ceph中cls 实现

Last updated on 2 months ago

概览

ceph提供了librados API来访问后端rados存储集群,上层业务都是基于 rados 提供的接口实现的

image-20240311220637974

随着项目规模的扩大,最初提供的简单增删改查 API 虽然能够满足基本业务需求,但在特殊场景下可能需要额外的 API 接口。逐渐增加的业务需求导致业务逻辑变得更加复杂,因此提供的 API 数量也会不断增加,这可能会使得底层的 RADOS 层变得臃肿,从解耦方面来看,特定的业务逻辑应该与基础架构分离,这样即使业务逻辑发生变化,也不会影响到基础架构的稳定性,实现业务逻辑与底层的解耦。

为此,ceph采用了动态插件的方式来处理这个问题;不同的应用程序可以根据自身的业务需求编写特定的模块(class),然后以插件的形式动态加载到 OSD上;OSD 进程会在启动时加载这些插件,当需要调用特定功能时再执行相应的插件。

本文主要介绍 ceph如何加载插件,以及业务层和rados 层是怎么通过 cls 交互。

先介绍几个关键函数

ceph 中cls 实现机制本质上是动态加载,简单来说是 使用 Linux 提供的 dlopen 函数,统一读取动态库,通过函数指针将这些函数持久到内存中,所以先介绍下 dlopen 使用

dlopen 打开动态链接库

1
void *dlopen(const char *filename, int flag);
  • filename:要打开的动态链接库的文件路径。
  • flag:打开动态链接库的标志
    • RTLD_LAZY(懒惰模式): 即在需要时才解析符号。这意味着在调用dlopen时并不会立即解析库中的符号,而是在第一次调用相关函数时才会解析

    • RTLD_NOW(立即模式): 表示立即加载,即在调用dlopen时立即解析库中的所有符号。这样会增加加载库的时间,但是在调用函数时不会出现未解析符号的错误

    • RTLD_GLOBAL: 表示将库中定义的符号加入全局符号表,这样加载同一库的其他代码也能够访问这些符号。如果不指定这个选项,则库中的符号只能被本库及其依赖项访问。使用这个选项可能会导致符号冲突和命名空间污染的问题,因此应该谨慎使用。

    • RTLD_LOCAL : 表示不将库中定义的符号加入全局符号表,而是使这些符号仅在加载该库的进程中可见。这是默认行为,除非指定了RTLD_GLOBAL选项。

dlsym 获取函数指针

1
void *dlsym(void *handle, const char *symbol);
  • handle:由 dlopen 返回的动态链接库句柄
  • symbol:要获取的函数或变量的名称

简单例子

创建一个共享库

1
2
3
4
5
6
7
8
9
10
11
12
[root@node86 tmp]# cat hello.c 
#include <stdio.h>

void hello() {
printf("tttttttt !\n");
}

[root@node86 tmp]# gcc -shared -o libhello.so hello.c -fPIC
[root@node86 tmp]# ll libhello.so
-rwxr-xr-x. 1 root root 8208 Mar 9 11:58 libhello.so
[root@node86 tmp]#

打开共享库

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
#include <stdio.h>
#include <dlfcn.h>

int main() {
void *handle;
void (*hello)();
// 打开动态链接库
handle = dlopen("./libhello.so", RTLD_LAZY);
if (!handle) {
fprintf(stderr, "%s\n", dlerror());
return 1;
}
// dlsym会通过 handle 找到符号地址,hello 也就是函数名字
hello = (void (*)()) dlsym(handle, "hello");
if (!hello) {
fprintf(stderr, "%s\n", dlerror());
dlclose(handle);
return 1;
}
// 此时 hello 就是个函数
// 调用函数 hello
hello();
// 关闭动态链接库
dlclose(handle);
return 0;
}

编译执行 , 符合 预期

1
2
3
4
5

[root@node86 tmp]# gcc main.c -o main -ldl
[root@node86 tmp]# ./main
tttttttt !
[root@node86 tmp]#

cls在那?

cls的实现代码都放在 源码目录下的cls 目录,并且根据不同模块的又划分了很多目录,(rgw 中很多实现 都是在用cls 方式实现的)

1
2
3
4
5
6
7
8
9
10
[root@node86 cls]# ls -l
total 8
drwxr-xr-x. 2 root root 94 Jan 23 22:22 cas
drwxr-xr-x. 2 root root 102 Jan 23 22:22 cephfs
-rw-r--r--. 1 root root 7491 Jan 23 22:22 CMakeLists.txt
.....
drwxr-xr-x. 2 root root 24 Jan 23 22:22 sdk
drwxr-xr-x. 2 root root 147 Jan 23 22:22 timeindex
drwxr-xr-x. 2 root root 170 Jan 23 22:22 user
drwxr-xr-x. 2 root root 165 Jan 23 22:22 version

若编译成功 , cls 全部模块都会编译成 .so动态库文文件存放编译目录下 lib 中,

1
2
3
4
5
6
7
8
9
10
11
/home/ceph/build/lib
[root@node86 lib]# ll |grep cls
-rw-r--r--. 1 root root 2876232 Jan 23 22:32 libcls_cas_client.a
lrwxrwxrwx. 1 root root 15 Jan 23 22:32 libcls_cas.so -> libcls_cas.so.1
....
....
lrwxrwxrwx. 1 root root 19 Jan 23 22:32 libcls_cas.so.1 -> libcls_cas.so.1.0.0
-rwxr-xr-x. 1 root root 4182408 Jan 23 22:32 libcls_cas.so.1.0.0
-rw-r--r--. 1 root root 8261516 Jan 23 22:33 libcls_cephfs_client.a
lrwxrwxrwx. 1 root root 18 Jan 23 22:32 libcls_cephfs.so -> libcls_cephfs.so.1

cls 代码是怎样的?

以为 ceph提供的 cls_hello daemon 为例

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
#include <algorithm>
#include <string>
#include <sstream>
#include <errno.h>
#include "objclass/objclass.h"
//cls 版本号
/*
#define CLS_VER(maj,min) \
int __cls_ver__## maj ## _ ##min = 0; \
int __cls_ver_maj = maj; \
int __cls_ver_min = min;
*/
CLS_VER(1,0)
//cls 名字
CLS_NAME(hello)
/*
#define CLS_NAME(name) \
int __cls_name__## name = 0; \
const char *__cls_name = #name;
*/
//要实现的函数
static int say_hello(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
{
if (in->length() > 100)
// 这里面函数是跑在 osd 上的
// 这不同于客户端,这是在存储集群进程osd直接访问
// this return value will be returned back to the librados caller
return 0;
}

/*
#define CLS_INIT(name) \
CEPH_CLS_API void __cls_init()
*/
// CLS_INIT 是必须的,加载 动态库后,会调用 init 函数,主要作用在于把cls中的函数指针持久到一个map中,为以后调用做准备
CLS_INIT(hello)
{
// this log message, at level 0, will always appear in the ceph-osd
// log file.
// cls的句柄,通过该句柄可以找到对应的库
cls_handle_t h_class;
// 动态库里面的函数句柄
cls_method_handle_t h_say_hello;
// 从系统 读到已加载的动态库
cls_register("hello", &h_class);

// There are two flags we specify for methods:
// RD : whether this method (may) read prior object state
// WR : whether this method (may) write or update the object
// A method can be RD, WR, neither, or both. If a method does
// neither, the data it returns to the caller is a function of the
// request and not the object contents.
//将 say_hello这个函数 注册到 h_say_hello
cls_register_cxx_method(h_class, "say_hello",
CLS_METHOD_RD,
say_hello, &h_say_hello);
}

怎么加载的?

ceph 中 有个 ClassHandler 类 专门管理 cls ,osd 在启动的过程中初始这个类,并且将全部动态库加入到内存中,具体实现在 open_all_classes

1
2
3
4
5
6
7
8
9
10
11
12
int OSD::init(){
//.....
class_handler = new ClassHandler(cct);
cls_initialize(class_handler);

if (cct->_conf->osd_open_classes_on_start) {
int r = class_handler->open_all_classes();
if (r)
dout(1) << "warning: got an error loading one or more classes: " << cpp_strerror(r) << dendl;
}
//....
}

读取lib目录下so文件

做一些字符匹配,找到符合cls 的动态文件

open_all_classes

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
int ClassHandler::open_all_classes()
{
ldout(cct, 10) << __func__ << dendl;
//一个目录,并返回一个指向目录流的指针 ,相当于 cd到某个目录
DIR *dir = ::opendir(cct->_conf->osd_class_dir.c_str());
if (!dir)
return -errno;

struct dirent *pde = nullptr;
int r = 0;
// 循环 读取目录中的下一个条目。相当于 ls 这个目录
while ((pde = ::readdir(dir))) {
if (pde->d_name[0] == '.')
continue;
//做一些字符匹配 , 找到 对应 符合 cls命名规则的 so库
if (strlen(pde->d_name) > sizeof(CLS_PREFIX) - 1 + sizeof(CLS_SUFFIX) - 1 &&
strncmp(pde->d_name, CLS_PREFIX, sizeof(CLS_PREFIX) - 1) == 0 &&
strcmp(pde->d_name + strlen(pde->d_name) - (sizeof(CLS_SUFFIX) - 1), CLS_SUFFIX) == 0) {
char cname[PATH_MAX + 1];
strncpy(cname, pde->d_name + sizeof(CLS_PREFIX) - 1, sizeof(cname) -1);
cname[strlen(cname) - (sizeof(CLS_SUFFIX) - 1)] = '\0';
ldout(cct, 10) << __func__ << " found " << cname << dendl;
//每个cls 动态库交由一个 ClassData管理
ClassData *cls;
// skip classes that aren't in 'osd class load list'
r = open_class(cname, &cls);
if (r < 0 && r != -EPERM)
goto out;
}
}
out:
closedir(dir);
return r;
}

注册并加载 cls

open_class

生成 cls 的hangle,并持久化到map 中

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
int ClassHandler::open_class(const string& cname, ClassData **pcls)
{
std::lock_guard lock(mutex);
// 在 map 找,看有没有相同的cname,没有生成一个 新的 ClassData 并保存到一个map 中
ClassData *cls = _get_class(cname, true);
if (!cls)
return -EPERM;
//如果没打开过 ,怎开始加载
if (cls->status != ClassData::CLASS_OPEN) {
int r = _load_class(cls);
if (r)
return r;
}
*pcls = cls;
return 0;
}

load_class 调用系统函数 dlopen 加载到内存中

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
int ClassHandler::_load_class(ClassData *cls)
{
// already open
if (cls->status == ClassData::CLASS_OPEN)
return 0;
.....
// RTLD_NOW 立即加载的方式
cls->handle = dlopen(fname, RTLD_NOW);
if (!cls->handle) {
struct stat st;
....
cls->status = ClassData::CLASS_MISSING;
return r;
}

cls_deps_t *(*cls_deps)();
//对于每个 shard obj都有可能依赖其他类,先统一加到 missing_dependencies 中
cls_deps = (cls_deps_t *(*)())dlsym(cls->handle, "class_deps");
if (cls_deps) {
cls_deps_t *deps = cls_deps();
while (deps) {
if (!deps->name)
break;
ClassData *cls_dep = _get_class(deps->name, false);
cls->dependencies.insert(cls_dep);
if (cls_dep->status != ClassData::CLASS_OPEN)
cls->missing_dependencies.insert(cls_dep);
deps++;
}
}
}

// resolve dependencies
//统一加载 依赖
set<ClassData*>::iterator p = cls->missing_dependencies.begin();
while (p != cls->missing_dependencies.end()) {
ClassData *dc = *p;
int r = _load_class(dc);
if (r < 0) {
cls->status = ClassData::CLASS_MISSING_DEPS;
return r;
}

ldout(cct, 10) << "_load_class " << cls->name << " satisfied dependency " << dc->name << dendl;
cls->missing_dependencies.erase(p++);
}

// initialize
// 对于每个cls 都有个 __cls_init 函数
// __cls_init 会将该动态库内的函数加入(注册)到一个全局map, 以后要是想调用,就根据 函数名字来找到对应函数
void (*cls_init)() = (void (*)())dlsym(cls->handle, "__cls_init");
if (cls_init) {
cls->status = ClassData::CLASS_INITIALIZING;
cls_init();
}

ldout(cct, 10) << "_load_class " << cls->name << " success" << dendl;
cls->status = ClassData::CLASS_OPEN;
return 0;


}

__cls_init 具体做了什么?

以为 ceph提供的 cls_hello daemon 为例, CLS_INIT 最主要的是实现cls 中的函数注册到 一个 cls 对应的map 中

以下以注册 say_hello为例子

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
 static int say_hello(cls_method_context_t hctx, bufferlist *in, bufferlist *out)
{
if (in->length() > 100)
// 这里面函数是跑在 osd 上的
// 这不同于客户端,这是在存储集群进程osd直接访问
// this return value will be returned back to the librados caller
return 0;
}

/*
#define CLS_INIT(name) \
CEPH_CLS_API void __cls_init()
*/
// CLS_INIT 是必须的,加载 动态库后,会调用 init 函数,主要作用在于把cls中的函数指针持久到一个map中,为以后调用做准备
CLS_INIT(hello)
{
// cls的句柄,通过该句柄可以找到对应的库
cls_handle_t h_class;
// 动态库里面的函数句柄
cls_method_handle_t h_say_hello;
// 从系统 读到已加载的动态库 , 通过 _get_class获取 对应的 ClassData,在 _load_class 已经保存过了
cls_register("hello", &h_class);

// There are two flags we specify for methods:
// RD : whether this method (may) read prior object state
// WR : whether this method (may) write or update the object
// A method can be RD, WR, neither, or both. If a method does
// neither, the data it returns to the caller is a function of the
// request and not the object contents.
//将 say_hello这个函数 注册到 h_say_hello
cls_register_cxx_method(h_class, "say_hello",
CLS_METHOD_RD,
say_hello, &h_say_hello);
}

cls_register_cxx_method

1
int cls_register_method(cls_handle_t hclass, const char *method, int flags,cls_method_call_t class_call, cls_method_handle_t *handle);
  • hclass : cls对应的 类(ClassData)

  • method: cls 函数名

  • flags: 指定cls 函数是什么操作类型的,有三种,三种模式可以并存

    • CLS_METHOD_RD
    • CLS_METHOD_WR
    • CLS_METHOD_PROMOTE
  • class_call: cls 函数的函数指针

  • handle: cls 函数handle

1
2
3
4
5
6
7
8
9
10
11
// cls_handle_t 是个空指针   typedef void *cls_handle_t;
//cls_method_handle_t 也是空指针

int cls_register_cxx_method(cls_handle_t hclass, const char *method,int flags,cls_method_cxx_call_t class_call, cls_method_handle_t *handle)
{
ClassHandler::ClassData *cls = (ClassHandler::ClassData *)hclass;
cls_method_handle_t hmethod = (cls_method_handle_t)cls->register_cxx_method(method, flags, class_call);
if (handle)
*handle = hmethod;
return (hmethod != NULL);
}

register_cxx_method

将 cls 函数指针,以及相关信息持久化到 map 中

1
2
3
4
5
6
7
8
9
ClassHandler::ClassMethod *ClassHandler::ClassData::register_cxx_method(const char *mname,int  flags, cls_method_cxx_call_t func)
{
ClassMethod& method = methods_map[mname];
method.cxx_func = func;
method.name = mname;
method.flags = flags;
method.cls = this;
return &method;
}

以上 是一个cls 模块代码 加载到osd内存中的流程
image-20240311232306333

上层业务怎么调用的?

cls 代码是跑在osd 上的,业务层要想调用cls模块的相关函数,需要将发送一个op 给osd ,告诉osd 要执行那个cls 函数,osd 收到op 请求后会执行从对应的cls handle 中找到相应函数函数指针并执行;

以gc remove 为例子,gc 流程后面需要生成 gc对象上的omap ,此时就是调用了cls 函数

1
2
3
4
5
6
7
int RGWGC::remove(int index, const std::vector<string>& tags, AioCompletion **pc)
{
ObjectWriteOperation op;
//cls 入口
cls_rgw_gc_remove(op, tags);
return store->gc_aio_operate(obj_names[index], &op, pc);
}

cls_rgw_gc_remove

1
2
3
4
5
6
7
8
9
10
11
12
void cls_rgw_gc_remove(librados::ObjectWriteOperation& op, const vector<string>& tags)
{
bufferlist in;
//cls函数需要的结构体,编码到 buferlist
cls_rgw_gc_remove_op call;
call.tags = tags;
encode(call, in);
// RGW_CLASS 说明是 rgw的 cls
// RGW_GC_REMOVE 是个宏定义
// #define RGW_GC_REMOVE "gc_remove"
op.exec(RGW_CLASS, RGW_GC_REMOVE, in);
}

op.exec

1
2
3
4
5
void librados::ObjectOperation::exec(const char *cls, const char *method, bufferlist& inbl, bufferlist *outbl, int *prval)
{
::ObjectOperation *o = &impl->o;
o->call(cls, method, inbl, outbl, NULL, prval);
}

o->call

1
2
3
4
5
void call(const char *cname, const char *method, bufferlist &indata)
{
add_call(CEPH_OSD_OP_CALL, cname, method, indata, NULL, NULL, NULL); //
}

add_call 后将 op 发送给osd(这里不在详细介绍怎么发,以及发个那个osd)

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
void add_call(int op, const char *cname, const char *method, bufferlist &indata, bufferlist *outbl, Context *ctx, int *prval)
{
OSDOp& osd_op = add_op(op); // 增加op
// 初始化op各成员
unsigned p = ops.size() - 1;
out_handler[p] = ctx;
out_bl[p] = outbl;
out_rval[p] = prval;
osd_op.op.op = op;
osd_op.op.cls.class_len = strlen(cname);
osd_op.op.cls.method_len = strlen(method);
osd_op.op.cls.indata_len = indata.length();
// 插件的所有输入参数,都是此次写操作的indata,包括插件名,方法名,输入参数
osd_op.indata.append(cname, osd_op.op.cls.class_len);
osd_op.indata.append(method, osd_op.op.cls.method_len);
osd_op.indata.append(indata);
}

假设osd 收到了 op , osd 会根据op的类型来做不同处理,刚才add_call 加入的是 CEPH_OSD_OP_CALL 类型的 op

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
int ReplicatedPG::do_osd_ops(OpContext *ctx, vector<OSDOp>& ops){
case CEPH_OSD_OP_CALL:
{
string cname, mname;
bufferlist indata;
//op里面携带了cls函数需要的参数,直接copy出来
try {
bp.copy(op.cls.class_len, cname);
bp.copy(op.cls.method_len, mname);
bp.copy(op.cls.indata_len, indata);
} catch (buffer::error& e) {
.....
}
// 根据 cname 找到对应的 Cla ClassData
ClassHandler::ClassData *cls;
result = osd->class_handler->open_class(cname, &cls);
ceph_assert(result == 0); // init_op_flags() already verified this works.
//在根据 函数名字,从map 找到对应 method (里面包含了函数相关信息)
ClassHandler::ClassMethod *method = cls->get_method(mname.c_str());
...
bufferlist outdata;
int prev_rd = ctx->num_read;
int prev_wr = ctx->num_write;
//执行 cls 函数
result = method->exec((cls_method_context_t)&ctx, indata, outdata);
//....
}

这里再看下 exec, 函数里面有分为两种执行方式,因为 cls 可能是以 c 或者 c++写的,比如 c写的 不能用直接用string ,所以要分开来

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
int ClassHandler::ClassMethod::exec(cls_method_context_t ctx, bufferlist& indata, bufferlist& outdata)
{
int ret;
if (cxx_func) {
// C++ call version
ret = cxx_func(ctx, &indata, &outdata);
} else {
// C version
char *out = NULL;
int olen = 0;
ret = func(ctx, indata.c_str(), indata.length(), &out, &olen);
if (out) {
// assume *out was allocated via cls_alloc (which calls malloc!)
buffer::ptr bp = buffer::claim_malloc(olen, out);
outdata.push_back(bp);
}
}
return ret;
}

以上就是cls 加载和实现流程; rgw 中有很多cls 的操作,所用需要深入的了解 cls的调用流程,如果有业务需求也可以写相关的cls 函数;
此外 这种加载函数的编码方式也是很有借鉴价值的;后续会再介绍下 cls 锁的实现。

相关引用

http://blog.wjin.org/posts/ceph-class-plugin.html
https://makedist.com/posts/2015/12/06/load-lua-rados-classes-from-local-file-system/