summaryrefslogtreecommitdiff
path: root/SECURITY.md
diff options
context:
space:
mode:
authorTom Scogland <scogland1@llnl.gov>2021-12-22 07:25:05 -0800
committerMassimiliano Culpo <massimiliano.culpo@gmail.com>2021-12-23 16:02:09 +0100
commit8e659f512e00534c6af427ef3832db68fcd2f2d3 (patch)
treed3324116f769b88ac8a2535a12237f83d9c8a635 /SECURITY.md
parent5daf023aecf3d72943cff0a010dfafd95edfdf3b (diff)
downloadspack-8e659f512e00534c6af427ef3832db68fcd2f2d3.tar.gz
spack-8e659f512e00534c6af427ef3832db68fcd2f2d3.tar.bz2
spack-8e659f512e00534c6af427ef3832db68fcd2f2d3.tar.xz
spack-8e659f512e00534c6af427ef3832db68fcd2f2d3.zip
locks: allow locks to work under high contention (#27846)
* locks: allow locks to work under high contention This is a bug found by Harshitha Menon. The `lock=None` line shouldn't be a release but should be ``` return (lock_type, None) ``` to inform the caller it couldn't get the lock type requested without disturbing the existing lock object in the database. There were also a couple of bugs due to taking write locks at the beginning without any checking or release, and not releasing read locks before requeueing. This version no longer gives me read upgrade to write errors, even running 200 instances on one box. * Change lock in check_deps_status to read, release if not installed, not sure why this was ever write, but read definitely is more appropriate here, and the read lock is only held out of the scope if the package is installed. * Release read lock before requeueing to reduce chance of livelock, the timeout that caused the original issue now happens in roughly 3 of 200 workers instead of 199 on average.
Diffstat (limited to 'SECURITY.md')
0 files changed, 0 insertions, 0 deletions