nginx之ngx_slab

nginx通过自己实现的slab机制来减少内存的碎片化。
而nginx的slab机制相对于linux内核的slab机制就显得相对的简单。通过nginx可以更快的理解slab机制。

ngx_slab.h:

typedef struct ngx_slab_page_s  ngx_slab_page_t;

struct ngx_slab_page_s {
    uintptr_t         slab;
    ngx_slab_page_t  *next;
    uintptr_t         prev;
};


typedef struct {
    ngx_shmtx_sh_t    lock;		//互斥锁

    size_t            min_size;
    size_t            min_shift;

    ngx_slab_page_t  *pages;
    ngx_slab_page_t  *last;
    ngx_slab_page_t   free;

    u_char           *start;
    u_char           *end;

    ngx_shmtx_t       mutex;

    u_char           *log_ctx;
    u_char            zero;

    unsigned          log_nomem:1;

    void             *data;
    void             *addr;
} ngx_slab_pool_t;

头文件很简单，定义了两个结构体，一个

ngx_slab.c:

void
ngx_slab_init(ngx_slab_pool_t *pool)
{
...
    pool->min_size = 1 << pool->min_shift;

    p = (u_char *) pool + sizeof(ngx_slab_pool_t);
    size = pool->end - p;

    ngx_slab_junk(p, size);

    slots = (ngx_slab_page_t *) p;
   /**
   	*这里的n通过移位来计算有多少个ngx_slab_page_t的结构体
    *由于均为2的整数，相当于做除法
	**/
	n = ngx_pagesize_shift - pool->min_shift;
	//初始化n个ngx_slab_page_t结构体
    for (i = 0; i < n; i++) {
        slots[i].slab = 0;
        slots[i].next = &slots[i];
        slots[i].prev = 0;
    }

    p += n * sizeof(ngx_slab_page_t);

    pages = (ngx_uint_t) (size / (ngx_pagesize + sizeof(ngx_slab_page_t)));

    ngx_memzero(p, pages * sizeof(ngx_slab_page_t));
	//将slab_page结构体挂载到ngx_slab_pool_t结构体上
    pool->pages = (ngx_slab_page_t *) p;

    pool->free.prev = 0;
    pool->free.next = (ngx_slab_page_t *) p;

    pool->pages->slab = pages;
    pool->pages->next = &pool->free;
    pool->pages->prev = (uintptr_t) &pool->free;
	//真正可以用作数据储存的开始地址，进行内存对齐，提高程序运行速度
    pool->start = (u_char *)
                  ngx_align_ptr((uintptr_t) p + pages * sizeof(ngx_slab_page_t),
                                 ngx_pagesize);
	//由于内存对齐可以会使pages减少，根据相应的减少量调整pages
    m = pages - (pool->end - pool->start) / ngx_pagesize;
    if (m > 0) {
        pages -= m;
        pool->pages->slab = pages;
    }
    
    pool->last = pool->pages + pages;
	//默认为1
    pool->log_nomem = 1;
    pool->log_ctx = &pool->zero;
    pool->zero = '\0';
}

可以看到nginx在初始化共享内存的内存池是比较简单的，而且为了提高程序的运行速度，nginx用了很多小技巧。

void *
ngx_slab_alloc(ngx_slab_pool_t *pool, size_t size)
{
    void  *p;
	//申请共享内存需要上互斥锁
    ngx_shmtx_lock(&pool->mutex);

    p = ngx_slab_alloc_locked(pool, size);
	//申请后立即释放
    ngx_shmtx_unlock(&pool->mutex);

    return p;
}

nginx对于申请的内存大小有着划分，在之前的ngx_slab_init可以看到计算slot 的n = ngx_pagesize_shift - pool->min_shift;(一般ngx_pagesize_shift为12，即一页大小为4096，min_shift为3，即最小为8，这里的单位为Byte)，nginx划分的slot有8，16，32，64，128，256，512，1024，2048这9个等级。(若超过2048Byte，nginx直接返回一个页的大小，即4096Byte)

void *
ngx_slab_alloc_locked(ngx_slab_pool_t *pool, size_t size)
{
...
    if (size > ngx_slab_max_size) {		
    	//这里的ngx_slab_max_size为2048
...
        page = ngx_slab_alloc_pages(pool, (size >> ngx_pagesize_shift)
                                          + ((size % ngx_pagesize) ? 1 : 0));
        if (page) {
            p = (page - pool->pages) << ngx_pagesize_shift;
            p += (uintptr_t) pool->start;

        } else {
            p = 0;
        }

        goto done;
    }
	//若是大于8，则计算应使用哪个slot下面的内存
    if (size > pool->min_size) {
        shift = 1;
       /**如申请33Byte
        *通过计算可以得到应使用块大小为64Byte的slot
        *即slot=6-3=3
        **/
        for (s = size - 1; s >>= 1; shift++) { /* void */ }
        slot = shift - pool->min_shift;

    } else {
    	//最小为8，即第一个slot
        size = pool->min_size;
        shift = pool->min_shift;
        slot = 0;
    }
...

    slots = (ngx_slab_page_t *) ((u_char *)pool + \
    			sizeof(ngx_slab_pool_t));
    page = slots[slot].next;
	//判断是否有剩余的页
    if (page->next != page) {
		
        if (shift < ngx_slab_exact_shift) {
		//当小于ngx_slab_exact_shift(64位系统为6，32位系统为5)
            do {
            	//根据偏移量计算page的地址
            	/**这里的(page-pool-pages)使用一个小技巧
            	 *即指针计算的时候会根据指针类型得到对应的数值
            	 *如page[1]的地址相对于page[0]偏移24个字节
            	 *&page[1]-&page[0]=1
            	 *而(char*)&page[1]-(char*)&page[0]=24
           		**/
                p = (page - pool->pages) << ngx_pagesize_shift;
                bitmap = (uintptr_t *) (pool->start + p);
				//计算bitmap的大小(占用多少个块)
                map = (1 << (ngx_pagesize_shift - shift))
                          / (sizeof(uintptr_t) * 8);

                for (n = 0; n < map; n++) {

                    if (bitmap[n] != NGX_SLAB_BUSY) {

                        for (m = 1, i = 0; m; m <<= 1, i++) {
                            if ((bitmap[n] & m)) {
                                continue;
                            }

                            bitmap[n] |= m;
							//计算未被占用的的块的位置
                            i = ((n * sizeof(uintptr_t) * 8) << \ 
                            	shift) + (i << shift);
							/*若此page里面的可用空间全满了
							*则将此page脱离该slot
							*即不能从page得到可用的内存
							*/
                            if (bitmap[n] == NGX_SLAB_BUSY) {
                                for (n = n + 1; n < map; n++) {
                                     if (bitmap[n] != NGX_SLAB_BUSY) {
                                         p = (uintptr_t) bitmap + i;

                                         goto done;
                                     }
                                }

                                prev = (ngx_slab_page_t *)
                                            (page->prev & ~NGX_SLAB_PAGE_MASK);
                                prev->next = page->next;
                                page->next->prev = page->prev;

                                page->next = NULL;
                                page->prev = NGX_SLAB_SMALL;
                            }

                            p = (uintptr_t) bitmap + i;

                            goto done;
                        }
                    }
                }

                page = page->next;

            } while (page);

        } else if (shift == ngx_slab_exact_shift) {
			//当要分配的size位于ngx_slab_exact_shift时
			//这些page不需要使用多个比bitmap来记录
			//仅需一个sizeof(uintptr_t)来记录即可
            do {
                if (page->slab != NGX_SLAB_BUSY) {

                    for (m = 1, i = 0; m; m <<= 1, i++) {
                        if ((page->slab & m)) {
                            continue;
                        }

                        page->slab |= m;

                        if (page->slab == NGX_SLAB_BUSY) {
                            prev = (ngx_slab_page_t *)
                                            (page->prev & ~NGX_SLAB_PAGE_MASK);
                            prev->next = page->next;
                            page->next->prev = page->prev;

                            page->next = NULL;
                            page->prev = NGX_SLAB_EXACT;
                        }

                        p = (page - pool->pages) << ngx_pagesize_shift;
                        p += i << shift;
                        p += (uintptr_t) pool->start;

                        goto done;
                    }
                }

                page = page->next;

            } while (page);

        } else { /* shift > ngx_slab_exact_shift */
			//由于是2的倍数，大于ngx_slab_exact_shift的
			//只需要用sizeof(uintptr_t)一半来储存就好
			//nginx使用低位记录块的大小，高位记录位图
            n = ngx_pagesize_shift - (page->slab & NGX_SLAB_SHIFT_MASK);
            n = 1 << n;
            n = ((uintptr_t) 1 << n) - 1;
            mask = n << NGX_SLAB_MAP_SHIFT;

            do {
                if ((page->slab & NGX_SLAB_MAP_MASK) != mask) {

                    for (m = (uintptr_t) 1 << NGX_SLAB_MAP_SHIFT, i = 0;
                         m & mask;
                         m <<= 1, i++)
                    {
                        if ((page->slab & m)) {
                            continue;
                        }

                        page->slab |= m;

                        if ((page->slab & NGX_SLAB_MAP_MASK) == mask) {
                            prev = (ngx_slab_page_t *)
                                            (page->prev & ~NGX_SLAB_PAGE_MASK);
                            prev->next = page->next;
                            page->next->prev = page->prev;

                            page->next = NULL;
                            page->prev = NGX_SLAB_BIG;
                        }

                        p = (page - pool->pages) << ngx_pagesize_shift;
                        p += i << shift;
                        p += (uintptr_t) pool->start;

                        goto done;
                    }
                }

                page = page->next;

            } while (page);
        }
    }	
	//当page不足的时候，重新分配一页
    page = ngx_slab_alloc_pages(pool, 1);

    if (page) {
        if (shift < ngx_slab_exact_shift) {
            p = (page - pool->pages) << ngx_pagesize_shift;
            bitmap = (uintptr_t *) (pool->start + p);
			//计算块的大小
            s = 1 << shift;
            //计算需要多少块存储bitmap
            n = (1 << (ngx_pagesize_shift - shift)) / 8 / s;

            if (n == 0) {
                n = 1;
            }
			//标志bitmap使用的块
            bitmap[0] = (2 << n) - 1;

            map = (1 << (ngx_pagesize_shift - shift)) / (sizeof(uintptr_t) * 8);
			//置位未使用的块
            for (i = 1; i < map; i++) {
                bitmap[i] = 0;
            }
			//挂载到对应的slot上
            page->slab = shift;
            page->next = &slots[slot];
            page->prev = (uintptr_t) &slots[slot] | NGX_SLAB_SMALL;

            slots[slot].next = page;
			//返回第一个未使用的块
            p = ((page - pool->pages) << ngx_pagesize_shift) + s * n;
            p += (uintptr_t) pool->start;

            goto done;

        } else if (shift == ngx_slab_exact_shift) {
			//使用slab作为位图
            page->slab = 1;
            page->next = &slots[slot];
            page->prev = (uintptr_t) &slots[slot] | NGX_SLAB_EXACT;

            slots[slot].next = page;

            p = (page - pool->pages) << ngx_pagesize_shift;
            p += (uintptr_t) pool->start;

            goto done;

        } else { /* shift > ngx_slab_exact_shift */
			//低位记录块大小
            page->slab = ((uintptr_t) 1 << NGX_SLAB_MAP_SHIFT) | shift;
            page->next = &slots[slot];
            page->prev = (uintptr_t) &slots[slot] | NGX_SLAB_BIG;
			//进行过内存对齐，因此至少低2位为0，即4的倍数
			//nginx利用低2位记录该page的属性
            slots[slot].next = page;

            p = (page - pool->pages) << ngx_pagesize_shift;
            p += (uintptr_t) pool->start;

            goto done;
        }
    }

    p = 0;

done:

    ngx_log_debug1(NGX_LOG_DEBUG_ALLOC, ngx_cycle->log, 0, "slab alloc: %p", p);

    return (void *) p;
}

举例：块大小为8Byte，需要用8个块来记录bitmap，分配bitmap如下：
（bitmap使用的块应该被标为已使用）

//calloc只是将申请到的page做清零的操作
void *
ngx_slab_calloc(ngx_slab_pool_t *pool, size_t size)
{
    void  *p;

    ngx_shmtx_lock(&pool->mutex);

    p = ngx_slab_calloc_locked(pool, size);

    ngx_shmtx_unlock(&pool->mutex);

    return p;
}


void *
ngx_slab_calloc_locked(ngx_slab_pool_t *pool, size_t size)
{
    void  *p;

    p = ngx_slab_alloc_locked(pool, size);
    if (p) {
        ngx_memzero(p, size);
    }

    return p;
}

内存释放逻辑跟申请内存逻辑相反，也需要上锁，由于内存对齐，可以根据指针p来计算相应的页的位置，slot的位置以及块的位置等等。

//释放内存上锁
void
ngx_slab_free(ngx_slab_pool_t *pool, void *p)
{
    ngx_shmtx_lock(&pool->mutex);

    ngx_slab_free_locked(pool, p);

    ngx_shmtx_unlock(&pool->mutex);
}


void
ngx_slab_free_locked(ngx_slab_pool_t *pool, void *p)
{
...
    if ((u_char *) p < pool->start || (u_char *) p > pool->end) {
        ngx_slab_error(pool, NGX_LOG_ALERT, "ngx_slab_free(): outside of pool");
        goto fail;
    }
	//计算需要释放的页的位置
    n = ((u_char *) p - pool->start) >> ngx_pagesize_shift;
    page = &pool->pages[n];
    //slab包含着块大小的信息
    slab = page->slab;
    //得到所释放的page的属性
    type = page->prev & NGX_SLAB_PAGE_MASK;

    switch (type) {

    case NGX_SLAB_SMALL:
		//得到块大小偏移
        shift = slab & NGX_SLAB_SHIFT_MASK;
        size = 1 << shift;
		//由于内存对齐，p一定为size的倍数
        if ((uintptr_t) p & (size - 1)) {
            goto wrong_chunk;
        }
		//n为要释放的块的位置，m为位图中要释放块的位置
        n = ((uintptr_t) p & (ngx_pagesize - 1)) >> shift;
        m = (uintptr_t) 1 << (n & (sizeof(uintptr_t) * 8 - 1));
        //计算使用的是哪个bitmap
        n /= (sizeof(uintptr_t) * 8);
        //bitmap的地址，由于内存对齐，只需要确定page的起始地址即可
        bitmap = (uintptr_t *)
                             ((uintptr_t) p & ~((uintptr_t) ngx_pagesize - 1));

        if (bitmap[n] & m) {
			//如果整个page已被使用，则将page挂载上对应的slot
            if (page->next == NULL) {
                slots = (ngx_slab_page_t *)
                                   ((u_char *) pool + sizeof(ngx_slab_pool_t));
                slot = shift - pool->min_shift;

                page->next = slots[slot].next;
                slots[slot].next = page;

                page->prev = (uintptr_t) &slots[slot] | NGX_SLAB_SMALL;
                page->next->prev = (uintptr_t) page | NGX_SLAB_SMALL;
            }
			//将要释放的块在bitmap对应的位置置0
            bitmap[n] &= ~m;
			//计算使用bitmap占了多少个块
            n = (1 << (ngx_pagesize_shift - shift)) / 8 / (1 << shift);

            if (n == 0) {
                n = 1;
            }
			//如果有还在使用 
            if (bitmap[0] & ~(((uintptr_t) 1 << n) - 1)) {
                goto done;
            }

            map = (1 << (ngx_pagesize_shift - shift)) / (sizeof(uintptr_t) * 8);

            for (n = 1; n < map; n++) {
                if (bitmap[n]) {
                    goto done;
                }
            }
			//不符合条件则直接释放page
            ngx_slab_free_pages(pool, page, 1);

            goto done;
        }

        goto chunk_already_free;

    case NGX_SLAB_EXACT:

        m = (uintptr_t) 1 <<
                (((uintptr_t) p & (ngx_pagesize - 1)) >> ngx_slab_exact_shift);
        size = ngx_slab_exact_size;

        if ((uintptr_t) p & (size - 1)) {
            goto wrong_chunk;
        }

        if (slab & m) {
            if (slab == NGX_SLAB_BUSY) {
                slots = (ngx_slab_page_t *)
                                   ((u_char *) pool + sizeof(ngx_slab_pool_t));
                slot = ngx_slab_exact_shift - pool->min_shift;

                page->next = slots[slot].next;
                slots[slot].next = page;

                page->prev = (uintptr_t) &slots[slot] | NGX_SLAB_EXACT;
                page->next->prev = (uintptr_t) page | NGX_SLAB_EXACT;
            }

            page->slab &= ~m;

            if (page->slab) {
                goto done;
            }

            ngx_slab_free_pages(pool, page, 1);

            goto done;
        }

        goto chunk_already_free;

    case NGX_SLAB_BIG:

        shift = slab & NGX_SLAB_SHIFT_MASK;
        size = 1 << shift;

        if ((uintptr_t) p & (size - 1)) {
            goto wrong_chunk;
        }

        m = (uintptr_t) 1 << ((((uintptr_t) p & (ngx_pagesize - 1)) >> shift)
                              + NGX_SLAB_MAP_SHIFT);

        if (slab & m) {

            if (page->next == NULL) {
                slots = (ngx_slab_page_t *)
                                   ((u_char *) pool + sizeof(ngx_slab_pool_t));
                slot = shift - pool->min_shift;

                page->next = slots[slot].next;
                slots[slot].next = page;

                page->prev = (uintptr_t) &slots[slot] | NGX_SLAB_BIG;
                page->next->prev = (uintptr_t) page | NGX_SLAB_BIG;
            }

            page->slab &= ~m;

            if (page->slab & NGX_SLAB_MAP_MASK) {
                goto done;
            }

            ngx_slab_free_pages(pool, page, 1);

            goto done;
        }

        goto chunk_already_free;

    case NGX_SLAB_PAGE:

        if ((uintptr_t) p & (ngx_pagesize - 1)) {
            goto wrong_chunk;
        }

        if (slab == NGX_SLAB_PAGE_FREE) {
            ngx_slab_error(pool, NGX_LOG_ALERT,
                           "ngx_slab_free(): page is already free");
            goto fail;
        }

        if (slab == NGX_SLAB_PAGE_BUSY) {
            ngx_slab_error(pool, NGX_LOG_ALERT,
                           "ngx_slab_free(): pointer to wrong page");
            goto fail;
        }

        n = ((u_char *) p - pool->start) >> ngx_pagesize_shift;
        size = slab & ~NGX_SLAB_PAGE_START;

        ngx_slab_free_pages(pool, &pool->pages[n], size);

        ngx_slab_junk(p, size << ngx_pagesize_shift);

        return;
    }

    /* not reached */

    return;

done:

    ngx_slab_junk(p, size);

    return;

wrong_chunk:

    ngx_slab_error(pool, NGX_LOG_ALERT,
                   "ngx_slab_free(): pointer to wrong chunk");

    goto fail;

chunk_already_free:

    ngx_slab_error(pool, NGX_LOG_ALERT,
                   "ngx_slab_free(): chunk is already free");

fail:

    return;
}

申请page逻辑也较为简单，首先从内存池里面查找有没有空闲的page，没有则从另一个内存池里面去找。

static ngx_slab_page_t *
ngx_slab_alloc_pages(ngx_slab_pool_t *pool, ngx_uint_t pages)
{
    ngx_slab_page_t  *page, *p;

    for (page = pool->free.next; page != &pool->free; page = page->next) {

        if (page->slab >= pages) {
			//从内存池里减去将分配的页数
            if (page->slab > pages) {
                page[page->slab - 1].prev = (uintptr_t) &page[pages];
				//重新将free节点指向未被分配的page
                page[pages].slab = page->slab - pages;
                page[pages].next = page->next;
                page[pages].prev = page->prev;

                p = (ngx_slab_page_t *) page->prev;
                p->next = &page[pages];
                page->next->prev = (uintptr_t) &page[pages];

            } else {
            	//刚好剩余则将free节点指向自己
                p = (ngx_slab_page_t *) page->prev;
                p->next = page->next;
                page->next->prev = page->prev;
            }
			//初始化出去的头一个page的属性
            page->slab = pages | NGX_SLAB_PAGE_START;
            page->next = NULL;
            page->prev = NGX_SLAB_PAGE;

            if (--pages == 0) {
                return page;
            }
			//初始化分配出去的page
            for (p = page + 1; pages; pages--) {
                p->slab = NGX_SLAB_PAGE_BUSY;
                p->next = NULL;
                p->prev = NGX_SLAB_PAGE;
                p++;
            }

            return page;
        }
    }

    if (pool->log_nomem) {
        ngx_slab_error(pool, NGX_LOG_CRIT,
                       "ngx_slab_alloc() failed: no memory");
    }

    return NULL;
}

申请一页的内存：

申请两页内存：

static void
ngx_slab_free_pages(ngx_slab_pool_t *pool, ngx_slab_page_t *page,
    ngx_uint_t pages)
{
    ngx_uint_t        type;
    ngx_slab_page_t  *prev, *join;

    page->slab = pages--;

    if (pages) {
        ngx_memzero(&page[1], pages * sizeof(ngx_slab_page_t));
    }
	//若有next节点则将page从链表里脱离
    if (page->next) {
        prev = (ngx_slab_page_t *) (page->prev & ~NGX_SLAB_PAGE_MASK);
        prev->next = page->next;
        page->next->prev = page->prev;
    }
	
    join = page + page->slab;
	//判断是否在pool的最后一页
    if (join < pool->last) {
        type = join->prev & NGX_SLAB_PAGE_MASK;

        if (type == NGX_SLAB_PAGE) {
			//若挂载在链表上则脱离
            if (join->next != NULL) {
                pages += join->slab;
                page->slab += join->slab;

                prev = (ngx_slab_page_t *) (join->prev & ~NGX_SLAB_PAGE_MASK);
                prev->next = join->next;
                join->next->prev = join->prev;

                join->slab = NGX_SLAB_PAGE_FREE;
                join->next = NULL;
                join->prev = NGX_SLAB_PAGE;
            }
        }
    }
	
    if (page > pool->pages) {
        join = page - 1;
        type = join->prev & NGX_SLAB_PAGE_MASK;
		//判断page类型
        if (type == NGX_SLAB_PAGE) {

            if (join->slab == NGX_SLAB_PAGE_FREE) {
                join = (ngx_slab_page_t *) (join->prev & ~NGX_SLAB_PAGE_MASK);
            }

            if (join->next != NULL) {
                pages += join->slab;
                join->slab += page->slab;

                prev = (ngx_slab_page_t *) (join->prev & ~NGX_SLAB_PAGE_MASK);
                prev->next = join->next;
                join->next->prev = join->prev;

                page->slab = NGX_SLAB_PAGE_FREE;
                page->next = NULL;
                page->prev = NGX_SLAB_PAGE;

                page = join;
            }
        }
    }
	//将要释放的page的最后一页指向开始的page
    if (pages) {
        page[pages].prev = (uintptr_t) page;
    }
	//加入free.next
    page->prev = (uintptr_t) &pool->free;
    page->next = pool->free.next;

    page->next->prev = (uintptr_t) page;

    pool->free.next = page;
}