I have three packages: cmake-bootstrap, cmake and jsoncpp. cmake-bootstrap does not depend on any of the two. cmake depends on jsoncpp. jsoncpp depends on || ( cmake-bootstrap cmake ). Now if I try to 'emerge cmake', I get: [nomerge ] dev-util/cmake-3.16.2::gentoo USE="ncurses test -doc -emacs -qt5" [ebuild N ] dev-libs/jsoncpp-1.9.2-r1:0/22::gentoo USE="test -doc" 0 KiB [ebuild N ~] dev-util/cmake-3.16.2::gentoo USE="ncurses test -doc -emacs -qt5" 0 KiB Total: 2 packages (2 new), Size of downloads: 0 KiB * Error: circular dependencies: (dev-libs/jsoncpp-1.9.2-r1:0/22::gentoo, ebuild scheduled for merge) depends on (dev-util/cmake-3.16.2:0/0::gentoo, ebuild scheduled for merge) (buildtime) (dev-libs/jsoncpp-1.9.2-r1:0/22::gentoo, ebuild scheduled for merge) (buildtime_slot_op) * Note that circular dependencies can often be avoided by temporarily * disabling USE flags that trigger optional dependencies. Emerge entirely ignores cmake-bootstrap package, even though it is listed as the first preference towards satisfying jsoncpp. The same happens if I try to emerge anything needing cmake. If I explicitly type 'emerge -1v cmake cmake-bootstrap', it works just fine. The relevant ebuilds can be found at: https://github.com/gentoo/gentoo/pull/14066
The bug has been referenced in the following commit(s): https://gitweb.gentoo.org/proj/portage.git/commit/?id=4e83e35467741535463c2f9284dcf21ab4229391 commit 4e83e35467741535463c2f9284dcf21ab4229391 Author: Zac Medico <zmedico@gentoo.org> AuthorDate: 2019-12-21 22:55:34 +0000 Commit: Zac Medico <zmedico@gentoo.org> CommitDate: 2019-12-21 23:54:15 +0000 test_circular_choices: add cases for bug 703440 The testCircularJsoncppCmakeBootstrapOrDeps method fails due to circular dependencies triggered when it ignores cmake-bootstrap in order to eliminate redundant packages. Meanwhile, the testVirtualCmakeBootstrapUseConditional method solves the problem by using a dependency conditional on the bootstrap USE flag. Bug: https://bugs.gentoo.org/703440 Signed-off-by: Zac Medico <zmedico@gentoo.org> .../tests/resolver/test_circular_choices.py | 75 ++++++++++++++++++++++ 1 file changed, 75 insertions(+)
If we use a dependency conditional on the bootstrap USE flag like in the testVirtualCmakeBootstrapUseConditional method, then the emerge error message will look like this: > * Error: circular dependencies: > > (dev-libs/jsoncpp-1.9.2:0/0::test_repo, ebuild scheduled for merge) depends on > (dev-util/cmake-3.16.2:0/0::test_repo, ebuild scheduled for merge) (buildtime) > (dev-libs/jsoncpp-1.9.2:0/0::test_repo, ebuild scheduled for merge) (buildtime_slot_op) > > It might be possible to break this cycle > by applying the following change: > - dev-libs/jsoncpp-1.9.2 (Change USE: +bootstrap) > > Note that this change can be reverted, once the package has been installed.
The bug has been referenced in the following commit(s): https://gitweb.gentoo.org/proj/portage.git/commit/?id=ae3e58e83b0a3e75990fd2cef04a0c14b3162650 commit ae3e58e83b0a3e75990fd2cef04a0c14b3162650 Author: Zac Medico <zmedico@gentoo.org> AuthorDate: 2019-12-22 00:24:47 +0000 Commit: Zac Medico <zmedico@gentoo.org> CommitDate: 2019-12-22 00:26:17 +0000 testVirtualCmakeBootstrapUseConditional: split out virtual/cmake If we split the dependency conditional on the bootstrap USE flag into a virtual, then it suffices to rebuild the virtual with USE=-bootstrap instead of rebuilding jsoncpp: * Error: circular dependencies: (dev-libs/jsoncpp-1.9.2:0/0::test_repo, ebuild scheduled for merge) depends on (virtual/cmake-0:0/0::test_repo, ebuild scheduled for merge) (buildtime) (dev-util/cmake-3.16.2:0/0::test_repo, ebuild scheduled for merge) (buildtime) (dev-libs/jsoncpp-1.9.2:0/0::test_repo, ebuild scheduled for merge) (buildtime_slot_op) It might be possible to break this cycle by applying the following change: - virtual/cmake-0 (Change USE: +bootstrap) Note that this change can be reverted, once the package has been installed. Bug: https://bugs.gentoo.org/703440 Signed-off-by: Zac Medico <zmedico@gentoo.org> lib/portage/tests/resolver/test_circular_choices.py | 10 +++++++--- 1 file changed, 7 insertions(+), 3 deletions(-)
The whole point of the exercise is to establish a solution that doesn't require the user to do anything and doesn't leave his system in bootstrap state until manually reset to normal state.
(In reply to Michał Górny from comment #4) > The whole point of the exercise is to establish a solution that doesn't > require the user to do anything and doesn't leave his system in bootstrap > state until manually reset to normal state. Sure, but how is the user supposed to eliminate the cmake-bootstrap package from their system if not by toggling a USE flag?
By --depclean.
Yes, and --depclean will not eliminate cmake-bootstrap if it's guided by the "explicit || preference" that's supposed to solve the circular dependency.
If both packages are installed, preference shouldn't be relevant. It should be only relevant when no packages are installed, and emerge is trying to figure out what to do. Trying to outsmart the developer doesn't help.
So, when --depclean sees this: || ( cmake-bootstrap cmake ) How's it supposed to know which package to eliminate?
It should eliminate cmake-bootstrap because revdeps of cmake-bootstrap are a (small) subset of revdeps of cmake. In other words, it can't eliminate cmake because there are packages that depend on cmake only. It can eliminate cmake-bootstrap because all its revdeps are happy with cmake.
Ok, so --depclean is supposed to perform a statistical analysis of all the || deps and determine which package to eliminate based on that? Sounds good to me.
The delayed || deps evaluation from bug 264434 should produce the desired --depclean behavor.
In lib/portage/tests/resolver/test_merge_order.py we've got this test case that might be sensitive to a change in || preference handling for installation actions: # Test swapping of providers for a new-style virtual package, # which relies on delayed evaluation of disjunctive (virtual # and ||) deps as required to solve bug #264434. Note that # this behavior is not supported for old-style PROVIDE virtuals, # as reported in bug #339164. ResolverPlaygroundTestCase( ["media-video/libav"], success=True, mergelist = ['media-video/libav-0.7_pre20110327', '[uninstall]media-video/ffmpeg-0.7_rc1', '!media-video/ffmpeg']),
I'd be happy enough if Portage retried different || solutions upon hitting a circular dep.
(In reply to Michał Górny from comment #14) > I'd be happy enough if Portage retried different || solutions upon hitting a > circular dep. This sort of retry will be an absolute necessity, since bugs of the form "emerge installs package only to have them removed by depclean" will otherwise be introduced. This sort of behavior is desirable only when it solves a circular dependency.
The bug has been referenced in the following commit(s): https://gitweb.gentoo.org/repo/gentoo.git/commit/?id=39293f8a4667fce2112792953dbc16f69b9fcb66 commit 39293f8a4667fce2112792953dbc16f69b9fcb66 Author: Zac Medico <zmedico@gentoo.org> AuthorDate: 2019-12-24 00:52:27 +0000 Commit: Zac Medico <zmedico@gentoo.org> CommitDate: 2019-12-24 00:57:38 +0000 sys-apps/portage: Bump to version 2.3.83 #384107 adjust || preference to break dependency cycles, which solves bug 382421 and bug 703440 #703348 emerge --with-test-deps: allow circular deps Bug: https://bugs.gentoo.org/701268 Bug: https://bugs.gentoo.org/382421 Bug: https://bugs.gentoo.org/384107 Bug: https://bugs.gentoo.org/703440 Bug: https://bugs.gentoo.org/703348 Package-Manager: Portage-2.3.83, Repoman-2.3.20 Signed-off-by: Zac Medico <zmedico@gentoo.org> sys-apps/portage/Manifest | 1 + sys-apps/portage/portage-2.3.83.ebuild | 276 +++++++++++++++++++++++++++++++++ 2 files changed, 277 insertions(+)
The bug has been referenced in the following commit(s): https://gitweb.gentoo.org/proj/portage.git/commit/?id=f78a91e44e3e82008e89f05fe3871e2cb03a8646 commit f78a91e44e3e82008e89f05fe3871e2cb03a8646 Author: Zac Medico <zmedico@gentoo.org> AuthorDate: 2019-12-23 05:42:16 +0000 Commit: Zac Medico <zmedico@gentoo.org> CommitDate: 2019-12-24 00:40:07 +0000 backtracking: adjust || preference to break dependency cycles Store dependency cycle edges as backtracking parameters, and use them to adjust || preferences in order to break dependency cycles. This extends direct cycle breaking to handle indirect dependency cycles, which solves the cmake-bootstrap test case for bug 703440. If any cycle(s) remain unsolved by the next backtracking run, then backtracking aborts and the cycle(s) are reported as usual. Note that backtracking is necessary in order to avoid bugs of the form "emerge installs packages only to have them removed by depclean", since this sort of behavior is desirable only when it eliminates a dependency cycle. Bug: https://bugs.gentoo.org/382421 Bug: https://bugs.gentoo.org/384107 Bug: https://bugs.gentoo.org/703440 Signed-off-by: Zac Medico <zmedico@gentoo.org> lib/_emerge/depgraph.py | 42 ++++++++++++++++++++-- lib/_emerge/resolver/backtracking.py | 11 ++++-- lib/portage/dep/dep_check.py | 10 ++++++ .../tests/resolver/test_circular_choices.py | 25 +++++++++++++ 4 files changed, 84 insertions(+), 4 deletions(-)
Making this block the latest release (bug 706142) since the fix for bug 705986 is needed for completeness.
