git 2.30.2-2: "git push" does not work

[lerdmann]

I recently installed git 2.30.2-2 via the RPMs. When I now try to push, I get this

[c:\libusb\git]git push
warning: Umleitung nach https://github.com/bitwiseworks/libusb1-os2/
fatal: Protokollfehler: ungültiges Zeichen für Zeilenlänge: ^5fc
fatal: ungültiger Commit '01c48e99ee093e1050de97b54a6ebfde5b15c'
fatal: Die Gegenseite hat unerwartet abgebrochen.
fatal: Die Gegenseite hat unerwartet abgebrochen.
error: Fehler beim Versenden einiger Referenzen nach 'http://github.com/bitwiseworks/libusb1-os2'

This used to work without any problem in the past.

I also am not able to downlevel "git" because I would need to downlevel 4 packages: "git","git-core","git-core-doc","git-p4" but these packages depend on each other and I neither can select all 4 of them for downlevelling nor can I downlevel them individually.

[SilvanScherrer]

Unfortunately its a known bug i need to search once.

[lerdmann]

How do I downlevel (to a specific version) individual packages from the commandline ?

[lerdmann]

Deinstalled all, then reinstalled 2.30.2-1 via RPMs. Led to same error. Deinstalled all again, then reinstalled 2.11.0-4. Works again.

[SilvanScherrer]

this bug can be worked around for now when redirecting the stderr output. with that it doesn't fail for me.

[altsan]

I have been experiencing the same issue.

I've found that setting GIT_TRACE=1 before running git push seems to make the problem go away.

[dryeo]

For me, redirecting stderr by itself never worked. At first redirecting both stdout and stderr did work. Now they don't. GIT_TRACE=1 does work

[SilvanScherrer]

as its a pure timing issue and GIT_TRACE slows it a bit down it helps. one day we might find the issue. who knows

[dmik]

@altsan re GIT_TRACE=1, it is just another proof that it's a timing issue… As it eventually works just by retrying several (or many) times without any additional env vars or options.

Another interesting observation is that with my new super fast PC (AMD Ryzen 7) I get these push problems much less often (but I still do get them once in a while). Another evidence it's timing. It's not clear where and there are lots of places in the transport layer where it may go wrong (especially in HTTPS mode which is my case at least).

I haven't tested the fresh git 2.50.1 a lot though. We will see if it changes anything.

[dmik]

I've been actively using the new git 2.50.1 for some time now (commit/push locally) and couldn't get a single push failure so far (even though git 2.30.2 gave me one or two). Either they redid HTTP(S) transport in this new git or I'm just lucky. I will keep testing.

Just as a side note, I really feel it has something to do with the TCP stack. Even something like the MTU size. And in old git it just somehow triggered races. If the new one is still the same (I didn't dig deep), then I'm sure I will get this failure again really soon.

[dmik]

With the latest git 2.50.1, the problem is unfortunately still there (but very rare here). Attaching two logs:

git-push.fail.log

git-push.ok.log

The first one - is the failed attempt, the second one - is the succeeded attempt immediately after. Needs some deep investigation.