Changes since 20.2.0:
Don't warn about a possibly redundant PKGNAME=${DISTNAME} assignment if
PKGNAME is defined somewhere else in the package Makefile.
Warn if NO_CONFIGURE=yes and REPLACE_* are combined.
Suggest to replace ${VAR:@l@-l${l}@} with the simpler ${VAR:S,^,-l,},
as well as ${VAR:@l@${l}suffix@} with the simpler ${VAR:=suffix}.
Allow lua in CATEGORIES.
Includes fixes to +REQUIRED_BY generation, performance improvements,
build fixes against newer libnetpgpverify, and better error messages.
When combined with newer pkgin releases, this should now eliminate the
various "pkg_add: Can't open +CONTENTS of depending package ..." errors
that users had frequently observed during upgrades (joyent/pkgsrc#158,
joyent/pkgsrc#190, joyent/pkgsrc#256, and many IRC logs).
In the pkg_admin front end, instead of adding +REQUIRED_BY entries as they
are found, which previously led to duplicate entries, cache the results and
write out the files at the end.
Underneath, add a caching version of iterate_pkg_db() that avoids the same
pkgdb directory lookup for every installed package, but is only suitable for
reads. Also add a cache for best_match lookups to avoid expensive matches
each time.
For all caches, use a simple hashing function to improve lookup performance.
In summary, as well as fixing +REQUIRED_BY files, these patches reduce the
wall/user/system time of "pkg_admin rebuild-tree" on a system with 12,762
packages installed down from 13m52s/11m20s/2m32s to just 1m4s/1m3s/0m1s.
The list of dependencies held by packages during recursive upgrades was not
refreshed after dependencies were themselves upgraded, leading to failures
attempting to read +REQUIRED_BY files in package directories that no longer
exist ("registration is incomplete!"). We now only perform the package
match after the upgrades have completed.
While here, hide the warning about dependencies not being fulfilled behind
ForceDepending, as the whole point of using that mode is to ignore such
issues with the assumption that the final state after updating will be
correct.
Changes since 20.1.17:
Fixed the algorithm for checking whether two patterns overlap, such as
MACHINE_PLATFORM:MNetBSD-[0-9].*-*.
Fixed wrong warning about foreign variable in SUBST block, as seen in
geography/qgis.
Changes since 20.1.15:
When a package adds an additional version requirement for another
package, it must do so using BUILDLINK_API_DEPENDS instead of
BUILDLINK_ABI_DEPENDS. Most packages already do this.
When a values is appended to an undefined variable using the += operator,
bmake does not add a space before, and pkglint caught up to do the same.
This change has no practical consequences though.
As always, a bit of refactoring. The method names of MkAssignChecker
contained the redundant word "varassign".
Changes since 20.1.14:
Fix confusing wording of diagnostic "should contain text". It's more
precise to say "this line should consist of this text". Otherwise it's
too easy to interpret it as "the text should occur somewhere in the
line".
Allow BUILDLINK_PREFIX.* to be used in helper files. Especially for
programming language packages, files like plugin.mk often include
buildlink3.mk and therefore may refer to the BUILDLINK_PREFIX of this
package.
Warn about redundant BUILDLINK_API_VERSION restrictions. Previously, the
check had only been enabled if the operators from the default dependency
pattern (buildlink3.mk) and the additional dependency pattern (package)
used the same operators, which already covered most practical cases.
Determine SuSE versions from the pkgsrc tree.
Changes since 20.1.13:
Packages that don't define DISTNAME probably don't download any files
and thus may not need a distinfo file. (There are several other
conditions involved in this, though.)
When reporting wrong distinfo hashes, always report them in the order in
which they appear in the distinfo file, not by hashmap order.
Fix panic when parsing the Makefile line "./=value", which according to
bmake is a variable assignment. This is not used in practice though.
Disallow a leading hyphen in package option names. There are only very
few packages that wrongly use these option selectors in
PKG_SUGGESTED_OPTIONS.
Distinguish between a tool dependency (USE_TOOLS) and a plain tool name
(TOOLS_NOOP, TOOLS_BROKEN, TOOLS_FAIL). Allow packages to add arbitrary
tools to these lists.
Changes since 20.1.11:
The file bsd.pkg.mk must only ever be included by package Makefiles
directly, not by other Makefile fragments. Seen in www/w3m.
The variable BUILDLINK_PREFIX.* should only be used for packages that
have actually been included by the package. This catches the use of
BUILDLINK_PREFIX.libiconv, which should have been iconv instead.
Allow comments before line 3 in buildlink3.mk files. This is necessary
for mariadb55-client since its buildlink identifier is mysql-client,
which is so non-obvious that it needs to be documented.
Changes since 20.1.10:
PKG_SYSCONFDIR and VARBASE must not appear in INSTALLATION_DIRS.
Patch files in which the line number have been edited manually are marked
with notes.
This is quite a disruptive change, but has been tested in various ways across
Darwin, Minix, and SmartOS, so I'm reasonably confident I caught most issues.
With the switch to automake the pkgsrc Makefile has been able to be simplified
a lot, and I also took the opportunity to remove the mostly useless MESSAGE
file.
## Version 20.5.1 (2020-05-25)
* Improve make portability.
## Version 20.5.0 (2020-05-25)
* Switch to automake and overhaul mandatory configure options.
* Remove unused code and pointless configure tests, cleaning up includes
and headers where appropriate.
* New version number scheme (yy.m.x). The previous scheme had no relevance,
this one at least gives users some clue about how up-to-date their version
is.
One change: use AC_PROG_CC_C99 and regen configure
This helps bootstrapping on legacy systems where the compiler doesn't
default to c99 without additional guidance.
Patch from ky0ko@disroot.org, tested on NetBSD and Alpine Linux by myself
and on IRIX with mipspro by ky0ko.
Changes since 20.1.8:
Recognize SUBST_NOOP_OK and SUBST_SHOW_DIFF.
Allow plain -Wl,-R and -Wl,-rpath in BUILDLINK_TRANSFORM, when they are
used to remove compiler options.
Recognize ggrep and other platform tools as valid tool names. Seen in
lang/rust.
Changes since 20.1.7:
There are about 300 cases where a package defines a PLIST conditional in
PLIST_VARS but none of its PLIST files actually uses that condition.
These cases get a warning.
In a pkgsrc setup with default paths, the simple substitutions are all
no-ops. The last two substitutions only apply to Makefile.in, which
means that they are no-ops for the other files. Since the whole sed
command is not considered an identity substitution, an explicit
SUBST_NOOP_OK is needed here.
https://mail-index.netbsd.org/pkgsrc-users/2020/05/17/msg031245.html
Changes since 20.1.5:
Category Makefiles must only list subdirectories in SUBDIR that actually
contain a package. There is no need to mention the other directories.
This is the same as in the top-level Makefile, where mk/ and regress/ are
not listed.
Packages from pkgsrc-wip may have a COMMIT_MSG file with a suggested
commit message for importing the package into main pkgsrc.
This allows to show the actual changes by setting SUBST_SHOW_DIFF=yes,
and with SUBST_NOOP_OK=no it also demonstrates that no substitutions are
needed for substplistbasedirs.
Another thing this demonstrates is this wrong patch:
-echo >>$sedrules "s|@@PKGVERSION@@|@PKGVERSION@|g"
+echo >>$sedrules "s|@3.17@|3.17|g"
Changes since 20.1.4:
No more wrong warnings about the Solaris /bin/sh. These warnings had been
there for 14 years, preventing pkgsrc developers from using the $$(...)
command substitution and negation in shell conditions.
https://mail-index.netbsd.org/pkgsrc-changes/2020/05/01/msg212194.html
Changes since 20.1.3:
For patches that patch a single file, the filename of the patch should
correspond to the patched file. There are a few different naming schemes
in action, therefore the check is relatively loose. Patches that are
called patch-[a-z][a-z] continue to be allowed for historic reasons.
Patches that are called patch-CVE-* are also allowed.
The entries in doc/CHANGES-* are checked for consistency. For example,
it doesn't make sense to add a package twice or "update" a package from
version 1.0 to version 1.0. All version numbers in these entries must
be valid pkgsrc versions, i.e. start with a digit and only use
characters from -.0-9A-Z_a-z.
Changes since 20.1.2:
Stricter check for Python version numbers. Before, 25 and 26 had not
been marked as wrong.
In assignments like PKGNAME=${DISTNAME:S,from,to,}, modifiers that don't
have any effect generate a note. Most of these modifiers are redundant
or outdated.
Patches must not add well-known absolute paths like /usr/pkg, /var and
/etc since these must be overridable by the pkgsrc user. Other absolute
paths continue to be allowed.
The "jperkin should have run the test suite before releasing 0.16.0" release.
## Version 0.16.1 (2020-04-29)
* Don't attempt to update the keep list if we're only downloading packages.
* Avoid C99isms for now, fixes CentOS 6 build.
* Ignore "pkg_admin rebuild-tree" output, removes spurious "Done." messages
from output.
* Some minor internal refactoring.
PKGPATH is one of the few variables that is guaranteed to be defined when
mk.conf is included. Conditions like ".if ${PKGPATH} == category/package"
trigger a "Malformed conditional" when this variable is not defined.
## Version 0.16.0 (2020-04-27)
* Rebuild +REQUIRED_BY files after each pkgdb modification, ensuring
consistency and avoiding "Can't open +CONTENTS of ..." errors during
subsequent upgrades.
* Improved output sorting.
* Turn off WARNS.
* Minor cleanups.
* Documentation fixes.
* Flush yes/no prompt before reading response (Michael Forney).
