前言

  本文从零到一，手把手实现一个内存池。

  比较出名的内存池有jemalloc和tcmalloc，这两个都是全局内存池，比较推荐使用tcmalloc。

本专栏知识点是通过零声教育的线上课学习，进行梳理总结写下文章，对c/c++linux课程感兴趣的读者，可以点击链接C/C++Linux服务器开发/后台架构师【零声教育】-学习视频教程-腾讯课堂介绍详细查看课程的服务。

为什么要用内存池

  为什么要用内存池？首先，在7 * 24h的服务器中如果不使用内存池，而使用malloc和free，那么就非常容易产生内存碎片，早晚都会申请内存失败；并且在比较复杂的代码或者继承的屎山中，非常容易出现内存泄漏导致mmo的问题。

为了解决这两个问题，内存池就应运而生了。内存池预先分配一大块内存来做一个内存池，业务中的内存分配和释放都由这个内存池来管理，内存池内的内存不足时其内部会自己申请。所以内存碎片的问题就交由内存池的算法来优化，而内存泄漏的问题只需要遵守内存池提供的api，就非常容易避免内存泄漏了。

即使出现了内存泄漏，排查的思路也很清晰。1.检查是不是内存池的问题；2.如果不是内存池的问题，就检查是不是第三方库的内存泄漏。

内存池的使用场景

1. 全局内存池
2. 一个连接一个内存池(本文实现这个场景的内存池)

设计一个内存池

总体介绍

  由于本文是一个连接一个内存池，所以后续介绍和代码都是以4k为分界线，大于4k的我们认为是大块内存；小于4k的我们认为是小块内存。并且注意这里的4k，并不是严格遵照4096，而是在描述上，用4k比较好描述。

在真正使用内存之前，内存池提前分配一定数量且大小相等的内存块以作备用，当真正被用户调用api分配内存的时候，直接从内存块中获取内存（指小块内存），当内存块不够用了，再有内存池取申请新的内存块。而如果是需要大块内存，则内存池直接申请大块内存再返回给用户。

内存池：就是将这些提前申请的内存块组织管理起来的数据结构，内存池实现原理主要分为分配，回收，扩容三部分。

内存池原理之小块内存：分配=> 内存池预申请一块4k的内存块，这里称为block，即block=4k内存块。当用户向内存池申请内存size小于4k时，内存池从block的空间中划分出去size空间，当再有新申请时，再划分出去。扩容=> 直到block中的剩余空间不足以分配size大小，那么此时内存池会再次申请一块block，再从新的block中划分size空间给用户。回收=> 每一次申请小内存，都会在对应的block中引用计数加1，每一次释放小内存时，都会在block中引用计数减1，只有当引用计数为零的时候，才会回收block使他重新成为空闲空间，以便重复利用空间。这样，内存池避免频繁向内核申请/释放内存，从而提高系统性能。

内存池原理之大块内存：分配=> 因为大块内存是大于4k的，所以内存池不预先申请内存，也就是用户申请的时候，内存池再申请内存，然后返回给用户。扩容=> 大块内存不存在扩容。回收=> 对于大块内存来说，回收就直接free掉即可。

  上面理论讲完了，下面来介绍如何管理小块内存和大块内存。

【文章福利】：小编整理了一些个人觉得比较好的学习书籍、视频资料共享在qun文件里面，有需要的可以自行添加哦！~点击832218493加入（需要自取）

小块内存的分配与管理

  在创建内存池的时候，会预先申请一块4k的内存，并且在起始处将pool的结构体和node的结构体放进去，

  在创建内存池的时候，会预先申请一块4k的内存，并且在起始处将pool的结构体和node的结构体放进去，从last开始一直到end都是空闲内存，<last , end >中间的区域就用来存储小块内存。每一次mp_malloc，就将last指针后移，直到 e n d − l a s t < s i z e end - last < sizeend−last<size 时，进行扩容，将新block的last后移即可。

初始状态

分配内存

扩容

​

大块内存的分配与管理

  对于大块内存，前面已经说了，用户申请的时候，内存池才申请

• 申请一块大内存

​

再申请一块大内存

内存池代码实现

向外提供的api

• mp_create_pool：创建一个线程池，其核心是创建struct mp_pool_s这个结构体，并申请4k内存，将各个指针指向上文初始状态的图一样。
• mp_destroy_pool：销毁内存池，遍历小块结构体和大块结构体，进行free释放内存
• mp_malloc：提供给用户申请内存的api
• mp_calloc：通过mp_malloc申请内存后置零，相当于calloc
• mp_free：释放由mp_malloc返回的内存
• mp_reset_pool：将block的last置为初始状态，销毁所有大块内存
• monitor_mp_poll：监控内存池状态

struct mp_pool_s *mp_create_pool(size_t size);void mp_destroy_pool(struct mp_pool_s *pool);void *mp_malloc(struct mp_pool_s *pool, size_t size);void *mp_calloc(struct mp_pool_s *pool, size_t size);void mp_free(struct mp_pool_s *pool, void *p);void mp_reset_pool(struct mp_pool_s *pool);void monitor_mp_poll(struct mp_pool_s *pool, char *tk);

相关结构体的定义

  mp_pool_s 就是整个内存池的管理结构，我们做的内存池是一个连接一个内存池，所以对于整个程序而言，内存池对象是有很多个的。

可能读者会有疑问，有了head，为什么还有current，是因为如果一个block剩余空间小于size超过一定次数后，将current指向下一个block，这样就加快内存分配效率，减少遍历次数。

//每4k一block结点struct mp_node_s { unsigned char *end;//块的结尾 unsigned char *last;//使用到哪了 struct mp_node_s *next;//链表 int quote;//引用计数 int failed;//失效次数};struct mp_large_s { struct mp_large_s *next;//链表 int size;//alloc的大小 void *alloc;//大块内存的起始地址};struct mp_pool_s { struct mp_large_s *large; struct mp_node_s *head; struct mp_node_s *current;};

内存对齐

访问速度是内存对齐的原因之一，另外一个原因是某些平台(arm)不支持未内存对齐的访问

 在4k里面划分内存，那么必然有很多地方是不对齐的，所以这里提供两个内存对齐的函数。那么为什么要内存对齐呢？其一：提高访问速度；其二：某些平台arm不支持未对其的内存访问，会出错。

#define mp_align(n, alignment) (((n)+(alignment-1)) & ~(alignment-1))#define mp_align_ptr(p, alignment) (void *)((((size_t)p)+(alignment-1)) & ~(alignment-1))

创建与销毁内存池

  创建一个线程池，其核心是创建struct mp_pool_s这个结构体，并申请4k内存，将各个指针指向上文初始状态的图一样。   销毁内存池，遍历小块结构体和大块结构体，进行free释放内存。

//创建内存池struct mp_pool_s *mp_create_pool(size_t size) { struct mp_pool_s *pool; if (size < PAGE_SIZE || size % PAGE_SIZE != 0) { size = PAGE_SIZE; } //分配4k以上不用malloc，用posix_memalign /* int posix_memalign (void **memptr, size_t alignment, size_t size); */ int ret = posix_memalign((void **) &pool, MP_ALIGNMENT, size); //4K + mp_pool_s if (ret) { return NULL; } pool->large = NULL; pool->current = pool->head = (unsigned char *) pool + sizeof(struct mp_pool_s); pool->head->last = (unsigned char *) pool + sizeof(struct mp_pool_s) + sizeof(struct mp_node_s); pool->head->end = (unsigned char *) pool + PAGE_SIZE; pool->head->failed = 0; return pool;}//销毁内存池void mp_destroy_pool(struct mp_pool_s *pool) { struct mp_large_s *large; for (large = pool->large; large; large = large->next) { if (large->alloc) { free(large->alloc); } } struct mp_node_s *cur, *next; cur = pool->head->next; while (cur) { next = cur->next; free(cur); cur = next; } free(pool);}

​

提供给用户的内存申请api

  申请的内存以size做区分，如果大于4k就分配大块内存，小于4k就去block里面划分。

//分配内存void *mp_malloc(struct mp_pool_s *pool, size_t size) { if (size <= 0) { return NULL; } if (size > PAGE_SIZE - sizeof(struct mp_node_s)) { //large return mp_malloc_large(pool, size); } else { //small unsigned char *mem_addr = NULL; struct mp_node_s *cur = NULL; cur = pool->current; while (cur) { mem_addr = mp_align_ptr(cur->last, MP_ALIGNMENT); if (cur->end - mem_addr >= size) { cur->quote++;//引用+1 cur->last = mem_addr + size; return mem_addr; } else { cur = cur->next; } } return mp_malloc_block(pool, size);// open new space }}void *mp_calloc(struct mp_pool_s *pool, size_t size) { void *mem_addr = mp_malloc(pool, size); if (mem_addr) { memset(mem_addr, 0, size); } return mem_addr;}

小块内存block扩容

   所有的block都 e n d − l a s t < s i z e end - last < sizeend−last<size 时，进行扩容，将新block的last后移即可。

//new block 4kvoid *mp_malloc_block(struct mp_pool_s *pool, size_t size) { unsigned char *block; int ret = posix_memalign((void **) &block, MP_ALIGNMENT, PAGE_SIZE); //4K if (ret) { return NULL; } struct mp_node_s *new_node = (struct mp_node_s *) block; new_node->end = block + PAGE_SIZE; new_node->next = NULL; unsigned char *ret_addr = mp_align_ptr(block + sizeof(struct mp_node_s), MP_ALIGNMENT); new_node->last = ret_addr + size; new_node->quote++; struct mp_node_s *current = pool->current; struct mp_node_s *cur = NULL; for (cur = current; cur->next; cur = cur->next) { if (cur->failed++ > 4) { current = cur->next; } } //now cur = last node cur->next = new_node; pool->current = current; return ret_addr;}

分配大块内存

//size>4kvoid *mp_malloc_large(struct mp_pool_s *pool, size_t size) { unsigned char *big_addr; int ret = posix_memalign((void **) &big_addr, MP_ALIGNMENT, size); //size if (ret) { return NULL; } struct mp_large_s *large; //released struct large resume int n = 0; for (large = pool->large; large; large = large->next) { if (large->alloc == NULL) { large->size = size; large->alloc = big_addr; return big_addr; } if (n++ > 3) { break;// 为了避免过多的遍历，限制次数 } } large = mp_malloc(pool, sizeof(struct mp_large_s)); if (large == NULL) { free(big_addr); return NULL; } large->size = size; large->alloc = big_addr; large->next = pool->large; pool->large = large; return big_addr;}

释放内存

  如果是大块内存，找到之后直接释放；如果是小块内存，将引用计数减1，如果引用计数为0则重置last。

//释放内存void mp_free(struct mp_pool_s *pool, void *p) { struct mp_large_s *large; for (large = pool->large; large; large = large->next) {//大块 if (p == large->alloc) { free(large->alloc); large->size = 0; large->alloc = NULL; return; } } //小块 引用-1 struct mp_node_s *cur = NULL; for (cur = pool->head; cur; cur = cur->next) {// printf("cur:%p p:%p end:%pn", (unsigned char *) cur, (unsigned char *) p, (unsigned char *) cur->end); if ((unsigned char *) cur <= (unsigned char *) p && (unsigned char *) p <= (unsigned char *) cur->end) { cur->quote--; if (cur->quote == 0) { if (cur == pool->head) { pool->head->last = (unsigned char *) pool + sizeof(struct mp_pool_s) + sizeof(struct mp_node_s); } else { cur->last = (unsigned char *) cur + sizeof(struct mp_node_s); } cur->failed = 0; pool->current = pool->head; } return; } }}

内存池测试

//// Created by 68725 on 2022/7/26.//#include <stdlib.h>#include <stdio.h>#include <string.h>#define PAGE_SIZE 4096#define MP_ALIGNMENT 16#define mp_align(n, alignment) (((n)+(alignment-1)) & ~(alignment-1))#define mp_align_ptr(p, alignment) (void *)((((size_t)p)+(alignment-1)) & ~(alignment-1))//每4k一block结点struct mp_node_s { unsigned char *end;//块的结尾 unsigned char *last;//使用到哪了 struct mp_node_s *next;//链表 int quote;//引用计数 int failed;//失效次数};struct mp_large_s { struct mp_large_s *next;//链表 int size;//alloc的大小 void *alloc;//大块内存的起始地址};struct mp_pool_s { struct mp_large_s *large; struct mp_node_s *head; struct mp_node_s *current;};struct mp_pool_s *mp_create_pool(size_t size);void mp_destroy_pool(struct mp_pool_s *pool);void *mp_malloc(struct mp_pool_s *pool, size_t size);void *mp_calloc(struct mp_pool_s *pool, size_t size);void mp_free(struct mp_pool_s *pool, void *p);void mp_reset_pool(struct mp_pool_s *pool);void monitor_mp_poll(struct mp_pool_s *pool, char *tk);void mp_reset_pool(struct mp_pool_s *pool) { struct mp_node_s *cur; struct mp_large_s *large; for (large = pool->large; large; large = large->next) { if (large->alloc) { free(large->alloc); } } pool->large = NULL; pool->current = pool->head; for (cur = pool->head; cur; cur = cur->next) { cur->last = (unsigned char *) cur + sizeof(struct mp_node_s); cur->failed = 0; cur->quote = 0; }}//创建内存池struct mp_pool_s *mp_create_pool(size_t size) { struct mp_pool_s *pool; if (size < PAGE_SIZE || size % PAGE_SIZE != 0) { size = PAGE_SIZE; } //分配4k以上不用malloc，用posix_memalign /* int posix_memalign (void **memptr, size_t alignment, size_t size); */ int ret = posix_memalign((void **) &pool, MP_ALIGNMENT, size); //4K + mp_pool_s if (ret) { return NULL; } pool->large = NULL; pool->current = pool->head = (unsigned char *) pool + sizeof(struct mp_pool_s); pool->head->last = (unsigned char *) pool + sizeof(struct mp_pool_s) + sizeof(struct mp_node_s); pool->head->end = (unsigned char *) pool + PAGE_SIZE; pool->head->failed = 0; return pool;}//销毁内存池void mp_destroy_pool(struct mp_pool_s *pool) { struct mp_large_s *large; for (large = pool->large; large; large = large->next) { if (large->alloc) { free(large->alloc); } } struct mp_node_s *cur, *next; cur = pool->head->next; while (cur) { next = cur->next; free(cur); cur = next; } free(pool);}//size>4kvoid *mp_malloc_large(struct mp_pool_s *pool, size_t size) { unsigned char *big_addr; int ret = posix_memalign((void **) &big_addr, MP_ALIGNMENT, size); //size if (ret) { return NULL; } struct mp_large_s *large; //released struct large resume int n = 0; for (large = pool->large; large; large = large->next) { if (large->alloc == NULL) { large->size = size; large->alloc = big_addr; return big_addr; } if (n++ > 3) { break;// 为了避免过多的遍历，限制次数 } } large = mp_malloc(pool, sizeof(struct mp_large_s)); if (large == NULL) { free(big_addr); return NULL; } large->size = size; large->alloc = big_addr; large->next = pool->large; pool->large = large; return big_addr;}//new block 4kvoid *mp_malloc_block(struct mp_pool_s *pool, size_t size) { unsigned char *block; int ret = posix_memalign((void **) &block, MP_ALIGNMENT, PAGE_SIZE); //4K if (ret) { return NULL; } struct mp_node_s *new_node = (struct mp_node_s *) block; new_node->end = block + PAGE_SIZE; new_node->next = NULL; unsigned char *ret_addr = mp_align_ptr(block + sizeof(struct mp_node_s), MP_ALIGNMENT); new_node->last = ret_addr + size; new_node->quote++; struct mp_node_s *current = pool->current; struct mp_node_s *cur = NULL; for (cur = current; cur->next; cur = cur->next) { if (cur->failed++ > 4) { current = cur->next; } } //now cur = last node cur->next = new_node; pool->current = current; return ret_addr;}//分配内存void *mp_malloc(struct mp_pool_s *pool, size_t size) { if (size <= 0) { return NULL; } if (size > PAGE_SIZE - sizeof(struct mp_node_s)) { //large return mp_malloc_large(pool, size); } else { //small unsigned char *mem_addr = NULL; struct mp_node_s *cur = NULL; cur = pool->current; while (cur) { mem_addr = mp_align_ptr(cur->last, MP_ALIGNMENT); if (cur->end - mem_addr >= size) { cur->quote++;//引用+1 cur->last = mem_addr + size; return mem_addr; } else { cur = cur->next; } } return mp_malloc_block(pool, size);// open new space }}void *mp_calloc(struct mp_pool_s *pool, size_t size) { void *mem_addr = mp_malloc(pool, size); if (mem_addr) { memset(mem_addr, 0, size); } return mem_addr;}//释放内存void mp_free(struct mp_pool_s *pool, void *p) { struct mp_large_s *large; for (large = pool->large; large; large = large->next) {//大块 if (p == large->alloc) { free(large->alloc); large->size = 0; large->alloc = NULL; return; } } //小块 引用-1 struct mp_node_s *cur = NULL; for (cur = pool->head; cur; cur = cur->next) {// printf("cur:%p p:%p end:%pn", (unsigned char *) cur, (unsigned char *) p, (unsigned char *) cur->end); if ((unsigned char *) cur <= (unsigned char *) p && (unsigned char *) p <= (unsigned char *) cur->end) { cur->quote--; if (cur->quote == 0) { if (cur == pool->head) { pool->head->last = (unsigned char *) pool + sizeof(struct mp_pool_s) + sizeof(struct mp_node_s); } else { cur->last = (unsigned char *) cur + sizeof(struct mp_node_s); } cur->failed = 0; pool->current = pool->head; } return; } }}void monitor_mp_poll(struct mp_pool_s *pool, char *tk) { printf("rnrn------start monitor poll------%srnrn", tk); struct mp_node_s *head = NULL; int i = 0; for (head = pool->head; head; head = head->next) { i++; if (pool->current == head) { printf("current==>第%d块n", i); } if (i == 1) { printf("第%02d块small block 已使用:%4ld 剩余空间:%4ld 引用:%4d failed:%4dn", i, (unsigned char *) head->last - (unsigned char *) pool, head->end - head->last, head->quote, head->failed); } else { printf("第%02d块small block 已使用:%4ld 剩余空间:%4ld 引用:%4d failed:%4dn", i, (unsigned char *) head->last - (unsigned char *) head, head->end - head->last, head->quote, head->failed); } } struct mp_large_s *large; i = 0; for (large = pool->large; large; large = large->next) { i++; if (large->alloc != NULL) { printf("第%d块large block size=%dn", i, large->size); } } printf("rnrn------stop monitor poll------rnrn");}int main() { struct mp_pool_s *p = mp_create_pool(PAGE_SIZE); monitor_mp_poll(p, "create memory pool");#if 0 printf("mp_align(5, %d): %d, mp_align(17, %d): %dn", MP_ALIGNMENT, mp_align(5, MP_ALIGNMENT), MP_ALIGNMENT, mp_align(17, MP_ALIGNMENT)); printf("mp_align_ptr(p->current, %d): %p, p->current: %pn", MP_ALIGNMENT, mp_align_ptr(p->current, MP_ALIGNMENT), p->current);#endif void *mp[30]; int i; for (i = 0; i < 30; i++) { mp[i] = mp_malloc(p, 512); } monitor_mp_poll(p, "申请512字节30个"); for (i = 0; i < 30; i++) { mp_free(p, mp[i]); } monitor_mp_poll(p, "销毁512字节30个"); int j; for (i = 0; i < 50; i++) { char *pp = mp_calloc(p, 32); for (j = 0; j < 32; j++) { if (pp[j]) { printf("calloc wrongn"); exit(-1); } } } monitor_mp_poll(p, "申请32字节50个"); for (i = 0; i < 50; i++) { char *pp = mp_malloc(p, 3); } monitor_mp_poll(p, "申请3字节50个"); void *pp[10]; for (i = 0; i < 10; i++) { pp[i] = mp_malloc(p, 5120); } monitor_mp_poll(p, "申请大内存5120字节10个"); for (i = 0; i < 10; i++) { mp_free(p, pp[i]); } monitor_mp_poll(p, "销毁大内存5120字节10个"); mp_reset_pool(p); monitor_mp_poll(p, "reset pool"); for (i = 0; i < 100; i++) { void *s = mp_malloc(p, 256); } monitor_mp_poll(p, "申请256字节100个"); mp_destroy_pool(p); return 0;}

​

nginx内存池对比分析

相关结构体定义对比

​

创建内存池对比

内存申请对比

​