| 1 |
Hi Fabian, |
| 2 |
|
| 3 |
On Fri, Feb 29, 2008 at 08:51:49AM +0100, Fabian Groffen wrote: |
| 4 |
|
| 5 |
> > > I guess, the next step would be to wrap src_build() and src_install() |
| 6 |
> > > into a sandbox that can't leave $EPREFIX/var/tmp/portage. |
| 7 |
> > Attached is a first attempt at this. |
| 8 |
> For the moment, no, as I haven't ever looked at how it is done in |
| 9 |
> Portage. I suspect it is part of ebuild.sh or misc-functions.sh though. |
| 10 |
> You could grep for "sandbox" in usr/lib/portage/bin for a start. |
| 11 |
|
| 12 |
Actually it's done in pym/portage/dbapi/vartree.py by functions |
| 13 |
treewalk() and mergeme(). This is why it can't actually be sandboxed - |
| 14 |
it's running in the same python interpreter and no new executable is |
| 15 |
spawned. So the whole emerge would have to be sandboxed which can't seem |
| 16 |
to be done since mac os x sandboxes can't be nested. :-( |
| 17 |
|
| 18 |
But you've already changed treewalk() to call mergme() with |
| 19 |
EPREFIX.lstrip(os.path.sep) instead of "". So it will ignore anything in |
| 20 |
the temporary install image that would be installed outside of prefix. |
| 21 |
This should be just fine. |
| 22 |
|
| 23 |
Attached is a new version of my sandboxing patch. It now doesn't need a |
| 24 |
temporary profile file. Instead it passes the profile as a command line |
| 25 |
parameter to sandbox-exec. Also, I've changed the profile to be |
| 26 |
black-listing instead of white-listing: Everything is allowed by default, |
| 27 |
except writing to anything outside of EPREFIX. This should minimise |
| 28 |
breakage caused by the sandboxing. |
| 29 |
|
| 30 |
There's a new feature "macosprefixsandbox". If set, any action that |
| 31 |
would normally run unconfined, will be confined to a sandbox limiting |
| 32 |
write access to EPREFIX. This way even setup and postrm and the like |
| 33 |
can't leave the prefix. |
| 34 |
|
| 35 |
I did an emerge -ateND world with and without sandboxing. The log |
| 36 |
revealed some differences caused by the sandboxing which I'm in the |
| 37 |
process of tracking down. Most notably, the getcwd() function seems to |
| 38 |
behave differently in a sandbox. See |
| 39 |
http://lists.apple.com/archives/darwin-kernel/2008/Mar/msg00004.html for |
| 40 |
this. |
| 41 |
|
| 42 |
Differences I can't explain right now: |
| 43 |
|
| 44 |
- bash_logout was missing from the sandboxed build |
| 45 |
- some DynaLoader writemain call was added in sandboxed build |
| 46 |
- readline's configure suddenly checks for xterm and write in sandboxed build |
| 47 |
- a pnmtops configure check cleanly fails in sandboxed build instead of throwing |
| 48 |
an error message |
| 49 |
- a sort configure check fails |
| 50 |
|
| 51 |
A diff of my logs is attached, in case you can spot some obvious cause |
| 52 |
for the differences. I'll give the build another go to rule out the |
| 53 |
possibility that those differences are caused by something else. |
| 54 |
|
| 55 |
I have two goals now: |
| 56 |
- see, how I can integrate sandboxing into bootstrapping |
| 57 |
- start using prefix as root and see how it behaves |
| 58 |
|
| 59 |
My guess is, that the first will boil down to automatically adding |
| 60 |
FEATURES="macossandbox macosprefixsandbox" to make.conf at an early |
| 61 |
point. Since portage itself is bootstrapped into EPREFIX/tmp everything |
| 62 |
after that should be properly sandboxed. Do you think, it'd make sense |
| 63 |
to work on sandboxing steps before that? |
| 64 |
|
| 65 |
As for the latter, I run into the following problem: |
| 66 |
|
| 67 |
root@esgaroth:~ $ cd ~michael/bin/gentoo/ |
| 68 |
root@esgaroth:/Users/michael/bin/gentoo $ ./startprefix |
| 69 |
Entering Gentoo Prefix /Users/michael/bin/gentoo |
| 70 |
root@esgaroth:/Users/michael/bin/gentoo $ emerge prefixtest |
| 71 |
emerge: superuser access is required. |
| 72 |
|
| 73 |
Can you give me a quick hint on that one? |
| 74 |
-- |
| 75 |
thanks, Micha |
Archives