Re: [PATCH] perf: Move fs.* to generic lib/lk/

[Arnaldo Carvalho de Melo <acme@ghostprotocols.net>]

Em Fri, Nov 22, 2013 at 01:27:01PM +0100, Ingo Molnar escreveu:
> * Arnaldo Carvalho de Melo <acme@infradead.org> wrote:
> > Em Thu, Nov 21, 2013 at 04:28:04PM +0100, Borislav Petkov escreveu:
> > > On Thu, Nov 21, 2013 at 12:05:24PM -0300, Arnaldo Carvalho de Melo wrote:
> > > > Naming is a bit hard, to keep it small, descriptive, as API can lead
> > > > people to think about other kinds of kernel APIs (syscalls?), "fskapi"
> > > > to mean "fs based kernel API" would perhaps be more descriptive? A
> > > > longer (more descriptive) possibility would be "linux-fskapi".
 
> > > Yeah, you can't have fskapi because we'll add other stuff to it 
> > > (see the diffstat I sent you last week) so not filesystem stuff 
> > > only. So I think "kapi" is as generic and as fitting as it gets. 
> > > We can use the "kernel-api" variant but I think the "k" is enough.

> > I think is that it is too generic, the other stuff you mention is 
> > not really "kapi" at all.

> > The rest, things like util.c, usage.c, rbtree.c, hash, strlist, etc 
> > are all, well, utilities that we got from the kernel, from git, or 
> > that were created for perf, could get a tools/lib/util/ generic name 
> > and be outside the one with the description agreed above.

> > But they are not "helper methods to interface with the Linux kernel" 
> > at all.
 
> I don't think those other bits should go into this library. rbtree 
> should go into lib/rbtree/, command-line bits into lib/cmdline/, build 
> system helpers into lib/build/, etc.

Agreed.
 
> Merging unrelated things into a single library is a user-space disease 
> we need not repeat.

Agreed.
 
> I'd also not expose any of this externally but straight link it into 
> the individual utilities - that way it does not matter that it's a 
> nice, topical, fine-grained set of functionality.

Agreed.
 
> I don't think we are ready for (nor do we want the overhead of) 
> maintaining a library ABI at this stage.

Agreed.
 
> Once things slow down and it's all so robust that we've had at most a 
> handful of commits in tools/lib/ in a full year we can think about 
> exporting it, maybe ...

Agreed.

:-)

Lets experiment at having things at the right granularity, even if it
involves many, directly linked, like libperf.a, libraries, one at a
time, starting with fskapi (or whatever name ends up being preferred for
this initial one).

- Arnaldo