[gentoo-dev] [PATCH multibuild.eclass] Use portable locking code from Fabian Groffen.

[Michał Górny]

The 'userland_*' flags have proven not good enough to determine
the availability of lock helpers. Fabian provided a nice portable
locking code instead.

Fixes: https://bugs.gentoo.org/show_bug.cgi?id=466554
---
gx86/eclass/multibuild.eclass | 29 +++++++++++++++--------------
1 file changed, 15 insertions(+), 14 deletions(-)

diff --git a/gx86/eclass/multibuild.eclass b/gx86/eclass/multibuild.eclass
index acfdbbd..819c814 100644
--- a/gx86/eclass/multibuild.eclass
+++ b/gx86/eclass/multibuild.eclass
@@ -251,38 +251,39 @@ multibuild_merge_root() {
local src=${1}
local dest=${2}

- local lockfile=${T}/multibuild_merge_lock
+ local lockfile=${T}/.multibuild_merge_lock
+ local lockfile_l=${lockfile}.${$}
local ret

+ # Lock the install tree for merge. The touch+ln method ensures race
+ # condition-free locking with maximum portability.
+ touch "${lockfile_l}" || die
+ until ln "${lockfile_l}" "${lockfile}" &>/dev/null; do
+ sleep 1
+ done
+ rm "${lockfile_l}" || die
+
if use userland_BSD; then
- # Locking is done by 'lockf' which can wrap a command.
# 'cp -a -n' is broken:
# http://www.freebsd.org/cgi/query-pr.cgi?pr=174489
# using tar instead which is universal but terribly slow.

tar -C "${src}" -f - -c . \
- | lockf "${lockfile}" tar -x -f - -C "${dest}"
+ | tar -x -f - -C "${dest}"
[[ ${PIPESTATUS[*]} == '0 0' ]]
ret=${?}
elif use userland_GNU; then
- # GNU has 'flock' which can't wrap commands but can lock
- # a fd which is good enough for us.
- # and cp works with '-a -n'.
-
- local lock_fd
- redirect_alloc_fd lock_fd "${lockfile}" '>>'
- flock ${lock_fd}
+ # cp works with '-a -n'.

cp -a -l -n "${src}"/. "${dest}"/
ret=${?}
-
- # Close the lock file when we are done with it.
- # Prevents deadlock if we aren't in a subshell.
- eval "exec ${lock_fd}>&-"
else
die "Unsupported userland (${USERLAND}), please report."
fi

+ # Remove the lock.
+ rm "${lockfile}" || die
+
if [[ ${ret} -ne 0 ]]; then
die "${MULTIBUILD_VARIANT:-(unknown)}: merging image failed."
fi
--
1.8.2.1

[Mike Frysinger]

On Wednesday 15 May 2013 15:22:21 Michał Górny wrote:
> The 'userland_*' flags have proven not good enough to determine
> the availability of lock helpers. Fabian provided a nice portable
> locking code instead.
>
> Fixes: https://bugs.gentoo.org/show_bug.cgi?id=466554
> ---
> gx86/eclass/multibuild.eclass | 29 +++++++++++++++--------------

this seems more appropriate for portability.eclass than other random places.
there you can keep flock when it's available, or fall back to manual hacks like
this.
-mike

[Zac Medico]

It can be tricky find out whether or not the underlying filesystem will
behave well with flock. For example, maybe it will just fail temporarily
due to lock exhaustion (ENOLCK), or maybe the filesystem doesn't support
locking at all (ENOSYS).

Given all of the variables, it may be safest to use hardlink locks in
any case, since they work reliably on practically any relevant
filesystem. So if we do end up putting something in portability.eclass,
maybe hardlink locks are the way to go. Alternatively, we could add some
locking functions in the next EAPI, and let the package manager handle
the implementation details.
--
Thanks,
Zac