Cache the compiler output of rust code with sccache, like we use ccache
for c code.
I've considered using sccache to completely replace ccache since it can
cache output of C/C++ code too. But let's not do it for now since ccache
doesn't need to run a daemon in the background that needs to be stopped
when shutting down / zapping. Also it would need more refactoring.
Reviewed-by: Luca Weiss <luca@z3ntu.xyz>
Link: https://lists.sr.ht/~postmarketos/pmbootstrap-devel/%3C20230806184729.4891-5-ollieparanoid@postmarketos.org%3E
With this code path, pmbootstrap would start a distccd + sshd in the
native chroot, and configure it so it runs the cross compiler. The
foreign arch chroots would then call this cross compiler from localhost
by calling the distcc client instead of gcc.
This code has been obsoleted by the much simpler crossdirect in 2019.
Let's finally remove it.
Fixes: issue 2179
Reviewed-by: Luca Weiss <luca@z3ntu.xyz>
Reviewed-by: Clayton Craft <clayton@craftyguy.net>
Link: https://lists.sr.ht/~postmarketos/pmbootstrap-devel/%3C20230613161437.570196-4-ollieparanoid@postmarketos.org%3E
Always set GOCACHE during build and let it point to pmbootstrap's work
dir. This has a similar effect as using ccache for C/C++.
Set GOMODCACHE conditionally: this is for caching the go modules (git
repositories) that get cloned during a build if they are not bundled
with the source. Usually APKBUILDs should cache them, but when using
pmbootstrap build --src, they would get downloaded over and over again.
Set GOMODCACHE automatically for --src and allow enabling/disabling this
manually with new --go-mod-cache / --no-go-mod-cache.
This speeds up multiple iterations of building the same package
significantly. I'm using it for:
$ pmbootstrap build postmarketos-mkinitfs --src=...
I've verified that using the same GOCACHE dir for multiple architectures
doesn't result in build artifacts for the wrong arch getting used.
Reviewed-by: Clayton Craft <clayton@craftyguy.net>
Link: https://lists.sr.ht/~postmarketos/pmbootstrap-devel/%3C20230223064743.1773-1-ollieparanoid@postmarketos.org%3E
The plan was to drop this workaround, once alpine releases a new gcc
version, and we regenerate our cross compilers against that version
_and_ drop options="!tracedeps", as latest pmbootstrap aportgen would
do. However !tracedeps wasn't dropped in the latest upgrade. This was
too easy to miss, I'll add a comment to the gcc-cross APKBUILDs so we
see it next time. Let's just update the workaround for this version.
Install isl25 for the specific gcc versions we currently have in
pmaports master that are missing it; this is easier than building new
packages and going out of sync with Alpine's versioning. This can be
reverted as soon as there is a newer gcc.
Related: pmaports issue 1732
Replace "args.cache" with a global variable in order to
avoid passing "args" to all functions. This is a step to get rid of this
args-passed-to-all-functions pattern in pmbootstrap.
Replace "args.arch_native" with the direct function call in order to
avoid passing "args" to all functions. This is a step to get rid of this
args-passed-to-all-functions pattern in pmbootstrap.
Fix the issue of having the postmarketOS binary repository key deleted
from the chroots after "abuild undeps" removes the postmarketos-keys
package. This happens if e.g. building two device packages in a row (as
they depend on postmarketos-base, which depends on postmarketos-keys in
current pmaports.git master).
I've also considered installing the postmarketos-keys next to
alpine-base in new chroots. But this would introduce a bootstrap
problem, since you can't install the postmarketos-keys package unless
it already exists in the repository. We'd run into that when building
the next release.
Before this commit, package folders were copied into the chroot one by
one in order to run apkbuild-lint on them. This logic is replaced by
mounting pmaports.git into the chroot and using a single apkbuild-lint
invocation to lint the supplied packages.
Both of these changes result in a performance improvement, especially
when linting multiple packages at once.
Before this change:
$ time ./pmbootstrap.py -q lint $(cd ../pmaports/cross; echo *) \
> /dev/null
real 0m5,261s
user 0m7,046s
sys 0m1,842s
Using the pmaports.git mount but calling apkbuild-lint in a loop:
$ time ./pmbootstrap.py -q lint $(cd ../pmaports/cross; echo *) \
> /dev/null
real 0m4,089s
user 0m6,418s
sys 0m1,219s
After this change:
$ time ./pmbootstrap.py -q lint $(cd ../pmaports/cross; echo *) \
> /dev/null
real 0m3,518s
user 0m5,968s
sys 0m0,959s
Additionally, running apkbuild-lint from the pmaports.git mount point
has the benefit that every printed violation contains a nice source
identifier à la "./cross/grub-x86/APKBUILD". This makes it possible to
differentiate between different packages even though only a single
apkbuild-lint invocation is used.
Relates: postmarketOS/pmaports#564
Let warnings like the following not get displayed in the regular
pmbootstrap output anymore, only in 'pmbootstrap log' if -v / --verbose
was used. This message informs the user that a package's dependencies
are newer than the package itself. But the WARNING makes it sound like
this is something to be concerned about, whereas in reality this is
fine. In this example, postmarketos-mkinitfs has gotten a new feature /
fix after postmarketos-base and there's no need to rebuild
postmarketos-base.
[18:02:59] WARNING: postmarketos-base depends on rebuilt package(s)
postmarketos-mkinitfs (use 'pmbootstrap build postmarketos-base --force'
if necessary!)
[skip ci] already built in CI, change is trivial
The sideload command runs the supplied names through the pmbootstrap
buildsystem to make sure they're up-to-date, then uses scp from the host
to copy the built apks to /tmp on the phone and installs them through
ssh.
If the --install-key option is set then it will also copy over the apk
key that's used for signing the packages built by pmbootstrap in case
the postmarketOS install on the device isn't build by the same machine
as you're sideloading from.
Migrate to workdir version 5 and move already built packages into the edge
channel subdir, for example:
$WORK/packages/x86_64/hello-world-1-r5.apk
to:
$WORK/packages/edge/x86_64/hello-world-1-r5.apk
The build.postmarketos.org code has already been adjusted to find built
packages in either directory structure.
The --no-depends option is supposed to stop pmbootstrap if it was
instructed to build a package, but a dependency must be built first. So
far, this only covers the case if there is no binary package for a dependency.
Make it stop if the binary package exists, but is outdated, too.
Fixes: #1895
While at it, also remove unnecessary "#!/usr/bin/env python3" in files
that only get imported, and adjust other empty/comment lines in the
beginnings of the files for consistency.
This makes files easier to read, and makes the pmbootstrap codebase more
consistent with the build.postmarketos.org codebase.
At the moment we have a simple subpkgdesc() function that can only
parse "pkgdesc" from subpackages, without support for any variables.
But we have a quite nice variable parser now that can be extended
to work for subpackages.
Simply put this works by:
- Finding the lines that belong to the subpackage function
- Stripping indentation (tab)
- Parsing relevant attributes similar to the apkbuild() function
The "subpackages" in the parsed APKBUILD are replaced by a dict
of subpkgname: {"pkgdesc": "...", "depends": "..."} which are
parsed from the subpackage function (if found).
This makes it possible to get the "depends" of a subpackage.
Install rust (rustc) in native chroot, if it is in the APKBUILD's
build dependencies. Add a test to verify that crossdirect + rust works
as expected.
Closes: #1861
Make /home/pmos/build/.git point to the .git dir from pmaports.git, with
a symlink so abuild does not fail.
abuild expects the current working directory to be a subdirectory of a
cloned git repository (e.g. main/openrc from aports.git). If git is
installed, it will try to get the last git commit from that repository,
and place it in the resulting apk (.PKGINFO) as well as use the date
from that commit as SOURCE_DATE_EPOCH (for reproducible builds).
With that symlink, we actually make it use the last git commit from
pmaports.git for SOURCE_DATE_EPOCH and have that in the resulting apk's
.PKGINFO.
Fixes: #1841
When 'pmbootstrap build' is started with '--no-depends', we make sure
that binary packages of all dependencies exist before proceeding with a
build. Make sure that we have downloaded the APKINDEX for the given arch
first, otherwise this may complain that a binary package is missing
although it does exist.
This happened when using the "native" cross compile method, e.g. when
building kernels for different architectures. It would complain that
there is no binary package for "bash" (which is in Alpine obviously):
https://builds.sr.ht/~postmarketos/job/103882
Packages can add pmb:strict to their options to enable the --strict
mode. This can be used if packages need to get build in a clean chroot
or their make dependencies need to get removed.
Depend on new pmaports version, where crossdirect uses the native ccache
binary instead of going through the foreign arch ccache first and then
going through crossdirect.
Old:
ccache (foreign) -> crossdirect (native) -> gcc (native)
New:
crossdirect (native) -> ccache (native) -> gcc (native)
Set the PATH to the crossdirect binaries, and don't set CCACHE_PATH or
CCACHE_COMPILERCHECK from pmbootstrap anymore. crossdirect sets the
CCACHE_PATH to /native/usr/bin now, along with all other required
environment variables. CCACHE_COMPILERCHECK isn't necessary anymore,
because ccache will call gcc directly and therefore be able to use the
file's timestamp and size directly. Also passing that would not work
with the current crossdirect package.
Launch native cross compilers inside foreign chroot. Enable by default,
but allow disabling with --no-crossdirect for now. This option and the
distcc-sshd related code will be removed in the future.
Aborts the build if any dependencies would have to be build first. This
is useful for build.postmarketos.org, because we want to build exactly
one package in one build job. If dependencies would need to be built, we
made a mistake earlier, and not aborting the build makes it harder to
find that orginal mistake.
When a pmaport can not be built for the desired architecture, fall back
to the binary package (from postmarketOS or Alpine) if it exists. This
allows us to provide an updated version of mesa for arm arches, but
using Alpine's mesa package on x86* arches.
Add a new action that lists all aports, for which no binary packages
exist. Only list packages that can be built for the relevant arch
(specified with --arch). This works recursively: when a package can be
built for a certain arch, but one of its dependencies
(or their depends) can not be built for that arch, then don't list it.
This action will be used for the new sr.ht based build infrastructure,
to figure out which packages need to be built ahead of time (so we can
trigger each of them as single build job). Determining the order of the
packages to be built is not determined with pmbootstrap, the serverside
code of build.postmarketos.org takes care of that.
For testing purposes, a single package can also be specified and the
action will list if it can be built for that arch with its
dependencies, and what needs to be built exactly.
Add pmb/helpers/package.py to hold functions that work on both pmaports
and (binary package) repos - in contrary to the existing
pmb/helpers/pmaports.py (see previous commit) and pmb/helpers/repo.py,
which only work with one of those.
Refactoring:
* pmb/helpers/pmaports.py: add a get_list() function, which lists all
aports and use it instead of writing the same glob loop over and over
* add pmb.helpers.pmaports.get(), which finds an APKBUILD and parses it
in one step.
* rename pmb.build._package.check_arch to ...check_arch_abort to
distinguish it from the other check_arch function
Move find_aport() and find_aport_guess_main() from pmb/build/other.py
to the new file pmb/helpers/pmaports.py.
Finding aports is not only needed when building packages, hence it
makes sense to move it out of pmb.build. The pmb/helpers/pmaports.py
file will have more pmaports related functions in a follow up commit.
Target arch is the system the package will run on. Host arch is the
system the package is compiled on.
For example kernel packages can be compiled on host arch x86_64 and
intended to run on target arch armhf. A build is necessary check
against host arch will always return True. The correct way is to check if
the package needs to be built for target arch.
Do not try to build and install dependencies for the package's
architecture when compiling in the "native" mode. That mode is
described here in detail:
https://wiki.postmarketos.org/wiki/Build_internals#Cross-compile_types
This makes it possible to cross compile kernels again, which need to
be built with GCC6. We have switched to Alpine's GCC6 package, but it is
not available for armhf/aarch64 on edge yet, because Alpine's build bots
are stuck (right now armhf is not even listed):
http://build.alpinelinux.org/
Huge thanks to ryang2478/Decatf for posting this patch in pmaports#138!
Clearly state which version is being used, and also display the message
when using build --force.
- Old:
WARNING: Package 'ubuntu-app-launch' in your aports folder has version
0_git20180604-r0, but the binary package repositories already have version
0_p20181101174257-r0! See also: <https://postmarketos.org/warning-repo2>
- New:
WARNING: package hello-world: aport version 1-r4 is lower than 1-r5 from
the binary repository. 1-r5 will be used when installing hello-world.
See also: <https://postmarketos.org/warning-repo2>
* aportgen: modify code to allow generating gcc6-armhf and other gcc6
cross compiler packages
* package: when 'gcc6' is in the depends of a package, and the cross
compiling mode is "native" (as we do it with kernels), install the
gcc6 cross compiler instead of the usual one (gcc8)
Related: pmaports#103
Overview:
Since Alpine updated to distcc 3.3 last week, pmbootstrap wasn't able to use
distcc for cross compilation anymore. It always falled back to running the
compiler in QEMU (which works, but is a lot slower). The reason for that is,
that distcc requires all compilers that are being used in a whitelist now.
This partially fixes CVE-2004-2687 in distccd, which allowed trivial remote
code execution by any process connecting to the distccd server. We only run
distccd on localhost, but still this can be used for privilege escalation of
sandboxed processes running on the host system (not part of pmbootstrap
chroots).
Because the CVE is only partially fixed (see the comment in
`pmb/chroot/distccd.py` for details), we make sure that only the building
chroots can talk to the distcc server by running distcc over ssh.
Details:
* Completely refactored `pmb/chroot/distccd.py` to run distcc over ssh
* Store the running distcc server's arguments as JSON now, not as INI
* Make debugging distcc issues easy:
* Set DISTCC_BACKOFF_PERIOD=0, so the distcc client will not ignore the
server after errors happened (this masks the original error!)
* New pmbootstrap parameters:
* `--distcc-nofallback`: avoids falling back to compiling with QEMU and not
throwing an error
* `--ccache-disable`: avoid ccache (when the compiler output is cached,
distcc does not get used)
* `--verbose` prints verbose output of the distcc too
* New test case, that uses the new pmbootstrap parameters to force
compilation through distcc, and shows the output of distcc and distccd in
verbose mode on error (as well as the log of sshd)
Building packages without git installed caused error messages from git
appearing on the screen, and I patched it upstream (see #1209).
However, this introduced a regression: when git *is* installed and the
aport folder *is not* part of a git repository (like we do it with
pmbootstrap), the build gets aborted now with git complaining that
the folder is not a valid git folder.
But the only case where this happens is, when abuild is trying to
include the git repositories' commit hash for the APKINDEX description.
This pmbootstrap commit hardcodes "postmarketOS" as APKINDEX
description, so it does not fail anymore.
## Introduction
In #1302 we noticed that `pmb.chroot.user()` does not escape commands
properly: When passing one string with spaces, it would pass them as
two strings to the chroot. The use case is passing a description with
a space inside to `newapkbuild` with `pmboostrap newapkbuild`.
This is not a security issue, as we don't pass strings from untrusted
input to this function.
## Functions for running commands in pmbootstrap
To put the rest of the description in context: We have four high level
functions that run commands:
* `pmb.helpers.run.user()`
* `pmb.helpers.run.root()`
* `pmb.chroot.root()`
* `pmb.chroot.user()`
In addition, one low level function that the others invoke:
* `pmb.helpers.run.core()`
## Flawed test case
The issue described above did not get detected for so long, because we
have a test case in place since day one, which verifies that all of the
functions above escape everything properly:
* `test/test_shell_escape.py`
So the test case ran a given command through all these functions, and
compared the result each time. However, `pmb.chroot.root()`
modified the command variable (passed by reference) and did the
escaping already, which means `pmb.chroot.user()` running directly
afterwards only returns the right output when *not* doing any escaping.
Without questioning the accuracy of the test case, I've escaped
commands and environment variables with `shlex.quote()` *before*
passing them to `pmb.chroot.user()`. In retrospective this does not
make sense at all and is reverted with this commit.
## Environment variables
By coincidence, we have only passed custom environment variables to
`pmb.chroot.user()`, never to the other high level functions. This only
worked, because we did not do any escaping and the passed line gets
executed as shell command:
```
$ MYENV=test echo test2
test 2
```
If it was properly escaped as one shell command:
```
$ 'MYENV=test echo test2'
sh: MYENV=test echo test2: not found
```
So doing that clearly doesn't work anymore. I have added a new `env`
parameter to `pmb.chroot.user()` (and to all other high level functions
for consistency), where environment variables can be passed as a
dictionary. Then the function knows what to do and we end up with
properly escaped commands and environment variables.
## Details
* Add new `env` parameter to all high level command execution functions
* New `pmb.helpers.run.flat_cmd()` function, that takes a command as
list and environment variables as dict, and creates a properly escaped
flat string from the input.
* Use that function for proper escaping in all high level exec funcs
* Don't escape commands *before* passing them to `pmb.chroot.user()`
* Describe parameters of the command execution functions
* `pmbootstrap -v` writes the exact command to the log that was
executed (in addition to the simplified form we always write down for
readability)
* `test_shell_escape.py`: verify that the command passed by reference
has not been modified, add a new test for strings with spaces, add
tests for new function `pmb.helpers.run.flat_cmd()`
* Remove obsolete commend in `pmb.chroot.distccd` about environment
variables, because we don't use any there anymore
* Add `TERM=xterm` to default environment variables in the chroot,
so running ncurses applications like `menuconfig` and `nano` works out of
the box
The `suffix` argument was not specified in chroot commands executed in
`pmb.build._package.override_source()`. Because of that, it was not
possible to use "build --src" when compiling in a non-native chroot,
for example:
```
$ pmbootstrap build hello-world --arch=armhf
...
(native) % rm /tmp/APKBUILD.append
rm: can't remove '/tmp/APKBUILD.append': No such file or directory
```
Oliver Smith
Johannes Marbach
BO41
Caio Fontes
Luca Weiss
Martijn Braam
Minecrell
Nick Reitemeyer
Oliver Smith
Zhuowei Zhang
Daniele Debernardi
Robert Yang
George Hopkins