berkeley db 内存池 LRU算法

标签：

priority based lru

in src/mp/mp_fget.c, __memp_fget(), 初始化 一个page buffer时, 设置其 priority:

    bhp->priority = MPOOL_LRU_REDZONE;
    

in src/mp/mp_fget.c, __memp_fput(), 对一个page buffer做put, 使其reference减一. bhp->ref--, 若reference减为0, 则调整其 priority. 下面的 bhp->priority = c_mp->lru_priority; 和 ++c_mp->lru_priority 保证LRU; 而且根据传入的 priority参数做相应的调整.
prioriry最大为UINT32_MAX, 所以有可能需要 wraparound. __memp_reset_lru() 降低 所有page buffer的priority.

/* Update priority values. */
if (priority == DB_PRIORITY_VERY_LOW ||
    mfp->priority == MPOOL_PRI_VERY_LOW)
    bhp->priority = 0;
else {
    /*
     * We don‘t lock the LRU priority or the pages field, if
     * we get garbage (which won‘t happen on a 32-bit machine), it
     * only means a buffer has the wrong priority.
     */
    bhp->priority = c_mp->lru_priority;

    switch (priority) {
    default:
    case DB_PRIORITY_UNCHANGED:
        pfactor = mfp->priority;
        break;
    case DB_PRIORITY_VERY_LOW:
        pfactor = MPOOL_PRI_VERY_LOW;
        break;
    case DB_PRIORITY_LOW:
        pfactor = MPOOL_PRI_LOW;
        break;
    case DB_PRIORITY_DEFAULT:
        pfactor = MPOOL_PRI_DEFAULT;
        break;
    case DB_PRIORITY_HIGH:
        pfactor = MPOOL_PRI_HIGH;
        break;
    case DB_PRIORITY_VERY_HIGH:
        pfactor = MPOOL_PRI_VERY_HIGH;
        break;
    }

    adjust = 0;
    if (pfactor != 0)
        adjust = (int)c_mp->pages / pfactor;

    if (F_ISSET(bhp, BH_DIRTY))
        adjust += (int)c_mp->pages / MPOOL_PRI_DIRTY;

    if (adjust > 0) {
        if (MPOOL_LRU_REDZONE - bhp->priority >=
            (u_int32_t)adjust)
            bhp->priority += adjust;
    } else if (adjust < 0)
        if (bhp->priority > (u_int32_t)-adjust)
            bhp->priority += adjust;
}

...

/*
 * On every buffer put we update the cache lru priority and check
 * for wraparound. The increment doesn‘t need to be atomic: occasional
 * lost increments are okay; __memp_reset_lru handles race conditions.
 */
if (++c_mp->lru_priority >= MPOOL_LRU_REDZONE &&
    (t_ret = __memp_reset_lru(env, infop)) != 0 && ret == 0)
    ret = t_ret;

return (ret);

src/mp/mp_alloc.c, __memp_alloc(), 无法分配新的 buffer的内存时, 则重用已有的 buffer内存. 算法很复杂, 单独写一个吧.

berkeley db 内存池 LRU算法

标签：

原文地址：http://www.cnblogs.com/brayden/p/5231249.html