In today's meeting the council has decided that "REQUIRED_USE" shall be part of EAPI 4. See URL for its specification.
I notice the spec's missing things like which metadata cache line this goes on. Or is it not cached?
(In reply to comment #1) > I notice the spec's missing things like which metadata cache line this > goes on. As far as I can see, portage uses line 12 for it. That line was formerly used for CDEPEND (whatever that was), but isn't any more since 2005, so I think that no confusion can arise. <http://archives.gentoo.org/gentoo-portage-dev/msg_8f0a8505658bb99f642d68c46fa533fc.xml>
Unfortunately, the spec also doesn't say how empty ^^ ( ) groups should behave. An empty "any-of" group counts as being matched (says PMS), i.e. "|| ( )" evaluates to true. Should it be the same for the ^^ operator, i.e. should "^^ ( )" evaluate to true? How is it currently implemented in portage?
(In reply to comment #3) > An empty "any-of" group counts as being matched (says PMS), i.e. "|| ( )" > evaluates to true. > > Should it be the same for the ^^ operator, i.e. should "^^ ( )" evaluate to > true? How is it currently implemented in portage? "|| ( )" and "^^ ( )" currently behave the same in portage. I've added a test case to show this: http://git.overlays.gentoo.org/gitweb/?p=proj/portage.git;a=commit;h=8b6a531874f170d5b45926ac3ee766a775a0c11a
Created attachment 256293 [details, diff] Proposed patch for PMS First attempt on a wording for PMS. Please review.
You need to state what is and isn't allowed to be done if REQUIRED_USE isn't met. In particular, is it legal to run pkg_pretend? Also, there's a huge quality issue here: how is the package mangler supposed to tell the user *why* it's not happy with flags? REQUIRED_USE provides no way for a good explanation to be provided.
(In reply to comment #6) > Also, there's a huge quality issue here: how is the package mangler supposed > to tell the user *why* it's not happy with flags? REQUIRED_USE provides no > way for a good explanation to be provided. Portage has a function "human_readable_required_use" which would output something like the following (example REQUIRED_USE taken from ferringb's proposal): This change violates use flag constraints defined by foo-bar/baz-1.0: 'client? ( !mips? ( any-of ( gtk qt motif ) ) mips? ( only-one-of ( gtk qt motif ) ) )' I agree that this is not optimal.
(In reply to comment #6) > You need to state what is and isn't allowed to be done if REQUIRED_USE isn't > met. In particular, is it legal to run pkg_pretend? So how about this: "If the package manager encounters a package version where REQUIRED_USE assertions are not met, it must treat this package version as if it was masked. No phase functions must be called."
I'm not sure that preventing pkg_pretend calls is necessarily the best thing to do, particularly given the difficulty of giving good explanations of unmet REQUIRED_USE constraints to the user.
That's how it is currently implemented in portage, but of course it can still be discussed. @ferringb: In <http://archives.gentoo.org/gentoo-dev/msg_b0e868626019f497eba47194c34e5421.xml> you had claimed: More importantly, they're also data rather then executable code. Via it being data, up front GUI's can tell you exactly what conflicts arise, and what needs to be adjusted. Can you give a sketch of the algorithm that you had in mind?
An algorithm that gives a suggestion isn't particularly tricky. An algorithm that gives a good suggestion and that never suggests something utterly horrible is impossible. The problem is that there's nowhere near enough information available about use flags for the package mangler to know which flags it can change safely, or which flags it should consider changing first. There's absolutely nothing telling the package mangler that turning on 'minimal' or that turning off 'acl' is a bad idea, for example.
I should say, "that gives a suggestion isn't particularly tricky, so long as you either don't require it always to produce a solution, or don't require it to run in P".
(In reply to comment #9) > I'm not sure that preventing pkg_pretend calls is necessarily the best thing to > do, particularly given the difficulty of giving good explanations of unmet > REQUIRED_USE constraints to the user. Maybe add a separate function to call if REQUIRED_USE is unsatisfied, analogous to pkg_nofetch.
What an excellent idea!
Okay, we can call the new function pkg_required_use or something like that. Obviously, it will only be called if required_use is in DEFINED_PHASES, and if it's not defined then the package manager will only show a generic message about unsatisfied REQUIRED_USE.
Created attachment 256799 [details, diff] Updated patch (part 1 of 2) Updated patches: - part 1 introduces the REQUIRED_USE variable, - part 2 introduces the pkg_required_use function.
Created attachment 256801 [details, diff] Updated patch (part 2 of 2)
All- sorry for the delay in responding, been busy. I don't actually want that function- it dead end's the full goals of this functionality. The plan I had was thus: 1) get the data in there. 2) PMs grew a dumb algo (as ciaran said, not too hard), that tells you the local conflicts that the resolver hit at that point. Specifically, dump the REQUIRED_USE in full, then a list of points where current configuration doesn't match- "needed to have either sqlite or mysql on", for example. If you think through portages previous behaviour for masks, or paludises, something akin. 3) now that the basics are there, deploy it (to allow the tree to be updated), but continue work- specifically enhancing what was written in #2, so that a global optimum can be suggested. Specifically, while resolving, if a conflict is detected, stop descent at that node, flag it, and keep resolving the rest of the request. Once the rest is completed, look at the failing/stopped nodes, grab the flag potentials, and map it back against what's actually enabled for the rest of the graph. Additionally, one can do a shallow look at what the flags actually enable/enforce- beating on mysql again, if the user has mysql merged, the flag is potentially a better choice- that info can be gleaned algorithmically without too much trickery (just do subresolution at that point and check if what the flag turns on is already satisfied). The point is, things like that make it possible for the resolver to give a suggestion- it still should tell you "conflict found, flags must be choosen", but then it can suggest *which* flags are optimal. Basically, for #2, it's possible to state "you must choose mysql or sqlite." If you go and do the function route, than users instead will get told "you must choose mysql or sqlite; ebuild authors recommend sqlite". #3 however can state "you must choose mysql or sqlite. You already are enabling mysql, so we suggest that- and here is the metadata.xml defined meaning of mysql/sqlite for this package". Grok? Algorithmically, it can pull the necessary info on it's own. You go the function route, it'll wind up duplicating the basics of it (or just stating "use mysql, sqlite is the suxor"), duplicating it slower, and less completely than what the mangler could do. This was my original plan, and best I can tell, the best end user experience without requiring each/every ebuild to try and duplicate it. I'm aware #3 takes some work (I've not even started it for my own for example), so if folk think it's truly impossible to pull it off in a sane time, an interim bit of duct tape might be useful. If that route is taken, I'm concerned it will get in the way of #3 long term- wind up stopping at a local maximum basically. Counter args?
Please explain why the algorithm isn't equivalent to 3SAT, and thus impossible for non-trivial cases.
(In reply to comment #19) > Please explain why the algorithm isn't equivalent to 3SAT, and thus impossible > for non-trivial cases. The vast majority of package resolution maps back to *sat. Notice we still manage to get things done despite having that hanging over our heads. As you noted, it gets algorithmically horrid if each/every node was conflicting in REQUIRED_USE. That's not the case, and that's an *extreme* degradation- that's frankly the boogeyman, not a real case. Being that *sat hovers over our head for all resolution anyways, your point is venturing into that territory. Note I stated that the resolver doesn't have to solve it- just mark the node as failed, and if in resolving it's solution it decides that node was actually needed (everything else was resolved), look at the nodes remaining, take their REQUIRED_USE, and check the flags first level deps it enables. Final note, as I'm sure you're aware, *sat worst cases are obviously there- not all degrade to it (fortunately a fair amount don't). If you still think it's impossible, I'm putting it to *you* to disprove it- you claim it's impossible, back up the claim with real world examples that would occur in the tree.
The idea behind the proposed pkg_required_use() function is that it's not possible for the dependency resolver to read the users's mind. In some cases, there are choices that need to be made by users, and pkg_required_use() can provide important information that will help them to make these choices.
Maybe we should have the council decide on this.
I understand both Brian's point about the "danger" of leaving the explaining at maintainer's hands and Zac's / Ciaran's point about being hard (impossible?) to the PM to provide a clear and correct answer in case it's not possible to satisfy REQUIRED_USE. What I don't understand as I don't have the required background is whether Brian's proposal for the dependency resolver can be implemented or not and whether there's a noticeable impact on performance or not.
REQUIRED_USE is solvable by the package manager only in extremely trivial cases. The whole "3SAT" thing we're discussing basically means that solving non-trivial cases takes longer than the lifetime of the universe. To make matters worse, there's no way for the package mangler to know whether the solution it proposes is safe. Given something like: REQUIRED_USE="!build? ( acl )" the package mangler might decide to turn on the 'build' flag, with horrible consequences.
Created attachment 257091 [details, diff] Updated patch (part 2 of 2) Small update: We don't need an extra EAPI dependent table for the function. (This doesn't imply that I'm in favour of pkg_required_use. I can live with it either way.)
Let me also clarify that I can live without pkg_required_use too. However, if we must provide some function to be called when REQUIRED_USE is unsatisfied (as Ciaran was suggesting to call pkg_pretend in comment #9), then I think it's best to have a specialized pkg_required_use function for this case. If we use pkg_pretend for both cases, then we'll have pkg_pretend being defined and called in cases when such a function call could have been entirely avoided (due to REQUIRED_USE already being satisfied).
REQUIRED_USE *only* allows the package mangler to figure out if a certain combination is legal; it doesn't give enough information to tell the user why something's illegal, and it doesn't give enough information to try to solve the problem. Thus, why not make pkg_required_use mandatory for anything using REQUIRED_USE, and avoid ever having the package mangler have to try to provide a decent error from REQUIRED_USE?
(In reply to comment #16) > Created an attachment (id=256799) [details] > Updated patch (part 1 of 2) There were no objections on the first part of the patch, therefore pushed to master.
Nrm, regarding the naming on that patch... "Exactly one" strikes me as being clearer than "Only one". "Only one may be selected" is sort of fuzzy as to whether zero is ok or not, but "Exactly one must be selected" is much more specific.
(In reply to comment #29) > Nrm, What does this abbreviation mean? > regarding the naming on that patch... "Exactly one" strikes me as being > clearer than "Only one". > > "Only one may be selected" is sort of fuzzy as to whether zero is ok or not, > but "Exactly one must be selected" is much more specific. The specification reads: "In an only-one-of group, exactly one immediate child element must be matched." The name "only-one-of" follows portage. Do you want to call it an "exactly-one-of group" instead?
(In reply to comment #30) > > Nrm, > > What does this abbreviation mean? Apparently, that I can't type "Hrm" any more. > The specification reads: "In an only-one-of group, exactly one immediate child > element must be matched." The name "only-one-of" follows portage. > > Do you want to call it an "exactly-one-of group" instead? Yeah, I think that's clearer. Consistently saying "exactly" rather than "only" might avoid a bit of confusion.
(In reply to comment #31) > > Do you want to call it an "exactly-one-of group" instead? > > Yeah, I think that's clearer. Consistently saying "exactly" rather than "only" > might avoid a bit of confusion. Go ahead.
Created attachment 257582 [details, diff] 0001-Exactly-rather-than-Only-one-of.patch Think this gets all of them
(In reply to comment #33) > Created an attachment (id=257582) [details] > 0001-Exactly-rather-than-Only-one-of.patch > > Think this gets all of them Looks good to me.
EAPI 4 has been approved by the council.
*** Bug 238887 has been marked as a duplicate of this bug. ***