setup: problems with local install

[Ken Brown]

On 3/7/2018 4:52 PM, Ken Brown wrote:
> On 3/6/2018 1:47 PM, Jon Turney wrote:
>> On 06/03/2018 15:18, Jon Turney wrote:
>>> So yeah, I guess putting some complexity back in accessible() would 
>>> work, or perhaps the attached?  (This doesn't do the right thing for 
>>> a few packages, for reasons I'm still looking into...)
>>
>> To be specific it was doing the wrong thing for those few packages 
>> with no source
>>
>>> Â  /* scan for local copies of package */
>>> -void
>>> +bool
>>> Â  packagemeta::scan (const packageversion &pkg, bool mirror_mode)
>>> Â  {
>>> -Â  /* Already have something */
>>> +Â  /* empty version */
>>> Â Â Â  if (!pkg)
>>> -Â Â Â  return;
>>> +Â Â Â  return true;
>>
>> So, this needs to be 'return false', as the empty version is always 
>> inaccessible, to get the same behaviour as before.
> 
> I've found another problem with local installs: If a package needs 
> upgrading, then the chooser will offer the upgraded version for install, 
> even if there's no archive available.  As a result, the current version 
> will get uninstalled, and then setup will discover that it doesn't have 
> the archive to install the new version.
> 
> The problem occurs because packagedb::defaultTrust() is called after 
> ScanDownloadedFiles() has already done its work.  solution.update() and 
> solution.trans2db() are called, and pkg->desired is set equal to an 
> inaccessible version pv (which has been previously removed from 
> pkg->versions).
> 
> I guess trans2db() should check that pv is in pkg->versions before 
> acting on an install transaction for pv.  And then we also have to make 
> sure to ignore the erase transaction for the current version of pkg.
> 
> Alternatively, can we just remove an inaccessible packageversion from 
> the libsolv pool, or at at least just tell libsolv that we don't want to 
> install it?

Still another alternative, and maybe the simplest, is to make sure that 
the chooser never shows a version that is not in pkg->versions. 
packagemeta::set_action (trusts const trust) almost does this, except 
for one useless desired.accessible() call.  So that should be fixed, as 
well as the setting of the initial action.

Ken