We already attest to the relevant dist-archive in inputs.SHA256SUMS,
which is recorded at build-time.
We use a SKIPATTEST.TAG file to indicate output directories which do not
require attestation (much like the CACHEDIR.TAG specification).
Generally, it's better to have build scripts declare properties of
directories instead of introducing name-based special cases in attest
scripts since build scripts have a more detailed context of what is
going on.
At build/codesigning-time, hash build inputs and output the digest to
${OUTDIR}/inputs.SHA256SUMS, which gets included in the final SHA256SUMS
constructed by guix-attest.
Example final SHA256SUMS:
ee832d2a35b7701bff581dea05a536118b118e3ad0a587a2855b6ee8cd6fba20 inputs/bitcoin-78199266af7b.tar.gz
ca765e70a0c12866dd63c0be228b675278a26329e5f8f5b5c52fd09200fedf21 bitcoin-78199266af7b-powerpc64le-linux-gnu-debug.tar.gz
dae95327d7f2c324e2728c4b73627be6cb2c0d2f2e5bea940d1d5e6463939327 bitcoin-78199266af7b-powerpc64le-linux-gnu.tar.gz
While files are being output to $OUTDIR, it will be under
${DISTSRC}/output, and only when everything is done, will
${DISTSRC}/output be moved to the actual $OUTDIR.
This makes it so that a Ctrl-C in the middle of a build is less likely
to result in a partially-constructed $OUTDIR. In fact, if I understand
correctly, if $OUTDIR and $DISTSRC reside on the same filesystem, the
move (rename) is likely atomic.
Also, since the "working $OUTDIR" is under ${DISTSRC}/output, it will be
cleaned properly by the guix-clean script.
c799a19b4b build, qt: No longer need to set QT_RCC_TEST=1 for determinism (Hennadii Stepanov)
Pull request description:
The Qt Resource Compiler (rcc) output order relies on [`QHash`](https://doc.qt.io/qt-5/qhash.html):
> This randomization of `QHash` is enabled by default. Even though programs should never depend on a particular `QHash` ordering, there may be situations where you temporarily need deterministic behavior, for example for debugging or regression testing. To disable the randomization, define the environment variable `QT_HASH_SEED` to have the value 0.
Since #3620 we use `QT_RCC_TEST=1` to achieve a deterministic output.
Since Qt 5.3.1 hash seeding is disabled for all of the bootstrapped tools, including rcc. Therefore, `QT_RCC_TEST=1` is no longer needed.
See commit [5283a6c87beac5a43f612786fefd6e43f2c70bf6](5283a6c87b).
ACKs for top commit:
fanquake:
ACK c799a19b4b
Tree-SHA512: 9d116ac1e8c605ee3e8ed7f618586f0de85d8b06bbbb70fe8c298939ce203d2a7e97264a9afac037179993ab54c5f69a65ebb9ab27ca7f45acb963011bd45743
Passing ADDITIONAL_GUIX_COMMON_FLAGS="--no-substitutes --bootstrap" as
suggested doesn't work:
```bash
...outputting in: '/bitcoin/guix-build-a1f0b8b62eb8/output/x86_64-linux-gnu'
...bind-mounted in container to: '/outdir-base/x86_64-linux-gnu'
guix time-machine: error: bootstrap: unrecognized option
```
and I think bootstrapping is more than covered in the preceding "Choose
your security model" section.
867a5e172a guix: Register garbage collector root for containers (Carl Dong)
8f8b96fb54 guix: Update hint messages to mention guix-clean (Carl Dong)
44f6d4f56b guix: Record precious directories and add guix-clean (Carl Dong)
84912d4b24 build: Remove spaces from variable-printing rules (Carl Dong)
Pull request description:
```
guix: Record precious directories and add guix-clean
Many users have reported problems that stem from having an unclean
working tree. To that end, I've written a guix-clean script which should
help reset the working tree while respecting user-specified precious
directories.
Precious directories, such as:
- SOURCES_PATH
- BASE_CACHE
- SDK_PATH
- OUTDIR
Should be preserved when cleaning the working tree, and are thus
recorded in ./contrib/guix/var/precious_dirs.
The ./contrib/guix/guix-clean script is able to parse that file and make
sure to avoid them when cleaning out the working tree.
```
ACKs for top commit:
laanwj:
ACK 867a5e172a
Tree-SHA512: c498fad781ff5e6406639df2b91b687fc528273fdf266bcdba8f6eec3b3b37ecce544b6da0252f0b9c6717f9d88e844e4c7b72d1877bdbabfc6871ddd0172af5
By registering the container profiles as garbage collector roots, it
will prevent `guix gc` from garbage collecting derivations which our
container needs and inconvieniencing the user with a rebuild.
Many users have reported problems that stem from having an unclean
working tree. To that end, I've written a guix-clean script which should
help reset the working tree while respecting user-specified precious
directories.
Precious directories, such as:
- SOURCES_PATH
- BASE_CACHE
- SDK_PATH
- OUTDIR
Should be preserved when cleaning the working tree, and are thus
recorded in ./contrib/guix/var/precious_dirs.
The ./contrib/guix/guix-clean script is able to parse that file and make
sure to avoid them when cleaning out the working tree.
This relatively easy change eliminates all runtime dependencies (except
for the kernel) for dmg, which is the only native build tool that gets
put in our output tarballs.
This allows much more flexibility when constructing the codesigning
environment, and is much more robust.
./windeploy is a "working directory", and therefore belongs inside
distsrc-*. Many people have noticed their Guix builds failing after
hours simply because they did not remove windeploy (but did remove the
distsrc-* directories).
In Guix, there are two flags for controlling parallelism:
Note: When I say "derivation," think "package"
--cores=n
- controls the number of CPU cores to build each derivation. This is
the value passed to `make`'s `--jobs=` flag.
- defaults to 0: as many cores as is available
--max-jobs=n
- controls how many derivations can be built in parallel
- defaults to 1
Therefore, if set --max-jobs=$MAX_JOBS and don't set --cores, Guix could
theoretically spin up $MAX_JOBS * $(nproc) number of threads, and that's
no good.
So we could either default to --cores=1, --max-jobs=$MAX_JOBS
- Pro: --cores=1 means that `make` will be invoked with `-j1`,
avoiding problems with package whose build systems and test
suites break when running multi-threaded.
- Con: There will be times when only 1 or 2 derivations can be built
at a time, because the rest of the dependency graph all depend
on those 1 or 2 derivations. During these times, the machine
will be severely under-utilized.
or --cores=$MAX_JOBS, --max-jobs=1
- Pro: We don't encounter prolonged periods of
severe under-utilization mentioned above.
- Con: Many packages' build systems and test suites break when running
multi-threaded.
or --cores=1, --max-jobs=1 and let the user override with
$ADDITIONAL_GUIX_COMMON_FLAGS
fbbb2d4fc1 lint: Fix spelling errors in comments (fyquah)
Pull request description:
Found some spelling errors while running spelling linter https://github.com/bitcoin/bitcoin/pull/21245
This PR fixes them.
ACKs for top commit:
fanquake:
ACK fbbb2d4fc1 - I thought we just fixed all of these.
Tree-SHA512: 95525040001f94e899b778c616cb66ebafb679dff88835b66fccf6349d8eb942d6b7374c536a44e393f13156bce9a32ed57e6a82bb02074d2b3cddb2696addb2
c33b199456 guix: Bump glibc and linux-headers (Carl Dong)
65363a1bd8 guix: Rebase on 95aca2991b (1.2.0-12.dffc918) (Carl Dong)
Pull request description:
On bumping the time-machine:
```
A few changes which are useful for us:
1. 'gnu: cross-gcc-arguments: Enable 128 bit long double for POWER9.' is
now merged into master.
2. gnutls is bumped to 3.6.15 and the temporal test failure in
status-request-revoked is fixed. Note that this does not fix the case
where one has installed Guix v1.2.0 and is running a substitute-less
bootstrap build, since the `guix time-machine` command itself has a
dependency on gnutls v3.6.12 (the one with the broken test) and will
thus try to build it before attempting to jump forwards in time. This
does however, mean that those who build a version of Guix that also
contains this fix will not go backwards in time to build the broken
gnutls v3.6.12.
```
On bumping the rest:
```
Bump glibc and linux-headers to match those of our Gitian counterparts.
We also require a glibc >= 2.28 for the test-symbol-check scripts to
work properly.
The default BASE-GCC-FOR-LIBC also has to be bumped since glibc 2.31
requires a gcc >= 6.2
```
This is a prerequisite for #20980
ACKs for top commit:
fanquake:
ACK c33b199456 - I think going ahead with this now and to sycn back up to gitian is fine. It will also unblock #20980. Potential code signing related issues can be sorted out in #21239 and later PRs.
Tree-SHA512: 31f022aadb93ba44813b0da005b1f2e5d67d76e8cdcdb53368924d1ea6cb076a21218c26831a6b0dcdcfe33507f54934330489ba557371d740f5587b7d727b95
a0a7a4337d guix, doc: Update default HOSTS value (Hennadii Stepanov)
Pull request description:
This is a #21089 follow up.
ACKs for top commit:
fanquake:
ACK a0a7a4337d
Tree-SHA512: c1813cc2b9212a79fd34d4e25cd0816b58264e1890daf777cd59411bd20fcc9affe312871d06fab1308b8f55c1a78ac1101e631882c18360a4709ecef4529f05
13a9fd11a5 guix: Passthrough SDK_PATH into container (Carl Dong)
Pull request description:
This is a usability improvement for Guix builders so that they don't have to extract the Xcode tarball into `depends/SDKs` every time.
Inspiration: https://github.com/bitcoin/bitcoin/pull/21089#issuecomment-778639698
ACKs for top commit:
laanwj:
Tested ACK 13a9fd11a5
Tree-SHA512: 63392d537e48a0da9f0ee04a929613b139bef1ac5643187871c9ea5376afd2a3d95df0f5e0950ae0eccd2813b166667be98401e5a248ae9c187fe4e84e54d427
Bump glibc and linux-headers to match those of our Gitian counterparts.
We also require a glibc >= 2.28 for the test-symbol-check scripts to
work properly.
The default BASE-GCC-FOR-LIBC also has to be bumped since glibc 2.31
requires a gcc >= 6.2
A few changes which are useful for us:
1. 'gnu: cross-gcc-arguments: Enable 128 bit long double for POWER9.' is
now merged into master.
2. gnutls is bumped to 3.6.15 and the temporal test failure in
status-request-revoked is fixed. Note that this does not fix the case
where one has installed Guix v1.2.0 and is running a substitute-less
bootstrap build, since the `guix time-machine` command itself has a
dependency on gnutls v3.6.12 (the one with the broken test) and will
thus try to build it before attempting to jump forwards in time. This
does however, mean that those who build a version of Guix that also
contains this fix will not go backwards in time to build the broken
gnutls v3.6.12.
a6a1b106dc guix: only download sources for hosts being built (fanquake)
Pull request description:
For example, if a user is only interested in building for Linux, this saves downloading the macOS compiler and additional dependencies, which is meaningful on a slow/poor connection. This will result in a few additional `make` invocations, for the Linux hosts, however this is low overhead, and time-wise irrelevant in terms of the overall build.
ACKs for top commit:
laanwj:
Code review ACK a6a1b106dc
Tree-SHA512: 34c916ae6f69fed0d5845690b39111a8bee37208fd727176f375cf5eb4860f512abe12bde2680d697c859b4d50a3bc5688ddca7c2f28f9968fcf358753cf3f6d
If a user is only interested in building for Linux, this saves
downloading the macOS compiler and additional dependencies.
This will result in a few additional `make` invocations, for the Linux
hosts, however this is quite low overhead.
Co-authored-by: Carl Dong <contact@carldong.me>
The new time-machine commit is Guix v1.2.0 with a yet-unupstreamed patch
for NSIS.
A few important changes:
1. Guix switched back from using CPATH to C{,PLUS}_INCLUDE_PATH as the
way to indicate #include search paths.
2. GCC's library is now split into a separate output, whereas before it
was included in the default output. This means that our gcc toolchain
packages need to propagate that output.
3. A few package versions were bumped
When building nsis, if VERSION is not specified, it defaults to
cvs_version which is non-deterministic as it includes the current date.
This patches nsis to default to SOURCE_DATE_EPOCH if it exists so that
nsis is reproducible.
Upstream change: https://github.com/kichik/nsis/pull/13
When using worktrees or submodules, you'll see a `.git' plain text file
at the root of your working tree instead of the usual `.git' directory.
This plain text file will point to the real GIT_DIR, under the
GIT_COMMON_DIR. From experimentation, the full GIT_COMMON_DIR is
required to exist for operations such as git-archive(1), so we expose it
as readonly inside the container.
Any -O argument will enable optimizations in GNU ld. We can use -O2
here, as this matches our compile flags. Note that this would also
enable additional optimizations if using the lld or gold linkers,
when compared to -O0.
The libtool unsorted 'find' determinism issue seemed to have been solved
in gcc-9's git: d41cd173e23ebea7c758644d6ad6e0fde1c2e3a6 or SVN: r262451
Furthermore, it seems that Ubuntu Focal 20.04 LTS is going to ship with
gcc 9 and mingw-w64 7, which will match what we have now.
-----
A note on this:
Careful observers will see that previously I stated that all released
versions of gcc were bootstrapped with a libtool 2.2.7a, meaning that
they all had the unsorted 'find' determinism issue first resolved in
libtool 2.2.7b.
However, I was mistaken, gcc's ltmain.sh CLAIMS it was generated by
libtool 2.2.7a, but it was in fact edited manually. It seems that gcc
maintains their own versions of ltmain.sh and libtool.m4, and only
sometimes backports patches from upstream.
Quite confusing.
This is no longer needed after 3bef7c22 in the mingw-w64 git repository,
which is first included in mingw-w64 v7.0.0.
As of the previous bump to our Guix time machine, we now use mingw-w64
v7.0.0.
Most of the mingw-w64 toolchain changes have now been upstreamed, we can
point to a commit that exists upstream.
NOTE: I'm not changing the URL yet until we see that Guix upstream will
accept all my patches for macOS.
-----
The Guix tree that's referred to by this commit contains the following
changes relevant to our mingw-w64 build:
b066c25026
Adds a PACKAGES-WITH-*PATCHES procedure which we can use in the future
to apply patches to packages if those patches are not considered
appropriate to upstream Guix
4719b71572
Adds mingw-w64 (the libc itself) reproducibility patches, taken from
debian.
79825bee07 + 401d28e433 + c1c50cb5b0
Add mingw-w64 specific binutils patches, taken from debian.
Specifically, the "Make DLL import libraries reproducible" patch made
libbitcoinconsensus.dll.a build reproducibly. The followup commits
were hotfixes for my mistakes.
0f864175dc
Bumps mingw-w64 to v7.0.0. This is the first release that enables
secure APIs by default (which we need), and gains _FORTIFY_SOURCE
support. This will also be what Ubuntu Focal 20.04 LTS releases with.
cdf00cf75d
Bumps NSIS to v3.05. This is the first release that includes a fix for
a reproducibility bug found by some of the electrum developers. See
details here: https://sourceforge.net/p/nsis/bugs/1230/
This bump will includes a couple of commits which improve the
reproducibility of the mingw-w64 toolchain. Most of which came from
debian. They will be upstreamed as upstream Guix release timeline
allows.
- Add "--no-insert-timestamp" LDFLAG for x86_64-w64-mingw32 builds
"The option --no-insert-timestamp can be used to insert a zero value for
the timestamp, this ensuring that binaries produced from identical
sources will compare identically." - ld(1)
- Set "SetDateSave off" in NSIS script
From https://nsis.sourceforge.io/Docs/Chapter4.html#flags
"This command sets the file date/time saving flag which is used by the
File command to determine whether or not to save the last write date and
time of the file, so that it can be restored on installation. Valid
flags are 'on' and 'off'. 'on' is the default."
- Add commented out NSIS options for reproducibility debugging in NSIS
script
- Make ZIPs deterministic by reseting file modification times to
SOURCE_DATE_EPOCH using touch(1) (Reference:
https://reproducible-builds.org/docs/archives/)
Previously, Guix would produce a gcc which did not know to use the SSP
function from glibc, and required a gcc make flag for it to do so, in my
attempt to fix it upstream I realized that this is no longer the case.
This can be verified by performing a Guix build and doing
readelf -s ... | grep __stack_chk
to check that symbols are coming from glibc, and doing
readelf -d ... | grep NEEDED | grep ssp
to see that libssp.so is not being depended on
- store_path() previously only worked for cross compilation packages, we
remove this assumption here
- Add CROSS_GCC_LIB variable which points to where gcc libs/headers are
located
- Add gcc libs/headers to our CROSS_*_PATH environment variables
- Check that all directories in CROSS_*_PATH are sane
- Clearer and more accurate prose
- Pin `guix pull' to commit rather than branch
- Just use `use-module' instead of `define-module'
- Use `bash-minimal' instead of `bash'
- Remove unneeded `tcsh' from manifest
- Explicitly use `python-3.7'
- Add comments about how {native,cross}-toolchains are produced and
why
