contrib: don't use the default datadir in gen-bitcoin-conf.sh

[kwsantiago]

Fixes #304.

gen-bitcoin-conf.sh invokes bitcoind --version and bitcoind --help. Both of these run common::InitConfig() (src/bitcoind.cpp:122) before the -help/-version early return (src/bitcoind.cpp:138), despite the comment there claiming help and version are processed "before taking care about datadir". So each invocation touches the default datadir:

• fs::create_directories(base_path / "wallets") (src/common/init.cpp:58) creates $HOME/.bitcoin
• fs::exists(base_path / BITCOIN_CONF_FILENAME) (src/common/init.cpp:68) stats $HOME/.bitcoin/bitcoin.conf
• WriteSettingsFile() (src/common/init.cpp:111) writes $HOME/.bitcoin/settings.json

If $HOME/.bitcoin exists but is not traversable, the fs::exists call throws filesystem_error on EACCES and both invocations fail, as @pdath reproduced in the issue.

Point bitcoind at a throwaway datadir instead. Note that -noconf alone is not sufficient: the fs::exists() call above runs regardless of whether a config file is in use, and -nosettings does not help either. Only -datadir avoids all three accesses. Verified by running each variant against a build of this tree.

Also add set -eo pipefail, per the second half of the issue. Previously a failing bitcoind was not detected at all: the VERSION_OUTPUT=$(...) assignment went unchecked, and the exit status of the bitcoind --help | sed | ... | sed pipeline was that of the trailing sed. The script printed bitcoind's errors and then exited 0 with a truncated bitcoin.conf, which is the failure mode seen in the issue report.

Bare mktemp -d matches the existing usage in contrib/devtools/check-deps.sh and is portable to the BSDs, whose mktemp behaves "as if -t tmp was supplied" when only -d is passed.

Testing

Built with -DWITH_ZMQ=ON -DWITH_MINIUPNPC=ON and confirmed the patched script regenerates share/examples/bitcoin.conf byte-for-byte identically to the committed file, so this change is output-neutral. bitcoind --help output is unchanged by -datadir.

Comparing against the script before this change:

Scenario	Before	After
Fresh $HOME	creates .bitcoin/, wallets/, settings.json	nothing created
Unwritable $HOME	fails	succeeds
chmod 444 ~/.bitcoin (issue #304)	Permission denied, but exits 0	succeeds, silent
bitcoind --version fails	exits 0	propagates exit status
bitcoind --help fails	exits 0	propagates exit status
bitcoind missing / not executable	exits 1	exits 1

The temporary datadir is removed on exit. test/lint/lint-shell.py passes.

One pre-existing wart is left alone: if bitcoind fails midway, the header has already been written and share/examples/bitcoin.conf is left truncated. That is now at least signalled by a non-zero exit status. Generating into a temporary file and renaming it into place on success would close it, if wanted.

[pdath]

ACK 49db8d1.

I reviewed the changes to the bash script and found them well written and robust.

I tested using the following commands on Ubuntu 24.04 on an ARM CPU:
git fetch origin pull/329/head:pr329
git switch pr329
chmod 444 ~/.bitcoin
contrib/devtools/gen-bitcoin-conf.sh

And then verified that the following file was created and contained the expected content:
./share/examples/bitcoin.conf

[luke-jr]

despite the comment there claiming help and version are processed "before taking care about datadir".

Shouldn't we fix this?

[kwsantiago]

Yes. Opened #341, which moves the common::InitConfig() call after ProcessInitCommands() so the comment becomes true. That matches the ordering bitcoin-cli already uses.

With that in, the -datadir here is redundant, since --help and --version no longer touch the datadir at all (verified against a build of #341: both exit 0 with chmod 444 ~/.bitcoin). The set -eo pipefail half of this PR is still worth having either way, so I am happy to trim this to just that if you would rather fix it only in bitcoind.

[kwsantiago]

Rebased onto v29.0 as a daggy fix and force-pushed.