error: git-svn died of signal 13

error: git-svn died of signal 13

[Jonathan Nieder]

Hi,

Just wanted to pass on a few reports about git-svn dying with SIGPIPE
when trying to fetch large repositories.  It seems to fetch a few
revisions, then die; the repositories can eventually be fetched
successfully by restarting 'git svn fetch' again and again.

Some bug reporters have been very patient, but I don’t really know
where to look or ask in solving this.  Attempts at tracing so far have
given either too little or too much information to make any sense.

The problem goes back at least to v1.5.6.5.

At the end of the git-svn process’s life, the last few function calls
are these:

 - SVN::Ra::Reporter::finish_report and related functions
 - _p_apr_pool_t::DESTROY (libsvn-perl cleaning up, presumably)
 - SVN::Ra::Callbacks::open_tmp_file, called from
   SVN::_Ra::svn_ra_do_update, called from
   Git::SVN::Ra::do_update, called from
   Git::SVN::Ra::gs_do_update
 - Git::SVN::_set_repos_root
 - starting a 'git update-index -z --index-info' to feed commands to

Maybe /tmp is filling up?  It is hard to know, since the SIGPIPE stops
everything and we do not get any nice, simple error message describing
what read or write would have received EPIPE.

See <http://bugs.debian.org/526989>, <http://bugs.debian.org/540150>,
<http://thread.gmane.org/gmane.comp.version-control.git/85868/>.

Any insights?

Jonathan

Re: error: git-svn died of signal 13

[Eric Wong]

Jonathan Nieder <jrnieder@gmail.com> wrote:
> Hi,
> 
> Just wanted to pass on a few reports about git-svn dying with SIGPIPE
> when trying to fetch large repositories.  It seems to fetch a few
> revisions, then die; the repositories can eventually be fetched
> successfully by restarting 'git svn fetch' again and again.
> 
> Some bug reporters have been very patient, but I don’t really know
> where to look or ask in solving this.  Attempts at tracing so far have
> given either too little or too much information to make any sense.
> 
> The problem goes back at least to v1.5.6.5.

Hi Jonathan,

Thanks for taking the time to look into this.  This has been a
problem for a long time and very difficult to track down.

One factor is that git svn continues happily afterwards so it's hard to
reproduce and there's less incentive to look for a fix.

> At the end of the git-svn process’s life, the last few function calls
> are these:
> 
>  - SVN::Ra::Reporter::finish_report and related functions
>  - _p_apr_pool_t::DESTROY (libsvn-perl cleaning up, presumably)
>  - SVN::Ra::Callbacks::open_tmp_file, called from
>    SVN::_Ra::svn_ra_do_update, called from
>    Git::SVN::Ra::do_update, called from
>    Git::SVN::Ra::gs_do_update
>  - Git::SVN::_set_repos_root
>  - starting a 'git update-index -z --index-info' to feed commands to
> 
> Maybe /tmp is filling up?  It is hard to know, since the SIGPIPE stops
> everything and we do not get any nice, simple error message describing
> what read or write would have received EPIPE.
> 
> See <http://bugs.debian.org/526989>, <http://bugs.debian.org/540150>,
> <http://thread.gmane.org/gmane.comp.version-control.git/85868/>.
> 
> Any insights?

Can somebody get an strace or core dump for the "git update-index"
call?  Or even an strace -f of the entire git svn invocation as well...

-- 
Eric Wong

Re: error: git-svn died of signal 13

[Jonathan Nieder]

CCing those who reported the problem.  There is some debugging advice below,
for anyone who wants to take a further stab at tracking this down.

Eric Wong wrote:
> Jonathan Nieder <jrnieder@gmail.com> wrote:

>> Just wanted to pass on a few reports about git-svn dying with SIGPIPE
>> when trying to fetch large repositories.  It seems to fetch a few
>> revisions, then die; the repositories can eventually be fetched
>> successfully by restarting 'git svn fetch' again and again.
>> 
>> Some bug reporters have been very patient, but I don’t really know
>> where to look or ask in solving this.  Attempts at tracing so far have
>> given either too little or too much information to make any sense.
>> 
>> The problem goes back at least to v1.5.6.5.
>
> Hi Jonathan,
>
> Thanks for taking the time to look into this.  This has been a
> problem for a long time and very difficult to track down.
>
> One factor is that git svn continues happily afterwards so it's hard to
> reproduce and there's less incentive to look for a fix.
>
>> At the end of the git-svn process’s life, the last few function calls
>> are these:
>> 
>>  - SVN::Ra::Reporter::finish_report and related functions
>>  - _p_apr_pool_t::DESTROY (libsvn-perl cleaning up, presumably)
>>  - SVN::Ra::Callbacks::open_tmp_file, called from
>>    SVN::_Ra::svn_ra_do_update, called from
>>    Git::SVN::Ra::do_update, called from
>>    Git::SVN::Ra::gs_do_update
>>  - Git::SVN::_set_repos_root
>>  - starting a 'git update-index -z --index-info' to feed commands to
>> 
>> Maybe /tmp is filling up?  It is hard to know, since the SIGPIPE stops
>> everything and we do not get any nice, simple error message describing
>> what read or write would have received EPIPE.
>> 
>> See <http://bugs.debian.org/526989>, <http://bugs.debian.org/540150>,
>> <http://thread.gmane.org/gmane.comp.version-control.git/85868/>.
>> 
>> Any insights?
>
> Can somebody get an strace or core dump for the "git update-index"
> call?  Or even an strace -f of the entire git svn invocation as well...

Re: error: git-svn died of signal 13

[Jason Felice]

Jonathan Nieder <jrnieder <at> gmail.com> writes:

> 
> CCing those who reported the problem.  There is some debugging advice below,
> for anyone who wants to take a further stab at tracking this down.
> 
> Eric Wong wrote:
> > Can somebody get an strace or core dump for the "git update-index"
> > call?  Or even an strace -f of the entire git svn invocation as well...
> 

I am having the same issue, pulling a HTTPS repo being served with Subversion
1.4.5.  Other info: The repo has many, many tags (~300?), and ~5000 revisions,
and I've been pulling it since Saturday!  Each iteration of git svn fetch takes
longer and longer before producing output.  It "feels" like  It The tags seem to
be what's slowing it down.)  I can't make this repo or shell access available
(sorry).

I got an strace:

jfelice@lx-cclamp:~/xpay-core$ strace -f -o /tmp/strace git svn fetch
r3176 = 816a8e7ffd28a2224a75037cf6a983883532c76e (refs/remotes/tags/v01_01_87)
error: git-svn died of signal 13

The strace output is at:

http://eraserhead.net/files/strace.bz2

I originally suspected that our crufty HTTPS server was closing the connection
on us due to some kind of timeout.