Memory application technology in Linux drivers: principles and methods

Memory is one of the most important resources in a Linux system. It can be used to store data, code, stacks, etc. Memory application and release is one of the most basic operations in Linux driver development. It involves concepts such as kernel space and user space, static allocation and dynamic allocation, continuous memory and non-contiguous memory. In this article, we will introduce the memory application technology in Linux drivers, including kmalloc, vmalloc, get_free_pages, dma_alloc_coherent and other functions, and give examples to illustrate their usage and precautions.

Let’s start with the basics. The picture below is the memory mapping model of Linux

1. Each process has its own process space. 0-3G of the process space is user space, and 3G-4G is kernel space
2. The user space of each process is not in the same physical memory page, but the kernel space of all processes corresponds to the same physical address
3. The address allocated by vmalloc can be high-end memory or low-end memory
4. Physical addresses of 0-896MB are linearly mapped to the physical mapping area.

“

”

Dynamic memory application

Like the application layer, the kernel program also needs to dynamically allocate memory. The difference is that the kernel process can control whether the allocated memory is in user space or kernel space. The former can be used to allocate memory to the heap area of ​​user space, eg , the malloc of the user space of the user process will eventually call back the memory allocation function of the kernel space through a system call. At this time, the memory allocation function belongs to the user process, and can allocate space in the heap area of ​​the user process and return it, ultimately making A user process obtains memory allocation in its own user space; the latter is only allocated in the kernel space, so the user process cannot directly access the space, so it is mostly used to meet the memory needs of the kernel program itself. The following is a common API for Linux kernel space application memory ：

kmalloc – kfree

The memory requested by kmalloc is continuous in physical memory. They have only a fixed offset from the real physical address, so there is a simple conversion relationship. This API is mostly used to apply for memory that is less than one page size. The bottom layer of kmalloc needs to call **__get_free_pages. The gtp_t flags indicating the memory type in the parameter is the abbreviation of this function. Commonly used memory types include GFP_USER, GFP_KERNEL, GFP_ATOMIC**.

• GFP_USER means allocating memory for user space pages and can be blocked;
• GFP_KERNEL is the most commonly used flag. Note that when using this flag to apply for memory, if it cannot be satisfied temporarily, it will cause the process to block. So, must not be used in interrupt processing functions, tasklets and kernel timers. Use GFP_KERNEL in non-process context! ! !
• GFP_ATOMIC can be used in the above three scenarios. This flag indicates that if the requested memory cannot be used, it will return immediately.

/**
 * kmalloc - allocate memory
 * @size: how many bytes of memory are required.
 * @flags: the type of memory to allocate.
 * The @flags argument may be one of:
 * %GFP_USER - Allocate memory on behalf of user.  May sleep.
 * %GFP_KERNEL - Allocate normal kernel ram.  May sleep.
 * %GFP_ATOMIC - Allocation will not sleep.  May use emergency pools.
 * 
 * For example, use this inside interrupt handlers.
 */
void *kmalloc(size_t size, gfp_t flags);
/**
 * kfree - free previously allocated memory
 * @objp: pointer returned by kmalloc.
 * If @objp is NULL, no operation is performed.
 */
void kfree(const void *objp);

The same series of APIs also have

void *kzalloc(size_t size, gfp_t flags)    

__get_free_pages – free_pages

__get_free_pages(), like kmalloc(), is physically continuous memory. This series of functions is the lowest-level method in the Linux kernel for obtaining free memory, because the underlying buddy algorithm is based on **(2^n )×PAGE_SIZE to manage memory, so theyalways allocate memory in units of pages**

unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)  
void free_pages(unsigned long addr, unsigned int order)  

The same series of APIs also have

unsigned long __get_free_page(gfp_t gfp)        
unsigned long get_zeroed_page(gfp_t gfp_mask)    
struct page *alloc_pages(gfp_t gfp_mask, unsigned int order)
void free_page(unsigned long addr)  

vmalloc – vfree

vmalloc在虚拟内存空间给出一块连续的内存区，实质上，这片连续的虚拟内存在物理内存中并不一定连续，所以vmalloc申请的虚拟内存和物理内存之间也就没有简单的换算关系，正因如此，vmalloc()通常用于分配远大于__get_free_pages()的内存空间，它的实现需要建立新的页表，此外还会调用使用GFP_KERN的kmalloc，so，一定不要在中断处理函数，tasklet和内核定时器等非进程上下文中使用vmalloc！

/**     
 * vmalloc  -  allocate virtually contiguous memory
 * @size:          allocation size
 * Allocate enough pages to cover @size from the page level allocator 
and map them into contiguous kernel virtual space.
 */
void *vmalloc(unsigned long size)   

/**
 *      vfree  -  release memory allocated by vmalloc()
 *      @addr:          memory base address
 */
void vfree(const void *addr)  

同系列的API还有

/**
 * vmalloc_32  -  allocate virtually contiguous memory (32bit addressable)
 * @size:          allocation size
 * Allocate enough 32bit PA addressable pages to cover @size from the 
page level allocator and map them into contiguous kernel virtual space.
 */
void *vmalloc_32(unsigned long size) 

slab缓存

我们知道，页是内存映射的基本单位，但内核中很多频繁创建的对象所需内存都不到一页，此时如果仍然按照页映射的方式，频繁的进行分配和释放就会造成资源的浪费，同时也会降低系统性能。为了解决的这样的问题，内核引入了slab机制，使对象在前后两次被使用时被分配在同一块内存或同一类内存空间，且保留了基本的数据结构，就可以大大提高效率。kmalloc的底层即是使用slab算法管理分配的内存的。注意，slab依然是以页为单位进行映射，只是映射之后分割这些页为相同的更小的单元，从而节省了内存。slab分配的单元不能小于32B或大于128K。

/**
 * kmem_cache_create - 创建slab缓存对象
 * @name:slab缓存区名字，
 * @size:slab分配的缓存区的每一个单元的大小
 * @align:缓存区内存的对齐方式，一般给0
 * @flags:控制分配的位掩码，
 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) to catch references 
to uninitialised memory.
 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check for buffer 
overruns.
 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware cacheline.  
This can be beneficial if you're counting cycles as closely as davem.
 * %SLAB_CACHE_DMA - Use GFP_DMA memory
 * %SLAB_STORE_USER - Store the last owner for bug hunting
 *define SLAB_PANIC - Panic if kmem_cache_create() fails 
 */
struct kmem_cache *kmem_cache_create(const char *name, size_t size,
 size_t align,unsigned long flags, void (*ctor)(void *))


/**
 * kmem_cache_alloc - Allocate an object from this cache. 
 * @cachep: The cache to allocate from.
 * @flags: See kmalloc().
 * The flags are only relevant if the cache has no available objects.
 */
void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)  


/**
 * kmem_cache_free - Deallocate an object
 * @cachep: The cache the allocation was from.
 * @objp: The previously allocated object.
 * Free an object which was previously allocated from this cache.
 */
void kmem_cache_free(struct kmem_cache *cachep, void *objp)  


void kmem_cache_destroy(struct kmem_cache *s)  

范例

//创建slab对象
struct kmem_cache_t *xj_sbcache;
xj_sbcache = kmem_cache_create("xjslab",sizeof(struct xj_unit_t),0,SLAB_CACHE_DMA|SLAB_PANIC，NULL,NULL);

//分配slab缓存
struct xj_unit_t *xj_unit;
xj_unit = kmem_cache_alloc(xj_sbcache,GFP_KERNEL);

/* 使用slab缓存 */

/* 释放slab缓存 */
kmem_cache_free(xj_sbcache, xj_unit);

/* 销毁slab缓存 */
kmem_cache_destroy(xj_sbcache);

内存池

除了slab机制，内核还提供了传统的内存池机制来管理小块内存的分配。内存池主要是用来解决可能出现的内存不足的情况，因为一个内存池在创建的时候就已经分配好了一内存，当我们用mempool_alloc向一个已经创建好的内存池申请申请内存时，该函数首先会尝试回调内存池创建时的分配内存函数，如果已经没有内存可以分配，他就会使用内存池创建时预先分配的内存，这样就可以避免因为无内存分配而陷入休眠，当然，如果预分配的内存也已经使用完毕，还是会陷入休眠。slab机制的目的是提高内存使用率以及内存管理效率，内存池的目的是避免内存的分配失败。下面是内核中提供的关于内存池的API

/**     
 * mempool_create - create a memory pool
 * @min_nr:    the minimum number of elements guaranteed to be  allocated for this pool.
 * @alloc_fn:  user-defined element-allocation function.
 * @free_fn:   user-defined element-freeing function.
 * @pool_data: optional private data available to the user-defined functions.
 *              
 * this function creates and allocates a guaranteed size, preallocated memory pool. 
The pool can be used from the mempool_alloc() and mempool_free() functions. 
 * This function might sleep. Both the alloc_fn() and the free_fn() functions
 might sleep - as long as the mempool_alloc() function is not called from IRQ contexts.
 */
mempool_t *mempool_create(int min_nr, mempool_alloc_t *alloc_fn, mempool_free_t 
*free_fn, void *pool_data)

/**     
 * mempool_alloc - allocate an element from a specific memory pool
 * @pool:      pointer to the memory pool which was allocated via mempool_create().
 * @gfp_mask:  the usual allocation bitmask.
 * this function only sleeps if the alloc_fn() function sleeps or returns NULL.
 Note that due to preallocation, this function never* fails when called from process
 contexts. (it might fail if called from an IRQ context.)
 */     
void * mempool_alloc(mempool_t *pool, gfp_t gfp_mask)    

/**
 * mempool_free - return an element to the pool.
 * @element:   pool element pointer.
 * @pool:      pointer to the memory pool which was allocated via mempool_create().
 *
 * this function only sleeps if the free_fn() function sleeps.
 */     
void mempool_free(void *element, mempool_t *pool)    

/**
 * mempool_destroy - deallocate a memory pool
 * @pool:      pointer to the memory pool which was allocated via mempool_create().
 *
 * Free all reserved elements in @pool and @pool itself. 
 This function only sleeps if the free_fn() function sleeps.
 */     
void mempool_destroy(mempool_t *pool)  

通过本文，我们了解了Linux驱动中的内存申请技术，它们各有优缺点和适用场景。我们应该根据实际需求选择合适的函数，并遵循一些基本原则，如匹配申请和释放函数，检查返回值是否为空，避免内存泄漏等。内存申请技术是Linux驱动开发中不可或缺的一部分，它可以保证驱动程序的正常运行和数据交换，也可以提升驱动程序的性能和稳定性。希望本文能够对你有所帮助和启发。

The above is the detailed content of Memory application technology in Linux drivers: principles and methods. For more information, please follow other related articles on the PHP Chinese website!