Ticket #250 (POST error when using Opera)

[Nick Loman]

Hi there

Just to update you on my experience with Varnish 1.1.2, I had exactly 
the same problem as reported by the user on this ticket:

  http://varnish.projects.linpro.no/ticket/250

In my case, the cause of the problem after some analysis of the log 
files was that the "Connection-Length" header was being removed, despite 
my VCL actually only amending headers, not deleting them. Hence the 
backend script could not process the POST request properly.

The application of the 3110 patch fixed the problem.

  http://varnish.projects.linpro.no/changeset/3110

However, as the patch supplied is for Varnish 2.0, I needed to adapt it 
to the 1.1.2 source code, as follows:

--- varnish-1.1.2/bin/varnishd/cache_http.c.orig        2008-10-11 
13:22:23.0000                                          00000 +0100
+++ varnish-1.1.2/bin/varnishd/cache_http.c     2008-10-11 
13:25:26.000000000 +0                                          100
@@ -767,6 +767,7 @@
         if (to->nhd < HTTP_HDR_MAX) {
                 to->hd[to->nhd].b = fm->hd[n].b;
                 to->hd[to->nhd].e = fm->hd[n].e;
+               to->hdf[to->nhd] = 0;
                 to->nhd++;
         } else  {
                 VSL_stats->losthdr++;
@@ -913,10 +914,12 @@
                 WSL(w, SLT_LostHeader, fd, "%s", string);
                 to->hd[field].b = NULL;
                 to->hd[field].e = NULL;
+               to->hdf[field] = 0;
         } else {
                 memcpy(p, string, l + 1);
                 to->hd[field].b = p;
                 to->hd[field].e = p + l;
+               to->hdf[field] = 0;
         }
  }

@@ -962,6 +965,7 @@
         } else {
                 to->hd[to->nhd].b = to->ws->f;
                 to->hd[to->nhd].e = to->ws->f + n;
+               to->hdf[to->nhd] = 0;
                 WS_Release(to->ws, n + 1);
                 to->nhd++;
         }
@@ -976,8 +980,10 @@
         for (v = u = HTTP_HDR_FIRST; u < hp->nhd; u++) {
                 if (http_IsHdr(&hp->hd[u], hdr))
                         continue;
-               if (v != u)
+               if (v != u) {
                         memcpy(&hp->hd[v], &hp->hd[u], sizeof hp->hd[v]);
+                       memcpy(&hp->hdf[v], &hp->hdf[u], sizeof hp->hdf[v]);
+               }
                 v++;
         }
         hp->nhd = v;

Hope this helps anyone else struggling with strange issues on Varnish 
1.1.2. I think the key point is that simply amending client headers in 
vcl_recv can also trigger missing headers being sent to the backend.

Regards,

Nick.