Re: 2.6.6-mm2

[Andrea Arcangeli <andrea@suse.de>]

On Fri, May 14, 2004 at 01:58:49PM -0700, Chris Wright wrote:
> gives the feel of cleaner hack), and is runtime safe (unless you care

what makes no sense to me is the "cleaner hack" approch. Since this is a
dirty hack anyways, trying to make it cleaner seems quite pointless, we
should keep it simple and localized instead, so it can be deleted with
minimal effort.  Having more than 1 hack (i.e. more than 1 sysctl) for
this as well seems pointless. Dealing with the groups as well seems
overkill and not needed.

The question is is if what you are proposing could be a long term
solution or not. If it cannot be a long term solution, then going with
a single disable_cap_mlock simplest of all hack is the best from my
point of view.

> that want safe gpg.  In fact, they probably aren't same machine, and I

they can or cannot be in the same machine, but the big question is if
the gpg user is "locally" trusted too or not. But this isn't just about
gpg. I had to put remap_file_pages under mlock too, not because of the
paging, paging of nonlinear VMAs works fine, but the truncate of the
nonlinear vmas doesn't work yet correctly. This will be eventually fixed
but in the short term I had to keep it under remap_file_pages under
mlock since you can mlock memory with remap_file_pages+truncate.

So if one group uses uml and the other group uses oracle, the group
approch won't work, only disable_cap_mlock will work. I can very well
imagine uml being run as nobody.nogroup or as wwwrun.www.

> [..] Well, anyway for gpg we only want rlimits, and this work is
> already done...

correct.