Index: stable/11/sys/kern/kern_racct.c =================================================================== --- stable/11/sys/kern/kern_racct.c (revision 314346) +++ stable/11/sys/kern/kern_racct.c (revision 314347) @@ -1,1343 +1,1346 @@ /*- * Copyright (c) 2010 The FreeBSD Foundation * All rights reserved. * * This software was developed by Edward Tomasz Napierala under sponsorship * from the FreeBSD Foundation. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * $FreeBSD$ */ #include __FBSDID("$FreeBSD$"); #include "opt_sched.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef RCTL #include #endif #ifdef RACCT FEATURE(racct, "Resource Accounting"); /* * Do not block processes that have their %cpu usage <= pcpu_threshold. */ static int pcpu_threshold = 1; #ifdef RACCT_DEFAULT_TO_DISABLED int racct_enable = 0; #else int racct_enable = 1; #endif SYSCTL_NODE(_kern, OID_AUTO, racct, CTLFLAG_RW, 0, "Resource Accounting"); SYSCTL_UINT(_kern_racct, OID_AUTO, enable, CTLFLAG_RDTUN, &racct_enable, 0, "Enable RACCT/RCTL"); SYSCTL_UINT(_kern_racct, OID_AUTO, pcpu_threshold, CTLFLAG_RW, &pcpu_threshold, 0, "Processes with higher %cpu usage than this value can be throttled."); /* * How many seconds it takes to use the scheduler %cpu calculations. When a * process starts, we compute its %cpu usage by dividing its runtime by the * process wall clock time. After RACCT_PCPU_SECS pass, we use the value * provided by the scheduler. */ #define RACCT_PCPU_SECS 3 struct mtx racct_lock; MTX_SYSINIT(racct_lock, &racct_lock, "racct lock", MTX_DEF); static uma_zone_t racct_zone; static void racct_sub_racct(struct racct *dest, const struct racct *src); static void racct_sub_cred_locked(struct ucred *cred, int resource, uint64_t amount); static void racct_add_cred_locked(struct ucred *cred, int resource, uint64_t amount); SDT_PROVIDER_DEFINE(racct); SDT_PROBE_DEFINE3(racct, , rusage, add, "struct proc *", "int", "uint64_t"); SDT_PROBE_DEFINE3(racct, , rusage, add__failure, "struct proc *", "int", "uint64_t"); SDT_PROBE_DEFINE3(racct, , rusage, add__buf, "struct proc *", "const struct buf *", "int"); SDT_PROBE_DEFINE3(racct, , rusage, add__cred, "struct ucred *", "int", "uint64_t"); SDT_PROBE_DEFINE3(racct, , rusage, add__force, "struct proc *", "int", "uint64_t"); SDT_PROBE_DEFINE3(racct, , rusage, set, "struct proc *", "int", "uint64_t"); SDT_PROBE_DEFINE3(racct, , rusage, set__failure, "struct proc *", "int", "uint64_t"); SDT_PROBE_DEFINE3(racct, , rusage, set__force, "struct proc *", "int", "uint64_t"); SDT_PROBE_DEFINE3(racct, , rusage, sub, "struct proc *", "int", "uint64_t"); SDT_PROBE_DEFINE3(racct, , rusage, sub__cred, "struct ucred *", "int", "uint64_t"); SDT_PROBE_DEFINE1(racct, , racct, create, "struct racct *"); SDT_PROBE_DEFINE1(racct, , racct, destroy, "struct racct *"); SDT_PROBE_DEFINE2(racct, , racct, join, "struct racct *", "struct racct *"); SDT_PROBE_DEFINE2(racct, , racct, join__failure, "struct racct *", "struct racct *"); SDT_PROBE_DEFINE2(racct, , racct, leave, "struct racct *", "struct racct *"); int racct_types[] = { [RACCT_CPU] = RACCT_IN_MILLIONS, [RACCT_DATA] = RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE, [RACCT_STACK] = RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE, [RACCT_CORE] = RACCT_DENIABLE, [RACCT_RSS] = RACCT_RECLAIMABLE, [RACCT_MEMLOCK] = RACCT_RECLAIMABLE | RACCT_DENIABLE, [RACCT_NPROC] = RACCT_RECLAIMABLE | RACCT_DENIABLE, [RACCT_NOFILE] = RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE, [RACCT_VMEM] = RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE, [RACCT_NPTS] = RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY, [RACCT_SWAP] = RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY, [RACCT_NTHR] = RACCT_RECLAIMABLE | RACCT_DENIABLE, [RACCT_MSGQQUEUED] = RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY, [RACCT_MSGQSIZE] = RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY, [RACCT_NMSGQ] = RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY, [RACCT_NSEM] = RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY, [RACCT_NSEMOP] = RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE, [RACCT_NSHM] = RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY, [RACCT_SHMSIZE] = RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY, [RACCT_WALLCLOCK] = RACCT_IN_MILLIONS, [RACCT_PCTCPU] = RACCT_DECAYING | RACCT_DENIABLE | RACCT_IN_MILLIONS, [RACCT_READBPS] = RACCT_DECAYING, [RACCT_WRITEBPS] = RACCT_DECAYING, [RACCT_READIOPS] = RACCT_DECAYING, [RACCT_WRITEIOPS] = RACCT_DECAYING }; static const fixpt_t RACCT_DECAY_FACTOR = 0.3 * FSCALE; #ifdef SCHED_4BSD /* * Contains intermediate values for %cpu calculations to avoid using floating * point in the kernel. * ccpu_exp[k] = FSCALE * (ccpu/FSCALE)^k = FSCALE * exp(-k/20) * It is needed only for the 4BSD scheduler, because in ULE, the ccpu equals to * zero so the calculations are more straightforward. */ fixpt_t ccpu_exp[] = { [0] = FSCALE * 1, [1] = FSCALE * 0.95122942450071400909, [2] = FSCALE * 0.90483741803595957316, [3] = FSCALE * 0.86070797642505780722, [4] = FSCALE * 0.81873075307798185866, [5] = FSCALE * 0.77880078307140486824, [6] = FSCALE * 0.74081822068171786606, [7] = FSCALE * 0.70468808971871343435, [8] = FSCALE * 0.67032004603563930074, [9] = FSCALE * 0.63762815162177329314, [10] = FSCALE * 0.60653065971263342360, [11] = FSCALE * 0.57694981038048669531, [12] = FSCALE * 0.54881163609402643262, [13] = FSCALE * 0.52204577676101604789, [14] = FSCALE * 0.49658530379140951470, [15] = FSCALE * 0.47236655274101470713, [16] = FSCALE * 0.44932896411722159143, [17] = FSCALE * 0.42741493194872666992, [18] = FSCALE * 0.40656965974059911188, [19] = FSCALE * 0.38674102345450120691, [20] = FSCALE * 0.36787944117144232159, [21] = FSCALE * 0.34993774911115535467, [22] = FSCALE * 0.33287108369807955328, [23] = FSCALE * 0.31663676937905321821, [24] = FSCALE * 0.30119421191220209664, [25] = FSCALE * 0.28650479686019010032, [26] = FSCALE * 0.27253179303401260312, [27] = FSCALE * 0.25924026064589150757, [28] = FSCALE * 0.24659696394160647693, [29] = FSCALE * 0.23457028809379765313, [30] = FSCALE * 0.22313016014842982893, [31] = FSCALE * 0.21224797382674305771, [32] = FSCALE * 0.20189651799465540848, [33] = FSCALE * 0.19204990862075411423, [34] = FSCALE * 0.18268352405273465022, [35] = FSCALE * 0.17377394345044512668, [36] = FSCALE * 0.16529888822158653829, [37] = FSCALE * 0.15723716631362761621, [38] = FSCALE * 0.14956861922263505264, [39] = FSCALE * 0.14227407158651357185, [40] = FSCALE * 0.13533528323661269189, [41] = FSCALE * 0.12873490358780421886, [42] = FSCALE * 0.12245642825298191021, [43] = FSCALE * 0.11648415777349695786, [44] = FSCALE * 0.11080315836233388333, [45] = FSCALE * 0.10539922456186433678, [46] = FSCALE * 0.10025884372280373372, [47] = FSCALE * 0.09536916221554961888, [48] = FSCALE * 0.09071795328941250337, [49] = FSCALE * 0.08629358649937051097, [50] = FSCALE * 0.08208499862389879516, [51] = FSCALE * 0.07808166600115315231, [52] = FSCALE * 0.07427357821433388042, [53] = FSCALE * 0.07065121306042958674, [54] = FSCALE * 0.06720551273974976512, [55] = FSCALE * 0.06392786120670757270, [56] = FSCALE * 0.06081006262521796499, [57] = FSCALE * 0.05784432087483846296, [58] = FSCALE * 0.05502322005640722902, [59] = FSCALE * 0.05233970594843239308, [60] = FSCALE * 0.04978706836786394297, [61] = FSCALE * 0.04735892439114092119, [62] = FSCALE * 0.04504920239355780606, [63] = FSCALE * 0.04285212686704017991, [64] = FSCALE * 0.04076220397836621516, [65] = FSCALE * 0.03877420783172200988, [66] = FSCALE * 0.03688316740124000544, [67] = FSCALE * 0.03508435410084502588, [68] = FSCALE * 0.03337326996032607948, [69] = FSCALE * 0.03174563637806794323, [70] = FSCALE * 0.03019738342231850073, [71] = FSCALE * 0.02872463965423942912, [72] = FSCALE * 0.02732372244729256080, [73] = FSCALE * 0.02599112877875534358, [74] = FSCALE * 0.02472352647033939120, [75] = FSCALE * 0.02351774585600910823, [76] = FSCALE * 0.02237077185616559577, [77] = FSCALE * 0.02127973643837716938, [78] = FSCALE * 0.02024191144580438847, [79] = FSCALE * 0.01925470177538692429, [80] = FSCALE * 0.01831563888873418029, [81] = FSCALE * 0.01742237463949351138, [82] = FSCALE * 0.01657267540176124754, [83] = FSCALE * 0.01576441648485449082, [84] = FSCALE * 0.01499557682047770621, [85] = FSCALE * 0.01426423390899925527, [86] = FSCALE * 0.01356855901220093175, [87] = FSCALE * 0.01290681258047986886, [88] = FSCALE * 0.01227733990306844117, [89] = FSCALE * 0.01167856697039544521, [90] = FSCALE * 0.01110899653824230649, [91] = FSCALE * 0.01056720438385265337, [92] = FSCALE * 0.01005183574463358164, [93] = FSCALE * 0.00956160193054350793, [94] = FSCALE * 0.00909527710169581709, [95] = FSCALE * 0.00865169520312063417, [96] = FSCALE * 0.00822974704902002884, [97] = FSCALE * 0.00782837754922577143, [98] = FSCALE * 0.00744658307092434051, [99] = FSCALE * 0.00708340892905212004, [100] = FSCALE * 0.00673794699908546709, [101] = FSCALE * 0.00640933344625638184, [102] = FSCALE * 0.00609674656551563610, [103] = FSCALE * 0.00579940472684214321, [104] = FSCALE * 0.00551656442076077241, [105] = FSCALE * 0.00524751839918138427, [106] = FSCALE * 0.00499159390691021621, [107] = FSCALE * 0.00474815099941147558, [108] = FSCALE * 0.00451658094261266798, [109] = FSCALE * 0.00429630469075234057, [110] = FSCALE * 0.00408677143846406699, }; #endif #define CCPU_EXP_MAX 110 /* * This function is analogical to the getpcpu() function in the ps(1) command. * They should both calculate in the same way so that the racct %cpu * calculations are consistent with the values showed by the ps(1) tool. * The calculations are more complex in the 4BSD scheduler because of the value * of the ccpu variable. In ULE it is defined to be zero which saves us some * work. */ static uint64_t racct_getpcpu(struct proc *p, u_int pcpu) { u_int swtime; #ifdef SCHED_4BSD fixpt_t pctcpu, pctcpu_next; #endif #ifdef SMP struct pcpu *pc; int found; #endif fixpt_t p_pctcpu; struct thread *td; ASSERT_RACCT_ENABLED(); /* * If the process is swapped out, we count its %cpu usage as zero. * This behaviour is consistent with the userland ps(1) tool. */ if ((p->p_flag & P_INMEM) == 0) return (0); swtime = (ticks - p->p_swtick) / hz; /* * For short-lived processes, the sched_pctcpu() returns small * values even for cpu intensive processes. Therefore we use * our own estimate in this case. */ if (swtime < RACCT_PCPU_SECS) return (pcpu); p_pctcpu = 0; FOREACH_THREAD_IN_PROC(p, td) { if (td == PCPU_GET(idlethread)) continue; #ifdef SMP found = 0; STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) { if (td == pc->pc_idlethread) { found = 1; break; } } if (found) continue; #endif thread_lock(td); #ifdef SCHED_4BSD pctcpu = sched_pctcpu(td); /* Count also the yet unfinished second. */ pctcpu_next = (pctcpu * ccpu_exp[1]) >> FSHIFT; pctcpu_next += sched_pctcpu_delta(td); p_pctcpu += max(pctcpu, pctcpu_next); #else /* * In ULE the %cpu statistics are updated on every * sched_pctcpu() call. So special calculations to * account for the latest (unfinished) second are * not needed. */ p_pctcpu += sched_pctcpu(td); #endif thread_unlock(td); } #ifdef SCHED_4BSD if (swtime <= CCPU_EXP_MAX) return ((100 * (uint64_t)p_pctcpu * 1000000) / (FSCALE - ccpu_exp[swtime])); #endif return ((100 * (uint64_t)p_pctcpu * 1000000) / FSCALE); } static void racct_add_racct(struct racct *dest, const struct racct *src) { int i; ASSERT_RACCT_ENABLED(); RACCT_LOCK_ASSERT(); /* * Update resource usage in dest. */ for (i = 0; i <= RACCT_MAX; i++) { KASSERT(dest->r_resources[i] >= 0, ("%s: resource %d propagation meltdown: dest < 0", __func__, i)); KASSERT(src->r_resources[i] >= 0, ("%s: resource %d propagation meltdown: src < 0", __func__, i)); dest->r_resources[i] += src->r_resources[i]; } } static void racct_sub_racct(struct racct *dest, const struct racct *src) { int i; ASSERT_RACCT_ENABLED(); RACCT_LOCK_ASSERT(); /* * Update resource usage in dest. */ for (i = 0; i <= RACCT_MAX; i++) { if (!RACCT_IS_SLOPPY(i) && !RACCT_IS_DECAYING(i)) { KASSERT(dest->r_resources[i] >= 0, ("%s: resource %d propagation meltdown: dest < 0", __func__, i)); KASSERT(src->r_resources[i] >= 0, ("%s: resource %d propagation meltdown: src < 0", __func__, i)); KASSERT(src->r_resources[i] <= dest->r_resources[i], ("%s: resource %d propagation meltdown: src > dest", __func__, i)); } if (RACCT_CAN_DROP(i)) { dest->r_resources[i] -= src->r_resources[i]; if (dest->r_resources[i] < 0) { KASSERT(RACCT_IS_SLOPPY(i) || RACCT_IS_DECAYING(i), ("%s: resource %d usage < 0", __func__, i)); dest->r_resources[i] = 0; } } } } void racct_create(struct racct **racctp) { if (!racct_enable) return; SDT_PROBE1(racct, , racct, create, racctp); KASSERT(*racctp == NULL, ("racct already allocated")); *racctp = uma_zalloc(racct_zone, M_WAITOK | M_ZERO); } static void racct_destroy_locked(struct racct **racctp) { struct racct *racct; int i; ASSERT_RACCT_ENABLED(); SDT_PROBE1(racct, , racct, destroy, racctp); RACCT_LOCK_ASSERT(); KASSERT(racctp != NULL, ("NULL racctp")); KASSERT(*racctp != NULL, ("NULL racct")); racct = *racctp; for (i = 0; i <= RACCT_MAX; i++) { if (RACCT_IS_SLOPPY(i)) continue; if (!RACCT_IS_RECLAIMABLE(i)) continue; KASSERT(racct->r_resources[i] == 0, ("destroying non-empty racct: " "%ju allocated for resource %d\n", racct->r_resources[i], i)); } uma_zfree(racct_zone, racct); *racctp = NULL; } void racct_destroy(struct racct **racct) { if (!racct_enable) return; RACCT_LOCK(); racct_destroy_locked(racct); RACCT_UNLOCK(); } /* * Increase consumption of 'resource' by 'amount' for 'racct', * but not its parents. Differently from other cases, 'amount' here * may be less than zero. */ static void racct_adjust_resource(struct racct *racct, int resource, int64_t amount) { ASSERT_RACCT_ENABLED(); RACCT_LOCK_ASSERT(); KASSERT(racct != NULL, ("NULL racct")); racct->r_resources[resource] += amount; if (racct->r_resources[resource] < 0) { KASSERT(RACCT_IS_SLOPPY(resource) || RACCT_IS_DECAYING(resource), ("%s: resource %d usage < 0", __func__, resource)); racct->r_resources[resource] = 0; } /* * There are some cases where the racct %cpu resource would grow * beyond 100% per core. For example in racct_proc_exit() we add * the process %cpu usage to the ucred racct containers. If too * many processes terminated in a short time span, the ucred %cpu * resource could grow too much. Also, the 4BSD scheduler sometimes * returns for a thread more than 100% cpu usage. So we set a sane * boundary here to 100% * the maxumum number of CPUs. */ if ((resource == RACCT_PCTCPU) && (racct->r_resources[RACCT_PCTCPU] > 100 * 1000000 * (int64_t)MAXCPU)) racct->r_resources[RACCT_PCTCPU] = 100 * 1000000 * (int64_t)MAXCPU; } static int racct_add_locked(struct proc *p, int resource, uint64_t amount, int force) { #ifdef RCTL int error; #endif ASSERT_RACCT_ENABLED(); /* * We need proc lock to dereference p->p_ucred. */ PROC_LOCK_ASSERT(p, MA_OWNED); #ifdef RCTL error = rctl_enforce(p, resource, amount); if (error && !force && RACCT_IS_DENIABLE(resource)) { SDT_PROBE3(racct, , rusage, add__failure, p, resource, amount); return (error); } #endif racct_adjust_resource(p->p_racct, resource, amount); racct_add_cred_locked(p->p_ucred, resource, amount); return (0); } /* * Increase allocation of 'resource' by 'amount' for process 'p'. * Return 0 if it's below limits, or errno, if it's not. */ int racct_add(struct proc *p, int resource, uint64_t amount) { int error; if (!racct_enable) return (0); SDT_PROBE3(racct, , rusage, add, p, resource, amount); RACCT_LOCK(); error = racct_add_locked(p, resource, amount, 0); RACCT_UNLOCK(); return (error); } /* * Increase allocation of 'resource' by 'amount' for process 'p'. * Doesn't check for limits and never fails. */ void racct_add_force(struct proc *p, int resource, uint64_t amount) { if (!racct_enable) return; SDT_PROBE3(racct, , rusage, add__force, p, resource, amount); RACCT_LOCK(); racct_add_locked(p, resource, amount, 1); RACCT_UNLOCK(); } static void racct_add_cred_locked(struct ucred *cred, int resource, uint64_t amount) { struct prison *pr; ASSERT_RACCT_ENABLED(); racct_adjust_resource(cred->cr_ruidinfo->ui_racct, resource, amount); for (pr = cred->cr_prison; pr != NULL; pr = pr->pr_parent) racct_adjust_resource(pr->pr_prison_racct->prr_racct, resource, amount); racct_adjust_resource(cred->cr_loginclass->lc_racct, resource, amount); } /* * Increase allocation of 'resource' by 'amount' for credential 'cred'. * Doesn't check for limits and never fails. */ void racct_add_cred(struct ucred *cred, int resource, uint64_t amount) { if (!racct_enable) return; SDT_PROBE3(racct, , rusage, add__cred, cred, resource, amount); RACCT_LOCK(); racct_add_cred_locked(cred, resource, amount); RACCT_UNLOCK(); } /* * Account for disk IO resource consumption. Checks for limits, * but never fails, due to disk limits being undeniable. */ void racct_add_buf(struct proc *p, const struct buf *bp, int is_write) { ASSERT_RACCT_ENABLED(); PROC_LOCK_ASSERT(p, MA_OWNED); SDT_PROBE3(racct, , rusage, add__buf, p, bp, is_write); RACCT_LOCK(); if (is_write) { racct_add_locked(curproc, RACCT_WRITEBPS, bp->b_bcount, 1); racct_add_locked(curproc, RACCT_WRITEIOPS, 1, 1); } else { racct_add_locked(curproc, RACCT_READBPS, bp->b_bcount, 1); racct_add_locked(curproc, RACCT_READIOPS, 1, 1); } RACCT_UNLOCK(); } static int racct_set_locked(struct proc *p, int resource, uint64_t amount, int force) { int64_t old_amount, decayed_amount, diff_proc, diff_cred; #ifdef RCTL int error; #endif ASSERT_RACCT_ENABLED(); /* * We need proc lock to dereference p->p_ucred. */ PROC_LOCK_ASSERT(p, MA_OWNED); old_amount = p->p_racct->r_resources[resource]; /* * The diffs may be negative. */ diff_proc = amount - old_amount; if (resource == RACCT_PCTCPU) { /* * Resources in per-credential racct containers may decay. * If this is the case, we need to calculate the difference * between the new amount and the proportional value of the * old amount that has decayed in the ucred racct containers. */ decayed_amount = old_amount * RACCT_DECAY_FACTOR / FSCALE; diff_cred = amount - decayed_amount; } else diff_cred = diff_proc; #ifdef notyet KASSERT(diff_proc >= 0 || RACCT_CAN_DROP(resource), ("%s: usage of non-droppable resource %d dropping", __func__, resource)); #endif #ifdef RCTL if (diff_proc > 0) { error = rctl_enforce(p, resource, diff_proc); if (error && !force && RACCT_IS_DENIABLE(resource)) { SDT_PROBE3(racct, , rusage, set__failure, p, resource, amount); return (error); } } #endif racct_adjust_resource(p->p_racct, resource, diff_proc); if (diff_cred > 0) racct_add_cred_locked(p->p_ucred, resource, diff_cred); else if (diff_cred < 0) racct_sub_cred_locked(p->p_ucred, resource, -diff_cred); return (0); } /* * Set allocation of 'resource' to 'amount' for process 'p'. * Return 0 if it's below limits, or errno, if it's not. * * Note that decreasing the allocation always returns 0, * even if it's above the limit. */ int racct_set(struct proc *p, int resource, uint64_t amount) { int error; if (!racct_enable) return (0); SDT_PROBE3(racct, , rusage, set__force, p, resource, amount); RACCT_LOCK(); error = racct_set_locked(p, resource, amount, 0); RACCT_UNLOCK(); return (error); } void racct_set_force(struct proc *p, int resource, uint64_t amount) { if (!racct_enable) return; SDT_PROBE3(racct, , rusage, set, p, resource, amount); RACCT_LOCK(); racct_set_locked(p, resource, amount, 1); RACCT_UNLOCK(); } /* * Returns amount of 'resource' the process 'p' can keep allocated. * Allocating more than that would be denied, unless the resource * is marked undeniable. Amount of already allocated resource does * not matter. */ uint64_t racct_get_limit(struct proc *p, int resource) { #ifdef RCTL uint64_t available; if (!racct_enable) return (UINT64_MAX); RACCT_LOCK(); available = rctl_get_limit(p, resource); RACCT_UNLOCK(); return (available); #else return (UINT64_MAX); #endif } /* * Returns amount of 'resource' the process 'p' can keep allocated. * Allocating more than that would be denied, unless the resource * is marked undeniable. Amount of already allocated resource does * matter. */ uint64_t racct_get_available(struct proc *p, int resource) { #ifdef RCTL uint64_t available; if (!racct_enable) return (UINT64_MAX); RACCT_LOCK(); available = rctl_get_available(p, resource); RACCT_UNLOCK(); return (available); #else return (UINT64_MAX); #endif } /* * Returns amount of the %cpu resource that process 'p' can add to its %cpu * utilization. Adding more than that would lead to the process being * throttled. */ static int64_t racct_pcpu_available(struct proc *p) { #ifdef RCTL uint64_t available; ASSERT_RACCT_ENABLED(); RACCT_LOCK(); available = rctl_pcpu_available(p); RACCT_UNLOCK(); return (available); #else return (INT64_MAX); #endif } /* * Decrease allocation of 'resource' by 'amount' for process 'p'. */ void racct_sub(struct proc *p, int resource, uint64_t amount) { if (!racct_enable) return; SDT_PROBE3(racct, , rusage, sub, p, resource, amount); /* * We need proc lock to dereference p->p_ucred. */ PROC_LOCK_ASSERT(p, MA_OWNED); KASSERT(RACCT_CAN_DROP(resource), ("%s: called for non-droppable resource %d", __func__, resource)); RACCT_LOCK(); KASSERT(amount <= p->p_racct->r_resources[resource], ("%s: freeing %ju of resource %d, which is more " "than allocated %jd for %s (pid %d)", __func__, amount, resource, (intmax_t)p->p_racct->r_resources[resource], p->p_comm, p->p_pid)); racct_adjust_resource(p->p_racct, resource, -amount); racct_sub_cred_locked(p->p_ucred, resource, amount); RACCT_UNLOCK(); } static void racct_sub_cred_locked(struct ucred *cred, int resource, uint64_t amount) { struct prison *pr; ASSERT_RACCT_ENABLED(); racct_adjust_resource(cred->cr_ruidinfo->ui_racct, resource, -amount); for (pr = cred->cr_prison; pr != NULL; pr = pr->pr_parent) racct_adjust_resource(pr->pr_prison_racct->prr_racct, resource, -amount); racct_adjust_resource(cred->cr_loginclass->lc_racct, resource, -amount); } /* * Decrease allocation of 'resource' by 'amount' for credential 'cred'. */ void racct_sub_cred(struct ucred *cred, int resource, uint64_t amount) { if (!racct_enable) return; SDT_PROBE3(racct, , rusage, sub__cred, cred, resource, amount); #ifdef notyet KASSERT(RACCT_CAN_DROP(resource), ("%s: called for resource %d which can not drop", __func__, resource)); #endif RACCT_LOCK(); racct_sub_cred_locked(cred, resource, amount); RACCT_UNLOCK(); } /* * Inherit resource usage information from the parent process. */ int racct_proc_fork(struct proc *parent, struct proc *child) { int i, error = 0; if (!racct_enable) return (0); /* * Create racct for the child process. */ racct_create(&child->p_racct); PROC_LOCK(parent); PROC_LOCK(child); RACCT_LOCK(); #ifdef RCTL error = rctl_proc_fork(parent, child); if (error != 0) goto out; #endif /* Init process cpu time. */ child->p_prev_runtime = 0; child->p_throttled = 0; /* * Inherit resource usage. */ for (i = 0; i <= RACCT_MAX; i++) { if (parent->p_racct->r_resources[i] == 0 || !RACCT_IS_INHERITABLE(i)) continue; error = racct_set_locked(child, i, parent->p_racct->r_resources[i], 0); if (error != 0) goto out; } error = racct_add_locked(child, RACCT_NPROC, 1, 0); error += racct_add_locked(child, RACCT_NTHR, 1, 0); out: RACCT_UNLOCK(); PROC_UNLOCK(child); PROC_UNLOCK(parent); if (error != 0) racct_proc_exit(child); return (error); } /* * Called at the end of fork1(), to handle rules that require the process * to be fully initialized. */ void racct_proc_fork_done(struct proc *child) { if (!racct_enable) return; PROC_LOCK_ASSERT(child, MA_OWNED); #ifdef RCTL RACCT_LOCK(); rctl_enforce(child, RACCT_NPROC, 0); rctl_enforce(child, RACCT_NTHR, 0); RACCT_UNLOCK(); #endif } void racct_proc_exit(struct proc *p) { struct timeval wallclock; uint64_t pct_estimate, pct, runtime; int i; if (!racct_enable) return; PROC_LOCK(p); /* * We don't need to calculate rux, proc_reap() has already done this. */ runtime = cputick2usec(p->p_rux.rux_runtime); #ifdef notyet KASSERT(runtime >= p->p_prev_runtime, ("runtime < p_prev_runtime")); #else if (runtime < p->p_prev_runtime) runtime = p->p_prev_runtime; #endif microuptime(&wallclock); timevalsub(&wallclock, &p->p_stats->p_start); if (wallclock.tv_sec > 0 || wallclock.tv_usec > 0) { pct_estimate = (1000000 * runtime * 100) / ((uint64_t)wallclock.tv_sec * 1000000 + wallclock.tv_usec); } else pct_estimate = 0; pct = racct_getpcpu(p, pct_estimate); RACCT_LOCK(); racct_set_locked(p, RACCT_CPU, runtime, 0); racct_add_cred_locked(p->p_ucred, RACCT_PCTCPU, pct); + KASSERT(p->p_racct->r_resources[RACCT_RSS] == 0, + ("process reaped with %ju allocated for RSS\n", + p->p_racct->r_resources[RACCT_RSS])); for (i = 0; i <= RACCT_MAX; i++) { if (p->p_racct->r_resources[i] == 0) continue; - if (!RACCT_IS_RECLAIMABLE(i)) + if (!RACCT_IS_RECLAIMABLE(i)) continue; racct_set_locked(p, i, 0, 0); } #ifdef RCTL rctl_racct_release(p->p_racct); #endif racct_destroy_locked(&p->p_racct); RACCT_UNLOCK(); PROC_UNLOCK(p); } /* * Called after credentials change, to move resource utilisation * between raccts. */ void racct_proc_ucred_changed(struct proc *p, struct ucred *oldcred, struct ucred *newcred) { struct uidinfo *olduip, *newuip; struct loginclass *oldlc, *newlc; struct prison *oldpr, *newpr, *pr; if (!racct_enable) return; PROC_LOCK_ASSERT(p, MA_NOTOWNED); newuip = newcred->cr_ruidinfo; olduip = oldcred->cr_ruidinfo; newlc = newcred->cr_loginclass; oldlc = oldcred->cr_loginclass; newpr = newcred->cr_prison; oldpr = oldcred->cr_prison; RACCT_LOCK(); if (newuip != olduip) { racct_sub_racct(olduip->ui_racct, p->p_racct); racct_add_racct(newuip->ui_racct, p->p_racct); } if (newlc != oldlc) { racct_sub_racct(oldlc->lc_racct, p->p_racct); racct_add_racct(newlc->lc_racct, p->p_racct); } if (newpr != oldpr) { for (pr = oldpr; pr != NULL; pr = pr->pr_parent) racct_sub_racct(pr->pr_prison_racct->prr_racct, p->p_racct); for (pr = newpr; pr != NULL; pr = pr->pr_parent) racct_add_racct(pr->pr_prison_racct->prr_racct, p->p_racct); } RACCT_UNLOCK(); #ifdef RCTL rctl_proc_ucred_changed(p, newcred); #endif } void racct_move(struct racct *dest, struct racct *src) { ASSERT_RACCT_ENABLED(); RACCT_LOCK(); racct_add_racct(dest, src); racct_sub_racct(src, src); RACCT_UNLOCK(); } /* * Make the process sleep in userret() for 'timeout' ticks. Setting * timeout to -1 makes it sleep until woken up by racct_proc_wakeup(). */ void racct_proc_throttle(struct proc *p, int timeout) { struct thread *td; #ifdef SMP int cpuid; #endif KASSERT(timeout != 0, ("timeout %d", timeout)); ASSERT_RACCT_ENABLED(); PROC_LOCK_ASSERT(p, MA_OWNED); /* * Do not block kernel processes. Also do not block processes with * low %cpu utilization to improve interactivity. */ if ((p->p_flag & (P_SYSTEM | P_KPROC)) != 0) return; if (p->p_throttled < 0 || (timeout > 0 && p->p_throttled > timeout)) return; p->p_throttled = timeout; FOREACH_THREAD_IN_PROC(p, td) { thread_lock(td); switch (td->td_state) { case TDS_RUNQ: /* * If the thread is on the scheduler run-queue, we can * not just remove it from there. So we set the flag * TDF_NEEDRESCHED for the thread, so that once it is * running, it is taken off the cpu as soon as possible. */ td->td_flags |= TDF_NEEDRESCHED; break; case TDS_RUNNING: /* * If the thread is running, we request a context * switch for it by setting the TDF_NEEDRESCHED flag. */ td->td_flags |= TDF_NEEDRESCHED; #ifdef SMP cpuid = td->td_oncpu; if ((cpuid != NOCPU) && (td != curthread)) ipi_cpu(cpuid, IPI_AST); #endif break; default: break; } thread_unlock(td); } } static void racct_proc_wakeup(struct proc *p) { ASSERT_RACCT_ENABLED(); PROC_LOCK_ASSERT(p, MA_OWNED); if (p->p_throttled != 0) { p->p_throttled = 0; wakeup(p->p_racct); } } static void racct_decay_callback(struct racct *racct, void *dummy1, void *dummy2) { int64_t r_old, r_new; ASSERT_RACCT_ENABLED(); RACCT_LOCK_ASSERT(); #ifdef RCTL rctl_throttle_decay(racct, RACCT_READBPS); rctl_throttle_decay(racct, RACCT_WRITEBPS); rctl_throttle_decay(racct, RACCT_READIOPS); rctl_throttle_decay(racct, RACCT_WRITEIOPS); #endif r_old = racct->r_resources[RACCT_PCTCPU]; /* If there is nothing to decay, just exit. */ if (r_old <= 0) return; r_new = r_old * RACCT_DECAY_FACTOR / FSCALE; racct->r_resources[RACCT_PCTCPU] = r_new; } static void racct_decay_pre(void) { RACCT_LOCK(); } static void racct_decay_post(void) { RACCT_UNLOCK(); } static void racct_decay(void) { ASSERT_RACCT_ENABLED(); ui_racct_foreach(racct_decay_callback, racct_decay_pre, racct_decay_post, NULL, NULL); loginclass_racct_foreach(racct_decay_callback, racct_decay_pre, racct_decay_post, NULL, NULL); prison_racct_foreach(racct_decay_callback, racct_decay_pre, racct_decay_post, NULL, NULL); } static void racctd(void) { struct thread *td; struct proc *p; struct timeval wallclock; uint64_t pct, pct_estimate, runtime; ASSERT_RACCT_ENABLED(); for (;;) { racct_decay(); sx_slock(&allproc_lock); LIST_FOREACH(p, &zombproc, p_list) { PROC_LOCK(p); racct_set(p, RACCT_PCTCPU, 0); PROC_UNLOCK(p); } FOREACH_PROC_IN_SYSTEM(p) { PROC_LOCK(p); if (p->p_state != PRS_NORMAL) { PROC_UNLOCK(p); continue; } microuptime(&wallclock); timevalsub(&wallclock, &p->p_stats->p_start); PROC_STATLOCK(p); FOREACH_THREAD_IN_PROC(p, td) ruxagg(p, td); runtime = cputick2usec(p->p_rux.rux_runtime); PROC_STATUNLOCK(p); #ifdef notyet KASSERT(runtime >= p->p_prev_runtime, ("runtime < p_prev_runtime")); #else if (runtime < p->p_prev_runtime) runtime = p->p_prev_runtime; #endif p->p_prev_runtime = runtime; if (wallclock.tv_sec > 0 || wallclock.tv_usec > 0) { pct_estimate = (1000000 * runtime * 100) / ((uint64_t)wallclock.tv_sec * 1000000 + wallclock.tv_usec); } else pct_estimate = 0; pct = racct_getpcpu(p, pct_estimate); RACCT_LOCK(); #ifdef RCTL rctl_throttle_decay(p->p_racct, RACCT_READBPS); rctl_throttle_decay(p->p_racct, RACCT_WRITEBPS); rctl_throttle_decay(p->p_racct, RACCT_READIOPS); rctl_throttle_decay(p->p_racct, RACCT_WRITEIOPS); #endif racct_set_locked(p, RACCT_PCTCPU, pct, 1); racct_set_locked(p, RACCT_CPU, runtime, 0); racct_set_locked(p, RACCT_WALLCLOCK, (uint64_t)wallclock.tv_sec * 1000000 + wallclock.tv_usec, 0); RACCT_UNLOCK(); PROC_UNLOCK(p); } /* * To ensure that processes are throttled in a fair way, we need * to iterate over all processes again and check the limits * for %cpu resource only after ucred racct containers have been * properly filled. */ FOREACH_PROC_IN_SYSTEM(p) { PROC_LOCK(p); if (p->p_state != PRS_NORMAL) { PROC_UNLOCK(p); continue; } if (racct_pcpu_available(p) <= 0) { if (p->p_racct->r_resources[RACCT_PCTCPU] > pcpu_threshold) racct_proc_throttle(p, -1); } else if (p->p_throttled == -1) { racct_proc_wakeup(p); } PROC_UNLOCK(p); } sx_sunlock(&allproc_lock); pause("-", hz); } } static struct kproc_desc racctd_kp = { "racctd", racctd, NULL }; static void racctd_init(void) { if (!racct_enable) return; kproc_start(&racctd_kp); } SYSINIT(racctd, SI_SUB_RACCTD, SI_ORDER_FIRST, racctd_init, NULL); static void racct_init(void) { if (!racct_enable) return; racct_zone = uma_zcreate("racct", sizeof(struct racct), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); /* * XXX: Move this somewhere. */ prison0.pr_prison_racct = prison_racct_find("0"); } SYSINIT(racct, SI_SUB_RACCT, SI_ORDER_FIRST, racct_init, NULL); #endif /* !RACCT */ Index: stable/11/sys/vm/vm_pageout.c =================================================================== --- stable/11/sys/vm/vm_pageout.c (revision 314346) +++ stable/11/sys/vm/vm_pageout.c (revision 314347) @@ -1,1856 +1,1859 @@ /*- * Copyright (c) 1991 Regents of the University of California. * All rights reserved. * Copyright (c) 1994 John S. Dyson * All rights reserved. * Copyright (c) 1994 David Greenman * All rights reserved. * Copyright (c) 2005 Yahoo! Technologies Norway AS * All rights reserved. * * This code is derived from software contributed to Berkeley by * The Mach Operating System project at Carnegie-Mellon University. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. All advertising materials mentioning features or use of this software * must display the following acknowledgement: * This product includes software developed by the University of * California, Berkeley and its contributors. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)vm_pageout.c 7.4 (Berkeley) 5/7/91 * * * Copyright (c) 1987, 1990 Carnegie-Mellon University. * All rights reserved. * * Authors: Avadis Tevanian, Jr., Michael Wayne Young * * Permission to use, copy, modify and distribute this software and * its documentation is hereby granted, provided that both the copyright * notice and this permission notice appear in all copies of the * software, derivative works or modified versions, and any portions * thereof, and that both notices appear in supporting documentation. * * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. * * Carnegie Mellon requests users of this software to return to * * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU * School of Computer Science * Carnegie Mellon University * Pittsburgh PA 15213-3890 * * any improvements or extensions that they make and grant Carnegie the * rights to redistribute these changes. */ /* * The proverbial page-out daemon. */ #include __FBSDID("$FreeBSD$"); #include "opt_vm.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include /* * System initialization */ /* the kernel process "vm_pageout"*/ static void vm_pageout(void); static void vm_pageout_init(void); static int vm_pageout_clean(vm_page_t m); static int vm_pageout_cluster(vm_page_t m); static bool vm_pageout_scan(struct vm_domain *vmd, int pass); static void vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage, int starting_page_shortage); SYSINIT(pagedaemon_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, vm_pageout_init, NULL); struct proc *pageproc; static struct kproc_desc page_kp = { "pagedaemon", vm_pageout, &pageproc }; SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start, &page_kp); SDT_PROVIDER_DEFINE(vm); SDT_PROBE_DEFINE(vm, , , vm__lowmem_scan); #if !defined(NO_SWAPPING) /* the kernel process "vm_daemon"*/ static void vm_daemon(void); static struct proc *vmproc; static struct kproc_desc vm_kp = { "vmdaemon", vm_daemon, &vmproc }; SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp); #endif int vm_pageout_deficit; /* Estimated number of pages deficit */ int vm_pageout_wakeup_thresh; static int vm_pageout_oom_seq = 12; bool vm_pageout_wanted; /* Event on which pageout daemon sleeps */ bool vm_pages_needed; /* Are threads waiting for free pages? */ #if !defined(NO_SWAPPING) static int vm_pageout_req_swapout; /* XXX */ static int vm_daemon_needed; static struct mtx vm_daemon_mtx; /* Allow for use by vm_pageout before vm_daemon is initialized. */ MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF); #endif static int vm_max_launder = 32; static int vm_pageout_update_period; static int defer_swap_pageouts; static int disable_swap_pageouts; static int lowmem_period = 10; static time_t lowmem_uptime; #if defined(NO_SWAPPING) static int vm_swap_enabled = 0; static int vm_swap_idle_enabled = 0; #else static int vm_swap_enabled = 1; static int vm_swap_idle_enabled = 0; #endif static int vm_panic_on_oom = 0; SYSCTL_INT(_vm, OID_AUTO, panic_on_oom, CTLFLAG_RWTUN, &vm_panic_on_oom, 0, "panic on out of memory instead of killing the largest process"); SYSCTL_INT(_vm, OID_AUTO, pageout_wakeup_thresh, CTLFLAG_RW, &vm_pageout_wakeup_thresh, 0, "free page threshold for waking up the pageout daemon"); SYSCTL_INT(_vm, OID_AUTO, max_launder, CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout"); SYSCTL_INT(_vm, OID_AUTO, pageout_update_period, CTLFLAG_RW, &vm_pageout_update_period, 0, "Maximum active LRU update period"); SYSCTL_INT(_vm, OID_AUTO, lowmem_period, CTLFLAG_RW, &lowmem_period, 0, "Low memory callback period"); #if defined(NO_SWAPPING) SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, CTLFLAG_RD, &vm_swap_enabled, 0, "Enable entire process swapout"); SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, CTLFLAG_RD, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); #else SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout"); SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); #endif SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts, CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem"); SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts, CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages"); static int pageout_lock_miss; SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss, CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout"); SYSCTL_INT(_vm, OID_AUTO, pageout_oom_seq, CTLFLAG_RW, &vm_pageout_oom_seq, 0, "back-to-back calls to oom detector to start OOM"); #define VM_PAGEOUT_PAGE_COUNT 16 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT; int vm_page_max_wired; /* XXX max # of wired pages system-wide */ SYSCTL_INT(_vm, OID_AUTO, max_wired, CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count"); static boolean_t vm_pageout_fallback_object_lock(vm_page_t, vm_page_t *); #if !defined(NO_SWAPPING) static void vm_pageout_map_deactivate_pages(vm_map_t, long); static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long); static void vm_req_vmdaemon(int req); #endif static boolean_t vm_pageout_page_lock(vm_page_t, vm_page_t *); /* * Initialize a dummy page for marking the caller's place in the specified * paging queue. In principle, this function only needs to set the flag * PG_MARKER. Nonetheless, it write busies and initializes the hold count * to one as safety precautions. */ static void vm_pageout_init_marker(vm_page_t marker, u_short queue) { bzero(marker, sizeof(*marker)); marker->flags = PG_MARKER; marker->busy_lock = VPB_SINGLE_EXCLUSIVER; marker->queue = queue; marker->hold_count = 1; } /* * vm_pageout_fallback_object_lock: * * Lock vm object currently associated with `m'. VM_OBJECT_TRYWLOCK is * known to have failed and page queue must be either PQ_ACTIVE or * PQ_INACTIVE. To avoid lock order violation, unlock the page queue * while locking the vm object. Use marker page to detect page queue * changes and maintain notion of next page on page queue. Return * TRUE if no changes were detected, FALSE otherwise. vm object is * locked on return. * * This function depends on both the lock portion of struct vm_object * and normal struct vm_page being type stable. */ static boolean_t vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next) { struct vm_page marker; struct vm_pagequeue *pq; boolean_t unchanged; u_short queue; vm_object_t object; queue = m->queue; vm_pageout_init_marker(&marker, queue); pq = vm_page_pagequeue(m); object = m->object; TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q); vm_pagequeue_unlock(pq); vm_page_unlock(m); VM_OBJECT_WLOCK(object); vm_page_lock(m); vm_pagequeue_lock(pq); /* * The page's object might have changed, and/or the page might * have moved from its original position in the queue. If the * page's object has changed, then the caller should abandon * processing the page because the wrong object lock was * acquired. Use the marker's plinks.q, not the page's, to * determine if the page has been moved. The state of the * page's plinks.q can be indeterminate; whereas, the marker's * plinks.q must be valid. */ *next = TAILQ_NEXT(&marker, plinks.q); unchanged = m->object == object && m == TAILQ_PREV(&marker, pglist, plinks.q); KASSERT(!unchanged || m->queue == queue, ("page %p queue %d %d", m, queue, m->queue)); TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q); return (unchanged); } /* * Lock the page while holding the page queue lock. Use marker page * to detect page queue changes and maintain notion of next page on * page queue. Return TRUE if no changes were detected, FALSE * otherwise. The page is locked on return. The page queue lock might * be dropped and reacquired. * * This function depends on normal struct vm_page being type stable. */ static boolean_t vm_pageout_page_lock(vm_page_t m, vm_page_t *next) { struct vm_page marker; struct vm_pagequeue *pq; boolean_t unchanged; u_short queue; vm_page_lock_assert(m, MA_NOTOWNED); if (vm_page_trylock(m)) return (TRUE); queue = m->queue; vm_pageout_init_marker(&marker, queue); pq = vm_page_pagequeue(m); TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q); vm_pagequeue_unlock(pq); vm_page_lock(m); vm_pagequeue_lock(pq); /* Page queue might have changed. */ *next = TAILQ_NEXT(&marker, plinks.q); unchanged = m == TAILQ_PREV(&marker, pglist, plinks.q); KASSERT(!unchanged || m->queue == queue, ("page %p queue %d %d", m, queue, m->queue)); TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q); return (unchanged); } /* * Scan for pages at adjacent offsets within the given page's object that are * eligible for laundering, form a cluster of these pages and the given page, * and launder that cluster. */ static int vm_pageout_cluster(vm_page_t m) { vm_object_t object; vm_page_t mc[2 * vm_pageout_page_count], p, pb, ps; vm_pindex_t pindex; int ib, is, page_base, pageout_count; vm_page_assert_locked(m); object = m->object; VM_OBJECT_ASSERT_WLOCKED(object); pindex = m->pindex; /* * We can't clean the page if it is busy or held. */ vm_page_assert_unbusied(m); KASSERT(m->hold_count == 0, ("page %p is held", m)); vm_page_unlock(m); mc[vm_pageout_page_count] = pb = ps = m; pageout_count = 1; page_base = vm_pageout_page_count; ib = 1; is = 1; /* * We can cluster only if the page is not clean, busy, or held, and * the page is inactive. * * During heavy mmap/modification loads the pageout * daemon can really fragment the underlying file * due to flushing pages out of order and not trying to * align the clusters (which leaves sporadic out-of-order * holes). To solve this problem we do the reverse scan * first and attempt to align our cluster, then do a * forward scan if room remains. */ more: while (ib != 0 && pageout_count < vm_pageout_page_count) { if (ib > pindex) { ib = 0; break; } if ((p = vm_page_prev(pb)) == NULL || vm_page_busied(p)) { ib = 0; break; } vm_page_test_dirty(p); if (p->dirty == 0) { ib = 0; break; } vm_page_lock(p); if (p->queue != PQ_INACTIVE || p->hold_count != 0) { /* may be undergoing I/O */ vm_page_unlock(p); ib = 0; break; } vm_page_unlock(p); mc[--page_base] = pb = p; ++pageout_count; ++ib; /* * We are at an alignment boundary. Stop here, and switch * directions. Do not clear ib. */ if ((pindex - (ib - 1)) % vm_pageout_page_count == 0) break; } while (pageout_count < vm_pageout_page_count && pindex + is < object->size) { if ((p = vm_page_next(ps)) == NULL || vm_page_busied(p)) break; vm_page_test_dirty(p); if (p->dirty == 0) break; vm_page_lock(p); if (p->queue != PQ_INACTIVE || p->hold_count != 0) { /* may be undergoing I/O */ vm_page_unlock(p); break; } vm_page_unlock(p); mc[page_base + pageout_count] = ps = p; ++pageout_count; ++is; } /* * If we exhausted our forward scan, continue with the reverse scan * when possible, even past an alignment boundary. This catches * boundary conditions. */ if (ib != 0 && pageout_count < vm_pageout_page_count) goto more; return (vm_pageout_flush(&mc[page_base], pageout_count, 0, 0, NULL, NULL)); } /* * vm_pageout_flush() - launder the given pages * * The given pages are laundered. Note that we setup for the start of * I/O ( i.e. busy the page ), mark it read-only, and bump the object * reference count all in here rather then in the parent. If we want * the parent to do more sophisticated things we may have to change * the ordering. * * Returned runlen is the count of pages between mreq and first * page after mreq with status VM_PAGER_AGAIN. * *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL * for any page in runlen set. */ int vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen, boolean_t *eio) { vm_object_t object = mc[0]->object; int pageout_status[count]; int numpagedout = 0; int i, runlen; VM_OBJECT_ASSERT_WLOCKED(object); /* * Initiate I/O. Bump the vm_page_t->busy counter and * mark the pages read-only. * * We do not have to fixup the clean/dirty bits here... we can * allow the pager to do it after the I/O completes. * * NOTE! mc[i]->dirty may be partial or fragmented due to an * edge case with file fragments. */ for (i = 0; i < count; i++) { KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, ("vm_pageout_flush: partially invalid page %p index %d/%d", mc[i], i, count)); vm_page_sbusy(mc[i]); pmap_remove_write(mc[i]); } vm_object_pip_add(object, count); vm_pager_put_pages(object, mc, count, flags, pageout_status); runlen = count - mreq; if (eio != NULL) *eio = FALSE; for (i = 0; i < count; i++) { vm_page_t mt = mc[i]; KASSERT(pageout_status[i] == VM_PAGER_PEND || !pmap_page_is_write_mapped(mt), ("vm_pageout_flush: page %p is not write protected", mt)); switch (pageout_status[i]) { case VM_PAGER_OK: case VM_PAGER_PEND: numpagedout++; break; case VM_PAGER_BAD: /* * Page outside of range of object. Right now we * essentially lose the changes by pretending it * worked. */ vm_page_undirty(mt); break; case VM_PAGER_ERROR: case VM_PAGER_FAIL: /* * If page couldn't be paged out, then reactivate the * page so it doesn't clog the inactive list. (We * will try paging out it again later). */ vm_page_lock(mt); vm_page_activate(mt); vm_page_unlock(mt); if (eio != NULL && i >= mreq && i - mreq < runlen) *eio = TRUE; break; case VM_PAGER_AGAIN: if (i >= mreq && i - mreq < runlen) runlen = i - mreq; break; } /* * If the operation is still going, leave the page busy to * block all other accesses. Also, leave the paging in * progress indicator set so that we don't attempt an object * collapse. */ if (pageout_status[i] != VM_PAGER_PEND) { vm_object_pip_wakeup(object); vm_page_sunbusy(mt); } } if (prunlen != NULL) *prunlen = runlen; return (numpagedout); } #if !defined(NO_SWAPPING) /* * vm_pageout_object_deactivate_pages * * Deactivate enough pages to satisfy the inactive target * requirements. * * The object and map must be locked. */ static void vm_pageout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object, long desired) { vm_object_t backing_object, object; vm_page_t p; int act_delta, remove_mode; VM_OBJECT_ASSERT_LOCKED(first_object); if ((first_object->flags & OBJ_FICTITIOUS) != 0) return; for (object = first_object;; object = backing_object) { if (pmap_resident_count(pmap) <= desired) goto unlock_return; VM_OBJECT_ASSERT_LOCKED(object); if ((object->flags & OBJ_UNMANAGED) != 0 || object->paging_in_progress != 0) goto unlock_return; remove_mode = 0; if (object->shadow_count > 1) remove_mode = 1; /* * Scan the object's entire memory queue. */ TAILQ_FOREACH(p, &object->memq, listq) { if (pmap_resident_count(pmap) <= desired) goto unlock_return; if (vm_page_busied(p)) continue; PCPU_INC(cnt.v_pdpages); vm_page_lock(p); if (p->wire_count != 0 || p->hold_count != 0 || !pmap_page_exists_quick(pmap, p)) { vm_page_unlock(p); continue; } act_delta = pmap_ts_referenced(p); if ((p->aflags & PGA_REFERENCED) != 0) { if (act_delta == 0) act_delta = 1; vm_page_aflag_clear(p, PGA_REFERENCED); } if (p->queue != PQ_ACTIVE && act_delta != 0) { vm_page_activate(p); p->act_count += act_delta; } else if (p->queue == PQ_ACTIVE) { if (act_delta == 0) { p->act_count -= min(p->act_count, ACT_DECLINE); if (!remove_mode && p->act_count == 0) { pmap_remove_all(p); vm_page_deactivate(p); } else vm_page_requeue(p); } else { vm_page_activate(p); if (p->act_count < ACT_MAX - ACT_ADVANCE) p->act_count += ACT_ADVANCE; vm_page_requeue(p); } } else if (p->queue == PQ_INACTIVE) pmap_remove_all(p); vm_page_unlock(p); } if ((backing_object = object->backing_object) == NULL) goto unlock_return; VM_OBJECT_RLOCK(backing_object); if (object != first_object) VM_OBJECT_RUNLOCK(object); } unlock_return: if (object != first_object) VM_OBJECT_RUNLOCK(object); } /* * deactivate some number of pages in a map, try to do it fairly, but * that is really hard to do. */ static void vm_pageout_map_deactivate_pages(map, desired) vm_map_t map; long desired; { vm_map_entry_t tmpe; vm_object_t obj, bigobj; int nothingwired; if (!vm_map_trylock(map)) return; bigobj = NULL; nothingwired = TRUE; /* * first, search out the biggest object, and try to free pages from * that. */ tmpe = map->header.next; while (tmpe != &map->header) { if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { obj = tmpe->object.vm_object; if (obj != NULL && VM_OBJECT_TRYRLOCK(obj)) { if (obj->shadow_count <= 1 && (bigobj == NULL || bigobj->resident_page_count < obj->resident_page_count)) { if (bigobj != NULL) VM_OBJECT_RUNLOCK(bigobj); bigobj = obj; } else VM_OBJECT_RUNLOCK(obj); } } if (tmpe->wired_count > 0) nothingwired = FALSE; tmpe = tmpe->next; } if (bigobj != NULL) { vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired); VM_OBJECT_RUNLOCK(bigobj); } /* * Next, hunt around for other pages to deactivate. We actually * do this search sort of wrong -- .text first is not the best idea. */ tmpe = map->header.next; while (tmpe != &map->header) { if (pmap_resident_count(vm_map_pmap(map)) <= desired) break; if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { obj = tmpe->object.vm_object; if (obj != NULL) { VM_OBJECT_RLOCK(obj); vm_pageout_object_deactivate_pages(map->pmap, obj, desired); VM_OBJECT_RUNLOCK(obj); } } tmpe = tmpe->next; } /* * Remove all mappings if a process is swapped out, this will free page * table pages. */ if (desired == 0 && nothingwired) { pmap_remove(vm_map_pmap(map), vm_map_min(map), vm_map_max(map)); } vm_map_unlock(map); } #endif /* !defined(NO_SWAPPING) */ /* * Attempt to acquire all of the necessary locks to launder a page and * then call through the clustering layer to PUTPAGES. Wait a short * time for a vnode lock. * * Requires the page and object lock on entry, releases both before return. * Returns 0 on success and an errno otherwise. */ static int vm_pageout_clean(vm_page_t m) { struct vnode *vp; struct mount *mp; vm_object_t object; vm_pindex_t pindex; int error, lockmode; vm_page_assert_locked(m); object = m->object; VM_OBJECT_ASSERT_WLOCKED(object); error = 0; vp = NULL; mp = NULL; /* * The object is already known NOT to be dead. It * is possible for the vget() to block the whole * pageout daemon, but the new low-memory handling * code should prevent it. * * We can't wait forever for the vnode lock, we might * deadlock due to a vn_read() getting stuck in * vm_wait while holding this vnode. We skip the * vnode if we can't get it in a reasonable amount * of time. */ if (object->type == OBJT_VNODE) { vm_page_unlock(m); vp = object->handle; if (vp->v_type == VREG && vn_start_write(vp, &mp, V_NOWAIT) != 0) { mp = NULL; error = EDEADLK; goto unlock_all; } KASSERT(mp != NULL, ("vp %p with NULL v_mount", vp)); vm_object_reference_locked(object); pindex = m->pindex; VM_OBJECT_WUNLOCK(object); lockmode = MNT_SHARED_WRITES(vp->v_mount) ? LK_SHARED : LK_EXCLUSIVE; if (vget(vp, lockmode | LK_TIMELOCK, curthread)) { vp = NULL; error = EDEADLK; goto unlock_mp; } VM_OBJECT_WLOCK(object); vm_page_lock(m); /* * While the object and page were unlocked, the page * may have been: * (1) moved to a different queue, * (2) reallocated to a different object, * (3) reallocated to a different offset, or * (4) cleaned. */ if (m->queue != PQ_INACTIVE || m->object != object || m->pindex != pindex || m->dirty == 0) { vm_page_unlock(m); error = ENXIO; goto unlock_all; } /* * The page may have been busied or held while the object * and page locks were released. */ if (vm_page_busied(m) || m->hold_count != 0) { vm_page_unlock(m); error = EBUSY; goto unlock_all; } } /* * If a page is dirty, then it is either being washed * (but not yet cleaned) or it is still in the * laundry. If it is still in the laundry, then we * start the cleaning operation. */ if (vm_pageout_cluster(m) == 0) error = EIO; unlock_all: VM_OBJECT_WUNLOCK(object); unlock_mp: vm_page_lock_assert(m, MA_NOTOWNED); if (mp != NULL) { if (vp != NULL) vput(vp); vm_object_deallocate(object); vn_finished_write(mp); } return (error); } /* * vm_pageout_scan does the dirty work for the pageout daemon. * * pass 0 - Update active LRU/deactivate pages * pass 1 - Free inactive pages * pass 2 - Launder dirty pages * * Returns true if pass was zero or enough pages were freed by the inactive * queue scan to meet the target. */ static bool vm_pageout_scan(struct vm_domain *vmd, int pass) { vm_page_t m, next; struct vm_pagequeue *pq; vm_object_t object; long min_scan; int act_delta, addl_page_shortage, deficit, error, inactq_shortage; int maxlaunder, maxscan, page_shortage, scan_tick, scanned; int starting_page_shortage, vnodes_skipped; boolean_t pageout_ok, queue_locked; /* * If we need to reclaim memory ask kernel caches to return * some. We rate limit to avoid thrashing. */ if (vmd == &vm_dom[0] && pass > 0 && (time_uptime - lowmem_uptime) >= lowmem_period) { /* * Decrease registered cache sizes. */ SDT_PROBE0(vm, , , vm__lowmem_scan); EVENTHANDLER_INVOKE(vm_lowmem, 0); /* * We do this explicitly after the caches have been * drained above. */ uma_reclaim(); lowmem_uptime = time_uptime; } /* * The addl_page_shortage is the number of temporarily * stuck pages in the inactive queue. In other words, the * number of pages from the inactive count that should be * discounted in setting the target for the active queue scan. */ addl_page_shortage = 0; /* * Calculate the number of pages that we want to free. This number * can be negative if many pages are freed between the wakeup call to * the page daemon and this calculation. */ if (pass > 0) { deficit = atomic_readandclear_int(&vm_pageout_deficit); page_shortage = vm_paging_target() + deficit; } else page_shortage = deficit = 0; starting_page_shortage = page_shortage; /* * maxlaunder limits the number of dirty pages we flush per scan. * For most systems a smaller value (16 or 32) is more robust under * extreme memory and disk pressure because any unnecessary writes * to disk can result in extreme performance degredation. However, * systems with excessive dirty pages (especially when MAP_NOSYNC is * used) will die horribly with limited laundering. If the pageout * daemon cannot clean enough pages in the first pass, we let it go * all out in succeeding passes. */ if ((maxlaunder = vm_max_launder) <= 1) maxlaunder = 1; if (pass > 1) maxlaunder = 10000; vnodes_skipped = 0; /* * Start scanning the inactive queue for pages that we can free. The * scan will stop when we reach the target or we have scanned the * entire queue. (Note that m->act_count is not used to make * decisions for the inactive queue, only for the active queue.) */ pq = &vmd->vmd_pagequeues[PQ_INACTIVE]; maxscan = pq->pq_cnt; vm_pagequeue_lock(pq); queue_locked = TRUE; for (m = TAILQ_FIRST(&pq->pq_pl); m != NULL && maxscan-- > 0 && page_shortage > 0; m = next) { vm_pagequeue_assert_locked(pq); KASSERT(queue_locked, ("unlocked inactive queue")); KASSERT(m->queue == PQ_INACTIVE, ("Inactive queue %p", m)); PCPU_INC(cnt.v_pdpages); next = TAILQ_NEXT(m, plinks.q); /* * skip marker pages */ if (m->flags & PG_MARKER) continue; KASSERT((m->flags & PG_FICTITIOUS) == 0, ("Fictitious page %p cannot be in inactive queue", m)); KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("Unmanaged page %p cannot be in inactive queue", m)); /* * The page or object lock acquisitions fail if the * page was removed from the queue or moved to a * different position within the queue. In either * case, addl_page_shortage should not be incremented. */ if (!vm_pageout_page_lock(m, &next)) goto unlock_page; else if (m->hold_count != 0) { /* * Held pages are essentially stuck in the * queue. So, they ought to be discounted * from the inactive count. See the * calculation of inactq_shortage before the * loop over the active queue below. */ addl_page_shortage++; goto unlock_page; } object = m->object; if (!VM_OBJECT_TRYWLOCK(object)) { if (!vm_pageout_fallback_object_lock(m, &next)) goto unlock_object; else if (m->hold_count != 0) { addl_page_shortage++; goto unlock_object; } } if (vm_page_busied(m)) { /* * Don't mess with busy pages. Leave them at * the front of the queue. Most likely, they * are being paged out and will leave the * queue shortly after the scan finishes. So, * they ought to be discounted from the * inactive count. */ addl_page_shortage++; unlock_object: VM_OBJECT_WUNLOCK(object); unlock_page: vm_page_unlock(m); continue; } KASSERT(m->hold_count == 0, ("Held page %p", m)); /* * We unlock the inactive page queue, invalidating the * 'next' pointer. Use our marker to remember our * place. */ TAILQ_INSERT_AFTER(&pq->pq_pl, m, &vmd->vmd_marker, plinks.q); vm_pagequeue_unlock(pq); queue_locked = FALSE; /* * Invalid pages can be easily freed. They cannot be * mapped, vm_page_free() asserts this. */ if (m->valid == 0) goto free_page; /* * If the page has been referenced and the object is not dead, * reactivate or requeue the page depending on whether the * object is mapped. */ if ((m->aflags & PGA_REFERENCED) != 0) { vm_page_aflag_clear(m, PGA_REFERENCED); act_delta = 1; } else act_delta = 0; if (object->ref_count != 0) { act_delta += pmap_ts_referenced(m); } else { KASSERT(!pmap_page_is_mapped(m), ("vm_pageout_scan: page %p is mapped", m)); } if (act_delta != 0) { if (object->ref_count != 0) { vm_page_activate(m); /* * Increase the activation count if the page * was referenced while in the inactive queue. * This makes it less likely that the page will * be returned prematurely to the inactive * queue. */ m->act_count += act_delta + ACT_ADVANCE; goto drop_page; } else if ((object->flags & OBJ_DEAD) == 0) goto requeue_page; } /* * If the page appears to be clean at the machine-independent * layer, then remove all of its mappings from the pmap in * anticipation of freeing it. If, however, any of the page's * mappings allow write access, then the page may still be * modified until the last of those mappings are removed. */ if (object->ref_count != 0) { vm_page_test_dirty(m); if (m->dirty == 0) pmap_remove_all(m); } if (m->dirty == 0) { /* * Clean pages can be freed. */ free_page: vm_page_free(m); PCPU_INC(cnt.v_dfree); --page_shortage; } else if ((object->flags & OBJ_DEAD) != 0) { /* * Leave dirty pages from dead objects at the front of * the queue. They are being paged out and freed by * the thread that destroyed the object. They will * leave the queue shortly after the scan finishes, so * they should be discounted from the inactive count. */ addl_page_shortage++; } else if ((m->flags & PG_WINATCFLS) == 0 && pass < 2) { /* * Dirty pages need to be paged out, but flushing * a page is extremely expensive versus freeing * a clean page. Rather then artificially limiting * the number of pages we can flush, we instead give * dirty pages extra priority on the inactive queue * by forcing them to be cycled through the queue * twice before being flushed, after which the * (now clean) page will cycle through once more * before being freed. This significantly extends * the thrash point for a heavily loaded machine. */ m->flags |= PG_WINATCFLS; requeue_page: vm_pagequeue_lock(pq); queue_locked = TRUE; vm_page_requeue_locked(m); } else if (maxlaunder > 0) { /* * We always want to try to flush some dirty pages if * we encounter them, to keep the system stable. * Normally this number is small, but under extreme * pressure where there are insufficient clean pages * on the inactive queue, we may have to go all out. */ if (object->type != OBJT_SWAP && object->type != OBJT_DEFAULT) pageout_ok = TRUE; else if (disable_swap_pageouts) pageout_ok = FALSE; else if (defer_swap_pageouts) pageout_ok = vm_page_count_min(); else pageout_ok = TRUE; if (!pageout_ok) goto requeue_page; error = vm_pageout_clean(m); /* * Decrement page_shortage on success to account for * the (future) cleaned page. Otherwise we could wind * up laundering or cleaning too many pages. */ if (error == 0) { page_shortage--; maxlaunder--; } else if (error == EDEADLK) { pageout_lock_miss++; vnodes_skipped++; } else if (error == EBUSY) { addl_page_shortage++; } vm_page_lock_assert(m, MA_NOTOWNED); goto relock_queue; } drop_page: vm_page_unlock(m); VM_OBJECT_WUNLOCK(object); relock_queue: if (!queue_locked) { vm_pagequeue_lock(pq); queue_locked = TRUE; } next = TAILQ_NEXT(&vmd->vmd_marker, plinks.q); TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_marker, plinks.q); } vm_pagequeue_unlock(pq); #if !defined(NO_SWAPPING) /* * Wakeup the swapout daemon if we didn't free the targeted number of * pages. */ if (vm_swap_enabled && page_shortage > 0) vm_req_vmdaemon(VM_SWAP_NORMAL); #endif /* * Wakeup the sync daemon if we skipped a vnode in a writeable object * and we didn't free enough pages. */ if (vnodes_skipped > 0 && page_shortage > vm_cnt.v_free_target - vm_cnt.v_free_min) (void)speedup_syncer(); /* * If the inactive queue scan fails repeatedly to meet its * target, kill the largest process. */ vm_pageout_mightbe_oom(vmd, page_shortage, starting_page_shortage); /* * Compute the number of pages we want to try to move from the * active queue to the inactive queue. */ inactq_shortage = vm_cnt.v_inactive_target - vm_cnt.v_inactive_count + vm_paging_target() + deficit + addl_page_shortage; pq = &vmd->vmd_pagequeues[PQ_ACTIVE]; vm_pagequeue_lock(pq); maxscan = pq->pq_cnt; /* * If we're just idle polling attempt to visit every * active page within 'update_period' seconds. */ scan_tick = ticks; if (vm_pageout_update_period != 0) { min_scan = pq->pq_cnt; min_scan *= scan_tick - vmd->vmd_last_active_scan; min_scan /= hz * vm_pageout_update_period; } else min_scan = 0; if (min_scan > 0 || (inactq_shortage > 0 && maxscan > 0)) vmd->vmd_last_active_scan = scan_tick; /* * Scan the active queue for pages that can be deactivated. Update * the per-page activity counter and use it to identify deactivation * candidates. Held pages may be deactivated. */ for (m = TAILQ_FIRST(&pq->pq_pl), scanned = 0; m != NULL && (scanned < min_scan || (inactq_shortage > 0 && scanned < maxscan)); m = next, scanned++) { KASSERT(m->queue == PQ_ACTIVE, ("vm_pageout_scan: page %p isn't active", m)); next = TAILQ_NEXT(m, plinks.q); if ((m->flags & PG_MARKER) != 0) continue; KASSERT((m->flags & PG_FICTITIOUS) == 0, ("Fictitious page %p cannot be in active queue", m)); KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("Unmanaged page %p cannot be in active queue", m)); if (!vm_pageout_page_lock(m, &next)) { vm_page_unlock(m); continue; } /* * The count for page daemon pages is updated after checking * the page for eligibility. */ PCPU_INC(cnt.v_pdpages); /* * Check to see "how much" the page has been used. */ if ((m->aflags & PGA_REFERENCED) != 0) { vm_page_aflag_clear(m, PGA_REFERENCED); act_delta = 1; } else act_delta = 0; /* * Perform an unsynchronized object ref count check. While * the page lock ensures that the page is not reallocated to * another object, in particular, one with unmanaged mappings * that cannot support pmap_ts_referenced(), two races are, * nonetheless, possible: * 1) The count was transitioning to zero, but we saw a non- * zero value. pmap_ts_referenced() will return zero * because the page is not mapped. * 2) The count was transitioning to one, but we saw zero. * This race delays the detection of a new reference. At * worst, we will deactivate and reactivate the page. */ if (m->object->ref_count != 0) act_delta += pmap_ts_referenced(m); /* * Advance or decay the act_count based on recent usage. */ if (act_delta != 0) { m->act_count += ACT_ADVANCE + act_delta; if (m->act_count > ACT_MAX) m->act_count = ACT_MAX; } else m->act_count -= min(m->act_count, ACT_DECLINE); /* * Move this page to the tail of the active or inactive * queue depending on usage. */ if (m->act_count == 0) { /* Dequeue to avoid later lock recursion. */ vm_page_dequeue_locked(m); vm_page_deactivate(m); inactq_shortage--; } else vm_page_requeue_locked(m); vm_page_unlock(m); } vm_pagequeue_unlock(pq); #if !defined(NO_SWAPPING) /* * Idle process swapout -- run once per second. */ if (vm_swap_idle_enabled) { static long lsec; if (time_second != lsec) { vm_req_vmdaemon(VM_SWAP_IDLE); lsec = time_second; } } #endif return (page_shortage <= 0); } static int vm_pageout_oom_vote; /* * The pagedaemon threads randlomly select one to perform the * OOM. Trying to kill processes before all pagedaemons * failed to reach free target is premature. */ static void vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage, int starting_page_shortage) { int old_vote; if (starting_page_shortage <= 0 || starting_page_shortage != page_shortage) vmd->vmd_oom_seq = 0; else vmd->vmd_oom_seq++; if (vmd->vmd_oom_seq < vm_pageout_oom_seq) { if (vmd->vmd_oom) { vmd->vmd_oom = FALSE; atomic_subtract_int(&vm_pageout_oom_vote, 1); } return; } /* * Do not follow the call sequence until OOM condition is * cleared. */ vmd->vmd_oom_seq = 0; if (vmd->vmd_oom) return; vmd->vmd_oom = TRUE; old_vote = atomic_fetchadd_int(&vm_pageout_oom_vote, 1); if (old_vote != vm_ndomains - 1) return; /* * The current pagedaemon thread is the last in the quorum to * start OOM. Initiate the selection and signaling of the * victim. */ vm_pageout_oom(VM_OOM_MEM); /* * After one round of OOM terror, recall our vote. On the * next pass, current pagedaemon would vote again if the low * memory condition is still there, due to vmd_oom being * false. */ vmd->vmd_oom = FALSE; atomic_subtract_int(&vm_pageout_oom_vote, 1); } /* * The OOM killer is the page daemon's action of last resort when * memory allocation requests have been stalled for a prolonged period * of time because it cannot reclaim memory. This function computes * the approximate number of physical pages that could be reclaimed if * the specified address space is destroyed. * * Private, anonymous memory owned by the address space is the * principal resource that we expect to recover after an OOM kill. * Since the physical pages mapped by the address space's COW entries * are typically shared pages, they are unlikely to be released and so * they are not counted. * * To get to the point where the page daemon runs the OOM killer, its * efforts to write-back vnode-backed pages may have stalled. This * could be caused by a memory allocation deadlock in the write path * that might be resolved by an OOM kill. Therefore, physical pages * belonging to vnode-backed objects are counted, because they might * be freed without being written out first if the address space holds * the last reference to an unlinked vnode. * * Similarly, physical pages belonging to OBJT_PHYS objects are * counted because the address space might hold the last reference to * the object. */ static long vm_pageout_oom_pagecount(struct vmspace *vmspace) { vm_map_t map; vm_map_entry_t entry; vm_object_t obj; long res; map = &vmspace->vm_map; KASSERT(!map->system_map, ("system map")); sx_assert(&map->lock, SA_LOCKED); res = 0; for (entry = map->header.next; entry != &map->header; entry = entry->next) { if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) continue; obj = entry->object.vm_object; if (obj == NULL) continue; if ((entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0 && obj->ref_count != 1) continue; switch (obj->type) { case OBJT_DEFAULT: case OBJT_SWAP: case OBJT_PHYS: case OBJT_VNODE: res += obj->resident_page_count; break; } } return (res); } void vm_pageout_oom(int shortage) { struct proc *p, *bigproc; vm_offset_t size, bigsize; struct thread *td; struct vmspace *vm; /* * We keep the process bigproc locked once we find it to keep anyone * from messing with it; however, there is a possibility of * deadlock if process B is bigproc and one of it's child processes * attempts to propagate a signal to B while we are waiting for A's * lock while walking this list. To avoid this, we don't block on * the process lock but just skip a process if it is already locked. */ bigproc = NULL; bigsize = 0; sx_slock(&allproc_lock); FOREACH_PROC_IN_SYSTEM(p) { int breakout; PROC_LOCK(p); /* * If this is a system, protected or killed process, skip it. */ if (p->p_state != PRS_NORMAL || (p->p_flag & (P_INEXEC | P_PROTECTED | P_SYSTEM | P_WEXIT)) != 0 || p->p_pid == 1 || P_KILLED(p) || (p->p_pid < 48 && swap_pager_avail != 0)) { PROC_UNLOCK(p); continue; } /* * If the process is in a non-running type state, * don't touch it. Check all the threads individually. */ breakout = 0; FOREACH_THREAD_IN_PROC(p, td) { thread_lock(td); if (!TD_ON_RUNQ(td) && !TD_IS_RUNNING(td) && !TD_IS_SLEEPING(td) && !TD_IS_SUSPENDED(td) && !TD_IS_SWAPPED(td)) { thread_unlock(td); breakout = 1; break; } thread_unlock(td); } if (breakout) { PROC_UNLOCK(p); continue; } /* * get the process size */ vm = vmspace_acquire_ref(p); if (vm == NULL) { PROC_UNLOCK(p); continue; } _PHOLD_LITE(p); PROC_UNLOCK(p); sx_sunlock(&allproc_lock); if (!vm_map_trylock_read(&vm->vm_map)) { vmspace_free(vm); sx_slock(&allproc_lock); PRELE(p); continue; } size = vmspace_swap_count(vm); if (shortage == VM_OOM_MEM) size += vm_pageout_oom_pagecount(vm); vm_map_unlock_read(&vm->vm_map); vmspace_free(vm); sx_slock(&allproc_lock); /* * If this process is bigger than the biggest one, * remember it. */ if (size > bigsize) { if (bigproc != NULL) PRELE(bigproc); bigproc = p; bigsize = size; } else { PRELE(p); } } sx_sunlock(&allproc_lock); if (bigproc != NULL) { if (vm_panic_on_oom != 0) panic("out of swap space"); PROC_LOCK(bigproc); killproc(bigproc, "out of swap space"); sched_nice(bigproc, PRIO_MIN); _PRELE(bigproc); PROC_UNLOCK(bigproc); wakeup(&vm_cnt.v_free_count); } } static void vm_pageout_worker(void *arg) { struct vm_domain *domain; int domidx, pass; bool target_met; domidx = (uintptr_t)arg; domain = &vm_dom[domidx]; pass = 0; target_met = true; /* * XXXKIB It could be useful to bind pageout daemon threads to * the cores belonging to the domain, from which vm_page_array * is allocated. */ KASSERT(domain->vmd_segs != 0, ("domain without segments")); domain->vmd_last_active_scan = ticks; vm_pageout_init_marker(&domain->vmd_marker, PQ_INACTIVE); vm_pageout_init_marker(&domain->vmd_inacthead, PQ_INACTIVE); TAILQ_INSERT_HEAD(&domain->vmd_pagequeues[PQ_INACTIVE].pq_pl, &domain->vmd_inacthead, plinks.q); /* * The pageout daemon worker is never done, so loop forever. */ while (TRUE) { mtx_lock(&vm_page_queue_free_mtx); /* * Generally, after a level >= 1 scan, if there are enough * free pages to wakeup the waiters, then they are already * awake. A call to vm_page_free() during the scan awakened * them. However, in the following case, this wakeup serves * to bound the amount of time that a thread might wait. * Suppose a thread's call to vm_page_alloc() fails, but * before that thread calls VM_WAIT, enough pages are freed by * other threads to alleviate the free page shortage. The * thread will, nonetheless, wait until another page is freed * or this wakeup is performed. */ if (vm_pages_needed && !vm_page_count_min()) { vm_pages_needed = false; wakeup(&vm_cnt.v_free_count); } /* * Do not clear vm_pageout_wanted until we reach our free page * target. Otherwise, we may be awakened over and over again, * wasting CPU time. */ if (vm_pageout_wanted && target_met) vm_pageout_wanted = false; /* * Might the page daemon receive a wakeup call? */ if (vm_pageout_wanted) { /* * No. Either vm_pageout_wanted was set by another * thread during the previous scan, which must have * been a level 0 scan, or vm_pageout_wanted was * already set and the scan failed to free enough * pages. If we haven't yet performed a level >= 2 * scan (unlimited dirty cleaning), then upgrade the * level and scan again now. Otherwise, sleep a bit * and try again later. */ mtx_unlock(&vm_page_queue_free_mtx); if (pass > 1) pause("psleep", hz / 2); pass++; } else { /* * Yes. Sleep until pages need to be reclaimed or * have their reference stats updated. */ if (mtx_sleep(&vm_pageout_wanted, &vm_page_queue_free_mtx, PDROP | PVM, "psleep", hz) == 0) { PCPU_INC(cnt.v_pdwakeups); pass = 1; } else pass = 0; } target_met = vm_pageout_scan(domain, pass); } } /* * vm_pageout_init initialises basic pageout daemon settings. */ static void vm_pageout_init(void) { /* * Initialize some paging parameters. */ vm_cnt.v_interrupt_free_min = 2; if (vm_cnt.v_page_count < 2000) vm_pageout_page_count = 8; /* * v_free_reserved needs to include enough for the largest * swap pager structures plus enough for any pv_entry structs * when paging. */ if (vm_cnt.v_page_count > 1024) vm_cnt.v_free_min = 4 + (vm_cnt.v_page_count - 1024) / 200; else vm_cnt.v_free_min = 4; vm_cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + vm_cnt.v_interrupt_free_min; vm_cnt.v_free_reserved = vm_pageout_page_count + vm_cnt.v_pageout_free_min + (vm_cnt.v_page_count / 768); vm_cnt.v_free_severe = vm_cnt.v_free_min / 2; vm_cnt.v_free_target = 4 * vm_cnt.v_free_min + vm_cnt.v_free_reserved; vm_cnt.v_free_min += vm_cnt.v_free_reserved; vm_cnt.v_free_severe += vm_cnt.v_free_reserved; vm_cnt.v_inactive_target = (3 * vm_cnt.v_free_target) / 2; if (vm_cnt.v_inactive_target > vm_cnt.v_free_count / 3) vm_cnt.v_inactive_target = vm_cnt.v_free_count / 3; /* * Set the default wakeup threshold to be 10% above the minimum * page limit. This keeps the steady state out of shortfall. */ vm_pageout_wakeup_thresh = (vm_cnt.v_free_min / 10) * 11; /* * Set interval in seconds for active scan. We want to visit each * page at least once every ten minutes. This is to prevent worst * case paging behaviors with stale active LRU. */ if (vm_pageout_update_period == 0) vm_pageout_update_period = 600; /* XXX does not really belong here */ if (vm_page_max_wired == 0) vm_page_max_wired = vm_cnt.v_free_count / 3; } /* * vm_pageout is the high level pageout daemon. */ static void vm_pageout(void) { int error; #ifdef VM_NUMA_ALLOC int i; #endif swap_pager_swap_init(); #ifdef VM_NUMA_ALLOC for (i = 1; i < vm_ndomains; i++) { error = kthread_add(vm_pageout_worker, (void *)(uintptr_t)i, curproc, NULL, 0, 0, "dom%d", i); if (error != 0) { panic("starting pageout for domain %d, error %d\n", i, error); } } #endif error = kthread_add(uma_reclaim_worker, NULL, curproc, NULL, 0, 0, "uma"); if (error != 0) panic("starting uma_reclaim helper, error %d\n", error); vm_pageout_worker((void *)(uintptr_t)0); } /* * Unless the free page queue lock is held by the caller, this function * should be regarded as advisory. Specifically, the caller should * not msleep() on &vm_cnt.v_free_count following this function unless * the free page queue lock is held until the msleep() is performed. */ void pagedaemon_wakeup(void) { if (!vm_pageout_wanted && curthread->td_proc != pageproc) { vm_pageout_wanted = true; wakeup(&vm_pageout_wanted); } } #if !defined(NO_SWAPPING) static void vm_req_vmdaemon(int req) { static int lastrun = 0; mtx_lock(&vm_daemon_mtx); vm_pageout_req_swapout |= req; if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { wakeup(&vm_daemon_needed); lastrun = ticks; } mtx_unlock(&vm_daemon_mtx); } static void vm_daemon(void) { struct rlimit rsslim; struct proc *p; struct thread *td; struct vmspace *vm; int breakout, swapout_flags, tryagain, attempts; #ifdef RACCT uint64_t rsize, ravailable; #endif while (TRUE) { mtx_lock(&vm_daemon_mtx); msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", #ifdef RACCT racct_enable ? hz : 0 #else 0 #endif ); swapout_flags = vm_pageout_req_swapout; vm_pageout_req_swapout = 0; mtx_unlock(&vm_daemon_mtx); if (swapout_flags) swapout_procs(swapout_flags); /* * scan the processes for exceeding their rlimits or if * process is swapped out -- deactivate pages */ tryagain = 0; attempts = 0; again: attempts++; sx_slock(&allproc_lock); FOREACH_PROC_IN_SYSTEM(p) { vm_pindex_t limit, size; /* * if this is a system process or if we have already * looked at this process, skip it. */ PROC_LOCK(p); if (p->p_state != PRS_NORMAL || p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) { PROC_UNLOCK(p); continue; } /* * if the process is in a non-running type state, * don't touch it. */ breakout = 0; FOREACH_THREAD_IN_PROC(p, td) { thread_lock(td); if (!TD_ON_RUNQ(td) && !TD_IS_RUNNING(td) && !TD_IS_SLEEPING(td) && !TD_IS_SUSPENDED(td)) { thread_unlock(td); breakout = 1; break; } thread_unlock(td); } if (breakout) { PROC_UNLOCK(p); continue; } /* * get a limit */ lim_rlimit_proc(p, RLIMIT_RSS, &rsslim); limit = OFF_TO_IDX( qmin(rsslim.rlim_cur, rsslim.rlim_max)); /* * let processes that are swapped out really be * swapped out set the limit to nothing (will force a * swap-out.) */ if ((p->p_flag & P_INMEM) == 0) limit = 0; /* XXX */ vm = vmspace_acquire_ref(p); _PHOLD_LITE(p); PROC_UNLOCK(p); if (vm == NULL) { PRELE(p); continue; } sx_sunlock(&allproc_lock); size = vmspace_resident_count(vm); if (size >= limit) { vm_pageout_map_deactivate_pages( &vm->vm_map, limit); + size = vmspace_resident_count(vm); } #ifdef RACCT if (racct_enable) { rsize = IDX_TO_OFF(size); PROC_LOCK(p); - racct_set(p, RACCT_RSS, rsize); + if (p->p_state == PRS_NORMAL) + racct_set(p, RACCT_RSS, rsize); ravailable = racct_get_available(p, RACCT_RSS); PROC_UNLOCK(p); if (rsize > ravailable) { /* * Don't be overly aggressive; this * might be an innocent process, * and the limit could've been exceeded * by some memory hog. Don't try * to deactivate more than 1/4th * of process' resident set size. */ if (attempts <= 8) { if (ravailable < rsize - (rsize / 4)) { ravailable = rsize - (rsize / 4); } } vm_pageout_map_deactivate_pages( &vm->vm_map, OFF_TO_IDX(ravailable)); /* Update RSS usage after paging out. */ size = vmspace_resident_count(vm); rsize = IDX_TO_OFF(size); PROC_LOCK(p); - racct_set(p, RACCT_RSS, rsize); + if (p->p_state == PRS_NORMAL) + racct_set(p, RACCT_RSS, rsize); PROC_UNLOCK(p); if (rsize > ravailable) tryagain = 1; } } #endif vmspace_free(vm); sx_slock(&allproc_lock); PRELE(p); } sx_sunlock(&allproc_lock); if (tryagain != 0 && attempts <= 10) goto again; } } #endif /* !defined(NO_SWAPPING) */ Index: stable/11 =================================================================== --- stable/11 (revision 314346) +++ stable/11 (revision 314347) Property changes on: stable/11 ___________________________________________________________________ Modified: svn:mergeinfo ## -0,0 +0,1 ## Merged /head:r313730