542 lines
16 KiB
Markdown
542 lines
16 KiB
Markdown
|
||
## usage
|
||
```shell
|
||
git submodule update --init --recursive
|
||
|
||
cd /home/lian/share/10.1-spdk/spdk
|
||
./configure --with-shared
|
||
make -j
|
||
|
||
make
|
||
|
||
# sometimes dd if=/dev/zero of=/dev/nvme0n1 bs=1M count=10
|
||
LD_PRELOAD=./libzvfs.so ./func_test
|
||
```
|
||
|
||
## 测试
|
||
### 总结
|
||
由于是目标是hook阻塞的API,相当于队列深度为1。
|
||
|
||
队列深度为1的情况下,spdk测试工具spdk_nvme_perf的测试结果:
|
||
1. iosize = 4K:100MiB/s
|
||
2. ioszie = 128K:1843MiB/s
|
||
|
||
zvfs的测试结果:
|
||
1. iosize = 4K:95MiB/s
|
||
2. ioszie = 128K:1662MiB/s
|
||
|
||
相当于spdk测试工具读写的90%性能。
|
||
|
||
对比系统调用:
|
||
1. O_DIRECT
|
||
1. 大块4K:43MiB/s
|
||
2. 小块128K:724MiB/s
|
||
2. !O_DIRECT
|
||
1. 大块4K:1460MiB/s
|
||
2. 小块128K:1266MiB/s
|
||
|
||
### spdk_nvme_perf 性能基准测试
|
||
```shell
|
||
cd /home/lian/share/10.1-spdk/spdk
|
||
|
||
export LD_LIBRARY_PATH=/home/lian/share/10.1-spdk/spdk/build/lib:/home/lian/share/10.1-spdk/spdk/dpdk/build/lib:$LD_LIBRARY_PATH
|
||
export PATH=/home/lian/share/10.1-spdk/spdk/build/bin:$PATH
|
||
|
||
./build/bin/spdk_nvme_perf \
|
||
-r 'trtype:PCIe traddr:0000:03:00.0' \
|
||
-q 1 -o 4096 -w randwrite -t 5
|
||
|
||
root@ubuntu:/home/lian/share/10.1-spdk/spdk# ./build/bin/spdk_nvme_perf -r 'trtype:PCIe traddr:0000:03:00.0' -q 1 -o 4096 -w randwrite -t 5
|
||
Initializing NVMe Controllers
|
||
Attached to NVMe Controller at 0000:03:00.0 [15ad:07f0]
|
||
Associating PCIE (0000:03:00.0) NSID 1 with lcore 0
|
||
Initialization complete. Launching workers.
|
||
========================================================
|
||
Latency(us)
|
||
Device Information : IOPS MiB/s Average min max
|
||
PCIE (0000:03:00.0) NSID 1 from core 0: 25765.92 100.65 38.77 16.58 802.32
|
||
========================================================
|
||
Total : 25765.92 100.65 38.77 16.58 802.32
|
||
|
||
|
||
./build/bin/spdk_nvme_perf \
|
||
-r 'trtype:PCIe traddr:0000:03:00.0' \
|
||
-q 32 -o 4096 -w randwrite -t 5
|
||
|
||
|
||
root@ubuntu:/home/lian/share/10.1-spdk/spdk# ./build/bin/spdk_nvme_perf -r 'trtype:PCIe traddr:0000:03:00.0' -q 32 -o 4096 -w randwrite -t 5
|
||
Initializing NVMe Controllers
|
||
Attached to NVMe Controller at 0000:03:00.0 [15ad:07f0]
|
||
Associating PCIE (0000:03:00.0) NSID 1 with lcore 0
|
||
Initialization complete. Launching workers.
|
||
========================================================
|
||
Latency(us)
|
||
Device Information : IOPS MiB/s Average min max
|
||
PCIE (0000:03:00.0) NSID 1 from core 0: 80122.94 312.98 399.36 36.31 2225.64
|
||
========================================================
|
||
Total : 80122.94 312.98 399.36 36.31 2225.64
|
||
|
||
|
||
./build/bin/spdk_nvme_perf \
|
||
-r 'trtype:PCIe traddr:0000:03:00.0' \
|
||
-q 1 -o 131072 -w write -t 5
|
||
|
||
root@ubuntu:/home/lian/share/10.1-spdk/spdk# export LD_LIBRARY_PATH=/home/lian/share/10.1-spdk/spdk/build/lib:/home/lian/share/10.1-spdk/spdk/dpdk/build/lib:$LD_LIBRARY_PATH
|
||
root@ubuntu:/home/lian/share/10.1-spdk/spdk# export PATH=/home/lian/share/10.1-spdk/spdk/build/bin:$PATH
|
||
root@ubuntu:/home/lian/share/10.1-spdk/spdk# ./build/bin/spdk_nvme_perf -r 'trtype:PCIe traddr:0000:03:00.0' -q 1 -o 131072 -w write -t 5
|
||
Initializing NVMe Controllers
|
||
Attached to NVMe Controller at 0000:03:00.0 [15ad:07f0]
|
||
Associating PCIE (0000:03:00.0) NSID 1 with lcore 0
|
||
Initialization complete. Launching workers.
|
||
========================================================
|
||
Latency(us)
|
||
Device Information : IOPS MiB/s Average min max
|
||
PCIE (0000:03:00.0) NSID 1 from core 0: 14746.80 1843.35 67.79 40.16 4324.96
|
||
========================================================
|
||
Total : 14746.80 1843.35 67.79 40.16 4324.96
|
||
|
||
|
||
./build/bin/spdk_nvme_perf \
|
||
-r 'trtype:PCIe traddr:0000:03:00.0' \
|
||
-q 32 -o 131072 -w write -t 5
|
||
|
||
root@ubuntu:/home/lian/share/10.1-spdk/spdk# ./build/bin/spdk_nvme_perf -r 'trtype:PCIe traddr:0000:03:00.0' -q 32 -o 131072 -w write -t 5
|
||
Initializing NVMe Controllers
|
||
Attached to NVMe Controller at 0000:03:00.0 [15ad:07f0]
|
||
Associating PCIE (0000:03:00.0) NSID 1 with lcore 0
|
||
Initialization complete. Launching workers.
|
||
========================================================
|
||
Latency(us)
|
||
Device Information : IOPS MiB/s Average min max
|
||
PCIE (0000:03:00.0) NSID 1 from core 0: 21997.40 2749.68 1455.09 96.64 26152.13
|
||
========================================================
|
||
Total : 21997.40 2749.68 1455.09 96.64 26152.13
|
||
```
|
||
### 系统调用
|
||
#### no O_DIRECT 小块
|
||
|
||
```shell
|
||
root@ubuntu:/home/lian/share/10.1-spdk/zvfs# ./func_test
|
||
|
||
=== test_single_file_perf ===
|
||
Path : /tmp/test.dat
|
||
IO size : 4 KB
|
||
Max file: 2048 MB
|
||
Duration: 10 sec
|
||
|
||
WRITE:
|
||
total : 12668.9 MB
|
||
time : 10.003 sec
|
||
IOPS : 324211 ops/sec
|
||
BW : 1266.45 MB/s
|
||
|
||
READ:
|
||
total : 7664.5 MB
|
||
time : 10.000 sec
|
||
IOPS : 196210 ops/sec
|
||
BW : 766.44 MB/s
|
||
|
||
=== all tests PASSED ===
|
||
```
|
||
#### no O_DIRECT 大块
|
||
|
||
```shell
|
||
root@ubuntu:/home/lian/share/10.1-spdk/zvfs# ./func_test
|
||
|
||
=== test_single_file_perf ===
|
||
Path : /tmp/test.dat
|
||
IO size : 128 KB
|
||
Max file: 2048 MB
|
||
Duration: 10 sec
|
||
|
||
WRITE:
|
||
total : 14609.5 MB
|
||
time : 10.000 sec
|
||
IOPS : 11688 ops/sec
|
||
BW : 1460.95 MB/s
|
||
|
||
READ:
|
||
total : 8138.6 MB
|
||
time : 10.000 sec
|
||
IOPS : 6511 ops/sec
|
||
BW : 813.85 MB/s
|
||
|
||
=== all tests PASSED ===
|
||
```
|
||
|
||
#### O_DIRECT 小块
|
||
```shell
|
||
root@ubuntu:/home/lian/share/10.1-spdk/zvfs# ./func_test
|
||
|
||
=== test_single_file_perf ===
|
||
Path : /tmp/test.dat
|
||
IO size : 4 KB
|
||
Max file: 2048 MB
|
||
Duration: 10 sec
|
||
|
||
WRITE:
|
||
total : 434.5 MB
|
||
time : 10.000 sec
|
||
IOPS : 11122 ops/sec
|
||
BW : 43.45 MB/s
|
||
|
||
READ:
|
||
total : 373.8 MB
|
||
time : 10.000 sec
|
||
IOPS : 9568 ops/sec
|
||
BW : 37.38 MB/s
|
||
|
||
=== all tests PASSED ===
|
||
```
|
||
#### O_DIRECT 大块
|
||
```shell
|
||
root@ubuntu:/home/lian/share/10.1-spdk/zvfs# ./func_test
|
||
|
||
=== test_single_file_perf ===
|
||
Path : /tmp/test.dat
|
||
IO size : 128 KB
|
||
Max file: 2048 MB
|
||
Duration: 10 sec
|
||
|
||
WRITE:
|
||
total : 7245.4 MB
|
||
time : 10.000 sec
|
||
IOPS : 5796 ops/sec
|
||
BW : 724.53 MB/s
|
||
|
||
READ:
|
||
total : 9006.5 MB
|
||
time : 10.000 sec
|
||
IOPS : 7205 ops/sec
|
||
BW : 900.64 MB/s
|
||
|
||
=== all tests PASSED ===
|
||
```
|
||
|
||
### SPDK
|
||
#### 非对齐
|
||
```shell
|
||
root@ubuntu:/home/lian/share/10.1-spdk/zvfs# LD_PRELOAD=./libzvfs.so ./func_test /zvfs
|
||
|
||
=== test_single_file_perf ===
|
||
Path : /zvfs/file.dat
|
||
IO size : 128 KB
|
||
Max file: 2048 MB
|
||
Duration: 10 sec
|
||
|
||
WRITE:
|
||
total : 10304.0 MB
|
||
time : 10.000 sec
|
||
IOPS : 8243 ops/sec
|
||
BW : 1030.40 MB/s
|
||
|
||
READ:
|
||
total : 17788.5 MB
|
||
time : 10.000 sec
|
||
IOPS : 14231 ops/sec
|
||
BW : 1778.85 MB/s
|
||
|
||
=== all tests PASSED ===
|
||
```
|
||
#### 全对齐大块
|
||
```shell
|
||
root@ubuntu:/home/lian/share/10.1-spdk/zvfs# LD_PRELOAD=./libzvfs.so ./func_test /zvfs
|
||
|
||
=== test_single_file_perf ===
|
||
Path : /zvfs/file.dat
|
||
IO size : 128 KB
|
||
Max file: 2048 MB
|
||
Duration: 10 sec
|
||
|
||
WRITE:
|
||
total : 16624.4 MB
|
||
time : 10.000 sec
|
||
IOPS : 13299 ops/sec
|
||
BW : 1662.43 MB/s
|
||
|
||
READ:
|
||
total : 16430.8 MB
|
||
time : 10.000 sec
|
||
IOPS : 13145 ops/sec
|
||
BW : 1643.07 MB/s
|
||
|
||
=== all tests PASSED ===
|
||
```
|
||
|
||
#### 全对齐小块
|
||
```shell
|
||
root@ubuntu:/home/lian/share/10.1-spdk/zvfs# LD_PRELOAD=./libzvfs.so ./func_test /zvfs
|
||
|
||
=== test_single_file_perf ===
|
||
Path : /zvfs/file.dat
|
||
IO size : 4 KB
|
||
Max file: 2048 MB
|
||
Duration: 10 sec
|
||
|
||
WRITE:
|
||
total : 944.5 MB
|
||
time : 10.000 sec
|
||
IOPS : 24179 ops/sec
|
||
BW : 94.45 MB/s
|
||
|
||
READ:
|
||
total : 982.8 MB
|
||
time : 10.000 sec
|
||
IOPS : 25159 ops/sec
|
||
BW : 98.28 MB/s
|
||
|
||
=== all tests PASSED ===
|
||
```
|
||
|
||
## SPDK
|
||
1. blob_store: blob仓库,管理多个blob对象。
|
||
2. blob: 存储对象,逻辑上连续,物理上不一定连续。相当于文件。
|
||
3. cluster: 分配单元,一个 blob 可以由多个 cluster 构成,扩容即分配新的 cluster。相当于文件系统的block group。
|
||
4. page: IO单元,一个 cluster 等于多个 page。
|
||
|
||
文件系统
|
||
|
||
## 架构设计
|
||
```scss
|
||
| 应用程序
|
||
| (POSIX API: open/read/write/close)
|
||
| LD_PRELOAD 拦截层
|
||
| (简单路径判断和转发到zvfs)
|
||
| zvfs 文件系统层
|
||
| (blob 操作)
|
||
| SPDK Blobstore
|
||
| 块设备 (Malloc0)
|
||
```
|
||
|
||
### 磁盘布局
|
||
```scss
|
||
BlobStore:
|
||
|—— Super Blob(元数据,使用SPDK的Super Blob锚定)
|
||
|——超级块
|
||
|——目录项/目录日志
|
||
|—— Blob 1 (文件A...)
|
||
|—— Blob 2 (文件B...)
|
||
|—— Blob N (文件C...)
|
||
```
|
||
|
||
|
||
### 数据结构
|
||
#### Super Blob(元数据)
|
||
```scss
|
||
[超级块]
|
||
- magic_number: 0x5A563146 (ZV1F)
|
||
- version: 1
|
||
|
||
[目录项]
|
||
- filename[256]: 文件名
|
||
- blob_id: 对应的数据blob ID
|
||
- file_size: 文件实际大小(字节)
|
||
- allocated_clusters: 已分配的cluster数量
|
||
- is_valid: 标记是否有效(用于删除)
|
||
```
|
||
|
||
```c
|
||
/* 目录项(内存中的目录) */
|
||
typedef struct {
|
||
char filename[256];
|
||
spdk_blob_id blob_id;
|
||
uint64_t file_size; // 文件逻辑大小(字节)
|
||
uint32_t allocated_clusters; // 已分配的cluster数量
|
||
bool is_valid; // false 表示已删除
|
||
int32_t open_count; // 打开的文件句柄数量
|
||
} zvfs_dirent_t;
|
||
|
||
/* 文件系统全局结构 */
|
||
typedef struct zvfs {
|
||
struct spdk_blob_store *bs;
|
||
struct spdk_io_channel *channel;
|
||
struct spdk_blob *super_blob; // 承载目录日志的blob
|
||
uint64_t io_unit_size; // page大小,单位字节
|
||
|
||
/* 目录 */
|
||
zvfs_dirent_t *dirents; // 目录项数组 #define ZVFS_MAX_FILES 1024
|
||
uint32_t dirent_count; // 当前有效项数
|
||
|
||
/* 伪FD表 */
|
||
struct zvfs_file *fd_table[ZVFS_MAX_FD]; // // e.g., #define ZVFS_MAX_FD 64
|
||
int fd_base; // 伪FD起始值,如10000
|
||
int openfd_count;
|
||
|
||
/* 元数据 */
|
||
uint32_t magic; // 0x5A563146 (ZV1F)
|
||
uint32_t version; // 1
|
||
} zvfs_t;
|
||
|
||
/* 打开的文件句柄 */
|
||
typedef struct zvfs_file {
|
||
zvfs_t *fs;
|
||
struct spdk_blob *blob;
|
||
zvfs_dirent_t *dirent; // 指回目录项 file_size/allocated_clusters
|
||
|
||
uint64_t current_offset; // 当前读写位置
|
||
int flags; // O_RDONLY / O_RDWR / O_CREAT 等
|
||
int pseudo_fd;
|
||
|
||
/* 临时DMA缓冲区(可选:每个file一个,避免每次malloc) */
|
||
void *dma_buf;
|
||
uint64_t dma_buf_size;
|
||
} zvfs_file_t;
|
||
```
|
||
|
||
### 工作流程
|
||
#### mount
|
||
hook POSIX API没有很好的调用时机,单线程目前采用懒加载。
|
||
```scss
|
||
1. [创建块设备]
|
||
- spdk_bdev_create_bs_dev_ext
|
||
2. [初始化文件系统]
|
||
- spdk_bs_init 或者 spdk_bs_load(已有数据时)
|
||
- spdk_bs_get_io_unit_size 获取io单元大小(page)
|
||
- spdk_bs_alloc_io_channel 分配blobstore的读写入口
|
||
3. [读取元数据]
|
||
- spdk_bs_get_super_blob 获取 Super Blob ID
|
||
- spdk_bs_open_blob 打开 Super Blob
|
||
- 读取超级块,校验 magic
|
||
- 读取目录项数组,加载到内存 dirents
|
||
4. [创建zvfs_t结构体]
|
||
- 创建 zvfs_t 结构体
|
||
- 填充 bs/channel/super_blob/dirents 等字段
|
||
```
|
||
#### open
|
||
##### O_RDONLY / O_RDWR
|
||
```scss
|
||
1. [文件名查找]
|
||
- 遍历 dirents,匹配 filename 且 is_valid=true
|
||
- 找不到返回 -ENOENT
|
||
2. [打开blob]
|
||
- spdk_bs_open_blob(dirent->blob_id)
|
||
- dirent->open_count++
|
||
- fs->openfd_count++
|
||
3. [分配文件句柄]
|
||
- 创建 zvfs_file_t,dirent 指针指向目录项
|
||
- 分配伪FD,写入 fd_table
|
||
5. [返回伪FD]
|
||
```
|
||
|
||
##### O_CREAT
|
||
```scss
|
||
1. [文件名查找]
|
||
- 遍历 dirents,若 filename 已存在且 is_valid=true,返回 -EEXIST
|
||
- 找一个 is_valid=false 的空槽位;没有空槽则追加(dirent_count < max_files)
|
||
2. [创建blob]
|
||
- spdk_bs_create_blob → 得到 blob_id
|
||
- spdk_bs_open_blob → 得到 blob 句柄
|
||
- spdk_blob_resize 初始分配空间
|
||
- spdk_blob_sync_md 持久化 cluster 分配
|
||
3. [写目录]
|
||
- 填充 filename/blob_id/file_size=0/is_valid=true
|
||
- dirent->open_count = 1
|
||
4. [创建文件句柄]
|
||
- 创建 zvfs_file_t
|
||
- 分配伪FD,写入 fd_table
|
||
5. [返回伪FD]
|
||
|
||
```
|
||
> 说明:目录变更只写内存,unmount 时统一持久化。
|
||
|
||
### read
|
||
读写都以字节为单位,offset / count 单位为字节;根据 io_unit_size 做对齐计算。
|
||
|
||
```scss
|
||
1. [参数]
|
||
- fd
|
||
- buffer
|
||
- count
|
||
- offset(隐含)
|
||
2. [边界检查]
|
||
- 实际可读 = min(count, dirent->file_size - current_offset)
|
||
- 实际可读为0则返回0
|
||
3. [计算Blob位置]
|
||
- start_page = current_offset / io_unit_size
|
||
- page_offset = current_offset % io_unit_size
|
||
- num_pages = (page_offset + 实际可读 + io_unit_size - 1) / io_unit_size
|
||
4. [DMA读取]
|
||
- 非对齐读(offset != 0 || count 不是整页)
|
||
- 需要DMA临时缓冲区(spdk_dma_zmalloc)
|
||
- spdk_blob_io_read(blob, channel, dma_buffer, start_page, num_pages, ...)
|
||
- 从 dma_buffer + page_offset 拷贝到用户 buffer
|
||
- 对齐
|
||
- 仍使用DMA缓冲区执行读取,再拷贝到用户buffer
|
||
5. [更新offset]
|
||
- current_offset += 实际可读
|
||
6. [返回实际读取字节数]
|
||
```
|
||
> 说明:SPDK需要DMA可用的内存,应用提供的用户缓冲区通常不满足要求。即便对齐也不能直接提交给spdk_blob_io_*,应使用DMA缓冲作为跳板;未来通过注册内存池可优化直传。
|
||
|
||
### write
|
||
```scss
|
||
1. [参数]
|
||
- fd
|
||
- buffer
|
||
- count
|
||
- offset(隐含)
|
||
2. [检查空间是否足够]
|
||
- 需要大小 = current_offset + count
|
||
- 若超过 allocated_clusters 对应容量:
|
||
- spdk_blob_resize 扩容
|
||
- spdk_blob_sync_md
|
||
- 更新 dirent->allocated_clusters
|
||
3. [计算写入位置]
|
||
- start_page / page_offset / num_pages(同read)
|
||
4. [DMA写入]
|
||
- 非对齐写(offset != 0 || count 不是整页)
|
||
- 读取涉及的首尾page到DMA临时缓冲区
|
||
- 修改对应位置的数据
|
||
- 写回:spdk_blob_io_write(blob, channel, dma_buffer, start_page, num_pages, ...)
|
||
- 对齐
|
||
- 仍通过DMA缓冲区提交写入
|
||
5. [更新状态]
|
||
- current_offset += count
|
||
- dirent->file_size = max(dirent->file_size, current_offset)
|
||
6. [返回写入字节数]
|
||
```
|
||
|
||
|
||
### close
|
||
```scss
|
||
1. [关闭Blob]
|
||
- spdk_blob_close(file->blob)
|
||
- dirent->open_count--
|
||
- fs->openfd_count++
|
||
- 若 open_count == 0 且 is_valid == false(已unlink):spdk_bs_delete_blob, 清空dirent
|
||
- 若 openfd_count == 0 则 unmount
|
||
2. [释放缓冲区]
|
||
- 释放 dma_buf
|
||
- 清除 fd_table[pseudo_fd]
|
||
- free(zvfs_file_t)
|
||
3. [返回0]
|
||
```
|
||
### unlink
|
||
```scss
|
||
1. [查找目录项]
|
||
- 遍历 dirents,匹配 filename 且 is_valid=true
|
||
- 找不到返回 -ENOENT
|
||
2. [标记删除]
|
||
- dirent->is_valid = false
|
||
3. [判断是否立即删除]
|
||
- open_count == 0:spdk_bs_delete_blob,清空该槽位
|
||
- open_count > 0:延迟,最后一个 close 负责删除
|
||
4. [返回0]
|
||
```
|
||
|
||
### unmount
|
||
```scss
|
||
1. [关闭channel]
|
||
- spdk_bs_free_io_channel
|
||
2. [关闭BlobStore]
|
||
- spdk_bs_unload
|
||
3. [释放FS]
|
||
- free(fs)
|
||
```
|
||
|
||
|
||
### 其他方案
|
||
如果不使用`LD_PRELOAD`hook,可以使用FUSE。\
|
||
FUSE是一种内核文件系统程序,挂载在文件目录上,对这个目录的访问,会使用这个文件系统程序。\
|
||
文件系统程序会将请求转发给应用层程序,这里的应用层程序可以是SPDK。这样就不用管其他的操作。
|