mirror of
https://github.com/eunomia-bpf/bpf-developer-tutorial.git
synced 2026-03-22 21:00:14 +08:00
100 lines
3.1 KiB
Markdown
100 lines
3.1 KiB
Markdown
## eBPF 入门实践教程：编写 eBPF 程序 Biostacks: 监控内核 I/O 操作耗时
|
||
|
||
|
||
### 背景
|
||
由于有些磁盘I/O操作不是直接由应用发起的，比如元数据读写，因此有些直接捕捉磁盘I/O操作信息可能
|
||
会有一些无法解释的I/O操作发生。为此，Biostacks 会直接追踪内核中初始化I/O操作的函数，并将磁
|
||
盘I/O操作耗时以直方图的形式展现。
|
||
|
||
### 实现原理
|
||
Biostacks 的挂载点为 fentry/blk_account_io_start, kprobe/blk_account_io_merge_bio 和
|
||
fentry/blk_account_io_done。fentry/blk_account_io_start 和 kprobe/blk_account_io_merge_bio
|
||
挂载点均时内核需要发起I/O操作中必经的初始化路径。在经过此处时，Biostacks 会根据 request queue ，将数据存入
|
||
map中。
|
||
```c
|
||
static __always_inline
|
||
int trace_start(void *ctx, struct request *rq, bool merge_bio)
|
||
{
|
||
struct internal_rqinfo *i_rqinfop = NULL, i_rqinfo = {};
|
||
struct gendisk *disk = BPF_CORE_READ(rq, rq_disk);
|
||
dev_t dev;
|
||
|
||
dev = disk ? MKDEV(BPF_CORE_READ(disk, major),
|
||
BPF_CORE_READ(disk, first_minor)) : 0;
|
||
if (targ_dev != -1 && targ_dev != dev)
|
||
return 0;
|
||
|
||
if (merge_bio)
|
||
i_rqinfop = bpf_map_lookup_elem(&rqinfos, &rq);
|
||
if (!i_rqinfop)
|
||
i_rqinfop = &i_rqinfo;
|
||
|
||
i_rqinfop->start_ts = bpf_ktime_get_ns();
|
||
i_rqinfop->rqinfo.pid = bpf_get_current_pid_tgid();
|
||
i_rqinfop->rqinfo.kern_stack_size =
|
||
bpf_get_stack(ctx, i_rqinfop->rqinfo.kern_stack,
|
||
sizeof(i_rqinfop->rqinfo.kern_stack), 0);
|
||
bpf_get_current_comm(&i_rqinfop->rqinfo.comm,
|
||
sizeof(&i_rqinfop->rqinfo.comm));
|
||
i_rqinfop->rqinfo.dev = dev;
|
||
|
||
if (i_rqinfop == &i_rqinfo)
|
||
bpf_map_update_elem(&rqinfos, &rq, i_rqinfop, 0);
|
||
return 0;
|
||
}
|
||
|
||
SEC("fentry/blk_account_io_start")
|
||
int BPF_PROG(blk_account_io_start, struct request *rq)
|
||
{
|
||
return trace_start(ctx, rq, false);
|
||
}
|
||
|
||
SEC("kprobe/blk_account_io_merge_bio")
|
||
int BPF_KPROBE(blk_account_io_merge_bio, struct request *rq)
|
||
{
|
||
return trace_start(ctx, rq, true);
|
||
}
|
||
|
||
```
|
||
在I/O操作完成后，fentry/blk_account_io_done 下的处理函数会从map中读取之前存入的信息，根据当下时间
|
||
记录时间差值，得到I/O操作的耗时信息，并更新到存储直方图数据的map中。
|
||
```c
|
||
SEC("fentry/blk_account_io_done")
|
||
int BPF_PROG(blk_account_io_done, struct request *rq)
|
||
{
|
||
u64 slot, ts = bpf_ktime_get_ns();
|
||
struct internal_rqinfo *i_rqinfop;
|
||
struct rqinfo *rqinfop;
|
||
struct hist *histp;
|
||
s64 delta;
|
||
|
||
i_rqinfop = bpf_map_lookup_elem(&rqinfos, &rq);
|
||
if (!i_rqinfop)
|
||
return 0;
|
||
delta = (s64)(ts - i_rqinfop->start_ts);
|
||
if (delta < 0)
|
||
goto cleanup;
|
||
histp = bpf_map_lookup_or_try_init(&hists, &i_rqinfop->rqinfo, &zero);
|
||
if (!histp)
|
||
goto cleanup;
|
||
if (targ_ms)
|
||
delta /= 1000000U;
|
||
else
|
||
delta /= 1000U;
|
||
slot = log2l(delta);
|
||
if (slot >= MAX_SLOTS)
|
||
slot = MAX_SLOTS - 1;
|
||
__sync_fetch_and_add(&histp->slots[slot], 1);
|
||
|
||
cleanup:
|
||
bpf_map_delete_elem(&rqinfos, &rq);
|
||
return 0;
|
||
}
|
||
```
|
||
在用户输入程序退出指令后，其用户态程序会将直方图map中的信息读出并打印。
|
||
|
||
### Eunomia中使用方式
|
||
|
||
|
||
### 总结
|
||
Biostacks 从源头实现了对I/O操作的追踪，可以极大的方便我们掌握磁盘I/O情况。 |