DPDK使用hugepage原理总结
hugepage原理参考http://blog.chinaunix.net/uid-28541347-id-5783934.html
DPDK版本：17.11.2
hugepage的作用：
1. 就是减少页的切换，页表项减少，产生缺页中断的次数也减少
2. 降低TLB的miss次数
1.DPDK使用前准备
- DPDK应用使用hugepage前，应保证系统已经配置hugepage
（配置参考https://blog.csdn.net/shaoyunzhe/article/details/54614077） - 将 hugetlbfs 特殊文件系统挂载到根文件系统的某个目录
mount-t hugetlbfs hugetlbfs /dev/hugepages （挂载默认的hugeage大小）mount-t hugetlbfs none /dev/hugepages_2mb-o pagesize=2MB(挂载2M的)
1G大页和2M大页必须挂载了才能使用。挂载其中一个，DPDK也能正常运行。
本测试时只设置了1G大页，具体信息如下：
挂载目录：cat /proc/mounts
2.DPDK使用hugepage代码分析
DPDK初始化函数rte_eal_init调用eal_hugepage_info_init初始化hugepage信息，
2.1. eal_hugepage_info_init初始化主要工作：
此函数主要收集可用hugepage信息（有多少页，挂载目录）。
- 进入”/sys/kernel/mm/hugepages“目录
- 寻找“hugepages-”开头的目录并获取此目录有后面的数字，就是hugepage大小，比如我系统下：
- 使用struct hugepage_info 结构体保存hugepage页面大小，挂载目录，可用页数。注意：如果对应大小hugepage没有挂载，此类hugepage则不会被DPDK程序使用
eg：比如我们没有执行mount-t hugetlbfs none /dev/hugepages_2mb-o pagesize=2MB，只挂载了mount-t hugetlbfs hugetlbfs /dev/hugepages，DPDK只会使用1G hugepage。
DPDK程序执行时打印“EAL: 2048 hugepages of size 2097152 reserved, but no mounted hugetlbfs found for that size”表明2M的没有挂载。 - 以下结构体就是保存hugepage信息的，这个信息后面初始化存储有用。
-
struct hugepage_info { uint64_t hugepage_sz; /**< size of a huge page */ const char *hugedir; /**< dir where hugetlbfs is mounted */ uint32_t num_pages[RTE_MAX_NUMA_NODES]; /**< number of hugepages of that size on each socket */ int lock_descriptor; /**< file descriptor for hugepage dir */ };本实验最后大页信息是：
hugepage_sz=1048576(1048576*1024)
hugedir="/dev/hugepages"
num_pages[0]=4
int
eal_hugepage_info_init(void)
{
const char dirent_start_text[] = "hugepages-";
const size_t dirent_start_len = sizeof(dirent_start_text) - 1;
unsigned i, num_sizes = 0;
DIR *dir;
struct dirent *dirent;
dir = opendir(sys_dir_path); //sys_dir_path[] = "/sys/kernel/mm/hugepages"
if (dir == NULL) {
RTE_LOG(ERR, EAL,
"Cannot open directory %s to read system hugepage info\n",
sys_dir_path);
return -1;
}
/*遍历/sys/kernel/mm/hugepages目录下以“hugepages-”开头的目录*/
for (dirent = readdir(dir); dirent != NULL; dirent = readdir(dir)) {
struct hugepage_info *hpi;
if (strncmp(dirent->d_name, dirent_start_text,
dirent_start_len) != 0)
continue;
if (num_sizes >= MAX_HUGEPAGE_SIZES)
break;
/*internal_config为DPDK全局变量*/
hpi = &internal_config.hugepage_info[num_sizes];
/*保存hugepage的大小，最多保存三种大小，一般也只用到了1G，2M*/
hpi->hugepage_sz =
rte_str_to_size(&dirent->d_name[dirent_start_len]);
/*get_hugepage_dir函数会到/proc/mounts里去寻找对应大小hugepage页挂载的目录 */
hpi->hugedir = get_hugepage_dir(hpi->hugepage_sz);
/* first, check if we have a mountpoint */
if (hpi->hugedir == NULL) {
uint32_t num_pages;
num_pages = get_num_hugepages(dirent->d_name);
if (num_pages > 0)
RTE_LOG(NOTICE, EAL,
"%" PRIu32 " hugepages of size "
"%" PRIu64 " reserved, but no mounted "
"hugetlbfs found for that size\n",
num_pages, hpi->hugepage_sz);
continue;
}
......
......
}2.2.rte_eal_hugepage_init初始化主要工作：
上面只是获取了hugepage信息，后面rte_eal_memory_init函数->rte_eal_hugepage_init->map_all_hugepages初始化每页具体虚拟地址，物理地址，大小等信息。
- 获取全局变量，存储分配内存相关信息
/* get pointer to global configuration */
mcfg = rte_eal_get_configuration()->mem_config;- 计算一共有多少页，并分配struct hugepage_file 结构管理所有页（如果设置了1G，2M，16G，nr_hugepages最后等于所有页数的总和，本测试nr_hugepages=4）
/* calculate total number of hugepages available. at this point we haven't
* yet started sorting them so they all are on socket 0 */
for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
/* meanwhile, also initialize used_hp hugepage sizes in used_hp */
used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
nr_hugepages += internal_config.hugepage_info[i].num_pages[0];
}
/*
* allocate a memory area for hugepage table.
* this isn't shared memory yet. due to the fact that we need some
* processing done on these pages, shared memory will be created
* at a later stage.
*/
tmp_hp = malloc(nr_hugepages * sizeof(struct hugepage_file));
if (tmp_hp == NULL)
goto fail;- 第一次调用map_all_hugepages创建内存映射文件。orig参数设置为1，下面解释了设置1或是0的作用
/* * Mmap all hugepages of hugepage table: it first open a file in * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to * map contiguous physical blocks in contiguous virtual blocks. */ static unsigned map_all_hugepages(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi, uint64_t *essential_memory __rte_unused, int orig)
eal_get_hugefile_path函数根据页的索引生成文件路径/dev/hugepages/rtemap_x（本测试是0，1，2，3），4个文件。然后调用open，mamp进行映射。然后把得到的虚拟地址存在hugepg_tbl[i].orig_va = virtaddr;
/* try to create hugepage file */
fd = open(hugepg_tbl[i].filepath, O_CREAT | O_RDWR, 0600);
if (fd < 0) {
RTE_LOG(DEBUG, EAL, "%s(): open failed: %s\n", __func__,
strerror(errno));
goto out;
}
/* map the segment, and populate page tables,
* the kernel fills this segment with zeros */
virtaddr = mmap(vma_addr, hugepage_sz, PROT_READ | PROT_WRITE,
MAP_SHARED | MAP_POPULATE, fd, 0);- 调用find_physaddrs函数获取每页虚拟地址对应的物理地址
/*
* For each hugepage in hugepg_tbl, fill the physaddr value. We find
* it by browsing the /proc/self/pagemap special file.
*/
static int
find_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
{
unsigned int i;
phys_addr_t addr;
for (i = 0; i < hpi->num_pages[0]; i++) {
addr = rte_mem_virt2phy(hugepg_tbl[i].orig_va);
if (addr == RTE_BAD_PHYS_ADDR)
return -1;
hugepg_tbl[i].physaddr = addr;
}
return 0;
}- 调用find_numasocket获取每页对应的socket ID。因为分配页内存时，在NUMA架构中会根据NUMA的内存分配策略决定在哪个NUMA节点分配。
if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
(unsigned)(hpi->hugepage_sz / 0x100000));
goto fail;
}- 根据每页的物理地址进行排序，排序的是struct hugepage_file *tmp_hp，tmp_hp存储了所有hugepage信息，是在一开始时初始化的。qsort排序的单位是一个struct hugepage_file结构体大小，排序依据是每页的物理地址大小。
qsort(&tmp_hp[hp_offset], hpi->num_pages[0],
sizeof(struct hugepage_file), cmp_physaddr);
static int
cmp_physaddr(const void *a, const void *b)
{
#ifndef RTE_ARCH_PPC_64
const struct hugepage_file *p1 = a;
const struct hugepage_file *p2 = b;
#else
/* PowerPC needs memory sorted in reverse order from x86 */
const struct hugepage_file *p1 = b;
const struct hugepage_file *p2 = a;
#endif
if (p1->physaddr < p2->physaddr)
return -1;
else if (p1->physaddr > p2->physaddr)
return 1;
else
return 0;
}
- 然后再次调用map_all_hugepages进行第二次映射。orig参数设置为0，这次和第一次调用有所区别。主要是是保证最大物理地址和最大虚拟地址都连续对应，此前已经保证物理地址是从小到大排序好了的。最后将新映射的地址保存到：hugepg_tbl[i].final_va = virtaddr；参考map_all_hugepages函数以下代码
else if (vma_len == 0) {
unsigned j, num_pages;
/* reserve a virtual area for next contiguous
* physical block: count the number of
* contiguous physical pages. */
for (j = i+1; j < hpi->num_pages[0] ; j++) {
#ifdef RTE_ARCH_PPC_64
/* The physical addresses are sorted in
* descending order on PPC64 */
if (hugepg_tbl[j].physaddr !=
hugepg_tbl[j-1].physaddr - hugepage_sz)
break;
#else
if (hugepg_tbl[j].physaddr !=
hugepg_tbl[j-1].physaddr + hugepage_sz)
break;
#endif
}
num_pages = j - i;
vma_len = num_pages * hugepage_sz;
/* get the biggest virtual memory area up to
* vma_len. If it fails, vma_addr is NULL, so
* let the kernel provide the address. */
vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
if (vma_addr == NULL)
vma_len = hugepage_sz;
}- 最后调用unmap_all_hugepages_orig取消第一次映射
/* unmap original mappings */
if (unmap_all_hugepages_orig(&tmp_hp[hp_offset], hpi) < 0)
goto fail;- 然后做一些清理工作，创建共享存储，umap不需要的页，最后将页信息保存到全局变量中
if (new_memseg) {
j += 1;
if (j == RTE_MAX_MEMSEG)
break;
mcfg->memseg[j].iova = hugepage[i].physaddr;
mcfg->memseg[j].addr = hugepage[i].final_va;
mcfg->memseg[j].len = hugepage[i].size;
mcfg->memseg[j].socket_id = hugepage[i].socket_id;
mcfg->memseg[j].hugepage_sz = hugepage[i].size;
}2.3.其他
rte_eal_hugepage_init只会被RTE_PROC_PRIMARY的进程调用（多进程情况下）。rte_eal_hugepage_init完成后只是将可用的大页内存物理地址，虚拟地址，socket id，大小信息保存到了全局变量中，怎么使用这些内存还需要进一步管理。