/* -*- Mode: C++; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
/* This Source Code Form is subject to the terms of the Mozilla Public
 * License, v. 2.0. If a copy of the MPL was not distributed with this
 * file, You can obtain one at http://mozilla.org/MPL/2.0/. */

/*
 *  w95cv.c -- Windows 95 Machine-Dependent Code for Condition Variables
 *
 *  We implement our own condition variable wait queue.  Each thread
 *  has a semaphore object (thread->md.blocked_sema) to block on while
 *  waiting on a condition variable.
 *
 *  We use a deferred condition notify algorithm.  When PR_NotifyCondVar
 *  or PR_NotifyAllCondVar is called, the condition notifies are simply
 *  recorded in the _MDLock structure.  We defer the condition notifies
 *  until right after we unlock the lock.  This way the awakened threads
 *  have a better chance to reaquire the lock.
 */
 
#include "primpl.h"

/*
 * AddThreadToCVWaitQueueInternal --
 *
 * Add the thread to the end of the condition variable's wait queue.
 * The CV's lock must be locked when this function is called.
 */

static void
AddThreadToCVWaitQueueInternal(PRThread *thred, struct _MDCVar *cv)
{
    PR_ASSERT((cv->waitTail != NULL && cv->waitHead != NULL)
            || (cv->waitTail == NULL && cv->waitHead == NULL));
    cv->nwait += 1;
    thred->md.inCVWaitQueue = PR_TRUE;
    thred->md.next = NULL;
    thred->md.prev = cv->waitTail;
    if (cv->waitHead == NULL) {
        cv->waitHead = thred;
    } else {
        cv->waitTail->md.next = thred;
    }
    cv->waitTail = thred;
}

/*
 * md_UnlockAndPostNotifies --
 *
 * Unlock the lock, and then do the deferred condition notifies.
 * If waitThred and waitCV are not NULL, waitThred is added to
 * the wait queue of waitCV before the lock is unlocked.
 *
 * This function is called by _PR_MD_WAIT_CV and _PR_MD_UNLOCK,
 * the two places where a lock is unlocked.
 */
static void
md_UnlockAndPostNotifies(
    _MDLock *lock,
    PRThread *waitThred,
    _MDCVar *waitCV)
{
    PRIntn index;
    _MDNotified post;
    _MDNotified *notified, *prev = NULL;

    /*
     * Time to actually notify any conditions that were affected
     * while the lock was held.  Get a copy of the list that's in
     * the lock structure and then zero the original.  If it's
     * linked to other such structures, we own that storage.
     */
    post = lock->notified;  /* a safe copy; we own the lock */

#if defined(DEBUG)
    ZeroMemory(&lock->notified, sizeof(_MDNotified));  /* reset */
#else
    lock->notified.length = 0;  /* these are really sufficient */
    lock->notified.link = NULL;
#endif

    /* 
     * Figure out how many threads we need to wake up.
     */
    notified = &post;  /* this is where we start */
    do {
        for (index = 0; index < notified->length; ++index) {
            _MDCVar *cv = notified->cv[index].cv;
            PRThread *thred;
            int i;
            
            /* Fast special case: no waiting threads */
            if (cv->waitHead == NULL) {
                notified->cv[index].notifyHead = NULL;
                continue;
            }

            /* General case */
            if (-1 == notified->cv[index].times) {
                /* broadcast */
                thred = cv->waitHead;
                while (thred != NULL) {
                    thred->md.inCVWaitQueue = PR_FALSE;
                    thred = thred->md.next;
                }
                notified->cv[index].notifyHead = cv->waitHead;
                cv->waitHead = cv->waitTail = NULL;
                cv->nwait = 0;
            } else {
                thred = cv->waitHead;
                i = notified->cv[index].times;
                while (thred != NULL && i > 0) {
                    thred->md.inCVWaitQueue = PR_FALSE;
                    thred = thred->md.next;
                    i--;
                }
                notified->cv[index].notifyHead = cv->waitHead;
                cv->waitHead = thred;
                if (cv->waitHead == NULL) {
                    cv->waitTail = NULL;
                } else {
                    if (cv->waitHead->md.prev != NULL) {
                        cv->waitHead->md.prev->md.next = NULL;
                        cv->waitHead->md.prev = NULL;
                    }
                }
                cv->nwait -= notified->cv[index].times - i;
            }
        }
        notified = notified->link;
    } while (NULL != notified);

    if (waitThred) {
        AddThreadToCVWaitQueueInternal(waitThred, waitCV);
    }

    /* Release the lock before notifying */
        LeaveCriticalSection(&lock->mutex);

    notified = &post;  /* this is where we start */
    do {
        for (index = 0; index < notified->length; ++index) {
            PRThread *thred;
            PRThread *next;

            thred = notified->cv[index].notifyHead;
            while (thred != NULL) {
                BOOL rv;

                next = thred->md.next;
                thred->md.prev = thred->md.next = NULL;

                rv = ReleaseSemaphore(thred->md.blocked_sema, 1, NULL);
                PR_ASSERT(rv != 0);
                thred = next;
            }
        }
        prev = notified;
        notified = notified->link;
        if (&post != prev) PR_DELETE(prev);
    } while (NULL != notified);
}

/*
 * Notifies just get posted to the protecting mutex.  The
 * actual notification is done when the lock is released so that
 * MP systems don't contend for a lock that they can't have.
 */
static void md_PostNotifyToCvar(_MDCVar *cvar, _MDLock *lock,
        PRBool broadcast)
{
    PRIntn index = 0;
    _MDNotified *notified = &lock->notified;

    while (1) {
        for (index = 0; index < notified->length; ++index) {
            if (notified->cv[index].cv == cvar) {
                if (broadcast) {
                    notified->cv[index].times = -1;
                } else if (-1 != notified->cv[index].times) {
                    notified->cv[index].times += 1;
                }
                return;
            }
        }
        /* if not full, enter new CV in this array */
        if (notified->length < _MD_CV_NOTIFIED_LENGTH) break;

        /* if there's no link, create an empty array and link it */
        if (NULL == notified->link) {
            notified->link = PR_NEWZAP(_MDNotified);
        }

        notified = notified->link;
    }

    /* A brand new entry in the array */
    notified->cv[index].times = (broadcast) ? -1 : 1;
    notified->cv[index].cv = cvar;
    notified->length += 1;
}

/*
 * _PR_MD_NEW_CV() -- Creating new condition variable
 * ... Solaris uses cond_init() in similar function.
 *
 * returns: -1 on failure
 *          0 when it succeeds.
 *
 */
PRInt32 
_PR_MD_NEW_CV(_MDCVar *cv)
{
    cv->magic = _MD_MAGIC_CV;
    /*
     * The waitHead, waitTail, and nwait fields are zeroed
     * when the PRCondVar structure is created.
     */
    return 0;
} 

void _PR_MD_FREE_CV(_MDCVar *cv)
{
    cv->magic = (PRUint32)-1;
    return;
}

/*
 *  _PR_MD_WAIT_CV() -- Wait on condition variable
 */
void _PR_MD_WAIT_CV(_MDCVar *cv, _MDLock *lock, PRIntervalTime timeout )
{
    PRThread *thred = _PR_MD_CURRENT_THREAD();
    DWORD rv;
    DWORD msecs = (timeout == PR_INTERVAL_NO_TIMEOUT) ?
            INFINITE : PR_IntervalToMilliseconds(timeout);

    /*
     * If we have pending notifies, post them now.
     */
    if (0 != lock->notified.length) {
        md_UnlockAndPostNotifies(lock, thred, cv);
    } else {
        AddThreadToCVWaitQueueInternal(thred, cv);
        LeaveCriticalSection(&lock->mutex);
    }

    /* Wait for notification or timeout; don't really care which */
    rv = WaitForSingleObject(thred->md.blocked_sema, msecs);

    EnterCriticalSection(&(lock->mutex));

    PR_ASSERT(rv != WAIT_ABANDONED);
    PR_ASSERT(rv != WAIT_FAILED);
    PR_ASSERT(rv != WAIT_OBJECT_0 || thred->md.inCVWaitQueue == PR_FALSE);

    if (rv == WAIT_TIMEOUT) {
        if (thred->md.inCVWaitQueue) {
            PR_ASSERT((cv->waitTail != NULL && cv->waitHead != NULL)
                    || (cv->waitTail == NULL && cv->waitHead == NULL));
            cv->nwait -= 1;
            thred->md.inCVWaitQueue = PR_FALSE;
            if (cv->waitHead == thred) {
                cv->waitHead = thred->md.next;
                if (cv->waitHead == NULL) {
                    cv->waitTail = NULL;
                } else {
                    cv->waitHead->md.prev = NULL;
                }
            } else {
                PR_ASSERT(thred->md.prev != NULL);
                thred->md.prev->md.next = thred->md.next;
                if (thred->md.next != NULL) {
                    thred->md.next->md.prev = thred->md.prev;
                } else {
                    PR_ASSERT(cv->waitTail == thred);
                    cv->waitTail = thred->md.prev;
                }
            }
            thred->md.next = thred->md.prev = NULL;
        } else {
            /*
             * This thread must have been notified, but the
             * ReleaseSemaphore call happens after WaitForSingleObject
             * times out.  Wait on the semaphore again to make it
             * non-signaled.  We assume this wait won't take long.
             */
            rv = WaitForSingleObject(thred->md.blocked_sema, INFINITE);
            PR_ASSERT(rv == WAIT_OBJECT_0);
        }
    }
    PR_ASSERT(thred->md.inCVWaitQueue == PR_FALSE);
    return;
} /* --- end _PR_MD_WAIT_CV() --- */

void _PR_MD_NOTIFY_CV(_MDCVar *cv, _MDLock *lock)
{
    md_PostNotifyToCvar(cv, lock, PR_FALSE);
    return;
}

void _PR_MD_NOTIFYALL_CV(_MDCVar *cv, _MDLock *lock)
{
    md_PostNotifyToCvar(cv, lock, PR_TRUE);
    return;
}

typedef BOOL (WINAPI *INITIALIZECRITICALSECTIONEX)(
    CRITICAL_SECTION *lpCriticalSection,
    DWORD dwSpinCount,
    DWORD Flags);

static INITIALIZECRITICALSECTIONEX sInitializeCriticalSectionEx;

void _PR_MD_INIT_LOCKS(void)
{
    /*
     * Starting with Windows Vista, every CRITICAL_SECTION allocates an extra
     * RTL_CRITICAL_SECTION_DEBUG object. Unfortunately, this debug object is
     * not reclaimed by DeleteCriticalSection(), causing an apparent memory
     * leak. This is a debugging "feature", not a bug. If we are running on
     * Vista or later, use InitializeCriticalSectionEx() to allocate
     * CRITICAL_SECTIONs without debug objects.
     */
    HMODULE hKernel32 = GetModuleHandle("kernel32.dll");
    PR_ASSERT(hKernel32);
    PR_ASSERT(!sInitializeCriticalSectionEx);
    sInitializeCriticalSectionEx = (INITIALIZECRITICALSECTIONEX)
            GetProcAddress(hKernel32, "InitializeCriticalSectionEx");
}

/*
 * By default, CRITICAL_SECTIONs are initialized with a spin count of 0.
 * Joe Duffy's "Concurrent Programming on Windows" book suggests 1500 is
 * a "reasonable starting point". On single-processor systems, the spin
 * count is ignored and the critical section spin count is set to 0.
 */
#define LOCK_SPIN_COUNT 1500

PRStatus _PR_MD_NEW_LOCK(_MDLock *lock)
{
    CRITICAL_SECTION *cs = &lock->mutex;
    BOOL ok;

    if (sInitializeCriticalSectionEx) {
        ok = sInitializeCriticalSectionEx(cs, LOCK_SPIN_COUNT,
                                          CRITICAL_SECTION_NO_DEBUG_INFO);
    } else {
        ok = InitializeCriticalSectionAndSpinCount(cs, LOCK_SPIN_COUNT);
    }
    if (!ok) {
        _PR_MD_MAP_DEFAULT_ERROR(GetLastError());
        return PR_FAILURE;
    }

    lock->notified.length = 0;
    lock->notified.link = NULL;
    return PR_SUCCESS;
}

void _PR_MD_UNLOCK(_MDLock *lock)
{
    if (0 != lock->notified.length) {
        md_UnlockAndPostNotifies(lock, NULL, NULL);
    } else {
        LeaveCriticalSection(&lock->mutex);
    }
}