On Tue, Nov 14, 2017 at 11:43 PM, Ingo Molnar <mi...@kernel.org> wrote:
>
> I also think that /proc/cpuinfo is a pretty bad interface for many uses - I
> personally only very rarely need the cpuinfo of _all_ CPUs.
>
> We we should eventually have /proc/cpu/N/info or so, so that 99% of the times
> cpuinfo is needed to report bugs we can do:
>
> cat /proc/cpu/0/info
I don't disagree in theory, and I used to enthusiastically support the
"let's make one value per file" model.
But in practice it's a pain, and I no longer really think it's viable
except for debugging.
As an example, I can do
cat /proc/cpuinfo | grep MHz
and get a nice overview of what's going on. If I were a system stat
program, I'd do that too.
In contrast, tell me where in /sys something useful is?
Honestly, the /sys directory structure _looks_ much more organized,
and it's lovely for some things, but it's a confusing jungle, and
finding things there is a complete and utter pain.
Yes, you can get the same thing:
cat /sys/devices/system/cpu/cpu*/cpufreq/scaling_cur_freq
but here's the deal: time that on your 120-core machine some day.
In other words, /proc/cpuinfo is not perfect. But the "let's split
things out" really doesn't work either. It's worse. It _sounds_ good
in theory, but it really really sucks.
I'm convinced that /sys is wonderful for management purposes. But no,
this whole "we should have individual files" is often a huge huge
mistake.
In fact, I think the current patch for CPU MHz shows exactly why
/proc/cpuinfo is actually hugely superior to the crazy "one file per
cpu" model: we could gather the statistics in parallel, all together,
and make it be reasonable.
Again, try that "cat" example again, and time it.
(And yes, I also know that "cpu*/cpufreq" is a symlink, but the direct
names are illogical, and no faster for me. Try it if you like:
time cat /sys/devices/system/cpu/cpufreq/policy*/scaling_cur_freq > /dev/null
and then look at /proc/cpuinfo again, and realize how lovely that
human-readable file actually is).
Linus
