- Explicitly stat(2) if mkdir failed. errno detection doesn't work e.g.
on Solaris (PR 45289)
- Provide a stable order for package names that only differe in the base
name, not the version number.
Problem analysis from Anthony:
On Wednesday, at 23:52, Anthony Mallet wrote:
> # rm -rf ~tmp && mkdir ~/tmp
> # pkg_add -K ~/tmp/var/db/pkg -p ~/tmp /usr/pkgsrc/packages/All/digest-20080510.tgz
> # pkg_add -U -K ~/tmp/var/db/pkg -p ~/tmp /usr/pkgsrc/packages/All/digest-20080510.tgz
> pkg_delete: Couldn't remove package directory in `/home/tho/tmp/var/db/pkg/digest-20080510'
> Also, the dir var/db/pkg/digest-20080510.xxxxxxx is never cleaned.
>
> The problem apparently is that pkg_delete finds some stalled NFS entries
> (.nfs*) in the var/db/pkg/digest-20080510 dir, so it does not delete the
> directory. Is this due to pkg_add not correctly closing file descriptors before
> exec'ing pkg_delete? For instance, I really don't understand the logic in
> check_already_installed() (add/perform.c:375) regarding the open() of +CONTENTS
> which is almost never closed (and never used as well...). Shouldn't this be
> closed before running pkg_delete?
ktrace shows that the +CONTENTS file is open() by pkg_add, I believe this is in
check_already_installed(), add/perform.c:381. Then pkg_delete is run and when
it comes to deleting the pkgdir entry, it finds that .nfs file and aborts.
Bump version to 20110805
all depending packages have their dependencies satisfied by the new
package. Essentially, split off this particular behavior as a special
case of -f, so that -f works as before, unforced works as before, and
one can give -D to override exactly this check, leaving all other
checks intact.
The -D flag is in support of make replace, as the workflow for make
replace is that inter-package dependencies are sometimes violated (but
then one must replace the depending packages, which is what
pkg_rolling-replace does via the unsafe_depends flags).
Bump PKGTOOLS_VERSION.
Add missing break statement in option parsing of "pkg_add -C", riding
the version bump.
Discussed extensively on tech-pkg@ over the summer.
OK pkgsrc-pmc@.
- Recognize -C for pkg_add like the rest of the tools do
- Do the existing check for pkg_add -f, it makes the combination of
-f and -U a bit less surprising
- Fix option handling for CHECK_VULNERABILITIES (from spz@)
- Make end-of-life check the default. pkg_install.conf(5) can be used to
override the default. The existing admin -e & co continue to work as
before.
Extract dependencies of libarchive from the pkgconfig file and thereby
drop knowledge of the needed libraries.
At least some versions of HP-UX are known to not support %zu, add a
workaround. This is using the black list approach for now.
Recognize xz as compression type for pkg_create.
The first time an error is hit while fetching packages, try to reget
from the same position. This works around the server closing the
connection while fetching dependencies.
- Try to detect common forms of pkgdb corruption and issue a warning in
that case.
- Refactor the pkg_vulnerabilities logic to use the compression support
from libarchive. This reduces the amount zlib/bzip2 interaction to the
linkage.
- Restore PKG_PREFIX in pkg_delete (PR 42731)
- Ensure that the current pkg_install version is at least as new as
the version used to build the package
- Refactor man pages:
- Move description of environmental variables into pkg_install.conf(5)
- Remove a bunch of dead references
- Avoid using hard-coded path names that can easily be overridden by
the user.
- Refactor pkgdb related logic, so that -K behaves consistently.
- Add some helper functions to isolate more places from the pkgdb
layout.
- Sync license list.
- Handle PKG_REFCOUNT_DBDIR consistently as option too.
- Utilize connection cache from libfetch.
- Add helper target for cat pages updates.
Do not overwrite a string with itself using snprintf. This breaks
setting the pkgdb directory internally on Linux. Explicitly check
if the string is the same and otherwise just use xstrdup.
- Cleanup configure.ac to not check for headers and functions not used
- Always use nbcompat/db.h on !NetBSD (PR 41865)
- Don't bother setting termwidth in pkg_info, it is never used.
Add a new command for pkg_admin: findbest. It takes one or more patterns
and searches for the best match in PKG_PATH, just like pkg_add would.
It prints the URLs of the best match for each pattern to stdout.
Rewrite the config file parser to read the file only once.
Make pkg_install WARNS=4 clean and fix a number of lint warnings.
Based on reports from veego about warnings with older GCC releases in
pkg_delete, where a variable is potentially used uninitialized.
- license handling: accept upper case letters. Keep license checks
case-sensitive as done in the older pkgsrc logic. Document this.
OK dillo@, schmonz@, wiz@
- pkg_info:
- fix handling of non-packages, that are valid archives
- invert order of pkg_info -r to better match the expectations of
make update.
Try harder to mimic the old pkg_delete behavior and provide a
topologically sorted output for pkg_info -r. Bail out on dependency
cycles based on recursion depth, currently 64K.
These will be available from pkg_summary(5) database.
To be used to find previous packages when a package is moved
or renamed. (Even though that is not recommended.)
Documented in pkgsrc Guide and in the pkg_summary(5) manual page.
Upcoming commit will have some packages modified to add them.
Add logic to evaluate license conditions. As frontend for pkgsrc,
pkg_admin gets two new commands (check-license and
check-single-license), which can be used to evaluate a given condition.
pkg_add will be changed to honour licenses at a later point.
Make pkg_delete -d the default behavior, remove the option.
Remove pkg_create -R support (reorder @dirrm entries).
Introduce new plist command @pkgdir, which makes pkg_add add
this directory if it doesn't exist already and pkg_delete remove it
only if no other package has a @pkgdir entry for it. Otherwise
directories are pruned when the last file or directory in them is
removed. @dirrm is now a pure hint, if the directory doesn't exist, it
will be silently skipped.
Consider EOF during signature scan a fatal error. At the very least, the
package is missing the +CONTENTS file at that point. Correctly reset the
entry pointer in that case and make the meta data extraction stricter,
avoiding NULL dereferences for invalid archives.
Fixes pkg_info -X crash on empty files as reported by Daniel Horecki.
Simplify archive handling by depending on archive_read_finish and the
close callback where needed. Fixes a file descriptor leak as side
effect as reported by wiz.
Integrate the PKG_PATH logic for tightly with the find_archive logic:
- remember initial current working directory from the time PKG_PATH is
processed, it will be used as reference for all relative entries
- remove now redundant fchdir dance in pkg_add
- pass down to find_archive if this is a top-level package (e.g.
requested on the command line) or not; the location of top-level
packages is searched for packages first and for URLs or path names the
PKG_PATH itself is not processed (e.g. pkg_add foo/bar not look into
PKG_PATH for bar). This addresses PR 33884 in a different way.
- document the rules for finding packages more explicitly
- allow specifying PKG_PATH in pkg_install.conf as well; environment
takes precendence.
- remove PKG_PATH related logic in pkg_info(1), it doesn't deal with
PKG_PATH anyway
- remove PKG_PATH, PKG_TMPDIR and TMPDIR description in pkg_info(1),
they are no longer used
