This commit is contained in:
Joey Hess 2020-10-05 15:19:48 -04:00
parent 471bcfaf37
commit 2daa28c05f
No known key found for this signature in database
GPG key ID: DB12DB0FF05F8F38

View file

@ -0,0 +1,34 @@
[[!comment format=mdwn
username="joey"
subject="""comment 1"""
date="2020-10-05T18:54:30Z"
content="""
That process tree seems to show git-annex with a bunch of child
git-annex processes. (Assuming that the first field is the pid, and
so those are not threads..)
But git-annex rarely exec's git-annex, and when it does, it's never with
the same params as the parent git-annex was run with. And, addurl in this
situation normally has some other child processes like `git cat-file
--batch` that this one is missing.
So it kind of looks like a fork/exec stalled before the exec somehow, I
don't know how. strace of a child process in this situation may shed some
light on this strange state.
If a fork/exec somehow stalled, it might explain why the jlog file
is incompletely written, since the haskell IO manager might have gotten
locked up.
It seems at least plasible that the race you then hit is somehow related to
it getting into this state. When I fixed that race, I noticed and fixed a
condition where runshell would run git-annex with the locales not
generated, which would lead to running with the system locales. Which, if
there were an incompatibility between them and the bundled libc, might
cause some kind of weird behavior.
Or, it could be something like
<https://gitlab.haskell.org/ghc/ghc/-/issues/4074>
But that was fixed in 2010, so unless this git-annex was built with an
old haskell library, unlikely that precise bug.
"""]]