, 24936af9869SEdward Tomasz Napierala [54] = FSCALE * 0.06720551273974976512, 25036af9869SEdward Tomasz Napierala [55] = FSCALE * 0.06392786120670757270, 25136af9869SEdward Tomasz Napierala [56] = FSCALE * 0.06081006262521796499, 25236af9869SEdward Tomasz Napierala [57] = FSCALE * 0.05784432087483846296, 25336af9869SEdward Tomasz Napierala [58] = FSCALE * 0.05502322005640722902, 25436af9869SEdward Tomasz Napierala [59] = FSCALE * 0.05233970594843239308, 25536af9869SEdward Tomasz Napierala [60] = FSCALE * 0.04978706836786394297, 25636af9869SEdward Tomasz Napierala [61] = FSCALE * 0.04735892439114092119, 25736af9869SEdward Tomasz Napierala [62] = FSCALE * 0.04504920239355780606, 25836af9869SEdward Tomasz Napierala [63] = FSCALE * 0.04285212686704017991, 25936af9869SEdward Tomasz Napierala [64] = FSCALE * 0.04076220397836621516, 26036af9869SEdward Tomasz Napierala [65] = FSCALE * 0.03877420783172200988, 26136af9869SEdward Tomasz Napierala [66] = FSCALE * 0.03688316740124000544, 26236af9869SEdward Tomasz Napierala [67] = FSCALE * 0.03508435410084502588, 26336af9869SEdward Tomasz Napierala [68] = FSCALE * 0.03337326996032607948, 26436af9869SEdward Tomasz Napierala [69] = FSCALE * 0.03174563637806794323, 26536af9869SEdward Tomasz Napierala [70] = FSCALE * 0.03019738342231850073, 26636af9869SEdward Tomasz Napierala [71] = FSCALE * 0.02872463965423942912, 26736af9869SEdward Tomasz Napierala [72] = FSCALE * 0.02732372244729256080, 26836af9869SEdward Tomasz Napierala [73] = FSCALE * 0.02599112877875534358, 26936af9869SEdward Tomasz Napierala [74] = FSCALE * 0.02472352647033939120, 27036af9869SEdward Tomasz Napierala [75] = FSCALE * 0.02351774585600910823, 27136af9869SEdward Tomasz Napierala [76] = FSCALE * 0.02237077185616559577, 27236af9869SEdward Tomasz Napierala [77] = FSCALE * 0.02127973643837716938, 27336af9869SEdward Tomasz Napierala [78] = FSCALE * 0.02024191144580438847, 27436af9869SEdward Tomasz Napierala [79] = FSCALE * 0.01925470177538692429, 27536af9869SEdward Tomasz Napierala [80] = FSCALE * 0.01831563888873418029, 27636af9869SEdward Tomasz Napierala [81] = FSCALE * 0.01742237463949351138, 27736af9869SEdward Tomasz Napierala [82] = FSCALE * 0.01657267540176124754, 27836af9869SEdward Tomasz Napierala [83] = FSCALE * 0.01576441648485449082, 27936af9869SEdward Tomasz Napierala [84] = FSCALE * 0.01499557682047770621, 28036af9869SEdward Tomasz Napierala [85] = FSCALE * 0.01426423390899925527, 28136af9869SEdward Tomasz Napierala [86] = FSCALE * 0.01356855901220093175, 28236af9869SEdward Tomasz Napierala [87] = FSCALE * 0.01290681258047986886, 28336af9869SEdward Tomasz Napierala [88] = FSCALE * 0.01227733990306844117, 28436af9869SEdward Tomasz Napierala [89] = FSCALE * 0.01167856697039544521, 28536af9869SEdward Tomasz Napierala [90] = FSCALE * 0.01110899653824230649, 28636af9869SEdward Tomasz Napierala [91] = FSCALE * 0.01056720438385265337, 28736af9869SEdward Tomasz Napierala [92] = FSCALE * 0.01005183574463358164, 28836af9869SEdward Tomasz Napierala [93] = FSCALE * 0.00956160193054350793, 28936af9869SEdward Tomasz Napierala [94] = FSCALE * 0.00909527710169581709, 29036af9869SEdward Tomasz Napierala [95] = FSCALE * 0.00865169520312063417, 29136af9869SEdward Tomasz Napierala [96] = FSCALE * 0.00822974704902002884, 29236af9869SEdward Tomasz Napierala [97] = FSCALE * 0.00782837754922577143, 29336af9869SEdward Tomasz Napierala [98] = FSCALE * 0.00744658307092434051, 29436af9869SEdward Tomasz Napierala [99] = FSCALE * 0.00708340892905212004, 29536af9869SEdward Tomasz Napierala [100] = FSCALE * 0.00673794699908546709, 29636af9869SEdward Tomasz Napierala [101] = FSCALE * 0.00640933344625638184, 29736af9869SEdward Tomasz Napierala [102] = FSCALE * 0.00609674656551563610, 29836af9869SEdward Tomasz Napierala [103] = FSCALE * 0.00579940472684214321, 29936af9869SEdward Tomasz Napierala [104] = FSCALE * 0.00551656442076077241, 30036af9869SEdward Tomasz Napierala [105] = FSCALE * 0.00524751839918138427, 30136af9869SEdward Tomasz Napierala [106] = FSCALE * 0.00499159390691021621, 30236af9869SEdward Tomasz Napierala [107] = FSCALE * 0.00474815099941147558, 30336af9869SEdward Tomasz Napierala [108] = FSCALE * 0.00451658094261266798, 30436af9869SEdward Tomasz Napierala [109] = FSCALE * 0.00429630469075234057, 30536af9869SEdward Tomasz Napierala [110] = FSCALE * 0.00408677143846406699, 30636af9869SEdward Tomasz Napierala }; 30736af9869SEdward Tomasz Napierala #endif 30836af9869SEdward Tomasz Napierala 30936af9869SEdward Tomasz Napierala #define CCPU_EXP_MAX 110 31036af9869SEdward Tomasz Napierala 31136af9869SEdward Tomasz Napierala /* 31236af9869SEdward Tomasz Napierala * This function is analogical to the getpcpu() function in the ps(1) command. 31336af9869SEdward Tomasz Napierala * They should both calculate in the same way so that the racct %cpu 31436af9869SEdward Tomasz Napierala * calculations are consistent with the values showed by the ps(1) tool. 31536af9869SEdward Tomasz Napierala * The calculations are more complex in the 4BSD scheduler because of the value 31636af9869SEdward Tomasz Napierala * of the ccpu variable. In ULE it is defined to be zero which saves us some 31736af9869SEdward Tomasz Napierala * work. 31836af9869SEdward Tomasz Napierala */ 31936af9869SEdward Tomasz Napierala static uint64_t 32036af9869SEdward Tomasz Napierala racct_getpcpu(struct proc *p, u_int pcpu) 32136af9869SEdward Tomasz Napierala { 32236af9869SEdward Tomasz Napierala u_int swtime; 32336af9869SEdward Tomasz Napierala #ifdef SCHED_4BSD 32436af9869SEdward Tomasz Napierala fixpt_t pctcpu, pctcpu_next; 32536af9869SEdward Tomasz Napierala #endif 32636af9869SEdward Tomasz Napierala #ifdef SMP 32736af9869SEdward Tomasz Napierala struct pcpu *pc; 32836af9869SEdward Tomasz Napierala int found; 32936af9869SEdward Tomasz Napierala #endif 33036af9869SEdward Tomasz Napierala fixpt_t p_pctcpu; 33136af9869SEdward Tomasz Napierala struct thread *td; 33236af9869SEdward Tomasz Napierala 3334b5c9cf6SEdward Tomasz Napierala ASSERT_RACCT_ENABLED(); 3344b5c9cf6SEdward Tomasz Napierala 33536af9869SEdward Tomasz Napierala swtime = (ticks - p->p_swtick) / hz; 33636af9869SEdward Tomasz Napierala 33736af9869SEdward Tomasz Napierala /* 33836af9869SEdward Tomasz Napierala * For short-lived processes, the sched_pctcpu() returns small 33936af9869SEdward Tomasz Napierala * values even for cpu intensive processes. Therefore we use 34036af9869SEdward Tomasz Napierala * our own estimate in this case. 34136af9869SEdward Tomasz Napierala */ 34236af9869SEdward Tomasz Napierala if (swtime < RACCT_PCPU_SECS) 34336af9869SEdward Tomasz Napierala return (pcpu); 34436af9869SEdward Tomasz Napierala 34536af9869SEdward Tomasz Napierala p_pctcpu = 0; 34636af9869SEdward Tomasz Napierala FOREACH_THREAD_IN_PROC(p, td) { 34736af9869SEdward Tomasz Napierala if (td == PCPU_GET(idlethread)) 34836af9869SEdward Tomasz Napierala continue; 34936af9869SEdward Tomasz Napierala #ifdef SMP 35036af9869SEdward Tomasz Napierala found = 0; 35136af9869SEdward Tomasz Napierala STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) { 35236af9869SEdward Tomasz Napierala if (td == pc->pc_idlethread) { 35336af9869SEdward Tomasz Napierala found = 1; 35436af9869SEdward Tomasz Napierala break; 35536af9869SEdward Tomasz Napierala } 35636af9869SEdward Tomasz Napierala } 35736af9869SEdward Tomasz Napierala if (found) 35836af9869SEdward Tomasz Napierala continue; 35936af9869SEdward Tomasz Napierala #endif 36036af9869SEdward Tomasz Napierala thread_lock(td); 36136af9869SEdward Tomasz Napierala #ifdef SCHED_4BSD 36236af9869SEdward Tomasz Napierala pctcpu = sched_pctcpu(td); 36336af9869SEdward Tomasz Napierala /* Count also the yet unfinished second. */ 36436af9869SEdward Tomasz Napierala pctcpu_next = (pctcpu * ccpu_exp[1]) >> FSHIFT; 36536af9869SEdward Tomasz Napierala pctcpu_next += sched_pctcpu_delta(td); 36636af9869SEdward Tomasz Napierala p_pctcpu += max(pctcpu, pctcpu_next); 36736af9869SEdward Tomasz Napierala #else 36836af9869SEdward Tomasz Napierala /* 36936af9869SEdward Tomasz Napierala * In ULE the %cpu statistics are updated on every 37036af9869SEdward Tomasz Napierala * sched_pctcpu() call. So special calculations to 37136af9869SEdward Tomasz Napierala * account for the latest (unfinished) second are 37236af9869SEdward Tomasz Napierala * not needed. 37336af9869SEdward Tomasz Napierala */ 37436af9869SEdward Tomasz Napierala p_pctcpu += sched_pctcpu(td); 37536af9869SEdward Tomasz Napierala #endif 37636af9869SEdward Tomasz Napierala thread_unlock(td); 37736af9869SEdward Tomasz Napierala } 37836af9869SEdward Tomasz Napierala 37936af9869SEdward Tomasz Napierala #ifdef SCHED_4BSD 38036af9869SEdward Tomasz Napierala if (swtime <= CCPU_EXP_MAX) 38136af9869SEdward Tomasz Napierala return ((100 * (uint64_t)p_pctcpu * 1000000) / 38236af9869SEdward Tomasz Napierala (FSCALE - ccpu_exp[swtime])); 38336af9869SEdward Tomasz Napierala #endif 38436af9869SEdward Tomasz Napierala 38536af9869SEdward Tomasz Napierala return ((100 * (uint64_t)p_pctcpu * 1000000) / FSCALE); 38636af9869SEdward Tomasz Napierala } 387097055e2SEdward Tomasz Napierala 388097055e2SEdward Tomasz Napierala static void 389097055e2SEdward Tomasz Napierala racct_add_racct(struct racct *dest, const struct racct *src) 390097055e2SEdward Tomasz Napierala { 391097055e2SEdward Tomasz Napierala int i; 392097055e2SEdward Tomasz Napierala 3934b5c9cf6SEdward Tomasz Napierala ASSERT_RACCT_ENABLED(); 3944c230cdaSEdward Tomasz Napierala RACCT_LOCK_ASSERT(); 395097055e2SEdward Tomasz Napierala 396097055e2SEdward Tomasz Napierala /* 397097055e2SEdward Tomasz Napierala * Update resource usage in dest. 398097055e2SEdward Tomasz Napierala */ 399097055e2SEdward Tomasz Napierala for (i = 0; i <= RACCT_MAX; i++) { 400097055e2SEdward Tomasz Napierala KASSERT(dest->r_resources[i] >= 0, 401baf85d0aSEdward Tomasz Napierala ("%s: resource %d propagation meltdown: dest < 0", 402baf85d0aSEdward Tomasz Napierala __func__, i)); 403097055e2SEdward Tomasz Napierala KASSERT(src->r_resources[i] >= 0, 404baf85d0aSEdward Tomasz Napierala ("%s: resource %d propagation meltdown: src < 0", 405baf85d0aSEdward Tomasz Napierala __func__, i)); 406097055e2SEdward Tomasz Napierala dest->r_resources[i] += src->r_resources[i]; 407097055e2SEdward Tomasz Napierala } 408097055e2SEdward Tomasz Napierala } 409097055e2SEdward Tomasz Napierala 410097055e2SEdward Tomasz Napierala static void 411097055e2SEdward Tomasz Napierala racct_sub_racct(struct racct *dest, const struct racct *src) 412097055e2SEdward Tomasz Napierala { 413097055e2SEdward Tomasz Napierala int i; 414097055e2SEdward Tomasz Napierala 4154b5c9cf6SEdward Tomasz Napierala ASSERT_RACCT_ENABLED(); 4164c230cdaSEdward Tomasz Napierala RACCT_LOCK_ASSERT(); 417097055e2SEdward Tomasz Napierala 418097055e2SEdward Tomasz Napierala /* 419097055e2SEdward Tomasz Napierala * Update resource usage in dest. 420097055e2SEdward Tomasz Napierala */ 421097055e2SEdward Tomasz Napierala for (i = 0; i <= RACCT_MAX; i++) { 42284c9193bSEdward Tomasz Napierala if (!RACCT_IS_SLOPPY(i) && !RACCT_IS_DECAYING(i)) { 423097055e2SEdward Tomasz Napierala KASSERT(dest->r_resources[i] >= 0, 424baf85d0aSEdward Tomasz Napierala ("%s: resource %d propagation meltdown: dest < 0", 425baf85d0aSEdward Tomasz Napierala __func__, i)); 426097055e2SEdward Tomasz Napierala KASSERT(src->r_resources[i] >= 0, 427baf85d0aSEdward Tomasz Napierala ("%s: resource %d propagation meltdown: src < 0", 428baf85d0aSEdward Tomasz Napierala __func__, i)); 429097055e2SEdward Tomasz Napierala KASSERT(src->r_resources[i] <= dest->r_resources[i], 430baf85d0aSEdward Tomasz Napierala ("%s: resource %d propagation meltdown: src > dest", 431baf85d0aSEdward Tomasz Napierala __func__, i)); 432097055e2SEdward Tomasz Napierala } 43336af9869SEdward Tomasz Napierala if (RACCT_CAN_DROP(i)) { 434097055e2SEdward Tomasz Napierala dest->r_resources[i] -= src->r_resources[i]; 4355c939660SEdward Tomasz Napierala if (dest->r_resources[i] < 0) 436097055e2SEdward Tomasz Napierala dest->r_resources[i] = 0; 437097055e2SEdward Tomasz Napierala } 438097055e2SEdward Tomasz Napierala } 439097055e2SEdward Tomasz Napierala } 440097055e2SEdward Tomasz Napierala 441097055e2SEdward Tomasz Napierala void 442097055e2SEdward Tomasz Napierala racct_create(struct racct **racctp) 443097055e2SEdward Tomasz Napierala { 444097055e2SEdward Tomasz Napierala 4454b5c9cf6SEdward Tomasz Napierala if (!racct_enable) 4464b5c9cf6SEdward Tomasz Napierala return; 4474b5c9cf6SEdward Tomasz Napierala 44836160958SMark Johnston SDT_PROBE1(racct, , racct, create, racctp); 449097055e2SEdward Tomasz Napierala 450097055e2SEdward Tomasz Napierala KASSERT(*racctp == NULL, ("racct already allocated")); 451097055e2SEdward Tomasz Napierala 452097055e2SEdward Tomasz Napierala *racctp = uma_zalloc(racct_zone, M_WAITOK | M_ZERO); 453097055e2SEdward Tomasz Napierala } 454097055e2SEdward Tomasz Napierala 455097055e2SEdward Tomasz Napierala static void 456097055e2SEdward Tomasz Napierala racct_destroy_locked(struct racct **racctp) 457097055e2SEdward Tomasz Napierala { 458097055e2SEdward Tomasz Napierala struct racct *racct; 459c1a43e73SEdward Tomasz Napierala int i; 460097055e2SEdward Tomasz Napierala 4614b5c9cf6SEdward Tomasz Napierala ASSERT_RACCT_ENABLED(); 4624b5c9cf6SEdward Tomasz Napierala 46336160958SMark Johnston SDT_PROBE1(racct, , racct, destroy, racctp); 464097055e2SEdward Tomasz Napierala 4654c230cdaSEdward Tomasz Napierala RACCT_LOCK_ASSERT(); 466097055e2SEdward Tomasz Napierala KASSERT(racctp != NULL, ("NULL racctp")); 467097055e2SEdward Tomasz Napierala KASSERT(*racctp != NULL, ("NULL racct")); 468097055e2SEdward Tomasz Napierala 469097055e2SEdward Tomasz Napierala racct = *racctp; 470097055e2SEdward Tomasz Napierala 471097055e2SEdward Tomasz Napierala for (i = 0; i <= RACCT_MAX; i++) { 4724fe84775SEdward Tomasz Napierala if (RACCT_IS_SLOPPY(i)) 473097055e2SEdward Tomasz Napierala continue; 4744fe84775SEdward Tomasz Napierala if (!RACCT_IS_RECLAIMABLE(i)) 475097055e2SEdward Tomasz Napierala continue; 476097055e2SEdward Tomasz Napierala KASSERT(racct->r_resources[i] == 0, 477097055e2SEdward Tomasz Napierala ("destroying non-empty racct: " 478097055e2SEdward Tomasz Napierala "%ju allocated for resource %d\n", 479097055e2SEdward Tomasz Napierala racct->r_resources[i], i)); 480097055e2SEdward Tomasz Napierala } 481097055e2SEdward Tomasz Napierala uma_zfree(racct_zone, racct); 482097055e2SEdward Tomasz Napierala *racctp = NULL; 483097055e2SEdward Tomasz Napierala } 484097055e2SEdward Tomasz Napierala 485097055e2SEdward Tomasz Napierala void 486097055e2SEdward Tomasz Napierala racct_destroy(struct racct **racct) 487097055e2SEdward Tomasz Napierala { 488097055e2SEdward Tomasz Napierala 4894b5c9cf6SEdward Tomasz Napierala if (!racct_enable) 4904b5c9cf6SEdward Tomasz Napierala return; 4914b5c9cf6SEdward Tomasz Napierala 4924c230cdaSEdward Tomasz Napierala RACCT_LOCK(); 493097055e2SEdward Tomasz Napierala racct_destroy_locked(racct); 4944c230cdaSEdward Tomasz Napierala RACCT_UNLOCK(); 495097055e2SEdward Tomasz Napierala } 496097055e2SEdward Tomasz Napierala 497097055e2SEdward Tomasz Napierala /* 498af1a7b25SEdward Tomasz Napierala * Increase consumption of 'resource' by 'amount' for 'racct', 499af1a7b25SEdward Tomasz Napierala * but not its parents. Differently from other cases, 'amount' here 500097055e2SEdward Tomasz Napierala * may be less than zero. 501097055e2SEdward Tomasz Napierala */ 502097055e2SEdward Tomasz Napierala static void 5033768a5dfSJeremie Le Hen racct_adjust_resource(struct racct *racct, int resource, 504a8e0740eSEdward Tomasz Napierala int64_t amount) 505097055e2SEdward Tomasz Napierala { 506097055e2SEdward Tomasz Napierala 5074b5c9cf6SEdward Tomasz Napierala ASSERT_RACCT_ENABLED(); 5084c230cdaSEdward Tomasz Napierala RACCT_LOCK_ASSERT(); 509097055e2SEdward Tomasz Napierala KASSERT(racct != NULL, ("NULL racct")); 510097055e2SEdward Tomasz Napierala 511097055e2SEdward Tomasz Napierala racct->r_resources[resource] += amount; 512097055e2SEdward Tomasz Napierala if (racct->r_resources[resource] < 0) { 51336af9869SEdward Tomasz Napierala KASSERT(RACCT_IS_SLOPPY(resource) || RACCT_IS_DECAYING(resource), 514baf85d0aSEdward Tomasz Napierala ("%s: resource %d usage < 0", __func__, resource)); 515097055e2SEdward Tomasz Napierala racct->r_resources[resource] = 0; 516097055e2SEdward Tomasz Napierala } 51736af9869SEdward Tomasz Napierala 51836af9869SEdward Tomasz Napierala /* 51936af9869SEdward Tomasz Napierala * There are some cases where the racct %cpu resource would grow 52009766dd5SJosh Paetzel * beyond 100% per core. For example in racct_proc_exit() we add 52109766dd5SJosh Paetzel * the process %cpu usage to the ucred racct containers. If too 52209766dd5SJosh Paetzel * many processes terminated in a short time span, the ucred %cpu 52309766dd5SJosh Paetzel * resource could grow too much. Also, the 4BSD scheduler sometimes 52409766dd5SJosh Paetzel * returns for a thread more than 100% cpu usage. So we set a sane 525a9bee9c7SGordon Bergling * boundary here to 100% * the maximum number of CPUs. 52636af9869SEdward Tomasz Napierala */ 52736af9869SEdward Tomasz Napierala if ((resource == RACCT_PCTCPU) && 52809766dd5SJosh Paetzel (racct->r_resources[RACCT_PCTCPU] > 100 * 1000000 * (int64_t)MAXCPU)) 52909766dd5SJosh Paetzel racct->r_resources[RACCT_PCTCPU] = 100 * 1000000 * (int64_t)MAXCPU; 530097055e2SEdward Tomasz Napierala } 531097055e2SEdward Tomasz Napierala 5322d8696d1SEdward Tomasz Napierala static int 5337cea96f6SEdward Tomasz Napierala racct_add_locked(struct proc *p, int resource, uint64_t amount, int force) 534097055e2SEdward Tomasz Napierala { 535097055e2SEdward Tomasz Napierala #ifdef RCTL 536097055e2SEdward Tomasz Napierala int error; 537097055e2SEdward Tomasz Napierala #endif 538097055e2SEdward Tomasz Napierala 5394b5c9cf6SEdward Tomasz Napierala ASSERT_RACCT_ENABLED(); 5404b5c9cf6SEdward Tomasz Napierala 541097055e2SEdward Tomasz Napierala /* 542097055e2SEdward Tomasz Napierala * We need proc lock to dereference p->p_ucred. 543097055e2SEdward Tomasz Napierala */ 544097055e2SEdward Tomasz Napierala PROC_LOCK_ASSERT(p, MA_OWNED); 545097055e2SEdward Tomasz Napierala 546097055e2SEdward Tomasz Napierala #ifdef RCTL 547097055e2SEdward Tomasz Napierala error = rctl_enforce(p, resource, amount); 548659e7466SEdward Tomasz Napierala if (error && !force && RACCT_IS_DENIABLE(resource)) { 54936160958SMark Johnston SDT_PROBE3(racct, , rusage, add__failure, p, resource, amount); 550097055e2SEdward Tomasz Napierala return (error); 551097055e2SEdward Tomasz Napierala } 552097055e2SEdward Tomasz Napierala #endif 5533768a5dfSJeremie Le Hen racct_adjust_resource(p->p_racct, resource, amount); 554097055e2SEdward Tomasz Napierala racct_add_cred_locked(p->p_ucred, resource, amount); 555097055e2SEdward Tomasz Napierala 556097055e2SEdward Tomasz Napierala return (0); 557097055e2SEdward Tomasz Napierala } 558097055e2SEdward Tomasz Napierala 5592d8696d1SEdward Tomasz Napierala /* 5602d8696d1SEdward Tomasz Napierala * Increase allocation of 'resource' by 'amount' for process 'p'. 5612d8696d1SEdward Tomasz Napierala * Return 0 if it's below limits, or errno, if it's not. 5622d8696d1SEdward Tomasz Napierala */ 5632d8696d1SEdward Tomasz Napierala int 5642d8696d1SEdward Tomasz Napierala racct_add(struct proc *p, int resource, uint64_t amount) 5652d8696d1SEdward Tomasz Napierala { 5662d8696d1SEdward Tomasz Napierala int error; 5672d8696d1SEdward Tomasz Napierala 5684b5c9cf6SEdward Tomasz Napierala if (!racct_enable) 5694b5c9cf6SEdward Tomasz Napierala return (0); 5704b5c9cf6SEdward Tomasz Napierala 5717cea96f6SEdward Tomasz Napierala SDT_PROBE3(racct, , rusage, add, p, resource, amount); 5727cea96f6SEdward Tomasz Napierala 5734c230cdaSEdward Tomasz Napierala RACCT_LOCK(); 5747cea96f6SEdward Tomasz Napierala error = racct_add_locked(p, resource, amount, 0); 5754c230cdaSEdward Tomasz Napierala RACCT_UNLOCK(); 5762d8696d1SEdward Tomasz Napierala return (error); 5772d8696d1SEdward Tomasz Napierala } 5782d8696d1SEdward Tomasz Napierala 5790b9f1ecbSEdward Tomasz Napierala /* 5800b9f1ecbSEdward Tomasz Napierala * Increase allocation of 'resource' by 'amount' for process 'p'. 5810b9f1ecbSEdward Tomasz Napierala * Doesn't check for limits and never fails. 5820b9f1ecbSEdward Tomasz Napierala */ 5830b9f1ecbSEdward Tomasz Napierala void 5840b9f1ecbSEdward Tomasz Napierala racct_add_force(struct proc *p, int resource, uint64_t amount) 5850b9f1ecbSEdward Tomasz Napierala { 5860b9f1ecbSEdward Tomasz Napierala 5870b9f1ecbSEdward Tomasz Napierala if (!racct_enable) 5880b9f1ecbSEdward Tomasz Napierala return; 5890b9f1ecbSEdward Tomasz Napierala 5900b9f1ecbSEdward Tomasz Napierala SDT_PROBE3(racct, , rusage, add__force, p, resource, amount); 5910b9f1ecbSEdward Tomasz Napierala 5924c230cdaSEdward Tomasz Napierala RACCT_LOCK(); 5930b9f1ecbSEdward Tomasz Napierala racct_add_locked(p, resource, amount, 1); 5944c230cdaSEdward Tomasz Napierala RACCT_UNLOCK(); 5950b9f1ecbSEdward Tomasz Napierala } 5960b9f1ecbSEdward Tomasz Napierala 597097055e2SEdward Tomasz Napierala static void 598097055e2SEdward Tomasz Napierala racct_add_cred_locked(struct ucred *cred, int resource, uint64_t amount) 599097055e2SEdward Tomasz Napierala { 600097055e2SEdward Tomasz Napierala struct prison *pr; 601097055e2SEdward Tomasz Napierala 6024b5c9cf6SEdward Tomasz Napierala ASSERT_RACCT_ENABLED(); 6034b5c9cf6SEdward Tomasz Napierala 6043768a5dfSJeremie Le Hen racct_adjust_resource(cred->cr_ruidinfo->ui_racct, resource, amount); 605097055e2SEdward Tomasz Napierala for (pr = cred->cr_prison; pr != NULL; pr = pr->pr_parent) 6063768a5dfSJeremie Le Hen racct_adjust_resource(pr->pr_prison_racct->prr_racct, resource, 607a7ad07bfSEdward Tomasz Napierala amount); 6083768a5dfSJeremie Le Hen racct_adjust_resource(cred->cr_loginclass->lc_racct, resource, amount); 609097055e2SEdward Tomasz Napierala } 610097055e2SEdward Tomasz Napierala 611097055e2SEdward Tomasz Napierala /* 612097055e2SEdward Tomasz Napierala * Increase allocation of 'resource' by 'amount' for credential 'cred'. 613097055e2SEdward Tomasz Napierala * Doesn't check for limits and never fails. 614097055e2SEdward Tomasz Napierala */ 615097055e2SEdward Tomasz Napierala void 616097055e2SEdward Tomasz Napierala racct_add_cred(struct ucred *cred, int resource, uint64_t amount) 617097055e2SEdward Tomasz Napierala { 618097055e2SEdward Tomasz Napierala 6194b5c9cf6SEdward Tomasz Napierala if (!racct_enable) 6204b5c9cf6SEdward Tomasz Napierala return; 6214b5c9cf6SEdward Tomasz Napierala 622bbe4eb6dSEdward Tomasz Napierala SDT_PROBE3(racct, , rusage, add__cred, cred, resource, amount); 623bbe4eb6dSEdward Tomasz Napierala 6244c230cdaSEdward Tomasz Napierala RACCT_LOCK(); 625097055e2SEdward Tomasz Napierala racct_add_cred_locked(cred, resource, amount); 6264c230cdaSEdward Tomasz Napierala RACCT_UNLOCK(); 627097055e2SEdward Tomasz Napierala } 628097055e2SEdward Tomasz Napierala 629ae34b6ffSEdward Tomasz Napierala /* 630ae34b6ffSEdward Tomasz Napierala * Account for disk IO resource consumption. Checks for limits, 631ae34b6ffSEdward Tomasz Napierala * but never fails, due to disk limits being undeniable. 632ae34b6ffSEdward Tomasz Napierala */ 633ae34b6ffSEdward Tomasz Napierala void 634ae34b6ffSEdward Tomasz Napierala racct_add_buf(struct proc *p, const struct buf *bp, int is_write) 635ae34b6ffSEdward Tomasz Napierala { 636ae34b6ffSEdward Tomasz Napierala 637ae34b6ffSEdward Tomasz Napierala ASSERT_RACCT_ENABLED(); 638ae34b6ffSEdward Tomasz Napierala PROC_LOCK_ASSERT(p, MA_OWNED); 639ae34b6ffSEdward Tomasz Napierala 640bbe4eb6dSEdward Tomasz Napierala SDT_PROBE3(racct, , rusage, add__buf, p, bp, is_write); 641bbe4eb6dSEdward Tomasz Napierala 642ae34b6ffSEdward Tomasz Napierala RACCT_LOCK(); 643ae34b6ffSEdward Tomasz Napierala if (is_write) { 644ae34b6ffSEdward Tomasz Napierala racct_add_locked(curproc, RACCT_WRITEBPS, bp->b_bcount, 1); 645ae34b6ffSEdward Tomasz Napierala racct_add_locked(curproc, RACCT_WRITEIOPS, 1, 1); 646ae34b6ffSEdward Tomasz Napierala } else { 647ae34b6ffSEdward Tomasz Napierala racct_add_locked(curproc, RACCT_READBPS, bp->b_bcount, 1); 648ae34b6ffSEdward Tomasz Napierala racct_add_locked(curproc, RACCT_READIOPS, 1, 1); 649ae34b6ffSEdward Tomasz Napierala } 650ae34b6ffSEdward Tomasz Napierala RACCT_UNLOCK(); 651ae34b6ffSEdward Tomasz Napierala } 652ae34b6ffSEdward Tomasz Napierala 653097055e2SEdward Tomasz Napierala static int 65410287198SEdward Tomasz Napierala racct_set_locked(struct proc *p, int resource, uint64_t amount, int force) 655097055e2SEdward Tomasz Napierala { 656c1a43e73SEdward Tomasz Napierala int64_t old_amount, decayed_amount, diff_proc, diff_cred; 657097055e2SEdward Tomasz Napierala #ifdef RCTL 658097055e2SEdward Tomasz Napierala int error; 659097055e2SEdward Tomasz Napierala #endif 660097055e2SEdward Tomasz Napierala 6614b5c9cf6SEdward Tomasz Napierala ASSERT_RACCT_ENABLED(); 6624b5c9cf6SEdward Tomasz Napierala 663097055e2SEdward Tomasz Napierala /* 664097055e2SEdward Tomasz Napierala * We need proc lock to dereference p->p_ucred. 665097055e2SEdward Tomasz Napierala */ 666097055e2SEdward Tomasz Napierala PROC_LOCK_ASSERT(p, MA_OWNED); 667097055e2SEdward Tomasz Napierala 66836af9869SEdward Tomasz Napierala old_amount = p->p_racct->r_resources[resource]; 66936af9869SEdward Tomasz Napierala /* 67036af9869SEdward Tomasz Napierala * The diffs may be negative. 67136af9869SEdward Tomasz Napierala */ 67236af9869SEdward Tomasz Napierala diff_proc = amount - old_amount; 673ae34b6ffSEdward Tomasz Napierala if (resource == RACCT_PCTCPU) { 67436af9869SEdward Tomasz Napierala /* 67536af9869SEdward Tomasz Napierala * Resources in per-credential racct containers may decay. 67636af9869SEdward Tomasz Napierala * If this is the case, we need to calculate the difference 67736af9869SEdward Tomasz Napierala * between the new amount and the proportional value of the 67836af9869SEdward Tomasz Napierala * old amount that has decayed in the ucred racct containers. 67936af9869SEdward Tomasz Napierala */ 68036af9869SEdward Tomasz Napierala decayed_amount = old_amount * RACCT_DECAY_FACTOR / FSCALE; 68136af9869SEdward Tomasz Napierala diff_cred = amount - decayed_amount; 68236af9869SEdward Tomasz Napierala } else 68336af9869SEdward Tomasz Napierala diff_cred = diff_proc; 684097055e2SEdward Tomasz Napierala #ifdef notyet 68536af9869SEdward Tomasz Napierala KASSERT(diff_proc >= 0 || RACCT_CAN_DROP(resource), 686baf85d0aSEdward Tomasz Napierala ("%s: usage of non-droppable resource %d dropping", __func__, 687097055e2SEdward Tomasz Napierala resource)); 688097055e2SEdward Tomasz Napierala #endif 689097055e2SEdward Tomasz Napierala #ifdef RCTL 690659e7466SEdward Tomasz Napierala if (diff_proc > 0) { 69136af9869SEdward Tomasz Napierala error = rctl_enforce(p, resource, diff_proc); 692659e7466SEdward Tomasz Napierala if (error && !force && RACCT_IS_DENIABLE(resource)) { 69336160958SMark Johnston SDT_PROBE3(racct, , rusage, set__failure, p, resource, 69436160958SMark Johnston amount); 695097055e2SEdward Tomasz Napierala return (error); 696097055e2SEdward Tomasz Napierala } 697097055e2SEdward Tomasz Napierala } 698097055e2SEdward Tomasz Napierala #endif 6993768a5dfSJeremie Le Hen racct_adjust_resource(p->p_racct, resource, diff_proc); 70036af9869SEdward Tomasz Napierala if (diff_cred > 0) 70136af9869SEdward Tomasz Napierala racct_add_cred_locked(p->p_ucred, resource, diff_cred); 70236af9869SEdward Tomasz Napierala else if (diff_cred < 0) 70336af9869SEdward Tomasz Napierala racct_sub_cred_locked(p->p_ucred, resource, -diff_cred); 704097055e2SEdward Tomasz Napierala 705097055e2SEdward Tomasz Napierala return (0); 706097055e2SEdward Tomasz Napierala } 707097055e2SEdward Tomasz Napierala 708097055e2SEdward Tomasz Napierala /* 709097055e2SEdward Tomasz Napierala * Set allocation of 'resource' to 'amount' for process 'p'. 710097055e2SEdward Tomasz Napierala * Return 0 if it's below limits, or errno, if it's not. 711097055e2SEdward Tomasz Napierala * 712097055e2SEdward Tomasz Napierala * Note that decreasing the allocation always returns 0, 713097055e2SEdward Tomasz Napierala * even if it's above the limit. 714097055e2SEdward Tomasz Napierala */ 715097055e2SEdward Tomasz Napierala int 716448db4f7SMateusz Guzik racct_set_unlocked(struct proc *p, int resource, uint64_t amount) 717448db4f7SMateusz Guzik { 718448db4f7SMateusz Guzik int error; 719448db4f7SMateusz Guzik 720448db4f7SMateusz Guzik ASSERT_RACCT_ENABLED(); 721448db4f7SMateusz Guzik PROC_LOCK(p); 722448db4f7SMateusz Guzik error = racct_set(p, resource, amount); 723448db4f7SMateusz Guzik PROC_UNLOCK(p); 724448db4f7SMateusz Guzik return (error); 725448db4f7SMateusz Guzik } 726448db4f7SMateusz Guzik 727448db4f7SMateusz Guzik int 728097055e2SEdward Tomasz Napierala racct_set(struct proc *p, int resource, uint64_t amount) 729097055e2SEdward Tomasz Napierala { 730097055e2SEdward Tomasz Napierala int error; 731097055e2SEdward Tomasz Napierala 7324b5c9cf6SEdward Tomasz Napierala if (!racct_enable) 7334b5c9cf6SEdward Tomasz Napierala return (0); 7344b5c9cf6SEdward Tomasz Napierala 73510287198SEdward Tomasz Napierala SDT_PROBE3(racct, , rusage, set__force, p, resource, amount); 73610287198SEdward Tomasz Napierala 7374c230cdaSEdward Tomasz Napierala RACCT_LOCK(); 73810287198SEdward Tomasz Napierala error = racct_set_locked(p, resource, amount, 0); 7394c230cdaSEdward Tomasz Napierala RACCT_UNLOCK(); 740097055e2SEdward Tomasz Napierala return (error); 741097055e2SEdward Tomasz Napierala } 742097055e2SEdward Tomasz Napierala 7430b9f1ecbSEdward Tomasz Napierala void 7440b9f1ecbSEdward Tomasz Napierala racct_set_force(struct proc *p, int resource, uint64_t amount) 7450b9f1ecbSEdward Tomasz Napierala { 7460b9f1ecbSEdward Tomasz Napierala 7470b9f1ecbSEdward Tomasz Napierala if (!racct_enable) 7480b9f1ecbSEdward Tomasz Napierala return; 7490b9f1ecbSEdward Tomasz Napierala 7500b9f1ecbSEdward Tomasz Napierala SDT_PROBE3(racct, , rusage, set, p, resource, amount); 7510b9f1ecbSEdward Tomasz Napierala 7524c230cdaSEdward Tomasz Napierala RACCT_LOCK(); 7530b9f1ecbSEdward Tomasz Napierala racct_set_locked(p, resource, amount, 1); 7544c230cdaSEdward Tomasz Napierala RACCT_UNLOCK(); 7550b9f1ecbSEdward Tomasz Napierala } 7560b9f1ecbSEdward Tomasz Napierala 757097055e2SEdward Tomasz Napierala /* 758097055e2SEdward Tomasz Napierala * Returns amount of 'resource' the process 'p' can keep allocated. 759097055e2SEdward Tomasz Napierala * Allocating more than that would be denied, unless the resource 760097055e2SEdward Tomasz Napierala * is marked undeniable. Amount of already allocated resource does 761097055e2SEdward Tomasz Napierala * not matter. 762097055e2SEdward Tomasz Napierala */ 763097055e2SEdward Tomasz Napierala uint64_t 764097055e2SEdward Tomasz Napierala racct_get_limit(struct proc *p, int resource) 765097055e2SEdward Tomasz Napierala { 766bbe4eb6dSEdward Tomasz Napierala #ifdef RCTL 767bbe4eb6dSEdward Tomasz Napierala uint64_t available; 768097055e2SEdward Tomasz Napierala 7694b5c9cf6SEdward Tomasz Napierala if (!racct_enable) 7704b5c9cf6SEdward Tomasz Napierala return (UINT64_MAX); 7714b5c9cf6SEdward Tomasz Napierala 772bbe4eb6dSEdward Tomasz Napierala RACCT_LOCK(); 773bbe4eb6dSEdward Tomasz Napierala available = rctl_get_limit(p, resource); 774bbe4eb6dSEdward Tomasz Napierala RACCT_UNLOCK(); 775bbe4eb6dSEdward Tomasz Napierala 776bbe4eb6dSEdward Tomasz Napierala return (available); 777097055e2SEdward Tomasz Napierala #else 778bbe4eb6dSEdward Tomasz Napierala 779097055e2SEdward Tomasz Napierala return (UINT64_MAX); 780097055e2SEdward Tomasz Napierala #endif 781097055e2SEdward Tomasz Napierala } 782097055e2SEdward Tomasz Napierala 783097055e2SEdward Tomasz Napierala /* 784097055e2SEdward Tomasz Napierala * Returns amount of 'resource' the process 'p' can keep allocated. 785097055e2SEdward Tomasz Napierala * Allocating more than that would be denied, unless the resource 786097055e2SEdward Tomasz Napierala * is marked undeniable. Amount of already allocated resource does 787097055e2SEdward Tomasz Napierala * matter. 788097055e2SEdward Tomasz Napierala */ 789097055e2SEdward Tomasz Napierala uint64_t 790097055e2SEdward Tomasz Napierala racct_get_available(struct proc *p, int resource) 791097055e2SEdward Tomasz Napierala { 792bbe4eb6dSEdward Tomasz Napierala #ifdef RCTL 793bbe4eb6dSEdward Tomasz Napierala uint64_t available; 794097055e2SEdward Tomasz Napierala 7954b5c9cf6SEdward Tomasz Napierala if (!racct_enable) 7964b5c9cf6SEdward Tomasz Napierala return (UINT64_MAX); 7974b5c9cf6SEdward Tomasz Napierala 798bbe4eb6dSEdward Tomasz Napierala RACCT_LOCK(); 799bbe4eb6dSEdward Tomasz Napierala available = rctl_get_available(p, resource); 800bbe4eb6dSEdward Tomasz Napierala RACCT_UNLOCK(); 801bbe4eb6dSEdward Tomasz Napierala 802bbe4eb6dSEdward Tomasz Napierala return (available); 803097055e2SEdward Tomasz Napierala #else 804bbe4eb6dSEdward Tomasz Napierala 805097055e2SEdward Tomasz Napierala return (UINT64_MAX); 806097055e2SEdward Tomasz Napierala #endif 807097055e2SEdward Tomasz Napierala } 808097055e2SEdward Tomasz Napierala 809097055e2SEdward Tomasz Napierala /* 81036af9869SEdward Tomasz Napierala * Returns amount of the %cpu resource that process 'p' can add to its %cpu 81136af9869SEdward Tomasz Napierala * utilization. Adding more than that would lead to the process being 81236af9869SEdward Tomasz Napierala * throttled. 81336af9869SEdward Tomasz Napierala */ 81436af9869SEdward Tomasz Napierala static int64_t 81536af9869SEdward Tomasz Napierala racct_pcpu_available(struct proc *p) 81636af9869SEdward Tomasz Napierala { 817bbe4eb6dSEdward Tomasz Napierala #ifdef RCTL 818bbe4eb6dSEdward Tomasz Napierala uint64_t available; 81936af9869SEdward Tomasz Napierala 8204b5c9cf6SEdward Tomasz Napierala ASSERT_RACCT_ENABLED(); 8214b5c9cf6SEdward Tomasz Napierala 822bbe4eb6dSEdward Tomasz Napierala RACCT_LOCK(); 823bbe4eb6dSEdward Tomasz Napierala available = rctl_pcpu_available(p); 824bbe4eb6dSEdward Tomasz Napierala RACCT_UNLOCK(); 825bbe4eb6dSEdward Tomasz Napierala 826bbe4eb6dSEdward Tomasz Napierala return (available); 82736af9869SEdward Tomasz Napierala #else 828bbe4eb6dSEdward Tomasz Napierala 82936af9869SEdward Tomasz Napierala return (INT64_MAX); 83036af9869SEdward Tomasz Napierala #endif 83136af9869SEdward Tomasz Napierala } 83236af9869SEdward Tomasz Napierala 83336af9869SEdward Tomasz Napierala /* 834097055e2SEdward Tomasz Napierala * Decrease allocation of 'resource' by 'amount' for process 'p'. 835097055e2SEdward Tomasz Napierala */ 836097055e2SEdward Tomasz Napierala void 837097055e2SEdward Tomasz Napierala racct_sub(struct proc *p, int resource, uint64_t amount) 838097055e2SEdward Tomasz Napierala { 839097055e2SEdward Tomasz Napierala 8404b5c9cf6SEdward Tomasz Napierala if (!racct_enable) 8414b5c9cf6SEdward Tomasz Napierala return; 8424b5c9cf6SEdward Tomasz Napierala 84336160958SMark Johnston SDT_PROBE3(racct, , rusage, sub, p, resource, amount); 844097055e2SEdward Tomasz Napierala 845097055e2SEdward Tomasz Napierala /* 846097055e2SEdward Tomasz Napierala * We need proc lock to dereference p->p_ucred. 847097055e2SEdward Tomasz Napierala */ 848097055e2SEdward Tomasz Napierala PROC_LOCK_ASSERT(p, MA_OWNED); 84936af9869SEdward Tomasz Napierala KASSERT(RACCT_CAN_DROP(resource), 850baf85d0aSEdward Tomasz Napierala ("%s: called for non-droppable resource %d", __func__, resource)); 851097055e2SEdward Tomasz Napierala 8524c230cdaSEdward Tomasz Napierala RACCT_LOCK(); 853097055e2SEdward Tomasz Napierala KASSERT(amount <= p->p_racct->r_resources[resource], 854baf85d0aSEdward Tomasz Napierala ("%s: freeing %ju of resource %d, which is more " 855baf85d0aSEdward Tomasz Napierala "than allocated %jd for %s (pid %d)", __func__, amount, resource, 856097055e2SEdward Tomasz Napierala (intmax_t)p->p_racct->r_resources[resource], p->p_comm, p->p_pid)); 857097055e2SEdward Tomasz Napierala 8583768a5dfSJeremie Le Hen racct_adjust_resource(p->p_racct, resource, -amount); 859097055e2SEdward Tomasz Napierala racct_sub_cred_locked(p->p_ucred, resource, amount); 8604c230cdaSEdward Tomasz Napierala RACCT_UNLOCK(); 861097055e2SEdward Tomasz Napierala } 862097055e2SEdward Tomasz Napierala 863097055e2SEdward Tomasz Napierala static void 864097055e2SEdward Tomasz Napierala racct_sub_cred_locked(struct ucred *cred, int resource, uint64_t amount) 865097055e2SEdward Tomasz Napierala { 866097055e2SEdward Tomasz Napierala struct prison *pr; 867097055e2SEdward Tomasz Napierala 8684b5c9cf6SEdward Tomasz Napierala ASSERT_RACCT_ENABLED(); 8694b5c9cf6SEdward Tomasz Napierala 8703768a5dfSJeremie Le Hen racct_adjust_resource(cred->cr_ruidinfo->ui_racct, resource, -amount); 871097055e2SEdward Tomasz Napierala for (pr = cred->cr_prison; pr != NULL; pr = pr->pr_parent) 8723768a5dfSJeremie Le Hen racct_adjust_resource(pr->pr_prison_racct->prr_racct, resource, 873a7ad07bfSEdward Tomasz Napierala -amount); 8743768a5dfSJeremie Le Hen racct_adjust_resource(cred->cr_loginclass->lc_racct, resource, -amount); 875097055e2SEdward Tomasz Napierala } 876097055e2SEdward Tomasz Napierala 877097055e2SEdward Tomasz Napierala /* 878097055e2SEdward Tomasz Napierala * Decrease allocation of 'resource' by 'amount' for credential 'cred'. 879097055e2SEdward Tomasz Napierala */ 880097055e2SEdward Tomasz Napierala void 881097055e2SEdward Tomasz Napierala racct_sub_cred(struct ucred *cred, int resource, uint64_t amount) 882097055e2SEdward Tomasz Napierala { 883097055e2SEdward Tomasz Napierala 8844b5c9cf6SEdward Tomasz Napierala if (!racct_enable) 8854b5c9cf6SEdward Tomasz Napierala return; 8864b5c9cf6SEdward Tomasz Napierala 887bbe4eb6dSEdward Tomasz Napierala SDT_PROBE3(racct, , rusage, sub__cred, cred, resource, amount); 888bbe4eb6dSEdward Tomasz Napierala 889bbe4eb6dSEdward Tomasz Napierala #ifdef notyet 890bbe4eb6dSEdward Tomasz Napierala KASSERT(RACCT_CAN_DROP(resource), 891bbe4eb6dSEdward Tomasz Napierala ("%s: called for resource %d which can not drop", __func__, 892bbe4eb6dSEdward Tomasz Napierala resource)); 893bbe4eb6dSEdward Tomasz Napierala #endif 894bbe4eb6dSEdward Tomasz Napierala 8954c230cdaSEdward Tomasz Napierala RACCT_LOCK(); 896097055e2SEdward Tomasz Napierala racct_sub_cred_locked(cred, resource, amount); 8974c230cdaSEdward Tomasz Napierala RACCT_UNLOCK(); 898097055e2SEdward Tomasz Napierala } 899097055e2SEdward Tomasz Napierala 900097055e2SEdward Tomasz Napierala /* 901097055e2SEdward Tomasz Napierala * Inherit resource usage information from the parent process. 902097055e2SEdward Tomasz Napierala */ 903097055e2SEdward Tomasz Napierala int 904097055e2SEdward Tomasz Napierala racct_proc_fork(struct proc *parent, struct proc *child) 905097055e2SEdward Tomasz Napierala { 906097055e2SEdward Tomasz Napierala int i, error = 0; 907097055e2SEdward Tomasz Napierala 9084b5c9cf6SEdward Tomasz Napierala if (!racct_enable) 9094b5c9cf6SEdward Tomasz Napierala return (0); 9104b5c9cf6SEdward Tomasz Napierala 911097055e2SEdward Tomasz Napierala /* 912097055e2SEdward Tomasz Napierala * Create racct for the child process. 913097055e2SEdward Tomasz Napierala */ 914097055e2SEdward Tomasz Napierala racct_create(&child->p_racct); 915097055e2SEdward Tomasz Napierala 916097055e2SEdward Tomasz Napierala PROC_LOCK(parent); 917097055e2SEdward Tomasz Napierala PROC_LOCK(child); 9184c230cdaSEdward Tomasz Napierala RACCT_LOCK(); 919097055e2SEdward Tomasz Napierala 920c0c09362SEdward Tomasz Napierala #ifdef RCTL 921c0c09362SEdward Tomasz Napierala error = rctl_proc_fork(parent, child); 922c0c09362SEdward Tomasz Napierala if (error != 0) 923c0c09362SEdward Tomasz Napierala goto out; 924c0c09362SEdward Tomasz Napierala #endif 925c0c09362SEdward Tomasz Napierala 92636af9869SEdward Tomasz Napierala /* Init process cpu time. */ 92736af9869SEdward Tomasz Napierala child->p_prev_runtime = 0; 92836af9869SEdward Tomasz Napierala child->p_throttled = 0; 92936af9869SEdward Tomasz Napierala 930097055e2SEdward Tomasz Napierala /* 931097055e2SEdward Tomasz Napierala * Inherit resource usage. 932097055e2SEdward Tomasz Napierala */ 933097055e2SEdward Tomasz Napierala for (i = 0; i <= RACCT_MAX; i++) { 934097055e2SEdward Tomasz Napierala if (parent->p_racct->r_resources[i] == 0 || 9354fe84775SEdward Tomasz Napierala !RACCT_IS_INHERITABLE(i)) 936097055e2SEdward Tomasz Napierala continue; 937097055e2SEdward Tomasz Napierala 938097055e2SEdward Tomasz Napierala error = racct_set_locked(child, i, 93910287198SEdward Tomasz Napierala parent->p_racct->r_resources[i], 0); 940ac6fafe6SEdward Tomasz Napierala if (error != 0) 941097055e2SEdward Tomasz Napierala goto out; 942097055e2SEdward Tomasz Napierala } 943097055e2SEdward Tomasz Napierala 9447cea96f6SEdward Tomasz Napierala error = racct_add_locked(child, RACCT_NPROC, 1, 0); 9457cea96f6SEdward Tomasz Napierala error += racct_add_locked(child, RACCT_NTHR, 1, 0); 9462d8696d1SEdward Tomasz Napierala 947097055e2SEdward Tomasz Napierala out: 9484c230cdaSEdward Tomasz Napierala RACCT_UNLOCK(); 949097055e2SEdward Tomasz Napierala PROC_UNLOCK(child); 950097055e2SEdward Tomasz Napierala PROC_UNLOCK(parent); 951097055e2SEdward Tomasz Napierala 952ab27d5d8SEdward Tomasz Napierala if (error != 0) 953ab27d5d8SEdward Tomasz Napierala racct_proc_exit(child); 954ab27d5d8SEdward Tomasz Napierala 955097055e2SEdward Tomasz Napierala return (error); 956097055e2SEdward Tomasz Napierala } 957097055e2SEdward Tomasz Napierala 95872a401d9SEdward Tomasz Napierala /* 95972a401d9SEdward Tomasz Napierala * Called at the end of fork1(), to handle rules that require the process 96072a401d9SEdward Tomasz Napierala * to be fully initialized. 96172a401d9SEdward Tomasz Napierala */ 96272a401d9SEdward Tomasz Napierala void 96372a401d9SEdward Tomasz Napierala racct_proc_fork_done(struct proc *child) 96472a401d9SEdward Tomasz Napierala { 96572a401d9SEdward Tomasz Napierala 9664b5c9cf6SEdward Tomasz Napierala if (!racct_enable) 9674b5c9cf6SEdward Tomasz Napierala return; 9684b5c9cf6SEdward Tomasz Napierala 969bbe4eb6dSEdward Tomasz Napierala #ifdef RCTL 970a5ac8272SMateusz Guzik PROC_LOCK(child); 9714c230cdaSEdward Tomasz Napierala RACCT_LOCK(); 97272a401d9SEdward Tomasz Napierala rctl_enforce(child, RACCT_NPROC, 0); 97372a401d9SEdward Tomasz Napierala rctl_enforce(child, RACCT_NTHR, 0); 9744c230cdaSEdward Tomasz Napierala RACCT_UNLOCK(); 975a5ac8272SMateusz Guzik PROC_UNLOCK(child); 97672a401d9SEdward Tomasz Napierala #endif 97772a401d9SEdward Tomasz Napierala } 97872a401d9SEdward Tomasz Napierala 979097055e2SEdward Tomasz Napierala void 980097055e2SEdward Tomasz Napierala racct_proc_exit(struct proc *p) 981097055e2SEdward Tomasz Napierala { 98236af9869SEdward Tomasz Napierala struct timeval wallclock; 983c1a43e73SEdward Tomasz Napierala uint64_t pct_estimate, pct, runtime; 984c1a43e73SEdward Tomasz Napierala int i; 985097055e2SEdward Tomasz Napierala 9864b5c9cf6SEdward Tomasz Napierala if (!racct_enable) 9874b5c9cf6SEdward Tomasz Napierala return; 9884b5c9cf6SEdward Tomasz Napierala 989097055e2SEdward Tomasz Napierala PROC_LOCK(p); 990097055e2SEdward Tomasz Napierala /* 991097055e2SEdward Tomasz Napierala * We don't need to calculate rux, proc_reap() has already done this. 992097055e2SEdward Tomasz Napierala */ 993097055e2SEdward Tomasz Napierala runtime = cputick2usec(p->p_rux.rux_runtime); 994097055e2SEdward Tomasz Napierala #ifdef notyet 995097055e2SEdward Tomasz Napierala KASSERT(runtime >= p->p_prev_runtime, ("runtime < p_prev_runtime")); 996097055e2SEdward Tomasz Napierala #else 997097055e2SEdward Tomasz Napierala if (runtime < p->p_prev_runtime) 998097055e2SEdward Tomasz Napierala runtime = p->p_prev_runtime; 999097055e2SEdward Tomasz Napierala #endif 100036af9869SEdward Tomasz Napierala microuptime(&wallclock); 100136af9869SEdward Tomasz Napierala timevalsub(&wallclock, &p->p_stats->p_start); 100284590fd8SEdward Tomasz Napierala if (wallclock.tv_sec > 0 || wallclock.tv_usec > 0) { 100336af9869SEdward Tomasz Napierala pct_estimate = (1000000 * runtime * 100) / 100436af9869SEdward Tomasz Napierala ((uint64_t)wallclock.tv_sec * 1000000 + 100536af9869SEdward Tomasz Napierala wallclock.tv_usec); 100684590fd8SEdward Tomasz Napierala } else 100784590fd8SEdward Tomasz Napierala pct_estimate = 0; 100836af9869SEdward Tomasz Napierala pct = racct_getpcpu(p, pct_estimate); 100936af9869SEdward Tomasz Napierala 10104c230cdaSEdward Tomasz Napierala RACCT_LOCK(); 101110287198SEdward Tomasz Napierala racct_set_locked(p, RACCT_CPU, runtime, 0); 101236af9869SEdward Tomasz Napierala racct_add_cred_locked(p->p_ucred, RACCT_PCTCPU, pct); 1013097055e2SEdward Tomasz Napierala 1014937c1b07SAndriy Gapon KASSERT(p->p_racct->r_resources[RACCT_RSS] == 0, 1015937c1b07SAndriy Gapon ("process reaped with %ju allocated for RSS\n", 1016937c1b07SAndriy Gapon p->p_racct->r_resources[RACCT_RSS])); 10172419d7f9SEdward Tomasz Napierala for (i = 0; i <= RACCT_MAX; i++) { 10182419d7f9SEdward Tomasz Napierala if (p->p_racct->r_resources[i] == 0) 10192419d7f9SEdward Tomasz Napierala continue; 10202419d7f9SEdward Tomasz Napierala if (!RACCT_IS_RECLAIMABLE(i)) 10212419d7f9SEdward Tomasz Napierala continue; 102210287198SEdward Tomasz Napierala racct_set_locked(p, i, 0, 0); 10232419d7f9SEdward Tomasz Napierala } 10242419d7f9SEdward Tomasz Napierala 1025097055e2SEdward Tomasz Napierala #ifdef RCTL 1026097055e2SEdward Tomasz Napierala rctl_racct_release(p->p_racct); 1027097055e2SEdward Tomasz Napierala #endif 1028bbe4eb6dSEdward Tomasz Napierala racct_destroy_locked(&p->p_racct); 1029bbe4eb6dSEdward Tomasz Napierala RACCT_UNLOCK(); 1030bbe4eb6dSEdward Tomasz Napierala PROC_UNLOCK(p); 1031097055e2SEdward Tomasz Napierala } 1032097055e2SEdward Tomasz Napierala 1033097055e2SEdward Tomasz Napierala /* 1034097055e2SEdward Tomasz Napierala * Called after credentials change, to move resource utilisation 1035097055e2SEdward Tomasz Napierala * between raccts. 1036097055e2SEdward Tomasz Napierala */ 1037097055e2SEdward Tomasz Napierala void 1038097055e2SEdward Tomasz Napierala racct_proc_ucred_changed(struct proc *p, struct ucred *oldcred, 1039097055e2SEdward Tomasz Napierala struct ucred *newcred) 1040097055e2SEdward Tomasz Napierala { 1041097055e2SEdward Tomasz Napierala struct uidinfo *olduip, *newuip; 1042097055e2SEdward Tomasz Napierala struct loginclass *oldlc, *newlc; 1043097055e2SEdward Tomasz Napierala struct prison *oldpr, *newpr, *pr; 1044097055e2SEdward Tomasz Napierala 10454b5c9cf6SEdward Tomasz Napierala if (!racct_enable) 10464b5c9cf6SEdward Tomasz Napierala return; 10474b5c9cf6SEdward Tomasz Napierala 1048f87beb93SAndriy Gapon PROC_LOCK_ASSERT(p, MA_OWNED); 1049097055e2SEdward Tomasz Napierala 1050097055e2SEdward Tomasz Napierala newuip = newcred->cr_ruidinfo; 1051097055e2SEdward Tomasz Napierala olduip = oldcred->cr_ruidinfo; 1052097055e2SEdward Tomasz Napierala newlc = newcred->cr_loginclass; 1053097055e2SEdward Tomasz Napierala oldlc = oldcred->cr_loginclass; 1054097055e2SEdward Tomasz Napierala newpr = newcred->cr_prison; 1055097055e2SEdward Tomasz Napierala oldpr = oldcred->cr_prison; 1056097055e2SEdward Tomasz Napierala 10574c230cdaSEdward Tomasz Napierala RACCT_LOCK(); 1058097055e2SEdward Tomasz Napierala if (newuip != olduip) { 1059097055e2SEdward Tomasz Napierala racct_sub_racct(olduip->ui_racct, p->p_racct); 1060097055e2SEdward Tomasz Napierala racct_add_racct(newuip->ui_racct, p->p_racct); 1061097055e2SEdward Tomasz Napierala } 1062097055e2SEdward Tomasz Napierala if (newlc != oldlc) { 1063097055e2SEdward Tomasz Napierala racct_sub_racct(oldlc->lc_racct, p->p_racct); 1064097055e2SEdward Tomasz Napierala racct_add_racct(newlc->lc_racct, p->p_racct); 1065097055e2SEdward Tomasz Napierala } 1066097055e2SEdward Tomasz Napierala if (newpr != oldpr) { 1067097055e2SEdward Tomasz Napierala for (pr = oldpr; pr != NULL; pr = pr->pr_parent) 1068a7ad07bfSEdward Tomasz Napierala racct_sub_racct(pr->pr_prison_racct->prr_racct, 1069a7ad07bfSEdward Tomasz Napierala p->p_racct); 1070097055e2SEdward Tomasz Napierala for (pr = newpr; pr != NULL; pr = pr->pr_parent) 1071a7ad07bfSEdward Tomasz Napierala racct_add_racct(pr->pr_prison_racct->prr_racct, 1072a7ad07bfSEdward Tomasz Napierala p->p_racct); 1073097055e2SEdward Tomasz Napierala } 10744c230cdaSEdward Tomasz Napierala RACCT_UNLOCK(); 1075097055e2SEdward Tomasz Napierala } 1076097055e2SEdward Tomasz Napierala 1077c34bbd2aSEdward Tomasz Napierala void 1078c34bbd2aSEdward Tomasz Napierala racct_move(struct racct *dest, struct racct *src) 1079c34bbd2aSEdward Tomasz Napierala { 1080c34bbd2aSEdward Tomasz Napierala 10814b5c9cf6SEdward Tomasz Napierala ASSERT_RACCT_ENABLED(); 10824b5c9cf6SEdward Tomasz Napierala 10834c230cdaSEdward Tomasz Napierala RACCT_LOCK(); 1084c34bbd2aSEdward Tomasz Napierala racct_add_racct(dest, src); 1085c34bbd2aSEdward Tomasz Napierala racct_sub_racct(src, src); 10864c230cdaSEdward Tomasz Napierala RACCT_UNLOCK(); 1087c34bbd2aSEdward Tomasz Napierala } 1088c34bbd2aSEdward Tomasz Napierala 1089c6d31b83SKonstantin Belousov static void 1090c6d31b83SKonstantin Belousov ast_racct(struct thread *td, int tda __unused) 109164cf6a62SMateusz Guzik { 1092c6d31b83SKonstantin Belousov struct proc *p; 109364cf6a62SMateusz Guzik 109464cf6a62SMateusz Guzik ASSERT_RACCT_ENABLED(); 109564cf6a62SMateusz Guzik 1096c6d31b83SKonstantin Belousov p = td->td_proc; 1097c6d31b83SKonstantin Belousov if (p->p_throttled == 0) 1098c6d31b83SKonstantin Belousov return; 1099c6d31b83SKonstantin Belousov 110064cf6a62SMateusz Guzik PROC_LOCK(p); 110164cf6a62SMateusz Guzik while (p->p_throttled != 0) { 110264cf6a62SMateusz Guzik msleep(p->p_racct, &p->p_mtx, 0, "racct", 110364cf6a62SMateusz Guzik p->p_throttled < 0 ? 0 : p->p_throttled); 110464cf6a62SMateusz Guzik if (p->p_throttled > 0) 110564cf6a62SMateusz Guzik p->p_throttled = 0; 110664cf6a62SMateusz Guzik } 110764cf6a62SMateusz Guzik PROC_UNLOCK(p); 110864cf6a62SMateusz Guzik } 110964cf6a62SMateusz Guzik 1110ae34b6ffSEdward Tomasz Napierala /* 1111ae34b6ffSEdward Tomasz Napierala * Make the process sleep in userret() for 'timeout' ticks. Setting 1112ae34b6ffSEdward Tomasz Napierala * timeout to -1 makes it sleep until woken up by racct_proc_wakeup(). 1113ae34b6ffSEdward Tomasz Napierala */ 1114ae34b6ffSEdward Tomasz Napierala void 1115ae34b6ffSEdward Tomasz Napierala racct_proc_throttle(struct proc *p, int timeout) 111636af9869SEdward Tomasz Napierala { 111736af9869SEdward Tomasz Napierala struct thread *td; 111836af9869SEdward Tomasz Napierala #ifdef SMP 111936af9869SEdward Tomasz Napierala int cpuid; 112036af9869SEdward Tomasz Napierala #endif 112136af9869SEdward Tomasz Napierala 1122ae34b6ffSEdward Tomasz Napierala KASSERT(timeout != 0, ("timeout %d", timeout)); 11234b5c9cf6SEdward Tomasz Napierala ASSERT_RACCT_ENABLED(); 112436af9869SEdward Tomasz Napierala PROC_LOCK_ASSERT(p, MA_OWNED); 112536af9869SEdward Tomasz Napierala 112636af9869SEdward Tomasz Napierala /* 112736af9869SEdward Tomasz Napierala * Do not block kernel processes. Also do not block processes with 112836af9869SEdward Tomasz Napierala * low %cpu utilization to improve interactivity. 112936af9869SEdward Tomasz Napierala */ 1130ae34b6ffSEdward Tomasz Napierala if ((p->p_flag & (P_SYSTEM | P_KPROC)) != 0) 113136af9869SEdward Tomasz Napierala return; 1132ae34b6ffSEdward Tomasz Napierala 1133ae34b6ffSEdward Tomasz Napierala if (p->p_throttled < 0 || (timeout > 0 && p->p_throttled > timeout)) 1134ae34b6ffSEdward Tomasz Napierala return; 1135ae34b6ffSEdward Tomasz Napierala 1136ae34b6ffSEdward Tomasz Napierala p->p_throttled = timeout; 113736af9869SEdward Tomasz Napierala 113836af9869SEdward Tomasz Napierala FOREACH_THREAD_IN_PROC(p, td) { 113916befafdSEdward Tomasz Napierala thread_lock(td); 1140c6d31b83SKonstantin Belousov ast_sched_locked(td, TDA_RACCT); 114130d158eeSEdward Tomasz Napierala 1142fa2528acSAlex Richardson switch (TD_GET_STATE(td)) { 114336af9869SEdward Tomasz Napierala case TDS_RUNQ: 114436af9869SEdward Tomasz Napierala /* 114536af9869SEdward Tomasz Napierala * If the thread is on the scheduler run-queue, we can 114636af9869SEdward Tomasz Napierala * not just remove it from there. So we set the flag 1147c6d31b83SKonstantin Belousov * TDA_SCHED for the thread, so that once it is 114836af9869SEdward Tomasz Napierala * running, it is taken off the cpu as soon as possible. 114936af9869SEdward Tomasz Napierala */ 1150c6d31b83SKonstantin Belousov ast_sched_locked(td, TDA_SCHED); 115136af9869SEdward Tomasz Napierala break; 115236af9869SEdward Tomasz Napierala case TDS_RUNNING: 115336af9869SEdward Tomasz Napierala /* 115436af9869SEdward Tomasz Napierala * If the thread is running, we request a context 1155c6d31b83SKonstantin Belousov * switch for it by setting the TDA_SCHED flag. 115636af9869SEdward Tomasz Napierala */ 1157c6d31b83SKonstantin Belousov ast_sched_locked(td, TDA_SCHED); 115836af9869SEdward Tomasz Napierala #ifdef SMP 115936af9869SEdward Tomasz Napierala cpuid = td->td_oncpu; 116036af9869SEdward Tomasz Napierala if ((cpuid != NOCPU) && (td != curthread)) 116136af9869SEdward Tomasz Napierala ipi_cpu(cpuid, IPI_AST); 116236af9869SEdward Tomasz Napierala #endif 116336af9869SEdward Tomasz Napierala break; 116436af9869SEdward Tomasz Napierala default: 116536af9869SEdward Tomasz Napierala break; 116636af9869SEdward Tomasz Napierala } 116716befafdSEdward Tomasz Napierala thread_unlock(td); 116836af9869SEdward Tomasz Napierala } 116936af9869SEdward Tomasz Napierala } 117036af9869SEdward Tomasz Napierala 117136af9869SEdward Tomasz Napierala static void 117236af9869SEdward Tomasz Napierala racct_proc_wakeup(struct proc *p) 117336af9869SEdward Tomasz Napierala { 11744b5c9cf6SEdward Tomasz Napierala 11754b5c9cf6SEdward Tomasz Napierala ASSERT_RACCT_ENABLED(); 11764b5c9cf6SEdward Tomasz Napierala 117736af9869SEdward Tomasz Napierala PROC_LOCK_ASSERT(p, MA_OWNED); 117836af9869SEdward Tomasz Napierala 1179ae34b6ffSEdward Tomasz Napierala if (p->p_throttled != 0) { 118036af9869SEdward Tomasz Napierala p->p_throttled = 0; 118136af9869SEdward Tomasz Napierala wakeup(p->p_racct); 118236af9869SEdward Tomasz Napierala } 118336af9869SEdward Tomasz Napierala } 118436af9869SEdward Tomasz Napierala 118536af9869SEdward Tomasz Napierala static void 1186862d03fbSEdward Tomasz Napierala racct_decay_callback(struct racct *racct, void *dummy1, void *dummy2) 118736af9869SEdward Tomasz Napierala { 118836af9869SEdward Tomasz Napierala int64_t r_old, r_new; 118936af9869SEdward Tomasz Napierala 11904b5c9cf6SEdward Tomasz Napierala ASSERT_RACCT_ENABLED(); 11914c230cdaSEdward Tomasz Napierala RACCT_LOCK_ASSERT(); 11924b5c9cf6SEdward Tomasz Napierala 1193ae34b6ffSEdward Tomasz Napierala #ifdef RCTL 1194ae34b6ffSEdward Tomasz Napierala rctl_throttle_decay(racct, RACCT_READBPS); 1195ae34b6ffSEdward Tomasz Napierala rctl_throttle_decay(racct, RACCT_WRITEBPS); 1196ae34b6ffSEdward Tomasz Napierala rctl_throttle_decay(racct, RACCT_READIOPS); 1197ae34b6ffSEdward Tomasz Napierala rctl_throttle_decay(racct, RACCT_WRITEIOPS); 1198ae34b6ffSEdward Tomasz Napierala #endif 1199ae34b6ffSEdward Tomasz Napierala 1200862d03fbSEdward Tomasz Napierala r_old = racct->r_resources[RACCT_PCTCPU]; 120136af9869SEdward Tomasz Napierala 120236af9869SEdward Tomasz Napierala /* If there is nothing to decay, just exit. */ 120336af9869SEdward Tomasz Napierala if (r_old <= 0) 120436af9869SEdward Tomasz Napierala return; 120536af9869SEdward Tomasz Napierala 120636af9869SEdward Tomasz Napierala r_new = r_old * RACCT_DECAY_FACTOR / FSCALE; 1207862d03fbSEdward Tomasz Napierala racct->r_resources[RACCT_PCTCPU] = r_new; 120815db3c07SEdward Tomasz Napierala } 120915db3c07SEdward Tomasz Napierala 121015db3c07SEdward Tomasz Napierala static void 121115db3c07SEdward Tomasz Napierala racct_decay_pre(void) 121215db3c07SEdward Tomasz Napierala { 121315db3c07SEdward Tomasz Napierala 12144c230cdaSEdward Tomasz Napierala RACCT_LOCK(); 121515db3c07SEdward Tomasz Napierala } 121615db3c07SEdward Tomasz Napierala 121715db3c07SEdward Tomasz Napierala static void 121815db3c07SEdward Tomasz Napierala racct_decay_post(void) 121915db3c07SEdward Tomasz Napierala { 122015db3c07SEdward Tomasz Napierala 12214c230cdaSEdward Tomasz Napierala RACCT_UNLOCK(); 122236af9869SEdward Tomasz Napierala } 122336af9869SEdward Tomasz Napierala 122436af9869SEdward Tomasz Napierala static void 1225097e0da7SEdward Tomasz Napierala racct_decay(void) 122636af9869SEdward Tomasz Napierala { 12274b5c9cf6SEdward Tomasz Napierala 12284b5c9cf6SEdward Tomasz Napierala ASSERT_RACCT_ENABLED(); 12294b5c9cf6SEdward Tomasz Napierala 1230862d03fbSEdward Tomasz Napierala ui_racct_foreach(racct_decay_callback, racct_decay_pre, 1231862d03fbSEdward Tomasz Napierala racct_decay_post, NULL, NULL); 1232862d03fbSEdward Tomasz Napierala loginclass_racct_foreach(racct_decay_callback, racct_decay_pre, 1233862d03fbSEdward Tomasz Napierala racct_decay_post, NULL, NULL); 1234862d03fbSEdward Tomasz Napierala prison_racct_foreach(racct_decay_callback, racct_decay_pre, 1235862d03fbSEdward Tomasz Napierala racct_decay_post, NULL, NULL); 123636af9869SEdward Tomasz Napierala } 123736af9869SEdward Tomasz Napierala 123836af9869SEdward Tomasz Napierala static void 1239097055e2SEdward Tomasz Napierala racctd(void) 1240097055e2SEdward Tomasz Napierala { 1241097055e2SEdward Tomasz Napierala struct thread *td; 1242097055e2SEdward Tomasz Napierala struct proc *p; 1243097055e2SEdward Tomasz Napierala struct timeval wallclock; 1244c1a43e73SEdward Tomasz Napierala uint64_t pct, pct_estimate, runtime; 1245097055e2SEdward Tomasz Napierala 12464b5c9cf6SEdward Tomasz Napierala ASSERT_RACCT_ENABLED(); 12474b5c9cf6SEdward Tomasz Napierala 1248097055e2SEdward Tomasz Napierala for (;;) { 1249862d03fbSEdward Tomasz Napierala racct_decay(); 125036af9869SEdward Tomasz Napierala 1251097055e2SEdward Tomasz Napierala sx_slock(&allproc_lock); 1252097055e2SEdward Tomasz Napierala 1253097055e2SEdward Tomasz Napierala FOREACH_PROC_IN_SYSTEM(p) { 125436af9869SEdward Tomasz Napierala PROC_LOCK(p); 125536af9869SEdward Tomasz Napierala if (p->p_state != PRS_NORMAL) { 125688cc62e5SMateusz Guzik if (p->p_state == PRS_ZOMBIE) 125788cc62e5SMateusz Guzik racct_set(p, RACCT_PCTCPU, 0); 125836af9869SEdward Tomasz Napierala PROC_UNLOCK(p); 1259097055e2SEdward Tomasz Napierala continue; 126036af9869SEdward Tomasz Napierala } 1261097055e2SEdward Tomasz Napierala 1262097055e2SEdward Tomasz Napierala microuptime(&wallclock); 1263097055e2SEdward Tomasz Napierala timevalsub(&wallclock, &p->p_stats->p_start); 12645c7bebf9SKonstantin Belousov PROC_STATLOCK(p); 12650a53cd57SEdward Tomasz Napierala FOREACH_THREAD_IN_PROC(p, td) 1266097055e2SEdward Tomasz Napierala ruxagg(p, td); 1267097055e2SEdward Tomasz Napierala runtime = cputick2usec(p->p_rux.rux_runtime); 12685c7bebf9SKonstantin Belousov PROC_STATUNLOCK(p); 1269097055e2SEdward Tomasz Napierala #ifdef notyet 1270097055e2SEdward Tomasz Napierala KASSERT(runtime >= p->p_prev_runtime, 1271097055e2SEdward Tomasz Napierala ("runtime < p_prev_runtime")); 1272097055e2SEdward Tomasz Napierala #else 1273097055e2SEdward Tomasz Napierala if (runtime < p->p_prev_runtime) 1274097055e2SEdward Tomasz Napierala runtime = p->p_prev_runtime; 1275097055e2SEdward Tomasz Napierala #endif 1276097055e2SEdward Tomasz Napierala p->p_prev_runtime = runtime; 127784590fd8SEdward Tomasz Napierala if (wallclock.tv_sec > 0 || wallclock.tv_usec > 0) { 127836af9869SEdward Tomasz Napierala pct_estimate = (1000000 * runtime * 100) / 127936af9869SEdward Tomasz Napierala ((uint64_t)wallclock.tv_sec * 1000000 + 128036af9869SEdward Tomasz Napierala wallclock.tv_usec); 128184590fd8SEdward Tomasz Napierala } else 128284590fd8SEdward Tomasz Napierala pct_estimate = 0; 128336af9869SEdward Tomasz Napierala pct = racct_getpcpu(p, pct_estimate); 12844c230cdaSEdward Tomasz Napierala RACCT_LOCK(); 1285ae34b6ffSEdward Tomasz Napierala #ifdef RCTL 1286ae34b6ffSEdward Tomasz Napierala rctl_throttle_decay(p->p_racct, RACCT_READBPS); 1287ae34b6ffSEdward Tomasz Napierala rctl_throttle_decay(p->p_racct, RACCT_WRITEBPS); 1288ae34b6ffSEdward Tomasz Napierala rctl_throttle_decay(p->p_racct, RACCT_READIOPS); 1289ae34b6ffSEdward Tomasz Napierala rctl_throttle_decay(p->p_racct, RACCT_WRITEIOPS); 1290ae34b6ffSEdward Tomasz Napierala #endif 129110287198SEdward Tomasz Napierala racct_set_locked(p, RACCT_PCTCPU, pct, 1); 129210287198SEdward Tomasz Napierala racct_set_locked(p, RACCT_CPU, runtime, 0); 1293097055e2SEdward Tomasz Napierala racct_set_locked(p, RACCT_WALLCLOCK, 129459f513cdSJaakko Heinonen (uint64_t)wallclock.tv_sec * 1000000 + 129510287198SEdward Tomasz Napierala wallclock.tv_usec, 0); 12964c230cdaSEdward Tomasz Napierala RACCT_UNLOCK(); 1297097055e2SEdward Tomasz Napierala PROC_UNLOCK(p); 1298097055e2SEdward Tomasz Napierala } 129936af9869SEdward Tomasz Napierala 130036af9869SEdward Tomasz Napierala /* 130136af9869SEdward Tomasz Napierala * To ensure that processes are throttled in a fair way, we need 130236af9869SEdward Tomasz Napierala * to iterate over all processes again and check the limits 130336af9869SEdward Tomasz Napierala * for %cpu resource only after ucred racct containers have been 130436af9869SEdward Tomasz Napierala * properly filled. 130536af9869SEdward Tomasz Napierala */ 130636af9869SEdward Tomasz Napierala FOREACH_PROC_IN_SYSTEM(p) { 130736af9869SEdward Tomasz Napierala PROC_LOCK(p); 130836af9869SEdward Tomasz Napierala if (p->p_state != PRS_NORMAL) { 130936af9869SEdward Tomasz Napierala PROC_UNLOCK(p); 131036af9869SEdward Tomasz Napierala continue; 131136af9869SEdward Tomasz Napierala } 131236af9869SEdward Tomasz Napierala 1313ae34b6ffSEdward Tomasz Napierala if (racct_pcpu_available(p) <= 0) { 1314ae34b6ffSEdward Tomasz Napierala if (p->p_racct->r_resources[RACCT_PCTCPU] > 1315ae34b6ffSEdward Tomasz Napierala pcpu_threshold) 1316ae34b6ffSEdward Tomasz Napierala racct_proc_throttle(p, -1); 1317ae34b6ffSEdward Tomasz Napierala } else if (p->p_throttled == -1) { 131836af9869SEdward Tomasz Napierala racct_proc_wakeup(p); 1319ae34b6ffSEdward Tomasz Napierala } 132036af9869SEdward Tomasz Napierala PROC_UNLOCK(p); 132136af9869SEdward Tomasz Napierala } 1322097055e2SEdward Tomasz Napierala sx_sunlock(&allproc_lock); 1323097055e2SEdward Tomasz Napierala pause("-", hz); 1324097055e2SEdward Tomasz Napierala } 1325097055e2SEdward Tomasz Napierala } 1326097055e2SEdward Tomasz Napierala 1327097055e2SEdward Tomasz Napierala static struct kproc_desc racctd_kp = { 1328097055e2SEdward Tomasz Napierala "racctd", 1329097055e2SEdward Tomasz Napierala racctd, 1330097055e2SEdward Tomasz Napierala NULL 1331097055e2SEdward Tomasz Napierala }; 13324b5c9cf6SEdward Tomasz Napierala 13334b5c9cf6SEdward Tomasz Napierala static void 13344b5c9cf6SEdward Tomasz Napierala racctd_init(void) 13354b5c9cf6SEdward Tomasz Napierala { 13364b5c9cf6SEdward Tomasz Napierala if (!racct_enable) 13374b5c9cf6SEdward Tomasz Napierala return; 13384b5c9cf6SEdward Tomasz Napierala 13394b5c9cf6SEdward Tomasz Napierala kproc_start(&racctd_kp); 13404b5c9cf6SEdward Tomasz Napierala } 13414b5c9cf6SEdward Tomasz Napierala SYSINIT(racctd, SI_SUB_RACCTD, SI_ORDER_FIRST, racctd_init, NULL); 1342097055e2SEdward Tomasz Napierala 1343097055e2SEdward Tomasz Napierala static void 1344097055e2SEdward Tomasz Napierala racct_init(void) 1345097055e2SEdward Tomasz Napierala { 13464b5c9cf6SEdward Tomasz Napierala if (!racct_enable) 13474b5c9cf6SEdward Tomasz Napierala return; 1348097055e2SEdward Tomasz Napierala 1349097055e2SEdward Tomasz Napierala racct_zone = uma_zcreate("racct", sizeof(struct racct), 135023e6fff2SEdward Tomasz Napierala NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 1351c6d31b83SKonstantin Belousov ast_register(TDA_RACCT, ASTR_ASTF_REQUIRED, 0, ast_racct); 1352c6d31b83SKonstantin Belousov 1353097055e2SEdward Tomasz Napierala /* 1354097055e2SEdward Tomasz Napierala * XXX: Move this somewhere. 1355097055e2SEdward Tomasz Napierala */ 1356a7ad07bfSEdward Tomasz Napierala prison0.pr_prison_racct = prison_racct_find("0"); 1357097055e2SEdward Tomasz Napierala } 1358097055e2SEdward Tomasz Napierala SYSINIT(racct, SI_SUB_RACCT, SI_ORDER_FIRST, racct_init, NULL); 1359