[Monotone-devel] cvssync invariant failure

[William Uther]

Hi,

  I'm not sure if this is interesting to everyone.  It's mostly for  
Christof Petig I expect.  BTW - there is no rush on this any more.   
Everyone has moved back to CVS for the moment until I get cvs-sync  
running stably for a while.  So far my record between glitches is 2  
whole days :).

  I'm using the cvssync.refactor branch to sync to a medium sized  
CVS repository.  Saturday morning was the latest hiccup.  One of the  
cvs users committed a file and then committed it again saying  
something about kde locks.  CVS itself has no problems checking out  
or updating.  But, when I try cvs pull with mtn_cvs, I get the  
"Inconsistency" message on line 450 of mtn_cvs/cvs_sync.cc and the  
invariant on the next line fails.  I'm trying to fix things, am a  
little out of my depth, and was wondering if anyone has any ideas...

  Interestingly, I can pull a fresh mtn repository from cvs  
repository.  It is only when I try to update my older mtn repository  
that I get the problem.  I tried various hacks like pulling a new  
repos and then syncing that with the old one to try to preserve  
history.  I didn't think it would work well, and it didn't.  There is  
something in the old repos that when synced with the freshly-pulled  
repository causes it to stop updating with the same problem.

  So I've been debugging the underlying problem, but I don't really  
understand this yet.  I feel like a medieval doctor applying leeches.

  The s->cvs_version string on line 450 of mtn_cvs/cvs_sync.cc is  
empty, which is weird.  That function is called from cvs_sync.cc: 
1523, and that function has a "#warning FIXME: changed files" at the  
top.  Is this going to be related?

  If I add a check, if (s->cvs_version.empty()) W(F("s- 
>cvs_version.empty()")); at line 1506 then it fires a while before  
the "Inconsistency".  I don't yet understand why that version is  
empty, and chasing that back for a while didn't seem to lead me  
anywhere.

  I then tried simply removing the invariant that was failing  
(drains a little more blood from the patient).  This continued and  
completed, but at the end dumped out this:

mtn_cvs: warning: blocking attempt to certify an empty CVS revision  
of 'CMakeLists.txt'
mtn_cvs: warning: (this is normal for a cvs_takeover of a locally  
modified tree)
mtn_cvs: warning: CMakeLists.txt   
f47402df16421a331af4c7f5535a175f4b200ef5
mtn_cvs: warning: DartConfig.cmake   
5ba1b26cc16d6042ddfd5e8e9ee53ede532ca687
mtn_cvs: warning: README  bef4771f00b278b8a5fb5a0c3ce4140fea21890c
[snip ~10 more lines of stuff]

This doesn't fill me with joy.  If I update again then I get the "s- 
>cvs_version.empty()" warning, but not the "Inconsistency" message.   
I also then get "Modules do not match: 'dgc-orca' != ''" which I  
didn't get before.

I have copies of the CVS repos, the mtn database that cannot up  
without breaking (I have it before the update, so I can re-run the  
update easily and reproduce the problem).  I'm not sure about the  
license on all the files in the repo - so it would be best if I don't  
send it to anyone right now, but I can chase that down if people  
really want to try.

Be well,

Will          :-}