On 10/19/2017 11:05 AM, Ken Brown wrote:
> On 10/18/2017 11:01 PM, Ken Brown wrote:
>> On 10/17/2017 3:31 PM, Ken Brown wrote:
>>> On 10/17/2017 2:43 PM, Jon Turney wrote:
>>>> On 16/10/2017 20:13, Ken Brown wrote:
>>>>> This reverts (the rest of) commit b43b697.  Part of that commit was
>>>>> already reverted in commit ff0bb3d.  The rest is not needed either
>>>>> since we no longer send the upgrade flag to the solver after the user
>>>>> has made their selections.
>>>>> ---
>>>>>   libsolv.cc     | 14 +++-----------
>>>>>   libsolv.h      |  1 -
>>>>>   package_meta.h |  2 --
>>>>>   3 files changed, 3 insertions(+), 14 deletions(-)
>>>>
>>>> Hmm... not sure about this.
>>>>
>>>> Say the initial upgrade solution had something like: package A 1.0 
>>>> -> 1.1, and A 1.1 has a dependency on package B 2.0, where currently 
>>>> B 1.0 is installed (so B 1.0 -> 2.0)
>>>>
>>>> If the user then changes B to 'keep' (at 1.0), we should report a 
>>>> conflict?
>>>
>>> Good point.  I wasn't thinking about dependencies with version 
>>> relations.
>>
>> I just tested this scenario, and it turns out that putting a lock on B 
>> 1.0 overrides the dependency of A 1.1 on B 2.0.  There's no problem 
>> reported.  On the other hand, if there's no lock, then the dependency 
>> overrides the Keep choice, again with no problem reported.
>>
>> Back to the drawing board.
> 
> Thinking about this further, using SOLVER_LOCK is probably the right 
> thing to do in this situation, even if no problem is reported.  We can't 
> insist that users install the recommended version.  For example, if I 
> choose to downgrade binutils while keeping the current gcc, then I'm 
> responsible for any breakage this might cause, with or without a warning 
> from setup.
> 
> I'll prepare a new patch that restores the SOLVER_LOCK functionality 
> that was inadvertently removed in commit ff0bb3d.

Attached.  I kept making mistakes while doing this, so I hope I got it 
right in the end.

Here's a related question.  Currently if libsolv decides I should 
install something and I choose Skip instead, it will get installed 
anyway (with no problem report).  Maybe we should have a taskSkip that 
generates a SOLVER_LOCK in that case, analogous to taskKeep?

Ken