IO::Socket::SSL and HTTP::Daemon's select-sysread loop don't work well together [rt.cpan.org #52602]

Migrated from rt.cpan.org#52602 (status was 'open')

Requestors:

• mark@aufflick.com

From mark@aufflick.com on 2009-12-09 11:12:49
:

The below bug raised for HTTP::Daemon::SSL is actually a bug in the 
base HTTP::Daemon. You can rewrite the test server in the Perlmonks 
node using only HTTP::Daemon and the result is the same.


HTTP::Daemon::SSL hangs on largish (e.g. >37k or >67k) POST request.

Quoting zwon on perlmonks.org:

"It looks like select-sysread loop in HTTP::Daemon doesn't work 
correctly with IO::Socket::SSL. That's because select in _need_more 
tests real filehandle and sysread reads from IO::Socket::SSL object 
which is buffered, so

sysread($self, $_[0], 2048, length($_[0]))

may actually read more than 2048 bytes from the socket and 
subsequent 
select on socket will hang.

Removing Timeout [which avoids the select() altogether] as proposed by 
derby solves the problem."

Complete description in: http://www.perlmonks.org/?node_id=761270

From mark@aufflick.com on 2009-12-09 11:16:31
:

sorry - my mistake :(

From mark@aufflick.com on 2009-12-09 11:18:56
:

ok, i shouldn't be raising RT tickets this late.

so: the test case does work fine with just HTTP::Daemon, but it *is* an 
issue in the underlying HTTP::Daemon and IO::Socket::SSL which prevents 
HTTP::Daemon::SSL from working with large posts.

From mark@aufflick.com on 2009-12-09 12:02:17
:

FYI - I'm testing re-introducing the _need_more method which was in 
HTTP::Daemon::SSL 1.03 to work around this. This has other issues (not 
least of which is breaking abstraction).

http://github.com/aufflick/p5-http-daemon-
ssl/commit/021ac9373855e7b99a1fd5b44c14537c950a0dd6

On Wed Dec 09 06:18:56 2009, AUFFLICK wrote:
> ok, i shouldn't be raising RT tickets this late.
> 
> so: the test case does work fine with just HTTP::Daemon, but it *is* an 
> issue in the underlying HTTP::Daemon and IO::Socket::SSL which 
prevents 
> HTTP::Daemon::SSL from working with large posts.

Start by reproducing the large POST hang described for HTTP::Daemon with IO::Socket::SSL, then inspect HTTP::Daemon's select-sysread loop and the _need_more behavior mentioned in the report. Add a regression test for a large POST and confirm the server completes the request without hanging.