[Devel] Re: [PATCH 1/2] [RFC] Add file f_flags to cgroup write_string ops

[Matt Helsley]

On Wed, Oct 28, 2009 at 12:50:14PM +0530, Balbir Singh wrote:
> * menage at google.com <menage at google.com> [2009-10-27 23:06:19]:
> 
> > On Tue, Oct 27, 2009 at 11:04 PM, Paul Menage <menage at google.com> wrote:
> > > On Tue, Oct 27, 2009 at 6:04 PM, Li Zefan <lizf at cn.fujitsu.com> wrote:
> > >>
> > >> I think maybe it's better to store struct file *file to struct cftype,
> > >> so we don't need to change write_string(), write(), write_u64()
> > >> and write_s64().
> > >
> > > We can't do that - multiple open files could be using the same cftype
> > > at the same time. I'd be inclined if necessary to just pass the struct
> > > file* in, rather than risk needing to change it to pass more
> > > parameters later.
> > 
> > And I imagine that the number of handlers that actually make use of
> > the cftype* is rather small. If we pass the file* to the handler
> > instead of passing the cftype*, and provide an inline function to get
> > from the file* to the cftype*, then we can avoid adding an extra
> > parameter to the handlers.
> >
> 
> This sounds like a reasonable approach except for the cost of
> indirection (which I assume is acceptable). 

Suprisingly few places require the struct cftype *. So they don't pay
any additional cost. Note that the code in kernel/cgroup.c always
fetches the op pointer by fetching and dereferencing the struct cftype *.
So fetching it with an inline shouldn't add much to the data cache
footprint.

The minisclue savings come in two places where struct file * was already
passed.

Cheers,
	-Matt Helsley
_______________________________________________
Containers mailing list
Containers at lists.linux-foundation.org
https://lists.linux-foundation.org/mailman/listinfo/containers