【MySQL】filesort.cc 源代码阅读笔记
最近阅读了部分MySQL排序的代码,把心得记录一下。
参考代码
MySQL: 文件: filesort.cc 函数: filesort()
排序过程伪代码
function filesort(table_sort) { param.init_for_filesort(); //是否可以使用优先队列加速? if (check_if_pq_applicable()) { // check_if_pq_applicable 已经分配了合适的sort_buffer // 这里初始化sort_buffer中的指针区域。sort_buffer中 // 数据的布局参考后文。 table_sort.init_record_pointers() } else { table_sort.alloc_sort_buffer(); } //扫描,初步处理,可能已经做了分片排序或者Top-N处理(PQ) find_all_keys(); if (all_records_is_memory) { //所有数据都在内存里面,直接排序 sort_in_memory_buffer(); } else { //使用了临时文件,需要多路合并排序 merge_many_buff(); } } function find_all_keys() { ; while((row=get_next_key()) { //get_next_key有可能读取的是排序字段+rowid //也有可能读取的是排序字段+addon_fields if (use_pq) { //使用优先队列 pq.push(row) } else { filesort_buffer.append(row); if (filesort_buffer.full()) { filesort_buffer.sort_buffer(); write_to_temp_file(filesort_buffer, temp_file); write_temp_file_num += ; } } } ) { filesort_buffer.sort_buffer(); write_to_temp_file(filesort_buffer, temp_file); write_temp_file_num += ; } } void Filesort_buffer::sort_buffer() { if (radixsort_is_appliccable()) //是否可以用基数排序? { radixsort(); return; } ) // 100个以内用快速排序 { my_qsort2(data, count, get_ptr_compare(sort_length)); } //排序的时候,直接使用内存比较 std::stable_sort(data, n, Mem_compare(sort_length)); } function merge_many_buff(buffs[], count) { //这两个数字是写死的 #define MERGEBUFF 7 #define MERGEBUFF2 15 if (count < MERGEBUFFF2) { return; } //归并排序:7合1,直到剩余总分片数量不多于15个 while (count > MERGEBUFF2) { int i; ; i <= count - MERGEBUFF * / ; i += MERGEBUFF) { first = buffs + i last = buffs + i + MERGEBUFF - merge_buffers(first, last); } first = buffs + i last = buffs + count merge_buffers(buffs + i, buffs + count); count = recount_buffs() } }
filesort.cc走读
/** Sort a table. Creates a set of pointers that can be used to read the rows in sorted order. This should be done with the functions in records.cc. Before calling filesort, one must have done table->file->info(HA_STATUS_VARIABLE) The result set is stored in table->io_cache or table->record_pointers. @param thd Current thread @param table Table to sort @param filesort How to sort the table @param sort_positions Set to TRUE if we want to force sorting by position (Needed by UPDATE/INSERT or ALTER TABLE or when rowids are required by executor) @param[out] examined_rows Store number of examined rows here This is the number of found rows before applying WHERE condition. @param[out] found_rows Store the number of found rows here. This is the number of found rows after applying WHERE condition. @note If we sort by position (like if sort_positions is 1) filesort() will call table->prepare_for_position(). @retval HA_POS_ERROR Error @retval \# Number of rows in the result, could be less than found_rows if LIMIT were provided. */ ha_rows filesort(THD *thd, TABLE *table, Filesort *filesort, bool sort_positions, ha_rows *examined_rows, ha_rows *found_rows);
第一步,初始化排序基本参数
//初始化排序参数 void Sort_param::init_for_filesort(uint sortlen, TABLE *table, ulong max_length_for_sort_data, ha_rows maxrows, bool sort_positions) { sort_length= sortlen; ref_length= table->file->ref_length; if (!(table->file->ha_table_flags() & HA_FAST_KEY_READ) && !table->fulltext_searched && !sort_positions) { /* Get the descriptors of all fields whose values are appended to sorted fields and get its total length in addon_length. */ /* get_addon_fields 构造查询字段的描述信息。如果返回0,表示应该使用回表排序,如果非0,用就地排序。 */ addon_field= get_addon_fields(max_length_for_sort_data, table->field, sort_length, &addon_length); } if (addon_field) res_length= addon_length; else { res_length= ref_length; /* The reference to the record is considered as an additional sorted field */ /* 这里的reference to the record 一般就是rowid, 这里会把rowid当作排序最后一个字段 */ sort_length+= ref_length; } rec_length= sort_length + addon_length; // 整个排序行的长度 max_rows= maxrows; }
第二步,判断是否可以使用优先队列
/** Test whether priority queue is worth using to get top elements of an ordered result set. If it is, then allocates buffer for required amount of records @param trace Current trace context. @param param Sort parameters. @param filesort_info Filesort information. @param table Table to sort. @param num_rows Estimate of number of rows in source record set. @param memory_available Memory available for sorting. DESCRIPTION Given a query like this: SELECT ... FROM t ORDER BY a1,...,an LIMIT max_rows; This function tests whether a priority queue should be used to keep the result. Necessary conditions are: - estimate that it is actually cheaper than merge-sort - enough memory to store the <max_rows> records. If we don't have space for <max_rows> records, but we *do* have space for <max_rows> keys, we may rewrite 'table' to sort with references to records instead of additional data. (again, based on estimates that it will actually be cheaper). @retval true - if it's ok to use PQ false - PQ will be slower than merge-sort, or there is not enough memory. */ bool check_if_pq_applicable(Opt_trace_context *trace, Sort_param *param, Filesort_info *filesort_info, TABLE *table, ha_rows num_rows, ulong memory_available);
第三步,全表扫描
/** Search after sort_keys, and write them into tempfile (if we run out of space in the sort_keys buffer). All produced sequences are guaranteed to be non-empty. @param param Sorting parameter @param select Use this to get source data @param sort_keys Array of pointers to sort key + addon buffers. @param buffpek_pointers File to write BUFFPEKs describing sorted segments in tempfile. @param tempfile File to write sorted sequences of sortkeys to. @param pq If !NULL, use it for keeping top N elements @param [out] found_rows The number of FOUND_ROWS(). For a query with LIMIT, this value will typically be larger than the function return value. @note Basic idea: @verbatim while (get_next_sortkey()) { if (using priority queue) push sort key into queue else { if (no free space in sort_keys buffers) { sort sort_keys buffer; dump sorted sequence to 'tempfile'; dump BUFFPEK describing sequence location into 'buffpek_pointers'; } put sort key into 'sort_keys'; } } if (sort_keys has some elements && dumped at least once) sort-dump-dump as above; else don't sort, leave sort_keys array to be sorted by caller. @endverbatim @retval Number of records written on success. @retval HA_POS_ERROR on error. */ static ha_rows find_all_keys(Sort_param *param, SQL_SELECT *select, Filesort_info *fs_info, IO_CACHE *buffpek_pointers, IO_CACHE *tempfile, Bounded_queue<uchar, uchar> *pq, ha_rows *found_rows)
第四步,排序
如果结果集全部在内存中,直接用Filesort_buffer::sort_buffer()。
如果需要合并排序,调用 merge_many_buffs进行合并。
Filesort_buffer
- 如果查询字段里面有BLOB格式,使用回表模式。
- 否则,如果排序字段加上查询字段总长度不超过max_length_for_sort_data,使用就地模式。
- 否则,如果超过,使用回表模式。
/** Get descriptors of fields appended to sorted fields and calculate its total length. The function first finds out what fields are used in the result set. Then it calculates the length of the buffer to store the values of these fields together with the value of sort values. If the calculated length is not greater than max_length_for_sort_data the function allocates memory for an array of descriptors containing layouts for the values of the non-sorted fields in the buffer and fills them. @param thd Current thread @param ptabfield Array of references to the table fields @param sortlength Total length of sorted fields @param[out] plength Total length of appended fields @note The null bits for the appended values are supposed to be put together and stored the buffer just ahead of the value of the first field. @return Pointer to the layout descriptors for the appended fields, if any @retval NULL if we do not store field values with sort data. */ static SORT_ADDON_FIELD * get_addon_fields(ulong max_length_for_sort_data, Field **ptabfield, uint sortlength, uint *plength) { // ... MY_BITMAP *read_set= (*ptabfield)->table->read_set; /* If there is a reference to a field in the query add it to the the set of appended fields. Note for future refinement: This this a too strong condition. Actually we need only the fields referred in the result set. And for some of them it makes sense to use the values directly from sorted fields. 目前实现:只要是read_set里面的字段就作为addon_fields,有点浪费 优化方向:只需要把结果集的字段加到addon_fields里面 例如 select a, b from table where c>0; c 这个字段也会在 addon_fields (?) 实际上,只需要a和b就可以 */ *plength= ; for (pfield= ptabfield; (field= *pfield) ; pfield++) { if (!bitmap_is_set(read_set, field->field_index)) continue; if (field->flags & BLOB_FLAG) // 有BLOB数据时,必定使用回表排序 ; length+= field->max_packed_col_length(field->pack_length()); if (field->maybe_null()) null_fields++; fields++; } if (!fields) ; length+= (null_fields+)/; //数据行长度超过了max_length_for_sort_data,使用回表排序 if (length+sortlength > max_length_for_sort_data || !(addonf= (SORT_ADDON_FIELD *) my_malloc(sizeof(SORT_ADDON_FIELD)* (fields+), MYF(MY_WME)))) ; // 构造addon_fields ... }
各种数据类型作为排序字段时的二进制表示
为了方便使用memcmp排序,各字段拷贝到排序字段时,都会进行特殊处理。以Decimal为例,前导空格和'+'号以及前导'0'都转换为空格字符,负号'-'转换为0x1,因为它比所有的‘数字’都要小。负数的数字部分,'0'转换为0x9, '1'转换为0x8,...依此类推。
例如 Decimal(5.2) A='12.34' B='-34.56' C='-44.56' 的二进制表示如下: A: 0x20 0x31 0x32 0x33 0x34 B: 0x01 0x06 0x05 0x04 0x03 C: 0x01 0x04 0x05 0x04 0x03 很显然A,B,C的二进制比较和A,B,C的逻辑比较结果是一致的。 void Field_decimal::make_sort_key(uchar *to, uint length) { uchar *str,*end; for (str=ptr,end=ptr+length; str != end && ((my_isspace(&my_charset_bin,*str) || *str == '+' || *str == ')) ; str++) *to++=' '; if (str == end) return; /* purecov: inspected */ if (*str == '-') { *to++=; // Smaller than any number str++; while (str != end) if (my_isdigit(&my_charset_bin,*str)) *to++= (' - *str++); else *to++= *str++; } else memcpy(to,str,(uint) (end-str)); }
参考资料
- MySQL排序内部原理探秘
- MySQL Server 5.6.34 Source Code
【MySQL】filesort.cc 源代码阅读笔记的更多相关文章
- Mongodb源代码阅读笔记:Journal机制
Mongodb源代码阅读笔记:Journal机制 Mongodb源代码阅读笔记:Journal机制 涉及的文件 一些说明 PREPLOGBUFFER WRITETOJOURNAL WRITETODAT ...
- CI框架源代码阅读笔记5 基准測试 BenchMark.php
上一篇博客(CI框架源代码阅读笔记4 引导文件CodeIgniter.php)中.我们已经看到:CI中核心流程的核心功能都是由不同的组件来完毕的.这些组件类似于一个一个单独的模块,不同的模块完毕不同的 ...
- CI框架源代码阅读笔记3 全局函数Common.php
从本篇開始.将深入CI框架的内部.一步步去探索这个框架的实现.结构和设计. Common.php文件定义了一系列的全局函数(一般来说.全局函数具有最高的载入优先权.因此大多数的框架中BootStrap ...
- CI框架源代码阅读笔记2 一切的入口 index.php
上一节(CI框架源代码阅读笔记1 - 环境准备.基本术语和框架流程)中,我们提到了CI框架的基本流程.这里再次贴出流程图.以备參考: 作为CI框架的入口文件.源代码阅读,自然由此開始. 在源代码阅读的 ...
- Spark源代码阅读笔记之DiskStore
Spark源代码阅读笔记之DiskStore BlockManager底层通过BlockStore来对数据进行实际的存储.BlockStore是一个抽象类,有三种实现:DiskStore(磁盘级别的持 ...
- Java Jdk1.8 HashMap源代码阅读笔记二
三.源代码阅读 3.元素包括containsKey(Object key) /** * Returns <tt>true</tt> if this map contains a ...
- ruby2.2.2 源代码阅读笔记
这是win32下的结构 从ruby_setup开始阅读 Ruby对象内存结构 RVALUE是一个union,内含ruby所有结构体(RBasic RObject RClass RFloat RStri ...
- Discuz!源代码阅读笔记之common.inc.php文件【1】
<?php /* [Discuz!] (C)2001-2007 Comsenz Inc. This is NOT a freeware, use is subject to license te ...
- Flask 源代码阅读笔记
我认为我已经养成了一个坏习惯.在使用一个框架过程中对它的内部原理非常感兴趣,有时候须要花不少精力才 明确,这也导致了学习的缓慢,但换来的是对框架的内部机理的熟悉,正如侯捷所说,源代码面前,了无秘密.这 ...
随机推荐
- Ubuntu下JDK+Tomcat+MySql环境的搭建
主机在阿里云上,所以网络的配置都省了,只剩下软件的安装和配置 1.安装mysql 1.1 apt-get install mysql-server-5.5 安装过程中,有两次提示输入 mysql 的 ...
- spring mvc后台接收中文乱码
可从如下几方面着手 1.jsp页面编码 2.web.xml配置字符过滤器,该字符过滤器最好放在开头 3.tomcat下server.xml添加URIEncoding="UTF-8" ...
- linux面试题集锦2《转》
一.填空题: 1. 在Linux系统中,以 文件 方式访问设备 . 2. Linux内核引导时,从文件 /etc/fstab 中读取要加载的文件系统. 3. Linux文件系统中每个文件用 索引节点来 ...
- CodeForces 632D Longest Subsequence
暴力. 虽然$a[i]$最大有${10^9}$,但是$m$最大只有${10^6}$,因此可以考虑暴力. 记$cnt[i]$表示数字$i$有$cnt[i]$个,记$p[i]$表示以$i$为倍数的情况下, ...
- How to use dt.Rows.Cast<System.Data.DataRow>().Take(n)
参考文章:http://stackoverflow.com/questions/2787458/how-to-select-top-n-rows-from-a-datatable-dataview-i ...
- 数据挖掘_wget整站下载
你应该了解的所有wget命令 翻译自All the Wget Commands You Should Know 如何下载整个网站用来离线浏览?怎样将一个网站上的所有MP3文件保存到本地的一个目录中?怎 ...
- 创建maven多模块项目(idea工具)
1.创建父项目模块(new 一个maven空项目模块)不勾选 create from archetype 删除src目录 2.创建子模块 webapp (该模块为web入口模块) 3.创建其他子模块 ...
- angularjs 中的iframe 标签 ng-src 路径
如果直接写路径到iframe标签里的ng-src中会出现报错: 解决方法: 1.ng里面有个属性是专门用来解决跨域问题的 $sce. 用法: $scope.someUrl = $sce.trustAs ...
- CoreJavaE10V1P3.3 第3章 Java的基本编程结构-3.3 数据类型
3.3 数据类型 这里所说的数据类型是指 Java的8中基本数据类型,是原生就存在的. 不同进制数的字面值表示方法 进制 字面值表示方法 例子 是否默认 JDK版本支持 2进制 0b或0B前缀(每4位 ...
- Linux 分区挂载方案
/boot 1G swap 2G(看内存决定) / 10-15G /home 5G