Re: [PATCH v2 4/5] bpf: Implement function bpf_perf_event_read() that get the selected hardware PMU conuter

[Alexei Starovoitov <ast@plumgrid.com>]

On 7/22/15 7:12 PM, xiakaixu wrote:
> So you mean like this?
>
> u64 bpf_perf_event_read(u64 r1, u64 index,...)
> {
>    struct bpf_perf_event_array *array = (void *) (long) r1;
>    struct perf_event *event;
>    ...
>    event = array->events[index];
>    ...
> }

yes. the only thing needed is to add:
if (index >= array->map.max_entries)
	return -E2BIG;
before accessing array->events[index];

>> >
>>> >>+const struct bpf_func_proto bpf_perf_event_read_proto = {
>>> >>+    .func        = bpf_perf_event_read,
>>> >>+    .gpl_only    = false,
>>> >>+    .ret_type    = RET_INTEGER,
>>> >>+    .arg1_type    = ARG_CONST_MAP_PTR,
>>> >>+    .arg2_type    = ARG_PTR_TO_MAP_KEY,
>> >
>> >make it arg2_type = ARG_ANYTHING then you'll just index
>> >into array the way prog_array does and similar to bpf_tail_call.

>   ARG_ANYTHING means any (initialized) argument is ok, but we here

correct.

>   really want is map key. So I'm not sure ARG_ANYTHING is suitable.
>   You know ARG_ANYTHING is not checked enough in verifier.

why? during perf_event_array creation time we check that key_size == u32
so we can accept any integer.
ARG_PTR_TO_MAP_KEY forces program author to use stack instead of
passing index directly. Direct index is obviously faster.