D5080.id12953.diff
No OneTemporary
Actions

Size

43 KB

Referenced Files

None

Subscribers

None

D5080.id12953.diff
View Options

	Index: sys/cddl/contrib/opensolaris/uts/common/fs/zfs/arc.c
	===================================================================
	--- sys/cddl/contrib/opensolaris/uts/common/fs/zfs/arc.c
	+++ sys/cddl/contrib/opensolaris/uts/common/fs/zfs/arc.c
	@@ -132,6 +132,7 @@
	#include <sys/multilist.h>
	#ifdef _KERNEL
	#include <sys/dnlc.h>
	+#include <sys/racct.h>
	#endif
	#include <sys/callb.h>
	#include <sys/kstat.h>
	@@ -4503,6 +4504,18 @@
	demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
	data, metadata, misses);
	#ifdef _KERNEL
	+#ifdef RACCT
	+ if (racct_enable) {
	+#if 0
	+ printf("%s: adding %jd bytes for %d (%s)\n",
	+ __func__, (uintmax_t)size, curproc->p_pid, curproc->p_comm);
	+#endif
	+ PROC_LOCK(curproc);
	+ racct_add_force(curproc, RACCT_READBPS, size);
	+ racct_add_force(curproc, RACCT_READIOPS, 1);
	+ PROC_UNLOCK(curproc);
	+ }
	+#endif /* RACCT */
	curthread->td_ru.ru_inblock++;
	#endif

	Index: sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dmu.c
	===================================================================
	--- sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dmu.c
	+++ sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dmu.c
	@@ -47,6 +47,7 @@
	#include <sys/sa.h>
	#include <sys/zfeature.h>
	#ifdef _KERNEL
	+#include <sys/racct.h>
	#include <sys/vm.h>
	#include <sys/zfs_znode.h>
	#endif
	@@ -427,6 +428,15 @@
	}
	dbp = kmem_zalloc(sizeof (dmu_buf_t ) nblks, KM_SLEEP);

	+#if defined(_KERNEL) && defined(RACCT)
	+ if (racct_enable && !read) {
	+ PROC_LOCK(curproc);
	+ racct_add_force(curproc, RACCT_WRITEBPS, length);
	+ racct_add_force(curproc, RACCT_WRITEIOPS, nblks);
	+ PROC_UNLOCK(curproc);
	+ }
	+#endif
	+
	zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
	blkid = dbuf_whichblock(dn, 0, offset);
	for (i = 0; i < nblks; i++) {
	@@ -1422,7 +1432,15 @@
	DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA) {
	#ifdef _KERNEL
	curthread->td_ru.ru_oublock++;
	-#endif
	+#ifdef RACCT
	+ if (racct_enable) {
	+ PROC_LOCK(curproc);
	+ racct_add_force(curproc, RACCT_WRITEBPS, blksz);
	+ racct_add_force(curproc, RACCT_WRITEIOPS, 1);
	+ PROC_UNLOCK(curproc);
	+ }
	+#endif /* RACCT */
	+#endif /* _KERNEL */
	dbuf_assign_arcbuf(db, buf, tx);
	dbuf_rele(db, FTAG);
	} else {
	Index: sys/fs/ext2fs/ext2_bmap.c
	===================================================================
	--- sys/fs/ext2fs/ext2_bmap.c
	+++ sys/fs/ext2fs/ext2_bmap.c
	@@ -42,6 +42,7 @@
	#include <sys/proc.h>
	#include <sys/vnode.h>
	#include <sys/mount.h>
	+#include <sys/racct.h>
	#include <sys/resourcevar.h>
	#include <sys/stat.h>

	@@ -243,6 +244,13 @@
	vfs_busy_pages(bp, 0);
	bp->b_iooffset = dbtob(bp->b_blkno);
	bstrategy(bp);
	+#ifdef RACCT
	+ if (racct_enable) {
	+ PROC_LOCK(curproc);
	+ racct_add_buf(curproc, bp, 0);
	+ PROC_UNLOCK(curproc);
	+ }
	+#endif
	curthread->td_ru.ru_inblock++;
	error = bufwait(bp);
	if (error) {
	Index: sys/kern/kern_physio.c
	===================================================================
	--- sys/kern/kern_physio.c
	+++ sys/kern/kern_physio.c
	@@ -27,6 +27,7 @@
	#include <sys/conf.h>
	#include <sys/malloc.h>
	#include <sys/proc.h>
	+#include <sys/racct.h>
	#include <sys/uio.h>
	#include <geom/geom.h>

	@@ -109,6 +110,22 @@
	prot \|= VM_PROT_WRITE; /* Less backwards than it looks */
	error = 0;
	for (i = 0; i < uio->uio_iovcnt; i++) {
	+#ifdef RACCT
	+ if (racct_enable) {
	+ PROC_LOCK(curproc);
	+ if (uio->uio_rw == UIO_READ) {
	+ racct_add_force(curproc, RACCT_READBPS,
	+ uio->uio_iov[i].iov_len);
	+ racct_add_force(curproc, RACCT_READIOPS, 1);
	+ } else {
	+ racct_add_force(curproc, RACCT_WRITEBPS,
	+ uio->uio_iov[i].iov_len);
	+ racct_add_force(curproc, RACCT_WRITEIOPS, 1);
	+ }
	+ PROC_UNLOCK(curproc);
	+ }
	+#endif /* RACCT */
	+
	while (uio->uio_iov[i].iov_len) {
	bzero(bp, sizeof(*bp));
	if (uio->uio_rw == UIO_READ) {
	Index: sys/kern/kern_racct.c
	===================================================================
	--- sys/kern/kern_racct.c
	+++ sys/kern/kern_racct.c
	@@ -35,6 +35,7 @@
	#include "opt_sched.h"

	#include <sys/param.h>
	+#include <sys/buf.h>
	#include <sys/systm.h>
	#include <sys/eventhandler.h>
	#include <sys/jail.h>
	@@ -66,21 +67,25 @@

	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
	+/*
	+ * Do not block processes that have their %cpu usage <= pcpu_threshold.
	+ */
	+static int pcpu_threshold = 1;
	+int racct_pctcpu_throttle_not_deny = 0;

	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.");
	+ 0, "\"pcpu:deny\" rules don't affect processes with %cpu below this value");
	+SYSCTL_UINT(_kern_racct, OID_AUTO, pcpu_throttle_not_deny, CTLFLAG_RD,
	+ &racct_pctcpu_throttle_not_deny,
	+ 0, "Use \"pcpu:throttle\" rules instead of \"pcpu:deny\"");

	/*
	* How many seconds it takes to use the scheduler %cpu calculations. When a
	@@ -90,7 +95,7 @@
	*/
	#define RACCT_PCPU_SECS 3

	-static struct mtx racct_lock;
	+struct mtx racct_lock;
	MTX_SYSINIT(racct_lock, &racct_lock, "racct lock", MTX_DEF);

	static uma_zone_t racct_zone;
	@@ -171,7 +176,16 @@
	[RACCT_WALLCLOCK] =
	RACCT_IN_MILLIONS,
	[RACCT_PCTCPU] =
	- RACCT_DECAYING \| RACCT_DENIABLE \| RACCT_IN_MILLIONS };
	+ RACCT_RECLAIMABLE \| 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;

	@@ -473,6 +487,8 @@
	for (i = 0; i <= RACCT_MAX; i++) {
	if (RACCT_IS_SLOPPY(i))
	continue;
	+ if (RACCT_IS_DECAYING(i))
	+ continue;
	if (!RACCT_IS_RECLAIMABLE(i))
	continue;
	KASSERT(racct->r_resources[i] == 0,
	@@ -498,8 +514,8 @@

	/*
	* Increase consumption of 'resource' by 'amount' for 'racct',
	- * but not its parents. Differently from other cases, 'amount' here
	- * may be less than zero.
	+ * 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,
	@@ -526,9 +542,10 @@
	* 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))
	+ if (!racct_pctcpu_throttle_not_deny && 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
	@@ -612,16 +629,11 @@
	mtx_unlock(&racct_lock);
	}

	-/*
	- * 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)
	+static void
	+racct_add_force_locked(struct proc *p, int resource, uint64_t amount)
	{

	- if (!racct_enable)
	- return;
	+ ASSERT_RACCT_ENABLED();

	SDT_PROBE3(racct, , rusage, add__force, p, resource, amount);

	@@ -630,12 +642,48 @@
	*/
	PROC_LOCK_ASSERT(p, MA_OWNED);

	- mtx_lock(&racct_lock);
	+ /*
	+ * Ignore the return value - we can't return error, but RCTL might
	+ * eg. throttle the process.
	+ */
	+ (void)rctl_enforce(p, resource, amount);
	+
	racct_adjust_resource(p->p_racct, resource, amount);
	racct_add_cred_locked(p->p_ucred, resource, amount);
	+}
	+
	+/*
	+ * Increase allocation of 'resource' by 'amount' for process 'p'.
	+ * Checks for limits, but never fails.
	+ */
	+void
	+racct_add_force(struct proc *p, int resource, uint64_t amount)
	+{
	+
	+ if (!racct_enable)
	+ return;
	+
	+ mtx_lock(&racct_lock);
	+ racct_add_force_locked(p, resource, amount);
	mtx_unlock(&racct_lock);
	}

	+void
	+racct_add_buf(struct proc p, const struct buf bp, int is_write)
	+{
	+
	+ ASSERT_RACCT_ENABLED();
	+ PROC_LOCK_ASSERT(p, MA_OWNED);
	+
	+ if (is_write) {
	+ racct_add_force(curproc, RACCT_WRITEBPS, bp->b_bcount);
	+ racct_add_force(curproc, RACCT_WRITEIOPS, 1);
	+ } else {
	+ racct_add_force(curproc, RACCT_READBPS, bp->b_bcount);
	+ racct_add_force(curproc, RACCT_READIOPS, 1);
	+ }
	+}
	+
	static int
	racct_set_locked(struct proc *p, int resource, uint64_t amount)
	{
	@@ -659,7 +707,7 @@
	* The diffs may be negative.
	*/
	diff_proc = amount - old_amount;
	- if (RACCT_IS_DECAYING(resource)) {
	+ if (!racct_pctcpu_throttle_not_deny && RACCT_IS_DECAYING(resource)) {
	/*
	* Resources in per-credential racct containers may decay.
	* If this is the case, we need to calculate the difference
	@@ -688,7 +736,7 @@
	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)
	+ else if (RACCT_IS_RECLAIMABLE(resource) && diff_cred < 0)
	racct_sub_cred_locked(p->p_ucred, resource, -diff_cred);

	return (0);
	@@ -735,7 +783,7 @@
	* The diffs may be negative.
	*/
	diff_proc = amount - old_amount;
	- if (RACCT_IS_DECAYING(resource)) {
	+ if (!racct_pctcpu_throttle_not_deny && RACCT_IS_DECAYING(resource)) {
	/*
	* Resources in per-credential racct containers may decay.
	* If this is the case, we need to calculate the difference
	@@ -750,7 +798,7 @@
	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)
	+ else if (RACCT_IS_RECLAIMABLE(resource) && diff_cred < 0)
	racct_sub_cred_locked(p->p_ucred, resource, -diff_cred);
	}

	@@ -850,7 +898,8 @@
	(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);
	+ if (RACCT_IS_RECLAIMABLE(resource))
	+ racct_sub_cred_locked(p->p_ucred, resource, amount);
	mtx_unlock(&racct_lock);
	}

	@@ -994,13 +1043,17 @@
	#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);
	+ if (!racct_pctcpu_throttle_not_deny) {
	+ 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);
	+ } else {
	+ pct = (runtime - p->p_prev_runtime) * 100;
	+ }

	mtx_lock(&racct_lock);
	racct_set_locked(p, RACCT_CPU, runtime);
	@@ -1085,14 +1138,15 @@
	mtx_unlock(&racct_lock);
	}

	-static void
	-racct_proc_throttle(struct proc *p)
	+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);

	@@ -1100,10 +1154,13 @@
	* 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))
	+ if ((p->p_flag & (P_SYSTEM \| P_KTHREAD)) != 0)
	+ return;
	+
	+ if (p->p_throttled < 0 \|\| (timeout > 0 && p->p_throttled > timeout))
	return;
	- p->p_throttled = 1;
	+
	+ p->p_throttled = timeout;

	FOREACH_THREAD_IN_PROC(p, td) {
	thread_lock(td);
	@@ -1144,30 +1201,37 @@

	PROC_LOCK_ASSERT(p, MA_OWNED);

	- if (p->p_throttled) {
	+ if (p->p_throttled != 0) {
	p->p_throttled = 0;
	wakeup(p->p_racct);
	}
	}

	static void
	-racct_decay_resource(struct racct racct, void res, void* dummy)
	+racct_decay_callback(struct racct racct, void res, void *dummy)
	{
	- int resource;
	int64_t r_old, r_new;

	ASSERT_RACCT_ENABLED();
	mtx_assert(&racct_lock, MA_OWNED);

	- resource = (int )res;
	- r_old = racct->r_resources[resource];
	+ rctl_throttle_decay(racct, RACCT_READBPS);
	+ rctl_throttle_decay(racct, RACCT_WRITEBPS);
	+ rctl_throttle_decay(racct, RACCT_READIOPS);
	+ rctl_throttle_decay(racct, RACCT_WRITEIOPS);
	+
	+ if (racct_pctcpu_throttle_not_deny) {
	+ rctl_throttle_decay(racct, RACCT_PCTCPU);
	+ } else {
	+ r_old = racct->r_resources[RACCT_PCTCPU];
	+
	+ /* If there is nothing to decay, just exit. */
	+ if (r_old <= 0)
	+ return;

	- /* If there is nothing to decay, just exit. */
	- if (r_old <= 0)
	- return;
	-
	- r_new = r_old * RACCT_DECAY_FACTOR / FSCALE;
	- racct->r_resources[resource] = r_new;
	+ r_new = r_old * RACCT_DECAY_FACTOR / FSCALE;
	+ racct->r_resources[RACCT_PCTCPU] = r_new;
	+ }
	}

	static void
	@@ -1185,17 +1249,17 @@
	}

	static void
	-racct_decay(int resource)
	+racct_decay(void)
	{

	ASSERT_RACCT_ENABLED();

	- ui_racct_foreach(racct_decay_resource, racct_decay_pre,
	- racct_decay_post, &resource, NULL);
	- loginclass_racct_foreach(racct_decay_resource, racct_decay_pre,
	- racct_decay_post, &resource, NULL);
	- prison_racct_foreach(racct_decay_resource, racct_decay_pre,
	- racct_decay_post, &resource, NULL);
	+ 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
	@@ -1206,18 +1270,22 @@
	struct timeval wallclock;
	uint64_t runtime;
	uint64_t pct, pct_estimate;
	+ bool throttle;
	+ //int error;

	ASSERT_RACCT_ENABLED();

	for (;;) {
	- racct_decay(RACCT_PCTCPU);
	+ racct_decay();

	sx_slock(&allproc_lock);

	- LIST_FOREACH(p, &zombproc, p_list) {
	- PROC_LOCK(p);
	- racct_set(p, RACCT_PCTCPU, 0);
	- PROC_UNLOCK(p);
	+ if (!racct_pctcpu_throttle_not_deny) {
	+ LIST_FOREACH(p, &zombproc, p_list) {
	+ PROC_LOCK(p);
	+ racct_set(p, RACCT_PCTCPU, 0);
	+ PROC_UNLOCK(p);
	+ }
	}

	FOREACH_PROC_IN_SYSTEM(p) {
	@@ -1241,24 +1309,63 @@
	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);
	+ if (!racct_pctcpu_throttle_not_deny) {
	+ 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);
	+ } else {
	+ pct = (runtime - p->p_prev_runtime) * 100;
	+ }
	mtx_lock(&racct_lock);
	- racct_set_force_locked(p, RACCT_PCTCPU, pct);
	+
	+ 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);
	+ if (racct_pctcpu_throttle_not_deny) {
	+ rctl_throttle_decay(p->p_racct, RACCT_PCTCPU);
	+ /*
	+ * Preferably we would bump the counters
	+ * from statclock. We cannot, however - to call
	+ * racct_add() we need the proc lock, and we
	+ * cannot acquire that in statclock().
	+ *
	+ * Note that 1% basically means 10 milliseconds.
	+ */
	+ racct_add_force_locked(p, RACCT_PCTCPU, pct);
	+ } else {
	+ racct_set_force_locked(p, RACCT_PCTCPU, pct);
	+ }
	+ p->p_prev_runtime = runtime;
	+
	+#if 0
	+ if (racct_pctcpu_throttle_not_deny) {
	+ /*
	+ * Ignore the return value; there cannot
	+ * be a "deny" rule there.
	+ */
	+ error = rctl_enforce(p, RACCT_PCTCPU, 0);
	+ KASSERT(error == 0,
	+ ("rctl_enforce() returned %d", error));
	+ }
	+#endif
	+
	racct_set_locked(p, RACCT_CPU, runtime);
	racct_set_locked(p, RACCT_WALLCLOCK,
	(uint64_t)wallclock.tv_sec * 1000000 +
	wallclock.tv_usec);
	+
	mtx_unlock(&racct_lock);
	PROC_UNLOCK(p);
	}

	+ if (racct_pctcpu_throttle_not_deny)
	+ goto done;
	+
	/*
	* To ensure that processes are throttled in a fair way, we need
	* to iterate over all processes again and check the limits
	@@ -1272,12 +1379,18 @@
	continue;
	}

	- if (racct_pcpu_available(p) <= 0)
	- racct_proc_throttle(p);
	- else if (p->p_throttled)
	+ throttle = false;
	+ if (racct_pcpu_available(p) <= 0 &&
	+ (p->p_racct->r_resources[RACCT_PCTCPU] > pcpu_threshold))
	+ throttle = true;
	+
	+ if (throttle)
	+ racct_proc_throttle(p, -1);
	+ else if (p->p_throttled == -1)
	racct_proc_wakeup(p);
	PROC_UNLOCK(p);
	}
	+done:
	sx_sunlock(&allproc_lock);
	pause("-", hz);
	}
	Index: sys/kern/kern_rctl.c
	===================================================================
	--- sys/kern/kern_rctl.c
	+++ sys/kern/kern_rctl.c
	@@ -77,11 +77,31 @@

	#define RCTL_PCPU_SHIFT (10 * 1000000)

	-unsigned int rctl_maxbufsize = RCTL_MAX_OUTBUFSIZE;
	+static unsigned int rctl_maxbufsize = RCTL_MAX_OUTBUFSIZE;
	+static int rctl_log_ratelimit = 10;
	+static int rctl_devctl_ratelimit = 10;
	+static unsigned int rctl_throttle_min = 0;
	+static unsigned int rctl_throttle_max = 0;
	+static unsigned int rctl_throttle_pct = 0;
	+static unsigned int rctl_throttle_pct2 = 0;

	SYSCTL_NODE(_kern_racct, OID_AUTO, rctl, CTLFLAG_RW, 0, "Resource Limits");
	SYSCTL_UINT(_kern_racct_rctl, OID_AUTO, maxbufsize, CTLFLAG_RWTUN,
	&rctl_maxbufsize, 0, "Maximum output buffer size");
	+SYSCTL_UINT(_kern_racct_rctl, OID_AUTO, rctl_log_ratelimit, CTLFLAG_RW,
	+ &rctl_log_ratelimit, 0, "Maximum number of log messages per second");
	+SYSCTL_UINT(_kern_racct_rctl, OID_AUTO, rctl_devctl_ratelimit, CTLFLAG_RW,
	+ &rctl_devctl_ratelimit, 0, "Maximum number of devctl messages per second");
	+SYSCTL_UINT(_kern_racct_rctl, OID_AUTO, throttle_min, CTLFLAG_RDTUN,
	+ &rctl_throttle_min, 0, "Shortest throttling duration, in hz");
	+SYSCTL_UINT(_kern_racct_rctl, OID_AUTO, throttle_max, CTLFLAG_RDTUN,
	+ &rctl_throttle_max, 0, "Longest throttling duration, in hz");
	+SYSCTL_UINT(_kern_racct_rctl, OID_AUTO, throttle_pct, CTLFLAG_RDTUN,
	+ &rctl_throttle_pct, 0,
	+ "Throttling penalty for process consumption, in percent");
	+SYSCTL_UINT(_kern_racct_rctl, OID_AUTO, throttle_pct2, CTLFLAG_RDTUN,
	+ &rctl_throttle_pct2, 0,
	+ "Throttling penalty for container consumption, in percent");

	/*
	* 'rctl_rule_link' connects a rule with every racct it's related to.
	@@ -128,6 +148,10 @@
	{ "shmsize", RACCT_SHMSIZE },
	{ "wallclock", RACCT_WALLCLOCK },
	{ "pcpu", RACCT_PCTCPU },
	+ { "readbps", RACCT_READBPS },
	+ { "writebps", RACCT_WRITEBPS },
	+ { "readiops", RACCT_READIOPS },
	+ { "writeiops", RACCT_WRITEIOPS },
	{ NULL, -1 }};

	static struct dict actionnames[] = {
	@@ -165,6 +189,7 @@
	{ "deny", RCTL_ACTION_DENY },
	{ "log", RCTL_ACTION_LOG },
	{ "devctl", RCTL_ACTION_DEVCTL },
	+ { "throttle", RCTL_ACTION_THROTTLE },
	{ NULL, -1 }};

	static void rctl_init(void);
	@@ -219,66 +244,96 @@
	panic("rctl_resource_name: unknown resource %d", resource);
	}

	-/*
	- * Return the amount of resource that can be allocated by 'p' before
	- * hitting 'rule'.
	- */
	-static int64_t
	-rctl_available_resource(const struct proc p, const struct rctl_rule rule)
	+static struct racct *
	+rctl_proc_rule_to_racct(const struct proc p, const struct rctl_rule rule)
	{
	- int resource;
	- int64_t available = INT64_MAX;
	struct ucred *cred = p->p_ucred;

	ASSERT_RACCT_ENABLED();
	rw_assert(&rctl_lock, RA_LOCKED);

	- resource = rule->rr_resource;
	switch (rule->rr_per) {
	case RCTL_SUBJECT_TYPE_PROCESS:
	- available = rule->rr_amount -
	- p->p_racct->r_resources[resource];
	- break;
	+ return (p->p_racct);
	case RCTL_SUBJECT_TYPE_USER:
	- available = rule->rr_amount -
	- cred->cr_ruidinfo->ui_racct->r_resources[resource];
	- break;
	+ return (cred->cr_ruidinfo->ui_racct);
	case RCTL_SUBJECT_TYPE_LOGINCLASS:
	- available = rule->rr_amount -
	- cred->cr_loginclass->lc_racct->r_resources[resource];
	- break;
	+ return (cred->cr_loginclass->lc_racct);
	case RCTL_SUBJECT_TYPE_JAIL:
	- available = rule->rr_amount -
	- cred->cr_prison->pr_prison_racct->prr_racct->
	- r_resources[resource];
	- break;
	+ return (cred->cr_prison->pr_prison_racct->prr_racct);
	default:
	- panic("rctl_compute_available: unknown per %d",
	- rule->rr_per);
	+ panic("%s: unknown per %d", __func__, rule->rr_per);
	}
	+}
	+
	+/*
	+ * Return the amount of resource that can be allocated by 'p' before
	+ * hitting 'rule'.
	+ */
	+static int64_t
	+rctl_available_resource(const struct proc p, const struct rctl_rule rule)
	+{
	+ int64_t available = INT64_MAX;
	+ const struct racct *racct;
	+
	+ ASSERT_RACCT_ENABLED();
	+ rw_assert(&rctl_lock, RA_LOCKED);
	+ mtx_assert(&racct_lock, MA_OWNED);
	+
	+ racct = rctl_proc_rule_to_racct(p, rule);
	+ available = rule->rr_amount - racct->r_resources[rule->rr_resource];

	return (available);
	}

	/*
	- * Return non-zero if allocating 'amount' by proc 'p' would exceed
	- * resource limit specified by 'rule'.
	+ * Called every second for proc, uidinfo, loginclass, and jail containers.
	+ * If the limit wasn't exceeded, it decreases the usage amount to zero.
	+ * Otherwise, it decreases it by the value of the limit. This way
	+ * resource consumption exceeding the limit "carries over" to the next
	+ * period.
	*/
	-static int
	-rctl_would_exceed(const struct proc p, const struct rctl_rule rule,
	- int64_t amount)
	+void
	+rctl_throttle_decay(struct racct *racct, int resource)
	{
	- int64_t available;
	+ struct rctl_rule *rule;
	+ struct rctl_rule_link *link;
	+ int64_t minavailable;

	- ASSERT_RACCT_ENABLED();
	+ mtx_assert(&racct_lock, MA_OWNED);

	- rw_assert(&rctl_lock, RA_LOCKED);
	+ minavailable = INT64_MAX;

	- available = rctl_available_resource(p, rule);
	- if (available >= amount)
	- return (0);
	+ rw_rlock(&rctl_lock);

	- return (1);
	+ LIST_FOREACH(link, &racct->r_rule_links, rrl_next) {
	+ rule = link->rrl_rule;
	+
	+ if (rule->rr_resource != resource)
	+ continue;
	+ if (rule->rr_action != RCTL_ACTION_THROTTLE)
	+ continue;
	+
	+ if (rule->rr_amount < minavailable)
	+ minavailable = rule->rr_amount;
	+ }
	+
	+ rw_runlock(&rctl_lock);
	+
	+ if (racct->r_resources[resource] < minavailable) {
	+ racct->r_resources[resource] = 0;
	+ } else {
	+ /*
	+ * Cap utilization counter at ten times the limit. Otherwise,
	+ * if we accumulated the resource before adding the limiting
	+ * rule, it could take unreasonably long time for the counter
	+ * value to drop.
	+ */
	+ if (racct->r_resources[resource] > minavailable * 10)
	+ racct->r_resources[resource] = minavailable * 10;
	+ else
	+ racct->r_resources[resource] -= minavailable;
	+ }
	}

	/*
	@@ -328,6 +383,38 @@
	return (minavailable);
	}

	+static uint64_t
	+xadd(uint64_t a, uint64_t b)
	+{
	+ uint64_t c;
	+
	+ c = a + b;
	+
	+ /*
	+ * Detect overflow.
	+ */
	+ if (c < a \|\| c < b)
	+ return (UINT64_MAX);
	+
	+ return (c);
	+}
	+
	+static uint64_t
	+xmul(uint64_t a, uint64_t b)
	+{
	+ uint64_t c;
	+
	+ if (a == 0 \|\| b == 0)
	+ return (0);
	+
	+ c = a * b;
	+
	+ if (c < a \|\| c < b)
	+ return (UINT64_MAX);
	+
	+ return (c);
	+}
	+
	/*
	* 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
	@@ -339,12 +426,16 @@
	struct rctl_rule *rule;
	struct rctl_rule_link *link;
	struct sbuf sb;
	+ int64_t available;
	+ uint64_t sleep_ms, sleep_ratio;
	int should_deny = 0;
	char *buf;
	- static int curtime = 0;
	- static struct timeval lasttime;
	+
	+ static int log_curtime = 0, devctl_curtime = 0;
	+ static struct timeval log_lasttime, devctl_lasttime;

	ASSERT_RACCT_ENABLED();
	+ mtx_assert(&racct_lock, MA_OWNED);

	rw_rlock(&rctl_lock);

	@@ -356,7 +447,9 @@
	rule = link->rrl_rule;
	if (rule->rr_resource != resource)
	continue;
	- if (!rctl_would_exceed(p, rule, amount)) {
	+
	+ available = rctl_available_resource(p, rule);
	+ if (available >= (int64_t)amount) {
	link->rrl_exceeded = 0;
	continue;
	}
	@@ -383,7 +476,8 @@
	if (p->p_state != PRS_NORMAL)
	continue;

	- if (!ppsratecheck(&lasttime, &curtime, 10))
	+ if (!ppsratecheck(&log_lasttime, &log_curtime,
	+ rctl_log_ratelimit))
	continue;

	buf = malloc(RCTL_LOG_BUFSIZE, M_RCTL, M_NOWAIT);
	@@ -408,7 +502,11 @@

	if (p->p_state != PRS_NORMAL)
	continue;
	-
	+
	+ if (!ppsratecheck(&devctl_lasttime, &devctl_curtime,
	+ rctl_devctl_ratelimit))
	+ continue;
	+
	buf = malloc(RCTL_LOG_BUFSIZE, M_RCTL, M_NOWAIT);
	if (buf == NULL) {
	printf("rctl_enforce: out of memory\n");
	@@ -427,6 +525,69 @@
	free(buf, M_RCTL);
	link->rrl_exceeded = 1;
	continue;
	+ case RCTL_ACTION_THROTTLE:
	+ if (p->p_state != PRS_NORMAL)
	+ continue;
	+
	+ /*
	+ * Make the process sleep for a fraction of second
	+ * proportional to the ratio of process' resource
	+ * utilization compared to the limit. The point is
	+ * to penalize resource hogs: processes that consume
	+ * more of the available resources sleep for longer.
	+ *
	+ * We're trying to defer division until the very end,
	+ * to minimize the rounding effects. The following
	+ * calculation could have been written in a clearer
	+ * way like this:
	+ *
	+ * sleep_ms = hz * p->p_racct->r_resources[resource] /
	+ * rule->rr_amount;
	+ * sleep_ms *= rctl_throttle_pct / 100;
	+ * if (sleep_ms < rctl_throttle_min)
	+ * sleep_ms = rctl_throttle_min;
	+ *
	+ */
	+ sleep_ms = xmul(hz, p->p_racct->r_resources[resource]);
	+ sleep_ms = xmul(sleep_ms, rctl_throttle_pct) / 100;
	+ if (sleep_ms < rctl_throttle_min * rule->rr_amount)
	+ sleep_ms = rctl_throttle_min * rule->rr_amount;
	+
	+ /*
	+ * Multiply that by the ratio of the resource
	+ * consumption for the container compared to the limit,
	+ * squared. In other words, a process in a container
	+ * that is two times over the limit will be throttled
	+ * four times as much for hitting the same rule. The
	+ * point is to penalize processes more if the container
	+ * itself (eg certain UID or jail) is above the limit.
	+ */
	+ if (available < 0)
	+ sleep_ratio = -available / rule->rr_amount;
	+ else
	+ sleep_ratio = 0;
	+ sleep_ratio = xmul(sleep_ratio, sleep_ratio);
	+ sleep_ratio = xmul(sleep_ratio, rctl_throttle_pct2) / 100;
	+ sleep_ms = xadd(sleep_ms, xmul(sleep_ms, sleep_ratio));
	+
	+ /*
	+ * Finally the division.
	+ */
	+ sleep_ms /= rule->rr_amount;
	+
	+ if (sleep_ms > rctl_throttle_max)
	+ sleep_ms = rctl_throttle_max;
	+#if 0
	+ printf("%s: pid %d (%s), %jd of %jd, will sleep for %ld ms (ratio %ld, available %ld)\n",
	+ __func__, p->p_pid, p->p_comm,
	+ p->p_racct->r_resources[resource],
	+ rule->rr_amount, sleep_ms, sleep_ratio, available);
	+#endif
	+
	+ KASSERT(sleep_ms >= rctl_throttle_min, ("%s: %ld < %d\n",
	+ __func__, sleep_ms, rctl_throttle_min));
	+ racct_proc_throttle(p, sleep_ms);
	+ continue;
	default:
	if (link->rrl_exceeded != 0)
	continue;
	@@ -642,6 +803,9 @@
	if ((size_t)(end - str) != strlen(str))
	return (EINVAL);

	+ if (*value <= 0)
	+ return (ERANGE);
	+
	return (0);
	}

	@@ -1008,8 +1172,13 @@
	error = str2int64(amountstr, &rule->rr_amount);
	if (error != 0)
	goto out;
	- if (RACCT_IS_IN_MILLIONS(rule->rr_resource))
	+ if (RACCT_IS_IN_MILLIONS(rule->rr_resource)) {
	+ if (rule->rr_amount > INT64_MAX / 1000000) {
	+ error = ERANGE;
	+ goto out;
	+ }
	rule->rr_amount *= 1000000;
	+ }
	}

	if (perstr == NULL \|\| perstr[0] == '\0')
	@@ -1048,20 +1217,27 @@
	KASSERT(rctl_rule_fully_specified(rule), ("rule not fully specified"));

	/*
	- * Some rules just don't make sense. Note that the one below
	- * cannot be rewritten using RACCT_IS_DENIABLE(); the RACCT_PCTCPU,
	- * for example, is not deniable in the racct sense, but the
	- * limit is enforced in a different way, so "deny" rules for %CPU
	- * do make sense.
	+ * Some rules just don't make sense, like "deny" rule for an undeniable
	+ * resource. The exception are the RSS and %CPU resources - they are
	+ * not deniable in the racct sense, but the limit is enforced in
	+ * a different way.
	*/
	if (rule->rr_action == RCTL_ACTION_DENY &&
	- (rule->rr_resource == RACCT_CPU \|\|
	- rule->rr_resource == RACCT_WALLCLOCK))
	+ !RACCT_IS_DENIABLE(rule->rr_resource) &&
	+ rule->rr_resource != RACCT_RSS &&
	+ rule->rr_resource != RACCT_PCTCPU) {
	return (EOPNOTSUPP);
	+ }
	+
	+ if (rule->rr_action == RCTL_ACTION_THROTTLE &&
	+ !RACCT_IS_DECAYING(rule->rr_resource)) {
	+ return (EOPNOTSUPP);
	+ }

	if (rule->rr_per == RCTL_SUBJECT_TYPE_PROCESS &&
	- RACCT_IS_SLOPPY(rule->rr_resource))
	+ RACCT_IS_SLOPPY(rule->rr_resource)) {
	return (EOPNOTSUPP);
	+ }

	/*
	* Make sure there are no duplicated rules. Also, for the "deny"
	@@ -1346,7 +1522,9 @@
	for (i = 0; i <= RACCT_MAX; i++) {
	if (sloppy == 0 && RACCT_IS_SLOPPY(i))
	continue;
	+ mtx_lock(&racct_lock);
	amount = racct->r_resources[i];
	+ mtx_unlock(&racct_lock);
	if (RACCT_IS_IN_MILLIONS(i))
	amount /= 1000000;
	sbuf_printf(sb, "%s=%jd,", rctl_resource_name(i), amount);
	@@ -1619,6 +1797,81 @@
	return (error);
	}

	+static void
	+rctl_rule_find_callback(struct racct racct, void arg2, void *arg3)
	+{
	+ struct rctl_rule filter = (struct rctl_rule )arg2;
	+ struct rctl_rule_link *link;
	+ int found = 0;
	+
	+ ASSERT_RACCT_ENABLED();
	+ rw_assert(&rctl_lock, RA_LOCKED);
	+
	+ LIST_FOREACH(link, &racct->r_rule_links, rrl_next) {
	+ if (!rctl_rule_matches(link->rrl_rule, filter))
	+ continue;
	+ found++;
	+ }
	+
	+ ((int )arg3) += found;
	+}
	+
	+/*
	+ * Try to set racct_pctcpu_throttle_not_deny to the value given.
	+ * It can only be set to '1' if there are no 'pcpu:deny' rules,
	+ * and to '0' only if there are no 'pcpu:throttle" rules.
	+ */
	+static int
	+rctl_set_pctcpu_throttle_not_deny(int val)
	+{
	+ struct rctl_rule *filter;
	+ struct proc *p;
	+ int found;
	+
	+ if (racct_pctcpu_throttle_not_deny == val)
	+ return (0);
	+
	+ /*
	+ * The purpose of this lock here is not only to be able to iterate
	+ * over processes, but also to avoid race with rctl_rule_add()
	+ * adding a clashing rule after we do the search, but before we set
	+ * the value. Besides, rctl_rule_add() holds this lock anyway.
	+ */
	+ sx_assert(&allproc_lock, SA_LOCKED);
	+
	+ filter = rctl_rule_alloc(M_WAITOK);
	+ filter->rr_resource = RACCT_PCTCPU;
	+ if (val == 1)
	+ filter->rr_action = RCTL_ACTION_DENY;
	+ else
	+ filter->rr_action = RCTL_ACTION_THROTTLE;
	+
	+ found = 0;
	+
	+ rctl_rule_pre_callback();
	+ FOREACH_PROC_IN_SYSTEM(p)
	+ rctl_rule_find_callback(p->p_racct, filter, &found);
	+ rctl_rule_post_callback();
	+
	+ loginclass_racct_foreach(rctl_rule_find_callback,
	+ rctl_rule_pre_callback, rctl_rule_post_callback,
	+ filter, &found);
	+ ui_racct_foreach(rctl_rule_find_callback,
	+ rctl_rule_pre_callback, rctl_rule_post_callback,
	+ filter, &found);
	+ prison_racct_foreach(rctl_rule_find_callback,
	+ rctl_rule_pre_callback, rctl_rule_post_callback,
	+ filter, &found);
	+
	+ rctl_rule_release(filter);
	+
	+ if (found != 0)
	+ return (EOPNOTSUPP);
	+
	+ racct_pctcpu_throttle_not_deny = val;
	+ return (0);
	+}
	+
	int
	sys_rctl_add_rule(struct thread td, struct rctl_add_rule_args uap)
	{
	@@ -1656,6 +1909,19 @@
	goto out;
	}

	+ if (rule->rr_resource == RACCT_PCTCPU) {
	+ if (rule->rr_action == RCTL_ACTION_DENY) {
	+ error = rctl_set_pctcpu_throttle_not_deny(0);
	+ if (error != 0)
	+ goto out;
	+ }
	+ if (rule->rr_action == RCTL_ACTION_THROTTLE) {
	+ error = rctl_set_pctcpu_throttle_not_deny(1);
	+ if (error != 0)
	+ goto out;
	+ }
	+ }
	+
	error = rctl_rule_add(rule);

	out:
	@@ -1935,6 +2201,15 @@
	UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
	rctl_rule_zone = uma_zcreate("rctl_rule", sizeof(struct rctl_rule),
	NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
	+
	+ if (rctl_throttle_min <= 0)
	+ rctl_throttle_min = 1;
	+ if (rctl_throttle_max <= 0)
	+ rctl_throttle_max = 2 * hz;
	+ if (rctl_throttle_pct <= 0)
	+ rctl_throttle_pct = 100;
	+ if (rctl_throttle_pct2 <= 0)
	+ rctl_throttle_pct2 = 100;
	}

	#else /* !RCTL */
	Index: sys/kern/subr_trap.c
	===================================================================
	--- sys/kern/subr_trap.c
	+++ sys/kern/subr_trap.c
	@@ -172,10 +172,14 @@
	(td->td_vnet_lpush != NULL) ? td->td_vnet_lpush : "N/A"));
	#endif
	#ifdef RACCT
	- if (racct_enable && p->p_throttled == 1) {
	+ if (racct_enable && p->p_throttled != 0) {
	PROC_LOCK(p);
	- while (p->p_throttled == 1)
	- msleep(p->p_racct, &p->p_mtx, 0, "racct", 0);
	+ while (p->p_throttled != 0) {
	+ msleep(p->p_racct, &p->p_mtx, 0, "racct",
	+ p->p_throttled < 0 ? 0 : p->p_throttled);
	+ if (p->p_throttled > 0)
	+ p->p_throttled = 0;
	+ }
	PROC_UNLOCK(p);
	}
	#endif
	Index: sys/kern/vfs_bio.c
	===================================================================
	--- sys/kern/vfs_bio.c
	+++ sys/kern/vfs_bio.c
	@@ -61,6 +61,7 @@
	#include <sys/kernel.h>
	#include <sys/kthread.h>
	#include <sys/proc.h>
	+#include <sys/racct.h>
	#include <sys/resourcevar.h>
	#include <sys/rwlock.h>
	#include <sys/smp.h>
	@@ -1784,8 +1785,16 @@
	rabp = getblk(vp, rablkno, rabsize, 0, 0, 0);

	if ((rabp->b_flags & B_CACHE) == 0) {
	- if (!TD_IS_IDLETHREAD(curthread))
	+ if (!TD_IS_IDLETHREAD(curthread)) {
	+#ifdef RACCT
	+ if (racct_enable) {
	+ PROC_LOCK(curproc);
	+ racct_add_buf(curproc, rabp, 0);
	+ PROC_UNLOCK(curproc);
	+ }
	+#endif /* RACCT */
	curthread->td_ru.ru_inblock++;
	+ }
	rabp->b_flags \|= B_ASYNC;
	rabp->b_flags &= ~B_INVAL;
	rabp->b_ioflags &= ~BIO_ERROR;
	@@ -1829,8 +1838,16 @@

	/* if not found in cache, do some I/O */
	if ((bp->b_flags & B_CACHE) == 0) {
	- if (!TD_IS_IDLETHREAD(curthread))
	+ if (!TD_IS_IDLETHREAD(curthread)) {
	+#ifdef RACCT
	+ if (racct_enable) {
	+ PROC_LOCK(curproc);
	+ racct_add_buf(curproc, bp, 0);
	+ PROC_UNLOCK(curproc);
	+ }
	+#endif /* RACCT */
	curthread->td_ru.ru_inblock++;
	+ }
	bp->b_iocmd = BIO_READ;
	bp->b_flags &= ~B_INVAL;
	bp->b_ioflags &= ~BIO_ERROR;
	@@ -1926,8 +1943,16 @@
	bp->b_runningbufspace = bp->b_bufsize;
	space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace);

	- if (!TD_IS_IDLETHREAD(curthread))
	+ if (!TD_IS_IDLETHREAD(curthread)) {
	+#ifdef RACCT
	+ if (racct_enable) {
	+ PROC_LOCK(curproc);
	+ racct_add_buf(curproc, bp, 1);
	+ PROC_UNLOCK(curproc);
	+ }
	+#endif /* RACCT */
	curthread->td_ru.ru_oublock++;
	+ }
	if (oldflags & B_ASYNC)
	BUF_KERNPROC(bp);
	bp->b_iooffset = dbtob(bp->b_blkno);
	Index: sys/kern/vfs_cluster.c
	===================================================================
	--- sys/kern/vfs_cluster.c
	+++ sys/kern/vfs_cluster.c
	@@ -45,6 +45,7 @@
	#include <sys/vnode.h>
	#include <sys/malloc.h>
	#include <sys/mount.h>
	+#include <sys/racct.h>
	#include <sys/resourcevar.h>
	#include <sys/rwlock.h>
	#include <sys/vmmeter.h>
	@@ -241,6 +242,13 @@
	BUF_KERNPROC(bp);
	bp->b_iooffset = dbtob(bp->b_blkno);
	bstrategy(bp);
	+#ifdef RACCT
	+ if (racct_enable) {
	+ PROC_LOCK(curproc);
	+ racct_add_buf(curproc, bp, 0);
	+ PROC_UNLOCK(curproc);
	+ }
	+#endif /* RACCT */
	curthread->td_ru.ru_inblock++;
	}

	@@ -294,6 +302,13 @@
	BUF_KERNPROC(rbp);
	rbp->b_iooffset = dbtob(rbp->b_blkno);
	bstrategy(rbp);
	+#ifdef RACCT
	+ if (racct_enable) {
	+ PROC_LOCK(curproc);
	+ racct_add_buf(curproc, rbp, 0);
	+ PROC_UNLOCK(curproc);
	+ }
	+#endif /* RACCT */
	curthread->td_ru.ru_inblock++;
	}

	Index: sys/sys/proc.h
	===================================================================
	--- sys/sys/proc.h
	+++ sys/sys/proc.h
	@@ -620,7 +620,7 @@
	after fork. */
	uint64_t p_prev_runtime; /* (c) Resource usage accounting. */
	struct racct p_racct; / (b) Resource accounting. */
	- u_char p_throttled; /* (c) Flag for racct pcpu throttling */
	+ int p_throttled; /* (c) Flag for racct pcpu throttling */
	struct vm_domain_policy p_vm_dom_policy; /* (c) process default VM domain, or -1 */
	/*
	* An orphan is the child that has beed re-parented to the
	Index: sys/sys/racct.h
	===================================================================
	--- sys/sys/racct.h
	+++ sys/sys/racct.h
	@@ -42,6 +42,7 @@
	#include <sys/stdint.h>
	#include <sys/sysctl.h>

	+struct buf;
	struct proc;
	struct rctl_rule_link;
	struct ucred;
	@@ -71,7 +72,11 @@
	#define RACCT_SHMSIZE 18
	#define RACCT_WALLCLOCK 19
	#define RACCT_PCTCPU 20
	-#define RACCT_MAX RACCT_PCTCPU
	+#define RACCT_READBPS 21
	+#define RACCT_WRITEBPS 22
	+#define RACCT_READIOPS 23
	+#define RACCT_WRITEIOPS 24
	+#define RACCT_MAX RACCT_WRITEIOPS

	/*
	* Resource properties.
	@@ -85,6 +90,7 @@

	extern int racct_types[];
	extern int racct_enable;
	+extern int racct_pctcpu_throttle_not_deny;

	#define ASSERT_RACCT_ENABLED() KASSERT(racct_enable, \
	("%s called with !racct_enable", __func__))
	@@ -98,7 +104,7 @@

	/*
	* Resource usage can drop, as opposed to only grow. When the process
	- * terminates, its resource usage is freed from the respective
	+ * terminates, its resource usage is subtracted from the respective
	* per-credential racct containers.
	*/
	#define RACCT_IS_RECLAIMABLE(X) (racct_types[X] & RACCT_RECLAIMABLE)
	@@ -126,8 +132,7 @@
	* When a process terminates, its resource usage is not automatically
	* subtracted from per-credential racct containers. Instead, the resource
	* usage of per-credential racct containers decays in time.
	- * Resource usage can olso drop for such resource.
	- * So far, the only such resource is RACCT_PCTCPU.
	+ * Resource usage can also drop for such resource.
	*/
	#define RACCT_IS_DECAYING(X) (racct_types[X] & RACCT_DECAYING)

	@@ -149,11 +154,14 @@

	SYSCTL_DECL(_kern_racct);

	+extern struct mtx racct_lock;
	+
	#ifdef RACCT

	int racct_add(struct proc *p, int resource, uint64_t amount);
	void racct_add_cred(struct ucred *cred, int resource, uint64_t amount);
	void racct_add_force(struct proc *p, int resource, uint64_t amount);
	+void racct_add_buf(struct proc p, const struct buf bufp, int is_write);
	int racct_set(struct proc *p, int resource, uint64_t amount);
	void racct_set_force(struct proc *p, int resource, uint64_t amount);
	void racct_sub(struct proc *p, int resource, uint64_t amount);
	@@ -171,6 +179,7 @@
	void racct_proc_ucred_changed(struct proc p, struct ucred oldcred,
	struct ucred *newcred);
	void racct_move(struct racct dest, struct racct src);
	+void racct_proc_throttle(struct proc *p, int timeout);

	#else

	Index: sys/sys/rctl.h
	===================================================================
	--- sys/sys/rctl.h
	+++ sys/sys/rctl.h
	@@ -129,7 +129,8 @@
	#define RCTL_ACTION_DENY (RCTL_ACTION_SIGNAL_MAX + 1)
	#define RCTL_ACTION_LOG (RCTL_ACTION_SIGNAL_MAX + 2)
	#define RCTL_ACTION_DEVCTL (RCTL_ACTION_SIGNAL_MAX + 3)
	-#define RCTL_ACTION_MAX RCTL_ACTION_DEVCTL
	+#define RCTL_ACTION_THROTTLE (RCTL_ACTION_SIGNAL_MAX + 4)
	+#define RCTL_ACTION_MAX RCTL_ACTION_THROTTLE

	#define RCTL_AMOUNT_UNDEFINED -1

	@@ -140,6 +141,7 @@
	int rctl_rule_add(struct rctl_rule *rule);
	int rctl_rule_remove(struct rctl_rule *filter);
	int rctl_enforce(struct proc *p, int resource, uint64_t amount);
	+void rctl_throttle_decay(struct racct *racct, int resource);
	int64_t rctl_pcpu_available(const struct proc *p);
	uint64_t rctl_get_limit(struct proc *p, int resource);
	uint64_t rctl_get_available(struct proc *p, int resource);
	Index: sys/ufs/ffs/ffs_inode.c
	===================================================================
	--- sys/ufs/ffs/ffs_inode.c
	+++ sys/ufs/ffs/ffs_inode.c
	@@ -41,6 +41,7 @@
	#include <sys/malloc.h>
	#include <sys/mount.h>
	#include <sys/proc.h>
	+#include <sys/racct.h>
	#include <sys/random.h>
	#include <sys/resourcevar.h>
	#include <sys/rwlock.h>
	@@ -616,6 +617,13 @@
	vp = ITOV(ip);
	bp = getblk(vp, lbn, (int)fs->fs_bsize, 0, 0, 0);
	if ((bp->b_flags & B_CACHE) == 0) {
	+#ifdef RACCT
	+ if (racct_enable) {
	+ PROC_LOCK(curproc);
	+ racct_add_buf(curproc, bp, 0);
	+ PROC_UNLOCK(curproc);
	+ }
	+#endif /* RACCT */
	curthread->td_ru.ru_inblock++; /* pay for read */
	bp->b_iocmd = BIO_READ;
	bp->b_flags &= ~B_INVAL;
	Index: sys/ufs/ffs/ffs_softdep.c
	===================================================================
	--- sys/ufs/ffs/ffs_softdep.c
	+++ sys/ufs/ffs/ffs_softdep.c
	@@ -69,6 +69,7 @@
	#include <sys/namei.h>
	#include <sys/priv.h>
	#include <sys/proc.h>
	+#include <sys/racct.h>
	#include <sys/rwlock.h>
	#include <sys/stat.h>
	#include <sys/sysctl.h>
	@@ -6229,6 +6230,13 @@
	vfs_busy_pages(bp, 0);
	bp->b_iooffset = dbtob(bp->b_blkno);
	bstrategy(bp);
	+#ifdef RACCT
	+ if (racct_enable) {
	+ PROC_LOCK(curproc);
	+ racct_add_buf(curproc, bp, 0);
	+ PROC_UNLOCK(curproc);
	+ }
	+#endif /* RACCT */
	curthread->td_ru.ru_inblock++;
	error = bufwait(bp);
	if (error) {
	Index: sys/ufs/ufs/ufs_bmap.c
	===================================================================
	--- sys/ufs/ufs/ufs_bmap.c
	+++ sys/ufs/ufs/ufs_bmap.c
	@@ -44,6 +44,7 @@
	#include <sys/proc.h>
	#include <sys/vnode.h>
	#include <sys/mount.h>
	+#include <sys/racct.h>
	#include <sys/resourcevar.h>
	#include <sys/stat.h>

	@@ -223,6 +224,13 @@
	vfs_busy_pages(bp, 0);
	bp->b_iooffset = dbtob(bp->b_blkno);
	bstrategy(bp);
	+#ifdef RACCT
	+ if (racct_enable) {
	+ PROC_LOCK(curproc);
	+ racct_add_buf(curproc, bp, 0);
	+ PROC_UNLOCK(curproc);
	+ }
	+#endif /* RACCT */
	curthread->td_ru.ru_inblock++;
	error = bufwait(bp);
	if (error) {
	Index: sys/vm/vm_fault.c
	===================================================================
	--- sys/vm/vm_fault.c
	+++ sys/vm/vm_fault.c
	@@ -83,6 +83,7 @@
	#include <sys/lock.h>
	#include <sys/mman.h>
	#include <sys/proc.h>
	+#include <sys/racct.h>
	#include <sys/resourcevar.h>
	#include <sys/rwlock.h>
	#include <sys/sysctl.h>
	@@ -994,6 +995,21 @@
	if (hardfault) {
	PCPU_INC(cnt.v_io_faults);
	curthread->td_ru.ru_majflt++;
	+#ifdef RACCT
	+ if (racct_enable && fs.object->type == OBJT_VNODE) {
	+ PROC_LOCK(curproc);
	+ if ((fault_type & (VM_PROT_COPY \| VM_PROT_WRITE)) != 0) {
	+ racct_add_force(curproc, RACCT_WRITEBPS,
	+ PAGE_SIZE + behind * PAGE_SIZE);
	+ racct_add_force(curproc, RACCT_WRITEIOPS, 1);
	+ } else {
	+ racct_add_force(curproc, RACCT_READBPS,
	+ PAGE_SIZE + ahead * PAGE_SIZE);
	+ racct_add_force(curproc, RACCT_READIOPS, 1);
	+ }
	+ PROC_UNLOCK(curproc);
	+ }
	+#endif
	} else
	curthread->td_ru.ru_minflt++;

	Index: usr.bin/rctl/rctl.8
	===================================================================
	--- usr.bin/rctl/rctl.8
	+++ usr.bin/rctl/rctl.8
	@@ -25,7 +25,7 @@
	.\"
	.\" $FreeBSD$
	.\"
	-.Dd November 29, 2015
	+.Dd January 30, 2016
	.Dt RCTL 8
	.Os
	.Sh NAME
	@@ -204,14 +204,22 @@
	.It Sy shmsize Ta "SysV shared memory size, in bytes"
	.It Sy wallclock Ta "wallclock time, in seconds"
	.It Sy pcpu Ta "%CPU, in percents of a single CPU core"
	+.It Sy readbps Ta "filesystem reads, in bytes per second"
	+.It Sy writebps Ta "filesystem writes, in bytes per second"
	+.It Sy readiops Ta "filesystem reads, in operations per second"
	+.It Sy writeiops Ta "filesystem writes, in operations per second"
	.El
	.Sh ACTIONS
	.Bl -column -offset 3n "pseudoterminals"
	.It Em action
	.It Sy deny Ta deny the allocation; not supported for
	-.Sy cputime
	+.Sy cputime ,
	+.Sy wallclock ,
	+.Sy readbps ,
	+.Sy writebps ,
	+.Sy readiops ,
	and
	-.Sy wallclock
	+.Sy writeiops
	.It Sy log Ta "log a warning to the console"
	.It Sy devctl Ta "send notification to"
	.Xr devd 8
	@@ -228,6 +236,19 @@
	See
	.Xr signal 3
	for a list of supported signals
	+.It Sy throttle Ta "slow down process execution"; only supported for
	+.Sy readbps ,
	+.Sy writebps ,
	+.Sy readiops ,
	+and
	+.Sy writeiops .
	+Note that one cannot have both
	+.Sy throttle
	+and
	+.Sy deny
	+rules for the
	+.Sy pcpu
	+resource at the same time.
	.El
	.Pp
	Not all actions are supported for all resources.
	@@ -287,3 +308,22 @@
	Limiting
	.Sy memoryuse
	may kill the machine due to thrashing.
	+.Pp
	+The
	+.Sy readiops
	+and
	+.Sy writeiops
	+counters are only approximations.
	+Like
	+.Sy readbps
	+and
	+.Sy writebps ,
	+they are calculated in the filesystem layer, where it is difficult
	+or even impossible to observe actual disk device operations.
	+.Pp
	+The
	+.Sy writebps
	+and
	+.Sy writeiops
	+resources generally account for writes to the filesystem cache,
	+not to actual devices.

File Metadata

Mime Type: text/plain
Expires: Wed, Apr 22, 8:48 AM (19 h, 35 m)
Storage Engine: blob
Storage Format: Raw Data
Storage Handle: 31966107
Default Alt Text: D5080.id12953.diff (43 KB)

D5080.id12953.diffNo OneTemporaryActions

D5080.id12953.diffView Options

File Metadata

Event Timeline

D5080.id12953.diff
No OneTemporary
Actions

D5080.id12953.diff
View Options