common/os/clock_highres.c

	clock_highres.c revision 7c478bd95313f5f23a4c958a745db2134aa03244
/*
 * CDDL HEADER START
 *
 * The contents of this file are subject to the terms of the
 * Common Development and Distribution License, Version 1.0 only
 * (the "License").  You may not use this file except in compliance
 * with the License.
 *
 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
 * or http://www.opensolaris.org/os/licensing.
 * See the License for the specific language governing permissions
 * and limitations under the License.
 *
 * When distributing Covered Code, include this CDDL HEADER in each
 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
 * If applicable, add the following below this CDDL HEADER, with the
 * fields enclosed by brackets "[]" replaced with your own identifying
 * information: Portions Copyright [yyyy] [name of copyright owner]
 *
 * CDDL HEADER END
 */
/*
 * Copyright 2003 Sun Microsystems, Inc.  All rights reserved.
 * Use is subject to license terms.
 */

#pragma ident   "%Z%%M% %I% %E% SMI"

#include <sys/timer.h>
#include <sys/systm.h>
#include <sys/param.h>
#include <sys/kmem.h>
#include <sys/debug.h>
#include <sys/cyclic.h>
#include <sys/cmn_err.h>
#include <sys/pset.h>
#include <sys/atomic.h>
#include <sys/policy.h>

static clock_backend_t clock_highres;

/*ARGSUSED*/
static int
clock_highres_settime(timespec_t *ts)
{
    return (EINVAL);
}

static int
clock_highres_gettime(timespec_t *ts)
{
    hrt2ts(gethrtime(), (timestruc_t *)ts);

    return (0);
}

static int
clock_highres_getres(timespec_t *ts)
{
    hrt2ts(cyclic_getres(), (timestruc_t *)ts);

    return (0);
}

/*ARGSUSED*/
static int
clock_highres_timer_create(itimer_t *it, struct sigevent *ev)
{
    /*
     * CLOCK_HIGHRES timers of sufficiently high resolution can deny
     * service; only allow privileged users to create such timers.
     * Sites that do not wish to have this restriction should
     * give users the "proc_clock_highres" privilege.
     */
    if (secpolicy_clock_highres(CRED()) != 0) {
        it->it_arg = NULL;
        return (EPERM);
    }

    it->it_arg = kmem_zalloc(sizeof (cyclic_id_t), KM_SLEEP);

    return (0);
}

static void
clock_highres_fire(void *arg)
{
    itimer_t *it = (itimer_t *)arg;
    hrtime_t *addr = &it->it_hrtime;
    hrtime_t old = *addr, new = gethrtime();

    do {
        old = *addr;
    } while (cas64((uint64_t *)addr, old, new) != old);

    timer_fire(it);
}

static int
clock_highres_timer_settime(itimer_t *it, int flags,
    const struct itimerspec *when)
{
    cyclic_id_t cyc, *cycp = it->it_arg;
    proc_t *p = curproc;
    kthread_t *t = curthread;
    cyc_time_t cyctime;
    cyc_handler_t hdlr;
    cpu_t *cpu;
    cpupart_t *cpupart;
    int pset;

    cyctime.cyt_when = ts2hrt(&when->it_value);
    cyctime.cyt_interval = ts2hrt(&when->it_interval);

    mutex_enter(&cpu_lock);
    if ((cyc = *cycp) != CYCLIC_NONE) {
        cyclic_remove(cyc);
        *cycp = CYCLIC_NONE;
    }

    if (cyctime.cyt_when == 0) {
        mutex_exit(&cpu_lock);
        return (0);
    }

    if (!(flags & TIMER_ABSTIME))
        cyctime.cyt_when += gethrtime();

    /*
     * Now we will check for overflow (that is, we will check to see
     * that the start time plus the interval time doesn't exceed
     * INT64_MAX).  The astute code reviewer will observe that this
     * one-time check doesn't guarantee that a future expiration
     * will not wrap.  We wish to prove, then, that if a future
     * expiration does wrap, the earliest the problem can be encountered
     * is (INT64_MAX / 2) nanoseconds (191 years) after boot.  Formally:
     *
     *  Given:  s + i < m   s > 0   i > 0
     *      s + ni > m  n > 1
     *
     *    (where "s" is the start time, "i" is the interval, "n" is the
     *    number of times the cyclic has fired and "m" is INT64_MAX)
     *
     *  Prove:
     *      (a)  s + (n - 1)i > (m / 2)
     *      (b)  s + (n - 1)i < m
     *
     * That is, prove that we must have fired at least once 191 years
     * after boot.  The proof is very straightforward; since the left
     * side of (a) is minimized when i is small, it is sufficient to show
     * that the statement is true for i's smallest possible value
     * (((m - s) / n) + epsilon).  The same goes for (b); showing that the
     * statement is true for i's largest possible value (m - s + epsilon)
     * is sufficient to prove the statement.
     *
     * The actual arithmetic manipulation is left up to reader.
     */
    if (cyctime.cyt_when > INT64_MAX - cyctime.cyt_interval) {
        mutex_exit(&cpu_lock);
        return (EOVERFLOW);
    }

    if (cyctime.cyt_interval == 0) {
        /*
         * If this is a one-shot, then we set the interval to assure
         * that the cyclic will next fire INT64_MAX nanoseconds after
         * boot (which corresponds to over 292 years -- yes, Buck Rogers
         * may have his 292-year-uptime-Solaris box malfunction).  If
         * this timer is never touched, this cyclic will simply
         * consume space in the cyclic subsystem.  As soon as
         * timer_settime() or timer_delete() is called, the cyclic is
         * removed (so it's not possible to run the machine out
         * of resources by creating one-shots).
         */
        cyctime.cyt_interval = INT64_MAX - cyctime.cyt_when;
    }

    it->it_itime = *when;

    hrt2ts(cyctime.cyt_when, &it->it_itime.it_value);

    hdlr.cyh_func = (cyc_func_t)clock_highres_fire;
    hdlr.cyh_arg = it;
    hdlr.cyh_level = CY_LOW_LEVEL;

    if (cyctime.cyt_when != 0)
        *cycp = cyc = cyclic_add(&hdlr, &cyctime);
    else
        *cycp = cyc = CYCLIC_NONE;

    /*
     * Now that we have the cyclic created, we need to bind it to our
     * bound CPU and processor set (if any).
     */
    mutex_enter(&p->p_lock);
    cpu = t->t_bound_cpu;
    cpupart = t->t_cpupart;
    pset = t->t_bind_pset;

    mutex_exit(&p->p_lock);

    cyclic_bind(cyc, cpu, pset == PS_NONE ? NULL : cpupart);

    mutex_exit(&cpu_lock);

    return (0);
}

static int
clock_highres_timer_gettime(itimer_t *it, struct itimerspec *when)
{
    /*
     * CLOCK_HIGHRES doesn't update it_itime.
     */
    hrtime_t start = ts2hrt(&it->it_itime.it_value);
    hrtime_t interval = ts2hrt(&it->it_itime.it_interval);
    hrtime_t diff, now = gethrtime();
    hrtime_t *addr = &it->it_hrtime;
    hrtime_t last;

    /*
     * We're using cas64() here only to assure that we slurp the entire
     * timestamp atomically.
     */
    last = cas64((uint64_t *)addr, 0, 0);

    *when = it->it_itime;

    if (!timerspecisset(&when->it_value))
        return (0);

    if (start > now) {
        /*
         * We haven't gone off yet...
         */
        diff = start - now;
    } else {
        if (interval == 0) {
            /*
             * This is a one-shot which should have already
             * fired; set it_value to 0.
             */
            timerspecclear(&when->it_value);
            return (0);
        }

        /*
         * Calculate how far we are into this interval.
         */
        diff = (now - start) % interval;

        /*
         * Now check to see if we've dealt with the last interval
         * yet.
         */
        if (now - diff > last) {
            /*
             * The last interval hasn't fired; set it_value to 0.
             */
            timerspecclear(&when->it_value);
            return (0);
        }

        /*
         * The last interval _has_ fired; we can return the amount
         * of time left in this interval.
         */
        diff = interval - diff;
    }

    hrt2ts(diff, &when->it_value);

    return (0);
}

static int
clock_highres_timer_delete(itimer_t *it)
{
    cyclic_id_t cyc;

    if (it->it_arg == NULL) {
        /*
         * This timer was never fully created; we must have failed
         * in the clock_highres_timer_create() routine.
         */
        return (0);
    }

    mutex_enter(&cpu_lock);

    if ((cyc = *((cyclic_id_t *)it->it_arg)) != CYCLIC_NONE)
        cyclic_remove(cyc);

    mutex_exit(&cpu_lock);

    kmem_free(it->it_arg, sizeof (cyclic_id_t));

    return (0);
}

static void
clock_highres_timer_lwpbind(itimer_t *it)
{
    proc_t *p = curproc;
    kthread_t *t = curthread;
    cyclic_id_t cyc = *((cyclic_id_t *)it->it_arg);
    cpu_t *cpu;
    cpupart_t *cpupart;
    int pset;

    if (cyc == CYCLIC_NONE)
        return;

    mutex_enter(&cpu_lock);
    mutex_enter(&p->p_lock);

    /*
     * Okay, now we can safely look at the bindings.
     */
    cpu = t->t_bound_cpu;
    cpupart = t->t_cpupart;
    pset = t->t_bind_pset;

    /*
     * Now we drop p_lock.  We haven't dropped cpu_lock; we're guaranteed
     * that even if the bindings change, the CPU and/or processor set
     * that this timer was bound to remain valid (and the combination
     * remains self-consistent).
     */
    mutex_exit(&p->p_lock);

    cyclic_bind(cyc, cpu, pset == PS_NONE ? NULL : cpupart);

    mutex_exit(&cpu_lock);
}

void
clock_highres_init()
{
    clock_backend_t *be = &clock_highres;
    struct sigevent *ev = &be->clk_default;

    ev->sigev_signo = SIGALRM;
    ev->sigev_notify = SIGEV_SIGNAL;
    ev->sigev_value.sival_ptr = NULL;

    be->clk_clock_settime = clock_highres_settime;
    be->clk_clock_gettime = clock_highres_gettime;
    be->clk_clock_getres = clock_highres_getres;
    be->clk_timer_create = clock_highres_timer_create;
    be->clk_timer_gettime = clock_highres_timer_gettime;
    be->clk_timer_settime = clock_highres_timer_settime;
    be->clk_timer_delete = clock_highres_timer_delete;
    be->clk_timer_lwpbind = clock_highres_timer_lwpbind;

    clock_add_backend(CLOCK_HIGHRES, &clock_highres);
}