[Varnish] #1500: Possible bug in 4.0.0 with Range+do_stream=true

Varnish varnish-bugs at varnish-cache.org
Sat May 10 18:40:25 CEST 2014


#1500: Possible bug in 4.0.0 with Range+do_stream=true
-------------------------+--------------------
 Reporter:  moseleymark  |       Type:  defect
   Status:  new          |   Priority:  normal
Milestone:               |  Component:  build
  Version:  4.0.0        |   Severity:  normal
 Keywords:               |
-------------------------+--------------------
 I noticed in testing Range requests (to an Apache 2.2 backend, Ubuntu
 Precise 64-bit) that in sufficiently large requests, Range requests on
 *uncached* content were often returning 200s with the full content. The
 larger the request, the more likely it was that this would happen. See
 https://www.varnish-cache.org/comment/1654#comment-1654 for details on
 some testing I did.

 In the VCL below, if I uncomment the do_stream=false, I seem to get 206s
 on uncached large files every time. I need to test this further, but so
 far, with it commented out (since I guess do_stream is true by default
 now), on a 10meg file, I was getting 200s on uncached content almost 100%
 of the time; with it uncommented, I haven't seen a 200 yet.

 Dropping in some debugging statements, it looks like that for streaming
 large-ish files, this 'if' statement in v1d_dorange
 (bin/varnishd/cache/cache_http1_deliver.c, about line 105 or so) fails:

 if (low >= req->obj->len)
             return;

 because req->obj->len is 0. I'm guessing that with a big enough file,
 streaming causes v1d_dorange to get called before len gets set to non-
 zero, since it's still downloading the file from the backend (which
 considering how streaming works, kind of makes sense).

 Without streaming (this is my debug statement, via varnishlog):

 -   Debug          "IN v1d_dorange 3: low=1, len=10240000"

 With streaming:

 -   Debug          "IN v1d_dorange 3: low=1, len=0"


 I'm still trying to trace why req->obj->len would be 0 in this case. I'm
 not a C guy, so I couldn't even begin to offer a patch that wouldn't make
 your eyes bleed :)


 Testcase is just using a very stripped down VCL:

 {{{
 vcl 4.0;
 import std;

 backend test_http {
         .host = "127.0.0.1";
         .port = "80";
 }

 sub vcl_backend_response {
         # Commented, I get 200s; uncommented, I get 206s
         # set beresp.do_stream = false;
         return( deliver );
 }

 sub vcl_recv {
         # Set backend dynamically
         set req.backend_hint = test_http;

         # Default action is to look up object in cache
         return( hash );
 }
 }}}

-- 
Ticket URL: <https://www.varnish-cache.org/trac/ticket/1500>
Varnish <https://varnish-cache.org/>
The Varnish HTTP Accelerator



More information about the varnish-bugs mailing list