kernel - Fix spinlock bug introduced with windowing (2)

[dragonfly.git] / sys / kern / kern_spinlock.c

/*
 * Copyright (c) 2005 Jeffrey M. Hsu.  All rights reserved.
 *
 * This code is derived from software contributed to The DragonFly Project
 * by Jeffrey M. Hsu. and Matthew Dillon
 *
 * 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. Neither the name of The DragonFly Project 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 COPYRIGHT HOLDERS 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
 * COPYRIGHT HOLDERS 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.
 */

/*
 * The implementation is designed to avoid looping when compatible operations
 * are executed.
 *
 * To acquire a spinlock we first increment counta.  Then we check if counta
 * meets our requirements.  For an exclusive spinlock it must be 1, of a
 * shared spinlock it must either be 1 or the SHARED_SPINLOCK bit must be set.
 *
 * Shared spinlock failure case: Decrement the count, loop until we can
 * transition from 0 to SHARED_SPINLOCK|1, or until we find SHARED_SPINLOCK
 * is set and increment the count.
 *
 * Exclusive spinlock failure case: While maintaining the count, clear the
 * SHARED_SPINLOCK flag unconditionally.  Then use an atomic add to transfer
 * the count from the low bits to the high bits of counta.  Then loop until
 * all low bits are 0.  Once the low bits drop to 0 we can transfer the
 * count back with an atomic_cmpset_int(), atomically, and return.
 */
#include <sys/param.h>
#include <sys/systm.h>
#include <sys/types.h>
#include <sys/kernel.h>
#include <sys/sysctl.h>
#ifdef INVARIANTS
#include <sys/proc.h>
#endif
#include <sys/priv.h>
#include <machine/atomic.h>
#include <machine/cpu.h>
#include <machine/cpufunc.h>
#include <machine/specialreg.h>
#include <machine/clock.h>
#include <sys/indefinite2.h>
#include <sys/spinlock.h>
#include <sys/spinlock2.h>
#include <sys/ktr.h>

#ifdef _KERNEL_VIRTUAL
#include <pthread.h>
#endif

struct spinlock pmap_spin = SPINLOCK_INITIALIZER(pmap_spin, "pmap_spin");

/*
 * Kernal Trace
 */
#if !defined(KTR_SPIN_CONTENTION)
#define KTR_SPIN_CONTENTION     KTR_ALL
#endif
#define SPIN_STRING     "spin=%p type=%c"
#define SPIN_ARG_SIZE   (sizeof(void *) + sizeof(int))

KTR_INFO_MASTER(spin);
#if 0
KTR_INFO(KTR_SPIN_CONTENTION, spin, beg, 0, SPIN_STRING, SPIN_ARG_SIZE);
KTR_INFO(KTR_SPIN_CONTENTION, spin, end, 1, SPIN_STRING, SPIN_ARG_SIZE);
#endif

#define logspin(name, spin, type)                       \
        KTR_LOG(spin_ ## name, spin, type)

#ifdef INVARIANTS
static int spin_lock_test_mode;
#endif

#ifdef DEBUG_LOCKS_LATENCY

static long spinlocks_add_latency;
SYSCTL_LONG(_debug, OID_AUTO, spinlocks_add_latency, CTLFLAG_RW,
    &spinlocks_add_latency, 0,
    "Add spinlock latency");

#endif

static long spin_backoff_max = 4096;
SYSCTL_LONG(_debug, OID_AUTO, spin_backoff_max, CTLFLAG_RW,
    &spin_backoff_max, 0,
    "Spinlock exponential backoff limit");
static long spin_window_shift = 8;      /* 1 << n clock cycles, approx */
SYSCTL_LONG(_debug, OID_AUTO, spin_window_shift, CTLFLAG_RW,
    &spin_window_shift, 0,
    "Spinlock TSC windowing");

/*
 * We contested due to another exclusive lock holder.  We lose.
 *
 * We have to unwind the attempt and may acquire the spinlock
 * anyway while doing so.
 */
int
spin_trylock_contested(struct spinlock *spin)
{
        globaldata_t gd = mycpu;

        /*
         * Handle degenerate case, else fail.
         */
        if (atomic_cmpset_int(&spin->counta, SPINLOCK_SHARED|0, 1))
                return TRUE;
        /*atomic_add_int(&spin->counta, -1);*/
        --gd->gd_spinlocks;
        crit_exit_raw(gd->gd_curthread);

        return (FALSE);
}

/*
 * The spin_lock() inline was unable to acquire the lock and calls this
 * function with spin->counta already incremented, passing (spin->counta - 1)
 * to the function (the result of the inline's fetchadd).
 *
 * Note that we implement both exclusive and shared spinlocks, so we cannot
 * use atomic_swap_int().  Instead, we try to use atomic_fetchadd_int()
 * to put most of the burden on the cpu.  Atomic_cmpset_int() (cmpxchg)
 * can cause a lot of unnecessary looping in situations where it is just
 * trying to increment the count.
 *
 * Similarly, we leave the SHARED flag intact and incur slightly more
 * overhead when switching from shared to exclusive.  This allows us to
 * use atomic_fetchadd_int() for both spinlock types in the critical
 * path.
 *
 * The exponential (n^1.5) backoff algorithm is designed to both reduce
 * cache bus contention between cpu cores and sockets, and to allow some
 * bursting of exclusive locks in heavily contended situations to improve
 * performance.
 *
 * The exclusive lock priority mechanism prevents even heavily contended
 * exclusive locks from being starved by shared locks
 */
void
_spin_lock_contested(struct spinlock *spin, const char *ident, int value)
{
        indefinite_info_t info;
        uint32_t ovalue;
        long expbackoff;
        long loop;

        /*
         * WARNING! Caller has already incremented the lock.  We must
         *          increment the count value (from the inline's fetch-add)
         *          to match.
         *
         * Handle the degenerate case where the spinlock is flagged SHARED
         * with only our reference.  We can convert it to EXCLUSIVE.
         */
        if (value == (SPINLOCK_SHARED | 1) - 1) {
                if (atomic_cmpset_int(&spin->counta, SPINLOCK_SHARED | 1, 1))
                        return;
        }
        /* ++value; value not used after this */
        info.type = 0;          /* avoid improper gcc warning */
        info.ident = NULL;      /* avoid improper gcc warning */
        expbackoff = 0;

        /*
         * Transfer our exclusive request to the high bits and clear the
         * SPINLOCK_SHARED bit if it was set.  This makes the spinlock
         * appear exclusive, preventing any NEW shared or exclusive
         * spinlocks from being obtained while we wait for existing
         * shared or exclusive holders to unlock.
         *
         * Don't tread on earlier exclusive waiters by stealing the lock
         * away early if the low bits happen to now be 1.
         *
         * The shared unlock understands that this may occur.
         */
        ovalue = atomic_fetchadd_int(&spin->counta, SPINLOCK_EXCLWAIT - 1);
        ovalue += SPINLOCK_EXCLWAIT - 1;
        if (ovalue & SPINLOCK_SHARED) {
                atomic_clear_int(&spin->counta, SPINLOCK_SHARED);
                ovalue &= ~SPINLOCK_SHARED;
        }

        for (;;) {
                expbackoff = (expbackoff + 1) * 3 / 2;
                if (expbackoff == 6)            /* 1, 3, 6, 10, ... */
                        indefinite_init(&info, ident, 0, 'S');
                if ((rdtsc() >> spin_window_shift) % ncpus != mycpuid)  {
                        for (loop = expbackoff; loop; --loop)
                                cpu_pause();
                }
                /*cpu_lfence();*/

                /*
                 * If the low bits are zero, try to acquire the exclusive lock
                 * by transfering our high bit reservation to the low bits.
                 *
                 * NOTE: Avoid unconditional atomic op by testing ovalue,
                 *       otherwise we get cache bus armageddon.
                 *
                 * NOTE: We must also ensure that the SHARED bit is cleared.
                 *       It is possible for it to wind up being set on a
                 *       shared lock override of the EXCLWAIT bits.
                 */
                ovalue = spin->counta;
                cpu_ccfence();
                if ((ovalue & (SPINLOCK_EXCLWAIT - 1)) == 0) {
                        uint32_t nvalue;

                        nvalue= ((ovalue - SPINLOCK_EXCLWAIT) | 1) &
                                ~SPINLOCK_SHARED;
                        if (atomic_fcmpset_int(&spin->counta, &ovalue, nvalue))
                                break;
                        continue;
                }
                if (expbackoff > 6 + spin_backoff_max)
                        expbackoff = 6 + spin_backoff_max;
                if (expbackoff >= 6) {
                        if (indefinite_check(&info))
                                break;
                }
        }
        if (expbackoff >= 6)
                indefinite_done(&info);
}

/*
 * The spin_lock_shared() inline was unable to acquire the lock and calls
 * this function with spin->counta already incremented.
 *
 * This is not in the critical path unless there is contention between
 * shared and exclusive holders.
 *
 * Exclusive locks have priority over shared locks.  However, this can
 * cause shared locks to be starved when large numbers of threads are
 * competing for exclusive locks so the shared lock code uses TSC-windowing
 * to selectively ignore the exclusive priority mechanism.  This has the
 * effect of allowing a limited number of shared locks to compete against
 * exclusive waiters at any given moment.
 *
 * Note that shared locks do not implement exponential backoff.  Instead,
 * the shared lock simply polls the lock value.  One cpu_pause() is built
 * into indefinite_check().
 */
void
_spin_lock_shared_contested(struct spinlock *spin, const char *ident)
{
        indefinite_info_t info;
        uint32_t ovalue;

        /*
         * Undo the inline's increment.
         */
        ovalue = atomic_fetchadd_int(&spin->counta, -1) - 1;

        indefinite_init(&info, ident, 0, 's');
        cpu_pause();

#ifdef DEBUG_LOCKS_LATENCY
        long j;
        for (j = spinlocks_add_latency; j > 0; --j)
                cpu_ccfence();
#endif

        for (;;) {
                /*
                 * Loop until we can acquire the shared spinlock.  Note that
                 * the low bits can be zero while the high EXCLWAIT bits are
                 * non-zero.  In this situation exclusive requesters have
                 * priority (otherwise shared users on multiple cpus can hog
                 * the spinlnock).
                 *
                 * NOTE: Reading spin->counta prior to the swap is extremely
                 *       important on multi-chip/many-core boxes.  On 48-core
                 *       this one change improves fully concurrent all-cores
                 *       compiles by 100% or better.
                 *
                 *       I can't emphasize enough how important the pre-read
                 *       is in preventing hw cache bus armageddon on
                 *       multi-chip systems.  And on single-chip/multi-core
                 *       systems it just doesn't hurt.
                 */
                cpu_ccfence();

                /*
                 * Ignore the EXCLWAIT bits if we are inside our window.
                 */
                if ((ovalue & (SPINLOCK_EXCLWAIT - 1)) == 0 &&
                    (rdtsc() >> spin_window_shift) % ncpus == mycpuid)  {
                        if (atomic_fcmpset_int(&spin->counta, &ovalue,
                                               ovalue | SPINLOCK_SHARED | 1)) {
                                break;
                        }
                        continue;
                }

                /*
                 * Check ovalue tightly (no exponential backoff for shared
                 * locks, that would result in horrible performance.  Instead,
                 * shared locks depend on the exclusive priority mechanism
                 * to avoid starving exclusive locks).
                 */
                if (ovalue == 0) {
                        if (atomic_fcmpset_int(&spin->counta, &ovalue,
                                              SPINLOCK_SHARED | 1)) {
                                break;
                        }
                        continue;
                }

                /*
                 * If SHARED is already set, go for the increment, improving
                 * the exclusive to multiple-readers transition.
                 */
                if (ovalue & SPINLOCK_SHARED) {
                        ovalue = atomic_fetchadd_int(&spin->counta, 1);
                        /* ovalue += 1; NOT NEEDED */
                        if (ovalue & SPINLOCK_SHARED)
                                break;
                        ovalue = atomic_fetchadd_int(&spin->counta, -1);
                        ovalue += -1;
                        continue;
                }
                if (indefinite_check(&info))
                        break;
                /*
                 * ovalue was wrong anyway, just reload
                 */
                ovalue = spin->counta;
        }
        indefinite_done(&info);
}

/*
 * If INVARIANTS is enabled various spinlock timing tests can be run
 * by setting debug.spin_lock_test:
 *
 *      1       Test the indefinite wait code
 *      2       Time the best-case exclusive lock overhead (spin_test_count)
 *      3       Time the best-case shared lock overhead (spin_test_count)
 */

#ifdef INVARIANTS

static int spin_test_count = 10000000;
SYSCTL_INT(_debug, OID_AUTO, spin_test_count, CTLFLAG_RW, &spin_test_count, 0,
    "Number of iterations to use for spinlock wait code test");

static int
sysctl_spin_lock_test(SYSCTL_HANDLER_ARGS)
{
        struct spinlock spin;
        int error;
        int value = 0;
        int i;

        if ((error = priv_check(curthread, PRIV_ROOT)) != 0)
                return (error);
        if ((error = SYSCTL_IN(req, &value, sizeof(value))) != 0)
                return (error);

        /*
         * Indefinite wait test
         */
        if (value == 1) {
                spin_init(&spin, "sysctllock");
                spin_lock(&spin);       /* force an indefinite wait */
                spin_lock_test_mode = 1;
                spin_lock(&spin);
                spin_unlock(&spin);     /* Clean up the spinlock count */
                spin_unlock(&spin);
                spin_lock_test_mode = 0;
        }

        /*
         * Time best-case exclusive spinlocks
         */
        if (value == 2) {
                globaldata_t gd = mycpu;

                spin_init(&spin, "sysctllocktest");
                for (i = spin_test_count; i > 0; --i) {
                    _spin_lock_quick(gd, &spin, "test");
                    spin_unlock_quick(gd, &spin);
                }
        }

        return (0);
}

SYSCTL_PROC(_debug, KERN_PROC_ALL, spin_lock_test, CTLFLAG_RW|CTLTYPE_INT,
        0, 0, sysctl_spin_lock_test, "I", "Test spinlock wait code");

#endif  /* INVARIANTS */