On Tue, Jun 12, 2012 at 10:51 AM, Justin Erenkrantz <jus...@erenkrantz.com> wrote: > It seems we've tweaked how we handle EAGAIN inside of ra_serf and are > treating it as a critical error. > > Right now, what Johan is seeing with his anti-virus software is that > we very often receive EAGAIN from the sockets. However, we're > treating EAGAIN as a fatal error inside of > svn_ra_serf__context_run_wait(). I believe we ought to be handling > EAGAIN specifically and just (re)run the serf context loop again. >
That seems strange, serf_context_run is not documented to return APR_EGAIN. While the response handler can return APR_EAGAIN, serf itself will translate this to APR_SUCCESS. As a result, ra_serf doesn't expect it and - rightfully IMO - treats it as an error. Are you trying to solve this issue: http://subversion.tigris.org/issues/show_bug.cgi?id=4175 ? If so, the solution - as discussed in http://svn.haxx.se/dev/archive-2012-05/0133.shtml - lies in getting the response handler not labeling the request as done until the complete response has been read, including retrying on APR_EAGAIN. So I don't think this patch is a good idea. If the issue you're seeing is not the above, I'm interested in knowing some more details. > The patch below fixes Johan's situation...but, I'd appreciate some > eyes on it before committing. -- justin > > Index: subversion/libsvn_ra_serf/util.c > =================================================================== > --- subversion/libsvn_ra_serf/util.c (revision 1349176) > +++ subversion/libsvn_ra_serf/util.c (working copy) > @@ -734,6 +734,13 @@ svn_ra_serf__context_run_wait(svn_boolean_t *done, > _("Connection timed out")); > } > > + if (err && APR_STATUS_IS_EAGAIN(err->apr_err)) > + { > + svn_error_clear(err); > + err = SVN_NO_ERROR; > + status = APR_SUCCESS; > + } > + > SVN_ERR(err); > if (status) > { Lieven
