| Summary: | sys-apps/portage: package moves should not cause a package to block itself (installed app-crypt/gnupg blocks itself) | ||
|---|---|---|---|
| Product: | Gentoo Linux | Reporter: | Christian Burger <c.burger> |
| Component: | New packages | Assignee: | Portage team <dev-portage> |
| Status: | RESOLVED FIXED | ||
| Severity: | normal | Keywords: | InVCS |
| Priority: | Normal | ||
| Version: | unspecified | ||
| Hardware: | All | ||
| OS: | Linux | ||
| Whiteboard: | |||
| Package list: | Runtime testing required: | --- | |
| Bug Depends on: | |||
| Bug Blocks: | 193766, 445274 | ||
| Attachments: |
emerge --info as requested
cave info gnupg shorter patch longer patch (but more elegant) patch to commit 228a860476d7543608b469c569ec1d4e70aa7f59 |
||
|
Description
Christian Burger
2011-05-14 09:29:40 UTC
Where did you get that copy of the ebuild? The RDEPEND you list there is NOT the same as the ebuild. Specifically, you have "!app-crypt/gnupg !<=app-crypt/gnupg-2.0.1", while the ebuild only has "!<=app-crypt/gnupg-2.0.1". Also, where is your emerge --info output? I cannot reproduce your problem. Thanks for your reply. (In reply to comment #1) > Where did you get that copy of the ebuild? The RDEPEND you list there is NOT > the same as the ebuild. No, it is not the same, you are right, it is the one from the /var/db/pkg after I installed the package via 'cave' the Paludis package manager. I am sorry if my explanation was too short. > Specifically, you have "!app-crypt/gnupg !<=app-crypt/gnupg-2.0.1", while the > ebuild only has "!<=app-crypt/gnupg-2.0.1". In my gnupg-2.0.17.ebuild (directly from the Portage tree) the depend is like you would expect it to be: RDEPEND="!static? ( ${COMMON_DEPEND_LIBS} ) ${COMMON_DEPEND_BINS} virtual/mta !app-crypt/gpg-agent !<=app-crypt/gnupg-2.0.1 selinux? ( sec-policy/selinux-gnupg ) nls? ( virtual/libintl )" It is just that cave adjusted the block on the old package name to reflect the new one -- according to a package move/rename done in Q2-2006: "!app-crypt/gpg-agent !<=app-crypt/gnupg-2.0.1" becomes "!app-crypt/gnupg !<=app-crypt/gnupg-2.0.1" > Also, where is your emerge --info output? I thought because of the nature of the problem, it would be irrelevant. But as I stated before I use Paludis' cave, so an emerge --info will only be mildly helpful, but I will attach it. > I cannot reproduce your problem. I understand that people using emerge do not have that problem. But according to http://www.gentoo.org/proj/en/devrel/handbook/handbook.xml?part=2&chap=5#doc_chap6 a block to the old package name is not required nor encouraged if a package is renamed or moved, simply let the package manager (pkgcore, emerge or paludis) do its magic -- at least that is my understanding for the moment. Created attachment 273687 [details]
emerge --info as requested
Portage: not sure if this is strictly a Paludis bug, or a case where Portage isn't picking up the rename in the same way (in the case that the rename does need to be removed). (In reply to comment #4) > Portage: not sure if this is strictly a Paludis bug, or a case where Portage > isn't picking up the rename in the same way (in the case that the rename does > need to be removed). If we fix portage so that the it doesn't create circular blockers when applying package moves, won't that solve the problem? (In reply to comment #5) > It is just that cave adjusted the block on the old package name to reflect the > new one -- according to a package move/rename done in Q2-2006: > > "!app-crypt/gpg-agent !<=app-crypt/gnupg-2.0.1" > > becomes > > "!app-crypt/gnupg !<=app-crypt/gnupg-2.0.1" This is a known bug in portage, since it's illogical to create a circular blocker when applying a package move. I'd like to fix portage's update_dbentry function so that it takes a parent argument and avoids creating circular blockers. Ok, so the gpg-agent block should be removed from the ebuilds? (In reply to comment #7) > Ok, so the gpg-agent block should be removed from the ebuilds? Well, I guess that's your decision. It wouldn't be a problem if portage's update_dbentry() function handled it better. The code that moves binary/installed packages should also be fixed to drop blockers that become circular as a result of the move. I have the same problem. I'm attaching 'cave info gnupg' as it seems is more relevant. Created attachment 282777 [details]
cave info gnupg
(In reply to comment #9) > I have the same problem. I'm attaching 'cave info gnupg' as it seems is more > relevant. We don't need any more info since the problem is well understood. The problem is that portage (and maybe paludis too) creates nonsensical circular blockers when doing package moves. The package move code (in portage, located in the update_dbentry() function) can be fixed to prevent creation of these kinds of nonsensical circular blockers. (In reply to comment #11) > (In reply to comment #9) > > I have the same problem. I'm attaching 'cave info gnupg' as it seems is more > > relevant. > > We don't need any more info since the problem is well understood. The > problem is that portage (and maybe paludis too) creates nonsensical circular > blockers when doing package moves. The package move code (in portage, > located in the update_dbentry() function) can be fixed to prevent creation > of these kinds of nonsensical circular blockers. Should this maybe be a Portage bug then? Or would you rather I closed it. Yeah, it's a Portage bug. Just another two cents from me: 1) I have to correct my initial description in retrospect: I later realized that (at least in my case) Paludis was not the cause of the problem. I did use the 'emerge' command shortly before the problem occurred, which caught up with a lot of package moves and so on from the last quarters -- Paludis did those already. Mixing Paludis and 'emerge' might sometimes not be the best idea. 2) Since version 0.74 Paludis ignores those self-blockers. http://ciaranm.wordpress.com/2012/04/15/paludis-0-74-0-released/ Though this helps the issue, it does not really solve the problem. 3) As far as I understand it, this is not a bug in Portage, but bad behavior by package maintainers. But I am not here to start a lengthy philosophical debate. :) 4) I will attach two patch proposals; either one should do the job. I did not test them, since I do not use 'emerge' nor do I have an environment set up where I could test intensively. Hope it helps. Created attachment 328744 [details, diff]
shorter patch
By regex search and replace, this patch resets obvious self-blockers after the update.
Created attachment 328746 [details, diff]
longer patch (but more elegant)
This patch prevents the offending self-block in the first place: "!cat/pkg" is not replaced.
Your patches don't apply to the latest (master branch): http://git.overlays.gentoo.org/gitweb/?p=proj/portage.git (In reply to comment #17) > Your patches don't apply to the latest (master branch): > > http://git.overlays.gentoo.org/gitweb/?p=proj/portage.git Ah, cool, I see. And there is a test environment. Nice. :) Wrote another patch and extended the test case for db-entry-updates. Attaching the output of "git log -u -1" A related question: Those db-updates only change the db-entries of the packages that are moved? It's not like another already installed package cat/A blocking a certain package cat/B will be updated if cat/B is moved to cat/C. Because then my patch would be harmful. I checked the source, but it appears to me that fixdbentries() is only used on the package to be moved. But then you (Zac) mentioned a additional "parent" argument: comment #6: > blocker when applying a package move. I'd like to fix portage's > update_dbentry function so that it takes a parent argument and avoids > creating circular blockers. That fits into my train of thought. Well, what do I know, first time looking at the Portage source ... But, I am a little bit confused about that, because if only db-entries of moved packages are updated, there might be dangling references to the old package name in the package database. Created attachment 329010 [details, diff]
patch to commit 228a860476d7543608b469c569ec1d4e70aa7f59
(In reply to comment #18) > A related question: Those db-updates only change the db-entries of the > packages that are moved? It's not like another already installed package > cat/A blocking a certain package cat/B will be updated if cat/B is moved to > cat/C. Because then my patch would be harmful. > > But, I am a little bit confused about that, because if only db-entries of > moved packages are updated, there might be dangling references to the old > package name in the package database. update_dbentry is called for *all* packages, in order to fix those dangling references. So, your patch needs to account for the parent. I've added a parent argument for this: http://git.overlays.gentoo.org/gitweb/?p=proj/portage.git;a=commit;h=0b4e6c27e87e57fac09fa28d3ed1320cb6a83ad9 Account for parent in update_dbentry: http://git.overlays.gentoo.org/gitweb/?p=proj/portage.git;a=commit;h=cae023ad64e403dd4246eab558bbffa9f611b724 Add tests based on patch from comment #19: http://git.overlays.gentoo.org/gitweb/?p=proj/portage.git;a=commit;h=3f289cb5dae80a801fb5b46a719939929ef454f9 (In reply to comment #21) > Account for parent in update_dbentry: > > http://git.overlays.gentoo.org/gitweb/?p=proj/portage.git;a=commit; > h=cae023ad64e403dd4246eab558bbffa9f611b724 Maybe I am wrong, but it seems to me that in pym/portage/dbapi/vartree.py: def move_ent(self, mylist, repo_match=None): [...] origmatches = self.match(origcp, use_cache=0) [...] for mycpv in origmatches: [...] fixdbentries([mylist], newpath, parent=mycpv) the old/original package name is given, but in pym/portage/update.py: def update_dbentry(update_cmd, mycontent, eapi=None, parent=None): [...] new_atom = Atom(token.replace(old_value, new_value, 1), eapi=eapi) [...] if new_atom.blocker and parent is not None and \ parent.cp == new_atom.cp and \ match_from_list(new_atom, [parent]): continue it is compared to the new package name. Thanks, you're right. This should fix it: http://git.overlays.gentoo.org/gitweb/?p=proj/portage.git;a=commit;h=c974a023882485b8eeae35bac35c1f00d1a0725b (In reply to comment #23) > Thanks, you're right. This should fix it: > > http://git.overlays.gentoo.org/gitweb/?p=proj/portage.git;a=commit; > h=c974a023882485b8eeae35bac35c1f00d1a0725b In that case, you might want to remove the method alltogether: $ grep fixdbentries portage/* -R pym/portage/update.py:def fixdbentries(update_iter, dbdir, eapi=None, parent=None): pym/portage/__init__.py: 'portage.update:dep_transform,fixdbentries,grab_updates,' + \ It looks to me like it isn't referenced anywhere anymore. Okay, I've added a deprecation warning to fixdbentries: http://git.overlays.gentoo.org/gitweb/?p=proj/portage.git;a=commit;h=a6a9b6dd6342a6009097944b38163ccc66908f0e This is fixed in 2.1.11.32 and 2.2.0_alpha143. |
