Re: [PATCH] CRED: Fix regression in cap_capable() as shown up by sys_faccessat() [ver #2]

[David Howells <dhowells@redhat.com>]

Serge E. Hallyn <serue@us.ibm.com> wrote:

> You have the 'acting_as' name for subj/eff, which I like.  Is there
> another name you could use in place of 'real' in the  name
> task_real_capable()?

Ummm...  'Actual' or 'Assigned' perhaps?

> I do find this version much easier to read.  It seems easier to
> track capable+current_cred() vs real_capable+get_task_cred().  Could
> you do a few benchmarks to gauge whether the difference the
> optimization makes?

Yeah...  My main objection is passing around two or three superfluous arguments
in the common case.  Most of the time, the only necessary argument to
sec->capable():

	int (*capable) (struct task_struct *tsk, const struct cred *cred,
			int cap, int audit);

is cap; tsk, cred and audit are all superfluous in the (very) common case.

How about:

	int (*fast_capable) (int cap);

which assumes current, current_cred() and SECURITY_CAP_AUDIT?

Benchmarking is tricky, given that the individual savings will be relatively
small in comparison to the code that calls them.

However, if I can get rid of three arguments passed into each of
security_capable(), selinux_capable() and cap_capable(), that really should
speed things up if you call it enough times, especially as current is held in a
register on some archs.

I'll see what I can do.

> I'm looking at a several-week-old linux-next, but only see one use of
> capable on another task which audits, and that is in commoncap for
> traceme, so it seems reasonable.

Should has_capability() be out of lines and have security_real_capable() merged
into it?  And the same for has_capability_noaudit() and
security_real_capable_noaudit()?

> So yeah, I do like this version better.

Perhaps a separate patch to optimise capable().  As I said, I'll see about
benchmarking it.

David