使用php trie_filter扩展，cpu高负载问题排查

       今天一同事在使用php trie_filter跑脚本，循环匹配关键词，发现cpu利用率达到了99%。经过strace命令跟踪，发现进程在反复的打开读取和关闭.tree文件，定位到trie_filter_load这个方法被反复调用。按正常的状况，这个文件应该会一次读到内存中，在内存中查询匹配关键词。为了证明，我追踪到扩展源码以下：

PHP_FUNCTION(trie_filter_load)
{
    Trie *trie;
    char *path;
    int path_len;

    if (zend_parse_parameters(ZEND_NUM_ARGS() TSRMLS_CC, "s",
                &path, &path_len) == FAILURE) {
        RETURN_NULL();
    }

    trie = trie_new_from_file(path);
    if (!trie) {
        php_error_docref(NULL TSRMLS_CC, E_WARNING,
                "Unable to load %s", path);
        RETURN_NULL();
    }

    ZEND_REGISTER_RESOURCE(return_value, trie, le_trie_filter);
}

 

trie变量保存的是一个Trie *结构体，还不能保证已读取到内存中，继续打开libdatrie源码，查找到.datrie/trie.c

Trie *
trie_new_from_file (const char *path)
{
    Trie       *trie;
    FILE       *trie_file;

    trie_file = fopen (path, "r");
    if (!trie_file)
        return NULL;

    trie = trie_fread (trie_file);
    fclose (trie_file);
    return trie;
}

粗看之下trie 结构体读取并保存了文件内容，再继续定位trie_fread

Trie *
trie_fread (FILE *file)
{
    Trie       *trie;

    trie = (Trie *) malloc (sizeof (Trie));
    if (!trie)
        return NULL;

    if (NULL == (trie->alpha_map = alpha_map_fread_bin (file)))
        goto exit_trie_created;
    if (NULL == (trie->da   = da_fread (file)))
        goto exit_alpha_map_created;
    if (NULL == (trie->tail = tail_fread (file)))
        goto exit_da_created;

    trie->is_dirty = FALSE;
    return trie;

exit_da_created:
    da_free (trie->da);
exit_alpha_map_created:
    alpha_map_free (trie->alpha_map);
exit_trie_created:
    free (trie);
    return NULL;
}

经过alpha_map_fread_bin方法读取二进制文件到trie->alpha_map

AlphaMap *
alpha_map_fread_bin (FILE *file)
{
    long        save_pos;
    uint32      sig;
    int32       total, i;
    AlphaMap   *alpha_map;

    /* check signature */
    save_pos = ftell (file);
    if (!file_read_int32 (file, (int32 *) &sig) || ALPHAMAP_SIGNATURE != sig)
        goto exit_file_read;

    if (NULL == (alpha_map = alpha_map_new ()))
        goto exit_file_read;

    /* read number of ranges */
    if (!file_read_int32 (file, &total))
        goto exit_map_created;

    /* read character ranges */
    for (i = 0; i < total; i++) {
        int32   b, e;

        if (!file_read_int32 (file, &b) || !file_read_int32 (file, &e))
            goto exit_map_created;
        alpha_map_add_range (alpha_map, b, e);
    }

    return alpha_map;

exit_map_created:
    alpha_map_free (alpha_map);
exit_file_read:
    fseek (file, save_pos, SEEK_SET);
    return NULL;
}

Bool
file_read_int32 (FILE *file, int32 *o_val)
{
    unsigned char   buff[4];

    if (fread (buff, 4, 1, file) == 1) {
        *o_val = (buff[0] << 24) | (buff[1] << 16) |  (buff[2] << 8) | buff[3];
        return TRUE;
    }

    return FALSE;
}

至此已肯定trie_filter_load方法在打开tree文件后读取到内存，所以将trie_filter_load的返回值赋给静态变量，就不用反复打开读取关闭tree文件了，修改后跟踪进程恢复正常。