Thanks Vijay for the detailed summary and the fix.
I committed it at r1088602.
With regards
Kamesh Jayachandran
On 04/04/2011 12:00 PM, vijay wrote:
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
