| 1 |
On 03/14/2013 11:18 PM, Robin H. Johnson wrote: |
| 2 |
> On Thu, Mar 14, 2013 at 10:32:30PM -0400, Michael Mol wrote: |
| 3 |
>>> As to how to accomplish this, it's either a throwaway sig, or poking the |
| 4 |
>>> agent protocol directly. |
| 5 |
>> The only trouble with that is if the agent is configured to only unlock |
| 6 |
>> keys for limited periods of time, then your initial check might catch |
| 7 |
>> the agent when the key is still unlocked, but your subsequent call to |
| 8 |
>> GPG comes after the timeout. I ran into this while trying to set up |
| 9 |
>> automated signing of debian packages I was building. |
| 10 |
> So Debian has a test-gpg function already? Do you know where in their |
| 11 |
> codebase it is? |
| 12 |
|
| 13 |
No idea; a build system I'd cobbled together at the time prodded |
| 14 |
gpg-agent to get an interactive auth. The build-and-package step took |
| 15 |
too long, leading to the timeout. And I apologize; I don't remember |
| 16 |
exactly what I was doing to prod gpg-agent, and since it was for a prior |
| 17 |
job, I did not retain any copies of any of the materials. |
| 18 |
|
| 19 |
> |
| 20 |
>> All it really means, in a practical procedural sense, is that you need |
| 21 |
>> to allow yourself a way to roll back anything you've been doing if that |
| 22 |
>> later check fails. |
| 23 |
> I think we'd do: |
| 24 |
> - All repoman checks |
| 25 |
> - initial file editing |
| 26 |
> if two-phase commit: |
| 27 |
> - test gpg |
| 28 |
> - commit1 |
| 29 |
> - gpg sign |
| 30 |
> - commit2 |
| 31 |
> if one-phase commit: |
| 32 |
> - gpg test |
| 33 |
> - gpg sign |
| 34 |
> - commit1 |
| 35 |
> |
| 36 |
> Unless commit1 took a really long time, the interval between the gpg |
| 37 |
> calls should be very small. |
| 38 |
> |
| 39 |
|
| 40 |
Murphy's going to call in a long I/O hang somewhere. Race conditions are |
| 41 |
hell manifest in systems. |
| 42 |
|
| 43 |
Since I haven't been paying close attention to this thread, I'm missing |
| 44 |
some context, but if you're commit1 and commit2 apply to a DVCS, you're |
| 45 |
probably fine so long as you don't do a push between commit1 and |
| 46 |
commit2, don't do a push if any earlier step failed, and do all this in |
| 47 |
a temporary branch that never (as a branch) gets pushed upstream. |
| 48 |
|
| 49 |
If you're not applying to a DVCS, then you risk interleaving commit1 and |
| 50 |
commit2 with others' work anyway. |
Archives