Add CPU percentage limit enforcement to RCTL. The resouce name is "pcpu".

It was implemented by Rudolf Tomori during Google Summer of Code 2012.

4 files changed, 563 insertions, 29 deletions

diff --git a/sys/kern/kern_racct.c b/sys/kern/kern_racct.c
index 534780f..df1c457 100644
--- a/sys/kern/kern_racct.c
+++ b/sys/kern/kern_racct.c
@@ -33,6 +33,7 @@
 __FBSDID("$FreeBSD$");
 
 #include "opt_kdtrace.h"
+#include "opt_sched.h"
 
 #include <sys/param.h>
 #include <sys/systm.h>
@@ -50,10 +51,13 @@ __FBSDID("$FreeBSD$");
 #include <sys/sbuf.h>
 #include <sys/sched.h>
 #include <sys/sdt.h>
+#include <sys/smp.h>
 #include <sys/sx.h>
+#include <sys/sysctl.h>
 #include <sys/sysent.h>
 #include <sys/sysproto.h>
 #include <sys/umtx.h>
+#include <machine/smp.h>
 
 #ifdef RCTL
 #include <sys/rctl.h>
@@ -63,6 +67,23 @@ __FBSDID("$FreeBSD$");
 
 FEATURE(racct, "Resource Accounting");
 
+/*
+ * Do not block processes that have their %cpu usage <= pcpu_threshold.
+ */
+static int pcpu_threshold = 1;
+
+SYSCTL_NODE(_kern, OID_AUTO, racct, CTLFLAG_RW, 0, "Resource Accounting");
+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
+
 static struct mtx racct_lock;
 MTX_SYSINIT(racct_lock, &racct_lock, "racct lock", MTX_DEF);
 
@@ -140,7 +161,217 @@ int racct_types[] = {
 	[RACCT_SHMSIZE] =
 		RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
 	[RACCT_WALLCLOCK] =
-		RACCT_IN_MILLIONS };
+		RACCT_IN_MILLIONS,
+	[RACCT_PCTCPU] =
+		RACCT_DECAYING | RACCT_DENIABLE | RACCT_IN_MILLIONS };
+
+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;
+
+	/*
+	 * 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)
@@ -180,7 +411,7 @@ racct_sub_racct(struct racct *dest, const struct racct *src)
 			KASSERT(src->r_resources[i] <= dest->r_resources[i],
 			    ("racct propagation meltdown: src > dest"));
 		}
-		if (RACCT_IS_RECLAIMABLE(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),
@@ -254,10 +485,23 @@ racct_alloc_resource(struct racct *racct, int resource,
 
 	racct->r_resources[resource] += amount;
 	if (racct->r_resources[resource] < 0) {
-		KASSERT(RACCT_IS_SLOPPY(resource),
+		KASSERT(RACCT_IS_SLOPPY(resource) || RACCT_IS_DECAYING(resource),
 		    ("racct_alloc_resource: usage < 0"));
 		racct->r_resources[resource] = 0;
 	}
+	
+	/*
+	 * There are some cases where the racct %cpu resource would grow
+	 * beyond 100%.
+	 * 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 boundary here to 100%.
+	 */
+	if ((resource == RACCT_PCTCPU) &&
+	    (racct->r_resources[RACCT_PCTCPU] > 100 * 1000000))
+		racct->r_resources[RACCT_PCTCPU] = 100 * 1000000;
 }
 
 static int
@@ -357,7 +601,8 @@ racct_add_force(struct proc *p, int resource, uint64_t amount)
 static int
 racct_set_locked(struct proc *p, int resource, uint64_t amount)
 {
-	int64_t diff;
+	int64_t old_amount, decayed_amount;
+	int64_t diff_proc, diff_cred;
 #ifdef RCTL
 	int error;
 #endif
@@ -369,15 +614,30 @@ racct_set_locked(struct proc *p, int resource, uint64_t amount)
 	 */
 	PROC_LOCK_ASSERT(p, MA_OWNED);
 
-	diff = amount - p->p_racct->r_resources[resource];
+	old_amount = p->p_racct->r_resources[resource];
+	/*
+	 * The diffs may be negative.
+	 */
+	diff_proc = amount - old_amount;
+	if (RACCT_IS_DECAYING(resource)) {
+		/*
+		 * 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 >= 0 || RACCT_IS_RECLAIMABLE(resource),
-	    ("racct_set: usage of non-reclaimable resource %d dropping",
+	KASSERT(diff_proc >= 0 || RACCT_CAN_DROP(resource),
+	    ("racct_set: usage of non-droppable resource %d dropping",
 	     resource));
 #endif
 #ifdef RCTL
-	if (diff > 0) {
-		error = rctl_enforce(p, resource, diff);
+	if (diff_proc > 0) {
+		error = rctl_enforce(p, resource, diff_proc);
 		if (error && RACCT_IS_DENIABLE(resource)) {
 			SDT_PROBE(racct, kernel, rusage, set_failure, p,
 			    resource, amount, 0, 0);
@@ -385,11 +645,11 @@ racct_set_locked(struct proc *p, int resource, uint64_t amount)
 		}
 	}
 #endif
-	racct_alloc_resource(p->p_racct, resource, diff);
-	if (diff > 0)
-		racct_add_cred_locked(p->p_ucred, resource, diff);
-	else if (diff < 0)
-		racct_sub_cred_locked(p->p_ucred, resource, -diff);
+	racct_alloc_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);
 }
@@ -412,10 +672,11 @@ racct_set(struct proc *p, int resource, uint64_t amount)
 	return (error);
 }
 
-void
-racct_set_force(struct proc *p, int resource, uint64_t amount)
+static void
+racct_set_force_locked(struct proc *p, int resource, uint64_t amount)
 {
-	int64_t diff;
+	int64_t old_amount, decayed_amount;
+	int64_t diff_proc, diff_cred;
 
 	SDT_PROBE(racct, kernel, rusage, set, p, resource, amount, 0, 0);
 
@@ -424,13 +685,35 @@ racct_set_force(struct proc *p, int resource, uint64_t amount)
 	 */
 	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 (RACCT_IS_DECAYING(resource)) {
+		/*
+		 * 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;
+
+	racct_alloc_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);
+}
+
+void
+racct_set_force(struct proc *p, int resource, uint64_t amount)
+{
 	mtx_lock(&racct_lock);
-	diff = amount - p->p_racct->r_resources[resource];
-	racct_alloc_resource(p->p_racct, resource, diff);
-	if (diff > 0)
-		racct_add_cred_locked(p->p_ucred, resource, diff);
-	else if (diff < 0)
-		racct_sub_cred_locked(p->p_ucred, resource, -diff);
+	racct_set_force_locked(p, resource, amount);
 	mtx_unlock(&racct_lock);
 }
 
@@ -469,6 +752,22 @@ racct_get_available(struct proc *p, int resource)
 }
 
 /*
+ * 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
+	return (rctl_pcpu_available(p));
+#else
+	return (INT64_MAX);
+#endif
+}
+
+/*
  * Decrease allocation of 'resource' by 'amount' for process 'p'.
  */
 void
@@ -481,8 +780,8 @@ racct_sub(struct proc *p, int resource, uint64_t amount)
 	 * We need proc lock to dereference p->p_ucred.
 	 */
 	PROC_LOCK_ASSERT(p, MA_OWNED);
-	KASSERT(RACCT_IS_RECLAIMABLE(resource),
-	    ("racct_sub: called for non-reclaimable resource %d", resource));
+	KASSERT(RACCT_CAN_DROP(resource),
+	    ("racct_sub: called for non-droppable resource %d", resource));
 
 	mtx_lock(&racct_lock);
 	KASSERT(amount <= p->p_racct->r_resources[resource],
@@ -504,8 +803,8 @@ racct_sub_cred_locked(struct ucred *cred, int resource, uint64_t amount)
 	    0, 0);
 
 #ifdef notyet
-	KASSERT(RACCT_IS_RECLAIMABLE(resource),
-	    ("racct_sub_cred: called for non-reclaimable resource %d",
+	KASSERT(RACCT_CAN_DROP(resource),
+	    ("racct_sub_cred: called for resource %d which can not drop",
 	     resource));
 #endif
 
@@ -551,6 +850,10 @@ racct_proc_fork(struct proc *parent, struct proc *child)
 		goto out;
 #endif
 
+	/* Init process cpu time. */
+	child->p_prev_runtime = 0;
+	child->p_throttled = 0;
+
 	/*
 	 * Inherit resource usage.
 	 */
@@ -602,6 +905,8 @@ racct_proc_exit(struct proc *p)
 {
 	int i;
 	uint64_t runtime;
+	struct timeval wallclock;
+	uint64_t pct_estimate, pct;
 
 	PROC_LOCK(p);
 	/*
@@ -614,8 +919,16 @@ racct_proc_exit(struct proc *p)
 	if (runtime < p->p_prev_runtime)
 		runtime = p->p_prev_runtime;
 #endif
+	microuptime(&wallclock);
+	timevalsub(&wallclock, &p->p_stats->p_start);
+	pct_estimate = (1000000 * runtime * 100) /
+	    ((uint64_t)wallclock.tv_sec * 1000000 +
+	    wallclock.tv_usec);
+	pct = racct_getpcpu(p, pct_estimate);
+
 	mtx_lock(&racct_lock);
 	racct_set_locked(p, RACCT_CPU, runtime);
+	racct_add_cred_locked(p->p_ucred, RACCT_PCTCPU, pct);
 
 	for (i = 0; i <= RACCT_MAX; i++) {
 		if (p->p_racct->r_resources[i] == 0)
@@ -692,23 +1005,124 @@ racct_move(struct racct *dest, struct racct *src)
 }
 
 static void
+racct_proc_throttle(struct proc *p)
+{
+	struct thread *td;
+#ifdef SMP
+	int cpuid;
+#endif
+
+	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_KTHREAD)) != 0) ||
+	    (p->p_racct->r_resources[RACCT_PCTCPU] <= pcpu_threshold))
+		return;
+	p->p_throttled = 1;
+
+	FOREACH_THREAD_IN_PROC(p, 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.
+			 */
+			thread_lock(td);
+			td->td_flags |= TDF_NEEDRESCHED;
+			thread_unlock(td);
+			break;
+		case TDS_RUNNING:
+			/*
+			 * If the thread is running, we request a context
+			 * switch for it by setting the TDF_NEEDRESCHED flag.
+			 */
+			thread_lock(td);
+			td->td_flags |= TDF_NEEDRESCHED;
+#ifdef SMP
+			cpuid = td->td_oncpu;
+			if ((cpuid != NOCPU) && (td != curthread))
+				ipi_cpu(cpuid, IPI_AST);
+#endif
+			thread_unlock(td);
+			break;
+		default:
+			break;
+		}
+	}
+}
+
+static void
+racct_proc_wakeup(struct proc *p)
+{
+	PROC_LOCK_ASSERT(p, MA_OWNED);
+
+	if (p->p_throttled) {
+		p->p_throttled = 0;
+		wakeup(p->p_racct);
+	}
+}
+
+static void
+racct_decay_resource(struct racct *racct, void * res, void* dummy)
+{
+	int resource;
+	int64_t r_old, r_new;
+
+	resource = *(int *)res;
+	r_old = racct->r_resources[resource];
+
+	/* If there is nothing to decay, just exit. */
+	if (r_old <= 0)
+		return;
+
+	mtx_lock(&racct_lock);
+	r_new = r_old * RACCT_DECAY_FACTOR / FSCALE;
+	racct->r_resources[resource] = r_new;
+	mtx_unlock(&racct_lock);
+}
+
+static void
+racct_decay(int resource)
+{
+	ui_racct_foreach(racct_decay_resource, &resource, NULL);
+	loginclass_racct_foreach(racct_decay_resource, &resource, NULL);
+	prison_racct_foreach(racct_decay_resource, &resource, NULL);
+}
+
+static void
 racctd(void)
 {
 	struct thread *td;
 	struct proc *p;
 	struct timeval wallclock;
 	uint64_t runtime;
+	uint64_t pct, pct_estimate;
 
 	for (;;) {
+		racct_decay(RACCT_PCTCPU);
+
 		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) {
-			if (p->p_state != PRS_NORMAL)
+			PROC_LOCK(p);
+			if (p->p_state != PRS_NORMAL) {
+				PROC_UNLOCK(p);
 				continue;
+			}
 
 			microuptime(&wallclock);
 			timevalsub(&wallclock, &p->p_stats->p_start);
-			PROC_LOCK(p);
 			PROC_SLOCK(p);
 			FOREACH_THREAD_IN_PROC(p, td)
 				ruxagg(p, td);
@@ -722,7 +1136,12 @@ racctd(void)
 				runtime = p->p_prev_runtime;
 #endif
 			p->p_prev_runtime = runtime;
+			pct_estimate = (1000000 * runtime * 100) /
+			    ((uint64_t)wallclock.tv_sec * 1000000 +
+			    wallclock.tv_usec);
+			pct = racct_getpcpu(p, pct_estimate);
 			mtx_lock(&racct_lock);
+			racct_set_force_locked(p, RACCT_PCTCPU, pct);
 			racct_set_locked(p, RACCT_CPU, runtime);
 			racct_set_locked(p, RACCT_WALLCLOCK,
 			    (uint64_t)wallclock.tv_sec * 1000000 +
@@ -730,6 +1149,26 @@ racctd(void)
 			mtx_unlock(&racct_lock);
 			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)
+				racct_proc_throttle(p);
+			else if (p->p_throttled)
+				racct_proc_wakeup(p);
+			PROC_UNLOCK(p);
+		}
 		sx_sunlock(&allproc_lock);
 		pause("-", hz);
 	}
diff --git a/sys/kern/kern_rctl.c b/sys/kern/kern_rctl.c
index 1d0a440..934327a 100644
--- a/sys/kern/kern_rctl.c
+++ b/sys/kern/kern_rctl.c
@@ -76,6 +76,8 @@ FEATURE(rctl, "Resource Limits");
 #define	RCTL_MAX_INBUFLEN	4096
 #define	RCTL_LOG_BUFSIZE	128
 
+#define	RCTL_PCPU_SHIFT		(10 * 1000000)
+
 /*
  * 'rctl_rule_link' connects a rule with every racct it's related to.
  * For example, rule 'user:X:openfiles:deny=N/process' is linked
@@ -120,6 +122,7 @@ static struct dict resourcenames[] = {
 	{ "nshm", RACCT_NSHM },
 	{ "shmsize", RACCT_SHMSIZE },
 	{ "wallclock", RACCT_WALLCLOCK },
+	{ "pcpu", RACCT_PCTCPU },
 	{ NULL, -1 }};
 
 static struct dict actionnames[] = {
@@ -271,6 +274,51 @@ rctl_would_exceed(const struct proc *p, const struct rctl_rule *rule,
 }
 
 /*
+ * Special version of rctl_available() function for the %cpu resource.
+ * We slightly cheat here and return less than we normally would.
+ */
+int64_t
+rctl_pcpu_available(const struct proc *p) {
+	struct rctl_rule *rule;
+	struct rctl_rule_link *link;
+	int64_t available, minavailable, limit;
+
+	minavailable = INT64_MAX;
+	limit = 0;
+
+	rw_rlock(&rctl_lock);
+
+	LIST_FOREACH(link, &p->p_racct->r_rule_links, rrl_next) {
+		rule = link->rrl_rule;
+		if (rule->rr_resource != RACCT_PCTCPU)
+			continue;
+		if (rule->rr_action != RCTL_ACTION_DENY)
+			continue;
+		available = rctl_available_resource(p, rule);
+		if (available < minavailable) {
+			minavailable = available;
+			limit = rule->rr_amount;
+		}
+	}
+
+	rw_runlock(&rctl_lock);
+
+	/*
+	 * Return slightly less than actual value of the available
+	 * %cpu resource.  This makes %cpu throttling more agressive
+	 * and lets us act sooner than the limits are already exceeded.
+	 */
+	if (limit != 0) {
+		if (limit > 2 * RCTL_PCPU_SHIFT)
+			minavailable -= RCTL_PCPU_SHIFT;
+		else
+			minavailable -= (limit / 2);
+	}
+
+	return (minavailable);
+}
+
+/*
  * Check whether the proc 'p' can allocate 'amount' of 'resource' in addition
  * to what it keeps allocated now.  Returns non-zero if the allocation should
  * be denied, 0 otherwise.
diff --git a/sys/kern/sched_4bsd.c b/sys/kern/sched_4bsd.c
index 7aaf37c..d1595c7 100644
--- a/sys/kern/sched_4bsd.c
+++ b/sys/kern/sched_4bsd.c
@@ -1585,6 +1585,40 @@ sched_pctcpu(struct thread *td)
 	return (ts->ts_pctcpu);
 }
 
+#ifdef	RACCT
+/*
+ * Calculates the contribution to the thread cpu usage for the latest
+ * (unfinished) second.
+ */
+fixpt_t
+sched_pctcpu_delta(struct thread *td)
+{
+	struct td_sched *ts;
+	fixpt_t delta;
+	int realstathz;
+
+	THREAD_LOCK_ASSERT(td, MA_OWNED);
+	ts = td->td_sched;
+	delta = 0;
+	realstathz = stathz ? stathz : hz;
+	if (ts->ts_cpticks != 0) {
+#if	(FSHIFT >= CCPU_SHIFT)
+		delta = (realstathz == 100)
+		    ? ((fixpt_t) ts->ts_cpticks) <<
+		    (FSHIFT - CCPU_SHIFT) :
+		    100 * (((fixpt_t) ts->ts_cpticks)
+		    << (FSHIFT - CCPU_SHIFT)) / realstathz;
+#else
+		delta = ((FSCALE - ccpu) *
+		    (ts->ts_cpticks *
+		    FSCALE / realstathz)) >> FSHIFT;
+#endif
+	}
+
+	return (delta);
+}
+#endif
+
 void
 sched_tick(int cnt)
 {
diff --git a/sys/kern/subr_trap.c b/sys/kern/subr_trap.c
index 66e3201..70df8a1 100644
--- a/sys/kern/subr_trap.c
+++ b/sys/kern/subr_trap.c
@@ -100,6 +100,9 @@ void
 userret(struct thread *td, struct trapframe *frame)
 {
 	struct proc *p = td->td_proc;
+#ifdef	RACCT
+	int sig;
+#endif
 
 	CTR3(KTR_SYSC, "userret: thread %p (pid %d, %s)", td, p->p_pid,
             td->td_name);
@@ -163,6 +166,16 @@ userret(struct thread *td, struct trapframe *frame)
 	    __func__, td, p->p_pid, td->td_name, curvnet,
 	    (td->td_vnet_lpush != NULL) ? td->td_vnet_lpush : "N/A"));
 #endif
+#ifdef	RACCT
+	PROC_LOCK(p);
+	while (p->p_throttled == 1) {
+		sig = msleep(p->p_racct, &p->p_mtx, PCATCH | PBDRY, "racct",
+		    hz);
+		if ((sig == EINTR) || (sig == ERESTART))
+			break;
+	}
+	PROC_UNLOCK(p);
+#endif
 }
 
 /*