Re: [PATCH v9 tip 8/9] samples: bpf: IO latency analysis (iosnoop/heatmap)

[Alexei Starovoitov <ast@plumgrid.com>]

On 3/23/15 12:40 AM, Ingo Molnar wrote:
>
> * Alexei Starovoitov <ast@plumgrid.com> wrote:
>
>> BPF C program attaches to blk_mq_start_request/blk_update_request kprobe events
>> to calculate IO latency.
>
>   ...
>
>> +/* kprobe is NOT a stable ABI
>> + * This bpf+kprobe example can stop working any time.
>> + */
>> +SEC("kprobe/blk_mq_start_request")
>> +int bpf_prog1(struct pt_regs *ctx)
>> +{
>> +	long rq = ctx->di;
>> +	u64 val = bpf_ktime_get_ns();
>> +
>> +	bpf_map_update_elem(&my_map, &rq, &val, BPF_ANY);
>> +	return 0;
>> +}
>
> So just to make sure the original BPF instrumentation model is still
> upheld: no matter in what way the kernel changes, neither the kprobe,
> nor the BPF program can ever crash or corrupt the kernel, assuming the
> kprobes, perf and BPF subsystem has no bugs, correct?

yes. of course. That was always #1 requirement.

> So 'stops working' here means that the instrumentation data might not
> be reliable if kernel internal interfaces change - but it won't ever
> make the kernel unreliable in any fashion. Right?

yes. of course.

The only situations where it can 'stop working':
- in-kernel blk_mq_start_request function is renamed, so kprobe cannot 
find it and cannot attach.
- arguments to blk_mq_start_request change. Then ctx->di can be 
meaningless and using it as key into map is useless.
- whole logic of blk_mq_start_request/blk_update_request pair changes.
then this sample code won't be measuring any useful io latency.

In all cases kernel will never crash (barring bugs in bpf, kprobe 
subsystems).