On Thursday 24 March 2011 07:41 PM, Kamesh Jayachandran wrote:
On 03/24/2011 05:54 PM, vijay wrote:
Just now I came to know that it fails in neon only.

I configured neon as a default ra_layer in my runtime configuration area.

When I use serf as ra_layer, the commit succeeds.

Yes that answers the failure I observed.

Please provide a testcase that exhibits the same failure for both neon and serf.

The testcase exhibits the following behaviour when tested with neon and serf.

RA_LAYER
        HTTP V2
        pre-HTTP V2
Neon
        Fail
        Fail
Serf
        Pass
        Fail


There are two options to make the testcase fail for both neon and serf.

1.Disabling "SVNAdvertiseV2Protocol" by making it explicitly off in the slave server's Location directive in Apache configuration.

2. Passing the command line option "--config-option servers:global:http-library=neon" to svnmucc for the particular commit operation as it fails in neon consistently.

I prefer to the second option(using neon for the particular commit) instead of disabling HTTP V2 in server side which may affect other testcases.

Meanwhile analysing why it succeeds in serf would teach something interesting.


First, let me tell why it fails in neon.

As in the test case, the slave repo is hosted in 127.0.0.1 and master repo is hosted in 127.0.0.2

1. For the particular commit which has space in its path, the CHECKOUT request from client reaches the slave repo.

CHECKOUT /repo/!svn/ver/5/branch%20new HTTP/1.1
User-Agent: SVN/1.7.0-dev (under development) neon/0.29.3

2. The request from the client is proxied by the slave to the master and the Master sends the response with Location header as follows.

Header Name: [location], Value: [http://127.0.0.2:30679/repo//!svn/wrk/69179a87-38cb-43c4-945c-de1e0b297aad/branch%20new]

3. The slave encodes the location header again(r917523) and forward it to the client.

Header Name: [location], Value: [http://127.0.0.1:26248/repo//!svn/wrk/5b604aca-a4b1-41fc-87ae-51299bf565b2/branch%2520new]

The above location header is used in subsequent dav request which fails as it has double-encoding of the space.

Here, PUT request fails as follows.

PUT /repo/!svn/wrk/5b604aca-a4b1-41fc-87ae-51299bf565b2/branch%2520new/file HTTP/1.1

svnmucc: E160013: File not found: transaction '5-5', path '/branch%20new/file'

But why didn't it fail in serf?

Serf has the implementations of HTTP V2 stuffs which does not use CHECKOUT request during commit.

It processes as follows.

OPTIONS -> POST -> PROPFIND -> PROPPATCH -> HEAD -> PUT -> MERGE -> DELETE

It gets the transaction id from POST request and directly put the contents there.

But I could see the same failure while committing with serf when I disable "SVNAdvertiseV2Protocol" , because it uses the CHECKOUT request there.

I think we can use neon for the particular commit operation instead of disabling HTTP V2 wholly in server side.

Please correct me if I am wrong.

Anyway we can enhance these tests further once we started to implement "make davmirrorautocheck" which I am going to take as my next activity.

Attaching the patch and log message.

Thanks & Regards,
Vijayaguru


With regards
Kamesh Jayachandran

Thanks & Regards,
vijayaguru

On Thursday 24 March 2011 05:12 PM, Kamesh Jayachandran wrote:
Thanks for the Patch Vijay.

In my FC14 your testcase passes both with and without patch.

I am investigating....

Will get back.

With regards
Kamesh Jayachandran
On 03/24/2011 04:11 PM, vijay wrote:
Hi,

This patch adds a testcase for the regression triggered by r917523 and fixes it.

The revision r917523 do some url-encodings to the paths and uris in subversion/mod_dav_svn/mirror.c.

<snip>
$svn log -r917523
------------------------------------------------------------------------ r917523 | kameshj | 2010-03-01 19:18:01 +0530 (Mon, 01 Mar 2010) | 26 lines

With the below apache configuration(See the <space> character "/svn 1/").

<Location "/svn 1/">
  DAV svn
  SVNParentPath /repositories
</Location>
<Location "/svn 2/">
  DAV svn
  SVNParentPath /repositories-slave
  SVNMasterURI "http://localhost/svn 1"
</Location>

Write through proxy is *not* happening and commit happens *directly* inside the slave.

* subversion/mod_dav_svn/mirror.c
  (proxy_request_fixup): URI encode the to be proxied file name.
(dav_svn__proxy_request_fixup): r->unparsed_uri is in url encoded form while root_dir is not in encoded form. So use r->uri to compare with root_dir. (dav_svn__location_in_filter): URL Encode the 'find & replace' urls as
    the request body has it in url encoded format.
(dav_svn__location_header_filter): Encode the master_uri as the response from master has the Location header url encoded already. Set the outgoing Location
    header url encoded.
(dav_svn__location_body_filter): URL Encode the 'find & replace' urls as
    the response body has it in url encoded format.

------------------------------------------------------------------------
</snip>

There is one extra url encoding in mirror.c:dav_svn__location_header_filter() which encodes the Location header response from master second time, which in turn causes failure while committing a path to slave repo which has space in it. The space in the path is encoded as "%20" first time, again it is encoded as "%2520", which fails with error "File not found" while committing.

I have added a testcase for this issue in subversion/tests/cmdline/dav-mirror-autocheck.sh. I have just used the existing master and slave repo setup for my new test case.

I have a plan of introducing "davmirrorautocheck" which we can execute like "make davmirrorautocheck", it will run all our python test suites via write-through proxy. All commits to the slave will be proxied to the master repo. The master repo's post commit hook will sync every commit to the slave repo. We can check whether slave and master repo are in sync in 'run_and_verify_commit'. We can extensively test our write-through proxy by making use of all our existing tests.

Attaching the patch and log message.

Thanks & Regards,
Vijayaguru










Fix the regression issue triggered by r917523. 
The revision r917523 do some url encodings to the paths and uris which are
not url-encoded. But there is one additional url-encoding of an uri which is
already encoded. With this extra encoding, committing a path to slave which has
space in it fails.

* subversion/tests/cmdline/dav-mirror-autocheck.sh
  Add a testcase for a regression issue triggered by r917523.

* subversion/mod_dav_svn/mirror.c
  (dav_svn__location_header_filter): Remove redundant url-encoding of new_uri.

Patch by: Vijayaguru G <vijay{_AT_}collab.net>                                     
Index: subversion/mod_dav_svn/mirror.c
===================================================================
--- subversion/mod_dav_svn/mirror.c     (revision 1087670)
+++ subversion/mod_dav_svn/mirror.c     (working copy)
@@ -241,7 +241,6 @@
                                                dav_svn__get_root_dir(r), "/",
                                                start_foo, (char *)NULL),
                                    r);
-        new_uri = svn_path_uri_encode(new_uri, r->pool);
         apr_table_set(r->headers_out, "Location", new_uri);
     }
     return ap_pass_brigade(f->next, bb);
Index: subversion/tests/cmdline/dav-mirror-autocheck.sh
===================================================================
--- subversion/tests/cmdline/dav-mirror-autocheck.sh    (revision 1087670)
+++ subversion/tests/cmdline/dav-mirror-autocheck.sh    (working copy)
@@ -416,10 +416,6 @@
 say "svn log on $BASE_URL : "
 $SVN --username jrandom --password rayjandom log -vq "$BASE_URL"
 
-# shut it down
-echo -n "${SCRIPT}: stopping httpd: "
-$HTTPD -f $HTTPD_CONFIG -k stop
-echo "."
 
 # verify result: should be at rev 4 in both repos
 # FIXME: do more rigorous verification here
@@ -438,5 +434,28 @@
 
 say "PASS: master, slave are both at r4, as expected"
 
+# The following test case is for the regression issue triggered by r917523.
+# The revision r917523 do some url encodings to the paths and uris which are
+# not url-encoded. But there is one additional url-encoding of an uri which is
+# already encoded. With this extra encoding, committing a path to slave which 
has
+# space in it fails. Please see this thread
+# http://svn.haxx.se/dev/archive-2011-03/0641.shtml for more info.
+
+say "Test case for regression issue triggered by r917523"
+
+$svnmucc cp 2 "$BASE_URL/trunk" "$BASE_URL/branch new"
+$svnmucc put /dev/null "$BASE_URL/branch new/file" --config-option 
servers:global:http-library=neon
+RETVAL=$?
+
+if [ $RETVAL -eq 0 ] ; then
+  say "PASS: committing a path which has space in it passes"
+else
+  say "FAIL: committing a path which has space in it fails as there are extra
+  url-encodings happening in server side"
+fi
+
+# shut it down
+echo -n "${SCRIPT}: stopping httpd: "
+$HTTPD -f $HTTPD_CONFIG -k stop
+echo "."
 exit 0
-

Reply via email to