mirror of
https://github.com/eunomia-bpf/bpf-developer-tutorial.git
synced 2026-05-05 21:01:27 +08:00
fix bpf programs with useless attach
This commit is contained in:
yunwei37
@@ -1,121 +0,0 @@
|
||||
## eBPF 入门实践教程：编写 eBPF 程序 Biolatency: 统计系统中发生的I/O事件
|
||||
|
||||
### 背景
|
||||
|
||||
Biolatency 可以统计在该工具运行后系统中发生的I/O事件个数，并且计算I/O事件在不同时间段内的分布情况，以
|
||||
直方图的形式展现给用户。
|
||||
|
||||
### 实现原理
|
||||
|
||||
Biolatency 主要通过 tracepoint 实现，其在 block_rq_insert, block_rq_issue,
|
||||
block_rq_complete 挂载点下设置了处理函数。在 block_rq_insert 和 block_rq_issue 挂载点下，
|
||||
Biolatency 会将IO操作发生时的request queue和时间计入map中。
|
||||
```c
|
||||
int trace_rq_start(struct request *rq, int issue)
|
||||
{
|
||||
if (issue && targ_queued && BPF_CORE_READ(rq->q, elevator))
|
||||
return 0;
|
||||
|
||||
u64 ts = bpf_ktime_get_ns();
|
||||
|
||||
if (filter_dev) {
|
||||
struct gendisk *disk = get_disk(rq);
|
||||
u32 dev;
|
||||
|
||||
dev = disk ? MKDEV(BPF_CORE_READ(disk, major),
|
||||
BPF_CORE_READ(disk, first_minor)) : 0;
|
||||
if (targ_dev != dev)
|
||||
return 0;
|
||||
}
|
||||
bpf_map_update_elem(&start, &rq, &ts, 0);
|
||||
return 0;
|
||||
}
|
||||
|
||||
SEC("tp_btf/block_rq_insert")
|
||||
int block_rq_insert(u64 *ctx)
|
||||
{
|
||||
if (filter_cg && !bpf_current_task_under_cgroup(&cgroup_map, 0))
|
||||
return 0;
|
||||
|
||||
if (LINUX_KERNEL_VERSION < KERNEL_VERSION(5, 11, 0))
|
||||
return trace_rq_start((void *)ctx[1], false);
|
||||
else
|
||||
return trace_rq_start((void *)ctx[0], false);
|
||||
}
|
||||
|
||||
SEC("tp_btf/block_rq_issue")
|
||||
int block_rq_issue(u64 *ctx)
|
||||
{
|
||||
if (filter_cg && !bpf_current_task_under_cgroup(&cgroup_map, 0))
|
||||
return 0;
|
||||
|
||||
if (LINUX_KERNEL_VERSION < KERNEL_VERSION(5, 11, 0))
|
||||
return trace_rq_start((void *)ctx[1], true);
|
||||
else
|
||||
return trace_rq_start((void *)ctx[0], true);
|
||||
}
|
||||
|
||||
```
|
||||
在block_rq_complete 挂载点下，Biolatency 会根据 request queue 从map中读取
|
||||
上一次操作发生的时间，然后计算与当前时间的差值来判断其在直方图中存在的区域，将该区域内的IO操作
|
||||
计数加一。
|
||||
```c
|
||||
SEC("tp_btf/block_rq_complete")
|
||||
int BPF_PROG(block_rq_complete, struct request *rq, int error,
|
||||
unsigned int nr_bytes)
|
||||
{
|
||||
if (filter_cg && !bpf_current_task_under_cgroup(&cgroup_map, 0))
|
||||
return 0;
|
||||
|
||||
u64 slot, *tsp, ts = bpf_ktime_get_ns();
|
||||
struct hist_key hkey = {};
|
||||
struct hist *histp;
|
||||
s64 delta;
|
||||
|
||||
tsp = bpf_map_lookup_elem(&start, &rq);
|
||||
if (!tsp)
|
||||
return 0;
|
||||
delta = (s64)(ts - *tsp);
|
||||
if (delta < 0)
|
||||
goto cleanup;
|
||||
|
||||
if (targ_per_disk) {
|
||||
struct gendisk *disk = get_disk(rq);
|
||||
|
||||
hkey.dev = disk ? MKDEV(BPF_CORE_READ(disk, major),
|
||||
BPF_CORE_READ(disk, first_minor)) : 0;
|
||||
}
|
||||
if (targ_per_flag)
|
||||
hkey.cmd_flags = rq->cmd_flags;
|
||||
|
||||
histp = bpf_map_lookup_elem(&hists, &hkey);
|
||||
if (!histp) {
|
||||
bpf_map_update_elem(&hists, &hkey, &initial_hist, 0);
|
||||
histp = bpf_map_lookup_elem(&hists, &hkey);
|
||||
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(&start, &rq);
|
||||
return 0;
|
||||
}
|
||||
|
||||
```
|
||||
当用户中止程序时，用户态程序会读取直方图map中的数据，并打印呈现。
|
||||
|
||||
### Eunomia中使用方式
|
||||
|
||||
|
||||
### 总结
|
||||
Biolatency 通过 tracepoint 挂载点实现了对IO事件个数的统计，并且能以直方图的
|
||||
形式进行展现，可以方便开发者了解系统I/O事件情况。
|
||||
@@ -1,100 +0,0 @@
|
||||
## 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。
|
||||
@@ -1,63 +0,0 @@
|
||||
## eBPF 入门实践教程：编写 eBPF 程序 Bitesize: 监控块设备 I/O
|
||||
|
||||
### 背景
|
||||
|
||||
为了能更好的获得 I/O 操作需要的磁盘块大小相关信息，Bitesize 工具被开发。它可以在启动后追踪
|
||||
不同进程所需要的块大小，并以直方图的形式显示分布
|
||||
|
||||
### 实现原理
|
||||
|
||||
Biteszie 在 block_rq_issue 追踪点下挂在了处理函数。当进程对磁盘发出了块 I/O 请求操作时，
|
||||
系统会经过此挂载点，此时处理函数或许请求的信息，将其存入对应的map中。
|
||||
```c
|
||||
static int trace_rq_issue(struct request *rq)
|
||||
{
|
||||
struct hist_key hkey;
|
||||
struct hist *histp;
|
||||
u64 slot;
|
||||
|
||||
if (filter_dev) {
|
||||
struct gendisk *disk = get_disk(rq);
|
||||
u32 dev;
|
||||
|
||||
dev = disk ? MKDEV(BPF_CORE_READ(disk, major),
|
||||
BPF_CORE_READ(disk, first_minor)) : 0;
|
||||
if (targ_dev != dev)
|
||||
return 0;
|
||||
}
|
||||
bpf_get_current_comm(&hkey.comm, sizeof(hkey.comm));
|
||||
if (!comm_allowed(hkey.comm))
|
||||
return 0;
|
||||
|
||||
histp = bpf_map_lookup_elem(&hists, &hkey);
|
||||
if (!histp) {
|
||||
bpf_map_update_elem(&hists, &hkey, &initial_hist, 0);
|
||||
histp = bpf_map_lookup_elem(&hists, &hkey);
|
||||
if (!histp)
|
||||
return 0;
|
||||
}
|
||||
slot = log2l(rq->__data_len / 1024);
|
||||
if (slot >= MAX_SLOTS)
|
||||
slot = MAX_SLOTS - 1;
|
||||
__sync_fetch_and_add(&histp->slots[slot], 1);
|
||||
|
||||
return 0;
|
||||
}
|
||||
|
||||
SEC("tp_btf/block_rq_issue")
|
||||
int BPF_PROG(block_rq_issue)
|
||||
{
|
||||
if (LINUX_KERNEL_VERSION >= KERNEL_VERSION(5, 11, 0))
|
||||
return trace_rq_issue((void *)ctx[0]);
|
||||
else
|
||||
return trace_rq_issue((void *)ctx[1]);
|
||||
}
|
||||
```
|
||||
|
||||
当用户发出中止工具的指令后，其用户态代码会将map中存储的数据读出并逐进程的展示追踪结果
|
||||
|
||||
### Eunomia中使用方式
|
||||
|
||||
|
||||
### 总结
|
||||
Bitesize 以进程为粒度，使得开发者可以更好的掌握程序对磁盘 I/O 的请求情况。
|
||||
Reference in New Issue
Block a user