Re: Bug report: Git pull hang occasionally

Kai Zhang <kai@netskope.com> writes:

(Not a solution)

In order to tell the client if HEAD is a symbolic ref and to what underlying ref it points at if it is a symbolic ref, at the very beginning of upload-pack, there is a call to head_ref_namespaced() that uses find_symref(). find_symref() gets "HEAD" and a boolean that says if it is a symbolic ref, but it does not get where the symbolic ref points at, so it does resolve_ref_unsafe() to learn that information.

Between the time head_ref_namespaced() checks the refs database and finds that HEAD is a symbolic ref, and the time find_symref() calls resolve_ref_unsafe() to find out where it leads to, if somebody else updates HEAD, resolve_ref_unsafe() can give an unexpected result, as all of these read-only operations are performed without any locking.

And the unexpected discrepancy is reported by find_symref() as fatal. The server side dies, and somehow that fact is lost between the upload-pack process and the client and somebody in the middle (e.g. fastcgi interface or nginx webserver on the server side, or the remote-curl helper on the client side) keeps the "git fetch" process waiting.

So there seem to be two issues.  

 - Because of the unlocked read, find_symref() can observe an
   inconsistent state.  Perhaps it should be updated not to die but
   to retry, expecting that transient inconsistency will go away.

 - A fatal error in upload-pack is not reported back to the client
   to cause it exit is an obvious one, and even if we find a way to
   make this fatal error in find_symref() not to trigger, fatal
   errors in other places in the code can trigger the same symptom.