| 1 |
Hello, |
| 2 |
|
| 3 |
On Sun, 11 May 2014 23:42:38 +0200 Michał Górny wrote: |
| 4 |
> Hi, everyone. |
| 5 |
> |
| 6 |
> Almost 9 months ago I've committed three new FEATURES for portage: |
| 7 |
> cgroup, ipc-sandbox and network-sandbox. Today I'd like to propose |
| 8 |
> enabling at least the latter two by default. |
| 9 |
> |
| 10 |
> |
| 11 |
> Firstly, I'd like to shortly remind you what they do: |
| 12 |
> |
| 13 |
> 1. cgroup -- puts all processes spawned by ebuild to cgroup, and kills |
| 14 |
> all of them once phase exits (prevents leaving orphans), |
| 15 |
> |
| 16 |
> 2. ipc-sandbox -- puts all processes spawned by ebuild to a separate |
| 17 |
> IPC namespace, preventing them from interfacing other system services |
| 18 |
> via IPC (message queues, semaphores, shared memory), |
| 19 |
> |
| 20 |
> 3. network-sandbox -- puts all processes spawned by ebuild to |
| 21 |
> a separate network namespace with a private loopback interface, |
| 22 |
> preventing them from interfacing other system services, local network |
| 23 |
> and the Internet. |
| 24 |
> |
| 25 |
> |
| 26 |
> Secondly, the benefits of using the latter two include: |
| 27 |
> |
| 28 |
> 1. improved tree quality through more complete testing |
| 29 |
> |
| 30 |
> In the past, usually users with no or shoddy Internet access were |
| 31 |
> the first to notice ebuilds breaking with no Internet access. With |
| 32 |
> network-sandbox, the ebuilds fail consistently for everyone including |
| 33 |
> developers. They can notice the issues first hand and fix them before |
| 34 |
> committing the ebuild to the tree. |
| 35 |
> |
| 36 |
> 2. protection of user privacy (and security) |
| 37 |
> |
| 38 |
> Currently, programs spawned by ebuilds can submit any data |
| 39 |
> to the Internet, often unnoticed. While committing ebuild performing |
| 40 |
> malicious activity is unlikely, such uses as test suites sending |
| 41 |
> pre-defined data and possibly some system-specific data to remote |
| 42 |
> services happen. This affects user's privacy and we should prevent |
| 43 |
> ebuilds from doing that without user's approval. |
| 44 |
> |
| 45 |
> 3. preventing unsolicited (and possibly costly) Internet access |
| 46 |
> |
| 47 |
> Bear that Gentoo can be run on a machine with byte-paid or otherwise |
| 48 |
> shoddy Internet access (possibly with local distfile host or alike). |
| 49 |
> Ebuilds using network unwisely can reduce throughput of other local |
| 50 |
> services or even cause higher Internet bill. |
| 51 |
> |
| 52 |
> 4. improving security through preventing unsolicited access |
| 53 |
> |
| 54 |
> The build process (and especially tests) can spawn daemons and bind |
| 55 |
> them to network interfaces. Without network sandbox, other processes |
| 56 |
> (or systems if 0.0.0.0/:: is used) can access the spawned services |
| 57 |
> and possibly perform malicious operations. With network-sandbox, even |
| 58 |
> local users can't access the spawned daemons (due to private loopback |
| 59 |
> interface). |
| 60 |
> |
| 61 |
> 5. preventing data loss through thoughtless access to local services |
| 62 |
> |
| 63 |
> I have seen test suites that used production database servers running |
| 64 |
> on the host. You can imagine how much damage such a test suite could |
| 65 |
> cause by mistake. network-sandbox prevents the ebuild from accessing |
| 66 |
> local services, requiring it to run a safe, separate instance. |
| 67 |
> |
| 68 |
> |
| 69 |
> What do you think? |
| 70 |
|
| 71 |
Please do not enable them prior rigorous testing. |
| 72 |
|
| 73 |
I tried network-sandbox — this is a disaster. It brokes distcc |
| 74 |
completely since distcc client can't connect to remote servers (and |
| 75 |
even to a local one if any). |
| 76 |
|
| 77 |
Best regards, |
| 78 |
Andrew Savchenko |
Archives