In the old days licenses were typically packaged with "%doc COPYING"
which when installed with --excludedocs would in many cases violate
the license. Using "%license COPYING" avoids the most common pitfall
but %license supports absolute paths too in which case those files
can have other special attributes attached too. Such as documentation
or language specificness, which must not cause exclusion for licenses.
And all sorts of funny corner cases...
On net shared paths we could not install anything anyway. Wrong
colored files would conflict. The common theme between these is that
the file is expected to be there via some other means, which makes
it ok to skip them. This covers --excludedocs, install langs etc,
ie the most obvious cases. Not covered here is skipping due
--excludepath which should not be allowed either (if a path is
inaccessible you can always relocate instead), leave a FIXME
reminder as it requires a bit more work whereas this is trival
to backport.
Currently, there's no callback type that would be issued per each
transaction element. RPMCALLBACK_TRANS_PROGRESS is only issued during
the prepare phase but not when packages are actually installed or
erased. Likewise, RPMCALLBACK_INST_ST* and RPMCALLBACK_UNINST_ST* won't
be issued if an install or erase operation is skipped for some reason (a
script or package upgrade failure).
Having such a callback would allow the Python API consumers to always
know upfront which element is about to be processed, before any other
callbacks are issued. This is important since not every callback type
carries enough data about the subject package; while the INST types
provide the user object passed to a former addInstall call, the UNINST
types only provide the package name (which may not be unique within the
transaction set).
This commit adds such a callback.
Now options --notriggers, --notriggerin, --notriggerun, --notriggerpostun
disable also file triggers and not only normal trigges. Transaction file
triggers can be also disabled with --nopretrans and --nopostrans options.
Before this patch file triggers weren't disabled when option --test was
specified. That caused unwanted multiple execution of file triggers
in dnf because dnf executes transaction with flag test before executing
transaction for real (rhbz:1282115).
File triggers defined with priority greater than or equal to 100000
will be executed before standard scriptlests. The other file triggers
will be executed after standard scriptlets.
Previously directory was skipped from installation if it contained only
skipped files. But it wasn't skipped if it contained some subdirectory.
Now if subdirectory contains also only skipped files then subdirectory
and also parent directory are skipped. It is achieved by solving
subdirectories at first and then solving parent directories.
(rhbz#1192625)
These warnings were caused by my recent file trigger commits. These
warnings are showed only with compiler option "-O2" and I used "-O0"
that's why I didn't notice them at first.
- After transaction this hash contains all packages installed in
transaction. This hash can be used as argument of rpmdbFilterIterator()
to get all packages installed in transaction from rpmdb.
- Rename removedHash to packageHash.
- If we process some file and we find another file with the same path
and the same content and this other file is skipped for color then
the currently being processed file has to be skipped for color too.
(RhBug:1170124)
- Always knew we'd need a plugin disabler flag sooner than later but
didn't realize enabled plugins would fail basically the entire
test-suite :)
- Enable --noplugins for entire test-suite for now, but eventually
we'll need to come up with ways to test plugins as well
- %__transaction_plugins style configuration is problematic for plugins
because we want plugins to be, well, pluggable. As in drop-in to
enable, which is not achievable with a single macro entry. Look up
all DSO's from the plugin dir and enable if a matching
%__transaction_foo macro is defined.
- This isn't optimal but it'll buy us the drop-in capability, which
is what matters most right now. We'll want to have forcability as
well later on (ie it should be possible to require given plugins
to be present)
- The basic concept is not without merit but what was implemented here
has been stuck in experimental state in middle of two sorta conflicting
goals for four years now, and world has moved onward in the meanwhile.
The sepolicy part is better handled in the new selinux plugin, and other
action business belongs to packages (in the form of some trigger-like
scripts or such) rather than rpm plugins.
- Deleted here, but the sepolicy plugin functionality still needs
merging into the new selinux plugin...
- RPMTAG_COLLECTIONS left in place but tagged unimplemented as per policy
to never actually remove tags
- When the world changes underneath us (eg by %pretrans hackery)
between test-transaction and actual transaction, in some cases
the file action does not get recomputed if already set.
Always reset file actions to FA_UNKNOWN before disposition
calculations to ensure have a clean starting point.
- We have FA_SAVE which results in .rpmsave suffix so use it, instead
of mapping FA_BACKUP to .rpmorig or .rpmsave depending on whether
we're installing or erasing. Doh.
- Goal is no longer needed by fsmMapPath()
- Account for packages removed in transaction when checking whether
package is already installed. This permits install + erasure of the
same package within a single transaction, aka reinstall, with
changing file policies (excludedocs, installangs etc) properly
taken into account. Related to RhBug:966715.
- This differs from --replacepkgs which is just a strange hack really
but one that's difficult to change because of how its presented
in the API. --replacepkgs behavior does not change here, and a proper
reinstall is NOT possible when replacepkg problem filter is active.
- Testing this is not possible with rpm cli :(
- Nothing outside rpmte.c should be able to muck around with the
file info set associated with the ts element. Except to free
memory, which is the only thing rpmteSetFI() is now being used for.
- Grab files and file states from the main loop and pass as arguments
to skipInstallFiles() and skipEraseFiles(), neither actually
needs the transaction element otherwise.
- Use local iterators in both skip*Files(), not that it actually
matters here other than eliminating rpmteFI() uses
- All the called functions expect rpmfiles anyway so might as well
grab that from the start. We could probably do some of this with
iterators now but as everything here is done with the random-access
functions, switching would take quite a bit of work...
- These all operate on rpmfiles, not rpmfi, now so make the point
clearer. All internal stuff so we're free to mess around.
- No functional changes, only a straightforward perl-assisted rename...
- The self-iterator in rpmfi prevents all sorts of sane uses of
file set iteration. Split the actual data into a separate data
type, changing the internal random-access functions to use the
new rpmfiles datatype directly and update internal callers minimally.
This should be entirely transparent to public API consumers who still
only see the braindamaged self-iterating version.
- Internal consumers dont directly benefit (yet), this is just an
early step on the road to rpmfi sanity. Much of the API and variables
will need renaming etc but avoiding that here to keep the changes
to minimum, this is a rather huge commit as it is.
- The DB backend is supposed to take care of syncing when necessary,
syncing all indexes "just in case" whether they were written to
doesn't make a whole lot of sense.
- Previously %pretrans and %posttrans were tied to --nopre and --nopost
disablers (since commit 0b2d7775c5)
because back then, there was no room new disablers in rpmtransFlags
bitfield. This is no longer the case as of rpm >= 4.9.x where
a bunch of obsolete flags were axed, so we can now add specific
--nopretrans and --noposttrans switches + corresponding flags.
- This is obviously a behavior change as --nopre and --nopost no
longer affect %pretrans and %posttrans, but --noscripts behavior
remains the same.
- Lazily initialize plugins when adding install or erase elements
to the transaction already. Since this ensures the setup will
always be performed, there's no need to do this from rpmtsRun()
anymore, empty transaction would not get that far anyway.
- This will allow adding hooks to the places where package header
is still available (which is required for various things) and
in general, before transaction is actually executed.
- The downside here is that adding or removing packages to a transaction
can now fail due to plugin initialization failure where its not
easily distinguishable from other errors, but then they can also
fail due to rpmdb open failure which kinda similar "shouldn't be
here" issue. The alternative would be initializing the plugins
in rpmtsCreate() but as that's not actually permitted to fail at
all, its not an option really.
- We hereby unceremoniously declare that from now on, SELinux and
other similar mechanisms are to live in plugins, rpm doesn't need
to know about every possible security etc mechanism there might be.
Its a big commit but as its really just removals...
- We need to disable sepolicy collection plugin for now as it relies on
built-in support for selinux, this not only makes no sense but
wont work now that there is no built-in support. The sepolicy
plugin needs to be merged into the selinux base plugin now.
- Another "breakage" is that --rebuilddb no longer relabels the database,
this needs plugins to called in a place where they currently cannot
be called. This needs to be resolved before next actual release.
- For now this doesn't do anything particularly interesting, it
just replaces the transaction lock handle we use internally.
We want to export this interface but that needs further changes.
- Explicitly test for shared file/dir being ghost when deciding whether
to skip create: we must never skip the creation of a real file/dir
because of a ghost as the ghosts do not get created.
- This fixes the shared ghost testcase, ie the thing that commit
a712252392 was supposed to fix,
without side-effects that I can see.
- Reverts the bad half of commit a712252392eca75443ca45c10a72873cabec7963:
treating files that wont be installed as skipped seems logical enough,
but rpm disagrees... making %ghosts FA_SKIP on install causes
upgrades on %ghost %config files to miss the necessary FA_SKIP getting
set on the erase half via handleInstInstalledFile() and
rpmfiDecideFate(), causing %ghost %config files to be nuked. Which isn't
very nice when your /var/lib/rpm/ contents are packaged as
%ghost %config...
- This fixes the %ghost %config testcase but breaks the shared %ghost
case in turn.
- Only skip shared file/dir creation if its actually being created
by another package. Previously we could've decided to skip an entry
where the other file was also being skipped for some other reason.
- Ensure %ghost entries always have FA_SKIP on install, previously
they often were FA_CREATE which makes no sense for %ghost. We dont
encounter %ghosts at all during install in the FSM as they dont
exist in the payload, but the file disposition calculations need
to take them into account now that we're avoiding redundant creates.
- Fixes a regression from commit f7f5f88f9f
which unearthed the %ghost FA_CREATE issue, reported here:
http://lists.fedoraproject.org/pipermail/buildsys/2013-January/004047.html
- Previously the pool would only get allocated on successfull addition
of install or erasure elements, causing assert() failures on
operations on empty transaction set, which should be just a no-op,
not an error.
- Make rpmtsPool() create the pool if it doesn't exist, update relevant
users to call rpmtsPool() instead of directly accessing tsmem->pool,
this avoids having to worry about pool existence in all the various cases.
- Also fix up the pool-related comment on rpmtsEmpty(): pools does not
and can not support emptying as it could break references to its
contents. Per-string refcount would be needed for emptying support.
-All post hooks take an additional rc argument that indicates general return code from operation.
-All post hooks are always called if correspoding pre hook has been called.
-The return value from post hooks is curently ignored
Signed-off-by: Panu Matilainen <pmatilai@redhat.com>
- Since commit f7f5f88f9f, only the first
instance of shared files is created, the rest are FA_SKIP which
consume no disk space and are not counted either.
- When updating packages, we first create them with a temporary names
and only after all files from payload have been created this way,
the files are renamed to the final target. This means that performing
an update temporarily requires roughly twice the disk space (and inodes)
compared to the final result on per-package level. Which matters
when space is tight, such as presumably in RhBug:872314.
- Simulate what happens on upgrades by adding block and inode delta
to the equation: installing a file always consumes an inode and
the specified amount of disk space. But when replacing files,
reduce the size-delta from disk consumption *after* checking for
problems in a given DSI.
- Also fixes inode accounting which has been broken for forever (since
commit a9a1fd866c more exactly)
Panu Matilainen