[Devel] Re: [PATCH v3 3/6] expose fine-grained per-cpu data for cpuacct stats

[Glauber Costa]

On 05/30/2012 04:48 PM, Paul Turner wrote:
> a) the information in /proc/stat is actually much denser since it's
> "cpu VAL VAL VAL VAL" as opposed to "cpuX.FIELD VAL"

easily fixable here. Less descriptive, but we can use a header line
with the description much like how /proc/slabinfo does, and we still
have an extensible interface that is dense, at the same time.

> b) If it became a problem the /proc/stat case is actually fairly
> trivially fixable by defining each cpu as a record and "everything
> else" as a magic im-out-of-cpus value.
>
>> >
>> >  Now, if you guys are okay with a file per-cpu, I can do it as well.
>> >  It pollutes the filesystem, but at least protects against the fact that this
>> >  is kmalloc-backed.
>> >
> As I prefaced, I'm not sure there's much that can be trivially done
> about it.  This is really a fundamental limitation of how read_map()
> works.
>
> What we really need is a proper seq_file exposed through cftypes.
That can be done.