summaryrefslogtreecommitdiff
path: root/doc/todo
diff options
context:
space:
mode:
authorJoey Hess2017-06-18 18:24:05 -0400
committerJoey Hess2017-06-18 18:24:05 -0400
commit01fc1375cece096ab2dec480b843ecdbc4f0d94e (patch)
treefb620f98f3e891eac092d31ed5ab43a8975bd3a0 /doc/todo
parent9b457ae66d5c5143f5896156cc4af8d5a0bc0ddc (diff)
Fix bug that sometimes made --spin fail with "fatal: Couldn't find remote ref HEAD"
Tricky stdin buffering problem. An easier fix would have been: hSetBuffering stdin NoBuffering But that approach is less robust; even with NoBuffering, anything that uses hLookAhead causes 1 byte of buffering. And, any reads from stdin before hSetBuffering would still cause the problem. Instead, I used a bigger hammer that will always work. It involves a bit more CPU work, but this is data that is already being fed through ssh; copying it one more time won't cause a measurable performance impact. This commit was sponsored by Jack Hill on Patreon.
Diffstat (limited to 'doc/todo')
-rw-r--r--doc/todo/spin_failure_HEAD.mdwn33
1 files changed, 30 insertions, 3 deletions
diff --git a/doc/todo/spin_failure_HEAD.mdwn b/doc/todo/spin_failure_HEAD.mdwn
index af525f61..f838e469 100644
--- a/doc/todo/spin_failure_HEAD.mdwn
+++ b/doc/todo/spin_failure_HEAD.mdwn
@@ -51,7 +51,6 @@ Sending privdata (73139 bytes) to kite.kitenet.net ... done
[2017-06-18 16:27:13 EDT] received marked GITPUSH
[2017-06-18 16:27:13 EDT] command line: GitPush 11 12
16:27:13.953638 pkt-line.c:80 packet: fetch< 3a3c8a731d169a2768dd243581803dcb7b275049 HEAD\0multi_ack thin-pack side-band side-band-64k ofs-delta shallow deepen-since deepen-not deepen-relative no-progress include-tag multi_ack_detailed symref=HEAD:refs/heads/joeyconfig agent=git/2.11.0
-16:27:13.953638 pkt-line.c:80 packet: fetch< 3a3c8a731d169a2768dd243581803dcb7b275049 HEAD\0multi_ack thin-pack side-band side-band-64k ofs-delta shallow deepen-since deepen-not deepen-relative no-progress include-tag multi_ack_detailed symref=HEAD:refs/heads/joeyconfig agent=git/2.11.0
16:27:13.953781 pkt-line.c:80 packet: fetch< 86b077b7a21efd5484dfaeee3c31fc5f3c151f6c refs/heads/confpairs
16:27:13.953789 pkt-line.c:80 packet: fetch< e03e4bf0f1e557f87d1fe7e01a6de7866296fce6 refs/heads/d-i
16:27:13.953795 pkt-line.c:80 packet: fetch< 3a3c8a731d169a2768dd243581803dcb7b275049 refs/heads/joeyconfig
@@ -94,7 +93,35 @@ Sending privdata (73139 bytes) to kite.kitenet.net ... done
> > * Could be in gitPushHelper, perhaps it's failing to write
> > some of the first lines somehow?
> > * Could be something on the remote side is consuming stdin
-> > that is not supposed to, and eats some of the protocol.a
+> > that is not supposed to, and eats some of the protocol.
+> >
> >
> > I added debug dumping to gitPushHelper, and it seems to be
-> > reading the same truncated data.
+> > reading the same truncated data, so it seems the problem is not there.
+> >
+> > Aha! The problem comes from stdin/stdInput confusion here:
+
+ req NeedGitPush gitPushMarker $ \_ -> do
+ hin <- dup stdInput
+ hout <- dup stdOutput
+ hClose stdin
+ hClose stdout
+
+> > A line read from stdin just before the dup gets the first line of the protocol
+> > as expected. But reading from stdInput starts with a later line.
+> > Apparently data is being buffered in the stdin Handle, so gitPushHelper,
+> > which reads from the Fd, does not see it.
+> >
+> > Here's a simple test case. Feeding this 2 lines on stdin will
+> > print the first and then fail with "hGetLine: end of file".
+> > The second line is lost in the buffer. This test case behaves
+> > like that reliably, so I'm surprised propellor only fails sometimes.
+
+ main = do
+ l <- hGetLine stdin
+ print l
+ bob <- fdToHandle stdInput
+ l2 <- hGetLine bob
+ print l2
+
+> > [[fixed|done]] --[[Joey]]